Estimating the Returns to Education: Models,

Methods and Results

Richard Blundell
Lorraine Dearden
Barbara Sianesi















October 2001
ISSN 2045-6557
Published by
Centre for the Economics of Education
London School of Economics
Houghton Street
London WC2A 2AE

Richard Blundell, Lorraine Dearden and Barbara Sianesi
October 2001
Individual copy price: 5

The Centre for the Economics of Education is an independent research centre funded by the
Department for Education and Skills. The views expressed in this work are those of the author
and do not reflect the views of the DfES. All errors and omissions remain the authors.
Estimating the Returns to Education: Models,
Methods and Results



Richard Blundell
Lorraine Dearden
Barbara Sianesi



1. I ntroduction 1

2. The Earnings-Education Relationship: Alternative Models 5
Earnings and education in a homogenous returns framework 5
The heterogeneous returns model 8

3. The Earnings-Education Relationship: Alternative Methods 10
Least squares 14
Instrumental variable methods 16
Control function methods 23
Matching methods 26

4. Measuring Returns in the NCDS: The Data 36
Single treatment models qualifications versus no qualifications 37
The multiple treatment model 42

5. Summary and Conclusions 46

References 48
Appendices 51
Acknowledgments


Richard Blundell is a Professor of Economics at the Department of Economics, University
College London, Research Director at the Institute for Fiscal Studies and a Research
Associate at the Centre for the Economics of Education. Lorraine Dearden is Director of the
Education, Employment and Evaluation Research sector, Institute of Fiscal Studies. Barbara
Sianesi is a Research Economist at the Institute for Fiscal Studies.

The authors would like to thank Bob Butcher, Paul J ohnson and participants at the RSS
meeting for comments on earlier drafts. Thanks also to thank Mario Fiorini for research
assistance.

Estimating the Returns to Education: Models,
Methods and Results

Richard Blundell, Lorraine Dearden and Barbara Sianesi

University College London and Institute for Fiscal Studies
October 2001
Abstract
This paper reviews appropriate non-experimental methods and microe-
conometric models for recovering the returns to education using individual
data. Three estimators are considered: matching methods, instrumental
variable methods and control function methods. The properties of these
methods are investigated for models with multiple treatments and hetero-
geneous returns. Data from the British 1958 NCDS birth cohort is used to
estimate returns to schooling and to illustrate the sensitivity of dierent
estimators to model specication and data availability.
1 Introduction
There are at least three distinct ways of dening the returns to education:
(a) the private return, (b) the social return and (c) the labour productivity
return. The rst of these is made up of the costs and benets to the individual
and is clearly net of any transfers from the state and any taxes paid. The
second denition highlights any externalities or spill-over eects and includes
transfers and taxes. The nal denition simply relates to the gross increase in
labour productivity (or growth). A key component of each of these measures

1
is the impact of education on earnings. This is perhaps the aspect of returns
to education measurement where statistical methods have been most developed
and most fruitfully deployed and is the central focus of this paper.
With extensive data available over time and individuals on schooling and on
earnings, the measurement of the education eect on earnings is one area where
we might expect agreement. However, a casual look through the literature on
the impact of education on earnings reveals a wide range of estimates and an
equally wide range of empirical approaches that have been adopted to estimate
the return. So why do the estimates vary so widely and what is the most appro-
priate empirical method to adopt? The answer to these two questions provides
the central motivation for this paper. It is illustrated using the sample of men
from the NCDS Birth Cohort data for the UK. This data source provides a
uniquely rich source of non-experimental data on family background, educa-
tional attainment and earnings. We argue that it is ideally suited to analyse
statistical methods for the measurement of the eect of education on earnings.
The appropriate statistical method to adopt will depend, in a rather obvious
way, on the chosen model for the relationship between education and earnings.
We distinguish two broad characterisations of this specication. The rst relates
to the measurement of education. In particular whether we can summarize edu-
cation, or human capital more generally, in a single measure - years of schooling,
for example. This is commonly referred to as a one factor model. It is a re-
strictive framework since it assumes that, as returns to education change over
time, it is only the single aggregate that matters and there are no dierential
trends in returns for dierent education levels. It is convenient though since
we can simply include a single education measure in an earnings equation. The
alternative to this is the multiple factor model where dierent educational levels
have separate eects on earnings.
We will refer to dierent education levels as dierent treatments borrowing
2
a common notation from the evaluation literature. A single treatment speci-
cation refers to the impact of a specic educational level - such as gaining a
qualication. A multiple treatment eect model will distinguish the impact of
many dierent education levels. In general the multiple treatment - multiple
factor model would seem a more attractive framework since we will typically
be interested in a wide range of education levels with very dierent returns.
However, we will also consider models with a single discrete treatment such as
the impact of a specic qualication and models with a single overall education
level such as years of schooling.
The second characterisation relates to the distinction between heterogeneous
and homogeneous returns. In simple terms - whether the response coecient
on the education variable(s) in the earnings equation is allowed to dier across
individuals. To allow this to happen according to observables is a relatively
straightforward extension of the homogeneous model, but to allow the het-
erogeneity to be unobservable completely changes the interpretation and the
properties of many common estimation approaches. We begin in Section 2 with
this distinction between model specications and use it to dene parameters of
interest in the earnings education relationship.
Even where there is agreement on the model specication there are alter-
native statistical methods which can be adopted. With experimental data the
standard comparison of control and treatment group recovers an estimate of
the average eect in the population under the assumption that the controls are
unaected by the treatment. Although in some studies of training, experimen-
tal design is possible and growing in popularity, for large reforms to schooling,
and for measuring the impact of existing educational systems, nonexperimental
methods are essential. There are broadly two nonexperimental methods: those
that attempt to control for correlation between individual factors and schooling
choices by way of an excluded instrument, and those that attempt to measure
3
all individual factors that may be the cause of such dependence and then match
on these observed variables. The implementation and the properties of these al-
ternative methods dier according to whether the model is one of heterogeneous
response and whether schooling is represented through a single or multiple mea-
sure. The dierent properties of these estimators and the drawbacks to each
method are discussed in Section 3.
The various models and non-experimental estimators are then compared in
Section 4 and 5 using the British NCDS data. In particular, we consider the
returns to education for three levels: 1) leaving after completing O levels or
its vocational equivalent 2) leaving after completing A levels or its vocational
equivalent 3) undertake some form of higher education (including sub-degree
level HE).
We only present results for men so as to conserve space and to focus on
the earnings eect versus the employment eect. Our results show signicant
returns to basic and higher qualications in education, even after controlling
for detailed family background and ability test measures. They also highlight
two important features: (a) the returns to education on earnings are sensitive
to the inclusion of controls. In particular we nd that controlling for ability test
scores at an early age signicantly reduces the returns by a factor of one third.
(b) there is strong evidence of heterogeneity of returns with quite high returns
found among subsets of those who do not stay on for higher qualications at
school. Although on average the returns among those that stay on for higher
qualications are higher than for those who do not.
4
2 The Earnings-Education Relationship: Alter-
native Models
This section highlights two central aspects in the empirical investigation
of the earnings return to human capital investments. First among these is the
distinction between the homogeneous returns and heterogeneous returns model.
In the homogeneous returns model the rate of return to gross earnings of a
particular human investment is the same for all individuals. Growing statis-
tical evidence and causal empiricism suggests that the homogeneous returns
restriction is unwarranted.
The second aspect is to distinguish between the one factor and multiple
factor models of human capital. In the one factor model all schooling can be
thought of as an investment in a single homogeneous construct called human
capital. Each additional unit has the same return. An example of an empirical
model that is both one factor and homogeneous returns is the popular linear
regression equation - log earnings regressed on years of schooling. The constant
parameter on the schooling variable is equivalent to homogeneous returns and
the use of years of schooling as a single measure of schooling is equivalent to a
single measure of human capital.
2.1 Earnings and Education in a Homogenous Returns
Framework
For each individual i = 1, ..., n, we let y
i
represents their earnings or hourly
wage opportunities in work. To begin with we will assume that we are measuring
earnings at one point in time for a sample of individuals who have completed
formal schooling. A good illustration to keep in mind is from the British cohort
studies where a single cohort is followed through education and employment
and sampled at specic intervals usually several years apart. We measure their
earnings when they are 33 years of age and ask: what is the impact on earnings
5
at age 33 of dierent schooling outcomes?
In the one factor human capital model it will always be possible to aggre-
gate schooling into a single measure S
i
- say, years of schooling. For example,
we may write
lny
i
=
i
+ S
i
+
i
(1)
where
i
represents diering relative levels of earnings across individuals for any
given level of schooling and measures the marginal return to schooling level
S
i
in terms of the particular denition of earnings y
i
. The error term
i
is
added to capture measurement error in earnings.
1
Since educational choices and educational levels are likely to dier according
to productivity (or expected earnings levels more generally), S
i
is very likely to
be positively correlated with
i
and this in turn will induce an upward bias in
the simple least squares estimation of . However, if S
i
is measured with error
there will be some o-setting attenuation bias and this trade-o was at the heart
of the early studies on measuring gross private returns (see Griliches (1977) and
Card (1999), for example). We will return to these estimation issues in more
detail below.
As mentioned above (1) assumes both homogeneous returns (the common
across all i) and one factor human capital (the single measure of schooling S
i
).
Although popular this seems unduly restrictive. In order to cover a fairly exible
representation of schooling we will consider the case of a nite set of schooling
levels. For example, in the application to the NCDS we refer to a number of
specic discrete educational levels - as obtaining a qualication, obtaining an
A level, or undertaking higher education. Using notation borrowed from the
experimental literature we will refer to these as multiple treatments. These
will typically be dened in some natural sequence of binary indicator variables:
1
Measurement error in the schooling or eduction variable S
i
is also likely to be important
and will be discussed in terms of the alternative approaches to estimation.
6
S
1i
= 1 if the individual completed the rst stage of schooling, S
2i
= 1 if the
next stage in the sequence is completed etc. For completeness S
0i
= 1 would
represent the base educational level. For example in the UK context S
1i
= 1
may refer to staying on after the minimum school leaving age, S
2i
= 1 might
represent achieving at least one A-level, S
2i
= 1 to achieving a rst degree etc.
Write the exhaustive set of J treatments (schooling levels) under examination
as S
1i
, S
2i
, ..,S
Ji
. In this case (1) might be adapted to become
lny
i
=
i
+
1
S
1i
+
2
S
2i
+.. +
J
S
Ji
+
i
(2)
where the
1
,
2
...
J
. now measure the marginal impact of a higher level of
schooling for some J distinct levels. Still the returns are homogeneous across
individuals but (2) can be seen to relax the one factor assumption and allow
dierent schooling levels to have quite dierent impact on earnings.
Of course, one can imagine a ner sequence and also possible set of non-
sequential outcomes. All the methods discussed below are easily extended to
more complicated situations but will typically require more demanding data
requirements and modelling assumptions to estimate the causal impact on
earnings. Indeed, to begin the discussion of the heterogeneous model we will
begin with a single treatment model and study the causal impact of a single
type of schooling level.
Before moving on to the more general models and the alternative methods
of estimation, it is worth pointing out that each of these equations will typically
be specic to a particular time period and location. For example, if (2) refers to
the impact of education levels on the earnings of British men aged 33 in 1991, it
will be unlikely to be stable across time periods and countries. The returns will
depend on the earnings set in the labour market and will in turn depend on the
demand and supply of individuals with these diering human capital attributes.
This point, although quite obvious, is often misunderstood in the context of
7
predicting returns to education.
2.2 The Heterogeneous Returns Model
The heterogeneity we are focussing on here is unobserved heterogeneity
across individuals in the response parameter . Consider the single treatment
model where we let schooling S
1i
for individual i be dened as a binary indicator
variable representing the successful achievement of a particular education level
- such as obtaining a qualication, obtaining an A level, or undertaking higher
education, for example. A completely general relationship between this level of
education and earnings in this single discrete treatment heterogeneous returns
model is then written
lny
i
=
i
+
i
S
1i
+
i
(3)
where
i
and
i
can be thought of as random coecients representing the
heterogeneous relationship between educational qualication S
1i
and earnings.
Typically we would assume the
i
and
i
have a nite population mean and
variance. Below the population means are labelled
0
and
0
respectively.
Despite the preponderance of the homogeneous returns model in the early
literature, the recent focus has been on the heterogeneous returns model (3)
2
.
This raises the immediate question in this model: what is the parameter of
interest. Is it the average of the

i
s? If so what average? Is it the average in the
population whether or not level S
1i
is achieved,
0
- the average treatment eect,
or the average among those who individuals actually observed with S
1i
= 1,
T
- the average treatment on the treated? In some cases a particular estimation
method will recover a local average treatment eect, measuring the impact of
S
1i
= 1 on an even smaller subgroup of individuals. We discuss all these in
greater detail in the next section.
2
See for example the papers by Heckman, J., Smith, J. And N. Clements, (1997), Dearden
(1999a) and (1999b) and Blundell, Dearden, Goodman and Reed (2000).
8
One interpretation of
i
is as a heterogeneous return to schooling level S
1i
for individual i since it measures the marginal proportional impact of this level
of education on earnings for individual i. Again this measure of returns is a gross
private measure since it ignores all costs of education and also taxes paid on
gross earnings y
i
. Note that in the homogeneous returns model
i
is constant
across all individuals. Even so, in this homogeneous returns model
i
is allowed
to vary across i to capture the diering productivities (or abilities), and diering
general levels of earnings, across individuals with the same education levels.
It should be pointed out that this model, and the others to be discussed
below, can be readily generalised to allow for observable heterogeneity in both

i
and
i
. For example, suppose there are a set of observed covariates X
i
(e.g.
early test scores, demographic variables, aspects of the local labor market).
The and parameters can be made to depend on these in a quite arbitrary
way. If they are assumed to depend on X
i
in a linear fashion then the levels
of X
i
and the interactions of X
i
with the education variable S
1i
will enter the
regression specication. The precise form chosen will depend on the richness of
the data set and the particular problem at hand
3
. But in what follows we shall
always assume that such levels and interactions are included in the specication.
Indeed, in the general nonparametric matching method described below a quite
general form of interaction is allowed. For the most part we will assume that
such observed heterogeneity terms are included, even if this is not explicitly
stated in the discussion of the properties of the various alternative estimators
described below.
As we saw in the homogeneous model (1), the dependence of the schooling
level S
i
on the unobserved ability component
i
is critical in understanding
the bias from standard least squares estimation. An additional central issue in
determining the properties of standard econometric estimators in the hetero-
3
This was done in Dearden (1999a) using the same NCDS data that is used in this paper.
9
geneous eects model is whether or not schooling choice S
1i
depends on the
unobservable determinants of the individualss marginal return from schooling

i
. If
i
were known when the individual makes his or her educational choices
then it would seem sensible to assume that choices will - in part, at least, reect
the return to earnings of that choice. But as mentioned before
i
is likely to
vary over time and will depend on the relative levels of demand and supply, so
the dependence of schooling choices on marginal returns is not clear-cut. Some
persistence in returns is however likely and so some correlation would seem more
likely than not.
The discussion of heterogeneous returns will extend easily to the multiple
treatment model (2). Writing the exhaustive set of J treatments (schooling
levels) under examination as S
1i
, S
2i
, ..,S
Ji
. The heterogeneous returns model
is then
lny
i
=
i
+
1i
S
1i
+
2i
S
2i
+.. +
Ji
S
Ji
+
i
. (4)
We will also want to discuss the one factor model in which S
i
enters as a single
continuous variable
lny
i
=
i
+
i
S
i
+
i
. (5)
In fact, the three basic specications (3), (2) and (5) will form the main alterna-
tives considered here. The single discrete treatment case (3) being the baseline
specication.
3 The Earnings-Education Relationship: Alter-
native Methods
The aim here is to investigate the properties of alternative estimation meth-
ods for each of the model specications considered above. The alternative meth-
ods we consider fall into three broad classes: the instrumental variable method,
10
the control function method and the method of matching. The rst two require
some excluded instrument which determines education choices but not earnings
while the matching method requires an extensive set of observable characteris-
tics on which to match. All place strong demands on data.
The initial setting for this discussion will be based on the biases that occur
from the simple application of ordinary least squares to the estimation of each
of the model specications described in the previous section. As mentioned the
primary model specication will be the single discrete treatment heterogeneous
returns model (3) but the extension to the multiple treatment model (4) will
also be considered and so will the specic issues that occur in the one factor
years of schooling specication (5). In each of these the complications that
are engendered by allowing the return parameter to be heterogeneous will be
central to the discussion.
It will also be useful to relate this choice of alternative methods to the eval-
uation and the selection literature (see in particular Heckman and Robb (1985)
and Heckman, LaLonde and Smith (1999)). The treatment eect parameters
referred to above already borrow heavily from this literature. It is easy to see
that even the estimation problem in the education returns framework is syn-
onymous with the construction of a counter-factual in the evaluation literature.
Indeed, some of the more recent developments in the in the returns literature,
for example those which use matching estimators or social experiments, relate
explicitly to similar approaches in evaluation.
It is also worth pointing out that we will not be looking at general spillover
eects or general equilibrium eects of education. In the statistical evaluation
literature this relates to the stable unit-treatment value assumption (SUTVA).
This assumption requires that an individuals potential outcomes depend only on
his own schooling participation, not on the schooling choice of other individuals
in the population (thus ruling out cross-eects or general equilibrium eects)
11
and that the education level chosen by an individual does not depend on the
schooling decisions of others (e.g. thus excluding peer eects in educational
choices).
To illustrate the importance of constructing the counterfactual, consider the
single discrete treatment model (3). For any individual i in the set of individuals
that receive S
1i
= 1 (i {S
1i
= 1}) the earnings outcome is
lny
1
i
=
i
+
i
+
i
for i {S
1i
= 1} (6)
where the superscript 1 refers to the case where individual i receives treatment
S
1i
= 1. Whereas if the same individual were not to receive this education level
their earnings outcome would be
lny
0
i
=
i
+
i
for i {S
1i
= 1}. (7)
now the superscript 0 refers to the counterfactual earnings of an individual i for
whom S
1i
= 1 in the observed data. Suppose that in a random sample of size n
there are n
1
individuals for whom S
1i
= 1. If we could observe both outcomes
for all individuals for whom S
1i
= 1 then the average

iS1i=1
lny
1
i
lny
0
i
n
1
(8)
would be a consistent estimate of the average treatment on the treated eect

T
. To recover the average treatment eect
0
across the whole population
under consideration further requires the counterfactual for the group for which
S
1i
= 0.
The focus here will be on non-experimental approaches. Of course, in a
randomized experiment the control group is chosen independently of the
i
,
i
and
i
by design. Consequently the average treatment eect can be measured
directly from a comparison of the control group and the treatment group. This
is the case even in the most general heterogeneous returns model described
12
above and whatever the underlying relationship between the education level
variables S
ji
etc., and the heterogeneous ability and returns parameters
i
and

i
parameters respectively. A truly randomised education experiment would
induce full independence between these individual heterogeneity parameters and
the education outcomes. However, pure education or schooling experiments are
very rare. It is dicult to persuade parents or the students themselves of the
virtues of being randomised out of an education programme - except for rather
minor programmes. Our application will be to the main stages of educational
level in the UK and randomised assignment is unavailable. Even if it were,
individuals enrolled may drop out - and systematically too.
However, the pure randomised experiment is useful as a basis for comparison.
Natural social experiments are more common. This is the case where some
educational rule or qualication level (say minimum schooling leaving age) is
exogenously changed for one group but not another. Provided the groups are
representative samples from the population then this simple comparison can
recover a parameter of interest like the average treatment eect. Where the
samples dier in their ability levels or other characteristics it may still be possible
to recover an average eect for those who experience the change in rules.
Many alternative methods are available and all have been widely used. Twins
and sibling data can help, an exogenous change in S
i
that only eects a subsam-
ple of the target group is also useful. The latter relates to instrumental variable
methods more generally, where some variable (or transformation of the data)
has to be found that can vary S
i
independently of the heterogeneity terms. An
alternative to using instruments is the matching method. This method seeks
to purge the relationship between schooling and earnings of any important ob-
served heterogeneity that would lead to bias, by matching individuals with and
without the schooling level according to some observed characteristic.
13
3.1 Least Squares
Consider rewriting the heterogeneous single treatment model (3) as
lny
i
=
0
+
0
S
1i
+ (
i

0
) + (
i

0
)S
1i
+
i
(9)
where
0
and
0
are the population means of
i
and
i
. This population may
be dened as all those individuals entering schooling at a particular date. In
this case a parameter of interest will be
0
itself which measures the average
returns to achieving education level S
1i
in this population. For example, in the
British context S
1i
= 1 may refer to those staying on after the minimum school
leaving age, or it may refer to those going on to a higher degree.
Gathering the unobservables together we have
lny
i
=
0
+
0
S
1i
+u
i
(10)
with
u
i
(
i

0
) + (
i

0
)S
1i
+
i
The least squares regression of log earnings on schooling produces a biased
estimator of
0
due to correlation between S
1i
and u
i
. There are three sources
of such bias:
(i) Ability bias: this occurs due to the likely correlation between S
1i
and
the (
i

0
) term. A positive correlation inducing an upward bias in the
estimated return.
(ii) Returns Bias: this occurs when the marginal returns (
i

0
) them-
selves are correlated with the schooling choice S
1i
. The direction of this bias
is less clear and will depend on the average returns among the sub-population
of those with S
1i
. Indeed, if ability bias is negligible (and the remaining ability
heterogeneity is unrelated to the unobserved return) and returns bias is the only
remaining bias present, then the least squares coecient estimate will recover
14
the average returns in the sample of those with S
1i
= 1, that is the average
treatment on the treated.
(iii) Measurement Error Bias: this refers to measurement error in the
schooling variable S
1i
. This may be due to random misclassication error. As
usual, measurement error of this kind will induce attenuation bias in the re-
gression coecient and an under-estimate of the returns parameter. For the
purposes of much of the discussion we can redene
i
to include measurement
error in the schooling variables(s).
In the homogeneous returns model the second bias is, by denition, absent.
This is the case that is much discussed in the literature - especially in the one
factor years of schooling model (1). Indeed, there is some evidence of a bal-
ancing of biases (see Card (1999), for example), in which case OLS fortuitously
consistently estimates the return coecient .
Much of the practical discussion of the properties of least squares bias de-
pends on the richness of other control variables that may be entered to capture
the omitted factors. Indeed, the method of matching, described below, takes
this one step further by trying to eliminate the imbalance of observables by
matching observations with similar covariates. One simple recommendation in
the use of least squares seems worth following - add in a rich set of controls
and try to nd separate measures of the education variable that may not suer
from the same measurement error. The rich set of controls may help reduce
the ability bias and the second measure of schooling may be used to purge the
measurement error. But, in the end, a comparison with the alternative meth-
ods: instrumental variables, control functions and matching is always helpful in
assessing how to interpret a least squares estimate of education returns.
15
3.2 Instrumental Variable Methods
The Instrumental Variable (IV) estimator seems a natural method to turn
to in estimating returns - at least in the homogeneous returns model. The biases
we have discussed in the case of least squares all stem from the correlation of
observable schooling measures with the unobservables in the earnings regres-
sion. If an instrument can be found that is correlated with the true measure
of schooling and uncorrelated with the unobservable ability, heterogeneity and
measurement error terms, then surely a consistent estimator of the returns is
achievable. This turns out to be true in the homogeneous returns model but not,
except for certain special cases described below, for the heterogeneous returns
model.
Even in the homogeneous returns model, nding a suitable instrument is
not easy. However, social and natural experiments can be useful - and many
such instruments have been used. Alternatively, parental background variables
are often chosen. But to be useful they must satisfy the Instrumental Variable
criteria of being correlated with the schooling choice and correctly excluded
from the earnings equation.
To more formally investigate the properties of the IV estimator, dene an
instrumental variable Z
i
and assume that it satises the orthogonality condi-
tions
4
:
IV: A1: E[(
i

0
)|Z
i
] = 0
IV: A2: E[(
i

0
)|Z
i
] = 0
IV: A3: E[
i
|Z
i
] = 0
Additionally assume that the instrumental variable is related to S
1i
through
IV: A4: E[S
1i
|Z
i
] = Z

4
In fact the IV estimator we will discuss will typically only require the weaker covariance
restrictions cov(
i
Z
i
) = cov(
i
Z
i
) = cov(
i
Z
i
) = 0 rather than these mean independence
assumptions.
16
where is a nite vector of unknown reduced form coecients.
To discuss the general properties of the IV estimator in the heterogeneous
returns model consider the conditional expectation of (9) under assumptions
IV: A1 - A4
E[lny
i
|Z
i
] =
0
+
0
Z

i
+E[(
i

0
)S
1i
|Z
i
]. (11)
Note that IV: A4 implies that can be estimated consistently from the least
squares regression of S
1i
on Z
i
- the rst stage, reduced form regression. There
is nothing in assumptions IV: A1 - A4 that makes the nal term in (11) disap-
pear. Further assumptions are required. Indeed, even IV: A4 is controversial
since in this specication S
1i
is discrete and this assumption is better suited to
the one factor years of schooling model. Indeed, in the homogeneous one
factor model
i
is constant across i and the latter term is zero by denition.
Consequently, instrumental variable estimation can produce a consistent esti-
mator of in this case.
One way to interpret the IV estimator is as the (weighted) least squares
estimator for the transformed regression model
Z
i
lny
i
= Z
i

i
+
i
Z
i
S
1i
+Z
i

i
(12)
with the weights depending on the sample covariance matrix of Z
i

i
. In the case
where there is a single instrument the IV estimator reduces to

IV
=
cov(lny
i
, Z
i
)
cov(S
i
, Z
i
)
(13)
Consequently, given the IV assumptions IV: A1 - A4, an IV transformation
Z
i
that eliminates
i
but leaves
i
S
1i
unaected, will estimate the average
i
among those individuals for whom S
1i
= 1. This is the treatment on the treated
parameter

IV
=
T
. A prime example of such a transformation is a dierencing
instrument. There are three such dierencing transformations commonly used.
17
The rst is the dierence in dierences estimator. This compares the group of
individuals with S
1i
= 1 to the group with S
1i
= 0 before and after the treatment
S
1i
occurs. This transformation sweeps away any common individual component
(
i
, for example) and just picks out the average of
i
among those individuals
with S
1i
= 1. Precisely the treatment on the treated parameter. The dierence
in dierences estimator is ideally suited for training programmes for which there
is an earnings observation before and after training but less applicable in the
schooling/formal education evaluation problem where a before observation
of earnings is very unlikely to be available, since for most individuals formal
education is completed before labour market participation.
A second such dierencing instrumental variable estimator is to use a sample
of twins
5
. The contrast in this case is between individuals within pairs of twins.
Dierencing within twins eliminates common gene determined ability or family
eects. These are often what is being represented by
i
. In the homogeneous
returns model this perfectly identies . However, it only uses information on
twins who have dierent S
1i
outcomes. This may only be a small subset of
twins. Typically twins separated at birth are used and this tends to increase
the probability of dierent values of S
1i
within any pair. In the heterogeneous
returns model the twin dierence estimator recovers the average marginal return
on those twins with dierent schooling outcomes. This may be a representa-
tive sample of the population but it may not. In general what is estimated
strictly reects the local average of the returns among the sample of twins who
experience both outcomes.
Finally, more general groups can be compared. For example we may compare
the outcomes among two groups that have a similar distribution of abilities but
who, from some exogenous reform, experience dierent schooling outcomes
6
.
5
See for example Ashenfelter and Krueger (1994), Altonji and Dunn (1996), Ashenfelter
and Zimmerman (1993), Behrman et. al. (1996), Bonjour et. al. (2000).
6
For example see the papers by Angrist and Krueger (1991) and (1992), Butcher and Case
18
A classic example of this is the comparison of adjacent cohorts one of which
experiences a school reform (say a change in the minimum school leaving age)
and the other who does not
7
. Again in the homogenous treatment eects model
this can be used to estimate , but in the heterogeneous model it will estimate
the average of returns among those induced to take more schooling by the reform.
This is the Local Average Treatment Eect and will not in general equal the
average returns among those with S
1i
- the treatment on the treated parameter,
or the average in the population - the average treatment eect. The way in
which the IV estimator for the heterogeneous returns model is related to a local
average of the returns is investigated further below. First we consider some
special cases.
Even where the IV estimator does produce a consistent estimate of there
remains the issue of eciency and of weak instruments. Eciency concerns the
imprecision induced in IV estimation when the instrument has a low correlation
with the schooling variable. The weak instrument case is an extreme version of
this where the sample correlation is very weak and the true correlation is zero.
In this case IV will tend to the biased OLS estimator even in very large samples
(see Bound [Link]. (1995) and Staiger and Stock (1997)).
3.2.1 IV in the Heterogeneous One Factor Model
Consider the one factor or years of schooling model (5). This is a case
where assumption IV A4 would seem plausible and note that this assumption
implies:
S
i
= Z

i
+v
i
where E(v
i
|Z
i
) = 0. (14)
and a consistent estimate of can be obtained from OLS on the reduced form
Now assume
(1994), Harmon and Walker (1995), Meghir and Palme (2000).
7
This was the reform exploited by Harmon and Walker (1995).
19
IV A5: v
i
=
v
(
i

0
) +
i
, with E[(
i

0
)
2
] =
2

.
This assumption implies
E[(
i

0
)S
i
|Z
i
] =
v

so that (11) becomes

E[lny
i
|Z
i
] =
0
+
0
Z

i
+
v

=
0
+
0
Z

i
(15)
so that the IV estimator will, under these assumptions, consistently estimate
the average return
0
but not the intercept.
Note that this is very specic to the continuous schooling measure S
i
in
the one factor model, since then the additively separable model for (14) is a
reasonable specication for the reduced form. Even so the homoscedasticity
assumption imbedded in IV A4 is strong. A useful generalisation of this IV
estimator for this specication occurs in the Control Variable approach discussed
below.
3.2.2 IV in the Homogeneous One Factor Model
By denition this is a specication in which
i
is constant for all i. Conse-
quently
2

= 0 and (11) becomes

E[lny
i
|Z
i
] =
0
+
0
Z

i
(16)
so that IV consistently estimates the
0
and
0
parameters.
Note that the exact same IV estimator can be computed from a regression of
log earnings on schooling S
i
including the reduced form error v
i
as an additional
regressor
lny
i
=
0
+
0
S
i
+
uv
v
i
+
i
where E[
i
|S
i
, v
i
] = 0 (17)
20
Replacing v
i
by
v
i
= S
i
Z

i

i.e. the residual from the reduced form, preserves the correspondence with IV.
This augmented regression framework for IV is popular for testing the exogene-
ity assumption (H
0
:
uv
= 0) and generalises to binary choice and censored
regression settings (see Smith and Blundell (1986)).
The estimation of and by the inclusion of v
i
in this homogeneous returns
specication is also exactly equivalent to the control function approach . As we
will show below, this analogy between IV and control function breaks down
outside the one factor homogeneous returns model.
3.2.3 IV in the Heterogeneous Single Treatment Model: Estimating
the Local Average Treatment Eect
Even in the general heterogeneous returns model with a single treatment
(3), it is still possible to provide an interesting interpretation of the IV estimator
even if it does not estimate the average treatment on the treated or average
treatment parameter. The interpretation of IV in this model specication was
precisely the motivation for the Local Average Treatment Eect of Imbens and
Angrist (1994).
Suppose there is a single discrete binary instrument Z
i
= {0, 1}. For example,
a discrete change in some educational ruling that is correlated with the schooling
level S
i
in the population. There will be four subgroups of individuals, one of
these is of particular interest and is made up of those individuals who are seen
with education level S
1i
= 1 after the rule change (Z
i
= 1) but who would not
have had this level of schooling in the absence of the rule change (Z
i
= 0). This
is the group induced to change behavior by the instrument. To be more precise
21
we dene the events
D
1i
{S
1i
|Z
i
= 1}
D
0i
{S
1i
|Z
i
= 0}
and assume
LATE: A1 For all i ether [D
1i
D
0i
] or [D
1i
D
0i
].
So that the instrument has the same directional eect on all whose behaviour
it changes. Assume for instance that D
1i
D
0i
; in this case the IV estimator
has the very simple form:

IV
=
cov(lny
i
, Z
i
)
cov(S
i
, Z
i
)
(18)
=
E[lny
i
|Z
i
= 1] E[lny
i
|Z
i
= 0]
E[S
i
|Z
i
= 1] E[S
i
|Z
i
= 0]
. (19)
Under independence of
i
,
i
and
i
from Z
i
this reduces to

IV
=
E[
i
S
i
|Z
i
= 1] E[
i
S
i
|Z
i
= 0]
Pr[D
1i
> D
0i
]
(20)
=
E[
i
D
1i
] E[
i
D
0i
]
Pr[D
1i
> D
0i
]
(21)
= E[
i
|D
1i
> D
0i
]. (22)
Which provides a useful interpretation for IV - it estimates the average returns
among those induced to change behaviour under the instrument - the Local
Average Treatment Eect, LATE. For example, suppose Z
i
= 1 reected a
bad nancial event for the family at the time the education decision was being
made. Then IV would pick out the average marginal return among those taking
schooling level S
1
relative to the average returns among those taking schooling
level S
1
whose family did not experience nancial diculties. This could be a
very high local average return.
22
3.2.4 Some Drawbacks of IV
The LATE discussion highlights the point that the IV estimate in the het-
erogeneous returns model will typically vary depending on which instrument is
used. Moreover, it could vary widely according to the local average it recovers.
For example, if the instrument is a change in the minimum school leaving age
which then induces a change in those achieving schooling level S
1i
, the IV esti-
mator will estimate the average returns among those induced to achieve S
1i
= 1
by the school leaving age reform. These could be a group with very high (or
very low) returns. If those who now achieve S
1i
were those who had little to gain
then the local average could be low. If on the other hand, they are individuals
who had previously left education earlier because of a lack of information or
family resources, the local average return for them could be quite high.
In any case the lesson to be learned from the discussion of IV in the het-
erogeneous returns model is that the nature of the incidence of the instrument
within the distribution of returns
i
is critical in understanding the estimated
coecient.
3.3 Control Function Methods
3.3.1 The Heterogeneous Single Treatment Model
Suppose S
1i
in (3) is determined according to the binary response model
CF: A1 S
1i
= 1(Z

i
+v
i
> 0) and v
i
N(0,
2
v
)
also assume
i
and
i
relate to S
1i
according to
CF: A2
i

0
=
v
v
i
+
i
CF: A3
i

0
=
v
v
i
+
i
Note that given CF: A1 - CF: A3
E[(
i

0
)|S
1i
= 1] =
v

1i
,
E[(
i

0
)|S
1i
= 0] =
v

0i
23
and
E[(
i

0
)|S
1i
= 1] =
v

1i
where
0i
and
1i
are the standard inverse Mills ratios from the normal selection
model (Heckman (1979)) - or control functions.
With these additional assumptions in place, the model (9) can be written
lny
i
=
0
+
0
S
1i
+
v
(1 S
1i
)
0i
+ (
v
+
v
)S
1i

1i
+
i
(23)
with
E[
i
|S
1i
, (1 S
1i
)
0i
, S
1i

1i
] = 0.
Consequently, least squares estimation of the augmented log earnings regres-
sion which includes the additional terms (1 S
1i
)
0i
and S
1i

1i
will produce a
consistent estimator of
0
. These additional terms are labelled control functions
and eliminate the bias induced by the endogeneity of schooling. Note that these
control function terms depend on the unknown reduced form parameter , but
this can be consistently estimated at a rst stage Probit step - again analogous
to the selection model
8
.
3.3.2 The Homogeneous Returns Model
In the special case where
i
is constant for all i, the control function terms
reduce to a single term

v
((1 S
1i
)
0i
+S
1i

1i
). (24)
3.3.3 The Multiple Treatment Model
The extension to the multiple treatment case is reasonably straightforward.
As in (4), write the exhaustive set of J treatments (schooling levels) under
8
An early example of this can be found in Willis and Rosen (1979).
24
examination as S
1i
, S
2i
, ..,S
Ji
. The heterogeneous returns model (9) is then
extended to
lny
i
=
i
+
J

j=1

ji
S
ji
+
J

j=0

j
S
ji

ji
.. +
Ji
S
Ji
+
i
(25)
with S
0i
= 1

J
j=1
S
ji
, . and
E[
i
|S
1i
, ..., S
Ji
] = 0
However, note that to avoid multicollinearity problems the
ji
terms will
need to have independent variation, suggesting that at least J 1 excluded
instruments will be required for identication. Typically nding such a large
set of good instruments is dicult. An alternative identication strategy is
to link the
ji
terms together. For example, if the schooling outcomes follow an
ordered sequence then it may be that a single ordered probit model could be
used for all
ji
terms.
3.3.4 The Heterogeneous One-Factor Model
Consider the one factor or years of schooling model (5). As mentioned
above in the discussion of IV for this case, assumption IV A4 would seem plau-
sible, so that we may write
S
i
= Z

i
+v
i
where E(v
i
|Z
i
) = 0. (26)
Now given the control function assumptions CF A1 - CF A3, we may write
lny
i
=
0
+
0
S
i
+
v
v
i
+
v
S
i
v
i
+
i
(27)
where
E[
i
|S
i
, v
i
, S
i
v
i
] = 0
and now the inclusion of control functions v
i
and S
i
v
i
, render least squares
consistent (see Garen (1984), for example).
25
Again note that this is very specic to the continuous schooling measure
S
i
in the one factor model and the additively separable model for S
i
. Finally,
as we noted above, in the homogeneous one factor model the control function
approach reduces to a regression of log earnings on schooling S
i
including the
reduced form error v
i
as an additional regressor
lny
i
=
0
+
0
S
i
+
v
v
i
+
i
where E[
i
|S
i
, v
i
] = 0 (28)
The estimation of and by the inclusion of v
i
in this homogeneous returns
specication is also exactly equivalent to the IV approach.
3.3.5 Some Drawbacks of CF
The control function approach allows for heterogeneity in a multiple treat-
ment model but at the cost of being able to construct a set of control function
- one for each treatment - that have independent variation. This places strong
demands on instrument availability. That is an excluded instrument is required
for each treatment. Moreover, a functional form assumption is typically made
on the control function. This is equivalent to making an assumption on the
distribution of unobservables.
It is true that the distributional assumptions can be relaxed, following the
recent developments in the semiparametric selection model literature, but the
requirement on excluded instruments can only be weakened by strengthening
the model for the treatment choices. For example, in the application below we
exploit the sequential nature of educational qualications to estimate an ordered
Probit model from which the control functions for each qualication level can
be derived. In that example a single instrument would be sucient.
3.4 Matching Methods
The matching method is a (non-parametric) approach to the problem of
identifying the treatment impact on outcomes. The main purpose of matching
26
is to re-establish the conditions of an experiment when no such data is available.
As discussed earlier, in the case of a social experiment, random assignment of
individuals to treatment ensures that potential outcomes are independent of
treatment status, which allows one to compare the treated and the non-treated
directly, without having to impose any structure on the problem.
The matching method attempts to mimic an experiment by choosing a com-
parison group from all the non-treated such that the selected group is as similar
as possible to the treatment group in terms of their observable characteristics.
Under the matching assumption that all the outcome-relevant dierences be-
tween any two individuals are captured in their observable attributes, the only
remaining dierence between the two groups is programme participation, so
that the outcomes of the matched non-treated individuals constitute the correct
sample counterpart for the missing information on the outcomes of the treated
had they not been treated
9
.
The central issue in the matching method is choosing the appropriate match-
ing variables. We will point out this is a knife edge decision as there can be too
many as well as too few to satisfy the assumptions for recovering a consistent
estimate of the treatment eect. In some ways this mirrors the issue in choosing
an appropriate excluded instrument in the IV and Control Function approaches
discussed above. However, it will become clear that instruments do not make
appropriate matching variables and visa versa. Instruments should satisfy an ex-
clusion condition in the outcome equation conditional on the treatment whereas
matching variables should impact on both the outcome and treatment equations.
3.4.1 General Matching Methods
To illustrate the matching solution for the average impact of the treatment
on the treated in a more formal way, consider a completely general specication
9
This discussion refers to the estimation of the average treatment eect on the treated; for
the average eect of the non-treated, a symmetric procedure applies.
27
of the earnings outcomes in the single discrete treatment case. Denote the
earnings outcome that would result if individual i were to receive the level
of education of interest as lny
1
i
, and let lny
0
i
be the earnings outcome if the
same individual were not to receive this education level. The actually observed
outcome lny
i
can thus be expressed in terms of the potential outcomes and
of the observed treatment indicator S
1i
as lny
i
= lny
0
i
+ S
1i
(lny
1
i
lny
0
i
).
The solution to the missing counterfactual advanced by matching is based on
a fundamental assumption of conditional independence between non-treatment
outcomes and the schooling variable S
1i
:
MM: A1 lny
0
S
1
| X
or its weaker version:
MM: A1 E(lny
0
| X, S
1
= 1) = E(lny
0
| X, S
1
= 0)
This assumption of selection on observables requires that, conditional on
observed attributes X, the distribution of the (counterfactual) outcome lny
0
in
the treated group is the same as the (observed) distribution of lny
0
in the non-
treated group. For each treated observation (lny
i
: i {S
1i
= 1}) we can look
for a non-treated (set of) observation(s) (lny
i
: i {S
1i
= 0}) with the same X-
realisation. Under the matching assumption that the chosen group of matched
controls - i.e. conditional on the Xs used to select them - does not dier from
the treatment group by any variable which is systematically linked to the non-
participation outcome lny
0
, this matched control group constitutes the required
counterfactual. Actually, this is a process of re-building an experimental data
set.
For the matching procedure to have empirical content, it is also required
that
MM: A2 P(S
1i
= 1 | X
i
) < 1 for X
i
C

which guarantees that all treated individuals have a counterpart on the non-
treated population over the set of X values over which we seek to make a com-
28
parison. Depending on the sample in use, this can be quite a strong requirement
(e.g. when the education level under consideration is directed to a well specied
group). If there are regions where the support of X does not overlap for the
treated and non-treated groups, matching has in fact to be performed over the
common support region; the estimated treatment eect has then to be rede-
ned as the mean treatment eect for those treated falling within the common
support.
Based on these conditions, a subset of comparable observations is formed
from the original sample, and with those a consistent estimator for the treat-
ment impact on the treated (within the common support C

) is the empirical
counterpart of:
E(lny
1
lny
0
|S
1
= 1, C

)
=
_
C

[E(lny
1
| X, S
1
= 1) E(lny
0
| X, S
1
= 1)] dF(X | S
1
= 1)
_
C

dF(X | S = 1)
=
_
C

[E(lny
1
| X, S
1
= 1) E(lny
0
| X, S
1
= 0)] dF(X | S
1
= 1)
_
C

dF(X | S = 1)
=
_
C

[E(lny | X, S
1
= 1) E(lny | X, S
1
= 0)] dF(X | S
1
= 1)
_
C

dF(X | S = 1)
If the second assumption is fullled and the two populations are large enough,
the common support is the entire support of both. Note that this estimator is,
simply, the mean dierence in earnings on the common support, appropriately
weighted by the distribution of participants.
If we are also interested in using matching to recover an estimate of the
treatment on the non-treated, as we do in our application to the NCDS data,
we need to extend MM: A1 to include lny
1
and MM: A2 to 0 < P(S
1i
= 1 | X
i
)
for X
i
C

.
As it should now be clear, the matching method avoids dening a specic
form for the outcome equation, decision process or either unobservable term.
We simply need to ensure that, given the right observables X, the observations
29
of non-participants are statistically what the participants observations would
have been had they not participated. Under a slightly dierent perspective, it
might be said that we are decomposing the conditional treatment eect in the
following way:
E(lny
1
lny
0
|X, S
1
= 1) =
{E(lny
1
|X, S
1
= 1) E(lny
0
|X, S
1
= 0)}
{E(lny
0
|X, S
1
= 1) E(lny
0
|X, S
1
= 0)}
The latter term is the bias conditional on X, which under the matching assump-
tion MM:A1 is zero.
Heckman, Ichimura, Smith and Todd (1998) use experimental data to pro-
vide a very useful breakdown of the bias which arises when the treatment on the
treated parameter is estimated using the earnings of the observed group with
S
1i
= 0 (lny
0
| X, S
1i
= 0) to construct the counterfactual, rather than the
true counterfactual (lny
0
| X, S
1i
= 1). They show that the bias term can be
decomposed in three distinct parts:
bias = E(lny
0
| X, S
1i
= 1) E(lny
0
| X, S
1i
= 0) = B
1
+B
2
+B
3
where B
1
represents the bias component due to non-overlapping support of
X; B
2
is the error part due to misweighting on the common support of X
as the resulting empirical distributions of treated and non-treated are not the
same even when restricted to the same support; and B
3
is the true econometric
selection bias resulting from selection on unobservables. Through the process
of choosing and re-weighting observations, matching corrects for the rst two
sources of bias. Arguing the importance of the remaining source of bias amounts
to arguing the inadequacy of the conditional independence assumption (MM:
A1) in the specic problem at hand, which should be done in relation to the
30
richness of the available observables (i.e. the data X) in connection to the
selection/outcome processes.
3.4.2 Propensity Score Matching
It is clear that when a wide range of variables X is in use, matching can
be very dicult to implement due to the high dimensionality of the problem. A
more feasible alternative based on the results of Rosenbaum and Rubin (1983)
is to match on a balancing score, that is a function of the observables X, b(X),
with the property: X S
1
| b(X). This is usually carried out on the propensity
score, the propensity to participate given the full set of observed characteristics:
p(X
i
) P(S
1i
= 1 | X
i
). By denition, treatment and non-treatment observa-
tions with the same value of the propensity score have the same distribution of
the full vector of regressors X. Rosenbaum and Rubin have further shown that
under MM: A1 and MM: A2, that is when
lny
1
, lny
0
S
1
| X and 0 < p(X) < 1
then
lny
1
, lny
0
S
1
| p(X)
In other words, the conditional independence remains valid if p(X) - a scalar
variable on the unit interval - is used for matching rather than the complete
vector of X. Under the two matching assumptions, a matched sample at each
propensity score p(X) is thus equivalent to a random sample: conditioning on
the propensity score, each individual has the same probability of assignment to
treatment, as in a randomised experiment, so that individuals with the same
value of p(X) but a dierent treatment status can act as controls for each
other. At any value of p(X), the dierence between the treatment and the
non-treatment averages is thus an unbiased estimate of the average treatment
31
eect at that value of p(X), and the estimate of matching can be thought of as
a weighted average of the estimates from a series of mini random experiments
at the dierent values of p(X).
3.4.3 Implementing Propensity Score Matching Estimators
The main idea of matching is to pair to each treated individual i some group of
comparable non-treated individuals and to then associate to the outcome lny
i
of treated i, a matched outcome

lny
i
given by the (weighted) outcomes of his
neighbours in the comparison group.
The general form of the matching estimator for the average eect of treat-
ment on the treated (within the common support) is then given by

MM
=

i{S
1i
=1C

}
_
lny
i

lny
i
_
w
i
with w
i
typically set equal to 1/N

1
(N

1
being the number of treated individuals
falling within the common support C

).
The general form for the outcome to be paired to treated is outcome is

lny
i
=

jC
0
(pi)
W
ij
lny
j
(29)
where
C
0
(p
i
) denes treated is neighbours in the comparison group (where prox-
imity is in terms of their propensity score to is propensity score, p
i
) and
W
ij
[0, 1] with

jC
0
(pi)
W
ij
= 1 is the weight placed on observation j
in forming a comparison with treated observation i.
The dierent matching estimators dier in how they construct the matched
outcome

lny, that is in how they dene the neighbourhood for the control group
for each treated observation. They also dier in how they choose the weights
for the control group.
32
The traditional and most intuitive form of matching is nearest-neighbour
matching, which associates to the outcome of treated unit i a matched outcome
given by the outcome of the most observably similar control unit k
i
. This
amounts to dening C
0
(p
i
) as a singleton:
C
0
(p
i
) =
_
k
i
{S
1
= 0} : |p
i
p
k
i
| = min
j{S
1
=0}
{|p
i
p
j
|}
_
and setting W
ij
= 1(j = k
i
) (ie. giving a unity weight to the closest control
observation and zero to any other).
In our application below we use a variant of nearest-neighbour matching,
caliper matching (see Cochran and Rubin (1973) and for a recent application,
Dehejia and Wahba (1999)). The caliper is used to exclude observations for
which there is no close match, thus allowing to better enforce common support
on the propensity score. This involves matching treated individual i with its
nearest-neighbour non-treated individual j provided that:
> |p
i
p
j
| = min
k{S1=0}
{|p
i
p
k
|}
If none of the non-treated individuals are within a certain predened absolute
distance or caliper of the treated individual i under consideration, individual
i is left unmatched.
A dierent class of matching estimators has been recently proposed by Heck-
man, Ichimura and Todd (1997) and (1998), Heckman, Ichimura, Smith and
Todd (1998). In kernel-based matching, the outcome lny
i
of treated individ-
ual i is matched to a weighted average of the outcomes of more (possibly all)
non-treated individuals, where the weight given to non-treated individual j is
in proportion to the closeness of the propensity scores of i and j. That is, the
weight in equation (29) above is set to:
W
ij
=
K
_
p
i
p
j
h
_

jC
0
(pi)
K
_
pipj
h
_
33
where h is the bandwidth and K(.) is the kernel.
With e.g. the Gaussian kernel, K(u) exp{u
2
/2} and all the non-treated
units are used to smooth at p
i
, that is C
0
(p
i
) = {j : S
1j
= 0}. By contrast,
with the Epanechnikov kernel, K(u) (1 u
2
) 1(|u| < 1) and thus only those
non-treated units whose propensity score falls within a xed caliper of h from
p
i
are used to smooth at p
i
, that is C
0
(p
i
) = {j {S
1
= 0} : |p
i
p
j
| < h}.
Typically with kernel-based matching common support is imposed on treated
individuals, that is those treated whose propensity score is larger than the largest
propensity score in the non-treated pool are left unmatched.
3.4.4 Multiple Treatments - Mahalanobis metric matching
The most attractive feature of propensity score matching methods is the fact
that they allow to select individuals based on a single variable (nearest-neighbour
matching), or to non-parametrically smooth outcomes on a single variable (kernel-
based matching). The idea underlying Mahalanobis metric matching is to reduce
the dimensionality of the matching problem by rst combining the variables one
wants to match on into a distance measure and to then match on the result-
ing scalar variable. Choosing the Mahalanobis distance dened below has the
advantage of oering a unit free metric, which is essential when the matching
variables have dierent units. Rubin (1979) and (1980) and Rosenbaum and
Rubin (1985) have looked at Mahalanobis metric matching as an alternative
to, as well as in combination with, propensity score matching, by comparing
the performance of matching estimators based on the propensity score alone,
on the Xs combined into a distance measure or on the Xs together with the
propensity score combined into a distance measure.
Even when one decides in favour of propensity score matching, however,
in some settings one may need to match on more than one propensity score.
In particular, when individuals can receive a set of dierent treatments, ade-
34
quately controlling for observed dierences may require matching on two scores
(for a very recent application to a framework of mutually exclusive multiple
treatments, see Lechner (2001)). In our application to the multiple treatment
framework below, we will need to match on three scores.
In such situations, the dierent scores can be rst combined into a distance
metric. Formally, if matching needs to be performed on k > 1 scores, the
formulae for the dierent types of estimators described above continue to apply
after replacing p
i
p
j
with the Mahalanobis distance d(i, j) dened as:
d(i, j) = (P
i
P
j
)

W
1
(P
i
P
j
)
where P
i
is the k1 vector of scores of individual i, P
j
is the k1 vector of scores
of individual j and W is the pooled within-sample (k k) covariance matrix
of P based on the sub-samples of the treated and complete non-treated pool.
This metric eectively weights each co-ordinate in proportion to the inverse of
the variance of that co-ordinate.
3.4.5 Some Drawbacks to Matching
The most obvious criticism that may be directed to the matching approach is
the fact that the identifying conditional independence assumption (MM: A1) on
which the method relies is in general a very strong one. As mentioned above,
the plausibility of such an assumption should always be discussed on a case-
by-case basis, with account being taken of the informational richness of the
available dataset (X) in relation to the institutional set-up where selection into
the treatment takes place.
Furthermore, the common support requirement implicit in MM: A2 may at
times prove quite restrictive. In the case of social experiments, randomisation
generates a comparison group for each X in the population of the treated,
so that the average eect of the treatment can be estimated over the entire
35
support. By contrast, under the conditional independence assumption matching
generates a comparison group, but only for those X values that satisfy MM: A2.
In some cases, matching may not succeed in nding a non-treated observation
with similar propensity score for all of the participants. If MM: A2 fails for
some subgroup(s) of the participants, the estimated treatment eect has then
to be redened as the mean treatment eect for those treated falling within the
common support.
If the impact of treatment is homogeneous, at least within the treated group,
no additional problems arises besides the loss of information. Note, however,
that the setting is general enough to include the heterogeneous case. If the
impact of participation diers across treated individuals, restricting to the com-
mon subset may actually change the parameter being estimated; in other words,
it is possible that the estimated impact does not represent the mean outcome
of the programme, so that we are unable to identify
T
. This is certainly a
drawback of matching in respect to randomised experiments; when compared
to standard parametric methods, though, it can be viewed as the price to pay
for not resorting to the specication of a functional relationship allowing to ex-
trapolate outside the common support. In fact, the absence of good overlap
may in general cast doubts as to the robustness of traditional methods relying
on functional form.
4 Measuring Returns in the NCDS: The Data
In this section we compare the dierent estimation approaches outlined in
Section 3. We begin by comparing the wage outcomes of those individuals who
obtain some qualications (qualications) with those who obtain no qualica-
tions (no qualications). We do this so that we can compare dierent regression
(parametric) based estimates of the returns to obtaining qualications with sin-
36
gle treatment matching techniques. We then go on to consider the sequence
of multiple treatments (O levels or equivalent, A levels or equivalent, higher
education) and again compare parametric based estimates of the returns to
these dierent education paths, with estimates derived from multiple treatment
matching approaches.
4.1 Single Treatment Models - qualications versus no
qualications
The estimated returns to obtaining qualications versus leaving education
with no qualications are shown in Table 1. The treated are all those obtain-
ing at least an O-level qualications or equivalent and we would expect a wide
range of returns among this group. We begin by comparing the dierent regres-
sion based estimates. Specication (i) gives the OLS estimate when we only
use minimal controls (region and ethnicity). We see from Table 1 that for this
basic specication the estimated return to staying on for men is 37.0%. When
we include ability and school type variables (specication (ii)) these estimates
fall. The estimate for is reduced to 27.5%. There is a further fall when we also
include standard family background variables
10
, with the estimated return for
this sample of men being 25.5%.
The Instrumental Variable and Control Function Results: The next
two specications in Table 1 allow the education variable to be endogenous. In
the rst of these, specication (iv), we include controls for region and ethnicity
alone (i.e. the same as specication (i)) and use the standard family background
variables (set I) as instruments for staying-on. The IV estimates of the returns
to staying on are signicantly higher at 78.6%. However, we should note that the
overidentifying restrictions are overwhelmingly rejected suggesting misspecica-
10
The family background variables that we include are parents education, age,
educationage, fathers social class when child was 16 (six dummies), mothers employment
status when the child was 16, and the number of siblings the child had at 16.
37
tion. In specication (v) we instead include these family background variables
as controls in the regression together with ability and school type variables(as
in specication (iii)) and use more credible instruments, set II. These are the
number of older siblings the child has (which controlling for total number of
siblings is exogenous), whether the family was experiencing nancial diculties
in 1969 or 1974, and nally parental interest in the childs education at the age
of seven as assessed by the childs teacher
11
. The IV estimates using this set
II instruments are still higher than least squares with controls but have fallen
back to 47.4%
12
.
The large increase in estimated returns over the corresponding OLS spec-
ications probably reect three factors. These factors are always important
to consider when interpreting IV estimates of returns. The rst relates to the
discussion of the Local Average Treatment Eect - LATE. In that discussion
we noted that the IV estimator, in the heterogeneous returns single treatment
model, will depend on the choice of instrument and the group of individuals
whose schooling is impacted by the instrument. Here the important instru-
ments for specication (iv) relate to parents education and social class, whereas
in specication (v) they relate to birth order, parents interest in the childs ed-
ucation and bad nancial problems for the family. It could well be that the
average return within this group who would otherwise have completed quali-
cations is high. The third relates to measurement error. This would also result
in an increase over the OLS estimates but it is unlikely that there is serious
measurement error in these recorded education levels. The third factor relates
to the use of invalid instruments. This will induce an upward bias in the return.
Some experiments we have conducted suggest that certain of the instruments
used in specication (iv)- fathers and mothers education, for example - may be
11
For both the mother and father we have 3 dummy variables corresponding to expects
too much, very interested and some interest.
12
The Sargan test does not reject the validity of the instruments: p-value 0.317.
38
invalid instruments.
Overall the high values for these IV estimates probably reect the extreme
heterogeneous nature of the group who obtain qualications and emphasise the
extreme caution with which instrumental variable estimates of rates of return
should be treated. This is conrmed by the results from the control function
model which specically allows for heterogeneity in the returns to qualications.
This control function model is simply a more general version of the endogenous
treatment eects model, in that the only dierence is that the inverse mill ratio is
now fully interacted with our qualications dummy. Our estimate of the return
to qualications is further reduced to 37.7%. We reject the null hypothesis of
homogeneous returns for those with qualications
13
. These estimates are almost
identical to those of specication (i) where we have only controlled for region
and ethnicity. In the full multiple treatment model below, the corresponding
IV estimates remain higher than the OLS results, however they appear more
reasonable and stable across specications.
The Matching Results: We now move on to our matching estimates. We
adopt two sorts of matching, one-to-one matching with replacement (nearest
neighbour matching) and kernel based matching using the Epanechnikov kernel.
We estimate a probit model of qualications versus no qualications and match
on the predicted probability of obtaining qualications from this model. We
estimate the eect of obtaining qualications for those who actually obtained
these qualications (treatment on the treated) and for those who did not obtain
qualications (treatment on the non-treated). The rst involves matching each
person who obtained a qualication, with somebody who looks like them, but
did not obtain any qualications. The dierences in wages between these two
groups (each of equal size) is the estimated eect of obtaining a qualication
13
The coecients on the control functions (see section 3) are -0.003 (0.047) for no quali-
cations and -0.278 (0.109) for qualications.
39
for those who did so (those who did not obtain a qualication can be used
more than once). With kernel density matching, the wage of each individual
who obtained a qualication is compared with the weighted average of wages
for all individuals who did obtain a qualication who fall within a specied
bandwidth. Again the dierence in wage outcomes gives us our estimate of the
eect of obtaining a qualication, for those who did undertake a qualication.
The second involves matching everyone who did not obtain a qualication, with
a person who looked like them, but did obtain a qualication. The dierence
between the outcomes for these two groups gives an estimate of what this group
would have received, if they had instead decided to undertake a qualication. To
ensure common support, we impose a caliper of 0.0025 for our nearest neighbour
matching and a bandwidth of 0.06 for our kernel based matching
14
. We also
report the estimate when common support is not imposed.
If we look at the results for men in Table 1 we see that the estimated return
to obtaining a qualication for those who actually undertook the qualication
(treatment on the treated) is around 31 to 35 %
15
, whereas the estimated return
for those who did not (treatment on the non-treated) is around 23 to 24 %, a
signicant dierence of around 10 percentage points. The treatment on the
treated results are somewhere between our highest and lowest OLS estimates
but below our IV and CF estimates whereas the treated on the non-treated
results are below all these estimates. These results suggest that if those who
have no qualications were to undertake them, they would receive a lower return
than the group who had already undertaken them.
14
We experimented with the sensitivity of our estimates to the caliper and bandwidths used
but this made little dierence to our estimates.
15
We focus on the results where common support has been imposed.
40
Table 1: The returns to obtaining a school qualication - NCDS Men
Estimation Coef (S.E.) No.
technique and specication: obs
OLS:
(i) region and ethnicity 0.370 (0.017) 3639
(ii) (i) + ability & school type variables 0.275 (0.017) 3639
(iii) (ii) + family background (excl instruments) 0.255 (0.017) 3639
Instrumental Variables:
(iv) (i) using set I instruments 0.786 (0.038) 3639
(v) (iii) using set II instruments 0.474 (0.082) 3639
Control Function:
(vi) (iii) using set II instruments 0.377 (0.094) 3639
Matching - Eect of treatment on the treated:
(vii) Nearest neighbour - all obs 0.346 (0.034) 2988
(viii) Nearest neighbour - caliper (0.0025) 0.352 (0.033) 2815
(ix) Kernel weighting - all obs 0.334 (0.025) 2988
(x) Kernel weighting - bandwidth(0.06) 0.313 (0.025) 2795
Matching - Eect of treatment on the non-treated:
(xi) Nearest neighbour - all obs 0.231 (0.028) 651
(xii) Nearest neighbour - caliper(0.0025) 0.242 (0.028) 619
(xiii) Kernel weighting - all obs 0.235 (0.020) 646
(xiv) Kernel weighting -bandwidth (0.06) 0.236 (0.019) 643
Note: Matching standard errors are bootstrapped (390 replications)
41
4.2 The Multiple Treatment Model
We now turn to a more disaggregated analysis that focuses on the sequential
nature of educational qualications. To this end we separate the qualications
variable into those who stopped education after completing O levels or equiva-
lent, those who stopped after completing A levels or equivalent, and those who
completed O levels, A levels and higher education (HE). In this sequential model
we exploit the ordering of our outcomes. For the control function estimates we
use an ordered probit model in our rst stage for these three levels of education
over the base
16
. Our matching estimator uses an adaption of the propensity
score matching method for multiple sequential treatments and this is described
in detail in Appendix A.
The Instrumental Variable and Control Function Results: The OLS,
IV and control function results are shown in Table 2. The return to higher
education versus no qualications is obtained by adding the return to O levels
or equivalent, the return to an A level or equivalent and the return to HE.
Overall the results have a distinct pattern. Controlling for ability and school
type is important and reduces the return to education at all levels. The IV
results are higher than the OLS results. The Control Function estimates are
somewhat smaller than the IV estimates but lie above the OLS estimates.
These results are enormously informative. The fact that OLS estimates of
the impact of education on earnings are lower than either the IV or CF estimates
suggests that there is some attenuation bias in OLS most likely due to measure-
ment error. The nding that IV is often higher than CF is strong evidence of
heterogeneous responses, the IV estimate picking out some local average eect
16
From this we calculate the control function terms described in Section 3. In this approach
the control functions are interacted with each of the qualication levels. Standard errors
are corrected to take into account the generated regressor(s) in the model as well as for
heterescedasticity.
42
Table 2: Regression estimates of the returns to qualications - NCDS Men
(i) (ii) (iii) (iv) (v)
Variable OLS Basic OLS Full set of IV IV CF
Specication controls Spec (i) Spec (ii) Spec (ii)
O level or equivalent 0.207 0.166 0.315 0.289 0.221
(0.018) (0.018) (0.053) (0.010) (0.010)
A level or equivalent 0.094 0.077 0.170 0.166 0.155
(0.016) (0.016) (0.035) (0.068) (0.137)
HE 0.298 0.249 0.393 0.373 0.331
(0.016) (0.016) (0.029) (0.063) (0.153)
Number of observation 3639 3639 3639 3639 3639
which, as we have already indicated, can be quite high in certain subpopulations.
Among these estimates the CF estimate is probably the most reliable estimate
of the average impact of each education level on gross earnings
17
. The results
show signicant overall returns to educational qualications at each stage of
the educational process even after correcting for detailed background variables
and ability dierences, as well as allowing for heterogeneity in the education
response parameters.
The Matching Results: To implement our adapted propensity score match-
ing method for this sequential treatment case, we estimated three propensity
scores. We used the full set of variables available for matching but do not include
our set II instruments which we argued impacted on the education treatment
level but not directly (conditional on treatment) on the earning outcome. Such
instruments are not appropriate matching variables. The rst was a probit of
qualications (O levels, A levels and HE) versus no qualications; the second
was a probit of A-levels and HE versus O levels; and the third was a probit of
higher education versus A levels. On the basis of these three probits we es-
timated the unconditional probability of undertaking each of our 3 treatments
and no qualications (which in Appendix A we denote by r(j, X), j 0, 1, 2, 3)
for our whole sample, including those who did not make particular qualication
17
The coecients on the control functions are -0.061 (0.070) for no qualications; -0.099
(0.056) for O levels; -0.121(0.063) for A levels; and -0.084 (0.067) for HE.
43
transitions. We then compared the outcomes across each of our four groups,
matching on the appropriate one dimensional propensity score for the particu-
lar transition in question which again is dened in Appendix A.
Our approach involves estimating the incremental return to each of our three
qualications by actual qualication. For those with no qualications, we es-
timate the returns they would have got if they had undertaken each of the
three qualications (treatment on the non-treated). For those with O level
qualications we estimate the return they obtained for taking that qualication
(treatment on the treated) and the returns they would have obtained if they
had progressed to A levels or HE (treatment on the non-treated). For those
with A levels we estimate the returns they obtained for undertaking O and A
level qualications (treatment on the treated) and the returns they would have
obtained if they had progressed to HE (treatment on the non-treated). For
those with HE all estimates are treatment on the treated. In all of our tables
we italicise treatment on the non-treated results.
We present three sets of results. The rst set of results uses nearest neighbour
matching but does not impose common support. The second set uses nearest
neighbour matching but only accepts matches within a caliper of 0.0025 and
only includes individuals who are matched for every possible transition (so that
we can make comparisons across the same sets of individuals). The third set
uses kernel density matching using a bandwidth of 0.06 and again only includes
individuals who are matched at every possible transition. The results of doing
this are given in Table 3.
Focusing on the results in Table 3, we see that the estimate of a return to an
O level is generally lower when we impose common support (with the exception
of the eect estimated using nearest neighbour for those with no qualications).
If we focus on the treatment on the treated results the estimates of the re-
turns are very close using both methods and rise with the actual qualication
44
Table 3: Multiple Treatment Matching Results - NCDS Men
Actual Estimated Returns Number
Qualication: O level A level HE of
Est (SE) Est (SE) Est (SE) observations
None:
Nearest neighbour - all obs 0.245 (0.032) 0.023 (0.042) 0.286 (0.052) 651
Nearest neighbour - cal (0.0025) 0.270 (0.038) 0.054 (0.045) 0.258 (0.047) 443
Kernel denisty - bdwith (0.06) 0.150 (0.021) 0.065 (0.023) 0.273 (0.035) 619
O level:
Nearest neighbour - all obs 0.189 (0.030) 0.179 (0.025) 0.226 (0.034) 993
Nearest neighbour - cal (0.0025) 0.177 (0.032) 0.185 (0.031) 0.235 (0.036) 716
Kernel denisty - bdwith (0.06) 0.174 (0.022) 0.066 (0.018) 0.257 (0.021) 970
A level:
Nearest neighbour - all obs 0.231 (0.039) 0.062 (0.026) 0.367 (0.028) 965
Nearest neighbour - cal (0.0025) 0.208 (0.039) 0.053 (0.031) 0.368 (0.033) 726
Kernel denisty - bdwith (0.06) 0.187 (0.026) 0.064 (0.018) 0.260 (0.020) 922
HE:
Nearest neighbour - all obs 0.267 (0.064) 0.076 (0.045) 0.240 (0.031) 1030
Nearest neighbour - cal (0.0025) 0.250 (0.049) 0.036 (0.045) 0.263 (0.036) 595
Kernel denisty - bdwith (0.06) 0.241 (0.041) 0.077 (0.024) 0.227 (0.021) 897
Note: Treatment on the non-treated estimates italicised. Matching standard errors are
bootstrapped (390 replications)
45
obtained. The estimates range from 17.4% for those with O levels to 25% for
those with HE qualications. They are very close (on average) to the results
we obtained from the control function method (22.1%). The results for treat-
ment on the non-treated are more ambiguous. The nearest neighbour method
estimates an eect of 27.0% compared to an estimate of 15.0% using the kernel
density method. This illustrates clearly the problems in choosing an appropriate
matching strategy and the sensitivity of some matching results to the estima-
tion technique chosen. In our results it appears to be a particular problem with
some of the treatment on the non-treated estimates.
If we now turn to the estimates of the returns to A levels we see that the
treatment on the treated returns range between 3.6% and 7.7% which are unam-
biguously lower than all our regression based estimates. Most of the treatment
on the non-treated estimates range between 5.4% to 6.6%, but there is an es-
timate of 18.5% for those with O levels using nearest neighbour techniques.
Finally if we look at the return to HE estimates we see that with the exception
of nearest neighbour matching for those with A levels (which has an estimate of
36.8%), the estimates range between 22.7% and 27.3%. These latter estimates
are close to the OLS estimates whereas the former result is closer to the IV
and CF results. Interestingly, for HE the matching results suggest the eect
of treatment on the non-treated may be higher than treatment on the treated
which has very interesting policy implications. This dimension was missed in
our single treatment model.
5 Summary and Conclusions
The aim of this paper has been to review alternative methods and models
for the estimation of the eect of education on earnings, and to apply these to
a high quality common data source. We have highlighted the importance of the
46
model specication. In particular, the distinction between single treatment and
multiple treatment models. Also the importance of allowing for heterogeneous
returns - that is returns that vary across individuals for the same educational
qualication. We considered three main estimation methods which rely on dif-
ferent identifying assumptions - instrumental variable methods, control function
methods and propensity score matching methods. In each case the properties
were analysed distinguishing between a single treatment model and a model
where there are a sequence of possible treatments. The sequential multiple
treatment model we argued is well suited to the education returns formulation
where educational qualication levels in formal schooling tend to be cumulative.
With heterogeneous returns dening the parameter of interest is central.
We distinguished four possible parameters of interest: the treatment on the
treated, the local average treatment eect, the average treatment eect and
the impact of treatment on the non-treated. In the homogeneous eects model
these were all equal but in the heterogeneous eects model they can dier sub-
stantially. Which is of interest will depend on the policy question. Moreover,
dierent estimation methods were shown to identify dierent parameters of in-
terest.
Our application was to NCDS 1958 birth cohort study for Britain. This data
is ideally suited for evaluating the impact of education on earnings using non-
experimental data and is suciently rich to allow the comparison of matching,
control function and instrumental variable methods. First there are extensive
and commonly administered ability tests at early ages. Second there are accu-
rately measured family background and school type variables ideal for matching.
Finally there are variables that are very likely to inuence schooling but not
wage outcomes, such us temporary nancial diculties among the parents and
the composition of siblings. These make good choices for excluded instruments
in the application of instrumental variables or control function methods.
47
The application showed the importance of allowing for schooling choices to
depend on family background variables, nancial constraints, etc. - as well as
ability. Among the estimators used we found that OLS was very sensitive to
the inclusion of ability measures and family background variables. In general
these tended to reduce the estimated return. But there was strong evidence
of heterogeneous returns with the instrumental variable estimator often much
higher than OLS and by more than could be explained from attenuation bias
caused by measurement error in the qualication level data.
References
[1] Altonji, J. and T. Dunn (1996), Using Siblings to Estimate the Eect
of Schooling Quality on Wages, Review of Economics and Statistics, 78,
665-671.
[2] Angrist, J. D. and Krueger, A. B. (1991), Does Compulsory Schooling
Attendance Aect Schooling Decisions, Quarterly Journal of Economics,
106(4), 970-1014.
[3] Angrist, J. D. and Krueger, A. B. (1992), Estimating the payo to school-
ing using the Vietnam-era draft lottery, National Bureau of Economic
Research, Working Paper no. 4067.
[4] Ashenfelter, O. and Zimmerman, D. (1993), Estimates of the returns to
schooling from sibling data: fathers, sons and brothers, Princeton Univer-
sity, Industrial Relations Section, Working Paper no. 318.
[5] Behrman, J.R., Rosenzwieg, M.R. and P. Taubman (1996), College Choice
and Wages: Estimates using Data on Female Twins, Review of Economics
and Statistics, 78, 665-671.
[6] Blundell, R., Dearden, L., Goodman, A. and Reed, H. (2000), The Returns
to Higher Education in Britain: Evidence from a British Cohort, Economic
Journal, 110, F82-F99.
48
[7] Bonjour, D., L. Cherkas, J. Haskel, D. Hawkes and T Spector (2000), Es-
timating Returns to Education using a New Sample of UK Twins, mimeo,
QMW, June.
[8] Bound, J., D. Jaeger and R. Baker, (1995), Problems with Instrumental
Variables Estimation when the Correlation between the Instruments and
the Endogenous Explanatory Variable is Weak, Journal of the American
Statistical Association, 90, June, 443-50.
[9] Butcher, K. and Case, A. (1994), The eect of sibling sex composition
on womens education and earnings, Quarterly Journal of Economics, vol.
109, pp. 531-63.
[10] Card, D. (1999),The Causal Eect of Education on Earnings, in O.
Ashenfelter and D. Card, Handbook of Labor Economics, Vol 3, Elsevier-
North Holland.
[11] Cochran, W. and [Link] (1973), Controlling Bias in Observational
Studies, Sankyha, 35, 417-446.
[12] Dearden, L. (1999a), The eects of families and ability on mens education
and earnings in Britain, Labour Economics, vol. 6, pp. 551-567.
[13] Dearden, L. (1999b), Qualications and Earnings in Britain: How reliable
are conventional OLS estimates of the returns to education?, IFS Working
Paper No. 99/7.
[14] Dehejia, R.H. and Wahba, S. (1999), Causal Eects in Non-Experimental
Studies: Re-Evaluating the Evaluation of Training Programmes, Journal
of the American Statistical Association, 94, 1053-1062.
[15] Garen, J. (1984),The returns to Schooling: A Selectivity Bias Approach
with a Continuous Choice Variable, Econometrica, 52(5), 1199-1218.
[16] Griliches, Z. (1977), Estimating the returns to schooling: some economet-
ric problems, Econometrica, vol. 45, pp. 1-22.
[17] Harmon, C. and Walker, I. (1995), Estimates of the economic return to
schooling for the UK, American Economic Review, 85, 1278-86.
49
[18] Heckman, J. J. (1979), Sample selection bias as a specication error,
Econometrica, vol. 47, pp. 153-61.
[19] Heckman, J.J., Ichimura, H. and Todd, P. (1997), Matching As An Econo-
metric Evaluation Estimator: Evidence from Evaluating a Job Training
Programme, Review of Economic Studies, 64, 605-654.
[20] Heckman, J.J., Ichimura, H., and Todd, P (1998), Matching as an Econo-
metric Evaluation Estimator, Review of Economic Studies, 65, 261-294.
[21] Heckman, J.J., Ichimura, H., Smith, J. and Todd, P. (1998), Characterizing
selection bias using experimental data, Econometrica, vol. 66, pp. 1017-98.
[22] Heckman, J.J., R. Lalonde and J. Smith (1999), The Economics and
Econometrics of Active Labor Market programs in O. Ashenfelter and
D. Card, Handbook of Labor Economics, Vol 3, Elsevier-North Holland.
[23] Heckman, J. And Robb, R. (1985), Alternative methods for Evaluating the
Impact of Interventions, in Longitudinal Analysis of Labour market Data
(New York: Wiley).
[24] Heckman, J., Smith, J. And N. Clements, (1997), Making the Most out of
Program Evaluations and Social Experiments: Accounting for Heterogene-
ity in program Impacts, Review of Economic Studies, 64, 487-536.
[25] Imbens, G. (2000), The Role of Propensity Score in Estimating Dose-
Response Functions, Biometrika, 87, 3, 706-710.
[26] Imbens, G. and Angrist, J. (1994), Identication and Estimation of Local
Average Treatment Eects, Econometrica, 62, 2, 467-476.
[27] Lechner, M. and Miquel, R. (2001), A Potential Outcome Approach to
Dynamic Programme Evaluation - Part I: Identication, discussion paper,
SIAW, Universitt St. Gallen.
[28] Lechner, M. (2001), Identication and Estimation of Causal Eects of
Multiple Treatments under the Conditional Independence Assumption, in:
Lechner, M., Pfeier, F. (eds), Econometric Evaluation of Labour Market
Policies, Heidelberg: Physica/Springer, p. 43-58.
50
[29] Meghir, C. and M. Palme (2000), Estimating the Eect of Schooling on
Earnings Using a Social Experiment, IFS Working Paper 99/12, March.
[30] Rosenbaum, P.R. and Rubin, D.B. (1983), The Central Role of the Propen-
sity Score in Observational Studies for Causal Eects, Biometrika, 70, 41-
55.
[31] Rosenbaum, P.R. and Rubin, D.B. (1985), Constructing a Comparison
Group Using Multivariate Matched Sampling Methods That Incorporate
the Propensity Score, The American Statistician, 39, 33-38.
[32] Rubin, D.B. (1979), Using Multivariate Matched Sampling and regres-
sion Adjustment to Control Bias in Observational Studies, Journal of the
American Statistical Association, 74, 318-329.
[33] Rubin, D.B. (1980), Bias Reduction Using Mahalanobis-Metric Matching,
Biometrics, 36, 293-298.
[34] Staiger, D. and J. H. Stock (1997), Instrumental Variables Regressions
with Weak Instruments, Econometrica, 65(3), May, 557-86.
[35] Smith, R, and Blundell, R.W. (1986), An exogeneity Test for a Simulta-
neous Tobit Model with an Application to Labour Supply, Econometrica,
54, 679685.
[36] Willis, R. J. and Rosen, S. (1979), Education and self-selection, Journal
of Political Economy, vol. 87, pp. S1-S36.
A Appendix A: A Balancing Score for Sequen-
tial Multiple Treatments
In the educational context, we can view the sequential treatments (basic edu-
cation, O-levels, A-levels, higher education) in a dose-response framework (cf.
Imbens, 2000). Like a drug which can be applied in dierent doses, the se-
quential treatments would thus correspond to ordered levels of a treatment
education (or investment in human capital). We focus on continuous education,
where individuals take uninterrupted sequential decisions of an incremental na-
ture: at each point, they can either stop or move on to the next educational
level.
51
We consider four treatments: D {0, 1, 2, 3}
D = 0 for stopping at basic (i.e. no qualications)
D = 1 for stopping at O-Levels (i.e. stay on and stop with O-levels)
D = 2 for stopping at A-Levels (i.e. stay on, take O-levels and stop with
A-levels)
D = 3 for stopping at Higher Education (i.e. stay on, take O-levels, take
A-levels and stop with HE)
Incidentally, we can link this analysis to the dynamic programme evaluation
framework recently suggested by Lechner and Miquel (2001). In our case of an
obliged chain of educational choices, we only have a restricted set of possible
sequences, four in fact. In addition, in each period there is only one type (and
in fact a dierent type) of treatment available.
Consider four periods: in period 0 everyone achieves basic qualications, in
period 1 the relevant choice is whether to take O-levels or not; in period 2 the
only treatment available is A-levels but only provided one has achieved O-levels
in the previous period; while the treatment in period 3 is higher education but
available only for those with A-levels. Outcome Y is then observed after period
3 (at age 33). The four possible sequences, corresponding to the four values of
D dened above, are thus:
t 0 1 2 3
D = 0 1 0 0 0
D = 1 1 1 0 0
D = 2 1 1 1 0
D = 3 1 1 1 1
Dene the sequence of treatments, each received at the beginning of each
period t, as S = (S
0
, S
1
, S
2
, S
3
), in our case, S
0
= 1 for all; for t = 1, 2, 3,
S
t
{0, 1} and S
t
= 1 =S
t1
= 1.
The sequential nature of the decision process is captured by the modelling
of the choice probabilities as follows:
Dene
52
r(j, x) Pr{D = j|X = x} for j = 0, 1, 2, 3.
Further dene the following probabilities:
P
1
(x) Pr{(OL AL HE) = 1|X = x} - the stay-on probability
P
2
(x) Pr{(AL HE) = 1|OL = 1, X = x}
P
3
(x) Pr{HE = 1|AL = 1, X = x}
Thus we have:
D =
_

_
0 r(0, x) = 1 P
1
(x)
1 r(1, x) = (1 P
2
(x))P
1
(x)
2 r(2, x) = (1 P
3
(x))P
2
(x)P
1
(x)
3 r(3, x) = P
3
(x)P
2
(x)P
1
(x)
Let Y
k
denote the outcome, Y lny, if the individual were to receive treatment
(or education level) k.
We are interested in the following 12 pairwise comparisons of the eects of
treatments (education levels) m and l, with m, l {0, 1, 2, 3} and m > l:
E(Y
m
Y
l
| D = j)
for j = m (eect on the treated) and j = l (eect on the non-treated).
Note that in the framework by Lechner and Miquel (2001), evaluating the
impact of stopping at one education level versus stopping at another one for
those individuals who have stopped at one of such levels amounts to evaluat-
ing the eect of one sequence of treatments compared to another sequence of
treatments (each of length four) for those individuals who have followed a given
sequence.
In our case, we need to identify all the counterfactuals of the type:
E(Y
m
| D = l) and E(Y
l
| D = m)
(More precisely, we need to identify E(Y
k
| D = j) for
(k, j) = (0, 1), (0, 2), (0, 3), (1, 2), (1, 3), (2, 3) (eect on the treated)
as well as for
(k, j) = (1, 0), (2, 0), (3, 0), (2, 1), (3, 1), (3, 2) (eect on the non-treated). )
An extension of the conditional independence assumption MM: A1 that
would allow to identify them is what Imbens (2000) termed strong unconfound-
edness:
{Y
0
, Y
1
, Y
2
, Y
3
} D| X
0
= x
0
53
This CIA corresponds to Lechner and Miquel (2001) full CIA Assumption 2-I or
2-II, and can be rewritten in terms of potential outcomes each one corresponding
to one of the sequences of treatments:
Y
1000
, Y
1100
, Y
1110
, Y
1111
S
t
| X
0
for all t = 1, 2, 3
In words, conditional on the information observed prior to period 0, X
0
, as-
signment to treatment in each period is independent of potential outcomes, in
particular it is not aected by any new information related to the outcomes that
may arrive in between schooling choices.
This implies that the complete treatment sequence, in our case the maximum
level of education attained, is chosen at the beginning of period 0 just as the
dose of a drug is decided at the start based on the information contained in
X
0
.
The assumption that subsequent schooling choices are not aected by the
outcomes of the schooling decisions in the previous periods hinges on the absence
of intermediate outcomes on which to possibly base future S decisions. This
amounts to ruling out intermittent educational choices - where an individual
achieves a level of education, drops out of the education system, observes the
corresponding outcomes (both in terms of Y and of possibly endogenous Xs)
and then possibly decides on re-entering the schooling system for investment in
the next level of education.
Note however that the weaker form (implied by strong unconfoundedness)
would suce to our purposes:
{Y
l
, Y
m
} D| X
0
= x
0
, D {l, m}
for (l, m) = (0, 1), (0, 2), (0, 3), (1, 2), (1, 3), (2, 3)
This relaxes the CIA above by requiring conditional independence to hold only
for the subpopulations receiving treatment m or treatment l (see Lechner 2001).
The common support assumption corresponding to MM: A2 is:
0 < r(j, x) Pr{D = j|X
0
= x} < 1 for x C

and j = 0, 1, 2, 3
which for the P
j
s implies the requirements:
0 < P
j
(x) < 1 for x C

and j = 1, 2, 3.
54
A.1 Looking for a balancing score
If we wanted the X

s (in what follows we drop the time-0 subscript from X)

to be simultaneously balanced in the four groups dened by the highest level
of education attained, i.e. if we required the same distribution in four selected
(matched) subgroups of the four types of treated, we would need to look for a
function of the X

s, b(X), such that (cf. Theorem 2 by Rosenbaum and Rubin,

1983 and Proposition 1 in Lechner 2001):
X D | b(X)
E(Pr{D = m|X} | b(X)) = Pr{D = m|X} m = 0, 1, 2, 3
Setting up the corresponding system:
_

_
E(Pr{D = 0|X} | b(X)) =
E(Pr{D = 1|X} | b(X)) =
E(Pr{D = 2|X} | b(X)) =
E(Pr{D = 3|X} | b(X)) =
r(0, x) =
r(1, x) =
r(2, x) =
r(3, x) =
1 P
1
(X)
(1 P
2
(X))P
1
(X)
(1 P
3
(X))P
2
(X)P
1
(X)
P
3
(X)P
2
(X)P
1
(X)
Choosing either b(X) = {r(1, x), r(2, x), r(3, x)} or b(X) = {P
1
(X), P
2
(X), P
3
(X)}
would solve the system. (Note that the dimensionality has been reduced by one;
this is allowed by the adding up of the treatment probabilities).
We are however just interested in the pairwise comparison of the various
levels of the treatment, so that the above balancing score may actually be more
restrictive than required for some type of comparison.
A.1.1 A balancing score for the pairwise comparisons
1. E(Y
1
Y
0
| D = j) for j = 0, 1
In this case, we just need
X D | b(X), D {0, 1}
which is veried if
E(Pr{D = 1|X, D {0, 1}}| b(X)) = Pr{D = 1|X, D {0, 1}}
One could use the propensity score:
Pr{D = 1|X, D {0, 1}} =
r(1, X)
r(1, X) +r(0, X)
55
which is itself a function of P
1
(X) and P
2
(X),
so that alternatively a balancing score for the problem is b(X) = {P
1
(X), P
2
(X)}.
For the other set of parameters, a ner score than the propensity score is
always the 3-dimensional b(X) = {P
1
(X), P
2
(X), P
3
(X)}:
When matching on the 1-dimensional propensity score, imposing common
support can be done in terms of this scalar; when matching on the other
balancing scores -> imposed on each element.
2. E(Y
2
Y
0
| D = j) for j = 0, 2
The propensity score is
r(2,X)
r(2,X)+r(0,X)
,
so that a balancing score for the problem is b(X) = {P
1
(X), P
2
(X), P
3
(X)}.
3. E(Y
2
Y
1
| D = j) for j = 1, 2
The propensity score is
r(2,X)
r(2,X)+r(1,X)
,
so that a balancing score for the problem is b(X) = {P
1
(X), P
2
(X), P
3
(X)}.
4. E(Y
3
Y
0
| D = j) for j = 0, 3
The propensity score is
r(3,X)
r(3,X)+r(0,X)
,
so that a balancing score for the problem is b(X) = {P
1
(X), P
2
(X), P
3
(X)}.
5. E(Y
3
Y
1
| D = j) for j = 1, 3
The propensity score is
r(3,X)
r(3,X)+r(1,X)
,
so that a balancing score for the problem is b(X) = {P
1
(X), P
2
(X), P
3
(X)}.
6. E(Y
3
Y
2
| D = j) for j = 2, 3
the propensity score is
r(3,X)
r(3,X)+r(2,X)
,
so that a balancing score for the problem is b(X) = {P
1
(X), P
2
(X), P
3
(X)}.
56
Table 4: Appendix B: Summary Statistics: NCDS Men
Variable 3639 Observations
Mean (S.D.)
Real log hourly wage 1991 2.040 (0.433)
Qualications:
O levels or equivalent 0.821 (0.383)
A levels or equivalent 0.548 (0.498)
Higher Education 0.283 (0.451)
White 0.969 (0.173)
Maths ability at 7:
5th quintile (highest) 0.212 (0.408)
4th quintile 0.190 (0.392)
3rd quintile 0.185 (0.389)
2nd quintile 0.158 (0.365)
1st quintile (lowest) 0.141 (0.348)
Reading ability at 7:
5th quintile (highest) 0.165 (0.371)
4th quintile 0.187 (0.390)
3rd quintile 0.188 (0.391)
2nd quintile 0.179 (0.383)
1st quintile (lowest) 0.166 (0.372)
Ability at 7 missing 0.115 (0.319)
Comprehensive school 1974 0.468 (0.499)
Secondary modern school 1974 0.162 (0.368)
Grammar school 1974 0.099 (0.299)
Private school 1974 0.052 (0.222)
Other school 1974 0.018 (0.134)
Fathers years of education 7.270 (4.827)
Fathers education missing 0.172 (0.377)
Mothers years of education 7.342 (4.606)
Mothers education missing 0.159 (0.366)
Fathers age 1974 43.171 (13.736)
Fathers age missing 0.075 (0.263)
Mothers age 1974 41.475 (10.864)
Mothers age missing 0.049 (0.216)
Fathers social class 1974:
Professional 0.044 (0.205)
Intermediate 0.145 (0.352)
Skilled non-manual 0.076 (0.265)
Skilled manual 0.297 (0.457)
Semi-skilled non-manual 0.010 (0.098)
Semi-skilled manual 0.095 (0.293)
missing 0.106 (0.308)
Mother employed 1974 0.513 (0.500)
Number of siblings 1.692 (1.789)
Number of siblings missing 0.106 (0.308)
Number of older siblings 0.821 (1.275)
Fathers interest in edn:
Expects too much 0.013 (0.114)
Very interested 0.252 (0.434)
Some interest 0.215 (0.411)
Mothers interest in edn:
Expects too much 0.032 (0.175)
Very interested 0.344 (0.475)
Some interest 0.354 (0.478)
Bad nances 1969 or 1974 0.159 (0.365)
Region 1974:
North Western 0.100 (0.300)
North 0.070 (0.256)
East & West Riding 0.079 (0.270)
North Midlands 0.072 (0.258)
Eastern 0.073 (0.261)
London & South East 0.143 (0.350)
Southern 0.057 (0.232)
South Western 0.061 (0.240)
Midlands 0.088 (0.283)
Wales 0.054 (0.227)
Scotland 0.096 (0.295)
57