Estimation of time-invariant effects in both balanced and

unbalanced static panel models: A discussion of the “xtfef ”

command for FEF and FEF-IV estimators
Yui Law
State University of New York at Binghamton

Qiankun Zhou
State University of New York at Binghamton

April 28, 2017

Abstract

This paper introduces the Stata command xtfef which implements the Fixed-Effects Filter
(FEF) and Fixed-Effects Filter IV (FEF-IV) estimation proposed by Pesaran and Zhou (2017)
in static panels. The FEF approach identifies and estimates the of time-invariant effects through
a two-step procedure. The first step is the usual FE estimation which filters out time-invariant
regressors, and the second step recovers the time-invariant effects by using time-series averages
of the residuals from the first step as dependent variables in the OLS regression. The FEF-IV
approach extends the FEF to the case when there are endogenous time-invariant regressors in
the model. The command xtfef facilitates the implementation of both FEF and FEF-IV in
Stata. Illustration example of how to use the command xtfef to empirical data is also provided.
Keywords: Static panel data models, fixed effects, time-invariant effects, instrumental vari-
ables
JEL classification: C01, C23, C33

1
1 Introduction
Identification and estimation of the impacts of time-invariant regressors, such as the effects of race
and gender, is often the main purpose of panel data studies. However, usual estimation approaches
such as fixed effects (FE) cannot be applied for estimation of the effects of time-invariant variables,
since the transformation of the FE eliminates all the impacts of time-invariant regressors. As a
result, estimation of time-invariant effects has been a challenging issue in panel data econometrics.
Recently, Plumper and Troeger (2011) propose the Fixed Effects Vector Decomposition (FEVD)
to estimate the time-invariant effects under a three-step procedure. As shown in Pesaran and Zhou
(2017) , although the FEVD approach with modification can provide consistent estimates of the
time-invariant effects, the variance estimator is inconsistent. More discussions are presented in
Greene (2011a) , Brusch et al. (2011b), Plumper and Troeger (2011), Greene (2011b), Brusch et al.
(2011a), and Beck (2011). Another procedure to estimate the time-invariant effects is proposed by
Hausman and Taylor (1981)(HT) . However, HT uses the time-varying variables as instrumental
variables to estimate the time-invariant effects. Therefore, the assumption that a known subset of
time-varying variables are exogenous is necessary. Most recently, Pesaran and Zhou (2017) propose
the Fixed-Effects Filter (FEF) approach to identify and estimate the of time-invariant effects in
static panel models under fairly general assumptions. The FEF approach is a two-step procedure,
where the first step is the usual FE estimation which filters out time-invariant regressors, and the
second step recovers the time-invariant effects by using time-series averages of the residuals from the
first step as dependent variables in the OLS regression. Pesaran and Zhou (2017) also extend the
FEF procedure to the model with endogenous time-invariant regressors, which they call it as Fixed-
Effects Filter IV (FEF-IV, for short) estimation. As shown by Pesaran and Zhou (2017), in contrast
to the FEVD and HT estimation, the FEF and FEF-IV provide not only consistent estimates of
the coefficients of the time-invariant effects, but also the variance is robust to heteroscedastic errors
and serial-correlated errors. In order to facilitate the implementation of both FEF and FEF-IV
in Stata, this paper introduces a new Stata command xtfef, which can be used to estimate the
time-invariant effects for static panels.
The rest of the paper is organized as follows. Estimation of time-invariant effects for both
balanced and unbalanced panel is briefly discussed in Section 2 and Section 3, respectively. Section
4 discusses the implementation of the xtfef command in Stata, and illustration of xtfef command
is provided in Section 5. Concluding remarks are provided in Section 6.

2 Balanced Panel data models with time-invariant effects

We consider the following panel data model that contains time-varying as well as time-invariant
regressors
yit = αi + z0i γ + x0it β + εit , i = 1, 2, . . . , N ; t = 1, 2, . . . , T, (1)

2
where
αi = α + ηi , (2)

Also let

vi = ηi + ε̄i

where,

T
X
ε̄i = T −1 εit
t=1

xit is a k × 1 vector of time-varying variables, zi is an m × 1 vector of observed individual-specific

variables that only vary over the cross section units, i, and T is the number of observations of the
cross section unit. The goal of the analysis is on estimation and inference involving the elements of
γ. In what follows we allow for ηi and xit to have any degree of dependence, but initially assume
that zi and vi are uncorrelated for identification of γ, and assume that xit and εis are uncorrelated
for all i,t and s, to identify β. This approach can be modified in cases where one or more instruments
are available for zi and/or xit .
For model (1) we make the following assumptions, which are quite standard in the literature
(e.g., Pesaran (2015) and Pesaran and Zhou (2017)).
Assumption P1: E(εit |xis ) = 0, for all i, t and s, and E ε4it < K < ∞, for all i and t.

Assumption P2: E (εit εjs |X ) = 0, for all i 6= j, and all t and s, where X = (xit ; i = 1, 2, ..., N ;
t = 1, 2, ..., T ).
Assumption P3: The errors, εit , are heteroscedastic and temporally dependent, namely

E (εit εis |X ) = γi (t, s), for all t and s,

where 0 < γi (t, t) = σi2 , and |γi (t, s)| < K, for all i, t and s.
Assumption P4 : The regressors, xit satisfy the moment conditions E kxit − x̄k4 < K < ∞,
and E kx̄i − x̄k4 < K < ∞, for all i and t, where x̄i = T −1 Tt=1 xit , and x̄ = N −1 N
P P
i=1 x̄i .
Assumption P5: The k × k matrices Qp,N T and QF E,N T defined by

N T
1 XX
Qp,N T = (xit − x̄) (xit − x̄)0 , (3)
NT
i=1 t=1

N T
1 XX
QF E,N T = (xit − x̄i ) (xit − x̄i )0 , (4)
NT
i=1 t=1

converge (in probability) to Qp,T and QF E,T for a fixed T and as N tends to infinity, λmin (QF E,N T ) >

3
1/K and λmin (Qp,N T ) > 1/K, for all N and T where K is a finite, non-zero constant.
Assumption P6: The m × m matrix, Qzz,N , and the m × k matrix Qz x̄,N defined by

N
1 X
Qzz,N = (zi − z̄) (zi − z̄)0 , (5)
N
i=1

N
1 X
Qz x̄,N = (zi − z̄) (x̄i − x̄)0 , (6)
N
i=1

converge (in probability) to the non-stochastic limits Qzz and Qz x̄ , and λmin (Qzz,N ) > 1/K, for
all N > m.
Assumption P7: The time-invariant regressors, zi , are independently distributed of vj =
ηj + ε̄j , for all i and j, and ηi and ε̄i are independently distributed. Also, zi satisfy the moment
conditions E k(zi − z̄)k4 < K, for all i.

2.1 Fixed effects filtered (FEF) estimator of time-invariant effects

Under the assumption that xit and εis are uncorrelated for all i, t and s, as it is well known β
can be estimated consistently under fairly general assumptions on temporal dependence and cross-
sectional heteroscedasticity of εit , and the distribution of the fixed effects, αi . Denoting the FE
0
estimator of β by β̂, γ can then be estimated by the regression of ȳi − β̂ x̄i on an intercept and zi .
We denote this estimator by γ̂ F EF and refer to it as the fixed effects filtered (FEF) estimator of γ.
Formally, the FEF estimator can be computed using the following two-step procedure:
Step 1: Using model (1), compute the fixed-effects estimator of β, denoted by β̂, and the
associated residuals ûit defined by
0
ûit = yit − β̂ xit . (7)
PT
Step 2: Compute the time averages of these residuals, ûi = T −1 t=1 ûit , and regress ûi on zi
with an intercept to obtain γ̂ F EF , namely
" N
#−1 N
0
X X

γ̂ F EF = (zi − z̄) (zi − z̄) (zi − z̄) ûi − û , (8)
i=1 i=1

and
α̂F EF = û − γ̂ 0F EF z̄, (9)
PN
where û = N −1 i=1 ûi .
For this FEF estimator (8), it is shown by Pesaran and Zhou (2017) that

Proposition 1 Consider the FEF estimator γ̂ F EF of γ in the panel data model (1) defined by
(A.24 of Pesaran and Zhou (2017)), and suppose that Assumptions P1-P7 hold. Then γ̂ F EF is an

4
unbiased and a consistent estimator of γ, and
√
N (γ̂ F EF − γ) →d N (0, Ωγ̂F EF ) , (10)

where
Ωγ̂F EF = Q−1 2

−1
zz ση Qzz + Ωξ̄ Qzz . (11)

and Qzz is defined in Assumption P6, Ωξ̄ = limN →∞ Ωξ̄,N , with Ωξ̄,N is defined in Pesaran and
Zhou (2017, eq A.24), and ση2 is the variance of the fixed effects defined by (2).

Given the asymptotic distribution of γ̂ F EF , Pesaran and Zhou (2017) show that
h   i
ar (γ̂ F EF ) = Q−1
Ω̂γ̂F EF = N Vd zz,N V̂ zz,N + Qz x̄,N N V
d ar( β̂) Q0
z x̄,N Q−1
zz,N , (12)

is a consistent estimator of Ωγ̂F EF defined in (11) for a fixed T and as N → ∞, where Qzz,N and
Qz x̄,N are defined by (5) and (6), respectively, and V̂zz,N is defined by (14),

N
!−1 N
! N
!−1
X X X
Vd
ar(β̂) = x0i· xi· x0i· ei e0i xi· x0i· xi· , (13)
i=1 i=1 i=1

where x0i· = (xi1 − x̄i , xi2 − x̄i , . . . , xiT − x̄i ) denotes the demeaned vector of xit , and the t-th
element of ei is given by eit = yit − ȳi − (xit − x̄i )0 β̂, and

N
1 X
2
V̂zz,N = ςˆi − ςˆ (zi − z̄)(zi − z̄)0 , (14)
N
i=1

where
ςˆi − ςˆ = ȳi − ȳ − (x̄i − x̄)0 β̂ − (zi − z̄)0 γ̂ F EF . (15)

2.2 FEF-IV estimation of time-invariant effects

In the above section, we consider the case when the time-invariant regressors are exogenous. How-
ever, this estimation procedure does not apply to the case where one or more time-invariant re-
gressors are endogenous. In this section, we provide a generalization of the FEF estimator that
allows for possible endogeneity of the time-invariant regressors, assuming the existence of the same
vector of time-invariant instruments, ri . We denote this estimator by FEF-IV and derive its asymp-
totic properties under the following assumptions that allow for the errors to be heteroscedastic and
serially correlated:
Assumption P8: There exists the s × 1 vector of instruments ri for zi , i = 1, 2, . . . , N , where
ri is distributed independently of ηj and ε̄j for all i and j, and s ≥ m, and satisfy the moment
condition E kri − r̄k4 < K < ∞, if it has unbounded support.

5
 Assumption P9: Let

N
X N
X N
X
Qrz,N = N −1 (ri −r̄)(zi −z̄)0 , Qrx̄,N = N −1 (ri −r̄)(x̄i −x̄)0 , Qrr,N = N −1 (ri −r̄)(ri −r̄)0 ,
i=1 i=1 i=1
(16)
PN PN PN
where r̄ = N −1 i=1 ri , z̄ = N −1 i=1 zi ,and x̄ = N −1 i=1 xi . The s × m matrix Qrz,N , and the
s × s matrix Qrr,N have full column ranks (namely rank (Qrz,N ) = m ≤ s and rank (Qrr,N ) = s),
and have finite probability limits as N → ∞ given by Qrz and Qrr , respectively. Matrices Qrx̄,N
and Qzz,N have finite probability limits given by Qrx̄ and Qzz , respectively, and in cases where xit
and zi are stochastic with unbounded supports, 1/K < λmin (Qrr,N ) < λmax (Qrr,N ) < K, for some
constant K, and for all N , and as N → ∞, with probability approaching unity.
Under the above assumptions and maintaining the earlier Assumptions P1-P6, a consistent
estimator of γ can be obtained as follows
 −1  
γ̂ F EF −IV = Qzr,N Q−1 0
rr,N Qzr,N Qzr,N Q−1
rr,N Qrû,N , (17)

where Qzr,N and Qrr,N are defined by (16),

N
1 X

Qrû,N = (ri − r̄) ûi − û ,
N
i=1

PN
û = 1
N i=1 ûi , ûi = ȳi − x̄0i β̂, and β̂ is the FE estimator of β.
For the FEF-IV estimator (17), it is shown by Pesaran and Zhou (2017) that

Proposition 2 Suppose that Assumptions P1-P6, P8 and P9 hold, and let the FEF-IV estimator
be defined as in (17). Then we have
√


N γ̂ F EF −IV − γ →d N 0, Ωγ̂F EF −IV ,

where Ωγ̂F EF −IV is given by

−1 −1 0 −1
Ωγ̂F EF −IV = Qzr Q−1 0
Qzr Q−1 2

−1 0

rr Qzr rr ση Qrr + Ωψ̄ Qrr Qzr Qzr Qrr Qzr . (18)

with Ωψ̄ is defined in Pesaran and Zhou (2017, eq 49).

The variance of γ̂ F EF −IV can now be estimated along similar lines as in the above section. We
have h   i
ar γ̂ F EF −IV = N −1 Hzr,N V̂rr,N + Qrx̄,N N Vd
ar(β̂) Q0rx̄,N H0zr,N ,


Vd (19)

where  −1
Hzr,N = Qzr,N Q−1
rr,N Q0
zr,N Qzr,N Q−1
rr,N ,

6
 N
1 X
Qrx̄,N = (ri − r̄) (x̄i − x̄)0 ,
N
i=1

N
1 X
2
V̂rr,N = (ri − r̄)(ri − r̄)0 υ̂i − υ̂ ,
N
i=1

υ̂i − υ̂ = ȳi − ȳ − (x̄i − x̄)0 β̂ − (zi − z̄)0 γ̂ F EF −IV ,

√ 
and as before Vd
ar(β̂) is given by (13). It is shown by Pesaran and Zhou (2017) that Vd
ar N γ̂ F EF −IV
is a consistent estimator of Ωγ̂F EF −IV defined by (18) under fairly general conditions.

3 Unbalanced Panel data models with time-invariant effects

For the above FEF and FEF-IV estimators, even if they are proposed for balanced panel, however,
they can be easily modified to apply to unbalanced panel with time varying and time-invariant
variables. Consider the following unbalanced panel data model

yit = αi + z0i γ + x0it β + εit , t = 1, 2, . . . , Ti , i = 1, 2, . . . , N ; (20)

where αi = α + ηi , xit and zi are defined in model (1). However, here we have Ti 6= Tj for some
i 6= j, i.e., the panel is unbalanced and we assume the unbalanced feasture is caused randomly.
For the unbalanced panel (20), we assume assumptions P1-P9 are valid, and we find the FEF
and FEF-IV estimator can be easily modified to estimate the time-invariant effects γ.

3.1 FEF of time-invariant effects in unbalanced panels

Under the assumption that xit and εis are uncorrelated for all i, t and s, the FEF estimator of time-
invariant effects in unbalanced panels can still be computed using the following two-step procedure:
Step 1: Using model (20), compute the fixed-effects estimator of β, denoted by β̂, and the
associated residuals ûit defined by

0
ûit = yit − β̂ xit , t = 1, 2, . . . , Ti ; i = 1, 2, . . . , N. (21)

∗
Step 2: Compute the time averages of these residuals for each cross-sectional units, ûi =
−1 PTi ∗
Ti t=1 ûit for i = 1, . . . , N , and regress ûi on zi with an intercept to obtain γ̂ F EF , namely

" N
#−1 N
∗ ∗
 
0
X X
γ̂ U B
F EF = (zi − z̄) (zi − z̄) (zi − z̄) ûi − û , (22)
i=1 i=1

and
∗
α̂F EF = û − z̄0 γ̂ U B
F EF , (23)
∗ PN ∗
where û = N −1 i=1 ûi .

7
 The limiting distribution and estimation of asymptotic variance of γ̂ U B
F EF can be reached by
following the similar argument in the above section.

3.2 FEF-IV of time-invariant effects in unbalanced panels

Similarly, if there is endogenous time-invariant variables in model (20), assuming the existence of
the same vector of time-invariant instruments, ri , we can still use the FEF-IV approach to estimate
the time-invariant effects.
The FEF-IV approach for unbalanced panel can be done in two steps, and the first step is
the same as above, and the second step is the usual 2SLS estimation using the instruments ri . A
consistent estimator of γ can be obtained as follows
 −1  
−1 0 −1 ∗
γ̂ U B
F EF −IV = Q Q Q
zr,N rr,N zr,N Qzr,N rr,N rû,N ,
Q Q (24)

where Qzr,N and Qrr,N are the same as above, and

N
1 X 
∗ ∗

Q∗rû,N = (ri − r̄) ûi − û ,
N
i=1

∗ PN ∗ ∗
ûi = Ti−1
1 PTi
û = N i=1 ûi , t=1 ûit , and ûit is residuals in the the FE estimation.
The limiting distribution and estimation of asymptotic variance of γ̂ U B
F EF −IV can be reached by
following the similar argument in the above section.

4 xtfef command in Stata

The syntaxes of the command xtfef which realizes the FEF and FEF-IV estimation in static panel
models is as follows.

For FEF,

xtfef depvar [if] [in] , invariant(varlist) varying(varlist) i(varname) t(varname)

For FEF-IV,

xtfef depvar [if] [in] , fefiv invariant(varlist) varying(varlist) endogenous(varlist) in-

strument(varlist) i(varname) t(varname)

8
 Options Description
The following options must be specified in FEF or FEF-IV
invariant(varlist) specifies the time-invariant variables
varying(varlist) specifies the time-varying variables
i(varname) specifies the cross-section variable
t(varname) specifies the time variable
Besides the options obove, the following options must be specified in FEF-IV
fefiv FEF-IV model if specified; FEF model if not
specified
endogenous(varlist) specifies the endogenous time-invariant
variables
instrument(varlist) specifies the instrumental variables of the
endogenous variables

In the FEF model, it is necessary to specify the dependent variable, time-invariant variables in
invariant(), time-varying variables in varying(), the cross-section identity in i(), and the time
variable in t(), respectively. The option xtfef should not be included.

For the FEF-IV, the option xtfef must be specified. It is also necessary to specify the depen-
dent variable, time-invariant variables in invariant(), time-varying variables in varying(), the
cross-section identity in i(), and the time variable in t(), respectively. Moreover, a subset of time-
invariant variables is endogenous and it is necessary to specify these variables in endogenous().
The instrumental variables for the endogenous variables must be specified in instrument(). The
instrumental variables can be variables other than the time-varying variables and time-invariant
variables. If the instrumental variables are time-varying variables in the model, the xtfef command
will take the within group time averages of the variables. The time-invariant variables which are
not endogenous should also be included in instrument(). They are the instrumental variables for
themselves.

The coefficients and variances of the first step of the FEF and FEF-IV models can be retrieved
by the ereturn functions e(beta) and e(V beta), respectively. The coefficients and variances
of the second step of the FEF model can be retrieved by e(gamma) and e(V gamma), respec-
tively. The coefficients and variances of the second step of the FEF-IV model can be retrieved by
e(gamma iv) and e(V gamma iv), respectively. The following are ereturn functions for other
saved results,

9
 The ereturn functions
Scalars
e(N) number of observations
e(n) number of groups
e(Tavg) average number of periods per group
Matrix
e(beta) coefficients in the first step of FEF and FEF-IV
e(V beta) variance-covariance matrix in the first step of FEF
e(gamma) coefficients in the second step of FEF
e(V gamma) variance-covariance matrix in the second step of FEF
e(gamma iv) coefficients in the second step of FEF-IV
e(V gamma iv) variance-covariance matrix in the second step of FEF-IV
Functions
e(sample) marks estimation sample

5 Empirical Application
In this application, we consider the return to schooling, where the wage equation is estimated on a
panel of N = 545 individuals from National Longitudinal Surveys of full-time working males over
the period 1980 to 1987. The equation includes five time-varying regressors, namely experience,
experience squared, three dummy variables for the marriage status, whether individual’s wage is
set by a union contract, and location (urban/rural), and three time-invariant regressors, namely
education, and two dummy variables for black and Hispanic. This data set was originally analyzed
by Vella and Verbeek (1998) and was recently used by in Example 59 of Pesaran (2015). The wage
equation to be estimated is given by

2
log (wit ) = αi + β1 U nionit + β2 experit + β3 experit + β4 Ruralit + β5 marriedit
+γ1 educi + γ2 blacki + γ3 hispi + uit ,

where the time-varying regressors are work-experience (experit ) , marriage status (marriedit = 1 if
married at time t), union coverage (U nionit = 1 if the individual’s wage is set by a union contract),
and location (Ruralit = 1 for rural area). The time-invariant variables are years of formal education
(educi ), and dummies for race, distinguishing between blacki and hispi .

5.1 Example 1: FEF estimation

Suppose all time-invariant variables are exogenous, the FEF model is estimated. The log of wage
is the dependent variable. Time-invariant variables include Education, Hispanic, and Black, while
time-varying variables include union, experience, experience squared, rural, and marriage status.
The cross-section identity is nr. The time variable is year. The command of the FEF model is as

10
follows,

. xtfef lnwage, invariant(education hispanic black) varying(union exper exp

> er_sq rural marriage_status) i(nr) t(year)
step1

Fixed-Effects-Filter estimation Number of obs = 4360

Group variable: nr Number of groups = 545

Time variable: year Obs per group: min = 8

avg = 8
max = 8

Time-varying coefficients
lnwage Coef. Std. Err. t

union .0815254 .0227376 3.59

exper .1177284 .0107802 10.92
exper_sq -.0043715 .0006933 -6.31
rural .0492558 .0391552 1.26
marriage_s~s .0453215 .0210285 2.16

Time-invariant coefficients
Coef. Std. Err. t

education .1035798 .0091296 11.35

hispanic .0330597 .0415021 0.80
black -.1398047 .0508217 -2.75

In the xtfef command if the estimation is FEF, i.e. all invariant variables are exogenous, it is
only necessary to specify the time-invariant variables education, Hispanic and black in the option
-invariant()-, time-varying variables union, experience, experience squared, rural, and marriage
status in the option -varying()-, the cross-section identity nr in the option -i()-, and the time vari-
able year in the option -t()-. The estimation result provides the coefficients, standard errors, and
t-statistics.

5.2 Example 2: FEF-IV estimation

Suppose now there is endogeneity in the wage equation, and the time-invariant variable Education
is endogenous. In order to estimate the effect of Education on the wage, we can use the FEF-IV
approach. We assume Hispanic is exogenous and use it as instrument for Education. The command
of FEF-IV for this specification is,

11
. xtfef lnwage, fefiv invariant(education black) varying(union exper exper_
> sq rural marriage_status) endogenous(education) instrument(hispanic black
> ) i(nr) t(year)

Fixed-Effects-Filter estimation Number of obs = 4360

Group variable: nr Number of groups = 545

Time variable: year Obs per group: min = 8

avg = 8
max = 8

Time-varying coefficients
lnwage Coef. Std. Err. t

union .0815254 .0227376 3.59

exper .1177284 .0107802 10.92
exper_sq -.0043715 .0006933 -6.31
rural .0492558 .0391552 1.26
marriage_s~s .0453215 .0210285 2.16

Time-invariant coefficients
Coef. Std. Err. t

education .0710674 .0396054 1.79

black -.1522397 .0515774 -2.95

In this example, the only time-invariant variables are Education and Black with Education being
endogenous. Therefore, Education and Black must be included in -invariant()-, while education
should be included in -endogenous()-. The time-invariant variable, Hispanic, is used as the instru-
mental variable for the endogenous variable Education. Therefore, Hispanic must be dropped from
the option -invariant()- and specified in -instrument()-. Besides Hispanic, the option -instrument()-
must include Black, which serves as an instrumental variable for itself.

6 Conclusion
In this paper, we discuss FEF and FEF-IV estimators of the time-invariant effects in static panels
proposed by Pesaran and Zhou (2017). We provide the command xtfef to implement the FEF and
FEF-IV estimation in Stata. Detailed instruction of the command xtfef is also presented. The
program provides the coefficients and standard errors. The vectors of coefficients and variance-
covariance matrices can be retrieved by using the e() functions. Illustration example of how to use
the command xtfef to empirical data is provided.

12
References
Beck, N. (2011). Of fixed-effects and time-invariant variables. Political Analysis 19, 119–122.

Brusch, T., M. Ward, H. Nguyen, and T. Kompas (2011a). Fevd: Just iv or just mistaken? Political
Analysis 19, 165–169.

Brusch, T., M. Ward, H. Nguyen, and T. Kompas (2011b). On the fixed-effects vector decomposi-
tion. Political Analysis 19, 123–134.

Greene, W. (2011a). Fixed effects vector decomposition: A marginal solution to the problem of
time-invariant variables in fixed effects models? Political Analysis 19, 135–146.

Greene, W. (2011b). Reply to rejoinder by plumper and troeger. Political Analysis 19, 170–172.

Hausman, B. Y. J. A. and W. Taylor (1981). Panel data and unobservale individual effects. Econo-
metrica 49 (1), 1377–1398.

Pesaran, M. H. (2015). Time series and panel data econometrics. Oxford: Oxford University Press.

Pesaran, M. H. and Q. Zhou (2017). Estiamtion of time-invariant effects in static panel data models.
Economic Review .

Plumper, T. and V. E. Troeger (2011). Fixed-effects vector decomposition: Properties, reliability,

and instruments. Political Analysis 19, 147–164.

Vella, F. and M. Verbeek (1998). Whose wages do unions raise? a dynamic model of unionism and
wage rate determination for young men. Journal of Applied Econometrics 13, 163–183.

13