Cluster and Stratified Sampling: 1. The Linear Model With Cluster Effects
Cluster and Stratified Sampling: 1. The Linear Model With Cluster Effects
, ,
(as G with M
g
fixed) and G -asymptotically normal.
Without more assumptions, a robust variance matrix is needed to account for correlation
within clusters or heteroskedasticity in Varv
gm
|x
g
, z
gm
, or both. When v
gm
has the form in
(1.2), the amount of within-cluster correlation can be substantial, which means the usual OLS
standard errors can be very misleading (and, in most cases, systematically too small). Write W
g
as the M
g
1 K L matrix of all regressors for group g. Then the
1 K L 1 K L variance matrix estimator is
Avarz
POLS
g1
G
W
g
W
g
1
g1
G
W
g
v
g
v
g
W
g
g1
G
W
g
W
g
1
(1.4)
where v
g
is the M
g
1 vector of pooled OLS residuals for group g. This asymptotic variance
is now computed routinely using cluster options.
Pooled OLS estimation of the parameters in (1.1) ignores the within-cluster correlation of
the v
gm
; even if the procedure is consistent (again, with G and the M
g
fixed), the POLS
estimators can be very inefficient. If we strengthen the exogeneity assumption to
Ev
gm
|x
g
, Z
g
0, m 1, . . . , M
g
; g 1, . . . , G, (1.5)
where Z
g
is the M
g
L matrix of unit-specific covariates, then we can exploit the presence of
c
g
in (1.2) in a generalized least squares (GLS) analysis. With true cluster samples, (1.5) rules
out the covariates from one member of the cluster affecting the outcomes on another, holding
own covariates fixed. In the panel data case, (1.5) is the strict exogeneity assumption on
z
gm
: m 1, . . . , M
g
that we discussed in the linear panel data notes The standard random
effects approach makes enough assumptions so that the M
g
M
g
variance-covariance matrix
of v
g
v
g1
, v
g2
, . . . , v
g,Mg
j
Mg
o
u
2
I
Mg
, (1.6)
where j
Mg
is the M
g
1 vector of ones and I
Mg
is the M
g
M
g
identity matrix. In the standard
setup, we also make the system homoskedasticity assumption,
Varv
g
|x
g
, Z
g
Varv
g
. (1.7)
It is important to understand the role of assumption (1,7): it implies that the conditional
variance-covariance matrix is the same as the unconditional variance-covariance matrix, but it
does not restrict Varv
g
; it can be any M
g
M
g
matrix under (1.7). The particular random
effects structure on Varv
g
is given by (1.6). Under (1.6) and (1.7), the resulting GLS
3
Imbens/Wooldridge, Lecture Notes 8, Summer 07
estimator is the well-known random effects (RE) estimator.
The random effects estimator z
RE
is asymptotically more efficient than pooled OLS under
(1.5), (1.6), and (1.7) as G with the M
g
fixed. The RE estimates and test statistics are
computed routinely by popular software packages. Nevertheless, an important point is often
overlooked in applications of RE: one can, and in many cases should, make inference
completely robust to an unknown form of Varv
g
|x
g
, Z
g
.
The idea in obtaining a fully robust variance matrix of RE is straightforward and we
essentially discussed it in the notes on nonlinear panel data models. Even if Varv
g
|x
g
, Z
g
does
not have the RE form, the RE estimator is still consistent and G -asymptotically normal under
(1.5), and it is likely to be more efficient than pooled OLS. Yet we should recognize that the
RE second moment assumptions can be violated without causing inconsistency in the RE
estimator. For panel data applications, making inference robust to serial correlation in the
idiosyncratic errors, especially with more than a few time periods, can be very important.
Further, within-group correlation in the idiosyncratic errors can arise for cluster samples, too,
especially if underlying (1.1) is a random coefficient model,
y
gm
o x
g
[ z
gm
,
g
v
gm
, m 1, . . . , M
g
; g 1, . . . , G. (1.8)
By estimating a standard random effects model that assumes common slopes ,, we effectively
include z
gm
,
g
, in the idiosyncratic error; this generally creates within-group correlation
because z
gm
,
g
, and z
gp
,
g
, will be correlated for m p, conditional on Z
g
. Also, the
idiosyncratic error will have heteroskedasticity that is a function of z
gm
. Nevertheless, if we
assume E,
g
|X
g
, Z
g
E,
g
, along with (1.5), the random effects estimator still
consistently estimates the average slopes, ,. Therefore, in applying random effects to panel
data or cluster samples, it is sensible (with large G) to make the variance estimator of random
effects robust to arbitrary heteroskedasticity and within-group correlation.
One way to see what the robust variance matrix looks like for z
RE
is to use the pooled OLS
characterization of RE on a transformed set of data. For each g, define
0
g
1 1/1 M
g
o
c
2
/o
u
2
1/2
, where o
c
2
and o
u
2
are estimators of the variances of c
g
and
u
gm
, respectively. Then the RE estimator is identical to the pooled OLS estimator of
y
gm
0
g
y g
on 1 0
g
, 1 0
g
x
g
, z
gm
0
g
zg
, m 1, . . . , M
g
; g 1, . . . , G; (1.9)
see, for example, Hsiao (2003). For fully robust inference, we can just apply the fully robust
variance matrix estimator in (1.4) but on the transformed data.
4
Imbens/Wooldridge, Lecture Notes 8, Summer 07
With panel data, it may make sense to estimate an unrestricted version of Varv
g
,
especially if G is large. Even in that case, it makes sense to obtain a variance matrix robust to
Varv
gm
|x
g
, Z
g
Varv
g
, as the GEE literature does. One can also specify a particular
structure, such as an AR(1) model for the idiosyncratic errors. In any case, fully robust
inference is still a good idea.
In summary, with large G and relatively small M
g
, it makes sense to compute fully robust
variance estimators even if we apply a GLS procedure that allows Varv
g
to be unrestricted.
Nothing ever guarantees Varv
gm
|x
g
, Z
g
Varv
g
. Because RE imposes a specific structure
on Varv
g
, there is a strong case for making RE inference fully robust. When c
g
is in the error
term, it is even more critical to use robust inference when using pooled OLS because the usual
standard errors ignore within-cluster correlation entirely.
If we are only interested in estimating ,, the fixed effects (FE) or within estimator is
attractive. The within transformation subtracts off group averages from the dependent variable
and explanatory variables:
y
gm
y g
z
gm
zg
, u
gm
g
, m 1, . . . , M
g
; g 1, . . . , G, (1.10)
and this equation is estimated by pooled OLS. (Of course, the x
g
get swept away by the
within-group demeaning.) Under a full set of fixed effects assumptions which, unlike
pooled OLS and random effects, allows arbitrary correlation between c
g
and the z
gm
inference is straightforward using standard software. Nevertheless, analogous to the random
effects case, it is often important to allow Varu
g
|Z
g
to have an arbitrary form, including
within-group correlation and heteroskedasticity. For panel data, the idiosyncratic errors can
always have serial correlation or heteroskedasticity, and it is easy to guard against these
problems in inference. Reasons for wanting a fully robust variance matrix estimator for FE
applied to cluster samples are similar to the RE case. For example, if we start with the model
(1.8) then z
gm
zg
,
g
, appears in the error term. As we discussed in the linear panel data
notes, the FE estimator is still consistent if E,
g
|z
g1
zg
, . . . , z
g,Mg
zg
E,
g
,, an
assumption that allows ,
g
to be correlated with zg
. Nevertheless, u
gm
, u
gp
will be correlated for
m p. A fully robust variance matrix estimator is
Avar,
FE
g1
G
Z
g
1
g1
G
Z
g
g1
G
Z
g
1
, (1.11)
where Z
g
is the matrix of within-group deviations from means and
g
is the M
g
1 vector of
5
Imbens/Wooldridge, Lecture Notes 8, Summer 07
fixed effects residuals. This estimator is justified with large-G asymptotics.
One benefit of a fixed effects approach, especially in the standard model with constant
slopes but c
g
in the composite error term, is that no adjustments are necessary if the c
g
are
correlated across groups. When the groups represent different geographical units, we might
expect correlation across groups close to each other. If we think such correlation is largely
captured through the unobserved effect c
g
, then its elimination via the within transformation
effectively solves the problem. If we use pooled OLS or a random effects approach, we would
have to deal with spatial correlation across g, in addition to within-group correlation, and this
is a difficult problem.
The previous discussion extends immediately to instrumental variables versions of all
estimators. With large G, one can afford to make pooled two stage least squares (2SLS),
random effects 2SLS, and fixed effects 2SLS robust to arbitrary within-cluster correlation and
heteroskedasticity. Also, more efficient estimation is possible by applying generalized method
of moments (GMM); again, GMM is justified with large G.
1.3. Should we Use the Large G Formulas with Large M
g
?
Until recently, the standard errors and test statistics obtained from pooled OLS, random
effects, and fixed effects were known to be valid only as G with each M
g
fixed. As a
practical matter, that means one should have lots of small groups. Consider again formula
(1.4), for pooled OLS, when the cluster effect, c
g
, is left in the error term. With a large number
of groups and small group sizes, we can get good estimates of the within-cluster correlations
technically, of the cluster correlations of the cross products of the regressors and errors even
if they are unrestricted, and that is why the robust variance matrix is consistent as G with
M
g
fixed. In fact, in this scenario, one loses nothing in terms of asymptotic local power (with
local alternatives shrinking to zero at the rate G
1/2
) if c
g
is not present. In other words, based
on first-order asymptotic analysis, there is no cost to being fully robust to any kind of
within-group correlation or heteroskedasticity. These arguments apply equally to panel data
sets with a large number of cross sections and relatively few time periods, whether or not the
idiosyncratic errors are serially correlated.
What if one applies robust inference in scenarios where the fixed M
g
, G asymptotic
analysis not realistic? Hansen (2007) has recently derived properties of the cluster-robust
variance matrix and related test statistics under various scenarios that help us more fully
understand the properties of cluster robust inference across different data configurations. First
6
Imbens/Wooldridge, Lecture Notes 8, Summer 07
consider how his results apply to true cluster samples. Hansen (2007, Theorem 2) shows that,
with G and M
g
both getting large, the usual inference based on (1.4) is valid with arbitrary
correlation among the errors, v
gm
, within each group. Because we usually think of v
gm
as
including the group effect c
g
, this means that, with large group sizes, we can obtain valid
inference using the cluster-robust variance matrix, provided that G is also large. So, for
example, if we have a sample of G 100 schools and roughly M
g
100 students per school,
and we use pooled OLS leaving the school effects in the error term, we should expect the
inference to have roughly the correct size. Probably we leave the school effects in the error
term because we are interested in a school-specific explanatory variable, perhaps indicating a
policy change.
Unfortunately, pooled OLS with cluster effects when G is small and group sizes are large
fall outside Hansens theoretical findings: the proper asymptotic analysis would be with G
fixed, M
g
, and persistent within-cluster correlation (because of the presence of c
g
in the
error). Hansen (2007, Theorem 4) is aimed at panel data where the time series dependence
satisfies strong mixing assumptions, that is, where the correlation within each group g is
weakly dependent. Even in this case, the variance matrix in (1.4) must be multiplied by
G/G 1 and inference based on the t
G1
distribution. (Conveniently, this adjustment is
standard in Statas calculation of cluster-robust variance matrices.) Interestingly, Hansen finds,
in simulations, that with G 10 and M
g
50 for all g, using the adjusted robust variance
matrix estimator with critical values from the t
G1
distribution produces fairly small size
distortions. But the simulation study is special (one covariate whose variance is as large as the
variance of the composite error).
We probably should not expect good properties of the cluster-robust inference with small
groups and very large group sizes when cluster effects are left in the error term. As an
example, suppose that G 10 hospitals have been sampled with several hundred patients per
hospital. If the explanatory variable of interest is exogenous and varies only at the hospital
level, it is tempting to use pooled OLS with cluster-robust inference. But we have no
theoretical justification for doing so, and reasons to expect it will not work well. In the next
section we discuss other approaches available with small G and large M
g
.
If the explanatory variables of interest vary within group, FE is attractive for a couple of
reasons. The first advantage is the usal one about allowing c
g
to be arbitrarily correlated with
the z
gm
. The second advantage is that, with large M
g
, we can treat the c
g
as parameters to
7
Imbens/Wooldridge, Lecture Notes 8, Summer 07
estimate because we can estimate them precisely and then assume that the observations are
independent across m (as well as g). This means that the usual inference is valid, perhaps with
adjustment for heteroskedasticity. Interestingly, the fixed G, large M
g
asymptotic results in
Theorem 4 of Hansen (2007) for cluster-robust inference apply in this case. But using
cluster-robust inference is likely to be very costly in this situation: the cluster-robust variance
matrix actually converges to a random variable, and t statistics based on the adjusted version of
(1.11) that is, multiplied by G/G 1 have an asymptotic t
G1
distribution. Therefore,
while the usual or heteroskedasticity-robust inference can be based on the standard normal
distribution, the cluster-robust inference is based on the t
G1
distribution (and the cluster-robust
standard errors may be larger than the usual standard errors). With small G, inference based on
cluster-robust statistics could be very conservative when it need not be. (Also, Hansens
Theorem 4 is not completely general, and may not apply with heterogeneous sampling across
groups.)
In summary, for true cluster sample applications, cluster-robust inference using pooled
OLS delivers statistics with proper size when G and M
g
are both moderately large, but they
should probably be avoided with large M
g
and small G. When cluster fixed effects are
included, the usual inference is often valid, perhaps made robust to heteroskedasticity, and is
likely to be much more powerful than cluster-robust inference.
For panel data applications, Hansens (2007) results, particularly Theorem 3, imply that
cluster-robust inference for the fixed effects estimator should work well when the cross section
(N) and time series (T) dimensions are similar and not too small. If full time effects are allowed
in addition to unit-specific fixed effects as they often should then the asymptotics must be
with N and T both getting large. In this case, any serial dependence in the idiosyncratic errors
is assumed to be weakly dependent. The similulations in Bertrand, Duflo, and Mullainathan
(2004) and Hansen (2007) verify that the fully robust cluster-robust variance matrix works
well.
There is some scope for applying the fully robust variance matrix estimator when N is
small relative to T when unit-specific fixed effects are included. Unlike in the true cluster
sampling case, it makes sense to treat the idiosyncratic errors as correlated with only weakly
dependent. But Hansens (2007, Theorem 4) does not allow time fixed effects (because the
asymptotics is with fixed N and T , and so the inclusion of time fixed effects means adding
more and more parameters without more cross section data to estimate them). As a practical
8
Imbens/Wooldridge, Lecture Notes 8, Summer 07
matter, it seems dangerous to rely on omitting time effects or unit effects with panel data.
Hansens result that applies in this case requires N and T both getting large.
2. Estimation with a Small Number of Groups and Large Group
Sizes
We can summarize the findings of the previous section as follows: fully robust inference
justified for large G (N) and small M
g
(T) can also be relied on when M
g
(T) is also large,
provided G N is also reasonably large. However, whether or not we leave cluster
(unobserved) effects in the error term, there are good reasons not to rely on cluster-robust
inference when G N) is small and M
g
(T) is large.
In this section, we describe approaches to inference when G is small and the M
g
are large.
These results apply primarily to the true cluster sample case, although we will draw on them
for difference-in-differences frameworks using pooled cross sections in a later set of notes.
In the large G, small M
g
case, it often makes sense to think of sampling a large number of
groups from a large population of clusters, where each cluster is relatively small. When G is
small while each M
g
is large, this thought experiment needs to be modified. For most data sets
with small G, a stratified sampling scheme makes more sense: we have defined G groups in the
population, and we obtain our data by randomly sampling from each group. As before, M
g
is
the sample size for group g. Except for the relative dimensions of G and M
g
, the resulting data
set is essentially indistinguishable from that described in Section 1.2.
The problem of proper inference when M
g
is large relative to G was brought to light by
Moulton (1990), and has been recently studied by Donald and Lang (2007). DL focus on
applications that seem well described by the stratified sampling scheme, but their approach
seems to imply a different sampling experiment. In particular, they treat the parameters
associated with the different groups as outcomes of random draws. One way to think about the
sampling in this case is that a small number of groups is drawn from a (large) population of
potential groups; therefore, the parameters common to all members of the group can be viewed
as random. Given the groups, samples are then obtained via random sampling within each
group.
To illustrate the issues considered by Donald and Lang, consider the simplest case, with a
single regressor that varies only by group:
y
gm
o [x
g
c
g
u
gm
o
g
[x
g
u
gm
, m 1, . . . , M
g
; g 1, . . . , G.
(2.1)
(2.2)
9
Imbens/Wooldridge, Lecture Notes 8, Summer 07
Notice how (2.2) is written as a model with common slope, [, but intercept, o
g
, that varies
across g. Donald and Lang focus on (2.1), where c
g
is assumed to be independent of x
g
with
zero mean. They use this formulation to highlight the problems of applying standard inference
to (2.1), leaving c
g
as part of the composite error term, v
gm
c
g
u
gm
. We know this is a bad
idea even in the large G, small M
g
case, as it ignores the persistent correlation in the errors
within each group. Further, from the discussion of Hansens (2007) results, using
cluster-robust inference when G is small is likely to produce poor inference.
One way to see the problem with the usual inference in applying standard inference is to
note that when M
g
M for all g 1, . . . , G, the pooled OLS estimator, [
, is identical to the
between estimator obtained from the regression
y g
on 1, x
g
, g 1, . . . , G. (2.3)
Conditional on the x
g
, [
y 3
y 4
/2 y 1
y 2
/2. (2.7)
(The pooled OLS regression using the disaggregated data results in the weighted average
p
3
y 3
p
4
y 4
p
1
y 1
p
2
y 2
, where p
1
M
1
/M
1
M
2
, p
2
M
2
/M
1
M
2
,
p
3
M
3
/M
3
M
4
, and p
4
M
4
/M
3
M
4
are the relative proportions within the control
and treatment groups, respectively.) With [
as in (2.7) and o y 1
y 2
/2. Using the MD
approach, we see there are two overidentifying restrictions, which are easily tested. But even if
we reject them, it simply implies at least one pair of means within each of the control and
treatment groups is different.
With large group sizes, and whether or not G is especially large, we can put the probably
generally into an MD framework, as done, for example, by Loeb and Bound (1996), who had
G 36 cohort-division groups and many observations per group. For each group g, write
y
gm
o
g
z
gm
,
g
u
gm
, m 1, . . . , M
g
, (2.8)
where we assume random sampling within group and independent sampling across groups.
We make the standard assumptions for OLS to be consistent (as M
g
) and
13
Imbens/Wooldridge, Lecture Notes 8, Summer 07
M
g
-asymptotically normal; see, for example, Wooldridge (2002, Chapter 4). The presence
of group-level variables x
g
in a structural model can be viewed as putting restrictions on the
intercepts, o
g
, in the separate group models in (2.8). In particular,
o
g
o x
g
[, g 1, . . . , G, (2.9)
where we now think of x
g
as fixed, observed attributes of heterogeneous groups. With K
attributes we must have G K 1 to determine o and [. If M
g
is large enough to estimate the
o
g
precisely, a simple two-step estimation strategy suggests itself. First, obtain the o
g
, along
with ,
g
, from an OLS regression within each group. If G K 1 then, typically, we can solve
for 0
o , [
:. 0
X
1
o
, where X is the
K 1 K 1 matrix with g
th
row 1, x
g
. If G K 1 then, in a second step, we can use a
minimum distance approach, as described in Wooldridge (2002, Section 14.6). If we use as the
weighting matrix I
G
, the G G identity matrix, then the minimum distance estimator can be
computed from the OLS regression
o
g
on 1, x
g
, g 1, . . . , G. (2.10)
Under asymptotics such that M
g
g
M where 0
g
1 and M , the minimum distance
estimator 0
g
is just y g
, the sample mean for group g. Let o
g
2
denote the usual sample variance for group g.
Because the y g
are independent across g, the efficient MD estimator uses a diagonal weighting
matrix. As a computational device, the minimum chi-square estimator can be computed by
using the weighted least squares (WLS) version of (2.10), where group g is weighted by M
g
/o
g
2
(groups that have more data and smaller variance receive greater weight). Conveniently, the
reported t statistics from the WLS regression are asymptotically standard normal as the group
sizes M
g
get large. (With fixed G, the WLS nature of the estimation is just a computational
device; the standard asymptotic analysis of the WLS estimator has G .). The minimum
distance approach works with small G provided G K 1 and each M
g
is large enough so that
normality is a good approximation to the distribution of the (properly scaled) sample average
within each group.
14
Imbens/Wooldridge, Lecture Notes 8, Summer 07
If z
gm
is present in the first-stage estimation, we use as the minimum chi-square weights the
inverses of the asymptotic variances for the g intercepts in the separate G regressions. With
large M
g
, we might make these fully robust to heteroskedasticity in Eu
gm
2
|z
gm
using the White
(1980) sandwich variance estimator. At a minimum we would want to allow different o
g
2
even
if we assume homoskedasticity within groups. Once we have the Avaro
g
which are just the
squared reported standard errors for the o
g
we use as weights 1/Avaro
g
in the
computationally simple WLS procedure. We are still using independence across g in obtaining
a diagonal weighting matrix in the MD estimation.
An important by-product of the WLS regression is a minimum chi-square statistic that can
be used to test the G K 1 overidentifying restrictions. The statistic is easily obtained as the
weighted sum of squared residuals, say SSR
w
. Under the null hypothesis in (2.9),
SSR
w
a
_
GK1
2
as the group sizes, M
g
, get large. If we reject H
0
at a reasonably small
significance level, the x
g
are not sufficient for characterizing the changing intercepts across
groups. If we fail to reject H
0
, we can have some confidence in our specification, and perform
inference using the standard normal distribution for t statistics for testing linear combinations
of the population averages.
We might also be interested in how one of the slopes in ,
g
depends on the group features,
x
g
. Then, we simple replace o
g
with, say ,
g1
, the slope on the first element of z
gm
. Naturally,
we would use 1/Avar,
g1
as the weights in the MD estimation.
The minimum distance approach can also be applied if we impose ,
g
, for all g, as in
the original model (1). Obtaining the o
g
themselves is easy: run the pooled regression
y
gm
on d1
g
, d2
g
, . . . , dG
g
, z
gm
, m 1, . . . , M
g
; g 1, . . . , G (2.11)
where d1
g
, d2
g
, . . . , dG
g
are group dummy variables. Using the o
g
from the pooled regression
(2.11) in MD estimation is complicated by the fact that the o
g
are no longer asymptotically
independent; in fact, o
g
y g
zg
,, where , is the vector of common slopes, and the presence
of , induces correlation among the intercept estimators. Let V
be the G G estimated
(asymptotic) variance matrix of the G 1 vector o
V
1
X
1
X
V
1
o
V
1
X
1
. If the OLS
regression (2.10) is used, or the WLS version, the resulting standard errors will be incorrect
because they ignore the across group correlation in the estimators. (With large group sizes the
errors might be small; see the next section.)
15
Imbens/Wooldridge, Lecture Notes 8, Summer 07
Intermediate approaches are available, too. Loeb and Bound (1996) (LB for short) allow
different group intercepts and group-specific slopes on education, but impose common slopes
on demographic and family background variable. The main group-level covariate is the
student-teacher ratio. Thus, LB are interested in seeing how the student-teach ratio affects the
relationship between test scores and education levels. LB use both the unweighted estimator
and the weighted estimator and find that the results differ in unimportant ways. Because they
impose common slopes on a set of regressors, the estimated slopes on education (say ,
g1
) are
not asymptotically independent, and perhaps using a nondiagonal estimated variance matrix V
g
o
g
O
p
Mg
1/2
, we can ingore the estimation error in o
g
for large M
g
(Recall that the
same large M
g
assumption underlies the minimum distance approach.) Then, it is as if we
are estimating the equation o
g
o x
g
[ c
g
, g 1, . . . , G by OLS. If the c
g
are drawn from a
normal distribution, classical analysis is applicable because c
g
is assumed to be independent of
x
g
. This approach is desirable when one cannot, or does not want to, find group-level
observables that completely determine the o
g
. It is predicated on the assumption that the other
factors in c
g
are not systematically related to x
g
, a reasonable assumption if, say, x
g
is a
randomly assigned treatment at the group level, a case considered by Angrist and Lavy (2002).
Beyond the treatment effect case, the issue of how to define parameters of interest appears
complicated, and deserves further research.
3. What if G and M
g
are Both Large?
Section 1 reviewed methods appropriate for a large number of groups and relatively small
group sizes. Section 2 considered two approaches appropriate for large group sizes and a small
number of groups. The DL and minimum distance approaches use the large group sizes
assumption differently: in its most applicable setting, DL use the large M
g
assumption to
ignore the first-stage estimation error entirely, with all uncertainty coming from unobserved
group effects, while the asymptotics underlying the MD approach is based on applying the
16
Imbens/Wooldridge, Lecture Notes 8, Summer 07
central limit theorem within each group. Not surprisingly, more flexibility is afforded if G and
M
g
are both large.
For example, suppose we adopt the DL specification (with an unobserved cluster effect),
o
g
o x
g
[ c
g
, g 1, . . . , G, (3.1)
and G 50 (say, states in the U.S.). Further, assume first that the group sizes are large enough
(or the cluster effects are so strong) that the first-stage estimation error can be ignored. Then,
it matters not whether we impose some common slopes or run separate regressions for each
group (state) in the first stage estimation. In the second step, we can treat the group-specific
intercepts, o
g
, g 1, . . . , G, as independent observations to be used in the second stage. This
means we apply regression (2.10) and apply the usual inference procedures. The difference
now is that with G 50, the usual t statistics have some robustness to nonnormality of the c
g
,
assuming the CLT approximation works well With small G, the exact inference was based on
normality of the c
g
.
Loeb and Bound (1996), with G 38, essentially use regression (2.10), but with estimated
slopes as the dependent variable in place of estimated intercepts. As mentioned in Section 2,
LB impose some common slopes across groups, which means the estimated parameters are
dependent across group. The minimum distance approach without cluster effects is one way to
account for the dependence. Alternatively, one can simply adopt the DL perspective and just
assume the estimation error is swamped by c
g
; then standard OLS analysis is approximately
justfied.
4. NONLINEAR MODELS
Many of the issues for nonlinear models are the same as for linear models. The biggest
difference is that, in many cases, standard approaches require distributional assumptions about
the unobserved group effects. In addition, it is more difficult in nonlinear models to allow for
group effects correlated with covariates, especially when group sizes differ. For the small G
case, we offer extensions of the Donald and Lang (2007) approach (with large group sizes) and
the minimum distance approach.
Rather than using a general, abstract setting, the issues for nonlinear models are easily
illustrated with the probit model. Wooldridge (2006) considers other models (which are also
covered in the nonlinear panel data notes).
4.1. Large Group Asymptotics
We can illustrate many issues using an unobserved effects probit model. Let y
gm
be a
17
Imbens/Wooldridge, Lecture Notes 8, Summer 07
binary response, with x
g
and z
gm
, m 1, . . . , M
g
, g 1, . . . , G defined as in Section 1. Assume
that
y
gm
1o x
g
[ z
gm
, c
g
u
gm
0
u
gm
|x
g
, Z
g
, c
g
~Normal0, 1
(4.1)
(4.2)
(where 1 is the indicator function). Equations (4.1) and (4.2) imply
Py
gm
1|x
g
, z
gm
, c
g
Py
gm
1|x
g
, Z
g
, c
g
o x
g
[ z
gm
, c
g
, (4.3)
where is the standard normal cumulative distribution function (cdf). We assume
throughout that only z
gm
affects the response probability of y
gm
conditional on x
g
and c
g
; the
outcomes of z
gp
for p m are assumed not to matter. This is captured in (4.3). For pooled
methods we could relax this restriction (as in the linear case), but, with the presence of c
g
, this
affords little generality in practice.
As in nonlinear panel data models, the presence of c
g
in (4.3) raises several important
issues, including how we estimate quantities of interest. As in the panel data case, we have
some interest in estimating average partial or marginal effects. For example, if the first element
of x
g
is continuous,
Py
gm
1|x
g
, z
gm
, c
g
x
g1
[
1
o x
g
[ z
gm
, c
g
, (4.4)
where is the standard normal density function. If
c
g
|x
g
, Z
g
~Normal0, o
c
2
, (4.5)
where the zero mean is without loss of generality because (4.1) contains an intercept, o, then
the APEs are obtained from the function
Py
gm
1|x
g
, Z
g
o x
g
[ z
gm
,/1 o
c
2
1/2
o
c
x
g
[
c
z
gm
,
c
, (4.6)
where o
c
o/1 o
c
2
1/2
, and so on. Conveniently, the scaled coefficients are exactly the
coefficients estimated by using a simple pooled probit procedure. So, for estimating the
average partial effects, pooled probit is perfectly acceptable. With large G and small group
sizes, we can easily make the standard errors and test statistics robust to arbitarary within
group correlation using standard sandwich variance estimators (robust to within-cluster
correlation).
Some authors prefer to call procedures such as pooled probit applied to cluster samples
pseudo maximum likelihood. Unfortunately, this term is used in contexts where only the
18
Imbens/Wooldridge, Lecture Notes 8, Summer 07
conditional mean is correctly specified in the context of the linear exponential family.
Wooldridge (2002, Chapter 13) calls such methods partial maximum likelihood to emphasize
that we have partially specified a distribution, namely the marginal distribution of y
gm
given
x
g
, Z
m
, without specifying a joint distribution y
g1
, . . . , y
g,Mg
conditional on x
g
, Z
g
.
If we supplement (4.1),.(4.2), and (4.5) with
u
g1
, . . . , u
g,Mg
are independent conditional on x
g
, Z
g
, c
g
(4.7)
then we have the so-called random effects probit model. Under the RE probit assumptions,
o, [, , and o
c
2
are all identified, and estimable by MLE, which means we can estimate the
APEs as well as the partial effects evaluated at the mean of c
g
, which is zero. We can also
compute partial effects at other values of c
g
that we might select from the normal distribution
with estimated standard deviation o
c
. The details for random effects probit in the balanced
panel data case are given in Wooldridge (2002, Chapter 15). The unbalanced case is similar.
As we discussed in the nonlinear panel data notes, minimum distance estimator or
generalized estimating equations can be used to obtain estimators (of the scaled coefficients)
more efficient than pooled probit but just as robust. (Remember, the RE probit estimator has no
known robustness properties to violation of assumption (4.7).)
A very challenging task, and one that appears not to have gotten much attention for true
cluster samples, is allowing correlation between the unobserved heterogeneity, c
g
, and the
covariates that vary within group, z
gm
. (For notational simplicity, we assume there are no
group-level controls in the model, but these can always be added.) For linear models, we know
that the within or fixed effects estimator allows arbitrary correlation, and does not restrict the
within-cluster dependence of u
g1
, . . . , u
g,Mg
. Unfortunately, allowing correlation between c
g
and z
g1
, z
g2
, . . . , z
gM
is much more difficult in nonlinear models. In the balanced case, where
the group sizes M
g
are the same for all g, the Chamberlain (1980) device can be used:
c
g
|Z
g
~Normalp zg
, o
a
2
, (4.8)
where o
a
2
is the conditional variance Varc
g
|Z
g
. If we use all random effects probit
assumptions but with (4.8) in place of (4.5), then we obtain a simple extension of the RE probit
model: simply add the group averages, zg
, as a set of additional explanatory variables. This is
identical to the balanced panel case we covered earlier. The marginal distributions are
Py
gm
1|Z
g
p z
gm
, zg
/1 o
a
2
1/2
p
a
z
gm
,
a
zg
a
(4.9)
where now the coefficients are scaled by a function of the conditional variance. This is just as
19
Imbens/Wooldridge, Lecture Notes 8, Summer 07
in the case of a balanced panel, and all calculations, including those for APEs, follow
immediately.
The Chamberlain-Mundlak needs to be modified for the unbalanced case. [One possibility
is to discard observations and balance the cluster sample under the assumption that the cluster
sizes are exogenous, and that might be desirable if there is not much variation in the cluster
sizes.] An alternative is to use the cluster setup and assuming a kind of exchangeability
assumption concerning the correlation between the cluster effect and the covariates. At a
minimum, (4.8) should be modified to allow the variances to depend on the cluster size, M
g
.
Under restrictive assumptions, such as joint normality of c
g
, z
g1
, . . . , z
g,Mg
, with the z
gm
independent and identically distributed within a cluster, one can derive Varc
g
|Z
g
. But these
are strong assumptions. We might just assume
c
g
|z
g1
, . . . , z
g,Mg
~Normalp zg
, o
a,Mg
2
, (4.10)
where o
a,Mg
2
denotes a different variance for each group size, M
g
. Then the marginal
distributions are
Py
gm
1|Z
g
p z
gm
, zg
/1 o
a,Mg
2
1/2
. (4.11)
Equation (4.11) can be estimated by pooled probit, allowing for different group variances. (A
normalization is also required.) A simpler approach is to estimate a different set of parameters,
p
Mg
,
Mg
, ,
Mg
, for each group size, and then to impose the restrictions in (4.11) by minimum
distance estimation. With very large G and little variation in M
g
, we might just use the
unrestricted estimates p
Mg
,
Mg
, ,
Mg
, estimate the APEs for each group size, and then average
these across group size. But more work needs to be done to see if such an approach loses too
much in terms of efficiency.
The methods of Altonji and Matzkin (2005) see also Wooldridge (2005) can be applied.
A completely nonparametric approach is based on
Py
gm
1|Z
g
, c
g
Py
gm
1|z
gm
, c
g
Fz
gm
, c
g
(4.12)
and
Dc
g
|z
g1
, z
g2
, . . . , z
g,Mg
Dc
g
|zg
. (4.13)
Define H
g
z
gm
, zg
Py
gm
1|z
gm
, zg
. As discussed in the nonlinear panel data notes, under
(4.12) and (4.13), it can be show that the APEs are obtained from
E
zg
H
g
z, zg
. (4.14)
20
Imbens/Wooldridge, Lecture Notes 8, Summer 07
If the group sizes differ, H
g
, generally depends on g. If there are relatively few group sizes,
it makes sense to estimate the H
g
, separately for each group size M
g
. Then, the APEs can
be estimated from
G
1
g1
G
g
z, zg
. (4.15)
As discussed before, as a practical matter we might just use flexible parametric models, such as
probit with flexible functional forms.
Other strategies are available for estimating APEs. We can apply fixed effects probit to
cluster samples just as with panel data and treat the c
g
as parameters to estimate in
Py
gm
1|Z
g
, c
g
Py
gm
1|z
gm
, c
g
z
gm
, c
g
. (4.16)
The same issues arise as in the panel data case, except with true cluster samples the conditional
indepdence assumption likely is more reasonable than in the panel data case. With small group
sizes M
g
(say, siblings or short panel data sets), treating the c
g
as parameters to estimate
creates an incidental parameters problem. As before, we might use
G
1
g1
G
z,
g
, (4.17)
to estimate the APEs.
The logit conditional MLE can be applied to cluster samples essentially without change,
which means we can estimate the parameters, ,, without restricting Dc
g
|Z
g
. This is especially
convenient in the unbalanced case.
4.2. A Small Number of Groups and Large Group Sizes
Unlike in the linear case, for nonlinear models exact inference is unavailable even under
the strongest set of assumptions. Nevertheless, if the group sizes M
g
are reasonably large, we
can extend the DL approach to nonlinear models and obtain approximate inference. In
addition, the the minimum distance approach carries over essentially without change.
We can apply the methods to any nonlinear model that has an index structure which
includes all of the common ones, and many other models besides, but we again consider the
probit case. With small G and random sampling of y
gm
, z
gm
: m 1, . . . , M
g
within each g,
write
Py
gm
1|z
gm
o
g
z
gm
,
g
, m 1, . . . , M
g
(4.18)
21
Imbens/Wooldridge, Lecture Notes 8, Summer 07
o
g
o x
g
[, g 1, . . . , G. (4.19)
As with the linear model, we assume the intercept, o
g
in (4.18), is a function of the group
features x
g
. With the M
g
moderately large, we can get good estimates of the o
g
. The
o
g
, g 1, . . . , G, are easily obtained by estimating a separate probit for each group. Or, we can
impose common ,
g
and just estimate different group intercepts (sometimes called group fixed
effects).
Under (4.19), we can apply the minimum distance approach just as before. Let Avaro
g
(so these shrink to zero at the rate 1/M
g
). If
the o
g
are obtained from G separate probits, they are independent, and the Avaro
g
are all we
need. As in the linear case, if a pooled method is used, the G G matrix Avaro
should be
obtained as the weighting matrix. For binary response, we use the usual MLE estimated
variance. If we are using fractional probit for a fractional response, these would be from a
sandwich estimate of the asymptotic variance. In the case where the o
g
are obtained from
separate probits, we can obtain the minimum distance estimates as the WLS estimates from
o
g
on 1, x
g
, g 1, . . . , G (4.20)
using weights 1/Avaro
g
are used as the weights. This is the efficient minimum distance
estimator and, conveniently, the proper asymptotic standard errors are reported from the WLS
estimation (even though we are doing large M
g
, not large G, asymptotics.) Generally, we can
write the MD estimator as in the linear case, 0
V
1
X
1
X
V
1
o
, where o
is the G 1
vector of o
g
and V
Avaro
g
o x
g
[ error
g
, g 1, . . . , G (4.21)
as approximately satisfying the classical linear model assumptions, provided G K 1, just as
before. As in the linear case, this approach is justified if o
g
o x
g
[ c
g
with c
g
independent of x
g
and c
g
drawn from a homoskedastic normal distribution. It assumes that we
can ignore the estimation error in o
g
, based on o
g
o
g
O1/ M
g
. Because the DL
approach ignores the estimation error in o
g
, it is unchanged if one imposes some constant
slopes across the groups, as with the linear model.
Once we have estimated o and [, the estimated effect on the response probability can be
obtained by averaging the response probability for a given x:
G
1
g1
G
M
g
1
m1
Mg
o x[
z
gm
,
g
, (4.22)
where derivatives or differences with respect to the elements of x can be computed. Here, the
minimum distance approach has an important advantage over the DL approach: the finite
sample properties of (4.22) are viritually impossible to obtain, whereas the large-M
g
asymptotics underlying minimum distance would be straightforward using the delta method.
How the bootstrap might work in this situation is an interesting question.
Particularly with binary response problems, the two-step methods described here are
problematical when the response does not vary within group. For example, suppose that x
g
is a
binary treatment equal to one for receiving a voucher to attend college and y
gm
is an
indicator of attending college. Each group is a high school class, say. If some high schools
have all students attend college, one cannot use probit (or logit) of y
gm
on z
gm
, m 1, . . . , M
g
.
A linear regression returns zero slope coefficients and intercept equal to unity. Of course, if
randomization occurs at the group level that is, x
g
is independent of group attributes then it
is not necessary to control for the z
gm
. Instead, the within-group averages can be used in a
simple minimum distance approach. In this case, as y
gm
is binary, the DL approximation will
not be valid, as the CLM assumptions will not even approximately hold in the model
y g
o x
g
[ e
g
(because y g
is always a fraction regardless of the size of M
g
).
4.3. Large G and Large M
g
As in the linear case, more flexibility is afforded if G is somewhat large along with large
M
g
. If we can ignore the estimation error in the o
g
, then, in implementing the DL approach
23
Imbens/Wooldridge, Lecture Notes 8, Summer 07
with or without common slopes imposed in the first stage one gains robustness to
nonnormality of c
g
if G is large enough so that G
1/2
g1
G
c
g
and G
1/2
g1
G
x
g
c
g
are
approximately normally distributed. The second step is the same as in the linear model case: o
g
is regressed on 1, x
g
, g 1, . . . , G; one can use heteroskedasticity-robust inference with large
G to partly account for the estimation error in the o
g
.
A version of the method proposed by Berry, Levinsohn, and Pakes (1995) for estimating
structural models using both individual-level and product-level data, or market-level data, or
both can be treated in the large G, large M
g
framework, where g indexes good or market and m
indexes individuals within a market. BLPs original application was where g indexes different
automobile models. Petrin and Train (2002) cover the case of about 170 television markets and
four TV services. To handle this case, assume that H products are available in each market.
Therefore, we now think of o
g
as an H-vector for each g, and so is c
g
. The main difference
here with the previous setup is that, for reasons discussed in BLP and Petrin and Train, we
must allow the c
gh
to be correlated with the x
gh
(which contains the price of good j in market g,
in addition to product/market attributes). BLP propose a two-step estimation strategy. In the
first step, a choice model, such as multinomial logit, is estimated using the individual-level
data pooled across markets. The key estimates are what we call the o
g
the market fixed
effects. Typically, most or all of the slope parameters in the multinomial logit estimation
are assumed to be constant across g, although, with many individuals per market, that is not
necessary.
In the second step, the o
gh
are used in place of o
gh
in the market/good-level equation
o
gh
o x
gh
[ c
gh
, h 1, . . . , H; g 1, . . . , G, (4.23)
where, say, w
g
is a set of instruments for x
gh
. (Typically, w
g
varies only by market, g, and not
by good, h.) This allows for market/good-specific unobservables in the individual choice
equations to be correlated with prices. If we could observe the o
gh
, then (4.23) would be a
standard problem in IV estimation for a cross section system of equations, provided G is large
enough to invoke the law of large numbers and central limit theorem. Replacing o
g
with o
g
is
justified if the M
g
are large because the variance of c
g
will dominate that of the o
g
. Further,
any correlation induced in the o
g
by pooling in the first-stage estimation shrinks to zero at the
rate 1/M, where we can think of M as the average group size. In other words, we just apply,
say, 2SLS in the second step.
24
Imbens/Wooldridge, Lecture Notes 8, Summer 07
Ignoring the estimation in o
g
, efficient estimation is obtained by writing the system of
equations as
o
g
X
g
0 c
g
(4.24)
where X
g
is the J K 1 matrix of attributes (including an intercept and prices). Because
(4.24) is a system of equations with instruments I
J
w
g
, we can use the 3SLS estimator or
GMM to efficiently account for the correlation across c
gh
: h 1, . . . , H.
5. Estimation of Population Parameters with Stratified Samples
We now provide a brief, modern treatment of estimation with stratified samples. The
emphasis here is in estimation parameters from a population that has been stratified. Typically,
with stratified sampling, some segments of the population are over- or underrepresented by the
sampling scheme. Fortunately, if we know enough information about the stratification scheme,
we can often modify standard econometric methods and consistently estimate population
parameters.
There are two common types of stratified sampling, standard stratified (SS) sampling and
variable probability (VP) sampling. A third type of sampling, typically called multinomial
sampling, is practically indistinguishable from SS sampling, but it generates a random sample
from a modified population (thereby simplifying certain theoretical analyses). See Cosslett
(1993), Imbens (1992), Imbens and Lancaster (1996), and Wooldridge (1999) for further
discussion. We focus on SS and VP sampling here.
SS sampling begins by partitioning the sample space (set of possible outcomes), say W,
into G non-overlapping, exhaustive groups, W
g
: g 1, . . . G. Then, a random sample is
taken from each group g, say w
gi
: i 1, . . . , N
g
, where N
g
is the number of observations
drawn from stratum g and N N
1
N
2
. . . N
G
is the total number of observations. If w is a
random vector representing the population, and taking values in W, then each random draw
from stratum g has the same distribution as w conditional on w belonging to W
g
:
Dw
gi
Dw|w W
g
, i 1, . . . , N
g
.
Therefore, the resulting sample across all strata consists of independent but not identically
distributed observations. Unless we are told, we have no way of knowing that our data came
from SS sampling.
What if we want to estimate the mean of w from an SS sample? It turns out we cannot get
an unbiased or consistent estimator of unless we have some additional information. Typically,
25
Imbens/Wooldridge, Lecture Notes 8, Summer 07
the information comes in the form of population frequencies for each of the strata. Specifically,
let m
g
Pw W
g
be the probability that w falls into stratum g; the m
g
are often called the
aggregate shares.
If we know the m
g
(or can consistently estimate them), then j
w
Ew is identified by a
weighted average of the expected values for the strata:
j
w
m
1
Ew|w W
1
. . . m
G
Ew|w W
G
. (5.1)
Because we can estimate each of the conditional means using the random sample from the
appropriate stratum, an unbiased estimator of is simply
j
w
m
1
w 1
m
2
w 2
. . . m
G
w G
, (5.2)
where w g
is the sample average from stratum g. As the strata sample sizes grow, j
w
is also a
consistent estimator of j
w
. The variance of j
w
is easily obtained because of independence
withing and between strata:
Varj
w
m
1
2
Varw 1
. . . m
G
2
Varw G
. (5.3)
Because Varw g
o
g
2
/N
f
, each of the variances can be estimated in an unbiased fashion by
using the usual unbiased variance estimator,
o
g
2
N
g
1
1
i1
Ng
w
gi
w g
2
. (5.4)
Sometimes it is useful to have a formula for j
w
that shows it is a weighted average across
all observations:
j
w
m
1
/h
1
N
1
i1
N
1
w
1i
. . . m
G
/h
G
N
1
i1
N
G
w
Gi
N
1
i1
N
m
g
i
/h
g
i
w
i
(5.5)
where h
g
N
g
/N is the fraction of observations in stratum g and in (5.5) we drop the strata
index on the observations and use the stratum for observation i, g
i
, to pick out the appropriate
weight, m
g
i
/h
g
i
. Formula (1.5) is useful because the sampling weights associated with SS
samples are reported as m
g
i
/h
g
i
, and so applying these weights in averaging across all N
produces an unbiased, consistent estimator. Nevertheless, the large sample properties of
estimators from SS samples are properly derived from (5.2) and its extensions.
A different sampling scheme is usually called variable probability (VP) sampling, which is
26
Imbens/Wooldridge, Lecture Notes 8, Summer 07
more convenient for telephone or email surveys, where little, if anything, is known ahead of
time about those being contacted. With VP sampling, each stratum g is assigned a nonzero
sampling probability, p
g
. Now, a random draw w
i
is taking from the population, and it is kept
with probability p
g
if w
i
W
g
. With VP sampling, the population is sampled N times. Often N
is not reported with VP samples (although, as we discuss latter, knowing how many times each
stratum was sampled can improve efficiency). Instead, we know how many data points were
kept, and we call this M. Because of the randomness in whether an observation is kept, M is
properly viewed as a random variable. With VP sampling, it is handy for each draw from the
population to have a selection indicator, s
i
, which is one of observation i is kept (and then its
stratum is also known). Then M
i1
N
s
i
. Let z
i
be a G-vector of stratum indicators, and let
pz
i
p
1
z
i1
. . . p
G
z
iG
be the function that delivers the sampling probability for any random
draw i.
A key assumption for VP sampling is that conditional on being in stratum g, the chance of
keeping an observation is p
g
. Statistically, this means that, conditional on z
i
, s
i
and w
i
are
independent. Using this assumption, we can show, just as in the treatment effect case,
Es
i
/pz
i
w
i
Ew
i
; (5.6)
that is, weighting a selected observation by the inverse of its sampling probability allows us to
recover the population mean. (We will use a more general version of this result when we
discuss missing data general. Like estimating counterfactual means in program evaluation, VP
sampling is, in effect, a missing data problem. But it is usually treated along with other
stratified sampling schemes.) It follows that
N
1
i1
N
s
i
/pz
i
w
i
(5.7)
is a consistent estimator of Ew
i
. We can also write this as
M/NM
1
i1
N
s
i
/pz
i
w
i
; (5.8)
if we define weights as v
i
/pz
i
where M/N is the fraction of observations retained
from the sampling scheme, then (5.8) is M
1
i1
M
v
i
w
i
, where only the observed points are
included in the sum. Thus, like in the SS case, we can write the esimator for the mean under
VP sampling as a weighted average of the observed data points. In the VP case, the weight is
(an estimate of) the probability of keeping an observation, Ps
i
1, over the probability
27
Imbens/Wooldridge, Lecture Notes 8, Summer 07
that an observation in its stratum is kept. If p
g
, the observations for stratum g are
underpresented in the eventual sample (asymptotically), and they receive weight greater than
one.
In both the SS and VP cases, one may replace the number of observed data points in the
average with the sum of the weights described in each case.
Virtually any estimation method can be used with SS or VP sampled data. Wooldridge
(1999, 2001) covers M-estimation for the VP and SS cases, respectively. This includes a wide
variety of estimators, including least squares, MLE, and quasi-MLE. There are several
interesting findings concerning asymptotic efficiency and estimating the asymptotic variances.
Consider the problem of linear regression for simplicity; analogous claims hold for MLE, NLS,
and many other estimators. The model in the population is
y x u, (5.9)
where may index the conditional mean, but consistency follows from Ex
u 0. Consider
SS sampling. Then a consistent estimator
. There are two assumptions that imply consistency of this widely used
variance matrix estimator: (i) Ey|x x, so that we are actually estimating a conditional
mean; and (ii) the strata are determined by the explanatory variables, x. It turns out that when
the White estimator is not consistent, it is actually conservative. In other words, the White
estimator converges to a matrix that is larger, in the matrix sense, than the correct asymptotic
variance.
To obtain the correct asymptotic variance, we need to use a more detailed formulation of
the estimation problem, which is
28
Imbens/Wooldridge, Lecture Notes 8, Summer 07
min
b
g1
G
m
g
N
g
1
i1
N
y
gi
x
gi
b
2
(5.11)
so that we are minimizing the a weighted average sum of squared residuals. Using this
formulation actually, the M-estimator version of it Wooldridge (2001) showed that a
consistent estimator of the asymptotic variance of
is
Avar
i1
N
m
g
i
/h
g
i
x
i
x
i
1
g1
G
m
g
/h
g
i1
Ng
x
gi
gi
x
g
g
x
gi
gi
x
g
i1
N
m
g
i
/h
g
i
x
i
x
i
1
.
(5.12)
This formula looks a bit daunting, but, it can be seen that the outer parts of the sandwich are
identical to the usual White sandwich estimator. The difference is in the middle. The usual
estimator ignores the information on the strata of the observations, which is the same as
dropping the within-strata averages, x
g
g
. Because a smaller sum of squared residuals (in a
matrix sense) is obtained by subtracting off the same average rather than centering around
zero the matrix in (5.12) is smaller than the usual White matrix. That happens
asymptotically, too, provided the means Ex
u|w W
g
, where w x, y, are nonzero. So, it
is the difference between subtracting off within-strata averages and not that produces the more
precise inference with stratified sampled data. Econometrics packages, such as Stata, will
compute (5.12) provided strata membership is included along with the weights. If only the
weights are provided, the larger asymptotic variance is computed.
One case where there is no gain from subtracting within-strata means is when Eu|x 0
and w W
g
is the same as x X
g
for some partition of the regressor space. In fact, if we add
the homoskedasticity assumption Varu|x o
2
, then the weighted estmator is less efficient
than the unweighted estimator, which, of course, is also consistent because Ey|x x and
stratification is based on x. So, the cost to weighting when the classical linear model
assumptions hold and stratification is exogenous is in terms of efficiency loss.
Some argue that even if stratification is based on x, one should use the weighted estimator.
The argument is based on consistently estimating the linear projection, Ly|x, even if the
29
Imbens/Wooldridge, Lecture Notes 8, Summer 07
conditional mean is not linear. If we can only assume Ly|x x, then the weighted estimator
consistently estiimates whether or not the stratification is based on x. The unweighted
estimator does not consistently estimate in either case.
The previous discussion applies to nonlinear least squares and maximum likelihood
problems, and others. Now, to exploit the stratification, strata means should be subtracted from
the gradient of the objective function when computing the asymptotic variance. This requires
knowing the stratum and its weight for each observation. A conservative estimate is obtained
by the Huber-White sandwich form for misspecified MLE but with sampling weights. This is
the proper formula for, say, MLE if the conditional density fy|x, is correctly specified and
stratification is based on x. But in that case the unweighted MLE is fully efficient, and the
usual variance matrix estimators can be used. The weighted estimator does consistently
estimate the solution to the population problem min
t
Elogfy|x, t if the density is
misspecified, and that is valuable in some situations.
The above findings have analogs for VP sampling. One interesting finding is that while the
Huber-White sandwich matrix applied to the weighted objective function (weighted by the
1/p
g
) is consistent when the known p
g
are used, an asymptotically more efficient estimator is
available when the retention frequencies, p
g
M
g
/N
g
, are observed, where M
g
is the number
of observed data points in stratum g and N
g
is the number of times stratum g was sampled. We
always know M
g
if we are given a stratum indicator with each observation. Generally, N
g
might not be known. If it is, we should use the p
g
in place of p
g
. Results such as this are
discussed in Imbens (1992), Imbens and Lancaster (1996), and Wooldridge (1999, 2007). The
VP sampling example in Wooldridge (2007) can be used to show that the following matrix is
valid:
Avar
i1
M
x
i
x
i
/p
g
i
1
g1
G
p
g
2
i1
Mg
x
gi
gi
x
g
g
x
gi
gi
x
g
i1
M
x
i
x
i
/p
g
i
1
,
(5.13)
where, remember, M
g
is the number of observed data points in stratum g, and the above sums
are over the observed data points. This formula is essentially the same as (5.12) in that the
30
Imbens/Wooldridge, Lecture Notes 8, Summer 07
quantities are weighted so that the sample represents the population and x
gi
gi
are centered
about the within-strata means. If we use the known sampling weights, we drop x
g
g
from
(5.13). If Eu|x 0 and the sampling is exogenous, we also drop this term because
Ex
u|w W
g
0 for all g, and this is whether or not we estimate the p
g
. See Wooldridge
(2007) for how these same claims carry over to general nonlinear models and estimation
methods.
References
(To be added.)
31