Methods in

Statistical Genomics
In the Context of Genome-Wide Association Studies

Edited by Philip Chester Cooley

 Methods in Statistical Genomics:
In the Context of Genome-Wide
Association Studies
Edited by
Philip Chester Cooley

RTI Press
2016 RTI International. All rights reserved. This The RTI Press mission is to disseminate
book is protected by copyright. Credit must be information about RTI research, analytic tools,
provided to the author and source of the book and technical expertise to a national and
when the content is quoted. No part of this book international audience. RTI Press publications
may be reproduced in any form by any electronic are peer-reviewed by at least two independent
or mechanical means (including photocopying, substantive experts and one or more Press
recording, or information storage and retrieval) editors.
without permission in writing from the publisher. RTI International is an independent, nonprofit
RTI International is a registered trademark and a research organization dedicated to improving
trade name of Research Triangle Institute. the human condition by turning knowledge
into practice. RTI offers innovative research
and technical services to governments and
businesses worldwide in the areas of health and
pharmaceuticals, education and training, surveys
and statistics, advanced technology, international
development, economic and social policy, energy
and the environment, and laboratory testing and
chemistry services.

Library of Congress Control Number: 2016949391

ISBN 978-1-934831-16-8
(refers to print version)
RTI Press publication No. BK-0016-1608
http://dx.doi.org/10.3768/rtipress.2016.bk.0016.1608
www.rti.org/rtipress
Cover design: John Theilgard

This publication is part of the RTI Press Book series.

RTI International
3040 East Cornwallis Road, PO Box 12194
Research Triangle Park, NC 27709-2194, USA
[email protected]
www.rti.org
 Contents
Chapter 1. Overview of Chapters 1
Philip Chester Cooley

Chapter 2. Genome-Wide Association Data:

Where Are the Standards? 17
Philip Chester Cooley

Chapter 3. Creating the Synthetic Gene Data 31

Philip Chester Cooley

Chapter 4. Genetic Inheritance and Genome-Wide Association

Statistical Test Performance Using Simulated Data 37
Philip Chester Cooley, Robert F. Clark, Ralph E. Folsom,
and Grier Page

Chapter 5. The Influence of Errors Inherent in Genome-Wide

Association Studies (GWAS) in Relation to Single-Gene Models 49
Philip Chester Cooley, Robert F. Clark, and Grier Page

Chapter 6. Conducting Genome-Wide Association Studies (GWAS):

Epistasis Scenarios 65
Philip Chester Cooley, Nathan Gaddis, Ralph E. Folsom,
and Diane Wagener

Chapter 7. Assessing Gene-Environment Interactions in

Genome-Wide Association Studies (GWAS): Statistical
Approaches 85
Philip Chester Cooley, Robert F. Clark, and Ralph E. Folsom

Chapter 8. Polygene Methods in Genome-Wide Association

Studies (GWAS) 117
Philip Chester Cooley and Ralph E. Folsom

Chapter 9. Conclusions and Recommendations 143

Philip Chester Cooley

Acknowledgment 149

Contributors 151

Index 153
 CHAPTER 1

Overview of Chapters
Philip Chester Cooley

Introduction
The objective of this book is to describe procedures for analyzing genome-wide
association studies (GWAS). Some of the material is unpublished and contains
commentary and unpublished research; other material (Chapters 4 through 7)
has been published previously. Each previously published chapter investigates
a different genomics model, but all focus on identifying the strengths and
limitations of various statistical procedures that have been applied to different
GWAS scenarios.
The distinction between genotype and phenotype was initially presented
by the Danish botanist, plant physiologist, and geneticist Wilhelm Johannsen
in a book he published in 1905, The Elements of Heredity. He distinguished
between the genotype of the organism (it is hereditary) and the ways in which
its heredity is demonstrated in phenotypes, or physical characteristics. This
distinction was an outgrowth of Johannsens experiments concerning heritable
variation in plants.1
Today, it is understood that the process leading from genes to proteins that
ultimately establish phenotypes is complex. Most proteins are the products
of multiple genes. Whether a protein is an enzyme, receptor, or hormone,
it functions in a specific environment that includes external factors like
temperature, rainfall, the amount of sunlight available, and nutrition, as well as
internal factors that can include other hormones, enzymes, and other proteins.
Further, biochemical pathways are not always linear; they can have multiple
positive and negative feedback loops and may involve multiple steps and the
products of hundreds of genes. In summary, the evolutionary forces producing
a phenotype may often involve many genes and can be influenced by a variety
of specific environmental factors.
2 Chapter 1

The Human Genome Project

The Human Genome Project (HGP) was an international scientific project
with the goals of determining the sequence of chemical base pairs that make
up human DNA and of identifying all of the physical and functional genes
of the human genome. The HGP produced the first complete sequences of
individual human genomes. As of 2012, thousands of human genomes had
been completely sequenced, and many more had been mapped at lower levels
of resolution. The resulting data have been used worldwide in biomedical
science, anthropology, forensics, and other branches of science. With the
mapping of the human genome near completion, researchers expected that
subsequent genomic studies would lead to advances in our understanding of
human evolution, and advances in many subfields of biology, particularly the
diagnosis and treatment of many diseases.
To that end, researchers have worked to identify genes that constitute
biomarkers using a combination of high-throughput experimental and
bioinformatics approaches; nevertheless, the identification of biological
functions of the protein and RNA products of DNA has only just begun.
Recent results suggest that most of the vast quantities of noncoding DNA
within the genome have biochemical activities that include regulating gene
expression, organizing chromosome architecture, and producing signals that
control epigenetic inheritance.2
A major aim of the HGP was to determine the functions of genes.
Researchers believed that once the complete genome sequence was developed,
interpreting the sequence by comparing the intermediate messenger RNA and
protein products would be straightforward and ultimately would identify the
genetic factors that influence important phenotypes such as predisposition to
certain diseases. The simple rationale behind GWAS is that if certain genetic
variations are more frequent in persons with a given disease, the variations are
said to be associated with the disease. The associated genetic variations serve
as pointers to regions of the human genome that may be involved in causing
the disease.

Genome-Wide Association Studies

GWAS compare the DNA of two groups of participants: subjects with the
phenotype of interest (cases, or persons with a particular disease) and
similar subjects without the phenotype (controls). Each subject provides a
sample of DNA, from which millions of genetic variants are read using single
 Overview of Chapters 3

polymorphism (SNP) arrays. If one type of the variant (one allele, i.e., the wild-
type allele) is more frequent in people with the disease, the SNP is said to be
associated with the disease. The associated SNPs are then considered to mark a
region of the human genome that influences the risk of the phenotype. Also, in
contrast to methods which specifically test one or a few genetic regions, GWAS
investigate the entire genome. The approach is therefore said to be noncandidate
driven, in contrast to gene-specific candidate-driven studies. GWAS identify tag
SNPs, which are defined as representative SNPs in a region of the genome with
high linkage disequilibrium and other variants in DNA associated with a disease.
Tag SNPs in isolation cannot specify which genes cause the phenotype.
The first successful GWAS investigated age-related macular degeneration
and was published in 2005.3 This study found two SNPs that had significantly
altered allele frequency when compared with healthy controls. As of 2015, The
Catalog of Published Genome-Wide Association Studies contained more than
2,141 catalog entries, 1,856 publications and 12,874 implicated SNPs.4 Prior to
the introduction of GWAS, the major method of investigation was via genetic
linkage studies in families. This approach was useful for identifying single-gene
disorders, many of which appear in the comprehensive compendium of human
genes and genetic phenotypes, the Online Mendelian Inheritance in Man
(OMIM) database.5
However, for both common and complex diseases, the results of genetic
linkage studies have been hard to reproduce.6,7 In contrast, GWAS seek
to identify whether the allele of a genetic variant is found more often than
expected in individuals with the phenotype of interest. The statistical methods
used in GWAS are based on traditional approaches, and early calculations of
statistical power indicated that GWAS could be better than linkage studies at
detecting weak genetic effects.8
In addition to a simple conceptual framework, the proliferation of GWAS
has also been driven by improvements in sequencing methods, reduced
computational costs, and the advent of biobanks, which are repositories of
human genetic material that greatly reduce the cost and difficulty of collecting
sufficient numbers of biological specimens for study.9,10 The development of
rapid genome-level sequencing techniques also permits researchers to assess
methods to mine this information to identify genetic associations with disease
and ultimately determine the biological basis of disease patterns. Knowing the
coding sequences of every nucleotide in an organism has permitted researchers
to study the collective influence of all genes simultaneously and their role in
structuring organism traits, including specific diseases.
4 Chapter 1

With improving genotyping technologies and the exponentially growing

number of available markers, case-control GWAS have become a key tool for
investigating complex diseases. To accommodate GWAS methods, researchers
have developed new procedures to ensure data quality, interpret GWAS
findings, and provide computationally tractable approaches when performing
hundreds of thousands of individual tests.
The promise of GWAS was anticipated in many quarters of the scientific
community.11 A 2007 fact sheet released by the National Human Genome
Research Institute, in the early days of GWAS, raised expectations that
personalized medicine, including individual risk prediction, disease
prevention, and specific treatment, was just around the corner. With the
first GWAS published in 2005, health professionals will be able to use such
tools to provide patients with individualized information about their risks of
developing certain diseases to tailor prevention programs to each persons
unique genetic makeup to select the treatments most likely to be effective
and least likely to cause adverse reactions.12 However, a number of critics of
GWAS argue that these expectations have not been met.13

GWAS: Useful or Misleading?

The overly high expectations were created in part because the early GWAS
success predicting age-related macular degeneration with a complement factor
H polymorphism was extreme, with an odds ratio (OR) of approximately 7.*
In contrast, most GWAS implemented after that success involved variants
conferring small effects, indicated by ORs slightly larger than 1.
Although medical science is still far from the GWAS-based personalized
medicine promised in the National Human Genome Research Institute 2007
fact sheet, at least three important considerations fuel legitimate hope that
genetics will become integral to a form of medicine more specifically tailored
to individual patients.11 First, important discoveries have already changed
medical practice and resulted in medical policy codes for some treatments. For
example, in the field of pharmacogenomics, researchers have already begun
using genetic testing to determine patients dosage of warfarin.
Second, genetic interaction studies are starting to provide useful data. For
example, researchers discovered that high-density lipoprotein cholesterol levels

* The OR is a measure of is a measure of association between an exposure and an outcome. The

OR represents the odds that an outcome will occur given a particular exposure, compared to the
odds of the outcome occurring in the absence of that exposure.
 Overview of Chapters 5

(HDL-C), one of the most important risk factors for coronary heart disease,
are significantly influenced by the interplay of multiple genes linked to GWAS
and involved gene-gene interaction effects.14
Third, GWAS are based on common variants (i.e., tag SNPs) that are
frequently in linkage disequilibrium with the actual causative variant, which
in turn may be associated with larger effect sizes than the common variant
included in the GWAS. For instance, fine mapping of loci associated with
low-density lipoprotein cholesterol (LDL-C) identified a rare nonsynonymous
variant gene that explained 5 times more of the contributed variance than the
initial GWAS finding. In this context, whole genome sequence data has the
potential to be a more accurate and powerful tool than SNPs to elucidate the
relationship between genetics and (common) diseases. However, even high-
resolution genetic variation will only explain a fraction of the heritability of
human diseases and traits. Thus, we are still searching for potential uses for
genetics in medical science beyond using simple genetics with gene-gene and
gene-environment interactions and identifying epigenetic effects as important
but complex targets.

Missing Heritability
Heritability is a genetic measure that identifies the observable differences in a
trait due to genetic factors between individuals within a population. Factors
including genetics, environment and random chance can all contribute to the
variation between individuals in their phenotypes.15 Heritability is a dynamic
measurement that identifies the fraction of phenotype variability that can be
attributed to genetic variation. The term missing heritability refers to the
low percentage of information about the overall genetic component and risk
of common diseases gleaned from GWAS. Common variants account for only
a small proportion of genetic components, and the missing heritability lies in
the huge class of rare genetic variants that GWAS do not see. Variants that are
primary drivers of disease are relatively rare in the human population.
Variants confer risk of disease, and natural selection acts against variants so
that they do not become too prevalent. Therefore, the issue of so-called missing
heritability becomes moot. We did not interrogate the whole genome; we
interrogated the common variants that pass through the filter of natural selection.
Many diseases, but not all, will involve rare variants not detected by GWAS.16
Autoimmune diseases may be an exception to this thinking. Some variants
that are major risk factors appear more commonly in the general population;
6 Chapter 1

for example, variants selected in response to infectious agents that have

consequences for autoimmune diseases.17
The inability of GWAS to replicate markers in this and other instances was
a major concern, and in a set of New England Journal of Medicine articles three
authors offer alternative opinions regarding the progress that GWAS methods
have made to date. Hirshhorn argued that the main goal of these studies is not
the prediction of genetic risks but rather the discovery of biological pathways
underlying polygenic diseases and traits.18 Goldstein countered that the
genetic burden of common diseases must be carried mainly by large numbers
of rare variants.19 He also suggested that most GWAS produced too many
associations with very small overall contributions toward explaining disease
risktoo many to provide any useful biologic insights. Kraft and colleagues
support the notion that the GWAS approach predicts that many, rather than
few, variant risk alleles are responsible for the majority of the inherited risk
of each common disease. It is possible that these initial GWAS have identified
only the strongest associations, with many more genes still to be identified.20

Organization of the Book

Chapter 2: Genome Wide Association Data: Where Are the Standards?
Chapter 2 is a previously unpublished manuscript based on an internal study
directed by Philip Cooley. Study participants included Huaqin Pan (RTI), Paul
S. Levy (deceased), Maureen K. Bunger (formerly of RTI), and Laxminarayana
Ganapathi (RTI Health Solutions).
We began our explorations into GWAS by applying statistical methods to
real genetic (SNP) data. We used standard statistical approaches that appear in
both GWAS and non-GWAS literature. The main objective of this venture was
to see if we could assess the performance of different statistical methods used
in a GWAS context. There have been many notable instances in which GWAS
provided inconsistent results and there have been a number of studies have
not been replicable. This has led to a question about the validity of the GWAS
approach. For example, in 2008, the online listing A Catalog of Published
Genome-Wide Association Studies identified a total of 8 amyotrophic lateral
sclerosis (ALS) GWAS.4 Each of these 8 studies identified candidate ALS
markers (or alternatively no markers), but none of the studies could replicate
the results of the other studies. As of 2015, the number of ALS GWAS has
grown to 16 studies that use a variety of international data and have identified
 Overview of Chapters 7

some possible genes linking ALS to genetic causes. By and large, no consistent
markers have been universally accepted as ALS genetic markers.
We obtained the SNP data for the initial ALS study appearing in the
literature21 and attempted to replicate some of the published results in
order to test statistical methods in genomics.21 We identified seven distinct
methods that have been or could be applied in GWAS studies. We also used
the method that Schymick et al. used to obtain their results. At the time we
performed these analyses, we were unaware of any comprehensive studies that
compared the performance of the different methods in the context of GWAS,
and we wanted to determine whether standards could be developed. Using
a previously conducted study would allow us to assess the performance of
specific statistical methods by using the study results as a yardstick by which to
measure the accuracy of our results.
What we learned from this effort is that either many of the algorithms
used in the literature in a GWAS context assume an additive gene model or
are agnostic with respect to the form of inheritance. We also documented
the inability of the ALS studies to replicate resultspart of the problem with
reproducing results is that ALS studies rely on having ALS patients as a study
population, which means that the original sample sizes were relatively small
due to circumstance. We were able to replicate the Schymick et al. study
results using the method they chose to measure associationsthe classic
epidemiology case-control method that uses the Pearson 2 test to text how
likely it is that an observed distribution of data fits with the distribution that is
expected if the variables are independent. However, our assessment indicated
that the reported results depended to an unknown degree on the statistical
method used to make the predictions. This suggested that the algorithms
selected could influence predictive outcome and further demonstrated the
need for developing GWAS standards.
In summary, the absence of both methodological standards and a process
for evaluating the statistical methods used in predicting associations between
genes and phenotypes suggested to us that we could examine these missing
elements more effectively by using simulation methods. Accordingly, we
created simulated data that were linked to known outcomes which therefore
constituted a truth set. The simulated data could be analyzed using different
statistical methods, and we could assess each methods predictive properties,
which could potentially reveal some or all of this missing information.
8 Chapter 1

Chapter 3: Creating the Synthetic Gene Data

Chapter 3 describes the generic data generation process we used for the
research described in Chapters 4 through 7. We used our simulated data in an
effort to exploit a process with a known outcome to identify those traits that
affect GWAS outcomes. These investigations took the form of (1) generating a
database of synthetic genes that incorporated a number of dynamic properties
that were varied for the explicit purpose of developing a truth set (i.e., a
database of known outcomes); (2) testing a number of statistical models
and competing algorithms that were developed to predict associations; and
(3) creating a compendium of outcomes that linked gene properties to statistical
model performance. This virtual gene resource enabled the power performance
of different single-gene statistical methods to be measured and recorded.
Consequently, this chapter describes the simulation model that generated the
synthetic SNP data for all subsequent assessments represented in this book.

Chapter 4: Genetic Inheritance and Genome-Wide Association Statistical Test

Performance Using Simulated Data
Chapter 4 is based on a study that was published previously.22 This chapter
focuses on single-gene models and the statistical methods that effectively
predict associations in a GWAS context. Here, we demonstrate that the choice
of a statistical method can affect the power profiles of GWAS predictions.
The initial step in a GWAS is to apply univariate statistical tests for each SNP
marker in the data set. Applying the tests is methodologically straightforward.
SNP-based tests are used to assess the likelihood of an association. In the
simulation-generated gene data, the probability of the occurrence of the
phenotype is regulated by an exogenously specified risk value that is a function
of the genotype. This allows the strength of the genotype-phenotype signal to
be controlled and power outcomes to be counted. Standard methods (e.g., 2
tests, logistic regressions) are commonly used in single-locus tests. In general,
the GWAS approach is a brute force method that scans the entire genome to
determine which genes demonstrate an association using a stringent threshold
level to compensate for the problem of multiple comparisons. The problem of
multiple test comparisons arises because as the number of tests (hypotheses)
increases, the likelihood of witnessing a rare event increases. Often the mode
of genetic inheritance (MOI) is assumed to be additive, which implies that the
allele causes the phenotype risk rather than the genotype.
 Overview of Chapters 9

Our simulation studies confirmed the results of others23 that the gene
model or MOI was a major influence on statistical power. This was no surprise
and it is well known that associations involving recessive MOI SNPs are much
more difficult to detect than other MOI types. Gene traits that influence
prediction accuracy had also been reported in other studies.24,25 They
demonstrated how the phenotype MOI assumption was a major influence on
association prediction accuracy.
We compared the power profiles of GWAS using a number of statistical
methods, including two that combine MOI-specific methods into multiple
test measures. Because most GWAS investigations have not determined the
specific gene model operating, many assume an additive model. Given that
most models cited in OMIM are either dominant or recessive gene models, we
investigated composite methods that did not make an additive assumption;
rather, they used three component tests with three distinct MOI properties
(dominant, recessive, and additive). We then compared the performance
(from a statistical power perspective) of the composite tests as contrasted with
methods that either assume a specific MOI gene model or are agnostic with
respect to MOI. By combining recessive, additive, and dominant individual
tests, we determined that if the MOI is not known, then a composite test
is more likely to make a correct association prediction. In that sense, it
constitutes a more powerful test and could have significant advantages with
respect to single test procedures.
Our findings did not provide a specific answer about which statistical
method is best. The best method depends on the MOI gene model associated
with the phenotype (diagnosis) in question and how common the traits are
that associate with the phenotype. However, our results do indicate that the
common additive assumption that the MOI of the locus is associated with the
diagnosis can have adverse consequences. It indicated that researchers should
consider a multitest procedure that combines the results of individual MOI-
based core tests as a statistical method for conducting the initial screen in a
GWAS. The process for combining the core tests into a single operational test
can occur in a number of ways. We identify two: the Bonferroni procedure
and the MAX procedure, each of which produces very similar statistical power
profiles.26,27 This was a surprising result because the Bonferroni procedure
was based on combining 2 tests and assumed that the tests were mutually
independent when they clearly are not, whereas the MAX procedure used
10 Chapter 1

normal distribution tests and adjusted for the covariance properties between
individual tests.
In summary, the focus of this chapter is on single-gene statistical methods
that predict associations in a GWAS context. We used simulation methods to
learn that there is no single, most powerful method. If the properties of the
gene model are not known, the most powerful approach is a composite test
that uses a recessive-dominant-additive composite model. We also found that
regardless of whether the MOI is known or not, there always exists a method
that outperforms (in a statistical power context) the Pearson 2 test.

Chapter 5: The Influence of Errors Inherent in GWAS in Relation to Single-

Gene Models
Chapter 5 is based on a study that was published elsewhere.28 This chapter
describes our investigation of the effects of errors in both genotype and
phenotype misclassifications. The central objective is to assess the impact that
these errors have on the additional sample size required to achieve a specific
power threshold, which for this study is set to 80 percent. Usually, GWAS
are conducted assuming that the study measurements are error free. This
chapter discusses the assembled evidence challenging that assumption and the
examples we used to assess the consequence of those assumptions.
We simulated the effects of genotype errors by intentionally mislabeling the
genotype X percent of the time, where X is an exogenously provided model
parameter. It was our assumption that genotype errors due to incorrect chip
assignments affect both genotypes by incorrectly switching the designation of
the disease versus nondisease genotype. We used a similar approach to process
phenotype errors that are assumed to affect disease diagnosis. In this process,
the simulation data will be recoded to simulate a diagnosis switch from a
positive to a negative outcome and vice versa Y percent of the time, where Y
is also an exogenously provided user parameter. In general, we assumed that
the value of X was less than 1 percent. We posed a much higher value for Y
because estimates in the literature report disease misdiagnoses could be as
high as 29 percent.29 These features allowed us to assess the individual or
combination of genotype and phenotype error levels and their statistical power
consequences. Studies of genotype error have led to a number of investigations
in the statistical genetics literature that we review in Chapter 5. Although the
accuracy of the genotyping process has improved, errors still occur.
 Overview of Chapters 11

Phenotypic misclassification errors are also a source of bias and can

reduce the power of detecting a statistical association between a phenotype
and a specific allele.30,31 To help provide insight into the influence of
simultaneous genotype and diagnosis errors affecting the accuracy of the
phenotype measure in a GWAS, we ran simulations with synthetically
generated data. We focused on assessing how statistical power was affected
by the influence of these frequently overlooked sources of errors in GWAS.
Our simulations demonstrated that genotype (even at low error rates) and
phenotype (diagnosis) errors produce substantial power losses for all MOIs,
with significant power losses for recessive MOIs. Because GWAS involving
recessive loci have additional power requirements relative to other MOI types,
researchers need to address these requirements in developing appropriate
sample sizes for their studies.
In summary, this chapter identifies the significant role that epigenetics
effects and diagnosis misclassifications can play in designing tests with realistic
power levels.

Chapter 6: Conducting GWAS Epistasis Scenarios

Chapter 6 is based on a study that was published previously.32 This chapter
presents the results of our investigations of analyzing epistatic scenarios in
GWAS. We used a qualitative association model to assess the statistical models
that reliably predict associations between a qualitative phenotype (i.e., a
disease diagnosis) and a pair of interacting genes. We employ the concept of
relative risk, which is the ratio of the probability of a positive diagnosis given
a mutated genotype divided by the probability with no risk present. We used
a simulation approach to generate synthetic data corresponding to a variety of
possible epistatic models (EMs). Our method took into account the strength of
association, disease prevalence in nonrisk populations, and most importantly,
the inheritance patterns of the pair of epistatic genes. We analyzed the
simulated gene data to assess how these individual factors influenced statistical
power in the context of GWAS.
The results indicated that the most powerful statistical methods for predicting
associations between phenotypes and genotypes in epistatic scenarios are
statistical models that simultaneously test for associations involving both
interacting loci. This has significant computational implications. The number
of single SNP evaluations is as large as 1,000,000 sets of calculations. The
number of SNPs pairs is approximately 51011 calculations. This result is not
12 Chapter 1

surprising and has been reported by others. An additional significance of our

study is that it incorporates new statistical methods as part of the comparison
analysis. We also documented the extent to which single-gene models fail to
predict associations involving interacting genes with phenotypes constructed
to be associated with low risk. Also, each gene MOI affects the ability to
identify the association, which further confounds the GWAS methodology.

Chapter 7: Assessing Gene-Environment Interactions in GWAS: Statistical

Approaches
Chapter 7 is based on a study that was previously published by RTI Press.33
Environmental influencing factors on GWAS are described in Chapter 7.
Classical statistical tests derived from case-control experiments can be used
to determine if two loci associate in a GWAS context. But this model depends
on a narrow range of environmental submodel formulations. In this scenario,
logistic regression models are versatile approaches because they are able to
examine main effects, pairwise interaction assumptions, or both. One early
study investigating gene-gene interactions showed that explicitly modeling
interactions between loci for GWAS with hundreds of thousands of markers is
computationally feasible.34 In this chapter, we also show that simple methods
explicitly considering interactions can actually achieve reasonably high power
with realistic sample sizes under different interaction models with some
marginal effects, even after adjusting for multiple testing using the Bonferroni
correction.
In this chapter, we also focus on low-effect/rare-variant loci with low
relative risks of association with disease diagnosis. The overarching goal
is to identify which statistical methods best identify genotype-phenotype
associations when environmental effects also influence the association.
Detecting such associations is particularly difficult for genetic variants
with modest impacts on risk. Consequently, our experiments specifically
investigated scenarios involving low-risk genetic variants and assessed whether
environmental influences with varied levels of risk could be a source of the
missing heritability observed using single-gene models.35 Not surprisingly,
our investigations demonstrated that the best statistical method (with respect
to statistical power) depends on whether there are interactions between the
genotype and environmental factors as well as how well the specified statistical
model matches the environmental effect associated with the phenotype.
 Overview of Chapters 13

Chapter 8: Polygene Methods in GWAS

Chapter 8 proposes a novel strategy for studying the association between large
sets of SNP predictors and groups of correlated phenotypes (i.e., outcomes).
These data commonly arise in GWAS, in which associations with a large number
of qualitative and quantitative phenotypes are investigated using hundreds of
thousands of genetic markers. Our strategy is formulated within the linearlog
linear models, a framework suited to the analysis of qualitative trait responses.
Our approach is based on the assumption that single-locus models do not
detect all of the markers that are part of the phenotype pathway. In Chapter 6,
we showed that for a given locus, single-locus tests are not as effective as
two-locus tests. Despite these empirical arguments, which are based on
computational considerations, single-gene models remain the core method
for detecting associations in a GWAS context. Accordingly, our general
strategy is to make an initial pass against all usable SNP autosomes and
apply a significance threshold to identify highly significant SNP-phenotype
associations, known as stage 1 SNPs. The second step statistically combines
the stage 1 SNPs with all original autosome SNPs to identify significant SNP
pairs that are phenotype-associated. This step uses a test for significance that is
conditional on the stage 1 SNP. We then propose to continue this process for
triple SNPs, quadruple SNPs, and so on until the combination of loci produces
no new SNP-phenotype associations. This process is analogous to a stepwise
regression process, in which networks of SNPs are combined stage by stage
until no new SNPs exceed the significance threshold.

Chapter 9: Conclusions
Our final chapter discusses statistical power properties in the GWAS context
and offers guidance on selecting best methods. Our early concerns with the
absence of standards led us to develop a synthetic gene database that recorded
known outcomes between synthetic phenotypes and genotype networks
provides a mechanism that was not possible using real genomics data. The
creation of this database enabled an evaluation of different statistical models
and methods specifically because the prediction outcomes were known and
statistical power profiles could be estimated. We also investigated methods
that examine combinations of genes acting in concert either with other
genes or environmental factors. Our assessments indicated that often single-
gene models will fail to identify markers in many types of gene-gene, gene-
environment networks.
14 Chapter 1

We also developed a general polygene test (Chapter 8) that builds up a

network of SNPs that link to a single phenotype. A reliable single-gene model
is still necessary to identify a starting SNP in the polygene process. However,
during this initial pass, the inheritance properties of the genes to which the
SNPs belong are predicted with high reliability. We then use the inheritance
properties of the SNPs in the step-by-step association process. Knowing the
inheritance improves the performance with respect to the statistical power of
our polygene process.

Chapter References
1. Johannsen WL. Arvelighedslrens elementer (the elements of heredity).
Copenhagen, Denmark: Gyldendalske Boghandel Nordisk Forlag; 1905.
2. Alberts B, Johnson A, Lewis J, et al. Molecular biology of the cell. 4th ed.
New York, NY: Garland Science; 2002.
3. Klein RJ, Zeiss C, Chew EY, et al. Complement factor H polymorphism in
age-related macular degeneration. Science. 2005;308(5720):385-9.
4. Burdett T, Hall P, Hasting E, et al. The NHGRI-EBI Catalog of published
genome-wide association studies. 2015 [cited 2015 Nov 2]; Available from:
www.ebi.ac.uk/gwas
5. McKusick-Nathans Institute of Genetic Medicine, Johns Hopkins
University School of Medicine. Online Medelian Inheritance in Man
(OMIM). 2016 [cited 2016 Feb 11]; Available from: http://www.ncbi.nlm.
nih.gov/omim
6. Altmuller J, Palmer LJ, Fischer G, et al. Genomewide scans of complex
human diseases: true linkage is hard to find. Am J Hum Genet.
2001;69(5):936-50.
7. Strachan T, Read A. Human molecular genetics. New York, NY: Garland
Science; 2010.
8. Risch N, Merikangas K. The future of genetic studies of complex human
diseases. Science. 1996;273(5281):1516-7.
9. Ardini MA, Pan H, Qin Y, et al. Sample and data sharing: observations
from a central data repository. Clin Biochem. 2014;47(4-5):252-7.
10. Greely HT. The uneasy ethical and legal underpinnings of large-scale
genomic biobanks. Annu Rev Genomics Hum Genet. 2007;8:343-64.
 Overview of Chapters 15

11. Klein C, Lohmann K, Ziegler A. The promise and limitations of genome-

wide association studies. JAMA. 2012;308(18):1867-1868.
12. National Human Genome Research Institute. Genome-wide association
studies. 2007 [cited 2014 Mar 22]; Available from: http://www.genome.
gov/20019523
13. Macarthur D. Bioscience Resource Project critique of modern genomics: a
missed opportunity. 2010 [cited 2015 Nov 2]; Available from: http://www.
wired.com/2010/12/bioscience-resource-project-critique-of-modern-
genomics-a-missed-opportunity/
14. Hindorff LA, Sethupathy P, Junkins HA, et al. Potential etiologic and
functional implications of genome-wide association loci for human
diseases and traits. Proc Natl Acad Sci U S A. 2009;106(23):9362-7.
15. Raj A, van Oudenaarden A. Nature, nurture, or chance: stochastic gene
expression and its consequences. Cell. 2008;135(2):216-26.
16. Wagner MJ. Rare-variant genome-wide association studies: a new frontier
in genetic analysis of complex traits. Pharmacogenomics. 2013;14(4):413-
24.
17. Hu X, Daly M. What have we learned from six years of GWAS
in autoimmune diseases, and what is next? Curr Opin Immunol.
2012;24(5):571-5.
18. Hirshhorn JN. Genomewide association studiesilluminating biologic
pathways. N Engl J Med. 2009;360(17):1699-1701.
19. Goldstein DB. Common genetic variation and human traits. N Engl J
Med. 2009;360(17):1696-1698.
20. Kraft P, Hunter DJ. Genetic risk prediction--are we there yet? N Engl J
Med. 2009;360(17):1701-1703.
21. Schymick JC, Scholz SW, Fung HC, et al. Genome-wide genotyping in
amyotrophic lateral sclerosis and neurologically normal controls: first
stage analysis and public release of data. Lancet Neurol. 2007;6(4):322-8.
22. Cooley P, Clark R, Folsom R, et al. Genetic inheritance and genome
wide association statistical test performance. J Proteomics Bioinform.
2010;3(12):321-325.
23. Bush WS, Moore JH. Chapter 11: Genome-wide association studies. PLoS
Comput Biol. 2012;8(12):e1002822.
16 Chapter 1

24. Freidlin B, Zheng G, Li Z, et al. Trend tests for case-control studies

of genetic markers: power, sample size and robustness. Hum Hered.
2002;53(3):146-52.
25. Sasieni PD. From genotypes to genes: doubling the sample size.
Biometrics. 1997;53(4):1253-61.
26. Johnson RC, Nelson GW, Troyer JL, et al. Accounting for multiple
comparisons in a genome-wide association study (GWAS). BMC
Genomics. 2010;11:724.
27. Li Q, Zheng G, Li Z, et al. Efficient approximation of P-value of the
maximum of correlated tests, with applications to genome-wide
association studies. Ann Hum Genet. 2008;72(Pt 3):397-406.
28. Cooley P, Clark RF, Page G. The influence of errors inherent in genome
wide association studies (GWAS) in relation to single gene models. J
Proteomics Bioinform. 2011;4:138-144.
29. Kircher T, Nelson J, Burdo H. The autopsy as a measure of accuracy of the
death certificate. N Engl J Med. 1985;313(20):1263-9.
30. Gordon D, Finch SJ, Nothnagel M, et al. Power and sample size
calculations for case-control genetic association tests when errors are
present: application to single nucleotide polymorphisms. Hum Hered.
2002;54(1):22-33.
31. Barendse W. The effect of measurement error of phenotypes on genome
wide association studies. BMC Genomics. 2011;12:232.
32. Cooley P, Gaddis N, Folsom R, et al. Conducting genome-wide association
studies: epistasis scenarios. J Proteomics Bioinform. 2012;5(10):245-251.
33. Cooley PC, Clark RP, Folsom RE. Statistical methods that identify
genotype-phenotype associations in the presence of environmental effects.
RTI Press Publication No. RR-0022-1405. Research Triangle Park, NC:
RTI Press; 2014.
34. Marchini J, Donnelly P, Cardon LR. Genome-wide strategies for
detecting multiple loci that influence complex diseases. Nat Genet.
2005;37(4):413-7.
35. Manolio TA, Collins FS, Cox NJ, et al. Finding the missing heritability of
complex diseases. Nature. 2009;461(7265):747-53.
 CHAPTER 2

Genome-Wide Association Data:

Where Are the Standards?
Philip Chester Cooley

Introduction
In this chapter we assess the predictive strength of a number of classical
statistical methods by applying them to a publically available set of
amyotrophic lateral sclerosis (ALS) data reported in a paper by Schymick
and colleagues (2007).1 We used these methods in the context of a single
locus genome wide association study (GWAS) experiment. The methods
are compared and the degree of similarity/dissimilarity between them is
empirically measured to determine if a combination of methods is more
predictive of phenotype genotype associations than a single method. All of the
methods in our assessment are single nucleotide polymorphism (SNP) based.
There are three types of ALS: classic sporadic, familial, and the Mariana
Island forms. Classic ALS accounts for 90 to 95 percent of ALS patients in the
United States and is called sporadic because it cannot be traced to ancestors
with the illness.2 The literature identifies variations and mutations in many
Mendelian loci and genes that potentially cause different forms and subtypes
of ALS, indicating that many complex and diverse molecular mechanisms are
involved in ALS pathogenesis. These genes include SOD1, ALS2, SETX, and
VAPB for familial ALS, and VEGF, ANG, HFE, SMN, and PON1 for sporadic
ALS. Research has also reported that the inheritance pattern varies with the
type of ALS, including autosomal dominant, autosomal recessive, X-linked
dominant and maternal through mitochondrial genes.3-5
Despite these complexities, ALS researchers have intensified their
investigations of sporadic ALS. Researchers are applying the GWAS approach,
looking for genes that increase susceptibility to sporadic ALS. By 2008 when
the authors investigations began, seven teams had reported results from ALS-
based GWAS,6-12 but none had highlighted genes already under suspicion,
18 Chapter 2

and each team had reported a different set of ALS markers. Subsequently, eight
additional studies13-20 have been reported in the literature. These results have
led to conjectures that using the GWAS method to search for ALS genes may
have to accommodate a spectrum of genes, each of which contribute to ALS
in some unknown manner. The failure of the different studies to replicate each
others results also suggests that GWAS may be inconsistently implemented, or
population admixtures are obscuring the findings.
The first GWAS for ALS found no SNPs significantly associated with the
disease.1 However, published studies that followed implicated several different
ALS genetic markers. For example, Dunckley and colleagues6 identified an
SNP located in the FLJ10986 gene. This study consisted of 1,152 patients
diagnosed with sporadic ALS and 1,297 controls. The initial discovery was
made on analysis done on 386 cases and 547 controls, all of whom were of
European descent and older than 65. A residual sample of 766 cases and 750
controls, as well as a subsample of the data identified in the Schymick study,
were used to replicate the discovery analysis.6 A third ALS study reported that
the inositol 1, 4, 5-triphosphate receptor 2 (ITPR2) marker was associated
with ALS. This study pooled three European populations with 1,337 ALS
patients and 1,356 controls.7 A fourth study (by the same team as the third
study) identified an SNP in the dipeptidyl-peptidase 6 (DPP6) gene that was
strongly associated with ALS susceptibility.8 However, the ITPR2 marker was
no longer significant. A fifth study examined an Irish population, which was
augmented with a Dutch and a US population consisting of 958 ALS cases
and 932 controls, and confirmed an association with the DPP6 marker.21
However, a sixth study sought to confirm the DPP6 marker finding by
examining the Irish ALS cohort data and augmenting it with a Polish cohort.
Cronin and colleagues,21 reported that their analysis of the combined cohorts
that consisted of 1,267 cases and 1,336 controls was unable to identify any
associations including the previously reported DPP6 marker.21 A seventh
study, performed by Chi et al.,9 used a two-stage analysis that consisted
of 553 cases and 2,338 controls: it identified two new markers, but markers
mentioned previously, including DPP6.
Our initial assessment concluded that clear and definitive disease
associations from these seven significant ALS studies was lacking and
suggested difficulties of using GWAS approaches for complex diseases like
ALS. Consequently, we made the decision at that time to pursue simulation-
based studies with known outcomes in an effort to explore the possibility of
developing standard procedures for conducting GWAS.
 Genome-Wide Association Data: Where Are the Standards? 19

Subsequent to our decision to pursue simulation methods to assess GWAS-

appropriate methods, nine additional ALS related GWAS were identified in the
literature. Blauw et al.10 investigated the role of copy number variants (CNVs)
as a source of genetic variation for 406 ALS cases and 406 controls and were
unable to identify a locus associated with sporadic ALS. Another study by
Landers et al.22 implicated the KIFAP3 gene that was associated with increased
survival in sporadic ALS. An additional GWAS by van Es et al.23 used two
cohorts to implicate an SNP (rs12608932) located in the UNC13A gene. This
gene is known to code proteins that are presynaptic proteins found in central
and neuromuscular synapses that regulate the release of neurotransmitters,
peptides, and hormones. This same study also showed genome-wide
significance for two additional SNPs located in chromosome 9p21. A GWAS
by Laaksovirta et al.24 was able to confirm that genes on chromosome 9p21
and suggested that it could be a major cause of familial ALS in a Finnish
population. A further study by Shatunov et al.25 provided additional evidence
that two SNPs in a locus on chromosome 9p21 were associated with ALS. The
International Consortium on Amyotrophic Lateral Sclerosis26 conducted a
meta-analysis of ALS samples consisting of 4,243 ALS cases and 5,112 controls
from 13 European and US cohorts and provided additional evidence for the
loci on UNC13A and chromosome 9p21. However, a study by Daoud et al.27
could not confirm this result in a relatively small population (285 cases and
285 controls) of French ALS subjects. As a result of increased incidence of ALS
in US veterans, Kwee et al.28 conducted a GWAS of ALS outcome and survival
time in a sample of US veterans. They report no SNPs reached genome-wide
significance in the discovery phase for either phenotype. A final GWAS
by Deng et al.29 on a Chinese population reported finding two additional
susceptibility loci that were not reported in other studies.

Methods
Our study assessed the statistical methods that have appeared in the GWAS
literature and included a number of established methods used by the cited
studies. We applied each of these methods to the ALS data of Schymick and
colleagues1 and provide a method of comparing their relative performance.

The Schymick et al. (2007) ALS data set

This SNP data was produced by the Laboratory of Neurogenetics of the
intramural program of the National Institute on Aging (NIA), National
Institutes of Health (NIH). The genotyping was performed using the Illumina
20 Chapter 2

Infinium assay humanhap550. Infinium assays assess haplotype tagging SNPs

based upon Phase I+II of the International HapMap Project. The genotype
data we used in this study consists of 555,352 SNPs from 276 ALS patients and
271 neurologically normal controls. These data are publicly available to the
scientific community.1

The Statistical Tests

In a simple GWAS analysis, we rely on individual statistical tests of each typed
SNP to identify potential associations. For any of the statistical methods
considered for measuring associations, we can represent the sample genotype
data or the sample allele data in a contingency array stratified by cases and
controls. Under a null hypothesis of no association with the phenotype, we
expect no difference between the frequencies across cases and controls. In
general, this is the strategy behind all of the statistical methods considered.
However, the methods make different operational assumptions that produce
very different measures of association.
Testing a persons DNA for more than half a million SNPs will produce
many spurious associations. For example, if there were no actual associations,
500,000 independent tests using a p<.0001 criterion would be expected to
identify 50 candidate genes. Statistical methods can correct for this, but they
can also obscure real associations and produce both false positive and false
negative associations. For this study, we are using the data generated by the
Illumina 550K chip, which has an SNP about every 5 kb. Also, loci near each
other might not be mutually independent. The general strategy for protecting
against type I errors is by setting a stringent statistical significance threshold.
Each locus is screened to establish if there is sufficient information to apply
the statistical procedures. The method we used is based on the procedure
defined by Zeggini and colleagues.30 SNPs are considered eligible if the minor
allele frequency (MAF) exceeds 1 percent in both cases and controls. These
restrictions help protect against computational difficulties caused by data
sparseness in our statistical calculations. Another purpose of any eligibility test
is to determine if there is sufficient representation of all genotypes to perform
an accurate statistical test.
We investigated a number of statistical tests to use in our assessment
including the following.
The Case-Control Genotype Method, Based on the Pearson 2 Test. This classic
test is used in many epidemiological studies. The procedure constructs a 23
genotype table (case control by the three genotypes) that uses a Pearson 2 test
 Genome-Wide Association Data: Where Are the Standards? 21

with 2 degrees of freedom to test the hypothesis that the cases and controls are
from the same distribution. Under the null hypothesis of no association with
disease, we expect the relative genotype frequencies to be the same in cases and
controls. These types of methods are commonly used in the GWAS context. A
case-control study uses the odds ratio to estimate the relative risk and assumes
that the disease under study has a low incidence. When the risk ratio is the
parameter of interest, the assumption of rarity is needed for the odds ratio to
be a consistent estimator.31,32 This method was used in the original Schymick
study.
Normal Approximation to Fishers Exact TestDominant and Recessive Models. The
null hypothesis behind Fishers test is that the rows (phenotype) and columns
(genotypes) are unrelated. The test calculates an exact probability value for the
relationship between three dichotomous variables, as found in a 23 table.
When N (the number of subjects) is large, the exact form of the Fisher test is
difficult to calculate. Therefore, a normal approximation is used. Because the
test estimates the probability of a given genotype using the marginal values and
assumes the probability is from a normal curve, an autosomal recessive and
dominant version of the test is easily implemented. We used both forms in our
assessment, listed as Fis-D and Fis-R in Table 2.1.
Logistic Regression Linear and Categorical Model Tests. The logistic regression
test (Log-A) assumes an additive mode of phenotype inheritance and regresses
case control outcomes using the number of minor alleles as the dependent
variable. In the Log-A test, the null hypothesis 1 = 0 is used to test if the
number of alleles associates with the phenotype variable.33

Table 2.1. Correlation values for eight statistical tests based on unadjusted p-values
Test
Name Fis-D Pea Tr-A All Fis-R Tr-D Log-A Tr-R
Fis-D 1.0 -.0472 -.0544 -.0535 0.2215 -.0947 -.0551 -.0041
Pea -.0472 1.0 0.4861 0.4832 -.0651 0.4818 0.4874 0.4793
Tr-A -.0544 0.4861 1.0 0.9843 -.0105 0.5294 0.9973 0.0821
All -.0535 0.4832 0.9843 1.0 -.0104 0.5269 0.9834 0.0811
Fis-R 0.2215 -.0651 -.0105 -.0104 1.0 -.0026 -.0107 -.1677
Tr-D -.0947 0.4818 0.5294 0.5269 -.0026 1.0 0.5299 0.0020
Log-A -.0551 0.4874 0.9973 0.9834 -.0107 0.5299 1.0 0.0822
Tr-R -.0041 0.4793 0.0821 0.0811 -.1677 0.0020 0.0822 1.0
Fis-D = Fisher dominant test; Pea = Pearson 2 test; Tr-A = trend additive test; All = allelic test; Fis-R = Fisher
recessive test; Tr-D = trend dominant test; Log-A = logistic linear test; Tr-R = trend recessive.
22 Chapter 2

Cochran-Armitage Trend Tests: Recessive, Additive, and Dominant Models. The

classical Cochran-Armitage test assumes an additive mode of inheritance
(MOI) and is typically used in categorical data analysis when some categories
are ordered. The test is sensitive to the linearity between phenotype and
genotype variables and detects trends that would not be noticed by other
tests.34 It also uses weights applied to each genotype variable to generalize the
method. For example, a weight of 0.0, X, or 1.0 (where 0 X 1) can be used
to assume different MOI assumptions: recessive (X = 0), additive (X = .5) or
dominant (X = 1) locus. The method is discussed by Zheng and Gastwirth.35
We use all three test assumptions in our assessment and identify them as Tr-R,
Tr-A, and Tr-D, respectively, in Table 2.1.
Allele Test. This is a commonly used test for association in a 22
contingency table, in which cases and controls are classified as carriers of the
minor, risk-carrying allele. This is a 1 degree of freedom (1df) test that assumes
dominance on a single allele. This test constructs a 22 allele table (case
control by the two alleles) and uses a Pearson 2 (CHI2) test (1df) to test the
hypothesis that the cases and controls are from the same distribution.32

Results
A fundamental question this study seeks to address is which method should
investigators use to assess candidate associations? No one has yet answered
this question. One obvious partial answer is that it depends on the gene
behavior, as well as a number of other biological factors yet to be determined.
Consequently, if we assume an additive gene model, then the likelihood of
establishing the association between genotype and phenotype will be nearly
the same whether or not one uses any of the three additive-based tests.

Test Performance
In an attempt to address the question above, we created a 23 table of counts
of case-control subjects by genotype counts for each locus. We then added
the marginal values and provided the necessary information to apply the
eligibility criteria to eliminate loci with low minor genotype representation.
The data provides information on 555,352 loci, of which 538,234 are autosome
loci. Applying a standard quality control method described by Zeggini et al.30
to the autosome loci identifies 55,304 loci (10 percent) that were screened
ineligible and not considered in the analysis due to low minor genotype
representation. Table 2.1 presents the 88 matrix of Pearson product-moment
correlation coefficients for all possible pairs of the eight statistical tests. Note
 Genome-Wide Association Data: Where Are the Standards? 23

that the statistical tests differ in the MOI assumption and therefore should
provide different results. However, Table 2.1 indicates that although some of
the MOI-based tests (i.e., the additive tests) produce consistent results, the
dominant based tests (Tr-D and Fis-D: correlation coefficient = -.0947), and
the recessive based tests (Tr-R and Fis-R correlation coefficient = -.1677) are
not correlated. The Pearson test is MOI agnostic and would be expected to
overlap with the other tests, and although they both capture 48 percent of the
correlation space represented by many of the other tests, they have only a small
negative correlation with both of the Fisher tests. Thus, Table 2.1 illustrates that
all of the tests historically used in GWAS incorporate different assumptions
and consequently have differences in the way they measure association. Based
on results presented here, we assert that if the MOI characteristics of the
associated loci are unknown, then researchers should consider multiple tests
treated in a nonhierarchal manner.
In addition, Table 2.1 suggests a correlation between the three tests (Tr-A,
Log-A, and All) that assume additive MOI properties. This suggests that no
additional predictive power for measuring associations is derived from using
more than one of these three tests; that weak correlation between the two tests
(Tr-R and Fis-R) that assume recessive MOI properties and weak negative
correlation between the Fisher dominant MOI test (Fis-D) and all of the other
tests except the Fisher recessive MOI test (Fis-R), which implies that these tests
measure a dimension that is different from all of the other tests.

Marker Assessment
This section compares our results directly to those of Schymick and colleagues1
and indirectly to other results reported in the literature.

Schymick et al. (2007). The Schymick and colleagues study1 reports using
six association tests: the genotypic test, two versions of the trend test (the
dominant and additive tests), a recessive model test, an allele-based test, and
a three-marker haplotype-association test. We included their five-loci-based
tests in the tests we ran.
We applied the Pearson test, the recessive version of the trend test, and the
additive version of the logistic regression test to the same data. A summary
of these results is presented in Table 2.2 below. It compares the top 34 SNPs
reported by Schymick and colleagues1 with our results. We found that all 34
were positive at the e-4 level of significance based on the Pearson test but that
none were significant at the 10-7 level. Also, 13 of the SNPs were positive at the
24 Chapter 2

Table 2.2. Comparison between the results of our study and the Schymick et al. study
SNP ID Chrom.$ Location Gene Pea Tr-R Log-A
rs4363506 10q26.13 129164493 Intergenic <.000001 <.005 <.000001
rs16984239 2p24 18097927 Intergenic <.00001 X <.00001
rs12680546 8q24.2 136940921 Intergenic <.0001 X <.001
rs6013382 20q13.2 50136040 ZFP64 <.00001 X <.01
rs2782931 9q31.3 113890011 SUSD1 <.00001 <.0001 X
rs11099864 4q31.3 154112804 KIAA1727 <.00001 <.0005 X
rs332389 3p14.1 $66493904 SLC25A26 <.0001 <.01 X
rs4964213 12q23.3 106274907 BTBD11 <.0001 <.0001 <.0001
rs10765118 10q26.13 129175173 Intergenic <0001 <.00001 <.0001
rs3733242 4q21.1 77894529 SHROOM3 <.0001 <.00001 <.0001
rs1037666 1q43 238425108 FMN2 <.0001 X <.001
rs1436918 15q14 32724213 LOC390569 <.0001 <.0001 X
rs4552942 8q24.2 136943505 Intergenic <.0001 X <.001
rs852801 1p32.2 58094497 DAB1 <.0001 <.00001 <.001
rs852802 1p32.2 58096531 DAB1 <.0001 <.00001 <.001
rs7250467 19q12 33261241 LOC727771 <.0001 <.0001 <.0001
rs10830099 10q26 129174355 Intergenic <.0001 <.0001 < 0001
rs10459680 15q26 91482474 Intergenic <.0001 X <.001
rs1752784 9q22.32 96217647 HIATL1 <.0001 <.00001 <.01
rs1202824 1p32.2 58121593 DAB1 <.0001 X <.01
rs5014235 5q14.1 77245417 Intergenic <.0001 <.01 <.0001
rs7201419 16q23.3 81887480 CDH13 <.0001 <.0001 <.01
rs11933187 4q31.3 175446507 KIAA1717 <.0001 <.0001 <.001
rs10773543 12q24.32 127489679 TMEM132C <.0001 X <.001
rs7976059 12q13 50537539 Intergenic <.0001 <.0001 <.001
rs9608416 22q12.1 24441018 ADRBK2 <.0001 <.01 X
rs2220999 12q12 40422035 Intergenic <.0001 <.01 <.0001
rs12632457 3p24 27995556 Intergenic <.0001 <.01 <.0001
rs2272519 2p24 18575231 Intergenic <.0001 <.05 X
rs2289599 5q14.1 77243905 Intergenic <.0001 <.05 <.0001
rs4478530 8p12 31517686 Intergenic <.0001 <.05 X
rs130110 22q13.32 47470399 FAM19A5 <.0001 X <.0001
rs9510982 13q12 23451861 Intergenic <.0001 <.0001 X
rs2767584 6p21 156964439 Intergenic <.0001 X X
Total < e-4 34 13 12
Fis-D = Fisher dominant test; Pea = Pearson 2 test; Tr-A = trend additive test; All = allelic test; Fis-R = Fisher
recessive test; Tr-D = trend dominant test; Log-A = logistic linear test; Tr-R = trend recessive.
$Chromosome location of single nucleotide polymorphism (SNP).

Some data from Schymick et al.1

 Genome-Wide Association Data: Where Are the Standards? 25

10-4 criteria according to the Trend-R test and 12 were positive according to
the Logistic-A test.
Thus, 12 of 34 association tests at the e-4 level overlap between the MOI-
agnostic Pearson test and both the recessive Tr-R test and the Log-A test. We
can therefore replicate the Schymick et al. results but only if we use the Pearson
test, which is commonly used for GWAS. The Trend-A test is equally popular.
However, if another MOI-specific test had been used instead, we could not
have replicated the results.

Other ALS GWAS. The year we first investigated the Schymick et al.1 ALS
GWAS study, there were seven other GWAS also published on ALS. These
studies were Dunckley et al.6; van Es et al.7,8; Blauw et al.10; Cronin et al.11,12;
and Chi et al.9 The focus of the Blauw et al.10 study was unique. It investigated
copy number variations, and the reported results were negative. We examined
the top markers reported in the remaining six studies and summarize that
information in Table 2.3. The DPP6 marker first identified by van Es et al.8
is worth mentioning because it was also reported in one of the Cronin et al.
studies.11

Table 2.3. Association test results of SNPs identified as significant in other studies
p-value p-value this
Study SNP ID Chr Gene reported study
Dunckley6 rs6700125 1 FLJ10986 1.8 10-5 3.1 10-3
Dunckley6 rs6690993 1 FLJ10986 2.0 10-4 NA
van Es7 rs2306677 12 ITPR2 7.0 10-4 X
van Es,8 Cronin11 rs10260404 7 DPP6 5.04 10-8 7.4 10-4
Chio9 rs2708909 7 SUNC1 6.98 10-7 4.5 10-3
Chio9 rs2708851 7 Intergenic 1.16 10-6 9.5 10-3
NA = SNP not included on chip; X = not significant at any level; SNP = single nucleotide polymorphism.

The results summarized in Table 2.3 identify a p-value < 10-3 for one of
the markers reported in other studies; however, all of the five others are less
significant. At this threshold, we are unable to replicate any other reported
results and therefore must hypothesize that the collective ALS studies have
identified no predisposing biological markers. We also performed a linkage
disequilibrium (LD) analysis of all SNPs in Table 2.3 to identify any other SNP
close to the index SNP in column 2 (i.e., < 200kb of the index SNP) with high
LD. We were unable to identify any.
26 Chapter 2

Finally, we examined the PON1 gene that was reported to be associated

with sporadic ALS in genetic studies by more than one study. 36,37 All tested
nonsignificant.

Conclusions
We used publicly available data that contained 276 cases and 271 controls.
This sample is very underpowered, with many SNPs containing few disease
alleles. The SNP coverage may also have been insufficiently dense. Although
the SNPs on the chip are tag SNPs that were selected to represent all the
genes in a comprehensive manner, the selection may not have included all the
potential markers for a specific disease. Other explanations are that the sample
of individuals was too heterogeneous, resulting in the study populations with
distinct disease penetrance traits; the strength on the association was too
weak to be detected; errors in the data perturbed the measurements; or the
phenotype definition differed across studies. Novel statistical methods cannot
overcome the problems inherent in poor quality data containing too few
subjects, or data that uses a poorly defined phenotype (i.e., an ALS diagnosis).
We found that different statistical tests varied significantly in estimating
the test association measurement, implying that GWAS results depend on the
chosen statistical method. We also find that there is no compelling standard
that establishes which statistical methods investigators should use in the
context of GWAS with unknown MOI properties. Although GWAS have
used and reported findings for a number of different tests, most tests have
unique properties and consequently can prescribe a different candidate set of
phenotype SNP associations. In general, the p-value threshold is Bonferroni
corrected to a very small value, which encourages high type II error rates.38
GWAS hold substantial promise, but for many phenotypes, the path
forward is complicated for many reasons not yet understood. The replicability
of results in some but not all studies suggests that researchers should present
any and all conclusions with caution. Identifying associations solely on the
basis of extreme p-values is likely to be misleading because an extreme p-value
alone does not identify the underlying biological mechanism that produces the
association. Furthermore, the rate of missing genotype measures suggests that
the genotype data contains an unknown number of errors (e.g., in the case of
ALS, obtaining an accurate diagnosis is notoriously difficult). Accurate error
rates are important components for assessing statistical power properties, and
their absence will lead to underpowered GWAS.
 Genome-Wide Association Data: Where Are the Standards? 27

Chapter References
1. Schymick JC, Scholz SW, Fung HC, et al. Genome-wide genotyping in
amyotrophic lateral sclerosis and neurologically normal controls: first
stage analysis and public release of data. Lancet Neurol. 2007;6(4):322-8.
2. MedicineNet.com. Amyotrophic lateral sclerosis (ALS or Lou Gehrigs
Disease). 2015 Nov 23 [cited 2015 Nov 23]; Available from: http://www.
medicinenet.com/amyotrophic_lateral_sclerosis/article.htm
3. Orrell RW. Understanding the causes of amyotrophic lateral sclerosis. N
Engl J Med. 2007;357(8):822-3.
4. Hardiman O, Greenway M. The complex genetics of amyotrophic lateral
sclerosis. Lancet Neurol. 2007;6(4):291-2.
5. Pasinelli P, Brown RH. Molecular biology of amyotrophic lateral sclerosis:
insights from genetics. Nat Rev Neurosci. 2006;7(9):710-23.
6. Dunckley T, Huentelman MJ, Craig DW, et al. Whole-genome analysis of
sporadic amyotrophic lateral sclerosis. N Engl J Med. 2007;357(8):775-88.
7. van Es MA, Van Vught PW, Blauw HM, et al. ITPR2 as a susceptibility
gene in sporadic amyotrophic lateral sclerosis: a genome-wide association
study. Lancet Neurol. 2007;6(10):869-77.
8. van Es MA, van Vught PW, Blauw HM, et al. Genetic variation in DPP6 is
associated with susceptibility to amyotrophic lateral sclerosis. Nat Genet.
2008;40(1):29-31.
9. Chio A, Schymick JC, Restagno G, et al. A two-stage genome-wide
association study of sporadic amyotrophic lateral sclerosis. Hum Mol
Genet. 2009;18(8):1524-32.
10. Blauw HM, Veldink JH, van Es MA, et al. Copy-number variation in
sporadic amyotrophic lateral sclerosis: a genome-wide screen. Lancet
Neurol. 2008;7(4):319-26.
11. Cronin S, Berger S, Ding J, et al. A genome-wide association study of
sporadic ALS in a homogenous Irish population. Hum Mol Genet.
2008;17(5):768-74.
12. Cronin S, Greenway MJ, Prehn JH, et al. Paraoxonase promoter and
intronic variants modify risk of sporadic amyotrophic lateral sclerosis.
J Neurol Neurosurg Psychiatry. 2007;78(9):984-6.
28 Chapter 2

13. Ahmeti KB, Ajroud-Driss S, Al-Chalabi A, et al. Age of onset of

amyotrophic lateral sclerosis is modulated by a locus on 1p34.1. Neurobiol
Aging. 2013;34(1):357 e7-19.
14. Deng M, Wei L, Zuo X, et al. Genome-wide association analyses in Han
Chinese identify two new susceptibility loci for amyotrophic lateral
sclerosis. Nat Genet. 2013;45(6):697-700.
15. Kwee LC, Liu Y, Haynes C, et al. A high-density genome-wide association
screen of sporadic ALS in US veterans. PLoS One. 2012;7(3):e32768.
16. Laaksovirta H, Peuralinna T, Schymick JC, et al. Chromosome 9p21 in
amyotrophic lateral sclerosis in Finland: a genome-wide association study.
Lancet Neurol. 2010;9(10):978-85.
17. Landers JE, Melki J, Meininger V, et al. Reduced expression of the
Kinesin-Associated Protein 3 (KIFAP3) gene increases survival in
sporadic amyotrophic lateral sclerosis. Proc Natl Acad Sci U S A.
2009;106(22):9004-9.
18. Shatunov A, Mok K, Newhouse S, et al. Chromosome 9p21 in sporadic
amyotrophic lateral sclerosis in the UK and seven other countries: a
genome-wide association study. Lancet Neurol. 2010;9(10):986-94.
19. Sha Q, Zhang Z, Schymick JC, et al. Genome-wide association reveals
three SNPs associated with sporadic amyotrophic lateral sclerosis through
a two-locus analysis. BMC Med Genet. 2009;10:86.
20. van Es MA, Veldink JH, Saris CG, et al. Genome-wide association study
identifies 19p13.3 (UNC13A) and 9p21.2 as susceptibility loci for sporadic
amyotrophic lateral sclerosis. Nat Genet. 2009;41(10):1083-7.
21. Cronin S, Tomik B, Bradley DG, et al. Screening for replication of
genome-wide SNP associations in sporadic ALS. Eur J Hum Genet.
2009;17(2):213-8.
22. Landers JE, Melki J, Meininger V, et al. Reduced expression of the
Kinesin-Associated Protein 3 (KIFAP3) gene increases survival in
sporadic amyotrophic lateral sclerosis. Proc Natl Acad Sci U S A.
2009;106(22):9004-9.
23. van Es MA, Veldink JH, Saris CG, et al. Genome-wide association study
identifies 19p13.3 (UNC13A) and 9p21.2 as susceptibility loci for sporadic
amyotrophic lateral sclerosis. Nat Genet. 2009;41(10):1083-7.
 Genome-Wide Association Data: Where Are the Standards? 29

24. Laaksovirta H, Peuralinna T, Schymick JC, et al. Chromosome 9p21 in

amyotrophic lateral sclerosis in Finland: a genome-wide association study.
Lancet Neurol. 2010;9(10):978-85.
25. Shatunov A, Mok K, Newhouse S, et al. Chromosome 9p21 in sporadic
amyotrophic lateral sclerosis in the UK and seven other countries: a
genome-wide association study. Lancet Neurol. 2010;9(10):986-94.
26. Ahmeti KB, Ajroud-Driss S, Al-Chalabi A, et al. Age of onset of
amyotrophic lateral sclerosis is modulated by a locus on 1p34.1. Neurobiol
Aging. 2013;34(1):357 e7-19.
27. Daoud H, Belzil V, Desjarlais A, et al. Analysis of the UNC13A gene
as a risk factor for sporadic amyotrophic lateral sclerosis. Arch Neurol.
2010;67(4):516-7.
28. Kwee LC, Liu Y, Haynes C, et al. A high-density genome-wide association
screen of sporadic ALS in US veterans. PLoS One. 2012;7(3):e32768.
29. Deng M, Wei L, Zuo X, et al. Genome-wide association analyses in Han
Chinese identify two new susceptibility loci for amyotrophic lateral
sclerosis. Nat Genet. 2013;45(6):697-700.
30. Zeggini E, Weedon MN, Lindgren CM, et al. Replication of genome-wide
association signals in UK samples reveals risk loci for type 2 diabetes.
Science. 2007;316(5829):1336-41.
31. Szklo M, Nieto FJ. Epidemiology: beyond the basics. Boston, MA: Jones
and Bartlett; 2004.
32. Balding DJ, Bishop M, Cannings C, editors. Handbook of statistical
genetics. 3rd ed. Chichester, United Kingdom: John Wiley & Sons; 2007.
33. Garson GD. Logistic regression, in Statnotes: Topics in multivariate
analysis. 2008. Available from: http://www2.chass.ncsu.edu/garson/pa765/
statnote.htm
34. Agresti A. Categorical data analysis. 2nd ed. Hoboken, NJ: John Wiley &
Sons; 2002.
35. Zheng G, Gastwirth JL. On estimation of the variance in Cochran-
Armitage trend tests for genetic association using case-control studies.
Stat Med. 2006;25(18):3150-9.
30 Chapter 2

36. Mackenzie IR, Bigio EH, Ince PG, et al. Pathological TDP-43 distinguishes
sporadic amyotrophic lateral sclerosis from amyotrophic lateral sclerosis
with SOD1 mutations. Ann Neurol. 2007;61(5):427-34.
37. Slowik A, Tomik B, Wolkow PP, et al. Paraoxonase gene polymorphisms
and sporadic ALS. Neurology. 2006;67(5):766-70.
38. Balding DJ. A tutorial on statistical methods for population association
studies. Nat Rev Genet. 2006;7(10):781-91.
 CHAPTER 3

Creating the Synthetic Gene Data

Philip Chester Cooley

Overview
We used simulation methods to compensate for the absence of both methodo
logical standards and a process for evaluating the statistical methods used in
predicting associations between genes and phenotypes. This required creating a
synthetic gene database of simulated data linked to known association outcomes
constituting a truth set. Using these data, we analyzed the simulated data
using candidate statistical methods for the purpose of assessing each methods
predictive properties in the context of an experimental setting that we create.
Our method for generating the synthetic marker data is based on Mendelian
concepts of inheritance and epidemiological concepts of relative risk (RR),
which is the ratio of the probability of an event occurring in an exposed group
to the probability of the event occurring in a comparison, nonexposed group.
For example, nonsmokers who inhale secondhand smoke may be more likely to
develop lung cancer than nonsmokers who have not been exposed. Individuals
also have genetic inheritance elements that include autosomal dominant and
autosomal recessive patterns conforming to single-gene inheritance effects. We
also incorporate additive and multiplicative inheritance patterns to represent
the actions of multifactorial inheritance processes.
We used a study by Iles to represent the contrast between a formal
disease diagnosis that stems from genetic causes and the concept of disease
penetrance.1 Penetrance in genetics is the proportion of individuals carrying
a particular variation of a gene (allele or genotype) that also express an
associated trait. We designate a as the risk allele, and A as the allele without
risk. Generating the synthetic gene data set was facilitated by defining the
relationships between penetrance and relative risk for different MOI categories.
This chapter describes a generic process to generate genotype-phenotype
data that we use in Chapters 4 through 8. All of the chapters use a version of
the data generation process that is derived from this generic process, but there
can be differences in detail that depend of the technical content of the chapter.
32 Chapter 3

The Data Generation Process

Specifically the steps were:
1. Preload the details that define the factor combinations for each MOI
category. The factors are specified in Table 3.1.

Table 3.1. Factors that define a synthetic gene data file

Factor Symbol Number of factors
Sample size N NCC
Penetrance P NP
Phenotype error rate Perr NY
Genotype error rate Gerr NX
Relative genetic risk NGR
Relative environmental risk NER

2. Draw a genotype distribution at random from the master set of genotype

distributions obtained from real distribution data (i.e., the study by
Schymick et al.2). At this stage, Chan et al. recommends that a minor
allele frequency (MAF) threshold not be applied.3 They argue that filtering
MAFs out of the process because of low frequencies or to maintain Hardy
Weinberg equilibrium (HWE) deviation has little effect on the overall false
positive rate and in some cases, filtering MAF only serves to exclude SNPs.
This step effectively selects a specific genotype distribution (at random)
from the master distribution.
3. Use Table 3.2 to assign a case (1) or a control (0) based on the selected
genetic relative risk (), penetrance (P) and MOI category. This step
converts the ratio value into the probability that the case occurs for the
MOI gene model of interest. This process is represented by the following
logic that was derived from Iles1:

Major Homozygote (AA). Assume that the AA genotype is selected. The

probability of a case given this selection is equal to the disease penetrance P,
or AA = P.

Minor Homozygote (aa): Liability Increasing Allele. Assume the aa genotype is

selected. The genetic relative risk (aa) can be expressed as a ratio of two
 Creating the Synthetic Gene Data 33

probabilities: the probability of a case for a minor homozygote divided by

the probability of a case for a major homozygote, or

aa = Prob(case/aa) / Prob(case/AA) = x/P. (3.1)

From (3.1) the probability of a case given the minor genotype = x =

aaP, (3.2)

where aa = one of the assigned risk factors and P is one of the assigned
penetrance factors.

Heterozygote (aA). Assume the aA genotype is selected. By the same

argument, the phenotype risk given a heterozygote is:

aA = Prob(case/aA) / Prob(case/AA) = y/P. (3.3)

By the same argument, the risk of a case given the heterozygote is

y = aAP, (3.4)

where aA = one of the assigned risk factors and P is one of the assigned
penetrance factors.
Using the estimate of x and y, assign a case or control at random
using the four different MOI models in conjunction with equations
(3.2) and (3.4) and Table 3.2. We assigned cases in proportion to x (y)
and controls in proportion to 1- x (1-y) for the minor homozygote
(heterozygote) genotypes respectively. For the MOI models that assume
an elevated risk from the minor and the hetero genotypes, we would
expect a higher proportion of cases to be more easily identified via the
statistical procedures. The specification of risk depends on specific and
unknown disease mechanisms. A relative risk of 1.7 is considered strong
and is associated with positive replication,4 and a risk of 1.3 is considered
by Ziegler et al.5 to be a realistic assumption for complex diseases. In
summary, individuals are either assigned as cases or controls according to
the probabilities given in Table 3.2.

4. Systematically select subjects. If the subject is a case (control), change its

phenotype designation to a control (case) at a rate determined by Perr.
5. Systematically select subjects. If the genotype is a disease (nondisease) allele
change the allele to a nondisease allele (disease) allele at a rate determined
by Gerr.
34 Chapter 3

6. Continue with the previously described process until n1 cases and n2

controls (N = n1 + n2) are generated (note that n1 and n2, are not required
to be equal).
7. Apply a set of statistical methods to predict associations and record the
results.
8. Generate NR (typically NR = 1,000) replicate experiments for each factor
combination.
9. Analyze the data.

Table 3.2. Relative risk assumptions, by mode of inheritance

Major homozygote Minor homozygote Heterozygote
Pr(case/aa) Pr(case/aA)
AA aa = Aa =
Inheritance model Pr(case/AA) Pr(case/AA)
Recessive 1 1
Dominant 1
Additive 1 2-1
Multiplicative 1
Source: Iles.1

aa is the relative risk of homozygous minor to homozygous major.

aA is the relative risk of heterozygote to homozygous major.

Figure 3.1 presents a flow description of the data generation process.

 Creating the Synthetic Gene Data 35

Figure 3.1. Schema of data-generation process

Preload factor values

Penetrance = p = Pr (Case/AA) = Low, Med, High Low, Med, High

Number of cases = N1 Low, Med, High
Number of controls = N2 Low, Med, High
Phenotype error rates = Perr Low, Med, High
Genotype error rates = Gerr Low, Med, High
Risk rates = Low, Med, High

Compute master genotype distributions for each SNP

G = number of wild type alleles

Loop
components:
P
MOI
N1
N1
Perr
Gerr

Replicates

Select a G: Genotype distribution

MOI G = 0 (AA) G = 2 (aa) G = 1 (Aa)
Pr(case/aa) Pr(case/aA)
AA aa = =
Pr(case/AA) Aa Pr(case/AA)
Select a genotype value g = (0,1,2) at random from G R 1 1
D 1
A 1 2-1
M 1
With g, p, , MOI Compute case (1) or control (0)
using Table 2

Loop
Alter genotype from error rate (0 to 1, 1 to 2, 2 to 0)

Record database entry

No Yes
Completed loops? End

AA = major genotype; Aa = heterozygote genotype; aa = minor genotype; MOI =

mode of inheritance.
36 Chapter 3

Computational Requirements
Note that the number of unique entity combinations being simulated (NS) as
described by the data generation process (see Table 3.1) is:

NS = NPNXNYNGRNER. (3.5)

Each of the NS combinations of traits consists of N cases + controls, and

each unique combination is replicated NR times. Thus, if the each trait of
interest had three risk levels (high, medium, and low) and if NR = 1,000,
the number of unique experimental combinations would equal 35 = 243,
the number of replicate simulations = 243,000, and the number of entries in
the database = 2.43105N. Depending on the value of N, the number of
hours of computations required by this process can vary from 2 to 80, and the
database entries from 2.43108 (N = 1,000) to 2.431011 (N = 100,000). Note
the space requirement on the later assumption is approximately 10 terabytes.

Chapter References
1. Iles MM. Effect of mode of inheritance when calculating the power of a
transmission/disequilibrium test study. Hum Hered. 2002;53(3):153-7.
2. Schymick JC, Scholz SW, Fung HC, et al. Genome-wide genotyping in
amyotrophic lateral sclerosis and neurologically normal controls: first stage
analysis and public release of data. Lancet Neurol. 2007;6(4):322-8.
3. Chan EK, Hawken R, Reverter A. The combined effect of SNP-marker and
phenotype attributes in genome-wide association studies. Anim Genet.
2009;40(2):149-56.
4. Sladek R, Rocheleau G, Rung J, et al. A genome-wide association study
identifies novel risk loci for type 2 diabetes. Nature. 2007;445(7130):881-5.
5. Ziegler A, Konig IR, Thompson JR. Biostatistical aspects of genome-wide
association studies. Biom J. 2008;50(1):8-28.
 CHAPTER 4

Genetic Inheritance and Genome-Wide

Association Statistical Test Performance
Using Simulated Data
Philip Chester Cooley, Robert F. Clark,
Ralph E. Folsom, and Grier Page

Overview
Choosing a particular statistical method for a study significantly affects
the power profiles of genome-wide association study (GWAS) predictions.
Previous simulation studies of a single synthetic phenotype marker
determined that the gene model or mode of inheritance (MOI) was a major
influence on power. In this chapter, we compare the power profiles of GWAS
statistical methods, ones that combine MOI specific methods into multiple
test scenarios, against individual methods that may or may not assume an
MOI gene model consistent with the marker that predicts the association.
Combining recessive, additive, and dominant individual tests are combined
and used with either the Bonferroni correction method or the MAX test2
with respect to single-test GWAS-based methods. If the gene model behind
the associated phenotype is not known, a multiple test procedure could have
significant advantages compared to single test procedures.
We found that the best statistical method for a study depends on the MOI
gene model associated with the phenotype (diagnosis) in question. Our results
also indicate that a common assumption that the MOI of the locus associated
with the diagnosis is additive will have adverse prediction consequences if the
assumption is incorrect.

Chapter 4 is based on a study that was published in the Journal of Proteomics & Bioinformatics.1
Copyright: 2010 Cooley P, et al. This is an open-access article distributed under the terms of
the Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original author and source are credited. Chapter 3 of
the current publication describes the generation of the synthetic gene database. This section was
removed from Chapter 4. The analysis of the gene data has not changed.
38 Chapter 4

Overall, our results indicate that researchers should consider a statistical

methodology that combines the results of individual MOI-based core tests for
conducting the initial screen in a GWAS. The core tests can be combined into
a single operational test in a number of ways. We identify two: the Bonferroni
procedure and the MAX procedure, each of which produce very similar
statistical power profiles.

Introduction
In this chapter, we examine the statistical methods used to perform GWAS.
GWAS usually apply univariate statistical tests to each gene marker or single
nucleotide polymorphism (SNP) as an initial step. This SNP based test is
statistically straightforward and the testing is done with standard methods
(e.g., 2 tests, regression) that have been studied outside of the GWAS context.
A paper by Kuo & Feingold3 described the most commonly used methods, and
the authors note the use of a compound procedure that combines two or more
statistical tests.
The literature also contains a number of papers that compare statistical
power among subsets of these methods.4,5 However, the question of which
method is best suited to univariate scanning in a GWAS remains an open issue.
The choice of method depends on the match between the true genetic model
underpinning the association and the type of model assumed by the method.
To investigate further, we used a multiple test procedure that combined the
most promising of the methods identified in the literature and applied them
to a set of synthetic marker data with known properties (the Schymick et al.
data set introduced in Chapter 2).6 Our goal was to identify marker properties
that could be linked to optimal methods (with reference to statistical power)
for predicting associations in GWAS. We know from prior studies that the
statistical procedure a researcher chooses influences GWAS prediction
accuracy and that there are specific properties of the underlying markers that
determine the optimization of the procedure choice.3
We also included the important properties that influence the association
prediction accuracy into our synthetic marker data via a Monte Carlo
simulation process, and we link the properties to the influencing marker to
study their individual and collective contributions to association prediction.
A synthetic marker data set allows us to assess the performance of different
statistical methods in a GWAS context. We applied a number of statistical
methods to the simulated data and used their statistical power profiles to
 Genetic Inheritance and Genome-Wide Association Statistical Test Performance 39

evaluate the performance of the methods. We also quantified the relationships

between locus traits and prediction accuracy.
This chapter identifies a number of these properties and quantifies the
loss in power if studies use nonoptimal methods. Similar results have been
reported in earlier studies.4,5 Both studies reinforce the view that the major
influence on prediction accuracy is the gene model of the locus associated with
the diagnosis.
We were particularly interested in assessing the consequences of applying
a statistical method that assumes an inherent additive mode of inheritance
(MOI) property to nonadditive SNP data. Our motivation for this was twofold.
First, the additive MOI model is commonly employed in GWAS; and second,
the answer to the question: what statistical methods should be used to
conduct GWAS? does not have a definitive answer. The best method typically
depends on what MOI gene model has been associated with the associated
diagnosis.
Our results show the major factors that influence association predictions.
They also indicate that a strategy based on predicting associations using
multiple statistical methods can be more accurate; much more accurate if
the governing marker is recessive, than those that assume a single, additive
mode. The multiple test procedure we developed and propose here combines
recessive, additive and dominant MOI-optimal statistical methods, all of
which are derived from the well-known Cochran-Armitage (CA) test. We also
examined different procedures for combining the tests.

Methods
We examined the accuracy of association detection by generating synthetic
data with properties that are known to influence statistical power. We used
a Monte Carlo process to generate the data from a set of random variables
described in Chapter 3. The main purpose of the synthetic data from Schymick
et al.6 is to act as a truth set to assess the performance of commonly used
statistical methods used in a GWAS context. Please see Chapter 3 for the
description of how the data were generated.
The simulated data set that was generated had the following characteristics:
The proportion of cases (controls) that are major homozygotes = 50.3 (63.0)
percent.
The proportion of cases (controls) that are heterozygotes = 39.2 (31.3)
percent.
40 Chapter 4

The proportion of cases (controls) that are minor homozygotes = 10.5 (5.7)
percent.
With MOI distribution:
recessive = 25 percent,
dominant = 25 percent,
additive = 25 percent, and
multiplicative = 25 percent.
We acknowledge that this distribution of MOI traits does not represent
how inheritance traits are distributed in humans. The Online Mendelian
Inheritance in Man (OMIM)7 provides the best source of information on the
MOI distribution (Table 4.1). However, OMIM is disproportionally populated
by genes linked to single Mendelian disorders. Therefore, genes associated with
multifactorial disorders are under-represented in OMIM. Because polygene
influences are assumed to be a major source of additive and multiplicative
SNP behavior, the distribution in Table 4.1 is likely biased. Accordingly, we
populated SNPs in our data with equal MOI representation and acknowledge
that it does not represent the true distribution.

Table 4.1. Distribution of genes in Online Mendelian Inheritance in Man, by mode of

inheritance (MOI)
MOI Frequency
Autosomal Dominant 3,805
Autosomal Additive 12
Autosomal Multiplicative 21
Autosomal Recessive 3,775

The three optimal MOI specific methods are the three variations of the
CA trend test described in Zheng and Gastwirth.8 We also included a fourth,
commonly used individual method, the 2-degrees-of-freedom (2df) genotype
association test. Using the notation in Table 4.2 to define the 23 table of
case-control counts stratified by genotype, a one-tailed test statistic (T2(x)) for
the three variations of the CA trend methods is defined as:

[0,1,2 { xi (s ri - r si)}]2
T2(x) = n . (4.1)
[r s (0,1,2 n {xi xi ni } {0,1,2 (xi ni)2 })]
 Genetic Inheritance and Genome-Wide Association Statistical Test Performance 41

The values represented in equations (4.1) and (4.2) are shown in Table 4.2,
and the value of xi defines the specific test x0 = 0, x2 = 1 and x1 = {0 recessive,
.5 additive, 1 dominant).

Table 4.2. Terms defined in equations (4.1) and (4.2)

AA Aa aa Total
Case r0 r1 r2 R
Control s0 s1 s2 S
Total n0 n1 n2 N
AA = major genotype; Aa = heterozygote genotype; aa = minor genotype.

Under the null hypothesis of no association, T2(x) has an asymptomatic

2 distribution with 1 degree of freedom. Please note that the power of the
trend tests may be affected by the variance estimator used. In particular, the
usual method of combining both cases and controls is not an asymptotically
unbiased estimator of the null variance when the alternative is true. The
authors note that at least two different estimates of the null variance are
available, which are consistent under both the null and alternative hypotheses.
In our calculations, we use a one-sided test.
As an alternative to equation (4.1), it is also possible to use a normally
distributed test statistic, per Li et al.2:

[0,1,2 { xi (s ri - r si)}]
N(x) = n1/2 . (4.2)
[r s (n 0,1,2 {xi xi ni} [0,1,2 {xi ni}2]1/2)]

Under the null hypothesis of no association, N(x) has an asymptotic normal

distribution N(0,1), which suggests a one-tailed test because the synthetic data
assumes that the minor allele conveys the risk of phenotype.
We used eight sample size assumptions with equal numbers of cases and
controls to perform our analysis, with N defined as the number of cases: = 100,
250, 500, 1,000, 2,000, 4,000, 8,000, or 9,500. We estimate statistical power by
statistical method and N using a significance threshold of = 10-7. In a GWAS,
researchers usually perform a single marker analysis as a starting point to
identify SNPs for additional and more comprehensive analysis. This initial pass
creates a large number of statistical tests as well as a high potential for false-
positive predictions, which has caused researchers to perceive the need for a
very low threshold. Accordingly, some studies have used type I threshold levels
on the order of 10-7. 9-11
42 Chapter 4

The multitest statistical methods we used in our comparisons are:

1. The Bonferroni (BON) method, shown in Holm,12 which is a simple form
of the Bonferroni correction results which uses n methods to test for an
association outcome. The correction involves dividing the alpha level by n.
For example, if the association of a given SNP involves using three different
statistical methods, the corrected alpha level () would be /3. This would
ensure that the overall chance of making a Type I error is still less than .
2. The MAX method from Li et al.2 that departs from the Bonferroni method.
Bonferroni assumes that the individual tests are mutually independent;
while Li et al. assume that the individual tests are correlated and incorporate
an approximation to the joint distributions.

Results: Statistical Method Assessment

Table 4.3 presents power estimates by statistical methods and sample size and
is based on a fixed alpha threshold ( = 10-7). All tests are one-sided, and the
tests included in this table are:
The additive 2 version of the CA (CA-A) test, which was the best method
for both additive and multiplicative gene models but was not particularly
effective when applied to recessive MOI data
The Bonferroni test (BON), which combines the 2 version of the recessive,
additive, and dominant MOI specific tests (CA-R, CA-A, CA-D) and
improves on the test performance of the three individual tests when the MOI
gene model is not known
The MAX test due to Li et al.,2 which combines the normal version of the
CA-R, CA-A, and CA-D tests to improve on the test performance of the
three individual tests if the MOI gene model is not known
The dominant 2 version of the CA test (CA-D), which was the best method
for the dominant gene models
The recessive 2 version of the CA test (CA-R), which was the best method
for recessive gene models but the least effective when applied to nonrecessive
MOI data
The 2df genotype test (2df-g), which is never an optimal method for any of
the scenarios we examined; in every scenario, a more powerful alternative
can be identified (see Table 4.3)
 Genetic Inheritance and Genome-Wide Association Statistical Test Performance 43

Table 4.3. Power results, by statistical method and number of cases: additive mode of
inheritance data
N CA-A BON MAX CA-D CA-R 2df-G
100 2.26* 1.73 1.91** 0.62 0.46 0.00
250 13.47* 12.22 12.25** 9.46 5.56 2.41
500 30.00* 28.12** 28.06 24.19 12.63 10.48
1,000 49.91* 48.31** 48.26 45.72 25.11 25.21
2,000 68.80* 67.40** 67.31 64.72 41.08 46.35
4,000 83.00* 81.86 81.94** 80.25 58.24 65.96
8,000 94.30* 93.69 93.72** 92.58 72.36 80.52
9,500 96.07* 95.63** 95.46 95.01 75.66 83.90
2df-G = 2-degrees-of-freedom genotype association test; BON = Bonferroni test; CA-A = autosomal additive
Cochran-Armitage test; CA-D = autosomal dominant Cochran-Armitage test; CA-R = autosomal recessive
Cochran-Armitage test; MAX = MAX combined test.
* Best power score ** Second best power score

Our results indicate that the best method in terms of statistical power is
CA-A, but that little is lost if the BON or MAX method is used instead. Also
there is little difference between the BON and MAX methods. Similarly, the
results in Tables 4.4, 4.5, and 4.6 indicate that the best method in terms of
statistical power for identifying dominant MOI loci is CA-D, and CA-R for
recessive MOI loci. For multiplicative MOI loci, the best method is CA-A.
In all four scenarios, little is lost if the MAX or BON method is used as a
replacement.

Table 4.4. Power results, by statistical method: dominant mode of inheritance gene data
N CA-A BON MAX CA-D CA-R 2df-G
100 0.05 0.07 0.40** 0.13* 0.00 0.00
250 2.09 2.94** 2.88 3.79* 0.00 0.00
500 12.89 15.01 15.02** 16.52* 0.01 1.21
1,000 32.75 34.75 34.84** 36.94* 0.19 12.79
2,000 54.80 56.33 56.55** 57.97* 2.42 32.75
4,000 71.01 73.48** 73.46 74.90* 11.24 54.50
8,000 85.15 86.95 87.18** 88.09* 26.71 71.98
9,500 88.27 90.15** 90.10 91.23* 31.55 76.39
2df-G = 2-degrees-of-freedom genotype association test; BON = Bonferroni test; CA-A = autosomal additive
Cochran-Armitage test; CA-D = autosomal dominant Cochran-Armitage test; CA-R = autosomal recessive
Cochran-Armitage test; MAX = MAX combined test.
* Best power score ** Second best power score
44 Chapter 4

Table 4.5. Power results, by statistical method: recessive mode of inheritance gene data
N CA-A BON MAX CA-D CA-R 2df-G
100 0.00 0.00 0.38** 0.00 0.00* 0.00
250 0.02 0.11 0.68** 0.00 0.19* 0.00
500 0.42 1.70 1.77** 0.00 2.11* 0.01
1,000 2.71 7.88 7.89** 0.01 8.95* 1.04
2,000 9.97 19.91 19.99** 0.13 21.40* 6.50
4,000 21.84 34.29 34.39** 1.50 35.95* 18.45
8,000 34.94 49.35 49.67** 7.17 51.01* 33.14
9,500 38.04 53.07 53.11** 9.32 54.62* 36.73
2df-G = 2-degrees-of-freedom genotype association test; BON = Bonferroni test; CA-A = autosomal additive
Cochran-Armitage test; CA-D = autosomal dominant Cochran-Armitage test; CA-R = autosomal recessive
Cochran-Armitage test; MAX = MAX combined test.
* Best power score ** Second best power score

Table 4.6. Power results, by statistical method: multiplicative mode of inheritance

gene data
N CA-A BON MAX CA-D CA-R 2df-G
100 4.00* 3.35 3.67** 0.98 1.65 0.01
250 17.16* 15.60 15.71** 11.07 8.82 4.74
500 35.65* 33.81** 33.77 27.25 19.4 12.50
1,000 53.84* 52.08** 51.78 48.11 33.65 30.72
2,000 71.59* 70.31** 70.27 66.44 49.68 50.69
4,000 85.12* 84.03** 83.96 81.40 64.32 67.66
8,000 95.10* 94.58** 94.58** 93.36 77.30 82.84
9,500 96.36* 95.99 96.08** 95.30 80.27 86.42
2df-G = 2-degrees-of-freedom genotype association test; BON = Bonferroni test; CA-A = autosomal additive
Cochran-Armitage test; CA-D = autosomal dominant Cochran-Armitage test; CA-R = autosomal recessive
Cochran-Armitage test; MAX = MAX combined test.
* Best power score ** Second best power score

However, regardless of MOI, power is lost if we use the CA-A method. Many
researchers use an additive model as the initial GWAS pass. What if the locus in
question is recessive or dominant? Table 4.7 indicates that although the CA-A
method is the optimal choice (by as much as 2 percent), if the MOI of the locus
is additive or multiplicative, there is a risk of more than 2 percent power loss
if the locus MOI is dominant and as much as 15 percent loss if the MOI of the
locus is recessive.
 Genetic Inheritance and Genome-Wide Association Statistical Test Performance 45

Table 4.7. Power results, by CA-A and MAX methods, for different MOI gene models
Loss using Loss using
N CA-A (D) MAX CA-A CA-A (R) MAX CA-A
100 0.05 0.40 .35 0.00 0.38 .38
250 1.45 2.88 1.43 0.02 0.68 .66
500 12.89 15.02 2.13 0.42 1.77 1.35
1,000 32.75 34.84 2.09 2.71 7.89 5.18
2,000 54.80 56.55 1.75 9.97 19.99 10.02
4,000 71.01 73.46 2.45 21.84 34.39 13.55
8,000 85.15 87.18 2.03 34.94 49.67 14.73
9,500 88.27 90.10 1.83 38.04 53.11 15.07
CA-A (D) = Cochran-Armitage method, additive model version applied to dominant mode of inheritance
(MOI) single nucleotide polymorphism (SNP) data; CA-A (R) = Cochran-Armitage method, additive model
version applied to recessive MOI SNP data; MAX = MAX combined test.

If we knew the distribution of the MOI property, we could assess the overall
risk of using an additive method such as CA-A for GWAS. However, without
a reliable estimate, researchers should exercise caution and apply a procedure
that limits the risk of incorrectly assessing the MOI inherent to the locus-
inducing diagnosis.

Discussion
In the literature, many statistical methods that have been used to perform
GWAS assume a MOI specific hypothesis. Our results confirm the work of
many others.9 In the context of a single-marker scenario, the best method
for predicting associations in recessive SNPs is the CA-R method; the best
method for dominant MOI SNPs is the CA-D method; and the best method
for additive and multiplicative SNPs is the CA-A method.
We also show that the 2df genotype method used in many studiesfor
example, the method used by Schymick et al.6is never optimal because there
are always other methods that provide greater statistical power. This statement
holds regardless of whether the MOI is known a priori or not. We also show
that in the context of a general method to use in the initial GWAS pass,
researchers may encounter adverse consequences if, for example, the MOI of
the operating locus is not consistent with the assumption employed by the
statistical method used. Therefore, 2df appears to be inappropriate to use for
GWAS under any circumstances.
46 Chapter 4

Consequently, we examined the possibility of employing an alternative

procedure that incorporates the three core tests defined above into two
multitest procedures: BON, a Bonferroni corrected procedure, and the MAX
test procedure developed by Li et al.2 Both methods are a composite of three
separate tests (additive, dominant, and recessive models). These procedures
are opposites in that they assume different underlying distributions of the test
statistics. The MAX method assumes that the three tests have dependencies
that can be accounted for, whereas the Bonferroni method assumes that the
three tests are mutually independent. We note that despite these differences,
the two methods produce very similar power profiles.
We generated our results using 1,000 replicates per parameter combination.
Our standard error estimate of power varies from .262 to .315. Consequently,
our 95 percent confidence interval around the mean will be approximately
plus or minus .019. Although we recognize that a larger number of replicates
will improve power precision, we believe that our conclusion would remain
as stated.
In summary, our results lead us to recommend that researchers use a
statistical methodology that combines the results of individual MOI-based
core tests as a statistical method for conducting GWAS rather than a 2df test.
Combining individual methods and comparing the individual and combined
results may help identify the MOI character of the gene. The actual process of
combining the core tests into a single operational test can be done in a number
of ways, all of which produce very similar statistical power profiles.

Chapter References
1. Cooley P, Clark R, Folsom R, et al. Genetic inheritance and genome
wide association statistical test performance. J Proteomics Bioinform.
2010;3(12):321-325.
2. Li Q, Zheng G, Li Z, et al. Efficient approximation of P-value of the
maximum of correlated tests, with applications to genome-wide
association studies. Ann Hum Genet. 2008;72(Pt 3):397-406.
3. Kuo CL, Feingold E. Whats the best statistic for a simple test of genetic
association in a case-control study? Genet Epidemiol. 2010;34(3):246-253.
4. Sasieni PD. From genotypes to genes: doubling the sample size.
Biometrics. 1997;53(4):1253-61.
 Genetic Inheritance and Genome-Wide Association Statistical Test Performance 47

5. Freidlin B, Zheng G, Li Z, et al. Trend tests for case-control studies

of genetic markers: power, sample size and robustness. Hum Hered.
2002;53(3):146-52.
6. Schymick JC, Scholz SW, Fung HC, et al. Genome-wide genotyping in
amyotrophic lateral sclerosis and neurologically normal controls: first
stage analysis and public release of data. Lancet Neurol. 2007;6(4):322-8.
7. McKusick-Nathans Institute of Genetic Medicine, Johns Hopkins
University School of Medicine. Online Medelian Inheritance in Man
(OMIM). 2016 [cited 2016 Feb 11]; Available from: http://www.ncbi.nlm.
nih.gov/omim
8. Zheng G, Gastwirth JL. On estimation of the variance in Cochran-
Armitage trend tests for genetic association using case-control studies.
Stat Med. 2006;25(18):3150-9.
9. Iles MM. Effect of mode of inheritance when calculating the power of a
transmission/disequilibrium test study. Hum Hered. 2002;53(3):153-7.
10. Ziegler A, Konig IR, Thompson JR. Biostatistical aspects of genome-wide
association studies. Biom J. 2008;50(1):8-28.
11. van Es MA, van Vught PW, Blauw HM, et al. Genetic variation in DPP6 is
associated with susceptibility to amyotrophic lateral sclerosis. Nat Genet.
2008;40(1):29-31.
12. Holm S. A simple sequentially rejective multiple test procedure. Scand J
Stat. 1979;6:65-70.
 CHAPTER 5

The Influence of Errors Inherent in

Genome-Wide Association Studies (GWAS)
in Relation to Single-Gene Models
Philip Chester Cooley, Robert F. Clark, and Grier Page

Overview
The influence of genotype and diagnosis errors present in genome-wide
association studies (GWAS) was assessed by analyzing a synthetic gene data
set incorporating factors known to influence association measurement. Monte
Carlo methods were used to generate the synthetic gene data that incorporated
factors that influenced including gene inheritance, relative risk levels, disease
penetrance, genotype distribution, sample size, as well as the two error factors
that are the focus of this study. The resulting data set provides a truth set for
assessing statistical method performance and association sensitivity.
Our results quantify the relationship between genotype and diagnosis
error measures and statistical power loss. The connection between these
relationships are understood, but we document their extent. Our results also
demonstrate that for low-risk nonrecessive loci, sample sizes in the range of
1,000 to 2,000 cases will achieve 80 percent power thresholds for type-I error
levels of 10-8, even with realistic genotype and phenotype error assumptions.
Nevertheless, increasing sample size is a viable method of compensating for
power loss caused by genotype and diagnosis errors. Our estimates indicate
that sample sizes should be increased by 20 to 40 percent, depending on the
gene inheritance model assumed.

Chapter 5 is based on a study that was published in the Journal of Proteomics & Bioinformatics.1
Copyright: 2011 Cooley P, et al. This is an open-access article distributed under the terms of
the Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original author and source are credited. Chapter 3 of
the current publication describes the generation of the synthetic gene database. This section was
removed from Chapter 5. The analysis of the gene data has not changed.
50 Chapter 5

Introduction
More than 2,306 human GWAS have examined more than 1,000 diseases
and traits and found more than 1,200 single nucleotide polymorphism (SNP)
associations.2 With improved genotyping technologies and the growing
number of available markers, case-control GWAS have become a key tool
for investigating complex diseases. Because GWAS have become a standard
primary investigative tool, researchers need to be aware of how errors
influence their studies and how to overcome or compensate for them. The
initial step in a GWAS is to apply univariate statistical tests for each SNP in
the data set. Applying the tests is statistically straightforward and uses several
standard approaches (e.g., 2 tests, regression methods).
Studies on the consequences of genotype error have led to a modest number
of investigations in the statistical genetics literature. Gordon and colleagues
investigated the effects of three published models of genotyping errors on the
2df genotype 2 test.3 In another study, Gordon and colleagues described a
statistical power calculator (PAWE-3D) that produces power and sample size
calculations that can support study designs for GWAS and compute power
and/or sample size requirements for a specified significance level.4 Zheng
and Tian, as well as Edwards and colleagues contributed to the development
of PAWE.5,6 Gordon and colleagues further analyzed the influence of both
random phenotype and genotype misclassification errors on statistical power
contrasting the Cochran-Armitage additive test (CA-A) with the 2-degrees-of-
freedom (2df) genotype test and concluded that the CA-A is more powerful.7
Ahn and colleagues addressed the effect of different types of genotyping
errors on statistical power in GWAS.8 Although their prior work focused on
non-differential genotype error rates, this study considered errors in each
of the three bi-allelic genotypes differentially. The methods were based on
a Taylor-series expansion of a noncentrality parameter of the asymptotic
distribution of the trend test. In a follow-up study, Ahn and colleagues
extended their work by developing a closed form analytic procedure for both
the 2df genotype and the CA-A tests.9 They reported that misclassifying the
heterozygote genotype is particularly detrimental when using the Cochran-
Armitage recessive trend test (CA-R) on data from a recessive mode of
inheritance (MOI) model.
Although the accuracy of the genotyping process has improved, data
errors still occur. Hao and colleagues reported an overall 0.5 percent error
rate imputation process, but they also reported a 2 percent error rate in
 The Influence of Errors Inherent in Genome-Wide Association Studies 51

underrepresented subpopulations.10 Miclaus and colleagues examined

genotype calling algorithms on HapMap samples and found that different
algorithms can produce genotyping errors that influence downstream
genotype calls.11 They reported a 2 to 3 percent error estimate attributable to
the genotype-calling algorithm. Laurie and colleagues estimated genotyping
error rates from duplicate sample discordance rates from addiction and lung
cancer projects genotyped on Illumina Human1Mv1_c and HumanHap550-
2v3_b arrays by the Center for Inherited Disease Research.12 The investigators
calculated genotyping error rates on the order of 10-4, which corresponds to
mean completion call rates of 99.7 and 99.8 percent, respectively, for the two
projects. If study samples are not duplicated, as in the type-2 diabetes project,
but with multiple replicates of the HapMap control sample, discordant rates of
1-4 10-3 lead to completion rates of 99.6 to 99.7 percent.
Phenotypic misclassification errors are also a source of bias and can reduce
the power to detect a statistical association between a phenotype and a specific
allele.13,14 Edwards and colleagues presented a quantification of the effect
of phenotypic error on power and sample size calculations for case-control
genetic association studies between a marker locus and a disease phenotype.6
Their process is specific to the standard case-control method used commonly
in epidemiology, from which they develop a process that quantifies power
loss and minimum sample size requirements in the presence of phenotypic
errors. Barendse described the effect of investigator measurement-error
on the phenotypesthe error was significant when looking at quantitative
traits.14 When the traits were coded as affected or unaffected, the error effect
sizably decreased (from 14.5 to 5.3 percent). For many diseases, the interrater
agreement for disease diagnosis can be quite low. For example, the sensitivity
and specificity of predeath Alzheimers diagnosis with most post mortem
autopsies can be as low as 0.83 and 0.84 respectively.15
Gene traits that influence prediction accuracy have been also reported
in other studies. For example, Sasienis work as well as that of Freidlin and
colleagues has demonstrated that the phenotype MOI model was a major
influence on association prediction accuracy.16,17
To provide additional insight into the influence of genotype and diagnosis
errors affecting the accuracy of the phenotype measure in a GWAS, we ran
simulations with synthetically generated data. We focused on assessing
the impact on statistical power caused by the influence of these two often-
overlooked errors. Our simulations demonstrated that genotype (even at low
52 Chapter 5

error rates) and phenotype (diagnosis) errors produce substantial power losses
for all MOIs, with significant power losses for recessive MOIs. Because GWAS
involving recessive loci have additional power requirements relative to other
MOI types, researchers need to address these requirements in developing
appropriate sample sizes for their studies.

Methods
Our approach is based entirely on a simulation framework. Chapter 3
describes in detail the data generation method that was used to produce our
synthetic data and how we used these data to assess the influence of errors on
statistical power loss.
We developed our assessments by analyzing a data set of synthetic gene data
that incorporates factors that we know influence association measurements in
GWAS. These include phenotypic errors (i.e., those caused by improper disease
diagnosis) and genotype errors (e.g., those caused by incorrect genotype
calls). We employed Monte Carlo methods to generate simulated gene data
that we analyzed to assess the influence of the individual factors on statistical
power in the context of GWAS. There are two advantages to using simulated
data. First, the association-affecting factors are isolated and can be linked to
the affecting locus. Second, we can choose any specific statistical method to
perform the association assessment. The simulated data set provides a truth
set for assessing the role of statistical methods on association sensitivity and
highlights the particular role of errors in disease diagnosis and incorrect
genotype assignments.

Results
Data Set Summary
Using the methods described in Chapter 3, we generated a synthetic gene
simulated data set with the following characteristics:
The proportion of cases (controls) that are major homozygotes = 50.3 (63.0)
percent.
The proportion of cases (controls) that are heterozygotes = 39.2 (31.3) percent.
The proportion of cases (controls) that are minor homozygotes = 10.5 (5.7)
percent.
With MOI distribution:
 The Influence of Errors Inherent in Genome-Wide Association Studies 53

recessive = 25 percent,
dominant = 25 percent,
additive = 25 percent, and
multiplicative = 25 percent.
Although this distribution of MOI traits does not represent a true
distribution, we currently know of no accurate way to obtain such a
distribution. Consequently, although we gave each of the four MOI traits equal
representation in the simulated data, we confined our examinations to within-
MOI assessments.

Factor Assessment
To check the influence of the factors (relative risk, penetrance, MOI, genotype
error rates, phenotype error rates) built into the data for each experiment, we
fitted the three distinct models to the four MOI subsets of the data: that is,
if the index i represents the ith (of 12) experiments, where the model can be
described as:

-log10(pi) = 0 + 1ni + 2dpi + 3ErrPi + 4ErrGi +

5 i + 6Gi(5.1)

With a single exception, the signs of the estimated phenotype error term
(3) and genotype error term (4) coefficients were consistent and the
magnitudes of the estimated coefficients similar. All estimates were significant
at the 10-7 level, suggesting that the factors incorporated in the simulated data
are all significant.

Error Analyses
To simulate the association estimation process in a GWAS experiment, we
applied three variations of the Cochran-Armitage (CA) trend test to each of
the 1,000 replicates of the 3,456 possible data subsets. Each of the variations of
the CA tests used a distinct genotype score vector: [0,0,1] for recessive, [0,0.5,1]
for additive and [0,1,1] for dominant. We applied each of the tests to all of the
replicates in each of the data subsets. This process allowed us to show that the
optimal strategy for maximizing statistical power is MOI specific. This strategy
posits that the recessive version (CA-R) be used to estimate associations
involving recessive loci data, the dominant version (CA-D) for the dominant
loci data, and the additive version (CA-A) for both additive and multiplicative
54 Chapter 5

loci data. This strategy is cited by others for single-gene models.18 Cooley and
colleagues provided a similar assessment and identified a multiple test strategy
that combines the three tests into an overall score that has merit if the MOI of
the causative loci is not known.19 For the assessment in this study, we assumed
that the MOI is known and selected the best statistical method to measure the
association. Consequently, our results tended to be optimistic.
Error rates of 0, 2, and 5 percent are incorporated into the simulated data
for the phenotype and 0, .5, and 1 percent for the genotype. Our approach
combines the three risk levels (mean risk = 1.3), three penetrance levels
(mean penetrance = 0.4), and groups the data into a with error (mean
phenotype error = 3.5 percent, and mean genotype error 0.75 percent) and
without error strata. Also, we also stratified the analysis by MOI. Figures
5.1a through 5.4b identify the four MOI-specific results. Each figure includes
a 0.75 percent genotype error curve, a 3.5 percent phenotype error curve, a
curve that includes both error sources, and a curve generated without either
source of error. Figure 5.1 presents the recessive loci analysis. The impact of a
0.75 percent average genotype error rate and a 3.5 percent average diagnosis
error rate with respect to power loss for recessive loci is nontrivial. However
each profile exhibits distinct behavior. The effect of the phenotype error
increases with N and peaks at N = 4,000 cases, whereas the genotype error
effect is constant across all N. Also observed at the peak is a genotype impact
of 6.04 percent power loss per 1.0 percent genotype error and a power loss
of 3.03 percent power loss per 1.0 percent phenotype error. Note that with
< 10-8 as the significance threshold, an 80 percent power target is far from
being realized even with N = 5,000 cases and controls.
Figures 5.2, 5.3, and 5.4 present the error effects of the dominant, additive,
and multiplicative loci respectively. All three figures indicate that power
loss is nontrivial for the MOI categories they represent but that the effect is
substantially less than recessive modes.
The pattern of the power profile for dominant loci is in sharp contrast to
the recessive loci profile. The diagnosis error pattern is constant across N for
dominant locithe recessive loci show an increasing pattern. The genotype
patterns are also different. As N increases, the power differences decline for the
dominant loci, whereas the patterns are constant for recessive loci.
The additive and the multiplicative loci show similar error profiles. The
impact of a 3.5 percent diagnosis error and the impact of a 0.75 percent
genotype error have a similar quantitative impact. Both have a declining power
loss as N increases.
 The Influence of Errors Inherent in Genome-Wide Association Studies 55

Figure 5.1a. The impact of genotype and diagnosis errors on power:

recessive loci
80

70
Statistical power (percent)

60

50 Perr = 3.5%
Gerr = .75%
40
No errors
30 Both errors

20

10

0
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Gerr = genotype error level Both errors = both genotype and phenotype errors
Perr = phenotype (diagnosis) error level No errors = neither error

Figure 5.1b. Power loss: total, genotype, and diagnosis: recessive loci
15

12
Statistical power (percent)

9 Loss diagnosis
Loss genotype
6 Loss total

0
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Loss genotype = power difference between Gerr = .75 and no errors assumption.
Loss phenotype (diagnosis) = power difference between Perr = 3.5 and no errors assumption.
Loss total = power difference between both Perr = 3.5 & Gerr = .75 and no errors assumption.
56 Chapter 5

Figure 5.2a. The impact of genotype and diagnosis errors on power:

dominant loci
100

90
Statistical power (percent)

80
Perr = 3.5%
Gerr = .75%
70
No errors
Both errors
60

50

40
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Gerr = genotype error level Both errors = both genotype and phenotype errors
Perr = phenotype (diagnosis) error level No errors = neither error

Figure 5.2b. Power loss: total, genotype, and diagnosis: dominant loci
7

6
Statistical power (percent)

4 Loss diagnosis
Loss genotype
3 Loss total

0
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Loss genotype = power difference between Gerr = .75 and no errors assumption.
Loss phenotype (diagnosis) = power difference between Perr = 3.5 and no errors assumption.
Loss total = power difference between both Perr = 3.5 & Gerr = .75 and no errors assumption.
 The Influence of Errors Inherent in Genome-Wide Association Studies 57

Figure 5.3a. The impact of genotype and diagnosis errors on power:

additive loci
100

90
Statistical power (percent)

Perr = 3.5%
80
Gerr = .75%
No errors
70
Both errors

60

50
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Gerr = genotype error level Both errors = both genotype and phenotype errors
Perr = phenotype (diagnosis) error level No errors = neither error

Figure 5.3b. Power loss: total, genotype, and diagnosis: additive loci
7

6
Statistical power (percent)

4 Loss diagnosis
Loss genotype
3 Loss total

0
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Loss genotype = power difference between Gerr = .75 and no errors assumption.
Loss phenotype (diagnosis) = power difference between Perr = 3.5 and no errors assumption.
Loss total = power difference between both Perr = 3.5 & Gerr = .75 and no errors assumption.
58 Chapter 5

Figure 5.4a. The impact of genotype and diagnosis errors on power:

multiplicative loci
100
Statistical power (percent)

90

Perr = 3.5%
Gerr = .75%
80
No errors
Both errors

70

60
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Gerr = genotype error level Both errors = both genotype and phenotype errors
Perr = phenotype (diagnosis) error level No errors = neither error

Figure 5.4b. Power loss: total, genotype, and diagnosis: multiplicative loci
7

6
Statistical power (percent)

4 Loss diagnosis
Loss genotype
3 Loss total

0
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Loss genotype = power difference between Gerr = .75 and no errors assumption.
Loss phenotype (diagnosis) = power difference between Perr = 3.5 and no errors assumption.
Loss total = power difference between both Perr = 3.5 & Gerr = .75 and no errors assumption.
 The Influence of Errors Inherent in Genome-Wide Association Studies 59

In summary, the four figures indicate that the genotype error versus the
diagnosis errors effects vary by MOI. For the recessive MOI, a 3.5 percent
diagnosis error has a larger impact than a 0.75 percent genotype error. This
result is reversed in the dominant MOI scenarios (Figure 2). The additive and
the multiplicative MOI scenarios represented in Figures 3 and 4 indicate that
a 0.75 percent genotype error is comparable in effect to a 3.5 percent diagnosis
error with respect to power loss.
These results are summarized in Table 5.1, which displays the power loss
for the smallest sample size (N = 500 cases) and the largest sample size (N =
5,000 cases). For example, row R (recessive) of Table 5.1 illustrates that error
loss due to genotype errors at N = 500 and N = 5,000 is flat, but that error loss
due to diagnosis error increases dramatically from N = 500 to N = 5,000 and
dominates the total error profile. The power loss pattern changes for the other
three MOIs where error loss patterns for both genotype and diagnosis sources
decline from N = 500 to N = 5,000.

Table 5.1. Maximum power loss due to genotype or diagnosis error

Genotype Errors Diagnosis Errors Both Errors
MOI N = 500 N = 5,000 N = 500 N = 5,000 N = 500 N = 5,000
R 3.86 3.96 3.11 9.98 5.23 13.45
D 5.23 3.43 1.21 1.06 6.46 3.26
A 4.57 1.27 1.97 1.36 6.51 2.69
M 4.09 0.98 1.81 1.40 5.58 2.35
A = additive; D = dominant; MOI = mode of inheritance; M = multiplicative; R = recessive.

We also examined the simultaneous influence of relative risk and error

effects on statistical power. As above, we analyzed our data set using all three
penetrance levels (mean penetrance = 0.4), but we also stratified the curves
by the low- (1.15), medium- (1.3) and high-risk (1.45) categories. Figure 5.5
displays the combined (genotype plus diagnosis) error effects for the three risk
categories using the CA-A method applied to the additive MOI data. Similar
curves can be generated for the dominant and multiplicative scenarios. This
figure suggests that, for additive inheritance scenarios, researchers can predict
associations in the context of GWAS with a type-I error threshold of < 10-8
and still achieve a power level greater than 80 percent. This statement applies
to low-risk loci even when diagnosis errors are 3.5 percent and genotype errors
are 0.75 percent.
60 Chapter 5

Figure 5.5. Genotype error: recessive mode of inheritance, by risk level

and N
100

80
Statistical power (percent)

1.15 no errors
1.15 both errors
60
1.3 no errors
1.3 both errors
40
1.45 no errors
1.45 both errors
20

0
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Loss genotype = power difference between Gerr = .75 and no errors assumption.
Loss phenotype (diagnosis) = power difference between Perr = 3.5 and no errors assumption.
Loss total = power difference between both Perr = 3.5 & Gerr = .75 and no errors assumption.

Figure 5.6. Phenotype error: recessive mode of inheritance, by risk level

and N
100

80
Statistical power (percent)

1.15 no errors
1.15 both errors
60
1.3 no errors
1.3 both errors
40
1.45 no errors
1.45 both errors
20

0
500 1,000 1,500 2,000 2,500 3,000 4,000 5,000
N (number of subjects)

Loss genotype = power difference between Gerr = .75 and no errors assumption.
Loss phenotype (diagnosis) = power difference between Perr = 3.5 and no errors assumption.
Loss total = power difference between both Perr = 3.5 & Gerr = .75 and no errors assumption.
 The Influence of Errors Inherent in Genome-Wide Association Studies 61

Figure 5.6 shows the same results for the recessive scenario. In this recessive
scenario, the likelihood of achieving an 80 percent power level is low and is
only possible for high-risk loci in the absence of genotype and diagnosis error
with a sample size N larger than attempted by our simulation experiments.

Summary/Discussion
We examined the influence of genotype and diagnosis errors that affect the
accuracy of association predictions in a GWAS and focused on assessing the
effect on statistical power loss caused by the influence of these two sources
of error. Our findings are MOI specific and indicate that both sources of
error can adversely affect power levels. This outcome is more pronounced
for recessive MOI and low-risk loci, which is common knowledge. What our
study shows is that the error magnitude depends on a variety of factors in
addition to MOI, especially relative risk and sample size; our study quantifies
this magnitude and indicates the significance of this impact. This loss can be
compensated for by increasing sample sizes. Gordon and colleagues reported
that a 1 percent increase in genotype error rates requires an increase in sample
size of 2 to 8 percent, which they also noted depends on the MOI scenario.3
Our estimates are much higher than those reported by Gordon and colleagues
and are based on achieving a power threshold of 80 percent. Using the additive
model, results at N = 1,000 (assuming no genotype errors) exceeds the 80
percent threshold (80.6 percent). Introducing a 1 percent genotype error,
power drops to 75 percent. An additional 405 cases are needed to compensate
for this loss to restore an 80.6 percent power level, which is a 40.5 percent
increase in sample size. Please note that we are not suggesting that 1 percent
error is standard operating procedure. In fact, genotype errors are improving
with the introduction of each new technology,and currently are likely less than
0.5 percent. Table 5.2 presents these results for all MOIs for both genotype and
diagnosis errors.

Table 5.2. Percent sample size increase to restore power caused by a 1 percent
genotype or diagnosis error
MOI Genotype percentage Diagnosis percentage
R 57.2 35.9
D 40.1 19.7
A 40.5 20.9
M 40.2 18.9
A = additive; D = dominant; MOI = mode of inheritance; M = multiplicative; R = recessive.
62 Chapter 5

In summary, our results quantify the relationship between genotype and

diagnosis error measures and statistical power loss. These relationships are
understood, but we document their extent. Our results also assume that
we know the MOI of the locus being analyzed; therefore, our results will
understate the true power loss and the compensating sample size increases.
Our results also demonstrate that for low-risk nonrecessive loci, sample sizes
in the range of 1,0002,000 cases will achieve 80 percent power thresholds for
typeI error levels of 10-8 even with realistic genotype and phenotype error
assumptions.
However, the recessive loci model remains problematic. Desirable power
thresholds for moderate risk levels can only be realized with sample sizes in the
tens of thousands, further complicated by accounting for power loss as a result
of genotype and diagnosis errors.

Chapter References
1. Cooley P, Clark RF, Page G. The influence of errors inherent in genome
wide association studies (GWAS) in relation to single gene models. J
Proteomics Bioinform. 2011;4:138-144.
2. Burdett T, Hall P, Hasting E, et al. The NHGRI-EBI Catalog of published
genome-wide association studies. 2015 [cited 2015 Nov 2]; Available from:
www.ebi.ac.uk/gwas
3. Gordon D, Finch SJ, Nothnagel M, et al. Power and sample size
calculations for case-control genetic association tests when errors are
present: application to single nucleotide polymorphisms. Hum Hered.
2002;54(1):22-33.
4. Gordon D, Haynes C, Blumenfeld J, et al. PAWE-3D: visualizing power
for association with error in case-control genetic studies of complex traits.
Bioinformatics. 2005;21(20):3935-7.
5. Zheng G, Tian X. The impact of diagnostic error on testing genetic
association in case-control studies. Stat Med. 2005;24(6):869-82.
6. Edwards BJ, Haynes C, Levenstien MA, et al. Power and sample size
calculations in the presence of phenotype errors for case/control genetic
association studies. BMC Genet. 2005;6:18.
7. Gordon D, Haynes C, Yang Y, et al. Linear trend tests for case-control
genetic association that incorporate random phenotype and genotype
misclassification error. Genet Epidemiol. 2007;31(8):853-70.
 The Influence of Errors Inherent in Genome-Wide Association Studies 63

8. Ahn K, Haynes C, Kim W, et al. The effects of SNP genotyping errors

on the power of the Cochran-Armitage linear trend test for case/control
association studies. Ann Hum Genet. 2007;71(Pt 2):249-61.
9. Ahn K, Gordon D, Finch SJ. Increase of rejection rate in case-control
studies with the differential genotyping error rates. Stat Appl Genet Mol
Biol. 2009;8:Article25.
10. Hao K, Chudin E, McElwee J, et al. Accuracy of genome-wide imputation
of untyped markers and impacts on statistical power for association
studies. BMC Genet. 2009;10:27.
11. Miclaus K, Chierici M, Lambert C, et al. Variability in GWAS analysis: the
impact of genotype calling algorithm inconsistencies. Pharmacogenomics
J. 2010;10(4):324-35.
12. Laurie CC, Doheny KF, Mirel DB, et al. Quality control and quality
assurance in genotypic data for genome-wide association studies. Genet
Epidemiol. 2010;34(6):591-602.
13. Gordon D. Gene mapping: balance among quality, quantity and cost of
data in the era of whole-genome mapping for complex disease. Euro J
Hum Genet. 2006;14:1147-1148.
14. Barendse W. The effect of measurement error of phenotypes on genome
wide association studies. BMC Genomics. 2011;12:232.
15. Blacker D, Albert MS, Bassett SS, et al. Reliability and validity of
NINCDS-ADRDA criteria for Alzheimers disease. The National Institute
of Mental Health Genetics Initiative. Arch Neurol. 1994;51(12):1198-204.
16. Sasieni PD. From genotypes to genes: doubling the sample size.
Biometrics. 1997;53(4):1253-61.
17. Freidlin B, Zheng G, Li Z, et al. Trend tests for case-control studies
of genetic markers: power, sample size and robustness. Hum Hered.
2002;53(3):146-52.
18. Kuo CL, Feingold E. Whats the best statistic for a simple test of genetic
association in a case-control study? Genet Epidemiol. 2010;34(3):246-53.
19. Cooley P, Clark R, Folsom R, et al. Genetic inheritance and genome
wide association statistical test performance. J Proteomics Bioinform.
2010;3(12):321-325.
 CHAPTER 6

Conducting Genome-Wide Association

Studies (GWAS): Epistasis Scenarios
Philip Chester Cooley, Nathan Gaddis, Ralph E. Folsom, and
Diane Wagener

Overview
In general, genome-wide association studies (GWAS) apply univariate
statistical tests to each gene marker or single nucleotide polymorphism (SNP)
as an initial step. This SNP-based test is statistically straightforward, and the
core tests for assessing the associations are standard methods (e.g., 2 tests,
regression) that have been studied outside of the GWAS context. Kuo and
Feingold describe the most commonly used statistical methods that are applied
to GWAS.2 All tests cited in the chapter are single-locus tests. If the genetic
inheritance properties are not known, we recommend combining two or more
statistical tests.3 In many cases, the SNPs associated with a disease are not
located in a region of DNA that codes for a protein. Instead, they are located
in the large noncoding regions between genes or in intron sequences, which
are edited out of mRNAs prior to translation to proteins. These regions are
presumably sequences of DNA that modify gene expression, but usually their
functions are unknown.4
The popularity of the GWAS approach belies its simplicity and obscures the
important issue of whether a single-gene model can illuminate the biosynthetic
pathways of a phenotype. In the path leading from gene to trait, factors such
as epigenetics, alternate splicing, gene expression levels, and protein-folding

Chapter 6 is based on a study that was published in the Journal of Proteomics & Bioinformatics.1
Copyright: 2012 Cooley P, et al. This is an open-access article distributed under the terms of
the Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original author and source are credited. Minor text
edits were made to the this chapter. The analysis of the gene data has not changed.
66 Chapter 6

processes create a great deal of complexity. Qualitative trait analysis, which

is the GWAS model most commonly reported in the literature, ignores these
factors. As of mid-2015, more than 2,300 human GWAS have examined
more than 210 diseases and traits and have reported more than 1,200 SNP
associations.5 Most of these GWAS employed a single-gene model that
assumes that each locus acts independently of the others.
Many researchers believe that complex diseases involve multiple genes and
their interactions.6,7 Although GWAS have had some success in identifying
genetic variants underlying complex diseases, most existing studies are based
on limited single-locus approaches, which detect SNPs based on their marginal
associations with a qualitative disease diagnosis.
Classical statistical tests derived from case-control experiments involving
two loci that use a Pearson 2 test or logistic regression are commonly used
as single-locus tests for GWAS and can be used in searching for pairwise
interactions. Marchini and colleagues showed that explicitly modeling
interactions between loci for GWAS with hundreds of thousands of markers
is computationally feasible.8 They also showed that simple methods that
explicitly consider interactions can actually achieve reasonably high power
with realistic sample sizes under different interaction models with some
marginal effects, even after adjusting for multiple testing using the Bonferroni
correction. However, the genotype-phenotype scenarios addressed in a study
by Marchini et al. had substantially larger effects than those that we examine
here.8 Specifically, we focus on low-effect locithose with low relative risk of
association with disease diagnosisbecause the evidence suggests they are
common.9 We also focus on theoretical examples of epistasis that are affected
by the mode of inheritance without assuming an additive inheritance model.
An overarching goal of this study was to review the evidence of whether
statistical methods based on single-gene models can effectively identify
genotype-phenotype associations for multigene processes. Detecting such
associations is particularly difficult for genetic variants with modest impacts
on risk. Consequently, our experiments specifically investigated scenarios
involving low-risk genetic variants and assessed whether multigene scenarios
could be a source of the missing heritability observed using single-gene
models.10 We also examined the impact of two recent studies that collaborated
in the development of novel tests for measuring interaction between two
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 67

linked (in epistasis) or unlinked loci.11,12 These studies purport to have higher
powers to detect interaction than classical logistic regression models.
Our investigations demonstrate that for low-effect loci, single-gene models
of association fail to identify many associations because the interacting locus
masks the effect on the index locus. For the scenarios we tested, our results
also support assessments by Wu and colleagues and Ueki and colleagues that
analytical methods that assume statistical interactions between loci are more
powerful than single-loci models.
In this chapter, we will refer to markers as loci, but more broadly, they
could also be viewed as genes, SNPs, or haplotypes.

Epistasis Analysis
One way to extend the single-gene model to accommodate multiple genes
involves studying gene pairs and their epistatic relationships. Epistasis analysis
is the genetic methodology used to identify which genes act in a particular
cellular process or pathway and to establish an order-of-function map that
reflects the sequence in which those genes act. The analysis typically involves
determining for a pair of genes whether the phenotype of a double mutant
resembles that of a single mutant or whether it is a novel phenotype. Knowing
what type of pathway is being investigated can help establish the type of
relationship between the two genes.
Two types of pathways can be defined: substrate-dependent and switch-
regulatory. Substrate-dependent pathways consist of a specific series of positive
reactions, each of which involves some gene product (e.g., an enzyme) acting
on a substrate produced in the previous step in the pathway and ultimately
producing some final outcome. Switch-regulatory pathways consist of genes
encoding negative or positive regulatory factors that alternate between on
and off states depending upon upstream signaling events, thereby affecting
some downstream response. Because substrate-dependent pathways comprise
only positive factors whereas switch-regulatory pathways can comprise both
positive and negative factors, interpreting results from epistatic studies is
typically less complex for substrate-dependent pathways. Therefore, for the
sake of simplicity, this analysis focuses on substrate-dependent pathways.
A number of studies argue that interacting loci may be the norm and not
the exception. For example, Templeton and colleagues report that experience
has revealed that most complex traits depend on more than one locus.13 Their
68 Chapter 6

study focuses on how often interactions among the loci play a significant role
in the mapping from genotype to phenotype, given that the phenotype is
influenced by two or more loci. They discuss a number of candidate scenarios,
including coronary artery disease, in which the ApoE gene has been shown
to affect males and females differently. Even the reported Mendelian trait
sickle-cell anemia is commonly presented as a single nucleotide trait. A study
by Gilbert-Diamond and colleagues indicates that gene-gene interactions
(epistasis) are a significant complicating factor in the search for disease
susceptibility genes.14

Objective
This chapter investigates epistatic interactions in a GWAS context using a
qualitative association model. The purpose of this exercise is to determine
the statistical methods and models that reliably predict associations between
a qualitative phenotype (specifically, a disease diagnosis, coded as case,
for a positive diagnosis, or control, for a negative diagnosis) and a pair of
interacting genes. As with our other work, we use the concept of relative risk,
the ratio of the probability of a positive diagnosis given a specific genotype and
epistatic model (EM) divided by the probability with no risk present (i.e., P).
The value of P is specified exogenously.

Methods
We employed a Monte Carlobased simulation method to generate synthetic
data corresponding to a variety of possible epistatic models for substrate-
dependent pathways. The method takes into account factors known to
influence association measurements in GWAS, including the relative risk
of association, disease prevalence in nonrisk populations, inheritance
properties of the simulated loci, and most important, the epistatic relationship
of the simulated loci. We then analyzed the simulated gene data to assess
the influence of these individual factors on statistical power in the context
of GWAS. There were two advantages to using simulated data. First, the
association-affecting factors were isolated and could be linked to the affecting
locus. Second, we could choose any specific statistical method to perform the
association assessment.
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 69

Epistatic Models of Inheritance

Table 6.1 defines four possible EMs for substrate-dependent pathways, as
described in the literature.15 Let gene1 and gene2 be distinct genes with varying
genotypes that affect the production of a common gene product, P, ultimately
influencing a phenotype (diagnosis of disease). Mutation of gene1 results in
a level of expression X of P and a relative risk a of exhibiting the disease
phenotype. Similarly, mutation of gene2 results in a level of expression Y of
P and a relative risk b of exhibiting the disease phenotype. The phenotype
of the gene1gene2 double mutant varies according to the EM. If gene1 acts
upstream of gene2 in the pathway leading to P (EM 1), the double mutant
exhibits the phenotype of the gene1 single mutant (gene1 is epistatic to gene2).
Conversely, if gene2 acts upstream of gene1 in the pathway leading to P (EM 2),
the double mutant exhibits the phenotype of the gene2 single mutant (gene2
is epistatic to gene1). If gene1 and gene2 function in parallel pathways leading
to P (EM 3), the double mutant exhibits a novel, more extreme level of P
expression, Z, with associated relative risk ab. Finally, if gene1 and gene2 act at
the same step in the pathway leading to P (EM 4), the observed phenotype can
be either one of the phenotypes of the single mutants or a novel phenotype.

Table 6.1. Epistatic models for substrate-dependent pathways

Phenotype of Phenotype of Phenotype of
gene1 single gene2 single gene1gene2
mutation mutation double mutation
gene1
Model 1 gene2 X (a) Y (b) X (a)

gene2
Model 2 gene1 X (a) Y (b) Y (b)

gene1

Model 3 X (a) Y (b) Z (ab)
gene2

gene1, X (a),
Model 4 gene2 X (a) Y (b) Y (b), or
Z (ab)
70 Chapter 6

To simulate the EM scenarios in Table 6.1 in terms of the contributing

locus genotypes, we referred to classical genetics material found in Klug and
Cummings.16 In each of the models, there are either two or three possible
phenotypes. In our scenarios, there are only two phenotypes (a positive or
negative diagnosis), but the risk of a diagnosis depends on the specific pairings
of the genotypes. Table 6.2 outlines the expected risks associated with each
possible combination of the wild-type (A and B) and mutant (a and b) alleles of
gene1 and gene2 for EM1, taking into account the mode of inheritance acting at
each locus. Because EM2 complements EM1, and EM3 and EM4 are subsumed
by EM1, we limited our analysis to EM1. We used Table 6.2 to generate
synthetic data sets representing the various scenarios and then examined our
ability to link (associate) the phenotypes with the contributing genotypes.

Table 6.2. Epistatic model 1 depicted in terms of risk associated with various
genotype combinations
gene1 D D R R
MOI gene2 D R D R
gene1 gene2 Risk () Risk () Risk () Risk ()
AA BB 1 1 1 1
AA Bb b 1 b 1
AA bB b 1 b 1
AA bb b b b b
Aa BB a a 1 1
Aa Bb a a b 1
Aa bB a a b 1
Aa bb a a b b
aA BB a a 1 1
aA Bb a a b 1
aA bB a a b 1
aA bb a a b b
aa BB a a a a
aa Bb a a a a
aa bB a a a a
aa bb a a a a
MOI = mode of inheritance.
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 71

Generation of Epistatic Synthetic SNP Data

The data generation method we used applies only to autosomal genes.
Furthermore, because our simulation process assumed epistatic behaviors
involving two interacting loci, we expect that the findings would apply to genes
exhibiting these types of interactions. We began generating data by considering
disease penetrance. We define P as the prevalence of a specific trait due to
nongenetic factors. We designate a as the risk allele and A as the allele without
risk for gene1. Similarly, we designate b as the risk allele and B as the allele
without risk for gene2. Following the procedure of Iles,17 we can then define
the risk of disease as the ratio of the probability of a case given a and/or b
divided by the probability of a case given no risk allele, which is P:

= Pr(case /a, b) / P . (6.1)

Generating the synthetic data set was straightforward, using the

relationships between P and risk for the different epistatic categories. Initially,
we assigned values to the following variables:
n = the target number of cases and controls in a given experiment;
P = the disease prevalence in subjects without genetic risk of a diagnosis;
a , b = the relative risks (1.10, 1.15); and
G = {g1, g2, g3}, a set of genotype distributions obtained from actual SNP
data.18
Our general strategy was to randomly select a genotype and assign a relative
risk (a, b) based on Table 6.2. Using the prevalence (P) assumption, we then
assigned a case or control code (1, 0). A detailed description of the process
follows:
1. Using the master genotype distribution G, draw at random a genotype (g1,
g2 or g3) for gene1.
2. Repeat this process for gene2; that is, draw at random a genotype (g1, g2 or
g3) for gene2.
3. Using Table 6.2, select the risk value of a case for the epistatic model being
considered.
4. Based on and P, define the probability of a case to be

x = P . (6.2)
72 Chapter 6

5. Using the estimate of x from equation (6.2), assign a case (0) or control
(1) designation at random. Note that using the 12 different EM/MOI
combinations outlined in Table 6.2 for EMs 13, cases should be linked to
both genetic loci, and this association should be identifiable via appropriate
statistical procedures. Disease risk depends on specific and unknown
disease mechanisms. A relative risk of 1.7 is considered strong and is
associated with positive replication.19 However, a risk of 1.3 is considered to
be a realistic assumption for complex diseases.20 However, many instances
of risk < 1.1 are reported in the literature. We limited our focus to a relative
risk range of 1.10 to 1.25 and were particularly interested in cases with low
relative risk. Note that implicit in equation (6.2) is a definition of prevalence
as the proportion of cases that are present where no genetic risk is assumed.
6. Continue the process until n1 cases and n2 controls have been generated
(note that in this example n1 = n2, but the procedure can be tailored to
specific n1/n2 targets).

Statistical Models
Using the assumptions presented in Table 6.2, we generated 1,000 replicates
of genotypic and phenotypic data for each MOI pair for EM 1 using different
sample sizes and risks. We then investigated the power of different statistical
models to detect genotype-phenotype associations. We analyzed models that
test each gene independently for association with the phenotype and models
that test pairs of genes with and without interaction terms for association.
Single-Gene Methods: Cochran-Armitage Trend Test. The Cochran-Armitage
(CA) trend test is often used as a genotype-based test for case-control genetic
association studies, as described by Purcell and colleagues.21 More generally,
it is used in categorical data analysis to detect the presence of an association
between a variable with two categories (e.g., a diagnosis) and a variable with k
categories (e.g., a genotype). The CA trend test modifies the chi-square test to
incorporate a suspected ordering in the effects of the k categories of the second
variable. For example, one could order the number of mutated alleles as zero,
one, and two and conjecture that the allele effect will not become smaller as
the dose increases.
As described by Zheng and Gastwirth, the CA trend test has three flavors:
dominant (CA-D), recessive (CA-R), and additive (CA-A).22 Using the
notation in Table 6.3 below to define the 23 table of case-control counts
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 73

stratified by genotype, a test statistic (T2(x)) for the three variations of the CA
trend methods can be defined as:

n [0,1,2 {xi (s ri - r si)}]2

T2(x) = . (6.3)
[r s (0,1,2 n {xi xi ni } {0,1,2 (xi ni)2 })]

The variables ri, si and ni in equation (6.3) are defined in Table 6.3. The
variable xi represents the specific test, namely x0 = 0, x2 = 1 and x1 = .5.

Table 6.3. Terms defined in equation (6.2)

AA Aa aa Total
Case r0 r1 r2 R
Control s0 s1 s2 S
Total n0 n1 n2 N
AA = major genotype; Aa = heterozygote genotype; aa = minor genotype.

Under the null hypothesis of no association, T2(x) has an asymptomatic 2

distribution with 1 degree of freedom. We applied the above test to both gene1
and gene2.
Two-Gene Models: Pearson Test. The two-gene, case-control test is derived
from the classical case-control test of epidemiology described by Jewell.23
As with all of the tests, this test compares subjects who have a condition (the
cases) with subjects who do not have the condition but are otherwise similar
(the controls). As in the CA test described previously, the Pearson 2 test is
used in categorical data analysis when testing for the presence of an association
between a variable with two categories (e.g., a positive or negative diagnosis)
and two variables with k categories (e.g., three genotypes). For this test, the
columns are the nine combinations of genotypes and the rows are the two case-
control designations. The central idea is to compute the theoretical frequencies
for all 18 cells from the marginal totals and then test for statistically significant
differences between the theoretical and observed frequencies. This test also
uses a 2 test with (nr 1)(nc 1) = 8 degrees of freedom.
Two-Gene Models: The Method of Wu et al., as Refined by Ueki et al. Wu et al.
developed two novel statistics, refined by Ueki et al., designed to test
interactions between linked or unlinked loci without including the influence
of main effects.11,12
74 Chapter 6

The 2-locus linked test, TIH

IH linked, is defined as:
The two-locus linked
The 2-locus linkedtest, TIHTIHlinked,
test, linked,isisdefined
definedasas:
T = ( I ) / V ( I )
22
H
H H
H

T = ( I ) / V ( I ) ,
IH
IH GH
GH 2 GH
GH
H H

IH GH GH

whereWhere
Where
( I=) [ M 1 M 2]
H
H
22

( I=) [ MP 1P M 2]
GH
GH
H 2

AA AA
P11 P22
GH

M 1 = log 11AA 22AA

P11A P22A
M 1 = log 1212AP2121A
P
P12NNP21NN
P11 P22
M 2 = log 11NN 22NN
P11N P22N
M 2 = log P1212N P2121N
P P
V IGH
HH
( )
= V121 +21V 2
( )
GH

V IGH= V 1 + V 2
H

1 1 1 1 1
=V1 + A + A + A
1 1AA 1A 1A 1A
=V 1 4 n AA P1111A + P1212A + P2121A + P2222A

4 n A P11 P12 P21 P22

1 1 1 1 1
=V2 + N + N + N
1 1NN 1N 1N 1N
=V 2 4 nGG P1111N + P1212N + P2121N + P2222N .

4 nG P11 P12 P21 P22

The second test,
The second test, T unlinked, assumes
IH TIH unlinked, assumes thatthat
thethe
twotwo
lociloci
areare
unlinked
unlinkedand
andisis defined as:
IH
defined asThe second test, TIH unlinked, assumes that the two loci are unlinked and is defined as:
T ( I ) ( I ) ,
22
H H
= H
/V H

( I ) ( I )
IH
IH GH
GH 2 GH
GH

T =
H H
/V
where
IH GH GH

where
where
( )
I = [ M 1]
H 2

( I ) = [ M 1]
GH
H 22
H

(VI( I ) =) [=MV11] .
GH
GH H 2
H 2

GH GH

P11AP22A
A A

M 1 = log 11AA 22AA

Results P11A P22A
M 1 = log P1212AP2121A
P12 P21
This study investigated the effect that polygene interactions have on association
And
predictionsAndin a GWAS context. We used statistical models that appear in
the literature to1 generate
1 1 predictions.
1 1 Some of the models were single-gene,
=V1 + + +
1 1models;
inheritance-specific 1 that
1 1 they assumed that a single additive or
is,
V 1 4n P + P + P + P
AA AA AA AA

=
recessive or dominant
4 n P gene
P
AA

produced
P
11
11 12
12 21
21

P the diagnosis. Other models were

A A A
22
22
A

A 11 12 21 22

inheritance agnostic and assumed that a pair of interacting genes produced the
diagnosis. To implement this investigation, we fixed the risk of the upstream
gene of EM1, gene1, to a low but detectable 1.10 risk level. Simultaneously, we
varied the risk on the downstream gene, gene2, from 1.00 (no risk) to 1.20,
a level that is twice as high as the risk of gene1. Note that a no-risk gene is
inconsistent with the purpose of Table 6.2, which identifies the interactions
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 75

between two genes; however, we use this scenario to describe an endpoint in

our assessment.
Table 6.4 presents a power analysis for gene1 of the simulated EM1 data
using six different statistical tests when the downstream gene has no risk of
disease (the single-gene scenario). The first three columns correspond to
three different versions of the single-gene CA test with different inheritance
assumptions: additive (CA-A), dominant (CA-D), or recessive (CA-R). Each
test was applied to both the upstream and the downstream gene. The last three
columns of Table 6.4 present the results for the two-gene tests.

Table 6.4. Model comparisons for EM 1, N = 12,500, P=.4, a = 1.10, b = 1.00

MOIs/stat CA-A CA-D CA-R TIH
model gene1 gene1 gene1 CC TIH linked unlinked
D-D 59.34 73.63 0.00 50.40 0.00 0.30
D-R 59.35 73.40 0.05 51.70 0.00 0.30
R-D 9.05 0.00 23.70 9.05 0.00 0.20
R-R 7.65 0.00 23.30 9.20 0.05 27.15
CA-A = Cochran-Armitage additive test; CA-D = Cochran-Armitage dominant test; CA-R = Cochran-
Armitage recessive test; CC = case-control association test ; MOI = mode of inheritance.

Table 6.4 indicates the following results:

Single-gene tests work better (from a statistical power perspective) than two-
gene tests for single-gene scenarios (i.e., low risk of disease for gene1, no risk
for gene2), because the additional degrees of freedom used by the two-gene
test provide no benefit when there is no additional risk of disease from the
second interacting gene.
In general, the MOI of the upstream gene determines which test is optimal
(optimal values are bolded in red) with the dominant version of the CA test
being optimal for dominant genes and the recessive version of the CA test
being optimal for recessive genes. Accordingly, the commonly used additive
CA test (CA-A) is never optimal unless the MOI of the gene is additive.3
Unexpectedly, when both genes are recessive, the unlinked refined test is
optimal,12 although the risk of the second locus is null.
In contrast, Table 6.5 presents the results for the case in which the risk from
the downstream gene is twice the risk of the upstream gene.
76 Chapter 6

Table 6.5. Model comparisons for EM 1, N = 12,500, P=.4, a = 1.10, b = 1.20

MOIs/stat CA-A CA-D CA-R TIH
model gene1 gene1 gene1 CC unlinked TIH linked
D-D 0.20 0.45 0.00 80.87 96.56 93.31
D-R 25.45 38.15 0.00 68.65 63.05 53.35
R-D 0.00 0.00 0.00 99.95 61.90 49.50
R-R 1.65 0.05 8.35 81.75 27.30 27.25
CA-A = Cochran-Armitage additive test; CA-D = Cochran-Armitage dominant test; CA-R = Cochran-
Armitage recessive test; CC == case-control association test ; MOI = mode of inheritance.

Table 6.5 indicates the following results:

The Pearson two-gene test is optimal for all MOI submodels of the EM1
model, except when both genes are dominant. In this case, the unlinked
refined Wu et al. test is optimal.
The risk conveyed by gene2 has apparently masked the contribution of
gene1 and the power to predict an association between gene1 and diagnosis
using single-gene models is very low, below 3 percent in all cases except
submodel D-R, where it is below 30 percent. This finding suggests that gene1
is unlikely to be associated with a diagnosis using single-gene models.
Figure 6.1 provides estimates of the statistical power (y-axis) to predict
association between gene1 and diagnosis given different risk values for gene2
(x-axis) for the dominant submodels (D-D and D-R) for EM1. The results
presented in Figure 6.1 correspond to the best single-locus and two-locus tests.
Note that the risk value for gene1 is fixed (1.10). Figure 6.2 presents the same
information for the recessive submodels (R-D and R-R). Both Figures 6.1 and
6.2 identify the crossover risk, which is the risk value at which the single-gene
(optimal model) and the two-gene model have the same power.
These figures suggest that beyond a risk value of 1.051.12 (depending
on the MOI), single-gene tests are no longer as effective (from a power
perspective) as two-gene tests. Furthermore, the power of two-gene tests
improves as the risk of the downstream gene increases, whereas the power of
single-gene tests progressively declines as the risk from gene2 increases.
We repeated the same analysis for EM 1 with the downstream gene (gene2)
risk fixed at 1.1. This time we varied the risk levels of gene1 and applied the
single-gene tests to gene2. Table 6.6 presents the results when the risk from the
upstream gene is null. Again we acknowledge that a no-risk gene is inconsistent
with the purpose of Table 6.2, but this scenario describes an endpoint.
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 77

Figure 6.1. Dominant gene1 analysis: D-D crossover risk = 1.07, D-R
crossover risk = 1.12
100

80
Statistical power (percent)

DD single
60
DD double
DR single
40 DR double

20

0
1.00 1.05 1.10 1.15 1.20
Relative risk
DD single = gene1 dominant, gene2 dominant, single-gene test; DD double = gene1 dominant,
gene2 dominant, two-gene test; DR single = gene1 dominant, gene2 recessive, single-gene test;
DR double = gene1 dominant, gene2 recessive, two-gene test

Figure 6.2. Recessive gene1 analysis: R-D crossover risk = 1.05, R-R
crossover risk = 1.07
100

80
Statistical power (percent)

RR single
60
RR double
RD single
40 RD double

20

0
1.00 1.05 1.10 1.15 1.20
Relative risk
RD single = gene1 recessive, gene2 dominant, single-gene test; RD double = gene1 recessive,
gene2 dominant, two-gene test; RR single = gene1 recessive, gene2 recessive, single-gene test;
RR double = gene1 recessive, gene2 recessive, two-gene test
78 Chapter 6

Surprisingly, the results indicate that single-gene tests do not universally

perform better (in a power sense) than the two-gene tests, even when there is
no risk of diagnosis from gene1. Specifically, if gene1 is recessive, the single-
gene tests do as well as or better than the 8df 2 two-gene tests, but if gene1 is
dominant, the two-gene unlinked refined Wu et al.12 outperforms the single-
gene tests.

Table 6.6. Model comparisons for EM 1, N = 12,500, P=.4, a = 1.00, b = 1.10

MOIs/ CA-A CA-D CA-R TIH
stat model gene2 gene2 gene2 CC unlinked TIH linked
DD 1.85 3.55 0.10 32.12 37.57 26.75
DR 0.10 0.00 0.65 1.15 11.10 8.45
RD 41.20 56.15 0.00 52.60 11.65 8.05
RR 4.20 0.00 14.65 8.10 3.95 23.90
CA-A = Cochran-Armitage additive test; CA-D = Cochran-Armitage dominant test; CA-R = Cochran-
Armitage recessive test; CC = ;case-control association test; MOI = mode of inheritance.

Table 6.7 presents the results for the case in which the risk from the
upstream gene is twice the risk of the downstream gene and demonstrates that
the two-locus, case-control test outperforms all single-gene tests and both of
the refined Wu et al.12 tests in this scenario.

Table 6.7. Model comparisons for EM 1, N = 12,500, P=.4, a = 1.20, b = 1.10

MOIs/ CA-A CA-D CA-R TIH
stat model gene2 gene2 gene2 CC unlinked TIH linked
DD 2.00 4.40 0.10 95.36 38.22 26.70
DR 5.00 0.00 0.55 100.0 11.15 8.05
RD 42.90 57.80 0.00 83.50 11.45 9.50
RR 3.70 0.05 14.00 83.80 3.60 25.90
CA-A = Cochran-Armitage additive test; CA-D = Cochran-Armitage dominant test; CA-R = Cochran-
Armitage recessive test; CC = = case-control association test ; MOI = mode of inheritance.

Figure 6.3 provides estimates of the power (y-axis) to detect association

between gene2 and disease diagnosis given different risk values for gene1 [1.0
(no risk) to 1.20 (double the risk of gene1)]. Note that the risk of gene2 is fixed
at 1.10. Figure 6.3 presents the results for the dominant gene2 submodels (D-D
and R-D). Figure 6.4 presents the results for the recessive gene2 submodels
(D-R and R-R).
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 79

Figure 6.3. Dominant gene2 analysis: D-D crossover risk = 1.0, R-D
crossover risk = 1.12
100

80
Statistical power (percent)

DD single
60
DD double
DR single
40 DR double

20

0
1.00 1.05 1.10 1.15 1.20
Relative risk
DD single = gene1 dominant, gene2 dominant, single-gene test; DD double = gene1 dominant,
gene2 dominant, two-gene test; RD single = gene1 recessive, gene2 dominant, single-gene test;
RD double = gene1 recessive, gene2 dominant, two-gene test

Figure 6.4. Recessive gene2 analysis: R-D crossover risk = 1.05, R-R
crossover risk = 1.07
100

80
Statistical power (percent)

RR single
60
RR double
RD single
40 RD double

20

0
1.00 1.05 1.10 1.15 1.20
Relative risk
DR single = gene1 dominant, gene2 recessive, single-gene test; DR double = gene1 dominant,
gene2 recessive, two-gene test; RR single = gene1 recessive, gene2 recessive, single-gene test;
RR double = gene1 recessive, gene2 recessive, two-gene test
80 Chapter 6

Figure 6.3 and 6.4 are consistent with the results from Figures 6.1 and 6.2
and further suggest that beyond risk value = 1.05, single-gene tests are no
longer as effective from a power perspective as two-gene tests. Furthermore,
the power of two-gene tests improves as the risk of the downstream gene
increases. This is exactly the opposite of the scenario for single-gene tests,
which decline in power as the risk of the downstream gene increases.

Discussion
Our investigation of epistatic scenarios involving low-risk loci indicates that
for a given locus, single-locus tests are not as effective as two-locus tests for
predicting associations if the risk value for a second interacting locus exceeds
1.051.12 (the crossover risk value varies depending on the genetic inheritance
properties of the pair of loci). In general, the power of two-locus tests to detect
associations improves as the risk value of the second locus increases, whereas
the power of single-locus tests progressively declines. Disturbingly, for certain
inheritance models and risk values, a true association between a locus and
phenotype can be entirely masked by a second interacting locus when using
single-locus tests. These finding are not unexpected and are consistent with
previous findings reported by others.6,24,25 However, single-gene models
continue to be used as the core methods for detecting associations in a GWAS
context. Our study is significant in that it provides a more exact estimate of the
risk scenarios in which single-locus models are inferior.
Comparing the performance of the three different two-locus tests evaluated
in this study, in most cases for EM1, the two-locus, case-control Pearson test is
optimal. In certain scenarios (i.e., when both genes have a dominant MOI), the
unlinked Wu et al. test (in which cases and controls are included) as refined
by Ueki et al. is optimal.11,12 This finding is somewhat surprising given that
the modified Wu et al. test measures interaction effects exclusively, whereas
the two-locus, case-control test includes main effects for both loci as well as
interaction effects.
Despite the widespread recognition that single-locus tests are likely to be
inferior to multilocus tests for GWAS of many diseases and phenotypes, an
unresolved issue is how to construct a computationally practical test that takes
into account interactions and enhances the detection of associations between
a specific locus and the phenotype of interest. Wang and colleagues conducted
an empirical comparison of five epistatic interaction detection methods,
including a number of two-pass methods.26 They indicate that each of the five
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 81

methods demonstrates unique utilities, but no single method is optimal, being

simultaneously the most powerful and the most scalable and having the lowest
type-1 error rate in every setting. When users want powerful results and are
not concerned with computation cost, Wang and colleagues cite Zhang and
colleagues TEAM method as the best-performing algorithm.26,27 However,
researchers should note that even when limiting the number of interacting
genes to two, n(n 1) / 2 association calculations are required. For n =
500,0001,000,000, the computational requirements of such an analysis are
daunting but readily parallelizable.

Chapter References
1. Cooley P, Gaddis N, Folsom R, et al. Conducting genome-wide association
studies: epistasis scenarios. J Proteomics Bioinform. 2012;5(10):245-251.
2. Kuo CL, Feingold E. Whats the best statistic for a simple test of genetic
association in a case-control study? Genet Epidemiol. 2010;34(3):246-253.
3. Cooley P, Clark R, Folsom R, et al. Genetic inheritance and genome
wide association statistical test performance. J Proteomics Bioinform.
2010;3(12):321-325.
4. Manolio TA. Genomewide association studies and assessment of the risk
of disease. N Engl J Med. 2010;363(2):166-76.
5. Burdett T, Hall P, Hasting E, et al. The NHGRI-EBI Catalog of published
genome-wide association studies. 2015 [cited 2015 Nov 2]; Available from:
www.ebi.ac.uk/gwas
6. Li J, Horstman B, Chen Y. Detecting epistatic effects in association studies
at a genomic level based on an ensemble approach. Bioinformatics.
2011;27(13):i222-9.
7. Carlson CS, Eberle MA, Kruglyak L, et al. Mapping complex disease loci
in whole-genome association studies. Nature. 2004;429(6990):446-52.
8. Marchini J, Donnelly P, Cardon LR. Genome-wide strategies for detecting
multiple loci that influence complex diseases. Nat Genet. 2005;37(4):413-7.
9. Suhre K, Shin SY, Petersen AK, et al. Human metabolic individuality in
biomedical and pharmaceutical research. Nature. 2011;477(7362):54-60.
10. Manolio TA, Collins FS, Cox NJ, et al. Finding the missing heritability of
complex diseases. Nature. 2009;461(7265):747-53.
82 Chapter 6

11. Ueki M, Cordell HJ. Improved statistics for genome-wide interaction

analysis. PLoS Genet. 2012;8(4):e1002625.
12. Wu X, Dong H, Luo L, et al. A novel statistic for genome-wide interaction
analysis. PLoS Genet. 2010;6(9):e1001131.
13. Templeton AR. Epistasis and complex traits. In: Wolf JB, Brodie EDI,
Wade MJ, editors. Epistasis and the evolutionary process. New York, NY:
Oxford University Press; 2000. p. 41-57.
14. Gilbert-Diamond D, Moore JH. Analysis of gene-gene interactions. Curr
Protoc Hum Genet. 2011;Chapter 1:Unit1 14.
15. Michels CA. Genetic techniques for biological research: a case study
approach. Chichester, United Kingdom: John Wiley & Sons; 2002.
16. Klug W, Cummings M. Concepts of genetics. 7th ed. New York, NY:
Prentice Hall; 2010.
17. Iles MM. Effect of mode of inheritance when calculating the power of a
transmission/disequilibrium test study. Hum Hered. 2002;53(3):153-7.
18. Schymick JC, Scholz SW, Fung HC, et al. Genome-wide genotyping in
amyotrophic lateral sclerosis and neurologically normal controls: first
stage analysis and public release of data. Lancet Neurol. 2007;6(4):322-8.
19. Sladek R, Rocheleau G, Rung J, et al. A genome-wide association study
identifies novel risk loci for type 2 diabetes. Nature. 2007;445(7130):881-5.
20. Ziegler A, Konig IR, Thompson JR. Biostatistical aspects of genome-wide
association studies. Biom J. 2008;50(1):8-28.
21. Purcell S, Neale B, Todd-Brown K, et al. PLINK: a tool set for whole-
genome association and population-based linkage analyses. Am J Hum
Genet. 2007;81(3):559-75.
22. Zheng G, Gastwirth JL. On estimation of the variance in Cochran-
Armitage trend tests for genetic association using case-control studies.
Stat Med. 2006;25(18):3150-9.
23. Jewell NP. Statistics for epidemiology. Boca Raton, FL: Chapman & Hall/
CRC; 2004.
24. Hoh J, Wille A, Zee R, et al. Selecting SNPs in two-stage analysis of disease
association data: a model-free approach. Ann Hum Genet. 2000;64(Pt
5):413-7.
 Conducting Genome-Wide Association Studies (GWAS): Epistasis Scenarios 83

25. Culverhouse R, Suarez BK, Lin J, et al. A perspective on epistasis: limits of

models displaying no main effect. Am J Hum Genet. 2002;70(2):461-71.
26. Wang Y, Liu G, Feng M, et al. An empirical comparison of several recent
epistatic interaction detection methods. Bioinformatics. 2011;27(21):2936-43.
27. Zhang X, Huang S, Zou F, et al. TEAM: efficient two-locus epistasis
tests in human genome-wide association study. Bioinformatics.
2010;26(12):i217-27.
 CHAPTER 7

Assessing Gene-Environment Interactions in

Genome-Wide Association Studies (GWAS):
Statistical Approaches
Philip Chester Cooley, Robert F. Clark, and Ralph E. Folsom

Overview
In this chapter, we address a scenario that uses synthetic genotype case-
control data that are influenced by environmental factors in the context of a
genome-wide association studies (GWAS). The precise way the environmental
influence contributes to a given phenotype is typically unknown. Therefore,
our study evaluates how to approach a GWAS that may have an environmental
component. Specifically, we assess different statistical models in the context of
a GWAS to make association predictions when the form of the environmental
influence is questionable. We used a simulation approach to generate
synthetic data corresponding to a variety of possible environmental-genetic
models, including a main effects only model as well as a main effects with
interactions model. Our method takes into account the strength of the
association between phenotype and both genotype and environmental factors,
but we focus on low-risk genetic and environmental risks that necessitate
using large sample sizes (N = 10,000 and 200,000) to predict associations with
high levels of confidence. We also simulated different Mendelian gene models,
and we analyzed how the collection of factors influences statistical power in
the context of a GWAS. Using simulated data provides a truth set of known
outcomes such that the association-affecting factors can be unambiguously
determined. We also test different statistical methods to determine their
performance properties. Our results suggest that the chances of predicting an
association in a GWAS is reduced if an environmental effect is present and
the statistical model does not adjust for that effect. This is especially true if the

Chapter 7 is based on a study that was published in the RTI Press.1 Minor text edits were made to
this chapter. The analysis of the gene data has not changed.
86 Chapter 7

environmental effect and genetic marker do not have an interaction effect. The
functional form of the statistical model also matters. The more accurately the
form of the environmental influence is portrayed by the statistical model, the
more accurate the prediction will be. Finally, even with very large samples sizes,
association predictions involving recessive markers with low risk can be poor.

Introduction
In recent years, scientists and researchers have increasingly used GWAS to
unravel the genetic factors that influence important phenotypes such as disease
presence and predisposition. The hypothesis GWAS follows is that if genetic
variations are more frequent in people with a given disease, the variations
are likely associated with the disease. In general, GWAS apply univariate
statistical tests to each gene marker or single nucleotide polymorphism (SNP)
as an initial step. This SNP-based test is statistically straightforward, and the
core tests for assessing the associations are standard methods (e.g., 2 tests,
regression) that have been studied outside of and within the GWAS context.
Kuo & Feingold describe the most commonly used statistical methods applied
to GWAS.2 All the tests they cite are single-locus tests.
The popularity of the GWAS approach is testimony to its simplicity;
however, it obscures the important issue of whether a single-gene model
is conducive to unraveling the workings of the biosynthetic pathways of a
phenotype. In the preceding chapter, we demonstrate that if two genes are in
epistasis the likelihood of identifying the weaker (in terms of risk) of the two is
diminished if a single-gene model is used in this context. We would anticipate
a similar finding involving genes and environmental interactions namely that
the risk weaker effect would be dominated by the stronger effect.
Researchers can use classical statistical tests derived from case-control
experiments to determine whether two loci associate in a GWAS context. Both
a Pearson 2 test and tests involving logistic regression can be used to examine
for pair-wise interaction assumptions.
In this study, we focused on low-effect loci with low relative risks of
association with disease diagnosis, because the evidence suggests these are
common.3 Most GWAS report only small changes in disease risk (1.1 to 1.5).
It has also been reported that relative risks underestimate the true risk and the
corresponding effect size.4
Note that this chapter does not assess the multilocus scenario, which is the
focus of Chapter 8. Nor does it account for the scenario involving multiple
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 87

loci that are tested simultaneously. This scenario, also discussed in Chapter 8,
requires an adjustment to the p-value threshold via a Bonferroni correction.5
This simple procedure (dividing the p-value threshold by the number of test)
assumes statistical independence between tests which is not true. Hence the
correction represents an overcorrection, leading to higher-than-necessary
type II error rates. We note that the correction does not affect our type I or
type II assessment because all of our examples have been generated with
positive genetic associations (even if some of these associations are very
difficult to detect). Thus, all scenarios are associated, and the issue is whether
that association can be detected and the statistical procedures that perform
most effectively.
The word risk can have a variety of meanings. In an environmental
context, it means a hazard based on an exposure to a chemical or pollutant
such as tobacco smoke. In another context, risk is interpreted more narrowly
to mean the probability of an adverse consequence (e.g., an adverse event
such as a disease). The term environmental risk in this study is used broadly;
we define it as any process that contributes to a disease diagnosis that is not
genetic in origin. For example, environmental risks can represent exposure to
chemicals or pollutantsor a subjects age.
Our overarching goal was to identify which statistical methods best identify
genotype-phenotype associations when environmental effects also influence
the association. Detecting such associations is particularly difficult for
genetic variants with modest impacts on risk. Consequently, our experiments
specifically investigated scenarios involving low-risk genetic variants and
assessed whether environmental influences with varied levels of risk could
be a source of the missing heritability observed using single-gene models.6
Not surprisingly, our investigations demonstrated that the best statistical
method (with respect to statistical power) depends on whether there are
interactions between the genotype and environmental factors and how well the
specified statistical model matches the environmental effect associated with
the phenotype. In summary, the simulated data set provides a truth set for
assessing the sensitivity of the effect of the statistical method and the predicted
association. Establishing the genotype-to-phenotype connections without
using a simulation approach is difficult to impossible. Although our study
results demonstrate a number of obvious truths, a number of unexpected
results may lead researchers to more powerful statistical approaches that can
establish the validity of the simulation approach.
88 Chapter 7

Background
Many complex diseases (e.g., diabetes, asthma, cancer) are affected in part by
interactions between genes and environmental factors. However, investigators
conducting GWAS typically do not investigate the influence of environmental
factors as part of the GWAS process.
There have been several notable exceptions. For example, a study by Terry
and colleagues showed a significant interaction between smoking status and
the specific gene for lung cancer.7 Another study, by Stern and colleagues
found smoking status to be an effect modifier of the association between
a codon and the risk of bladder cancer.8 Understanding the relationship
between genetic polymorphisms and environmental exposures can greatly aid
investigators in detecting high-risk subgroups in the population and provide
better insight into pathway mechanisms for complex diseases.
Current GWAS methods are designed to detect main effects, that is, direct
associations of an SNP or clusters of SNPs with disease.9,10 In the context of
complex diseases, examining main effects only could miss important genetic
variants specific to subgroups of the population.

Gene-Gene Interaction Studies

Lichtenstein and colleagues studied twins and sought to connect hereditary
factors to the causes of sporadic cancer.11 The study assessed the risks of
cancer at 28 anatomical sites for twin children of a parent who has cancer.
Statistical modeling was used to estimate the relative importance of heritable
and environmental factors in causing cancer at 11 of those sites. A major
finding was that inherited genetic factors make a minor contribution to
susceptibility for most types of neoplasms, indicating that the environment
plays the principal role in causing sporadic cancer. The relatively large effect
of heritability in cancer at a few sites (such as prostate and colorectal cancer)
suggests major gaps in our knowledge of the genetics of cancer.
Another large study, by Pearce and colleagues that also focused on cancer
attempted to link several well-established environmental risk factors for
ovarian cancer and the results of a recent GWAS that identified six variants
that influence disease risk.12 They pooled data from 14 ovarian cancer case-
control studies, and then conducted stratified analyses of each environmental
risk factor to evaluate the presence of interactions for all histological
subtypes. They fit a multivariate model to examine the association between
all environmental risk factors and genetic risk score on ovarian cancer risk.
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 89

The results indicated no strong statistical evidence of interaction between the

six SNPs or genetic risk score and the environmental risk factors on ovarian
cancer risk.
A large bladder cancer study reported by Rothman and colleagues
demonstrated interactions due to smoking using a logistic regression (LR)
adjusted for age.13 This study coded the genotype variable as a count of minor
alleles conforming to our Models 1, 2, and 3 described here. Lindstrom and
colleagues, in a study involving prostate cancer, found no contribution from a
number of environmental factors.14 This study used a number of LR models
similar to those we used in our analysis. Another study by Yu and colleagues
developed a Bayesian framework to investigate the influence of multiple loci
simultaneously on disease risk.15 Yu and colleagues full model consisted of
a standard LR model that treats the genotype variable as a categorical variable
and specifies a main effect with interactions model.
Patel and colleagues used GWAS to examine type 2 diabetes, a second
disease with a strong interplay between environmental and genetic factors.16
Genetic loci discovered through GWAS in this and other studies explained
only a small portion of the disease risk variance; some of the unexplained risk
was likely due to gene-environment interactions. The study suggested that
the adverse effect of several type 2 diabetes loci may be abolished or at least
attenuated by higher physical activity levels or healthy lifestyle, whereas low
physical activity and the typical Western diet may augment it. This study used
data from two surveys from the Centers for Disease Control and Preventions
National Health and Nutrition Examination Survey (NHANES). They used a
GWAS to screen 18 genetic loci and type 2 diabetes for statistical interactions
that were associated with disease. They describe their investigation as an
environment-wide association study (EWAS), and they used data sets from
four cohorts from the NHANES. Because the four cohorts were analyzed
individually, the number of environmental factors varied among them.
The models used by Patel and colleagues were logistic regression examples
that were adjusted for age, sex, body mass index, and race.16 The results
identified eight potential disease geneenvironmental factor interactions. One
interaction (trans--carotene) was particularly significant. The perrisk-allele
effect sizes, after adjusting for age, sex, body mass index, and race for subjects
with low trans--carotene levels, were 40 percent greater than the marginal
genetic effect size of the SNP. They also found a strong interaction between an
90 Chapter 7

SNP and a nutrient found in corn oil, which conveyed a 20 percent higher risk
than the SNP alone did.
A study by Murcray and colleagues performed a general methodological
study that focused on identifying SNPs that demonstrate heterogeneity
between subgroups defined by some environmental exposure.17 They
describe a two-step approach for detecting loci involved in gene-environment
interactions that is performed independently of any initial scans for
main effects. They expanded on the traditional test for gene-environment
interaction in a case-control study by incorporating a preliminary screening
step constructed to efficiently use all available information in the data. They
claim that their two-step approach is more powerful than the standard
test of interaction across a wide range of models and consequently is more
robust to changes in environmental exposure and minor allele frequency
than the traditional one-step test for identifying highly significant SNPs. The
difficulty with most methods, including theirs, is that it is not a data mining
method. The specific environmental factor and the form of that factor have
to be established prior to analysis. This has proven to be a difficulty with our
methods as well. The specific environmental factor or factors to include in the
model greatly affect the power of the tests. Specifically, researchers should use
some combination of the literature and/or data mining activities to establish
the form of the environmental effect model (step function or linear) on the
logistic scale.
A study by Cornelis and colleagues provides a comparative study of
several logistic regressionbased tests of gene-environment (G-E) and GE
interactions.18 All seven methods compared in their paper assumed a log-
additive mode-of-inheritance model for each SNP. This differs from our
methods, in which the mode of inheritance was agnostic. Cornelius et al. do
not identify a preference for any of the seven methods and instead indicate that
preference would depend on the goal of the study. They also explored methods
investigating environment effects only in subjects with a positive phenotype
case (i.e., case-only studies).
Finally, Kraft and colleagues performed a study, which was similar in
content to the Cornelius and colleagues study in that it also focused on log-
additive gene models, that formulated a likelihood ratio test of association
between disease and locus with the possibility that the genetic effect may be
modified by an environmental factor.19 The specific environment model they
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 91

investigated was similar to one of the experiments examined in our study

namely, a chemical spillbecause it was an all-or-nothing type of exposure.
In summary, all of the methods cited previously use logistic regression
models. This is a result of the flexibility of that approach with respect to
treating multiple genetic, environmental, and interaction variables as
simultaneous effects. Also, all of the studies attempted to characterize how
the environmental effects influenced the association outcome. In general, the
effects could be characterized as single environmental exposures that triggered
the risk of association immediately, or as a single risk that accumulated as the
subject aged. Because each of these two types of risks could depend (or not) on
a genetic predisposition, we decided to investigate these four categories of risk
as separate possibilities. Because our study uses synthetic gene-environmental
data with specified (known) risk, we are able to characterize data mining
strategies in terms of their statistical power.

Methods
Overview
We simulated genetic and environmental interactions in a GWAS context using
a qualitative association framework to determine which statistical methods
and models reliably predict associations between a qualitative phenotype
(specifically, a disease diagnosis, coded as case for a positive diagnosis or
control for a negative diagnosis) and a gene paired with an environmental
influence. As with our previous work, the concept of relative risk is the basis
for this investigation.20 We define the genetic relative risk () of a wild-type
genotype to be the ratio of the probability of a positive diagnosis given an
occurrence of a (wild-type) genotype divided by the probability of disease in
the absence of the disease genotype. We also define the environmental risk ()
as the ratio of the probability of a positive diagnosis given an exposure divided
by the probability of a positive diagnosis in unexposed subjects. The values of
and are specified exogenously and vary from low-risk to not-so-low-risk.
We generated 1,000 replicates of simulation data that depended on the two
risk values ( and ) for each of three gene models using a standard Bernoulli
process and analyzed them in terms of the observed power profiles for a low
alpha error ( 10-8). The distribution of the number of alleles per genotype
was randomized across replicates and was based on real data from the study by
Schymick and colleagues used in previous chapters.21 We biased the risk levels
92 Chapter 7

to the low end of the risk continuum because these are more difficult scenarios
and are typical of what has been observed in the literature.3 To support these
low risk levels, we fixed our sample size to N = 10,000 (5,000 cases and 5,000
controls) and N = 200,000 (100,000 cases and 100,000 controls) to determine
whether it is possible to measure associations in low-risk, recessive inheritance
scenarios. Other studies have used smaller values (N = 6,000) for comparable
investigations.13

Generating the Synthetic SNP Data

We derived our data generation method from a study by Iles and Mendelian
concepts of inheritance.22 We specifically incorporated autosomal dominant,
recessive, and additive inheritance patterns into the data. These data also
depend on factors known to influence association measurements in the context
of GWAS. Our simulation process assumes Mendelian-type inheritance
patterns.
Penetrance was defined as the proportion of individuals without the risk
allele who have a definable trait (phenotype). In other words, penetrance was a
genotype-specific probability of being affected with the trait. We designated a
as the risk allele and A as the allele without risk. Generating the synthetic data
set using the relationships between penetrance and risk for different mode of
inheritance (MOI) categories was straightforward; Chapter 3 and Cooley and
colleagues provide additional detail.23
Initially, we supply as input data the following variables:
n = the target number of cases and controls in a given experiment,
P = the disease penetrance,
= the genotype relative risk (1.10, 1.15, 1.20),
= the environmental relative risk, and
the distribution of genotypes, which were drawn at random from a master
set of genotype distributions obtained from real SNP data.23
In screening samples from the master set of genotype distributions, Chan
and colleagues recommend that a minor allele frequency (MAF) threshold
not be applied as a filter.24 They argue that filtering MAFs out of the process
because of low frequencies or to maintain HardyWeinberg equilibrium
deviation has little effect on the overall false positive rate and, in some cases,
filtering MAFs excludes SNPs. The effect of this step is to select a specific
genotype distribution at random from the master distribution.
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 93

From the selected relative risk (), penetrance (P), and MOI assumptions,
we used the formulas in Table 7.1 to assign a case (1) or control code (0). This
step converts the relative risk ratio () into the probability of a case (disease),
given the MOI gene model assumed.

Table 7.1. Relative risk assumptions, by mode of inheritance

Major homozygote Minor homozygote risk Heterozygote risk
Inheritance risk Pr(case/aa) Pr(case/aA)
aa = ____________ Aa = ____________
model AA Pr(case/AA) Pr(case/AA)
Recessive 1 1
Dominant 1
Additive 1 2 1
Multiplicative 1
Pr = probability. = genetic inheritance risk.
Source: Iles (2002).22

This genotype-specific process can be represented by the following logic:

Major homozygote (AA)
If the AA (nondisease) genotype is selected, the probability of a case equals
the disease penetrance, Pj.
Minor homozygote (aa)
aa is the exogenous risk and represents the ratio of two probabilities: the
probability of a case for a minor homozygote divided by the probability (Pr)
of a case for a major homozygote. In other words,
aa = Pr(case/aa) / Pr(case/AA) = x/Pj.  (7.1)
Thus, the probability of a case given the minor genotype is
x = aaPj.  (7.2)
Heterozygote (aA)
By the same argument, the phenotype risk given a heterozygote is
aA = Pr(case/aA) / Pr(case/AA) = y/Pj.(7.3)
Thus, the risk of a case given the heterozygote genotype is
y = aAPj,(7.4)
where aA is the assumed risk factor and Pj is the assumed penetrance.
94 Chapter 7

Implicit in equations (7.1) through (7.4) is a consistent definition of

penetrance defined as the proportion of cases that are present in the major
genotype AA.
Using the estimate of x from equation (7.2) and y from equation (7.4),
we specified a subject as a case (1) or control (0) at random using the four
different MOI models from Table 7.1. For the MOI models that assume an
elevated risk from the minor and the heterozygote genotypes, we would
expect a higher proportion of cases to be more easily identified via the
statistical procedures. Specifying risk depends on known and unknown disease
mechanisms. Some consider a relative risk of 1.7 high and a risk of 1.3 to be a
more realistic assumption for complex diseases.25,26 We limited our focus to
relative risks in the range of 1.1 to 1.2.
Note that we assigned cases and controls so that there would be no
possibility for the introduction of bias. We chose to ignore errors in both
genotype and the phenotype measurements, which in a real experiment
could be a source of bias (we examined both sources of error in an earlier
study20). This process continued until we created n1 cases and n2 controls. We
then applied a set of statistical methods (identified subsequently) to predict
associations, and recorded and tracked the results. For each set of unique
factor combinations (i.e., penetrance, sample sizes, relative risk levels, and
MOI categories), we generated 1,000 replicate experiments.
Exogenously, we specified the genetic inheritance (GI) relative risk of
disease as 1.10, 1.15, and 1.20 and defined it in the overview as the ratio of the
probability of a disease diagnosis for subjects, dividing the wild-type gene by
the probability of disease, based on all genetic and nongenetic causes. We also
defined a second relative risk component based on a specific environmental
exposure (EE). We defined this ratio as the probability of a disease given the
EE divided by the probability of a diagnosis given no EE. In discussion of these
experiments, we use the notation to represent GI and to represent EE.
The form of the EE relative risk can be specified using a variety of
assumptions. In all scenarios, the genetic risk is first used to determine the
phenotype status (case or control). Then the environmental risk calculation
determines whether the phenotype status is altered from control to case
according to the EE assumptions. We assume that the form of the EE
effect is not known but that the specific variable is known. In the following
experiments, we use E = age as a proxy for the different assumed forms of
exposure, and we assign E a value obtained from a uniform distribution of
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 95

30 to 70. The value of E controls the EE risk according to different experiment

designs. The main objective of this assessment is to identify whether one
statistical model outperforms all other models and how much variation occurs
across the different experiments.
For all experiments, we used the GI as described previously.

Experiment 1The Main Effects Model. For the first experiment, half of
the population (selected at random and assigned 50 < E < 71) incurred an
EE relative risk (). The assigned risk value was 1.10, 1.20, 1.30, or 1.40. The
other half of the population (assigned 29 < E < 51) incurred no risk ( = 1.0).
Thus, Experiment 1 simulates a fixed EE. When the determinant risk variable,
E, exceeds a threshold, a positive diagnosis is more likely to occur. This is
identified as the fixed risk, main effects, no interaction model.

Experiment 2The Interaction Effects Model. For the second scenario,

again half of the population (selected at random and assigned 50 < E < 71)
incurred an EE relative risk (). This risk value was 1.10, 1.20, 1.30, or 1.40,
but only if the subject also had a wild-type allele (i.e., a heterozygote or minor
homozygote genotype). The other component of the population (50 < E < 71
and genotype = AA) incurred no EE risk ( = 1.0). Experiment 2 also
simulates a fixed EE but only if the genotype contains a wild-type allele. This is
identified as the fixed-risk, main effect with interaction model.

Experiment 3The Main Effects Log-Linear Risk Model. For the third
scenario, the entire population (randomly assigned 30 E 70) incurred an
EE relative risk () which was related to E in the following manner:
y = (E 30)/40.
= Xy (X to the y power), where X = {1.10, 1.20, 1.30, 1.40).
Experiment 3 simulates a log-linear variable risk model, with larger values of
E conveying additional risk levels. As in Experiment 1, there is no interaction
between the GI and EE risks.

Experiment 4The Interaction Effects Log-Linear Risk Model. The fourth

scenario is the same as the third scenario, but the risk applies only if the
subject has a wild-type allele.
Experiment 4 simulates a variable-risk scenario with larger values of E
conveying higher risk levelsbut only if the genotype contains a wild-type
allele. This is the log-linear variable risk main effect with genotype interaction
model.
96 Chapter 7

For each experiment type, we varied the gene model to determine the
relative power differences across model specification. Overall, Experiment 1
data have a step function relationship to EE and no interaction or difference
in slopes (or EE step heights) across the three genotypes. In contrast, the
Experiment 2 data has a step function relationship with EE where the aa and
aA genotypes have the same slope (step height) but different intercepts. The
AA genotype relationship to the EE is flat or has zero slope (no step up). In
Experiment 3, the relationship to EE is log-linear, with equal slopes for all
three genotypes. Finally, in Experiment 4, the relationship to EE is log-linear;
the aa and aA genotypes have the same slope but different intercepts; and the
AA genotype relationship to the EE is flat, or has zero slope.

Statistical Models
All models tested assumed a logistic regression (LR) specification. This form is
commonly used in association studies involving environmental interactions.21
Table 7.2 shows the variables used in the different models.

Table 7.2. Descriptions of variables used in the logistic regression models

Variable category Name Form Values
Genotype G Continuous 0, 1, 2
Genotype g1, g2 Categorical 0, 1
Environmental E Continuous 3070
Environmental e1 Categorical 0, 1
Interaction g1E Mixed 0, 3070
Interaction g1e1, g2e1 Categorical 0, 1
Notes: G = the number of wild-type alleles for the genotype (0, 1, 2).
g1 = 1 if the subjects genotype is a heterozygote, otherwise g1 = 0.
g2 = 1 if the subjects genotype is a minor or wild homozygote, otherwise g2 = 0.
E = a variate from a uniform distribution (3070) that suggests it is an age.
e1 = an indicator variable set to 0 if E < 50. Otherwise e1 = 1.

The difference between the experiments is straightforward. For subjects

younger than 50 years old, there is no risk from environmental exposure
(i.e., the relative risk = 1.0) in Experiments 1 and 2; subjects older than 50
have an environmental exposure (i.e., the risk is greater than 1.0). However,
in Experiment 2, only subjects aged 50 or older who have a wild-type allele
are assumed to have the assigned risk. The main discriminator between
Experiments 1 and 3 (and Experiments 2 and 4) is the risk characterization. For
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 97

Experiments 1 and 2, the risk is intended to be an all-or-nothing process akin

to a toxic exposure that occurs sometime after the subject reaches age 50. For
Experiments 3 and 4, the risk due to an environmental exposure is present in all
subjects and increases as age increases. Table 7.3 summarizes the experiments.

Table 7.3. Experiment description

Scenario
Experiment Risk type description Exposure action
1 Fixed EE Chemical exposure Risk applies to half of the
population
2 Fixed EE with Chemical exposure Risk applies to the half of the
interaction affects genotype population who have the wild-
type allele
3 Variable EE Advancing age Risk applies to half of the
population and increases with age
4 Variable Advancing age Risk applies to the half of the
EE with affects genotype population who have the wild-
interaction type allele, and risk increases with
age
EE = environmental exposure.

We used three specific statistical models to assess the data generated by the
four experiments. Each assumed an intercept term and had the following form:
Model 1 is a logistic regression model with a single variable genotype (G)
main effect (2df). This is a candidate model if no environmental exposure
were suspected.
Model 2 is a logistic regression mixed main effects and interaction model
(g1, g2, E, g1E, g2E) (6df). This is a fully specified model that assumes
that the environmental exposure is a continuous variable.
Model 3 is also a logistic regression mixed main effects and interaction
model (g1, g2, e1, g1e1, g2e1) (6df). This is the fully specified
categorical model and assumes that the environmental exposure has a
specific (all-or-nothing) categorical variable form.
Table 7.4 summarizes the specific regression models we used in this study.
Note that we initially compared six models. Two were gene-only modelsa 1df
(log-additive test) and a 2df testand four were main effects plus interaction
models. We had two environmental exposure specifications (E and e1) and two
genetic inheritance specifications (G and g1, g2). From the six initial models,
98 Chapter 7

we selected the three models that dominated the other three: M-1, M-2, and
M-3. We dropped the other three models (M-1a, M-2a, and M-3a) from our
assessment.

Table 7.4. Statistical models assessed

Model Main effects Interactions df Test statistic
M-1 G NA 1 LLH[log(, G)] LLH[log()]
M-1a g1, g2 NA 2 LLH[log(, g1, g2)]
LLH[log()]
M-2 g1, g2, E g1E, g2E 5 LLH[log(, g1, g2, E, g1*E,
g2*E)] LLH[log()]
M-2a G, E GE 3 LLH[log(, G, E, GE)]
LLH[log()]
M-3 g1, g2, e1 g1e1, g2e1 5 LLH[log(, g1, g2, e1, g1*e1,
g2*e1)] LLH[log()]
M-3a G, e1 Ge1 3 LLH[log(, G, e1, Ge1)]
LLH[log()]
df = degrees of freedom; NA = not applicable; LLH = log-likelihood. = the logit scale intercept for the line
relating environmental exposure (EE) to the log-odds risk among those subjects with the nondisease
genotype AA.
Notes: G = the number of wild-type alleles for the genotype(0, 1, 2).
g1 = 1 if the subjects genotype is a heterozygote, otherwise g1 = 0.
g2 = 1 if the subjects genotype is a minor or wild homozygote, otherwise g2 = 0.
E = age, 3070.

The test statistics we used in our analyses are defined as the difference
between two log-likelihood (LLH) statistics. The first is specific to the model
used, and the second is based on a model with only the intercept term.

Results
Association Analysis
In this section, we describe the power profiles that result by applying the
models described in Table 7.4 to the data generated according to the four
different experiments in Table 7.3. We focus on detecting the associations
between the combined genotype-environmental factors on phenotype
outcome (disease diagnosis). We assess the importance of model specification
in predicting the presence of association with a phenotype of interest and to
what degree the gene model and genotype environment interactions influence
power. In the following section on genotype associations, we assess the role
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 99

of the genotype alone in predicting association while controlling for the

environmental influence.
Note that in calculating all power results in this section, we assumed that
the Type I error rate was 10-8. However, because all combined environmental
exposure and genetic inheritance risk values are greater than 1.0 in all of our
experiments, only Type II errors were possible.
Table 7.5 shows the data generated using the protocol for Experiment 1.
Note that for this and all subsequent tables in this section, the highest power
value for each risk profile within the three MOI categories is bolded to highlight
the optimal model. For each genetic inheritance (GI) risk level () there is an
environmental exposure (EE) risk level () equal to 1.0, indicating no EE risk.

Table 7.5. Power values, by statistical model, , and : Experiment 1all gene
models
Additive Dominant Recessive
M-1 M-2 M-3 M-1 M-2 M-3 M-1 M-2 M-3
1.10 1.00 .002 .004 .004 .000 .004 .004 .000 .000 .000
1.10 1.05 .002 .016 .022 .000 .040 .044 .000 .000 .000
1.10 1.10 .000 .160 .316 .000 .214 .354 .000 .050 .122
1.10 1.15 .002 .654 .882 .000 .728 .924 .000 .488 .810
1.10 1.20 .006 .986 1.00 .000 .992 1.00 .000 .968 1.00
1.15 1.00 .024 .046 .042 .000 .028 .024 .000 .000 .000
1.15 1.05 .028 .102 .138 .000 .082 .102 .000 .000 .000
1.15 1.10 .042 .378 .538 .000 .336 .468 .000 .038 .144
1.15 1.15 .052 .838 .948 .000 .832 .954 .000 .528 .806
1.15 1.20 .064 .996 1.00 .000 .994 1.00 .000 .958 .998
1.20 1.00 .246 .210 .206 .004 .100 .104 .000 .000 .000
1.20 1.05 .274 .308 .338 .002 .214 .240 .000 .002 .004
1.20 1.10 .308 .630 .746 .006 .552 .684 .000 .054 .142
1.20 1.15 .350 .908 .982 .016 .912 .972 .002 .586 .852
1.20 1.20 .394 .996 1.00 .028 .994 1.00 .004 .972 .998
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level.
Note: Table 7.4 defines the statistical models (M-1, M-2, and M-3).

Figure 7.1 shows the data generated using the protocol for Experiment 1
for the additive gene model. Figure 7.1 includes the optimal model (M-3,
identified by the boldfaced cells in Table 7.5) and the model that does not
include an EE variable in its specification (M-1). The results presented in
100 Chapter 7

Table 7.5 and Figure 7.1 indicate that there is little difference in performance
between models when the risk of EE is not present.

Figure 7.1. Power curves, by statistical model, , and : Experiment 1

additive gene model
Model M-1 Model M-3
1.0

0.8
Statistical power

0.6

0.4

0.2

0.0
1.10 1.10 1.10 1.10 1.10 1.15 1.15 1.15 1.15 1.15 1.20 1.20 1.20 1.20 1.20
1.00 1.05 1.10 1.15 1.20 1.00 1.05 1.10 1.15 1.20 1.00 1.05 1.10 1.15 1.20
Relative risk combination (upper number = ; lower number = )

= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level.
Note: Table 7.4 defines the statistical models (M-1 and M-3).

The results that Figure 7.1 and Table 7.5 show indicate the following:
The power profile of M-1 is substantially less than that of models M-2 and
M-3. M-1 represents a typical single-locus method used in a GWAS that
ignores environmental influences. We conclude that not including an EE
reduces the likelihood of the locus being associated with the phenotype.
M-3 is the most powerful of the three models. This is expected because the
Experiment 1 protocol should generate data consistent with the M-3 model
formulation.
The difference between the profiles of M-2 and M-3 is a result of the
manner used to characterize the EE functional form. Because the data were
generated in a manner compatible with the e1 variable used in M-3, that
model generated more accurate power predictions.
Note that in the full M-3, the overall intercept is the log of the intercept
for the line that relates EE to the log-odds risk among those subjects with the
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 101

nondisease genotype AA. The coefficient associated with the g1 main effect is
testing for the difference between intercepts for the subjects with genotype aA
and those with genotype AA. Similarly, the g2 main effect coefficient is testing
for the difference between the intercepts for subjects with the aa genotype and
those with the AA genotype.
The EE main effect coefficient is the height of the step in the step function
relating EE to log-odds-risk for subjects with the AA genotype, and it,
therefore, tests for a common EE step height across all three genotypes. The
g1E interaction coefficient is the difference between the step heights for
the aA subjects and the AA subjects. Similarly, the g2E coefficient is the
difference between the step heights for the aa subjects and the AA subjects.
Because the AA, aA, and aa step heights/slopes associated with the EE
environmental effect are all equal in Experiments 1 and 3, only the common
main effect (ME) associated with EE contributes to association prediction in
those data sets, and the interaction terms are superfluous.
Table 7.6 and Figure 7.2 show the results of applying the three models
that Table 7.4 describes to the data generated according to the Experiment 2
protocol (see Table 7.3). The results from Experiment 2 indicate the following:
Although M-1 does not adjust for EE, the observed (relatively) high power
profiles for high EE risk levels suggest that the GI-EE interaction effect is
embedded in the M-1 power values, and the high power profiles are credited
as a genotype main effect.
As in Experiment 1, M-3 outperforms all other models because the variable
e1 properly characterizes EE behavior. This clearly demonstrates the value
of preprocessing (i.e., mining) the data before committing to a specific
association model.
Table 7.7 and Figure 7.3 show the results of applying the three statistical
models described in Table 7.4 to the data generated according to the
Experiment 3 protocol (see Table 7.3). The results from Experiment 3 indicate
the following:
M-1 consistently performs below M-2 and M-3, indicating that not
including an EE term limits the association assessment.
In general, model M-2 produces better power profiles than M-3. This is
expected given that the EE incremental risk is linearly related to the log of
EE. Thus, model M-2 is more consistent with the protocol used to generate
the data in Experiment 3.
102 Chapter 7

Table 7.6. Power values, by statistical model, , and : Experiment 2all gene models
Additive Dominant Recessive
M-1 M-2 M-3 M-1 M-2 M-3 M-1 M-2 M-3
1.10 1.00 .002 .004 .004 .000 .004 .004 .000 .000 .000
1.10 1.05 .012 .086 .106 .000 .086 .094 .000 .000 .002
1.10 1.10 .114 .394 .462 .010 .408 .442 .002 .072 .116
1.10 1.15 .340 .702 .744 .078 .690 .738 .022 .376 .462
1.10 1.20 .640 .856 .914 .318 .844 .896 .104 .642 .714
1.15 1.00 .024 .046 .042 .000 .028 .024 .000 .000 .000
1.15 1.05 .166 .258 .290 .006 .212 .222 .000 .000 .004
1.15 1.10 .478 .634 .676 .082 .570 .590 .002 .090 .126
1.15 1.15 .710 .808 .832 .380 .846 .872 .082 .398 .464
1.15 1.20 .880 .952 .968 .730 .934 .954 .230 .654 .740
1.20 1.00 .246 .210 .206 .004 .100 .104 .000 .000 .000
1.20 1.05 .544 .518 .520 .084 .404 .426 .008 .002 .006
1.20 1.10 .760 .776 .784 .336 .752 .784 .052 .120 .162
1.20 1.15 .918 .926 .936 .736 .920 .944 .184 .418 .496
1.20 1.20 .978 .990 .996 .916 .984 .986 .345 .640 .725
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level.
Note: Table 7.4 defines the statistical models (M-1, M-2, and M-3).

Figure 7.2. Power curves, by statistical model, , and : Experiment 2

additive gene model
Model M-1 Model M-3
1.0

0.8
Statistical power

0.6

0.4

0.2

0.0
1.10 1.10 1.10 1.10 1.10 1.15 1.15 1.15 1.15 1.15 1.20 1.20 1.20 1.20 1.20
1.00 1.05 1.10 1.15 1.20 1.00 1.05 1.10 1.15 1.20 1.00 1.05 1.10 1.15 1.20
Relative risk combination (upper number = ; lower number = )

= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level.
Note: Table 7.4 defines the statistical models (M-1 and M-3).
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 103

Table 7.7. Power values, by statistical model, , and : Experiment 3all gene models
Additive Dominant Recessive
M-1 M-2 M-3 M-1 M-2 M-3 M-1 M-2 M-3
1.10 1.00 .002 .000 .004 .000 .006 .004 .000 .000 .000
1.10 1.05 .002 .008 .006 .000 .014 .014 .000 .000 .000
1.10 1.10 .002 .034 .030 .000 .044 .030 .004 .004 .000
1.10 1.15 .002 .152 .096 .000 .190 .120 .062 .024 .018
1.10 1.20 .004 .472 .296 .000 .500 .260 .258 .228 .066
1.15 1.00 .024 .036 .042 .000 .034 .024 .000 .000 .000
1.15 1.05 .030 .068 .058 .000 .048 .052 .000 .000 .000
1.15 1.10 .040 .152 .106 .000 .162 .118 .002 .002 .002
1.15 1.15 .046 .390 .278 .000 .384 .302 .048 .034 .012
1.15 1.20 .052 .650 .524 .000 .670 .508 .308 .278 .074
1.20 1.00 .246 .174 .206 .004 .114 .104 .000 .000 .000
1.20 1.05 .270 .274 .250 .002 .166 .150 .000 .000 .000
1.20 1.10 .286 .402 .376 .002 .344 .260 .004 .002 .002
1.20 1.15 .348 .658 .548 .012 .520 .490 .098 .058 .032
1.20 1.20 .378 .862 .718 .024 .796 .660 .299 .289 .107
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level.
Note: Table 7.4 defines the statistical models (M-1, M-2, and M-3).

Figure 7.3. Power curves, by statistical model, , and : Experiment 3

additive gene model
Model M-1 Model M-2
1.0

0.8
Statistical power

0.6

0.4

0.2

0.0
1.10 1.10 1.10 1.10 1.10 1.15 1.15 1.15 1.15 1.15 1.20 1.20 1.20 1.20 1.20
1.00 1.05 1.10 1.15 1.20 1.00 1.05 1.10 1.15 1.20 1.00 1.05 1.10 1.15 1.20
Relative risk combination (upper number = ; lower number = )

= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level.
Note: Table 7.4 defines the statistical models (M-1 and M-2).
104 Chapter 7

Table 7.8 and Figure 7.4 show the results for Experiment 4. They indicate
the results of applying the three models described in Table 7.4 to the data
generated according to the Experiment 4 protocol (see Table 7.3).

Table 7.8. Power values, by statistical model, , and : Experiment 4all gene models
Additive Dominant Recessive
M-1 M-2 M-3 M-1 M-2 M-3 M-1 M-2 M-3
1.10 1.00 .002 .006 .004 .000 .006 .004 .000 .000 .000
1.10 1.05 .008 .054 .006 .000 .068 .054 .000 .000 .002
1.10 1.10 .078 .236 .030 .006 .284 .226 .002 .014 .008
1.10 1.15 .220 .478 .096 .048 .500 .476 .008 .110 .078
1.10 1.20 .510 .672 .678 .148 .624 .632 .046 .314 .294
1.15 1.00 .024 .074 .042 .000 .046 .024 .000 .000 .000
1.15 1.05 .118 .208 .164 .002 .162 .130 .000 .000 .000
1.15 1.10 .384 .514 .476 .040 .454 .410 .000 .018 .016
1.15 1.15 .618 .688 .658 .232 .698 .684 .046 .162 .132
1.15 1.20 .820 .830 .838 .570 .802 .792 .144 .354 .328
1.20 1.00 .246 .250 .206 .004 .138 .104 .000 .000 .000
1.20 1.05 .464 .466 .406 .046 .354 .290 .004 .002 .000
1.20 1.10 .714 .714 .692 .222 .642 .624 .026 .040 .022
1.20 1.15 .892 .876 .864 .622 .816 .824 .136 .216 .170
1.20 1.20 .954 .940 .944 .848 .916 .930 .257 .317 .343
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level.
Note: Table 7.4 defines the statistical models (M-1, M-2, and M-3).

Table 7.8 and Figure 7.4 show results that indicate the following:
Consistent with Experiment 2s results, model M-1 does not adjust for
EE, but because of the influence of GI-EE interaction effects, M-1 displays
higher power profiles for large EE risk levels.
As in Experiment 3, M-2 outperforms M-3 because it better characterizes
the EE by using the variable E (age) and further demonstrates the value
of preprocessing (i.e., mining) the data before committing to a specific
association model.
In the presence of GI-EE interaction effects, the genetic-only model (M-1)
performs better than anticipated.
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 105

Figure 7.4. Power curves, by statistical model, , and : Experiment 4

additive gene model
Model M-1 Model M-2
1.0

0.8
Statistical power

0.6

0.4

0.2

0.0
1.10 1.10 1.10 1.10 1.10 1.15 1.15 1.15 1.15 1.15 1.20 1.20 1.20 1.20 1.20
1.00 1.05 1.10 1.15 1.20 1.00 1.05 1.10 1.15 1.20 1.00 1.05 1.10 1.15 1.20
Relative risk combination (upper number = ; lower number = )

= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level.
Note: Table 7.4 defines the statistical models (M-1 and M-2).

Genotype Associations
The analysis in the previous section focused exclusively on composite associa-
tionsthat is, whether a specific gene plus an environmental factor associates
with a phenotype. As we noted earlier, our main interest was separating main
genetic effects from environmental effects and their interactions. To accomplish
this, we defined a total effect test (TOT) that adjusts for EE where

TOT = LLH [log (, g1, g2, e1, g1e1, g2e1)] LLH [log (, e1)] (7.5)

is the test we applied to the data generated by Experiment 1 and 2 protocols, and

TOT = LLH [log (, g1, g2, E, g1E, g2E)] LLH [log (, E)] (7.5a)

is the test that we applied to the data generated by Experiment 3 and 4 protocols.
TOT is the association test that measures genetic effects (main and inter
active) and is adjusted for the environmental effect.27 TOT simultaneously
measures whether the aA and aa intercepts are different from the AA intercept
and whether the aA and aa slopes are nonzero, given that the AA slope on EE
is zero. This test was used to test for association from all causes.
106 Chapter 7

We also define two additional tests for genotype-environment interactions,

INT, as follows:

INT = LLH [log (, e1, g1, g2, g1e1, g2e1)] LLH [log (, e1, g1, g2)] (7.6)

and

INT = LLH [log (, E, g1, g2, g1E, g2E)] LLH [log (, E, g1, g2)]. (7.6a)

The INT test subtracts the main effects for g1, g2, and EE from the TOT and
tests whether the EE steps (or slopes) for the aA and aa genotypes are different
from the corresponding EE step (slope) for genotype AA.
The final test measures the influence of the genetic main effects (ME).

ME = LLH [log (, e1, g1, g2)] LLH [log (, e1)] (7.7)

is the test applied to the data generated by Experiments 1 and 2 protocols and

ME = LLH [log (, E, g1, g2)] LLH [log (, E)] (7.7a)

is the corresponding test for data from Experiments 3 and 4 protocols.

The ME tests check whether the estimated aA and aa intercepts differ from
the AA intercept, conditioned on the EE step sizes (e1 in experiments 1and 2)
or the EE slopes (E in experiments 3 and 4) being equal for all three genotypes.
Note that for Experiments 2 and 4, both the AA step (coefficient of e1) and
slope (coefficient of E) on EE are zero, and therefore the coefficient for the EE
main effect (assuming that the M-3 is operating) is estimating zero; the two
interaction columns are estimating the aA step/slope minus zero and the aa
step/slope minus zero, respectively.
Typically, these three tests would be applied sequentially: TOT followed
by INT, then ME. Assessing whether an interactive or noninteractive genetic
association is obtained would depend on the result of the preceding test.
For example, if TOT is nonsignificant, the process stops, and we conclude that
there is no connection between the genetic locus and the phenotype. Otherwise,
we would apply the INT test. If INT was significant, we could conclude that
the locus and the phenotype are significantly related, with the caveat that the
strength of the genotype effect varies by the EE risk level. The ME test would
only be applied if the TOT is significant and INT test is not significant. In this
case, the ME test would be applied to affirm that the genetic and environmental
effects are operating independently of each other and to assert that a common
genotype main effect exists that applies to all EE levels. Tables 7.9 through 7.13
show the results of running the three tests (TOT, INT, and ME).
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 107

Table 7.9. Total effects test (TOT) power values, by risk profile, , and all
experiments and gene models, N = 200,000
Experiment 1 Experiment 2 Experiment 3 Experiment 4
TOT^ TOT^ TOT^ TOT^ TOT^ TOT^ TOT* TOT* TOT* TOT* TOT* TOT*
Rec Dom Add Rec Dom Add Rec Dom Add Rec Dom Add
1.10 1.00 .319 .601 .574 .319 .601 .574 .328 .573 .538 .328 .573 .538
1.10 1.05 .328 .602 .612 .523 .777 .827 .340 .619 .608 .719 .958 .968
1.10 1.10 .343 .604 .577 .806 .949 .953 .377 .636 .596 .990 1.00 1.00
1.10 1.15 .344 .613 .625 .958 .994 .999 .408 .684 .662 1.00 1.00 1.00
1.10 1.20 .358 .622 .637 .993 .999 1.00 .492 .709 .704 1.00 1.00 1.00
1.15 1.00 .341 .725 .739 .341 .725 .739 .336 .725 .713 .336 .725 .713
1.15 1.05 .363 .745 .769 .534 .918 .919 .367 .750 .766 .764 .995 .997
1.15 1.10 .375 .770 .773 .837 .986 .993 .427 .796 .828 .992 1.00 1.00
1.15 1.15 .358 .751 .776 .934 .999 1.00 .480 .832 .870 1.00 1.00 1.00
1.15 1.20 .351 .775 .787 .995 .999 1.00 .487 .861 .893 1.00 1.00 1.00
1.20 1.00 .444 .889 .915 .444 .889 .915 .439 .851 .904 .439 .851 .904
1.20 1.05 .456 .900 .918 .631 .986 .982 .490 .912 .950 .809 1.00 .999
1.20 1.10 .477 .897 .943 .854 .998 1.00 .514 .949 .949 .995 1.00 1.00
1.20 1.15 .471 .917 .935 .973 1.00 1.00 .594 .946 .975 1.00 1.00 1.00
1.20 1.20 .514 .925 .945 .996 1.00 1.00 .632 .961 .986 1.00 1.00 1.00
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level; TOT = total effects test;
gene models: Rec = recessive, Dom = dominant, Add = additive.
Note: ^ = TOT from equation 5; * = TOT from equation 5a; 10-8.

Consider that every replicate in every cell produced by the simulation

experiments is designed to generate a genotype-phenotype association (albeit
at low risk). Some of these replicates influenced by an EE also contribute
toward association. However, in a perfect statistical world, all are generated to
predict an association with the phenotype. The fact that they do not indicates
the limitations of the GWAS process.
In addition, Table 7.9 suggests the following:
Association detection involving recessive genes is difficult to identify and
accordingly requires a larger sample size than we used in our experiments.
Scenarios involving gene-environment interactions (Experiments 2 and 4)
greatly influence whether genetic influences can be detected by a gene-only
model.
108 Chapter 7

The type of EE process influences the ability to detect an association,

whether the effect is caused by a chemical-type exposure (Experiments 1
and 2) or by aging (Experiments 3 and 4).
Table 7.10 presents the results of applying the INT test to all experiments
and all gene models. Not shown are the results for Experiments 1 and 3,
which generated data without interaction effects. They estimate no interaction
between GI and EE (as they should), so those results are not shown. Note that
the Type 1 thresholds in Table 7.11 for generating power estimates for all
cells are 10-2.

Table 7.10. Genotype-environment interactions (INT) power values, by risk profile

( and )all experiments and gene models, N = 10,000
Experiment 2 Experiment 4
INT^ INT^ INT^ INT* INT* INT*
Rec Dom Add Rec Dom Add
1.10 1.00 .319 .601 .574 .319 .601 .574
1.10 1.05 .328 .602 .612 .523 .777 .827
1.10 1.10 .343 .604 .577 .806 .949 .953
1.10 1.15 .344 .613 .625 .958 .994 .999
1.10 1.20 .358 .622 .637 .993 .999 1.00
1.15 1.00 .341 .725 .739 .341 .725 .739
1.15 1.05 .363 .745 .769 .534 .918 .919
1.15 1.10 .375 .770 .773 .837 .986 .993
1.15 1.15 .358 .751 .776 .934 .999 1.00
1.15 1.20 .351 .775 .787 .995 .999 1.00
1.20 1.00 .444 .889 .915 .444 .889 .915
1.20 1.05 .456 .900 .918 .631 .986 .982
1.20 1.10 .477 .897 .943 .854 .998 1.00
1.20 1.15 .471 .917 .935 .973 1.00 1.00
1.20 1.20 .514 .925 .945 .996 1.00 1.00
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level; INT = genotype-
environment interactions; gene models: Rec = recessive, Dom = dominant, Add = additive.
Note: ^ = INT from equation 6; * = INT from equation 6a; 10-2.

Table 7.10 reaffirms the results of Table 7.9:

Power values for recessive genes are very low and accordingly were more
difficult to identify than other gene models.
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 109

Gene-environment interactions influence association outcomes. This is

evidenced by all cells of the no-interaction experiments (Experiments 1
and 3) having power values <.004.
The type of EE process influences the detection of an association, whether
the effect is due to an exposure (Experiments 1 and 2) or is due to an aging
mechanism (Experiments 3 and 4).
Interaction effects achieve significant levels in Experiment 2 for risk values
of EE 1.2 only.

Table 7.11. Main effects (ME) power values, by risk profile ( and )all gene models,
N = 10,000
Experiment 1 Experiment 3
ME^ ME^ ME^ ME* ME* ME*
Rec Dom Add Rec Dom Add
1.10 1.00 .123 .463 .434 .139 .450 .417
1.10 1.05 .131 .478 .476 .156 .479 .437
1.10 1.10 .138 .491 .459 .138 .476 .399
1.10 1.15 .138 .523 .500 .149 .498 .478
1.10 1.20 .154 .507 .495 .167 .506 .487
1.15 1.00 .153 .626 .628 .147 .608 .585
1.15 1.05 .179 .642 .664 .146 .619 .659
1.15 1.10 .193 .676 .660 .177 .654 .675
1.15 1.15 .163 .670 .670 .191 .653 .675
1.15 1.20 .171 .668 .682 .182 .676 .695
1.20 1.00 .204 .805 .850 .239 .755 .834
1.20 1.05 .244 .824 .850 .236 .807 .861
1.20 1.10 .250 .829 .887 .239 .817 .866
1.20 1.15 .264 .840 .873 .274 .816 .867
1.20 1.20 .315 .854 .893 .268 .830 .878
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level; ME = main effects; gene
models: Rec = recessive, Dom = dominant, Add = additive.
Note:^ = ME from equation 7; * = ME from equation 7a; 10-2.

Table 7.11 presents the results of the ME test for Experiments 1 and 3. ME
results for Experiments 2 and 4 are not shown because they were generated
by a protocol that produced EE and GI interactions, and if the INT test
demonstrated significance (as it should have), the ME tests would have been
unnecessary. In all cases, the alpha threshold was set to 10-2.
110 Chapter 7

Table 7.11 reaffirms the results of Tables 7.9 and 7.10:

Associations involving recessive genes are more difficult to identify.
Gene-environment interactions influence association outcomes.
The type of process influences the detection of an association, as shown by
differences between power values for exposure mechanisms such as those
resembling chemical spills (Experiments 1 and 2) and those recognizing
aging mechanisms (Experiments 3 and 4).
Main effects are only considered significant for larger risk values of genetic
inheritance (those with a risk of 1.2 or greater).
Note that the power threshold values are set to low ( 10-2) for the
interaction and ME tables. To investigate the effect of a very large N, we
repeated the simulation process with N =200,000 and reduced the threshold to
( 10-8). Tables 7.12 and 7.13 present the results.

Table 7.12. Genotype-environment interactions (INT) power values, by risk profile (

and )all gene models, N = 200,000
Experiment 2 Experiment 4
INT^ INT^ INT* INT* INT* INT*
Rec Dom Add Rec Dom Add
1.10 1.00 .00 .00 .00 .00 .00 .00
1.10 1.05 .00 .00 .03 .00 .00 .00
1.10 1.10 .75 .77 .88 .35 .50 .56
1.10 1.15 1.0 1.0 1.0 1.0 1.0 1.0
1.10 1.20 1.0 1.0 1.0 1.0 1.0 1.0
1.15 1.00 .00 .00 .00 .00 .00 .00
1.15 1.05 .00 .01 .03 .00 .00 .01
1.15 1.10 .78 .86 .91 .44 .52 .60
1.15 1.15 1.0 1.0 1.0 .96 1.0 1.0
1.15 1.20 1.0 1.0 1.0 1.0 1.0 1.0
1.20 1.00 .00 .00 .00 .00 .00 .00
1.20 1.05 .01 .04 .05 .00 .00 .00
1.20 1.10 .75 .92 .96 .45 .65 .74
1.20 1.15 .99 1.0 1.0 1.0 1.0 1.0
1.20 1.20 1.0 1.0 1.0 1.0 1.0 1.0
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level; INT = genotype-
environment interactions; gene models: Rec = recessive, Dom = dominant, Add = additive.
Note: ^ = INT from equation 6; * = INT from equation 6a; 10-8.
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 111

The results shown in Table 7.12 suggest that the GI-EE interactions are
very sensitive to low EE levels ( < 1.10). They also accurately characterize an
interaction power value of zero when = 1.00, (i.e., no EE risk).

Table 7.13. Main effects (ME) power values, by risk profile ( and )all gene models,
N = 200,000
Experiment 1 Experiment 3
ME^ ME^ ME^ ME* ME* ME*
Rec Dom Add Rec Dom Add
1.10 1.00 .68 .68 .68 .68 .68 .68
1.10 1.05 .67 .70 .73 .70 .70 .73
1.10 1.10 .65 .75 .68 .75 .75 .68
1.10 1.15 .65 .68 .72 .68 .68 .71
1.10 1.20 .77 .81 .69 .78 .79 .68
1.15 1.00 .60 .88 .94 .60 .88 .94
1.15 1.05 .65 .84 .92 .67 .85 .92
1.15 1.10 .61 .84 .92 .61 .84 .92
1.15 1.15 .75 .92 .97 .75 .92 .96
1.15 1.20 .65 .87 .94 .66 .85 .94
1.20 1.00 .75 1.0 1.0 .75 1.0 1.0
1.20 1.05 .73 1.0 1.0 .75 1.0 1.0
1.20 1.10 .81 1.0 1.0 .80 1.0 1.0
1.20 1.15 .81 1.0 1.0 .81 1.0 1.0
1.20 1.20 .85 1.0 1.0 .85 1.0 1.0
= genetic inheritance (GI) risk level; = environmental exposure (EE) risk level; ME = main effects; gene
models: Rec = recessive, Dom = dominant, Add = additive.
Note:^ = ME from equation 7; * = ME from equation 7a; 10-8.

These results suggest that for very large studies it is possible to predict
positive associations between recessive genes linked to phenotypes with low to
moderate risk.

Conclusions
In summary, the chances of predicting an association in a GWAS are reduced if
an environmental effect is present and the statistical model does not adjust for
it. This is especially true if the environmental effect and genetic marker do not
have an interaction effect. The functional form of the model also matters. The
more accurately the form of the environmental influence is portrayed by the
112 Chapter 7

statistical model, the more accurate the prediction will be. Even with very large
sample sizes, association predictions involving recessive markers are low.
This study focused on one important methodological step involved in
conducting a GWAS: selecting a statistical method and a supporting model
that reliably predict associations. This study does not address the broader issue
of the supporting experimental design that employs the statistical methods as
part of an overall solution strategy. Those combined issues and their mutual
interconnections are described by Cordell.28
The specific scenarios we address here involve genetic associations that
have environmental influences. Our assumption is that the environmental
influence that contributes to a given phenotype is in question and the precise
form of that influence is unknown. A separate analysis to characterize the
functional form to proxy the mechanism behind the environmental exposure
is required. These approaches should focus on case-only data similar to the
methods described in the Cornelis and colleagues study.18 These approaches
involve investigating different environmentally related functional relationships
between the suspected environmental influence and the phenotype in the
cases-only subpopulation. For example, if gene effects and environmental
effects are independently significant with respect to disease prevalence, a
polynomial model could be used to characterize the relationship between
environmental effects and the log-odds of disease prevalence. This would
allow us to test whether the nonlinear parameterization would be required
to characterize the environmental effect. Alternatively, if the environmental
effect has multiple levels such as age, researchers could investigate a cubic
polynomial to assess whether the effect stayed low initially then rose at some
point and flattened out toward the end of the environmental effects range.
If this analysis suggests an appropriate polynomial level for environmental
effects, researchers should also investigate a similar assessment using the gene-
environment interaction variable.
We have used this simulation scenario in previous studies. We reviewed
single-gene models and evaluated a wide class of statistical methods.23 Our
results indicated that researchers should consider a multitest procedure that
combines individual gene-based (dominant, recessive, additive) core tests as a
composite statistical method for conducting the initial screen in a GWAS. The
tests can be combined into a single operational test in a number of ways. Two
such tests are Holms Bonferroni procedure and the MAX procedure described
by Li and colleagues.29,30 Of course, if the gene model under investigation is
known, a single test that assumes the implied inheritance form is better than
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 113

a combined test. For this study, all patterns across gene models are consistent
and only vary by degree.
In Chapter 5, we have also evaluated the effect of phenotype errors that
resulted from inaccurate diagnoses and genotype errors that resulted from
gene-chip errors or occurrences of DNA methylation altering gene expression
that associate a wild-type gene with the wrong phenotype outcome.20 Our
results quantify the relationship between genotype and diagnosis error
measures and sample size to achieve a .80 statistical power level. Our results
also demonstrate that researchers should not underestimate the need to
increase sample size to compensate for power loss due to the presence of
genotype and diagnosis errors.
In Chapter 6, we also investigated epistatic scenarios involving two genes.31
The results showed that the most powerful statistical methods for predicting
associations between phenotypes and genotypes in epistatic scenarios are
statistical models that simultaneously test for associations involving both
interacting loci. This is consistent with the results we present here. This result
is not surprising and has been reported by others. We reported that if two
genes contribute to a phenotype, the weaker gene will be obscured by the
stronger gene and often will not be identified as a contributor to the phenotype
when a single-gene model is used. Again, this result is similar to showing
that the effect of an environmental exposure can obscure the influence of a
genotype-phenotype association if the model does not account for the GI and
the EE simultaneously. In this sense, two-gene models (or alternatively a gene-
environment model) produce better predictions of association than single-
gene models do.
We acknowledge that our results could possibly depend on the particular
experiments we devised to investigate how the statistical models performed. In
light of this, we are reviewing other scenarios to establish the robustness of our
findings. Nevertheless, establishing the genotype-to-phenotype connections
without using a simulation approach is limited.
For the gene-environment interaction scenarios addressed here, the results
across all gene models lead us to conclude that using a composite test that
supports distinct underlying statistical modelsthat is, a main effectsonly
model and a main effects with interactions modelis likely to be more
effective than single-model tests. This result does not depend on the gene
model and thus differs from the single-gene and epistatic scenarios, in which
each different gene model assumption (i.e., recessive, dominant, or additive)
requires representation in the composite test.20,29
114 Chapter 7

Chapter References
1. Cooley PC, Clark RF, Folsom RE. Statistical methods that identify
genotype-phenotype associations in the presence of environmental effects.
RTI Press Publication No. RR-0022-1405. Research Triangle Park, NC:
RTI Press; 2014.
2. Kuo CL, Feingold E. Whats the best statistic for a simple test of genetic
association in a case-control study? Genet Epidemiol. 2010;34(3):246-253.
3. Suhre K, Shin SY, Petersen AK, et al. Human metabolic individuality in
biomedical and pharmaceutical research. Nature. 2011;477(7362):54-60.
4. Spencer C, Hechter E, Vukcevic D, et al. Quantifying the underestimation
of relative risks from genome-wide association studies. PLoS Genet.
2011;7(3):e1001337.
5. Dunn OJ. Multiple Comparisons among Means. Journal of the American
Statistical Association. 1961;56(293):52-&.
6. Manolio TA, Collins FS, Cox NJ, et al. Finding the missing heritability of
complex diseases. Nature. 2009;461(7265):747-53.
7. Terry PD, Umbach DM, Taylor JA. APE1 genotype and risk of bladder
cancer: evidence for effect modification by smoking. Int J Cancer.
2006;118(12):3170-3.
8. Stern MC, Johnson LR, Bell DA, et al. XPD codon 751 polymorphism,
metabolism genes, smoking, and bladder cancer risk. Cancer Epidemiol
Biomarkers Prev. 2002;11(10 Pt 1):1004-11.
9. Browning BL, Browning SR. Efficient multilocus association testing for
whole genome association studies using localized haplotype clustering.
Genet Epidemiol. 2007;31(5):365-75.
10. Zhao J, Jin L, Xiong M. Nonlinear tests for genomewide association
studies. Genetics. 2006;174(3):1529-38.
11. Lichtenstein P, Holm NV, Verkasalo PK, et al. Environmental and heritable
factors in the causation of cancer--analyses of cohorts of twins from
Sweden, Denmark, and Finland. N Engl J Med. 2000;343(2):78-85.
12. Pearce CL, Rossing MA, Lee AW, et al. Combined and interactive effects
of environmental and GWAS-identified risk factors in ovarian cancer.
Cancer Epidemiol Biomarkers Prev. 2013;22(5):880-90.
 Assessing Gene-Environment Interactions in Genome-Wide Association Studies 115

13. Rothman N, Garcia-Closas M, Chatterjee N, et al. A multi-stage genome-

wide association study of bladder cancer identifies multiple susceptibility
loci. Nat Genet. 2010;42(11):978-84.
14. Lindstrom S, Schumacher F, Siddiq A, et al. Characterizing associations
and SNP-environment interactions for GWAS-identified prostate cancer
risk markers--results from BPC3. PLoS One. 2011;6(2):e17142.
15. Yu K, Wacholder S, Wheeler W, et al. A flexible Bayesian model for
studying gene-environment interaction. PLoS Genet. 2012;8(1):e1002482.
16. Patel CJ, Chen R, Kodama K, et al. Systematic identification of interaction
effects between genome- and environment-wide associations in type 2
diabetes mellitus. Hum Genet. 2013;132(5):495-508.
17. Murcray CE, Lewinger JP, Gauderman WJ. Gene-environment interaction
in genome-wide association studies. Am J Epidemiol. 2009;169(2):219-26.
18. Cornelis MC, Tchetgen EJ, Liang L, et al. Gene-environment interactions
in genome-wide association studies: a comparative study of tests applied
to empirical studies of type 2 diabetes. Am J Epidemiol. 2012;175(3):191-
202.
19. Kraft P, Yen YC, Stram DO, et al. Exploiting gene-environment interaction
to detect genetic associations. Hum Hered. 2007;63(2):111-9.
20. Cooley P, Clark RF, Page G. The influence of errors inherent in genome
wide association studies (GWAS) in relation to single gene models. J
Proteomics Bioinform. 2011;4:138-144.
21. Schymick JC, Scholz SW, Fung HC, et al. Genome-wide genotyping in
amyotrophic lateral sclerosis and neurologically normal controls: first
stage analysis and public release of data. Lancet Neurol. 2007;6(4):322-8.
22. Iles MM. Effect of mode of inheritance when calculating the power of a
transmission/disequilibrium test study. Hum Hered. 2002;53(3):153-7.
23. Cooley P, Clark R, Folsom R, et al. Genetic inheritance and genome
wide association statistical test performance. J Proteomics Bioinform.
2010;3(12):321-325.
24. Chan EK, Hawken R, Reverter A. The combined effect of SNP-marker and
phenotype attributes in genome-wide association studies. Anim Genet.
2009;40(2):149-56.
116 Chapter 7

25. Sladek R, Rocheleau G, Rung J, et al. A genome-wide association study

identifies novel risk loci for type 2 diabetes. Nature. 2007;445(7130):881-5.
26. Ziegler A, Konig IR, Thompson JR. Biostatistical aspects of genome-wide
association studies. Biom J. 2008;50(1):8-28.
27. Lehmann EL, Romano JP. Testing statistical hypotheses. 3rd ed. New York:
Springer; 2005.
28. Cordell HJ. Detecting gene-gene interactions that underlie human
diseases. Nat Rev Genet. 2009;10(6):392-404.
29. Holm S. A simple sequentially rejective multiple test procedure. Scand J
Stat. 1979;6:65-70.
30. Li Q, Zheng G, Li Z, et al. Efficient approximation of P-value of the
maximum of correlated tests, with applications to genome-wide
association studies. Ann Hum Genet. 2008;72(Pt 3):397-406.
31. Cooley P, Gaddis N, Folsom R, et al. Conducting genome-wide association
studies: epistasis scenarios. J Proteomics Bioinform. 2012;5(10):245-251.
 CHAPTER 8

Polygene Methods in
Genome-Wide Association Studies (GWAS)
Philip Chester Cooley and Ralph E. Folsom

Overview
The hope that the Human Genome Project would pave the way to detect
associations between genetic markers and common diseases such as heart
disease, diabetes, auto-immune diseases, and psychiatric disorders has not
achieved expectations.1 Genome-wide association studies (GWAS) have been
the dominant tool used in the exploration of the genome to determine the
associations between genomic regions and complex traits in the population.
There are two ways to improve the GWAS process. The first is to improve the
quality of the data; the next generation of sequence data will provide greater
genomic coverage and will soon be available to the scientific community. The
second is by improving the computational process. Issues such as multiple
testing and the influence of external (nondisease) variables that confound
experimental results limit the interpretability of the statistical results.
This chapter examines the utility of a computation process that considers
multiple markers that simultaneously associate with a phenotypic trait. This
computational process builds a single nucleotide polymorphism (SNP)
network using a stepwise association method that proceeds in stages adding a
single SNP at each stage until no further additions are called for.
Initially, the model used in a GWAS consisted of a series of single-locus
statistical tests, examining each SNP independently for association to the
phenotype. In Chapter 4, we examined this single-gene model and showed
that, if the mode of inheritance (MOI) modelwhether genes are dominant,
recessive, additive, or multiplicativeis unknown, then a Cochran-Armitage
composite (CA-C) test, which pools the possibility of three MOI outcomes into

This chapter has not been published elsewhere.

118 Chapter 8

a single test, is the most powerful single-gene model test. We also showed that
the classical case-control method of epidemiology (an MOI-agnostic test) will
never be as powerful as the CA-C test.
Because most GWAS assume that the phenotype and the genotype data are
error free, we examined the consequences of this assumption in Chapter 5 and
found that adherence to the assumption forecasts sample size requirements too
small to achieve reliable association predictions and undermines replication
across studies.
Because genes do not always act as single triggers but act in concert with
other genes, polygene analysis is an important methodology that represents the
next step forward if GWAS are to remain a viable exploratory tool. Single-gene
GWAS have detected associations with numerous diseases but explain little
about disease heritability. Polygenic models assume that thousands of genetic
variants could impact the phenotype. For example, in humans, height, skin
color, eye color, and weight are all phenotypes linked to polygenes.
Chapter 6 discusses epistatic effects in association studies. We describe a
simple polygene method: two-gene models that have been designed to act in
ways that depend on the pair of MOIs and how the pair are positioned relative
to each other. We show that in contrast to the single-gene model, two-gene
models that are MOI agnostic (i.e., the two-gene case-control model) are
more powerful than either single-gene models or models that treat gene-gene
interactions only.
Chapter 7 uses a similar approach as that discussed in Chapter 6, but we
replace one of the genes with a variable representing an external environmental
influence such as a chemical exposure or the effects of aging. We show that
detection of the genetic marker is more likely to occur if the external process is
represented in the statistical model even if the characterization is inaccurate.
This chapter also builds on the work of previous chapters by confronting the
issue of how to process general polygene models with an unknown number of
significant traits. We propose stepwise polygene analysis as a new strategy for
studying the association between large sets of SNP predictors and groups of
correlated phenotypes (i.e., outcomes). These data commonly arise in GWAS,
in which associations with a large number of qualitative and quantitative
phenotypes are investigated using hundreds of thousands of genetic markers.
Our strategy uses the log linear-logistic model framework suited to analyzing
qualitative trait responses. In the two-gene epistatic scenario, we showed
that single-locus models do not detect all of the markers that are part of
 Polygene Methods in Genome-Wide Association Studies 119

the phenotype pathway. Other studies have also reported this finding.2
However, from a combinatorial perspective, polygene models can create an
overwhelming number of comparisons when using more than two loci. A
number of methods exist that have developed approaches to examine all-
possible two-locus combinations.3,4 However, the computational requirements
of these approaches is daunting.
Chapter 8 defines a new polygene approach: a multistep process that is
analogous to a stepwise association process. A major new contribution is
that the initial stage of our process examines all usable SNP autosomes using
the CA-C testwhich, as we showed in Chapter 4, is the most powerful of
the conventional tests used to predict associations in GWAS5to predict
each SNPs genetic inheritance properties. The accuracy of these inheritance
predictions determine the effectiveness of the method because this assumption
is used to estimate the type of wild-type alleles for each loci. Subsequent stages
statistically combine the SNPs from previous stages and examine whether any
of the unassigned SNPs are associated with the phenotype based on the SNPs
from previous stages. Because each stage builds on the previous one, we refer
to this method as the stepwise association method, and we use it to predict
polygene networks for a given phenotype.
To test the method, we generated sets of simulated data described in the
Generating SNP Data section. These data sets consist of SNP (genotype)
data for each subject and an associated phenotype (case-control) measure.
The interclass correlation of the genotype data is controlled to determine
how strongly the different SNPs resemble each other, and each SNP has a
designated MOI and risk level. Using a combination of SNP and genotype
information, we estimate the value that each SNP contributes to the probability
of a case and use Monte Carlo methods to assign a case-control measure.
Collectively, this constitutes a truth set of known outcomes and the factors
that affect genotype-phenotype association.6
Hence, we can assess the performance of each method to predict known
outcomes. In general, our results suggest that the process is effective for main
effect and interaction SNPs with modest effects, as measured by their odds
ratios (ORs), but variables with small effects present a challenge that will be
difficult to overcome even with very large sample sizes.
120 Chapter 8

Background
In the past 7 years, GWAS have been the most widely used tool in genetic
research. Yet over this period, the tools of our research efforts have failed to
unravel the complete biological architecture of genetic diseases.7 Researchers
have learned from the myriad GWAS conducted during the past few years
that the biggest effect sizes for associations between genes and phenotypes are
much smaller than anticipated. For example, in a genome-wide meta-analysis
of intelligence quotient (IQ) involving nearly 18,000 children, the largest
effect size accounted for only 0.2 percent of the variance.8 This suggests that,
in general, smaller effects will be barely detectable in GWAS and extremely
difficult to replicate. However, this finding conflicts with hundreds of candidate
gene and gene-environment interaction studies that find significant effects
using modest sample sizes. The finding also implies that many of the reported
markers are false positives that cannot be replicated, although some journals
now require that candidate gene papers include an independent replication.9
Identifying additional loci of small effect can be partially accomplished
by meta-analysis of multiple GWAS using stringent significance thresholds.
Furthermore, it is unlikely that GWAS will ever be powered to identify the
full spectrum of small effects. Several recent analytical approaches have
been developed to test whether common variants of extremely small effect
size could be combined to explain phenotype variation. This approach was
successfully applied to the schizophrenia gene SCZ. A polygene small-effect
SNP model explained approximately 35 percent of disease diagnosis variance
of an estimated total of 80 percent.10 In a second example, Yang and colleagues
estimated that 67 percent of the heritability of human height could be
explained by a polygenic model.11 Although both models fail to estimate total
heritability, the models do account for far more than that of known, validated
associations. However, these approaches are limited because they are unable
to identify the proportion that each marker contributes to trait variation: the
marginal effects contributed by each variable are impossible to estimate using
models with many genetic variables.12 Nevertheless, a polygenic analysis
defines a large set of variants with an unknown subset that affects phenotype.
Together, these represent the true underlying biology.
Performing polygenic analysis to understand the genetic basis of complex
traits leads to a systems biology perspective in which many perturbations of a
complex network contribute to the outcome of a complex trait phenotype. The
complexity is difficult or impossible to disentangle on a per variant basis. To
 Polygene Methods in Genome-Wide Association Studies 121

establish the biology of the complex trait directly, researchers need a systems
genetics approach in which large sets of genetic variants and/or genes are
analyzed in an integrated way and incorporate functional data.13

Disease-Scoring Methods
Disease-scoring methods are one such approach that researchers have
explored. Combining multiple genetic markers into a single score that predicts
disease risk is a relatively new approach for associating SNPs with disease
in the context of GWAS. This approach has shown that some diseases have
a strong genetic basis, even if few actual genes have been identified. Disease
scoring has also revealed a common genetic basis for distinct diseases. The
scoring approach has been used to obtain evidence of genetic effects when no
single markers are significant, establish a common genetic basis for related
disorders, and construct risk prediction models. Published studies have
demonstrated that significant associations of polygenic scores have occurred
only in well-powered studies and that useful levels of prediction may occur
when predictors are estimated from very large samples, up to an order of
magnitude greater than any currently available. This suggests the need for
studies to use larger sample sizes.14
Many geneticists believe that better polygenic scoring methods will be
developed. Quantitative genetic techniques that estimate heritability using
DNA suggest that about half of the heritability can be detected using the
common SNPs that are currently genotyped on commercially available DNA
arrays, given sufficiently large samples.15 Economic improvements in whole-
genome sequencing will make it cost-effective to identify DNA sequence
variation of every kind throughout the genome.7
Studying environment effects is more difficult than studying genes because,
as Cooley and colleagues have shown, the environment is more complex than
DNA with its genotype 0,1,2 designation code.16 However, environmental
research likely could capitalize on the advances in whole-genome technology
to identify biomarkers of environmental influence and therefore help detect
genetic associations.7

Statistical Approaches to Polygene Analysis

Although disease scoring is one viable approach to polygene analysis, this
chapter focuses on using statistical analyses to approach the problem. In an
earlier study of epistatic loci, we showed that for a given locus, single-locus
tests are not as effective as two-locus tests for predicting associations if the risk
122 Chapter 8

value for a second interacting locus exceeds a relative risk of 1.051.12.1 The
crossover risk value varies, depending on the genetic inheritance properties of
the pair of interacting loci. In general, the power of two-locus tests to detect
associations improves as the risk value of the second locus increases, whereas
the power of single-locus tests progressively declines.
For certain inheritance models and risk values, a true association between
a locus and phenotype can be masked by a second interacting locus when
using single-locus tests. However, these findings are not unexpected and
are consistent with the findings of others.17-19 We also note that a study of
ALS subjects identified SNPs that, when paired with other SNPs, became
significant markers of ALS, although there were no indications of association
using single-gene models.20 Thus, some markers can only be identified if they
are paired with other markers within a polygene network that is part of the
phenotypes biological pathway.*
Multilocus analyses are not as straightforward as conducting single-
locus tests and present numerous computational, statistical, and logistical
challenges.21 Examining all pairwise combinations of SNPs using gene
chips that generate 500,000 to 1,000,000 SNPs is currently computationally
infeasible. One approach to this issue is to preprocess the SNPs set to
eliminate redundant SNPs. A way to accomplish this is to select a set of results
from a single-SNP analysis based on an arbitrary significance threshold and
exhaustively evaluate interactions in that subset. We incorporate this approach
into our stepwise polygene analysis method but acknowledge that selecting
SNPs to analyze based on main effects will prevent certain multilocus models
from being detected where the heritability is concentrated in the interaction
rather than in the main effects.
Our stepwise polygene analysis approach looks for a main effect at Stage 1
but also considers the possibility of interactions. The benefit of our method
is that it performs an unbiased analysis for interactions within the selected
set of SNPs. It is also far more computationally and statistically tractable than
analyzing all possible combinations of markers.
Bush and colleagues describe another strategy: restricting examination of
SNP combinations to those that fall within an established biological context,

* We use the term biological pathway to represent the biological reactions and the interaction
network in a cell where each reaction is identified with its enzyme, which in turn is coded by
certain genes. Some of the most common biological pathways are involved in metabolism,
regulating gene expression, and transmitting signals. Pathway analysis plays a key role in the
advanced studies of genomics.
 Polygene Methods in Genome-Wide Association Studies 123

such as a biochemical pathway or a protein family.22 Because both these

techniques rely on electronic repositories of structured biomedical knowledge,
they use a tool that generates SNP-SNP combinations with a statistical method
that evaluates combinations in the GWAS data set. For example, the Biofilter
approach uses public data sources in conjunction with logistic regression and
multifactor dimensionality reduction methods.23 Similarly, the INTERSNP
tool for genome-wide interaction analysis (GWIA) of case-control SNP
data and quantitative traits selects SNPs for joint analysis using a priori
information. Sources of information to define meaningful strategies are based
on statistical evidence (e.g., single-marker association studies computed from
different data sources). INTERSNP uses logistic regression and contingency
table approaches to assess SNP-SNP interaction models.24
Cooley and colleagues examined the performance of three different
two-locus tests.25 In most scenarios, the two-locus, case-control 8-degree of
freedom (df ) Pearson test was the most powerful. In certain scenarios (i.e.,
when both genes have a dominant MOI), the unlinked, cases-only version of
the Wu and colleagues test as refined by Ueki and colleagues was optimal. The
version of the Ueki test in which cases and controls are included was never
optimal.26,27 These findings are not surprising because Wu and colleagues
modified test measures interaction effects exclusively, whereas the two-locus,
case-control test includes main effects for both loci as well as interaction
effects.
One unresolved issue is how to construct a computationally practical test
that takes into account interactions and enhances the detection of associations
between a specific locus and the phenotype of interest. As mentioned
previously, Wang and colleagues conducted an empirical comparison of five
epistatic interaction detection methods, which examines all combinations of
two-gene methods.3 Each of the five methods demonstrated unique utilities,
but no single method was simultaneously the most powerful and the most
scalable and had the lowest type-1 error rate in every setting. When users
want powerful results and are not concerned with computation cost, Wang
and colleagues cite the TEAM method of Zhang and colleagues as having
the highest performing algorithm.3,4 However, researchers should note that
even limiting the number of interacting genes to two requires n(n - 1) / 2
association calculations. For n = 500,0001,000,000, the computational
requirements are daunting but readily parallelizable.
124 Chapter 8

Methods
The Stepwise Algorithm
Our method proceeds in stages. At each stage, we add an unassigned SNP
to the network unless the test score is below the significance threshold. The
first step (Stage 1) makes an initial pass against all usable SNP autosomes to
identify the most highly significant SNP-phenotype association. The second
stage statistically combines the Stage 1 SNP with all original autosome SNPs
to identify the most significant SNP pair associated with the phenotype. This
stage uses a test for significance that is conditional on the Stage 1 SNP. We
then continue this process for triple SNPs, quadruple SNPs, and so on until
combining the loci produces no new SNP-phenotype associations. This process
is analogous to a stepwise regression process, in which networks of SNPs are
connected stage by stage until no new SNPs can be identified. Our process is
not an exhaustive polygene analysis, but it does assume that at least one loci
associated with the phenotype can be identified via a single-gene model.

Step 1 Details. The initial pass uses the CA-C test to examine all SNP
autosomes and apply a restrictive significance threshold to identify the most
highly significant SNP-phenotype association, known as the Stage 1 SNP.
The single-locus model used in Stage 1 is based on the CA-C test that used
an MOI agnostic model that combined the three versions (recessive, dominant,
and additive) of the classical CA-C method into a single test. Our previous
results indicate a multitest procedure that combines the results of individual
MOI-based tests is an effective method for initially filtering a GWAS. Note
that the CA-C method can be used to predict the SNP-specific MOIs.
Consequently, the first stage not only selects the Stage 1 SNP, it also tags all
other SNPs with an MOI, which is then used to predict ensuing stage-specific
SNPs.

Ensuing Steps. Stage i (i = 2,3,4, ) is defined as progressively adding one

SNP, if appropriate, into the process; the stages advance until it is no longer
possible to add SNPs and advance. Stage i applies a model of all of the Stage i-1
SNPs and computes a test score. Stage i (i > 0), i-1 SNPs have been associated
with the phenotype and are assumed to be part of the SNP/phenotype
network. We then systematically add remaining unassigned SNPs to the set of
stage i-1 SNPs and compute a new test score. The test score is defined as twice
the difference between the natural log of the likelihood including the Stage i
 Polygene Methods in Genome-Wide Association Studies 125

SNP and the natural log of the likelihood including only the Stage i-1 SNPs.
This test score has approximately a 2 distribution with one degree of freedom
if the log odds ratio for the Stage i SNP is actually 0.
Note that the test used to identify the Stage i (i>1) SNP uses a logistic
regression (LR) approach and the construction of the test depends on the
stage. A log likelihood (LLH) criteria is used to estimate the LR parameters
and measure whether the Stage i SNP sufficiently differentiates between cases
and controls given the set of Stage i-1 SNPs at a specified significance level T.
Formally, for si to be a candidate SNP at Stage i, it must satisfy:

ti = LLH[si / s1 , s2 , , si-1 ] T for i = 2,3,4, N, (8.1)

where s1 s2 si-1. If more than one SNP satisfies equation (8.1),

si is defined as the SNP with the largest value of ti. If ti < T the process is
terminated at the i-1 stage. Also note that T will be Bonferroni corrected at
each stage in the process to account for the multiple number of tests carried
out for each stage.
To summarize, the process begins with the best single SNP and continues
for SNP pairs, triple SNPs, quadruple SNPs, and so forth, until the process
is unable to identify any new SNPs to add to the network. At Stage i, a test of
significance is used that is conditional on the prior selected set of SNPs. The
process continues to define a networks of SNPs that are connected stage by
stage, until no new SNPs can be identified.

Generating SNP Data

To assess the method we propose, we first had to generate a synthetic set of
SNP data. There are a number of features that characterize these data that also
characterize real-life properties of SNP data:
1. The SNP inheritance properties
2. The distribution of genotype alleles
3. The correlation between SNP pairs (as a linkage disequilibrium measure)
4. The contributions of the individual SNPs toward their association with the
phenotype
We discuss each of these properties in subsequent sections. Note that
alternative methods to generate data are presented in the Appendix to this
chapter.
126 Chapter 8

Mode of Inheritance. There are two main properties of genetic inheritance

formulated by Mendel: the principle of segregation and the principle of
independent assortment. In segregation, for any particular trait, the pair of
alleles of each parent separate and only one allele passes from each parent on
to an offspring. In independent assortment, different pairs of alleles are passed
to offspring independently of each other. This principle indicates why the
human inheritance of a particular eye color does not increase or decrease the
likelihood of having six fingers on each hand, for example. This is because the
genes for independently assorted traits are located on different chromosomes.
The Online Mendelian Inheritance in Man (OMIM)28 provides the best source
of information on the MOI distribution (see Table 8.1). However, OMIM is
disproportionally populated by genes linked to single Mendelian disorders,
and genes associated with multifactorial disorders are underrepresented.
Because polygene influences are assumed to be a major source of additive and
multiplicative SNP behavior, the distribution in Table 8.1 is likely biased.

Table 8.1. Distribution of genes in Online Mendelian Inheritance in Man, by mode of

inheritance (MOI)
MOI Frequency
Autosomal Dominant 3,805
Autosomal Additive 12
Autosomal Multiplicative 21
Autosomal Recessive 3,775

Genotype Distributions. The genotype distribution data we used in this study

was first described by Schymick et al.6 It consists of 555,352 SNPs from 276
ALS patients and 271 neurologically normal controls. These data are publicly
available to the scientific community and were produced by the Laboratory of
Neurogenetics of the intramural program of the National Institute on Aging
(NIA), National Institutes of Health (NIH). The genotyping was performed using
the Illumina Infinium assay humanhap550. Infinium assays assess haplotype
tagging SNPs based upon Phases I+II of the International HapMap Project.

Linkage Disequilibrium (LD). LD is a property of SNP pairs that describes the

degree to which an allele of one SNP is inherited (correlated) with an allele
of another SNP. The concept of LD was developed by population geneticists
in an attempt to describe changes in genetic variation within a population.
 Polygene Methods in Genome-Wide Association Studies 127

According to the theory, recombination events occur within a family. This

effect is amplified through generations, and in a population of fixed size
undergoing random mating, repeated random recombination events will
eventually produce linkage equilibrium/independence. Linkage between
markers on a population scale is referred to as linkage disequilibrium.
In other words, LD is the nonrandom association of alleles at two or more
loci that descend from single, ancestral chromosomes. It is also a measurement
of distance between SNP locations and represents the combination of alleles in
a population more often or less often than would be expected from a random
draw. The degree of LD depends on the difference between observed allelic
frequencies and those expected from a homogenous, randomly distributed
model.29 Linkage equilibrium occurs in populations in which combinations of
alleles or genotypes can be found in the expected proportions.30
However, biology is not the only factor at play with respect to LD
measurements. Humans tend to associate with other humans who have similar
genetic characteristics. Studies have shown that populations tend to stratify
according to nonrandom mating patterns in which subjects with similar
genotypes and/or phenotypes mate more frequently than would be expected
under a random mating pattern. For example, it is common for individuals of
similar body size to mate. Evidence for this process has been reported from
the National Longitudinal Study of Adolescent Health.31 One feature of this
study tested genetic similarity between friends. Using friendship networks
clusters of genotypes demonstrated genotype positive correlation (homophily)
and negative correlation (heterophily). These results were replicated in an
independent sample from the Framingham Heart Study. These unique results
suggest that association tests should take into account the fact that human
genes are influenced by the humans of their social networks.
A natural way to measure LD is by the correlation coefficient.29
Characterizing LD structure is useful for designing genetic association
studies because we can control for genetically homogeneous populations
and their influence on population-specific parameters. However, in reality, a
study sample can be genetically heterogeneous with substructure even if the
recruiting criterion requires a specific ethnic group.
In this study, we specify the degree of correlation (using the product
moment or Pearson correlation coefficient) between each SNP pair during the
generation of the SNP database as a measure of LD.
128 Chapter 8

Designating the Case-Control Assignments. The final procedure in creating

the SNP data set uses the SNP data (with the appropriate LD genomic spacing)
and the SNP odds ratio (ODDS) data to determine the case (1) control (0)
designation codes that are used to predict association. In Table 8.2, we present
three examples of data we will use to test our method.
We assume that these three examples are real gene networks that are known
with certainty. Our goal is to examine the genotype data that describe these
networks to assess method performance. To accomplish this, we will create
data sets that reproduce the specifications that define the networks. These
specifications include the MOI, which we limit to dominant, recessive, and
additive; sample size (N); the LD characteristics; and the ODDS relative to
the penetrance odds. If P0 is the disease penetrance probability, then the
penetrance odds is P0/(1-P0)=exp(B0). If P1 is the probability of disease when
a given dominant MOI SNP (SNP1) has one or two wild-type alleles, then P1/
(1-P1)=exp(B0+B1) is the disease odds with the addition of SNP1. The SNP1
ODDS divided by the penetrance odds or the SNP-1 odds ratio (relative to the
penetrance odds) is therefore exp(B1).
By knowing these specifications, we can reverse-engineer a data set that
replicates (with varying degrees of certainty) the starting specifications. This
approach thus generates replicates with varying degrees of goodness of fit. We
specify the total number of replicates to calculate and then store the best-fitting
replicates as the data to analyze, because their LR coefficients best conform to
the specifications and serve to represent the methods average performance
over a large number of replicated samples of the specified size.
Example 1 consists of 7 high-risk SNPs (ODDS > 1.2), 4 low-risk SNPs
(ODDS > 1.0 & ODDS < 1.2), and 4 no-risk SNPs (ODDS = 1.0). The
penetrance is .25. There are 5 recessive SNPs, and 10 dominant SNPs.
Example 2 is identical to Example 1, except it includes two additional effects
that both represent interactions. The first is the interaction between SNPs 3
and 6. The second represents the interaction between SNPs 1 and 9. All four
SNPs that compromise the interaction terms have positive main effects.
Example 3 comprises three SNPs with ODDS = 1.10, three no-risk SNPs,
and nine SNPs with risks between 1.01 and 1.05. This example is designed to
test the feasibility of performing association studies with low-effect SNPS, and
thus will require a much larger sample size to detect an effect.
In summary, the X (design) matrix consists of genotype data that contains
an LD structure specified exogenously. We use these data, in conjunction with
 Polygene Methods in Genome-Wide Association Studies 129

the MOI and the ODDS parameters, to determine the probability of a case.
Using a random number compared with the case probability, we assign the
case-control designation vector Z and the number of subjects N, which we also
specify exogenously and use to determine the number of rows in X and Z, as
Table 8.2 shows.

Table 8.2. SNP data set design datathree examples

E(value) = ODDS

E(value) = ODDS

E(value) = ODDS
Parameter

Example 1

Example 2

Example 3
Value

Value

Value
Variable MOI
Intercept B0# -1.386 .25 -1.386 .25 -0.916 .40
SNP1 D B1 .2231 1.25 .2231 1.25 .0010 1.01
SNP2 R B2 .0 1.00 .0 1.00 .0953 1.10
3 D B3 .3365 1.40 .3365 1.40 .0953 1.10
4 D B4 .0953 1.10 .0953 1.10 .0010 1.01
5 R B5 .4055 1.50 .4055 1.50 .0 1.00
6 D B6 .1823 1.20 .1823 1.20 .0488 1.05
7 D B7 .1397 1.15 .1397 1.15 .0392 1.04
8 R B8 .0 1.00 .0 1.00 .0953 1.10
9 R B9 .2852 1.33 .2852 1.33 .0488 1.05
10 R B10 .0953 1.10 .0953 1.10 .0198 1.02
11 R B11 .2231 1.25 .2231 1.25 .0010 1.01
12 D B12 .3001 1.35 .3001 1.35 .0392 1.04
13 D B13 .0953 1.10 .0953 1.10 .0 1.00
14 R B14 .0 1.00 .0 1.00 .0 1.00
15 D B15 .0 1.00 .0 1.00 .0296 1.03
SNP1 SNP9 I B16 X X .2523 1.30 X X
SNP3 SNP4 I B17 X X .1823 1.20 X X
MOI = mode of inheritance; ODDS = SNP odds ratios; SNP= single nucleotide polymorphism.
#B
0 = log odds of penetrance.

We assign a constant genomic distance between SNP pairs as measured by the

Pearson correlation coefficient, R. We begin by selecting SNP1 at random from
the Schymick et al.6 distribution of SNPs. We define D as the genomic distance,
which will be reflected in R. We then select D percent of the SNP1 genotypes and
assign them to SNP2; the remaining 1-D percent are selected at random. This
130 Chapter 8

process performs the same calculations with SNP2 and SNP3, and so forth. Thus,
the genomic distance between SNP1 and SNP2 will be approximately D and the
genomic distance between SNP1 and SNP3 will be D squared.
Completion of this step generates a design matrix, X, consisting of genotype
values (0,1,2). The next step, which calculates the case-control designation
proceeds according to the following steps:
1. Given that the Xijs are = {0,1,2}F, define a set of variables Yijs derived from
the Xijs in the following way:
a. If Xij = 2 and j is a recessive SNP, Yij = 1 otherwise Yij = 0.
b. If Xij > 0 and j is a dominant SNP, Yij = 1 otherwise Yij = 0.
c. Ignore (for the time being) additive and multiplicative SNPs.
2. Calculate the case score Wi = B0 + SUM (BjYij) for the ith subject.
3. Convert the score into a probability pi = exp(Wi) / (1.0+exp(Wi) for the ith
subject (pi is the predicted probability of a case).
4. Use pi to generate Zi .the designation of the case-control value i.e. if random
number (01) < pi Zi = 1 otherwise Zi = 0.
5. After the Z vector is generated, determine whether the calculation of the
Bjs are sufficiently close to the parameters in Table 8.2 that were used to
generate the data set by performing a logistic regression of the generated
data to see how well the estimated coefficients reproduce the values in Table
8.2.
6. Generate a measure of fit closeness with respect to the estimated Bis
relative to the parameters in Table 8.2 and save the corresponding design
matrix with the closeness measures. This will generate a number of
replicates that exhibit closeness scores, with the last one having the best
overall correspondence.
At this point in the process we have generated an X matrix and a
companion Y matrix, which indicates whether the genotype has a wild-type
allele. We also generated a vector Z that accounts for the collective ODDS
of the SNPS. Furthermore, the process has been checked by estimating the
coefficients (i.e., step 5 above) of the logistic regression model for a number of
replicates. We save only the replicates that match original ODDS specifications
for future analysis, which in our case will be to assess how the proposed
method performs.
 Polygene Methods in Genome-Wide Association Studies 131

Results
Model Comparisons and MOI Predictions
The first example illustrates the approach using genotype data consisting of
SNP pairs that are paired at rates D = 0.0, 0.2, and 0.4. We assume that the
15 SNPs have risk levels that are consistent with those that Table 8.2 presents.
We then use two LR models to apply our method. The first model uses the
0,1,2 genotype codes as the independent variable codes. The second model
uses the MOI predictions to stratify the genotype data into wild-type and non
wild-type allele. We will refer to each step as a stage.
The Stage 1 MOI predictions are shown in Table 8.3. The missed predictions
predominate for those SNPs with low ODDS values and for LD = .4. For
example, with ODDS > 1.0, the missed predictions occur twice in the LD = .0
column (SNPs 10 and 13), once in the LD = .2 column (SNP 6) and three times
in the LD = .4 column (SNPs 5, 7, 11).

Table 8.3. Stage 1 summary

MOI MOI MOI
Predicted Predicted Predicted
SNP ODDS MOIActual LD = .0 LD = .2 LD = .4
1 1.25 D D D D
2 1.00 R R D D
3 1.40 D D D D
4 1.10 D D D D
5 1.50 R R R A
6 1.20 D D A D
7 1.15 D D D A
8 1.00 R D A A
9 1.33 R R R R
10 1.10 R A R R
11 1.25 R R R A
12 1.35 D D D D
13 1.10 D A D D
14 1.00 R D A D
15 1.00 D R D A
SNP = single nucleotide polymorphism; ODDS = SNP odds ratios; MOI = mode of inheritance; LD = linkage
disequilibrium; N = 10,000.
Note: Missed predictions are shown in boldface.
132 Chapter 8

Having predicted MOIs reasonably accurately, the question remains: is there

value in using this information to predict associations? Figures 8.1A, 8.1B,
and 8.1C compare the two models for each LD level. The three figures display
the log of the test p-value scores by stage. The log values of the corrected .05
and .01 threshold levels are also included on each figure. They make a clear
case that Model 2, which uses the MOI information to map the three genotype
values to the two allele levels, is more powerful than Model 1, which treats the
genotype levels as a single variable. Both models use the LR framework. In
each of the figures, the p-value of the Model 2 (orange) is less than the p-value
of Model 1 (blue).
Table 8.4 translates the p-value measures in the figures into a performance
measure describing the composition of the network. This table summarizes the
overall performance by model for each LD level. For example, the comparison
for LD = .2 indicates that at the .01 (.05) threshold, Model 1 missed 6 (3) SNPs
(of 11 SNPs with ODDS > 1.0) and associated 5 (8) correctly. By comparison,
Model 2 predicted correctly 10 (9) SNPs at the .01 (.05) significance level and
missed 1 (2).

Figure 8.1a. Model comparison for LD = 0.0 SNP data

Stage of assessment
3 4 5 6 7 8 9 10 11 14
0
Log of statistical power (0.01.0)

-5

-10
Model 1
-15 Model 2
log(.05)
-20
log(.01)
-25

-30

-35

LD = linkage disequilibrium; SNP = single nucleotide polymorphism; MOI = mode of inheritance;

Model 1 = model without predicted MOI; Model 2 = model with predicted MOI;
log(.05) = .05 corrected confidence threshold; log(.01) = .01 corrected confidence threshold.
 Polygene Methods in Genome-Wide Association Studies 133

Figure 8.1b. Model comparison for LD = 0.2 SNP data

Stage of assessment
3 4 5 6 7 8 9 10 11 14
0
Log of statistical power (0.01.0)

-5

-10
Model 1
-15 Model 2
log(.05)
-20
log(.01)
-25

-30

-35
LD = linkage disequilibrium; SNP = single nucleotide polymorphism; MOI = mode of inheritance;
Model 1 = model without predicted MOI; Model 2 = model with predicted MOI;
log(.05) = .05 corrected confidence threshold; log(.01) = .01 corrected confidence threshold.

Figure 8.1c. Model performance comparison for LD = 0.4 SNP data

Stage of assessment
3 4 5 6 7 8 9 10 11 14
0
Log of statistical power (0.01.0)

-5

-10
Model 1
-15 Model 2
log(.05)
-20
log(.01)
-25

-30

-35
LD = linkage disequilibrium; SNP = single nucleotide polymorphism; MOI = mode of inheritance;
Model 1 = model without predicted MOI; Model 2 = model with predicted MOI;
log(.05) = .05 corrected confidence threshold; log(.01) = .01 corrected confidence threshold.
134 Chapter 8

Table 8.4. Prediction accuracy results, by model and LD criteria

Criteria\LD .0 .0 .2 .2 .4 .4
Model 1 2 1 2 1 2
Correct at .01 8 8 5 9 7 7
Correct at .05 8 8 8 10 8 10
Missed at .01 3 3 6 2 4 4
Missed at .05 3 3 3 1 3 0
ODDs .01 missed at .01 4 4 3 3 2 1
ODDS .01 missed at .05 3 3 2 0 2 0
LD = linkage disequilibrium; ODDS = SNP odds ratios; SNP = single nucleotide polymorphism; N = 10,000.

The results of Table 8.4 and Figures 8.1A through 8.1C suggest collectively
that differences exist between statistical models and that Model 2 outperforms
Model 1. Also, with the subject size N at 10,000, the predictive accuracy of
both models appears unreliable for those SNPs with an effect size measured by
ODDS 1.10.

Models With Interactions

Our investigation of interaction effects used Example 2, as defined in Table 8.2.
Example 2 is the same as Example 1, with two additional interactions terms.
SNP1 and SNP9 form an interaction with ODDS = 1.30, and SNP3 and SNP
form an interaction with ODDS = 1.2. SNP9 is recessive; therefore, we would
anticipate that this interaction term would be more difficult to detect than the
interaction involving SNPs 3 and 4, which are both dominant SNPs.
The first task is to build a data set that conforms to the Example 2
specifications. We do this for each of the three SNP intercorrelation effects that
varied from .0 to .2 to .4. Table 8.5 shows the build results.
The results shown in Table 8.5 are based on assigning N = 10,000, running
1,000 replicates, and selecting the run that best fits the specification from the
1,000 replicates. In reconstructing the data, fitting the specifications is not an
exact process and even small deviations could influence the accuracy of the
association prediction phase. However, we controlled this by adding replicates
and we generated sufficient replicates to demonstrate that the methods
performance is consistently applied. The LD = .0 and .2 runs do not identify
one main effect SNP (with an ODDS < .10) and both interactions (type II
errors). For the LD = .4 run, the MOI of SNP 5 was inaccurately predicted as
an additive SNP.
 Polygene Methods in Genome-Wide Association Studies 135

Table 8.5. Build results for interaction example (Example 2) by LD levels = 0.0, 0.2, and 0.4
Corr =.0 Corr = .2 Corr = .4
ODDS ODDS ODDS
SNP ODDS MOI reconstructed reconstructed reconstructed
1 1.25 D 1.26 1.29 1.23
2 1.00 R .96 .99 1.02
3 1.40 D 1.41 1.42 1.40
4 1.10 D 1.13 1.16 1.03
5 1.50 R 1.48 1.55 1.52
6 1.20 D 1.24 1.24 1.19
7 1.15 D 1.15 1.15 1.16
8 1.00 R 1.02 .98 .97
9 1.33 R 1.33 1.35 1.36
10 1.10 R 1.10 1.11 1.09
11 1.25 R 1.20 1.25 1.30
12 1.35 D 1.34 1.37 1.38
13 1.10 D 1.12 1.08 1.11
14 1.00 R .97 1.00 .97
15 1.00 D 1.00 1.02 1.03
19 1.30 DR 1.31 1.27 1.26
34 1.20 DD 1.18 1.24 1.19
Correct Calls 10 10 8
Missed SNP 10 & SNP 13 & SNPs 5,11,13 &
Interactions Interactions Interactions
SNP = single nucleotide polymorphism; ODDS = SNP odds ratios; MOI = mode of inheritance.

Overall, these errors contributed to the exclusion of SNP5 as a member

of the network. The central question is can the interactions be predicted for
larger sample sizes? Figure 8.2 results partially address this issue by graphing
the log of the p-value scores by algorithm stages as a function of N (i.e., N =
10,000, 20,000, 30,000 ,and 40,000).
Table 8.6 summarizes the results of Figure 8.2 for N = 40,000 subjects. An
N = 40,000 is sufficient to predict all SNPs with ODDS > 1.00. Only one of the
two interactions was predicted to associate with the phenotype. The interaction
involving recessive SNP9 was undetected.
136 Chapter 8

Figure 8.2. Influence on sample size N on model performance

Stage of assessment
3 4 5 6 7 8 9 10 11
0
Log of statistical power (0.01.0)

-20
N = 10,000
-40 N = 20,000
N = 30,000
-60
N = 40,000
log(.05)
-80
log(.01)
-100

-120

LD = linkage disequilibrium; SNP = single nucleotide polymorphism.

Table 8.6. Results by stage: correlated SNP pairs LD = .200

LD = .20
Stage ODDS MOI SNP p-value
1 1.40 D 3 -269.2
2 1.35 D 12 -132.3
3 1.33 R 9 -116.2
4 1.50 R 5 -100.6
5 1.25 D 1 -69.02
6 1.20 D 6 -50.95
7 1.10 D 4 -33.88
8 1.25 R 11 -31.83
9 1.15 D 7 -26.31
10 1.10 D 13 -15.49
11 1.20 DD 34 -10.20
12 1.10 R 10 -6.135
ODDS = SNP odds ratios; MOI = mode of inheritance; SNP = single nucleotide polymorphism; LD = linkage
disequilibrium; N = 40,000, note all significant at the .01 (Bonferroni-corrected) level. Only the SNP1 by
SNP9 interaction was not detected. All SNPs with ODDs = 1.00 were not associated (correctly).
 Polygene Methods in Genome-Wide Association Studies 137

The results of Figure 8.2 and Table 8.6 suggest that more than N = 10,000
samples are needed before the interaction is detected. However, one of the
interaction terms involving the recessive SNP (SNP9) was undetected even
with an N = 40,000.

Detecting Low-Effect SNPs

Table 8.7 presents results using Example 2. The results are intended to examine
the sample-size requirement for Model 2 with SNP data that has low levels of
risk. Recall that the Example 2 network consists of three SNPs with ODDS =
1.10, three no-risk SNPs, and nine SNPs with risks between 1.01 and 1.05. We
expanded our sample to 300,000 to detect the low-level effects contained in
the example.

Table 8.7. Results of low-effect genes by stage and correlated SNP pairs LD = .200 for
Model 2
log(p-value)

log(p-value)

log(p-value)
N = 100,000

N = 200,000

N = 400,000
ODDS

ODDS

ODDS
Stage

MOI

MOI

MOI
SNP

SNP

1 1.10 D 3 -43.45 1.10 D 3 -59.77 1.10 D SNP

3 -115.0
2 1.05 D 2 -18.16 1.04 D 7 -31.03 1.10 R 8 -75.33
3 1.05 R 7 -15.39 1.10 R 8 -24.94 1.10 R 2 -43.84
4 1.10 R 8 -10.77 1.10 R 2 -21.03 1.05 D 6 -37.24
5 1.10 R 6 -8.504 1.05 R 9 -13.72 1.04 D 12 -24.97
6 - - - - 1.05 D 6 -12.98 1.05 R 9 -23.39
7 - - - - 1.04 D 12 -11.35 1.03 D 15 -10.42
8 - - - - 1.03 D 15 -7.835 1.04 D 7 -9.722
9 - - - - - - - - 1.01 D 4 -6.457
Missed 1.04 D 4 -5.854 1.01 R 11 -5.101 1.02 R 10 -1.740
1.03 D 9 -3.677 1.02 R 10 -1.740 1.01 D 1 -1.274
1.04 D 12 -3.614 1.01 D 1 -1.274 1.00 R 5 -.8376
1.02 R 15 -2.412 1.00 R 5 -.8376 1.00 D 13 -.80
1.01 D 1 -2.046 1.00 D 13 -.8067 1.01 D 4 -.5392
1.00 D 10 -1.729 1.01 D 4 -.5392 1.00 R 14 -.1960
1.01 R 14 -1.488 1.00 R 14 -.1960 - - - -
1.01 D 11 -1.246 - - - - - - - -
1.00 R 13 -1.062 - - - - - - - -
1.00 R 5 -.3823 - - - - - - - -
ODDS = SNP odds ratios; MOI = mode of inheritance; SNP = single nucleotide polymorphism.
138 Chapter 8

As Table 8.7 indicates, increasing N affects the performance of the stepwise

algorithm. For N = 100,000, the algorithm identifies five significant loci at the
.01 corrected significance level. For N = 200,000, eight loci are identified, and
for N = 400,000, nine significant loci are identified. Overall, the method missed
seven small-effect loci when N = 100,000, four loci when N = 200,000, and
three loci when N = 400,000.
Table 8.7 also demonstrates that a very large N is necessary to detect very
low risk levels (<1.03), especially if the SNP is recessive. Also, much of the
N = 400,000 model advantage occurs at its higher stages, which suggests that
even models with larger Ns may still be ineffective for low-risk loci.
Overall, these results suggest there is little value in extending N to levels
higher than 400,000.

Discussion
This study has demonstrated that creating a set of synthetic genes with known
properties that can be analyzed in the context of GWAS experiments is
possible and informative. Because the properties of the genes are determined,
and it is possible to establish exploratory protocols that define computational
best practices. At least seven other similar approaches have been reported
in the literature; however, the creation process we used in these experiments
controls for LD, sample size, MOI, the single locus main effect risks, and the
interaction loci risks, which other similar approaches do not.
Our synthetic gene procedure is designed to examine the properties of
defining a network of genes that link to a single phenotype. The method for
establishing this gene network is straightforward. It proceeds in stages and
uses logistic regression methods at each stage except Stage 1. Each stage uses
a maximum likelihood ratio test to identify the optimal locus to include in
the network, given the loci that have accumulated in the network in prior
stages. The Stage 1 calculation comprises the composite test that uses the three
variations of the Cochran-Armitage test. Each variation assumes one of three
MOIs and permits an estimate of the individual loci MOI. The MOI prediction
is then used in subsequent stages to improve statistical power.
The results indicate that
a network of SNPs that links to a given phenotype can be identified if (and
only if) the study size is sufficient;
interaction effects can be addressed in this process but may require a larger
N in their detection;
 Polygene Methods in Genome-Wide Association Studies 139

the statistical power of the method may be sensitive to the level of LD; and
a network consisting of very low-effect loci can be predicted accurately if the
study size, N, is sufficient, but the study size needed is far larger than that
employed by traditional GWAS. For example, to detect very low-effect SNPs,
we used N = 400,000. This is well beyond what is conceived as practical, but
future research and improvements might make it possible.
Many polygene methods seek to investigate a large number of SNP
combinations. These approaches are computationally demanding and often
impractical to implement. Even with widespread recognition that single-locus
tests are likely to be inferior to multilocus tests for GWAS of many diseases
and phenotypes, an unresolved issue is how to construct a computationally
practical test that takes into account interactions and enhances the detection of
associations between multiple loci and the phenotype of interest.
The main value of our study is that it proposes computations that are
practical and straightforward and require less computational effort than those
reviewed by others.3 Furthermore, our simulation studies suggest that if the
associations are real, they can be found with properly powered studies. An
additional feature of our study is that as part of the Stage 1 SNP selection, the
prediction defines the inheritance properties of all SNPs, which our simulation
experiments have shown to be accurate and can be used to improve the
reliability of ensuing stage predictions.

Appendix: DNA Simulator Tools

A number of different coalescent simulators that generate simulated DNA
fragments for different evolutionary models are discussed in the literature
that differ from the approach we describe in the main text. These applications
simulate different standard neutral evolutionary models with recombination,
variable population size, and migration. They also allow for spatial and
temporal environmental heterogeneity. Carvajal-Rodrguez provide a
comprehensive description of many of these models.32 The most recent and
one of the most versatile models is not included in his analysis. GWAsimulator
is a program that simulates genotype data for SNP chips that are used in
GWAS.33 It creates whole genome case-control or population samples. It also
can simulate specific genomic regions. For case-control data, the program can
be linked to sample cases and controls according to a user-specified multilocus
disease model. The program requires phased data as input, and the simulated
data will have similar LD patterns as the input data.
140 Chapter 8

A second program worth mentioning is GENOME.34 It simulates a wider

range of scenarios including recombination hotspots. As well as whole genome
data. In addition to features of standard coalescent simulators, the program
allows for recombination rates to vary along the genome and for flexible
population histories. The program and C++ source code are available online at
http://www.sph.umich.edu/csg/liang/genome/.

Chapter References
1. Visscher PM, Brown MA, McCarthy MI, et al. Five years of GWAS
discovery. Am J Hum Genet. 2012;90(1):7-24.
2. Chatterjee N, Wheeler B, Sampson J, et al. Projecting the performance of
risk prediction based on polygenic analyses of genome-wide association
studies. Nat Genet. 2013;45(4):400-5, 405e1-3.
3. Wang Y, Liu G, Feng M, et al. An empirical comparison of several recent
epistatic interaction detection methods. Bioinformatics. 2011;27(21):2936-
43.
4. Zhang X, Huang S, Zou F, et al. TEAM: efficient two-locus epistasis tests in
human genome-wide association study. Bioinformatics. 2010;26(12):i217-
27.
5. Cooley P, Clark R, Folsom R, et al. Genetic inheritance and genome
wide association statistical test performance. J Proteomics Bioinform.
2010;3(12):321-325.
6. Schymick JC, Scholz SW, Fung HC, et al. Genome-wide genotyping in
amyotrophic lateral sclerosis and neurologically normal controls: first
stage analysis and public release of data. Lancet Neurol. 2007;6(4):322-8.
7. Plomin R, Simpson MA. The future of genomics for developmentalists.
Dev Psychopathol. 2013;25(4 Pt 2):1263-78.
8. Benyamin B, Pourcain B, Davis OS, et al. Childhood intelligence is
heritable, highly polygenic and associated with FNBP1L. Mol Psychiatry.
2014;19(2):253-8.
9. Stranger BE, Stahl EA, Raj T. Progress and promise of genome-
wide association studies for human complex trait genetics. Genetics.
2011;187(2):367-83.
 Polygene Methods in Genome-Wide Association Studies 141

10. International Schizophrenia Consortium, Purcell SM, Wray NR, et al.

Common polygenic variation contributes to risk of schizophrenia and
bipolar disorder. Nature. 2009;460(7256):748-52.
11. Yang J, Benyamin B, McEvoy BP, et al. Common SNPs explain a
large proportion of the heritability for human height. Nat Genet.
2010;42(7):565-9.
12. Gibson G. Hints of hidden heritability in GWAS. Nat Genet.
2010;42(7):558-60.
13. Mackay TF, Stone EA, Ayroles JF. The genetics of quantitative traits:
challenges and prospects. Nat Rev Genet. 2009;10(8):565-77.
14. Dudbridge F. Power and predictive accuracy of polygenic risk scores. PLoS
Genet. 2013;9(3):e1003348.
15. Plomin R, Haworth CM, Meaburn EL, et al. Common DNA markers can
account for more than half of the genetic influence on cognitive abilities.
Psychol Sci. 2013;24(4):562-8.
16. Cooley PC, Clark RF, Folsom RE. Statistical methods that identify
genotype-phenotype associations in the presence of environmental effects.
RTI Press Publication No. RR-0022-1405. Research Triangle Park, NC:
RTI Press; 2014.
17. Culverhouse R, Suarez BK, Lin J, et al. A perspective on epistasis: limits of
models displaying no main effect. Am J Hum Genet. 2002;70(2):461-71.
18. Hoh J, Wille A, Zee R, et al. Selecting SNPs in two-stage analysis of disease
association data: a model-free approach. Ann Hum Genet. 2000;64(Pt
5):413-7.
19. Li J, Horstman B, Chen Y. Detecting epistatic effects in association studies
at a genomic level based on an ensemble approach. Bioinformatics.
2011;27(13):i222-9.
20. Sha Q, Zhang Z, Schymick JC, et al. Genome-wide association reveals
three SNPs associated with sporadic amyotrophic lateral sclerosis through
a two-locus analysis. BMC Med Genet. 2009;10:86.
21. Moore JH, Ritchie MD. STUDENTJAMA. The challenges of whole-
genome approaches to common diseases. JAMA. 2004;291(13):1642-3.
22. Bush WS, Moore JH. Chapter 11: Genome-wide association studies. PLoS
Comput Biol. 2012;8(12):e1002822.
142 Chapter 8

23. Bush WS, Dudek SM, Ritchie MD. Biofilter: a knowledge-integration

system for the multi-locus analysis of genome-wide association studies.
Pac Symp Biocomput. 2009:368-79.
24. Herold C, Steffens M, Brockschmidt FF, et al. INTERSNP: genome-wide
interaction analysis guided by a priori information. Bioinformatics.
2009;25(24):3275-81.
25. Cooley P, Gaddis N, Folsom R, et al. Conducting genome-wide association
studies: epistasis scenarios. J Proteomics Bioinform. 2012;5(10):245-251.
26. Wu X, Dong H, Luo L, et al. A novel statistic for genome-wide interaction
analysis. PLoS Genet. 2010;6(9):e1001131.
27. Ueki M, Cordell HJ. Improved statistics for genome-wide interaction
analysis. PLoS Genet. 2012;8(4):e1002625.
28. McKusick-Nathans Institute of Genetic Medicine, Johns Hopkins
University School of Medicine. Online Medelian Inheritance in Man
(OMIM). 2016 [cited 2016 Feb 11]; Available from: http://www.ncbi.nlm.
nih.gov/omim
29. Lin CY, Xing G, Xing C. Measuring linkage disequilibrium by the partial
correlation coefficient. Heredity (Edinb). 2012;109(6):401-2.
30. Abecasis GR, Noguchi E, Heinzmann A, et al. Extent and distribution
of linkage disequilibrium in three genomic regions. Am J Hum Genet.
2001;68(1):191-197.
31. Eunice Kennedy Shriver National Institute of Child Health and Human
Development. Add Health: The National Longitudinal Study of Adolescent
ot Adult Health. 2015 [cited 2015 July 20]; Available from: http://www.cpc.
unc.edu/projects/addhealth
32. Carvajal-Rodriguez A. Simulation of genomes: a review. Curr Genomics.
2008;9(3):155-9.
33. Li C, Li M. GWAsimulator: a rapid whole-genome simulation program.
Bioinformatics. 2008;24(1):140-2.
34. Liang L, Zollner S, Abecasis GR. GENOME: a rapid coalescent-based
whole genome simulator. Bioinformatics. 2007;23(12):1565-7.
 CHAPTER 9

Conclusions and Recommendations

Philip Chester Cooley

Purpose of the Manuscript

Our explorations into genome-wide association studies (GWAS) led us to
believe that a number of governing principles were absent from the manner in
which GWAS were conducted. The descriptions of methods that researchers
have provided in the literature commonly assumed an additive inheritance
model or were agnostic with respect to an inheritance model assumption.
Furthermore, there were many examples in which GWAS provided
inconsistent results and researchers could not replicate study findings. Our
own experiences investigating GWAS indicated that different statistical
methods assuming specific inheritance properties also produced different
association results. Because the inheritance properties are generally not
known a priori, we favored a statistical model that was agnostic with respect to
inheritance, and we turned to simulation studies to confirm our beliefs.
We chose simulation because we noted the absence of both methodological
standards and a process for evaluating statistical methods used in predicting
associations between genes and phenotypes. We believed that that we could
examine these missing elements more effectively with a simulation approach.
Accordingly, we created a simulated data set of virtual genes that were linked
to known outcomes. These data constituted a truth set. We analyzed the
simulated data using different statistical methods and used it to assess each
methods predictive properties. This process revealed a number of principles
described subsequently.
144 Chapter 9

Summary of the Methodology

We developed a database of synthetic/virtual genetic loci. We assigned
Mendelian inheritance properties to those loci and connected different
combinations/networks of loci to virtual phenotypes. The phenotypes were
qualitative traits intended to represent the positive or negative diagnoses of
specific diseases. We also introduced errors into the recording of the genotypes
and the assignment of diagnoses, and we simulated the strength of the
genotype to phenotype signal via a relative risk process. For each inheritance,
error, and relative-risk setting combination, we assigned a targeted number of
cases and controls to include in an experimental GWAS design. The number
of cases was varied in an effort to encapsulate a statistical power region
containing the value of 0.8.
Having generated these data, we then applied a number of distinct GWAS
statistical methods and models that were germane to the specific GWAS, and
we used them to predict an association. We replicated each unique GWAS
design by statistical prediction model 1,000 times.
Each database entry, including performance measures, was catalogued
with respect to the individual data entries and the properties that we used
to generate GWAS data. An analysis of these measures across the replicates
permitted performance assessment of the different statistical methods to
predict genotype-to-phenotype associations.
Our experiments were not limited to genetic mechanisms. We also
generated database entries that incorporated environmental influences into the
GWAS experimental designs that were used to simulate different mechanisms
such as aging and chemical exposure.
We also investigated loci that depended on the properties of other loci to
produce phenotypic elevated risk levels. The effect of one gene that depends on
the presence of one or more modifier genes is known as epistasis, and others
have shown that epistatic mutations have different effects in combination than
individually because of interactions between genes or within them that lead to
nonadditive effects.1
By design, every entry contained in the database was produced with a
positive genetic association risk. The risks were varied between 1.01 and 1.50
and were kept small by design to test the sensitivity of the statistical models. To
assess method performance, we applied many different and varied statistical
methods to determine which class of models exhibited the best power
properties. Because all entries had a positive risk, each and every method
 Conclusions and Recommendations 145

should have predicted an association. Statistical power was determined by

computing the percentage of positive predictions across the 1,000 replicates.
Graphing the power profiles as a function of GWAS subject participants
enabled us to estimate the number of subjects required to achieve a specific
power goal and also illustrated how errors in the data, signal strength,
environment, and polygene behaviors affected model predictions.

Findings and Recommendations

The analysis described previously has led us to formulate the following
assertions:
Developing a synthetic gene database that recorded known outcomes
between synthetic phenotypes and genotype networks provides a
mechanism that was not possible using real genomics data. Of course it has
been asserted that real data and data generated by simulation experiments
may not have the same qualities (i.e., real data will contain more noise than
simulated data). Our response was to provide a mechanism that generates
data that includes a level of noise that can be specified a priori.
The creation of this database enabled an evaluation of different statistical
models and methods specifically because the prediction outcomes were
known and statistical power profiles could be estimated.
Methods that treat combinations of genes acting in concert either with other
genes or environmental factors are important investigations in a GWAS
context because a single-gene model will fail to identify markers in many
types of gene-gene, and gene-environment networks.
In response to this principle, we proposed a general polygene test (Chapter 8)
that is based on a procedure that accepts (or not) a positive association of
the n+1 SNP given the presence of a set of n previously confirmed SNPs. In
the absence of a test that considers the combined effect of multiple genes, the
association process will be dominated/obscured by the strongest (dominant)
associated locus.
A reliable single-gene model is still necessary to identify a starting point/
filter in a polygene process, but the influence of the inheritance properties
of the locus should be considered in the selection of the statistical model.
There can be a substantial power loss if the inheritance property of the loci
is assumed to be additive (or log-odds additive), and the locus is dominant
(which results in a modest power loss) or recessive (which results in a
146 Chapter 9

substantial power loss). Of course, if the gene model is additive, a model that
assumes additive behavior is optimal with respect to power loss.
Making an informed assessment of the inheritance properties of the locus
is possible. Applying multiple tests that each assume a distinct inheritance
property and selecting the property that produces the lowest p-value is one
possibility. Knowing the inheritance will improve the performance of the
polygene process.
If we assume error-free phenotype data, we may need to supplement
the number of case/control subjects to achieve desired power targets
(substantially, if the locus is recessive). Adding noise to both phenotype and
genotype measures is informative in this context.
The GWAS process is exploratory in nature, and various tools, such as odds
ratio confidence regions and pseudo R-squared (R2) measures, are useful to
the exploration. However, odds ratio confidence regions and pseudo R2 are
limited to logistic regression model explorations.
It is possible to use the public databases that link SNPs to genes to diseases/
conditions and to develop interesting and plausible explanations of
genotype-phenotype linkages. In our experience, computing cell counts
often exposes potential biases due to low frequency of disease alleles.

Future Directions
GWAS have identified many new genetic risk factors for a number of
common human diseases, but much work remains to be done and can only be
accomplished by using new approaches. Furthermore, new technologies are
coming, such as whole-genome sequencing, which will replace the 1 million
SNP chip data with the entire genomic sequence of 3 billion nucleotides.
This will have a huge impact on data storage, manipulation, quality control,
and data analysis processes. New computer science and bioinformatics tools
will be needed, and cloud operations will become the new infrastructure.
Also, high-throughput technologies for measuring the transcriptome, the
proteome, the environment, and the whole-genome sequences will become
standard operating procedures. New phenotypes from the development of new
technologies such as neuroimaging will also become available and add to the
level of complexity.
 Conclusions and Recommendations 147

In this environment, unravelling genotype-phenotype relationships for the

purpose of improving health care will expand. Integrating these varied and
complex biomedical data and findings is the future of human genetics.

Chapter References
1. Otto SP, Feldman MW. Deleterious mutations, variable epistatic
interactions, and the evolution of recombination. Theor Popul Biol.
1997;51(2):134-47.
 Acknowledgment
The development of this book was made possible by the generous support
of the RTI Fellow program.
 Contributors
Philip Cooley, MSc, is a senior fellow at RTI estimation research. In addition to his innovative
International in Computational Biology and High work on many complex survey efforts, Dr. Folsom
Performance Computing. He has more than has made significant contributions to the
47 years of experience in developing computer development of RTIs computer software for
models used in the study of environmental survey data analysis, SUDAAN.
health and disease transmission scenarios. He has Nathan Gaddis, PhD, is a research programmer/
developed infectious disease models to study the analyst in RTIs Research Computing Division. He
transmission of a number of diseases including has a diverse background in molecular biology
malaria, tuberculosis, HIV/AIDS, and influenza. He and bioinformatics, combining laboratory
is currently developing a process to forecast research experience in microbiology,
stunting in Indonesia, and his research focus is immunology, and genetics with genome-scale
shifting to the development and application of analyses, including genetic analyses for GWAS
models that explore and predict the impending and transcriptome-wide analyses of alternative
epidemic of chronic diseases. splicing. Currently, Dr. Gaddis is a co-investigator
Robert F. Clark, PhD, is a senior genetic on several GWAS projects, the lead developer for
epidemiologist in RTI Internationals Genetic the National Heart, Lung, and Blood Institute-
Epidemiology and Omics research program. funded LungMAP project, and a developer for the
Throughout most of his career, he has focused on National Human Genome Research Institute-
multidisciplinary work in the omics and systems funded PhenX project.
biology of various multifactorial disorders, and Grier Page, PhD, is a senior statistical geneticist
since 1992, he has conducted many genetic in RTIs Genomics and Statistical Genetics
studies of neurodegenerative diseases; breast, Research unit. He has been conducting research
bone, and brain cancers; nicotine, heroin, and in all areas of statistical genetics since 1993 and
cocaine addiction; and aortic aneurysms. in systems biology since 2002. Dr. Page has
Ralph E. Folsom, PhD, chief scientist in RTIs developed methods for the analysis of linkage
Division of Statistical and Data Sciences, is an and association data as well as microarray,
expert in the design and analysis of complex proteomic, and next-generation sequencing
probability samples. Working on the nations methods. Dr. Page has also developed numerous
largest household survey (the National Survey on bioinformatics tools to implement the methods
Drug Use and Health or NSDUH), Dr. Folsom he has developed, including the PowerAtlas
initiated innovative weight adjustment methods (http://www.PowerAtlas.org), HDBStat! (http://
based on his logistic response propensity and www.ssg.uab.edu/HDBStat), and CressExpress
exponential poststratification models. This (http://www.cressexpress.org).
pioneering work led to the sophisticated GEM Diane Wagener, PhD, was an epidemiologist at
weight adjustment methods currently employed RTI International. She has 37 years of experience
for NSDUH. Dr. Folsom also introduced in academia, government, and consulting
model-based imputations for missing frequency studying the causes, genetics, and social impact
of use and income data items, and he has been on a number of diseases, both in the United
an influential collaborator in the development of States and internationally. Her expertise is in
NSDUHs current Predictive Mean Neighborhoods research design and statistical and computational
(PMN) imputation methodology. Dr. Folsom has analysis of the data.
recently led RTIs innovative work in small area
 Index

A ALS2, gene linked to ALS pathogenesis in

A (allele without risk) familial ALS and, 17
environmental influencing factors studies, 92 alternate splicing, qualitative trait analysis and,
epistatic model generation and, 7172 6566
epistatic models, 70t amyotrophic lateral sclerosis (ALS)
penetrance in genetics and, 31 applying classical statistical methods to,
synthetic gene database generation and, 31 1719
a (risk allele) GWAS replicability issues, 67
environmental influencing factors studies, 92 other GWAS studies, 2526
epistatic model generation and, 7172 results compared to Schymick data, 2225
epistatic models, 70t Schymick data set, 1920
penetrance in genetics and, 31 single-gene models and, 122
synthetic gene database generation and, 31 statistical tests, 2022
additive Cochran-Armitage (CA-A) (Tr-A) test ANG, gene linked to ALS pathogenesis in
errors assessment and, 5354 sporadic ALS and, 17
genotype and diagnosis errors impact on ApoE, gene linked to coronary artery disease,
power and, 57f 68
GWAS statistical test performance, 4243,
43t B
power results, for different MOI gene B (allele without risk), epistatic model
models, 45t generation and, 7172
predicting additive and multiplicative SNPs b (risk allele), epistatic model generation and,
and, 45 7172
prediction accuracy, 21t, 22 Barendse W, 51
Schymick vs. Cooley results, 23, 24t, 25 bias, environmental influencing factors studies
single-gene epistatic models, 7273, 75t and, 94
2-degrees-of-freedom genotype association biobanks, for genetic studies, 3
test vs., 50 biochemical pathways, examining SNP
additive mode of inheritance SNPs, prediction combinations within, 12223
accuracy and, 9 Biofilter approach, to SNP-SNP combinations,
age as environmental risk (E) 123
environmental exposure risk and, 9697, biological pathways
97t, 101 of polygenic diseases and traits, 6, 122
environmental influence statistical models, use of term, 122
98 biomarkers. See also environmental influencing
logistic regression studies adjusted for, 89 factors studies
macular degeneration-related to, 3, 4 disease scoring and, 121
main effect (ME) test and, 110 growing number of, case-control GWAS
Ahn K, 50 and, 4
allele test identifying, 2
results comparison, 23, 24t, 25 Schymick vs. Cooley results, 23, 24t, 25
on Schymick ALS data set, 21t, 22 variants, 56

Note: Page numbers followed by t and f indicate tables and figures.

154 Index

bladder cancer correlation coefficient and linkage

a codon and, 88 disequilibrium, 127
smoking and, 89 Cronin S, 25
Blauw HM, 19, 25 crossover risk
body mass index, logistic regression studies dominant vs. recessive gene, 77f
adjusted for, 89 genetic inheritance of interacting loci and,
Bonferroni correction procedure, 37, 42 122
genetic inheritance and GWAS test Cummings M, 70
performance and, 4243, 43t
MAX test vs., 46 D
multitest studies of environmental Daoud H, 19
influencing factors and, 112 data mining, 90, 101, 104
polygene analysis, 125 Deng M, 19
predicting associations in GWAS context diabetes, type 2, gene-environment
and, 9 interactions and, 8990
Bush WS, 12223 diagnosis errors. See phenotype errors
dipeptidyl-peptidase 6 (DPP6) gene, ALS and,
C 18
Carvajal-Rodrguez A, 139 disease. See also phenotypes
case-control assignments, polygene analysis determining genetic variations associated
and, 12830, 129t with, 2
case-control genotype method, based on disease penetrance (P)
Pearson 2 test, 2021 environmental influencing factors studies,
case-control GWAS, errors and, 5052 92
Catalog of Published Genome-Wide Association epistatic model generation and, 7172
Studies, 3, 67 errors assessment, 53, 54
Center for Inherited Disease Research, 51 odds, polygene analysis and, 128, 129t
Centers for Disease Control and Prevention, 89 synthetic gene database and, 31
Chan EK, 32, 92 synthetic gene database generation and,
chemical spill 3234
environmental exposure, 91 disease prevalence in nonrisk populations,
environmental influencing factors studies epistatic models and, 68
and, 97, 97t disease-scoring methods, 121
main effect (ME) test and, 110 DNA simulator tools, 13940
Chi A, 18, 25 dominant Cochran-Armitage (CA-D) (Tr-D)
chromosome 9p21, ALS and, 19 test
Cochran-Armitage (CA) trend test error assessment and, 5354
epistatic models, 7273, 73t genotype and diagnosis errors impact on
errors assessment and, 5354 power and, 56f
GWAS statistical test performance, 4041, prediction accuracy, 21t, 22
42 Schymick vs. Cooley results, 23, 24t, 25
multiple test procedure, 39 single-gene epistatic models, 7273, 75t
Schymick ALS data set, 21t, 22 X2, in GWAS statistical test performance,
Cochran-Armitage composite (CA-C) test 4243, 43t
pooling three MOI outcomes and, 11718, X2, predicting dominant MOI SNPs and, 45
138 dominant Fishers exact test (Fis-D), normal
in Stage 1 polygene analysis, 124 approximation to
Cooley P, 54, 92, 123 results comparison, 23
Cordell, HJ, 112 Schymick ALS data set, 21, 21t, 23, 24t, 25
Cornelis MC, 90, 112 dominant mode of inheritance SNPs,
coronary artery disease, ApoE gene in, 68 prediction accuracy and, 9
Dunckley T, 18, 25
 Index 155

E interaction (INT) test on, 108t, 110t

Edwards BJ, 50, 51 power curves of genetic relative risk and
The Elements of Heredity (Johannsen), 1 environmental exposure, 101, 103f
environmental exposure (EE) power values, genetic relative risk, and
genetic inheritance risk by statistical model environmental risk by, 103t
and, 9798, 98t total effect test on, 105, 107t
low levels of gene by environmental variables used in, 9697, 96t
exposure (GI-EE) interactions and, environmental risk interaction effects log-
110t, 111 linear risk model
power values, by statistical model, genetic influence by gene-only model and, 107
relative risk and, 102f genotype-environment interactions (INT)
step height, log-odds-risk and, 101, 102f test on, 108t, 110t
total effect test, 107t power curves of genetic relative risk and
type of, association detection and, 108, 109 environmental exposure, 104, 105f
environmental influencing factors studies power values, genetic relative risk, and
association analysis, 98101, 99t, 100f, 102f, environmental risk by, 104t
102t, 103f, 103t, 104, 104t, 105f total effect test on, 105, 107t
gene-gene interaction studies, 8891 variables used in, 9697, 96t
generating synthetic SNP data for, 9296 epistatic model
genotype associations, 10511, 107t, 108t, Cochran-Armitage trend test on, 7273,
109t, 110t, 111t 75t, 76t, 78, 80
GWAS experimental designs and, 12, 144 dominant crossover risk, 77f, 79f
statistical models, 9698, 96t, 97t, 98t computational models, 69, 69t, 7172
environmental risk main effects model recessive crossover risk, 77f, 79f
experiment description 95 double mutant, in epistatic models, 69, 69t,
fixed risk, main effects, no interaction 7172
model, 95, 97t environment-wide association study (EWAS),
fixed-risk, main effect with interaction of type 2 diabetes, 89
model, 95, 96, 97t epigenetics, qualitative trait analysis and, 6566
gene-environment interactions and, 109, epistatic models (EMs)
109t as analytic tool, 6768, 144
gene by environmental exposure (GI-EE) environmental influencing factors studies
interaction effect and, 104 and, 113
genotype-environment interactions (INT) generation of synthetic SNP data, 7172
test on, 108t methods, 6874, 69t, 70t
power values, by statistical model, genetic results, 7476, 75t, 76t, 77f, 78, 78t, 79f, 80
relative risk, and environmental risk, statistical models, 7274
99101, 99t, 100f testing for associations with interacting loci,
total effect test on, 105, 107t 1112
variables used in, 9697, 96t errors
environmental risk interaction effects model genotype and phenotype misclassifications,
influence by gene-only model and, 107 power threshold and, 1011
power curves of genetic relative risk and GWAS, 118
environmental exposure, 102f polygene analysis, 144, 146
power values, genetic relative risk, and polygene analysis examples, 12830, 129t,
environmental risk by, 102t 132
total effect test on, 105, 107t low-effect loci in, 13738, 137t
variables used in, 9697, 96t model with interactions, 13435, 135t
environmental risk main effects log-linear risk exogenous risk (aa), in environmental
model influencing factors studies, 93, 94
gene-environment interactions and, 109,
109t
156 Index

F genetic main effects (ME) test

familial amyotrophic lateral sclerosis, 17 elements in, 106
Feingold E, 38, 65, 86 power values, by risk profile for all gene
Fishers exact test, normal approximation to models, 109t, 11011
results comparison, 23, 24t, 25 genetic markers. See biomarkers
goodness of fit, in polygene analysis, 130 genetic relative risk (). See also mode of
FLJ10986 gene, as ALS genetic marker, 18 genetic inheritance
Framingham Heart Study, 127 definition, 91
Freidlin B, 51 environmental exposure (EE) in single gene
statistical model and, 101
G environmental exposure risk by statistical
Gastwirth JL, 22, 40, 72 model and, 99t
gene by age environmental effect (M-2) environmental exposure risk level and, 99
statistical model environmental influencing factors studies,
assessed, 98t 92, 93t, 94, 95
environmental exposure and, 100 epistatic model generation and, 7172
M-3 power profiles in Experiment 3 vs., 101 heterozygote (aA), 34t, 93t, 94
M-3 power profiles in Experiment 4 vs., 104 minor homozygote (aa), 3233, 34t, 93, 93t
power values, genetic relative risk, and in polygene analysis, 14445
environmental risk by, 99t power values, by statistical model,
gene by toxic exposure environmental effect environmental exposure risk and, 102f
(M-3) statistical model synthetic gene database generation and, 32
additive gene model power curves, 100f total effect test, 107t
assessed, 98t genetic testing, warfarin dosages and, 4
power profile, 100 genetic variants
power values, genetic relative risk, and epistatic models and, 66
environmental risk by, 99t marginal effects studies, 12021
gene expression levels, qualitative trait analysis GENOME, simulator tool 140
and, 6566 genome-wide association studies (GWAS)
gene properties, statistical model performance determining disease associations using, 24
and, 8, 9 errors in, single-gene models and, 1011
gene-environment interactions. See improvement methods, 117
environmental influencing factors studies inheritance model assumptions, 143
gene-only (M-1) statistical model limitations, environmental influencing
additive gene model power curves, 100f factors and, 107
assessed, 98t missing heritability and, 56
GI-EE interaction effect and, 101 polygene methods in, 13
power profile, 100 power profiles comparison, 910
power profiles for large EE risk levels in, 104 statistical detection of weak genetic effects, 3
power values, genetic relative risk, and as useful or misleading? 45
environmental risk by, 99t genomic distance (D), in polygene analysis,
genes. See also loci 12930
biochemical pathway to proteins from, 1 genotype. See also single polymorphism arrays
genetic inheritance and GWAS test diagnosis errors and, 113
performance distribution set from SNP data, 71
discussion, 4546 phenotype vs., 1
introduction, 3839 similarities, mating and, 127
methods, 3942 genotype errors
overview, 3738 GWAS simulated data and, 53, 54
results (assessment), 4245 maximum power loss due to, 59t
genetic linkage studies, reproduction models of, 5051
difficulties, 3 percent sample size increase to restore
power and, 61t
 Index 157

genotype errors (continued) inheritance. See also mode of genetic

recessive MOI, statistical power by relative inheritance
risk and, 60f ALS type and, 17
single-gene models and, 1011 GWAS on predicting, 6
statistical power and, 55f properties, in epistatic models, 68
genotype-phenotype linkages, cell counts and, sporadic cancer and, 88
146 synthetic gene database and, 31
Gilbert-Diamond D, 68 inositol 1, 4, 5-triphosphate receptor 2
Goldstein DB, 6 (ITPR2) marker, ALS and, 18
Gordon D, 50, 61 genotype-environment interactions (INT) test,
GWAsimulator, 139 106, 1089, 108t
intelligence quotient (IQ), genome-wide meta-
analysis of, 120
H
International Consortium on Amyotrophic
Hao K, 5051
Lateral Sclerosis, 19
HardyWeinberg equilibrium (HWE)
International HapMap Project, 20, 51, 126
deviation
INTERSNP tool, for genome-wide interaction
environmental influencing factors studies
analysis (GWIA), 123
and, 92
synthetic gene database generation and, 32
height heritability, polygenic model of, 120 J
Jewell NP, 73
heritability. See inheritance; missing
Johannsen, Wilhelm, 1
heritability; mode of genetic inheritance
heterozygotes (aA) (g1)
as controls in GWAS statistical performance K
KIFAP3, gene linked to ALS pathogenesis in
test, 39
sporadic ALS and, 19
in environmental influence statistical
Klug W, 70
models, 98
Kraft P, 6, 90
environmental influencing factors studies
Kuo CL, 38, 65, 86
and, 93
Kwee LC, 19
for errors assessment, 52
synthetic gene database generation and,
3334, 34t L
Laaksovirta H, 19
HFE, gene linked to ALS pathogenesis in
Laboratory of Neurogenetics, 126
sporadic ALS and, 17
Landers JE, 19
high-density lipoprotein cholesterol (HDL-C),
Laurie CC, 51
multiple genes linked to GWAS and, 45
Li Q, 41, 42, 46, 112
Hirshhorn JN, 6
Lichtenstein P, 88
Holm S, 42, 112
likelihood ratio test, disease and locus
homozygotes. See major homozygotes; minor
association study, 9091
homozygotes; wild homozygotes
Lindstrom S, 89
Human Genome Project (HGP), 2, 117
linkage disequilibrium (LD) analysis
other ALS GWAS studies, 25
I
performance by 0.0 SNP data and, 132f
Iles MM, 31, 32, 71, 92
performance by 0.2 SNP data and, 133f
Illumina Human1Mv1_c and HumanHap550-
performance by 0.4 SNP data and, 133f
2v3_b arrays, 51
polygene analysis and, 12627, 128, 129t,
Illumina Infinium assay humanhap550, 1920,
139
126
prediction accuracy by model and, 134t
independent assortment principle, in genetic
loci
inheritance, 126
epistatic relationship of, 68
use of term, 67
158 Index

logistic regression (LR) models major homozygotes (AA) (BB)

of environmental influencing factors, 8991 controls in GWAS statistical performance
INTERSNP use of, 123 test, 39
in polygene analysis, 125, 138 environmental influencing factors studies
results comparison, 23, 24t, 25 and, 93
on Schymick ALS data set, 21, 21t synthetic gene database generation and,
log-likelihood (LLH) statistics 3233, 34t
environmental influencing factors studies, Marchini J, 66
98 Mariana Island form amyotrophic lateral
polygene analysis, 125 sclerosis, 17
log-linear variable risk main effects model, masking effect on index locus
95, 96 low-effect loci in single-gene models and,
log-linear variable risk main effects with 67, 76, 80
genotype interaction model, 95, 96, 97t in single-gene models, 122
log-linear-logistic model framework, of MAX test (method)
qualitative trait responses, 11819 Bonferroni procedure vs., 46
log-odds-risk, environmental exposure risk genetic inheritance and GWAS test
and, 101 performance and, 37
low-density lipoprotein cholesterol (LDL-C), GWAS statistical test performance, 4243,
nonsynonymous variant gene identified, 5 43t
low-effect loci multitest studies of environmental
detecting, 13738, 137t, 139 influencing factors and, 112
epistatic models and, 66 power results, for different MOI gene
masking effect on index locus in single-gene models, 45t
models and, 67 predicting associations in GWAS context
lung cancer, smoking and, 88 and, 910
Mendel, Gregor, 126. See also mode of genetic
M inheritance
M-1 statistical model. See gene-only (M-1) Miclaus K, 51
statistical model minor allele frequency (MAF) threshold
M-2 statistical model. See gene by age environmental influencing factors studies
environmental effect (M-2) statistical and, 92
model synthetic gene database generation and, 32
M-3 statistical model. See gene by toxic minor homozygotes (aa) (g2) (bb)
exposure environmental effect (M-3) controls in GWAS statistical performance
statistical model test, 40
macular degeneration, age-related environmental influence statistical models,
GWAS of, 3 98
odds ratio for GWAS identification of, 4 environmental influencing factors studies
main effects (ME) test and, 93
elements in, 106 for errors assessment, 52
power values, by risk profile for all gene synthetic gene database generation and,
models, 109t, 11011 3234, 34t
main effects only model missing heritability
current GWAS methods and, 88 autoimmune diseases and, 56
environmental influencing factors, 85, 95, environmental influencing factors and, 12
96, 97t, 113 epistatic models and, 6667
main effects with interactions model mode of genetic inheritance (MOI)
environmental influencing factors, 85, 95, additive model, nonadditive SNP data and,
96, 97t, 113 39
single-gene models vs., 145 assumptions, in GWAS, 143
Cochran-Armitage test and, 21t, 22
 Index 159

mode of genetic inheritance (MOI) (continued) Pearson correlation coefficient 22, 127, 128
environmental influencing factors studies, Pearson 2 test
92, 93t allele test and, 22
epistatic models of, 6970, 69t, 70t case-control genotype based on, 2021, 21t
for errors assessment, 5253, 54 replicating Schymick et al using, 7
genotype errors vs. phenotype errors and, results comparison, 23, 24t, 25
59 on two-gene epistatic models, 73, 76, 80
GWAS statistical test performance, 8-10, penetrance (P). See disease penetrance
40, 40t personalized medicine predictions, 4
log-additive model, 90 phenotype errors
polygene analysis, 128, 129t, 144, 146 GWAS simulated data and, 53, 54
predictions, in polygene analysis, 13132, maximum power loss due to, 59t
131t, 132f, 133f, 134, 134t, 13839 models of, 5152
SNP-specific, CA-C prediction of, 124 percent sample size increase to restore
synthetic gene database generation and, power and, 61t
3234 random misclassification, 50
multiplicative MOI gene data recessive MOI, statistical power by relative
error effects of GWAS simulated data and, risk and, 60f
53, 54 single-gene models and, 11
genotype and diagnosis errors impact on statistical power and, 55f
power and, 58f phenotypes
GWAS statistical test performance, 44t association detection of recessive genes, 107
Murcray CE, 90 combined genotype-environmental factors
and, 98101, 99t, 100f
environmental effects only studies of, 90
N
genotype vs., 1
National Health and Nutrition Examination
MOI gene model associated with, 9
Survey (NHANES), 89
poorly defined, ALS as, 25
National Human Genome Research Institute, 4
similarities, mating and, 127
National Institute on Aging (NIA), National
polygene methods (analysis)
Institutes of Health (NIH), 126
background, 12023
National Longitudinal Study of Adolescent
detecting low-effect SNPs, 13738
Health, 127
discussion, 13839
neuroimaging, 146
disease-scoring methods, 12122
New England Journal of Medicine, 6
general test, 14
generating SNP data, 12530
O
methods, 12430
odds ratio confidence regions, 146
model comparisons and MOI predictions,
odds ratio (OR). See also SNP odds ratio data
13134
definition, 4
models with interactions, 13437
in polygene methods, 119
overview, 11719
1 degree of freedom (1df) test, 22
positive association of n+1 SNP test, 145
gene-only log-additive test, 97, 98t
results, 13138
Online Mendelian Inheritance in Man
statistical approaches to, 12123
(OMIM), 3, 40, 126, 126t
stepwise algorithm, 13, 12425
ovarian cancer, variants influencing risk of,
polynomial level for environmental effects, 112
8889
PON1, gene linked to ALS pathogenesis in
sporadic ALS and, 17, 25
P
positive genetic association risk, in polygene
Patel CJ, 8990
analysis, 144, 145
PAWE-3D (statistical power calculator), 50
product moment, as correlation coefficient. See
Pearce CL, 88
Pearson correlation coefficient, 127
160 Index

protein families, examining SNP combinations S

within, 12223 sample sizes
protein-folding processes, qualitative trait environmental influencing factors studies
analysis and, 6566 and, 9192
pseudo R-squared (R2) measures, 146 genotype and diagnosis errors and, 49
Purcell S, 72 larger, polygene analysis and, 135, 136f,
136t, 137, 138
Q low-effect SNPs and, 13738, 137t, 139
qualitative association framework, of Sasieni PD, 51
environmental influencing factors, 9192 schizophrenia gene SCZ, polygene small-effect
qualitative trait analysis, GWAS and, 6566 SNP model of, 120
quantitative genetic techniques, of heritability Schymick JC et al synthetic ALS data set
using DNA, 121 comparison to results of, 23
polygene analysis using, 126
R replicating, 7
race, logistic regression studies adjusted for, 89 statistical predictive strength and, 17, 18,
recessive Cochran-Armitage (CA-R) (Tr-R) 1920
test segregation principle, in genetic inheritance,
error effects on statistical power by relative 126
risk, 60f sequencing, improved methods in, 3, 121
errors assessment and, 5354 SETX, gene linked to ALS pathogenesis in
genotype and diagnosis errors impact on familial ALS and, 17
power and, 55f sex, logistic regression studies adjusted for, 89
in GWAS statistical test performance, Shatunov A, 19
4243, 44t sickle-cell anemia, as single nucleotide trait, 68
predicting associations in recessive SNPs simulated data. See synthetic gene database
and, 45 single nucleotide polymorphism (SNP) arrays
prediction accuracy of, 21t, 22 ALS associations and, 18
Schymick vs. Cooley results, 23 generating combinations of, 123
in single-gene epistatic models, 7273, 75t of genetic variants of sample DNA, 23
recessive Fishers exact test (Fis-R), normal polygene methods and, 117
approximation to whole genome sequence data vs., 5
results comparison, 23, 24t, 25 single-gene disorders, genetic linkage studies
in Schymick ALS data set, 21, 21t, 23 of, 3
recessive genes single-gene models. See also environmental
association detection of, main effect (ME) influencing factors studies
test and, 110 Cochran-Armitage trend test on, 7273,
association detection of, total effects test 75t, 76t, 78
and, 107 epistatic models and, 6768
power values in genotype-environment in Stage 1 polygene analysis, 124, 14546
interactions (INT) test and, 108 6-degrees-of-freedom (6df) genotype 2 test
recessive mode of inheritance SNPs, prediction mixed main effects and interaction model
accuracy and, 9 and, 97
redundant SNPs, preprocessing to eliminate, small effects, GWAS detection and replication
122 of, 120
relative risk (). See genetic relative risk SMN, gene linked to ALS pathogenesis in
replicability, GWAS statistical methods and, sporadic ALS and, 17
67, 143 smoking
risk, use of term, 87 bladder cancer and, 89
Rothman N, 89 lung cancer and, 88
 Index 161

SNP odds ratio (ODDS) data, 128 T

SOD1, gene linked to ALS pathogenesis in tag SNPs
familial ALS and, 17 Cooley analysis of Schymick data and, 25
sporadic amyotrophic lateral sclerosis, 17 GWAS identification of, 3
sporadic cancer, twin studies of, 88 TEAM method, 81, 123
standards Templeton AR, 6768
conclusions, 26 Terry PD, 88
Genome-wide Association Studies (GWAS), test scores, in polygene analysis, 12425
67 Tian X, 50
statistical tests, 2022 total effect test (TOT)
statistical models. See also M-1 statistical elements in, 105
model; M-2 statistical model; M-3 power values, by risk profile, 107t
statistical model transcriptome, high-throughput technologies
polygene analysis using, 12123 for measuring, 146
predicting environmental influence and, trans--carotene levels, type 2 diabetes and, 89
11112 truth set
simulated data analysis using, 8 developing, 8
simulated data in epistatic models and, 68 as known outcomes, 85, 87
statistical power. See also errors for polygene methods, 119, 143
properties, best methods, 13 Schymick synthetic data as, 39
stepwise polygene analysis simulated data analysis using, 7, 52
algorithm, 12425 synthetic gene database as, 31, 49
main effects and interactions with, 122 2-degrees-of-freedom (2df) genotype
SNP predictors and phenotype groups, 118, association test. See also Pearson 2 test
119 environmental influencing factors studies,
Stern MC, 88 98t
substrate-dependent pathways genotyping error models on, 50
description of, 67 GWAS statistical test performance, 40, 42,
epistatic models, 69t 43t
switch-regulatory pathways, 67 no environmental exposure, 97
synthetic gene database as not useful, 45
computational requirements, 36 two-gene epistatic models
creating synthetic gene data using, 8 linked and unlinked tests of, 7374
data generation in polygene analysis, 125, risk of second interacting gene and, 75
138, 144 risk value of downstream gene and, 76, 80
data generation process, 8, 3234, 35f single-gene tests vs., 78, 11819
environmental influencing factors and, 85 two-locus, case-control 8-degree of freedom
in epistatic models, 68 (df) Pearson test, 123
for errors assessment, 5253 two-locus, case-control test, 123
genetic inheritance and GWAS test Type I errors
performance using, 810 alpha
known outcomes of, as truth set, 13942, Bonferroni method and, 42
91, 43, 145 environmental influencing factors
noise level specifications, 145 studies and, 9192, 109
statistical analysis of, 7 statistical method assessment and, 42
systems genetics approach, biology of complex Bonferroni method and, 42
traits and, 12021 Experiment 1 (main effects model), 99
low-effect nonrecessive loci and, 62
in polygene analysis, 123
predictions and, 59
sample sizes and, 49
statistical significance threshold and, 20
162 Index

U
Ueki M, 67. See also Wu et al test refined by
Ueki
UNC13A gene, ALS and, 19

V
validity, GWAS statistical methods and, 6
van Es MA, 19, 25
VAPB, gene linked to ALS pathogenesis in
familial ALS and, 17
variants
genetic burden of common diseases and, 6
GWAS detection of, 56
VEGF, gene linked to ALS pathogenesis in
sporadic ALS and, 17

W
Wang Y, 8081, 123
warfarin, genetic testing and dosages of, 4
whole-genome sequencing, 146
wild homozygotes (g2), in environmental
influence statistical models, 98
wild-type alleles (G), in statistical models, 98
Wu et al test refined by Ueki
two-gene epistatic models, 7374, 75, 76,
78, 80
two-locus models, 123
Wu X, 67

Y
Yang J, 120
Yu K, 89

Z
Zeggini E, 20, 22
Zhang X, 81, 123
Zheng G, 22, 40, 50, 72
Ziegler A, 33
 This groundbreaking work uses a simulated data set to evaluate new analytic
methods in genome-wide association studies (GWAS). The human genome
is very complex, and the effect of a genetic variant depends on many factors
including where the gene is expressed, when it is expressed, how it interacts
with other genes that themselves may harbor variants, and the effect of
the environment. GWAS have identified many new genetic risk factors for a
number of common human diseases, but much work remains to be done and
can only be accomplished by using new approaches. The role of this book is to
help jump-start the investigation of such new approaches. Identifying which
computational strategy is best suited for investigating a specific aspect of
genomics is a daunting task. Using simulated data to evaluate new analytic
methods provides a truth set against which to assess methods predictive
properties. In this book, Cooley and colleagues use simulated data to test a
variety of analytic methods, starting with single-gene models and progressing to
more complex polygene and gene-by-environment scenarios. The methods that
Cooley and colleagues use are straightforward, easily applied, and thoroughly
documented.

RTI Press

www.rti.org/rtipress RTI Press publication BK-0016-1608