Panel Data Econometrics in R: The plm Package

Yves Croissant Giovanni Millo

Université de La Réunion University of Trieste and Generali SpA

Abstract
This introduction to the plm package is a slightly modified version of Croissant and
Millo (2008), published in the Journal of Statistical Software.
Panel data econometrics is obviously one of the main fields in the profession, but most
of the models used are difficult to estimate with R. plm is a package for R which intends
to make the estimation of linear panel models straightforward. plm provides functions to
estimate a wide variety of models and to make (robust) inference.

Keywords: panel data, covariance matrix estimators, generalized method of moments, R.

1. Introduction
Panel data econometrics is a continuously developing field. The increasing availability of
data observed on cross-sections of units (like households, firms, countries etc.) and over time
has given rise to a number of estimation approaches exploiting this double dimensionality to
cope with some of the typical problems associated with economic data, first of all that of
unobserved heterogeneity.
Timewise observation of data from different observational units has long been common in
other fields of statistics (where they are often termed longitudinal data). In the panel data
field as well as in others, the econometric approach is nevertheless peculiar with respect to
experimental contexts, as it is emphasizing model specification and testing and tackling a
number of issues arising from the particular statistical problems associated with economic
data.
Thus, while a very comprehensive software framework for (among many other features) max-
imum likelihood estimation of linear regression models for longitudinal data, packages nlme
(Pinheiro, Bates, DebRoy, and the R Core team 2007) and lme4 (Bates 2007), is available in
the R (R Development Core Team 2008) environment and can be used, e.g., for estimation
of random effects panel models, its use is not intuitive for a practicing econometrician, and
maximum likelihood estimation is only one of the possible approaches to panel data econo-
metrics. Moreover, economic panel datasets often happen to be unbalanced (i.e., they have a
different number of observations between groups), which case needs some adaptation to the
methods and is not compatible with those in nlme. Hence the need for a package doing panel
data “from the econometrician’s viewpoint” and featuring at a minimum the basic techniques
econometricians are used to: random and fixed effects estimation of static linear panel data
models, variable coefficients models, generalized method of moments estimation of dynamic
models; and the basic toolbox of specification and misspecification diagnostics.
2 Panel Data Econometrics in R: The plm Package

Furthermore, we felt there was a need for automation of some basic data management tasks
such as lagging, summing and, more in general, applying (in the R sense) functions to the
data, which, although conceptually simple, become cumbersome and error-prone on two-
dimensional data, especially in the case of unbalanced panels.
This paper is organized as follows: Section 2 presents a very short overview of the typical
model taxonomy1 . Section 3 discusses the software approach used in the package. The next
three sections present the functionalities of the package in more detail: data management
(Section 4), estimation (Section 5) and testing (Section 6), giving a short description and
illustrating them with examples. Section 7 compares the approach in plm to that of nlme
and lme4, highlighting the features of the latter two that an econometrician might find most
useful. Section 8 concludes the paper.

2. The linear panel model

The basic linear panel models used in econometrics can be described through suitable restric-
tions of the following general model:

yit = αit + βit> xit + uit (1)

where i = 1, . . . n is the individual (group, country . . . ) index, t = 1, . . . T is the time index

and uit a random disturbance term of mean 0.
Of course uit is not estimable with N = n × T data points. A number of assumptions are
usually made about the parameters, the errors and the exogeneity of the regressors, giving
rise to a taxonomy of feasible models for panel data.
The most common one is parameter homogeneity, which means that αit = α for all i, t and
βit = β for all i, t. The resulting model

yit = α + β > xit + uit (2)

is a standard linear model pooling all the data across i and t.

To model individual heterogeneity, one often assumes that the error term has two separate
components, one of which is specific to the individual and doesn’t change over time2 . This is
called the unobserved effects model:

yit = α + β > xit + µi + it (3)

The appropriate estimation method for this model depends on the properties of the two error
components. The idiosyncratic error it is usually assumed well-behaved and independent of
both the regressors xit and the individual error component µi . The individual component
may be in turn either independent of the regressors or correlated.
1
Comprehensive treatments are to be found in many econometrics textbooks, e.g. Baltagi (2005, 2013) or
Wooldridge (2002, 2010): the reader is referred to these, especially to the first 9 chapters of Baltagi (2005,
2013).
2
For the sake of exposition we are considering only the individual effects case here. There may also be time
effects, which is a symmetric case, or both of them, so that the error has three components: uit = µi + λt + it .
 Yves Croissant, Giovanni Millo 3

If it is correlated, the ordinary least squares (ols) estimator of β would be inconsistent, so

it is customary to treat the µi as a further set of n parameters to be estimated, as if in the
general model αit = αi for all t. This is called the fixed effects (a.k.a. within or least squares
dummy variables) model, usually estimated by ols on transformed data, and gives consistent
estimates for β.
If the individual-specific component µi is uncorrelated with the regressors, a situation which is
usually termed random effects, the overall error uit also is, so the ols estimator is consistent.
Nevertheless, the common error component over individuals induces correlation across the
composite error terms, making ols estimation inefficient, so one has to resort to some form
of feasible generalized least squares (gls) estimators. This is based on the estimation of the
variance of the two error components, for which there are a number of different procedures
available.
If the individual component is missing altogether, pooled ols is the most efficient estimator
for β. This set of assumptions is usually labelled pooling model, although this actually refers
to the errors’ properties and the appropriate estimation method rather than the model itself.
If one relaxes the usual hypotheses of well-behaved, white noise errors and allows for the
idiosyncratic error it to be arbitrarily heteroskedastic and serially correlated over time, a more
general kind of feasible gls is needed, called the unrestricted or general gls. This specification
can also be augmented with individual-specific error components possibly correlated with the
regressors, in which case it is termed fixed effects gls.
Another way of estimating unobserved effects models through removing time-invariant indi-
vidual components is by first-differencing the data: lagging the model and subtracting, the
time-invariant components (the intercept and the individual error component) are eliminated,
and the model

∆yit = β > ∆xit + ∆uit (4)

(where ∆yit = yit − yi,t−1 , ∆xit = xit − xi,t−1 and, from (3), ∆uit = uit − ui,t−1 = ∆it for
t = 2, ..., T ) can be consistently estimated by pooled ols. This is called the first-difference,
or fd estimator. Its relative efficiency, and so reasons for choosing it against other consistent
alternatives, depends on the properties of the error term. The fd estimator is usually preferred
if the errors uit are strongly persistent in time, because then the ∆uit will tend to be serially
uncorrelated.
Lastly, the between model, which is computed on time (group) averages of the data, discards
all the information due to intragroup variability but is consistent in some settings (e.g., non-
stationarity) where the others are not, and is often preferred to estimate long-run relationships.
Variable coefficients models relax the assumption that βit = β for all i, t. Fixed coefficients
models allow the coefficients to vary along one dimension, like βit = βi for all t. Random
coefficients models instead assume that coefficients vary randomly around a common average,
as βit = β + ηi for all t, where ηi is a group– (time–) specific effect with mean zero.
The hypotheses on parameters and error terms (and hence the choice of the most appropriate
estimator) are usually tested by means of:

• pooling tests to check poolability, i.e. the hypothesis that the same coefficients apply
across all individuals,
4 Panel Data Econometrics in R: The plm Package

• if the homogeneity assumption over the coefficients is established, the next step is to
establish the presence of unobserved effects, comparing the null of spherical residuals
with the alternative of group (time) specific effects in the error term,

• the choice between fixed and random effects specifications is based on Hausman-type
tests, comparing the two estimators under the null of no significant difference: if this is
not rejected, the more efficient random effects estimator is chosen,

• even after this step, departures of the error structure from sphericity can further affect
inference, so that either screening tests or robust diagnostics are needed.

Dynamic models and in general lack of strict exogeneity of the regressors, pose further prob-
lems to estimation which are usually dealt with in the generalized method of moments (gmm)
framework.
These were, in our opinion, the basic requirements of a panel data econometrics package
for the R language and environment. Some, as often happens with R, were already fulfilled
by packages developed for other branches of computational statistics, while others (like the
fixed effects or the between estimators) were straightforward to compute after transforming
the data, but in every case there were either language inconsistencies w.r.t. the standard
econometric toolbox or subtleties to be dealt with (like, for example, appropriate computation
of standard errors for the demeaned model, a common pitfall), so we felt there was need for an
“all in one” econometrics-oriented package allowing to make specification searches, estimation
and inference in a natural way.

3. Software approach

3.1. Data structure

Panel data have a special structure: each row of the data corresponds to a specific individual
and time period. In plm the data argument may be an ordinary data.frame but, in this
case, an argument called index has to be added to indicate the structure of the data. This
can be:

• NULL (the default value), it is then assumed that the first two columns contain the
individual and the time index and that observations are ordered by individual and by
time period,

• a character string, which should be the name of the individual index,

• a character vector of length two containing the names of the individual and the time
index,

• an integer which is the number of individuals (only in case of a balanced panel with
observations ordered by individual).

The pdata.frame function is then called internally, which returns a pdata.frame which is
a data.frame with an attribute called index. This attribute is a data.frame that contains
the individual and the time indexes.
 Yves Croissant, Giovanni Millo 5

It is also possible to use directly the pdata.frame function and then to use the pdata.frame
in the estimation functions.

3.2. Interface

Estimation interface
plm provides four functions for estimation:

• plm: estimation of the basic panel models, i.e. within, between and random effect
models. Models are estimated using the lm function to transformed data,

• pvcm: estimation of models with variable coefficients,

• pgmm: estimation of generalized method of moments models,

• pggls: estimation of general feasible generalized least squares models.

The interface of these functions is consistent with the lm() function. Namely, their first two
arguments are formula and data (which should be a data.frame and is mandatory). Three
additional arguments are common to these functions:

• index: this argument enables the estimation functions to identify the structure of the
data, i.e. the individual and the time period for each observation,

• effect: the kind of effects to include in the model, i.e. individual effects, time effects
or both3 ,

• model: the kind of model to be estimated, most of the time a model with fixed effects
or a model with random effects.

The results of these four functions are stored in an object which class has the same name
of the function. They all inherit from class panelmodel. A panelmodel object contains:
coefficients, residuals, fitted.values, vcov, df.residual and call and functions that
extract these elements are provided.

Testing interface
The diagnostic testing interface provides both formula and panelmodel methods for most
functions, with some exceptions. The user may thus choose whether to employ results stored
in a previously estimated panelmodel object or to re-estimate it for the sake of testing.
Although the first strategy is the most efficient one, diagnostic testing on panel models mostly
employs ols residuals from pooling model objects, whose estimation is computationally in-
expensive. Therefore most examples in the following are based on formula methods, which
are perhaps the cleanest for illustrative purposes.
3
Although in most models the individual and time effects cases are symmetric, there are exceptions: es-
timating the fd model on time effects is meaningless because cross-sections do not generally have a natural
ordering, so trying effect="time" stops with an error message as does effect="twoways" which is not defined
for fd models.
6 Panel Data Econometrics in R: The plm Package

3.3. Computational approach to estimation

The feasible gls methods needed for efficient estimation of unobserved effects models have
a simple closed-form solution: once the variance components have been estimated and hence
the covariance matrix of errors V̂ , model parameters can be estimated as

β̂ = (X > V̂ −1 X)−1 (X > V̂ −1 y) (5)

Nevertheless, in practice plain computation of β̂ has long been an intractable problem even
for moderate-sized datasets because of the need to invert the N × N V̂ matrix. With the
advances in computer power, this is no more so, and it is possible to program the “naive”
estimator (5) in R with standard matrix algebra operators and have it working seamlessly for
the standard “guinea pigs”, e.g. the Grunfeld data. Estimation with a couple of thousands
of data points also becomes feasible on a modern machine, although excruciatingly slow and
definitely not suitable for everyday econometric practice. Memory limits would also be very
near because of the storage needs related to the huge V̂ matrix. An established solution
exists for the random effects model which reduces the problem to an ordinary least squares
computation.

The (quasi–)demeaning framework

The estimation methods for the basic models in panel data econometrics, the pooled ols, ran-
dom effects and fixed effects (or within) models, can all be described inside the ols estimation
framework. In fact, while pooled ols simply pools data, the standard way of estimating fixed
effects models with, say, group (time) effects entails transforming the data by subtracting the
average over time (group) to every variable, which is usually termed time-demeaning. In the
random effects case, the various feasible gls estimators which have been put forth to tackle
the issue of serial correlation induced by the group-invariant random effect have been proven
to be equivalent (as far as estimation of βs is concerned) to ols on partially demeaned data,
where partial demeaning is defined as:

yit − θȳi = (Xit − θX̄i )β + (uit − θūi ) (6)

where θ = 1−[σu2 /(σu2 +T σe2 )]1/2 , ȳ and X̄ denote time means of y and X, and the disturbance
vit − θv̄i is homoskedastic and serially uncorrelated. Thus the feasible re estimate for β may
be obtained estimating θ̂ and running an ols regression on the transformed data with lm().
The other estimators can be computed as special cases: for θ = 1 one gets the fixed effects
estimator, for θ = 0 the pooled ols one.
Moreover, instrumental variable estimators of all these models may also be obtained using
several calls to lm().
For this reason the three above estimators have been grouped inside the same function.
On the output side, a number of diagnostics and a very general coefficients’ covariance matrix
estimator also benefits from this framework, as they can be readily calculated applying the
standard ols formulas to the demeaned data, which are contained inside plm objects. This
will be the subject of Subsection 3.4.
 Yves Croissant, Giovanni Millo 7

The object oriented approach to general GLS computations

The covariance matrix of errors in general gls models is too generic to fit the quasi-demeaning
framework, so this method calls for a full-blown application of gls as in (5). On the other
hand, this estimator relies heavily on n–asymptotics, making it theoretically most suitable
for situations which forbid it computationally: e.g., “short” micropanels with thousands of
individuals observed over few time periods.
R has general facilities for fast matrix computation based on object orientation: particular
types of matrices (symmetric, sparse, dense etc.) are assigned the relevant class and the
additional information on structure is used in the computations, sometimes with dramatic
effects on performance (see Bates 2004) and packages Matrix (see Bates and Maechler 2016)
and SparseM (see Koenker and Ng 2016). Some optimized linear algebra routines are available
in the R package bdsmatrix (see Therneau 2014) which exploit the particular block-diagonal
and symmetric structure of V̂ making it possible to implement a fast and reliable full-matrix
solution to problems of any practically relevant size.
The V̂ matrix is constructed as an object of class bdsmatrix. The peculiar properties of this
matrix class are used for efficiently storing the object in memory and then by ad-hoc versions
of the solve and crossprod methods, dramatically reducing computing times and memory
usage. The resulting matrix is then used “the naive way” as in (5) to compute β̂, resulting in
speed comparable to that of the demeaning solution.

3.4. Inference in the panel model

General frameworks for restrictions and linear hypotheses testing are available in the R en-
vironment4 . These are based on the Wald test, constructed as β̂ > V̂ −1 β̂, where β̂ and V̂ are
consistent estimates of β and V (β), The Wald test may be used for zero-restriction (i.e., signifi-
cance) testing and, more generally, for linear hypotheses in the form (Rβ̂−r)> [RV̂ R> ]−1 (Rβ̂−
r)5 . To be applicable, the test functions require extractor methods for coefficients’ and covari-
ance matrix estimates to be defined for the model object to be tested. Model objects in plm
all have coef() and vcov() methods and are therefore compatible with the above functions.
In the same framework, robust inference is accomplished substituting (“plugging in”) a robust
estimate of the coefficient covariance matrix into the Wald statistic formula. In the panel
context, the estimator of choice is the White system estimator. This called for a flexible
method for computing robust coefficient covariance matrices à la White for plm objects.
A general White system estimator for panel data is:

n
V̂R (β) = (X > X)−1 Xi> Ei Xi (X > X)−1
X
(7)
i=1

where Ei is a function of the residuals êit , t = 1, . . . T chosen according to the relevant het-
eroskedasticity and correlation structure. Moreover, it turns out that the White covariance
matrix calculated on the demeaned model’s regressors and residuals (both part of plm ob-
jects) is a consistent estimator of the relevant model’s parameters’ covariance matrix, thus the
4
See packages lmtest (Hothorn, Zeileis, Farebrother, Cummins, Millo, and Mitchell 2015) and car (Fox
2016).
5
Moreover, coeftest() provides a compact way of looking at coefficient estimates and significance diag-
nostics.
8 Panel Data Econometrics in R: The plm Package

method is readily applicable to models estimated by random or fixed effects, first difference
or pooled ols methods. Different pre-weighting schemes taken from package sandwich (see
Zeileis 2004; Lumley and Zeileis 2015) are also implemented to improve small-sample per-
formance. Robust estimators with any combination of covariance structures and weighting
schemes can be passed on to the testing functions.

4. Managing data and formulae

The package is now illustrated by application to some well-known examples. It is loaded using

R> library("plm")

The four datasets used are EmplUK which was used by Arellano and Bond (1991), the Grunfeld
data (Kleiber and Zeileis 2008) which is used in several econometric books, the Produc data
used by Munnell (1990) and the Wages used by Cornwell and Rupert (1988).

R> data("EmplUK", package="plm")

R> data("Produc", package="plm")
R> data("Grunfeld", package="plm")
R> data("Wages", package="plm")

4.1. Data structure

As observed above, the current version of plm is capable of working with a regular data.frame
without any further transformation, provided that the individual and time indexes are in the
first two columns, as in all the example datasets but Wages. If this weren’t the case, an index
optional argument would have to be passed on to the estimating and testing functions.

R> head(Grunfeld)

firm year inv value capital

1 1 1935 317.6 3078.5 2.8
2 1 1936 391.8 4661.7 52.6
3 1 1937 410.6 5387.1 156.9
4 1 1938 257.7 2792.2 209.2
5 1 1939 330.8 4313.2 203.4
6 1 1940 461.2 4643.9 207.2

R> E <- pdata.frame(EmplUK, index=c("firm","year"), drop.index=TRUE, row.names=TRUE)

R> head(E)

sector emp wage capital output

1-1977 7 5.041 13.1516 0.5894 95.7072
1-1978 7 5.600 12.3018 0.6318 97.3569
1-1979 7 5.015 12.8395 0.6771 99.6083
1-1980 7 4.715 13.8039 0.6171 100.5501
1-1981 7 4.093 14.2897 0.5076 99.5581
1-1982 7 3.166 14.8681 0.4229 98.6151
 Yves Croissant, Giovanni Millo 9

R> head(attr(E, "index"))

firm year
1 1 1977
2 1 1978
3 1 1979
4 1 1980
5 1 1981
6 1 1982

Two further arguments are logical: drop.index = TRUE drops the indexes from the data.frame
and row.names = TRUE computes “fancy” row names by pasting the individual and the time
indexes. While extracting a series from a pdata.frame, a pseries is created, which is the
original series with the index attribute. This object has specific methods, like summary and
as.matrix. The former indicates the total variation of the variable and the shares of this
variation due to the individual and the time dimensions. The latter gives the matrix repre-
sentation of the series, with, by default, individuals as rows and times as columns.

R> summary(E$emp)

total sum of squares: 261539.4

id time
0.980765381 0.009108488

R> head(as.matrix(E$emp))

1976 1977 1978 1979 1980 1981 1982 1983 1984

1 NA 5.041 5.600 5.015 4.715 4.093 3.166 2.936 NA
2 NA 71.319 70.643 70.918 72.031 73.689 72.419 68.518 NA
3 NA 19.156 19.440 19.900 20.240 19.570 18.125 16.850 NA
4 NA 26.160 26.740 27.280 27.830 27.169 24.504 22.562 NA
5 86.677 87.100 87.000 90.400 89.200 82.700 73.700 NA NA
6 0.748 0.766 0.762 0.729 0.731 0.779 0.782 NA NA

4.2. Data transformation

Panel data estimation requires to apply different transformations to raw series. If x is a series
of length nT (where n is the number of individuals and T is the number of time periods), the
transformed series x̃ is obtained as x̃ = M x where M is a transformation matrix. Denoting
j a vector of one of length T and In the identity matrix of dimension n, we get:

• the between transformation: P = T1 In ⊗ jj 0 returns a vector containing the individ-

ual means. The Between and between functions perform this operation, the first one
returning a vector of length nT , the second one a vector of length n,

• the within transformation: Q = InT − P returns a vector containing the values in

deviation from the individual means. The Within function performs this operation.
10 Panel Data Econometrics in R: The plm Package

• the first difference transformation D = In ⊗ d where

 
1 −1 0 0 ... 0 0

 0 1 −1 0 . . . 0 0 

0 0 1 −1 . . . 0 0
 
d= 

.. .. .. .. . . .. .. 
. . . . . . .
 
 
0 0 0 0 ... 1 −1
is of dimension (T − 1, T ).
Note that R’s diff() and lag() functions don’t compute correctly these transformations
for panel data because they are unable to identify when there is a change in individual
in the data. Therefore, specific methods for pseries objects have been written in order
to handle correctly panel data. Note that compared to the lag() method for ts objects,
the order of lags are indicated by a positive integer. Moreover, 0 is a relevant value and
a vector argument may be provided:

R> head(lag(E$emp, 0:2))

0 1 2
1-1977 5.041 NA NA
1-1978 5.600 5.041 NA
1-1979 5.015 5.600 5.041
1-1980 4.715 5.015 5.600
1-1981 4.093 4.715 5.015
1-1982 3.166 4.093 4.715

Further functions called Between, between and Within are also provided to compute the
between and the within transformation. The between returns unique values, whereas
Between duplicates the values and returns a vector which length is the number of ob-
servations.

R> head(diff(E$emp), 10)

1-1977 1-1978 1-1979 1-1980 1-1981 1-1982 1-1983

NA 0.5590000 -0.5850000 -0.2999997 -0.6220003 -0.9270000 -0.2299998
2-1977 2-1978 2-1979
NA -0.6760020 0.2750010

R> head(lag(E$emp, 2), 10)

1-1977 1-1978 1-1979 1-1980 1-1981 1-1982 1-1983 2-1977 2-1978 2-1979
NA NA 5.041 5.600 5.015 4.715 4.093 NA NA 71.319

R> head(Within(E$emp))

1-1977 1-1978 1-1979 1-1980 1-1981 1-1982

0.6744285 1.2334285 0.6484285 0.3484288 -0.2735715 -1.2005715

R> head(between(E$emp), 4)
 Yves Croissant, Giovanni Millo 11

1 2 3 4
4.366571 71.362428 19.040143 26.035000

R> head(Between(E$emp), 10)

1 1 1 1 1 1 1 2
4.366571 4.366571 4.366571 4.366571 4.366571 4.366571 4.366571 71.362428
2 2
71.362428 71.362428

4.3. Formulas
In some circumstances, standard formulas are not very useful to describe a model, notably
while using instrumental variable like estimators: to deal with these situations, we use the
Formula package.
The Formula package provides a class which enables to construct multi-part formula, each
part being separated by a pipe sign. plm provides a pFormula object which is a Formula with
specific methods.
The two formulas below are identical:

R> emp~wage+capital|lag(wage,1)+capital

emp ~ wage + capital | lag(wage, 1) + capital

R> emp~wage+capital|.-wage+lag(wage,1)

emp ~ wage + capital | . - wage + lag(wage, 1)

In the second case, the . means the previous parts which describes the covariates and this
part is “updated”. This is particularly interesting when there are a few external instruments.

5. Model estimation

5.1. Estimation of the basic models with plm

Several models can be estimated with plm by filling the model argument:

• the fixed effects model ("within"),

• the pooling model ("pooling"),

• the first-difference model ("fd"),

• the between model ("between"),

• the error components model ("random").

12 Panel Data Econometrics in R: The plm Package

The basic use of plm is to indicate the model formula, the data and the model to be estimated.
For example, the fixed effects model and the random effects model are estimated using:

R> grun.fe <- plm(inv~value+capital, data = Grunfeld, model = "within")

R> grun.re <- plm(inv~value+capital, data = Grunfeld, model = "random")

R> summary(grun.re)

Oneway (individual) effect Random Effect Model

(Swamy-Arora's transformation)

Call:
plm(formula = inv ~ value + capital, data = Grunfeld, model = "random")

Balanced Panel: n = 10, T = 20, N = 200

Effects:
var std.dev share
idiosyncratic 2784.46 52.77 0.282
individual 7089.80 84.20 0.718
theta: 0.8612

Residuals:
Min. 1st Qu. Median 3rd Qu. Max.
-177.6063 -19.7350 4.6851 19.5105 252.8743

Coefficients:
Estimate Std. Error t-value Pr(>|t|)
(Intercept) -57.834415 28.898935 -2.0013 0.04674 *
value 0.109781 0.010493 10.4627 < 2e-16 ***
capital 0.308113 0.017180 17.9339 < 2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Total Sum of Squares: 2381400

Residual Sum of Squares: 548900
R-Squared: 0.7695
Adj. R-Squared: 0.76716
F-statistic: 328.837 on 2 and 197 DF, p-value: < 2.22e-16

For a random model, the summary method gives information about the variance of the com-
ponents of the errors. Fixed effects may be extracted easily using fixef. An argument type
indicates how fixed effects should be computed: in levels type = "level" (the default), in
deviations from the overall mean type = "dmean" or in deviations from the first individual
type = "dfirst".

R> fixef(grun.fe, type = "dmean")

 Yves Croissant, Giovanni Millo 13

1 2 3 4 5 6
-11.552778 160.649753 -176.827902 30.934645 -55.872873 35.582644
7 8 9 10
-7.809534 1.198282 -28.478333 52.176096

The fixef function returns an object of class fixef. A summary method is provided, which
prints the effects (in deviation from the overall intercept), their standard errors and the test
of equality to the overall intercept.

R> summary(fixef(grun.fe, type = "dmean"))

Estimate Std. Error t-value Pr(>|t|)

1 -11.5528 49.7080 -0.2324 0.8164700
2 160.6498 24.9383 6.4419 9.627e-10 ***
3 -176.8279 24.4316 -7.2377 1.130e-11 ***
4 30.9346 14.0778 2.1974 0.0292129 *
5 -55.8729 14.1654 -3.9443 0.0001129 ***
6 35.5826 12.6687 2.8087 0.0054998 **
7 -7.8095 12.8430 -0.6081 0.5438694
8 1.1983 13.9931 0.0856 0.9318489
9 -28.4783 12.8919 -2.2090 0.0283821 *
10 52.1761 11.8269 4.4116 1.725e-05 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

In case of a two-ways effect model, an additional argument effect is required to extract fixed
effects:

R> grun.twfe <- plm(inv~value+capital,data=Grunfeld,model="within",effect="twoways")

R> fixef(grun.twfe,effect="time")

1935 1936 1937 1938 1939 1940 1941

-32.83632 -52.03372 -73.52633 -72.06272 -102.30660 -77.07140 -51.64078
1942 1943 1944 1945 1946 1947 1948
-53.97611 -75.81394 -75.93509 -88.51936 -64.00560 -72.22856 -76.55283
1949 1950 1951 1952 1953 1954
-106.33142 -108.73243 -95.31723 -97.46866 -100.55428 -126.36254

5.2. More advanced use of plm

Random effects estimators

As observed above, the random effect model is obtained as a linear estimation on quasi-
demeaned data. The parameter of this transformation is obtained using preliminary estima-
tions.
Four estimators of this parameter are available, depending on the value of the argument
random.method:
14 Panel Data Econometrics in R: The plm Package

• "swar": from Swamy and Arora (1972), the default value,

• "walhus": from Wallace and Hussain (1969),

• "amemiya": from Amemiya (1971),

• "nerlove": from Nerlove (1971).

For example, to use the amemiya estimator:

R> grun.amem <- plm(inv~value+capital, data=Grunfeld,

+ model="random", random.method="amemiya")

The estimation of the variance of the error components are performed using the ercomp
function, which has a method and an effect argument, and can be used by itself:

R> ercomp(inv~value+capital, data=Grunfeld, method = "amemiya", effect = "twoways")

var std.dev share

idiosyncratic 2644.13 51.42 0.256
individual 7452.02 86.33 0.721
time 243.78 15.61 0.024
theta: 0.868 (id) 0.2787 (time) 0.2776 (total)

Introducing time or two-ways effects

The default behavior of plm is to introduce individual effects. Using the effect argument,
one may also introduce:

• time effects (effect="time"),

• individual and time effects (effect="twoways").

For example, to estimate a two-ways effect model for the Grunfeld data:

R> grun.tways <- plm(inv~value+capital, data = Grunfeld, effect = "twoways",

+ model = "random", random.method = "amemiya")
R> summary(grun.tways)

Twoways effects Random Effect Model

(Amemiya's transformation)

Call:
plm(formula = inv ~ value + capital, data = Grunfeld, effect = "twoways",
model = "random", random.method = "amemiya")

Balanced Panel: n = 10, T = 20, N = 200

 Yves Croissant, Giovanni Millo 15

Effects:
var std.dev share
idiosyncratic 2644.13 51.42 0.256
individual 7452.02 86.33 0.721
time 243.78 15.61 0.024
theta: 0.868 (id) 0.2787 (time) 0.2776 (total)

Residuals:
Min. 1st Qu. Median 3rd Qu. Max.
-176.9062 -18.0431 3.2697 17.1719 234.1735

Coefficients:
Estimate Std. Error t-value Pr(>|t|)
(Intercept) -63.767791 29.851537 -2.1362 0.0339 *
value 0.111386 0.010909 10.2102 <2e-16 ***
capital 0.323321 0.018772 17.2232 <2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Total Sum of Squares: 2066800

Residual Sum of Squares: 518200
R-Squared: 0.74927
Adj. R-Squared: 0.74673
F-statistic: 294.358 on 2 and 197 DF, p-value: < 2.22e-16

In the “effects” section of the result, the variance of the three elements of the error term and
the three parameters used in the transformation are now printed. The two-ways effect model
is for the moment only available for balanced panels.

Unbalanced panels
Most of the features of plm are implemented for panel models with some limitations:

• the random two-ways effect model is not implemented,

• the only estimator of the variance of the error components is the one proposed by Swamy
and Arora (1972)

The following example is using data used by (Harrison and Rubinfeld 1978) to estimate an
hedonic housing prices function. It is reproduced in (Baltagi 2005), p.174/Baltagi (2013),
p.197.

R> data("Hedonic", package = "plm")

R> Hed <- plm(mv~crim+zn+indus+chas+nox+rm+age+dis+rad+tax+ptratio+blacks+lstat,
+ data = Hedonic, model = "random", index = "townid")
R> summary(Hed)
16 Panel Data Econometrics in R: The plm Package

Oneway (individual) effect Random Effect Model

(Swamy-Arora's transformation)

Call:
plm(formula = mv ~ crim + zn + indus + chas + nox + rm + age +
dis + rad + tax + ptratio + blacks + lstat, data = Hedonic,
model = "random", index = "townid")

Unbalanced Panel: n = 92, T = 1-30, N = 506

Effects:
var std.dev share
idiosyncratic 0.01696 0.13025 0.562
individual 0.01324 0.11505 0.438
theta:
Min. 1st Qu. Median Mean 3rd Qu. Max.
0.2505 0.5483 0.6284 0.6141 0.7147 0.7976

Residuals:
Min. 1st Qu. Median Mean 3rd Qu. Max.
-0.62902 -0.06712 -0.00156 -0.00216 0.06858 0.54973

Coefficients:
Estimate Std. Error t-value Pr(>|t|)
(Intercept) 9.6859e+00 1.9751e-01 49.0398 < 2.2e-16 ***
crim -7.4120e-03 1.0478e-03 -7.0738 5.211e-12 ***
zn 7.8877e-05 6.5001e-04 0.1213 0.9034662
indus 1.5563e-03 4.0349e-03 0.3857 0.6998718
chasyes -4.4247e-03 2.9212e-02 -0.1515 0.8796662
nox -5.8425e-03 1.2452e-03 -4.6921 3.510e-06 ***
rm 9.0552e-03 1.1886e-03 7.6182 1.331e-13 ***
age -8.5787e-04 4.6793e-04 -1.8333 0.0673581 .
dis -1.4442e-01 4.4094e-02 -3.2753 0.0011301 **
rad 9.5984e-02 2.6611e-02 3.6069 0.0003415 ***
tax -3.7740e-04 1.7693e-04 -2.1331 0.0334132 *
ptratio -2.9476e-02 9.0698e-03 -3.2499 0.0012336 **
blacks 5.6278e-01 1.0197e-01 5.5188 5.529e-08 ***
lstat -2.9107e-01 2.3927e-02 -12.1650 < 2.2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Total Sum of Squares: 987.94

Residual Sum of Squares: 8.9988
R-Squared: 0.99091
Adj. R-Squared: 0.99067
F-statistic: 4117.13 on 13 and 492 DF, p-value: < 2.22e-16
 Yves Croissant, Giovanni Millo 17

Instrumental variable estimators

All of the models presented above may be estimated using instrumental variables. The in-
struments are specified at the end of the formula after a | sign.
The instrumental variables estimator used is indicated with the inst.method argument:

• "bvk", from Balestra and Varadharajan-Krishnakumar (1987), the default value,

• "baltagi", from Baltagi (1981).

• "am", from Amemiya and MaCurdy (1986)

• "bms", from Breusch, Mizon, and Schmidt (1989)

as illustrated in the following example from Baltagi (2005), p.120/Baltagi (2013), p.137
(”G2SLS”).

R> data("Crime", package = "plm")

R> cr <- plm(log(crmrte) ~ log(prbarr) + log(polpc) + log(prbconv) +
+ log(prbpris) + log(avgsen) + log(density) + log(wcon) +
+ log(wtuc) + log(wtrd) + log(wfir) + log(wser) + log(wmfg) +
+ log(wfed) + log(wsta) + log(wloc) + log(pctymle) + log(pctmin) +
+ region + smsa + factor(year) | . - log(prbarr) - log(polpc) +
+ log(taxpc) + log(mix), data = Crime,
+ model = "random")
R> summary(cr)

Oneway (individual) effect Random Effect Model

(Swamy-Arora's transformation)
Instrumental variable estimation
(Balestra-Varadharajan-Krishnakumar's transformation)

Call:
plm(formula = log(crmrte) ~ log(prbarr) + log(polpc) + log(prbconv) +
log(prbpris) + log(avgsen) + log(density) + log(wcon) + log(wtuc) +
log(wtrd) + log(wfir) + log(wser) + log(wmfg) + log(wfed) +
log(wsta) + log(wloc) + log(pctymle) + log(pctmin) + region +
smsa + factor(year) | . - log(prbarr) - log(polpc) + log(taxpc) +
log(mix), data = Crime, model = "random")

Balanced Panel: n = 90, T = 7, N = 630

Effects:
var std.dev share
idiosyncratic 0.02227 0.14923 0.326
individual 0.04604 0.21456 0.674
theta: 0.7458
18 Panel Data Econometrics in R: The plm Package

Residuals:
Min. 1st Qu. Median 3rd Qu. Max.
-0.7485123 -0.0710015 0.0040742 0.0784401 0.4756493

Coefficients:
Estimate Std. Error t-value Pr(>|t|)
(Intercept) -0.4538241 1.7029840 -0.2665 0.789955
log(prbarr) -0.4141200 0.2210540 -1.8734 0.061498 .
log(polpc) 0.5049285 0.2277811 2.2167 0.027014 *
log(prbconv) -0.3432383 0.1324679 -2.5911 0.009798 **
log(prbpris) -0.1900437 0.0733420 -2.5912 0.009796 **
log(avgsen) -0.0064374 0.0289406 -0.2224 0.824052
log(density) 0.4343519 0.0711528 6.1045 1.847e-09 ***
log(wcon) -0.0042963 0.0414225 -0.1037 0.917426
log(wtuc) 0.0444572 0.0215449 2.0635 0.039495 *
log(wtrd) -0.0085626 0.0419822 -0.2040 0.838456
log(wfir) -0.0040302 0.0294565 -0.1368 0.891220
log(wser) 0.0105604 0.0215822 0.4893 0.624798
log(wmfg) -0.2017917 0.0839423 -2.4039 0.016520 *
log(wfed) -0.2134634 0.2151074 -0.9924 0.321421
log(wsta) -0.0601083 0.1203146 -0.4996 0.617544
log(wloc) 0.1835137 0.1396721 1.3139 0.189383
log(pctymle) -0.1458448 0.2268137 -0.6430 0.520458
log(pctmin) 0.1948760 0.0459409 4.2419 2.565e-05 ***
regionwest -0.2281780 0.1010317 -2.2585 0.024272 *
regioncentral -0.1987675 0.0607510 -3.2718 0.001129 **
smsayes -0.2595423 0.1499780 -1.7305 0.084046 .
factor(year)82 0.0132140 0.0299923 0.4406 0.659676
factor(year)83 -0.0847676 0.0320008 -2.6489 0.008286 **
factor(year)84 -0.1062004 0.0387893 -2.7379 0.006366 **
factor(year)85 -0.0977398 0.0511685 -1.9102 0.056587 .
factor(year)86 -0.0719390 0.0605821 -1.1875 0.235512
factor(year)87 -0.0396520 0.0758537 -0.5227 0.601345
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Total Sum of Squares: 30.168

Residual Sum of Squares: 12.418
R-Squared: 0.59228
Adj. R-Squared: 0.5747
F-statistic: 33.1494 on 26 and 603 DF, p-value: < 2.22e-16

The Hausman-Taylor model (see Hausman and Taylor 1981) may be estimated with the pht
function. The following example is from Baltagi (2005), p.130/Baltagi (2013), p.146.

R> ht <- pht(lwage~wks+south+smsa+married+exp+I(exp^2)+

+ bluecol+ind+union+sex+black+ed |
 Yves Croissant, Giovanni Millo 19

+ sex+black+bluecol+south+smsa+ind,
+ data=Wages,index=595)
R> summary(ht)

Oneway (individual) effect Hausman-Taylor Model

Call:
pht(formula = lwage ~ wks + south + smsa + married + exp + I(exp^2) +
bluecol + ind + union + sex + black + ed | sex + black +
bluecol + south + smsa + ind, data = Wages, index = 595)

T.V. exo : NA, sex, black, bluecol, south, smsa, ind

T.V. endo : wks, married, exp, I(exp^2), union, ed
T.I. exo :
T.I. endo :

Balanced Panel: n = 595, T = 7, N = 4165

Effects:
var std.dev share
idiosyncratic 0.02304 0.15180 0.134
individual 0.14913 0.38618 0.866
theta: 0.853

Residuals:
Min. 1st Qu. Median 3rd Qu. Max.
-2.070475 -0.116106 0.013176 0.125657 2.139104

Coefficients:
Estimate Std. Error z-value Pr(>|z|)
(Intercept) 2.9317e+00 1.7849e-01 16.4256 < 2.2e-16 ***
wks 8.3787e-04 7.7790e-04 1.0771 0.2814400
southyes 2.9987e-02 3.2519e-02 0.9221 0.3564628
smsayes -3.7427e-02 2.2243e-02 -1.6826 0.0924499 .
marriedyes -3.0798e-02 2.4596e-02 -1.2522 0.2105119
exp 1.1284e-01 3.2032e-03 35.2261 < 2.2e-16 ***
I(exp^2) -4.2043e-04 7.0808e-05 -5.9376 2.892e-09 ***
bluecolyes -1.7773e-02 1.7855e-02 -0.9954 0.3195249
ind -8.9816e-03 1.8608e-02 -0.4827 0.6293372
unionyes 3.3535e-02 1.9281e-02 1.7393 0.0819856 .
sexfemale -1.3980e-01 7.1167e-02 -1.9643 0.0494919 *
blackyes -2.9270e-01 8.4004e-02 -3.4844 0.0004932 ***
ed 1.3698e-01 1.2456e-02 10.9973 < 2.2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Total Sum of Squares: 886.9

Residual Sum of Squares: 161.45
20 Panel Data Econometrics in R: The plm Package

F-statistic: 1554.72 on 12 and 4152 DF, p-value: < 2.22e-16

5.3. Variable coefficients model

The pvcm function enables the estimation of variable coefficients models. Time or individual
effects are introduced if effect is fixed to "time" or "individual" (the default value).
Coefficients are assumed to be fixed if model="within" or random if model="random". In the
first case, a different model is estimated for each individual (or time period). In the second
case, the Swamy model (see Swamy 1970) model is estimated. It is a generalized least squares
model which uses the results of the previous model. Denoting β̂i the vectors of coefficients
obtained for each individual, we get:

n 
!
−1  −1
ˆ + σ̂i2 (Xi> Xi )−1 ˆ + σ̂i2 (Xi> Xi )−1
X
β̂ = ∆ ∆ β̂i (8)
i=1

where σ̂i2 is the unbiased estimator of the variance of the errors for individual i obtained from
the preliminary estimation and:

n n
! n
!> n
ˆ = 1 1X 1X 1X
σ̂ 2 (X > Xi )−1
X
∆ β̂i − β̂i β̂i − β̂i − (9)
n − 1 i=1 n i=1 n i=1 n i=1 i i

If this matrix is not positive-definite, the second term is dropped.

With the Grunfeld data, we get:

R> grun.varw <- pvcm(inv~value+capital, data=Grunfeld, model="within")

R> grun.varr <- pvcm(inv~value+capital, data=Grunfeld, model="random")
R> summary(grun.varr)

Oneway (individual) effect Random coefficients model

Call:
pvcm(formula = inv ~ value + capital, data = Grunfeld, model = "random")

Balanced Panel: n = 10, T = 20, N = 200

Residuals:
total sum of squares: 2177914
id time
0.67677732 0.02974195

Estimated mean of the coefficients:

Estimate Std. Error z-value Pr(>|z|)
(Intercept) -9.629285 17.035040 -0.5653 0.5718946
value 0.084587 0.019956 4.2387 2.248e-05 ***
capital 0.199418 0.052653 3.7874 0.0001522 ***
 Yves Croissant, Giovanni Millo 21

---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Estimated variance of the coefficients:

(Intercept) value capital
(Intercept) 2344.24402 -0.6852340 -4.0276612
value -0.68523 0.0031182 -0.0011847
capital -4.02766 -0.0011847 0.0244824

Total Sum of Squares: 474010000

Residual Sum of Squares: 2194300
Multiple R-Squared: 0.99537

5.4. Generalized method of moments estimator

The generalized method of moments is mainly used in panel data econometrics to estimate
dynamic models (Arellano and Bond 1991; Holtz-Eakin, Newey, and Rosen 1988).

yit = ρyit−1 + β > xit + µi + it (10)

The model is first differenced to get rid of the individual effect:

∆yit = ρ∆yit−1 + β > ∆xit + ∆it (11)

Least squares are inconsistent because ∆it is correlated with ∆yit−1 . yit−2 is a valid, but
weak instrument (see Anderson and Hsiao 1981). The gmm estimator uses the fact that the
number of valid instruments is growing with t:

• t = 3: y1 ,

• t = 4: y1 , y2 ,

• t = 5: y1 , y2 , y3

For individual i, the matrix of instruments is then:

 
y1 0 0 0 0 0 ... 0 0 0 0 xi3

 0 y1 y2 0 0 0 ... 0 0 0 0 xi4 

0 0 0 y1 y2 y3 ... 0 0 0 0 xi5
 
Wi =   (12)

.. .. .. .. .. .. .. .. .. .. .. .. 
. . . . . . . . . . . .
 
 
0 0 0 0 ... ... . . . y1 y2 . . . yt−2 xiT −2

The moment conditions are: ni=1 Wi> ei (β) where ei (β) is the vector of residuals for individual
P

i. The gmm estimator minimize:

n n
! !
>
Wi> ei (β)
X X
ei (β) Wi A (13)
i=1 i=1
22 Panel Data Econometrics in R: The plm Package

where A is the weighting matrix of the moments.

One-step estimators are computed using a known weighting matrix. For the model in first
differences, one uses:

n
!−1
Wi> H (1) Wi
X
(1)
A = (14)
i=1

with:
 
2 −1 0 . . . 0

 −1 2 −1 . . . 0 

(1) > 0 −1 2 . . . 0
 
H =d d=  (15)

.. .. .. .. .. 
. . . . .
 
 
0 0 0 −1 2
(2) Pn (1) (1)> (1)
Two-steps estimators are obtained using Hi = i=1 ei ei where ei are the residuals of
the one step estimate.
Blundell and Bond (1998) show that with weak hypothesis on the data generating process,
supplementary moment conditions exist for the equation in level:

yit = γyit−1 + µi + ηit

More precisely, they show that ∆yit−2 = yit−2 − yit−3 is a valid instrument. The estimator is
obtained using the residual vector in difference and in level:

e+
i = (∆ei , ei )

and the matrix of augmented moments:

 
Zi 0 0 ... 0
 0 ∆yi2 0 ... 0 
Zi+ = 
 
0 0 ∆yi3 ... 0

 
0 0 0 . . . ∆yiT −1
The moment conditions are then

n
!!> n n n
ēi (β)
Zi+>
X X X X
= yi1 ēi3 , yi1 ēi4 , yi2 ēi4 , . . . ,
i=1
ei (β) i=1 i=1 i=1
n
X n
X n
X n X
X T
yi1 ēiT , yi2 ēiT , . . . , yiT −2 ēiT , xit ēit
i=1 i=1 i=1 i=1 t=3
n n n
!>
X X X
ei3 ∆yi2 , ei4 ∆yi3 , . . . , eiT ∆yiT −1
i=1 i=1 i=1

The gmm estimator is provided by the pgmm function. It’s main argument is a dynformula
which describes the variables of the model and the lag structure.
 Yves Croissant, Giovanni Millo 23

In a gmm estimation, there are “normal” instruments and “gmm” instruments. gmm instru-
ments are indicated in the second part of the formula. By default, all the variables of the
model that are not used as gmm instruments are used as normal instruments, with the same
lag structure; “normal” instruments may also be indicated in the third part of the formula.
The effect argument is either NULL, "individual" (the default), or "twoways". In the first
case, the model is estimated in levels. In the second case, the model is estimated in first
differences to get rid of the individuals effects. In the last case, the model is estimated in first
differences and time dummies are included.
The model argument specifies whether a one-step or a two-steps model is required ("onestep"
or "twosteps").
The following example is from Arellano and Bond (1991). Employment is explained by past
values of employment (two lags), current and first lag of wages and output and current value
of capital.

R> emp.gmm <- pgmm(log(emp)~lag(log(emp), 1:2)+lag(log(wage), 0:1)+log(capital)+

+ lag(log(output), 0:1)|lag(log(emp), 2:99),
+ data = EmplUK, effect = "twoways", model = "twosteps")
R> summary(emp.gmm)

Twoways effects Two steps model

Call:
pgmm(formula = log(emp) ~ lag(log(emp), 1:2) + lag(log(wage),
0:1) + log(capital) + lag(log(output), 0:1) | lag(log(emp),
2:99), data = EmplUK, effect = "twoways", model = "twosteps")

Unbalanced Panel: n = 140, T = 7-9, N = 1031

Number of Observations Used: 611

Residuals:
Min. 1st Qu. Median Mean 3rd Qu. Max.
-0.6190677 -0.0255683 0.0000000 -0.0001339 0.0332013 0.6410272

Coefficients:
Estimate Std. Error z-value Pr(>|z|)
lag(log(emp), 1:2)1 0.474151 0.185398 2.5575 0.0105437 *
lag(log(emp), 1:2)2 -0.052967 0.051749 -1.0235 0.3060506
lag(log(wage), 0:1)0 -0.513205 0.145565 -3.5256 0.0004225 ***
lag(log(wage), 0:1)1 0.224640 0.141950 1.5825 0.1135279
log(capital) 0.292723 0.062627 4.6741 2.953e-06 ***
lag(log(output), 0:1)0 0.609775 0.156263 3.9022 9.530e-05 ***
lag(log(output), 0:1)1 -0.446373 0.217302 -2.0542 0.0399605 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
24 Panel Data Econometrics in R: The plm Package

Sargan test: chisq(25) = 30.11247 (p-value = 0.22011)

Autocorrelation test (1): normal = -1.53845 (p-value = 0.12394)
Autocorrelation test (2): normal = -0.2796829 (p-value = 0.77972)
Wald test for coefficients: chisq(7) = 142.0353 (p-value = < 2.22e-16)
Wald test for time dummies: chisq(6) = 16.97046 (p-value = 0.0093924)

The following example is from Blundell and Bond (1998). The “sys” estimator is obtained
using transformation = "ld" for level and difference. The robust argument of the summary
method enables to use the robust covariance matrix proposed by Windmeijer (2005).

R> z2 <- pgmm(log(emp) ~ lag(log(emp), 1)+ lag(log(wage), 0:1) +

+ lag(log(capital), 0:1) | lag(log(emp), 2:99) +
+ lag(log(wage), 2:99) + lag(log(capital), 2:99),
+ data = EmplUK, effect = "twoways", model = "onestep",
+ transformation = "ld")
R> summary(z2, robust = TRUE)

Twoways effects One step model

Call:
pgmm(formula = log(emp) ~ lag(log(emp), 1) + lag(log(wage), 0:1) +
lag(log(capital), 0:1) | lag(log(emp), 2:99) + lag(log(wage),
2:99) + lag(log(capital), 2:99), data = EmplUK, effect = "twoways",
model = "onestep", transformation = "ld")

Unbalanced Panel: n = 140, T = 7-9, N = 1031

Number of Observations Used: 1642

Residuals:
Min. 1st Qu. Median Mean 3rd Qu. Max.
-0.7530341 -0.0369030 0.0000000 0.0002882 0.0466069 0.6001503

Coefficients:
Estimate Std. Error z-value Pr(>|z|)
lag(log(emp), 1) 0.935605 0.026295 35.5810 < 2.2e-16 ***
lag(log(wage), 0:1)0 -0.630976 0.118054 -5.3448 9.050e-08 ***
lag(log(wage), 0:1)1 0.482620 0.136887 3.5257 0.0004224 ***
lag(log(capital), 0:1)0 0.483930 0.053867 8.9838 < 2.2e-16 ***
lag(log(capital), 0:1)1 -0.424393 0.058479 -7.2572 3.952e-13 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Sargan test: chisq(100) = 118.763 (p-value = 0.097096)

Autocorrelation test (1): normal = -4.808434 (p-value = 1.5212e-06)
Autocorrelation test (2): normal = -0.2800133 (p-value = 0.77947)
 Yves Croissant, Giovanni Millo 25

Wald test for coefficients: chisq(5) = 11174.82 (p-value = < 2.22e-16)

Wald test for time dummies: chisq(7) = 14.71138 (p-value = 0.039882)

5.5. General FGLS models

General fgls estimators are based on a two-step estimation process: first an ols model is
estimated, then its residuals ûit are used to estimate an error covariance matrix more general
than the random effects one for use in a feasible-gls analysis. Formally, the estimated error
covariance matrix is V̂ = In ⊗ Ω̂, with
n
X ûit û>
it
Ω̂ =
i=1
n

(see Wooldridge 2002, 10.4.3 and 10.5.5).

This framework allows the error covariance structure inside every group (if effect="individual")
of observations to be fully unrestricted and is therefore robust against any type of intragroup
heteroskedasticity and serial correlation. This structure, by converse, is assumed identical
across groups and thus general fgls is inefficient under groupwise heteroskedasticity. Cross-
sectional correlation is excluded a priori.
Moreover, the number of variance parameters to be estimated with N = n × T data points is
T (T + 1)/2, which makes these estimators particularly suited for situations where n >> T ,
as e.g. in labour or household income surveys, while problematic for “long” panels, where V̂
tends to become singular and standard errors therefore become biased downwards.
In a pooled time series context (effect="time"), symmetrically, this estimator is able to
account for arbitrary cross-sectional correlation, provided that the latter is time-invariant
(see Greene 2003, 13.9.1–2, pp.321–2). In this case serial correlation has to be assumed away
and the estimator is consistent with respect to the time dimension, keeping n fixed.
The function pggls estimates general fgls models, with either fixed of “random” effects6 .
The “random effect” general fgls is estimated by:

R> zz <- pggls(log(emp)~log(wage)+log(capital), data=EmplUK, model="pooling")

R> summary(zz)

NA

Call:
pggls(formula = log(emp) ~ log(wage) + log(capital), data = EmplUK,
model = "pooling")

Unbalanced Panel: n = 140, T = 7-9, N = 1031

Residuals:
Min. 1st Qu. Median Mean 3rd Qu. Max.
6
The “random effect” is better termed “general fgls” model, as in fact it does not have a proper random
effects structure, but we keep this terminology for general language consistency.
26 Panel Data Econometrics in R: The plm Package

-1.80696 -0.36552 0.06181 0.03230 0.44279 1.58719

Coefficients:
Estimate Std. Error z-value Pr(>|z|)
(Intercept) 2.023480 0.158468 12.7690 < 2.2e-16 ***
log(wage) -0.232329 0.048001 -4.8401 1.298e-06 ***
log(capital) 0.610484 0.017434 35.0174 < 2.2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Total Sum of Squares: 1853.6
Residual Sum of Squares: 402.55
Multiple R-squared: 0.78283

The fixed effects pggls (see Wooldridge 2002, p.276) is based on the estimation of a within
model in the first step; the rest follows as above. It is estimated by:

R> zz <- pggls(log(emp)~log(wage)+log(capital), data=EmplUK, model="within")

The pggls function is similar to plm in many respects. An exception is that the estimate
of the group covariance matrix of errors (zz$sigma, a matrix, not shown) is reported in the
model objects instead of the usual estimated variances of the two error components.

6. Tests
As sketched in Section 2, specification testing in panel models involves essentially testing
for poolability, for individual or time unobserved effects and for correlation between these
latter and the regressors (Hausman-type tests). As for the other usual diagnostic checks, we
provide a suite of serial correlation tests, while not touching on the issue of heteroskedasticity
testing. Instead, we provide heteroskedasticity-robust covariance estimators, to be described
in Subsection 6.7.

6.1. Tests of poolability

pooltest tests the hypothesis that the same coefficients apply to each individual. It is a
standard F test, based on the comparison of a model obtained for the full sample and a model
based on the estimation of an equation for each individual. The first argument of pooltest is
a plm object. The second argument is a pvcm object obtained with model=within. If the first
argument is a pooling model, the test applies to all the coefficients (including the intercepts),
if it is a within model, different intercepts are assumed.
To test the hypothesis that all the coefficients in the Grunfeld example, excluding the inter-
cepts, are equal, we use :

R> znp <- pvcm(inv~value+capital, data=Grunfeld, model="within")

R> zplm <- plm(inv~value+capital, data=Grunfeld, model="within")
R> pooltest(zplm,znp)
 Yves Croissant, Giovanni Millo 27

F statistic

data: inv ~ value + capital

F = 5.7805, df1 = 18, df2 = 170, p-value = 1.219e-10
alternative hypothesis: unstability

The same test can be computed using a formula as first argument of the pooltest function:

R> pooltest(inv~value+capital, data=Grunfeld, model="within")

6.2. Tests for individual and time effects

plmtest implements Lagrange multiplier tests of individual or/and time effects based on the
results of the pooling model. Its main argument is a plm object (the result of a pooling model)
or a formula.
Two additional arguments can be added to indicate the kind of test to be computed. The
argument type is one of:

• "honda": Honda (1985), the default value,

• "bp": Breusch and Pagan (1980),

• "kw": King and Wu (1997)7 ,

• "ghm": Gourieroux, Holly, and Monfort (1982).

The effects tested are indicated with the effect argument (one of "individual", "time" or
"twoways"). The test statistics implemented are also suitable for unbalanced panels.8
To test the presence of individual and time effects in the Grunfeld example, using the Gourier-
oux et al. (1982) test, we use:

R> g <- plm(inv ~ value + capital, data=Grunfeld, model="pooling")

R> plmtest(g, effect="twoways", type="ghm")

Lagrange Multiplier Test - two-ways effects (Gourieroux, Holly and

Monfort) for balanced panels

data: inv ~ value + capital

chibarsq = 798.16, df0 = 0.00, df1 = 1.00, df2 = 2.00, w0 = 0.25, w1 =
0.50, w2 = 0.25, p-value < 2.2e-16
alternative hypothesis: significant effects
7
NB: Oneway King-Wu ("kw") statistics ("individual" and "time") coincide with the respective Honda
statistics ("honda"); however, the twoway statistics of "kw" and "honda" differ.
8
The "bp" test for unbalanced panels was derived in Baltagi and Li (1990), the "kw" test for unbalanced
panels in Baltagi, Chang, and Li (1998). The "ghm" test and the "kw" test were extended to two–way effects
in Baltagi, Chang, and Li (1992). For a concise overview of all these statistics see Baltagi (2013, Sec. 4.2, pp.
68–76 (for balanced panels) and Sec. 9.5, pp. 200–203 (for unbalanced panels)).
28 Panel Data Econometrics in R: The plm Package

or

R> plmtest(inv~value+capital, data=Grunfeld, effect="twoways", type="ghm")

pFtest computes F tests of effects based on the comparison of the within and the pooling
models. Its main arguments are either two plm objects (the results of a pooling and a within
model) or a formula.

R> gw <- plm(inv ~ value + capital,data=Grunfeld,effect="twoways",model="within")

R> gp <- plm(inv ~ value + capital,data=Grunfeld,model="pooling")
R> pFtest(gw,gp)

F test for twoways effects

data: inv ~ value + capital

F = 17.403, df1 = 28, df2 = 169, p-value < 2.2e-16
alternative hypothesis: significant effects

R> pFtest(inv~value+capital, data=Grunfeld,effect="twoways")

6.3. Hausman test

phtest computes the Hausman test which is based on the comparison of two sets of estimates
(see Hausman 1978). Its main arguments are two panelmodel objects or a formula. A classical
application of the Hausman test for panel data is to compare the fixed and the random effects
models:

R> gw <- plm(inv~value+capital, data=Grunfeld, model="within")

R> gr <- plm(inv~value+capital, data=Grunfeld, model="random")
R> phtest(gw, gr)

Hausman Test

data: inv ~ value + capital

chisq = 2.3304, df = 2, p-value = 0.3119
alternative hypothesis: one model is inconsistent

6.4. Tests of serial correlation

A model with individual effects has composite errors that are serially correlated by definition.
The presence of the time-invariant error component9 gives rise to serial correlation which does
not die out over time, thus standard tests applied on pooled data always end up rejecting
9
Here we treat fixed and random effects alike, as components of the error term, according with the modern
approach in econometrics (see Wooldridge 2002, 2010).
 Yves Croissant, Giovanni Millo 29

the null of spherical residuals10 . There may also be serial correlation of the “usual” kind in
the idiosyncratic error terms, e.g. as an AR(1) process. By “testing for serial correlation” we
mean testing for this latter kind of dependence.
For these reasons, the subjects of testing for individual error components and for serially
correlated idiosyncratic errors are closely related. In particular, simple (marginal) tests for one
direction of departure from the hypothesis of spherical errors usually have power against the
other one: in case it is present, they are substantially biased towards rejection. Joint tests are
correctly sized and have power against both directions, but usually do not give any information
about which one actually caused rejection. Conditional tests for serial correlation that take
into account the error components are correctly sized under presence of both departures from
sphericity and have power only against the alternative of interest. While most powerful if
correctly specified, the latter, based on the likelihood framework, are crucially dependent on
normality and homoskedasticity of the errors.
In plm we provide a number of joint, marginal and conditional ml-based tests, plus some semi-
parametric alternatives which are robust vs. heteroskedasticity and free from distributional
assumptions.

Unobserved effects test

The unobserved effects test à la Wooldridge (see Wooldridge 2002, 10.4.4), is a semiparametric
test for the null hypothesis that σµ2 = 0, i.e. that there are no unobserved effects in the
residuals. Given that under the null the covariance matrix of the residuals for each individual
is diagonal, the test statistic is based on the average of elements in the upper (or lower)
P −1 PT
triangle of its estimate, diagonal excluded: n−1/2 ni=1 Tt=1
P
s=t+1 ûit ûis (where û are the
pooled ols residuals), which must be “statistically close” to zero under the null, scaled by its
standard deviation: Pn PT −1 PT
s=t+1 ûit ûis
W = Pn i=1 t=1
PT −1 PT
[ i=1 ( t=1 s=t+1 ûit ûis )2 ]1/2

This test is (n-) asymptotically distributed as a standard Normal regardless of the distribution
of the errors. It does also not rely on homoskedasticity.
It has power both against the standard random effects specification, where the unobserved
effects are constant within every group, as well as against any kind of serial correlation. As
such, it “nests” both random effects and serial correlation tests, trading some power against
more specific alternatives in exchange for robustness.
While not rejecting the null favours the use of pooled ols, rejection may follow from serial
correlation of different kinds, and in particular, quoting Wooldridge (2002), “should not be
interpreted as implying that the random effects error structure must be true”.
Below, the test is applied to the data and model in Munnell (1990):

R> pwtest(log(gsp)~log(pcap)+log(pc)+log(emp)+unemp, data=Produc)

Wooldridge's test for unobserved individual effects

10
Neglecting time effects may also lead to serial correlation in residuals (as observed in Wooldridge 2002,
10.4.1).
30 Panel Data Econometrics in R: The plm Package

data: formula
z = 3.9383, p-value = 8.207e-05
alternative hypothesis: unobserved effect

Locally robust tests for serial correlation or random effects

The presence of random effects may affect tests for residual serial correlation, and the opposite.
One solution is to use a joint test, which has power against both alternatives. A joint LM
test for random effects and serial correlation under normality and homoskedasticity of the
idiosyncratic errors has been derived by Baltagi and Li (1991) and Baltagi and Li (1995) and
is implemented as an option in pbsytest:

R> pbsytest(log(gsp)~log(pcap)+log(pc)+log(emp)+unemp, data=Produc, test="j")

Baltagi and Li AR-RE joint test - balanced panel

data: formula
chisq = 4187.6, df = 2, p-value < 2.2e-16
alternative hypothesis: AR(1) errors or random effects

Rejection of the joint test, though, gives no information on the direction of the departure
from the null hypothesis, i.e.: is rejection due to the presence of serial correlation, of random
effects or of both?
Bera, Sosa-Escudero, and Yoon (2001) derive locally robust tests both for individual random
effects and for first-order serial correlation in residuals as “corrected” versions of the standard
LM test (see plmtest). While still dependent on normality and homoskedasticity, these
are robust to local departures from the hypotheses of, respectively, no serial correlation or
no random effects. The authors observe that, although suboptimal, these tests may help
detecting the right direction of the departure from the null, thus complementing the use of
joint tests. Moreover, being based on pooled ols residuals, the BSY tests are computationally
far less demanding than likelihood-based conditional tests.
On the other hand, the statistical properties of these “locally corrected” tests are inferior
to those of the non-corrected counterparts when the latter are correctly specified. If there
is no serial correlation, then the optimal test for random effects is the likelihood-based LM
test of Breusch and Godfrey (with refinements by Honda, see plmtest), while if there are no
random effects the optimal test for serial correlation is, again, Breusch-Godfrey’s test11 . If the
presence of a random effect is taken for granted, then the optimal test for serial correlation
is the likelihood-based conditional LM test of Baltagi and Li (1995) (see pbltest).
The serial correlation version is the default:

R> pbsytest(log(gsp)~log(pcap)+log(pc)+log(emp)+unemp, data=Produc)

Bera, Sosa-Escudero and Yoon locally robust test - balanced panel

11
LM3 in Baltagi and Li (1995).
 Yves Croissant, Giovanni Millo 31

data: formula
chisq = 52.636, df = 1, p-value = 4.015e-13
alternative hypothesis: AR(1) errors sub random effects

The BSY test for random effects is implemented in the one-sided version12 , which takes heed
that the variance of the random effect must be non-negative:

R> pbsytest(log(gsp)~log(pcap)+log(pc)+log(emp)+unemp, data=Produc, test="re")

Bera, Sosa-Escudero and Yoon locally robust test (one-sided) -

balanced panel

data: formula
z = 57.914, p-value < 2.2e-16
alternative hypothesis: random effects sub AR(1) errors

Conditional LM test for AR(1) or MA(1) errors under random effects

Baltagi and Li (1991) and Baltagi and Li (1995) derive a Lagrange multiplier test for serial
correlation in the idiosyncratic component of the errors under (normal, heteroskedastic) ran-
dom effects. Under the null of serially uncorrelated errors, the test turns out to be identical
for both the alternative of AR(1) and MA(1) processes. One- and two-sided versions are
provided, the one-sided having power against positive serial correlation only. The two-sided
is the default, while for the other one must specify the alternative option to onesided:

R> pbltest(log(gsp)~log(pcap)+log(pc)+log(emp)+unemp,
+ data=Produc, alternative="onesided")

Baltagi and Li one-sided LM test

data: log(gsp) ~ log(pcap) + log(pc) + log(emp) + unemp

z = 21.69, p-value < 2.2e-16
alternative hypothesis: AR(1)/MA(1) errors in RE panel model

As usual, the LM test statistic is based on residuals from the maximum likelihood estimate of
the restricted model (random effects with serially uncorrelated errors). In this case, though,
the restricted model cannot be estimated by ols any more, therefore the testing function
depends on lme() in the nlme package for estimation of a random effects model by maximum
likelihood. For this reason, the test is applicable only to balanced panels.
No test has been implemented to date for the symmetric hypothesis of no random effects in
a model with errors following an AR(1) process, but an asymptotically equivalent likelihood
ratio test is available in the nlme package (see Section 7)..

General serial correlation tests

A general testing procedure for serial correlation in fixed effects (fe), random effects (re) and
pooled-ols panel models alike can be based on considerations in (Wooldridge 2002, 10.7.2).
12
Corresponding to RSOµ∗ in the original paper.
32 Panel Data Econometrics in R: The plm Package

Recall that plm model objects are the result of ols estimation performed on “demeaned” data,
where, in the case of individual effects (else symmetric), this means time-demeaning for the
fe (within) model, quasi-time-demeaning for the re (random) model and original data, with
no demeaning at all, for the pooled ols (pooling) model (see Section 3).
For the random effects model, Wooldridge (2002) observes that under the null of homoskedas-
ticity and no serial correlation in the idiosyncratic errors, the residuals from the quasi-
demeaned regression must be spherical as well. Else, as the individual effects are wiped
out in the demeaning, any remaining serial correlation must be due to the idiosyncratic com-
ponent. Hence, a simple way of testing for serial correlation is to apply a standard serial
correlation test to the quasi-demeaned model. The same applies in a pooled model, w.r.t. the
original data.
The fe case needs some qualification. It is well-known that if the original model’s errors are
uncorrelated then fe residuals are negatively serially correlated, with cor(ûit , ûis ) = −1/(T −
1) for each t, s (see Wooldridge 2002, 10.5.4). This correlation clearly dies out as T increases,
so this kind of AR test is applicable to within model objects only for T “sufficiently large”13 .
On the converse, in short panels the test gets severely biased towards rejection (or, as the
induced correlation is negative, towards acceptance in the case of the one-sided DW test with
alternative="greater"). See below for a serial correlation test applicable to “short” fe
panel models.
plm objects retain the “demeaned” data, so the procedure is straightforward for them. The
wrapper functions pbgtest and pdwtest re-estimate the relevant quasi-demeaned model by
ols and apply, respectively, standard Breusch-Godfrey and Durbin-Watson tests from package
lmtest:

R> pbgtest(grun.fe, order = 2)

Breusch-Godfrey/Wooldridge test for serial correlation in panel models

data: inv ~ value + capital

chisq = 42.587, df = 2, p-value = 5.655e-10
alternative hypothesis: serial correlation in idiosyncratic errors

The tests share the features of their ols counterparts, in particular the pbgtest allows testing
for higher-order serial correlation, which might turn useful, e.g., on quarterly data. Analo-
gously, from the point of view of software, as the functions are simple wrappers towards
bgtest and dwtest, all arguments from the latter two apply and may be passed on through
the ‘. . . ’ argument.

Wooldridge’s test for serial correlation in “short” FE panels

For the reasons reported above, under the null of no serial correlation in the errors, the
it , ˆis ) = −1/(T − 1)
residuals of a fe model must be negatively serially correlated, with cor(ˆ
13
Baltagi and Li derive a basically analogous T-asymptotic test for first-order serial correlation in a fe panel
model as a Breusch-Godfrey LM test on within residuals (see Baltagi and Li 1995, par. 2.3 and formula
12). They also observe that the test on within residuals can be used for testing on the re model, as “the
within transformation [time-demeaning, in our terminology] wipes out the individual effects, whether fixed
or random”. Generalizing the Durbin-Watson test to fe models by applying it to fixed effects residuals is
documented in Bhargava, Franzini, and Narendranathan (1982).
 Yves Croissant, Giovanni Millo 33

for each t, s. Wooldridge suggests basing a test for this null hypothesis on a pooled regression
of fe residuals on themselves, lagged one period:

ˆi,t = α + δˆ
i,t−1 + ηi,t

Rejecting the restriction δ = −1/(T − 1) makes us conclude against the original null of no
serial correlation.
The building blocks available in plm make it easy to construct a function carrying out this
procedure: first the fe model is estimated and the residuals retrieved, then they are lagged
and a pooling AR(1) model is estimated. The test statistic is obtained by applying the above
restriction on δ and supplying a heteroskedasticity- and autocorrelation-consistent covariance
matrix (vcovHC with the appropriate options, in particular method="arellano")14 .

R> pwartest(log(emp) ~ log(wage) + log(capital), data=EmplUK)

Wooldridge's test for serial correlation in FE panels

data: plm.model
F = 312.3, df1 = 1, df2 = 889, p-value < 2.2e-16
alternative hypothesis: serial correlation

The test is applicable to any fe panel model, and in particular to “short” panels with small
T and large n.

Wooldridge’s first-difference-based test

In the context of the first difference model, Wooldridge (2002, 10.6.3) proposes a serial corre-
lation test that can also be seen as a specification test to choose the most efficient estimator
between fixed effects (within) and first difference (fd).
The starting point is the observation that if the idiosyncratic errors of the original model
uit are uncorrelated, the errors of the (first) differenced model15 eit ≡ uit − ui,t−1 will be
correlated, with cor(eit , ei,t−1 ) = −0.5, while any time-invariant effect, “fixed” or “random”,
is wiped out in the differencing. So a serial correlation test for models with individual effects
of any kind can be based on estimating the model

ûi,t = δ ûi,t−1 + ηi,t

and testing the restriction δ = −0.5, corresponding to the null of no serial correlation. Drukker
(2003) provides Monte Carlo evidence of the good empirical properties of the test.
On the other extreme (see Wooldridge 2002, 10.6.1), if the differenced errors eit are uncorre-
lated, as by definition uit = ui,t−1 + eit , then uit is a random walk. In this latter case, the
most efficient estimator is the first difference (fd) one; in the former case, it is the fixed effects
one (within).
The function pwfdtest allows testing either hypothesis: the default behaviour h0="fd" is to
test for serial correlation in first-differenced errors:
14
see Subsection 6.7.
15
Here, eit for notational simplicity (and as in Wooldridge): equivalent to ∆it in the general notation of
the paper.
34 Panel Data Econometrics in R: The plm Package

R> pwfdtest(log(emp) ~ log(wage) + log(capital), data=EmplUK)

Wooldridge's first-difference test for serial correlation in panels

data: plm.model
F = 0.9316, df1 = 1, df2 = 749, p-value = 0.3348
alternative hypothesis: serial correlation in differenced errors

while specifying h0="fe" the null hypothesis becomes no serial correlation in original errors,
which is similar to the pwartest.

R> pwfdtest(log(emp) ~ log(wage) + log(capital), data=EmplUK, h0="fe")

Wooldridge's first-difference test for serial correlation in panels

data: plm.model
F = 131.01, df1 = 1, df2 = 749, p-value < 2.2e-16
alternative hypothesis: serial correlation in original errors

Not rejecting one of the two is evidence in favour of using the estimator corresponding to
h0. Should the truth lie in the middle (both rejected), whichever estimator is chosen will
have serially correlated errors: therefore it will be advisable to use the autocorrelation-robust
covariance estimators from the Subsection 6.7 in inference.

6.5. Tests for cross-sectional dependence

Next to the more familiar issue of serial correlation, over the last years a growing body of
literature has been dealing with cross-sectional dependence (henceforth: xsd) in panels, which
can arise, e.g., if individuals respond to common shocks (as in the literature on factor models)
or if spatial diffusion processes are present, relating individuals in a way depending on a
measure of distance (spatial models).
The subject is huge, and here we touch only some general aspects of misspecification testing
and valid inference. If xsd is present, the consequence is, at a minimum, inefficiency of the
usual estimators and invalid inference when using the standard covariance matrix16 .The plan
is to have in plm both misspecification tests to detect xsd and robust covariance matrices to
perform valid inference in its presence, like in the serial dependence case. For now, though,
only misspecification tests are included.

CD and LM-type tests for global cross-sectional dependence

The function pcdtest implements a family of xsd tests which can be applied in different
settings, ranging from those where T grows large with n fixed to “short” panels with a big n
dimension and a few time periods. All are based on (transformations of–) the product-moment
correlation coefficient of a model’s residuals, defined as
16
This is the case, e.g., if in an unobserved effects model when xsd is due to an unobservable factor structure,
with factors that are uncorrelated with the regressors. In this case the within or random estimators are still
consistent, although inefficient (see De Hoyos and Sarafidis 2006).
 Yves Croissant, Giovanni Millo 35

PT
t=1 ûit ûjt
ρ̂ij = PT
( t=1 ûit )1/2 ( Tt=1 û2jt )1/2
2 P

i.e., as averages over the time dimension of pairwise correlation coefficients for each pair of
cross-sectional units.
The Breusch-Pagan (Breusch and Pagan 1980) LM test, based on the squares of ρij , is valid
for T → ∞ with n fixed; defined as

n−1
X n
X
LM = Tij ρ̂2ij
i=1 j=i+1

where in the case of an unbalanced panel only pairwise complete observations are considered,
and Tij = min(Ti , Tj ) with Ti being the number of observations for individual i; else, if the
panel is balanced, Tij = T for each i, j. The test is distributed as χ2n(n−1)/2 . It is inappropriate
whenever the n dimension is “large”. A scaled version, applicable also if T → ∞ and then
n → ∞ (as in some pooled time series contexts), is defined as

n−1 n q
s
1 X X
SCLM = ( Tij ρ̂2ij )
n(n − 1) i=1 j=i+1

and distributed as a standard Normal.

Pesaran’s (Pesaran 2004, 2015) CD test

n−1 n q
s
2 X X
CD = ( Tij ρ̂ij )
n(n − 1) i=1 j=i+1

based on ρij without squaring (also distributed as a standard Normal) is appropriate both in
n– and in T –asymptotic settings. It has remarkable properties in samples of any practically
relevant size and is robust to a variety of settings. The only big drawback is that the test
loses power against the alternative of cross-sectional dependence if the latter is due to a factor
structure with factor loadings averaging zero, that is, some units react positively to common
shocks, others negatively.
The default version of the test is "cd". These tests are originally meant to use the residuals
of separate estimation of one time-series regression for each cross-sectional unit, so this is the
default behaviour of pcdtest.

R> pcdtest(inv~value+capital, data=Grunfeld)

Pesaran CD test for cross-sectional dependence in panels

data: inv ~ value + capital

z = 5.3401, p-value = 9.292e-08
alternative hypothesis: cross-sectional dependence
36 Panel Data Econometrics in R: The plm Package

If a different model specification (within, random, ...) is assumed consistent, one can resort
to its residuals for testing17 by specifying the relevant model type. The main argument of
this function may be either a model of class panelmodel or a formula and a data.frame; in
the second case, unless model is set to NULL, all usual parameters relative to the estimation
of a plm model may be passed on. The test is compatible with any consistent panelmodel
for the data at hand, with any specification of effect. E.g., specifying effect="time" or
effect="twoways" allows to test for residual cross-sectional dependence after the introduction
of time fixed effects to account for common shocks.

R> pcdtest(inv~value+capital, data=Grunfeld, model="within")

Pesaran CD test for cross-sectional dependence in panels

data: inv ~ value + capital

z = 4.6612, p-value = 3.144e-06
alternative hypothesis: cross-sectional dependence

If the time dimension is insufficient and model=NULL, the function defaults to estimation of a
within model and issues a warning.

CD(p) test for local cross-sectional dependence

A local variant of the CD test, called CD(p) test (Pesaran 2004), takes into account an
appropriate subset of neighbouring cross-sectional units to check the null of no xsd against
the alternative of local xsd, i.e. dependence between neighbours only. To do so, the pairs
of neighbouring units are selected by means of a binary proximity matrix like those used in
spatial models. In the original paper, a regular ordering of observations is assumed, so that
the m-th cross-sectional observation is a neighbour to the (m − 1)-th and to the (m + 1)-th.
Extending the CD(p) test to irregular lattices, we employ the binary proximity matrix as a
selector for discarding the correlation coefficients relative to pairs of observations that are not
neighbours in computing the CD statistic. The test is then defined as

s n−1 n
1 X X q
CD = Pn−1 Pn ( [w(p)]ij Tij ρ̂ij )
i=1 j=i+1 w(p)ij i=1 j=i+1

where [w(p)]ij is the (i, j)-th element of the p-th order proximity matrix, so that if h, k are
not neighbours, [w(p)]hk = 0 and ρ̂hk gets “killed”; this is easily seen to reduce to formula
(14) in Pesaran (Pesaran 2004) for the special case considered in that paper. The same can
be applied to the LM and SCLM tests.
Therefore, the local version of either test can be computed supplying an n × n matrix (of any
kind coercible to logical), providing information on whether any pair of observations are
neighbours or not, to the w argument. If w is supplied, only neighbouring pairs will be used in
computing the test; else, w will default to NULL and all observations will be used. The matrix
17
This is also the only solution when the time dimension’s length is insufficient for estimating the heteroge-
neous model.
 Yves Croissant, Giovanni Millo 37

needs not really be binary, so commonly used “row-standardized” matrices can be employed
as well: it is enough that neighbouring pairs correspond to nonzero elements in w 18 .

6.6. Unit root tests

Preliminary results
We consider the following model:

pi
X
yit = δyit−1 + θi ∆yit−L + αmi dmt + it
L=1

The unit root hypothesis is ρ = 1. The model can be rewritten in difference:

pi
X
∆yit = ρyit−1 + θi ∆yit−L + αmi dmt + it
L=1

So that the unit-root hypothesis is now ρ = 0.

Some of the unit-root tests for panel data are based on preliminary results obtained by running
the above Augmented Dickey-Fuller (ADF) regression.
First, we have to determine the optimal number of lags pi for each time-series. Several
possibilities are available. They all have in common that the maximum number of lags have
to be chosen first. Then, pi can be chosen using:

• the Swartz information criteria (SIC),

• the Akaike information criteria (AIC),

• the Hall method, which consist in removing the higher lags while it is not significant.

The ADF regression is run on T − pi − 1 observations for each individual, so that the total
number of observations is n × T̃ where T̃ = T − pi − 1
p̄ is the average number of lags. Call ei the vector of residuals.
Estimate the variance of the i as:
PT 2
t=pi +1 eit
σ̂2i =
dfi

Levin-Lin-Chu model
Then, compute artificial regressions of ∆yit and yit−1 on ∆yit−L and dmt and get the two
vectors of residuals zit and vit .
18
The very comprehensive package spdep for spatial dependence analysis (see Bivand 2008) contains features
for creating, lagging and manipulating neighbour list objects of class nb, that can be readily converted to and
from proximity matrices by means of the nb2mat function. Higher orders of the CD(p) test can be obtained
lagging the corresponding nbs through nblag.
38 Panel Data Econometrics in R: The plm Package

Standardize these two residuals and run the pooled regression of zit /σ̂i on vit /σ̂i to get ρ̂, its
standard deviation σ̂(ρ̂) and the t-statistic tρ̂ = ρ̂/σ̂(ρ̂).
Compute the long run variance of yi :
 
T K̄ T
2 1 X 2
X 1 X
σ̂yi = ∆yit +2 wK̄L  ∆yit ∆yit−L 
T − 1 t=2 L=1
T − 1 t=2+L

Define s̄i as the ratio of the long and short term variance and s̄ the mean for all the individuals
of the sample

σ̂yi
si =
σ̂i
Pn
i=1 si
s̄ =
n

tρ − nT̄ s̄σ̂˜−2 σ̂(ρ̂)µ∗mT̃

t∗ρ = ∗
σm T̃

follows a normal distribution under the null hypothesis of stationarity. µ∗mT̃ and σm
∗
T̃
are
given in table 2 of the original paper and are also available in the package.

Im, Pesaran and Shin test

This test does not require that ρ is the same for all the individuals. The null hypothesis is
still that all the series have an unit root, but the alternative is that some may have a unit
root and others have different values of ρi < 0.
The test is based on the average of the student statistic of the ρ obtained for each individual:

n
1X
t̄ = tρi
n i=1

The statistic is then:

√
n (t̄ − E(t̄))
z= p
V (t̄)

µ∗mT̃ and σm
∗
T̃
are given in table 2 of the original paper and are also available in the package.

6.7. Robust covariance matrix estimation

Robust estimators of the covariance matrix of coefficients are provided, mostly for use in
Wald-type tests. vcovHC estimates three “flavours” of White’s heteroskedasticity-consistent
covariance matrix19 (known as the sandwich estimator). Interestingly, in the context of panel
data the most general version also proves consistent vs. serial correlation.
19
See White (1980) and White (1984).
 Yves Croissant, Giovanni Millo 39

All types assume no correlation between errors of different groups while allowing for het-
eroskedasticity across groups, so that the full covariance matrix of errors is V = In ⊗ Ωi ; i =
1, .., n. As for the intragroup error covariance matrix of every single group of observations,
"white1" allows for general heteroskedasticity but no serial correlation, i.e.
 
2
σi1 ... ... 0

 0 2
.. 
σi2 . 
Ωi =  . (16)
 
 . .. 
 . . 0  
0 2
. . . . . . σiT

while "white2" is "white1" restricted to a common variance inside every group, estimated
as σi2 = Tt=1 û2it /T , so that Ωi = IT ⊗ σi2 (see Greene (2003, 13.7.1–2) and Wooldridge (2002,
P

10.7.2); "arellano" (see ibid. and the original ref. Arellano 1987) allows a fully general
structure w.r.t. heteroskedasticity and serial correlation:

2
 
σi1 σi1,i2 . . . ... σi1,iT

 σ 2
.. 
 i2,i1 σi2 . 


Ωi =  .. .. .. 
(17)
 . . .


 .. 2

. σiT σiT −1,iT 
 
 −1
σiT,i1 ... . . . σiT,iT −1 2
σiT

The latter is, as already observed, consistent w.r.t. timewise correlation of the errors, but on
the converse, unlike the White 1 and 2 methods, it relies on large n asymptotics with small
T.
The fixed effects case, as already observed in Section 6.4 on serial correlation, is complicated
by the fact that the demeaning induces serial correlation in the errors. The original White
estimator (white1) turns out to be inconsistent for fixed T as n grows, so in this case it is
advisable to use the arellano version (see Stock and Watson 2008).
The errors may be weighted according to the schemes proposed by MacKinnon and White
(1985) and Cribari-Neto (2004) to improve small-sample performance20 .
The main use of vcovHC is together with testing functions from the lmtest and car packages.
These typically allow passing the vcov parameter either as a matrix or as a function (see
Zeileis 2004). If one is happy with the defaults, it is easiest to pass the function itself:

R> library("lmtest")
R> re <- plm(inv~value+capital, data=Grunfeld, model="random")
R> coeftest(re,vcovHC)

t test of coefficients:

Estimate Std. Error t value Pr(>|t|)

(Intercept) -57.834415 23.449626 -2.4663 0.01451 *
20
The HC3 and HC4 weighting schemes are computationally expensive and may hit memory limits for nT
in the thousands, where on the other hand it makes little sense to apply small sample corrections.
40 Panel Data Econometrics in R: The plm Package

value 0.109781 0.012984 8.4551 6.186e-15 ***

capital 0.308113 0.051889 5.9379 1.284e-08 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

else one may do the covariance computation inside the call to coeftest, thus passing on a
matrix:

R> coeftest(re, vcovHC(re, method="white2", type="HC3"))

For some tests, e.g. for multiple model comparisons by waldtest, one should always provide
a function21 . In this case, optional parameters are provided as shown below (see also Zeileis
2004, p.12):

R> waldtest(re, update(re,.~.-capital),

+ vcov=function(x) vcovHC(x, method="white2", type="HC3"))

Wald test

Model 1: inv ~ value + capital

Model 2: inv ~ value
Res.Df Df Chisq Pr(>Chisq)
1 197
2 198 -1 87.828 < 2.2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Moreover, linearHypothesis from package car may be used to test for linear restrictions:

R> library("car")
R> linearHypothesis(re, "2*value=capital", vcov.=vcovHC)

Linear hypothesis test

Hypothesis:
2 value - capital = 0

Model 1: restricted model

Model 2: inv ~ value + capital

Note: Coefficient covariance matrix supplied.

Res.Df Df Chisq Pr(>Chisq)

1 198
21
Joint zero-restriction testing still allows providing the vcov of the unrestricted model as a matrix, see the
documentation of package lmtest.
 Yves Croissant, Giovanni Millo 41

2 197 1 3.4783 0.06218 .

---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

A specific vcovHC method for pgmm objects is also provided which implements the robust
covariance matrix proposed by Windmeijer (2005) for generalized method of moments esti-
mators.

7. plm versus nlme/lme4

The models termed panel by the econometricians have counterparts in the statistics literature
on mixed models (or hierarchical models, or models for longitudinal data), although there are
both differences in jargon and more substantial distinctions. This language inconsistency
between the two communities, together with the more complicated general structure of sta-
tistical models for longitudinal data and the associated notation in the software, is likely to
scare some practicing econometricians away from some potentially useful features of the R
environment, so it may be useful to provide here a brief reconciliation between the typical
panel data specifications used in econometrics and the general framework used in statistics
for mixed models22 .
R is particularly strong on mixed models’ estimation, thanks to the long-standing nlme pack-
age (see Pinheiro et al. 2007) and the more recent lme4 package, based on S4 classes (see
Bates 2007)23 . In the following we will refer to the more established nlme to give some ex-
amples of “econometric” panel models that can be estimated in a likelihood framework, also
including some likelihood ratio tests. Some of them are not feasible in plm and make a useful
complement to the econometric “toolbox” available in R.

7.1. Fundamental differences between the two approaches

Econometrics deal mostly with non-experimental data. Great emphasis is put on specifica-
tion procedures and misspecification testing. Model specifications tend therefore to be very
simple, while great attention is put on the issues of endogeneity of the regressors, dependence
structures in the errors and robustness of the estimators under deviations from normality.
The preferred approach is often semi- or non-parametric, and heteroskedasticity-consistent
techniques are becoming standard practice both in estimation and testing.
For all these reasons, although the maximum likelihood framework is important in testing24
and sometimes used in estimation as well, panel model estimation in econometrics is mostly
accomplished in the generalized least squares framework based on Aitken’s Theorem and,
when possible, in its special case ols, which are free from distributional assumptions (although
these kick in at the diagnostic testing stage). On the contrary, longitudinal data models in
22
This discussion does not consider gmm models. One of the basic reasons for econometricians not to choose
maximum likelihood methods in estimation is that the strict exogeneity of regressors assumption required for
consistency of the ml models reported in the following is often inappropriate in economic settings.
23
The standard reference on the subject of mixed models in S/R is Pinheiro and Bates (2000).
24
Lagrange Multiplier tests based on the likelihood principle are suitable for testing against more general
alternatives on the basis of a maintained model with spherical residuals and find therefore application in testing
for departures from the classical hypotheses on the error term. The seminal reference is Breusch and Pagan
(1980).
42 Panel Data Econometrics in R: The plm Package

nlme and lme4 are estimated by (restricted or unrestricted) maximum likelihood. While under
normality, homoskedasticity and no serial correlation of the errors ols are also the maximum
likelihood estimator, in all the other cases there are important differences.
The econometric gls approach has closed-form analytical solutions computable by standard
linear algebra and, although the latter can sometimes get computationally heavy on the ma-
chine, the expressions for the estimators are usually rather simple. ml estimation of longitudi-
nal models, on the contrary, is based on numerical optimization of nonlinear functions without
closed-form solutions and is thus dependent on approximations and convergence criteria. For
example, the “gls” functionality in nlme is rather different from its “econometric” counter-
part. “Feasible gls” estimation in plm is based on a single two-step procedure, in which an
inefficient but consistent estimation method (typically ols) is employed first in order to get a
consistent estimate of the errors’ covariance matrix, to be used in gls at the second step; on
the converse, “gls” estimators in nlme are based on iteration until convergence of two-step
optimization of the relevant likelihood.

7.2. Some false friends

The fixed/random effects terminology in econometrics is often recognized to be misleading, as
both are treated as random variates in modern econometrics (see e.g. Wooldridge 2002, 10.2.1).
It has been recognized since Mundlak’s classic paper (Mundlak 1978) that the fundamental
issue is whether the unobserved effects are correlated with the regressors or not. In this last
case, they can safely be left in the error term, and the serial correlation they induce is cared
for by means of appropriate gls transformations. On the contrary, in the case of correlation,
“fixed effects” methods such as least squares dummy variables or time-demeaning are needed,
which explicitly, although inconsistently25 , estimate a group– (or time–) invariant additional
parameter for each group (or time period).
Thus, from the point of view of model specification, having fixed effects in an econometric
model has the meaning of allowing the intercept to vary with group, or time, or both, while
the other parameters are generally still assumed to be homogeneous. Having random effects
means having a group– (or time–, or both) specific component in the error term.
In the mixed models literature, on the contrary, fixed effect indicates a parameter that is
assumed constant, while random effects are parameters that vary randomly around zero ac-
cording to a joint multivariate Normal distribution.
So, the fe model in econometrics has no counterpart in the mixed models framework, unless
reducing it to ols on a specification with one dummy for each group (often termed least
squares dummy variables, or lsdv model) which can trivially be estimated by ols. The re
model is instead a special case of a mixed model where only the intercept is specified as a
random effect, while the “random” type variable coefficients model can be seen as one that has
the same regressors in the fixed and random sets. The unrestricted generalized least squares
can in turn be seen, in the nlme framework, as a standard linear model with a general error
covariance structure within the groups and errors uncorrelated across groups.

25
For fixed effects estimation, as the sample grows (on the dimension on which the fixed effects are specified)
so does the number of parameters to be estimated. Estimation of individual fixed effects is T – (but not n–)
consistent, and the opposite.
 Yves Croissant, Giovanni Millo 43

7.3. A common taxonomy

To reconcile the two terminologies, in the following we report the specification of the panel
models in plm according to the general expression of a mixed model in Laird-Ware form (see
the web appendix to Fox 2002) and the nlme estimation commands for maximum likelihood
estimation of an equivalent specification26 .

The Laird-Ware representation for mixed models

A general representation for the linear mixed effects model is given in Laird and Ware (1982).

yit = β1 x1ij + . . . + βp xpij

b1 z1ij + . . . + bp zpij + ij
bik ∼ N (0, ψk2 ), Cov(bk , bk0 ) = ψkk0
ij ∼ N (0, σ 2 λijj ), Cov(ij , ij 0 ) = σ 2 λijj 0
where the x1 , . . . xp are the fixed effects regressors and the z1 , . . . zp are the random effects
regressors, assumed to be normally distributed across groups. The covariance of the random
effects coefficients ψkk0 is assumed constant across groups and the covariances between the
errors in group i, σ 2 λijj 0 , are described by the term λijj 0 representing the correlation structure
of the errors within each group (e.g., serial correlation over time) scaled by the common error
variance σ 2 .

Pooling and Within

The pooling specification in plm is equivalent to a classical linear model (i.e., no random effects
regressor and spherical errors: biq = 0 ∀i, q, λijj = σ 2 for j = j 0 , 0 else). The within one is
the same with the regressors’ set augmented by n − 1 group dummies. There is no point in
using nlme as parameters can be estimated by ols which is also ml.

Random effects
In the Laird and Ware notation, the re specification is a model with only one random effects
regressor: the intercept. Formally, z1ij = 1 ∀i, j, zqij = 0 ∀i, ∀j, ∀q 6= 1 λij = 1 for i = j,
0 else). The composite error is therefore uij = 1bi1 + ij . Below we report coefficients of
Grunfeld’s model estimated by gls and then by ml

R> library(nlme)
R> reGLS <- plm(inv~value+capital, data=Grunfeld, model="random")
R> reML <- lme(inv~value+capital, data=Grunfeld, random=~1|firm)
R> coef(reGLS)

(Intercept) value capital

-57.8344149 0.1097812 0.3081130
26
In doing so, we stress that “equivalence” concerns only the specification of the model, and neither the
appropriateness nor the relative efficiency of the relevant estimation techniques, which will of course be depen-
dent on the context. Unlike their mixed model counterparts, the specifications in plm are, strictly speaking,
distribution-free. Nevertheless, for the sake of exposition, in the following we present them in the setting which
ensures consistency and efficiency (e.g., we consider the hypothesis of spherical errors part of the specification
of pooled ols and so forth).
44 Panel Data Econometrics in R: The plm Package

R> summary(reML)$coefficients$fixed

(Intercept) value capital

-57.8644245 0.1097897 0.3081881

Variable coefficients, “random”

Swamy’s variable coefficients model (Swamy 1970) has coefficients varying randomly (and
independently of each other) around a set of fixed values, so the equivalent specification
is zq = xq ∀q, i.e. the fixed effects and the random effects regressors are the same, and
ψkk0 = σµ2 IN , and λijj = 1, λijj 0 = 0 for j 6= j 0 , that’s to say they are not correlated.
Estimation of a mixed model with random coefficients on all regressors is rather demanding
from the computational side. Some models from our examples fail to converge. The below
example is estimated on the Grunfeld data and model with time effects.

R> vcm <- pvcm(inv~value+capital, data=Grunfeld, model="random", effect="time")

R> vcmML <- lme(inv~value+capital, data=Grunfeld, random=~value+capital|year)
R> coef(vcm)

y
(Intercept) -18.5538638
value 0.1239595
capital 0.1114579

R> summary(vcmML)$coefficients$fixed

(Intercept) value capital

-26.3558395 0.1241982 0.1381782

Variable coefficients, “within”

This specification actually entails separate estimation of T different standard linear models,
one for each group in the data, so the estimation approach is the same: ols. In nlme this
is done by creating an lmList object, so that the two models below are equivalent (output
suppressed):

R> vcmf <- pvcm(inv~value+capital, data=Grunfeld, model="within", effect="time")

R> vcmfML <- lmList(inv~value+capital|year, data=Grunfeld)

Unrestricted fgls
The general, or unrestricted, feasible gls, pggls in the plm nomenclature, is equivalent to
a model with no random effects regressors (biq = 0 ∀i, q) and an error covariance structure
which is unrestricted within groups apart from the usual requirements. The function for
estimating such models with correlation in the errors but no random effects is gls().
This very general serial correlation and heteroskedasticity structure is not estimable for the
original Grunfeld data, which have more time periods than firms, therefore we restrict them
to firms 4 to 6.
 Yves Croissant, Giovanni Millo 45

R> sGrunfeld <- Grunfeld[Grunfeld$firm%in%4:6,]

R> ggls <- pggls(inv~value+capital, data=sGrunfeld, model="pooling")
R> gglsML <- gls(inv~value+capital, data=sGrunfeld,
+ correlation=corSymm(form=~1|year))
R> coef(ggls)

(Intercept) value capital

1.19679342 0.10555908 0.06600166

R> summary(gglsML)$coefficients

(Intercept) value capital

-2.4156266 0.1163550 0.0735837

The within case is analogous, with the regressors’ set augmented by n − 1 group dummies.

7.4. Some useful “econometric” models in nlme

Finally, amongst the many possible specifications estimable with nlme, we report a couple
cases that might be especially interesting to applied econometricians.

AR(1) pooling or random effects panel

Linear models with groupwise structures of time-dependence27 may be fitted by gls(), spec-
ifying the correlation structure in the correlation option28 :

R> Grunfeld$year <- as.numeric(as.character(Grunfeld$year))

R> lmAR1ML <- gls(inv~value+capital,data=Grunfeld,
+ correlation=corAR1(0,form=~year|firm))

and analogously the random effects panel with, e.g., AR(1) errors (see Baltagi 2005, 2013,
chap 5), which is a very common specification in econometrics, may be fit by lme specifying
an additional random intercept:

R> reAR1ML <- lme(inv~value+capital, data=Grunfeld,random=~1|firm,

+ correlation=corAR1(0,form=~year|firm))

The regressors’ coefficients and the error’s serial correlation coefficient may be retrieved this
way:

R> summary(reAR1ML)$coefficients$fixed

(Intercept) value capital

-40.27650822 0.09336672 0.31323330
27
Take heed that here, in contrast to the usual meaning of serial correlation in time series, we always speak
of serial correlation between the errors of each group.
28
note that the time index is coerced to numeric before the estimation.
46 Panel Data Econometrics in R: The plm Package

R> coef(reAR1ML$modelStruct$corStruct, unconstrained=FALSE)

Phi
0.823845

Significance statistics for the regressors’ coefficients are to be found in the usual summary
object, while to get the significance test of the serial correlation coefficient one can do a
likelihood ratio test as shown in the following.

An LR test for serial correlation and one for random effects

A likelihood ratio test for serial correlation in the idiosyncratic residuals can be done as a
nested models test, by anova(), comparing the model with spherical idiosyncratic residuals
with the more general alternative featuring AR(1) residuals. The test takes the form of a zero
restriction test on the autoregressive parameter.
This can be done on pooled or random effects models alike. First we report the simpler case.
We already estimated the pooling AR(1) model above. The gls model without correlation in
the residuals is the same as ols, and one could well use lm() for the restricted model. Here
we estimate it by gls().

R> lmML <- gls(inv~value+capital, data=Grunfeld)

R> anova(lmML, lmAR1ML)

Model df AIC BIC logLik Test L.Ratio p-value

lmML 1 4 2400.217 2413.350 -1196.109
lmAR1ML 2 5 2094.936 2111.352 -1042.468 1 vs 2 307.2813 <.0001

The AR(1) test on the random effects model is to be done in much the same way, using the
random effects model objects estimated above:

R> anova(reML, reAR1ML)

Model df AIC BIC logLik Test L.Ratio p-value

reML 1 5 2205.851 2222.267 -1097.926
reAR1ML 2 6 2094.802 2114.501 -1041.401 1 vs 2 113.0496 <.0001

A likelihood ratio test for random effects compares the specifications with and without random
effects and spherical idiosyncratic errors:

R> anova(lmML, reML)

Model df AIC BIC logLik Test L.Ratio p-value

lmML 1 4 2400.217 2413.350 -1196.109
reML 2 5 2205.851 2222.267 -1097.926 1 vs 2 196.366 <.0001

The random effects, AR(1) errors model in turn nests the AR(1) pooling model, therefore
a likelihood ratio test for random effects sub AR(1) errors may be carried out, again, by
comparing the two autoregressive specifications:
 Yves Croissant, Giovanni Millo 47

R> anova(lmAR1ML, reAR1ML)

Model df AIC BIC logLik Test L.Ratio p-value

lmAR1ML 1 5 2094.936 2111.352 -1042.468
reAR1ML 2 6 2094.802 2114.501 -1041.401 1 vs 2 2.134349 0.144

whence we see that the Grunfeld model specification doesn’t seem to need any random effects
once we control for serial correlation in the data.

8. Conclusions
With plm we aim at providing a comprehensive package containing the standard functionali-
ties that are needed for the management and the econometric analysis of panel data. In partic-
ular, we provide: functions for data transformation; estimators for pooled, random and fixed
effects static panel models and variable coefficients models, general gls for general covariance
structures, and generalized method of moments estimators for dynamic panels; specification
and diagnostic tests. Instrumental variables estimation is supported. Most estimators allow
working with unbalanced panels. While among the different approaches to longitudinal data
analysis we take the perspective of the econometrician, the syntax is consistent with the basic
linear modeling tools, like the lm function.
On the input side, formula and data arguments are used to specify the model to be estimated.
Special functions are provided to make writing formulas easier, and the structure of the data
is indicated with an index argument.
On the output side, the model objects (of the new class panelmodel) are compatible with
the general restriction testing frameworks of packages lmtest and car. Specialized methods
are also provided for the calculation of robust covariance matrices; heteroskedasticity- and
correlation-consistent testing is accomplished by passing these on to testing functions, together
with a panelmodel object.
The main functionalities of the package have been illustrated here by applying them on some
well-known datasets from the econometric literature. The similarities and differences with
the maximum likelihood approach to longitudinal data have also been briefly discussed.
We plan to expand the methods in this paper to systems of equations and to the estimation
of models with autoregressive errors. Addition of covariance estimators robust vs. cross-
sectional correlation are also in the offing. Lastly, conditional visualization features in the R
environment seem to offer a promising toolbox for visual diagnostics, which is another subject
for future work.

Acknowledgments
While retaining responsibility for any error, we thank Jeffrey Wooldridge, Achim Zeileis and
three anonymous referees for useful comments. We also acknowledge kind editing assistance
by Lisa Benedetti.
48 Panel Data Econometrics in R: The plm Package

References

Amemiya T (1971). “The Estimation of the Variances in a Variance–Components Model.”

International Economic Review, 12(1), 1–13.

Amemiya T, MaCurdy T (1986). “Instrumental–Variable Estimation of an Error Components

Model.” Econometrica, 54(4), 869–880.

Anderson T, Hsiao C (1981). “Estimation of Dynamic Models With Error Components.”

Journal of the American Statistical Association, 76(375), 598–606.

Arellano M (1987). “Computing Robust Standard Errors for Within–Group Estimators.”

Oxford Bulletin of Economics and Statistics, 49(4), 431–434.

Arellano M, Bond S (1991). “Some Tests of Specification for Panel Data: Monte Carlo
Evidence and an Application to Employment Equations.” Review of Economic Studies,
58(2), 277–297.

Balestra P, Varadharajan-Krishnakumar J (1987). “Full Information Estimations of a System

of Simultaneous Equations With Error Components.” Econometric Theory, 3(2), 223–246.

Baltagi B (1981). “Simultaneous Equations With Error Components.” Journal of Economet-

rics, 17(2), 189–200.

Baltagi B (2005). Econometric Analysis of Panel Data. 3rd edition. John Wiley and Sons
ltd.

Baltagi B (2013). Econometric Analysis of Panel Data. 5th edition. John Wiley and Sons
ltd.

Baltagi B, Chang Y, Li Q (1992). “Monte Carlo results on several new and existing tests for
the error component model.” Journal of Econometrics, 54(1–3), 95–120.

Baltagi B, Chang Y, Li Q (1998). “Testing for random individual and time effects using
unbalanced panel data.” Advances in Econometrics, 13, 1–20.

Baltagi B, Li Q (1990). “A lagrange multiplier test for the error components model with
incomplete panels.” Econometric Reviews, 9(1), 103–107.

Baltagi B, Li Q (1991). “A Joint Test for Serial Correlation and Random Individual Effects.”
Statistics and Probability Letters, 11(3), 277–280.

Baltagi B, Li Q (1995). “Testing AR(1) Against MA(1) Disturbances in an Error Component

Model.” Journal of Econometrics, 68(1), 133–151.

Bates D (2004). “Least Squares Calculations in R.” R–news, 4(1), 17–20.

Bates D (2007). lme4: Linear Mixed–Effects Models Using S4 Classes. R package version
0.99875-9, URL https://cran.r-project.org/package=lme4.

Bates D, Maechler M (2016). Matrix: Sparse and Dense Matrix Classes and Methods. R
package version 1.2-7.1, URL https://cran.r-project.org/package=Matrix.
 Yves Croissant, Giovanni Millo 49

Bera A, Sosa-Escudero W, Yoon M (2001). “Tests for the Error Component Model in the
Presence of Local Misspecification.” Journal of Econometrics, 101(1), 1–23.

Bhargava A, Franzini L, Narendranathan W (1982). “Serial Correlation and the Fixed Effects
Model.” Review of Economic Studies, 49(4), 533–554.

Bivand R (2008). spdep: Spatial Dependence: Weighting Schemes, Statistics and Models. R
package version 0.4-17, URL https://cran.r-project.org/package=spdep.

Blundell R, Bond S (1998). “Initial Conditions and Moment Restrictions in Dynamic Panel
Data Models.” Journal of Econometrics, 87(1), 115–143.

Breusch T, Mizon G, Schmidt P (1989). “Efficient Estimation Using Panel Data.” Economet-
rica, 57(3), 695–700.

Breusch T, Pagan A (1980). “The Lagrange Multiplier Test and Its Applications to Model
Specification in Econometrics.” Review of Economic Studies, 47(1), 239–253.

Cornwell C, Rupert P (1988). “Efficient Estimation With Panel Data: An Empirical Com-
parison of Instrumental Variables Estimators.” Journal of Applied Econometrics, 3(2),
149–155.

Cribari-Neto F (2004). “Asymptotic Inference Under Heteroskedasticity of Unknown Form.”

Computational Statistics & Data Analysis, 45(2), 215–233.

Croissant Y, Millo G (2008). “Panel Data Econometrics in R: The plm Package.” Journal of
Statistical Software, 27(2). URL http://www.jstatsoft.org/v27/i02/.

De Hoyos R, Sarafidis V (2006). “Testing for Cross–Sectional Dependence in Panel–Data

Models.” The Stata Journal, 6(4), 482–496.

Drukker D (2003). “Testing for Serial Correlation in Linear Panel–Data Models.” The Stata
Journal, 3(2), 168–177.

Fox J (2002). An R and S–plus Companion to Applied Regression. Sage.

Fox J (2016). car: Companion to Applied Regression. R package version 2.1-3, URL https:
//cran.r-project.org/package=car,http://socserv.socsci.mcmaster.ca/jfox/.

Gourieroux C, Holly A, Monfort A (1982). “Likelihood Ratio Test, Wald Test, and Kuhn–
Tucker Test in Linear Models With Inequality Constraints on the Regression Parameters.”
Econometrica, 50(1), 63–80.

Greene W (2003). Econometric Analysis. 5th edition. Prentice Hall.

Harrison D, Rubinfeld D (1978). “Hedonic housing prices and the demand for clean air.”
Journal of Environmental Economics and Management, 5(1), 81–102.

Hausman J (1978). “Specification Tests in Econometrics.” Econometrica, 46(6), 1251–1271.

Hausman J, Taylor W (1981). “Panel Data and Unobservable Individual Effects.” Economet-
rica, 49(6), 1377–1398.
50 Panel Data Econometrics in R: The plm Package

Holtz-Eakin D, Newey W, Rosen H (1988). “Estimating Vector Autoregressions With Panel

Data.” Econometrica, 56(6), 1371–1395.

Honda Y (1985). “Testing the Error Components Model With Non–Normal Disturbances.”
Review of Economic Studies, 52(4), 681–690.

Hothorn T, Zeileis A, Farebrother RW, Cummins C, Millo G, Mitchell D (2015). lmtest: Test-
ing Linear Regression Models. R package version 0.9-34, URL https://cran.r-project.
org/package=lmtest.

King M, Wu P (1997). “Locally Optimal One–Sided Tests for Multiparameter Hypotheses.”

Econometric Reviews, 16(2), 131–156.

Kleiber C, Zeileis A (2008). Applied Econometrics with R. Springer-Verlag, New York. ISBN
978-0-387-77316-2, URL https://cran.r-project.org/package=AER.

Koenker R, Ng P (2016). SparseM: Sparse Linear Algebra. R package version 1.72, URL
https://cran.r-project.org/package=SparseM.

Laird N, Ware J (1982). “Random–Effects Models for Longitudinal Data.” Biometrics, 38(4),
963–974.

Lumley T, Zeileis A (2015). sandwich: Robust Covariance Matrix Estimators. R package

version 2.3-4, URL https://cran.r-project.org/package=sandwich.

MacKinnon J, White H (1985). “Some Heteroskedasticity–Consistent Covariance Matrix Esti-

mators With Improved Finite Sample Properties.” Journal of Econometrics, 29(3), 305–325.

Mundlak Y (1978). “On the Pooling of Time Series and Cross Section Data.” Econometrica,
46(1), 69–85.

Munnell A (1990). “Why Has Productivity Growth Declined? Productivity and Public In-
vestment.” New England Economic Review, pp. 3–22.

Nerlove M (1971). “Further Evidence on the Estimation of Dynamic Economic Relations from
a Time Series of Cross Sections.” Econometrica, 39(2), 359–382.

Pesaran M (2004). “General Diagnostic Tests for Cross Section Dependence in Panels.” CESifo
Working Paper Series, 1229.

Pesaran M (2015). “Testing Weak CrossâĂŞSectional Dependence in Large Panels.” Econo-

metric Reviews, 34(6-10), 1089–1117.

Pinheiro J, Bates D (2000). Mixed–Effects Models in S and S-plus. Springer-Verlag.

Pinheiro J, Bates D, DebRoy S, the R Core team DS (2007). nlme: Linear and Nonlinear
Mixed Effects Models. R package version 3.1-86, URL https://cran.r-project.org/
package=nlme.

R Development Core Team (2008). R: A Language and Environment for Statistical Computing.
R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL http:
//www.r-project.org/.
 Yves Croissant, Giovanni Millo 51

Stock JH, Watson MW (2008). “Heteroskedasticity–Robust Standard Errors for Fixed Effects
Panel Data Regression.” Econometrica, 76(1), 155–174.

Swamy P (1970). “Efficient Inference in a Random Coefficient Regression Model.” Economet-

rica, 38(2), 311–323.

Swamy P, Arora S (1972). “The Exact Finite Sample Properties of the Estimators of Coeffi-
cients in the Error Components Regression Models.” Econometrica, 40(2), 261–275.

Therneau T (2014). bdsmatrix: Routines for Block Diagonal Symmetric matrices. R package
version 1.3-2, URL https://cran.r-project.org/package=bdsmatrix.

Wallace T, Hussain A (1969). “The Use of Error Components Models in Combining Cross
Section With Time Series Data.” Econometrica, 37(1), 55–72.

White H (1980). Asymptotic Theory for Econometricians. Academic Press, Orlando.

White H (1984). “A Heteroskedasticity–Consistent Covariance Matrix and a Direct Test for

Heteroskedasticity.” Econometrica, 48(4), 817–838.

Windmeijer F (2005). “A Finite Sample Correction for the Variance of Linear Efficient Two–
Step GMM Estimators.” Journal of Econometrics, 126(1), 25–51.

Wooldridge J (2002). Econometric Analysis of Cross Section and Panel Data. MIT Press.

Wooldridge J (2010). Econometric Analysis of Cross Section and Panel Data. 2nd edition.
MIT Press.

Zeileis A (2004). “Econometric Computing With HC and HAC Covariance Matrix Estimators.”
Journal of Statistical Software, 11(10), 1–17. URL http://www.jstatsoft.org/v11/i10/.

Affiliation:
Yves Croissant
CEMOI
Faculté de Droit et d’Economie
Université de La Réunion
15 avenue René Cassin
CS 92003
F-97744 Saint-Denis Cedex 9
Telephone: +262262938446
E-mail: [email protected]

Giovanni Millo
DiSES, Un. of Trieste and R&D Dept., Generali SpA
Via Machiavelli 4
34131 Trieste (Italy)
Telephone: +39/040/671184
52 Panel Data Econometrics in R: The plm Package

Fax: +39/040/671160
E-mail: [email protected]