Scribd
Upload
78 views

Panel Guidelines

This document provides guidance on panel data analysis. It discusses different panel data models including fixed effects, random effects, and two-way models. It recommends conducting tests like the F test, Breusch-Pagan LM test, and Hausman test to determine the appropriate model. The document also provides guidance on reporting results, interpreting coefficients, and considering alternative functional forms like log-linear models.

Uploaded by

PatofnaPatofnic
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
78 views

Panel Guidelines

This document provides guidance on panel data analysis. It discusses different panel data models including fixed effects, random effects, and two-way models. It recommends conducting tests like the F test, Breusch-Pagan LM test, and Hausman test to determine the appropriate model. The document also provides guidance on reporting results, interpreting coefficients, and considering alternative functional forms like log-linear models.

Uploaded by

PatofnaPatofnic
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Practical Guides To Panel Data Analysis

Hun Myoung Park
05/16/2010

1. Which effect? Group vs. Time? Fixed vs. Random?


Panel data models examine cross-sectional (group) and/or time-series (time) effects. These effects may
be fixed and/or random. Fixed effects assume that individual group/time have different intercept in the
regression equation, while random effects hypothesize individual group/time have different disturbance.
When the type of effects (group versus time) and property of effects (fixed versus random) combined,
there are several specific models: fixed group effect model (one-way), fixed time effect model (one-way),
fixed group and time effect model (two-way), random group effect model (one-way), random time effect
(one-way), and random group and time effect model (two-way).
As a result, you have to select one specific model. Presenting all possible models sounds quite absurd;
strictly speaking, if one model is right, the other models are wrong. It must be ridiculous to present wrong
models as well unless you want to go “data-fishing.”

2. Determine Your Model


In order to determine your model, conduct necessary tests: F test for the fixed effect model and
Breusch-Pagan Lagrange Multiplier (LM) test. Hausman test is needed if you find both fixed and random
effects. See the next section for the test. Once you finish F and LM tests, determine your model as shown in
the following table.
Fixed effect Random effect Your model
(F test or Wald test) (Breusch-Pagan LM test)

H0 is not rejected H0 is not rejected Data are poolable. Pooled OLS


(No fixed effect) (No random effect)

H0 is rejected H0 is not rejected Fixed effect model


(fixed effect) (No random effect)

H0 is not rejected H0 is rejected Random effect model


(No fixed effect) (random effect)
1
H0 is rejected H0 is rejected (1) Fixed and random effect model or
(fixed effect) (random effect) (2) choose one of the two depending on the result
of Hausman test (recommended direction).
If you have a small number of observations in you data set, avoid two-way models especially when a
group or time has many different values (e.g., 100 countries in group and 400 quarters in time). Why? The
null hypothesis of fixed or random effect model is almost useless; almost all cases end up rejecting H 0.

1 In theory, it is not allowed to impose both fixed and random effects for a group or time variable. For instance, it is
not possible to have fixed group (say, urban versus rural) and random group effect simultaneously. Imposing two
effects of a group or time is contradictory conceptually. However, you may fit a model with a fixed group effect and
random time effect (or vice versa) using both least squares dummy variable (LSDV) model and a random effect
model. This model is possible but least recommended largely due to the loss of parsimony and degrees of freedom.

1/3
3. F Test (Wald Test) for Fixed Effects
F test reported in the output of the fixed effect model is for overall goodness-of-fit, not for the test of the
fixed effect. In order to test fixed effect, run .test command in Stata after fitting the least squares dummy
variable model with .regress (not .xtreg). For example, if you have 3 dummies variables for 4 groups,
you may run the command as, “.test d1 d2 d3”

4. Breusch-Pagan Lagrange Multiplier (LM) Test for Random Effect


.xtreg conducts Wald test to report the goodness-of-fit but this is not for random effect. Use
.xttest0 command to test if there is any random effect.

5. Hausman Test for Comparing Fixed and Random Effects


Hausman test compares the fixed and random effect models. If both fixed and random effects turn out
significant, Hausman test will give you a good idea when choosing one between the two. The null
hypothesis is that the (fixed or random) effect is not correlated with other regressors (independent
variables). “A random effect model is better than the fixed effect model” and “A random effect model is
consistent” are NOT CORRECT null hypotheses for the Hausman test.
If the null hypothesis is rejected, a random effect model will be suffering from the violation of the Gauss-
Markov theorem and end up with biased and inconsistent estimates; by contrast, a fixed effect model still
remains unbiased and consistent. To make it simple, if the null hypothesis is rejected, use the fixed effect
model; otherwise, go for the random effect model.

2
6. Reporting R (Stata Specific)
Some software packages do not report correct statistics. Stata (up to version 11) .xtreg returns wrong R 2
in the fixed effect model because the command fits the within model (running OLS on transformed data
2 2 2
with the intercept suppressed) without adjusting R . And other two R (between and overall R ) are almost
2
meaningless. .xtreg also report incorrect (a bit different) R in random effect models. Therefore, DO NOT
2
REPORT any R from the output of the fixed effect model that Stata produces unless Stata revises the
2 2
command to report the correct R . In order to get correct R for the fixed effect model, use .areg command
in Stata, SAS TSCSREG or PANEL produce, or LIMDEP.

7. Presenting the Results


You need to report parameter estimates and their standard errors. Of course, present proper goodness-
2
of-fit measures (e.g., F, R ) as well. In a random effect model, you also need to report estimated theta and
variance component for error. In F-test and LM test, do not forget to report degrees of freedom.
If you fit a fixed effect model, you may report parameter estimate for constant term of each group or
time. This is not always required, but it depends on your research questions. For instance, if you want to
show which group (or year) has the largest initial status and this issue is essential in your research,
individual intercepts in addition to parameter estimates of regressors should be reported. In this case, do
not simply present estimated constant term and dummy parameters.

2/3
8. Interpretation
2
Evaluate the overall goodness-of-fit measures (e.g., F, R ) of the model. If the model does not fit the data
well. You must stop here and report no significant effects of independent variables.
Interpret parameter estimates (coefficients) in a substantive way. Do not simply report signs and
magnitude of coefficients. Do not simply say “significant,” “negatively (or positively) related”, or
“insignificant relationship between...” One recommended example is “For one unit increase in a IV, DV is
expected to increase by OO units, holding all other variables constant.”

9. Comments on Functional Forms


Many students use log linear regression model to take advantage of getting growth rate. I guess this
biased practice is pervasive in economics. In fact, it is a nonlinear model, however. You SHOULD report
the original equation. Many studies have unstructured equations that may be estimated by simple OLS
without log transformation. In this case, estimation and interpretation become much easier; growth rate
needs to be computed after estimation though. Think about this functional form as an alternative.
However, I am not saying that log linear regression is wrong though.

• ln_Y = b0 + b1*ln_X1 + b2*ln_X2 + b3*ln_X3 (log linear equation)


• Y=(b0)*(X1^b1)*(X2^b2)*(X3^b3) (original nonlinear equation)
• Y=b0 + b1*X1 + b2*X2 + b3*X3 (alternative equation)

Some References for Panel Data Analysis


Baltagi, Badi H. 2001. Econometric Analysis of Panel Data. Wiley, John & Sons.
Greene, William H. 2003. Econometric Analysis, 5th ed. Upper Saddle River, NJ: Prentice Hall.
Wooldridge, Jeffrey M. 2002. Econometric Analysis of Cross Section and Panel Data . Cambridge, MA: MIT
Press.
http://www.indiana.edu/~statmath/stat/all/panel/panel.pdf

3/3

You might also like