Statistics

Chap ma n & Ha ll/ C RC Chapm an & Hall/C R C

In terdi sc i p l i n a r y St a t is t ics Series I nterdis c iplinar y S tatis tic s S eries

MENDELIAN RANDOMIZATION
MENDELIAN
Mendelian Randomization: Methods for Using Genetic Variants in Causal Es-
timation provides thorough coverage of the methods and practical elements of
Mendelian randomization analysis. It brings together diverse aspects of Mendelian
randomization spanning epidemiology, statistics, genetics, and econometrics.

RANDOMIZATION
Through several examples, the first part of the book shows how to perform simple
applied Mendelian randomization analyses and interpret their results. The second
part addresses specific methodological issues, such as weak instruments, multiple
instruments, power calculations, and meta-analysis, relevant to practical applica-
tions of Mendelian randomization. In this part, the authors draw on data from the
C-reactive protein Coronary heart disease Genetics Collaboration (CCGC) to illus-
trate the analyses. They present the mathematics in an easy-to-understand way by
using nontechnical language and reinforcing key points at the end of each chapter.
Methods for Using
The last part of the book examines the potential of Mendelian randomization in the
future, exploring both methodological and applied developments. Genetic Variants
in Causal Estimation
Features
• Offers first-hand, in-depth guidance on Mendelian randomization from
leaders in the field
• Makes the diverse aspects of Mendelian randomization understandable to
newcomers
• Illustrates the technical details using data from a large collaborative study
• Includes other real-world examples that show how Mendelian randomization
is used in studies involving inflammation, heart disease, and more
• Discusses possible future directions for research involving Mendelian
randomization

This book gives you the foundation to understand issues concerning the use of Burgess • Thompson Stephen Burgess
genetic variants as instrumental variables. It will get you up to speed in undertak-
ing and interpreting Mendelian randomization analyses. Chapter summaries, paper
summaries, web-based applications, and software code for implementing the sta- Simon G. Thompson
tistical techniques are available on a supplementary website.

K16638

w w w. c rc p r e s s . c o m

K16638_cover.indd 1 12/16/14 8:27 AM

 MENDELIAN
RANDOMIZATION
Methods for Using
Genetic Variants
in Causal Estimation
CHAPMAN & HALL/CRC
Interdisciplinar y Statistics Series
Series editors: N. Keiding, B.J.T. Morgan, C.K. Wikle, P. van der Heijden
Published titles
AGE-PERIOD-COHORT ANALYSIS: Y. Yang and K. C. Land
NEW MODELS, METHODS, AND
EMPIRICAL APPLICATIONS

ANALYSIS OF CAPTURE-RECAPTURE DATA R. S. McCrea and B. J.T. Morgan

AN INVARIANT APPROACH TO S. Lele and J. Richtsmeier

STATISTICAL ANALYSIS OF SHAPES

ASTROSTATISTICS G. Babu and E. Feigelson

BAYESIAN ANALYSIS FOR R. King, B. J.T. Morgan,

POPULATION ECOLOGY O. Gimenez, and S. P. Brooks

BAYESIAN DISEASE MAPPING: A. B. Lawson

HIERARCHICAL MODELING IN SPATIAL
EPIDEMIOLOGY, SECOND EDITION

BIOEQUIVALENCE AND STATISTICS S. Patterson and B. Jones

IN CLINICAL PHARMACOLOGY

CLINICAL TRIALS IN ONCOLOGY, S. Green, J. Benedetti,

THIRD EDITION A. Smith, and J. Crowley

CLUSTER RANDOMISED TRIALS R.J. Hayes and L.H. Moulton

CORRESPONDENCE ANALYSIS M. Greenacre

IN PRACTICE, SECOND EDITION

DESIGN AND ANALYSIS OF D.L. Fairclough

QUALITY OF LIFE STUDIES
IN CLINICAL TRIALS, SECOND EDITION

DYNAMICAL SEARCH L. Pronzato, H. Wynn, and A. Zhigljavsky

FLEXIBLE IMPUTATION OF MISSING DATA S. van Buuren

GENERALIZED LATENT VARIABLE A. Skrondal and S. Rabe-Hesketh

MODELING: MULTILEVEL, LONGITUDINAL,
AND STRUCTURAL EQUATION MODELS

GRAPHICAL ANALYSIS OF K. Basford and J. Tukey

MULTI-RESPONSE DATA

INTRODUCTION TO COMPUTATIONAL M. Waterman

BIOLOGY: MAPS, SEQUENCES, AND
GENOMES
Published titles

MARKOV CHAIN MONTE CARLO W. Gilks, S. Richardson, and

IN PRACTICE D. Spiegelhalter

MEASUREMENT ERROR AND P. Gustafson

MISCLASSIFICATION IN STATISTICS
AND EPIDEMIOLOGY: IMPACTS AND
BAYESIAN ADJUSTMENTS

MEASUREMENT ERROR: J. P. Buonaccorsi

MODELS, METHODS, AND APPLICATIONS

MENDELIAN RANDOMIZATION: METHODS S.Burgess and S.G. Thompson

FOR USING GENETIC VARIANTS IN CAUSAL
ESTIMATION

META-ANALYSIS OF BINARY DATA USING D. Böhning, R. Kuhnert, and

PROFILE LIKELIHOOD S. Rattanasiri

STATISTICAL ANALYSIS OF GENE T. Speed

EXPRESSION MICROARRAY DATA

STATISTICAL AND COMPUTATIONAL R. Wu and M. Lin

PHARMACOGENOMICS

STATISTICS IN MUSICOLOGY J. Beran

STATISTICS OF MEDICAL IMAGING T. Lei

STATISTICAL CONCEPTS AND J. Aitchison, J.W. Kay, and I.J. Lauder

APPLICATIONS IN CLINICAL MEDICINE

STATISTICAL AND PROBABILISTIC P.J. Boland
METHODS IN ACTUARIAL SCIENCE

STATISTICAL DETECTION AND P. Rogerson and I.Yamada

SURVEILLANCE OF GEOGRAPHIC
CLUSTERS

STATISTICS FOR ENVIRONMENTAL A. Bailer and W. Piegorsch

BIOLOGY AND TOXICOLOGY

STATISTICS FOR FISSION TRACK ANALYSIS R.F. Galbraith

VISUALIZING DATA PATTERNS D.B. Carr and L.W. Pickle

WITH MICROMAPS
 C h ap man & H all/CRC
I n t e rd i s c i p l in ar y St atistics Series

MENDELIAN
RANDOMIZATION
Methods for Using
Genetic Variants
in Causal Estimation

Stephen Burgess
Department of Public Health and Primary Care
University of Cambridge, UK

Simon G. Thompson
Department of Public Health and Primary Care
University of Cambridge, UK
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2015 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Version Date: 20141229

International Standard Book Number-13: 978-1-4665-7318-5 (eBook - PDF)

This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor-
age or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copy-
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that pro-
vides licenses and registration for a variety of users. For organizations that have been granted a photo-
copy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Contents

Preface xi

Abbreviations xiii

Notation xiv

I Using genetic variants as instrumental variables to

assess causal relationships 1
1 Introduction and motivation 3

1.1 Shortcomings of classical epidemiology . . . . . . . . . . . . 3

1.2 The rise of genetic epidemiology . . . . . . . . . . . . . . . . 5
1.3 Motivating example: The inflammation hypothesis . . . . . . 6
1.4 Other examples of Mendelian randomization . . . . . . . . . 9
1.5 Overview of book . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 What is Mendelian randomization? 13

2.1 What is Mendelian randomization? . . . . . . . . . . . . . . 13

2.2 Why use Mendelian randomization? . . . . . . . . . . . . . . 18
2.3 A brief overview of genetics . . . . . . . . . . . . . . . . . . . 20
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3 Assumptions for causal inference 25

3.1 Observational and causal relationships . . . . . . . . . . . . . 25

3.2 Finding a valid instrumental variable . . . . . . . . . . . . . 28
3.3 Testing for a causal relationship . . . . . . . . . . . . . . . . 39
3.4 Estimating a causal effect . . . . . . . . . . . . . . . . . . . . 41
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4 Methods for instrumental variable analysis 45

4.1 Ratio of coefficients method . . . . . . . . . . . . . . . . . . 45

4.2 Two-stage methods . . . . . . . . . . . . . . . . . . . . . . . 56

vii
viii Contents

4.3 Likelihood-based methods . . . . . . . . . . . . . . . . . . . . 60

4.4* Semi-parametric methods . . . . . . . . . . . . . . . . . . . . 63
4.5 Efficiency and validity of instruments . . . . . . . . . . . . . 67
4.6 Computer implementation . . . . . . . . . . . . . . . . . . . 69
4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

5 Examples of Mendelian randomization analysis 75

5.1 Fibrinogen and coronary heart disease . . . . . . . . . . . . . 75

5.2 Adiposity and blood pressure . . . . . . . . . . . . . . . . . . 77
5.3 Lipoprotein(a) and myocardial infarction . . . . . . . . . . . 80
5.4 High-density lipoprotein cholesterol and myocardial infarction 82
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

6 Generalizability of estimates from Mendelian randomization 87

6.1 Internal and external validity . . . . . . . . . . . . . . . . . . 87

6.2 Comparison of estimates . . . . . . . . . . . . . . . . . . . . 90
6.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

II Statistical issues in instrumental variable analysis

and Mendelian randomization 97
7 Weak instruments and finite-sample bias 99

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.2 Demonstrating the bias of IV estimates . . . . . . . . . . . . 100
7.3 Explaining the bias of IV estimates . . . . . . . . . . . . . . 102
7.4 Properties of IV estimates with weak instruments . . . . . . 106
7.5 Bias of IV estimates with different choices of IV . . . . . . . 109
7.6 Minimizing the bias of IV estimates . . . . . . . . . . . . . . 112
7.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.8 Key points from chapter . . . . . . . . . . . . . . . . . . . . 121

8 Multiple instruments and power 123

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

8.2 Allele scores . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.3 Power of IV estimates . . . . . . . . . . . . . . . . . . . . . . 126
8.4 Multiple variants and missing data . . . . . . . . . . . . . . . 131
8.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.6 Key points from chapter . . . . . . . . . . . . . . . . . . . . 137
Contents ix

9 Multiple studies and evidence synthesis 139

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

9.2 Assessing the causal relationship . . . . . . . . . . . . . . . . 140
9.3 Study-level meta-analysis . . . . . . . . . . . . . . . . . . . . 140
9.4 Summary-level meta-analysis . . . . . . . . . . . . . . . . . . 140
9.5 Individual-level meta-analysis . . . . . . . . . . . . . . . . . 147
9.6 Example: C-reactive protein and fibrinogen . . . . . . . . . . 150
9.7 Binary outcomes . . . . . . . . . . . . . . . . . . . . . . . . . 152
9.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
9.9 Key points from chapter . . . . . . . . . . . . . . . . . . . . 156

10 Example: The CRP CHD Genetics Collaboration 157

10.1 Overview of the dataset . . . . . . . . . . . . . . . . . . . . . 157

10.2 Single study: Cardiovascular Health Study . . . . . . . . . . 164
10.3 Meta-analysis of all studies . . . . . . . . . . . . . . . . . . . 165
10.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
10.5 Key points from chapter . . . . . . . . . . . . . . . . . . . . 172

III Prospects for Mendelian randomization 173

11 Future directions 175

11.1 Methodological developments . . . . . . . . . . . . . . . . . . 175

11.2 Applied developments . . . . . . . . . . . . . . . . . . . . . . 180
11.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

Bibliography 185

Index 207
Preface

The quantity of research into the genetics of common diseases has exploded
over the last 20 years. While many genetic variants related to various diseases
have been identified, their usefulness may lie more in what they offer to our
understanding of the biological mechanisms leading to disease rather than
to, for example, predicting disease risk. To understand mechanisms, we need
to separate the relationships of risk factors with diseases into those that are
causal and those that are not. This is where Mendelian randomization can
play an important role.
The technique of Mendelian randomization itself has undergone rapid de-
velopment, mostly in the last 10 years, and applications now abound in current
medical and epidemiological journals. Its basis is that of instrumental vari-
able analysis, which has a much longer history in statistics and particularly
in econometrics. Relevant papers on Mendelian randomization are therefore
dispersed across the multiple fields of genetics, epidemiology, statistics and
econometrics. The intention of this book is to bring together this literature
on the methods and practicalities of Mendelian randomization, especially to
help those who are relatively new to this area.
In writing this book, we envisage the target audience comprising two main
groups, Epidemiologists and Medical Statisticians, who want to perform ap-
plied Mendelian randomization analyses or understand how to interpret their
results. We therefore assume a familiarity with basic epidemiological terminol-
ogy, such as prospective and case-control studies, and basic statistical meth-
ods, such as ordinary least squares and logistic regression. Meanwhile, we have
tried to make the perhaps alien terminology of econometrics accessible to our
intended readership.
While we hope that this book will be accessible to a wide audience, a
geneticist may baulk at the simplistic explanations of Mendelian inheritance,
a statistician may yearn for a deeper level of technical exposition, and an
epidemiologist may wonder why we don’t just cut to the chase of how to
perform the analyses. Our hope is that enough detail is given for those who
need it, references are available for those who want more, and a section can
simply be glossed over by those for whom it is redundant.
While we have included relevant statistical methodology available up to
the publication date of the book, our focus has been on methods and issues
which are of practical relevance for applied Mendelian randomization analy-
ses, rather than those which are of more theoretical interest, or ‘cutting-edge’
developments which may not stand the test of time. As such, to a research

xi
xii Mendelian Randomization

statistician, the book will provide a background to current areas of method-

ological debate, but it will generally not offer opinions on controversial topics
which are likely to become out-of-date quickly as further investigations are
performed. Where possible, sections with technical content in the first part of
the book are marked with asterisks (*), and are written in such a way that
they can be omitted without interrupting the flow of the book.
A website to complement this book, as well as the authors’ ongoing re-
search on this topic, is available at www.mendelianrandomization.com. This
contains chapter summaries, paper summaries, web-based applications, and
software code for implementing some of the statistical techniques discussed in
the book.
We would like to express our thanks to all those who commented on chap-
ters of this book, whether in chapter or book form. We thank Frank Dudbridge,
Brandon Pierce, Dylan Small, Maria Glymour, Stephen Sharp, Mary School-
ing, Tom Palmer, George Davey Smith, Debbie Lawlor, John Thompson, Jack
Bowden, Shaun Seaman, Lucas Tittmann, Daniel Freitag, Peter Willeit, Ed-
mund Jones, Angela Wood and Adam Butterworth. Further individuals com-
mented as anonymous referees, and so we cannot thank them by name. We
also thank Rob Calver, our editor, for being knowledgeable, supportive, and
open to our ideas. We are also grateful to the principal investigators of the
studies in the CRP CHD Genetics Collaboration who have allowed us to use
their data in this book, as well as to the study participants for giving their
time and consent to participate in this research.
In short, while we realize that we will not be able to please all of our readers
all of the time, we hope that this book will enable a wide range of people to
better understand what is an important, but complex and multidisciplinary,
area of research.

Stephen Burgess, Simon G. Thompson

University of Cambridge, UK
August 2014
Abbreviations

2SLS two-stage least squares

ACE average causal effect
BMI body mass index
CCGC CRP CHD Genetics Collaboration
CRP C-reactive protein
CHD coronary heart disease
CI confidence interval
COR causal odds ratio
CRR causal risk ratio
DIC deviance information criterion
DNA deoxyribonucleic acid
FIML full information maximum likelihood
FTO a gene associated with obesity
GMM generalized method of moments
GWAS genome-wide association study (or studies)
HDL-C high-density lipoprotein cholesterol
IL6 interleukin-6
IPD individual participant data
IV instrumental variable
LIML limited information maximum likelihood
LIVAE linear IV average effect
LD linkage disequilibrium
LDL-C low-density lipoprotein cholesterol
lp(a) lipoprotein(a)
MAR missing at random
MCMC Monte Carlo Markov chain
MI myocardial infarction
OLS ordinary least squares
OR odds ratio
RCT randomized controlled trial
RNA ribonucleic acid
SE standard error
SMM structural mean model
SNP single nucleotide polymorphism
SUTVA stable unit treatment value assumption

Abbreviations for the various studies in the CCGC are given in Table 10.1.

xiii
xiv Mendelian Randomization

Notation

Throughout this book, we use the notation:

X exposure: the risk factor (or protective factor, or inter-

mediate phenotype) of interest
Y outcome
U (sufficient) confounder of the X–Y association
G instrumental variable
α parameter of genetic association: regression parameter
in the G–X regression
β regression parameter in the X–Y regression
β1 causal effect of X on Y : the main parameter of interest
ρ correlation parameter between X and Y
ρGX correlation parameter between G and X
σ2 variance parameter
τ2 between-study heterogeneity variance parameter
F F statistic from regression of X on G
i subscript indexing individuals
j subscript indexing genetic subgroups
J total number of genetic subgroups
k subscript indexing genetic variants (SNPs)
K total number of genetic variants
m subscript indexing studies in a meta-analysis
M total number of studies
N total number of individuals
n total number of cases (individuals with a disease event)
N normal distribution
N2 bivariate normal distribution

We follow the usual convention of using upper-case letters for random

variables and lower-case letters for data values with X, Y , U , and G.
 Part I

Using genetic variants as

instrumental variables to
assess causal relationships
1
Introduction and motivation

This book concerns making inferences about causal effects based on ob-
servational data using genetic instrumental variables, a concept known as
Mendelian randomization. In this chapter, we introduce the basic idea of
Mendelian randomization, giving examples of when the approach can be used
and why it may be useful. We aim in this chapter only to give a flavour of
the approach; details about its conditions and requirements are reserved for
later chapters. Although the examples given in this book are mainly in the
context of epidemiology, Mendelian randomization can address questions in
a variety of fields of study, and the majority of the material in this book is
equally relevant to problems in different research areas.

1.1 Shortcomings of classical epidemiology

Epidemiology is the study of patterns of health and disease at the population
level. We use the term ‘classical epidemiology’ meaning epidemiology without
the use of genetic factors, to contrast with genetic epidemiology. A fundamen-
tal problem in epidemiological research, in common with other areas of social
science, is the distinction between correlation and causation. If we want to
address important medical questions, such as to determine disease aetiology
(what is the cause of a disease?), to assess the impact of a medical or public
health intervention (what would be the result of a treatment?), to inform pub-
lic policy, to prioritize healthcare resources, to advise clinical practice, or to
counsel on the impact of lifestyle choices, then we have to answer questions of
cause and effect. The optimal way to address these questions is by appropriate
study design, such as the use of prospective randomized trials.

1.1.1 Randomized trials and observational studies

The ‘gold standard’ for the empirical testing of a scientific hypothesis in clin-
ical research is a randomized controlled trial. This design involves the alloca-
tion of different treatment regimes at random to experimental units (usually
individuals) in a population. In its simplest form, one ‘active treatment’ (for
example, intervention on a risk factor) is compared against a ‘control treat-

3
4 Mendelian Randomization

ment’ (no intervention), and the average outcomes in each of the arms of the
trial are contrasted. Here the risk factor (which we will often refer to as the
“exposure” variable) is a putative causal risk factor. We seek to assess whether
the risk factor is a cause of the outcome, and estimate (if appropriate) the
magnitude of the causal effect.
While randomized trials are in principle the best way of determining the
causal status of a particular risk factor, they have some limitations. Random-
ized trials are expensive and time-consuming, especially when the outcome
is rare or requires a long follow-up period to be observed. Additionally, in
some cases, a targeted treatment which has an effect only on the risk factor
of interest may not be available. Moreover, many risk factors cannot be ran-
domly allocated for practical or ethical reasons. For example, in assessing the
impact of drinking red wine on the risk of coronary heart disease, it would
not be feasible to recruit participants to be randomly assigned to either drink
or abstain from red wine over, say, a 20-year period. Alternative approaches
for judging causal relationships are required.
Scientific hypotheses are often assessed using observational data. Rather
than by intervening on the risk factor, individuals with high and low levels of
the risk factor are compared. In many cases, differences between the average
outcomes in the two groups have been interpreted as evidence for the causal
role of the risk factor. However, such a conclusion confuses correlation with
causation. There are many reasons why individuals with elevated levels of the
risk factor may have greater average outcome levels, without the risk factor
being a causal agent.
Interpreting an association between an exposure and a disease outcome in
observational data as a causal relationship relies on untestable and usually
implausible assumptions, such as the absence of unmeasured confounding (see
Chapter 2) and of reverse causation. This has led to several high-profile cases
where a risk factor has been widely promoted as an important factor in disease
prevention based on observational data, only to be later discredited when evi-
dence from randomized trials did not support a causal interpretation [Taubes
and Mann, 1995]. For example, observational studies reported a strong inverse
association between vitamin C and risk of coronary heart disease, which did
not attenuate on adjustment for a variety of risk factors [Khaw et al., 2001].
However, results of experimental data obtained from randomized trials showed
a non-significant association in the opposite direction [Collins et al., 2002].
The confidence interval for the observational association did not include the
randomized trial estimate [Davey Smith and Ebrahim, 2003]. Similar stories
apply to the observational and experimental associations between β-carotene
and smoking-related cancers [Peto et al., 1981; Hennekens et al., 1996], and
between vitamin E and coronary heart disease [Hooper et al., 2001]. More
worrying is the history of hormone-replacement therapy, which was previously
advocated as being beneficial for the reduction of breast cancer and cardio-
vascular mortality on the basis of observational data, but was subsequently
shown to increase mortality in randomized trials [Rossouw et al., 2002; Beral
Introduction and motivation 5

et al., 2003]. More robust approaches are therefore needed for assessing causal
relationships using observational data. Mendelian randomization is one such
approach.

1.2 The rise of genetic epidemiology

Genetic epidemiology is the study of the role of genetic factors in health and
disease for populations. We sketch the history and development of genetic
epidemiology, indicating why it is an important area of epidemiological and
scientific research.

1.2.1 Historical background

Although the inheritance of characteristics from one generation to the next
has been observed for millennia, the mechanism for inheritance was long un-
known. When Charles Darwin proposed his theory of evolution in 1859, one
of its major problems was the lack of an underlying mechanism for heredity
[Darwin, 1871]. Gregor Mendel in 1866 proposed two laws of inheritance: the
law of segregation, that when any individual produces gametes (sex cells),
the two copies of a gene separate so that each gamete receives only one copy;
and the law of independent assortment, that ‘unlinked or distantly linked seg-
regating gene pairs assort independently at meiosis [cell division]’ [Mendel,
1866]. These laws are summarized by the term “Mendelian inheritance”, and
it is this which gives Mendelian randomization its name [Davey Smith and
Ebrahim, 2003]. The two areas of evolution and Mendelian inheritance were
brought together through the 1910s-30s in the “modern evolutionary synthe-
sis”, by amongst others Ronald Fisher, who helped to develop population
genetics [Fisher, 1918]. A specific connection between genetics and disease
was established by Linus Pauling in 1949, who linked a specific genetic mu-
tation in patients with sickle-cell anaemia to a demonstrated change in the
haemoglobin of the red-blood cells [Pauling et al., 1949]. The discovery of
the structure of deoxyribonucleic acid (DNA) in 1953 gave rise to the birth
of molecular biology, which led to greater understanding of the genetic code
[Watson and Crick, 1953]. The Human Genome Project was established in
1990, leading to the publication of the entirety of the human genetic code
by the early 2000s [Roberts et al., 2001; McPherson et al., 2001]. Recently,
technological advances have reduced the cost of DNA sequencing to the level
where it is now economically viable to measure genetic information for a large
number of individuals [Shendure and Ji, 2008].
6 Mendelian Randomization

1.2.2 Genetics and disease

As the knowledge of the human genome has developed, the search for genetic
determinants of disease has expanded from monogenic disorders (disorders
which are due to a single mutated gene, such as sickle-cell anaemia), to poly-
genic and multifactorial disorders, where the burden of disease risk is not due
to a single gene, but to multiple genes combined with lifestyle and environ-
mental factors. These diseases, such as cancers, diabetes and coronary heart
disease, tend to cluster within families, but also depend on modifiable risk fac-
tors, such as diet and blood pressure. Several genetic factors have been found
which relate to these diseases, especially through the increased use of genome-
wide association studies (GWAS), in which the associations of thousands or
even millions of genetic variants with a disease outcome are tested. In some
cases, these discoveries have added to the scientific understanding of disease
processes and the ability to predict disease risk for individuals. Nevertheless,
they are of limited immediate interest from a clinical perspective, as an in-
dividual’s genome cannot generally be changed. However, genetic discoveries
provide opportunities for Mendelian randomization: a technique for using ge-
netic data to assess and estimate causal effects of modifiable (non-genetic)
risk factors based on observational data.

1.3 Motivating example: The inflammation hypothesis

We introduce the approach of Mendelian randomization using an example. The
‘inflammation hypothesis’ is an important question in the understanding of
cardiovascular disease. Inflammation is one of the body’s response mechanisms
to a harmful stimulus. It is characterized by redness, swelling, heat, pain and
loss of function in the affected body area. Cases can be divided into acute
inflammation, which refers to the initial response of the body, and chronic
inflammation, which refers to more prolonged changes. Examples of conditions
classified as inflammation include appendicitis, chilblains, and arthritis.
Cardiovascular disease is a term covering a range of diseases including coro-
nary heart disease (in particular myocardial infarction or a ‘heart attack’) and
stroke. It is currently the biggest cause of death worldwide. The inflamma-
tion hypothesis states that there is some aspect of the inflammation response
mechanism which leads to cardiovascular disease events, and that intervening
on this pathway will reduce the risk of cardiovascular disease.

1.3.1 C-reactive protein and coronary heart disease

As part of the inflammation process, several chemicals are produced by the
body, known as (positive) acute-phase proteins. These represent the body’s
Introduction and motivation 7

first line of defence against infection and injury. There has been particular in-
terest in one of these, C-reactive protein (CRP), and the role of elevated levels
of CRP in the risk of coronary heart disease (CHD). It is known that CRP is
observationally associated with the risk of CHD [Kaptoge et al., 2010], but,
prior to robust Mendelian randomization studies, it was not known whether
this association was causal [Danesh and Pepys, 2009]. The specific question in
this example (a small part of the wider inflammation hypothesis) is whether
long-term elevated levels of CRP lead to greater risk of CHD.

1.3.2 Alternative explanations for association

In our example, there are many factors that increase both levels of CRP
and the risk of CHD. These factors, known as confounders, may be measured
and accounted for by statistical analysis, for instance multivariable regression.
However, it is not possible to know whether all such factors have been identi-
fied. Also, CRP levels increase in response to sub-clinical disease, giving the
possibility that the observed association is due to reverse causation.
One of the potential confounders of particular interest is fibrinogen, a sol-
uble blood plasma glycoprotein, which enables blood-clotting. It is also part
of the inflammation pathway. Although CRP is observationally positively as-
sociated with CHD risk, this association was shown to reduce on adjustment
for various conventional risk factors (such as age, sex, body mass index, and
diabetes status), and to attenuate to near null on further adjustment for fib-
rinogen [Kaptoge et al., 2010]. It is important to assess whether elevated levels
of CRP are causally related to changes in fibrinogen, since if so conditioning
the CRP–CHD association on fibrinogen would represent an over-adjustment,
which would attenuate a true causal effect.

1.3.3 Instrumental variables

To address the problems of confounding and reverse causation in conventional
epidemiology, we introduce the concept of an instrumental variable. An instru-
mental variable is a measurable quantity (a variable) which is associated with
the exposure of interest, but not associated with any other competing risk
factor that is a confounder. Neither is it associated with the outcome, except
potentially via the hypothesized causal pathway through the exposure of in-
terest. A potential example of an instrumental variable for health outcomes is
geographic location. We imagine that two neighbouring regions have different
policies on how to treat patients, and assume that patients who live on one
side of the border are similar in all respects to those on the other side of the
border, except that they receive different treatment regimes. By comparing
these groups of patients, geographic location acts like the random allocation
to treatment assignment in a randomized controlled trial, influencing the ex-
posure of interest without being associated with competing risk factors. It
therefore is an instrumental variable, and gives rise to a natural experiment
8 Mendelian Randomization

in the population, from which causal inferences can be obtained. Other plausi-
ble non-genetic instrumental variables include government policy changes (for
example, the introduction of a smoking ban in public places, or an increase
in cigarette tax, which might decrease cigarette smoking prevalence without
changing other variables) and physician prescribing preference (for example,
the treatment a doctor chose to prescribe to the previous patient, which will be
representative of the doctor’s preferred treatment, but should not be affected
by the current patient’s personal characteristics or case history).

1.3.4 Genetic variants as instrumental variables

A genetic variant is a section of genetic code that differs between individu-
als. In Mendelian randomization, genetic variants are used as instrumental
variables. Individuals in a population can be divided into subgroups based on
their genetic variants. On the assumption that the genetic variants are ‘ran-
domly’ distributed in the population, that is independently of environmental
and other variables, then these genetic subgroups do not systematically differ
with respect to any of these variables. Additionally, as the genetic code for
each individual is determined before birth, there is no way that a variable
measured in a mature individual can be a ‘cause’ of a genetic variant. Re-
turning to our example, if we can find a suitable genetic variant (or variants)
associated with CRP levels, then we can compare the genetically-defined sub-
group of individuals with lower average levels of CRP to the subgroup with
higher average levels of CRP. In effect, we are exploiting a natural experi-
ment in the population, whereby nature has randomly given some individuals
a genetic ‘treatment’ which increases their CRP levels. If individuals with a
genetic variant, which is associated with elevated average levels of CRP and
satisfies the instrumental variable assumptions, exhibit greater incidence of
CHD, then we can conclude that CRP is a causal risk factor for CHD, and
that lowering CRP is likely to lead to reductions in CHD rates. Under further
assumptions about the statistical model for the relationship between CRP
and CHD risk, a causal parameter can be estimated. Although Mendelian
randomization uses genetic variants to answer inferential questions, these are
not questions about genetics, but rather about modifiable risk factors, such
as CRP, and their causal effect on outcomes (usually disease outcomes).

1.3.5 Violations of instrumental variable assumptions

It is impossible to test whether there is a causal relationship between two
variables on the basis of observational data alone. All empirical methods for
making causal claims by necessity rely on untestable assumptions. Instrumen-
tal variable methods are no exception. Taking the example of Section 1.3.3,
if geographic location is associated with other factors, such as socioeconomic
status, then the assumption that the distribution of the outcome would be
the same for both populations under each policy regime would be violated.
Introduction and motivation 9

Or if the genetic variant(s) associated with CRP levels used in a Mendelian

randomization analysis were also independently associated with, say, blood
pressure, the comparison of genetic subgroups would not be a valid test of
the causal effect of CRP on CHD risk. The validity of the instrumental vari-
able assumptions is crucial to the interpretation of a Mendelian randomization
investigation, and is discussed at length in later chapters.

1.3.6 The CRP CHD Genetics Collaboration

The statistical methods and issues discussed in this book are illustrated us-
ing the example of the causal relationships of CRP on the outcomes CHD
risk and fibrinogen. Data are taken from the CRP CHD Genetic Collabora-
tion (CCGC), a consortium of 47 studies comprising cohort, case-control and
nested case-control studies [CCGC, 2008]. Most of these studies recorded data
on CRP levels, on incident CHD events (or history of CHD events in retro-
spective or cross-sectional studies), and on up to 20 genetic variants associated
with CRP levels. Of these, we will focus on four, which were pre-specified as
the variants to be used as instrumental variables in the main applied analysis
from the collaboration and are located in and around the CRP gene region
on chromosome 1. Some studies did not measure all four of these variants;
others did not measure CRP levels in some or all participants. Several stud-
ies measured a range of additional covariates, including fibrinogen, many of
which are potential confounders in the association between CRP and CHD
risk. A full analysis of the data from the CCGC for the causal effect of CRP
on CHD risk is given in Chapter 10. While the aim of the book is not to prove
or disprove the causal role of CRP for CHD, the epidemiological implications
of the analyses are explored.

1.4 Other examples of Mendelian randomization

Although the initial applications of Mendelian randomization were in the field
of epidemiology [Youngman et al., 2000], the use of genetic instrumental vari-
ables is becoming widespread in a number of different fields. A systematic
review of applied Mendelian randomization studies was published in 2010
[Bochud and Rousson, 2010]. A list of the exposures and outcomes of some
causal relationships which have been assessed using Mendelian randomization
is given in Table 1.1. The list includes examples from the fields of epidemi-
ology, nutrition, sociology, psychology, and economics: the only limitation in
the use of Mendelian randomization to assess the causal effect of an exposure
on an outcome is the availability of a suitable genetic variant to use as the
instrumental variable.
10 Mendelian Randomization

The reasons to use Mendelian randomization outside of epidemiology are

similar to those in epidemiology. In many fields, randomized experiments are
difficult to perform and instrumental variable techniques represent one of the
few ways of assessing causal relationships in the absence of complete knowledge
of confounders. Although the language and context of this book will generally
be that of epidemiology, much applies equally to other areas of research. More
detailed expositions of some examples of applied Mendelian randomization
analyses are given in Chapter 5.

1.5 Overview of book

Although there has been much research into the use of instrumental variables
in econometrics and epidemiology since they were first proposed [Wright,
1928], several barriers existed in applying this to the context of Mendelian
randomization. These include differences in terminology, where the same con-
cept is referred to in various disciplines by different names, and differences
in theoretical concepts, particularly relating to the definition and interpreta-
tion of causal relationships. Additionally, several methodological issues have
been posed by the use of genetic variants as instrumental variables that had
not previously been considered in the instrumental variables literature, and
required (and still require) methodological development. A major motivation
in writing this book is to provide an accessible resource to those coming from
different academic disciplines to understand issues relevant to the use of ge-
netic variants as instrumental variables, and particularly for those wanting to
undertake and interpret Mendelian randomization analyses.

1.5.1 Structure
This book is divided into three parts. The first part, comprising Chapters 1
to 6, is entitled “Using genetic variants as instrumental variables to assess
causal relationships”. This part contains the essential information for a prac-
titioner interested in Mendelian randomization (Chapters 1 and 2), including
definitions of causal relationships and instrumental variables (Chapter 3), and
methods for the estimation of causal effects (Chapter 4). With the exception
of some of the technical details about statistical methods marked as ‘starred’,
these sections should be fully accessible to most epidemiologists. Issues sur-
rounding the application of Mendelian randomization in practice are explored
by presenting examples of Mendelian randomization investigations from the
literature (Chapter 5). Also addressed is the question of how to interpret a
Mendelian randomization estimate, and how it may compare to the effect of
an intervention on the exposure of interest in practice (Chapter 6).
The second part, comprising Chapters 7 to 10, is entitled “Statistical is-
sues with instrumental variable analysis in Mendelian randomization”. This
Introduction and motivation 11

Nature of exposure Exposure Outcome Reference

apolipoprotein E cancer [1]
CRP insulin resistance [2]
CRP CIMT [3]
CRP cancer [4]
folate blood pressure [5]
Biomarker
HDL-C myocardial infarction [6]
homocysteine stroke [7]
lipoprotein(a) myocardial infarction [8–9]
SHBG CHD [10]
BMI CIMT [11]
BMI early menarche [12]
Physical
BMI labour market outcomes [13]
characteristic
fat mass academic achievement [14]
alcohol intake blood pressure [15]
caffeine intake stillbirth [16]
Dietary factor
milk intake metabolic syndrome [17]
alcohol abuse drug abuse [18]
Pathological
ADHD education [19]
behaviour
depression education [19]
Inter-generational interuterine
neural tube defects [20]
effect folate

TABLE 1.1
Examples of causal relationships assessed by Mendelian randomization in ap-
plied research.

Abbreviations:
CRP = C-reactive protein, CIMT = carotid intima-media thickness, CHD = coronary
heart disease, SHBG = sex-hormone binging globulin, HDL-C = high-density lipoprotein
cholesterol, BMI = body mass index, ADHD = attention deficit hyperactivity disorder.
References:
1. Trompet et al., 2009, 11. Kivimäki et al., 2007,
2. Timpson et al., 2005, 12. Mumby et al., 2011,
3. Kivimäki et al., 2008, 13. Norton and Han, 2008,
4. Allin et al., 2010, 14. Von Hinke et al., 2010,
5. Thompson et al., 2005, 15. Chen et al., 2008,
6. Voight et al., 2012, 16. Bech et al., 2006,
7. Casas et al., 2005, 17. Almon et al., 2010,
8. Kamstrup et al., 2009, 18. Irons et al., 2007,
9. Clarke et al., 2009, 19. Ding et al., 2009b,
10. Ding et al., 2009a, 20. Ebrahim and Davey Smith, 2008
12 Mendelian Randomization

consists of comparisons of methods for using instrumental variables to esti-

mate a causal effect, and matters concerning the behaviour of instrumental
variable estimates, such as potential biases. In particular, we consider the is-
sue of weak instrument bias (Chapter 7), and the problems of estimating a
single causal effect using data on multiple instrumental variables (Chapter
8) and data from multiple studies (Chapter 9). Estimates from instrumental
variable methods typically have wide confidence intervals, often necessitating
the synthesis of evidence from multiple sources to obtain an estimate pre-
cise enough to be clinically relevant. As part of the discussion on the use of
multiple instruments, we address questions relating to the power and sample
size requirements of Mendelian randomization studies. This part of the book
is illustrated throughout using data from the CCGC, and a comprehensive
analysis of the CCGC dataset for the causal effect of CRP on CHD risk is
provided (Chapter 10). Although the details in this part require a greater
depth of mathematical understanding, each chapter is introduced using non-
technical language, and concludes with a set of key points to convey the main
messages of the chapter.
Finally, we conclude with the final part, Chapter 11, by discussing possible
future directions for research involving Mendelian randomization.

1.6 Summary
Distinguishing between a factor which is merely associated with an outcome
and one which has a causal effect on the outcome is problematic outside of
the context of a randomized controlled trial. Instrumental variables provide
a way of assessing causal relationships in observational data, and Mendelian
randomization is the use of genetic variants as instrumental variables.
In the next chapter, we provide more detail of what Mendelian random-
ization is, and when and why it may be useful.
2
What is Mendelian randomization?

In this chapter, we illustrate the conceptual framework and motivation for

Mendelian randomization, explaining how Mendelian randomization offers op-
portunities to address some of the problems of conventional epidemiology.
We describe the specific characteristics of genetic data which give rise to the
Mendelian randomization approach.

2.1 What is Mendelian randomization?

Mendelian randomization is the use of genetic variants in non-experimental
data to make causal inferences about the effect of an exposure on an outcome.
We use the word “exposure” throughout this book to refer to the putative
causal risk factor, sometimes called an intermediate phenotype, which can be a
biomarker, an anthropometric measure, or any other risk factor that may affect
the outcome. Usually the outcome is disease, although there is no methodolog-
ical restriction as to what outcomes can be considered. Non-experimental data
encompass all observational studies, including cross-sectional and longitudi-
nal, cohort and case-control designs – any study where there is no intervention
applied by the researcher.

2.1.1 Motivation
A foundational aim of epidemiological research is the estimation of the effect
of changing an exposure on an outcome. This is known as the causal effect
of the exposure on the outcome, and typically differs from the observational
association between the exposure and outcome, for example due to confound-
ing. Correlation between the exposure and the outcome cannot be reliably
interpreted as evidence of a causal relationship. For example, those who drink
red wine regularly have a lower incidence of heart disease. But socio-economic
status is a common predictor of both wine consumption and better coronary
health, and so it may be that socio-economic status rather than wine con-
sumption underlies the risk of heart disease. Observational associations may
also arise as a result of reverse causation. For example, those who regularly
take headache tablets are likely to have more headaches than those who do

13
14 Mendelian Randomization

not, but taking headache tablets is unlikely to be a cause of the increased inci-
dence of headaches. Another example is vitamin D levels, which may decrease
in individuals who are ill and therefore do not go outside, rather than vitamin
D being a cause of illness.
The idea of Mendelian randomization is to find a genetic variant (or vari-
ants) associated with the exposure, but not associated with any other risk fac-
tor which affects the outcome, and not directly associated with the outcome.
This means that any association of the genetic variant with the outcome must
come via the variant’s association with the exposure, and therefore implies a
causal effect of the exposure on the outcome. Such a genetic variant would
satisfy the assumptions of an instrumental variable (IV) [Greenland, 2000a;
Sussman and Hayward, 2010]. As the theory of IVs was initially developed
in the field of econometrics, a number of terms commonly used in the IV lit-
erature derive from this field and are not always well understood by medical
statisticians or epidemiologists. Table 2.1 is a glossary of terms which are used
in each field.

2.1.2 Instrumental variables

A technical definition of Mendelian randomization is “instrumental variable
analysis using genetic instruments” [Wehby et al., 2008]. In Mendelian ran-
domization, genetic variant(s) are used as IVs for assessing the causal effect
of the exposure on the outcome [Thomas and Conti, 2004].
The fundamental conditions for a genetic variant to satisfy to be an IV
are summarized as:
i. the variant is associated with the exposure,
ii. the variant is not associated with any confounder of the exposure–outcome
association,
iii. the variant does not affect the outcome, except possibly via its association
with the exposure.
Although Mendelian randomization analyses often involve a single genetic
variant, multiple variants can be used either as separate IVs or combined into
a single IV. More detail on the IV assumptions, which are key to the validity
of Mendelian randomization investigations, is given in Chapter 3.

2.1.3 Confounding and endogeneity

One of the reasons why there may be a correlation between the exposure
and outcome in an observational study is confounding, or the related concept,
endogeneity of the exposure.
Confounding is defined as the presence of inherent differences between
groups with different levels of the exposure [Greenland and Robins, 1986]. It
What is Mendelian randomization? 15

Econometrics Epidemiological Notes

term term
Endogenous / Confounded / A variable is confounded / endogenous
endogeneity confounding in a regression model if it is correlated
Exogenous / Unconfounded / with the error term, meaning that the
exogeneity no confounding regression coefficient is a biased estimate of
the causal effect. A variable is unconfounded
/ exogenous if it is not correlated with the
error term (see Section 2.1.3).

Outcome Outcome Denoted Y in this text.

Endogenous Exposure Denoted X in this text; the causal effect

regressor of X on Y cannot be estimated by simple
regression of Y on X if there is unmeasured
confounding.

Instrumental Instrumental Denoted G in this text; the instrument

variable / variable / is called ‘excluded’ because it is not
excluded instrument included in the second-stage of the two-stage
instrument regression method often used for calculating
IV estimates.

Included Measured A covariate that is included in a model, such

regressor covariate as a multivariable regression.

OLS Least-squares OLS stands for ordinary least squares.

regression The OLS estimate is the observational
association, as opposed to the IV estimate,
which is an estimate of the causal effect.

Concentrate Profile out To exclude a nuisance parameter from an

out equation by forming a profile likelihood with
its maximum likelihood estimate given the
other variables.

Panel data Longitudinal Data on items at multiple timepoints.

data

TABLE 2.1
A summary of instrumental variable terms used in the fields of econometrics
and epidemiology.
16 Mendelian Randomization

is often considered to result from the distribution of particular variables in

the population, known as confounders. A confounder is a variable which is
a common cause of both the exposure and the outcome. When confounders
are recognized, measured and adjusted for, for example by multivariable re-
gression, the remaining association between the exposure and outcome will
often still be a biased estimate of the causal effect, due to the existence of un-
known or unmeasured confounders or imprecision in measured confounders.
Confounding not adjusted for in an analysis is termed ‘residual confounding’.
Endogeneity means that there is a correlation between the regressor and
the error term in a regression model. The words ‘exogenous’ and ‘endogenous’
are rarely used in epidemiology (see Table 2.1), but the terms have rigorous
definitions that are useful in understanding confounding. Endogeneity literally
means “coming from within”. The opposite of endogenous is exogenous; an
exogenous variable “comes from outside” of the regression model. The term
endogeneity encompasses confounding, but also includes phenomena that are
traditionally thought of as separate from confounding, such as measurement
error and reverse causation. If the exposure in a model is endogenous in a
regression model, then the regression coefficient for the exposure will be biased
for the causal effect. An IV can be understood as an exogenous variable,
associated with an endogenous exposure, which is used to estimate the causal
effect of changing the exposure while keeping all other factors equal [Martens
et al., 2006].
Mendelian randomization has also been named ‘Mendelian deconfound-
ing’ [Tobin et al., 2004] as it aims to give estimates of the causal effect free
from biases due to confounding. The correlations between risk factors make
it impossible in an observational study to look at the increase in one variable
keeping all others equal, as changes in one factor will always be accompanied
by changes in other factors. While we can measure individual confounders and
adjust for them in our analysis, we can never be certain that all confounders
have been identified or measured precisely, leading to residual confounding.
Additionally, if we adjust for a variable that lies on the true causal pathway
between the exposure of interest and outcome (a mediator), this represents an
over-adjustment and attenuates the estimate of the causal effect [Christenfeld
et al., 2004]. By finding a genetic variant which satisfies the IV assumptions,
we can estimate the unconfounded association between the exposure and the
outcome.

2.1.4 Analogy with a randomized controlled trial

Mendelian randomization is analogous to a randomized controlled trial (RCT)
[Nitsch et al., 2006]. An RCT, considered to provide the “gold standard” of
medical evidence, involves dividing a set of individuals into two or more sub-
groups in a random way. These subgroups are each given different treatments.
Randomization is preferred over any other assignment to subgroups as all
What is Mendelian randomization? 17

Randomized trial Mendelian randomization

Randomization into groups Randomization by genetic variant

Variant allele Variant allele

Control Treatment
absent present

Exposure higher Exposure lower

Competing risk factors assumed equal by design

Outcome higher Outcome lower

FIGURE 2.1
Comparison of a randomized controlled trial and Mendelian randomization.

possible confounders, known and unknown, are on average balanced between

the subgroups.
In Mendelian randomization, we use a genetic variant to form subgroups
analogous to those in an RCT, as shown in Figure 2.1. From the IV assump-
tions (Section 2.1.2), these subgroups differ systematically in the exposure,
but not in any other factor except for those causally ‘downstream’ of the ex-
posure. A difference in outcomes between these subgroups would therefore
indicate a causal effect of the exposure on the outcome [Hernán and Robins,
2006]. Inferring a causal effect of the exposure on the outcome from an asso-
ciation between the genetic variant and the outcome is analogous to inferring
an intention-to-treat effect from an association between randomization and
the outcome in an RCT (that is, assignment to the treatment group affects
the outcome).
Genetic variants for an individual are inherited from their parents, and
so are not randomly assigned. For example, if neither of an individual’s par-
ents carry a particular genetic mutation, there is no way that the individual
will carry that mutation. Nonetheless, under fairly realistic conditions the dis-
tribution of genetic variants in the population can be thought of as random
with respect to environmental and social factors which may be important con-
founders. The necessary assumptions for a variant to be randomly distributed
are random mating and lack of selection effects relating to the variant of inter-
est. While there will be some departures from these assumptions, studies have
shown that the distribution of most genetic variants is fairly uniform across the
population, at least for example in a Western European context [Davey Smith,
2011]. Considerable departures from the random mating assumptions which
18 Mendelian Randomization

may invalidate the use of a genetic variant can be assessed by performing a

test of Hardy–Weinberg equilibrium, to see if the frequency of heterozygotes
and homozygotes (see Section 2.3) in the population is in line with what is ex-
pected. A variable which is distributed as if being randomly assigned despite
the lack of true randomness in the assignment is known as quasi-randomized.
Most natural experiments rely on quasi-randomization rather than the strict
randomization of experimental units.
A recent observational study showed that linear regression gave a p-value
less than 0.01 in 45% of 4560 associations between all pairs of 96 non-genetic
variables [Davey Smith et al., 2007]. This suggests that many observed asso-
ciations between environmental variables may not have a true causal inter-
pretation. In contrast, the proportion of associations between genetic variants
and these 96 variables with p-values less than 0.01 was not significantly higher
than would be expected by chance. This gives plausibility to the assumption
that genetic variants used as IVs will be distributed independently from many
potential confounders, and so in many cases assignment to a genetic subgroup
can be regarded as analogous to randomization in an RCT.
However, Mendelian randomization differs from a randomized trial in an-
other respect. The aim of Mendelian randomization is not to estimate the size
of a genetic effect, but the causal effect of the exposure on the outcome. The
average change in the outcome associated with a genetic variant may differ
in magnitude from that resulting from an intervention in the exposure (see
Chapter 6). Additionally, even if the association of a genetic variant with the
outcome is small in magnitude, the population attributable risk of the expo-
sure is not necessarily low, as the exposure may vary to a considerably larger
extent than that which can be explained by the variant. It may be possible
to change the exposure by a greater amount than the difference in the mean
exposure between genetic subgroups. For example, the effect of statin drug
use on low-density lipoprotein cholesterol levels is several times larger than
the association of low-density lipoprotein cholesterol levels with variants in the
HMGCR gene, and consequently the effect on subsequent outcomes is greater.

2.2 Why use Mendelian randomization?

Although the main reason to use Mendelian randomization is to avoid
the problem of residual confounding, there are additional reasons for using
Mendelian randomization in specific contexts: with case-control data and with
exposures that are difficult to measure.
What is Mendelian randomization? 19

2.2.1 Reverse causation and case-control studies

Reverse causation occurs when an association between the exposure and the
outcome is not due to the exposure causing a change in the outcome, but
the outcome causing a change in the exposure. This could happen if the ex-
posure increased in response to pre-clinical disease, for example from cancer
before it becomes clinically apparent or from atherosclerosis prior to clinical
manifestations of coronary heart disease. As the genotype of an individual is
determined at conception and cannot be changed, there is no possibility of
reverse causation being responsible for an association between genotype and
disease.
For this reason, Mendelian randomization has great strengths in a retro-
spective setting where genetic variants are measured after the disease outcome,
such as in a case-control study. Many exposures of interest cannot be reliably
measured in cases, that is in individuals who have already experienced an
outcome event, as the event may distort the measurement. In this case, the
genetic variant can be used as a proxy for the exposure, and the genetic asso-
ciation with the outcome can be assessed retrospectively. As the genotype of
an individual can be measured in diseased individuals, causal inferences can
be obtained using Mendelian randomization in a case-control setting.

2.2.2 Exposures that are expensive or difficult to measure

Mendelian randomization can be a useful technique when the exposure of in-
terest is expensive or difficult to measure. For example, gold standard assays
for biomarkers such as water-soluble vitamins may cost too much to be af-
fordable for a large sample, or measurement of fasting blood glucose, which
requires overnight fasting, may be impractical. If the genetic variant is asso-
ciated with the exposure (this can be verified in a subsample or a separate
dataset) and is a valid IV for the exposure, a causal relationship between the
exposure and outcome can be inferred from an association between the genetic
variant and the outcome even in the absence of measurement of the exposure.
Additionally, instrumental variable estimates do not attenuate due to clas-
sical measurement error (including within-individual variation) in the expo-
sure [Pierce and VanderWeele, 2012]. This contrasts with observational stud-
ies, in which measurement error in the exposure usually leads to the attenu-
ation of regression coefficients towards the null (known as regression dilution
bias) [Frost and Thompson, 2000].
A further example is where the risk factor is not only difficult to mea-
sure, but also difficult to define. For example, a variant in the IL6R gene
region that is associated with serum interleukin-6 concentrations (as well as
levels of downstream inflammatory markers, including C-reactive protein and
fibrinogen) was shown to be associated with coronary heart disease (CHD)
risk [Swerdlow et al., 2012]. However, from knowledge about the functional
role of the variant, the causal effect assessed is not thought to operate through
20 Mendelian Randomization

elevated serum interleukin-6 concentrations, but rather through changes in sig-

nalling in interleukin-6 receptor pathways. This is a cellular phenotype which
varies over time, and so a representative measurement for an individual is not
straightforward to define. However, as the genetic variant can be measured,
the causal role of interleukin-6 receptor-related pathways on CHD risk can be
assessed by Mendelian randomization [Sarwar et al., 2012].

2.3 A brief overview of genetics

In order to understand Mendelian randomization, it is necessary to have at
least a cursory understanding of genetics. We here provide a brief overview
of genetics, only covering the information necessary to understand Mendelian
randomization. A glossary of genetic terminology, adapted from a Mendelian
randomization review paper [Lawlor et al., 2008] is provided in Table 2.2. Fur-
ther information on genetic terminology related to Mendelian randomization
can be found in other papers [Davey Smith and Ebrahim, 2003; Sheehan et al.,
2008].

2.3.1 Reading the genetic code

The genetic information (or genome) of many living organisms consists of
long strings of genetic code in the form of DNA (deoxyribonucleic acid), the
molecule that encodes life, packaged up into chromosomes. Humans have 23
pairs of chromosomes, with one chromosome in each of the pairs coming from
the mother and one from the father. Chromosomes contain genes, which are
locatable regions of the genetic code that encode a unit of heritable infor-
mation. Not all of the genetic sequence falls into a gene region, and much of
the chromosome consists of intermediate genetic material known as noncoding
DNA.
A single chromosome has two strands, each consisting of a sequence of
nucleotide bases which can represented by letters. There are four possible
nucleotide bases (adenine, thymine, cytoside and guanine) represented by the
letters A, T, C and G. These nucleotide bases pair up in such a way that the
strands contain complementary sequences. Wherever the first strand has A,
the other will have T – and vice versa. Wherever the first strand has C, the
other will have G – and vice versa. In this way, each of the strands contains
the same information, and so only one of the strands is considered. Suppose
that a chromosome at a given locus (position) in the DNA sequence on one of
its strands reads:
What is Mendelian randomization? 21

• Alleles are the variant forms of a single nucleotide polymorphism (SNP). For
a diallelic SNP where there are two possible alleles, the more common allele is
called the major allele or wildtype allele, and the less common allele is the minor
allele or variant allele.
• Canalization (also known as developmental compensation) is the process by
which potentially disruptive influences on normal development from genetic (and
environmental) variation are damped or buffered by compensatory processes.
• A chromosome carries a collection of genes located on a long string of DNA.
Humans have 22 pairs of autosomal (non-sex) chromosomes and 1 pair of sex
chromosomes.
• A copy number variant (or variation) is a (possibly) repeating section of DNA
where the number of copies of the section varies between individuals.
• DNA (deoxyribonucleic acid) is a molecule that contains the genetic instruc-
tions used in the development and functioning of all living organisms. The main
role of DNA is the long-term storage of information. It contains the instructions
needed to construct other components of cells, including proteins and ribonucleic
acid (RNA) molecules. DNA has four nucleotide bases labelled A, T, C and G.
• A gene is a section of a chromosome comprising DNA which encodes infor-
mation relevant to the function of an organism.
• The genotype of an individual at a particular locus refers to the two alleles at
that locus. If the alleles are the same, the genotype is homozygous; if different,
heterozygous.
• A haplotype describes a particular combination of alleles from linked loci found
on a single chromosome.
• Linkage disequilibrium (LD) is the correlation between allelic states at dif-
ferent loci within the population. The term LD describes a state that represents
a departure from the hypothetical situation in which all loci exhibit complete
independence (linkage equilibrium).
• A locus (plural: loci ) is the position in a DNA sequence and can be a SNP, a
region of DNA sequence, or a whole gene.
• Meiosis is the process of cell division leading to gametes (sex cells) which
contain half of the genetic material from the original cell.
• Pleiotropy is the potential for genes or genetic variants to have more than one
independent phenotypic effect.
• Polymorphism is the existence of two or more variants at a locus. The term
polymorphism is usually restricted to moderately common genetic variants, with
at least two alleles having frequencies of greater than 1% in the population. A
less common variant allele is called a mutation.
• Single nucleotide polymorphisms (SNPs) are genetic variations in which one
base in the DNA is altered, for example a T instead of an A.

TABLE 2.2
A glossary of genetic terminology, adapted from Lawlor et al., 2008.
22 Mendelian Randomization

...ATTACGCTTCCGAGCTTCCGCAG...
and that same locus on the paired chromosome reads:
...ATTACGCCTCCGAGCTTCCGCAG...
The underlined nucleotide represents a nucleotide at a particular locus that is
polymorphic: it exists in various forms. All individuals contain many genetic
mutations, where the DNA code has changed from that generally seen in the
population. A single nucleotide polymorphism (SNP) is a mutation where a
single nucleotide base at a particular locus has been replaced with a differ-
ent nucleotide. The different possible nucleotides which may appear at each
locus are known as alleles. For example, at the highlighted locus above, one
chromosome has the letter T, and the other has the letter C: so T and C are
alleles of this particular SNP. If these are the only two possibilities, this is a
diallelic SNP; triallelic and quadrallelic SNPs are far less common, but have
also been observed.
For a diallelic SNP, it is conventional to denote the more common allele,
known as the wildtype or major allele, by an upper case letter (for example,
A) and the less common allele, the variant or minor allele, by a lower case
letter (for example, a). The choice of letter is arbitrary; there is no connection
between the letter A commonly used for the first variant considered, and the
nucleotide base adenine represented by letter A. The proportion of minor
alleles in a population for a given SNP is called the ‘minor allele frequency’.
Although some genetic mutations seem to be specific to particular individuals,
others are more widespread, showing up in a substantial proportion of the
population. SNPs occur on average about once in every 300 nucleotides along
the genome, and extensive catalogues of SNPs have been compiled.
As people have two copies of each chromosome (one from each parent),
individuals can be categorized for each diallelic SNP into three possible sub-
groups corresponding to their combination of alleles (their genotype). These
subgroups are the major homozygotes (AA), heterozygotes (Aa) and minor
homozygotes (aa). We shall denote these subgroups as 0, 1 and 2, correspond-
ing to the number of minor alleles for that SNP. For a more complicated
genetic variant, such as a triallelic SNP where there are three possible alleles
at one locus, there is no natural ordering of the six possible subgroups given
by the SNP.
When multiple SNPs on a single chromosome are considered, the combi-
nation of alleles on each of the chromosomes is known as a haplotype. For
example, if an individual has one chromosome reading:
...GCACCTTAC...GTAGAATC...TCAACTGTCAT
and the other reading:
...GCACCGTAC...GTAAAATC...TCAACTGTCAT
then the individual is a heterozygote for the first two SNPs, and a homozygote
for the final SNP. The haplotypes are TGT and GAT. One of these haplotypes
What is Mendelian randomization? 23

is inherited from each of the individual’s parents. As a haplotype is a series of

alleles on the same chromosome, haplotype patterns, especially for SNPs that
are physically close together, are often inherited together. This means that ge-
netic variants are not always independently distributed. Using patterns which
have been observed in a large number of individuals, haplotypes can sometimes
be inferred from SNP data using computer software, as generally not all possi-
ble combinations of alleles will be present on a chromosome in a population. In
some cases, haplotypes can be determined uniquely from SNP data, whereas
in other cases, there is uncertainty in this determination. If the SNPs satisfy
the IV assumptions, then the haplotypes will also satisfy the IV assumptions.
Other patterns of genetic variation can also be used as IVs, such as copy
number variations where a section of genetic material is repeated a variable
number of times. Generally, throughout this book we shall assume that IVs
are diallelic SNPs, although the majority of methods and findings discussed
will apply similarly in other cases. SNPs are given numbers by which they can
be uniquely referenced. Reference numbers begin “rs” (standing for “reference
SNP”), such as rs1205.

2.3.2 Using a genetic variant as an instrumental variable

The use of any particular genetic variant as an IV requires caution as the IV
assumptions cannot be fully tested and may be violated for various epidemi-
ological and biological reasons (see Chapter 3). As a plausible example of a
valid genetic IV, in the Japanese population, a common genetic mutation in
the ALDH2 gene affects the processing of alcohol, causing excess production of
a carcinogenic by-product, acetaldehyde, as well as nausea and headaches. We
can use this genetic variant as an IV to assess the causal relationship between
alcohol consumption and oesophageal cancer. Here, alcohol consumption is
the exposure and oesophageal cancer the outcome.
Assessment of the causal relationship using classical epidemiological stud-
ies is hindered by the strong association between alcohol and tobacco smoking,
another risk factor for oesophageal cancer [Davey Smith and Ebrahim, 2004].
Individuals with two copies of the ALDH2 polymorphism tend to avoid alco-
hol, due to the severity of the short-term symptoms. Their risk of developing
oesophageal cancer is one-third of the risk of those with no copies of the muta-
tion [Lewis and Davey Smith, 2005]. Carriers of a single copy of this mutation
exhibit only a mild intolerance to alcohol. They are still able to drink, but
they cannot process the alcohol efficiently and have an increased exposure to
acetaldehyde. Carriers of a single mutated allele are at three times the risk
of developing oesophageal cancer compared to those without the mutation,
with up to 12 times the risk in studies of heavy drinkers. This is an exam-
ple of a gene–environment interaction (here between the genotype and alcohol
consumption). The conclusion is that alcohol consumption causes oesophageal
cancer, since there is no association between this genetic variant and many
other risk factors, and any single risk factor would have to have a massive
24 Mendelian Randomization

Genetic Effect on alcohol metabolism Genetic association

subgroup with oesophageal
cancer
Major No effect – can metabolize alcohol (Reference group)
homozygotes
Heterozygotes Mild effect – individuals can drink, Increased disease risk
but alcohol stays in bloodstream for
longer
Minor Severe effect – cannot metabolize Decreased disease risk
homozygotes alcohol, individuals tend to abstain
from alcohol

TABLE 2.3
Example: alcohol intake and the ALDH2 polymorphism in the Japanese pop-
ulation.

effect on oesophageal cancer risk as well as a strong association with the ge-
netic variant to provide an alternative explanation for these results.
These associations are summarized in Table 2.3. The genetic mutation
provides a fair test to compare three populations who differ systematically
only in their consumption of alcohol and exposure to acetaldehyde, and who
have vastly differing risks of the outcome. The evidence for a causal link be-
tween alcohol consumption, exposure to acetaldehyde and oesophageal cancer
is compelling [Schatzkin et al., 2009]. However, in other cases, particularly if
the genetic variant(s) do not explain much of the variation in the exposure, the
power to detect a causal effect may be insufficient to provide such a convincing
conclusion.

2.4 Summary
Mendelian randomization has the potential to be a useful tool in a range
of scientific contexts to investigate claims of causal relationships. It must be
applied with care, as its causal claims come at the price of assumptions which
are not empirically testable. Its methods must be refined, as often data on
multiple genetic variants or data taken from several study populations are
required to achieve meaningful findings. But, when properly used, it gives an
insight into the underlying causal relationships between variables which few
other approaches can rival.
3
Assumptions for causal inference

In the previous chapters, we repeatedly used the word ‘causal’ to describe the
inferences obtained by Mendelian randomization. In this chapter, we clarify
what is meant by the causal effect of an exposure on an outcome. We give
a more detailed explanation of the theory of instrumental variables, and ex-
plain in biological terms various situations that may lead to violations of the
instrumental variable assumptions and thus misleading causal inferences. We
conclude by discussing the difference between testing for the presence of a
causal relationship and estimating a causal effect, and the additional assump-
tions necessary for causal effect estimation.

3.1 Observational and causal relationships

As the saying goes, “association is not causation” or in its more widely quoted
form “correlation does not imply causation”. Naive interpretation of an ob-
served relationship between two variables as causal is a well-known logical
fallacy. However, precise definitions of causality which correspond to our in-
tuitive understanding have eluded philosophers for centuries [Pearl, 2000a].
Definitions are also complicated by the fact that, in many epidemiological
contexts, causation is probabilistic rather than deterministic: for example,
smoking does not always lead to lung cancer.

3.1.1 Causation as the result of manipulation

The fundamental concept in thinking about causal relationships is the idea of
intervention on, or manipulation of, a variable. This is often cited as “no cau-
sation without manipulation”, reflecting that direct experimentation is neces-
sary to demonstrate a causal effect [Holland, 1986]. A causal effect is present
if the outcome is different when the exposure is set to two different levels.
This differs from an observational association, which represents the difference
in the outcome when the exposure is observed at two different levels. If there
are variables which are correlated with the exposure, the observational asso-
ciation reflects differences not only in the exposure of interest, but also in the

25
26 Mendelian Randomization

variables correlated with the exposure. With the causal effect, setting the value
of the exposure only alters the exposure and variables on the causal pathway
downstream of the exposure, not variables on alternative causal pathways.
The outcome variable Y for different observed values x of the exposure X
is written as Y |X = x, read as Y conditional on X equalling x. Causal effects
cannot be expressed in terms of probability distributions and so additional
notation is required [Pearl, 2010]. The outcome variable Y when the exposure
X is set to a given value x is written as Y |do(X = x), where the do operator
indicates that the variable is manipulated to be set to the given value.

3.1.2 Causation as a counterfactual contrast

One common definition of a causal effect is that of a counterfactual contrast
[Maldonado and Greenland, 2002]. Counterfactual, literally meaning counter
or contrary to fact, refers to a potential situation which could have happened,
but did not [Greenland, 2000b]. For example, in the morning, Adam has a
headache. He may or may not take an aspirin tablet. At the point of decision,
we can conceive that there are two potential universes where Adam makes
different choices about whether to take the aspirin or not. Associated with each
universe is a potential outcome – does he still have a headache that afternoon?
Once he has made this decision, one of these universes and outcomes becomes
counterfactual; both outcomes cannot be observed. A causal effect is present
if the two outcomes are different; if he still had a headache in the universe
where he did not take the aspirin, but did not have a headache in the universe
where he did take the aspirin, then the aspirin has caused the alleviation of
the headache. With a probabilistic interpretation, assuming that the outcome
is stochastic rather than deterministic, if the probability that he would still
have a headache is lower in the aspirin universe than in the no-aspirin universe,
then taking aspirin has a causal effect on alleviating the headache.
There are several conceptual difficulties with the counterfactual approach
[Dawid, 2000]. The main difficulty is that the causal effect of an exposure for
an individual can never be measured, as at least one of the two outcomes in
the causal contrast is always unobserved. This is referred to as the ‘fundamen-
tal problem of causal inference’ [Holland, 1986]. It means that a counterfac-
tual causal estimate is not the answer to any real experiment that could be
conducted, but the answer to a hypothetical experiment requiring two parallel
universes. However, the counterfactual approach has many appealing features.
Chiefly, it gives a precise framework for defining causal effects, aiding both
informal and mathematical thinking about causal relationships.
In terms of notation, the potential outcomes Y |do(X = x) that the out-
come variable can take are written as Y (x). If the exposure is binary, the two
potential outcomes for an individual are Y (1) and Y (0), and the causal effect
of increasing X from X = 0 to X = 1 is Y (1) − Y (0).
Assumptions for causal inference 27

3.1.3 Causation using graphical models

Graphical models, and in particular directed acyclic graphs, can provide a
helpful way of thinking about and expressing causal relationships. A graphical
model comprises a set of nodes representing variables, and arrows representing
causal effects. An arrow from variable A to variable B indicates that there is
a causal effect of A on B. A graphical model need not contain all intermediate
variables (such as C if A → C → B), but must contain all common causes of
variables included in the graph (such as D if A ← D → B). Relations between
variables are expressed by directed arrows, indicating a (direct) causal effect
(conditional dependence), or without an arrow, indicating no direct effect
(conditional independence). A direct causal effect is only ‘direct’ with respect
to the variables included in the graph and as such is not direct in an absolute
sense, but could act via an intermediate variable. A directed acyclic graph
(DAG) is a graph that does not contain any complete cycles, such as A ↔ B
or A → B → C → A; a cycle would imply that a variable is its own cause.
As an example, Figure 3.1 shows the instrumental variable (IV) assump-
tions (Section 2.1.2) in the form of a graph. To simplify the graph, all con-
founding variables are subsumed into a single ‘confounder’, which has effects
on both the exposure and outcome. We see that there are arrows from the
IV to the exposure (assumption i.), from the exposure to the outcome, and
from the confounder to the exposure and to the outcome. Just as importantly,
there is no pathway between the IV and the confounder (assumption ii.) and
no pathway from the IV to the outcome apart from that passing through
the exposure (assumption iii.), indicating that a hypothetical intervention to
change the value of the IV without varying the exposure or the confounder
would not affect the outcome.
A pathway does not necessarily mean a route consisting only of directed
arrows. For there to be no pathway from the IV G to the outcome Y (except via
the exposure), there cannot be a sequence consisting of chains (G → C → Y )
or forks (G ← D → Y ) of variables not including the exposure. There can
be inverted forks (G → E ← Y ), provided neither E nor a descendent of E
is adjusted for in the analysis (E may be referred to as a collider). In these
examples, C may represent a competing risk factor to the exposure, and D may
represent a selection variable, such as ethnicity, which must be accounted for
in the analysis to prevent bias due to population stratification (Section 3.2.5).
Formally, the genetic variants and outcome must be d-separated by the risk
factor and confounders [Geiger et al., 1990].

3.1.4 Causation based on multivariable adjustment

Multivariable adjustment is often undertaken in the analysis of observational
data in order to try to account for confounding. A set of covariates which, if
known and conditioned on, would give an estimate of association equal to the
causal effect, is referred to as ‘sufficient’. Assuming that a set of covariates
28 Mendelian Randomization

FIGURE 3.1
Directed acyclic graph illustrating instrumental variable (IV) assumptions.

is sufficient is necessary to interpret the result of a multivariable-adjusted

regression analysis as a causal effect. On conditioning for a sufficient set of
covariates, the counterfactual outcomes at different values of the exposure
should be independent of the exposure, a property known as “conditional
exchangeability” [Greenland and Robins, 1986].
If the causal relationships between all the variables in a model representing
the generating mechanism for observational data were known, a set of covari-
ates can be assessed as sufficient or otherwise using the “back-door criterion”
[Pearl, 2000b]. For simple causal networks, a set of covariates is sufficient if
it includes all common causes of the exposure and the outcome and does not
include variables on the causal pathway from the exposure to the outcome,
nor common effects of the exposure and outcome. In practice, neither the un-
derlying network of associations between variables nor the sets of all common
causes and all common effects of exposure and outcome are known, and so
the use of multivariable-adjusted regression analyses to assess causal effects
is unreliable. It is not possible to know if adjustment for a sufficient set of
covariates has been made, or if there is residual confounding due to unmea-
sured covariates, or if the set of covariates includes variables on the causal
pathway between the exposure and outcome, whose inclusion in a regression
model also biases regression coefficients. This highlights the need to consider
other methods for assessing causal relationships.

3.2 Finding a valid instrumental variable

Instrumental variable (IV) techniques represent one of the few ways available
for estimating causal effects without complete knowledge of all confounders
Assumptions for causal inference 29

of the exposure–outcome association. We continue by recalling and discussing

the properties of an IV, and how the IV assumptions may be violated in
practice.

3.2.1 Instrumental variable assumptions

In order for a genetic variant to be used to estimate a causal effect, it must
satisfy the assumptions of an instrumental variable (Section 2.1.2), which we
repeat here:
i. the variant is associated with the exposure,
ii. the variant is not associated with any confounder of the exposure–outcome
association,
iii. the variant does not affect the outcome, except possibly via its association
with the exposure.
These conditions can be understood intuitively. The first assumption guar-
antees that genetic subgroups defined by the variant will have different average
levels of the exposure. This ensures that there is a systematic difference be-
tween the subgroups. If the genetic variant is not strongly associated with
the exposure (in the sense of its statistical strength of association), then it is
referred to as a weak instrument (see Chapter 7). A weak instrument differs
from an invalid instrument in that a weak instrument can be made stronger
by collecting more data. If a single genetic variant is a weak instrument, then
it will still give a valid test of the null hypothesis of no causal effect, but the
power to detect a true causal effect may be low. However, combining multiple
weak instruments in an analysis model to obtain a single effect estimate can
lead to misleading inferences.
The second assumption can be understood as ensuring that the compari-
son between the genetic subgroups is a fair test, that is, all other variables are
distributed equally between the subgroups. The third assumption is often ex-
pressed using the concept of conditional independence as “the genetic variant
is not associated with the outcome conditional on the value of the exposure
and confounders of the exposure–outcome association”. It ensures that the
only causal pathway(s) from the genetic variant to the outcome are via the
exposure. This means that the genetic variant is not directly associated with
the outcome, nor is there any alternative pathway by which the variant is
associated with the outcome other than that through the exposure.

3.2.2 Validity of the IV assumptions

The counterfactual framework for causation helps understanding of when and
why a randomized controlled trial (RCT) can estimate a causal effect – thought
of in a counterfactual sense as a contrast between parallel universes. The ran-
domized subgroups in an RCT can be regarded as exchangeable. This means
30 Mendelian Randomization

that the same distribution of outcomes would be expected if each of the sub-
groups were exposed to the treatment or the control regime. Although an in-
dividual can only be exposed to one of the two treatment regimes (and so only
observed in one universe), by exposing each subgroup to a different treatment
regime, in effect we observe the population in each of the two counterfactual
parallel universes, and the average outcomes in each of the universes (sub-
groups) can be compared [Greenland and Robins, 1986]. A causal effect can
be consistently estimated which represents the average effect of being assigned
to the treatment group as opposed to the control group. This means that an
RCT can estimate an average causal effect for the population as the contrast
between the average levels of the outcome in the randomized subgroups of
the population (which will have the same characteristics on average as the
overall population due to the random assignment into subgroups). An indi-
vidual causal effect cannot be estimated, as an individual cannot in general
be subjected to both the treatment and control regimes [Rubin, 1974].
For Mendelian randomization, the similar key property of an IV is that the
division of the population into genetic subgroups is independent of competing
risk factors, and so genetic subgroups defined by the IV are exchangeable. For
a genetic variant to be an IV, it is necessary that same distribution of out-
comes would be observed if individuals with no copies of the genetic variant
instead had one copy of the genetic variant (and the exposure distributions
were unchanged), and vice versa. However, empirical testing of the exchange-
ability criterion is not possible.
We return to the question of assessing the validity of genetic variants as
IVs later in this section; firstly we consider reasons why a genetic variant may
not be a valid IV. These include issues of biological mechanism, genetic co-
inheritance, and population effects. Invalid IVs lead to unreliable inferences
for the causal effect of an exposure. The situations discussed here represent
potential lack of internal validity of estimates; the question of the external
validity of an IV estimate as an estimate of the effect of a clinical intervention
is discussed in Chapter 6.

3.2.3 Violations of IV assumptions: biological mechanisms

The first category of ways that we consider by which the IV assumptions may
be violated is because of an underlying biological mechanism.
Pleiotropy: Pleiotropy refers to a genetic variant being associated with
multiple risk factors. If a genetic variant used as an IV is additionally associ-
ated with another risk factor for the outcome, then either the second or the
third IV assumption is violated (depending on whether the risk factor is a
confounder of the exposure–outcome association or not), and the variant is
not a valid IV.
If the genetic variant is associated with an additional variable solely due
to mediation of the genetic association via the exposure of interest (some-
times called vertical pleiotropy), that this is not regarded as pleiotropy for our
Assumptions for causal inference 31

purposes. For example, the FTO gene is a determinant of satiety (how full of
food a person feels) [Wardle et al., 2008]. If satiety affects body mass index
(BMI), then a variant in the FTO gene can be used as an IV for BMI if the
two variables are on the same causal pathway, and if there is no alternative
causal pathway from the genetic variant to the outcome not via BMI. How-
ever, if the FTO gene was also associated with (say) blood pressure, and this
association was not completely mediated by the association of the gene with
BMI, then it would be misleading to use a variant in the FTO gene to make
specific inferences about the causal effect of BMI on an outcome.
Concerns about pleiotropy can be alleviated by using genetic variants lo-
cated in genes, the biological function of which are well-understood. For ex-
ample, for C-reactive protein (CRP), we can use genetic variants in the CRP
gene which are known to have functional relevance in the regulation of CRP
levels. Associations of a variant with measured covariates can be assessed to
investigate potential pleiotropy, although such associations may also reflect
mediation, particularly if the associations are consistent across independent
variants.
Canalization: Canalization, or developmental compensation, is the phe-
nomenon by which an individual adapts in response to genetic change in such
a way that the expected effect of the change is reduced or absent [Debat and
David, 2001]. It is most evident in knockout studies, where a gene is rendered
completely inactive in an organism, typically a mouse. Often the organism de-
velops a compensatory mechanism to allow for the missing gene such that the
functionality of the gene is expressed via a different biological pathway. This
buffering of the genetic effect may have downstream effects on other variables.
Canalization may be a problem in Mendelian randomization if groups with dif-
ferent levels of the genetic variants differ with respect not only to the exposure
of interest, but also to other risk factors via a canalization mechanism.
In a sense, canalization is not a violation of the IV assumptions, but merely
an (often unwanted) consequence. Canalization is the same process as that
assessed by Mendelian randomization, as any change in other risk factors
from canalization occurs as a causal effect of the genetic variant. However,
the aim of Mendelian randomization is not simply to describe the effects of
genetic change, but to assess the causal effect of the (non-genetic) exposure.
If there is substantial canalization, Mendelian randomization estimates may
be unrepresentative of clinical interventions on the exposure performed in a
mature cohort.

3.2.4 Violations of IV assumptions: non-Mendelian

inheritance
The second category of ways that we consider by which the IV assumptions
may be violated is because of non-Mendelian inheritance. Although Mendelian
principles state that separate characteristics are inherited separately, this is
not always true in practice. Non-Mendelian inheritance refers to patterns of
32 Mendelian Randomization

Confounder

Measured variant

Causal variant Exposure Outcome

FIGURE 3.2
Graph of instrumental variable assumptions where a variant in linkage dise-
quilibrium with the causal variant has been measured. Such a variant would
still be a valid instrumental variable. The dashed line connecting the genetic
variants indicates correlation without a causal interpretation.

inheritance which do not correspond to Mendel’s laws, specifically the law of

independent assortment.
Linkage disequilibrium: One particular reason for genetic variants to
be inherited together is the physical proximity of the variants on the same
chromosome. Variants whose distributions are correlated are said to be in
linkage disequilibrium (LD). The opposite of LD is linkage equilibrium.
LD has both desirable and undesirable consequences. If genetic variants
were truly independently distributed, then only the genetic variant which was
causally responsible for variation in the exposure could be used as an IV, as all
other genetic variants would not be associated with the exposure. In reality, it
is not necessary for the genetic variant used as the IV to be the causal variant,
merely to be correlated with the causal variant [Hernán and Robins, 2006].
This is because an IV must simply divide the population into subgroups which
differ systematically only with respect to the exposure. This is illustrated in
Figure 3.2.
An undesirable consequence of LD is that genetic variants correlated with
the variant used in the analysis may have effects on competing risk factors.
This would lead to the violation of the second or the third IV assumption
(similar to violations due to pleiotropy). Concerns about invalid inferences
due to LD can be alleviated by empirical testing of the association of known
potential confounders with the measured variant.
Effect modification: Effect modification is a separate phenomenon from
confounding, and relates to a statistical interaction between the effect of a
variable (usually an effect of the exposure) and the value of a covariate, lead-
ing to the causal effect of the exposure varying across strata defined by the
covariate. Factors that may lead to effect modification include (but are not
limited to) issues of non-Mendelian inheritance, such epigenetic variation [Og-
buanu et al., 2009] and parent-of-origin effects [Bochud et al., 2008].
Assumptions for causal inference 33

Effect modification alone is unlikely to represent a violation of the IV

assumptions; however, it may lead to difficulties in interpreting Mendelian
randomization investigations. Taking the example from Section 2.3.2 of the
effect of alcohol intake on oesophageal cancer risk, in the Japanese population
only men tend to drink alcohol. Hence, genetic associations with the outcome
may be observed only in men and may not be present in women. If there are
biological reasons for genetic associations to be stronger or weaker (or even
absent) in some strata of the population, then associations measured in that
stratum of the population would not be representative of the effect in the
population as a whole. However, this may also provide an opportunity for
verifying the IV assumptions; Japanese women are a natural control group for
Japanese men. If the same genetic associations of alcohol-related variants with
oesophageal cancer risk seen in Japanese men are not observed in Japanese
women, this provides further evidence that the genetic associations with dis-
ease risk are driven by alcohol consumption, and not by violations of the IV
assumptions.

3.2.5 Violations of IV assumptions: population effects

The final category of ways that we consider by which the IV assumptions may
be violated is because of population effects.
Population stratification: Population stratification occurs when the
population under investigation can be divided into distinct subpopulations.
This may occur, for example, when the population is a mixture of individuals
of different ethnic origins. If the frequency of the genetic variant and the distri-
bution of the exposure are different in the different subpopulations, a spurious
association between the variant and the exposure will be induced which is due
to subpopulation differences, not the effect of the genetic variant. Violations
of the IV assumptions may also occur if there is continuous variation in the
structure of the population rather than distinct subpopulations.
Concerns about population stratification can be alleviated by restricting
the study population to those with the same ethnic background (although
there may still may be differences associated with ancestry in broadly-defined
ethnic groups). In a genome-wide association study (GWAS), genomic control
approaches, such as adjustment for genetic principal components, are possible.
However, the use of Mendelian randomization in a population with a large
amount of genetic heterogeneity is not advised.
Ascertainment effects: If the genetic variant is associated with recruit-
ment into the study, then the relative proportions of individuals in each genetic
subgroup are not the same as those in the population, and so a genetic as-
sociation with the outcome in the sample may not be present in the original
population. If the study cohort is taken from the general population, ascer-
tainment effects are unlikely to be a major problem in practice. However, if,
for example, the study cohort is pregnant mothers, and the genetic variant is
associated with fertility, then the distributions of the covariates in the genetic
34 Mendelian Randomization

subgroups will differ and not be the same as those in the general population.
This may introduce bias in the estimation of causal effects, as there is a path-
way opened up from the genetic variant to the outcome by conditioning on
a common cause of the variant and the outcome (sometimes called collider
bias).
This would also be a problem in studies looking at genetic associations
in populations of diseased individuals, such as clinical trials of secondary dis-
ease prevention. Individuals with greater genetically determined disease risk
are less likely to survive to study recruitment, and so the randomization of
individuals into genetic subgroups at conception would not hold in the study
population, leading to biased genetic associations.

3.2.6 Statistical assessment of the IV assumptions

Although it is not possible to demonstrate conclusively the validity of the IV
assumptions, several tests and assessments are possible to increase or decrease
confidence in the use of genetic variants as IVs.
The simplest assessment of instrument validity is to test the association
between the genetic variant and known confounders. Association of the variant
with a covariate associated with the outcome which is not on the causal path-
way between the exposure and outcome would violate the second IV assump-
tion. However, there is no definitive way to tell whether the association with
the covariate is due to violation of the IV assumptions (such as by pleiotropy or
linkage disequilibrium), or due to mediation through the exposure of interest.
Additionally, there is no way of testing whether or not the variant is associ-
ated with an unmeasured confounder. If there are multiple covariates and/or
genetic variants, then any hypothesis testing approach needs to account for
the multiple comparisons of each covariate, leading to a lack of power to detect
any specific association. Additionally, as several covariates may be correlated,
a simple Bonferroni correction may be an over-correction. A sensible way to
proceed is to combine a hypothesis testing approach with a quantitative and
qualitative assessment of the imbalance of the covariates between genetic sub-
groups and the degree to which this may bias the IV estimate.
A further approach for testing instrument validity is to see whether the
association of a genetic variant with the outcome attenuates on adjustment
for the risk factor [Glymour et al., 2012]. Although attenuation may not be
complete even when the instrumental variable assumptions are satisfied for the
risk factor due to confounding and measurement error [Didelez and Sheehan,
2007], if the attenuation is not substantial then the risk factor is unlikely to
be on the causal pathway from the variant to the outcome.
If multiple genetic variants are available, each of which is a valid IV,
then a separate IV estimate can be calculated using each of the instruments
in turn. Assuming that each variant affects the exposure in a similar way,
even if the genetic associations with the exposure are of different magnitude,
the separate IV estimates should be similar, as they are targeting the same
Assumptions for causal inference 35

quantity. This can be assessed graphically by plotting the genetic associations

for an additional variant allele with the exposure and outcome for multiple
variants: a straight line through the origin is expected, as in Figure 6.1. For-
mally, heterogeneity between variants can be tested using an overidentification
test (Section 4.5.3). Failure of an overidentification test may be due to one
or more of the IVs being invalid. However, the power of such tests may be
limited in practice, and so testing should not be relied on for justification of
the IV assumptions.
Other mathematical results for testing IV validity are available [Glymour
et al., 2012], but these are only likely to detect gross violations of the IV
assumptions. Biological knowledge rather than statistical testing should form
the backbone of any justification of the use of a particular genetic variant as an
IV in Mendelian randomization. The Bradford Hill criteria form a systematic
summary of common-sense principles for assessing causality in epidemiolog-
ical investigations [Hill, 1965]. In Table 3.1, we apply the relevant Bradford
Hill criteria for causation to Mendelian randomization as a checklist to judge
whether the validity of genetic variant(s) as an IV is plausible.

3.2.7 Summary of issues relating to IV validity

The validity of IVs is of vital importance to Mendelian randomization. It is our
view that the choice of genetic variants as IVs should be justified mainly by
basic biological knowledge but can be verified by empirical statistical testing.
Appropriate caution should be attached to the interpretation of Mendelian
randomization findings depending on the plausibility of the IV assumptions,
and particularly to those where the justification of the IV assumptions is
mainly empirical. This suggests that variants from candidate gene investiga-
tions, where the function of the genetic variant(s) is well-understood, will have
more credibility for use in Mendelian randomization studies than variants out-
side of gene coding regions, such as those discovered in genome-wide associa-
tion studies. However, it should be remembered that all statistical methods for
assessing causal effects rely on untestable assumptions, and as such, Mendelian
randomization has an important role in building up the case for the causal
nature of a given exposure even if the validity of the IV assumptions can be
challenged.
On a more positive note, a British study into the distribution of genetic
variants and non-genetic factors (such as environmental exposures) in a group
of blood donors and a representative sample from the population showed
marked differences in the non-genetic factors, but no more difference than
would be expected by chance in the genetic factors [Ebrahim and Davey Smith,
2008], indicating that genetic factors seem to be distributed independently of
possible confounders in the population of the United Kingdom [Davey Smith,
2011]. This gives plausibility to the general suitability of genetic variants as
IVs, but in each specific case, justification of the assumptions relies on biolog-
ical knowledge about the genetic variants in question.
36 Mendelian Randomization

• Strength: If a genetic association with the outcome is slight, then the

association could be explained by only a small imbalance in a covariate
associated with the genetic variant. A small violation of the instrumental
variable assumptions is less likely to be detected by testing the association
of the variant with known covariates.
• Consistency: A causal relationship is more plausible if multiple genetic
variants associated with the same exposure are all concordantly associated
with the outcome, especially if the variants are located in different gene
regions and/or have different mechanisms of association with the outcome.

• Biological gradient: Further, a causal relationship is more plausible if

the genetic associations with the outcome and with the exposure for each
variant are proportional (for example, as in Figure 6.1).
• Specificity: A causal relationship is more plausible if the genetic vari-
ant(s) are associated with a specific risk factor and outcome, and do not
have associations with a wide range of covariates and outcomes. A specific
association is most likely if the genetic variant(s) are biologically proxi-
mal to the exposure, and not biologically distant. This is most likely for
risk factors that are biomarkers (such as C-reactive protein and low-density
lipoprotein cholesterol), rather than generic risk factors (such as body mass
index and blood pressure).
• Plausibility: If the function of the genetic variant(s) is known, a causal
relationship is more plausible if the mechanism by which the variant acts
is credibly and specifically related to the exposure.
• Coherence: If an intervention on the exposure has been performed (for
example, if a drug has been developed that acts on the exposure), associa-
tions with intermediate outcomes (covariates) observed in the experimental
context should also be present in the genetic context; directionally con-
cordant genetic associations should be observed with the same covariates.
For example, associations of genetic variants in the IL6R gene region with
C-reactive protein and fibrinogen should be similar to those observed for
tocilizumab, an interleukin-6 receptor inhibitor [Swerdlow et al., 2012].

TABLE 3.1
Bradford Hill criteria applied to Mendelian randomization for judging the
biological plausibility of a genetic variant as an instrumental variable.
Assumptions for causal inference 37

3.2.8* Definition of an IV as a random variable

For the more mathematically inclined, we give a further characterization of an
IV in terms of random variables. We assume that we have an outcome Y that
is a function of a measured exposure X and an unmeasured confounder U ;
that the confounding factors can be summarized by a single random variable U
[Palmer et al., 2008], which satisfies the requirements of a sufficient covariate
(Section 3.1.4); and that the exposure X can be expressed as a function of the
confounder U and the genetic variant G. G may be a single genetic variant
or a matrix corresponding to several genetic variants. The IV assumptions of
Section 3.2.1 are rewritten here in terms of random variables:
i. G is not independent of X (G 6⊥
⊥ X),
ii. G is independent of U (G ⊥
⊥ U ),
iii. G is independent of Y conditional on X and U (G ⊥
⊥ Y |X, U ).
This implies that the joint distribution of Y, X, U, G factorizes as

p(y, x, u, g) = p(y|u, x)p(x|u, g)p(u)p(g) (3.1)

which corresponds to the directed acyclic graph (DAG) in Figure 3.3 [Dawid,
2002; Didelez and Sheehan, 2007].
It is a common mistake to think that the third IV assumption should read
not G ⊥ ⊥ Y |X, U , but G ⊥
⊥ Y |X, that is conditioning on U is not necessary.
As X is a common descendent of G and U , conditioning on X induces an
association between G and U , and therefore between G and Y . For example,
if X and U are positively correlated and both have positive causal effects on
Y , then conditional on X taking a value around the middle of its distribution,
a large value of Y is associated with a low value of G. This is because the large
value of Y is associated with a large value of U , and so G is more likely to be

G X Y
FIGURE 3.3
Directed acyclic graph of Mendelian randomization assumptions as random
variables.
38 Mendelian Randomization

low so that the value of X is moderate and not large. The lack of independence
(G 6⊥
⊥ Y |X) means that, in the regression of Y on X and G, the coefficient
for G will generally be close to, but not equal to zero in a large sample (and
especially if X is measured with error).
In order to interpret the unconfounded estimates produced by IV analysis
as causal estimates, we require the additional structural assumption:

p(y, u, g, x|do(X = x0 )) = p(y|u, x0 )1(X = x0 )p(u)p(g) (3.2)

where 1(.) is the indicator function. This ensures that intervening on X does
not affect the distributions of any other variables except the conditional dis-
tribution of Y [Didelez et al., 2010].

3.2.9* Definition of an IV in potential outcomes

In the “potential outcomes” or counterfactual causal framework (Sec-
tion 3.1.2), a set of outcomes Y (x), x ∈ X are considered to exist, where Y (x)
is the outcome which would be observed if the exposure were set to X = x
and X is the set of possible values of the exposure. At most one of these out-
comes is ever observed. The three assumptions of Section 3.2.1 necessary for
the assessment of a causal relationship can be expressed in the language of
potential outcomes as follows [Angrist et al., 1996]:
i’. Causal effect of IV on exposure: p(x|g) is a non-trivial function of g
ii’. Independence of the potential exposures and outcomes from the IV:
X(g), Y (x, g) ⊥
⊥ G.
iii’. Exclusion restriction: Y (x, g) = Y (x)
where p(x|g) is the probability distribution function of X conditional on G =
g, Y (x, g) is the potential outcome that would be observed if X were set to
x and G were set to g, Y (x) is the potential outcome observed when X = x,
and X(g) is the potential value of the exposure when G = g. Assumption ii’.
states that the potential values of the exposure and outcome for each value
of the IV do not depend on the actual value of the IV. This would not be
true if, for example, the IV were associated with a confounder. Assumption
iii’. is named ‘exclusion restriction’ and states that the observed outcome for
each value of the exposure is the same for each possible value of the IV. This
means that the IV can only affect the outcome through its association with
the exposure [Clarke and Windmeijer, 2010].
Assumptions for causal inference 39

3.3 Testing for a causal relationship

Mendelian randomization studies are able to address two related questions:
whether there is a causal effect of the exposure on the outcome, and what is
the size of the causal effect [Tobin et al., 2004].
Under the assumption that the genetic variant is a valid IV, the hypothesis
of a causal effect of the exposure on the outcome can be assessed by testing
for independence of the variant and the outcome. A non-zero association is
indicative of a causal relationship [Hernán and Robins, 2006]. The presence
and direction of effect can be tested statistically by straightforward regression
of the outcome on the genetic variant to see whether the estimated association
is compatible with no causal effect based on a chosen threshold for statistical
significance.

3.3.1 Converse of the test

The converse statement to the test for a causal relationship is that if the
correlation between the outcome and variant is zero, then there is no causal
effect of the exposure on the outcome. Although this converse statement is
not always true, as there may be zero linear correlation between the variant
and outcome without independence [Spirtes et al., 2000], it is true for most
biologically plausible models of the exposure–outcome association.

3.3.2 Does Mendelian randomization really assess a causal

relationship?
In a natural experiment such as Mendelian randomization, as there is no in-
tervention or manipulation of the exposure, use of the label ‘causal’ relies
on the assumption that the observational relationships between the genetic
variant(s), exposure, and outcome are informative about the structural rela-
tionship between the exposure and the outcome (structural meaning relating
to the distribution of the variables under intervention). Put simply, this as-
sumption states that the effect on the outcome of the unconfounded observed
difference in the exposure due to the genetic variant would be similar (same
direction of effect) if the genetic variant (or equivalently, the exposure) were
manipulated to take different values, rather than being observed at different
values. Hence although Mendelian randomization is an observational rather
than an experimental technique, under this assumption it does assess a causal
relationship.
40 Mendelian Randomization

3.3.3 Interpreting a null result

A difficulty faced by practitioners of Mendelian randomization is how to inter-
pret a ‘null’ (for example p > 0.05) finding. In such cases, above all, caution
must be exercised against the overinterpretation of a null finding which may
simply be due to low power.
One common approach is to compare the observed and ‘expected’ asso-
ciation between the exposure and the outcome; the latter is based on trian-
gulating the associations between the genetic variant and the exposure and
between the variant and the outcome (Figure 3.4). This ‘expected’ associa-
tion is calculated as the coefficient from the regression of the outcome on the
variant divided by the coefficient from the regression of the exposure on the
variant. This is a ratio estimate (Section 4.1), and is the change in the out-
come expected for a unit change in the exposure if there were no confounding
in the observational association between the exposure and the outcome. While
there is some merit in comparing the ‘expected’ and observed association es-
timates of the exposure with the outcome, this comparison should be seen
as a guide rather than a conclusive statistical test. (The formal test of com-
parison of these estimates is known as an endogeneity test; reasons why we
discourage reliance on an endogeneity test are given in Section 4.5.4.) If the
expected and observed association estimates are similar, then a null finding
may give little evidence as to the causal nature of the exposure. Even if the
estimates are different, there may be good biological reasons for a difference
other than residual confounding (Chapter 6). A better approach is to consider
an estimate of the causal effect and of its precision using an IV method.

Genetic variant

Exposure Outcome

FIGURE 3.4
Triangle of associations: an ‘expected’ association estimate between the ex-
posure and the outcome can be calculated by dividing the coefficient for the
association between the genetic variant and the outcome by the coefficient for
the association between the exposure and the variant. The dashed line is the
association which is estimated.
Assumptions for causal inference 41

3.4 Estimating a causal effect

Although testing for a causal relationship is useful and may be sufficient in
some cases, there are several reasons why it is desirable to go beyond this
and to estimate the size of a causal effect. First, this is usually the parameter
representing the answer to the question of interest. Secondly, with multiple
genetic variants, greater power can be achieved. If several independent IVs
all show a concordant causal effect, the overall estimate of causal effect using
all the IVs may give statistical significance at a given level even if none of
the estimates from the individual IVs achieve significance. Thirdly, often a
null association is expected. By estimating a confidence interval for the causal
effect, we obtain bounds on its plausible size. Although it is not statistically
possible to prove the null hypothesis, it may be possible to obtain a sample
size large enough such that the confidence interval bounds for the causal effect
are narrow enough that the range of plausible causal effect values excludes a
minimally clinically relevant causal effect.
In this section, we consider technical issues associated with parameter esti-
mation: the assumptions necessary to estimate a causal effect, and definitions
of the causal parameters to be estimated. Having discussed these points, we
proceed in the next chapter to consider methods for constructing different IV
estimators.

3.4.1* Additional IV assumptions for estimating a causal

effect
In order to estimate a causal effect, it is necessary to make further assumptions
to the ones listed in Section 3.2.1 [Angrist et al., 1996]:
1. The stable unit treatment value assumption (SUTVA), which states
that the potential outcomes for each individual should be unaffected
by how the exposure was assigned, and unaffected by variables in
the model relating to other individuals [Cox, 1958];
2. Strong monotonicity, which means that varying the IV should alter
the exposure for at least one individual in the population, and that
any change in the exposure from varying the IV should be in the
same direction (an increase or a decrease) for all individuals.
The monotonicity assumption is credible for most biologically plausible
situations in which Mendelian randomization investigations for estimating a
causal effect are undertaken. It would not be plausible in the example from
Section 2.3.2 of the effect of alcohol intake on oesophageal cancer risk, as the
average levels of alcohol intake and the associated disease risk are not mono-
tone in the number of variant alleles. If the monotonicity assumption is not
plausible (for example, if the IV is an unweighted allele score, Section 8.2),
42 Mendelian Randomization

then a causal effect can be identified under a homogeneity assumption that

the causal effect has the same magnitude in all individuals [Swanson and
Hernán, 2013]. If the monotonicity assumption is plausible, then an IV anal-
ysis typically estimates an average causal effect, known as the local average
treatment effect or complier-average causal effect. This is the average causal
effect (see below) amongst individuals whose exposure value is influenced by
the IV. This may be the whole population; an example where it is a subset
of the population is for the exposure of alcohol intake, where individuals who
abstain from alcohol for cultural or religious reasons would do so regardless
of their IV value.
If there is a single IV, then this could be used to calculate the average causal
effect of changing the value of the IV from one value to another. However, it
is usually desired to express a causal effect in terms of the exposure. For
this, it is necessary to assume a parametric relationship between the exposure
and outcome. For a continuous outcome, this is usually a linear model; the
expected value of the outcome is a linear function of the exposure. For a binary
outcome, this may be a linear model for the probability of an outcome, but is
more often a log-linear model or a logistic-linear model; the log-transformed
(or logit-transformed) probability of an outcome is a linear function of the
exposure. Non-linear parametric models have been considered for IV analysis;
however, inference from such models has been shown to be highly dependent
on the parametric form considered [Mogstad and Wiswall, 2010; Horowitz,
2011]. Non-parametric models are discussed in Section 11.1.2.
The SUTVA is generally not plausible in Mendelian randomization, as the
effect on an outcome associated with a genetic variant is likely to be different
to the effect from intervention on the exposure in a number of qualitative and
quantitative ways (Chapter 6) – for example, due to the duration of the inter-
vention (life-long or short-term), the timing of the intervention (on long-term
levels of the exposure or on acute levels), the magnitude of the intervention (ge-
netic effects are usually small, clinical interventions are typically larger), and
the mechanism of the intervention (genetic effects and clinical interventions
may operate via different pathways). Estimates from Mendelian randomiza-
tion should therefore not be interpreted naively as the expected outcome of
an intervention in the risk factor of interest (Section 6.3.3).

3.4.2* Causal parameters

Generally, the desired causal parameter of interest is that which corresponds
to a population-based intervention, equivalent to a randomized controlled trial
(RCT) [Greenland, 1987].
The average causal effect (ACE) [Didelez and Sheehan, 2007] under inter-
vention on the exposure is the expected difference in the outcome when the
Assumptions for causal inference 43

exposure is set to two different values:

ACE(x0 , x1 ) = Expected outcome when the exposure is set at x1

− expected outcome when the exposure is set at x0 . (3.3)

This can be written as:

ACE(x0 , x1 ) = E(Y |do(X = x1 )) − E(Y |do(X = x0 )). (3.4)

The ACE is zero when there is conditional independence between Y and X

given U , but the converse is not generally true [Didelez and Sheehan, 2007].
With a binary outcome (Y = 0 or 1), the ACE is also called the causal
risk difference. However, it is often more natural to consider a causal risk ratio
(CRR) or causal odds ratio (COR):

Probability of outcome when the exposure is set at x1

CRR(x0 , x1 ) = ,
Probability of outcome when the exposure is set at x0
(3.5)
Odds of outcome when the exposure is set at x1
COR(x0 , x1 ) = . (3.6)
Odds of outcome when the exposure is set at x0
These can be written as:
P(Y = 1|do(X = x1 ))
CRR(x0 , x1 ) = , (3.7)
P(Y = 1|do(X = x0 ))
P(Y = 1|do(X = x1 ))P(Y = 0|do(X = x0 ))
COR(x0 , x1 ) = . (3.8)
P(Y = 1|do(X = x0 ))P(Y = 0|do(X = x1 ))

3.5 Summary
The instrumental variable assumptions make assessment of causation in an ob-
servational setting possible without complete knowledge of all the confounders
of the exposure–outcome association. Genetic variants have good theoretical
and empirical plausibility for use as instrumental variables in general, but the
instrumental variable assumptions may be violated for a number of reasons.
We continue in the next chapter to consider methods for estimating the
magnitude of a causal effect using instrumental variables.
4
Methods for instrumental variable analysis

In this chapter, we discuss methods for the estimation of causal effects using
instrumental variables (IVs) with both continuous and binary outcomes. We
focus attention on the case of a single continuous exposure variable, as this
is the usual situation in Mendelian randomization studies; although the same
methods could be used in the case of a single binary exposure. We explain for
each method how to estimate a causal effect, and describe specific properties
of the estimator. In turn, we consider the ratio of coefficients method, two-
stage methods, likelihood-based methods, and semi-parametric methods. This
order corresponds roughly to the complexity of the methods, with the simplest
ones first. These methods are contrasted in terms of bias, coverage, efficiency,
power, robustness to misspecification, and existence of finite moments. We
have included a simple explanation of each method at first, and then further
details for more technical readers. Also discussed are implementations of the
methods using standard statistical software packages.

4.1 Ratio of coefficients method

The ratio of coefficients method, or the Wald method [Wald, 1940], is the sim-
plest way of estimating the causal effect of the exposure (X) on the outcome
(Y ). The ratio method uses a single IV. If more than one variant is available
which is an IV then the causal estimates from the ratio method using each
variant can be calculated separately, or the variants can be combined into a
single IV in an allele score approach (Section 8.2). Otherwise, other estimation
methods in this chapter can be used.

4.1.1 Continuous outcome, dichotomous IV

We initially assume that we have an IV G which takes the values 0 or 1, di-
viding the population into two genetic subgroups. The IV can be thought of
as a single nucleotide polymorphism (SNP) where two of the three subgroups
are merged together, for example reflecting a dominant or recessive genetic
model, or because there are very few individuals in the least common genetic

45
46 Mendelian Randomization

subgroup (the minor homozygotes). In a recessive model, a single copy of the

major (wildtype) allele A is sufficient to mask a minor (variant) allele; the
genetic subgroups are AA/Aa (major homozygote/heterozygote) and aa (mi-
nor homozygote). A dominant model is similar, except that the heterozygotes
are combined with the minor homozygotes; the two genetic subgroups are AA
and Aa/aa.
From the IV assumptions, the distribution of the exposure differs in the
two genetic subgroups. If the distribution of the outcome also differs, then
there is a causal effect of the exposure on the outcome. We define ȳj for
j = 0, 1 as the average value of outcome for all individuals with genotype
G = j, and define x̄j similarly for the exposure. Figure 4.1 displays the mean
exposure and outcome in the two genetic subgroups in a fictitious example
with a positive causal effect of X on Y .
IV estimates are usually expressed as the change in the outcome result-
ing from a unit change in the exposure, although changes in the outcome
0.4
0.3

∆Y 0.4
= = 0.4
Outcome

∆X 1.0
∆Y = 0.4
0.2
0.1
0.0

∆X= 1.0

3.0 3.2 3.4 3.6 3.8 4.0 4.2

Exposure

FIGURE 4.1
Points representing mean exposure and outcome in two genetic subgroups
with IV ratio estimate.
Methods for instrumental variable analysis 47

corresponding to different magnitudes of change in the exposure could be

quoted instead. If the exposure has been (natural) log-transformed, a unit
increase in the log-transformed exposure corresponds to a exp(1) = 2.72-fold
multiplicative in the untransformed exposure. The effect of a (say) 20% in-
crease in the exposure can be considered by multiplying the causal estimate
by log(1.2) = 0.182, or of a 30% decrease by multiplying by log(0.7) = −0.357
(where log is the natural logarithm or ln). If an IV estimate is expressed for
a change in the exposure much greater than that associated with the genetic
variant, this extrapolation may not be justified and the IV estimate may not
be realistic. However, some extrapolation is often necessary to convert the
genetic association to a clinically relevant causal effect of the exposure.
We see that an average difference in the exposure between the two sub-
groups of ∆X = x̄1 − x̄0 results in an average difference in the outcome of
∆Y = ȳ1 − ȳ0 . Assuming that the effect of the exposure on the outcome is
linear, the ratio estimate for the change in outcome due to a unit increase in
the exposure is:
∆Y ȳ1 − ȳ0
Ratio method estimate (dichotomous IV) = = . (4.1)
∆X x̄1 − x̄0
In the example shown (Figure 4.1), ∆Y = 0.4 and ∆X = 1.0, giving a ratio
estimate of 0.4
1.0 = 0.4.
The numerator and denominator in the ratio estimate are the average
causal effects on the outcome and exposure respectively of being in genetic
subgroup 1 versus being in genetic subgroup 0. If we assume that the effect of
the exposure on the outcome is linear, then the ratio estimate is the average
causal effect on the outcome of an exposure of x + 1 units versus an exposure
of x units. (Under the linearity assumption, the causal effect of a unit increase
in the exposure is equal for all values of x.) If the effect is not linear, then a
ratio estimate approximates the average causal effect of a population inter-
vention in the exposure [Burgess et al., 2014b]. (Non-linear exposure–outcome
relationships are discussed further in Section 11.1.2.)

4.1.2 Continuous outcome, polytomous or continuous IV

Alternatively, the IV may not be dichotomous, but polytomous (takes more
than two distinct values). This is the usual case for a diallelic SNP; the three
levels AA (major homozygote), Aa (heterozygote), and aa (minor homozy-
gote) will be referred to as 0, 1, and 2, corresponding to the number of minor
alleles. In a linear ‘per allele’ model, we assume that the association of the
genetic variant with the exposure is proportional to the number of variant
alleles. The IV could also be a continuous allele score (Section 8.2), under the
assumption that the association of the score with the exposure is also linear.
The coefficient of G in the regression of X on G is written as β̂X|G , and
represents the change in X for a unit change in G. Similarly, the coefficient
of G in the regression of Y on G is written as β̂Y |G . The ratio estimate of the
48 Mendelian Randomization

causal effect is:

β̂Y |G
Ratio method estimate (polytomous/continuous IV) = . (4.2)
β̂X|G
Intuitively, we can think of the ratio method as saying that the change in Y
for a unit increase in X is equal to the change in Y for a unit increase in G,
scaled by the change in X for a unit increase in G.
Illustrative data are shown in Figure 4.2. Each of the graphs is plotted on
the same scale. The top-left panel shows that the exposure and outcome are
negatively correlated, with the line showing the observational association from
linear regression. However, as shown in the top-right panel, where individuals
in different genetic subgroups are marked with different plotting symbols,
individuals in the subgroup marked with circles tend to congregate towards the
south-west of the graph and individuals in the subgroup marked with squares
tend towards the north-east of the graph. The bottom-left panel shows the
mean values of the exposure and outcome in each genetic subgroup with lines
representing 95% confidence intervals intervals for the means. The bottom-
right panel includes the individual data points, the subgroup means and the
causal estimate from the ratio method. We see that the causal estimate is
positive. The 95% confidence intervals for the lines passing through the points
show that the uncertainty in the ratio IV estimate is greater than that of the
observational estimate.
From a technical point of view, the ratio estimator is valid under the as-
sumption of monotonicity of the genetic effect on the exposure and linearity
of the causal X–Y association [Angrist et al., 2000]. Because of this, the ratio
estimate has been named the linear IV average effect (LIVAE) [Didelez et al.,
2010]. Monotonicity means that the exposure for each individual would be in-
creased (or alternatively for each individual would be decreased) or unchanged
if that person had G = g1 compared to if they had G = g0 for all g1 > g0 .
We note that it is not necessary for the genetic effect on the exposure to be
constant in magnitude for all individuals, merely consistent in direction (that
is, there may be effect modification), or for the exposure effect on the outcome
to be constant in magnitude. If the monotonicity assumption is not satisfied,
then the causal effect of the exposure on the outcome can only be estimated
consistently if it is constant for all individuals across the population.
The linearity assumption is that the expected value of the outcome Y
conditional on the exposure X and confounders U is:

E(Y |X = x, U = u) = β0 + β1 x + h(u) (4.3)

where h(u) is a function of U . Hence, there is no interaction term between

X and U in the conditional expectation of Y . It is also required that the
structural model:
E(Y |do(X = x)) = β0′ + β1 x (4.4)
holds, where the causal effect β1 is the same as in the equation above. This is
Methods for instrumental variable analysis 49

Outcome

Outcome
Exposure Exposure
Outcome

Outcome

Exposure Exposure

FIGURE 4.2
Illustration of ratio method for polytomous IV taking three values with a
continuous outcome in a fictitious dataset: (top-left) exposure and outcome
for all individuals, observational estimate with 95% confidence interval; (top-
right) individuals divided into genetic subgroups by plot symbol; (bottom-left)
mean exposure and outcome in each genetic subgroup (lines represent 95%
confidence intervals); (bottom-right) ratio IV estimate with 95% confidence
interval.
50 Mendelian Randomization

similar to a consistency assumption, which states that the outcome for an indi-
vidual would be the same if the value of the exposure were observed naturally
or set due to an intervention [VanderWeele, 2009]. Although confounding is
represented by a single variable U , this is simply for presentation; U represents
the combined effect of all confounding variables.
We note that the ratio estimate can be calculated simply from the coef-
ficients β̂X|G and β̂Y |G , and as such only requires the availability of summa-
rized data, not individual-level data. Methods for obtaining IV estimates using
summarized data are discussed further in Section 9.4. The two coefficients can
also be estimated in different groups of individuals. Common examples include
where the IV–outcome association is measured on the whole sample and the
IV–exposure association on a subsample (subsample Mendelian randomiza-
tion, see Section 8.5.2), or the associations are estimated on non-overlapping
datasets (two-sample Mendelian randomization, see Section 9.8.2).

4.1.3 Binary outcome

Generally in epidemiological applications, disease is the outcome of interest.
Disease outcomes are often dichotomous. We use the epidemiological termi-
nology of referring to an individual with an outcome event as a case (Y = 1),
and an individual with no event as a control (Y = 0).
With a binary outcome and a dichotomous IV, the ratio estimate is defined
similarly as with a continuous outcome:
∆Y
Ratio method log risk ratio estimate (dichotomous IV) = (4.5)
∆X
ȳ1 − ȳ0
=
x̄1 − x̄0
where ȳj is commonly the log of the probability of an event, or the log odds
of an event, in genetic subgroup j. The term “risk ratio” is used as a generic
term meaning relative risk (for the log of the probability) or odds ratio (for
the log odds) as appropriate.
With a polytomous or continuous IV, the coefficient β̂Y |G in the ratio
estimate (equation 4.2) is taken from regression of Y on G. The regression
model used could in principle be linear, where the IV estimate represents
the change in the probability of an event for a unit change in the exposure.
However, with a dichotomous outcome, log-linear or logistic regression models
are generally preferred, where the IV estimate represents the log relative risk
or log odds ratio, respectively, for a unit change in the exposure. With logistic
models, the odds ratio being estimated depends on the choice of covariates
included in the model (Section 4.2.3*).
The ratio estimate is also commonly quoted in its exponentiated form:
Ratio method risk ratio estimate (dichotomous IV) = R1/∆X (4.6)
where R is the estimated risk ratio between the two genetic subgroups.
Methods for instrumental variable analysis 51

As in the continuous case, this estimator is valid under the assumption

of monotonicity of X on G and a log-linear or logistic-linear model [Didelez
et al., 2010]. In the log-linear case, the association model is:

log(E(Y |X = x, U = u)) = β0 + β1 x + h(u) (4.7)

and the structural model is:

log(E(Y |do(X = x))) = β0′ + β1 x (4.8)

for some β0 , β0′ , β1 , h(u) as above.

4.1.4 Retrospective and case-control data

In Mendelian randomization, when retrospective data are available, it is usual
to make inferences on the gene–exposure association using only non-diseased
individuals, such as the control population in a case-control study [Minelli
et al., 2004]. This makes the assumption that the distribution of the exposure
in the controls is similar to that of the general population, which is true for a
rare disease [Bowden and Vansteelandt, 2011]. This is necessary to prevent bias
of the causal estimate for two reasons. The first reason is reverse causation,
whereby post-event measurements of the exposure may be distorted by the
outcome event. Secondly, in a case-control setting, over-recruitment of cases
into the study means that the distribution of confounders in the ascertained
population is different to that in the general population. An association is
then induced between the IV and the confounders, leading to possible bias in
the IV estimate [Didelez and Sheehan, 2007]. This affects not only the ratio
method, but all IV methods.
If the outcome is common and its prevalence in the population from which
the case-control sample was taken is known, such as in a nested case-control
study, then inferences on the gene–exposure association can be obtained using
both cases and controls, provided that measurements of the exposure in cases
were taken prior to the outcome event. This analysis can be performed by
weighting the sample so that the proportions of cases and controls in the
reweighted sample match those in the underlying population [Bowden and
Vansteelandt, 2011].

4.1.5 Confidence intervals

Confidence intervals for the ratio estimate can be calculated in several ways.
Normal approximation: The simplest way is to use a normal approx-
imation. With a continuous outcome, standard errors (SEs) and confidence
intervals from the two-stage least squares method, introduced below, are given
in standard software commands (Section 4.6). Alternatively, the following ap-
proximation can be used, based on the first two terms of the delta method
52 Mendelian Randomization

expansion for the variance of a ratio:

v
u
u se(β̂Y |G ) 2 β̂Y |G 2 se(β̂X|G ) 2
Standard error of ratio estimate ≃ t + (4.9)
β̂X|G 2 β̂X|G 4

This approximation assumes that the numerator and denominator of the ratio
estimator are uncorrelated; such correlation could be accounted for by includ-
ing a third term of the delta expansion [Thomas et al., 2007], but is unlikely
to have a considerable impact on the estimate of the standard error.
However, asymptotic (large sample) normal approximations may result in
overly narrow confidence intervals, especially if the sample size is not large or
the IV is ‘weak’. This is because IV estimates are not normally distributed.
Fieller’s theorem: If the regression coefficients in the ratio method β̂Y |G
and β̂X|G are assumed to be normally distributed, critical values and confi-
dence intervals for the ratio estimator may be calculated using Fieller’s theo-
rem [Fieller, 1954; Lawlor et al., 2008]. We assume that the correlation between
β̂Y |G and β̂X|G is zero; other values can be used, but the impact on the confi-
dence interval is usually small [Minelli et al., 2004]. If the standard errors are
se(β̂Y |G ) and se(β̂X|G ) and the sample size is N , then we define:

f0 = β̂Y |G 2 − tN (0.975)2 se(β̂Y |G )2 (4.10)

2 2 2
f1 = β̂X|G − tN (0.975) se(β̂X|G )
f2 = β̂Y |G β̂X|G
D = f2 2 − f0 f1

where tN (0.975) is the 97.5th percentile point of a t-distribution with N de-

grees of freedom (for N > 100, tN (0.975) ≈ 1.96).
√ If D > 0 and √ f1 > 0, then the 95% confidence interval is from (f2 −
D)/f1 to (f2 + D)/f1 . The confidence interval is more likely to be a closed
interval like this if we have a ‘strong’ instrument, that is, an instrument which
explains a large proportion of the variation of the exposure in the population.
Confidence intervals of size α can be similarly constructed by using the (1 −
α/2) point of the t-distribution.
If D < 0, then there is no interval which covers the true parameter with
95% confidence. This occurs when there is little differentiation in both the
exposure and outcome distributions between the genetic subgroups (due to a
weak instrument), and so a gradient corresponding to any size of causal effect
is plausible. The only valid 95% confidence interval is the unbounded interval
from minus infinity to plus infinity. An example where Fieller’s theorem would
give an unbounded confidence interval is displayed in Figure 4.3. This situation
is likely to occur when the IV explains little of the variation in the exposure;
it is a weak instrument.
If D > 0 and f1 < 0, then the 95% confidence
√ interval is the union
√ of
two intervals from minus infinity to (f2 + D)/f1 and from (f2 − D)/f1
Methods for instrumental variable analysis 53

0.4
0.3
0.2
Outcome
0.1
0.0
−0.1
−0.2
−0.3

2.6 2.8 3.0 3.2 3.4 3.6

Exposure

FIGURE 4.3
Points representing mean exposure and outcome (lines are 95% confidence
intervals) in two genetic subgroups where the confidence interval from Fieller’s
theorem for the IV ratio estimate is unbounded.

to plus infinity.
√ All possible values
√ are included in the interval except those
between (f2 + D)/f1 and (f2 − D)/f1 . An example where Fieller’s theorem
would give such a confidence interval including infinity but excluding zero, is
displayed in Figure 4.4. This suggests that the differences in the outcome are
not caused solely by differences in the exposure, and so the IV assumptions
are violated.
To summarize, Fieller’s theorem gives confidence intervals that have one
of three possible forms [Buonaccorsi, 2005]:
i. The interval may be a closed interval [a, b],

ii. The interval may be the complement of a closed interval (−∞, b] ∪ [a, ∞),
iii. The interval may be unbounded.
√ √
where a = (f2 − D)/f1 , b = (f2 + D)/f1 . Confidence intervals from Fieller’s
theorem are preferred to those from an asymptotic normal approximation
when the IV is weak. A tool to calculate confidence intervals from Fieller’s
theorem based on the gene–exposure and gene–outcome associations is avail-
able online (http://spark.rstudio.com/sb452/fieller/).
54 Mendelian Randomization

0.8
0.6
Outcome
0.4
0.2
0.0
−0.2

2.6 2.8 3.0 3.2 3.4 3.6

Exposure

FIGURE 4.4
Points representing mean exposure and outcome (lines are 95% confidence
intervals) in two genetic subgroups where the confidence interval from Fieller’s
theorem for the IV ratio estimate is compatible with an infinite (vertical)
association, but not a null (horizontal) association.

Bootstrapping: As an alternative approach, also applicable to any of the

following methods, confidence intervals can be calculated by bootstrapping
[Efron and Tibshirani, 1993]. The simplest way of constructing a bootstrapped
confidence interval is by taking several random samples with replacement from
the data of the same sample size. The empirical distribution of the IV estima-
tor in the bootstrapped samples approximates the true distribution of the IV
estimator [Imbens and Rosenbaum, 2005]. However, there are some concerns
about the behaviour of bootstrapped confidence intervals with weak instru-
ments [Moreira et al., 2009].
Other approaches: Alternative approaches for inference with weak in-
struments not discussed further here are confidence based on inverting a test
statistic, such as the Anderson–Rubin test statistic [Anderson and Rubin,
1949] or the conditional likelihood ratio test statistic [Moreira, 2003]. These in-
tervals give appropriate confidence levels under the null hypothesis with weak
instruments, but may be underpowered with stronger instruments. They have
been discussed in detail elsewhere [Mikusheva, 2010; Davidson and MacKin-
non, 2014] and implemented in Stata [Mikusheva and Poi, 2006] and R [Small,
2014].
Methods for instrumental variable analysis 55

4.1.6 Absence of finite moments

One peculiar property of the ratio estimator is that its mean (also known
as its first moment) is not finite. This implies that, if you generated data
on the exposure and outcome from a model with a valid IV and calculated
the ratio IV estimate a large number of times, the mean value of these IV
estimates could become arbitrarily high (or low). This is due to the fact that
there is a finite probability that the denominator in the ratio estimate (∆X
or β̂X|G ) is very close to zero, leading to a large IV estimate. In practice,
this is unlikely to be a serious issue since, if ∆X were close to zero, the IV
would be considered invalid as assumption i. (Section 3.2.1) would appear to
be violated. Theoretically, the absence of a finite mean makes comparison of
IV methods more difficult, as the (mean) bias of the ratio estimate, defined
as the difference between the mean IV estimate (the expected value of the
IV estimate) and the true value of the causal effect, cannot be calculated for
any finite sample size. We therefore additionally consider the median bias, the
difference between the median of the estimator over its distribution and the
true causal effect, when comparing different methods for IV estimation.
Central moments (often simply called the moments) are the expectations
of the powers of a random variable with its mean subtracted. The kth moment
of the random variable Z with mean µ is E((Z − µ)k ), for k = 1, 2, . . . . All of
the central moments of the ratio IV estimator are infinite. In particular, its
mean and variance are undefined.

4.1.7 Coverage and efficiency

The coverage of a confidence interval is the probability that the confidence
interval contains the true parameter value. By definition, a 95% confidence
interval should contain the true parameter value 95% of the time. However,
in practice, this may not be true, due to approximations and distributional
approximations made in constructing the interval. By simulating data where
the true parameter values are known, the coverage properties of differently
estimated confidence intervals can be investigated.
Efficiency is a property of an estimator relating to its variance. An efficient
estimator has low variance and therefore a narrow confidence interval. A de-
sirable estimator has a narrow confidence interval, but maintains the correct
coverage. The coverage and efficiency of various IV estimators are discussed
in this chapter, but addressed in more detail in Chapter 7 onwards.

4.1.8 Reduced power of IV analyses

Figure 4.2 illustrates the wider confidence interval of an IV estimate com-
pared with that of an observational estimate. As in many areas of applied
statistics, there is a trade-off in choice of estimation procedure between bias
and variance. The observational estimate is precisely estimated, but typically
56 Mendelian Randomization

biased for the causal effect, whereas the IV estimate is unbiased, but typically
imprecisely estimated. The loss of precision in the IV estimate is the cost of
unbiased estimation. When making causal assessments, we would argue that
no appreciable amount of bias should be introduced in order to reduce the
variance of the estimate [Zohoori and Savitz, 1997].
However, the sample size required to obtain precise enough causal esti-
mates to be clinically relevant can be very large [Ebrahim and Davey Smith,
2008]. A rule of thumb for power is that the sample size for a conventional
analysis should be divided by the coefficient of determination (R2 ) of the IV
on the exposure (Section 8.3) [Wooldridge, 2009]. For example, if the sample
size for an observational regression analysis of the outcome on the exposure
to detect a given effect size requires a sample size of 400, and the IV explains
2% of the variation in the exposure, then the sample size required for an IV
analysis is approximately 400/0.02 = 20000. For this reason, while for some
researchers the ratio method may be sufficient for the analysis in question, we
are motivated to consider methods which can incorporate data on more than
one IV, and hence give more precise estimates of causal effects.

4.2 Two-stage methods

A two-stage method comprises two regression stages: the first-stage regression
of the exposure on the genetic IVs, and the second-stage regression of the
outcome on the fitted values of the exposure from the first stage.

4.2.1 Continuous outcome – Two-stage least squares

With continuous outcomes and a linear model, the two-stage method is known
as two-stage least squares (2SLS). It can be used with multiple IVs. In the
first-stage (G–X) regression, the exposure is regressed on the IV(s) to give
fitted values of the exposure (X̂|G). In the second-stage (X–Y ) regression,
the outcome is regressed on the fitted values for the exposure from the first
stage regression. The causal estimate is this second-stage regression coefficient
for the change in outcome caused by a unit change in the exposure.
With a single IV, the 2SLS estimate is the same as the ratio estimate (with
a continuous and with a binary outcome). With multiple IVs, the 2SLS esti-
mator may be viewed as a weighted average of the ratio estimates calculated
using the instruments one at the time, where the weights are determined by
the relative strengths of the instruments in the first-stage regression [Angrist
et al., 2000; Angrist and Pischke, 2009].
Suppose we have K instrumental variables available. With data on individ-
uals indexed by i = 1, . . . , N who have exposure xi , outcome yi and assuming
an additive per allele model for the IVs gik indexed by k = 1, . . . , K, the
Methods for instrumental variable analysis 57

first-stage regression model is:

X
xi = α0 + αk gik + εXi . (4.11)
k
P
The fitted values x̂i = α̂0 + k α̂k gik are then used in the second-stage re-
gression model:
yi = β0 + β1 x̂i + εY i (4.12)
where εXi and εY i are independent error terms. The causal parameter of in-
terest is β1 . If both models are estimated by standard least-squares regression,
both the error terms are implicitly assumed to be normally distributed.
Although estimation of the causal effect in two stages (a sequential re-
gression method) gives the correct point estimate, the standard error from
the second-stage regression (equation 4.12) is not correct. This is because it
does not take into account the uncertainty in the first-stage regression. Under
homoscedasticity of the error term in the equation:
yi = β0 + β1 xi + ε′Y i (4.13)
the asymptotic variance of the 2SLS estimator is:
σ̂ 2 (X T G(GT G)−1 GT X)−1 = σ̂ 2 (X̂ T X̂)−1 (4.14)
where σ̂ 2 is an estimate of the variance of the residuals from equation (4.13),
and the matrices G of IVs and X for the exposure contain constant terms. The
use of 2SLS software is recommended for estimation (Section 4.6) [Angrist and
Pischke, 2009]. Robust standard errors are often used in practice, as estimates
are sensitive to heteroscedasticity and misspecification of the equations in the
model.
When all the associations are linear and the error terms normally dis-
tributed, the 2SLS estimator has a finite kth moment when there are at least
(k + 1) IVs [Kinal, 1980]. Therefore the mean of a 2SLS estimator is only
defined when there are at least 2 IVs, and the variance is only defined when
there are at least 3 IVs.

4.2.2 Binary outcome

The analogue of 2SLS with binary outcomes is a two-stage estimator where the
second-stage (X–Y ) regression uses a log-linear or logistic regression model.
This can be implemented using a sequential regression method by performing
the two regression stages in turn (also known as two-stage predictor substitu-
tion). Estimates from such an approach will be overly precise, as uncertainty
in the first-stage regression is not accounted for; however, this over-precision
may be slight if the standard error in the first-stage coefficients is low. This
can be resolved by the use of a likelihood-based method (Section 4.3.4*) or a
bootstrap method, such as that implemented in Stata using the qvf command
(Section 4.6.2).
58 Mendelian Randomization

As with the ratio IV estimator, in a case-control study it is important to

undertake the first-stage regression only in the controls, not the cases (Sec-
tion 4.1.4). Fitted exposure values for the cases are be obtained by substituting
their genetic variants into the first-stage regression model.
Two-stage regression methods with non-linear second-stage regression
models (such as with binary outcomes) have been criticized and called “for-
bidden regressions” [Angrist and Pischke, 2009, page 190]. This is because the
non-linear model does not guarantee that the residuals from the second-stage
regression are uncorrelated with the instruments [Foster, 1997]. There is cur-
rent debate about the interpretation and validity of such estimates, especially
when the measure of association is non-collapsible.

4.2.3* Non-collapsibility
Several measures of association, including odds ratios, differ depending on
whether they are considered conditional or marginal on a covariate. For ex-
ample in the left half of Table 4.1, the odds ratio of an outcome for exposed
versus unexposed individuals is equal to 2 for men and 2 for women. Even
under the assumption of no confounding (that the proportion of exposed and
non-exposed individuals is the same in both men and women), the odds ratio
for a population with equal numbers of men and women is not 2. In contrast,
as the example in the right half of Table 4.1 shows, a relative risk is the same
whether considered conditional or marginal on sex.
A measure of association, such as an odds ratio or relative risk, would be
termed collapsible if, when it is constant across the strata of the covariate, this
constant value equals the value obtained from the overall (marginal) analysis.
Non-collapsibility is the violation of this property [Greenland et al., 1999]. The
relative risk and absolute risk difference are collapsible measures of association.
Odds ratios are generally non-collapsible [Ducharme and LePage, 1986]. This
means that the conditional model:

logit(E(Y |X = x, U = u)) = β0 + β1 x + h(u) (4.15)

Probability of event Odds Probability of event Relative

Unexposed Exposed ratio Unexposed Exposed risk
3 3
Men 13 8
2 0.3 0.6 2
1 1
Women 21 11
2 0.05 0.1 2
Overall 0.139 0.233 1.88 0.175 0.35 2

TABLE 4.1
Illustrative examples of collapsing an effect estimate over a covariate: non-
equality of conditional and marginal odds ratios and equality of relative risks.
Methods for instrumental variable analysis 59

and the structural model:

log(E(Y |do(X = x))) = β0′ + β1 x (4.16)

for a logistic model of association cannot in general both be true simultane-

ously for the same value of β1 .
An odds ratio estimated in an observational study by conventional mul-
tivariable logistic regression is conditional on those covariates adjusted for in
the analysis. Unless adjustment is made in the instrumental variable analysis,
an odds ratio estimated in a Mendelian randomization study is marginal on
these covariates. The odds ratio from a ratio or two-stage analysis method
is conditional on the IV, but marginal in all other variables, including the
exposure itself if it is continuous [Burgess and CCGC, 2013].
This has several consequences for Mendelian randomization. First, the pa-
rameter estimated by a two-stage analysis is best interpreted as a population-
averaged causal effect. This approximates the effect estimated by a RCT,
without adjustment for covariates, where the intervention is to change the dis-
tribution of the exposure by increasing the exposure uniformly for all individ-
uals in the population [Stukel et al., 2007]. Generally, a population-averaged
causal effect marginal across all covariates is the estimate of interest for a
policy-maker as it represents the effect of intervention on the exposure at a
population level [Vansteelandt et al., 2011].
Secondly, naive comparison of odds ratio estimates from multivariable re-
gression and from two-stage IV analysis is not strictly valid, as the two odds
ratios represent different quantities. The degree of attenuation of the IV es-
timate depends on the prevalence of the outcome (greater attenuation for
common outcomes), the magnitude of the causal effect (greater proportional
attenuation for larger effects) and the heterogeneity in individual risks (greater
attenuation for more heterogeneous populations). Epidemiological data for
coronary heart disease risk has shown attenuation towards unity of 5–14% for
odds ratios of around 1.2 to 1.4, although this is likely to be an underestimate
of the true attenuation as not all predictors in the risk model are known in
practice [Burgess, 2012a]. Attenuation is more substantial when the odds ratio
estimate is further from the null.
Thirdly, the apparent inconsistency of estimates from two-stage methods
with a non-collapsible measure of association can be explained as manifesta-
tion of non-collapsibility. For example, the estimate from a two-stage method
with a logistic second-stage model is not in general consistent for the param-
eter β1 in equation (4.15) or equation (4.16). This is discussed further below.
Despite the consequences of non-collapsibility, the two-stage estimator
with a logistic second-stage model still provides a valid test of the null hy-
pothesis [Vansteelandt et al., 2011].
60 Mendelian Randomization

4.2.4* Adjusted two-stage method

An adjusted two-stage method has been proposed, where the residuals from
the first-stage regression of the exposure on the IV are included in the second-
stage regression of the outcome on the fitted values of the exposure. This
has been referred to as a control function approach [Nagelkerke et al., 2000],
or two-stage residual inclusion (2SRI) method [Terza et al., 2008]. If we
have a first-stage regression of X on G with fitted values X̂|G and resid-
uals R̂|G = X − X̂|G, then the adjusted two-stage estimate comes from a
second-stage regression additively on X̂|G and R̂|G (or equivalently on X and
R̂|G). The residuals from the first-stage regression incorporate information on
confounders.
If the second-stage regression is linear, as with a continuous outcome,
then inclusion of these residuals in the second-stage regression model does not
change the estimate, as the residuals are orthogonal to the fitted values. If the
outcome is binary, inclusion of these residuals in a second-stage logistic regres-
sion model means that the IV estimate will be conditional on these residuals.
Numerically, it brings the IV estimate closer to the conditional log odds ratio,
the parameter β1 in the logistic-linear model (equation 4.15) [Palmer et al.,
2008]. Some investigators have therefore recommended the adjusted two-stage
method when the second-stage regression is logistic on the premise that it is
less biased than the unadjusted two-stage method.
Under a particular choice of mathematical model, the adjusted two-stage
estimate is consistent for the parameter β1 [Terza et al., 2008]. However, this
mathematical model is unrealistic, and in general the adjusted two-stage es-
timate is biased for this parameter [Cai et al., 2011]. Further, when the con-
founders are unknown, as is usual in an IV analysis, it is not clear what
variable is represented by the first-stage residuals, and so which covariates
the adjusted two-stage estimate is conditional on and which it is marginal
across. It is uncertain what odds ratio is being estimated by an adjusted two-
stage approach, that is, to what question is the adjusted two-stage estimate
the answer. This is in contrast to the unadjusted two-stage method, which
consistently estimates an odds ratio which is marginal across all covariates
except for the IV itself [Burgess and Thompson, 2012]. We therefore do not
recommend adjustment for the first-stage residuals in a two-stage method.

4.3 Likelihood-based methods

The above methods are not likelihood-based and do not provide maximum
likelihood estimates, which have the desirable properties of asymptotic unbi-
asedness, normality and efficiency. So we next consider likelihood-based meth-
ods.
Methods for instrumental variable analysis 61

4.3.1 Full information maximum likelihood

If we have the same situation as for the two-stage model equations (4.11) and
(4.12), such that each individual i = 1, . . . , N has exposure xi , continuous
outcome yi and IVs gik indexed by k = 1, . . . , K, we can assume the following
model:
X
xi = α0 + αk gik + εXi (4.17)
k
yi = β0 + β1 xi + εY i

where the error terms ε = (εX , εY )T have a bivariate normal distribution

ε ∼ N (0, Σ). (These error terms differ from those defined in equations 4.11
and 4.12.) The causal parameter of interest is β1 . Correlation between εX
and εY is due to confounding. We can simultaneously calculate the maximum
likelihood estimates of β1 and each of the other parameters in the model.
This is known as full information maximum likelihood (FIML) [Davidson and
MacKinnon, 1993].
Confidence intervals can be obtained by the assumption of asymptotic
normality of the parameter estimates.

4.3.2 Limited information maximum likelihood

A disadvantage of FIML is that all the parameters in each of the equations
are estimated. This means that each of the regression equations has to be
correctly specified to give a consistent estimate of β1 . In practice, we are
only interested in β1 , and not in the other parameters. In limited information
maximum likelihood (LIML), we maximize the likelihood substituting for and
profiling out (referred to by economists as ‘concentrating out’) each of the
parameters except β1 .
LIML has been called the ‘maximum likelihood counterpart of 2SLS’
[Hayashi, 2000, page 227] and gives the same causal estimate as the 2SLS
and ratio methods with a single IV. As with 2SLS, estimates are sensitive to
heteroscedasticity and misspecification of the equations in the model. Use of
LIML has been strongly discouraged by some, as LIML estimates do not have
defined moments for any number of instruments [Hahn et al., 2004]. However,
use has also been encouraged by others, especially with weak instruments
(Section 4.5.2), as the median of the distribution of the estimator is close to
unbiased even with weak instruments [Angrist and Pischke, 2009]. With large
numbers of IVs (10 or more), standard confidence intervals from the LIML
method with weak instruments are too narrow and a correction is needed
(known as Bekker standard errors) [Bekker, 1994]. Although this correction
is required to maintain nominal coverage levels, the efficiency of the LIML
estimator is reduced, and it may be outperformed by a simple allele score
approach [Davies et al., 2014].
62 Mendelian Randomization

The LIML estimate can be intuitively understood as the effect β1 that

minimizes the residual sum of squares from the regression of the component
of Y not caused by X, (yi − β1 xi ), on G. Informally, the LIML estimator is
the causal parameter for which the component of Y due to confounding is as
badly predicted by G as possible.

4.3.3 Bayesian methods

Inference from a similar likelihood model can be undertaken in a Bayesian
framework. For each individual i, we model the measured exposure xi and
outcome yi as coming from a bivariate normal distribution for (Xi , Yi )T with
mean (ξi , ηi )T and variance-covariance matrix Σ. The mean of the exposure
distribution ξi is assumed to be a linear function of the instruments gik , k =
1, . . . , K, and the mean of the outcome distribution ηi is assumed to be a
linear function of the mean exposure [Jones et al., 2012].
    
Xi ξi
∼ N2 ,Σ (4.18)
Yi ηi
X
ξi = α0 + αk gik
k
ηi = β0 + β1 ξi

This model is similar to that in the FIML and LIML methods, except that
the causal parameter β1 represents the causal effect between the true means ξi
and ηi rather than the measured values of outcome and exposure. The model
can be estimated in a Markov chain Monte Carlo (MCMC) framework, such
as that implemented in WinBUGS [Spiegelhalter et al., 2003]. The output
is a posterior distribution, from which the posterior mean or median can
be interpreted as a point estimate, and the 2.5th and 97.5th percentiles as
a ‘95% confidence interval’. Under certain choices of prior, estimates based
on the Bayesian posterior distribution are similar to those from the 2SLS
or LIML methods [Kleibergen and Zivot, 2003]. With vague priors, the joint
posterior distribution is similar to the frequentist likelihood function [Burgess
and Thompson, 2012].
An advantage of the Bayesian approach is that no distributional assump-
tion is made for the posterior distribution of the causal parameter. Inference is
therefore more robust using weak instruments [Burgess and Thompson, 2012].

4.3.4* Likelihood-based methods with binary outcomes

Maximum likelihood and Bayesian estimates can be estimated with bi-
nary outcomes. If we assume a linear model of association between the
logit-transformed probability of an event (πi ) and the exposure (a logistic-
linear model), and a Bernoulli distribution for the outcome event, as in the
Methods for instrumental variable analysis 63

following model:
2
xi ∼ N (ξi , σX ) (4.19)
yi ∼ Bernoulli(πi )
K
X
ξi = α0 + αk gik
k=1
logit(πi ) = β0 + β1 xi
then the joint likelihood L is given by:
Y  y 1 1

1−yi 2
L= πi (1 − πi )
i
√ {exp(− 2 (xi − ξi ) )} . (4.20)
2πσX σX
i=1,...,N

Estimates can be obtained by maximization of the joint likelihood. As all coef-

ficients are simultaneously estimated, this is a full information maximum like-
lihood (FIML) approach. Alternatively, model parameters can be estimated
in a Bayesian framework, obtaining posterior distributions from the model by
MCMC methods.
Log-linear models can in principle be estimated in the same way, although
care is needed to ensure that the probabilities πi do not exceed 1 at any point
in the estimation.

4.3.5 Comparison of two-stage and likelihood-based methods

In the two-stage methods, the two stages are performed sequentially. The out-
put from the first-stage regression is fed into the second-stage regression with
no acknowledgement of uncertainty. In the likelihood-based methods, the two
stages are performed simultaneously: the α and β parameters are estimated
at the same time. Uncertainty in the first-stage parameters is acknowledged
and feedback between the regression stages is possible. The uncertainty in
the estimate of the causal parameter β1 is therefore better represented in
the likelihood-based approaches if there is non-negligible uncertainty in the
first-stage regression model.

4.4* Semi-parametric methods

A semi-parametric model has both parametric and non-parametric compo-
nents. Typically semi-parametric estimators with IVs assume a parametric
form for the equation relating the outcome and exposure, but make no assump-
tion on the distribution of the errors. Semi-parametric models are designed to
be more robust to model misspecification than fully parametric models [Clarke
and Windmeijer, 2010].
64 Mendelian Randomization

4.4.1* Generalized method of moments

The generalized method of moments (GMM) is a semi-parametric estimator
designed as a more flexible form of 2SLS to deal with problems of heteroscedas-
ticity of error distributions and non-linearity in the two-stage structural equa-
tions [Foster, 1997; Johnston et al., 2008]. With a single instrument, the esti-
mator is chosen to give orthogonality between the instrument and the residuals
from the second-stage regression. Using bold face to represent vectors, if we
have
E(Y ) = f (X; β) (4.21)
then the GMM estimate is the value of β such that:
X
(yi − f (xi ; β)) = 0 (4.22)
i
X
and gi (yi − f (xi ; β)) = 0
i

where the summation is across i, which indexes study participants. In

the linear (or additive) case, f (xi ; β) = β0 + β1 xi ; in the log-linear (or
multiplicative) case, f (xi ; β) = exp(β0 + β1 xi ); and in the logistic case,
f (xi ; β) = expit(β0 + β1 xi ); where β1 is our causal parameter of interest
and expit(z) = (1 − exp(−z))−1 , the inverse of the logit function. These two
equations can be solved numerically [Palmer et al., 2011b].
GMM estimates are sensitive to the parametrization of the model used.
For example, estimates from the estimating equations (4.22) and from:
X
yi f (xi ; β)−1 − 1 = 0


(4.23)
i
X
gi yi f (xi ; β)−1 − 1 = 0


and
i

may be different in finite samples, although they each assume the same struc-
tural model between Y and X.
When there is more than one instrument, gi becomes gik and we have a
separate estimating equation for each instrument k = 1, . . . , K. The orthog-
onality conditions for each instrument cannot generally be simultaneously
satisfied. The estimate is taken as the minimizer of the objective function

(y − f (x; β))T G(GT ΩG)−1 GT (y − f (x; β)) (4.24)

where G = (1 g1 . . . gK ) is the N by K + 1 matrix of instruments, including a

column of 1s for the constant term in the G–X association. Although this gives
consistent estimation for general matrix Ω, efficient estimation is achieved
when Ωij = cov(εi , εj ) (i, j = 1, . . . , N ), where εi is the residual yi − f (xi ; β)
[Hansen, 1982].
Methods for instrumental variable analysis 65

As the estimation of Ω requires knowledge of the unknown β, a two-step

approach is suggested. We firstly estimate β ∗ using (GT ΩG) = I, where I is
the identity matrix, which gives consistent but not P efficient estimation of β.
We then use ei = yi −f (xi ; β ∗ ) to estimate GT ΩG = i gi gi T ε2i as i gi gi T e2i
P
in a second-stage estimation [Johnston et al., 2008].

4.4.2* Structural mean models

The structural mean model (SMM) approach is another semi-parametric esti-
mation method designed in the context of randomized trials with incomplete
compliance [Robins, 1994; Fischer-Lapp and Goetghebeur, 1999]. (Technically,
g-estimation is the method by which a structural mean model is fitted, but
we refer to the approach as SMM [Robins, 1986; Greenland et al., 2008].) We
recall that the potential outcome Y (x) is the outcome which would have been
observed if the exposure X were set to x. In particular, the exposure-free out-
come Y (0)|X = x is the outcome which would have been observed if we had
set X to 0 rather than it taking its observed value of x [Clarke and Windmei-
jer, 2010]. Conditioning is performed on X = x so that no other variable is
changed from the value it would take if X = x. We note that the expectation
E(Y (0)|X = x) is typically different from the expected outcome if X = 0 had
been observed, as intervening on X alone would not change the confounder
distribution. An explicit parametric form is assumed for the expected differ-
ence in potential outcomes between the outcome for the observed X = x and
the potential outcome for X = 0. In the continuous case, the linear or additive
SMM is:
E(Y (x)) − E(Y (0)|X = x) = β1 x (4.25)
and β1 is taken as the causal parameter of interest. In the context of non-
compliance in randomized trials, this is referred to as the ‘effect of treatment
on the treated’ [Dunn et al., 2005].
As the expected exposure-free outcome E(Y (0)|X = x) is statistically
independent of G, the causal effect is estimated as the value of β1 which
gives zero covariance between E(Y (0)|X = x) = E(Y (x) − β1 x) and G. The
estimating equations are:
X
(gik − ḡk )(yi − β1 xi ) = 0 k = 1, . . . , K (4.26)
i

1
P
where ḡk = N i gik and the summation is across i, which indexes study
participants.
Where the model for the expected outcomes is non-linear, this is known
as a generalized structural mean model. With a binary outcome, it is natural
to use a log-linear (or multiplicative) SMM:

log E(Y (x)) − log E(Y (0)|X = x) = β1 x (4.27)

Due to non-collapsibility of the odds ratio, the logistic SMM cannot be

66 Mendelian Randomization

estimated in the same way, as the expectation logit E(Y (x)) depends on the
distribution of the IV [Robins, 1999]. This problem can be addressed by es-
timating Y (x) assuming an observational model [Vansteelandt and Goetghe-
beur, 2003]:
logit E(Y (x)) = β0a + β1a x (4.28)
where the subscripts a indicate associational, as well as a structural model:

logit E(Y (x)) − logit E(Y (0)|X = x) = β1c x (4.29)

where the subscript c indicates causal. The associational parameters can be

estimated by logistic regression, leading to estimating equations:
X
(gik − ḡk ) expit(Ŷ (x) − β1c xi ) = 0 k = 1, . . . , K (4.30)
i

where logit Ŷ (x) = β̂0a + β̂1a x [Vansteelandt et al., 2011].

We note that the choice of estimating equations presented here is not the
most efficient, but leads to consistent estimates [Vansteelandt and Goetghe-
beur, 2003]. In the general case, the linear (additive) and log-linear (multi-
plicative) GMM and SMM approaches give rise to the same estimates. This
is not true in the logistic case [Clarke and Windmeijer, 2010].

4.4.3* Lack of identification with binary outcomes

An issue with semi-parametric IV estimation in practice is lack of identification
of the causal parameter. A parameter in a statistical model is identified if an
estimate of its value can be uniquely determined on the basis of the data.
For a semi-parametric instrumental variable analysis with a binary outcome,
the causal parameter of interest may not be identified; there may be multiple
or no parameter values which satisfy the estimating equations [Burgess et al.,
2014c]. This is especially likely if the IV is weak (Section 4.5.2). Consequently,
estimates and standard errors reported by automated commands for GMM or
SMM estimation can be misleading.
It is recommended that investigators wanting to use a GMM or SMM
approach should plot the relevant estimating equations for a large range of
values of the parameter of interest to check if there is a unique solution. If
there is not, this should be reported as an indication that there is a lack of
information on the parameter in the data. An alternative estimation technique
can be used, such as a two-stage method, but identification will be rely on
stronger assumptions.
Methods for instrumental variable analysis 67

4.5 Efficiency and validity of instruments

Having discussed the methods for IV estimation, we present statistical ap-
proaches to improve the efficiency of estimates, and to test the validity of
IVs.

4.5.1 Use of measured covariates

If we can find measured covariates which explain variation in the exposure or
a continuous outcome, and which are not correlated with the IV nor on the
causal pathway between exposure and outcome, then we can incorporate such
covariates into our analysis. In econometrics, such a variable is called an ex-
ogenous regressor or included instrument, as opposed to an IV, which is called
an excluded instrument [Baum et al., 2003]. This is because the covariate is
included in the second-stage regression model for the outcome. Incorporation
of covariates generally increases efficiency and hence the precision of the causal
estimate. However, it may lead to bias in the causal estimate if the covariate
is on the causal pathway between exposure and outcome, or if the analysis
model including the covariate is misspecified. In a two-stage estimation, any
covariate adjusted for in the first-stage regression should also be adjusted for
in the second-stage regression [Wooldridge, 2009]; failure to do so can cause
associations between the IV and confounders leading to bias [Angrist and
Pischke, 2009, page 189].

4.5.2 Weak instruments

A ‘weak instrument’ is defined as an IV for which the statistical evidence
of association with the exposure is not strong [Lawlor et al., 2008]. An IV is
weak if it explains only a small amount of the variation of the exposure, where
the amount defined as ‘small’ depends on the sample size. The F statistic in
the regression of the exposure on the IV (also known as the Cragg–Donald F
statistic [Baum et al., 2007]) is usually quoted as a measure of the strength of
an instrument [Stock et al., 2002]. Although IV methods are asymptotically
unbiased, they typically demonstrate systematic finite sample bias, typically
in the direction of the observational (confounded) association between the ex-
posure and outcome. The bias of the IV estimate from the two-stage method
with a continuous outcome is approximately 1/E(F ) of the bias of the observa-
tional association, where E(F ) is the expected F statistic from the first-stage
regression.
IVs with an F statistic less than 10 are often labelled as ‘weak instruments’
[Staiger and Stock, 1997]. The value 10 was chosen as this limits the bias of the
two-stage IV estimate to 10% of the bias of the observational association. Such
characterization of IVs is misleading for several reasons. First, it gives a binary
68 Mendelian Randomization

classification of IVs as either weak or strong based on an arbitrarily chosen

threshold F statistic, whereas the true magnitude of bias relates to instrument
strength in a continuous way. Secondly, the F statistic is not simply a measure
of the intrinsic strength of the IV (unlike the coefficient of determination R2 )
as it depends on the sample size. Labelling an IV as a weak instrument
leads researchers to think that ‘weak instrument bias’ is due to an intrinsic
property of the instrument, whereas any instrument can be made stronger
by increasing the sample size. Thirdly, the measured F statistic in a given
dataset is an unreliable guide to the true strength of an instrument, due to
the large sampling variability of the F statistic. Fourthly, the use of rules
for the post hoc selection of data based on measured F statistics can lead
to more bias than it prevents [Burgess et al., 2011b]. Fifthly, the threshold
was determined based on the 2SLS method, and is not necessarily relevant to
other IV methods. Indeed, the F statistic may not even be a reliable measure
of instrument strength for obtaining identification in a semi-parametric model
[Burgess et al., 2014c].
It is the authors’ view that weakness in instruments is best combatted
through a priori specification of the variable(s) used as IVs in the analysis
and careful choice of analysis method. Further advice on weak instruments is
given in Chapter 7.

4.5.3 Overidentification tests

When more than one instrument is used, an overidentification test, such as the
Basmann test [Basmann, 1960] or Sargan test [Sargan, 1958], can be carried
out to test whether the instruments have additional effects on the outcome
beyond that mediated by the exposure. Overidentification means that the
number of instruments used is greater than the number of exposures measured.
The latter is almost always one in Mendelian randomization, so when there
is more than one IV, separate causal estimates can be calculated using each
IV in turn. The overidentification test assesses whether these IV estimates are
compatible, or equivalently whether the IVs have residual associations with
the outcome once the main effect of the exposure has been removed [Wehby
et al., 2008]. Such a residual association may indicate that at least one of the
IVs has a pathway of association with the outcome not via the exposure (such
as via another risk factor), meaning that the IV assumptions may be violated.
Overidentification tests are omnibus tests, where the alternative hypoth-
esis includes failure of the IV assumptions for one IV, failure for all IVs, a
non-linear relationship between the exposure and outcome, and that different
variants identify different magnitudes of causal effect (treatment effect het-
erogeneity, Section 8.5.1) [Baum et al., 2003]. They generally have low power
and so have limited practical use in detecting violations of the IV assumptions
[Glymour et al., 2012].
Methods for instrumental variable analysis 69

4.5.4 Endogeneity tests

Some applied Mendelian randomization analyses have reported on whether
there is a difference between the observational and IV estimates as the primary
outcome of interest [Hingorani and Humphries, 2005]. This can be formally
tested using the Durbin–Wu–Hausman test [Baum et al., 2003]. This is a test
of equality of the observational and IV estimates, where a significant result
indicates disagreement between the two estimates. Such a test is known as an
endogeneity test (see Table 2.1).
While an informal comparison of the observational and causal estimates
may be reasonable, there are several reasons why reliance on an endogene-
ity test as a primary analysis result is not recommended in practice. If a
non-significant result is achieved, it would be fallacious to assume that the
exposure was exogenous, that is to assume that the observational associa-
tion is unconfounded. A non-significant result may simply reflect the limited
power of the test. If a significant result is achieved, this does not imply that
there is no causal effect. There may be a causal effect, but this may be dif-
ferent in magnitude to the observational association. The conclusion from a
significant endogeneity test is that the exposure is endogenous, and so there is
confounding. If a researcher believed that there was no (residual) confounding,
then they would be content with interpreting the observational association as
causal, and IV analysis would be unnecessary. For this reason, it is more ap-
propriate to consider the presence or absence of a causal effect as the subject
of investigation [Thomas et al., 2007]. The confidence interval of the causal es-
timate gives the researcher bounds on the plausible size of any possible causal
effect.

4.6 Computer implementation

Several commands are available in statistical software packages for IV estima-
tion, such as Stata [StataCorp, 2009], SAS [SAS, 2004], and R [R Development
Core Team, 2011]. We assume that the reader is familiar enough with the
software to calculate the ratio method ‘by hand’ (that is without the use of
pre-written software commands). Code is also given below for the estimation
of Bayesian models in WinBUGS [Spiegelhalter et al., 2003].

4.6.1 IV analysis of continuous outcomes in Stata

The commands in Stata ivreg, ivreg2, ivhettest, overid, and ivendog,
have been written to implement the 2SLS, LIML and GMM methods, with
estimators and tests, including the Cragg–Donald F statistic (weak instru-
ments) and the Sargan statistic (overidentification) [Baum et al., 2003]. The
70 Mendelian Randomization

main command in Stata for IV analysis is ivreg2. If the exposure is x, the

outcome is y and the IV is g, the syntax for a 2SLS analysis is:
ivreg2 y (x=g)
The syntax for a LIML analysis is:
ivreg2 y (x=g), liml
The syntax for a (linear) GMM analysis is:
ivreg2 y (x=g), gmm
In the output of the ivreg2 command, several additional results are displayed,
as follows:

The underidentification test assesses whether the IV is sufficiently associ-

ated with the exposure to give reliable identification of the causal parameter.
A parameter is formally identified if the data-generating model corresponds
to a unique set of parameter values. Poor identification means that there are
multiple parameter values which fit the data well. Underidentification tests
are rarely performed in practice, and are unlikely to be useful in Mendelian
randomization, as there is typically only one parameter of interest, and un-
deridentification would be reflected in a wide confidence interval for this pa-
rameter.
Critical values for the F statistic in determining the potential impact of
the weakness of the IV are provided. The values cited are from a simulation
study [Stock and Yogo, 2002], where the authors sought to improve on the
general arbitrary threshold of 10 for a weak instrument to give more accurate
bias and coverage thresholds for different numbers of IVs. With a single IV,
there is no threshold limiting relative bias in the 2SLS method. The coverage
thresholds cited relate to the coverage of the IV estimate. For example, with
a single IV, an F statistic of 16.38 or greater is needed to guarantee that the
95% confidence interval will exclude the true parameter value no more than
10% of the time, compared to the nominal 5%. The coverage thresholds were
calculated, however, assuming an unrealistically large correlation between the
exposure and the outcome. This means that the coverage levels should be more
conservative than the upper bounds cited from Stock and Yogo, although there
may well be some undercoverage of confidence intervals from the 2SLS method
with weak instruments (Section 4.1.5). The large sampling variability in the
F statistic, and the selection bias induced by data-driven procedures based on
the measured value of the F statistic (see Chapter 7), mean that the apparent
precision of these cited threshold values should not be relied on to protect
against bias.
Overidentification tests are described above (Section 4.5.3). With a single
IV, an overidentification test is not possible. In the 2SLS analysis, the Sargan
test alone is given. In the LIML and GMM analyses, different overidentification
tests are computed.
Methods for instrumental variable analysis 71

The commands ivregress 2sls y (x=g) and ivreg y (x=g) give the
same estimates as ivreg2 y (x=g), but a more limited output. The command
ivhettest performs a test of heteroscedasticity of the errors in the second-
stage regression. If heteroscedasticity is present, a GMM analysis with robust
standard errors is preferred to a 2SLS analysis. The command overid gives
more information about overidentification tests. The command ivendog gives
more information about endogeneity tests. The command qvf has been written
to implement a fast bootstrap estimation of standard errors for IV analysis
[Hardin et al., 2003]. Each of these commands can be used with multiple
instruments, for example ivreg2 y (x=g1 g2 g3).

4.6.2 IV analyses of binary outcomes in Stata

With a binary outcome and a logistic-linear model, the two-stage estimates
can be obtained by the commands:

reg x g
predict xhat
logit y xhat, robust
where robust standard errors are calculated in the second-stage regression.
Generic estimating equations for GMM or SMM analyses can be solved in
Stata using the gmm command [Drukker, 2009]. For example, a linear GMM
estimate can be obtained using:

gmm (y - {beta0} - x*{beta1}), instruments(g)

A logistic GMM estimate can be obtained using:

gmm (y - invlogit({beta0} + x*{beta1})), instruments(g)

A log-linear (multiplicative) GMM estimate can be obtained using the com-
mand ivpois:

ivpois y, endog(x) exog(g)

The same log-linear GMM estimate can be obtained using the gmm command:

gmm (y*exp(-x*{beta1}) - {beta0}), instruments(g)

An alternative log-linear GMM estimate can be estimated using:

gmm (y - exp({beta0}+x*{beta1})), instruments(g)

These GMM models assume the same structural relationship between the
exposure and outcome, but give different answers in finite samples (Sec-
tion 4.4.1*). The first formulation of the log-linear GMM model is equivalent
to a log-linear SMM [Palmer et al., 2011b].
72 Mendelian Randomization

Useful notes are available for the estimation of SMMs with a binary out-
come [Clarke et al., 2011]. Each of these commands can be used with multiple
instruments: for example gmm (y - beta0 - x*beta1), instruments(g1
g2). The command qvf can also be used in non-linear cases [Hardin et al.,
2003], such as a two-stage analysis with a non-linear second-stage model, to
prevent the over-precision of estimates resulting from a sequential regression
method. A probit IV model (not considered in this book) can be estimated
using the command ivprobit.

4.6.3 IV analysis in SAS

The command proc syslin in SAS has been written to implement the 2SLS,
FIML, and LIML methods:

proc syslin data=in 2sls;

endogenous x;
instruments g;
model y=x;
run;
where 2sls can be replaced by liml or fiml as appropriate.

4.6.4 IV analysis in R
The R command tsls in the library sem carries out a 2SLS procedure [Fox,
2006]. Care must be taken as the constant term usually used in regression
equations is not included by default. If the exposure is x, the outcome is y
and the IV is g, the syntax for a two-stage (2SLS) analysis is:

tsls(y, cbind(x, rep(1, length(x))), cbind(g, rep(1, length(g))),

w=rep(1, length(x)))
where w are the weights, here set to 1 for each individual. Also available are
the function ivreg in the aer package [Kleiber and Zeileis, 2014], and the
ivpack package with some additional functions, such as implementation of the
Anderson–Rubin confidence intervals [Small, 2014].
In a sequential regression two-stage analysis of a binary outcome in a case-
control setting, inference on the controls only (where Y = 0) can be made
using the predict function:

g0=g[y==0]
glm(y~predict(lm(x[y==0]~g0), newdata=list(g0=g)), family=binomial)
Generic estimating equations for GMM or SMM can be solved in R us-
ing the gmm package [Chaussé, 2010]; details and sample code are available
[Clarke et al., 2011].
Methods for instrumental variable analysis 73

4.6.5 IV analysis in WinBUGS

Bayesian analyses can be performed in WinBUGS. The following annotated
code can be used with a continuous outcome. We here assume vague pri-
ors for all the parameters: Normal(0,106 ) for the regression parameters, Uni-
form(0,20) for standard deviations, and Uniform(−1,1) for correlations, to
mimic a likelihood-based analysis. These could be changed for particular
datasets, to be better represent ‘non-informative’ priors or to alternatively
to represent prior information.

model {
beta1 ~ dnorm(0, 1E-6)
beta0 ~ dnorm(0, 1E-6)
alpha0 ~ dnorm(0, 1E-6)
for (k in 1:K) { alpha1[k] ~ dnorm(0, 1E-6) }
xsd ~ dunif(0, 20)
ysd ~ dunif(0, 20)
rho ~ dunif(-1, 1)
# priors for the parameters
xtau <- pow(xsd, -2)
ytau <- pow(ysd, -2)
tauy <- ytau/(1-pow(rho,2))
# tauy is the precision of y conditional on x
for (i in 1:N) {
ksi[i] <- alpha0 + inprod(alpha1[1:K], g[i,1:K])
x[i] ~ dnorm(ksi[i], xtau)
eta[i] <- beta0 + beta1 * ksi[i]
muy[i] <- eta[i] + sqrt(xtau/ytau)*rho*(x[i]-ksi[i])
# muy[i] is the mean of y[i] conditional on x[i]
y[i] ~ dnorm(muy[i], tauy)
} }
In the above, the bivariate normal distribution of (X, Y )T from equation (4.18)
has been equivalently replaced by the marginal distribution of X and the
conditional distribution of Y |X = x [Burgess and Thompson, 2012].
In a case-control study with a binary outcome and a logistic model of
association, the following code can be used (the first P individuals in the
dataset are the controls):

model {
beta1 ~ dnorm(0, 1E-6)
beta0 ~ dnorm(0, 1E-6)
alpha0 ~ dnorm(0, 1E-6)
for (k in 1:K) { alpha1[k] ~ dnorm(0, 1E-6) }
74 Mendelian Randomization

xtau <- pow(xsd, -2)

xsd ~ dunif(0, 20)
for (i in 1:P) { x[i] ~ dnorm(ksi[i], xtau) }
# where P is the number of controls
for (i in 1:N) {
ksi[i] <- alpha0 + inprod(alpha1[1:K], g[i,1:K])
logit(pi[i]) <- beta0 + beta1 * ksi[i]
y[i] ~ dbern(pi[i]) } }

4.7 Summary
Methods for IV analysis range from the very simple (calculate the difference
between two pairs of numbers and divide one by the other) to the more com-
plicated. The development of complex methods has been driven by the desire
to produce efficient estimates, for example by integrating data on multiple
IVs, to allow for more flexible modelling assumptions, or to provide robust-
ness against misspecification of modelling assumptions. Each method has its
own advantages and disadvantages. The properties of many of these estima-
tors will be discussed in the chapters to come in the specific contexts of weak
instruments, binary outcomes, and evidence synthesis.
In the next chapter, we consider examples of the use of Mendelian ran-
domization, focusing particularly on practical aspects of the analyses, such as
study design, and their impact on methods.
5
Examples of Mendelian randomization
analysis

Having discussed several of the statistical issues regarding Mendelian random-

ization analyses, in this chapter we present four published examples of the use
of Mendelian randomization from the literature, commenting on interesting
features of the analysis which help to clarify the methodology and aid readers
performing similar investigations.

5.1 Fibrinogen and coronary heart disease

The paper entitled “Fibrinogen and coronary heart disease: test of causal-
ity by ‘Mendelian randomization’ ” [Keavney et al., 2006] assesses the causal
effect of fibrinogen on risk of coronary heart disease (CHD). Fibrinogen is
observationally associated with CHD risk, although the magnitude of associ-
ation attenuates on adjustment for age and sex, and further on adjustment
for other covariates such as smoking and body mass index. When the plasma
apolipoprotein B/A1 ratio is additionally adjusted for, the observational as-
sociation is compatible with the null. However, it may be that some of the
variables adjusted for are on the causal pathway between fibrinogen and CHD,
and so this may represent an over-adjustment (Section 3.1.4).

5.1.1 Study design

Two approaches are proposed for the assessment of causality using Mendelian
randomization. First, the authors analyse individual participant data from
a case-control study, the International Studies of Infarct Survival (ISIS).
ISIS contains 4685 cases with confirmed myocardial infarction (MI) and 3460
disease-free control participants with measurements of fibrinogen levels. Sec-
ondly, they conduct a meta-analysis for the association between a particular
genetic variant and the risk of CHD, following a literature-based search for
relevant summary genetic estimates. The meta-analysis contains 20 studies

75
76 Mendelian Randomization

measuring beta-fibrinogen genotypes, including the original study, comprising

a total of 12 220 CHD cases and 18 716 controls.
In the context of a disease outcome, a case-control design may be necessary
if the outcome of interest is not common, as the power of the analysis depends
on the precision of the estimate of the gene–outcome association, which in turn
depends on the number of cases. Although measurement of the exposure in the
cases is unreliable due to possible reverse causation, the genetic instrumental
variable is not affected by the outcome, and so the gene–outcome association
can be reliably estimated in a case-control study (Section 2.2.1).

5.1.2 Genetic instruments

A single genetic variant is used as an instrumental variable (IV). This vari-
ant is a single nucleotide polymorphism (SNP) in the beta-fibrinogen gene
promoter which regulates fibrinogen production, giving some biological cred-
ibility to its specific association with fibrinogen, and therefore its validity as
an IV. Tests of association between the variant and a range of confounders
show no strong associations, except for that with plasma apolipoprotein B/A1
ratio. Although the p-value of 0.01 is not particularly extreme in view of the
multiple comparisons, and would not be judged conventionally significant us-
ing a threshold of p = 0.05 and a standard Bonferroni correction procedure,
the result may indicate a pleiotropic association of the variant with fibrinogen
and the plasma apolipoprotein B/A1 ratio. This would be problematic if the
Mendelian randomization estimate indicated a causal relationship, as it would
not be possible empirically to distinguish between causal effects of fibrinogen
and of the plasma apolipoprotein B/A1 ratio on CHD risk. Alternatively, it
may be that changes in the plasma apolipoprotein B/A1 ratio associated with
the genetic variant are not directly associated with the genetic variant, but
occur as a result of the increase in fibrinogen levels. This would mean that
the Mendelian randomization analysis was valid, as a clinical intervention on
fibrinogen levels would also increase the plasma apolipoprotein B/A1 ratio. If
the plasma apolipoprotein B/A1 ratio is a mediator on the causal pathway
from fibrinogen to CHD risk, it should not be adjusted for in the observational
analysis.
It is not possible to differentiate between the association of the variant
with the plasma apolipoprotein B/A1 ratio being a chance finding, evidence
of pleiotropy of the genetic variant, or evidence of a causal pathway from
fibrinogen.

5.1.3 Statistical methodology

In both the single study and meta-analysis, the causal effect of fibrinogen
on CHD risk is assessed, but no causal parameter is estimated. In the single
study, the association of fibrinogen levels with the genetic variant is estimated
in control participants using linear regression, and the association of CHD
Examples of Mendelian randomization analysis 77

risk with the variant is estimated in the whole study population using lo-
gistic regression. In the meta-analysis, summary-level data from each study
on the number of cases and controls in each genetic subgroup are used to
estimate the association of the variant with the risk of CHD in each study.
The study-specific estimates are then combined using a fixed-effect inverse-
variance weighted meta-analysis (Chapter 9). A per allele genetic model is
used, as this is best supported by the data on fibrinogen levels. In both cases,
the result is cited as the risk ratio of CHD per additional variant allele.

5.1.4 Results
The analyses show a null association of the variant with CHD risk, with a
narrow confidence interval (CI) for the genetic association with disease risk:
risk ratios of 1.06 (95% CI 0.96 to 1.16) per fibrinogen-increasing allele in ISIS
alone and of 1.00 (95% CI 0.95 to 1.04) in the meta-analysis. On the basis
of this, the authors conclude that “these genetic results provide strong evi-
dence that long-term differences in fibrinogen concentrations are not a major
determinant of coronary disease risk”.

5.1.5 Commentary
A weakness of the presentation of the results is that a causal estimate of the
effect of fibrinogen on CHD risk is not presented. Although the risk ratio
estimate of 1.00 (95% CI, 0.95 to 1.04) per additional allele appears to be a
small effect, each additional allele is only associated with a small increase in
fibrinogen levels (0.14 (standard error 0.024) g/l), meaning that a standard
deviation increase in fibrinogen (0.81 g/l, estimated in control participants)
could still lead to an approximate 25% increase in CHD risk based on the upper
bound of the 95% CI (assuming a log-linear relationship between fibrinogen
and the risk of CHD).

5.2 Adiposity and blood pressure

The paper “Does greater adiposity increase blood pressure and hypertension
risk? Mendelian randomization using the FTO /MC4R genotype” [Timpson
et al., 2009] considers the causal effect of adiposity on blood pressure. Adipos-
ity is observationally associated with blood pressure, although there are many
potential confounders that may bias the observational estimate. Randomized
trials of weight reduction have shown related decreases in blood pressure, but
such interventions may additionally affect other variables, such as physical
activity and diet. Although the prevalence of obesity has increased over time,
the secular trend in blood pressure and hypertension has been in the opposite
78 Mendelian Randomization

direction, leading some to question whether the observational association is

in fact causal. Hypertension (severe hypertension) was defined as a systolic
blood pressure of over 140 mmHg (over 160 mmHg for severe hypertension),
a diastolic blood pressure of over 90 mmHg (over 100 mmHg), or (in both
cases) the taking of antihypertensive drugs.

5.2.1 Study design

The authors analyse cross-sectional data on 37 027 unrelated individuals from
a population-based study, the Copenhagen General Population Study. All par-
ticipants are of the same ethnic background (Danish), and were selected to
reflect the composition of the general population of Copenhagen.
For an outcome that is a continuous trait rather than a disease outcome,
a cross-sectional study is able to provide all the information necessary for
a Mendelian randomization experiment without necessitating the expense of
following up participants over a period of time. A further advantage of a well-
designed population study is increased external validity, whereby an estimate
from a Mendelian randomization study represents an effect estimate for a
cohort similar to the population on whom an intervention could be performed.

5.2.2 Genetic instruments

Two genetic variants are used as IVs. The SNPs are located in the FTO
and MC4R loci, which have been shown to be associated with body mass
index (BMI) in a number of previous studies. The precise functions of the two
genetic regions are unknown, although variation in the FTO gene is known to
be linked with food intake [Wardle et al., 2008].
Although knowledge of the function of genetic variants is not necessary for
Mendelian randomization, instrumental variable analysis with variants of un-
known function can be problematic to interpret. As the instrumental variable
assumptions are scientifically more uncertain, a conclusion that the specific
risk factor of interest is in fact the causal agent is less reliable. This is especially
true for a risk factor such as BMI, in the same way that a single causal agent in
a randomized trial for weight loss is difficult to isolate. Unlike a biomarker such
as fibrinogen, there is no single regulatory gene for “BMI production” or for a
“BMI receptor”. There are additional difficulties in comparing the Mendelian
randomization estimate to the effect of a potential clinical intervention, as the
intervention and genetic effect on BMI reduction may have different pathways
of action. It may be that there is heterogeneity in the proportional effect of
changes in BMI on blood pressure and hypertension as instrumented by dif-
ferent variants resulting from differences between pathways (treatment effect
heterogeneity). For example, if there were several genetic variants used in the
analysis, it is possible for some of the variants to be associated with changes
in BMI that do affect blood pressure, and some to be associated with changes
that do not.
Examples of Mendelian randomization analysis 79

5.2.3 Statistical methodology

The causal effect of adiposity on blood pressure is estimated using the gen-
eralized method of moments (GMM). Adiposity is represented by ‘relative
BMI’, calculated as the ratio of an individual’s observed BMI to predicted
BMI from a linear regression model on age, sex and height. Results are also
calculated using the two-stage least squares (2SLS) and limited information
maximum likelihood (LIML) methods; similar results are obtained from each
method. The observational and IV estimates of association are compared us-
ing a Durbin–Wu–Hausmann test of the equality of the observational and IV
estimates. As we discuss in Section 4.5.4, we do not advocate such tests, as
there is a multitude of reasons for the estimates to be different unrelated to
the question of causality, and neither a significant nor a non-significant finding
is directly interpretable as evidence for or against a causal effect.

5.2.4 Results
The IV analysis shows a positive causal effect of BMI on blood pressure and
hypertension of similar magnitude to the observational association. For exam-
ple, the estimate for the increase in systolic blood pressure associated with a
10% increase in BMI is 2.75 mmHg (95% CI 2.62 to 2.88) from the observa-
tional analysis with adjustment for age, sex and height, and 2.54 mmHg (95%
CI 2.39 to 2.69) with further adjustment for socio-behavioural factors. The
corresponding estimate of the increase in systolic blood pressure caused by
a 10% increase in BMI from the IV analysis is 3.85 mmHg (95% CI 1.88 to
5.83). The FTO SNP has statistically robust associations with BMI (1.18%
[95% CI 0.96 to 1.41] increase in BMI on a multiplicative scale per additional
allele) and with blood pressure (0.63 mmHg [95% CI 0.33 to 0.93] increase in
systolic blood pressure per additional allele), whereas the MC4R SNP has a
smaller magnitude of association with BMI (0.78%, 95% CI 0.53 to 1.04), and
an association with blood pressure compatible with the null (0.20 mmHg, 95%
CI -0.14 to 0.54). This may be due to the MC4R SNP’s reduced association
with BMI and the statistical uncertainty in the association estimates, but it
may reflect heterogeneity of the causal effects identified by the two variants.
The association of the FTO SNP with severe hypertension does not fully
attenuate on adjustment for BMI: attenuation from an odds ratio of 1.07 (95%
1.04 to 1.11) on adjustment for age and sex, to 1.07 (95% 1.03 to 1.11) on
additional adjustment for socio-behavioural factors, and to 1.04 (95% 1.01 to
1.08) on additional adjustment for log(BMI). Although a complete attenuation
is not expected, the limited attenuation suggests that the causal effect of
adiposity on hypertension may not simply be explained as a function of BMI.
The Durbin–Wu–Hausmann tests for each variant are not significant, indi-
cating no difference between the observational and IV estimates beyond that
compatible with chance.
80 Mendelian Randomization

5.2.5 Commentary
Although the Mendelian randomization analysis suggests that adiposity is
causally associated with blood pressure, the unknown function of the genetic
variants limits the certainty of the conclusions that can be drawn.

5.3 Lipoprotein(a) and myocardial infarction

The paper “Genetically elevated lipoprotein(a) and increased risk of myocar-
dial infarction” [Kamstrup et al., 2009] examines the causal effect of lipopro-
tein(a) [denoted lp(a)] on the risk of myocardial infarction (MI). Lipopro-
tein(a) is an assembly of a lipid, essentially a low-density lipoprotein (LDL)
particle, and a protein, known as apolipoprotein(a). Concentrations of lp(a)
vary widely between individuals and are highly heritable.

5.3.1 Study design

The authors analyse data from three related studies of Danish participants:
a prospective study with 16 years of follow-up, the Copenhagen City Heart
Study, comprising 9867 participants with genetic data of whom 4514 have
a lp(a) plasma level measurement and 599 suffered a MI event during the
follow-up period; a cross-sectional study, the Copenhagen General Population
Study, comprising 29 388 participants with genetic data of whom 5543 have
a lp(a) plasma level measurement and 994 suffered a MI event in a defined
period prior to study entry; and a case-control study, the Copenhagen Ischemic
Heart Disease Study, comprising 1231 participants with genetic data and a
MI event, and 1230 matched controls taken from the Copenhagen City Heart
Study (which reduces the effective size of that study to 8637 participants).
By combining evidence from prospective, cross-sectional and case-control
designs, the advantages of each approach are exploited. The prospective study
measured lp(a) levels at a range of timepoints, enabling assessment of the
long-term associations of genetic variation. The cross-sectional study is the
simplest study design, enabling assessment of the genetic association with
the exposure in a large population. The case-control study design has known
potential weaknesses, including selection bias, but enables more precise esti-
mation of the genetic association with the outcome in a sample enriched for
cases. Although lp(a) levels were not measured in all participants, this does
not invalidate findings of the Mendelian randomization experiment, and may
even be a worthwhile design strategy if the exposure is difficult or expensive
to measure (see Section 8.5.2).
Examples of Mendelian randomization analysis 81

5.3.2 Genetic instruments

In this study, the genetic variant is not a SNP, but a copy number variant in
the LPA gene, the kringle IV type 2 (KIV-2) size polymorphism. Individuals
have a variable number of repeating sections of DNA known as kringle repeats,
and this number correlates inversely with lp(a) concentration. There is good
biological plausibility for the use of the polymorphism as an IV. (The IV
in kringle IV type 2 is the Roman numeral 4, rather than the abbreviation
for instrumental variable.) While the two variants in the previous example
explained less than 1% of the variation in BMI, the KIV-2 polymorphism here
explains more than 20% of the variation in lp(a).

5.3.3 Statistical methodology

Two approaches are taken to assess and estimate the causal effect of lp(a)
on MI risk. First, the association between the IV and MI risk is assessed in
each of the datasets. To address potential non-linearity, the IV is defined by
dividing the population into quartiles based on the number of kringle repeats.
In the prospective study, the association is assessed using Cox proportional
hazards regression with adjustment for a range of covariates. In the cross-
sectional and case-control studies, logistic and matched logistic regression are
used. Adjustment is made for a limited set of covariates which would not be
thought to be affected by potential reverse causation, such as age, sex and
diabetes status. Secondly, a formal IV method is conducted in the prospective
study only, using the average level of lp(a) and the risk of MI in the top and
bottom quartiles of the IV to construct a ratio estimate. Confidence intervals
are evaluated using Fieller’s theorem (Section 4.1.5).

5.3.4 Results
The analyses show a positive causal effect of lp(a) on MI risk. The odds
ratios of MI in the quartiles of the IV (fourth quartile is reference group)
were 1.3 (95% CI, 1.1 to 1.5) in the first quartile, 1.1 (95% CI, 0.9 to 1.3)
in the second quartile, and 0.9 (95% CI, 0.8 to 1.1) in the third quartile in
the Copenhagen General Population Study (p = 0.005 for trend), and 1.4
(95% CI, 1.1 to 1.7), 1.2 (95% CI, 1.0 to 1.6), and 1.3 (95% CI, 1.0 to 1.6)
in the Copenhagen Ischemic Heart Disease Study (p = .01 for trend). In the
Copenhagen City Heart Study, the IV estimate for the hazard ratio (HR)
of MI per doubling of lp(a) (HR 1.22, 95% CI 1.09 to 1.37) is considerably
larger than the observational estimate (HR 1.08, 95% CI 1.03 to 1.12). This
finding, which was replicated in a similar study [Clarke et al., 2009], may
reflect the increased effect of lifelong differences in lp(a) levels, similar to that
observed for low-density lipoprotein cholesterol (LDL-C) (Section 6.2.1). It
also may result from the association of the KIV-2 polymorphism with both
the concentration of lp(a) and the lp(a) particle size, which is also implicated
82 Mendelian Randomization

as a potential risk factor for MI. In the absence of further evidence, it is

difficult to disentangle these two variables.

5.3.5 Commentary
A limitation of the interpretation of the IV estimate is the non-linear associa-
tion of the number of kringle repeats with lp(a) levels. The IV estimate should
be interpreted as a population-averaged effect, comparing genetic subgroups
which have different average levels of the exposure. With non-linear relation-
ships, the IV estimate does not necessary represent the effect of intervening
on lp(a) for an individual (Section 11.1.2).

5.4 High-density lipoprotein cholesterol and myocardial

infarction
The paper “Plasma HDL cholesterol and risk of myocardial infarction: a
Mendelian randomisation study” [Voight et al., 2012] examines the causal
effect of high-density lipoprotein cholesterol (HDL-C) on risk of MI. As a
proof of concept, the causal effect of LDL-C on risk of MI is also assessed.

5.4.1 Study design

The authors analyse individual participant data from six prospective studies
and 14 cross-sectional studies, comprising 20 913 MI cases and 95 407 controls,
although assessment of the assumptions for IV analysis is performed in a larger
set of studies.

5.4.2 Genetic instruments

Two approaches are proposed for the assessment and estimation of the causal
effect of HDL-C on MI risk. First, a single SNP is used as an IV. This SNP is a
loss-of-function coding variant at the endothelial lipase gene which has known
functional association with HDL-C concentration, and does not show any as-
sociation with LDL-C or triglycerides in the dataset (p > 0.05). Secondly,
an allele score (or gene score, Section 8.2) is used, comprising 14 variants
associated with HDL-C (p < 5 × 10−8 ), but not associated with LDL-C or
triglycerides (p > 0.01). For comparison, an allele score comprising 13 variants
associated with LDL-C, but not HDL-C or triglycerides, is also constructed.
The reason for the two approaches is that the first is more scientifically rig-
orous, as the function of the variant used as an IV is known, whereas the
second gives more statistical power, as the allele score explains more of the
Examples of Mendelian randomization analysis 83

variation in the exposure than any of the constituent variants individually

(bias–variance trade-off). Another practical reason for including both analy-
ses is that the second analyses is performed in a smaller subset of participants,
comprising 12 482 MI cases and 41 331 controls, due to missing genetic data
on one or more variants (Section 8.4).
The dilemma between only including genetic variants where there is strong
evidence of their validity as instrumental variables, risking an underpowered
estimate, and including all variants even if their function is not fully known,
risking a biased estimate, is an example of a bias–variance trade-off. A sensible
compromise in practice is to present the estimate using fewer “safer” variants
as the primary analysis result, acknowledging the statistical imprecision in
the estimate, and to present the estimate using more variants as a secondary
analysis result, acknowledging both the statistical imprecision and the scien-
tific uncertainty in the assumptions necessary to interpret the estimate as a
causal effect.

5.4.3 Statistical methodology

In the first approach using a single variant, causal estimates from each of
the prospective studies are calculated using the qvf command in Stata to fit
two-stage logistic models with robust standard errors. In two of the studies,
a two-stage method is employed with sequential regression using generalized
estimating equations in the first-stage of the analysis, to account for related
individuals. These study-level causal estimates are combined in a fixed-effect
inverse-variance weighted meta-analysis. In the second approach, a weighted
allele score is constructed for both HDL-C and LDL-C using coefficients as
weights for the association of each variant with the exposure of interest taken
from a large meta-analysis. The association of the allele score with MI case
status is assessed using logistic regression in the cross-sectional studies. The
data source for the weights is not entirely independent from the data under
analysis, as some studies are included in both analyses (Section 8.2.1).

5.4.4 Results
From observational epidemiology, the expected odds ratio (OR) for each vari-
ant allele in the endothelial lipase gene is 0.87 (95% CI 0.84 to 0.91). This
is obtained by triangulation of the observed estimate of the association of
HDL-C on MI risk from multivariable adjusted logistic regression with the
observed genetic association of the variant with HDL-C (Section 3.3.3). How-
ever, the variant is not associated with risk of myocardial infarction (OR 0.99,
95% CI 0.88 to 1.11). With the allele score, the expected OR for a 1 standard
deviation increase in HDL-C from observational epidemiology (OR 0.62, 95%
CI 0.58 to 0.66) is not compatible with the estimated OR for a 1 SD increase in
HDL cholesterol from Mendelian randomization (OR 0.93, 95% CI 0.68-1.26).
For a 1 standard deviation increase in LDL-C, the observational epidemiology
84 Mendelian Randomization

(OR 1.54, 95% CI 1.45 to 1.63) and Mendelian randomization (OR 2.13, 95%
CI 1.69 to 2.69) estimates are directionally concordant.
The authors conclude that “some genetic mechanisms that raise plasma
HDL-C do not seem to lower risk of myocardial infarction”. This tentative
conclusion reflects the limited power of the single SNP analysis, where the
CIs for the causal effect and the observational estimate substantially overlap,
and the limited knowledge of the specific function of the allele score, which
may contain variants not exclusively or not directly associated with HDL-C.

5.4.5 Commentary
Aside from the limitations of the conclusions stated above, this paper demon-
strates the statistical difficulty of applied Mendelian randomization analysis
where the studies under analysis are heterogeneous. Although Mendelian ran-
domization investigations can be undertaken in a number of study designs,
differences between studies and specific features of each study may make in-
tegrated analysis of the entirety of the data available challenging.
The authors choose a pragmatic approach, combining a more conservative
analysis using a single genetic variant with a more speculative analysis using
an allele score. This is contrasted with a parallel analysis of LDL-C, which
provides plausibility of the allele score approach, as a positive causal effect of
LDL-C on MI risk is estimated.

5.5 Discussion
A question of interpretation relating to each of these analyses, and to
Mendelian randomization more widely, is how much weight of evidence to
attach to the result of a Mendelian randomization investigation. In a hierar-
chy of evidence, Mendelian randomization has been advocated as providing
“critical evidence” on exposure–outcome relationships [Gidding et al., 2012].
However, the true weight of evidence in each case depends strongly on the
plausibility of the instrumental variable assumptions for the genetic variants.
If the function of the genetic variants is poorly understood, then a causal con-
clusion is in doubt, particularly if there are multiple genetic variants and there
is little consistency in the causal effect estimates using each of the variants.
Equally, if the genetic variants explain a small proportion of the variance in the
exposure, then a Mendelian randomization investigation using those variants
will be inconclusive unless the sample size is very large.
For exposures which are biomarkers, variants can be employed as IVs which
are located in the gene coding the biomarker (such as for fibrinogen in the
example above). For exposures which are complex multifactorial traits, such
as body mass index and blood pressure, the association between the genetic
Examples of Mendelian randomization analysis 85

variants and exposure is less proximal, giving more opportunities for viola-
tions of the IV assumptions. A non-null Mendelian randomization estimate
is indicative that genetic predictors of the exposure are also associated with
the outcome, but there may be an alternative causal pathway other than that
through the exposure of interest. An analogous situation is inferring a causal
effect of a specific biomarker based on a pharmaceutical intervention with
multiple effects, such as those of statins on lipid fractions and inflammation
markers. The existence of pleiotropic associations of variants is particularly
likely if a large number of variants is included in the analysis.
In conclusion, the reliability of the findings from a Mendelian randomiza-
tion study depends heavily on both the validity of the variant(s) used as an
IV, and the power of the analysis to detect a clinically relevant causal effect.
Before further considering the statistical properties of IV estimators, in the
next chapter we consider a more fundamental question: what does a Mendelian
randomization estimate represent?
6
Generalizability of estimates from Mendelian
randomization

In the previous chapters, we have discussed the meaning of causation and pre-
sented methods and examples of estimating causal effects using instrumental
variables (IVs). In this chapter, we consider the interpretation of causal effects
assessed and estimated in Mendelian randomization, and address the question
of under what circumstances a Mendelian randomization estimate may be a
reliable guide to the effect of an intervention on the exposure of interest in
practice.

6.1 Internal and external validity

From the first discussions of Mendelian randomization, researchers have em-
phasized that the assumptions leading to the assertion of a causal relationship
may be invalid for many genetic variants. Violations in the assumptions of no
direct effect of the genetic variant on the outcome or of no association with a
confounding risk factor may occur for several reasons, as discussed in Chapter
3. Such violations of internal validity can potentially lead to misleading con-
clusions. An aspect of Mendelian randomization which is less well appreciated
is the issue of external validity. If the IV assumptions about the genetic vari-
ant are true and a valid estimate is made which corresponds to a causal effect,
what questions are raised in generalizing this estimate to an experimental
context? For example, is the estimate of lowered risk derived from considering
genetically reduced levels of cholesterol the same as the lowered risk conferred
by an intervention that reduces levels of cholesterol?
Mendelian randomization is different from a randomized trial in a funda-
mental way which impacts on questions of external validity [Rothwell, 2010].
In a randomized trial, the intervention applied to the treatment group is usu-
ally identical or similar to the intervention which is proposed to be applied
in clinical practice. In Mendelian randomization, the “intervention” leading
to differences between genetically-defined subgroups within the study is the
presence of a genetic variant. The question of external validity is whether the

87
88 Mendelian Randomization

causal effect due to the change in the exposure as a result of the presence of
the genetic variant is similar to the causal effect due to the proposed inter-
vention on the exposure. There are several reasons why these effects may be
unequal, as we now discuss.

6.1.1 Time-scale and developmental compensation

First, the presence or absence of the genetic variant in an individual is deter-
mined at conception. This means that the Mendelian randomization estimate
represents the result of a life-long difference in the exposure between the ge-
netic subgroups [Davey Smith, 2006]. In contrast, most clinical interventions
are performed on mature individuals. For some exposures, an individual may
develop compensatory mechanisms in response to long-term elevated (or low-
ered) levels of the exposure, known as canalization (Section 3.2.3).
Secondly, it may be that a stage of disease progression is irreversible. There
may be no intervention on the exposure in a mature cohort which can imitate
the genetic effect. This may be especially relevant if the genetic change in the
exposure affects intra-uterine or early-stage development.

6.1.2 Usual versus pathological levels

Secondly, the genetic variant would be expected to affect average or “usual”
levels of the exposure. This is often the target of interest for epidemiologists
interested in disease prevention. Mendelian randomization has a particular
role to play here, as typically life-long randomized trials affecting usual levels
of exposures cannot be undertaken. However, Mendelian randomization stud-
ies are unlikely to be informative about the acute response behaviour of an
individual to a stimulus, such as a sudden large increase in an inflammation
biomarker. It is plausible that long-term elevated average levels of an exposure
for an individual do not affect the outcome, but acute response of the expo-
sure does. The efficacy of short-term targeted interventions on pathological
levels of an exposure cannot be validly assessed by a Mendelian randomization
approach.
An example is that of C-reactive protein (CRP). Genetic variants which are
associated with usual levels of CRP have been used to assess the causal effect
of long-term elevated average levels of CRP on cardiovascular risk [Elliott et
al., 2009; CCGC, 2011]. Although the causal effect of CRP on cardiovascular
risk appears to be null, this does not preclude the efficacy of a therapeutic
intervention on acute levels of CRP.

6.1.3 Extrapolation of small differences

Thirdly, the change in an exposure due to genetic variants is generally small.
For evolutionary reasons, genetic variants associated with substantial changes
Generalizability of estimates from Mendelian randomization 89

in clinically relevant exposures are uncommon. Most genetic variants which

have been used in Mendelian randomization studies have explained in the
region of 1 to 4% of the variation in the exposure [Schatzkin et al., 2009;
Davey Smith, 2011]. If the target of interest for the epidemiologist is an in-
tervention lowering (or raising) the exposure uniformly by a small amount
for everyone in the population, then a Mendelian randomization study may
provide a relevant estimate of the effect of the intervention. However, if the
proposed intervention effect is more substantial, then the Mendelian random-
ization estimate relies on extrapolation beyond the genetic change in the ex-
posure observed. Estimates relying on a linear assumption for the effect of the
exposure on the outcome may not be valid; moreover this assumption may
not be testable from empirical data.

6.1.4 Different pathways of genetic and intervention effects

Fourthly, the genetic variant and the proposed intervention will not, in general,
have the same specific mechanism of effect on the exposure. The genetic change
in the exposure may be associated with another variable, as in the case of a
variant in the FTO gene which has been used to study body mass index (BMI)
[Brennan et al., 2009]. The effect of variation in the FTO gene on BMI is not
direct; rather the genetic variant affects satiety, which in turn affects BMI
[Wardle et al., 2008]. An intervention on BMI which is not based on reducing
food intake may have a different effect on the outcome to the estimate from
a Mendelian randomization study using a variant in the FTO gene as an IV.
Equivalently, the effect of the intervention may not be limited to the ex-
posure of interest. For example, bariatric surgery aimed at reducing BMI may
also result in dietary and lifestyle changes. It is difficult to assess which changes
in covariates are direct results of a decrease in BMI and so are on the causal
pathway from BMI to disease, and which are separate consequences of the
intervention. Even when both the genetic change and proposed intervention
specifically target the exposure, it may be that they are on different biological,
biochemical or physiological pathways, and so the genetic and clinical changes
in exposure may affect the outcome in different ways and to different extents.

6.1.5 Differences in populations

Fifthly, the genetic variant potentially affects all members of a population.
If the proposed intervention is to be made across the whole population, then
Mendelian randomization using a population-based cohort may give a valid
estimate of its potential effect. However, if the intervention is intended to be
made in a particular subpopulation, it may not be possible to choose a cohort
for a Mendelian randomization study which would give a relevant estimate.
For example, an intervention on blood pressure may only be applied to those
with clinically-determined hypertension, whereas a genetic variant associated
with blood pressure would potentially affect the whole population.
90 Mendelian Randomization

6.2 Comparison of estimates

We give some examples to illustrate the differences between Mendelian ran-
domization estimates and those from other epidemiological approaches, such
as effect estimates from randomized controlled trials (RCTs), and observa-
tional associations from multivariable adjusted regression models.

6.2.1 Cholesterol and coronary heart disease

Coronary heart disease (CHD) is the result of a build-up of atheromatous
plaques in the coronary arteries. A major component of such plaques is choles-
terol, and low-density lipoprotein cholesterol (LDL-C) is an established causal
risk factor for CHD. We here use the available literature to assess the magni-
tude of the effect of LDL-C on CHD risk as estimated from Mendelian ran-
domization, and from RCTs where statin drugs have been used as a clinical
intervention to lower LDL-C.
A recent meta-analysis of genome-wide association studies reported five
SNPs associated with LDL-C, but not with high-density lipoprotein choles-
terol (HDL-C) nor triglycerides [Waterworth et al., 2010]. Table 6.1 gives the
SNPs and relevant genes, the estimates of association of each SNP with log-
transformed LDL-C and risk of CHD, and estimates using each SNP of the
causal odds ratio of CHD per 30% decrease in LDL-C using the ratio method
(Section 4.1). We note that this relies on a log-linear assumption of the effect
of log(LDL-C) on CHD risk, and between eight- and twenty-fold extrapola-
tion of the genetic effects on log(LDL-C). Although further SNPs associated
with LDL-C are known, these five were chosen as they represent variants with
known strong associations with LDL-C, where there is some biological knowl-
edge to justify the assumption of the specific effect of the SNP on LDL-C.
The dose–response relationship between the genetic associations with LDL-C
and with CHD risk is noted; this gives plausibility that LDL-C is a causal risk
factor for CHD risk (Table 3.1).
Odds ratio estimates for each SNP individually range from 0.27 to 0.45.
If we assume that the five estimates of causal effect in Table 6.1 are inde-
pendent, then a fixed-effect inverse-variance weighted meta-analysis method
(see Section 9.4.1) gives a combined odds ratio of 0.33 (95% CI 0.24 to 0.46)
[Thompson et al., 2005]. The estimates will not be strictly independent, as
they are derived from the same data and affect related pathways, but as the
SNPs are on different chromosomes and therefore independently distributed,
the correlation between estimates can reasonably be assumed to be small
[Burgess et al., 2013].
In comparison, RCTs of statins have given lesser estimates of the benefits
of reducing LDL-C levels. A meta-analysis of 9 trials of the effect of statin
use on CHD comprising 69 139 participants with 6406 CHD events gave a
Generalizability of estimates from Mendelian randomization 91

Per allele Per allele odds Odds ratio of CHD

SNP change in ratio of CHD per 30% decrease
(relevant gene) log(LDL-C) (SE) (95% CI) in LDL-C (95% CI) 1
rs11206510 (PCSK9 ) 0.026 (0.004) 1.07 (1.01–1.13) 0.40 (0.15–0.85)
rs660240 (SORT1 ) −0.044 (0.004) 0.85 (0.80–0.90) 0.27 (0.15–0.44)
rs515135 (APOB ) −0.038 (0.004) 0.90 (0.85–0.96) 0.37 (0.19–0.66)
rs12916 (HMGCR) −0.023 (0.003) 0.94 (0.90–0.99) 0.38 (0.16–0.80)
rs2738459 (LDLR) −0.018 (0.004) 0.96 (0.89–1.03) 0.45 (0.07–1.95)

TABLE 6.1
Association of five SNPs with log-transformed low-density lipoprotein choles-
terol (LDL-C) and coronary heart disease (CHD) risk. Causal estimates of
odds ratio for 30% reduction in LDL-C on coronary heart disease from
Mendelian randomization using each SNP in turn.

1A 30% decrease in LDL-C is equivalent to a change in log(LDL-C) of -0.357.

relative risk of 0.73 (95% CI 0.70 to 0.77) based on a reduction of around

30% in LDL-C over an average follow-up time of at least 3 years [Cheung
et al., 2004]. A more focused meta-analysis examining the effect of statin use
for primary disease prevention, comprising around 27 969 individuals without
a history of coronary heart disease with 1677 events, gave a similar relative
risk of 0.72 (95% CI 0.65 to 0.79) over 1.5 to three years’ follow-up [Taylor
et al., 2013]. The data on the genetic variants, together with the combined
Mendelian randomization estimate and the proportional estimate of the effect
of statins (assuming the relative risk of 0.73 approximates an odds ratio of
the same magnitude) are displayed in Figure 6.1.
The Mendelian randomization estimate of the effect of LDL-C reduction
is greater (further from the null) than the estimate from RCTs of LDL-C
reduction using statins. It is known that the effect of statins in reducing CHD
increases over time [Law et al., 2003]. As atherosclerosis is a chronic condition
which develops progressively, it is not surprising that the estimates of the
effect of the life-long lowering of LDL-C associated with the SNPs considered
corresponds to a greater proportional change in cardiovascular risk than the
effect on LDL-C due to statin usage. Further possible reasons for differences
between the estimates include the non-specific action of statins, which also
reduce inflammatory response [Davignon and Laaksonen, 1999]. However, any
effects of statins on inflammatory response may further lessen the causal role
of low-density lipoprotein cholesterol, and make the contrast with the genetic
effects more extreme.
92 Mendelian Randomization

1.00
Per allele odds ratio of CHD
0.95
0.90
0.85

Estimate from Mendelian

randomization
Estimate of effect of
statin use
0.80

−4% −2% 0
Per allele percentage change in LDL−C

FIGURE 6.1
Estimates of percentage change in low-density lipoprotein cholesterol (LDL-C)
and odds ratio of coronary heart disease (CHD) per LDL-C decreasing allele
for five SNPs (point estimates with 95% confidence intervals), plus estimate of
causal effect of LDL-C on CHD risk from Mendelian randomization using all
5 SNPs (solid line) with 95% confidence interval (dotted lines). Proportionate
effect from meta-analysis of statin use on CHD risk in RCTs (dashed line with
dotted lines for 95% confidence interval) is displayed for comparison.
Generalizability of estimates from Mendelian randomization 93

6.2.2 Blood pressure and coronary heart disease

A similar example can be observed in the association between blood pressure
and CHD. An allele score associated with a 1.6mmHg decrease in systolic
blood pressure is associated with a odds ratio for CHD of 0.91 (95% CI 0.89 to
0.92) [Ehret et al., 2011]. Assuming a log-linear association, this corresponds
to an odds ratio of 0.55 (95% CI 0.47 to 0.61) for a 10mmHg decrease in
systolic blood pressure, compared to the relative risks from a meta-analysis
of 0.78 (95% CI 0.73 to 0.83) in clinical trials and 0.75 (95% CI 0.73 to 0.77)
in cohort studies [Law et al., 2009].
The Mendelian randomization estimate of the effect of blood pressure low-
ering is greater (further from the null) than the estimates from both RCTs
and observational studies. However, unlike in the lipids example above, the
mechanisms of the effects of all the 29 SNPs included in the allele score are
not well-known. Hence there is a possibility of pleiotropic effects (or equiva-
lent, such as linkage with other variants) leading to violation of the Mendelian
randomization assumptions and lack of internal validity of the causal estimate
(Section 3.2.3). However, the pleiotropic associations with alternative risk fac-
tors would have to be reasonably strong to explain a sizeable proportion of
the reduction in CHD risk [Martens et al., 2006]. As none of the variants in
the allele score are known to be strongly associated with other known CHD
risk factors, it would seem that the reduction in CHD risk is most plausibly
due to the effect of the blood pressure reduction and not due to other factors.

6.3 Discussion
External validity in epidemiology is often thought of in terms of generaliz-
ability to a population other than the one considered in the original study
[Dekkers et al., 2010]. Although variation in populations may cause some dif-
ficulties, the differences between the change in exposure levels associated with
natural genetic variation and with any proposed clinical intervention on the
exposure lead to inescapable problems in generalizing Mendelian randomiza-
tion estimates to clinical questions of interest.
Mendelian randomization is a useful tool for exploring causal relationships
between modifiable exposures and outcomes of interest. It is one of the few
methodologies that can aid the selection of targets for therapeutic interven-
tion. However, it would be misleading to assume that the estimate from a
Mendelian randomization study gave the definitive answer to every question
of causal relevance of an exposure. Mendelian randomization estimates are es-
pecially relevant when the effect of interest is that of a long-term population-
based intervention; otherwise, although a Mendelian randomization approach
94 Mendelian Randomization

may be qualitatively informative, the quantity estimated may not correspond

to the clinical effect of interest.

6.3.1 Using Mendelian randomization in drug assessment

Questions of generalizability of results are important when using Mendelian
randomization to prioritize or de-prioritize targets for drug development, es-
pecially in the context of the primary prevention of disease. A considerable
proportion of large-scale and expensive clinical trials of drugs targeting sus-
pected novel mechanisms of action fail to demonstrate efficacy. A prudent
approach to drug development would be to only go forward with research on
targets where there is evidence on the causal nature of the exposure and/or
mechanism from human genetics [Plenge et al., 2013]. Where suitable genetic
variants for the application of Mendelian randomization on a given exposure
are known and available in a large enough sample, assessment of the associ-
ation between the variants and the disease outcome, and consequently of the
causal effect of the exposure on the outcome, is simple, quick, and relatively
inexpensive to perform.
Association between a relevant genetic variant affecting the exposure and
the outcome may be taken as evidence for the potential efficacy of a drug
affecting the exposure pathway. However, absence of evidence for such an
association does not necessarily imply lack of efficacy. A drug which blocks
a particular biological pathway may have a profound effect on downstream
markers, which may lead to a substantially different effect on outcome com-
pared to the slight changes associated with a genetic variant. Although we may
expect Mendelian randomization in many circumstances to provide a good
qualitative indication of the efficacy of clinical intervention, the magnitude of
the Mendelian randomization estimate will not necessarily be a reliable guide
to the potential benefit of a drug. Additionally, in many cases the drug will be
aimed at secondary disease prevention or targeted at a particular population
group (such as individuals who have high or low levels of the exposure) rather
than at the general population. As the randomization of genetic variants is
valid only in the population, testing the genetic association with the disease
in a sample population chosen according to their disease status or exposure
value may lead to misleading inference (ascertainment bias, Section 3.2.5).
The Mendelian randomization paradigm can also aid in target re-
assessment (drug repositioning). If an existing drug has a genetic variant that
mimics its effect, then an association of the variant with another outcome
may indicate that the drug is also an effective treatment for that outcome.
For example, the effects of anakinra, an interleukin-1 receptor antagonist and
licensed treatment for rheumatoid arthritis, can be assessed for a range of
further auto-immune diseases by considering the association of variants in
the IL1RN with those disease outcomes. Equally, genetic associations can in-
form the assessment of mechanism-associated safety. For example, a variant in
the GCKR gene is associated with lower plasma glucose levels, but also with
Generalizability of estimates from Mendelian randomization 95

higher triglyceride levels [Beer et al., 2009]. This observation may suggest that
additional monitoring of triglyceride levels would be advisable in clinical trials
of glucokinase activators.

6.3.2 Using Mendelian randomization in drug discovery

‘Reverse Mendelian randomization’ can be used when a SNP is found in
genome-wide association data to be associated with a disease outcome, but
the mechanism for the association is not known. An exposure is sought which
is associated with the variant and could explain the gene–disease association.
This concept is closely related to that of functional genomics. For example,
associations between variants in the HMGCR gene and risk of coronary heart
disease are indicative of the causal role of low-density lipoprotein cholesterol
in cardiovascular pathogenesis, and also point towards the potential efficacy
of drugs to inhibit HMG-CoA reductase (statins). This approach should pro-
vide a fruitful source of targets for ongoing pharmacological research, and
has already been used successfully in the discovery of the PCSK9 enzyme
for cholesterol lowering (a variant in the PCSK9 gene having been previously
shown to be associated with CHD risk [Cohen et al., 2006]). PCSK9 inhibitors
have already been demonstrated to be effective in lowering LDL-C [Robinson
et al., 2014] and lipoprotein(a) [Raal et al., 2014], and phase III trials for the
secondary prevention of cardiovascular endpoints were underway at the time
of writing [Farnier, 2013].

6.3.3 Relevance of causal estimation in Mendelian

randomization
As has been emphasized throughout this book, Mendelian randomization in-
vestigations can assess a causal relationship, or estimate a causal effect. The
arguments in this chapter suggest that the magnitude of causal effect estimates
using Mendelian randomization should not be taken too literally. While they
provide some indication of the potential relevance of an exposure, the direc-
tion of the causal effect and whether it is compatible or not with the null are
more important.
One reason for this is that the true causal risk factor may be difficult to de-
fine, and so the measured exposure may only be a surrogate (proxy) measure
of the underlying risk factor. For example, in a Mendelian randomization
analysis of the causal effect of BMI, genetic variants associated with BMI
are likely to be associated with the outcome by causal pathways via other
adiposity-related variables. In this case, formally the IV assumptions are vi-
olated [Glymour et al., 2012]. However, if the investigation is interpreted not
narrowly as estimating the causal effect of BMI on the outcome, but more
broadly as estimating the causal effect of adiposity (for which BMI is used
as a proxy measure) on the outcome, then the estimate may still a valid test
96 Mendelian Randomization

of the causal null hypothesis if there is no causal pathway from the genetic
variant(s) to the outcome not via adiposity, in spite of the IV assumptions for
BMI being violated.
For these reasons, some authors have questioned whether causal effect esti-
mates should ever be considered as part of a Mendelian randomization analysis
[VanderWeele et al., 2014]. Although there is a danger of estimates being over-
interpreted, there are several reasons why causal estimates are useful. First,
in epidemiology generally, estimates with confidence intervals are preferred
to hypothesis tests with p-values, as they are more informative [Sterne and
Davey Smith, 2001]. For instance, if a p-value does not achieve conventional
levels of statistical significance, a point estimate with a confidence interval al-
lows the reader to judge in a quantitative way whether the null result reflects
a lack of evidence or a genuine negative finding in comparison with either
the observational association, or with a minimal clinically relevant effect. Sec-
ondly, if several genetic variants are valid instrumental variables for the same
exposure, greater power to detect a clinically relevant causal effect can be ob-
tained using information on all of the variants simultaneously rather than that
using the variants individually. Causal estimates from multiple variants also
enable the quantitative comparison of the consistency of genetic associations,
using a heterogeneity or overidentification test, as a statistical assessment of
pleiotropy (Section 4.5.3). Finally, although the causal estimate in a Mendelian
randomization analysis may not be equal to the effect of an intervention in
the exposure, it does have a well-defined interpretation as the effect of an
intervention in the genetic code at conception. Hence, although assessment
of causation should be the primary outcome of a Mendelian randomization
investigation, the estimate of a causal effect also has considerable utility.

6.4 Summary
In Mendelian randomization, differences in the exposure distribution due to
genetic variation are materially distinct from the change due to any proposed
therapeutic intervention on the exposure, and so may affect the outcome dif-
ferently. Consequently, it may be misleading to generalize the magnitude of a
Mendelian randomization association to the effect of a potential intervention
on the exposure in practice. Awareness of this is important for the use of
Mendelian randomization in target-based drug development.
In this chapter, we have considered qualitative and quantitative issues
relating to the interpretation of causal effects using Mendelian randomization
and their relationship to the effects of interventions. In the next part of this
book, we consider statistical aspects of Mendelian randomization analyses
relating to the topics discussed in this and previous chapters.
 Part II

Statistical issues in
instrumental variable
analysis and Mendelian
randomization
7
Weak instruments and finite-sample bias

In this chapter, we consider the effect of weak instruments on instrumen-

tal variable (IV) analyses. Weak instruments, which were introduced in Sec-
tion 4.5.2, are those that do not explain a large proportion of the variation in
the exposure, and so the statistical association between the IV and the expo-
sure is not strong. This is of particular relevance in Mendelian randomization
studies since the associations of genetic variants with exposures of interest are
often weak. This chapter focuses on the impact of weak instruments on the
bias and coverage of IV estimates.

7.1 Introduction
Although IV techniques can be used to give asymptotically unbiased estimates
of causal effects in the presence of confounding, these estimates suffer from bias
when evaluated in finite samples [Nelson and Startz, 1990]. A weak instrument
(or a weak IV) is still a valid IV, in that it satisfies the IV assumptions, and es-
timates using the IV with an infinite sample size will be unbiased; but for any
finite sample size, the average value of the IV estimator will be biased. This
bias, known as weak instrument bias, is towards the observational confounded
estimate. Its magnitude depends on the strength of association between the
IV and the exposure, which is measured by the F statistic in the regression
of the exposure on the IV [Bound et al., 1995]. In this chapter, we assume
the context of ‘one-sample’ Mendelian randomization, in which evidence on
the genetic variant, exposure, and outcome are taken on the same set of indi-
viduals, rather than subsample (Section 8.5.2) or two-sample (Section 9.8.2)
Mendelian randomization, in which genetic associations with the exposure
and outcome are estimated in different sets of individuals (overlapping sets in
subsample, non-overlapping sets in two-sample Mendelian randomization).
We illustrate this chapter using data from the CRP CHD Genetics Col-
laboration (CCGC) to estimate the causal effect of blood concentrations of C-
reactive protein (CRP) on plasma fibrinogen concentrations (Section 1.3). As
the distribution of CRP is positively skewed, we take its logarithm and assume
a linear relationship between log(CRP) and fibrinogen. Although log(CRP)

99
100 Mendelian Randomization

and fibrinogen are highly positively correlated (r = 0.45 to 0.55 in the ex-
amples below), it is thought that long-term elevated levels of CRP are not
causally associated with an increase in fibrinogen.
We first demonstrate the direction and magnitude of weak instrument bias
for IV estimates from real and simulated data (Section 7.2). We explain why
this bias comes about, why it acts in the direction of the confounded observa-
tional association, and why it is related to instrument strength (Section 7.3).
We discuss simulated results that quantify the size of this bias for different
strengths of instruments and different analysis methods (Section 7.4). When
multiple IVs are available, we show how the choice of IV affects the variance
and bias of IV estimators (Section 7.5). We propose ways of designing and
analysing Mendelian randomization studies to minimize bias (Section 7.6).
We conclude with a discussion of this bias from both theoretical and practi-
cal viewpoints, ending with a summary of recommendations aimed at applied
researchers on how to design and analyse a Mendelian randomization study
to minimize bias from weak instruments (Section 7.7).

7.2 Demonstrating the bias of IV estimates

First, we demonstrate the existence and nature of weak instrument bias in IV
estimation using both real and simulated data.

7.2.1 Bias of IV estimates in small studies

As a motivating example, we consider the Copenhagen General Population
Study [Zacho et al., 2008], a cohort study from the CCGC with complete
cross-sectional baseline data for 35 679 participants on CRP, fibrinogen, and
three SNPs from the CRP gene region: rs1205, rs1130864, and rs3093077. We
calculate the observational estimate by regressing fibrinogen on log(CRP),
and the IV estimate by the two-stage least squares (2SLS) method using all
three SNPs as IVs in a per allele additive model (Section 4.2.1). We then
analyse the same data as if it came from multiple studies by dividing the data
randomly into substudies of equal size, calculating estimates of association
in each substudy, and combining the results using inverse-variance weighted
fixed-effect meta-analysis. We divide the whole study into, in turn, 5, 10, 16,
40, 100, and 250 substudies. We recall that the F statistic from the regres-
sion of the exposure on the IV is used as a measure of instrument strength
(Section 4.5.2).
We see from Table 7.1 that the observational estimate stays almost un-
changed whether the data are analysed as one study or as several studies.
However, as the number of substudies increases, the pooled IV estimate in-
creases from near zero until it approaches the observational estimate. At the
Weak instruments and finite-sample bias 101

Substudies Observational estimate 2SLS IV estimate Mean F statistic

1 1.68 (0.01) −0.05 (0.15) 152.0
5 1.68 (0.01) −0.01 (0.15) 31.4
10 1.68 (0.01) 0.09 (0.14) 16.4
16 1.68 (0.01) 0.23 (0.14) 10.8
40 1.68 (0.01) 0.46 (0.13) 4.8
100 1.67 (0.01) 0.83 (0.11) 2.5
250 1.67 (0.01) 1.27 (0.08) 1.6

TABLE 7.1
Estimates of effect (standard error) of log(CRP) on fibrinogen (µmol/l) from
the Copenhagen General Population Study (N = 35 679) divided randomly
into substudies of equal size and combined using fixed-effect meta-analysis:
observational estimates using unadjusted linear regression, IV estimates using
2SLS. Mean F statistics averaged across substudies from linear regression of
log(CRP) on three genetic variants.

same time, the standard error of the pooled IV estimates decreases. We can
see that even where the number of substudies is 16 and the average F statistic
is around 10, there is a serious bias. The causal estimate with 16 substudies
is positive (p = 0.09) despite the causal estimate with the data analysed as
one study being near to zero.

7.2.2 Distribution of the ratio IV estimate

In order to investigate the distribution of IV estimates with weak instruments,
we use a simulation exercise, taking a simple example of a confounded associa-
tion with a single dichotomous IV [Burgess and Thompson, 2011]. Parameters
are chosen such that the causal effect is null, but simply regressing the outcome
on the exposure yields a strong positive confounded observational association
of close to 0.5. We took 6 different values of the strength of the IV–exposure
association, corresponding to mean F statistic values between 1.1 and 8.7.
Causal estimates are calculated using the ratio method, although with a
single IV the estimates from the ratio, 2SLS and limited information maxi-
mum likelihood (LIML) methods are the same (Section 4.3.2). The resulting
distributions for the estimate of the causal parameter are shown in Figure 7.1.
For weaker IVs, there is a marked bias in the median of the distribution in
the positive direction and the distribution of the IV estimate has long tails.
For the weakest IV considered, the mean F statistic is barely above its null
expectation of 1 and the median IV estimate is close to the confounded ob-
servational estimate of 0.5. For stronger IVs, the median of the distribution
of IV estimates is close to zero. The distribution is skew with more extreme
causal estimates tending to take negative values.
102 Mendelian Randomization

The analyses in Table 7.1 and simulations in Figure 7.1 show that IV
estimates can be biased. This bias has two notable features: it is larger when
the F statistic for the IV–exposure relationship is smaller, and it is in the
direction of the confounded observational estimate.

7.3 Explaining the bias of IV estimates

We now try to provide some more intuitive understanding of why weak in-
strument bias occurs. We give three separate explanations for its existence, in
terms of the definition of the ratio estimator, finite-sample violation of the IV
assumptions, and sampling variation of IV estimators.

7.3.1 Correlation of IV associations

First, there is a correlation between the numerator (estimate of the G–Y
association) and denominator (estimate of the G–X association) in the ratio
estimator. To understand this, we consider a simple model of confounded
association with causal effect β1 of X on Y , with a dichotomous IV G =
0 or 1, and further correlation between X and Y due to association with a
confounder U :

X = α1 G + α2 U + εX (7.1)
Y = β1 X + β2 U + ε Y
2 2
U ∼ N (0, σU ); εX ∼ N (0, σX ); εY ∼ N (0, σY2 ) independently.
2
We initially assume that σX = σY2 = 0 for ease of explanation.
If ūj is the average confounder level for the subgroup with G = j (where
j = 0, 1), an expression for the causal effect from the ratio method is:
∆Y β1 ∆X + β2 ∆U β2 ∆U
β1R = = = β1 + (7.2)
∆X ∆X α1 + α2 ∆U
where ∆U = ū1 − ū0 is normally distributed with expectation zero; ∆X and
∆Y are defined similarly. When the instrument is strong, α1 is large compared
to α2 ∆U . Then the expression β1R will be close to β1 . When the instrument is
weak, α1 may be small compared to β2 ∆U and α2 ∆U . Then the bias β1R − β1
is close to αβ22 , which is approximately the bias of the confounded observational
association (it is exactly this if α1 is zero). This is true whether ∆U is positive
or negative. Figure 7.2 (top panel) shows how the IV estimate bias varies with
∆U . Although for any non-zero α1 the IV estimator will be an asymptotically
consistent estimator as sample size increases and ∆U tends towards zero, a
bias in the direction of the confounded association will be present in finite
samples. From Figure 7.2 (top panel), the median bias will be positive, as the
Weak instruments and finite-sample bias 103

Expected F statistic = 1.1 Expected F statistic = 1.6

Median estimate = 0.47 Median estimate = 0.29

0.4
Density

Density
0.2

0.2
0.0

0.0
−2 −1 0 1 2 −2 −1 0 1 2

Expected F statistic = 2.6 Expected F statistic = 4.1

Median estimate = 0.13 Median estimate = 0.05
0.6
Density

Density
0.4
0.3
0.0

0.0

−2 −1 0 1 2 −2 −1 0 1 2

Expected F statistic = 6.1 Expected F statistic = 8.7

Median estimate = 0.01 Median estimate = 0.00
0.8
Density

Density
0.6
0.4
0.0

0.0

−2 −1 0 1 2 −2 −1 0 1 2

FIGURE 7.1
Histograms of IV estimates of a null causal effect using weak instruments
from simulated data for six strengths of the IV–exposure association. Average
F statistics and median IV estimates for each scenario are shown.
104 Mendelian Randomization
α1
estimate is greater than β1 when ∆U > 0 or ∆U < − α 2
, which happens with
probability greater than 0.5.
This also explains the heavier negative tail in the histograms in Figure 7.1.
The estimator takes extreme values when the denominator α1 +α2 ∆U is close
to zero. Taking parameters α1 , α2 and β2 as positive, as in the example of Sec-
tion 7.2.2, this is associated with a negative value of ∆U , whence the numer-
ator β2 ∆U will be negative. As ∆U has expectation zero, the denominator
is more likely to be small and positive than small and negative, giving more
negative extreme values of β1R than positive ones.
2
If there is independent error in X and Y (that is, σX and σY2 in equation
(7.1) are non-zero), then the picture is similar, but more noisy, as seen in
Figure 7.2 (bottom panel). The expression for the IV estimator is:

β2 ∆U + ∆εY
β1R = β1 +
α1 + α2 ∆U + ∆εX
where ∆εX = ε̄X1 − ε̄X0 and ∆εY = ε̄Y 1 − ε̄Y 0 are defined analogously to ∆U
above.

7.3.2 Finite-sample violation of IV assumptions

An alternative explanation of weak instrument bias is in terms of violation
of the second IV assumption in a finite sample. Although a valid instrument
will be asymptotically independent of all confounders, in a finite sample there
will be a non-zero correlation between the instrument and confounders. This
correlation biases the IV estimator towards the observational confounded as-
sociation.
If the instrument is strong, then the difference in mean exposure between
genetic subgroups will be mainly due to the genetic instrument, and the differ-
ence in outcome (if any) will be due to this difference in exposure. However if
the instrument is weak, that is it explains little variation in the exposure, the
chance difference in confounders may explain more of the difference in mean
exposure between genetic subgroups than the instrument. If the effect of the
instrument is near zero, then the estimate of the “causal effect” approaches
the association between exposure and outcome resulting from changes in the
confounders, which is the observational confounded association [Bound et al.,
1995].

7.3.3 Sampling variation within genetic subgroups

Finally, we offer a graphical explanation of weak instrument bias. To do
this, we simulate data with a negative causal effect of the exposure on the
outcome, but with positive confounding giving a strong positive observa-
tional association between the exposure and outcome. We generate 1000 sim-
ulated datasets with 600 subjects divided equally into three genetic subgroups
Weak instruments and finite-sample bias 105

8
6
1 − β1
4
2
Bias: βR
0
−2
−4
−6

−1.0 −0.5 0.0 0.5 1.0

Difference in confounder: ∆U
8
6
1 − β1
4
2
Bias: βR
0
−2
−4
−6

−1.0 −0.5 0.0 0.5 1.0

Difference in confounder: ∆U
FIGURE 7.2
Bias in IV estimator as a function of the difference in mean confounder be-
tween groups (α1 = 0.25, α2 = β2 = 1). Horizontal dotted line is at the
confounded association αβ22 , and the vertical dotted line at ∆U = − α
α2 where
1

β1R is not defined. Top panel: no independent error in X or Y ; bottom panel:

∆εX , ∆εY ∼ N (0, 0.12 ) independently.
106 Mendelian Randomization

(G = 0, 1, or 2):

xi = α1 gi + ui + εXi (7.3)
yi = β1 xi + ui + εY i
2 2
ui ∼ N (0, σU ); εXi ∼ N (0, σX ); εY i ∼ N (0, σY2 ) independently.
2
We set β1 = −0.4, σU = 12 , σX2
= 0.22 , and σY2 = 0.22 , and take four val-
ues for the strength of the IV (α1 = 0.5, 0.2, 0.1, and 0.05) corresponding
to expected F statistics of 100, 16, 4.7, and 2.0. The mean levels of exposure
and outcome for each genetic subgroup from each simulated dataset are plot-
ted (Figure 7.3), representing joint density functions for each subgroup. To
examine the sampling distribution of the IV estimate, we draw one point at
random from each of these distributions; the gradient of the line through these
three points is the 2SLS IV estimate. When the instrument is strong, the large
differences in exposure between the subgroups due to variation in the IV will
generally lead to estimating a negative effect of exposure on outcome. When
the instrument is weak, the differences in exposure between the subgroups due
to the IV are small and the positively confounded observational association is
more likely to be recovered.

7.4 Properties of IV estimates with weak instruments

In the previous section, we showed that IV estimates are biased in finite sam-
ples. In this section, we consider the magnitude of the bias in IV estimates,
as well as the coverage of IV methods with weak instruments.

7.4.1 Bias of IV estimates

The bias of an estimator is the difference between the expectation of the
estimator and the true value of the parameter. In IV analysis, the relative
mean bias is the ratio of the bias of the IV estimator (β̂IV ) to the bias of the
observational association (β̂OBS ) found by linear regression of the outcome on
the exposure:
E(β̂IV ) − β1
Relative mean bias = . (7.4)
E(β̂OBS ) − β1
The relative mean bias from the 2SLS method is asymptotically approxi-
mately equal to 1/E(F ), where E(F ) is the expected F statistic in the regres-
sion of the exposure on the IV [Staiger and Stock, 1997]. This approximation
is only valid when the number of IVs is at least three. The rule-of-thumb of
F < 10 indicating weak instruments (Section 4.5.2) derives from this expres-
sion. This rule approximately limits the bias in the IV estimate to less than
Weak instruments and finite-sample bias 107

Strong instrument: E(F)=100 Moderate instrument: E(F)=16

x
xxx
x xx
x xxxxxxxxxxx xxxx x
x x xx xxx xx x x
x x xxxx xx
x x
x xx x x xx xx xxxxxxxxxxx x
x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xx x
xxx xxxxxxxxxxx oo
Outcome

Outcome
xxxxx xx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xx x o
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxx x x
x o o oo ooo o
x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx x o oo o o
xxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx x
x x ooooooooo o o
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx x xx xxx x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxx o ooooo o o o o
xxxxxxxxxxxxxxxxxxxxxxxxxxx x x x x x x x x x xxx x x o o
xxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxx xxx xxxxxxxxxxxxxxx xxxxxxxxxxx xxxxx x o ooo
ooooooooo ooooo
ooooo
oo
xxxxxxxxxxxxxxxxxxxxxxxxxxx x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx o ooooooo ooooo ooooo
o
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx x x xxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx x o o oo
o
oooo
o o o
oo o
o
o o o
oo
o o o o
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx x xx xxxxxxxxxxxxxxxxxxxxxxxx x xx x x
x x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxx
oooooo o oo
oooooooooo oo
o
ooo
ooooooooooooo
oo
oo ooooooooo
ooo ooo
x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xx o x x xxxxxxxxxx xx oooooo
ooo oo
oooooooooo ooooooooooo oooooooooo oo
ooo
x xx xxxxxxxxxxxxxxxx o xx xxxxxxxxxxxxxxxxxxxxxxx xx ooooo ooooo o
o o
oooo ooo oooo
oo
oo
ooo
oo
o oooo oo o
xx xxxxxxxxxxxx x ooo o ooooo xx x xxxxxxxx xxxx
x oo
oooooo
oo ooo
oo oo
ooo ooooo
o o
ooooo oo
x o oooo oooooooooooo x xxxxxxxxxxxxxxxxx xx
oooooo oo
ooo o
oo
oooo oooo
oo
o
o oo oo
oooo oo ooooooo
xx ooooo ooooooooo oooooooo ooooooo oooo oooooo oooooooooo
ooooooo ooo
ooo
oooooooo oooo
o
oo xxx xxxxx x o o o ooo o o oo
ooo
ooo
o ooo
oo o
o
o o o
x x
ooo oooo oo
o o
o ooo oo
o
ooooo
oo
oo
ooo
o
oooo
oooooooooo o xx x o ooooo ooooooo
oo
o oooo
oooo
oo
oooo
oo
oooooo
o
ooo
oo oo
oooo
oooooo
xx x o
ooo o
o
oooo o
ooo
o
oo
oo oo o
oo
o
ooooooooo o o o x x oooooo
ooooo
ooo oo ooooooo
o
oooo
o
oo o
oo oo
oooooo o
oo oo oooo
oo oooo oo
oooo oo
oo
oo oooo oooooo ooo x ooooo o
oooo
o ooooo ooooooo o
o oo
o
oooo o
o oo oooooo
o oooo
oo
oo ooo
oooooo
o
oo o
oo
oooooooooooo oooo o ooo oo
o o
oo oooooo
ooooooooooo
o oo
oooo
o
oooo
ooooo o
ooo
o o
oo o
ooooo
oo oo
o o
oooo
oooo oo
oooooo oo ooo
o
ooooooooooo oo
oo oooooooo
oo oooooo ooooo ooooo
o oooo o
oooooooooooooo
oo o oo
ooo ooo oo o oo o
oooo o
oo
oooo
ooo
oo oo
oooooo
o oo oooo
ooo oooooooo o ooooo ooooo ooo o o
o o o oo
oo o oo o oo o
ooooooooooooooo
o oo o o oooo
oo o o o
ooooooo o o
ooo o oooo
o

Exposure Exposure

Weak instrument: E(F)=4.7 Very weak instrument: E(F)=2.0

x
x x
x x o xx
x x o x x o
x x x x
xxx xxxx xx x xx
x xx x x x o x x xx xxxxxx xxo o o o
x x x x x x x x x x xx ox ox
xxxx xxxxx xxx x o o x
xxxx xx x xxx xxxxx o o ooo ooo
x x xxxx xxxxxxxxxxxxxxxxxxxxx o oo x xxxx xxxxxxo xx xxx x oxoo x oo x
ox o o o
x x x x x x x
x x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xx x ooooo oooooooo ooo oo
x x x x o x x xxxxxxxxxxxxo xxxxooxoxxoxxooooo ooooooo oo o
xoo
Outcome

Outcome

xx xxxxoxxxxxo
xx xxxx xx xxxxxxxxxxxxx xxx oo oo oo oo oo xxxxxxxxxxxxxxxxxxxxxxo x xxxxooo xxoox oo xoooo
o
xooooooooooo oo o
o
x xxxxxxxxxxxxxxxxx xxxxxxxxxxx x xx oxooo o oooooooo o oo ooooo o x x x x xxxxxxo
x
x xx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx oooooo oooo xx x x x
xxxxxxxxxxxxxxxxxxxxoxo
x x xxxxxxxxo xxxxoxxxoxxoxxo
x xxx x x o x ooo oxo
oooo o o
oooo ooooooo ooooooo
o oo
o oo x xox o xox xooxoooo o o o o
xx xxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxx x xx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xx oooo oooooooo
ooo
ooo ooo
oooo o ooo
oo ooo
o oooo
x
x x x xxxxxxxxxxxxxxxxxxoxxxxxxxxo x x x x
xxxxxoxxo x x
o o o
xooooo o
xooooooo oo ooo
oo ooooooo o o
x ooooooooooooooo oo oooooo ooo ooo x xxxxxxxxxxxxxxxxxxxxxxxxo xxoxoxxxoxxo xxxxxoooo xooxoooooxoooooo o ooooooo
x x xx
x xxx xxxxxxxxxxxxxxxxxxxxxxxxxxx xxxx ooooooooo x ooo
ooo oooooooooooo o o xxxx xx xxxxxxxxxxx xxxxoo xooxooooooo
xo ooooo
oo oooooo
oooooo ooo o
x x x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xx oxooooo o oooooo oooo o o xxx xxxxxoxxxoxo xooxxo xooxoooo oooooooooooo
o ooo oooo o o
oooooo
oooo
ooo oo o
ooooo o
ooooo o o x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxoxxxxxxxxxxxxxxooxo ooxoo xoxoo xooo ooo ooo
oooo oo o
xx xx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx x
x x x x xxx ooooooooooooo o
ooo ooooooo
oo ooooooooooo
o o o ooooo o x xxxxxxxxxxxxxxxxxxxoxxxxxxxxxxoxxoo xooxooxxo oxooxo ooxo oo
oooo oo
oxoooooo
o ooo
o
ooooooo
ooooooo oo o
x xxxxxxxxxxxxxxxxxxxxxxxxxx x x o oo o oooooo oooooo oo o oooooo
ooo
o oo
oooo oooooooo oo
ooo o x xxxxxxx xxoxxxoxxxxxxxoxxoxoxxxxoxoooxooxoxxoxxo oxxooooooooooo
o oooooxoooooooooooooooooooo
xx xxx xxxxxxxxxx xxxxxxxxxxxxxxxxxxooxooooooooooo oooooo
o oo oooooo
ooo oooooooo
o x x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxo
xxxxxoxoxoxxooxoxo x x x
o o ooo o o o oo
o o
x xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xx oooooooooo ooooooo oooooooo oooo oooooooo oooo o oo
x x o
oxxoxooooxo o
x
ooxo
o
xoxoooooo o
oo
ooooo
o o
oooooooooooo oo oo o o o
x xx xxx xxxxxxxxxxxxxxx x xoo
x o
oooo oooooo
ooo ooooo
ooo oooooo ooo
oo ooooooo
o ooooo oooo o x xxxxxxxxxxxxxxxxxxxxxxxxoxxxxxxxoxxoxxoxoxo xoxo xxxoooxo
ooxooooo oooooooo ooooo
o o
ooo o
x x xxxxxxxxxxxxxxxxxxxxxxxxxxxx x ox xo
ooooo o
o ooo o oo ooo x xxxoxxo xo oo xxo xxoxxooxoxo oxoo
ooo oo
xoooooo
ooooooo
o oo o o
o o ooo o oo o o
ooooooo oo
o ooooooo oo
o oo o x x xxxxxxx xxxxxxxxxxxxxoxxxo
xxxxoxoxxoooxooo
o
xoo ooooooo oooo
o
oooooooo
oooo oooooo oo
o
x xxxxxx xx xx x x o oooooooo o ooooooooo
oooooo
o o o o
oooooooooooo oo oo o x xxxxxxxxxxxxxoxxxxo xxoooxoooxoxo xo o
o
ooooo oo
o oooo o o
xx xxxxx xxx xx x xo o ooooooo o ooooooooo ooo
ooooo oo oo xx x xxx xxxxxxxxxxxxxxxxxxxxxxxxo
o xo ooo
xoooooooooooooo
ooo oooo oooo ooooo ooooooo ooo x xx xxxxx xxoxxoxoxoo ox xo ooo
oxxoxo oooooo
oooooooo
x xxxxxx x xx x o o ooo oo
oo ooooooo oo x xx x oooo xxoooxooo o
ooo
oo o o
o o o
xx o ooooo o o xxxxxoxxxox xoo oxo
o
oooooooooo ooooooo
o
x o oooo ooooooooo oo o oo x x xxx ooxox oooo ooxooo oooooooo o
o
o oo oo o oo x x x o o
xx xx
x x
oo oo ooo
o o ooo ooooo
o xx x x o o ooox oo oo o ooo
x x o oo o o
o oo x o o oo o
o o x
x o o
o o

Exposure Exposure

FIGURE 7.3
Distribution of mean outcome and mean exposure levels in three genetic
subgroups (indicated by different symbols and shades of grey) for various
strengths of the instrument, with expected values of the F statistic. One point
of each colour comes from each of 1000 simulated datasets. The IV estimate
in each simulation is the gradient of the line through the three points.
108 Mendelian Randomization

10% of the bias in the observational association estimate. However, weak in-
strument bias depends in a graded way on the F statistic, and such cut-offs are
not always helpful or sensible. Biases less than this, corresponding to greater
F statistics, can be important in practice. Moreover, as explained later in this
chapter, there is an important distinction between the expected F statistic,
on which the magnitude of the bias depends, and the F statistic observed in
a particular dataset.
With a single IV, the expected value of the 2SLS estimate, and hence the
bias, is undefined (Section 4.1.6). Simulations have shown that the median
bias of the 2SLS method with a single IV (or equivalently the ratio or LIML
method) is close to zero even for IVs with expected F statistics around 5, where
the median bias is defined as the difference between the median estimate and
the true value [Burgess and Thompson, 2011].
Other methods, such as likelihood-based methods, are less susceptible to
bias. Although the mean bias of the LIML estimate is undefined (Section
4.3.2), the median bias is close to zero [Angrist and Pischke, 2009]. Simulations
for Bayesian methods for IVs with expected F statistics around 5 have shown
mean and median bias close to zero [Burgess and Thompson, 2012].

7.4.2 Coverage of IV estimates

In addition to problems of bias, IV estimates with weak instruments can have
underestimated coverage [Stock and Yogo, 2002; Mikusheva and Poi, 2006].
As seen in Figure 7.1, the distribution of the IV estimate has long tails, and so
is poorly approximated by a normal distribution. This means that asymptot-
ically derived confidence intervals may underestimate the true uncertainty in
the causal effect. This underestimation is especially severe when confounding
is strong. Simulations for the 2SLS method have shown coverage as low as 75%
for a nominal 95% confidence interval [Burgess and Thompson, 2012]. Similar
results have been observed for the LIML method when there is a large number
of IVs; while a correction is available (Bekker standard errors [Bekker, 1994]),
this leads to inefficient estimates [Davies et al., 2014]. Confidence intervals
from Fieller’s theorem (Section 4.1.5), which are not constrained to be sym-
metric (or even finite), or those which do not rely on asymptotic assumptions,
such as credible intervals from a Bayesian posterior distribution drawn from
Monte Carlo Markov chain (MCMC) sampling, result in better coverage prop-
erties [Imbens and Rosenbaum, 2005]. Alternatively, confidence intervals from
inverting a test statistic, such as the Anderson–Rubin test statistic [Anderson
and Rubin, 1949] or the conditional likelihood ratio test statistic [Moreira,
2003] give appropriate confidence levels under the null hypothesis with weak
instruments [Mikusheva, 2010].
Weak instruments and finite-sample bias 109

7.4.3 Lack of identification

For semi-parametric approaches to IV analysis, such as the generalized method
of moments (GMM) or structural mean models (SMM), there is no guarantee
that a unique parameter estimate will be obtained, as the estimating equa-
tions may have no or multiple solutions (Section 4.4.3*). This is a common
problem when the instrument is weak. Even when there is a unique solution, if
the gradient of the graph of the objective function from the estimating equa-
tions against parameter values is close to zero in the neighbourhood of the
parameter estimate, or if the objective function cannot be well approximated
by a quadratic function, then identification is said to be weak, and problems of
bias and coverage as explained above are likely to occur. Simulations suggest
that the probability of obtaining a unique solution to the estimating equa-
tions with a binary outcome and a log-linear model is not especially sensitive
to the sample size, depending more on the coefficient of determination (R2 ,
the proportion of variance in the exposure explained by the IV(s)) than the
F statistic. With R2 of 2% or less, lack of identification in a multiplicative
GMM (or equivalently a multiplicative SMM) model was observed in over
50% of simulated datasets even when the F statistic was in the hundreds or
even thousands [Burgess et al., 2014c].

7.5 Bias of IV estimates with different choices of IV

Including more instruments, where each instrument explains extra variation
in the exposure, should give more information on the causal parameter (see
Chapter 8). However, bias may increase, due to the weakening of the set of
instruments. In this section, we consider the impact of choice of instrument
on the bias of IV estimates.

7.5.1 Multiple candidate IVs in simulated data

In order to investigate how using multiple instruments affects the bias of IV
estimates, we perform simulations in a model [Burgess and Thompson, 2011]
where, for each participant indexed by i, the exposure xi depends linearly
on six dichotomous IVs (gik , k = 1, . . . , 6), a normally distributed confounder
ui , and an independent normally distributed error term εXi . Outcome yi is
a linear combination of exposure, confounder, and an independent error term
110 Mendelian Randomization

εY i :
6
X
xi = α1k gik + α2 ui + εXi (7.5)
k=1
yi = β1 xi + β2 ui + εY i
ui , εXi , εY i ∼ N (0, 12 ) independently.
We set β1 = 0, α2 = 1, β2 = 1 so that X is observationally strongly posi-
tively associated with Y , but the causal effect is null. We take parameters for
the genetic association α1k = 0.4 for each genetic instrument k, corresponding
to a mean F statistic of 10.2. We used a sample size of 512 divided equally
between the 26 = 64 genetic subgroups. The IVs are uncorrelated, so that the
variation in X explained by each IV is independent, and the mean F statistics
do not depend greatly on the number of IVs (mean 10.2 using 1 IV, 11.3 using
6 IVs).
Table 7.2 shows the median and 95% range of the estimates from the
2SLS and LIML methods and the mean estimate for the 2SLS method using
all combinations of all numbers of IVs as the instrument, with the mean
across simulations of the F statistic for all the instruments used. We also give
results using the IV with the greatest and lowest observed F statistics in each
simulation, as well as using all IVs with an F statistic greater than 10 in
univariate regressions of exposure on each IV.
Using 2SLS, as the number of IVs increases, the bias increases, despite the
mean F statistic remaining fairly constant. This is because there is a greater
risk of imbalances in confounders between the greater number of genetic sub-
groups defined by the instruments. The data are being subdivided in more
different ways, and so there is more chance of these divisions giving genetic
subgroups with different average levels of confounders. However, the variabil-
ity of the IV estimator decreases. This is because a greater proportion of the
variance in the exposure is modelled. The greatest increase in median bias is
from one IV to two IVs, and coincides with the greatest increase in precision.
With the 2SLS method, we therefore have a bias–variance trade-off in deciding
how many IVs to use [Zohoori and Savitz, 1997].
While LIML provides estimates which are slightly more variable than
2SLS, a similar increase in precision with the number of IVs is observed, but
no increase in bias. For 2SLS, the mean estimates are slightly smaller than the
median estimates presented. In the case of a single IV, the theoretical mean
is infinite (Section 4.1.6). For LIML, the mean bias is infinite for all numbers
of IVs (Section 4.3.2).
Using the single IV with the greatest F statistic gives markedly biased
results, despite a mean F statistic of 23.9. There is a similar bias only using
IVs with F > 10. In the simulation, each IV in truth explains the same
amount of variation in the exposure. If the IVs are chosen to be included in
an analysis because they explain a large proportion of the variation in the
exposure in the data under analysis, then the estimate using these IVs is
Weak instruments and finite-sample bias 111

additionally biased. This is because the IVs explaining the most variation will
be overestimating the proportion of true variation explained, due to chance
correlation with confounders. In the notation of Section 7.3.1, ∆U is large
and, having the same sign as α1 , leads to an estimate biased in the direction
of αβ22 . Conversely, if the IV with the least F statistic is used as an instrument,
the IV estimator will be biased in the opposite direction to the observational
association, as shown in Table 7.2.
So we see that if the F statistic is used either to choose between instru-
ments, or via a rule such as only including an IV in the analysis if F > 10, this
procedure itself introduces a selection bias which can be greater in magnitude
than the bias from weak instruments [Hall et al., 1996]. In a more realistic
example, IVs would not all have the same true strength. However, the large
sampling variation in F statistics means that choosing between IVs on the
basis of a single measured F statistic is unreliable. One solution to this in
practice is to use the strength of the IVs in an independent dataset to de-
termine the IVs to include in an applied analysis, or to use an allele score to
summarize multiple variants as a single IV (see Chapter 8).

Median 2.5% to 97.5% quantiles Mean Mean F

IVs used 2SLS LIML 2SLS 1 statistic
1 IV 0.00 −1.12 to 0.53 − 10.2
2 IVs 0.02 −0.54 to 0.39 0.00 −0.64 to 0.39 0.00 10.4
3 IVs 0.03 −0.39 to 0.33 0.00 −0.48 to 0.32 0.02 10.6
4 IVs 0.03 −0.31 to 0.30 0.00 −0.40 to 0.28 0.02 10.8
5 IVs 0.04 −0.26 to 0.27 0.00 −0.34 to 0.26 0.03 11.0
6 IVs 0.04 −0.23 to 0.26 0.00 −0.31 to 0.23 0.03 11.3
Greatest F 0.14 −0.30 to 0.52 − 23.9
Least F −0.32 −2.57 to 0.58 − 6.7
IVs with F > 10 0.11 −0.20 to 0.39 0.10 −0.22 to 0.39 0.11 16.4

TABLE 7.2
Evaluation of bias: Median and 95% range of estimates of β1 = 0 using 2SLS
and LIML methods, mean estimate using 2SLS method and mean F statistic
across 100 000 simulations using combinations of six uncorrelated instruments,
using the instrument with the greatest/least F statistic, and using all instru-
ments with univariate F statistics greater than 10.

1 Mean estimate is reported only when it is not theoretically infinite

112 Mendelian Randomization

7.5.2 Multiple candidate IVs in the Framingham Heart

Study
As a further illustration, we consider the Framingham Heart Study, a cohort
study measuring CRP and fibrinogen at baseline with complete data on 1500
participants for nine SNPs in the CRP gene. The observational estimate of
the log(CRP)–fibrinogen (µmol/l) association is 1.13 (95% CI 1.05 to 1.22).
We calculate the causal estimate of the association using the 2SLS method
with different numbers of SNPs as an instrument, using a per allele additive
model. Figure 7.4 shows a plot of the 2SLS IV estimates against number of
instruments, where each point represents the causal estimate calculated using
the 2SLS method with a different combination of SNPs. The range of point
estimates of the causal effect reduces as we include more instruments, but
the median causal estimate across the different combinations of IVs increases.
The 2SLS estimate using all nine SNPs in an per allele additive model is
−0.01 (95% CI: −0.72 to 0.71, p = 0.99, F9,1490 = 3.34). If we relax the
assumptions of a per allele genetic model with additivity between SNPs to
instead use a fully saturated model with one coefficient for each of the 49
genotypes represented in the data, the 2SLS estimate is 0.79 (95% CI 0.42 to
1.16, p < 0.001, F48,1451 = 1.66). Using LIML, the estimate from the saturated
genetic model is 0.05 (95% CI −0.71 to 0.81, p = 0.89) – much less biased
than the 2SLS estimate.
This illustrates the bias in the 2SLS method due to the use of multiple
instruments, showing how an estimate close to the observational association
can be obtained by injudicious choice of instrument. In the extreme case, if
each of the individuals in a study were placed into separate genetic subgroups,
then the IV estimate would be exactly the observational association. The
LIML method with the saturated genetic model gives a substantially different
answer to the 2SLS method, an indication that the 2SLS estimate may be
biased.

7.6 Minimizing the bias of IV estimates

To provide guidance for epidemiological applications, we now list specific ways
by which bias from weak instruments can be minimized in the design and
analysis of Mendelian randomization studies.

7.6.1 Increasing the F statistic

As stated previously, the bias in 2SLS IV estimates depends on the expected
F statistic in the regression of the exposure on the IV. This means that bias
can be reduced by increasing the expected F statistic. The F statistic is
Weak instruments and finite-sample bias 113

Mean F: 8.6 7.3 6.4 5.6 5.0

0.8
0.6
0.4
IV estimates
0.2
0.0
−0.6 −0.4 −0.2

1 2 3 4 5
Number of instruments
FIGURE 7.4
2SLS IV estimates for causal effect in the Framingham Heart Study of
log(CRP) on fibrinogen (µmol/l) using all combinations of varying numbers
of SNPs as IVs. Point estimates, associated box plots (median, inter-quartile
range, range) and mean F statistics across combinations are displayed.
114 Mendelian Randomization

related to the proportion of variance in the exposure explained by the ge-

netic variants (R2 ), sample size (N ) and number of instruments (K) by the
R2
formula F = ( N −K−1
K ) ( 1−R 2 ). As the F statistic depends on the sample size,

bias can be reduced by increasing the sample size. Similarly, if there are in-
struments that are not contributing much to explaining the variation in the
exposure, then excluding these instruments will increase the F statistic. In
general, employing fewer degrees of freedom to model the genetic association,
that is using parsimonious models, will increase the F statistic and reduce
weak instrument bias, provided that the model does not misrepresent the
data [Pierce et al., 2011; Palmer et al., 2011a]. Simulations have shown that,
even when the true model is only approximately linear in the IV, a per allele
genetic model reduces bias [Burgess and Thompson, 2011].
However, it is not enough to simply rely on an F statistic measured from
data to inform us about bias [Hall et al., 1996]. Returning to the example from
Section 7.2.1 where we divided the Copenhagen General Population Study into
16 equally sized substudies with mean F statistic 10.8, Figure 7.5 shows the
estimates of these 16 substudies using the 2SLS method with their correspond-
ing F statistics. We see that the substudies which have greater estimates are
the ones with larger F statistics; the correlation between F statistics and point
estimates is 0.83. The substudies with higher F statistics also have tighter CIs
and so receive more weight in the meta-analysis. If we exclude from the meta-
analysis substudies with an F statistic less than 10, then the pooled estimate
increases from 0.23 (SE 0.14, p = 0.09) to 0.43 (SE 0.16, p = 0.006). Equally,
if we only use as instruments in each substudy the IVs with an F statistic
greater than 10 when regressed in a univariate regression on the exposure,
then the pooled estimate increases to 0.28 (SE 0.15, p = 0.06). So neither of
these approaches are useful in reducing bias.
Although the expectation of the F statistic is a good indicator of bias,
the observed F statistic shows considerable variation. In the 16 substudies of
Figure 7.5, the measured F statistic ranges from 3.4 to 22.6. In more realistic
examples, assuming similar instruments in each study, larger studies would
have higher expected F statistics which would correspond to truly stronger
instruments and less bias. However, the sampling variation of causal effects
and observed F statistics in each study would still tend to follow the pattern
of Figure 7.5, with larger observed F statistics corresponding to more biased
causal estimates.
So while it is desirable to use strong instruments, the measured strength
of instruments in data is not a good guide to the true instrument strength.
Echoing the comments of Section 7.5 regarding the inclusion of IVs in a model,
any guidance that relies on providing a threshold (such as F > 10) as an
inclusion criterion is flawed and may introduce more bias than it prevents.
Weak instruments and finite-sample bias 115

Study Effect (95% CI)

F statistic: 4.4 −0.91 ( −2.90 , 1.08 )

F statistic: 5.9 −0.80 ( −2.54 , 0.93 )
F statistic: 8.6 −0.67 ( −2.04 , 0.70 )
F statistic: 3.4 −0.44 ( −2.61 , 1.73 )
F statistic: 8.2 −0.41 ( −1.77 , 0.95 )
F statistic: 6.9 −0.27 ( −1.64 , 1.10 )
F statistic: 4.2 −0.14 ( −1.98 , 1.69 )
F statistic: 14.2 −0.12 ( −1.11 , 0.87 )
F statistic: 7.4 −0.01 ( −1.28 , 1.26 )
F statistic: 11.4 0.02 ( −1.06 , 1.10 )
F statistic: 16.5 0.05 ( −0.87 , 0.97 )
F statistic: 10.4 0.18 ( −0.95 , 1.32 )
F statistic: 12.4 0.33 ( −0.65 , 1.30 )
F statistic: 17.2 0.59 ( −0.22 , 1.40 )
F statistic: 22.6 0.74 ( 0.05 , 1.42 )
F statistic: 19.4 0.83 ( 0.11 , 1.54 )

Pooled estimate 0.23 ( −0.04 , 0.50 )

−3.0 −1.8 −0.6 0.4 1.4

FIGURE 7.5
Forest plot of causal estimates of log(CRP) on fibrinogen (µmol/l) using data
from the Copenhagen General Population Study divided randomly into 16
equally sized substudies (each N ≃ 2230). Studies ordered by causal esti-
mate. F statistic from regression of exposure on three IVs. Size of markers is
proportional to weight in a fixed-effect meta-analysis.
116 Mendelian Randomization

7.6.2 Adjustment for measured covariates

If we can find measured covariates that explain variation in the exposure, and
that are not on the causal pathway between exposure and outcome, then we
can incorporate these covariates in our model. This will increase precision in
the genetic association with the exposure and reduce weak instrument bias.
Simulations have shown that we may also see an increase in the precision of
the IV estimator if these covariates are additionally used to explain variation
in the outcome [Burgess et al., 2011b].
As an example, we consider data on interleukin-6 (IL6), a cytokine which is
involved in the inflammation process upstream of CRP and fibrinogen [Hans-
son, 2005]. Elevated levels of IL6 lead to elevated levels of both CRP and
fibrinogen, so IL6 is correlated with short-term variation in CRP [Kaptoge et
al., 2010], but is independent of underlying genetic variation in CRP [CCGC,
2011]. We assume that it is a confounder in the association of CRP with fib-
rinogen and not on the causal pathway (if such a pathway exists). As IL6 has
a positively skewed distribution, we take its logarithm.
We use data from the Cardiovascular Health Study, a cohort study from
the CCGC measuring CRP, IL6 and fibrinogen at baseline, as well as three
SNPs (rs1205, rs1417938, and rs1800947) on the CRP gene, with complete
data for 4137 subjects. The proportion of variance in log(CRP) explained by
log(IL6) is 26%. We calculate the 2SLS IV estimate of the CRP–fibrinogen as-
sociation for each SNP separately and for all the SNPs together in an per allele
additive model, both without and with adjustment for log(IL6) in the first-
and second-stage regressions. Results are given in Table 7.3. We see that after
adjusting for log(IL6) the causal estimate in each case has decreased (reflect-
ing reduced weak instrument bias), its standard error has reduced (reflecting
increased precision), and the F statistic has increased. With adjustment for
a covariate, the relevant F statistic is a partial F statistic, representing the
variation in the exposure explained by the IVs once the variation explained
by the covariate has been accounted for. This is calculated from an analysis
of variance (ANOVA) model.

7.6.3 Borrowing information across studies

The IV estimator would be unbiased if we knew the true values for the average
exposure in different genetic subgroups. In a meta-analysis context [Thompson
et al., 2005], we can combine the estimates of genotype–exposure association
from different studies to give more precise estimates of exposure levels in each
genetic subgroup. In the 2SLS method, an individual participant data (IPD)
fixed-effect meta-analysis for data on individual i in study m with exposure
xim , outcome yim and gikm for number of minor alleles (0, 1, or 2) of genetic
Weak instruments and finite-sample bias 117

Not adjusted Adjusted

IV estimate Estimate (SE) F statistic Estimate (SE) F statistic
Using rs1205 0.219 (0.201) 79.6 0.173 (0.196) 100.2
Using rs1417938 −0.457 (0.407) 27.6 −0.458 (0.362) 37.2
Using rs1800947 0.354 (0.325) 28.6 0.324 (0.316) 36.5
Using all 3 SNPs 0.186 (0.194) 24.4 0.127 (0.188) 32.2

TABLE 7.3
2SLS estimates and standard errors (SE) of the causal effect of log(CRP)
on fibrinogen, and F statistic for regression of log(CRP) on IVs, calculated
using each SNP separately and all SNPs together in per allele additive model,
without and with adjustment for log(IL6) in the Cardiovascular Health Study.

variant k (k = 1, 2, . . . Km ) is:
Km
X
xim = α0m + αkm gikm + εXim (7.6)
k=1
yim = β0m + β1 x̂im + εY im
2
εXim ∼ N (0, σX ); εY im ∼ N (0, σY2 ) independently.
The exposure levels are regressed on the IVs using a per allele additive linear
model separately in each study, and then the outcome levels are regressed on
the fitted values of exposure (x̂im ). The terms α0m and β0m are study-specific
intercept terms. Here we assume homogeneity of variances across studies; we
can use Bayesian methods to allow for possible heterogeneity (see Section 9.6).
If the same genetic variants are measured in each study and are assumed
to have the same effect on the exposure, we can use common genetic effects
(i.e. αkm = αk ) across studies by replacing the first line in equation (7.6) with:
K
X
xim = α0m + αk gikm + εXim (7.7)
k=1

If the assumption of common genetic effects is correct, this will improve the
precision of the fitted values (x̂im ) and reduce weak instrument bias.
To illustrate this, we consider the Copenhagen City Heart Study (CCHS),
Edinburgh Artery Study (EAS), Health Professionals Follow-up Study
(HPFS), Nurses Health Study (NHS), and Stockholm Heart Epidemiology
Program (SHEEP), which are cohort studies or case-control studies measuring
CRP and fibrinogen levels at baseline [CCGC, 2008]. In case-control studies,
we use the data from controls alone since these better represent cross-sectional
population studies. These five studies measured the same three SNPs on the
CRP gene: rs1205, rs1130864 and rs3093077 (or rs3093064, which is in com-
plete linkage disequilibrium with rs3093077). We estimate the causal effect
118 Mendelian Randomization

Causal Observational
Study N F df estimate (SE) estimate (SE)
CCHS 7999 29.6 (3, 7995) −0.286 (0.373) 1.998 (0.030)
EAS 650 6.9 (3, 646) 0.754 (0.327) 1.115 (0.056)
HPFS 405 5.3 (3, 401) 0.758 (0.423) 1.048 (0.081)
NHS 385 6.1 (3, 381) −0.906 (0.636) 0.562 (0.114)
SHEEP 1044 10.5 (3, 1040) 0.088 (0.345) 1.078 (0.051)
Different genetic effects 14.4 (15, 10463) 0.021 (0.195)
Common genetic effects 56.6 ( 3, 10475) −0.093 (0.225)
Study-level estimates 0.234 (0.174)

TABLE 7.4
Estimates of effect of log(CRP) on fibrinogen (µmol/l) from each of five stud-
ies separately and from meta-analysis of studies: number of participants (N ),
F statistic (F ) with degrees of freedom (df) from per allele additive regression
of exposure on three SNPs used as IVs, causal estimate using 2SLS with stan-
dard error (SE), observational estimate with SE. Fixed-effect meta-analyses
conducted using individual-level data with different study-level genetic ef-
fects, common pooled genetic effects, and combining study-level estimates
with inverse-variance weighting.

using the 2SLS method with different genetic effects (model 7.6), common ge-
netic effects (model 7.7) and by a fixed-effect meta-analysis of estimates from
each study.
Table 7.4 shows that the studies analysed separately have apparently dis-
parate causal estimates with large SEs. The meta-analysis estimate assuming
common genetic effects across studies is further from the confounded observa-
tional estimates and closer to the IV estimate from the largest study with the
strongest instruments (CCHS) than the model with different genetic effects,
suggesting that the latter suffers bias from weak instruments.
The pooled estimate from the study-level meta-analysis is greater than
those from the individual-level meta-analyses. Although the CCHS study has
about 8 times the number of participants as SHEEP and 12 times as many as
EAS, its causal estimate has a larger standard error. The standard errors in
the 2SLS method are known to be underestimated when the correlation due to
confounding is strong, especially with weak instruments (Section 7.4.2) [Stock
and Yogo, 2002]. Also, Figure 7.5 showed that causal estimates nearer to the
observational association have lower variance. So a study-level meta-analysis
may be biased due to overestimated weights in the studies with more biased
estimates.
Returning to the example of data from the Copenhagen General Popu-
lation Study considered in Section 7.2.1, if we use the IPD (model 7.6) to
Weak instruments and finite-sample bias 119

Different Common
genetic genetic
Substudies Meta-analysis p-value effects p-value effects p-value
1 −0.05 (0.15) 0.76
5 −0.01 (0.15) 0.95 −0.03 (0.15) 0.85 −0.05 (0.15) 0.75
10 0.09 (0.14) 0.54 0.04 (0.14) 0.80 −0.05 (0.15) 0.76
16 0.23 (0.14) 0.09 0.15 (0.14) 0.26 −0.05 (0.15) 0.75
40 0.46 (0.13) < 0.001 0.30 (0.13) 0.02 −0.04 (0.15) 0.77
100 0.83 (0.11) < 0.001 0.68 (0.11) < 0.001 −0.04 (0.15) 0.77
250 1.27 (0.08) < 0.001 1.15 (0.08) < 0.001 −0.04 (0.15) 0.78

TABLE 7.5
2SLS estimates of causal effect (standard error) of log(CRP) on fibrinogen
from the Copenhagen General Population Study divided randomly into sub-
studies and combined: using fixed-effect meta-analysis of substudy estimates,
and using individual patient data (IPD) with different or common genetic
effects across substudies.

combine the substudies in the meta-analysis rather than combining estimates

from each substudy, then the pooled estimates are somewhat less biased (Ta-
ble 7.5). If we additionally assume common genetic effects across studies
(model 7.7), then we recover close to the original estimate based on analysing
the full dataset as one study: weak instrument bias has been eliminated.

7.7 Discussion
This chapter has demonstrated the effect of weak instrument bias on causal
estimates in real and simulated data. The magnitude of this bias depends on
the statistical strength of the association between instrument and exposure.
Weak instrument bias can reintroduce the problem that IVs were developed
to solve. It is misleading not solely because it biases estimates, but because
estimates suffering from the bias do not provide a valid test of the null hy-
pothesis. Weak instruments may convince a researcher that an observational
association that they have estimated is in fact causal. The reason for the bias
is that the variation in the exposure explained by the IV is not large enough to
dominate the variation in the exposure caused by chance correlation between
the IV and confounders.
While the magnitude of the bias depends on the instrument strength
through the expected or mean F statistic, for a study of fixed size and under-
lying instrument strength, an observed F statistic greater than its expected
value corresponds to an estimate closer to the observational association with
120 Mendelian Randomization

greater precision; conversely an observed F statistic less than the expected

value corresponds with an estimate further from the observational association
with less precision. Simply relying on an F statistic from an individual study
is over-simplistic and threshold rules such as ensuring F > 10 may cause more
bias than they prevent.

7.7.1 Bias–variance trade-off

Using the 2SLS method, we demonstrated a bias–variance trade-off for the
number of instruments used in IV estimation. For a fixed mean F statistic,
as the number of instruments increases, the precision of the IV estimator
increases, but the bias also increases. Using the LIML method, bias did not
increase with the number of instruments, but the precision was slightly lower
than for 2SLS. When using 2SLS, we seek parsimonious models of genetic
association, for example using per allele additive models and including only
IVs with a known association with the exposure, based on biological knowledge
and external information. Provided the data are not severely misrepresented,
these should provide the best estimates of the causal effect. Again, post hoc
use of observed F statistics to choose between instruments may cause more
bias than it prevents.

7.7.2 Combatting weak instrument bias in practice

Ideally, issues of weak instrument bias should be addressed prior to data col-
lection, by specifying sample sizes, instruments, and genetic models using the
best prior evidence available, to ensure that the expected values of F statis-
tics are large. Where this is not possible, our advice would be to conduct
sensitivity analyses using different IV methods, numbers of instruments and
genetic models to investigate the impact of different assumptions on the causal
estimate.
Testing the association between the outcome and each IV in turn (without
estimating a causal effect) is a valid test of a causal relationship even with
weak instruments. If there is a single IV, then an expected F statistic of 5
corresponds to a p-value in the regression of the exposure on the IV of around
0.03. It is perhaps unlikely that an IV would be considered for use in a dataset
if the expected p-value were much greater than 0.03, and so bias from weak
instruments would not be expected to be an issue in practice with a single
IV. If there are multiple IVs, LIML or Bayesian methods could be used in the
analysis, as the estimates from these are less biased than the 2SLS estimate.
A difference between the 2SLS and LIML IV estimates is evidence of possible
bias from weak instruments. The use of Fieller’s theorem, the Anderson–Rubin
test statistic or a Bayesian posterior distribution for inference is recommended.
It is also possible to summarize multiple SNPs into a single variable to
reduce weak instrument bias using an allele score. Details about how to con-
struct such a score are given in Chapter 8.
Weak instruments and finite-sample bias 121

Adjustment for covariates helps reduce weak instrument bias. Including

predictors of the exposure in the first-stage regression, or predictors of the
outcome in the second-stage regression, also increases precision of the causal
estimate. The former will also increase the F statistic for the IVs, and thus
reduce weak instrument bias.
This chapter has considered bias in a one-sample Mendelian randomiza-
tion setting. If the genetic associations with the exposure and outcome are
estimated in non-overlapping sets of individuals, then bias from weak instru-
ments will act in the direction of the null (Section 9.8.2). Although bias is
never welcome, the direction of bias in a two-sample Mendelian randomiza-
tion analysis means that a non-null causal effect estimate will not simply be
an artefact of weak instrument bias.

7.7.3 Bias in study-level meta-analysis

In a meta-analysis context, bias is a more serious issue, as it arises not only
from the bias in the individual studies, but also from the correlation between
causal effect estimates and their variances which results in studies with effects
closer to the observational estimate being over-weighted. By using a single
IPD model, we can reduce the second source of bias. Additionally, we can
pool information on the genetic association across studies to strengthen the
instruments. The assumptions of homogeneity of variances and common ge-
netic effects across studies made in Section 7.6.3 are overly restrictive in prac-
tice; more reasonable extensions of IV methods to a meta-analysis context are
discussed in Chapter 9.

7.7.4 Caution about validity of IVs

Finally, we recall that the use of a genetic instrument in Mendelian randomiza-
tion relies on certain assumptions. In this chapter we have assumed, although
these may fail in finite samples, that they hold asymptotically. If these as-
sumptions do not hold, for example if there were a true correlation between
the instrument and a confounder, then IV estimates can be entirely misleading
[Small and Rosenbaum, 2008].

7.8 Key points from chapter

• Bias from weak instruments can result in seriously misleading estimates of
causal effects. Studies with instruments having large expected F statistics
are less biased on average. However, if a study by chance has a larger
122 Mendelian Randomization

observed F statistic than expected, then the causal estimate will be more
biased.
• Coverage levels with weak instruments can be poorly estimated by methods
which rely on assumptions of asymptotic normality.
• Data-driven choice of instruments or analysis can exacerbate bias. In partic-
ular, any threshold guideline such as ensuring that an observed F statistic
is greater than 10 is misleading. Methods, instruments, and data to be
used should be specified prior to data analysis. Meta-analyses based on
study-specific estimates of causal effect are susceptible to bias.
• Bias can be alleviated by use of measured covariates and parsimonious
modelling of the genetic association (such as a per allele additive SNP
model rather than one coefficient per genotype). This should be accom-
panied by sensitivity analyses to assess potential bias, for example from
model misspecification.

• Bias can be reduced substantially by using LIML, Bayesian and allele score
(see next chapter) methods rather than 2SLS, and bias in practice with a
single IV should be minimal. Nominal coverage levels can be maintained
by the use of Fieller’s theorem with a single IV, and confidence intervals
from the Anderson–Rubin test statistic or Bayesian MCMC methods with
multiple IVs.
8
Multiple instruments and power

In the next two chapters, we consider extensions to IV methods to efficiently

analyse data typically available in Mendelian randomization investigations.
The first extension is the inclusion of multiple instrumental variables in a
single analysis model, and the statistical issues arising. We consider the impact
on statistical power, and discuss the practical issue of missing data, which can
limit power gains.

8.1 Introduction
Although instrumental variable (IV) methods give estimates which are con-
sistent for the causal effect, their variance is typically much larger than the
variance of the estimate from an observational analysis [Davey Smith and
Ebrahim, 2004]. This is because the variation in the exposure explained by
the IV is usually small. If there are multiple IVs available, a more precise
causal effect estimate can be obtained by incorporating data on all the IVs
simultaneously to estimate a single causal effect [Palmer et al., 2011a]. How-
ever, two problems arising from including multiple IVs in an analysis are weak
instruments and missing data.
When there are large numbers of genetic variants, several IV methods
give estimates which are biased in the direction of the observational estimate
with incorrectly sized confidence intervals (see Chapter 7). Allele scores are a
convenient way of summarizing a large number of genetic variants associated
with an exposure. Using a univariate allele score as a single IV rather than
each genetic variant as a separate IV helps resolve problems in IV estimation
resulting from weak instruments.
Sporadically missing genetic data typically arise due to difficulty in in-
terpreting the output of genotyping platforms. If the output is not clear, a
“missing” result is recorded. Hence, although efficiency will be gained from
using multiple instruments, this may be offset in a complete-case analysis due
to more participants with missing data being omitted. Rather than omitting
participants, methods for incorporating participants with partially missing
data can be employed.

123
124 Mendelian Randomization

In this chapter, we address the construction and use of allele scores in

Mendelian randomization (Section 8.2). We investigate the power of an IV
analysis, demonstrating the gain in power from using IVs which explain a
greater proportion of the variance in the exposure (which can be achieved by
including more genetic variants in an analysis) (Section 8.3). We then show
that subjects with partially missing genetic data can be included in an anal-
ysis, enabling multiple IVs to be employed without reducing the available
sample size even if data on each IV is incomplete (Section 8.4). Finally, we
discuss other issues relating to the use of multiple IVs in Mendelian random-
ization analyses (Section 8.5).

8.2 Allele scores

As explained in Chapter 7, the use of large numbers of IVs can result in
bias and poor coverage properties. An allele score (also called a genetic risk
score, gene score, or genotype score) is a single variable summarizing multiple
genetic variants in a univariate score. An unweighted allele score is constructed
as the total number of exposure-increasing alleles present in the genotype of
an individual. A weighted allele score can also be considered, where each allele
contributes a weight reflecting the effect of the corresponding genetic variant
on the exposure. These weights can be derived internally from the data under
analysis, or externally from prior knowledge or an independent data source.
If an individual i has gik copies of the exposure-increasing
PK allele for each
variant k = 1, . . . , K, then their unweighted score is k=1 gik . This score
takes integer values between
PK 0 and 2K. If the weight for variant k is wk , then
their weighted score is k=1 wk gik . Either score can then be used in an IV
analysis using the ratio method, which, as we saw in Chapter 7, has median
bias close to zero.
Another reason for using an allele score is simplicity. With large numbers of
variants, the validity of the IV assumptions can be partially assessed by testing
the association of each variant with a set of measured covariates. Additionally,
the association of the allele score with the covariates can be tested. Assessment
of IV violations will be clearer with a single score variable, rather than many
variants, and power to detect a violation will be improved if several variants
have pleiotropic associations with the same risk factor.
The use of an allele score in Mendelian randomization requires the assump-
tion that the allele score is an instrumental variable. This means that each
variant which contributes to the allele score must satisfy the assumptions of
an instrumental variable, except that it is not necessary for all the variants
to be associated with the exposure (a variant not associated with the expo-
sure but satisfying the second and third IV assumptions will not invalidate
the score, but neither will it add any information to the score). Additionally,
Multiple instruments and power 125

several parametric assumptions are made in specifying the allele score, such
as additivity in the genetic model with no interactions between variants.

8.2.1 Choosing variants to include in an allele score

In Section 7.5.1, we saw that criteria for selecting IVs based on the data
under analysis led to bias. The phenomenon that the magnitude of effect of
the variant with the strongest association is typically over-estimated is known
as the “winner’s curse” (also the Beavis effect) [Taylor et al., 2014]. The choice
of variants to include in an allele score should be made prior to analysis, or
on the basis of external (independent) data. This is particularly important
if there are several candidate variants with similar magnitudes of association
with the exposure. Additionally, the inclusion of variants which are highly
correlated with each other (in high linkage disequilibrium) will not give extra
information compared to including any one of these variants, and may lead
to inefficiency if the correlation is not taken into account in determining the
weights.

8.2.2 Choosing weights in a weighted allele score

If the weights in a weighted allele score are the estimates from a regression of
the exposure on the genetic variants using the data under analysis, then an IV
analysis using an allele score gives precisely the same answer as a two-stage
least squares (2SLS) analysis using each of the variants as separate IVs. In this
case, there is no advantage in using an allele score over a conventional multiple
IV analysis. Weights can be derived from external data, although simulations
have shown that estimates using an unweighted score are unbiased [Burgess
and Thompson, 2013]. Although there is some loss of power associated with
using an unweighted rather than a weighted score, this loss is not large if
the genetic variants have fairly similar magnitudes of association with the
exposure. A similar loss of power is suffered in using a weighted score approach
if the weights are imprecisely estimated.
If external data are not available, weights can instead be estimated using
the data under analysis in a cross-validation approach, by dividing the data
into equal sized parts, and constructing an allele score using weights in each
part estimated using the data from all the other parts. For example, in a
10-fold cross-validation, 10 sets of weights are estimated. Weights used for
constructing the allele score in each tenth of the sample are obtained from the
remaining 90% of the sample. In this way, there is no correlation between the
weights and the data for each individual, and a weighted allele score can be
assigned to each individual in the study using the appropriate set of weights.
A single IV estimate can be obtained using the weighted allele score across
the whole dataset. Alternatively, separate IV estimates can be obtained for
each tenth of the data, and then these estimates can be combined, for example
using a fixed-effect meta-analysis model.
126 Mendelian Randomization

In general, a cross-validation approach would be preferred if external

weights are not available or are thought to be not fully relevant to the data
under analysis. Otherwise, whichever approach gave more precisely estimated
weights would be preferred.

8.2.3 Performance of an allele score in IV estimation

Simulations have shown that the use of an allele score improves bias and cov-
erage properties of IV estimates compared with estimates from the 2SLS and
LIML methods, especially when large numbers of variants are included in the
score [Burgess and Thompson, 2013]. They were also more efficient than 2SLS
and LIML estimates in a simulation example [Davies et al., 2014]. The bias
and coverage properties seem to be robust to misspecifications of the score,
such as the presence of gene–gene and gene–environment interactions, depar-
tures from additivity in the genetic model, and mismeasurement of weights
in a weighted score approach. However, as stated above, they are not robust
to naive procedures which use the data under analysis to construct the score,
nor to the inclusion of invalid IVs in a score.
One important conclusion from this is that the procedure for construct-
ing an allele score in an applied Mendelian randomization analysis should be
described fully and clearly (see Section 5.4).

8.3 Power of IV estimates

In this section, we initially investigate the power of an IV analysis with a
single IV, and then consider the potential benefits of using multiple IVs.

8.3.1 Power with a single IV, continuous outcome

With a single IV and a continuous outcome, the asymptotic variance of the
IV estimate of the causal effect of the exposure X on the outcome Y with a
single IV G is given by the formula:

var(RYIV )
var(β̂1 ) = (8.1)
N var(X) ρ2GX

where N is the sample size, RYIV = Y − β1 X is the residual of the outcome on

subtraction of the causal effect of the exposure, and ρ2GX is the square of the
correlation between the exposure X and the IV G [Nelson and Startz, 1990].
The coefficient of determination (R2 ) in the regression of the exposure on the
IV is an estimate of ρ2GX . The IV in these calculations could either be a single
genetic variant or an allele score. This formula corresponds to the first term
Multiple instruments and power 127

from the delta method expansion (equation 4.9), and ignores uncertainty in
the IV–exposure association; the subsequent calculations therefore represent
the power to detect an association between the IV and the outcome.
The asymptotic variance of the conventional regression (ordinary least
squares, OLS) estimate of the association between the exposure X and the
outcome Y is given by the formula:

var(RYOLS )
var(β̂OLS ) = (8.2)
N var(X)

where RYOLS = Y − βOLS X is the residual of the outcome on subtraction of

the observational association of the exposure. The sample size necessary for
an IV analysis to demonstrate a given magnitude of causal effect is therefore
approximately equal to that for a conventional epidemiological analysis to
demonstrate the same magnitude of association divided by the parameter
ρ2GX for the IV [Wooldridge, 2009].
If the two-sided significance level is α and the power desired to test the null
hypothesis is β, then (assuming approximate normality of the IV estimate)
the sample size required to test a causal effect of size β1 is [Freeman et al.,
2013]:
(z(1− α2 ) + zβ )2 var(RYIV )
N= (8.3)
var(X) β12 ρ2GX
where the quantile function za is the 100a percentile point of a standard
normal distribution. If the significance level is 0.05 and the power is 0.8, then
the sample size required to test a standardized causal effect of β1s (measured
in units of standard deviations in Y per standard deviation increase in X) is
approximately:
7.848
N= 2 2 . (8.4)
β1s ρGX
This assumes that the variance of Y is approximately equal to the variance of
RYIV , which will be true if the causal effect of X does not explain much of the
variation in Y .
For a given sample size N , the power to detect a standardized causal effect
(in the same direction as the true effect) can be calculated as:
√
Power = Φ(β1s ρGX N − z(1− α2 ) ) (8.5)

where Φ is the cumulative distribution function of the standard normal dis-

tribution. This is the inverse function of the quantile function (Φ(za ) = a).
We use these formulae to construct power curves for Mendelian random-
ization using a two-sided significance level α = 0.05. In Figure 8.1, we fix
the squared correlation ρ2GX at 0.02, meaning the variant explains 2% of the
variance of the exposure, and vary the size of the standardized causal effect
β1s = 0.05 to 0.3 and the sample size N = 1000 to 10 000. In Figure 8.2,
we fix the size of the standardized causal effect at β1s = 0.2 and vary the
128 Mendelian Randomization

squared correlation ρ2GX = 0.005 to 0.03 and the sample size as before. In
each of the figures, the power to detect a positive causal effect is displayed;
this tends to 0.025 as the sample size tends to zero. Sample sizes of several
thousands are required to achieve adequate power in settings typical for many
Mendelian randomization studies (modest causal effects, low correlation of
genetic variants with exposure). Code to customize these calculations for a
given scenario is available [Brion et al., 2013] together with an online calcula-
tor (http://glimmer.rstudio.com/kn3in/mRnd/).
100

β1s=0.30
β1s=0.25
β1s=0.20
80

β1s=0.15
β1s=0.10
β1s=0.05
60
Power (%)
40
20
0

0 2000 4000 6000 8000 10000

Sample size

FIGURE 8.1
Power curves with two-sided significance level α = 0.05 varying the sample
size for a fixed value of the IV strength (ρ2GX = 0.02) and different values of
the size of the standardized causal effect (β1s = 0.05 to 0.3) with a single IV.
Multiple instruments and power 129

100

ρ2GX=0.030
=0.005
ρ2GX=0.025
=0.010
ρ2GX=0.020
=0.015
80

ρ2GX=0.015
=0.020
ρ2GX=0.010
=0.025
ρ2GX=0.005
=0.030
60
Power (%)
40
20
0

0 2000 4000 6000 8000 10000

Sample size

FIGURE 8.2
Power curves with two-sided significance level α = 0.05 varying the sample
size for a fixed size of standardized causal effect (β1s = 0.2) and varying the
value of the IV strength (ρ2GX = 0.005 to 0.3) with a single IV.
130 Mendelian Randomization

8.3.2 Power with a single IV, binary outcome

With a single IV, the asymptotic variance of the IV estimate of the causal
effect of the exposure X on the outcome Y with a single IV G can be approx-
imated using the delta method for the ratio estimate. The leading term in the
expansion is:
var(β̂Y |G )
var(β̂1 ) = (8.6)
2
β̂X|G

where β̂Y |G and β̂X|G are the genetic association estimates with the outcome
and exposure respectively.
The sample size for an IV analysis can therefore be approximated by con-
sidering the variance of the coefficient β̂Y |G . Assuming the outcome is binary
(Y = 0 or 1) and using a logistic regression model to obtain β̂Y |G , the variance
of the IV estimate is approximately:
1
var(β̂1 ) = (8.7)
N var(X) ρ2GX P(Y = 1) P(Y = 0)

where P(Y = 1) and P(Y = 0) are the probabilities of the two outcomes for
Y in the sample population (so the proportions of cases and controls in a
case-control study).
The sample size required to detect a standardized causal effect of size β1s
(the log odds ratio per standard deviation increase in X) with 80% power and
a two-sided significance level of α = 0.05 is therefore:
7.848
N= 2 ρ2 . (8.8)
β1s GX P(Y = 1) P(Y = 0)

If there are to be an equal number of cases and controls, P(Y = 1) = P(Y =

0) = 0.5, and:
31.39
N= 2 2 . (8.9)
β1s ρGX
The corresponding power to detect a standardized causal effect of size β1s
with a two-sided significance level of 0.05 is:
p
Power = Φ(β1s ρGX N P(Y = 1) P(Y = 0) − 1.96). (8.10)

We use these formulae to calculate the number of cases needed to obtain

80% power at α = 0.05 in a Mendelian randomization analysis with a binary
outcome for different values of β1s and ρ2GX , assuming a 1:1 ratio of cases to
controls. The results are displayed in Figure 8.3. It is evident that in most
realistic Mendelian randomization contexts (moderate causal odds ratio, low
correlation of genetic variants with exposure), many thousands of cases are
required to achieve adequate power.
These formulae can be used by investigators planning a Mendelian ran-
domization study, or to assess whether their study has adequate power to
Multiple instruments and power 131

detect a causal effect of a given magnitude. Code to customize these calcula-

tions for a given scenario is available [Burgess, 2014] together with an online
calculator (http://spark.rstudio.com/sb452/power/).

8.3.3 Power with multiple IVs

Simulation studies have been performed to estimate power and sample sizes
required with multiple IVs using the 2SLS method [Pierce et al., 2011]. An
advantage of the use of simulation studies in this context is the reliance of
analytical methods on simplifications and approximations. For example, the
expression (8.6) does not take into account uncertainty in the genetic associ-
ation with the exposure. Asymptotic approximations for the variance of the
ratio estimator assume that IV estimates follow normal distributions. This
is known to underestimate the variability of estimates, particularly if the IV
is weak. However, comparisons between analytical and simulation approaches
have generally shown a good level of agreement [Freeman et al., 2013]. In
the absence of confounding, when the coefficient of determination (R2 ) in
the regression of the exposure on the IVs is constant, varying the number of
variants does not seem to affect the power [Pierce et al., 2011]. When there is
confounding, using additional variants increases weak instrument bias, making
the comparison of power levels using the 2SLS method problematic.
However, in practice, using multiple variants will also increase the pro-
portion of the variance in the exposure explained by the IVs. As shown in
the previous two sections, gains in power from increasing the strength of the
IV are substantial, giving motivation to researchers to find and use multiple
variants in Mendelian randomization analyses.

8.4 Multiple variants and missing data

We illustrate the gain in precision from using multiple genetic variants and
the problems of missing data (particularly genetic data) using the British
Women’s Heart and Health Study (BWHHS), one of the constituent studies
of the CRP CHD Genetics Collaboration (CCGC).

8.4.1 Data from the British Women’s Heart and Health

Study
We examine the causal effect of C-reactive protein (CRP) on fibrinogen using
three single nucleotide polymorphisms (SNPs) in the CRP gene coding region
as IVs: rs1205, rs1130864, and rs1800947. Although log(CRP) and fibrinogen
are positively correlated (r = 0.45), it is not thought that long-term variation
132 Mendelian Randomization

100K
Number of cases required for 80% power
ρ2GX=0.005
ρ2GX=0.010

80K
ρ2GX=0.015
ρ2GX=0.020
ρ2GX=0.025
60K
40K
20K
5K ρ2GX=0.030

1.1 1.2 1.3 1.4 1.5

Odds ratio per SD increase in risk factor

50K
Number of cases required for 80% power

ρ2GX=0.01
ρ2GX=0.02
40K

ρ2GX=0.03
ρ2GX=0.05
ρ2GX=0.08
30K
20K
10K
2K

1.1 1.2 1.3 1.4 1.5

Odds ratio per SD increase in risk factor

FIGURE 8.3
Number of cases required (assuming an equal number of controls) in a
Mendelian randomization analysis with a binary outcome and a single in-
strumental variable for 80% power with a 5% significance level varying the
size of the standardized causal effect (odds ratio per standard deviation in-
crease in exposure) for different values of IV strength (ρ2GX : 0.005 to 0.03 in
top panel, 0.01 to 0.08 in bottom panel).
Multiple instruments and power 133

in CRP is causally associated with levels of fibrinogen. As CRP has a skewed

distribution, a linear association is assumed between log-transformed CRP
and fibrinogen.
Each of the SNPs has some missing data. We use cross-sectional baseline
data on 3693 participants with CRP and fibrinogen data, who have complete
or partial data for the three SNPs. There is missingness in 10.8% of partic-
ipants for rs1205, 1.9% for rs1130864, and 2.6% for rs1800947. Genotyping
was undertaken on two separate occasions for SNP rs1205, and then for SNPs
rs1130864 and rs1800947. Although it is unusual to see so much more missing
data in one SNP than in another, this may be due to the individual character-
istics of that SNP or region of the DNA. 3188 participants have complete data
on all the SNPs. In these complete data, the F statistic in a multiple regres-
sion of log(CRP) on all the SNPs is 16.7. The Sargan overidentification test
(Section 4.5.3) gives p = 0.72, indicating that there is no more heterogeneity
between the causal estimates using different IVs than would be expected by
chance.
Table 8.1 gives the estimates of causal effect from a Bayesian method us-
ing an additive per allele genetic model (Section 4.3.3). For each SNP, the
causal effect is given both using all participants with data on the given SNP,
and for the 3188 individuals with complete data on all three SNPs. We see
that, considering the data on participants with complete data, using all the
SNPs as the IV gives the most precise estimate, with at least a 34% reduction
in standard error compared to the estimate using any of the SNPs individu-
ally. However, a substantial proportion of the data has been discarded in the
complete-case analysis. If we only use SNP rs1130864 as the IV, an additional
421 participants can be included in the analysis, resulting in about a 20%
reduction in the standard error of the causal estimate. Although the gain in
precision is not uniform across all SNPs, with a slight loss of precision in the
causal estimate using SNP rs1800947 as the IV despite a sample size increase
of 396, these results motivate us to use methods for incorporating individuals
with missing data.

8.4.2 Power and missing data

Power can be increased in a study by including individuals with partially
missing data in an analysis. Although missing data is not a problem which is
unique to Mendelian randomization, missing genetic data is a specific problem
in this context. Mendelian randomization studies often have limited power,
and so excluding participants due to the presence of missing data is not the
best strategy if they provide information on the causal effect. Additionally, if
there are multiple genetic variants which can be used as IVs, the aim would
be to include all available genetic variants, but not to exclude participants
with missing data on some of the available SNPs.
Genetic data may be missing for several reasons: an individual may fail to
provide a sample for analysis, consent may not be given for genetic testing,
134 Mendelian Randomization

Participants with Participants with

complete data on SNP complete data on all SNPs
SNP N (sample size = N ) (sample size = 3188)
rs1205 3283 0.03 (0.40) 0.02 (0.49)
rs1130864 3609 −0.15 (0.34) −0.27 (0.43)
rs1800947 3584 −0.22 (0.43) −0.17 (0.41)
All three 3188 −0.10 (0.27)

TABLE 8.1
Estimate and standard error of causal effect of unit increase in log(CRP) on
fibrinogen (µmol/l) using various SNPs as IVs: analyses for participants (N )
with complete data on SNP used as IV in analysis, and for participants with
complete data on all SNPs.

DNA extracted may be of insufficient quality or quantity for analysis, or the

reading from a genotyping platform may be difficult to interpret. In the first
three cases, no genetic data would be available for the individual, and they
would not contribute greatly to the estimation of the causal effect. In the
fourth case, data may be available for several individuals on some variants,
but a missing result may be reported for one or more variants. By imputing
missing genetic data, we can include all participants in an IV analysis using
all the genetic variants as IVs, while appropriately acknowledging uncertainty
in the imputation. If the genetic variants are highly correlated (in high linkage
disequilibrium, LD), then the imputation of missing genetic data may be pos-
sible with little uncertainty, and there may be little loss of precision compared
to a hypothetical complete-data analysis if all the data were available (or little
over-precision if the uncertainty in the imputation procedure is ignored).

8.4.3 Methods for incorporating missing data

Here we present a brief description of IV methods for handling missing data;
further details for interested readers are available elsewhere [Burgess et al.,
2011a]. A difficulty with the two-stage method is that uncertainty in the first-
stage regression is not acknowledged in the second-stage regression even with-
out missing data (Section 4.3.5). There is no clear way to account for uncer-
tainty in imputed data in the first-stage regression. This is not a difficulty in
likelihood-based methods, such as in a Bayesian framework. Likelihood-based
methods usually assume that data are “missing at random” (MAR), meaning
that the probability that a data value is missing depends only on the observed
data values of the measured variables [Little and Rubin, 2002].
Genetic data can be imputed using many software packages, including Bea-
gle [Browning, 2006; Browning and Browning, 2007] and fastPHASE [Scheet
and Stephens, 2006]. Output from these packages can be obtained in the form
Multiple instruments and power 135

of posterior probabilities of the number of variant alleles for each SNP in indi-
viduals, or as imputed datasets randomly drawn from the same posterior dis-
tributions. Either multiple imputed datasets (multiple imputations method),
or the posterior probabilities (SNP imputation method) can be used as inputs
in an analysis model. Both approaches acknowledge uncertainty in the impu-
tation process; however, as the imputation and analysis models are performed
separately, there is no feedback between the two stages. Alternatively, impu-
tation can be performed as part of the analysis model using a latent variable
approach, modelling the haplotypes using a multivariate normal distribution
(latent variable method) [Lunn et al., 2006], or by modelling the probability of
an individual having a given set of haplotypes directly using knowledge about
the structure of the data and the prevalence of known haplotype patterns
(haplotype imputation method).

8.4.4 Results of missing data analyses

We apply the four imputation methods sketched out above. Each of the meth-
ods gives fairly similar answers; the point estimates are all nearer zero than
that from the complete-case analysis (Table 8.2). The reduction in the stan-
dard error for all missing data methods compared to the complete-case analysis
is 8–12%. Assuming that the precision (= 1/variance) of the causal estimate
increases proportionally to the sample size, this corresponds to a 17–29% in-
crease in effective sample size, slightly more than the true increase in sample
size of 16% (3693 compared to 3188 individuals).
It is perhaps surprising to find a gain in precision more than anticipated
from the gain in sample size. However, the increase in sample size within each
of the genetic subgroups is not uniform. In this example, individuals with
imputed data fall disproportionately into the smaller subgroups. This means
that most of the smaller subgroups increase in size by more than 16%, giving
rise to a greater than expected increase in precision. Although this may be
simply good fortune, heterozygotes and minor homozygotes are less easy to
determine from the output of genotyping platforms, and so this may not be
an isolated case.

8.5 Discussion
In this chapter, we have considered using multiple instruments in IV analyses.
Using multiple instruments has the potential to reduce the standard error
of causal estimates, but if there are sporadically missing genetic data, this
increase is offset by a decrease in sample size in a complete-case analysis.
136 Mendelian Randomization

Imputation method Effect (SE) 95% confidence interval

Complete case analysis −0.10 (0.27) −0.70, 0.38
Multiple imputations −0.09 (0.25) −0.62, 0.36
SNP imputation −0.07 (0.25) −0.61, 0.37
Latent variable method −0.04 (0.24) −0.55, 0.40
Haplotype imputation −0.06 (0.25) −0.59, 0.39

TABLE 8.2
Estimate, standard error (SE) and 95% confidence interval of the causal effect
for a unit increase in log(CRP) on fibrinogen (µmol/l) in a complete-case
analysis (N = 3188) and in the entire study population (N = 3693) using
different imputation methods for missing genetic data in the British Women’s
Heart and Health Study.

8.5.1 Heterogeneity and supplementary analyses

In using multiple genetic variants to estimate a single causal effect, the as-
sumption is made that the causal effect identified by each of these variants is
the same (known as ‘no treatment effect heterogeneity’). This may not be true,
even if the variants are all valid IVs. Differences may occur if there are multiple
mechanisms by which the exposure affects the outcome. For example, variants
may be associated with body mass index (BMI) by various mechanisms, such
as suppressing appetite or increasing metabolic rate. If genetic variants can be
categorized as associated with one or other of these mechanisms, then sepa-
rate Mendelian randomization estimates can be obtained using each category
of variants. A Mendelian randomization estimate constructed using variants
associated with BMI through appetite suppression more closely represents
the causal effect of intervening on BMI via appetite suppression [Hernán and
Taubman, 2008]. Differences in the causal estimates using genetic variants as-
sociated with different mechanisms may be informative in understanding the
aetiology of the disease, and may highlight specific mechanisms to prioritize
for pharmacological intervention (Section 6.3.1).

8.5.2 Subsample Mendelian randomization

Especially when the outcome is binary, the sample size required in a Mendelian
randomization experiment may be prohibitively large due to the expense of
collecting data on the exposure. In this case, a subsample IV approach may be
a cost-effective approach. Rather than collecting data on the exposure from the
entire study sample, exposure data can be measured for a random subsample
of (control) participants. As the association between the IV and the exposure
is typically stronger than that between the IV and the outcome, the precision
of the IV estimate may not be noticeably affected by reducing the sample size
Multiple instruments and power 137

on which the exposure is measured. Simulations have shown that a subsample

IV analysis with exposure data on only 10% of participants may retain 90% of
the power of the full-sample IV analysis [Pierce and Burgess, 2013]. Estimates
and confidence intervals with a single IV can be calculated using the ratio
method and Fieller’s theorem (Section 4.1.5). With multiple IVs, a modified
version of the two-stage least squares method can be used [Inoue and Solon,
2010].

8.5.3 Relevance to epidemiological practice

The conclusion of Chapter 7 was that problems due to weak instruments,
while potentially serious, are surmountable. Bearing in mind the advice of
Chapter 7, multiple instrumental variables provide an opportunity to obtain
more precise estimates of causal effects. One particular way of incorporating
multiple instrumental variables into an analysis which avoids the danger of
weak instrument bias is the use of an allele score, although care must be taken
in the construction of the score so as not to introduce bias.

8.6 Key points from chapter

• Use of multiple instrumental variables in Mendelian randomization leads
to more precise estimates of causal effects.
• Sporadically missing genetic data may offset this gain, but missing data
methods can recover much of the loss.
• Parsimonious models of genetic association, and in particular allele scores,
can alleviate the problems of weak instruments which may arise when using
large numbers of instrumental variables.
• The procedure for constructing an allele score to be used in an analysis
should be made clear, and in particular how variants and weights for the
score are chosen, as this has a considerable impact on bias.
9
Multiple studies and evidence synthesis

In this chapter, we consider extensions to a simple Mendelian randomization

analysis to include data from multiple studies. We provide methods for com-
bining the information provided by each study in an efficient way to produce
a single causal estimate. Also, we consider how to combine summarized data
on genetic associations from multiple variants in a single study.

9.1 Introduction
In general, the variation in the exposure of interest explained by genetic vari-
ants in Mendelian randomization is small, and so adequately powered inves-
tigations typically require large sample sizes. This often demands synthesis of
evidence from multiple, possibly heterogeneous studies.
In this chapter, we first consider assessment of the causal relationship us-
ing data from multiple studies (Section 9.2) before proceeding to methods for
estimating a pooled causal effect. Methods are presented in order of the homo-
geneity and detail of data required from each constituent study. A study-level
meta-analysis requires the least detail, combining the causal effect estimates
obtained in each study (Section 9.3). However, in order to estimate such a
pooled causal effect, each study needs to measure data on genetic variants,
the exposure and the outcome. A summary-level meta-analysis requires more
detailed information from studies, including information which may not be
routinely reported in a published paper but is increasingly being made avail-
able by large consortia (Section 9.4). An individual-level meta-analysis re-
quires individual participant data (IPD) from studies (Section 9.5). However,
individual-level models are the most flexible for addressing the heterogeneity
of data available in each study. The methods are illustrated and compared us-
ing real data (Section 9.6). We discuss extensions to the meta-analysis model
for binary outcomes (Section 9.7), and conclude with a discussion of applica-
tion of the methods presented in practice (Section 9.8).

139
140 Mendelian Randomization

9.2 Assessing the causal relationship

In Section 3.3, we drew a distinction between assessing a causal relationship
and estimating a causal effect. If assessment of a causal relationship is suffi-
cient, with a single genetic variant or an allele score as the sole instrumental
variable (IV), a causal relationship can be inferred by undertaking a meta-
analysis of the IV–outcome regression coefficients from each of the studies.
Standard inverse-variance weighted methods for meta-analysis are described
in many basic texts; software for performing such analyses is available and
well-documented [Borenstein et al., 2009]. A pooled estimate away from the
null is indicative of a causal relationship.
If genetic variants used as IVs (G) in each study have different magnitudes
of association with the exposure (X), the causal effect of the exposure on the
outcome (Y ) can be examined visually by plotting a graph of the regression
estimates for the G–Y association against the regression estimates for the G–
X association [Minelli et al., 2004]. The points on this graph will be subject to
error in both associations and the gradient of the graph will show the causal
X–Y association. This will be similar to Figure 6.1, although the points will
represent different genetic variants in multiple studies.

9.3 Study-level meta-analysis

If it is possible to estimate the causal effect in each study, a study-level meta-
analysis can be performed directly on these estimated causal effects, for ex-
ample using inverse-variance weighting. However, if the genetic association
with the exposure is small or is measured imprecisely, the asymptotic vari-
ance estimates from each study used in the meta-analysis may be unreliable
measures of uncertainty (Section 7.4.2). Additionally, meta-analysis based on
study-level causal effect estimates tends to exaggerate weak instrument bias
(Section 7.7.3).

9.4 Summary-level meta-analysis

It may be that some studies only provide data on one of the exposure or the
outcome, and so a causal effect cannot be estimated from that study alone.
Even if some study-level causal estimates can be calculated, a more precise
estimate of the pooled causal effect can be obtained using summary-level data,
Multiple studies and evidence synthesis 141

such as estimates of the associations between the genetic variant(s) and each
of the exposure and outcome, or the mean level of the exposure and outcome
in each genetic subgroup. A pooled estimate can be evaluated by combining
data in a hierarchical model, which we now describe.

9.4.1 Multiple genetic variants in a single study

Before considering the possibility of using summary-level data in a meta-
analysis, we explore their use in a single study with multiple genetic variants.
We assume that the estimate of association for genetic variant k = 1, . . . , K
with the exposure is β̂Xk with standard error σXk , and the estimate of associ-
ation with the outcome is β̂Y k with standard error σY k . (The standard error
parameters are assumed to be estimated without uncertainty.) When a study
is used in estimating both the gene–exposure and the gene–outcome associ-
ations, these estimates will be correlated. We assume that these association
estimates can be modelled by a bivariate normal distribution, with correlation
ρ assumed to be the same for each variant:
2
     
β̂Xk ξk σXk ρ σXk σY k
∼ N2 , . (9.1)
β̂Y k ηk ρ σXk σY k σY2 k

A linear association is assumed between the underlying unmeasured means ξk

and ηk . As the genetic association with the outcome is assumed to be zero if
the association with the exposure is zero (from the IV assumptions), we have:

ηk = β1 ξk . (9.2)

The causal effect β1 is assumed to be the same for all variants (Section 8.5.1).
This and subsequent models in this chapter can be estimated either by nu-
merical maximization of the log-likelihood function or by Bayesian methods
[Thompson et al., 2005], the latter for example using WinBUGS [Spiegelhalter
et al., 2003] or MLwiN [Rasbash et al., 2009].
The correlation ρ can be estimated as part of the analysis, but there is
likely to be little information on the parameter in the data [Riley et al., 2007].
We recommend that the value of the parameter be specified as part of the
model, and a sensitivity analysis performed to assess the effect of varying this
parameter value on estimates; its value should be similar to the observational
correlation between the exposure and outcome.
By combining the estimates of association from multiple variants into a
single estimate of the causal effect, an assumption is made that the variants
provide independent information on the causal effect. If the association es-
timates are derived from the same data, then they will not be independent.
However, if the variants are independently distributed (that is, they are not
in linkage disequilibrium, LD), correlation between these estimates should be
low unless the sample size is particularly small. Simulations using indepen-
dently distributed variants have shown that estimates from the summary-level
142 Mendelian Randomization

data model (9.1) and (9.2) are well-behaved even in the presence of statisti-
cal interactions between the effects of the genetic variants (gene–gene inter-
actions). Similar weak instrument bias was observed to estimates from the
two-stage least squares (2SLS) method, and confidence intervals were appro-
priately sized with correct coverage rates. The efficiency of estimates based on
summary-level data was similar to that of estimates based on individual-level
data [Burgess et al., 2013].
If the genetic variants are correlated in their distributions (that is, they
are in LD), then the association estimates β̂Xk (k = 1, . . . , K) will be corre-
lated, as will β̂Y k (k = 1, . . . , K). This can be accounted for in the likelihood
model by a multivariate normal distribution for the genetic association es-
timates from each variant using estimates of the correlations between the
variants (which will be the same as the correlations between the association
estimates) [Burgess et al., 2014e]. This extension is not discussed further here.
If the variants are correlated, then estimates from equation (9.1) will overstate
precision.
For a single study, a further method has been developed for combining
summary-level data on multiple genetic variants not in LD [Johnson, 2011].
This method combines the ratio estimates β̂β̂Y k from each variant in an inverse-
Xk
variance weighted meta-analysis using asymptotic variances calculated from
the delta method for the ratio of two random variables [Dastani et al., 2012].
This variance is:
σY2 k
. (9.3)
2
β̂Xk
This differs from the formula for the standard error of the ratio estimate
given in equation (4.9) as the uncertainty in the genetic association with the
exposure is assumed to be zero.
The combined inverse-variance weighted (IVW) estimate β̂IV W is:
P −2
k β̂Xk β̂Y k σY k
β̂IV W = −2
. (9.4)
2
P
k β̂Xk σY k

The approximate standard error of the estimate is:

s
1
se(β̂IV W ) = P −2
(9.5)
2
k β̂Xk σY k

As the method assumes the ratio estimates are normally distributed, and as
the uncertainty in the genetic associations with the exposure is not accounted
for, the precision of IVW estimates is overstated. However, simulations have
shown that the underestimation of confidence intervals may be slight, with an
average 93% coverage probability for a nominal 95% confidence interval in a
plausibly realistic scenario [Burgess et al., 2013]. Therefore the IVW method
may be a reasonable simpler alternative to a likelihood-based model when
Multiple studies and evidence synthesis 143

the genetic associations with the exposure are estimated precisely. However,
likelihood-based models should be preferred for use in practice where possi-
ble, particularly if the genetic associations with the exposure are estimated
imprecisely.

9.4.2 Single genetic variant in multiple studies

If a single genetic variant (potentially different between studies) is measured
in multiple studies, the same likelihood-based model as above can be used to
combine the genetic association estimates from each study into a single pooled
causal estimate.
For each study m = 1, . . . , M , the estimated G–X association β̂Xm is
2
assumed to be normally distributed with mean ξm and variance σXm and the
estimated G–Y association β̂Y m is normally distributed with mean ηm and
variance σY2 m . The correlation ρ between β̂Xm and β̂Y m is assumed to be
independent of m. This is identical to equations (9.1) and (9.2) except for the
change in the subscripted index.
2
     
β̂Xm ξm σXm ρ σXm σY m
∼ N2 , (9.6)
β̂Y m ηm ρ σXm σY m σY2 m
ηm = β1 ξm
Alternatively, the IVW method can be used, although this is equivalent to
a study-level meta-analysis rather than a summary-level meta-analysis.

9.4.3 Single common genetic variant in multiple studies

If the same single genetic variant has been measured in each of the studies
then, in principle, the same within-study model (9.6) could be used to com-
bine the genetic association estimates from each study. However, this would
not take into account the fact that the genetic variant is the same in each
study. A hierarchical model is therefore proposed, whereby the G–X and G–Y
association parameters are additionally pooled in a second-level (or between-
study) model. This approach also allows the inclusion of studies where only
one of the exposure or outcome have been measured.
Initially, we assume a fixed-effect meta-analysis model; random-effects
models are considered later. For each study m = 1, . . . , M measuring both
the G–X and G–Y associations, the estimated G–X association β̂Xm is as-
sumed to be normally distributed with mean ξ (the same for each study) and
2
variance σXm and the estimated G–Y association β̂Y m is normally distributed
with mean η = β1 ξ and variance σY2 m .
2
     
β̂Xm ξ σXm ρ σXm σY m
∼ N2 , (9.7)
β̂Y m η ρ σXm σY m σY2 m
To include studies where only one of the G–X and G–Y associations has
144 Mendelian Randomization

been reported, we use the marginal distribution of β̂Xm or β̂Y m as appropriate.

For example:
2
β̂Xm ∼ N (ξ, σXm ). (9.8)
Estimation proceeds by direct maximization of the log-likelihood function or
by Bayesian methods, as before [Thompson et al., 2005].

9.4.4 Multiple genetic variants in multiple studies – Genetic

associations
If multiple, potentially different genetic variants are measured in multiple
studies, equation (9.6) can be extended to a hierarchical model:
2
     
β̂Xkm ξkm σXkm ρ σXkm σY km
∼ N2 , (9.9)
β̂Y km ηkm ρ σXkm σY km σY2 km
ηkm = β1 ξkm

where k = 1, . . . , Km indexes genetic variants (first level of the hierarchical

model) and m indexes studies (second level). We initially assume that the
causal effect β1 takes the same value in each study (fixed-effect model); in
other words the same parameter β1 is estimated regardless of which genetic
variants are measured and of how many variants are available in each study.

9.4.5 Multiple genetic variants in multiple studies – Genetic

subgroups
An alternative way of modelling based on summary-level data is to partition
the population into genetic subgroups, each of which contains all the individ-
uals in the study with the same genotype for the measured variants. We index
genetic subgroups by the subscript j = 1, . . . , Jm . For each study m, the mean
of the exposure (X̄jm ) in each subgroup is assumed to come from a normal
2
distribution with mean ξjm and known variance σXjm . Similarly, the mean
of the outcome (Ȳjm ) in each subgroup is assumed to come from a normal
distribution with mean ηjm and known variance σY2 jm .
2
     
X̄jm ξjm σXjm ρ σXjm σY jm
∼ N2 , (9.10)
Ȳjm ηjm ρ σXjm σY jm σY2 jm
ηjm = β0m + β1 ξjm

This model is appropriate even if the genetic variants are in LD, as the
genetic subgroups are defined using all the variants, and so the mean values of
the exposure and outcome in the genetic subgroups are independent. However,
unlike estimates of genetic association, data on the mean values of the exposure
and outcome in genetic subgroups are unlikely to be routinely reported in
publications.
Multiple studies and evidence synthesis 145

9.4.6 Fixed- and random-effects meta-analysis

The models given so far in this chapter represent fixed-effect meta-analyses, as
the same value of ξ (equation 9.7) or β1 (equations 9.9 and 9.10) is assumed
for each study. For a random-effects meta-analysis, we allow study-specific
parameters ξm or β1m to vary between studies, but model them as coming from
a common distribution. This acknowledges the possibility that the parameters
are somewhat different across studies, as is plausible due to the influences
of different population characteristics, but that they are expected to have
generally similar values.
Equation (9.7) can be extended to a random-effects model by allowing the
study-specific parameters ξm and ηm to come from normal distributions:
2
     
β̂Xm ξm σXm ρ σXm σY m
∼ N2 , (9.11)
β̂Y m ηm ρ σXm σY m σY2 m
ξm ∼ N (µξ , τξ2 )
ηm ∼ N (µη , τη2 ).

Pooling of these parameters assumes that the gene–exposure and gene–

outcome associations are similar in each study. The causal effect estimate
is the ratio of the means of the random-effects distributions for ξm and ηm :
µ̂
β̂1 = µ̂ηξ [Thompson et al., 2005]. The variance parameters τξ2 and τη2 are
measures of between-study heterogeneity.
If different genetic variants are measured in each study, then a more gener-
alizable way of modelling heterogeneity between studies is by allowing study-
specific causal effect parameters β1m to come from a common distribution; in
particular, a normal distribution with mean β1 and variance τ 2 . In equation
(9.9), a random-effects meta-analysis model is achieved by replacing the last
line by:

ηkm = β1m ξkm (9.12)

2
β1m ∼ N (β1 , τ ).

In equation (9.10), for a random-effects meta-analysis the last line is re-

placed by:

ηjm = β0m + β1m ξjm (9.13)

2
β1m ∼ N (β1 , τ ).

Additionally, the correlation parameter ρ could be replaced by study-specific

parameters ρm , which could be specified or estimated separately in each study,
and combined in a random-effects distribution if required. If τ = 0, then a
fixed-effect model is recovered.
We generally advocate random-effects models rather than fixed-effect mod-
els in applied investigations, as the assumption of homogeneity for a param-
eter across studies is usually unrealistic. If there is not much heterogeneity
146 Mendelian Randomization

between the studies, then the value of the heterogeneity parameter will be
close to zero, and the random-effects analysis will approximate a fixed-effect
analysis. If there is considerable heterogeneity, then this provides evidence
against the fixed-effect model and in favour of the random-effects model. A
fixed-effect model may be used if there is a strong argument why a parameter
may be similar across studies (for example, if the separate studies were in
fact centres in a clustered investigation using the same protocol and sampling
individuals from the same population). If there are few studies then it may be
difficult to obtain a precise estimate of heterogeneity, and either a fixed-effect
model or an informative prior on the heterogeneity parameter in a Bayesian
analysis may be employed. A fixed-effect model may also be useful in compar-
ing between meta-analysis coefficients from different analyses to ensure that
the differences were not simply due to changes in the heterogeneity estimate.
The hierarchical nature of the above models is now clear: at the first
(within-study) level, a causal parameter (β1m ) is specified in each study, and
at the second (between-study) level, these causal parameters are pooled to
provide a single causal estimate (β1 ). By evaluating the estimates using a
likelihood-based method in a single model, the meta-analysis is performed in
a single step. This is in contrast with a two-step meta-analysis, as described
in Section 9.3, in which the causal estimates are first estimated in each study,
and then the estimates are combined. By performing the analysis in a sin-
gle step, uncertainty in the model is correctly acknowledged and feedback is
allowed between the two levels of the model.

9.4.7 Using published summary-level data

Several consortia with large numbers of participants, such as CARDIoGRAM-
plusC4D for coronary artery disease [Schunkert et al., 2011] and DIAGRAM
for type 2 diabetes [Morris et al., 2012], have published summary-level data
on the association of catalogues of genetic variants with either risk factors or
disease status. These provide precise estimates of genetic associations which
can be used to obtain causal estimates, provided the genetic variants included
in the analysis are restricted to those for which the IV assumptions are valid.
The main advantages of using published data in Mendelian randomiza-
tion are their size and scope. Large meta-analyses of genome-wide association
studies (GWAS) have discovered many genetic variants associated with vari-
ous risk factors which are candidate instrumental variables. The associations
of these variants with the exposure and outcome in large consortia are likely
to be more precisely estimated than in a single study or a more limited meta-
analysis of available studies. However, it is unlikely that published data are
available on the genetic associations with the exposure and with the outcome
on the same set of studies. This may necessitate a two-sample Mendelian ran-
domization analysis strategy (see Section 9.8.2), in which data on the genetic
associations with the exposure and with the outcome are estimated on non-
overlapping sets of individuals [Angrist and Krueger, 1992]. This simplifies
Multiple studies and evidence synthesis 147

the models such as equation (9.1), as the correlation between the genetic as-
sociation estimates with the exposure and with the outcome (ρ) would be
zero.
It may be that gene–exposure and gene–outcome association estimates
taken from the literature are not estimated from a single study, but themselves
represent pooled estimates from meta-analyses. These can be combined across
variants using the methods of Section 9.4.1. However, the heterogeneity across
studies will not be modelled as faithfully as in a hierarchical model using the
study-specific association estimates.

9.4.8 Advantages of summary-level meta-analysis

Summary-level meta-analysis provides a compromise between study-level and
individual-level meta-analysis. In some cases, it may not be possible or feasi-
ble for a researcher to share individual-level data. By sharing summary-level
data, evidence can be included even if a study only provides data on genetic
association estimates or on genetic subgroups. These data may be available
from published work, and should contribute information on the parameter
representing the causal effect, as well as helping to avoid weak instrument
bias by providing an alternative to a study-level analysis.

9.5 Individual-level meta-analysis

If individual participant data (IPD) are available, then rather than modelling
summary-level data, we can model the individual-level data on the exposure
and outcome directly. This enables us to consider the model of genetic asso-
ciation between the genetic variants and the exposure in more detail.

9.5.1 Modelling in a single study

We drop the subscript m where possible to improve readibility. We initially
consider estimates using individual-level data in a single study, although there
is a natural extension to a meta-analysis identical to the summary-level models
considered in the previous section by pooling studies in a hierarchical model on
the causal effect parameter. We index individuals by the subscript i = 1, . . . , N
or i = 1, . . . , Nj as appropriate.
In the summary-level analyses, the assumption of exact knowledge of vari-
ances for each genetic association or mean value in a genetic subgroup is
not strictly appropriate. Indeed, using genetic subgroups if Nj = 1, a group-
specific estimate of variance cannot even be calculated. It is then preferable
2
to base the analysis on the variance of the exposure (σX ) and the outcome
148 Mendelian Randomization

(σY2 ) in the whole population, using an individual-based model. For example,

equation (9.10) becomes
2
     
Xij ξj σX ρ σX σY
∼ N2 , (9.14)
Yij ηj ρ σX σY σY2
ηj = β0 + β1 ξj

where j indexes genetic subgroups and ρ denotes the observational correlation

between the exposure and outcome.

9.5.2 Model of genetic association

In equation (9.14), separate parameters ξj and ηj are included for each ge-
netic subgroup. If a specific model of genetic association is assumed, such as
an additive per allele model, this can be included in the analysis model. If
the model is correct, it should help to provide more precise estimates of the
unknown parameters in the model and should reduce weak instrument bias
(Section 7.5.2). If gik (= 0, 1, 2) is the number of copies of the minor allele for
genetic variant k = 1, . . . , K for individual i, we can write the model as:
2
     
Xi ξi σX ρ σX σY
∼ N2 , (9.15)
Yi ηi ρ σX σY σY2
K
X
ξi = α0 + αk gik
k=1
ηi = β0 + β1 ξi

dropping the subscript j and the division into genetic subgroups. In model
(9.14), there are up to 3K genetic subgroups and an equal number of ξj pa-
rameters in the model of genetic association with the exposure (assuming each
variant is a biallelic SNP and so takes three values). In model (9.15), there
are K + 1 αk parameters (one parameter for each genetic variant and an in-
tercept parameter), meaning that the genetic association with the exposure is
modelled more parsimoniously.

9.5.3 Common genetic variants

In a meta-analysis context, an additive per allele model for genetic association
in each study m can be written as:
Km
X
ξim = α0m + αkm gikm (9.16)
k=1

When the same set of genetic variants has been used in several studies, we
can combine the estimates of genetic association αkm across studies, in the
Multiple studies and evidence synthesis 149

same way as the parameter ξm was combined in equation (9.11). This should
give a more precise model of association in smaller studies and should reduce
weak instrument bias, as instrument strength will be combined across the
studies (Section 7.6.3). Due to possible heterogeneity between populations,
we propose a random-effects model, where we impose a multivariate normal
distribution on the study level parameters αm = (αkm , k = 1, . . . , Km ) with
mean vector µα and variance-covariance matrix Ψ. Note that the intercept
parameters α0m are not pooled, as these depend on the characteristics of each
study population and would not necessarily be similar across studies.
K
X
ξim = α0m + αkm gikm (9.17)
k=1
αm ∼ NK (µα , Ψ)

9.5.4 Lack of exposure or outcome data

Where a study has not measured the exposure but has genetic data in com-
mon with other studies, we can use the random-effects distributions for the
genetic association parameters defined above as a predictive distribution or
implicit prior for the unknown parameters. This requires an assumption that
the mean difference in exposure per additional allele is similar (i.e. can be
drawn from the same random-effects distribution) to that in the other stud-
ies. For identifiability, we set α0m = 0 as with no data on the exposure, this
parameter cannot be estimated. Alternatively, the exposure and IVs could be
centered in each study, so that the intercept parameter α0m would be equal
to zero in each study by design. Studies without data on the outcome can be
included in a meta-analysis in a similar way.

9.5.5 Advantages of individual-level meta-analysis

There are several advantages of analysing individual-level data in Mendelian
randomization studies. The most important reasons are not related to the
estimation of a causal effect, but rather concern the assessment of the as-
sumptions necessary for the validity of the genetic variants as IVs. For exam-
ple, with individual-level data on covariates, the associations of each genetic
variant with measured potential confounders that might bias causal estimates
can be tested. While these assessments can be performed using summary-level
data, it is not usually possible to do this in a systematic way with a range of
covariates.
Individual-level data enable more complex modelling of the genetic as-
sociation with the exposure. This allows the pooling of genetic association
parameters across studies and inclusion of studies without complete informa-
tion on all of the genetic variants, exposure and outcome. As we shall see
in the subsequent applied analysis of the CCGC dataset (Chapter 10), this
150 Mendelian Randomization

enables large gains in precision of causal effect estimates by the inclusion of

additional data.

9.5.6 Combining summary- and individual-level data

In a practical setting, it may be the case that some studies are able to provide
individual-level data and others are able to provide only summary-level data.
The parameters for the genetic association with the exposure (α parameters)
and the parameters for the exposure association with the outcome (β parame-
ters) are the same in the summary- and individual-level models [Sutton et al.,
2008]. Hence, if for example summary-level data are available for each of the
genetic subgroups, the hierarchical model can include individual-level data
from those studies for which they are available, and summary-level data from
all other studies.

9.6 Example: C-reactive protein and fibrinogen

In Section 7.6.3, we provided estimates of the causal effect of C-reactive protein
(CRP) on fibrinogen from five studies. However, the model used for analysis re-
quired the homogeneity of variance of the exposure and outcome in each study.
We re-evaluate the same data using a hierarchical meta-analysis model. Mean
levels of log-transformed CRP and fibrinogen in each of the genetic groups for
the studies are shown in Figure 9.1. Due to the small number of studies, we use
fixed-effect meta-analysis models for the causal effect parameters. An additive
per allele genetic model was used throughout. For the individual-level mod-
els, the parameters of genetic association were estimated in three ways: with
different, study-specific parameters; with parameters common across studies;
and with parameters drawn from a random-effects distribution.
Estimation of the hierarchical models (for summary-level data, equation
(9.10); for individual-level data, equation (9.15) with different genetic effects,
equation (9.16) with common genetic effects, and equation (9.17) with ran-
dom genetic effects) was performed in WinBUGS using vague priors (normal
with mean zero and variance 10002, uniform on the interval [0,20] for posi-
tive valued parameters), except for the standard deviation parameter in the
random-effects distributions for the parameters of genetic association, where
a uniform prior distribution on the interval [0,1] was used.
Results are shown in Table 9.1. For comparison, the confounded observa-
tional association was 1.568 from a fixed-effect meta-analysis of the study-
level observational estimates. We see that the study-level meta-analyses give
the most positive causal effect estimates with the narrowest confidence inter-
vals, in line with the comments of Section 7.6.3 on the effect of weak instru-
ments on study-level meta-analysis. Using the LIML method rather than 2SLS
Multiple studies and evidence synthesis 151

CCHS EAS
9.4

8.5
9.0

8.0
8.6

7.5
8.2

0.6 0.7 0.8 0.9 1.0 0.4 0.6 0.8 1.0 1.2

HPFS NHS
11
12.5

10
fibrinogen
12.0

9
11.5

8
11.0

0.0 0.5 1.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4

SHEEP
11.0
10.6
10.2
9.8

0.0 0.2 0.4 0.6

log−transformed CRP

FIGURE 9.1
Summary plot of mean fibrinogen (µmol/l) against mean log(CRP) (lines are
95% confidence intervals) in each genetic subgroup for five studies. Subgroups
with less than 5 subjects have been omitted; the size of the shaded squares is
proportional to the number of subjects in each subgroup.
152 Mendelian Randomization

Meta-analysis model Estimate 95% CI DIC

Study-level model (2SLS) 0.234 −0.107 to 0.575
Study-level model (LIML) 0.182 −0.172 to 0.536
Summary-level model 0.058 −0.394 to 0.437
Individual-level: different genetic effects 0.108 −0.301 to 0.479 70119
Individual-level: common genetic effects −0.123 −0.733 to 0.348 70125
Individual-level: random genetic effects 0.072 −0.352 to 0.440 70112

TABLE 9.1
Estimates of causal effect (95% confidence intervals, CI) of log(CRP) on fib-
rinogen (µmol/l) from meta-analysis of five studies using study-level (data
on study-specific causal effects combined by inverse-variance weighting),
summary-level (data on genetic subgroups), and individual-level (without and
with pooling of parameters of genetic association): hierarchical models with
deviance information criterion (DIC) for individual-level models.

results in a pooled estimate further from the observational estimate, although

the estimate is still closer to the positive confounded association than those
of the individual-level models. The point estimates in the summary-level and
individual-level model move away from the confounded association towards the
slight negative association estimated in the largest study of the collaboration
(Table 7.4) as the pooling of the parameters of genetic association becomes
more restrictive. The deviance information criterion (DIC) is a Bayesian mea-
sure of model adequacy, where a lower value indicates a better predictive fit
[Spiegelhalter et al., 2002]. Out of the models considered, the model with the
lowest DIC best predicts a replicate dataset which has the same structure as
that currently observed. Only models with the same data structure can be
compared, hence the DIC is only given for individual-level models. A differ-
ence in DIC of 5 to 10 is considered substantial. Using the DIC to assess model
adequacy, the model with random genetic effects is preferred.

9.7 Binary outcomes

Often in Mendelian randomization the outcome of interest is binary. We can
modify the above methods to assume a logistic-linear association between the
outcome and the exposure, thus estimating an odds ratio parameter.
Multiple studies and evidence synthesis 153

9.7.1 Using summary-level data

We again drop the study-level subscript m for clarity. If we have summary-
level data on the genetic associations with the exposure from linear regression,
and with the outcome from logistic regression, then these coefficients can be
included in a model such as equation (9.9). In this case, the causal effect
parameter β1 represents a causal log odds ratio.
If we have summary-level data on genetic subgroups, a binomial distribu-
tion in the outcome model can be assumed for the number of individuals in
genetic subgroup j with events nj (that is, the number with Y = 1) out of the
total number of individuals in the subgroup (Nj ). A linear association can be
assumed between the mean level of the exposure (ξj ) and the linear predictor
(ηj ), which in the example below is the logit of the probability of an event in
the subgroup (logit(πj )):
2
X̄j ∼ N (ξj , σXj ) (9.18)
nj ∼ Binomial(Nj , πj )
ηj = logit(πj ) = β0 + β1 ξj .

A log-linear regression model for the genetic associations with the outcome,
or a log-linear model for relating the outcome and exposure could also be
considered; in this case a causal log relative risk parameter would be estimated.

9.7.2 Using individual-level data

Similarly, we can consider modelling the probability of an event (πi ) for each
individual i. The outcome Yi takes the values 0 (no event) or 1 (event):

Xi ∼ N (ξi , σx2 ) (9.19)

Yi ∼ Binomial(1, πi )
ηi = logit(πi ) = β0 + β1 ξi .

A hierarchical model for meta-analysis can be introduced as in the contin-

uous outcome case.

9.7.3 Combining incident and prevalent cases in a

longitudinal study
In a longitudinal cohort study with a binary outcome, if individuals are not
excluded from study entry at baseline due to history of disease, each partici-
pant has two windows of opportunity to have an event: one before study entry
and one after. We want to include participants in such longitudinal studies
up to twice in the analysis, once in the study viewed retrospectively and once
prospectively. A retrospective analysis is performed by viewing the baseline
data as a cross-sectional case-control study with cases taken as individuals
154 Mendelian Randomization

with previous history of disease (prevalent cases) and controls as all non-
diseased individuals. A prospective analysis excludes all prevalent cases and
considers new incident events within the reporting period. An individual who
is censored at the end of the follow-up period is taken as a control in both
the retrospective and prospective analyses. However, while we do not want to
include the individual’s exposure measurement twice, we want to ensure that
the same odds ratio parameter is estimated in both analyses.
In the corresponding model (9.20), we consider genetic subgroup j, con-
taining N1j individuals, n1j of whom are prevalent cases, and N2j (= N1j −n1j )
non-prevalent individuals, n2j of whom have incident events.
Xij ∼ N (ξj , σ 2 ) for i = 1, . . . N2j non-prevalent individuals (9.20)
n1j ∼ Binomial(N1j , π1j )
n2j ∼ Binomial(N2j , π2j )
logit(π1j ) = η1j = β01 + β1 ξj
logit(π2j ) = η2j = β02 + β1 ξj

This model ensures that the same fitted values of the exposure are used in
both logistic regressions without including individuals twice in the regression
of the exposure on the genetic variants. The causal log odds ratio parameter
is β1 , which is assumed to be the same in the retrospective and prospective
analyses.

9.8 Discussion
In this chapter, we have presented a flexible set of models for meta-analysis of
multiple studies in a hierarchical framework. This allows for the efficient syn-
thesis of summary-level and/or individual-level data from different sources.
Although study-level causal estimates can be combined in a conventional
inverse-variance weighted meta-analysis, such an analysis has a number of
technical deficiencies, and does not allow for the inclusion of extra information
from studies where a causal estimate cannot be obtained. More detailed data
present a number of advantages to the researcher, including the incorporation
of information from studies where the exposure or outcome is not measured,
and the efficient estimation of the genetic model of association where the same
genetic variants are measured in multiple studies.
An advantage of the hierarchical structure is that the whole meta-analysis
can be performed in one step. This keeps each study distinct within the hier-
archical model, only combining information from studies at the top level. This
is more effective at dealing with heterogeneity, both statistical and in study
design, than performing separate meta-analyses on each of the gene–exposure
and gene–outcome associations [Thompson et al., 2005].
Multiple studies and evidence synthesis 155

9.8.1 Precision of the causal estimate

To obtain a precise estimate of the causal effect, one needs to have precise
estimates of both the gene–exposure and gene–outcome associations. A precise
estimate of the gene–exposure association comes from a study with many
participants, such as baseline data in a cohort study. For a binary outcome,
a precise estimate of the gene–outcome association comes from a study with
many events, such as a case-control study. The proposed hierarchical methods
are able to borrow strength across such studies measuring common genetic
variants to provide precise estimates of the genetic associations in all studies,
and therefore obtain a more precise estimate of the causal effect.

9.8.2 Two-sample Mendelian randomization

An extreme example of the above is two-sample Mendelian randomization,
in which the associations between the genetic variant(s) and exposure and
between the variant(s) and outcome are estimated from non-overlapping sets
of individuals. Although this may simply reflect the absence of information on
the exposure and outcome associations in the same participants, it is also a
potentially efficient design strategy for Mendelian randomization, particularly
in view of the increasing public availability of summarized data on genetic
associations with risk factors and disease outcomes from large consortia.
An important assumption therefore, to ensure the validity of the analysis,
is that the two sets of individuals represent samples taken from the same un-
derlying population. If this is not the case, then inferences may be misleading,
as the association of the genetic variants with the exposure may not be repli-
cated in the set of individuals in which the association with the outcome is
estimated.
A further feature of a two-sample analysis is that any bias due to weak
instruments does not act in the direction of the observational confounded
association, but rather in the direction of the null [Inoue and Solon, 2010]. This
means that the use of large numbers of genetic variants should not result in
misleading causal claims. If the data sources for the gene–exposure and gene–
outcome association estimates are partially overlapping (subsample Mendelian
randomization, see Section 8.5.2), then the direction of bias will depend on
the degree of overlap. If the overlap is substantial, then bias will be similar
to a one-sample Mendelian randomization analysis, in the direction of the
observational association. If the overlap is not substantial, then bias will be
similar to a two-sample Mendelian randomization analysis, in the direction of
the null.

9.8.3 Relevance to epidemiological practice

Evidence synthesis is particularly necessary in Mendelian randomization to
obtain sufficiently precise estimates of causal effects to be clinically relevant.
156 Mendelian Randomization

Hierarchical models can be used to combine evidence from multiple sources in

an efficient way. An example of such an analysis using a Bayesian model with
vague priors is given in Chapter 10 for the pooled association of C-reactive
protein on coronary heart disease risk.

9.9 Key points from chapter

• A pooled causal effect estimate can be obtained by combining study-level,
summary-level or individual-level data.
• A single causal effect can be estimated from published data on genetic asso-
ciations with the exposure and with the outcome, either taken from a single
study or from separate sources.
• If the same genetic variants have been measured in several studies, the pa-
rameters of genetic association can be pooled in a hierarchical model across
studies.
• Studies with common genetic variants can contribute to a pooled causal
effect estimate even if data on one of the exposure or the outcome has not
been measured.
10
Example: The CRP CHD Genetics
Collaboration

Much of this book has been motivated and illustrated by data collected by the
CRP CHD Genetics Collaboration [CCGC, 2008]. In this chapter, we anal-
yse the entirety of the CCGC data to estimate the causal effect of C-reactive
protein (CRP) on coronary heart disease (CHD) risk as an illustration of the
Mendelian randomization approach, as well as several of the methodological
issues highlighted in this book. We first give an overview of the complete
dataset and address the validity of the genetic variants as instrumental vari-
ables (IVs) (Section 10.1). We then analyse a single study, exemplifying some
features of the data (Section 10.2), before continuing to present an analysis
of the full dataset (Section 10.3). We conclude this chapter with a discussion,
including the interpretation of the results of this analysis (Section 10.4). A
more detailed analysis of these data is available in a published paper [Burgess
et al., 2012].

10.1 Overview of the dataset

The CCGC collated data from 47 epidemiological studies seeking to ascertain
the causal role of CRP on CHD using a Mendelian randomization approach.
CRP is an acute-phase protein found in the blood which is commonly mea-
sured as a marker of systemic inflammation. As discussed in Section 1.3, it
is known that CRP is observationally associated with CHD, but it is not es-
tablished whether this association is causal. Studies from the collaboration
measured CRP levels, genetic variants relating to CRP, and CHD events.
We restrict attention to participants of European descent, excluding the four
studies with no European descent participants from the analysis. This is to
ensure greater homogeneity of the genetic associations in the different study
populations and to mitigate potential violations of the IV assumptions due to
population stratification. A list of study abbreviations for the 43 studies with
European-descent participants in the CCGC is provided in Table 10.1.
Table 10.2 lists the major statistical features of the studies of the CCGC.

157
158 Mendelian Randomization

AGES The Reykjavik Study of Healthy Aging for the New Millennium
ARIC Atherosclerosis Risk in Communities Study
BHF-FHS British Heart Foundation Family Heart Study
BRHS British Regional Heart Study
BWHHS British Women’s Heart and Health Study
CAPS Caerphilly Study
CCHS Copenhagen City Heart Study
CGPS Copenhagen General Population Study
CHAOS Cambridge Heart Antioxidant Study
CHS Cardiovascular Health Study
CIHDS Copenhagen Ischaemic Heart Disease Study
CUDAS Carotid Ultrasound Disease Assessment Study
CUPID Carotid Ultrasound in Patients with Ischaemic Heart Disease
DDDD Die Deutsche Diabetes Dialyse (4D) Trial
EAS Edinburgh Artery Study
ELSA English Longitudinal Aging Study
EPICNL European Prospective Investigation in Cancer and Nutrition,
Netherlands Centre
EPICNOR European Prospective Investigation in Cancer and Nutrition, Nor-
folk Centre
FRAMOFF Framingham Offspring Study
GISSI Gruppo Italiano per lo Studio della Sopravvivenza nell’Infarto Mio-
cardico
HEALTHABC Health Aging and Body Composition Study
HIFMECH The Hypercoagulability and Impaired Fibrinolytic Function Mech-
anisms Study
HIMS Health in Men Study
HPFS Health Professionals Follow Up Study
HVHS Heart and Vascular Health Study
INTHEART INTERHEART Study
ISIS International Study of Infarct Survival
LURIC The Ludwigshafen Risk and Cardiovascular Health Study
MALMO Malmo Diet and Cancer Study
MONICA/ Monitoring of Trends and Determinants in Cardiovascular Disease/
KORA Cooperative Health Research in the Region of Augsburg Study
NHS Nurses Health Study
NPHSII Northwick Park Heart Study II
NSC Northern Swedish Cohort Study
PENNCATH University of Pennsylvania Catheterization Study
PROCARDIS Precocious Coronary Artery Disease Study
PROSPER Prospective Study of Pravastatin in the Elderly at Risk
ROTT Rotterdam Study
SHEEP Stockholm Heart Epidemiology Program
SPEED Speedwell Study
UCP Utrecht Cardiovascular Pharmacogenetics Study
WHIOS Women’s Health Initiative Observational Study
WHITE2 Whitehall II Study
WOSCOPS West of Scotland Coronary Prevention Study

TABLE 10.1
Abbreviations for the 43 studies with subjects of European descent in the
CCGC.
Example: The CRP CHD Genetics Collaboration 159

Further details on the individual studies can be found in the main published
paper from the collaboration [CCGC, 2011]. We discuss below issues relating
to the study design for studies in the collaboration, as well as relevant details
for the analysis about the exposure, genetic variants, outcome, and various
covariates.

10.1.1 Study design

The collaboration includes prospective studies: cohort studies, and nested
case-control studies (both matched and unmatched); and retrospective studies:
case-control studies (unmatched). In some prospective studies, CRP measure-
ments were not made at recruitment, but rather at a later occasion, which we
have defined as our baseline. Hence, some of the individuals who had incident
events in the original study have prevalent events in the baseline-transformed
study. Four of the studies in the collaboration did not provide individual-level
but only summary-level data on the numbers of individuals with and without
CHD events in each genetic subgroup.

10.1.2 Exposure data: C-reactive protein

The exposure CRP was measured in each study using a high-sensitivity assay.
Some of the studies did not measure CRP for all individuals, and others did
not measure it for any individuals. In retrospective case-control studies, CRP
measurements for cases were excluded from the analysis, as they were mea-
sured after the CHD event, to prevent bias in the causal effect due to reverse
causation. In nested (prospective) case-control studies, blood was drawn and
stored at baseline, to enable pre-CHD event measurement of CRP. However,
in both nested and retrospective case-control studies, oversampling of cases
into the study population compared to the general population biases the as-
sociations between the genetic variant(s) and the exposure. Hence analysis of
CRP measurements is restricted to the controls, who form a more represen-
tative sample of the population as a whole [Bowden and Vansteelandt, 2011].
In prospective cohort studies where individuals with a CHD event at baseline
were not excluded from the study due to the study design, CRP measure-
ments for individuals with prevalent CHD were excluded from the analysis of
the exposure. Table 10.2 lists the number of individuals in each study with a
CRP measurement suitable for use in the IV analysis according to the criteria
above. As CRP has a skewed distribution, log-transformed CRP is used as the
exposure.

10.1.3 Genetic data

The 43 studies in the collaboration with European descent participants
measured different genetic information in the form of single nucleotide
 160
Total Number of subjects with: SNP data 1
Study Study type participants Incident CHD Prevalent CHD CRP data 2 g1 g2 g3 g4
BRHS Cohort with prevalent cases 3824 379 151 3516 X X X X
BWHHS Cohort with prevalent cases 3771 43 236 2970 X X X
CCHS Cohort with prevalent cases 10 259 680 241 9503 X X X
CGPS Cohort with prevalent cases 32 038 188 899 30 491 X X X
CHS Cohort with prevalent cases 4511 793 447 4051 X P X
EAS Cohort with prevalent cases 907 61 28 644 X X X
ELSA Cohort with prevalent cases 5496 71 241 4504 X X X
FRAMOFF Cohort with prevalent cases 1680 46 81 1479 X X X X
PROSPER Cohort with prevalent cases 5777 476 768 4876 X P X
ROTT Cohort with prevalent cases 5406 259 614 4524 X X P
NPHSII Cohort without prevalent cases 2282 99 2158 X X X
WOSCOPS Cohort without prevalent cases 1451 279 1334 X
EPICNOR Nested matched case-control 3298 1074 2126 X X X
HPFS Nested matched case-control 737 200 403 X X X P
NHS Nested matched case-control 684 196 387 X X X P
NSC Nested matched case-control 1673 577 969 X X X X
CAPS Nested unmatched case-control 1157 198 783 X X X
DDDD Nested unmatched case-control 897 269 614 X X X P
EPICNL Nested unmatched case-control 3478 426 3215 X X X P
WHIOS Nested unmatched case-control 3756 1339 1725 X X X P
MALMO Nested unmatched case-control with prevalent cases 2148 530 398 139 X X X X
SPEED Nested unmatched case-control with prevalent cases 854 71 19 564 X X X X
ARIC Unmatched case-control 2261 632 859 X P P
CUDAS Unmatched case-control 1107 56 983 X X
CUPID Unmatched case-control 555 340 193 X X
HIFMECH Unmatched case-control 1006 490 495 X
HIMS Unmatched case-control 3946 522 3077 X X X
ISIS Unmatched case-control 3618 2075 1258 (see Section 10.1.3)
LURIC Unmatched case-control 2747 1137 1599 X X X P
PROCARDIS Unmatched case-control 6464 3126 3302 X X X P
SHEEP Unmatched case-control 2671 1113 1083 X X X
WHITE2 Unmatched case-control 5515 31 4800 X X X
CIHDS Unmatched case-control (CRP in controls only) 6716 2236 4415 X X X
BHF-FHS Unmatched case-control (no CRP data) 4548 2146 0 X X X P
CHAOS Unmatched case-control (no CRP data) 2475 623 0 X X X
GISSI Unmatched case-control (no CRP data) 4034 3054 0 X X X X

Mendelian Randomization
HVHS Unmatched case-control (no CRP data) 4407 1040 0 X P X X
INTHEART Unmatched case-control (no CRP data) 4188 1883 0 X X X X
UCP Unmatched case-control (no CRP data) 2011 922 0 X X X P
AGES Tabular data 3219 800 0 X X X X
HEALTHABC Tabular data 1660 584 0 X X X X
MONICA/KORA Tabular data 1675 272 0 X X X X
PENNCATH Tabular data 1509 1022 0 X X X X
Total 162 416 8392 28 089 103 039

TABLE 10.2
Summary of studies in the CRP CHD Genetics Collaboration with subjects of European descent.

1 g1 = rs1205, g2 = rs1130864, g3 = rs1800947, g4 = rs3093077 or equivalent proxies (P indicates use of a proxy).

2 In case-control studies, CRP data was taken in controls only; in prospective cohort studies, in subjects without prevalent CHD.
Example: The CRP CHD Genetics Collaboration 161

polymorphisms (SNPs) in the CRP gene region. Only SNPs located in this
region were considered as potential IVs to ensure maximal plausibility of the
IV assumptions. The region is on chromosome 1 and is responsible for the pro-
duction of CRP and its regulation. The number of relevant SNPs measured
in each study varied from 1 to 13. Over 20 SNPs in total were measured in at
least one study. Four SNPs were pre-specified in the study protocol as the in-
strumental variables to be used in the analysis: rs1205, rs1130864, rs1800947,
and rs3093077 [CCGC, 2008]. These four SNPs show varying degrees of cor-
relation and give rise to five haplotypes which comprise at least 99% of the
genetic variation exhibited in European descent populations. Indeed, over 99%
of individuals in the CCGC had a genotype which was compatible with these
haplotypes. Only 11 studies measured all four of the pre-specified SNPs. Some
studies measured SNPs which are in complete linkage disequilibrium (LD)
with one of the pre-specified SNPs (r2 > 0.97 in European populations in the
HapMap database), and which are used as proxies for these SNPs. 20 mea-
sured all four SNPs or proxies thereof and an additional 17 measured some
three out of these four. Five of the remaining studies measured fewer than
this, and the final study ISIS measured no SNPs which correspond to any of
these four (in ISIS a single SNP rs2808628, also in the CRP gene region, was
used as an IV).
Proxy SNPs are treated as if they are the SNP of interest. We denote rs1205
(or proxies thereof) as g1, rs1130864 (or proxies thereof) as g2, rs1800947 (or
proxies thereof) as g3, and rs3093077 (or proxies thereof) as g4. Overall minor
allele frequencies were 0.34 for g1, 0.30 for g2, 0.06 for g3, and 0.06 for g4.
There was some sporadic missingness in the genetic data in most of the
studies, although this was rarely greater than 10% per SNP and usually much
less. Table 10.2 lists the pre-specified SNPs measured in each study. We found
that an additive per allele model of association was the most appropriate,
with similar coefficients for the per allele increase in the exposure in each
study [Burgess et al., 2012].

10.1.4 Outcome data: coronary heart disease

The outcome CHD was defined as fatal coronary heart disease (based on Inter-
national Classification of Diseases codings) or nonfatal myocardial infarction
(using World Health Organization criteria). In five studies, coronary stenosis
(more than 50% narrowing of at least one coronary artery assessed by angiog-
raphy) was also included as a disease outcome as it could not be separated
from other CHD events in the data available. Only the first CHD event was
included, so an individual could not contribute more than one event to the
analysis. We use the term ‘prevalent’ to refer to a CHD event prior to blood
draw for CRP measurement and ‘incident’ to refer to a CHD event subsequent
to blood draw.
162 Mendelian Randomization

10.1.5 Covariate data

Data on various covariates were measured in the individual studies, includ-
ing physical variables such as body mass index (BMI), systolic and diastolic
blood pressure; lipid measurements, such as total cholesterol, high-density
lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C),
triglycerides, apolipoprotein A1 (apo A1), and apolipoprotein B (apo B); and
inflammation markers, such as fibrinogen and interleukin-6. Figure 10.1 sum-
marizes the pooled associations of the four SNPs with CRP levels and with a
wide range of 21 covariates from meta-analyses across all studies in the col-
laboration reporting measurements on each SNP and covariate in turn. The
associations represent the standard deviation change in the covariate per allele
change in the SNP. These show strong associations for CRP (p < 10−30 for
each of the four SNPs), but no more significant associations with any other
covariates than would be expected by chance. Out of 84 tested associations
between a covariate and SNP, one had p < 0.01 (p = 0.003 for association be-
tween height and rs1205), and three had p < 0.05. We conclude that there is
no indication of violation of the IV assumptions due to pleiotropic associations
with measured covariates for any of the SNPs.

10.1.6 Validity of the SNPs used as IVs

Although conclusive proof is never possible, there is strong evidence for the
validity of the SNPs as IVs [CCGC, 2011]. First, the SNPs are taken from the
CRP gene region. Scientific knowledge about this genetic region gives strong
plausibility to the specific association of the SNPs with CRP. Secondly, the
genes from the CRP gene region are not known to be in linkage disequilib-
rium (LD) with functional variants in genes outside this region. Thirdly, the
empirical associations of the SNPs with a range of potential confounders are
no stronger than would be expected by chance. These potential confounders
comprise the major known predictors of CHD risk. Fourthly, the genetic as-
sociations with CRP are consistent (up to chance variation) across studies.
Fifthly, the frequencies of genetic variants are consistent (up to chance vari-
ation) across studies. The last two observations support the homogeneity of
the European descent populations, and hence that combining estimates across
different populations is reasonable and meaningful. They are also consistent
with the associations of the genetic variants being due to the effects of the
variants themselves, and not due to the distributions of confounders, which
may differ in each population.
Example: The CRP CHD Genetics Collaboration 163

FIGURE 10.1
Pooled estimates of standard deviation change in covariate per CRP-increasing
allele change in SNP for a range of covariates and the SNPs used for IV anal-
ysis. Estimates and 95% confidence intervals presented are based on random-
effects meta-analyses of study-specific associations.
164 Mendelian Randomization

10.2 Single study: Cardiovascular Health Study

We first analyse a prospective cohort study, the Cardiovascular Health Study
(CHS) [Fried et al., 1991], in detail as a worked example before considering
the other studies. As some individuals entered the study having suffered a
previous CHD event, we analyse the study in two ways for illustrative pur-
poses: retrospectively as a case-control study, where cases are those with a
prevalent CHD event and controls are all other individuals; and prospectively
as a longitudinal study, excluding those with prevalent events and including
only healthy individuals at baseline, where cases are those with an incident
CHD event. In all analyses, we use a logistic model of association so that
an odds ratio parameter is estimated in prospective and retrospective anal-
yses; limitations of this approach in the prospective setting are discussed in
Section 10.4.2.

10.2.1 Results
We analyse the data separately retrospectively and prospectively using some
of the methods of Chapter 4: two-stage, Bayesian, generalized method of mo-
ments (GMM), and structural mean model (SMM) methods. Results are given
in Table 10.3 using each SNP individually as an IV (analyses using an addi-
tional SNP rs2808630, labelled g5, are also presented here, although this SNP
is not used in the overall meta-analysis). We see that the results from differ-
ent methods are similar throughout, with differences between estimates small
compared to their uncertainty. CHS suggests a significantly positive causal
effect of CRP in some of the prospective analyses; this is not representative
of the totality of the data (Section 10.3).

10.2.2 Posterior distributions from Bayesian methods

To illustrate the Bayesian method, the prior (normal with mean zero and
variance 10002) and posterior distributions of the causal effect (β1 ) for the
retrospective logistic analyses using SNPs g1, g2 and g3 separately as IVs are
shown in Figure 10.2, and the distributions using g5 in Figure 10.3. We see
that the posterior distributions using g1, g2 and g3 are very different to the
prior distribution, but that in the case of g5, much of the information in the
posterior distribution comes from the prior. Variant g5 is only weakly asso-
ciated with CRP in the CHS dataset (F statistic = 0.1, p = 0.70). Indeed,
due to the weakness of the variant, convergence in the Monte Carlo Markov
chain (MCMC) algorithm for g5 was not achieved even after a million itera-
tions, as can be seen by the heavy tails of the posterior distribution. Conver-
gence was assessed by sampling from multiple chains using different starting
values in the MCMC algorithm, and examining the Gelman–Rubin plots to
Example: The CRP CHD Genetics Collaboration 165

Prospective analyses (N = 4064, n = 793)

SNP used as IV Two-stage Bayesian GMM SMM
rs1205 (g1) 0.758 (0.295) 0.784 (0.320) 0.844 (0.438) 0.773 (0.319)
rs1417938 (g2) 0.671 (0.475) 0.728 (0.559) 0.721 (0.625) 0.680 (0.494)
rs1800947 (g3) 0.723 (0.556) 0.830 (0.704) 0.834 (0.894) 0.726 (0.579)
rs2808630 (g5) 1 1.889 (6.546)
all 0.725 (0.252) 0.717 (0.264) 0.791 (0.355) 0.737 (0.272)
Retrospective analyses (N = 4511, n = 447)
rs1205 (g1) 0.388 (0.366) 0.408 (0.382) 0.388 (0.388) 0.388 (0.388)
rs1417938 (g2) −0.527 (0.671) −0.531 (0.696) −0.553 (0.766) −0.506 (0.687)
rs1800947 (g3) 0.627 (0.620) 0.864 (0.893) 0.666 (0.806) 0.634 (0.669)
rs2808630 (g5) 1 3.521 (2.614)
all 0.352 (0.322) 0.309 (0.326) 0.314 (0.329) 0.342 (0.330)

TABLE 10.3
Causal log odds ratios (standard errors) of CHD per unit increase in log(CRP)
in prospective and retrospective analyses of the Cardiovascular Health Study
(N = sample size, n = number of events) using two-stage, Bayesian, general-
ized method of moments (GMM), and structural mean model (SMM) meth-
ods.
1 In the Bayesian, GMM, and SMM analyses, the estimates using g5 as an IV failed to

converge.

compare between- and within-chain variance. Similarly, a single causal esti-

mate was not obtained in the semi-parametric (GMM and SMM) approaches.
This corresponds to the example of Figure 4.3, for which the confidence in-
terval from Fieller’s theorem would be unbounded. The two-stage estimates
using g5 should therefore be viewed with suspicion, as the data using g5 as
an IV appear to give little information on a causal effect.

10.3 Meta-analysis of all studies

Having discussed causal estimation in a single dataset, we proceed to consider
the causal estimate based on the whole CCGC dataset, applying the meta-
analysis methods of Chapter 9. First, we look at estimation of the causal
effect using a single SNP as the IV; then we present results using all the
pre-specified SNPs from study-level meta-analyses of two-stage estimates and
from individual-level meta-analyses using Bayesian hierarchical models.
166 Mendelian Randomization

g1
1.0

g2
g3
Prior
0.8
Density
0.6
0.4
0.2
0.0

−4 −2 0 2 4

β1
FIGURE 10.2
Prior and posterior distributions of causal log odds ratio parameter (β1 ) for
retrospective logistic IV analyses of the Cardiovascular Health Study using
SNPs rs1205 (g1), rs1417938 (g2) and rs1800947 (g3). On this horizontal scale,
the prior appears as a flat line at close to zero density.
Example: The CRP CHD Genetics Collaboration 167

g5

0.008
Prior
Density
0.004
0.000

−1000 −500 0 500 1000

β1
FIGURE 10.3
Prior and posterior distributions of causal log odds ratio parameter (β1 ) for
retrospective logistic IV analysis of the Cardiovascular Health Study using
SNP rs2808630 (g5).

10.3.1 Using SNPs one at a time

We calculate causal estimates using each SNP in turn as the sole IV
(Table 10.4). Pooled estimates for the G–X and G–Y associations (beta-
coefficients and standard errors) are obtained from inverse-variance weighted
meta-analyses using a moment estimate for the heterogeneity parameter. The
causal X–Y effect estimates (odds ratios and 95% confidence intervals) are
obtained from the summary-level study-specific G–X and G–Y association
estimates by IV analyses in a Bayesian analysis framework using the hierar-
chical model of equation (9.11) and allowing for heterogeneity in the genetic
association parameters using random-effects models. The correlation param-
eter ρ is taken as 0 and the point estimate is the mean of the posterior
distribution.
The causal estimates using each SNP are similar and all compatible with a
null effect; heterogeneity in the causal effect estimates would be potential evi-
dence against the validity of one or more of the genetic variants as IVs. As the
genetic variants are correlated (in linkage disequilibrium), the causal estimates
are correlated, and so cannot be naively combined in a meta-analysis without
considering the individual- or summary-level data. As none of these analy-
ses uses the totality of the genetic data, an integrated approach is preferred
including all of the SNPs in a single analysis.
168 Mendelian Randomization

SNP Number of Pooled effect p-value Heterogeneity

studies (SE) (I 2 and 95% CI)
g1 29 0.170 (0.010) 2 × 10−78 58% (37–72%)
G–X

g2 32 0.128 (0.007) 1 × 10−75 29% (0–54%)

g3 17 0.263 (0.019) 7 × 10−43 14% (0–51%)
g4 24 0.198 (0.012) 3 × 10−57 8% (0–41%)
g1 39 0.014 (0.013) 0.29 31% (0–54%)
g2 42 0.001 (0.010) 0.91 2% (0–37%)
G–Y

g3 26 0.004 (0.024) 0.86 0% (0–41%)

g4 34 −0.003 (0.023) 0.90 4% (0–32%)
SNP Number of studies Causal estimate (95% CI)
g1 39 1.08 (0.93, 1.26)
g2 42 1.00 (0.84, 1.19)
X–Y

g3 26 1.02 (0.83, 1.24)

g4 34 0.99 (0.78, 1.25)

TABLE 10.4
Pooled estimates from univariate inverse-variance weighted random-effects
meta-analysis of per allele effect on log(CRP) (G–X association) and log odds
of CHD (G–Y association) in regression on each SNP in turn, and hetero-
geneity (I 2 represents the percentage of the variability in effect estimates that
is due to heterogeneity rather than sampling error [Higgins et al., 2003]);
causal estimates (X–Y association) for odds ratio of CHD per unit increase in
log(CRP) from meta-analysis using each SNP as the sole IV from the method
of equation (9.11).

10.3.2 Using all SNPs

We perform meta-analyses of both study-level and individual-level data. In
the study-level data meta-analysis, we undertake a two-stage analysis in each
study with information on genetic variants, the exposure and the outcome
using all the pre-specified SNPs measured in that study as IVs. In cohort
studies, two separate estimates are calculated using the two-stage method
with prevalent and with incident events; the two estimates are then combined
for each study using an inverse-variance weighted fixed-effect meta-analysis to
give a study-specific effect estimate. The study-level estimates are combined
in an inverse-variance weighted random-effects meta-analysis.
In the individual-level meta-analyses, a Bayesian hierarchical model with
vague priors (uniform on the interval [0, 10] for standard deviation and hetero-
geneity parameters, normal with mean zero and variance 10002 for all other
parameters) is used as described in Section 9.5 (in particular, equations 9.15
and 9.12). Analyses are presented based on the same data as the study-level
meta-analysis, as well as on the totality of the data. By pooling the parameters
Example: The CRP CHD Genetics Collaboration 169

of genetic association (Section 9.5.3), an additional 10 studies and over 10 000

extra CHD cases were able to be included in the analysis. These additional
studies either did not measure CRP levels, or only provided summary-level
data. Studies were divided into four groups based on the SNPs measured
in that study, and the parameters of genetic association were pooled across
studies within these groups. Prospective and retrospective analyses of cohort
studies were combined as described in Section 9.7.3. Heterogeneity was ac-
knowledged by the use of random-effects models in both the genetic associa-
tion and causal effect parameters.
Table 10.5 shows the pooled estimates of association. We see that the
causal effect in each analysis is close to the null. When the same data are
used, the point estimates in the two-stage and Bayesian methods are very
similar and the 95% confidence/credible intervals (CIs) are of similar width,
with the Bayesian interval slightly wider. The analysis with pooled genetic
association parameters based on the same data here gave a slight reduction in
precision of the causal estimate because of an increase in the between-study
heterogeneity, but for the analysis based on all of the data, the precision of
the causal effect estimate increased.
These analyses rule out even a small causal effect of long-term CRP levels
on CHD risk. The upper bound of the 95% CI in the final analysis using the
totality of the data available corresponds to an odds ratio of 1.10 for a unit
increase in log(CRP), which is close to a 1 standard deviation increase in
log(CRP).

Method used Studies Events Causal estimate τ̂

Study-level meta-analysis of 1.02
33 24 135 0.121
two-stage estimates (0.91 to 1.15)
Individual-level Bayesian meta- 1.02
33 24 135 0.132
analysis without pooling (0.89 to 1.16)
Individual-level Bayesian meta- 1.01
33 24 135 0.153
analysis with pooling (same data) (0.87 to 1.16)
Individual-level Bayesian meta- 0.99
43 36 463 0.106
analysis with pooling (all data) (0.89 to 1.10)

TABLE 10.5
Causal estimates of odds ratio of CHD per unit increase in log(CRP) using all
available pre-specified SNPs as IVs in random-effects meta-analyses: number
of studies and CHD events included in analysis, estimate of causal effect (95%
confidence/credible interval), heterogeneity estimate (τ̂ , the between-study
standard deviation of the causal log odds ratios); pooling refers to pooling of
the genetic associations with log(CRP) across studies.
170 Mendelian Randomization

10.4 Discussion
This chapter has illustrated methods for the synthesis of Mendelian random-
ization data comprising a variety of study designs and measuring a variety
of genetic variants. Studies with differing design can be analysed separately
and then combined in a study-level meta-analysis, or alternatively analysed
together in an individual-level meta-analysis using a hierarchical model.

10.4.1 Precision of the causal estimate

The individual-level hierarchical method is able to include an additional 10
studies and 50% more events compared to a study-level meta-analysis. A more
precise estimate of the causal effect is obtained. This is illustrated by the width
of the 95% CI of the causal parameter on the log odds ratio scale reducing
from 0.306, 0.343, 0.401 and 0.468 using a single SNP as the IV (Table 10.4),
or 0.232 and 0.260 using the two-stage or hierarchical methods with data on
33 studies (Table 10.5), down to 0.209 in the final hierarchical method with
data on 43 studies (Table 10.5) due to the borrowing of information across
studies and inclusion of studies without measured exposure levels. The use of
the final method represents more than a 110% gain in efficiency compared to
the single SNP analyses of Table 10.4, and more than a 20% gain compared
to the two-stage estimate in this example.

10.4.2 Limitations of this analysis

The main limitation of the methods used is the reliance on parametric as-
sumptions and explicit specification of distributions, such as normality of the
exposure, homogeneity of its variance across genetic subgroups, and additive
per allele models of the IV–exposure association. While there is no evidence
against these assumptions in this example, sensitivity analyses can be used to
quantify the potential impact of violation of these assumptions. For example,
with the CHS study, the semi-parametric approaches (GMM and SMM) gave
similar estimates to the fully-parametric methods.
One particular assumption was that all of the studies could be analysed
using a logistic model of association, although sensitivity analyses have been
performed using alternative models (such as a log-linear model and a Cox
model for a survival outcome, see Section 11.1.1) [Burgess, 2012b]. Studies
with different designs could be analysed using different regression models, such
as conditional logistic regression for matched case-control studies, or propor-
tional hazards regression models for prospective cohort studies. If this were
done, an additional assumption would have to be made in the meta-analysis,
that estimates of somewhat different parameters from studies of different de-
signs can be combined in a single meta-analysis model.
Example: The CRP CHD Genetics Collaboration 171

10.4.3 Assessing the IV assumptions

The combination of individual-level data from multiple studies enables more
detailed assessment of the IV assumptions than summary-level data or data
from a single study. With individual-level data, the associations of the SNPs
used as IVs with numerous measured covariates can be tested in a systematic
way. With data from multiple studies, in addition to the gains in power from
the increased sample size, the IV assumptions can be assessed by inspection
of the homogeneity of genetic associations and haplotype frequencies across
studies. The combination of statistical assessment and scientific knowledge
helps to justify the validity of the genetic variants as IVs in this analysis.

10.4.4 Interpretation of the results

The concept of causation has different meanings to different people. For exam-
ple, to a biochemist, the question of causality is one of function. The question
“Is CRP causally implicated in atherosclerosis?” can be seen as equivalent
to “In the absence of CRP, can atherosclerosis take place?”. If the presence
of CRP is necessary for the formation of atherosclerotic plaques then, on a
biochemical level, CRP is causal for CHD. However, the epidemiological in-
terpretation of the causal question of interest is: “What is the impact of an
increase (or decrease) in CRP levels on CHD risk?”. This is the relevant ae-
tiological question from a clinical point of view where the primary concern is
public health and patient risk. It may be that the level of CRP necessary for
the formation of atherosclerotic plaques is so small that no practical interven-
tion can lower CRP to a level where the CHD risk is reduced. The biochemical
notion of causation is not necessarily relevant to the consequences of an inter-
vention targeted at CRP. The interpretation of the Mendelian randomization
estimate is in terms of the effect of a long-term change in usual levels of CRP
on CHD risk.
While the null association from the Mendelian randomization analysis of
CRP on CHD risk does not preclude a causal effect of a small magnitude,
the rationale for proposing CRP as a target for clinical intervention to re-
duce CHD risk is diminished. The results from this analysis add to a growing
body of evidence that CRP is a bystander of CHD, rather than a causal agent
[Keavney, 2011]. The estimate from a Mendelian randomization analysis is
not attenuated by measurement error or within-individual variation, and rep-
resents the effect of long-term exposure to elevated levels of CRP, so may be
greater in magnitude than that of any potential intervention even if there were
a small causal effect of CRP. However, the results do not preclude the possi-
bility of short-term acutely elevated levels of CRP being part of the process
leading to a CHD event.
172 Mendelian Randomization

10.4.5 Relevance to epidemiological practice

This analysis demonstrates the feasibility of Mendelian randomization for ad-
dressing a clinically important question, but also the great efforts required
to achieve a clinically relevant estimate. The gains in efficiency of the more
sophisticated analyses do not come from additional assumptions, but from the
synthesis of evidence from multiple IVs and multiple studies to give a single
causal estimate based on the totality of the data available. Although it may
seem disappointing to go to all this effort to demonstrate a null finding, not all
analyses using Mendelian randomization have given negative results (Chapter
5), and in many cases, including this one, the demonstration of no clinically
relevant causal effect still has considerable scientific importance.

10.5 Key points from chapter

• The analyses presented in this chapter exemplify the assessment of the
assumptions required to perform a Mendelian randomization analysis and
the estimation of an overall causal effect.
• The integrated analyses presented based on the totality of available data
give a precise enough causal estimate to rule out even a moderately-sized
causal effect of C-reactive protein on coronary heart disease risk.
 Part III

Prospects for Mendelian

randomization
11
Future directions

In this final chapter, we consider the future of Mendelian randomization within

the wider context of genetic epidemiology. We divide the chapter into two sec-
tions. First, we discuss methodological developments in instrumental variable
techniques which enable more sophisticated Mendelian randomization analy-
ses. Secondly, we discuss applied developments, such as advances in genotyping
and other high-throughput cell biology techniques, which widen the scope for
future Mendelian randomization analyses.

11.1 Methodological developments

We consider here areas in need of further methodological development, along-
side recent innovations in instrumental variable (IV) methods, as well as IV
methods which are established in the econometrics literature but have not yet
been applied to the context of Mendelian randomization.

11.1.1 Survival data

For disease incidence in a longitudinal study, rather than the outcome being a
binary indicator of the presence or absence of disease, survival data (also called
time-to-event data) may be available on the length of time each individual was
enrolled in the study prior to a disease event [Collett, 2003]. Typically, such
data are analysed using a proportional hazards model (known as a Cox model)
to investigate the relationship between covariates and disease risk, although
other approaches are also available. The relevant estimate from a proportional
hazards model is a (log) hazard ratio.
As with other forms of data, a causal relationship can be assessed by testing
for an association between the IV and the outcome. With survival data, this
may be performed in a Cox regression model of the survival outcome on the
IV. Parameter estimation is more troublesome, particularly in view of the
non-collapsibility of the hazard ratio (Section 4.2.3*). The precise definition
of a causal hazard ratio, and so the target parameter to be estimated in an

175
176 Mendelian Randomization

IV analysis using survival data, is not clear. Additionally, issues of competing

risks or informative censoring of follow-up times may have to be addressed.
Although ad hoc methods for IV estimation with survival data have been
considered (such as a ratio estimate, the coefficient from the gene–outcome
Cox regression model divided by the coefficient from the gene–exposure linear
regression model, as in Section 5.3.3), a more principled approach should be
possible. A potential approach for this is an accelerated failure-time model,
as this has proved to be a good choice in other aspects of causal modelling
[Robins, 1992]. A pragmatic alternative is to ignore the time-to-event compo-
nent and just consider a binary outcome with a log-linear or logistic model
(as in Chapter 10).

11.1.2 Non-linear exposure–outcome relationships

Although semi-parametric methods (Section 4.4*) are able to weaken the dis-
tributional assumptions made in a fully parametric IV analysis model, a para-
metric model is still necessary for the association between the exposure and
the outcome. In many cases, such as with the association of obesity with all-
cause mortality, the relationship between the exposure and the outcome is
non-linear and may even not be monotone. Non-linear parametric methods
have been considered for use in IV analyses, but inferences based on such
methods have been shown to be highly sensitive to the choice of parameteri-
zation [Mogstad and Wiswall, 2010; Horowitz, 2011].
One reason for this is that most genetic variants have a small effect on
the exposure. For example, the variant (located in the FTO gene region)
explaining the most variation in body mass index (BMI) has an association
of less than 1 kg/m2 per additional allele [Speliotes et al., 2010], whereas
BMI in most populations typically ranges from about 17 to 40 kg/m2 across
individuals. So a Mendelian randomization analysis using this genetic variant
would compare subgroups which differ only slightly in their average level of
BMI. Non-linearities on this reduced scale would not be apparent, and would
not address the clinically relevant question of whether being underweight leads
to an increased risk of death.
The estimate from a linear IV analysis, such as using the ratio method
(Section 4.1), approximates a population-averaged causal effect [Angrist et al.,
1996]. With a linear exposure–outcome relationship, this is the average change
in the outcome for a uniform change (usually a 1 unit increase) in the distribu-
tion of the exposure across the whole population. For a non-linear relationship,
a linear IV estimate approximates the same population-averaged causal effect
when the change in the distribution of the exposure associated with the IV is
small, and the linear IV estimate is scaled to represent the effect of a change
in the exposure of similar magnitude to that associated with a change in the
IV. For example, in the case of BMI, each additional copy of the variant in the
FTO gene region in a European population was estimated to be associated
with a 0.4 kg/m2 increase in BMI [Burgess et al., 2014b]. Hence the linear IV
Future directions 177

estimate using this variant, expressed as the causal effect on the outcome of a
0.4 kg/m2 change in BMI, would approximate the average effect of increasing
the BMI of every individual in the population by 0.4 kg/m2 .
If the exposure–outcome relationship is not monotone (for example, it is
J- or U-shaped), then the true change in the outcome for a given change in
the exposure may be in different directions for various members of the pop-
ulation; but the IV estimate is of the average change in the outcome across
the population [Angrist et al., 2000]. Hence, standard IV methods can still be
used to test for the presence of a causal effect even if the exposure–outcome
relationship is non-linear, and the estimated parameter has a natural inter-
pretation, but any single effect estimate will not tell the whole story of the
causal relationship.
If the shape of the exposure–outcome causal relationship is of interest,
local IV estimates can be obtained within strata of the exposure, such as
deciles or quintiles. By plotting these estimates against the average level of
the exposure in the strata, the shape of the causal relationship can be assessed
graphically. However, if the exposure is stratified on directly, misleading re-
sults may be obtained. This is because the exposure lies on the causal pathway
between the IV and the outcome, and so conditioning on the exposure induces
an association between the IV and confounders. This can be circumvented by
initially subtracting the effect of the IV on the exposure from the exposure
measurement, to obtain the ‘IV-free exposure’. This quantity, representing
the expected value of the exposure for an individual if their IV took the value
zero, can then be safely conditioned on. For the approach to be valid, it is
necessary for the average genetic association with the exposure in the popu-
lation to remain constant at different levels of the exposure [Burgess et al.,
2014b]. Further methodological work is required to assess the robustness of
this approach to violation of this assumption, as well as to stratifying directly
on the exposure in situations where calculating the IV-free exposure may be
problematic, such as if the exposure takes discrete values or has a natural
maximum or minimum value.

11.1.3 Untangling the causal effects of related exposures

Low-density lipoprotein cholesterol (LDL-C) and high-density lipoprotein
cholesterol (HDL-C) are both lipid fractions which have been shown to be ob-
servationally associated with coronary heart disease (CHD) risk. The causal
effects of both lipid fractions on CHD risk have been estimated in Mendelian
randomization investigations using genetic variants specifically associated
with each of LDL-C and HDL-C in turn, and neither associated with the
other, nor with another lipid fraction, triglycerides [Voight et al., 2012] (see
Section 5.4). However, these analyses exclude the majority of variants asso-
ciated with HDL-C and LDL-C, and cannot be performed for triglycerides,
due to a lack of variants associated with triglycerides and not associated with
either HDL-C and LDL-C.
178 Mendelian Randomization

Instrumental variable analyses can be performed for multiple exposure

variables simultaneously. If variants that describe variation in the exposures
of interest have pleiotropic effects, but these effects are restricted to the set of
exposures under investigation, then the causal effects of each of the exposures
can be estimated.
Formally, the assumptions necessary are:
i. the set of genetic variants must be associated with each of the exposures
(it is not necessary for each variant to be associated with every exposure),
ii. each variant must not be associated with confounders of any exposure–
outcome association, and
iii. all causal pathways from a variant to the outcome must pass through one
of the exposures.
This situation is analogous to a factorial randomized controlled trial, where
multiple randomized interventions are simultaneously assessed, and has been
named ‘multivariable Mendelian randomization’ [Burgess and Thompson,
2014]. In the case of the lipid fractions above, such an analysis may have
greater power to detect causal effects than analyses restricted to variants only
associated with each of the individual lipid fractions in turn.
A multivariable Mendelian randomization analysis can be performed us-
ing the two-stage least squares method (Section 4.2). This is undertaken by
first regressing the exposures on the genetic variants in a multivariate mul-
tiple linear regression (first stage; multiple dependent variables and multiple
explanatory variables), and then by regressing the outcome linearly on the fit-
ted values of each of the exposures in a univariate multiple regression (second
stage; one dependent variable and multiple explanatory variables). Alterna-
tively, a multivariable Mendelian randomization analysis can be performed
using summarized data, as in Section 9.4.1, except considering a multivariate
normal distribution for the genetic associations with each of the exposures
and with the outcome. If there are causal effects between the exposures (say,
of one exposure on another), then the causal effect estimates from a multivari-
able Mendelian randomization analysis represent direct causal effects of each
exposure variable on the outcome, not including indirect effects via another
exposure variable.
Inference of a causal effect in multivariable Mendelian randomization relies
on the differential associations of multiple genetic variants with the exposures.
Consequently, the intuitive appeal of using Mendelian randomization to infer
a causal effect from a variant’s sole associations with an exposure and outcome
is somewhat reduced. Additionally, the assumption that the pleiotropic effects
of variants can be completely characterized may be unrealistic. This approach
should therefore only be considered for closely-related exposure variables, and
is not a general purpose way of attempting to deal with pleiotropy.
Future directions 179

11.1.4 Elucidating the direction of causal effect

If two distinct sets of genetic variants are available, each of which consists of
valid IVs for a separate variable, then the direction of causal effect between
the two variables (if any) can be judged by assessing whether each variable in
turn has a causal effect on the other. For example, C-reactive protein (CRP)
and BMI are observationally correlated. A genetic variant in the CRP gene
region has been shown not to be associated with BMI, suggesting that elevated
CRP is not a cause of changes in BMI levels; but a variant in the FTO gene
region has been shown to be associated with CRP, suggesting that elevated
BMI is a cause of increased CRP levels [Timpson et al., 2011]. In this way,
the direction of the causal relationship between obesity and inflammation (in
particular, between BMI and CRP) can be assessed [Welsh et al., 2010].
In additional to there existing no causal relationship or a unidirectional
causal relationship, it is also possible for there to be a reciprocal causal re-
lationship, where each of the variables is a cause of the other [Grassi et al.,
2007]. This may occur due to variation in the causal effects of the variables
across different periods of the life-span [Burgess et al., 2014a].

11.1.5 Investigating indirect and direct effects

Furthermore, if the direction of causation is known for two exposure variables
(and one is the cause of the other), the direct and indirect effects of the pri-
mary exposure (in the example above, BMI) on an outcome via the secondary
exposure (in the example above, CRP) can be considered. The indirect effect
of BMI via CRP represents the effect of BMI on the outcome mediated by
the effect of BMI on CRP. The direct effect of BMI represents the effect of
BMI on the outcome via all other causal pathways, but not via CRP. If IVs
for both the primary exposure and the secondary exposure (referred to as a
mediator) are available, then direct and indirect effects can be calculated in
the presence of unmeasured confounding on the assumption that all effects
are linear without interactions [Burgess et al., 2014a].
It is also possible to consider the indirect and direct effects of a genetic
variant on an outcome, with the exposure as a mediator. For example, genetic
variants linked with smoking have also been shown to be associated with
lung cancer, suggesting a causal effect of smoking on lung cancer risk. But
the indirect effect of the variant mediated by a measure of smoking intensity,
the number of cigarettes smoked per day, was close to zero and the direct
effect via other pathways was similar in magnitude to the overall causal effect
[VanderWeele et al., 2012]. Such a scenario indicates a violation of the IV
assumptions, as a direct effect of the genetic variant on the outcome in a
Mendelian randomization analysis is precluded. However, a more reasonable
interpretation is that the genetic association with the outcome is mediated via
a different pathway through another measure of smoking behaviour, such as
via the amount of nicotine extracted from each cigarette [Le Marchand et al.,
180 Mendelian Randomization

2008]. More generally, mediation analysis can suggest the pathway by which
a genetic variant is associated with the outcome, and hence be informative
about causal mechanisms linking an exposure measure to the outcome.
There has been considerable recent research on mediation analysis, in-
cluding technical definitions of direct and indirect effects [Pearl, 2001], and
investigations into the assumptions necessary for valid estimation of these
effects, in particular relating to unmeasured confounding [VanderWeele and
Vansteelandt, 2009]. When the genetic variant can be assumed to be ran-
domly assigned, as in Mendelian randomization for a valid genetic instrumen-
tal variable, the “no unmeasured confounding” assumptions relating to the
associations between the genetic variant and the exposure and between the
genetic variant and the outcome are automatically satisfied; however addi-
tional assumptions such as no unmeasured confounding between the mediator
and exposure and no post-treatment confounding are still required [Emsley
et al., 2010].

11.2 Applied developments

In this final section, we consider advances in genetic epidemiology leading to
emerging directions for applied Mendelian randomization analyses.

11.2.1 High-throughput cell biology: -omics data

The term “-omics” covers a broad range of fields of study in cell biology
and beyond resulting from developments in high-throughput analytical tech-
niques. Examples of -omics data include gene expression data (genomics),
methylation data (epigenomics), protein data (proteomics), transcription
data (transcriptomics), and metabolites (metabolomics/metabonomics) [Rel-
ton and Davey Smith, 2012a]. Integration of multiple types of -omics data may
give insight into the relations between basic biological biomarkers. Examples
of such approaches have been named ‘genetical genomics’ (integration of ge-
netic variants and gene expression data) [Jansen and Nap, 2001] and ‘genetical
epigenomics’ (integration of genetic variants and epigenetic data) [Relton and
Davey Smith, 2010]. A practical application of the integration of -omics data
with phenotypic and disease data is an investigation into associations between
cigarette smoking behaviours and disease outcomes with DNA methylation to
search for mechanisms by which an increased risk of smoking-related diseases
may persist even after cessation of smoking [Wan et al., 2012].
Relationships between epigenetic markers, proteins, transcription factors
and metabolites can be affected by confounding and reverse causation in the
same way as relationships between phenotypic exposures and outcomes. Al-
though the causal network is generally high-dimensional and unknown, the
Future directions 181

direction of potential causal relationships between types of -omics data can

often be deduced from external biological knowledge (for example, from a ge-
netic variant to gene expression to a protein). A similar analytical approach
to the investigation of indirect and direct effects in Mendelian randomization
has been proposed under the name ‘two-step epigenetic Mendelian randomiza-
tion’ using separate genetic variants as instrumental variables for a phenotype
(exposure) and an epigenetic marker (mediator), to investigate mediation of
the causal effect of the exposure on the outcome [Relton and Davey Smith,
2012b].
A key difficulty here is finding separate genetic variants specifically asso-
ciated with the phenotype and with the epigenetic marker if the two variables
are closely biologically related. Additionally, obtaining relevant data may be
problematic, as several variables (such as methylation and transcription data)
are tissue-specific, and so must be measured in a specific cell type. However,
as technologies for measuring -omics data improve, Mendelian randomization
will be an important tool for understanding biological pathways, particularly
as genetic variants are closer biologically to these cellular variables than they
are to phenotypes such as BMI or blood pressure, and so genetic associations
may be stronger.

11.2.2 Mendelian randomization with GWAS data

A genome-wide association study (GWAS) is a hypothesis-free examination
of the whole genome of individuals in a study population to discover genetic
variants associated with a particular trait. Such studies present difficulties
due to the sheer number of genetic variants analysed and the corresponding
number of association tests. Stringent levels for p-values, such as p < 5 × 10−8,
have been used as a threshold for statistical significance to control the number
of false positive findings. Such a stringent p-value means that the power to
detect relevant variants may be low. However, GWAS investigations have been
successful, discovering hundreds of genetic variants associated with exposures
and disease outcomes [Manolio, 2010], and providing evidence regarding novel
causal pathways and risk factors [Klein et al., 2005].
A GWAS can be used as a source of genetic variants for a Mendelian
randomization analysis. However, when the function of these genetic variants
is unknown, it may be that causal estimates are biased due to violations of
the IV assumptions for one or more variants. Although it is possible for the
associations of multiple genetic variants with the outcome all to be biased due
to pleiotropy, if several genetic variants associated with an exposure are all
concordantly associated with the outcome, then it is more implausible for all
to be due to pleiotropy, particularly if there is a dose–response relationship in
the genetic associations with the exposure and outcome (Section 3.2.6). This
would increase confidence in the conclusion that the exposure is a cause of the
outcome, or at least is a proxy measure of such a cause.
182 Mendelian Randomization

Although a hypothesis-free (agnostic) approach, in which the function of

variants in an analysis is unknown, may give an indication of whether an
exposure is a causal risk factor, neither positive nor negative results should be
over-interpreted. This is particularly relevant when large numbers of genetic
variants are included in the analysis. Analyses using variants from the whole
genome of individuals regardless of the strength of association of the variant
or its function are common in estimating the heritability of traits [Yang et al.,
2011] and in risk prediction [Dudbridge, 2013], but have been shown to give
misleading findings in Mendelian randomization. For example, such analyses
have suggested that CRP is a causal risk factor for CHD risk, but BMI is not
[Evans et al., 2013]. There is no strong justification for using large numbers of
variants from genome-wide data for Mendelian randomization investigations.
Indeed, investigations have shown that the proportion of variance in exposures
explained by externally-derived allele scores can decrease (rather than increase
as might be expected) as the p-value threshold for including a variant in such
a score becomes more liberal [Burgess et al., 2014d].

11.2.3 Whole-genome sequencing and rare variants

Advances in genotyping technology known as ‘next-generation sequencing’
are enabling the measurement of increasing numbers of genetic markers, up to
and including the whole genome of an individual (whole-genome sequencing).
The measurement of rare genetic variants provides opportunities to discover
‘better’ tools for Mendelian randomization: such as variants with stronger as-
sociations with the exposure, or more specific associations with the exposure
if a candidate gene region is pleiotropic. However, Mendelian randomization
simply relies on the genetic variant being specifically associated with the ex-
posure of interest; it is not necessary to find the ‘causal variant’ to perform a
valid Mendelian randomization analysis (Figure 3.2).
The use of rare genetic variants in Mendelian randomization may pose
problems. Genetic variants are not truly randomized in a population, but
rather passed on through Mendelian inheritance. If the variants are fairly
common, then it may reasonable to assume that the variants are ran-
domly distributed with regard to potential confounding variables, and so can
be regarded as being randomized (known as quasi-randomization, see Sec-
tion 2.1.4). But rare variants will be clustered in families, and hence cannot
be regarded as randomly distributed in the population. So while rare genetic
variants are useful for functional genomics, their use in Mendelian randomiza-
tion should be viewed with some caution. Additionally, if the variant is rare,
the power to detect a causal effect may be low.
Future directions 183

11.2.4 Published data and two-sample Mendelian

randomization
Two-sample Mendelian randomization (Section 9.8.2) is the use of separate
datasets to estimate the gene–exposure and gene–outcome associations in a
Mendelian randomization analysis. While such analyses are not altogether
novel, the increasing availability of published resources of genetic associations
with traits (both exposures and disease outcomes) enables the assessment of
causality for an exposure with many outcomes. For example, a functional vari-
ant in the IL6R gene region associated with interleukin-6 receptor is associated
with CHD risk [Swerdlow et al., 2012]. Two-sample Mendelian randomization
can be used to see whether the variant is also associated with a number of
other outcomes, and so whether the interleukin-6 pathway is causal for those
outcomes. For example, the variant is also associated with psoriatic arthritis
and asthma, suggesting common causal risk factors and pathways underlying
the disease outcomes. The availability of genetic associations in public repos-
itories makes these ‘phenome scans’ of genetic associations with multiple risk
factors and disease outcomes a practical option [Burgess et al., 2014e].

11.3 Conclusion
In conclusion, there are still areas of ongoing methodological research in
Mendelian randomization, and work is needed to translate existing and future
methodological developments into the context of Mendelian randomization for
applied researchers. This is fueled to a large extent by increasing data avail-
ability: new exposure variables, increasing detail of genetic measurements, and
publicly-available data resources. These are likely to provide further insights
into causal mechanisms, and further scope for methodological and applied
developments in the future.
Bibliography

Allin, K., Nordestgaard, B., Zacho, J., Tybjærg-Hansen, A., and Bojesen, S.
2010. C-reactive protein and the risk of cancer: a Mendelian randomization
study. Journal of the National Cancer Institute, 102(3):202–206. (Cited on
page 11.)
Almon, R., Álvarez-Leon, E., Engfeldt, P., Serra-Majem, L., Magnuson, A.,
and Nilsson, T. 2010. Associations between lactase persistence and the
metabolic syndrome in a cross-sectional study in the Canary Islands. Eu-
ropean Journal of Nutrition, 49(3):141–146. (Cited on page 11.)
Anderson, T. and Rubin, H. 1949. Estimators of the parameters of a single
equation in a complete set of stochastic equations. Annals of Mathematical
Statistics, 21(1):570–582. (Cited on pages 54 and 108.)
Angrist, J., Graddy, K., and Imbens, G. 2000. The interpretation of instru-
mental variables estimators in simultaneous equations models with an appli-
cation to the demand for fish. Review of Economic Studies, 67(3):499–527.
(Cited on pages 48, 56, and 177.)
Angrist, J., Imbens, G., and Rubin, D. 1996. Identification of causal effects us-
ing instrumental variables. Journal of the American Statistical Association,
91(434):444–455. (Cited on pages 38, 41, and 176.)
Angrist, J. and Krueger, A. 1992. The effect of age at school entry on ed-
ucational attainment: An application of instrumental variables with mo-
ments from two samples. Journal of the American Statistical Association,
87(418):328–336. (Cited on page 146.)
Angrist, J. and Pischke, J. 2009. Mostly harmless econometrics: an empiri-
cist’s companion. Chapter 4: Instrumental variables in action: sometimes
you get what you need. Princeton University Press. (Cited on pages 56, 57,
58, 61, 67, and 108.)
Basmann, R. 1960. On finite sample distributions of generalized classical linear
identifiability test statistics. Journal of the American Statistical Associa-
tion, 55(292):650–659. (Cited on page 68.)
Baum, C., Schaffer, M., and Stillman, S. 2003. Instrumental variables and
GMM: Estimation and testing. Stata Journal, 3(1):1–31. (Cited on pages 67,
68, and 69.)

185
186 Mendelian Randomization

Baum, C., Schaffer, M., and Stillman, S. 2007. Enhanced routines for instru-
mental variables/generalized method of moments estimation and testing.
Stata Journal, 7(4):465–506. (Cited on page 67.)
Bech, B., Autrup, H., Nohr, E., Henriksen, T., and Olsen, J. 2006. Stillbirth
and slow metabolizers of caffeine: comparison by genotypes. International
Journal of Epidemiology, 35(4):948–953. (Cited on page 11.)
Beer, N., Tribble, N., McCulloch, L., et al. 2009. The P446L variant in GCKR
associated with fasting plasma glucose and triglyceride levels exerts its effect
through increased glucokinase activity in liver. Human Molecular Genetics,
18(21):4081–4088. (Cited on page 95.)
Bekker, P. 1994. Alternative approximations to the distributions of instrumen-
tal variable estimators. Econometrica: Journal of the Econometric Society,
62(3):657–681. (Cited on pages 61 and 108.)
Beral, V., Banks, E., Bull D., Reeves, G. (Million Women Study Collabora-
tors) 2003. Breast cancer and hormone-replacement therapy in the Million
Women Study. The Lancet, 362(9382):419–427. (Cited on page 4.)
Bochud, M., Chiolero, A., Elston, R., and Paccaud, F. 2008. A cautionary note
on the use of Mendelian randomization to infer causation in observational
epidemiology. International Journal of Epidemiology, 37(2):414–416. (Cited
on page 32.)
Bochud, M. and Rousson, V. 2010. Usefulness of Mendelian randomization
in observational epidemiology. International Journal of Environmental Re-
search and Public Health, 7(3):711–728. (Cited on page 9.)
Borenstein, M., Hedges, L., Higgins, J., and Rothstein, H. 2009. Introduc-
tion to meta-analysis. Chapter 34: Generality of the basic inverse-variance
method. Wiley. (Cited on page 140.)
Bound, J., Jaeger, D., and Baker, R. 1995. Problems with instrumental vari-
ables estimation when the correlation between the instruments and the en-
dogenous explanatory variable is weak. Journal of the American Statistical
Association, 90(430):443–450. (Cited on pages 99 and 104.)
Bowden, J. and Vansteelandt, S. 2011. Mendelian randomisation analysis
of case-control data using structural mean models. Statistics in Medicine,
30(6):678–694. (Cited on pages 51 and 159.)
Brennan, P., McKay, J., Moore, L., et al. 2009. Obesity and cancer: Mendelian
randomization approach utilizing the FTO genotype. International Journal
of Epidemiology, 38(4):971–975. (Cited on page 89.)
Brion, M.-J., Shakhbazov, K., and Visscher, P. 2013. Calculating statistical
power in Mendelian randomization studies. International Journal of Epi-
demiology, 42(5):1497–1501. (Cited on page 128.)
Bibliography 187

Browning, S. 2006. Multilocus association mapping using variable-length

Markov chains. American Journal of Human Genetics, 78(6):903–913.
(Cited on page 134.)
Browning, S. and Browning, B. 2007. Rapid and accurate haplotype phas-
ing and missing-data inference for whole-genome association studies by use
of localized haplotype clustering. American Journal of Human Genetics,
81(5):1084–1097. (Cited on page 134.)
Buonaccorsi, J. 2005. Encyclopedia of Biostatistics, chapter Fieller’s theorem,
pages 1951–1952. Wiley. (Cited on page 53.)
Burgess, S. 2012a. Statistical issues in Mendelian randomization: use of ge-
netic instrumental variables for assessing causal associations. Chapter 4:
Collapsibility for IV analyses of binary outcomes. PhD thesis, Univer-
sity of Cambridge. Available at http://www.dspace.cam.ac.uk/handle/
1810/242184. (Cited on page 59.)
Burgess, S. 2012b. Statistical issues in Mendelian randomization: use of ge-
netic instrumental variables for assessing causal associations. Chapter 8:
Meta-analysis of Mendelian randomization studies of C-reactive protein and
coronary heart disease. PhD thesis, University of Cambridge. Available at
http://www.dspace.cam.ac.uk/handle/1810/242184. (Cited on page 170.)
Burgess, S. 2014. Sample size and power calculations in Mendelian randomiza-
tion with a single instrumental variable and a binary outcome. International
Journal of Epidemiology, 43(3):922–929. (Cited on page 131.)
Burgess, S., Butterworth, A., and Thompson, S. 2013. Mendelian randomiza-
tion analysis with multiple genetic variants using summarized data. Genetic
Epidemiology, 37(7):658–665. (Cited on pages 90 and 142.)
Burgess, S. and CCGC (CHD CRP Genetics Collaboration) 2013. Identifying
the odds ratio estimated by a two-stage instrumental variable analysis with
a logistic regression model. Statistics in Medicine, 32(27):4726–4747. (Cited
on page 59.)
Burgess, S., Daniel, R., Butterworth, A., Thompson, S., and EPIC-InterAct
Consortium 2014a. Network Mendelian randomization: extending instru-
mental variable techniques. International Journal of Epidemiology, avail-
able online before print. (Cited on page 179.)
Burgess, S., Davies, N., Thompson, S., and EPIC-InterAct Consortium 2014b.
Instrumental variable analysis with a non-linear exposure–outcome relation-
ship. Epidemiology, available online before print. (Cited on pages 47, 176,
and 177.)
Burgess, S., Granell, R., Palmer, T., Didelez, V., and Sterne, J. 2014c. Lack of
identification in semi-parametric instrumental variable models with binary
188 Mendelian Randomization

outcomes. American Journal of Epidemiology, 180(1):111–119. (Cited on

pages 66, 68, and 109.)
Burgess, S., Howson, J., Surendran, P., Thompson, S., and EPIC-CVD Con-
sortium 2014d. Mendelian randomization in the post-GWAS era: the perils
of causal inference without biological knowledge. Submitted for publication.
(Cited on page 182.)
Burgess, S., Scott, R., Timpson, N., Davey Smith, G., Thompson, S., and
EPIC-InterAct Consortium 2014e. Using published data in Mendelian ran-
domization: a blueprint for efficient identification of causal risk factors.
European Journal of Epidemiology, available online before print. (Cited
on pages 142 and 183.)
Burgess, S., Seaman, S., Lawlor, D., Casas, J., and Thompson, S. 2011a. Miss-
ing data methods in Mendelian randomization studies with multiple instru-
ments. American Journal of Epidemiology, 174(9):1069–1076. (Cited on
page 134.)
Burgess, S. and Thompson, S. 2011. Bias in causal estimates from Mendelian
randomization studies with weak instruments. Statistics in Medicine,
30(11):1312–1323. (Cited on pages 101, 108, 109, and 114.)
Burgess, S. and Thompson, S. 2012. Improvement of bias and coverage in
instrumental variable analysis with weak instruments for continuous and
binary outcomes. Statistics in Medicine, 31(15):1582–1600. (Cited on
pages 60, 62, 73, and 108.)
Burgess, S. and Thompson, S. 2013. Use of allele scores as instrumental vari-
ables for Mendelian randomization. International Journal of Epidemiology,
42(4):1134–1144. (Cited on pages 125 and 126.)
Burgess, S. and Thompson, S. 2014. Multivariable Mendelian randomization:
the use of pleiotropic genetic variants to estimate causal effects. American
Journal of Epidemiology, available online before print. (Cited on page 178.)
Burgess, S., Thompson, S., and CRP CHD Genetics Collaboration 2011b.
Avoiding bias from weak instruments in Mendelian randomization studies.
International Journal of Epidemiology, 40(3):755–764. (Cited on pages 68
and 116.)
Burgess, S., Thompson, S., and CRP CHD Genetics Collaboration 2012.
Methods for meta-analysis of individual participant data from Mendelian
randomization studies with binary outcomes. Statistical Methods in Medical
Research, available online before print. (Cited on pages 157 and 161.)
Cai, B., Small, D., and Ten Have, T. 2011. Two-stage instrumental variable
methods for estimating the causal odds ratio: Analysis of bias. Statistics in
Medicine, 30(15):1809–1824. (Cited on page 60.)
Bibliography 189

Casas, J., Bautista, L., Smeeth, L., Sharma, P., and Hingorani, A. 2005. Ho-
mocysteine and stroke: evidence on a causal link from Mendelian randomi-
sation. The Lancet, 365(9455):224–232. (Cited on page 11.)
CCGC (CRP CHD Genetics Collaboration) 2008. Collaborative pooled analy-
sis of data on C-reactive protein gene variants and coronary disease: judging
causality by Mendelian randomisation. European Journal of Epidemiology,
23(8):531–540. (Cited on pages 9, 117, 157, and 161.)
CCGC (CRP CHD Genetics Collaboration) 2011. Association between C reac-
tive protein and coronary heart disease: Mendelian randomisation analysis
based on individual participant data. British Medical Journal, 342:d548.
(Cited on pages 88, 116, 159, and 162.)
Chaussé, P. 2010. Computing generalized method of moments and generalized
empirical likelihood with R. Journal of Statistical Software, 34(11):1–35.
(Cited on page 72.)
Chen, L., Davey Smith, G., Harbord, R., and Lewis, S. 2008. Alcohol in-
take and blood pressure: a systematic review implementing a Mendelian
randomization approach. PLoS Medicine, 5(3):e52. (Cited on page 11.)
Cheung, B., Lauder, I., Lau, C., and Kumana, C. 2004. Meta-analysis of large
randomized controlled trials to evaluate the impact of statins on cardiovas-
cular outcomes. British Journal of Clinical Pharmacology, 57(5):640–651.
(Cited on page 91.)
Christenfeld, N., Sloan, R., Carroll, D., and Greenland, S. 2004. Risk factors,
confounding, and the illusion of statistical control. Psychosomatic Medicine,
66(6):868–875. (Cited on page 16.)
Clarke, P., Palmer, T., and Windmeijer, F. 2011. Estimating structural mean
models with multiple instrumental variables using the generalised method of
moments. The Centre for Market and Public Organisation 11/266, Centre
for Market and Public Organisation, University of Bristol, UK. (Cited on
page 72.)
Clarke, P. and Windmeijer, F. 2010. Instrumental variable estimators for bi-
nary outcomes. The Centre for Market and Public Organisation 10/239,
Centre for Market and Public Organisation, University of Bristol, UK.
(Cited on pages 38, 63, 65, and 66.)
Clarke, R., Peden, J., Hopewell, J., et al. 2009. Genetic variants associated
with Lp(a) lipoprotein level and coronary disease. New England Journal of
Medicine, 361(26):2518–2528. (Cited on pages 11 and 81.)
Cohen, J., Boerwinkle, E., Mosley Jr, T., and Hobbs, H. 2006. Sequence vari-
ations in PCSK9, low LDL, and protection against coronary heart disease.
New England Journal of Medicine, 354(12):1264–1272. (Cited on page 95.)
190 Mendelian Randomization

Collett, D. 2003. Modelling survival data in medical research. Chapman and

Hall/CRC Press. (Cited on page 175.)
Collins, R., Armitage, J., Parish, S., Sleight, P., Peto, R. (Heart Protection
Study Collaborative Group) 2002. MRC/BHF Heart Protection Study of
antioxidant vitamin supplementation in 20536 high-risk individuals: a ran-
domised placebo-controlled trial. The Lancet, 360(9326):23–33. (Cited on
page 4.)
Cox, D. 1958. Planning of experiments. Section 2: Some key assumptions.
Wiley. (Cited on page 41.)
Danesh, J. and Pepys, M. 2009. C-reactive protein and coronary disease: is
there a causal link? Circulation, 120(21):2036–2039. (Cited on page 7.)
Darwin, C. 1871. The descent of man and selection in relation to sex. Murray,
London. (Cited on page 5.)
Dastani, Z., Hivert, M.-F., Timpson, N., et al. 2012. Novel loci for adiponectin
levels and their influence on type 2 diabetes and metabolic traits: A multi-
ethnic meta-analysis of 45,891 individuals. PLoS Genetics, 8(3):e1002607.
(Cited on page 142.)
Davey Smith, G. 2006. Randomised by (your) god: robust inference from
an observational study design. Journal of Epidemiology and Community
Health, 60(5):382–388. (Cited on page 88.)

Davey Smith, G. 2011. Use of genetic markers and gene-diet interactions for
interrogating population-level causal influences of diet on health. Genes &
Nutrition, 6(1):27–43. (Cited on pages 17, 35, and 89.)
Davey Smith, G. and Ebrahim, S. 2003. ‘Mendelian randomization’: can ge-
netic epidemiology contribute to understanding environmental determinants
of disease? International Journal of Epidemiology, 32(1):1–22. (Cited on
pages 4, 5, and 20.)
Davey Smith, G. and Ebrahim, S. 2004. Mendelian randomization:
prospects, potentials, and limitations. International Journal of Epidemi-
ology, 33(1):30–42. (Cited on pages 23 and 123.)
Davey Smith, G., Lawlor, D., Harbord, R., Timpson, N., Day, I., and Ebrahim,
S. 2007. Clustered environments and randomized genes: a fundamental
distinction between conventional and genetic epidemiology. PLoS Medicine,
4(12):e352. (Cited on page 18.)
Davidson, R. and MacKinnon, J. 1993. Estimation and inference in economet-
rics. Chapter 18: Simultaneous equation models. Oxford University Press.
(Cited on page 61.)
Bibliography 191

Davidson, R. and MacKinnon, J. 2014. Confidence sets based on inverting

Anderson–Rubin tests. The Econometrics Journal, 17(2):S39–S58. (Cited
on page 54.)
Davies, N., von Hinke Kessler Scholder, S., Farbmacher, H., Burgess, S., Wind-
meijer, F., and Davey Smith, G. 2014. The many weak instrument problem
and Mendelian randomization. Statistics in Medicine, available online be-
fore print. (Cited on pages 61, 108, and 126.)
Davignon, J. and Laaksonen, R. 1999. Low-density lipoprotein-independent
effects of statins. Current Opinion in Lipidology, 10(6):543–559. (Cited on
page 91.)
Dawid, A. 2000. Causal inference without counterfactuals. Journal of the
American Statistical Association, 95(450):407–424. (Cited on page 26.)
Dawid, A. 2002. Influence diagrams for causal modelling and inference. In-
ternational Statistical Review, 70(2):161–189. (Cited on page 37.)
Debat, V. and David, P. 2001. Mapping phenotypes: canalization, plasticity
and developmental stability. Trends in Ecology & Evolution, 16(10):555–
561. (Cited on page 31.)
Dekkers, O., von Elm, E., Algra, A., Romijn, J., and Vandenbroucke, J. 2010.
How to assess the external validity of therapeutic trials: a conceptual ap-
proach. International Journal of Epidemiology, 39(1):89–94. (Cited on
page 93.)
Didelez, V., Meng, S., and Sheehan, N. 2010. Assumptions of IV methods
for observational epidemiology. Statistical Science, 25(1):22–40. (Cited on
pages 38, 48, and 51.)
Didelez, V. and Sheehan, N. 2007. Mendelian randomization as an instru-
mental variable approach to causal inference. Statistical Methods in Medical
Research, 16(4):309–330. (Cited on pages 34, 37, 42, 43, and 51.)
Ding, E., Song, Y., Manson, J., et al. 2009a. Sex hormone-binding globulin
and risk of type 2 diabetes in women and men. New England Journal of
Medicine, 361(12):1152–1163. (Cited on page 11.)
Ding, W., Lehrer, S., Rosenquist, J., and Audrain-McGovern, J. 2009b. The
impact of poor health on academic performance: new evidence using genetic
markers. Journal of Health Economics, 28(3):578–597. (Cited on page 11.)
Drukker, D. 2009. Generalized method of moments estimation in Stata 11.
Technical report, Stata Corp. (Cited on page 71.)
Ducharme, G. and LePage, Y. 1986. Testing collapsibility in contingency
tables. Journal of the Royal Statistical Society: Series B (Methodological),
48(2):197–205. (Cited on page 58.)
192 Mendelian Randomization

Dudbridge, F. 2013. Power and predictive accuracy of polygenic risk scores.

PLoS Genetics, 9(3):e1003348. (Cited on page 182.)
Dunn, G., Maracy, M., and Tomenson, B. 2005. Estimating treatment effects
from randomized clinical trials with noncompliance and loss to follow-up:
the role of instrumental variable methods. Statistical Methods in Medical
Research, 14(4):369–395. (Cited on page 65.)
Ebrahim, S. and Davey Smith, G. 2008. Mendelian randomization: can ge-
netic epidemiology help redress the failures of observational epidemiology?
Human Genetics, 123(1):15–33. (Cited on pages 11, 35, and 56.)
Efron, B. and Tibshirani, R. 1993. An introduction to the bootstrap. Chapman
& Hall/CRC Press. (Cited on page 54.)
Ehret, G., Munroe, P., Rice, K., et al. (The International Consortium for
Blood Pressure Genome-Wide Association Studies) 2011. Genetic variants
in novel pathways influence blood pressure and cardiovascular disease risk.
Nature, 478:103–109. (Cited on page 93.)
Elliott, P., Chambers, J., Zhang, W., et al. 2009. Genetic loci associated with
C-reactive protein levels and risk of coronary heart disease. Journal of the
American Medical Association, 302(1):37–48. (Cited on page 88.)
Emsley, R., Dunn, G., and White, I. 2010. Mediation and moderation of
treatment effects in randomised controlled trials of complex interventions.
Statistical Methods in Medical Research, 19(3):237–270. (Cited on page 180.)
Evans, D., Brion, M.-J., Paternoster, L., et al. 2013. Mining the human phe-
nome using allelic scores that index biological intermediates. PLoS Genetics,
9(10):e1003919. (Cited on page 182.)
Farnier, M. 2013. PCSK9 inhibitors. Current Opinion in Lipidology,
24(3):251–258. (Cited on page 95.)
Fieller, E. 1954. Some problems in interval estimation. Journal of the Royal
Statistical Society: Series B (Statistical Methodology), 16(2):175–185. (Cited
on page 52.)
Fischer-Lapp, K. and Goetghebeur, E. 1999. Practical properties of some
structural mean analyses of the effect of compliance in randomized trials.
Controlled Clinical Trials, 20(6):531–546. (Cited on page 65.)
Fisher, R. 1918. The correlation between relatives on the supposition of
Mendelian inheritance. Transactions of the Royal Society of Edinburgh,
52(2):399–433. (Cited on page 5.)
Foster, E. 1997. Instrumental variables for logistic regression: an illustration.
Social Science Research, 26(4):487–504. (Cited on pages 58 and 64.)
Bibliography 193

Fox, J. 2006. Teacher’s Corner: Structural Equation Modeling with the sem
Package in R. Structural Equation Modeling: A Multidisciplinary Journal,
13(3):465–486. (Cited on page 72.)
Freeman, G., Cowling, B., and Schooling, M. 2013. Power and sample size
calculations for Mendelian randomization studies. International Journal of
Epidemiology, 42(4):1157–1163. (Cited on pages 127 and 131.)
Fried, L., Borhani, N., Enright, P., et al. 1991. The Cardiovascular Health
Study: Design and rationale. Annals of Epidemiology, 1(3):263–276. (Cited
on page 164.)
Frost, C. and Thompson, S. 2000. Correcting for regression dilution bias:
comparison of methods for a single predictor variable. Journal of the Royal
Statistical Society: Series A (Statistics in Society), 163(2):173–189. (Cited
on page 19.)
Geiger, D., Verma, T., and Pearl, J. 1990. Identifying independence in
Bayesian networks. Networks, 20(5):507–534. (Cited on page 27.)
Gidding, S., Daniels, S., Kavey, R., and Expert Panel on Cardiovascular Health
and Risk Reduction in Youth 2012. Developing the 2011 integrated pediatric
guidelines for cardiovascular risk reduction. Pediatrics, 129(5):e1311–e1319.
(Cited on page 84.)
Glymour, M., Tchetgen Tchetgen, E., and Robins, J. 2012. Credible Mendelian
randomization studies: approaches for evaluating the instrumental variable
assumptions. American Journal of Epidemiology, 175(4):332–339. (Cited
on pages 34, 35, 68, and 95.)
Grassi, M., Assanelli, D., and Pezzini, A. 2007. Direct, reverse or reciprocal
causation in the relation between homocysteine and ischemic heart disease.
Thrombosis Research, 120(1):61–69. (Cited on page 179.)
Greenland, S. 1987. Interpretation and choice of effect measures in epidemio-
logic analyses. American Journal of Epidemiology, 125(5):761–768. (Cited
on page 42.)
Greenland, S. 2000a. An introduction to instrumental variables for epidemi-
ologists. International Journal of Epidemiology, 29(4):722–729. (Cited on
page 14.)
Greenland, S. 2000b. Causal analysis in the health sciences. Journal of the
American Statistical Association, 95(449):286–289. (Cited on page 26.)
Greenland, S., Lanes, S., and Jara, M. 2008. Estimating effects from random-
ized trials with discontinuations: the need for intent-to-treat design and
G-estimation. Clinical Trials, 5(1):5–13. (Cited on page 65.)
194 Mendelian Randomization

Greenland, S. and Robins, J. 1986. Identifiability, exchangeability, and

epidemiological confounding. International Journal of Epidemiology,
15(3):413–419. (Cited on pages 14, 28, and 30.)
Greenland, S., Robins, J., and Pearl, J. 1999. Confounding and collapsibility
in causal inference. Statistical Science, 14(1):29–46. (Cited on page 58.)
Hahn, J., Hausman, J., and Kuersteiner, G. 2004. Estimation with weak in-
struments: accuracy of higher-order bias and MSE approximations. Econo-
metrics Journal, 7(1):272–306. (Cited on page 61.)
Hall, A., Rudebusch, G., and Wilcox, D. 1996. Judging instrument rele-
vance in instrumental variables estimation. International Economic Review,
37(2):283–298. (Cited on pages 111 and 114.)
Hansen, L. 1982. Large sample properties of generalized method of moments
estimators. Econometrica: Journal of the Econometric Society, 50(4):1029–
1054. (Cited on page 64.)
Hansson, G. 2005. Inflammation, atherosclerosis, and coronary artery dis-
ease. The New England Journal of Medicine, 352(16):1685–1695. (Cited on
page 116.)
Hardin, J., Schmiediche, H., and Carroll, R. 2003. Instrumental variables,
bootstrapping, and generalized linear models. Stata Journal, 3(4):351–360.
(Cited on pages 71 and 72.)
Hayashi, F. 2000. Econometrics. Princeton University Press. (Cited on
page 61.)
Hennekens, C., Buring, J., Manson, J., et al. 1996. Lack of effect of long-
term supplementation with beta carotene on the incidence of malignant
neoplasms and cardiovascular disease. New England Journal of Medicine,
334(18):1145–1149. (Cited on page 4.)
Hernán, M. and Robins, J. 2006. Instruments for causal inference: an epi-
demiologist’s dream? Epidemiology, 17(4):360–372. (Cited on pages 17, 32,
and 39.)
Hernán, M. and Taubman, S. 2008. Does obesity shorten life? The importance
of well-defined interventions to answer causal questions. International Jour-
nal of Obesity, 32:S8–S14. (Cited on page 136.)
Higgins, J., Thompson, S., Deeks, J., and Altman, D. 2003. Measuring in-
consistency in meta-analyses. British Medical Journal, 327(7414):557–560.
(Cited on page 168.)
Hill, A. B. 1965. The environment and disease: association or causation?
Proceedings of the Royal Society of Medicine, 58(5):295–300. (Cited on
page 35.)
Bibliography 195

Hingorani, A. and Humphries, S. 2005. Nature’s randomised trials. The

Lancet, 366(9501):1906–1908. (Cited on page 69.)
Holland, P. 1986. Statistics and causal inference. Journal of the American
Statistical Association, 81(396):945–960. (Cited on pages 25 and 26.)
Hooper, L., Ness, A., and Davey Smith, G. 2001. Antioxidant strategy for car-
diovascular diseases. The Lancet, 357(9269):1705–1706. (Cited on page 4.)
Horowitz, J. L. 2011. Applied nonparametric instrumental variables estima-
tion. Econometrica, 79(2):347–394. (Cited on pages 42 and 176.)
Imbens, G. and Rosenbaum, P. 2005. Robust, accurate confidence intervals
with a weak instrument: quarter of birth and education. Journal of the
Royal Statistical Society: Series A (Statistics in Society), 168(1):109–126.
(Cited on pages 54 and 108.)
Inoue, A. and Solon, G. 2010. Two-sample instrumental variables estimators.
The Review of Economics and Statistics, 92(3):557–561. (Cited on pages 137
and 155.)
Irons, D., McGue, M., Iacono, W., and Oetting, W. 2007. Mendelian ran-
domization: A novel test of the gateway hypothesis and models of gene–
environment interplay. Development and Psychopathology, 19(4):1181–1195.
(Cited on page 11.)
Jansen, R. and Nap, J.-P. 2001. Genetical genomics: the added value from
segregation. Trends in Genetics, 17(7):388–391. (Cited on page 180.)
Johnson, T. 2011. Conditional and joint multiple-SNP analysis of GWAS
summary statistics identifies additional variants influencing complex traits.
Technical report, Queen Mary University of London. (Cited on page 142.)
Johnston, K., Gustafson, P., Levy, A., and Grootendorst, P. 2008. Use of
instrumental variables in the analysis of generalized linear models in the
presence of unmeasured confounding with applications to epidemiological
research. Statistics in Medicine, 27(9):1539–1556. (Cited on pages 64
and 65.)
Jones, E., Thompson, J., Didelez, V., and Sheehan, N. 2012. On the choice
of parameterisation and priors for the Bayesian analyses of Mendelian ran-
domisation studies. Statistics in Medicine, 31(14):1483–1501. (Cited on
page 62.)
Kamstrup, P., Tybjaerg-Hansen, A., Steffensen, R., and Nordestgaard, B.
2009. Genetically elevated lipoprotein(a) and increased risk of myocardial
infarction. Journal of the American Medical Association, 301(22):2331–
2339. (Cited on pages 11 and 80.)
196 Mendelian Randomization

Kaptoge, S., Di Angelantonio, E., Lowe, G. et al. (Emerging Risk Factors

Collaboration) 2010. C-reactive protein concentration and risk of coronary
heart disease, stroke, and mortality: an individual participant meta-analysis.
The Lancet, 375(9709):132–140. (Cited on pages 7 and 116.)
Keavney, B. 2011. C reactive protein and the risk of cardiovascular disease.
British Medical Journal, 342:d144. (Cited on page 171.)
Keavney, B., Danesh, J., Parish, S., et al. 2006. Fibrinogen and coronary
heart disease: test of causality by ‘Mendelian randomization’. International
Journal of Epidemiology, 35(4):935–943. (Cited on page 75.)
Khaw, K., Bingham, S., Welch, A., et al. 2001. Relation between plasma
ascorbic acid and mortality in men and women in EPIC-Norfolk prospec-
tive study: a prospective population study. The Lancet, 357(9257):657–663.
(Cited on page 4.)
Kinal, T. 1980. The existence of moments of k-class estimators. Econometrica,
48(1):241–249. (Cited on page 57.)
Kivimäki, M., Lawlor, D., Davey Smith, G., et al. 2008. Does high C-reactive
protein concentration increase atherosclerosis? The Whitehall II Study.
PLoS ONE, 3(8):e3013. (Cited on page 11.)
Kivimäki, M., Lawlor, D., Eklund, C., et al. 2007. Mendelian randomiza-
tion suggests no causal association between C-reactive protein and carotid
intima-media thickness in the young Finns study. Arteriosclerosis, Throm-
bosis, and Vascular Biology, 27(4):978–979. (Cited on page 11.)
Kleiber, C. and Zeileis, A. 2014. AER: Applied Econometrics with R. R
package version 1.2-2. (Cited on page 72.)
Kleibergen, F. and Zivot, E. 2003. Bayesian and classical approaches to instru-
mental variable regression. Journal of Econometrics, 114(1):29–72. (Cited
on page 62.)
Klein, R., Zeiss, C., Chew, E., et al. 2005. Complement factor H polymorphism
in age-related macular degeneration. Science, 308(5720):385–389. (Cited on
page 181.)
Law, M., Morris, J., and Wald, N. 2009. Use of blood pressure lowering drugs
in the prevention of cardiovascular disease: meta-analysis of 147 randomised
trials in the context of expectations from prospective epidemiological stud-
ies. British Medical Journal, 338:b1665. (Cited on page 93.)
Law, M., Wald, N., and Rudnicka, A. 2003. Quantifying effect of statins on
low density lipoprotein cholesterol, ischaemic heart disease, and stroke: sys-
tematic review and meta-analysis. British Medical Journal, 326(7404):1423.
(Cited on page 91.)
Bibliography 197

Lawlor, D., Harbord, R., Sterne, J., Timpson, N., and Davey Smith, G. 2008.
Mendelian randomization: using genes as instruments for making causal
inferences in epidemiology. Statistics in Medicine, 27(8):1133–1163. (Cited
on pages 20, 21, 52, and 67.)
Le Marchand, L., Derby, K., Murphy, S., et al. 2008. Smokers with the
CHRNA lung cancer–associated variants are exposed to higher levels of
nicotine equivalents and a carcinogenic tobacco-specific nitrosamine. Can-
cer Research, 68(22):9137–9140. (Cited on page 179.)
Lewis, S. and Davey Smith, G. 2005. Alcohol, ALDH2, and esophageal can-
cer: a meta-analysis which illustrates the potentials and limitations of a
Mendelian randomization approach. Cancer Epidemiology Biomarkers &
Prevention, 14(8):1967–1971. (Cited on page 23.)
Little, R. and Rubin, D. 2002. Statistical analysis with missing data (2nd
edition). Wiley. (Cited on page 134.)

Lunn, D., Whittaker, J., and Best, N. 2006. A Bayesian toolkit for genetic asso-
ciation studies. Genetic Epidemiology, 30(3):231–247. (Cited on page 135.)
Maldonado, G. and Greenland, S. 2002. Estimating causal effects. Interna-
tional Journal of Epidemiology, 31(2):422–429. (Cited on page 26.)
Manolio, T. 2010. Genomewide association studies and assessment of the risk
of disease. New England Journal of Medicine, 363(2):166–176. (Cited on
page 181.)
Martens, E., Pestman, W., de Boer, A., Belitser, S., and Klungel, O. 2006. In-
strumental variables: application and limitations. Epidemiology, 17(3):260–
267. (Cited on pages 16 and 93.)
McPherson, J., et al. (The International Human Genome Mapping Consor-
tium) 2001. A physical map of the human genome. Nature, 409(6822):934–
941. (Cited on page 5.)
Mendel, G. 1866. Versuche über Pflanzen-hybriden. Verhandlungen des natur-
forschenden Vereines in Brünn [Proceedings of the Natural History Society
of Brünn], 4:3–47. (Cited on page 5.)
Mikusheva, A. 2010. Robust confidence sets in the presence of weak in-
struments. Journal of Econometrics, 157(2):236–247. (Cited on pages 54
and 108.)
Mikusheva, A. and Poi, B. 2006. Tests and confidence sets with correct size
when instruments are potentially weak. Stata Journal, 6(3):335–347. (Cited
on pages 54 and 108.)
198 Mendelian Randomization

Minelli, C., Thompson, J., Tobin, M., and Abrams, K. 2004. An integrated
approach to the meta-analysis of genetic association studies using Mendelian
randomization. American Journal of Epidemiology, 160(5):445–452. (Cited
on pages 51, 52, and 140.)
Mogstad, M. and Wiswall, M. 2010. Linearity in instrumental variables esti-
mation: Problems and solutions. Technical report, Forschungsinstitut zur
Zukunft der Arbeit. Bonn, Germany. (Cited on pages 42 and 176.)
Moreira, M. 2003. A conditional likelihood ratio test for structural models.
Econometrica, 71(4):1027–1048. (Cited on pages 54 and 108.)
Moreira, M., Porter, J., and Suarez, G. 2009. Bootstrap validity for the score
test when instruments may be weak. Journal of Econometrics, 149(1):52–
64. (Cited on page 54.)
Morris, A., Voight, B., Teslovich, T., et al. 2012. Large-scale association
analysis provides insights into the genetic architecture and pathophysiology
of type 2 diabetes. Nature Genetics, 44(9):981–990. (Cited on page 146.)
Mumby, H., Elks, C., Li, S., et al. 2011. Mendelian randomisation study of
childhood BMI and early menarche. Journal of Obesity, available online
before print. (Cited on page 11.)
Nagelkerke, N., Fidler, V., Bernsen, R., and Borgdorff, M. 2000. Estimat-
ing treatment effects in randomized clinical trials in the presence of non-
compliance. Statistics in Medicine, 19(14):1849–1864. (Cited on page 60.)
Nelson, C. and Startz, R. 1990. The distribution of the instrumental variables
estimator and its t-ratio when the instrument is a poor one. Journal of
Business, 63(1):125–140. (Cited on pages 99 and 126.)
Nitsch, D., Molokhia, M., Smeeth, L., DeStavola, B., Whittaker, J., and Leon,
D. 2006. Limits to causal inference based on Mendelian randomization: a
comparison with randomized controlled trials. American Journal of Epi-
demiology, 163(5):397–403. (Cited on page 16.)
Norton, E. and Han, E. 2008. Genetic information, obesity, and labor market
outcomes. Health Economics, 17(9):1089–1104. (Cited on page 11.)
Ogbuanu, I., Zhang, H., and Karmaus, W. 2009. Can we apply the Mendelian
randomization methodology without considering epigenetic effects? Emerg-
ing Themes in Epidemiology, 6(1):3. (Cited on page 32.)
Palmer, T., Lawlor, D., Harbord, R., et al. 2011a. Using multiple genetic
variants as instrumental variables for modifiable risk factors. Statistical
Methods in Medical Research, 21(3):223–242. (Cited on pages 114 and 123.)
Bibliography 199

Palmer, T., Sterne, J., Harbord, R., et al. 2011b. Instrumental variable estima-
tion of causal risk ratios and causal odds ratios in Mendelian randomization
analyses. American Journal of Epidemiology, 173(12):1392–1403. (Cited on
pages 64 and 71.)
Palmer, T., Thompson, J., Tobin, M., Sheehan, N., and Burton, P. 2008.
Adjusting for bias and unmeasured confounding in Mendelian randomiza-
tion studies with binary responses. International Journal of Epidemiology,
37(5):1161–1168. (Cited on pages 37 and 60.)
Pauling, L., Itano, H., Singer, S., and Wells, I. 1949. Sickle cell anemia, a
molecular disease. Science, 110(2865):543–548. (Cited on page 5.)
Pearl, J. 2000a. Causality: models, reasoning, and inference. Cambridge Uni-
versity Press. (Cited on page 25.)
Pearl, J. 2000b. Causality: models, reasoning, and inference. Chapter 3, Sec-
tion 3.1: The back-door criterion. Cambridge University Press. (Cited on
page 28.)
Pearl, J. 2001. Direct and indirect effects. In Proceedings of the Seventeenth
Conference on Uncertainty in Artificial Intelligence, pages 411–420. (Cited
on page 180.)
Pearl, J. 2010. An introduction to causal inference. The International Journal
of Biostatistics, 6(2):1–60. (Cited on page 26.)

Peto, R., Doll, R., Buckley, J., and Sporn, M. 1981. Can dietary beta-carotene
materially reduce human cancer rates? Nature, 290:201–208. (Cited on
page 4.)
Pierce, B., Ahsan, H., and VanderWeele, T. 2011. Power and instrument
strength requirements for Mendelian randomization studies using multi-
ple genetic variants. International Journal of Epidemiology, 40(3):740–752.
(Cited on pages 114 and 131.)
Pierce, B. and Burgess, S. 2013. Efficient design for Mendelian randomiza-
tion studies: subsample and two-sample instrumental variable estimators.
American Journal of Epidemiology, 178(7):1177–1184. (Cited on page 137.)
Pierce, B. and VanderWeele, T. 2012. The effect of non-differential mea-
surement error on bias, precision and power in Mendelian randomization
studies. International Journal of Epidemiology, 41(5):1383–1393. (Cited on
page 19.)
Plenge, R., Scolnick, E., and Altshuler, D. 2013. Validating therapeutic targets
through human genetics. Nature Reviews Drug Discovery, 12(8):581–594.
(Cited on page 94.)
200 Mendelian Randomization

R Development Core Team 2011. R: A language and environment for statis-

tical computing. R Foundation for Statistical Computing, Vienna, Austria.
(Cited on page 69.)
Raal, F., Giugliano, R., Sabatine, M., et al. 2014. Reduction in lipoprotein(a)
with PCSK9 monoclonal antibody evolocumab (AMG 145): a pooled analy-
sis of more than 1,300 patients in 4 phase II trials. Journal of the American
College of Cardiology, 63(13):1278–1288. (Cited on page 95.)
Rasbash, J., Steele, F., Browne, W., and Goldstein, H. 2009. A user’s guide
to MLwiN, v2. 10. Centre for Multilevel Modelling, University of Bristol.
(Cited on page 141.)
Relton, C. and Davey Smith, G. 2010. Epigenetic epidemiology of common
complex disease: Prospects for prediction, prevention, and treatment. PLoS
Medicine, 7(10):e1000356. (Cited on page 180.)
Relton, C. and Davey Smith, G. 2012a. Is epidemiology ready for epigenetics?
International Journal of Epidemiology, 41(1):5–9. (Cited on page 180.)
Relton, C. and Davey Smith, G. 2012b. Two-step epigenetic Mendelian ran-
domization: a strategy for establishing the causal role of epigenetic processes
in pathways to disease. International Journal of Epidemiology, 41(1):161–
176. (Cited on page 181.)
Riley, R., Abrams, K., Sutton, A., Lambert, P., and Thompson, J. 2007.
Bivariate random-effects meta-analysis and the estimation of between-
study correlation. BMC Medical Research Methodology, 7(1):3. (Cited on
page 141.)
Roberts, L., Davenport, R., Pennisi, E., and Marshall, E. 2001. A history of
the Human Genome Project. Science, 291(5507):1195. (Cited on page 5.)
Robins, J. 1986. A new approach to causal inference in mortality studies with
a sustained exposure period–application to control of the healthy worker
survivor effect. Mathematical Modelling, 7(9-12):1393–1512. (Cited on
page 65.)
Robins, J. 1992. Semiparametric estimation of an accelerated failure time
model with time-dependent covariates. Biometrika, 79(2):311–334. (Cited
on page 176.)
Robins, J. 1994. Correcting for non-compliance in randomized trials using
structural nested mean models. Communications in Statistics – Theory
and Methods, 23(8):2379–2412. (Cited on page 65.)
Robins, J. 1999. Statistical models in epidemiology: the environment and clini-
cal trials, chapter Marginal structural models versus structural nested mod-
els as tools for causal inference, pages 95–134. Springer. (Cited on page 66.)
Bibliography 201

Robinson, J., Nedergaard, B., Rogers, W., et al. 2014. Effect of evolocumab
or ezetimibe added to moderate-or high-intensity statin therapy on LDL-
C lowering in patients with hypercholesterolemia: the LAPLACE-2 ran-
domized clinical trial. Journal of the American Medical Association,
311(18):1870–1882. (Cited on page 95.)
Rossouw, J., et al. (Writing Group for the Women’s Health Initiative Investi-
gators) 2002. Risks and benefits of estrogen plus progestin in healthy post-
menopausal women: principal results from the Women’s Health Initiative
randomized controlled trial. Journal of the American Medical Association,
288(3):321–333. (Cited on page 4.)
Rothwell, P. 2010. Commentary: External validity of results of randomized
trials: disentangling a complex concept. International Journal of Epidemi-
ology, 39(1):94–96. (Cited on page 87.)
Rubin, D. 1974. Estimating causal effects of treatments in randomized and
nonrandomized studies. Journal of Educational Psychology, 66(5):688–701.
(Cited on page 30.)
Sargan, J. 1958. The estimation of economic relationships using instrumental
variables. Econometrica, 26(3):393–415. (Cited on page 68.)
Sarwar, N., Butterworth, A., Freitag, D., et al. (IL6R Genetics Consortium
and Emerging Risk Factors Collaboration) 2012. Interleukin-6 receptor
pathways in coronary heart disease: a collaborative meta-analysis of 82 stud-
ies. Lancet, 379(9822):1205–1213. (Cited on page 20.)
SAS (SAS Institute and SAS Publishing Staff) 2004. SAS/STAT 9.1 User’s
Guide. SAS Institute Inc, Cary, NC. (Cited on page 69.)
Schatzkin, A., Abnet, C., Cross, A., et al. 2009. Mendelian randomization:
how it can – and cannot – help confirm causal relations between nutrition
and cancer. Cancer Prevention Research, 2(2):104–113. (Cited on pages 24
and 89.)
Scheet, P. and Stephens, M. 2006. A fast and flexible statistical model
for large-scale population genotype data: applications to inferring missing
genotypes and haplotypic phase. American Journal of Human Genetics,
78(4):629–644. (Cited on page 134.)
Schunkert, H., König, I., Kathiresan, S., et al. 2011. Large-scale association
analysis identifies 13 new susceptibility loci for coronary artery disease.
Nature Genetics, 43(4):333–338. (Cited on page 146.)
Sheehan, N., Didelez, V., Burton, P., and Tobin, M. 2008. Mendelian randomi-
sation and causal inference in observational epidemiology. PLoS Medicine,
5(8):e177. (Cited on page 20.)
202 Mendelian Randomization

Shendure, J. and Ji, H. 2008. Next-generation DNA sequencing. Nature

Biotechnology, 26(10):1135–1145. (Cited on page 5.)
Small, D. 2014. ivpack: Instrumental Variable Estimation. R package version
1.1. (Cited on pages 54 and 72.)
Small, D. and Rosenbaum, P. 2008. War and wages: the strength of instru-
mental variables and their sensitivity to unobvserved biases. Journal of the
American Statistical Association, 103(483):924–933. (Cited on page 121.)
Speliotes, E., Willer, C., Berndt, S., et al. 2010. Association analyses of 249,796
individuals reveal 18 new loci associated with body mass index. Nature
Genetics, 42(11):937–948. (Cited on page 176.)
Spiegelhalter, D., Best, N., Carlin, B., and Linde, A. 2002. Bayesian measures
of model complexity and fit. Journal of the Royal Statistical Society: Series
B (Statistical Methodology), 64(4):583–639. (Cited on page 152.)
Spiegelhalter, D., Thomas, A., Best, N., and Lunn, D. 2003. WinBUGS version
1.4 user manual. Technical report, MRC Biostatistics Unit, Cambridge, UK.
(Cited on pages 62, 69, and 141.)
Spirtes, P., Glymour, C., and Scheines, R. 2000. Causation, prediction, and
search. MIT Press. (Cited on page 39.)
Staiger, D. and Stock, J. 1997. Instrumental variables regression with weak
instruments. Econometrica, 65(3):557–586. (Cited on pages 67 and 106.)
StataCorp 2009. Stata Statistical Software: Release 11. College Station, TX.
(Cited on page 69.)
Sterne, J. and Davey Smith, G. 2001. Sifting the evidence – What’s wrong
with significance tests? British Medical Journal, 322:226–231. (Cited on
page 96.)
Stock, J., Wright, J., and Yogo, M. 2002. A survey of weak instruments and
weak identification in generalized method of moments. Journal of Business
and Economic Statistics, 20(4):518–529. (Cited on page 67.)
Stock, J. and Yogo, M. 2002. Testing for weak instruments in linear IV re-
gression. SSRN eLibrary, 11:T0284. (Cited on pages 70, 108, and 118.)
Stukel, T., Fisher, E., Wennberg, D., Alter, D., Gottlieb, D., and Vermeulen,
M. 2007. Analysis of observational studies in the presence of treatment
selection bias. Journal of the American Medical Association, 297(3):278–
285. (Cited on page 59.)
Sussman, J. and Hayward, R. 2010. An IV for the RCT: using instrumental
variables to adjust for treatment contamination in randomised controlled
trials. British Medical Journal, 340:c2073. (Cited on page 14.)
Bibliography 203

Sutton, A., Kendrick, D., and Coupland, C. 2008. Meta-analysis of individual-

and aggregate-level data. Statistics in Medicine, 27(5):651–669. (Cited on
page 150.)
Swanson, S. and Hernán, M. 2013. Commentary: how to report instrumen-
tal variable analyses (suggestions welcome). Epidemiology, 24(3):370–374.
(Cited on page 42.)
Swerdlow, D., Holmes, M., Kuchenbaecker, K., et al. (The Interleukin-6 Recep-
tor Mendelian Randomisation Analysis Consortium) 2012. The interleukin-
6 receptor as a target for prevention of coronary heart disease: a Mendelian
randomisation analysis. Lancet, 379(9822):1214–1224. (Cited on pages 19,
36, and 183.)
Taubes, G. and Mann, C. 1995. Epidemiology faces its limits. Science,
269(5221):164–169. (Cited on page 4.)
Taylor, A., Davies, N., Ware, J., VanderWeele, T., Davey Smith, G., and Mu-
nafò, M. 2014. Mendelian randomization in health research: Using appro-
priate genetic variants and avoiding biased estimates. Economics & Human
Biology, 13:99–106. (Cited on page 125.)
Taylor, F., Ward, K., Moore, T., et al. 2013. Statins for the primary prevention
of cardiovascular disease. Cochrane Database of Systematic Reviews, 2013:1.
(Cited on page 91.)
Terza, J., Basu, A., and Rathouz, P. 2008. Two-stage residual inclusion es-
timation: addressing endogeneity in health econometric modeling. Journal
of Health Economics, 27(3):531–543. (Cited on page 60.)
Thomas, D. and Conti, D. 2004. Commentary: the concept of ‘Mendelian Ran-
domization’. International Journal of Epidemiology, 33(1):21–25. (Cited on
page 14.)
Thomas, D., Lawlor, D., and Thompson, J. 2007. Re: Estimation of bias
in nongenetic observational studies using “Mendelian triangulation” by
Bautista et al. Annals of Epidemiology, 17(7):511–513. (Cited on pages 52
and 69.)
Thompson, J., Minelli, C., Abrams, K., Tobin, M., and Riley, R. 2005. Meta-
analysis of genetic studies using Mendelian randomization – a multivariate
approach. Statistics in Medicine, 24(14):2241–2254. (Cited on pages 11, 90,
116, 141, 144, 145, and 154.)

Timpson, N., Harbord, R., Davey Smith, G., Zacho, J., Tybjærg-Hansen, A.,
and Nordestgaard, B. 2009. Does greater adiposity increase blood pressure
and hypertension risk? Mendelian randomization using the FTO /MC4R
genotype. Hypertension, 54(1):84–90. (Cited on page 77.)
204 Mendelian Randomization

Timpson, N., Lawlor, D., Harbord, R., et al. 2005. C-reactive protein and its
role in metabolic syndrome: mendelian randomisation study. The Lancet,
366(9501):1954–1959. (Cited on page 11.)
Timpson, N., Nordestgaard, B., Harbord, R., et al. 2011. C-reactive protein
levels and body mass index: elucidating direction of causation through recip-
rocal Mendelian randomization. International Journal of Obesity, 35:300–
308. (Cited on page 179.)
Tobin, M., Minelli, C., Burton, P., and Thompson, J. 2004. Commentary: De-
velopment of Mendelian randomization: from hypothesis test to ‘Mendelian
deconfounding’. International Journal of Epidemiology, 33(1):26–29. (Cited
on pages 16 and 39.)
Trompet, S., Jukema, J., Katan, M., et al. 2009. Apolipoprotein E genotype,
plasma cholesterol, and cancer: a Mendelian randomization study. American
Journal of Epidemiology, 170(11):1415–1421. (Cited on page 11.)

VanderWeele, T. 2009. Concerning the consistency assumption in causal in-

ference. Epidemiology, 20(6):880–883. (Cited on page 50.)
VanderWeele, T., Asomaning, K., Tchetgen Tchetgen, E., et al. 2012. Genetic
variants on 15q25.1, smoking, and lung cancer: an assessment of mediation
and interaction. American Journal of Epidemiology, 175(10):1013–1020.
(Cited on page 179.)
VanderWeele, T., Tchetgen Tchetgen, E., Cornelis, M., and Kraft, P. 2014.
Methodological challenges in Mendelian randomization. Epidemiology,
25(3):427–435. (Cited on page 96.)
VanderWeele, T. and Vansteelandt, S. 2009. Conceptual issues concern-
ing mediation, interventions and composition. Statistics and its Interface,
2(4):457–468. (Cited on page 180.)
Vansteelandt, S., Bowden, J., Babanezhad, M., and Goetghebeur, E. 2011. On
instrumental variables estimation of causal odds ratios. Statistical Science,
26(3):403–422. (Cited on pages 59 and 66.)
Vansteelandt, S. and Goetghebeur, E. 2003. Causal inference with generalized
structural mean models. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 65(4):817–835. (Cited on page 66.)
Voight, B., Peloso, G., Orho-Melander, M., et al. 2012. Plasma HDL choles-
terol and risk of myocardial infarction: a mendelian randomisation study.
The Lancet, 380(9841):572–580. (Cited on pages 11, 82, and 177.)
von Hinke Kessler Scholder, S., Davey Smith, G., Lawlor, D., Propper, C.,
and Windmeijer, F. 2010. Genetic markers as instrumental variables: An
application to child fat mass and academic achievement. The Centre for
Bibliography 205

Market and Public Organisation 10/229, University of Bristol, UK. (Cited

on page 11.)
Wald, A. 1940. The fitting of straight lines if both variables are subject to
error. Annals of Mathematical Statistics, 11(3):284–300. (Cited on page 45.)
Wan, E., Qiu, W., Baccarelli, A., et al. 2012. Cigarette smoking behaviors
and time since quitting are associated with differential DNA methylation
across the human genome. Human Molecular Genetics, 21(13):3073–3082.
(Cited on page 180.)
Wardle, J., Carnell, S., Haworth, C., Farooqi, I., O’Rahilly, S., and Plomin, R.
2008. Obesity associated genetic variation in FTO is associated with dimin-
ished satiety. Journal of Clinical Endocrinology & Metabolism, 93(9):3640–
3643. (Cited on pages 31, 78, and 89.)
Waterworth, D., Ricketts, S., Song, K., et al. 2010. Genetic variants influencing
circulating lipid levels and risk of coronary artery disease. Arteriosclerosis,
Thrombosis, and Vascular Biology, 30(11):2264–2276. (Cited on page 90.)
Watson, J. and Crick, F. 1953. Molecular structure of nucleic acids: a structure
for deoxyribose nucleic acid. Nature, 171(4356):737–738. (Cited on page 5.)
Wehby, G., Ohsfeldt, R., and Murray, J. 2008. “Mendelian randomization”
equals instrumental variable analysis with genetic instruments. Statistics
in Medicine, 27(15):2745–2749. (Cited on pages 14 and 68.)

Welsh, P., Polisecki, E., Robertson, M., et al. 2010. Unraveling the direc-
tional link between adiposity and inflammation: a bidirectional Mendelian
randomization approach. Journal of Clinical Endocrinology & Metabolism,
95(1):93–99. (Cited on page 179.)
Wooldridge, J. 2009. Introductory econometrics: A modern approach. Chap-
ter 5: Multiple regression analysis – OLS asymptotics. South-Western,
Nashville, TN. (Cited on pages 56, 67, and 127.)
Wright, P. 1928. The tariff on animal and vegetable oils. Appendix B. Macmil-
lan, New York. (Cited on page 10.)
Yang, J., Lee, S., Goddard, M., and Visscher, P. 2011. GCTA: a tool for
genome-wide complex trait analysis. The American Journal of Human Ge-
netics, 88(1):76–82. (Cited on page 182.)
Youngman, L., Keavney, B., Palmer, A., et al. 2000. Plasma fibrinogen and
fibrinogen genotypes in 4685 cases of myocardial infarction and in 6002 con-
trols: test of causality by ‘Mendelian randomization’. Circulation, 102(Suppl
II):31–32. (Cited on page 9.)
206 Mendelian Randomization

Zacho, J., Tybjaerg-Hansen, A., Jensen, J., Grande, P., Sillesen, H., and
Nordestgaard, B. 2008. Genetically elevated C-reactive protein and ischemic
vascular disease. New England Journal of Medicine, 359(18):1897–1908.
(Cited on page 100.)
Zohoori, N. and Savitz, D. 1997. Econometric approaches to epidemiologic
data: Relating endogeneity and unobserved heterogeneity to confounding.
Annals of Epidemiology, 7(4):251–257. (Cited on pages 56 and 110.)
 Statistics

Chap ma n & Ha ll/ C RC Chapm an & Hall/C R C

In terdi sc i p l i n a r y St a t is t ics Series I nterdis c iplinar y S tatis tic s S eries

MENDELIAN RANDOMIZATION
MENDELIAN
Mendelian Randomization: Methods for Using Genetic Variants in Causal Es-
timation provides thorough coverage of the methods and practical elements of
Mendelian randomization analysis. It brings together diverse aspects of Mendelian
randomization spanning epidemiology, statistics, genetics, and econometrics.

RANDOMIZATION
Through several examples, the first part of the book shows how to perform simple
applied Mendelian randomization analyses and interpret their results. The second
part addresses specific methodological issues, such as weak instruments, multiple
instruments, power calculations, and meta-analysis, relevant to practical applica-
tions of Mendelian randomization. In this part, the authors draw on data from the
C-reactive protein Coronary heart disease Genetics Collaboration (CCGC) to illus-
trate the analyses. They present the mathematics in an easy-to-understand way by
using nontechnical language and reinforcing key points at the end of each chapter.
Methods for Using
The last part of the book examines the potential of Mendelian randomization in the
future, exploring both methodological and applied developments. Genetic Variants
in Causal Estimation
Features
• Offers first-hand, in-depth guidance on Mendelian randomization from
leaders in the field
• Makes the diverse aspects of Mendelian randomization understandable to
newcomers
• Illustrates the technical details using data from a large collaborative study
• Includes other real-world examples that show how Mendelian randomization
is used in studies involving inflammation, heart disease, and more
• Discusses possible future directions for research involving Mendelian
randomization

This book gives you the foundation to understand issues concerning the use of Burgess • Thompson Stephen Burgess
genetic variants as instrumental variables. It will get you up to speed in undertak-
ing and interpreting Mendelian randomization analyses. Chapter summaries, paper
summaries, web-based applications, and software code for implementing the sta- Simon G. Thompson
tistical techniques are available on a supplementary website.

K16638

w w w. c rc p r e s s . c o m

K16638_cover.indd 1 12/16/14 8:27 AM