Title stata.

com
xtpcse — Linear regression with panel-corrected standard errors

Syntax Menu Description Options

Remarks and examples Stored results Methods and formulas Acknowledgments
References Also see

Syntax
         
xtpcse depvar indepvars if in weight , options

options Description
Model
noconstant suppress constant term
correlation(independent) use independent autocorrelation structure
correlation(ar1) use AR1 autocorrelation structure
correlation(psar1) use panel-specific AR1 autocorrelation structure
rhotype(calc) specify method to compute autocorrelation parameter;
seldom used
np1 weight panel-specific autocorrelations by panel sizes
hetonly assume panel-level heteroskedastic errors
independent assume independent errors across panels
by/if/in
casewise include only observations with complete cases
pairwise include all available observations with nonmissing pairs
SE
nmk normalize standard errors by N − k instead of N
Reporting
level(#) set confidence level; default is level(95)
detail report list of gaps in time series
display options control column formats, row spacing, line width, display of omitted
variables and base and empty cells, and factor-variable labeling
coeflegend display legend instead of statistics
A panel variable and a time variable must be specified; use xtset; see [XT] xtset.
indepvars may contain factor variables; see [U] 11.4.3 Factor variables.
depvar and indepvars may contain time-series operators; see [U] 11.4.4 Time-series varlists.
by and statsby are allowed; see [U] 11.1.10 Prefix commands.
iweights and aweights are allowed; see [U] 11.1.6 weight.
coeflegend does not appear in the dialog box.
See [U] 20 Estimation and postestimation commands for more capabilities of estimation commands.

1
 2 xtpcse — Linear regression with panel-corrected standard errors

Menu
Statistics > Longitudinal/panel data > Contemporaneous correlation > Regression with panel-corrected standard
errors (PCSE)

Description
xtpcse calculates panel-corrected standard error (PCSE) estimates for linear cross-sectional time-
series models where the parameters are estimated by either OLS or Prais–Winsten regression. When
computing the standard errors and the variance–covariance estimates, xtpcse assumes that the
disturbances are, by default, heteroskedastic and contemporaneously correlated across panels.
See [XT] xtgls for the generalized least-squares estimator for these models.

Options

 Model
noconstant; see [R] estimation options.
correlation(corr) specifies the form of assumed autocorrelation within panels.
correlation(independent), the default, specifies that there is no autocorrelation.
correlation(ar1) specifies that, within panels, there is first-order autocorrelation AR(1) and
that the coefficient of the AR(1) process is common to all the panels.
correlation(psar1) specifies that, within panels, there is first-order autocorrelation and that the
coefficient of the AR(1) process is specific to each panel. psar1 stands for panel-specific AR(1).
rhotype(calc) specifies the method to be used to calculate the autocorrelation parameter. Allowed
strings for calc are
regress regression using lags; the default
freg regression using leads
tscorr time-series autocorrelation calculation
dw Durbin–Watson calculation
All above methods are consistent and asymptotically equivalent; this is a rarely used option.
np1 specifies that the panel-specific autocorrelations be weighted by Ti rather than by the default
Ti − 1 when estimating a common ρ for all panels, where Ti is the number of observations in
panel i. This option has an effect only when panels are unbalanced and the correlation(ar1)
option is specified.
hetonly and independent specify alternative forms for the assumed covariance of the disturbances
across the panels. If neither is specified, the disturbances are assumed to be heteroskedastic (each
panel has its own variance) and contemporaneously correlated across the panels (each pair of
panels has its own covariance). This is the standard PCSE model.
hetonly specifies that the disturbances are assumed to be panel-level heteroskedastic only with
no contemporaneous correlation across panels.
independent specifies that the disturbances are assumed to be independent across panels; that
is, there is one disturbance variance common to all observations.
 xtpcse — Linear regression with panel-corrected standard errors 3


 by/if/in
casewise and pairwise specify how missing observations in unbalanced panels are to be treated
when estimating the interpanel covariance matrix of the disturbances. The default is casewise
selection.
casewise specifies that the entire covariance matrix be computed only on the observations (periods)
that are available for all panels. If an observation has missing data, all observations of that period
are excluded when estimating the covariance matrix of disturbances. Specifying casewise ensures
that the estimated covariance matrix will be of full rank and will be positive definite.
pairwise specifies that, for each element in the covariance matrix, all available observations
(periods) that are common to the two panels contributing to the covariance be used to compute
the covariance.
The casewise and pairwise options have an effect only when the panels are unbalanced and
neither hetonly nor independent is specified.

 SE
nmk specifies that standard errors be normalized by N − k , where k is the number of parameters
estimated, rather than N , the number of observations. Different authors have used one or the other
normalization. Greene (2012, 280) remarks that whether a degree-of-freedom correction improves
the small-sample properties is an open question.

 Reporting
level(#); see [R] estimation options.
detail specifies that a detailed list of any gaps in the series be reported.
display options: noomitted, vsquish, noemptycells, baselevels, allbaselevels, nofvla-
bel, fvwrap(#), fvwrapon(style), cformat(% fmt), pformat(% fmt), sformat(% fmt), and
nolstretch; see [R] estimation options.

The following option is available with xtpcse but is not shown in the dialog box:
coeflegend; see [R] estimation options.

Remarks and examples stata.com

xtpcse is an alternative to feasible generalized least squares (FGLS)—see [XT] xtgls —for fitting
linear cross-sectional time-series models when the disturbances are not assumed to be independent
and identically distributed (i.i.d.). Instead, the disturbances are assumed to be either heteroskedastic
across panels or heteroskedastic and contemporaneously correlated across panels. The disturbances
may also be assumed to be autocorrelated within panel, and the autocorrelation parameter may be
constant across panels or different for each panel.
We can write such models as
yit = xit β + it
where i = 1, . . . , m is the number of units (or panels); t = 1, . . . , Ti ; Ti is the number of periods in
panel i; and it is a disturbance that may be autocorrelated along t or contemporaneously correlated
across i.
4 xtpcse — Linear regression with panel-corrected standard errors

This model can also be written panel by panel as

y1 X1 1
     
 y 2   X2   
 .  =  .  β +  .2 
 .   ..   .. 
.
ym Xm m

For a model with heteroskedastic disturbances and contemporaneous correlation but with no autocor-
relation, the disturbance covariance matrix is assumed to be

σ11 I11 σ12 I12 ··· σ1m I1m

 
σ21 I21 σ22 I22 ··· σ2m I2m 
E[0 ] = Ω = 

.. .. .. .. 
 . . . .

σm1 Im1 σm2 Im2 · · · σmm Imm

where σii is the variance of the disturbances for panel i, σij is the covariance of the disturbances
between panel i and panel j when the panels’ periods are matched, and I is a Ti by Ti identity
matrix with balanced panels. The panels need not be balanced for xtpcse, but the expression for the
covariance of the disturbances will be more general if they are unbalanced.
This could also be written as

E[0 ] = Σm×m ⊗ ITi ×Ti

where Σ is the panel-by-panel covariance matrix and I is an identity matrix.

See [XT] xtgls for a full taxonomy and description of possible disturbance covariance structures.
xtpcse and xtgls follow two different estimation schemes for this family of models. xtpcse
produces OLS estimates of the parameters when no autocorrelation is specified, or Prais–Winsten (see
[TS] prais) estimates when autocorrelation is specified. If autocorrelation is specified, the estimates
of the parameters are conditional on the estimates of the autocorrelation parameter(s). The estimate
of the variance–covariance matrix of the parameters is asymptotically efficient under the assumed
covariance structure of the disturbances and uses the FGLS estimate of the disturbance covariance
matrix; see Kmenta (1997, 121).
xtgls produces full FGLS parameter and variance–covariance estimates. These estimates are condi-
tional on the estimates of the disturbance covariance matrix and are conditional on any autocorrelation
parameters that are estimated; see Kmenta (1997), Greene (2012), Davidson and MacKinnon (1993),
or Judge et al. (1985).
Both estimators are consistent, as long as the conditional mean (xit β) is correctly specified. If the
assumed covariance structure is correct, FGLS estimates produced by xtgls are more efficient. Beck
and Katz (1995) have shown, however, that the full FGLS variance–covariance estimates are typically
unacceptably optimistic (anticonservative) when used with the type of data analyzed by most social
scientists—10–20 panels with 10–40 periods per panel. They show that the OLS or Prais–Winsten
estimates with PCSEs have coverage probabilities that are closer to nominal.
Because the covariance matrix elements, σij , are estimated from panels i and j , using those
observations that have common time periods, estimators for this model achieve their asymptotic
behavior as the Ti s approach infinity. In contrast, the random- and fixed-effects estimators assume
a different model and are asymptotic in the number of panels m; see [XT] xtreg for details of the
random- and fixed-effects estimators.
 xtpcse — Linear regression with panel-corrected standard errors 5

Although xtpcse allows other disturbance covariance structures, the term PCSE, as used in the
literature, refers specifically to models that are both heteroskedastic and contemporaneously correlated
across panels, with or without autocorrelation.

Example 1: Controlling for heteroskedasticity and cross-panel correlation

Grunfeld and Griliches (1960) analyzed a company’s current-year gross investment (invest) as
determined by the company’s prior year market value (mvalue) and the prior year’s value of the
company’s plant and equipment (kstock). The dataset includes 10 companies over 20 years, from
1935 through 1954, and is a classic dataset for demonstrating cross-sectional time-series analysis.
Greene (2012, 1112) reproduces the dataset.
To use xtpcse, the data must be organized in “long form”; that is, each observation must represent
a record for a specific company at a specific time; see [D] reshape. In the Grunfeld data, company
is a categorical variable identifying the company, and year is a variable recording the year. Here are
the first few records:
. use http://www.stata-press.com/data/r13/grunfeld
. list in 1/5

company year invest mvalue kstock time

1. 1 1935 317.6 3078.5 2.8 1

2. 1 1936 391.8 4661.7 52.6 2
3. 1 1937 410.6 5387.1 156.9 3
4. 1 1938 257.7 2792.2 209.2 4
5. 1 1939 330.8 4313.2 203.4 5

To compute PCSEs, Stata must be able to identify the panel to which each observation belongs and
be able to match the periods across the panels. We tell Stata how to do this matching by specifying
the panel and time variables with xtset; see [XT] xtset. Because the data are annual, we specify the
yearly option.
. xtset company year, yearly
panel variable: company (strongly balanced)
time variable: year, 1935 to 1954
delta: 1 year

We can obtain OLS parameter estimates for a linear model of invest on mvalue and kstock
while allowing the standard errors (and variance–covariance matrix of the estimates) to be consistent
when the disturbances from each observation are not independent. Specifically, we want the standard
errors to be robust to each company having a different variance of the disturbances and to each
company’s observations being correlated with those of the other companies through time.
 6 xtpcse — Linear regression with panel-corrected standard errors

This model is fit in Stata by typing

. xtpcse invest mvalue kstock
Linear regression, correlated panels corrected standard errors (PCSEs)
Group variable: company Number of obs = 200
Time variable: year Number of groups = 10
Panels: correlated (balanced) Obs per group: min = 20
Autocorrelation: no autocorrelation avg = 20
max = 20
Estimated covariances = 55 R-squared = 0.8124
Estimated autocorrelations = 0 Wald chi2(2) = 637.41
Estimated coefficients = 3 Prob > chi2 = 0.0000

Panel-corrected
invest Coef. Std. Err. z P>|z| [95% Conf. Interval]

mvalue .1155622 .0072124 16.02 0.000 .101426 .1296983

kstock .2306785 .0278862 8.27 0.000 .1760225 .2853345
_cons -42.71437 6.780965 -6.30 0.000 -56.00482 -29.42392

Example 2: Comparing the FGLS and PCSE approaches

xtgls will produce more efficient FGLS estimates of the models’ parameters, but with the
disadvantage that the standard error estimates are conditional on the estimated disturbance covariance.
Beck and Katz (1995) argue that the improvement in power using FGLS with such data is small and
that the standard error estimates from FGLS are unacceptably optimistic (anticonservative).
The FGLS model is fit by typing
. xtgls invest mvalue kstock, panels(correlated)
Cross-sectional time-series FGLS regression
Coefficients: generalized least squares
Panels: heteroskedastic with cross-sectional correlation
Correlation: no autocorrelation
Estimated covariances = 55 Number of obs = 200
Estimated autocorrelations = 0 Number of groups = 10
Estimated coefficients = 3 Time periods = 20
Wald chi2(2) = 3738.07
Prob > chi2 = 0.0000

invest Coef. Std. Err. z P>|z| [95% Conf. Interval]

mvalue .1127515 .0022364 50.42 0.000 .1083683 .1171347

kstock .2231176 .0057363 38.90 0.000 .2118746 .2343605
_cons -39.84382 1.717563 -23.20 0.000 -43.21018 -36.47746

The coefficients between the two models are close; the constants differ substantially, but we are
generally not interested in the constant. As Beck and Katz observed, the standard errors for the FGLS
model are 50%–100% smaller than those for the OLS model with PCSE.
If we were also concerned about autocorrelation of the disturbances, we could obtain a model with
a common AR(1) parameter by specifying correlation(ar1).
 xtpcse — Linear regression with panel-corrected standard errors 7

. xtpcse invest mvalue kstock, correlation(ar1)

(note: estimates of rho outside [-1,1] bounded to be in the range [-1,1])
Prais-Winsten regression, correlated panels corrected standard errors (PCSEs)
Group variable: company Number of obs = 200
Time variable: year Number of groups = 10
Panels: correlated (balanced) Obs per group: min = 20
Autocorrelation: common AR(1) avg = 20
max = 20
Estimated covariances = 55 R-squared = 0.5468
Estimated autocorrelations = 1 Wald chi2(2) = 93.71
Estimated coefficients = 3 Prob > chi2 = 0.0000

Panel-corrected
invest Coef. Std. Err. z P>|z| [95% Conf. Interval]

mvalue .0950157 .0129934 7.31 0.000 .0695492 .1204822

kstock .306005 .0603718 5.07 0.000 .1876784 .4243317
_cons -39.12569 30.50355 -1.28 0.200 -98.91154 20.66016

rho .9059774

The estimate of the autocorrelation parameter is high (0.906), and the standard errors are larger
than for the model without autocorrelation, which is to be expected if there is autocorrelation.

Example 3: Controlling for cross-panel correlation and autocorrelation

Let’s estimate panel-specific autocorrelation parameters and change the method of estimating the
autocorrelation parameter to the one typically used to estimate autocorrelation in time-series analysis.
. xtpcse invest mvalue kstock, correlation(psar1) rhotype(tscorr)
Prais-Winsten regression, correlated panels corrected standard errors (PCSEs)
Group variable: company Number of obs = 200
Time variable: year Number of groups = 10
Panels: correlated (balanced) Obs per group: min = 20
Autocorrelation: panel-specific AR(1) avg = 20
max = 20
Estimated covariances = 55 R-squared = 0.8670
Estimated autocorrelations = 10 Wald chi2(2) = 444.53
Estimated coefficients = 3 Prob > chi2 = 0.0000

Panel-corrected
invest Coef. Std. Err. z P>|z| [95% Conf. Interval]

mvalue .1052613 .0086018 12.24 0.000 .0884021 .1221205

kstock .3386743 .0367568 9.21 0.000 .2666322 .4107163
_cons -58.18714 12.63687 -4.60 0.000 -82.95496 -33.41933

rhos = .5135627 .87017 .9023497 .63368 .8571502 ... .8752707

Beck and Katz (1995, 121) make a case against estimating panel-specific AR parameters, as opposed
to one AR parameter for all panels.
 8 xtpcse — Linear regression with panel-corrected standard errors

Example 4: Controlling for heteroskedasticity only; not quite PCSEs

We can also diverge from PCSEs to estimate standard errors that are panel corrected, but only
for panel-level heteroskedasticity; that is, each company has a different variance of the disturbances.
Allowing also for autocorrelation, we would type
. xtpcse invest mvalue kstock, correlation(ar1) hetonly
(note: estimates of rho outside [-1,1] bounded to be in the range [-1,1])
Prais-Winsten regression, heteroskedastic panels corrected standard errors
Group variable: company Number of obs = 200
Time variable: year Number of groups = 10
Panels: heteroskedastic (balanced) Obs per group: min = 20
Autocorrelation: common AR(1) avg = 20
max = 20
Estimated covariances = 10 R-squared = 0.5468
Estimated autocorrelations = 1 Wald chi2(2) = 91.72
Estimated coefficients = 3 Prob > chi2 = 0.0000

Het-corrected
invest Coef. Std. Err. z P>|z| [95% Conf. Interval]

mvalue .0950157 .0130872 7.26 0.000 .0693653 .1206661

kstock .306005 .061432 4.98 0.000 .1856006 .4264095
_cons -39.12569 26.16935 -1.50 0.135 -90.41666 12.16529

rho .9059774

With this specification, we do not obtain what are referred to in the literature as PCSEs. These
standard errors are in the same spirit as PCSEs but are from the asymptotic covariance estimates of
OLS without allowing for contemporaneous correlation.
 xtpcse — Linear regression with panel-corrected standard errors 9

Stored results
xtpcse stores the following in e():
Scalars
e(N) number of observations
e(N g) number of groups
e(N gaps) number of gaps
e(n cf) number of estimated coefficients
e(n cv) number of estimated covariances
e(n cr) number of estimated correlations
e(n sigma) observations used to estimate elements of Sigma
e(mss) model sum of squares
e(df) degrees of freedom
e(df m) model degrees of freedom
e(rss) residual sum of squares
e(g min) smallest group size
e(g avg) average group size
e(g max) largest group size
e(r2) R-squared
e(chi2) χ2
e(p) significance
e(rmse) root mean squared error
e(rank) rank of e(V)
e(rc) return code
Macros
e(cmd) xtpcse
e(cmdline) command as typed
e(depvar) name of dependent variable
e(ivar) variable denoting groups
e(tvar) variable denoting time within groups
e(wtype) weight type
e(wexp) weight expression
e(title) title in estimation output
e(panels) contemporaneous covariance structure
e(corr) correlation structure
e(rhotype) type of estimated correlation
e(rho) ρ
e(cons) noconstant or ""
e(missmeth) casewise or pairwise
e(balance) balanced or unbalanced
e(chi2type) Wald; type of model χ2 test
e(vcetype) title used to label Std. Err.
e(properties) b V
e(predict) program used to implement predict
e(marginsok) predictions allowed by margins
e(asbalanced) factor variables fvset as asbalanced
e(asobserved) factor variables fvset as asobserved
Matrices
e(b) coefficient vector
e(Sigma) Σ
b matrix
e(rhomat) vector of autocorrelation parameter estimates
e(V) variance–covariance matrix of the estimators
Functions
e(sample) marks estimation sample
 10 xtpcse — Linear regression with panel-corrected standard errors

Methods and formulas

If no autocorrelation is specified, the parameters β are estimated by OLS; see [R] regress. If
autocorrelation is specified, the parameters β are estimated by Prais–Winsten; see [TS] prais.
When autocorrelation with panel-specific coefficients of correlation is specified (by using option
correlation(psar1)), each panel-level ρi is computed from the residuals of an OLS regression
across all panels; see [TS] prais. When autocorrelation with a common coefficient of correlation is
specified (by using option correlation(ar1)), the common correlation coefficient is computed as

ρ1 + ρ2 + · · · + ρm
ρ=
m
where ρi is the estimated autocorrelation coefficient for panel i and m is the number of panels.
The covariance of the OLS or Prais–Winsten coefficients is

Var(β) = (X0 X)−1 X0 ΩX(X0 X)−1

where Ω is the full covariance matrix of the disturbances.

When the panels are balanced, we can write Ω as

Ω = Σm×m ⊗ ITi ×Ti

where Σ is the m by m panel-by-panel covariance matrix of the disturbances; see Remarks and
examples.
xtpcse estimates the elements of Σ as
0
b ij = i j
Σ
Tij

where i and j are the residuals for panels i and j , respectively, that can be matched by period, and
where Tij is the number of residuals between the panels i and j that can be matched by time period.
When the panels are balanced (each panel has the same number of observations and all periods
are common to all panels), Tij = T , where T is the number of observations per panel.
When panels are unbalanced, xtpcse by default uses casewise selection, in which only those
residuals from periods that are common to all panels are used to compute Sbij . Here Tij = T ∗ ,
where T ∗ is the number of periods common to all panels. When pairwise is specified, each Sbij is
computed using all observations that can be matched by period between the panels i and j .

Acknowledgments
We thank the following people for helpful comments: Nathaniel Beck of the Department of
Politics at New York University, Jonathan Katz of the Division of the Humanities and Social Science
at California Institute of Technology, and Robert John Franzese Jr. of the Center for Political Studies
at the Institute for Social Research at the University of Michigan.
 xtpcse — Linear regression with panel-corrected standard errors 11

References
Beck, N. L., and J. N. Katz. 1995. What to do (and not to do) with time-series cross-section data. American Political
Science Review 89: 634–647.
Blackwell, J. L., III. 2005. Estimation and testing of fixed-effect panel-data systems. Stata Journal 5: 202–207.
Davidson, R., and J. G. MacKinnon. 1993. Estimation and Inference in Econometrics. New York: Oxford University
Press.
Greene, W. H. 2012. Econometric Analysis. 7th ed. Upper Saddle River, NJ: Prentice Hall.
Grunfeld, Y., and Z. Griliches. 1960. Is aggregation necessarily bad? Review of Economics and Statistics 42: 1–13.
Hoechle, D. 2007. Robust standard errors for panel regressions with cross-sectional dependence. Stata Journal 7:
281–312.
Judge, G. G., W. E. Griffiths, R. C. Hill, H. Lütkepohl, and T.-C. Lee. 1985. The Theory and Practice of Econometrics.
2nd ed. New York: Wiley.
Kmenta, J. 1997. Elements of Econometrics. 2nd ed. Ann Arbor: University of Michigan Press.

Also see
[XT] xtpcse postestimation — Postestimation tools for xtpcse
[XT] xtset — Declare data to be panel data
[XT] xtgls — Fit panel-data models by using GLS
[XT] xtreg — Fixed-, between-, and random-effects and population-averaged linear models
[XT] xtregar — Fixed- and random-effects linear models with an AR(1) disturbance
[R] regress — Linear regression
[TS] newey — Regression with Newey–West standard errors
[TS] prais — Prais – Winsten and Cochrane – Orcutt regression
[U] 20 Estimation and postestimation commands