Panel Data
Panel Data
When using panel data a complete set of equations for the system comprises an
equation specified for each of the N cross sections. Such a complete system may
be estimated jointly.
Even though the combination of time series and cross section data produces a
large number, N × T , of observations, panel models also involve a large number of
unknown parameters. These may relate to
1 specific characteristics of each unit or individual in the cross section; or
2 common forms of financial agent responses.
Panel data models lead to large numbers of potentially nuisance parameters –
called incidental parameters – and a much smaller number of homogeneous
parameters that result from commonality restrictions that apply across the
parameters of the N equations for each cross section unit.
Panel models are often classified according to restrictions that are imposed on the
parameters.
1 No common effects model where no restrictions at all are imposed on the coefficients of
the equations.
2 Common effects where all equations have common coefficients imposed.
3 Intermediate cases are where individual specific incidental parameters – usually the
intercept coefficients in the equations – are allowed to differ but parameters on
explanatory variables are constrained.
The common effects model involves extreme pooling of information. Each equation
of this model is assumed to have the same intercept and same slope coefficients
where the regression coefficients are the same for each equation j = 1, · · · , N but
equation error variances σj2 are heterogeneous.
Tthe common effects CAPM has the explicit form
To estimate the common effects model with the restrictions imposed the approach
adopted is to stack the data. This is equivalent to expressing the data in long form.
For the CAPM example, stacking requires that the dependent variable rjt − rft of
each stock are stacked on top of each other to produce a new variable that is an
(NT × 1) vector of excess returns on all N stocks over all time periods.
Stacking the explanatory variable is achieved in exactly the same way except that
the excess returns on the market are the same for all stocks, so the stacked
explanatory variable simply stacks the same variable on itself N times to produce
an (NT × 1) vector of excess returns of the market.
Estimation then involves a linear regression of the stacked dependent variable on
a constant vector – represented by an (NT × 1) vector of ones – and the stacked
explanatory variable.
This process of stacking the data and applying least squares to the stacked
system imposes the cross-equation restrictions that the α-risks and the β -risks
are the same across all N stocks.
The simple bivariate panel regression then has the following form
where the simplifying assumption of constant equation error variance over the
cross sections is made, so that σ12 = σ22 = · · · = σN2 = σ 2 .
In this specification the αj are considered fixed parameters by assumption, thereby
giving the model its name. The fixed effects model is characterised by N
intercepts (αj : j = 1, · · · , N) and 1 slope parameter (β). As discussed earlier, the
intercepts are known as incidental parameters.
The fixed effects model imposes N − 1 restrictions on the no common effects
model by requiring the slope coefficient β to be common in each equation.
Alternatively, a comparison of the fixed effects model and the common effects
model suggests that the common effects model imposes N − 1 restrictions on the
fixed effects model by requiring the constant term to α to be common across all
equations.
An alternative way to specify the fixed effects model is to define a set of dummy
variables as
1 cross section j
Djt =
0 otherwise,
The fixed effects model is now rewritten as
or more compactly
N
X
yjt = αj Djt + βxjt + ujt .
j=1
The fixed effects model is estimated by an ordinary least squares regression of the
(NT × 1) stacked dependent variable yjt on the (NT × 1) stacked explanatory
variable xjt and the N stacked dummy variables, D1t , D2t , · · · , DNt . Alternatively, if a
constant is included in the regression equation then one of the dummy variables is
excluded to avoid the dummy variable trap of collinear regressors.
The dummy variable approach is only feasible in long panels as the number of
dummy variables required by a wide panel would make the approach inefficient.
yj = αj + βx j + u j .
bj = y i − βx
α b j,
Estimation of the random effects model is somewhat more involved than the fixed
effects model. The complexity arises from the composite form of the equation
disturbance term wjt .
The covariance structure of wjt is
cov(wjt wjs ) σ2
ρts = p = 2 v 2.
var(wjt )var(wjs ) σ + σv
The steps involved to estimate the parameters of the random effects model in this more
general case are as follows.
1 Choose starting values for σ 2 and σv2 .
2 Define the quasi-difference parameter
σ
λ=1− p ,
σ 2 + T σv2
and compute the quasi-differenced data
in which w
ejt is a disturbance term, by ordinary least squares.
A formal test of fixed effects versus random effects in modelling panel data can be
based on the fundamental assumption that the random effects, vj , are considered
to be independent of the explanatory variables, xjt , that is E(vj xjt ) = 0.
Formally the hypotheses are
H0 : random effects
H1 : fixed effects.
To perform the test, let βbFE and βbRE represent, respectively, the fixed and random
effects estimates of the slope parameters.
The general form of the test is based on the Wald statistic
h i0 h i−1 h i
WD = βbFE − βbRE cov(βbFE ) − cov(βbRE ) βbFE − βbRE ,
where cov(βbFE ) and cov(βbRE ) are the estimated covariances of βbFE and βbRE .
The test statistic WD is distributed asymptotically under the null hypothesis as χ2R
where R represents the number of explanatory variables in the model excluding
the constant.
The panel data models discussed so far specify a relationship between the
dependent variable, yjt , and the explanatory variable, xjt , both at time t.
Dynamics may be introduced into the model by including lags of both yjt and xjt .
The respective fixed effects and random effects versions of the so-called panel
autoregressive model are as follows:
where ujt is the disturbance term which is assumed to be independent over time
E(ujt ujs ) = 0, for all t 6= s.
The autoregressive parameter ρ controls the strength of the dynamics and, in this
formulation of the model, is assumed to take a common value across section.
1 If ρ = 0 the model reduces to the non-dynamic or static panel data model.
2 If −1 < ρ < 1 the autoregressive dynamics are stationary with the successive effects of
shocks on the dependent variable dissipating over time.
3 If ρ = 1 the shocks have persistent effects and do not dissipate over time, resulting in a
nonstationary panel model.
σ2
cov(y jt−1 , ujt ) = 6= 0, for all t.
T
This correlation violates one of the key conditions needed for the validity of least
squares. In particular, this condition is needed to ensure consistent estimation of
the autoregressive coefficient ρ as N → ∞ with finite time series sample size T .
This non-zero correlation is present even when the disturbance term, ujt , is itself
not autocorrelated.
The resultant bias in estimating ρ is commonly referred to as Nickell bias, following
its original analysis by Nickell (1981). It has the asymptotic form
1+ρ
ρ − ρ) ' −
plim (b .
N→∞ T
The higher the value of σs2 the more useful xjt is in explaining yjt .
The settings T = {10, 20, 30} reveal the performance of the fixed effects estimator
in short, medium and long panels; and the settings N = {20, 100} illustrate
estimator performance in narrow and wide panels. Each parameter combination is
assessed using 1000 replications.
Results should demonstrate that the fixed effects estimator of ρ suffers from
significant bias and that the bias of the fixed effects estimator increases with ρ and
decreases as T increases.
Financial Econometric Modeling c
Hurn, Martin, Phillips & Yu Oxford University Press, 2020 21 / 31
Results
N T ρ = 0.2 ρ = 0.8
Mean Std. Dev. Mean Std. Dev.
20 10 0.138 0.051 0.560 0.070
20 20 0.175 0.034 0.693 0.040
20 30 0.181 0.027 0.731 0.029
100 10 0.140 0.023 0.567 0.030
100 20 0.172 0.015 0.694 0.018
100 30 0.182 0.012 0.733 0.013
If the autoregressive coefficient ρ in a dynamic panel model does not satisfy the
stability restriction |ρ| < 1, the variables for each cross section are nonstationary
and may be cointegrated with common long run parameters.
An example is the present value model applied to a cross section of countries.
The question is whether these countries have a common cointegrating vector that
embodies the same present value relationship.
A general specification of the present value model for a set of N, countries is
written as
pjt = αj + βj djt + ujt , j = 1, 2, · · · , N,
where pjt is the log equity price of country j, djt is the log dividend of country j and
ujt is a disturbance term.
Assuming that pjt and djt are nonstationary time series and ujt is stationary, the
parameters αj and βj represent the cointegrating parameters which in this
formulation may vary for each country. T
The parameter αj is a function of the relevant discount factor that relates to
country j.
This model is comparable to the no common effects panel model with the
parameters for each cross section being estimated separately using appropriate
methods.
Financial Econometric Modeling c
Hurn, Martin, Phillips & Yu Oxford University Press, 2020 25 / 31
Panel Scatter Plots of the Present Value Model for 12 Countries
Log Price
Log Price
8.5 10.0 9.0
8.0 9.0 8.5
7.5 8.0 8.0
4 4.5 5 5.5 3 4 5 6 7 4 4.5 5 5.5 6
Log Dividends Log Dividends Log Dividends
China France Germany
7.0 9.0 9.5
6.5 8.8
Log Price
Log Price
Log Price
8.6 9.0
6.0
8.4 8.5
5.5 8.2
5.0 8.0 8.0
Log Price
Log Price
8.0 4.8 7.5
4.6 7.0
7.5
4.4 6.5
7.0 4.2 6.0
3.5 4 4.5 5 -.5 0 .5 1 2.5 3 3.5 4 4.5
Log Dividends Log Dividends Log Dividends
South Korea U.K. U.S.
1.0 9.6 7.6
0.5 9.4 7.4
Log Price
Log Price
β1 = β2 = · · · = βN = β,
imposes a common long run relationship between equity prices and dividends.
The result is a present value fixed effects cointegrating model with homogeneous
slopes of the form
pjt = αj + βdjt + ujt .
The fixed effects are given by the parameters {α1 , α2 , · · · , αN }, which allow the
discount factor for each country to differ.
A further set of restrictions is to assume that the discount factors are also the
same across all N countries by imposing the restrictions
α1 = α2 = · · · = αN = α.
Rather than testing each variable separately for a unit root, a test statistic is
constructed to assess the presence of a unit root jointly among all N variables (or
cross sections) simultaneously. This strategy can be interpreted as an efficient
procedure of testing for the presence of a time series unit root in a variable that
varies not only over time but also over N cross sections.
The Levin, Lin and Chu (2002) test is based on the panel unit root regression
equation
XL
∆yjt = αj + βyj t−1 + γjk ∆yj t−k + ujt .
k=1
The null and alternative hypotheses of the panel unit root test are given by
H0 : β=0 [unit root]
H1 : β<0 [stationary].
Note that the null hypothesis requires that all time series in the panel have unit
roots (ρ := β + 1 = 1) and the alternative similarly requires that all time series
have stationary roots (|ρ| < 1).
This model is the panel analogue of the augmented Dickey-Fuller regression
equation with β = ρ − 1. Note however that the panel unit root t statistic is
asymptotically distributed as N(0, 1) under the null hypothesis.
Financial Econometric Modeling c
Hurn, Martin, Phillips & Yu Oxford University Press, 2020 28 / 31
Panel Unit Root Tests - IPS
A less restrictive panel unit root test that allows for some stationary and
nonstationary elements in the cross section under the alternative hypothesis was
proposed by Im, Pesaran and Shin (2003).
This test involves the weaker formulation
XL
∆yjt = αj + βj yj t−1 + γjk ∆yj t−k + ujt ,
k=1
As with cointegration testing in non-panel data models there exists a range of test
procedures for cointegration in dynamic panel models. Many of these tests rely on
Fisher’s (1932) idea of combining individual tests to produce a meta-analysis test.
If πj is the p value of the test for cross section j for a particular hypothesis, the joint
statistic based on the N cross sections is
N
X
−2 log (πj ) → χ22N ,
j=1