"The Apple iPod iTunes Anti-Trust Litigation"

Filing 753

***ERRONEOUS ENTRY, PLEASE REFER TO DOCUMENT NO. 754 *** EXHIBITS re 752 Opposition/Response to Motion, filed byApple Inc.. (Attachments: # 1 Exhibit 2, # 2 Exhibit 3, # 3 Exhibit 4, # 4 Exhibit 5, # 5 Exhibit 6, # 6 Exhibit 7, # 7 Exhibit 8, # 8 Exhibit 9, # 9 Exhibit 11, # 10 Exhibit 12, # 11 Proposed Order)(Related document(s) 752 ) (Kiernan, David) (Filed on 1/14/2014) Modified on 1/14/2014 (jlmS, COURT STAFF).

Exhibit 8

NBER WORKING PAPER SERIES

WHAT ARE WE WEIGHTING FOR?
Gary Solon
Steven J. Haider
Jeffrey Wooldridge
Working Paper 18859
http://www.nber.org/papers/w18859
NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue
Cambridge, MA 02138
February 2013

We are grateful for helpful comments from David Autor, Raj Chetty, John DiNardo, Todd Elder, Mike
Elsby, Osborne Jackson, Fabian Lange, Jason Lindo, Jim Ziliak, and seminar participants at the University
of Kentucky. The views expressed herein are those of the authors and do not necessarily reflect the
views of the National Bureau of Economic Research.
NBER working papers are circulated for discussion and comment purposes. They have not been peerreviewed or been subject to the review by the NBER Board of Directors that accompanies official
NBER publications.
© 2013 by Gary Solon, Steven J. Haider, and Jeffrey Wooldridge. All rights reserved. Short sections
of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full
credit, including © notice, is given to the source.

What Are We Weighting For?
Gary Solon, Steven J. Haider, and Jeffrey Wooldridge
NBER Working Paper No. 18859
February 2013
JEL No. C1
ABSTRACT
The purpose of this paper is to help empirical economists think through when and how to weight the
data used in estimation. We start by distinguishing two purposes of estimation: to estimate population
descriptive statistics and to estimate causal effects. In the former type of research, weighting is called
for when it is needed to make the analysis sample representative of the target population. In the latter
type, the weighting issue is more nuanced. We discuss three distinct potential motives for weighting
when estimating causal effects: (1) to achieve precise estimates by correcting for heteroskedasticity,
(2) to achieve consistent estimates by correcting for endogenous sampling, and (3) to identify average
partial effects in the presence of unmodeled heterogeneity of effects. In each case, we find that the
motive sometimes does not apply in situations where practitioners often assume it does. We recommend
diagnostics for assessing the advisability of weighting, and we suggest methods for appropriate inference.
Gary Solon
Department of Economics
Michigan State University
East Lansing, MI 48824-1038
and NBER
[e-mail]
Steven J. Haider
Department of Economics
Michigan State University
101 Marshall Hall
East Lansing, MI 48824
[e-mail]

Jeffrey Wooldridge
Department of Economics
Michigan State University
[e-mail]

What Are We Weighting For?

I. Introduction
At the beginning of their textbook’s section on weighted estimation of regression
models, Angrist and Pischke (2009, p. 91) acknowledge, “Few things are as confusing to
applied researchers as the role of sample weights. Even now, 20 years post-Ph.D., we
read the section of the Stata manual on weighting with some dismay.” After years of
discussing weighting issues with fellow economic researchers, we know that Angrist and
Pischke are in excellent company. In published research, top-notch empirical scholars
make conflicting choices about whether and how to weight, and often provide little or no
rationale for their choices. And in private discussions, we have found that accomplished
researchers sometimes own up to confusion or declare demonstrably faulty reasons for
their weighting choices.
Our purpose in writing this paper is to dispel confusion and dismay by clarifying
the issues surrounding weighting. Our central theme is that the confusion stems from a
lack of clarity about which among multiple potential motives for weighting pertains to
the research project at hand. Once one specifies the particular motive for weighting, it
becomes straightforward to consider whether the purpose for weighting really does apply,
to use appropriate diagnostics to check whether it does, and then to proceed with
appropriate estimation and inference methods. Hence the title of our paper: “What Are
We Weighting For?”
In the next section, we pose a prior question: “What Are We Trying to Estimate?”
In some projects, the purpose is to estimate descriptive statistics for a particular
population. In those cases, whether weighting is called for depends simply on whether

1

weighting is necessary to make the analysis sample representative of the target
population. But in many other projects, the purpose is to estimate causal effects. In those
cases, the weighting issue becomes more nuanced.
In Sections III, IV, and V, we successively discuss three distinct potential motives
for weighting when estimating causal effects: (1) to achieve more precise estimates by
correcting for heteroskedasticity, (2) to achieve consistent estimates by correcting for
endogenous sampling, and (3) to identify average partial effects in the presence of
heterogeneous effects. 1 In each case, after explaining the potential relevance of the
motive, we will note that the motive sometimes does not apply in situations where
practitioners often assume it does. We will recommend diagnostics for assessing the
advisability of weighting, and we will suggest methods for appropriate inference. In
Section VI, we will summarize our analysis and our recommendations for empirical
practice.

II. What Are We Trying to Estimate?
A. Descriptive Statistics for a Population
Sometimes the purpose of a research project is to estimate descriptive statistics of
interest for a population. Consider, for example, the 1967 poverty rate for the United
States, which was officially measured as 13 percent based on the Current Population
Survey (U.S. Bureau of the Census, 1968). But suppose that one sought to estimate that
rate on the basis of the reports of 1967 income in the first wave of the Panel Study of
Income Dynamics (PSID) in 1968. The complication is that the PSID began with a

1

The use of propensity-score weighting to control for covariates when estimating treatment effects is
discussed elsewhere in this symposium by Imbens ( ).

2

sample that purposefully overrepresented low-income households by incorporating a
supplementary sample drawn from households that had reported low income to the
Survey of Economic Opportunity in 1967. As in other surveys that purposefully sample
with different probabilities from different parts of the population, the point of the
oversampling was to obtain more precise information on a subpopulation of particular
interest, in this case the low-income population. 2
If one estimated the 1967 poverty rate for the United States population with the
poverty rate for the full PSID sample, without any weighting to adjust for the low-income
oversample, one would estimate the U.S. poverty rate at 26 percent.3 That, of course, is
an upward-biased estimate because the PSID, by design, overrepresents the poor. But
one might achieve unbiased and consistent estimation by using the PSID sample’s
weighted poverty rate, weighting by the inverse probabilities of selection. 4

A

visualization of how this works is that the PSID sample design views the U.S. population
through a funhouse mirror that exaggerates the low-income population.

Weighted

estimation views the sample through a reverse funhouse mirror that undoes the original
exaggeration. It turns out that the PSID’s weighted poverty rate is 12 percent, a more
reasonable estimate than the 26 percent figure.
The poverty-rate example illustrates the simple case of estimating a population
mean on the basis of a sample that systematically fails to represent the target population,
but can be made to represent it by weighting. Much economic research, however, seeks
2

Similarly, the Current Population Survey oversamples in less populous states, and the first wave of the
Health and Retirement Study oversampled blacks, Mexican-Americans, and residents of Florida.
3
This calculation is based on approximating the official poverty line by dividing the PSID-reported “needs
standard” by 1.25.
4
For simplicity, we are overlooking complications from nonresponse, including mishaps in the PSID’s
implementation of the low-income oversample. For discussion of the latter and further references, see Shin
and Solon (2011, footnote 11)

3

to estimate more complex population statistics.

Suppose, for example, that the

population descriptive statistic one wishes to estimate is the 1967 earnings gap between
black and white men with the same years of schooling and potential work experience (age
minus years of schooling minus 5). A typical approach is to attempt to estimate the
population linear projection of log earnings on a dummy variable that equals 1 for blacks
along with controls for years of schooling and a quartic in potential experience. 5
Now suppose that one estimates that population regression by performing
ordinary least squares (OLS) estimation of the regression of log earnings on the race
dummy, years of schooling, and a quartic in potential earnings for black and white male
household heads in the PSID sample. Doing so estimates the coefficient of the dummy
variable for blacks at -0.344. Because exp (-0.344) = 0.71, this estimate seems to imply
that, among male household heads with the same education and potential experience,
blacks tended to earn only 71 percent as much as whites.
As in the example of estimating the poverty rate, however, this estimate might be
distorted by the PSID’s oversampling of low-income households, which surely must lead
to an unrepresentative sample with respect to male household heads’ earnings. But again,
one can apply a reverse funhouse mirror by using weights. In particular, instead of
applying ordinary (i.e., equally weighted) least squares to the sample regression, one can
use weighted least squares (WLS), minimizing the sum of squared residuals weighted by
the inverse probabilities of selection.

Doing so leads to an estimated coefficient of

− 0.260 for the dummy variable for blacks, implying that, among male household heads

with the same education and potential experience, blacks tended to earn 77 percent as
5

Alternatively, one could control for race differences in covariates through propensity-score weighting.
See Imbens ( ) for a general discussion and Elder, Goddeeris, and Haider (2011) for an application to the
black/white difference in infant mortality rates.

4

much as whites. This is still a large shortfall, but not as large as implied by the OLS
estimate. A likely reason is that the particular way that the PSID overrepresented the
low-income population involved an especially concentrated oversampling of low-income
households in nonmetropolitan areas of the South. The unweighted PSID therefore may
understate typical income for blacks even more than for whites.
What both our examples have in common is that they involve estimating
descriptive statistics for a population on the basis of sample data. If the sample is
representative of the target population (the most straightforward case being a simple
random sample drawn from that population), the population statistic is consistently
estimated by the analogous sample statistic.

If the sample is systematically

unrepresentative of the population in a known manner, the population statistic generally
is not consistently estimated by the analogous sample statistic, but it can be consistently
estimated by reweighting the sample statistic with the inverse probabilities of selection. 6
This point is intuitive and not at all controversial. So why does the issue of
weighting provoke confusion and dismay among economic researchers? The answer,
which will occupy the rest of this paper, is that much economic research is directed not at
estimating population descriptive statistics, but at estimating causal effects.

B. Causal Effects
In the microeconometrics textbooks of both Wooldridge (2010) and Angrist and
Pischke (2009), the very first page describes the estimation of causal effects as the
principal goal of empirical microeconomists. According to Angrist and Pischke, “In the
beginning, we should ask, What is the causal relationship of interest? Although purely
6

For a general formal demonstration, see Wooldridge (1999).

5

descriptive research has an important role to play, we believe that the most interesting
research in social science is about questions of cause and effect, such as the effect of
class size on children’s test scores. . . .” Similarly, the first sentences in the Wooldridge
textbook are, “The goal of most empirical studies in economics and other social sciences
is to determine whether a change in one variable, say w, causes a change in another
variable, say y. For example, does having another year of education cause an increase in
monthly salary? Does reducing class size cause an improvement in student performance?
Does lowering the business property tax rate cause an increase in city economic
activity?”
In contrast to the case of estimating population descriptive statistics, when
economists perform estimation of causal effects, the question of whether to weight the
data is complex. There are several distinct reasons that we may (or, as we will stress,
sometimes may not) prefer to use weights in our estimation. We will take up these
distinct reasons separately in each of the next three sections.

III. Correcting for Heteroskedasticity
One motivation for weighting, taught for decades in undergraduate and graduate
econometrics classes, is to correct for heteroskedastic error terms and thereby achieve
more precise estimation of coefficients in linear or nonlinear regression models of causal
effects. A nice example of this motivation comes from the literature on the impact of
unilateral divorce laws on the divorce rate. During the 1970s, many states in the United
States adopted laws allowing unilateral divorce, instead of requiring mutual consent of
both spouses. Were these laws responsible for the rise in divorce rates that occurred

6

during that period? In two insightful and influential articles published in the American
Economic Review, Leora Friedberg (1998) and Justin Wolfers (2006) reported
differences-in-differences estimates of the impact of unilateral divorce laws on divorce
rates. In particular, using a panel of annual state divorce rates over time, they estimated
linear regressions of the divorce rate on dummy variables for unilateral divorce laws with
controls for state fixed effects and secular time trends. Following the practice of many
other top-notch empirical economists, 7 both Friedberg and Wolfers weighted by
state/year population in the estimation of their regression models. Friedberg justified the
weighting as a correction for population-size-related heteroskedasticity in the state/year
error terms.
Table 1 presents examples from a wide set of variations on the Friedberg/Wolfers
regressions reported in Lee and Solon (2011).

The regressions are estimated with

Wolfers’s 1956-1988 data on annual divorce rates by state. The main point of Wolfers’s
article was that the short-run and long-run effects of unilateral divorce may differ, so the
regressions in Table 1 follow Wolfers in representing unilateral divorce with a set of
dummy variables for whether unilateral divorce had been in place for up to 2 years, 3-4
years, 5-6 years, …, 13-14 years, or at least 15 years. The dependent variable is the
logarithm of the annual divorce rate by state, and the regressions include controls for
state fixed effects, year fixed effects, and state-specific linear time trends.
The table’s first column follows Friedberg and Wolfers in estimating by weighted
least squares with weighting by state/year population. The second column uses ordinary
least squares, which weights all observations equally. In both instances, to maintain
7

Some other prominent examples of similarly weighted estimation are Card and Krueger (1992), Autor,
Katz, and Krueger (1998), Levitt (1998), Donohue and Levitt (2001), Borjas (2003), and Dehejia and
Lleras-Muney (2004).

7

agnosticism about which weighting approach – if either – comes close to producing a
homoskedastic error term, Table 1 reports standard error estimates robust to
heteroskedasticity (as well as to serial correlation over time within the same state). 8
Setting aside other interesting aspects of these results (for example, the absence in
this specification of any evidence for a positive effect of unilateral divorce on divorce
rates), notice this striking pattern:

Even though Friedberg’s expressed purpose in

weighting was to improve the precision of estimation, the robust standard error estimates
are smaller for OLS than for WLS. For the estimated effects over the first eight years
after adoption of unilateral divorce, the robust standard error estimates for OLS are only
about half those for WLS. Apparently, weighting by population made the estimates much
less precise! And as discussed by Dickens (1990), this is quite a common phenomenon.
To see what’s going on here, let’s start with the classic heteroskedasticity-based
argument for weighting when the dependent variable is a group average and the averages
for different groups are based on widely varying within-group sample sizes. Suppose the
model to be estimated is
y i = X i β + vi

(1)

where y i is a group-level average outcome observed for group i and the error term is
fully independent of the explanatory variables. The group-average error term vi equals
Ji

∑v
j =1

ij

/ J i , where vij is the micro-level error term for individual j in group i and J i

denotes the number of individuals observed in group i. If vij is independently and

8

Lee and Solon (2011) show that, for both the OLS and WLS results, naïve standard error estimates that
correct for neither heteroskedasticity nor serial correlation are far smaller than the robust estimates. This
occurs mainly because the error term is highly serially correlated.

8

identically distributed with variance σ 2 , then elementary statistics shows that the
variance of the group-average error term vi is σ 2 / J i . Thus, if J i varies widely across
groups (e.g., if many more individuals are observed in California than in Wyoming), the
group-average error term vi is highly heteroskedastic. Then, as taught in almost every
introductory econometrics course, OLS estimation of β in equation (1) is inefficient and
also leads to inconsistent standard error estimation if nothing is done to correct the
standard error estimates for heteroskedasticity. The WLS estimator that applies least
squares to the reweighted equation
(2)

J i y i = J i X i β + J i vi

is the minimum-variance linear unbiased estimator and also generates consistent standard
error estimation.
This presumably is the line of thinking that led Friedberg and Wolfers to use WLS
to estimate their divorce-rate regressions. Compared to Wyoming, California offers
many more observations of the individual-level decision of whether or not to divorce, and
therefore it seems at first that weighting by state population should lead to more precise
coefficient estimation. And yet, for the specification shown in Table 1, it appears that
weighting by population harms the precision of estimation.
What is going on here is explained in Dickens’s (1990) excellent article subtitled
“Is It Ever Worth Weighting?” Dickens points out that, in many practical applications,
the assumption that the individual-level error terms vij are independent is wrong.
Instead, the individual-level error terms within a group are positively correlated with each
other because they have unobserved group-level factors in common. In current parlance,

9

the individual-level error terms are “clustered.” Dickens illustrates with the simple
example of an error components model for the individual-level error term:
(3)

vij = ci + u ij

where each of the error components, ci and u ij , is independently and identically
distributed (including independence of each other), with respective variances σ c2 and σ u2 .
In this scenario, the variance of the group-average error term vi is not σ 2 / J i , but
rather is
(4)

Var (vi ) = σ c2 + (σ u2 / J i ) .

If σ c2 is substantial and the sample size J i is sufficiently large in every group (e.g., a lot
of people live in Wyoming, even if not nearly as many as in California), the variance of
the group-average error term may be well approximated by σ c2 , which is homoskedastic.
In that case, OLS applied to equation (1) is nearly the best linear unbiased estimator. In
contrast, if one weights by

J i , as in equation (2), the reweighted error term has

variance J iσ c2 + σ u2 , which could be highly heteroskedastic.

This provides an

explanation for why weighting by the within-group sample size sometimes leads to less
precise estimation than OLS. 9 On the other hand, if σ c2 is small and the within-group
sample size J i is highly variable and small in some groups, weighting by the within-

9

An important related point is that, if one has access to the individual-level data on y ij and applies OLS to

the regression of y ij on X i , this is numerically identical to the group-average WLS of equation (2), and
hence suffers from the same inefficiency associated with ignoring the clustered nature of the error term.
For more discussion of the mapping between individual-level and group-average regressions, see
Wooldridge (2003) and Donald and Lang (2007).

10

group sample size may indeed improve the precision of estimation, sometimes by a great
deal.
So what is a practitioner to do? Fortunately, as Dickens points out, it is easy to
approach the heteroskedasticity issue as an empirical question. One way to go is to start
with OLS estimation of equation (1), and then use the OLS residuals to perform the
standard heteroskedasticity diagnostics we teach in introductory econometrics.

For

example, in this situation, the modified Breusch-Pagan test described in Wooldridge
(2013, pp. 276-8) comes down to just applying OLS to a simple regression of the squared
OLS residuals on the inverse within-group sample size 1 / J i . The significance of the tratio for the coefficient on 1 / J i indicates whether the OLS residuals display significant
evidence of heteroskedasticity. The test therefore provides some guidance for whether
weighted estimation seems necessary.

A remarkable feature of this test is that the

estimated intercept consistently estimates σ c2 , and the estimated coefficient of 1 / J i
consistently estimates σ u2 . This enables an approximation of the variance structure in
equation (4), which then can be used to devise a more refined weighting procedure that,
unlike the simple weighting scheme in equation (2), takes account of the group error
component ci .
So our first recommendation to practitioners in this situation is not to assume that
heteroskedasticity is (or is not) an issue, but rather to perform appropriate diagnostics
before deciding.

We wish to make two additional recommendations.

One is that,

regardless of whether one uses weighted or unweighted estimation, the inevitable
uncertainty about the true variance structure means that some heteroskedasticity may

11

remain in the error term. We therefore recommend reporting heteroskedasticity-robust
standard error estimates.
Finally, it often is good practice to report both weighted and unweighted
estimates. For one thing, as in our divorce example, a comparison of the robust standard
error estimates is instructive about which estimator is more precise. But there is an
additional consideration. Under exogenous sampling and correct specification of the
conditional mean of y in equation (1), both OLS and WLS are consistent for estimating
the regression coefficients. On the other hand, under either the endogenous sampling
discussed in the next section or model misspecification (an example of which is the
failure to model heterogeneous effects, to be discussed in Section V), OLS and WLS
generally have different probability limits. Therefore, as suggested by DuMouchel and
Duncan (1983), the contrast between OLS and WLS estimates can be used as a diagnostic
for model misspecification or endogenous sampling. 10
In truth, of course, the parametric models we use for estimating causal effects are
nearly always misspecified at least somewhat. Thus, the practical question is not whether
a chosen specification is exactly the true data-generating process, but rather whether it is
a good enough approximation to enable nearly unbiased and consistent estimation of the
causal effects of interest. When weighted and unweighted estimates contradict each
other, this may be a red flag that the specification is not a good enough approximation to
the true form of the conditional mean. For example, Lee and Solon (2011) find that,
when the dependent variable used in the divorce-rate regressions is specified not in logs,
but in levels (as in both the Friedberg and Wolfers studies), the OLS and WLS estimates
10

See Deaton (1997, p. 72) for a clear exposition of how to assess the statistical significance of the contrast
between OLS and WLS estimates of a linear regression model when, under the null hypothesis, OLS is
efficient. For a more general treatment, see Wooldridge (2001, pp. 463-4).

12

are dramatically different from each other. This in itself does not pinpoint exactly what is
wrong with the linear-in-levels model specification, but it is a valuable warning sign that
the issue of functional form specification warrants further attention.

IV. Correcting for Endogenous Sampling
An altogether different motive for weighting in research on causal effects is to
achieve consistent estimation in the presence of endogenous sampling. A nice example
comes from the classic paper on choice-based sampling by Manski and Lerman (1977).
Suppose one is studying commuters’ choice of transit mode, such as the choice between
driving to work and taking the bus. One might be particularly interested in how certain
explanatory variables, such as bus fare and walking distance to and from bus stops, affect
the probability of choosing one mode versus the other. Given a random sample of
commuters, most empirical researchers would perform maximum likelihood estimation of
a probit or logit model for the binary choice between transit modes.
But suppose the sample is drawn not as a random sample of commuters, but as a
choice-based sample. As Manski and Lerman explain, “in studying choice of mode for
work trips, it is often less expensive to survey transit users at the station and auto users at
the parking lot than to interview commuters at their homes.” Manski and Lerman show
that, if the resulting sample overrepresents one mode and underrepresents the other
relative to the population distribution of choices, maximizing the conventional log
likelihood (which is an incorrect log likelihood because it fails to account for the
endogenous sampling) generally results in inconsistent parameter estimation. 11 And if

11

They also note a quirky exception: In a logit model that includes mode-specific intercepts in the
associated random-utility model, the coefficients of the other explanatory variables are consistently

13

instead one maximizes the quasi-log likelihood that weights each observation’s
contribution to the conventional log likelihood by its inverse probability of selection from
the commuter population (thus using a reverse funhouse mirror to make the sample
representative of the population), consistent estimation of the parameters is restored.
Another example is estimating the earnings return to an additional year of
schooling. Most labor economists would frame their analysis within a linear regression
of log earnings on years of schooling with controls for other variables such as years of
work experience. Although that regression model has been estimated countless times by
OLS, researchers cognizant of the endogeneity of years of schooling often have sought to
devise instrumental variables (IV) estimators of the regression.

In any case, if the

regression were estimated with the full PSID without any correction for the oversampling
of the low-income population, this would lead to inconsistent estimation of the regression
parameters. The sampling would be endogenous because the sampling criterion, family
income, is related to the error term in the regression for log earnings. Again, however,
for an estimation strategy that would be consistent if applied to a representative sample,
suitably weighted estimation would achieve consistency. For example, if the schooling
variable somehow were exogenous so that OLS estimation with a representative sample
would be consistent, then applying WLS to the endogenously selected sample (weighting
each contribution to the sum of squares by its inverse probability of selection) also would
be consistent. This could be achieved by applying least squares to an equation that looks
like equation (2), but now with J i standing for the inverse probability of selection.

estimated. That is a peculiar feature of the logit specification, and it does not carry over to other
specifications such as the probit model. Furthermore, without a consistent estimate of the intercept, one
cannot obtain consistent estimates of the average partial effects, which are commonly reported in empirical
studies.

14

Similarly, if one were to perform IV estimation, one would need to weight the IV
orthogonality conditions by the inverse probabilities of selection.
These examples illustrate a more general point, analyzed formally in Wooldridge
(1999) for the entire class of M-estimation. In the presence of endogenous sampling,
estimation that ignores the endogenous sampling generally will be inconsistent. But if
instead one weights the criterion function to be minimized (a sum of squares, a sum of
absolute deviations, the negative of a log likelihood, a distance function for orthogonality
conditions, etc.) by the inverse probabilities of selection, the estimation becomes
consistent.
An important point stressed in Wooldridge (1999) is that, if the sampling
probabilities vary exogenously instead of endogenously, weighting might be unnecessary
for consistency and harmful for precision. In the case of a linear regression model that
correctly specifies the conditional mean, the sampling would be exogenous if the
sampling probabilities are independent of the error term in the regression equation. This
would be the case, for example, if the sampling probabilities vary only on the basis of
explanatory variables. More generally, the issue is whether the sampling is independent
of the dependent variable conditional on the explanatory variables.
For example, suppose one estimates a linear regression model with a sample that
overrepresents certain states (as in the Current Population Survey), but the model
includes state dummy variables among the explanatory variables. Then, if the model is
correctly specified (more about that soon), the error term is not related to the sampling
criterion, and weighting is unnecessary. If the error term obeys the ideal conditions, then

15

OLS estimation is optimal in the usual way. 12 Is there any cost to using WLS instead of
OLS when weighting is unnecessary for consistency? Yes, there can be an efficiency
cost. If the error term was homoskedastic prior to weighting, the weighting will induce
heteroskedasticity, with the usual consequence of imprecise estimation. More generally,
when the error term may have been heteroskedastic to begin with, the efficiency
comparison between weighting or not weighting by inverse probabilities of selection
becomes less clear. Again, as in Section III, we recommend using standard diagnostics
for heteroskedasticity as a guide in the search for an efficient estimator.
Of course, we need again to acknowledge that, in practice, one’s model is almost
never perfectly specified. At best, it is a good approximation to the data-generating
process. As a result, just as theorems in microeconomic theory based on unrealistically
strong assumptions provide only rough guidance about what is going on in the actual
economy, theorems from theoretical econometrics provide inexact (though valuable, in
our view) guidance about how to do empirical research. In that light, let’s reconsider the
example in the previous paragraph. If the sampling probability varies only across states
and the regression model that controls for state dummies is a good, though imperfect,
approximation to the true model for the conditional mean, then one might reasonably
hope that OLS estimation would come close to unbiased and consistent estimation of the
effects of the explanatory variables. The same goes for WLS estimation (which also

12

Another practical example is where the survey organization provides sampling weights to adjust for
differential non-response, including attrition from a panel survey. If the weights are based only on
observable characteristics that are controlled for in the regression model (perhaps gender, race, age,
location), it is not clear that there is an advantage to using such weights when estimating that model. For
more on this topic, see Wooldridge (2002) and Fitzgerald, Gottschalk, and Moffitt (1998).

16

would fall short of perfect unbiasedness and consistency 13), but WLS might be less
precise.
In the end, what is our advice to practitioners? First, if the sampling rate varies
endogenously, estimation weighted by the inverse probabilities of selection is needed on
consistency grounds. Second, the weighted estimation should be accompanied by robust
estimation of standard errors. For example, in the case of a linear regression model, the
heteroskedasticity induced by the weighting calls for the use of White (1980)
heteroskedasticity-robust standard error estimates. 14 Finally, when the variation in the
sampling rate is exogenous, both weighted and unweighted estimation are consistent for
the parameters of a correctly specified model, but unweighted estimation may be more
precise.

Even then, as in the previous section, we recommend reporting both the

weighted and unweighted estimates because the contrast serves as a useful joint test
against model misspecification and/or misunderstanding of the sampling process.

V. Identifying Average Partial Effects
To consider a third motivation for weighted estimation of causal effects, let’s
return to the example of divorce-rate regressions. Recall that, when Lee and Solon
followed Friedberg and Wolfers in using the level, rather than the log, of the divorce rate
13

In footnote 15 in the next section, we will mention special cases of model misspecification where,
although OLS may be inconsistent for estimating particular causal effects, certain weighted estimators do
achieve consistency.
14
Wooldridge (1999) presents the appropriate “sandwich” estimator of the asymptotic variance-covariance
matrix for the general case of M-estimation under endogenous sampling. Wooldridge (2001) analyzes a
subtly different sort of sampling that he calls “standard stratified sampling.” In standard stratified
sampling, the survey selects a deterministically set number of observations per stratum. In this case, the
asymptotic variance-covariance matrix is more complex, and is strictly smaller than the one analyzed in
Wooldridge (1999). Intuitively, sampling variability is reduced by not leaving the within-stratum sample
sizes to chance. In the example of a linear regression model, the White heteroskedasticity-robust standard
error estimates then become conservative in the sense that they are slightly upward-inconsistent estimates
of the true standard errors.

17

as the dependent variable, they found that the OLS and WLS estimates differed
dramatically from each other, with the WLS results showing more evidence of a positive
impact of unilateral divorce on divorce rates. One possible explanation is that, if the
impact of unilateral divorce is heterogeneous – i.e., if it interacts with other state
characteristics – then OLS and WLS estimates that do not explicitly account for those
interactions may identify different averages of the heterogeneous effects. For example, if
unilateral divorce tends to have larger effects in more populous states, then WLS
estimation that places greater weight on more populous states will tend to estimate larger
effects than OLS does. Indeed, Lee and Solon found that, when they redid WLS with
California omitted from the sample, the estimated effects of unilateral divorce came out
smaller and more similar to the OLS estimates, which gave the same weight to California
as to any other state.
This raises the question of whether one might want to weight in order to identify a
particular average of heterogeneous effects, such as the population average partial effect.
Indeed, we have the impression that many empirical practitioners believe that, by
performing WLS with weights designed to reflect population shares, they do achieve
consistent estimation of population average partial effects (e.g., the average impact of
unilateral divorce on divorce rates for the U.S. population). This belief may be based on
the fact, discussed above in Section II.A, that this WLS approach does consistently
estimate the population linear projection of the dependent variable on the explanatory
variables. That, however, is not the same thing as identifying the population average

18

partial effects. 15 For a previous demonstration of this point, see Deaton (1997, pp. 6770).
Here, we illustrate with a simple cross-sectional example. Suppose the true model
for an individual-level outcome y i is
(5)

y i = β1 + β 2 X i + β 3 Di + β 4 X i Di + vi

where D is a dummy variable indicating urban (rather than rural) location and the error
term v is fully independent of all the explanatory variables. Then the effect of X on y is
heterogeneous, with β 2 as the rural effect and β 2 + β 4 as the urban effect. The average
effect for the population is the population-weighted average of these two effects, which is

β 2 + β 4π where π represents the urban share of the population.
Suppose that one fails to model the heterogeneity of effects and instead estimates
the regression of y on just X and D, with the interaction term omitted. And suppose that
one does so with data from a survey that oversampled in the urban sector, so that the
urban fraction of the sample is p > π . The OLS estimator of the coefficient of X does
identify a particular weighted average of the rural and urban effects, but no one would
expect that weighted average to be the same as the population average effect. After all,
the sample systematically overrepresents the urban sector. And, as we soon will show,
that is indeed one of the reasons that the probability limit of the OLS estimator differs
from the population average partial effect. But the math also will reveal a second reason.
In least squares estimation, observations with extreme values of the explanatory variables
15

One exception in which it is the same thing is in a simple regression on one dummy regressor, that is, a
simple contrast between the means for two subpopulations. And this exception extends to the case of a
“fully saturated” regression on a set of category dummies, which is a contrast among means for multiple
subpopulations. Another case in which using suitably weighted estimators to identify population linear
projections identifies a population average causal effect is the “doubly robust” estimator of treatment
effects introduced by Robins, Rotnitzky, and Zhao (1994) and analyzed by Wooldridge (2007).

19

have particularly large influence on the estimates. As a result, the weighted average of
the rural and urban effects identified by OLS depends not only on the sample shares of
the two sectors, but also on how the within-sector variance of X differs between the two
sectors.
Now suppose that instead one estimates the regression of y on X and D by WLS
with weighting by the inverse probabilities of selection. By reweighting the sample to
get the sectoral shares in line with the population shares, WLS eliminates the first reason
that OLS fails to identify the population average partial effect, but it does not eliminate
the second. As a result, the WLS estimator and the OLS estimator identify different
weighted averages of the heterogeneous effects, and neither one identifies the population
average effect.

ˆ
To be precise, let β 2,OLS denote the OLS estimator of the coefficient of X when
ˆ
the interaction term is omitted, and let β 2,WLS denote the corresponding WLS estimator.
It is straightforward to show that the probability limit of the latter is what one would get
from the corresponding population linear projection:
(6)



πσ 12
ˆ
plim β 2,WLS = β 2 + β 4  2
2 
 πσ 1 + (1 − π )σ 0 

2
where σ 0 and σ 12 respectively denote the within-sector variances of X for the rural and

urban sectors. In contrast, the probability limit of the OLS estimator is
(7)



pσ 12
ˆ
plim β 2,OLS = β 2 + β 4 
.
2
2 
 pσ 1 + (1 − p )σ 0 

If the effect of X were homogeneous (i.e., if β 4 = 0 ), then both estimators would be
consistent for the homogeneous effect β 2 . Which estimator is preferable would depend

20

on which is more precise, the question we already discussed in Section III’s analysis of
heteroskedasticity.
The point of the present section, however, is to consider the heterogeneous-effects
case where β 4 ≠ 0 . In that case, equations (6) and (7) imply that the inconsistencies of
the two estimators with respect to the true population average partial effect β 2 + β 4π are
(8)



πσ 12
ˆ
plim β 2,WLS − ( β 2 + β 4π ) = β 4  2
−π
2
 πσ 1 + (1 − π )σ 0


and
(9)



pσ 12
ˆ
plim β 2,OLS − ( β 2 + β 4π ) = β 4 
−π.
2
2
 pσ 1 + (1 − p )σ 0


2
In the knife-edge special case where σ 0 = σ 12 , WLS is consistent for the population

average effect and OLS is not. More generally, though, both estimators are inconsistent
for the population average effect (or any other average effect that researchers commonly
consider interesting).

With either over- or undersampling of the urban sector ( p ≠ π ),

WLS and OLS are inconsistent in different ways, and neither strictly dominates the other.
It is easy to concoct examples in which each is subject to smaller inconsistency than the
other.
Here are the lessons we draw from this example. First, we urge practitioners not
to fall prey to the fallacy that, in the presence of unmodeled heterogeneous effects,
weighting to reflect population shares generally identifies the population average partial
effect.

Second, we reiterate the usefulness of the contrast between weighted and

unweighted estimates.

We said before that the contrast can serve as a test for

misspecification, and the failure to model heterogeneous effects is one sort of

21

misspecification that can generate a significant contrast. Third, where heterogeneous
effects are salient, we urge researchers to study the heterogeneity, not just try to average
it out. Typically, the average partial effect is not the only quantity of interest, and
understanding the heterogeneity of effects is important.

For example, unless one

understands the heterogeneity, it is impossible to extrapolate from even a well-estimated
population average effect in one setting to what the average effect might be in a different
setting. In the simple example above, this recommendation just amounts to advising the
practitioner to include the interaction term instead of omitting it. We understand that, in
most empirical studies, studying the heterogeneity is more complex, but we still consider
it worthwhile.

VI. Summary and General Recommendations for Empirical Practice
In Section II, we distinguished between two types of empirical research: (1)
research directed at estimating population descriptive statistics and (2) research directed
at estimating causal effects. For the former, weighting is called for when it is needed to
make the analysis sample representative of the target population. For the latter, the
question of whether and how to weight is more nuanced.
In Sections III-V, we proceeded to discuss three distinct potential motives for
weighting when estimating causal effects: (1) to achieve more precise estimates by
correcting for heteroskedasticity, (2) to achieve consistent estimates by correcting for
endogenous sampling, and (3) to identify average partial effects in the presence of
unmodeled heterogeneity of effects. In our detailed discussion of each case, we have
noted instances in which weighting is not as good an idea as empirical researchers

22

sometimes think. Our overarching recommendation therefore is to take seriously the
question in our title: What are we weighting for? Be clear about the reason that you are
considering weighted estimation, think carefully about whether the reason really applies,
and double-check with appropriate diagnostics.
A couple of other recurring themes also bear repeating. In situations in which you
might be inclined to weight, it often is useful to report both weighted and unweighted
estimates and to discuss what the contrast implies for the interpretation of the results.
And, in many of the situations we have discussed, it is advisable to use robust standard
error estimates.

23

Table 1. Estimated Effects of Unilateral Divorce Laws
(1)
Dependent
variable:
Estimation
method:

(2)

Log of
Log of
divorce rate divorce rate
WLS

OLS

First 2 years

-0.022
(0.063)

-0.017
(0.026)

Years 3-4

-0.049
(0.063)

-0.014
(0.031)

Years 5-6

-0.051
(0.064)

-0.022
(0.034)

Years 7-8

-0.033
(0.065)

-0.013
(0.039)

Years 9-10

-0.052
(0.067)

-0.030
(0.046)

Years 11-12

-0.051
(0.074)

-0.015
(0.052)

Years 13-14

-0.043
(0.077)

-0.005
(0.060)

Years 15+

0.006
(0.084)

0.026
(0.073)

Notes: These results are drawn from Lee and Solon (2011, Table 2). The divorce rate is
the number of divorces per 1,000 persons by state and year. The standard error estimates
in parentheses are robust to heteroskedasticity and serial correlation. Both regressions
include controls for state fixed effects, year fixed effects, and state-specific linear time
trends.

24

References
Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An
Empiricist’s Companion. Princeton, NJ: Princeton University Press.
Autor, David H., Lawrence F. Katz, and Alan B. Krueger. 1998. “Computing Inequality:
Have Computers Changed the Labor Market?” Quarterly Journal of Economics,
113(4): 1169-213.
Borjas, George J. 2003. “The Labor Demand Curve Is Downward Sloping:
Reexamining the Impact of Immigration on the Labor Market.” Quarterly
Journal of Economics, 118(4): 1335-74.
Card, David, and Alan B. Krueger. 1992. “Does School Quality Matter? Returns to
Education and the Characteristics of Public Schools in the United States.”
Journal of Political Economy, 100(1): 1-40.
Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric
Approach to Development Policy. Baltimore: The Johns Hopkins University
Press.
Dehejia, Rajeev, and Adriana Lleras-Muney. 2004. “Booms, Busts, and Babies’
Health.” Quarterly Journal of Economics, 119(3): 1091-130.
Dickens, William T. 1990. “Error Components in Grouped Data: Is It Ever Worth
Weighting?” Review of Economics and Statistics, 72(2): 328-33.
Donald, Stephen G., and Kevin Lang. 2007. “Inference with Difference-in-Differences
and Other Panel Data.” Review of Economics and Statistics, 89(2): 221-33.
Donohue, John J., III, and Steven D. Levitt. 2001. “The Impact of Legalized Abortion
on Crime.” Quarterly Journal of Economics, 116(2): 379-420.

25

DuMouchel, William H., and Greg J. Duncan. 1983. “Using Sample Survey Weights in
Multiple Regression Analyses of Stratified Samples.” Journal of the American
Statistical Association, 78(383): 535-43.
Elder, Todd E., John H. Goddeeris, and Steven J. Haider. 2011. “A Deadly Disparity: A
Unified Assessment of the Black-White Infant Mortality Gap.” B.E. Journal of
Economic Analysis and Policy (Contributions), 11(1).
Fitzgerald, John, Peter Gottschalk, and Robert Moffitt. 1998. “An Analysis of Sample
Attrition in Panel Data: The Michigan Panel Study of Income Dynamics.”
Journal of Human Resources, 33(2): 251-99.
Friedberg, Leora. 1998. “Did Unilateral Divorce Raise Divorce Rates? Evidence from
Panel Data.” American Economic Review, 88(3): 608-27.
Imbens, Guido W.….
Lee, Jin Young, and Gary Solon. 2011. “The Fragility of Estimated Effects of Unilateral
Divorce Laws on Divorce Rates.” B.E. Journal of Economic Analysis and Policy
(Contributions), 11(1).
Levitt, Steven D. 1998. “Juvenile Crime and Punishment.” Journal of Political
Economy, 102(4): 1156-85.
Manski, Charles F., and Steven R. Lerman. 1977. “The Estimation of Choice
Probabilities from Choice Based Samples.” Econometrica, 45(8): 1977-88.
Robins, James M., Andrea Rotnitzky, and Lue Ping Zhao. 1994. “Estimation of
Regression Coefficients When Some Regressors Are Not Always Observed.”
Journal of the American Statistical Association, 89 (427): 846-66.
Shin, Donggyun, and Gary Solon. 2011. “Trends in Men’s Earnings Volatility: What

26

Does the Panel Study of Income Dynamics Show?” Journal of Public Economics
95(7-8): 973-82.
U.S. Bureau of the Census. 1968. Current Population Reports: Consumer Income, P-60
(55). Washington, DC: U.S. Government Printing Office.
White, Halbert. 1980. “A Heteroskedasticity-Consistent Covariance Matrix Estimator
and a Direct Test for Heteroskedasticity.” Econometrica, 48(4): 817-38.
Wolfers, Justin. 2006. “Did Unilateral Divorce Laws Raise Divorce Rates? A
Reconciliation and New Results.” American Economic Review, 96(5): 1802-20.
Wooldridge, Jeffrey M. 1999. “Asymptotic Properties of Weighted M-Estimators for
Variable Probability Samples.” Econometrica, 67(6): 1385-406.
Wooldridge, Jeffrey M. 2001. “Asymptotic Properties of Weighted M-Estimators for
Standard Stratified Samples.” Econometric Theory, 17(2): 451-70.
Wooldridge, Jeffrey M. 2002. “Inverse Probability Weighted M-Estimators for Sample
Selection, Attrition, and Stratification.” Portuguese Economic Journal, 1(2): 11739.
Wooldridge, Jeffrey M. 2003. “Cluster-Sample Methods in Applied Econometrics.”
American Economic Review, 93(2): 133-8.
Wooldridge, Jeffrey M. 2007. “Inverse Probability Weighted Estimation for General
Missing Data Problems.” Journal of Econometrics, 141 (2): 1281-1301.
Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data.
2nd ed. Cambridge, MA: MIT Press.
Wooldridge, Jeffrey M. 2013. Introductory Econometrics: A Modern Approach. 5th ed.
Mason, OH: South-Western.

27