EC327: Advanced Econometrics, Spring 2007

Wooldridge, Introductory Econometrics (3rd ed, 2006)

Chapter 14:
Advanced panel data methods

Fixed effects estimators

We discussed the first difference (FD) model

as one solution to the problem of unobserved
heterogeneity in the context of panel data. It is
not the only solution; the leading alternative is
the fixed effects model, which will be a better
solution under certain assumptions.

For a model with a single explanatory variable,

yit = β1xit + ai + uit (1)

If we average this equation over time for each
unit i, we have

ȳit = β1x̄it + āi + ūit (2)

Subtracting the second equation from the first,
we arrive at

yit − ȳit = β1(xit − x̄it) + (uit − ūit) (3)

defining the demeaned data on [y, x] as the ob-
servations of each panel with their mean values
per individual removed. This algebra is known
as the within transformation, and the estima-
tor we derive is known as the within estimator.
Just as OLS in a cross-sectional context only
“explains” the deviations of y from its mean
ȳ, the within estimator’s explanatory value is
derived from the comovements of y around
its individual-specific mean with x around its
individual-specific mean. Thus, it matters not
if a unit has consistently high or low values of
y and x. All that matters is how the variations
around those mean values are correlated.
Just as in the case of the FD estimator, the
within estimator will be unbiased and consis-
tent if the explanatory variables are strictly ex-
ogenous: independent of the distribution of u
in terms of their past, present and future val-
ues. Correlation with ai is allowable since that
term will be removed in the within transforma-
tion.

The within estimator is implemented by Stata’s

command xtreg, fe (fe for fixed effects). The
within transformation implements what has of-
ten been called the LSDV (least squares dummy
variable) model because the regression on de-
meaned data yields the same results as esti-
mating the model from the original data and a
set of (N − 1) indicator variables for all but one
of the panel units. It is often not workable to
estimate that LSDV model directly because we
may have hundreds or thousands of individual
panel units in our dataset. We can always per-
form the within transformation for any number
of units, though, and implement the FE model.
Note that the degrees of freedom for such a
model will take account of the N means that
were estimated, one for each individual. Thus,
unlike pooled OLS where the number of de-
grees of freedom would be (N T − k), the de-
grees of freedom for the FE estimator will be
(N (T − 1) − k). In Stata’s implementation in
xtreg, fe, a constant term is included and a F -
test is provided for the null hypothesis that all
0 0
coefficients ai are zero, where ai are deviations
from the mean value āi.

Analogous to the FD model, we cannot include

time-invariant variables in the FE model, since
the demeaning process will cause their value
to be zero for all time periods. We can inter-
act such variables with time-varying variables,
though. We could interact a gender indicator
with time dummies, which would allow us to
estimate how the effect of gender has changed
over the time periods. We cannot estimate the
effect of gender in the base period, though,
since that is subsumed in the ai term.

If we introduce a full set of (T − 1) time dum-

mies (one for each period but the first), we
cannot include any explanatory variables that
have a constant difference over time for each
individual: e.g., age in an annual dataset. The
same would be true if we introduced a linear
time trend rather than time dummies: it ab-
sorbs all time-constant effects.

Fixed effects or first differencing?

Two competing methods: first differencing and

fixed effects. Which should we use? If T=2,
it does not matter, since FD and FE meth-
ods are identical in that case. When T ≥
3, the two methods do not yield the same
results, but they are both unbiased estima-
tors of the underlying coefficient vector. Both
are consistent with T fixed as N → ∞. For
large N and small T (a common setup in many
datasets) we might be concerned with relative
efficiency. When the uit are serially uncorre-
lated (given that they are homoskedastic, this
amounts to saying they are i.i.d.) FE will be
more efficient than FD, and the standard er-
rors reported from FE are valid. We often may
assume serially uncorrelated errors, but there
is no reason why that condition will necessarily
hold in the data. If uit follows a random walk
process, then its differences will be uncorre-
lated, and first differencing will be the appro-
priate estimator. But we may often encounter
an error process with some serial correlation,
but not necessarily a random walk process.

When T is large and N is not very large (for

instance, when we have many time periods
of data on each of a small number of units)
we must be careful in using the FE estima-
tor, since its large-sample justification relies on
N → ∞, not T. If FE and FD give substan-
tively different results, it might be very hard
to choose between them, and we might want
to report them both.

One consideration arises when we are using an

unbalanced panel—especially one in which the
missing observations on some units do not ap-
pear at the beginning or end of their time se-
ries, but create “gaps” in the time series. The
FE estimator has no problem with this, but the
FD estimator will lose two observations when
there is a single period missing in the sequence
of observations for that unit. One thing we
must consider is why the data are missing. If
they can be considered “missing at random”,
this may not be problematic, but if there is
some pattern to missingness we must be con-
cerned about it.
One issue that often arises with individuals or
firms is attrition: units leaving the sample.
Individuals can die; firms can liquidate or be
taken over. Are these events related to the
variables we are using in the regression model?
If so, we may want to worry about the sample
selection problem that this entails. Neverthe-
less, one advantage of fixed effects is that it
allows the attrition to be correlated with ai,
the unobserved fixed effect.

Two-way fixed effects

Stata lacks a command to estimate two-way

fixed effects models. If the number of time
periods is reasonably small, you may estimate
a two-way FE model by creating a set of time
indicator variables and including all but one in
the regression.

The joint test that all of the coefficients on

those indicator variables are zero will be a test
of the significance of time fixed effects. Just
as the individual fixed effects (LSDV) model
requires regressors’ variation over time within
each unit, a time fixed effect (implemented
with a time indicator variable) requires regres-
sors’ variation over units within each time pe-
riod. If we are estimating an equation from
individual or firm microdata, this implies that
we cannot include a “macro factor” such as
the rate of GDP growth or price inflation in
a model with time fixed effects, since those
factors do not vary across individuals.

Random effects models

As an alternative to the individual fixed effects

model, we may consider a random effects for-
mulation. For a single explanatory variable,
this becomes

yit = β0 + β1xit + [ai + uit] (4)

where we explicitly include an intercept so that
we can make the assumption that the unob-
served effect, ai, has a zero mean. The brack-
eted term is the composite error term, which
is now assumed to have an individual-specific
component and an idiosyncratic component.

In the fixed effects formulation, ai is treated

as an unknown “nuisance parameter” which, if
ignored, causes bias and inconsistency in our
estimators because it is correlated with one or
more of the regressors. In the fixed effects
within transformation we get rid of ai. But if
we can assume that ai is a random variable dis-
tributed independently of x (or generally inde-
pendent of all of the regressors), we can derive
a more efficient estimator of the problem than
fixed effects.

The random effects model then proceeds by

using the form of the composite error term
in an optimal manner. For instance, with the
assumption of independence, we could use a
single cross section to optimally estimate the
β vector; there would be no need for panel
data. That would discard information from
other cross sections, so we might rather want
to use pooled OLS, which will be consistent
in this case. But pooled OLS will not make
optimal use of the assumed structure in the
error term. The composite error term assumes
that the errors arise for two reasons: one of
them common to all observations on a single
individual, the other purely idiosyncratic.

If we define the composite error term vit =

ai + uit as the regression error, the vit series
must be serially correlated. Under the random
effects assumptions,
σa2
corr(vit, vis) = 2 2
∀ t 6= s (5)
(σa + σu )
and this correlation will be positive whenever
σa2 is nontrivial. Indeed, in a case where a large
fraction of the variation in the composite er-
ror term is due to the individual-specific com-
ponent, this correlation can be quite substan-
tial. Since the standard errors of a pooled OLS
model ignore it, they will be biased.

Just as in other cases where we can explicitly

model the form of serial correlation in the error
process, we can use generalized least squares
(GLS) to solve this problem. Although the
algebra to derive the GLS estimator is quite
complex, the GLS transformation itself is sim-
ple. If we define
v
σu2
u !
u
λ=1−t (6)
(σu2 + T σa2)

we will generate a weight λ, 0 ≤ λ ≤ 1. The

transformed equation is then defined in the
quasi-demeaned y and x variables:

(yit − λȳt) = β0(1 − λ) + β1(xit − λx̄t) + (vit − λv̄t)

(7)
where the overbar denotes the averages over
time for each panel unit, just as in the fixed
effects model.

The fixed effects (FE) model arbitrarily sets

λ = 1 and fully demeans the data. As we
can see from equation (??), that would be ap-
propriate in this context iff the idiosyncratic
error variance was very small relative to the
variance of the individual effect ai. In other
cases, where both sources of variance are non-
negligible, the optimal λ will be less than one.
We can also consider the pooled OLS model
in this context; it corresponds to a λ = 0 in
which we do not transform the data at all.
Arbitrarily setting λ = 1 à la FE leads to a
consistent estimator of the equation, but it is
inefficient relative to the RE alternative. Be-
cause the FE model is equivalent to the LSDV
formulation, it involves the loss of N degrees
of freedom. Given that the ai may be consid-
ered as nuisance parameters, if we do not care
about their values, we might rather apply RE
and substantially reduce the degrees of free-
dom lost in estimation: especially important if
T is small.

We do not know λ, of course, so we must con-

sistently estimate it. The ability to do so in-
volves the crucial assumption that cov(xit, ai) =
0: the unobservable individual effects must be
independently distributed of the regressors. If
our estimate of λ is close to zero, the RE esti-
mates will be similar to those of a pooled OLS
model. If our estimate of λ is close to one,
the RE estimates will be similar to those of a
FE model. The RE estimator may be chosen
in Stata by giving the command xtreg depvar
indepvars, re. The estimated results will dis-
play an estimate of λ. In the example of a
wage equation given in the textbook (14.4), a
λ̂ = 0.643 is displayed, indicating that the RE
estimates are likely to differ considerably from
both pooled OLS and FE counterparts.

One interesting feature of the random effects

estimator, evident in that example: since it
involves quasi-demeaning of the data, a vari-
able without time variation within the individ-
ual may be included in the model. Thus, if
we are estimating a wage equation, we can in-
clude gender or race in a RE model, whereas it
cannot be included in a FE model. However,
we must ensure than any such variable satisfies
the assumption that cov(xit, ai) = 0.

Random effects or fixed effects?

To justify RE, the necessary assumption that

an individual effect can be considered indepen-
dent of all regressors is often problematic. If
we are interested in testing the effect of a time-
invariant variable, RE can yield such an esti-
mate, but we should include all available time-
invariant variables as controls to try to ensure
that the independence assumption is satisfied.

If we are interested in evaluating the effect of a

time-varying explanatory variable, can we jus-
tify the use of RE? Yes, but in realistic terms
probably only in the case where the key vari-
able is set randomly. For instance, if students
are assigned randomly to sections of a course
or home rooms in a K-12 context, RE would
be appropriate given that the assignment vari-
able would not be correlated with unobserv-
ables such as aptitude. On the other hand, if
students are grouped by ability or test scores
and assigned to home rooms accordingly, the
assignment variable will not be independent of
the unobservable individual aptitude, and RE
will be inconsistent.
We can formally evaluate the appropriateness
of the RE estimator in a given context with a
Hausman test. A Hausman test compares the
coefficient vectors from two estimators. If they
are both consistent estimators, then their point
estimates should not differ greatly, whereas if
one of the estimators is inconsistent, its point
estimates are likely to differ widely from those
of a consistent estimator. In the current con-
text, the FE estimator is always consistent,
but inefficient under the null hypothesis that
cov(xit, ai) = 0. RE is both consistent and rel-
atively efficient under that null hypothesis, but
inconsistent under the alternative. To evaluate
the null hypothesis, we give the commands

xtreg depvar indepvars1, fe

estimates store fe
xtreg depvar indepvars2, re
estimates store re
hausman fe re, sigmamore
where we note that indepvars1 may not contain
all of the regressors in indepvars2 because the
RE estimator may also estimate time-invariant
effects. It is crucial that the two sets of es-
timates’ names be given in the order shown,
with the always-consistent estimator first in
the hausman command.

The null hypothesis for the Hausman test is

that RE is consistent and should be preferred.
If we reject that null, RE is inappropriate and
FE should be used instead. However, like many
tests, the Hausman test is performed condi-
tional on proper specification of the underly-
ing model. If we have omitted an important
explanatory variable from both forms of the
model, then we are comparing two inconsistent
estimators of the population model. When a
rejection is received, specification tests should
be used to try to rule out this possibility.
We might consider RE as more appropriate
when applied to a random sample of individ-
uals (such as a sample of workers, or the un-
employed, or those who have completed a job
training program), and FE the better choice
when we consider observations corresponding
to a mutually exhaustive set of units: e.g.,
states of the US. If we have a dataset con-
taining all 50 states’ values, it is not a random
sample; it encompasses the entire population.
We may want to allow for a state-specific in-
tercept term, and the FE (a/k/a LSDV) esti-
mator is a simple way to accomplish this.

Panel data methods for other data structures

We have considered the FD, FE and RE es-

timators as appropriate for strict panel data:
those possessing both individual and time sub-
scripts. But we may have datasets that do
not possess a time element at all, but rather
a cross-sectional clustering variable (such as
siblings within each family, or workers within
each plant). Conceptually, we can apply any
of these three panel data estimators in this
context to take account of a common “fam-
ily effect” or “plant effect”. We cannot use
tsset to declare such data as being panel data
in Stata, but we can use the i(panelvar) op-
tion on any form of xtreg to designate the
panel identifier. Just as the standard panel
setup considers the likelihood that the individ-
ual’s identity will give rise to unobserved het-
erogeneity in the form of ai, we may consider
it as likely that belonging to a particular family
or working in a specific plant may have its own
effect.

An alternative, available in most estimation

commands in Stata, is the notion of cluster-
ing. We may consider families or plants in the
prior example of clusters: groups within which
errors are likely to be correlated with one an-
other. The cluster covariance matrix estima-
tor allows for error variances to differ between
clusters (but not within clusters), as well as
allowing for correlations between errors in the
same cluster (but not between clusters). Ig-
noring these correlations will cause estimated
standard errors to be biased and inconsistent.

It may be invoked in regress and many other

commands with the ,cluster(id) option, where
id specifies the name of an integer variable
denoting cluster membership. The values of
id need not be consecutive. When estimat-
ing cluster standard errors, it is important that
there are more clusters than regressors in the
model. In practical terms, this rules out the
case that a panel identifier is specified as the
cluster id and individual-specific constant terms
are estimated. However, that does not rule out
use of the cluster option in a FE mode because
that model does not literally estimate the N
fixed effects among the regressors.

Seemingly unrelated regressions (SURE)

We often have a situation in which we want to

estimate a similar specification for a number of
different units: for instance, the estimation of
a production function or cost function for each
industry. If the equation to be estimated for
a given unit meets the zero conditional mean
assumption, we may estimate each equation
independently. However, we may want to es-
timate the equations jointly: first, to allow
cross-equation restrictions to be imposed or
tested, and second, to gain efficiency, since
we might expect the error terms across equa-
tions to be contemporaneously correlated.Such
equations are often called seemingly unrelated
regressions, and Zellner proposed an estimator
for this problem: the SU R estimator. Unlike
the fixed effects and random effects estima-
tors, whose large-sample justification is based
on “small T, large N” datasets as N → ∞, the
SU R estimator is based on the large-sample
properties of “large T, small N” datasets as
T → ∞. In that context, it may be considered
a multiple time series estimator.

Equation i of the SU R model is:

yi = Xiβi + i, i = 1, . . . , N (8)

where yi is the ith equation’s dependent vari-
able and Xi is the matrix of regressors for the
ith equation, on which we have T observations.
The disturbance process  = [01, 02, . . . , 0N ]0 is
assumed to have an expectation of zero and
a covariance matrix of Ω. We will only con-
sider the case where we have T observations
per equation, although it is feasible to estimate
the model with an unbalanced panel. Note also
that although each Xi matrix will have T rows,
it may have Ki columns. Each equation may
have a differing set of regressors, and apart
from the constant term, there might be no
variables in common across the Xi. Note that
the application of SU R requires that the T ob-
servations per unit must exceed N , the number
of units, in order to render Ω of full rank and
invertible. If this constraint is not satisfied,
SU R cannot be employed.

We assume that E[itjs] = σij , t = s, other-

wise zero. This implies that we are allowing
for the error terms in different equations to
be contemporaneously correlated, but assum-
ing that they are not correlated at other points
(including within a unit: they are assumed in-
dependent). Thus for any two error vectors,

E[i0j ] = σij IT
Ω = Σ ⊗ IT (9)
where Σ is the VCE of the N error vectors and
⊗ is the Kronecker matrix product (For any
matrices AK×L, BM ×N , A ⊗ B = CKM ×LN . To
form the product matrix, each element of A
scalar multiplies the entire matrix B).

The efficient estimator for this problem is gen-

eralized least squares (GLS), in which we may
write y as the stacked set of yi vectors, and X
as the block-diagonal matrix of Xi. Since the
GLS estimator is

bGLS = [X 0Ω−1X][X 0Ω−1y] (10)

and
Ω−1 = Σ−1 ⊗ I (11)
We can write the (infeasible) GLS estimator
as

bGLS = [X 0(Σ−1 ⊗ I)X]−1[X 0(Σ−1 ⊗ I)y] (12)

which if expanded demonstrates that each block
of the Xi0Xj matrix is weighted by the scalar
−1
σij . The large-sample VCE of bGLS is the first
term of this expression.
When will this estimator provide a gain in effi-
ciency over equation-by-equation OLS? First,
if the σij , i 6= j are actually zero, there is no
gain. Second, if the Xi matrices are identical
across equations—not merely having the same
variable names, but containing the same nu-
merical values—then GLS is identical to equation-
by-equation OLS, and there is no gain. Be-
yond these cases, the gain in efficiency de-
pends on the magnitude of the cross-equation
contemporaneous correlations of the residuals.
The higher are those correlations, the greater
the gain. Furthermore, if the Xi matrices’
columns are highly correlated across equations,
the gains will be smaller.

The feasible SU R estimator requires a consis-

tent estimate of Σ, the N × N contempora-
neous covariance matrix of the equations’ dis-
turbance processes. The representative ele-
ment σij , the contemporaneous correlation be-
tween i, j , may be estimated from equation-
by-equation OLS residuals as
e0iej
sij = (13)
T
assuming that each unit’s equation is estimated
from T observations. These estimates are then
used to perform the “Zellner step”, where the
algebra of partitioned matrices will show that
the Kronecker products may be rewritten as
products of the blocks in the expression for
bGLS . The estimator may be iterated. The
GLS estimates will produce a new set of resid-
uals, which may be used in a second Zell-
ner step, and so on. Iteration will make the
GLS estimates equivalent to maximum likeli-
hood estimates of the system.

The SU R estimator is available in Stata via the

sureg command. It is a panel data estimator
applicable to data in the wide format. If the
data are set up in the long format more com-
monly used with panel data, the reshape com-
mand may be used to place them in the “wide”
format. It is an attractive estimator relative to
pooled OLS, or even in comparison with fixed
effects, in that SU R allows each unit to have
its own coefficient vector. Not only the con-
stant term differs from unit to unit, but each
of the slope parameters differ as well across
units, as does σ2, which is constrained to be
equal across units in pooled OLS, fixed effects
or random effects estimators.

Standard F -tests may be used to compare the

unrestricted SU R results with those that may
be generated in the presence of linear con-
straints, such as cross-equation restrictions (see
constraint). Cross-equation constraints corre-
spond to the restriction that a particular re-
gressor’s effect is the same for each panel unit.
The isure option may be used to iterate the
estimates, as described above.
SUR with identical regressors

The second case discussed above in which SU R

will generate the same point and interval estimates—
the case of numerically identical regressors—
arises quite often in economic theory and finan-
cial theory. For instance, the demand for each
good should depend on the set of prices and
income, or the portfolio share of assets held in
a given class should depend on the returns to
each asset and on total wealth. In this case,
there is no reason to use anything other than
OLS in terms of efficiency. However, SU R es-
timation is often employed in this case, since it
allows for tests of cross-equation constraints,
or estimation with those constraints in place.

If we try to apply SU R to a system with (nu-

merically) identical regressors, such as a com-
plete set of cost share or portfolio share equa-
tions, the SU R estimator will fail because the
error covariance matrix is singular. This holds
not only for the unobservable errors, but also
for the least squares residuals. A bit of alge-
bra will show that if there are adding-up con-
straints across equations—for instance, if the
set of yi variables are a complete set of port-
folio shares or demand shares—then the OLS
residuals will sum to zero across equations, and
their empirical covariance matrix will be singu-
lar by construction.

We may still want to utilize systems estima-

tion in order to impose the cross-equation con-
straints arising from economic theory. In this
case, the appropriate estimation strategy is to
drop one of the equations and estimate the
system of (N − 1) equations with SU R. The
parameters of the N th equation, in point and
interval form, can be algebraically derived from
those estimates. The feasible GLS estimates
will be sensitive to which equation is dropped,
but iterated SU R will restore the invariance
property of the maximum likelihood estimator
of the problem.

Dynamic panel data models

A serious difficulty arises with the one-way fixed

effects model in the context of a dynamic panel
data (DPD) model: one containing a lagged
dependent variable (and possibly other regres-
sors), particularly in the “small T , large N ”
context. As Nickell (1981) shows, this arises
because the demeaning process which subtracts
the individual’s mean value of y and each X
from the respective variable creates a correla-
tion between regressor and error. The mean of
the lagged dependent variable contains obser-
vations 0 through (T − 1) on y, and the mean
error—which is being conceptually subtracted
from each it—contains contemporaneous val-
ues of  for t = 1 . . . T . The resulting correla-
tion creates a bias in the estimate of the coeffi-
cient of the lagged dependent variable which is
not mitigated by increasing N , the number of
individual units. In the simplest setup of a pure
AR(1) model without additional regressors:

yit = β + ρyi,t−1 + ui + it (14)

yit − yi· = ρ(yi,t−1 − yi,·−1) + (it − i·)
The demeaning operation creates a regressor
which cannot be distributed independently of
the error term. Nickell demonstrates that the
inconsistency of ρ̂ as N → ∞ is of order 1/T ,
which may be quite sizable in a “small T ” con-
text. If ρ > 0, the bias is invariably negative,
so that the persistence of y will be underesti-
mated. For reasonably large values of T , the
limit of (ρ̂ − ρ) as N → ∞ will be approxi-
mately −(1 + ρ)/(T − 1): a sizable value, even
if T = 10. With ρ = 0.5, the bias will be -
0.167, or about 1/3 of the true value. The
inclusion of additional regressors does not re-
move this bias. Indeed, if the regressors are
correlated with the lagged dependent variable
to some degree, their coefficients may be se-
riously biased as well. Note also that this bias
is not caused by an autocorrelated error pro-
cess . The bias arises even if the error pro-
cess is i.i.d. If the error process is autocorre-
lated, the problem is even more severe given
the difficulty of deriving a consistent estimate
of the AR parameters in that context. The
same problem affects the one-way random ef-
fects model. The ui error component enters
every value of yit by assumption, so that the
lagged dependent variable cannot be indepen-
dent of the composite error process.

A solution to this problem involves taking first

differences of the original model. Consider a
model containing a lagged dependent variable
and a single regressor X:

yit = β1 + ρyi,t−1 + Xitβ2 + ui + it (15)

The first difference transformation removes both
the constant term and the individual effect:

∆yit = ρ∆yi,t−1 + ∆Xitβ2 + ∆it (16)

There is still correlation between the differ-
enced lagged dependent variable and the dis-
turbance process (which is now a first-order
moving average process, or M A(1)): the for-
mer contains yi,t−1 and the latter contains i,t−1.
But with the individual fixed effects swept out,
a straightforward instrumental variables esti-
mator is available. We may construct instru-
ments for the lagged dependent variable from
the second and third lags of y, either in the
form of differences or lagged levels. If  is i.i.d.,
those lags of y will be highly correlated with the
lagged dependent variable (and its difference)
but uncorrelated with the composite error pro-
cess. Even if we had reason to believe that 
might be following an AR(1) process, we could
still follow this strategy, “backing off” one pe-
riod and using the third and fourth lags of y
(presuming that the timeseries for each unit is
long enough to do so).

The DP D (Dynamic Panel Data) approach of

Arellano and Bond (1991) is based on the no-
tion that the instrumental variables approach
noted above does not exploit all of the informa-
tion available in the sample. By doing so in a
GMM context, we may construct more efficient
estimates of the dynamic panel data model.
The Arellano–Bond estimator can be thought
of as an extension of the Anderson–Hsiao es-
timator implemented by xtivreg, fd. Arellano
and Bond argue that the Anderson–Hsiao es-
timator, while consistent, fails to take all of
the potential orthogonality conditions into ac-
count. Consider the equations

yit = Xitβ1 + Witβ2 + vit

vit = ui + it (17)
where Xit includes strictly exogenous regres-
sors, Wit are predetermined regressors (which
may include lags of y) and endogenous regres-
sors, all of which may be correlated with ui, the
unobserved individual effect. First-differencing
the equation removes the ui and its associated
omitted-variable bias. The Arellano–Bond es-
timator sets up a generalized method of mo-
ments (GM M ) problem in which the model is
specified as a system of equations, one per
time period, where the instruments applicable
to each equation differ (for instance, in later
time periods, additional lagged values of the
instruments are available). The instruments
include suitable lags of the levels of the en-
dogenous variables (which enter the equation
in differenced form) as well as the strictly ex-
ogenous regressors and any others that may be
specified. This estimator can easily generate
an immense number of instruments, since by
period τ all lags prior to, say, (τ − 2) might be
individually considered as instruments. If T is
nontrivial, it is often necessary to employ the
option which limits the maximum lag of an in-
strument to prevent the number of instruments
from becoming too large. This estimator is
available in Stata as xtabond.

A potential weakness in the Arellano–Bond DP D

estimator was revealed in later work by Arel-
lano and Bover (1995) and Blundell and Bond
(1995). The lagged levels are often rather
poor instruments for first differenced variables,
especially if the variables are close to a random
walk. Their modification of the estimator in-
cludes lagged levels as well as lagged differ-
ences. The original estimator is often entitled
difference GMM, while the expanded estimator
is commonly termed System GMM. The cost
of the System GMM estimator involves a set of
additional restrictions on the initial conditions
of the process generating y.
Both the difference GMM and System GMM
estimators have one-step and two-step vari-
ants. The two-step estimates of the differ-
ence GMM standard errors have been shown to
have a severe downward bias. If the precision
of the two-step estimators is to be evaluated
for hypothesis tests, we should ensure that
the “Windmeijer finite-sample correction” (see
Windmeijer (2005)). to these standard errors
has been applied. All of the features described
above are available in David Roodman’s im-
proved version of official Stata’s estimator. His
version, xtabond2, offers a much more flexible
syntax than official Stata’s xtabond, which does
not allow the same specification of instrument
sets, nor does it provide the System GMM
approach or the Windmeijer correction to the
standard errors of the two-step estimates.