Panel Data 4: Fixed Effects vs Random Effects Models

Richard Williams, University of Notre Dame, https://www3.nd.edu/~rwilliam/

Last revised March 20, 2018

These notes borrow very heavily, sometimes verbatim, from Paul Allison’s book, Fixed Effects Regression Models
for Categorical Data. The Stata XT manual is also a good reference. This handout tends to make lots of assertions;
Allison’s book does a much better job of explaining why those assertions are true and what the technical details
behind the models are.

Overview. With panel/cross sectional time series data, the most commonly estimated models are
probably fixed effects and random effects models. Population-Averaged Models and Mixed
Effects models are also sometime used. In this handout we will focus on the major differences
between fixed effects and random effects models.

Several considerations will affect the choice between a fixed effects and a random effects model.

1. What is the nature of the variables that have been omitted from the model?
a. If you think there are no omitted variables – or if you believe that the omitted
variables are uncorrelated with the explanatory variables that are in the model –
then a random effects model is probably best. It will produce unbiased estimates
of the coefficients, use all the data available, and produce the smallest standard
errors. More likely, however, is that omitted variables will produce at least some
bias in the estimates.
b. If there are omitted variables, and these variables are correlated with the variables
in the model, then fixed effects models may provide a means for controlling for
omitted variable bias. In a fixed-effects model, subjects serve as their own
controls. The idea/hope is that whatever effects the omitted variables have on the
subject at one time, they will also have the same effect at a later time; hence their
effects will be constant, or “fixed.” HOWEVER, in order for this to be true, the
omitted variables must have time-invariant values with time-invariant effects.
i. By time-invariant values, we mean that the value of the variable does not
change across time. Gender and race are obvious examples, but this can
also include things like the Educational Level of the Respondent’s Father.
ii. By time-invariant effects, we mean the variable has the same effect across
time, e.g. the effect of gender on the outcome at time 1 is the same as the
effect of gender at time 5.
iii. If either of these assumptions is violated, we need to have explicit
measurements of the variables in question and include them in our models.
In the case of time-varying effects, we can include things like the
interaction of gender with time. We also need explicit measurements of
time-invariant variables if they are thought to interact with other variables
in the model, e.g. we think the effect of SES differs by race.
2. How much variability is there within subjects?
a. If subjects change little, or not at all, across time, a fixed effects model may not
work very well or even at all. There needs to be within-subject variability in the
variables if we are to use subjects as their own controls. If there is little variability

Panel Data 4: Fixed Effects vs Random Effects Models Page 1

 within subjects then the standard errors from fixed effects models may be too
large to tolerate.
b. Conversely, random effects models will often have smaller standard errors. But,
the trade-off is that their coefficients are more likely to be biased.
3. Do we wish to estimate the effects of variables whose values do not change across time,
or do we merely wish to control for them?
a. With fixed effects models, we do not estimate the effects of variables whose
values do not change across time. Rather, we control for them or “partial them
out.” This is similar to an experiment with random assignment. We may not
measure variables like SES, but whatever effects those variable have are (subject
to sampling variability) assumed to be more or less the same across groups
because of random assignment.
b. Random effects models will estimate the effects of time-invariant variables, but
the estimates may be biased because we are not controlling for omitted variables.

Fixed effects models. Allison says “In a fixed effects model, the unobserved variables are
allowed to have any associations whatsoever with the observed variables.” Fixed effects models
control for, or partial out, the effects of time-invariant variables with time-invariant effects. This
is true whether the variable is explicitly measured or not. Exactly how they do so varies by the
statistical technique being used. The optional appendix discusses these methods further.
Unfortunately, the effects of time-invariant variables that are measured cannot be estimated.
. use https://www3.nd.edu/~rwilliam/statafiles/teenpovxt, clear
. *fixed effects
. xtlogit pov i.mother i.spouse i.school hours i.year i.black age, fe nolog
note: multiple positive outcomes within groups encountered.
note: 324 groups (1,620 obs) dropped because of all positive or
all negative outcomes.
note: 1.black omitted because of no within-group variance.
note: age omitted because of no within-group variance.

Conditional fixed-effects logistic regression Number of obs = 4,135

Group variable: id Number of groups = 827

Obs per group:

min = 5
avg = 5.0
max = 5

LR chi2(8) = 97.28
Log likelihood = -1520.1139 Prob > chi2 = 0.0000

------------------------------------------------------------------------------
pov | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
1.mother | .5824322 .1595831 3.65 0.000 .269655 .8952094
1.spouse | -.7477585 .1753466 -4.26 0.000 -1.091431 -.4040854
1.school | .2718653 .1127331 2.41 0.016 .0509125 .4928181
hours | -.0196461 .0031504 -6.24 0.000 -.0258208 -.0134714
|
year |
2 | .3317803 .1015628 3.27 0.001 .132721 .5308397
3 | .3349777 .1082496 3.09 0.002 .1228124 .547143
4 | .4327654 .1165144 3.71 0.000 .2044013 .6611295
5 | .4025012 .1275277 3.16 0.002 .1525514 .652451
|
1.black | 0 (omitted)
age | 0 (omitted)
------------------------------------------------------------------------------

Panel Data 4: Fixed Effects vs Random Effects Models Page 2

Random Effects Models. Quoting Allison, “In a random effects model, the unobserved
variables are assumed to be uncorrelated with (or, more strongly, statistically independent of) all
the observed variables.” That assumption will often be wrong but, for the reasons given above
(e.g. standard errors may be very high with fixed effects, RE lets you estimate effects for time-
invariant variables), an RE model may still be desirable under some circumstances. RE models
can be estimated via Generalized Least Squares (GLS). Here is an example of a random effects
logistic regression model.

. *random effects
. xtlogit pov i.mother i.spouse i.school hours i.year i.black age, re nolog

Random-effects logistic regression Number of obs = 5,755

Group variable: id Number of groups = 1,151

Random effects u_i ~ Gaussian Obs per group:

min = 5
avg = 5.0
max = 5

Integration method: mvaghermite Integration pts. = 12

Wald chi2(10) = 266.60

Log likelihood = -3403.7655 Prob > chi2 = 0.0000

------------------------------------------------------------------------------
pov | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
1.mother | 1.009877 .118372 8.53 0.000 .7778724 1.241882
1.spouse | -1.171833 .1512544 -7.75 0.000 -1.468286 -.8753802
1.school | -.1145721 .0990775 -1.16 0.248 -.3087604 .0796163
hours | -.0259014 .0028771 -9.00 0.000 -.0315403 -.0202624
|
year |
2 | .2830958 .1000437 2.83 0.005 .0870138 .4791778
3 | .213423 .1040523 2.05 0.040 .0094842 .4173618
4 | .2415184 .1090094 2.22 0.027 .0278639 .455173
5 | .1447937 .1161395 1.25 0.212 -.0828355 .372423
|
1.black | .6093942 .0975653 6.25 0.000 .4181698 .8006186
age | -.0627952 .0472163 -1.33 0.184 -.1553373 .029747
_cons | -.0045847 .7620829 -0.01 0.995 -1.49824 1.48907
-------------+----------------------------------------------------------------
/lnsig2u | .3086358 .1008833 .1109083 .5063634
-------------+----------------------------------------------------------------
sigma_u | 1.166862 .0588584 1.057021 1.288117
rho | .2927197 .0208864 .2535175 .3352612
------------------------------------------------------------------------------
LR test of rho=0: chibar2(01) = 327.62 Prob >= chibar2 = 0.000

Among other things, according to this model, blacks are significantly more likely to be in
poverty than are whites. The highly significant likelihood ratio test at the end tells us it would not
be appropriate to use regular logistic regression instead. Note too that there are some major
differences in the coefficients for the fixed and random effects models, which might reflect the
importance of omitted variable bias in the latter.

Panel Data 4: Fixed Effects vs Random Effects Models Page 3

Mixed Effects Model. Give or take a few decimal places, a mixed-effects model (aka multilevel
model or hierarchical model) replicates the above results. Again, it is ok if the data are xtset
but it is not required. We will (hopefully) explain mixed effects models more later.

. * Equivalent mixed-effects model

. xtset, clear
. melogit pov i.mother i.spouse i.school hours i.year i.black age || id:, nolog

Mixed-effects logistic regression Number of obs = 5,755

Group variable: id Number of groups = 1,151

Obs per group:

min = 5
avg = 5.0
max = 5

Integration method: mvaghermite Integration pts. = 7

Wald chi2(10) = 266.64

Log likelihood = -3403.7637 Prob > chi2 = 0.0000
------------------------------------------------------------------------------
pov | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
|
1.mother | 1.009935 .1183721 8.53 0.000 .7779301 1.24194
1.spouse | -1.171859 .1512457 -7.75 0.000 -1.468295 -.8754231
1.school | -.114617 .0990711 -1.16 0.247 -.3087927 .0795587
hours | -.0259016 .0028769 -9.00 0.000 -.0315403 -.0202629
|
year |
2 | .2830838 .1000419 2.83 0.005 .0870052 .4791624
3 | .2134042 .10405 2.05 0.040 .00947 .4173385
4 | .2414921 .1090061 2.22 0.027 .027844 .4551401
5 | .144759 .1161351 1.25 0.213 -.0828617 .3723796
|
1.black | .6094854 .0975621 6.25 0.000 .4182672 .8007036
age | -.0628037 .0472134 -1.33 0.183 -.1553403 .029733
_cons | -.0045483 .7620352 -0.01 0.995 -1.49811 1.489013
-------------+----------------------------------------------------------------
id |
var(_cons)| 1.361483 .1371712 1.117513 1.658715
------------------------------------------------------------------------------
LR test vs. logistic model: chibar2(01) = 327.62 Prob >= chibar2 = 0.0000

Panel Data 4: Fixed Effects vs Random Effects Models Page 4

 Appendix (Optional): Estimation methods for fixed-effects models
Fixed effects models control for, or partial out, the effects of time-invariant variables with time-
invariant effects. This is true whether the variable is explicitly measured or not. Exactly how
they do so varies by the statistical technique being used. Some of the methods used include
• Demeaning variables. The within-subject means for each variable (both the Xs and the
Y) are subtracted from the observed values of the variables. Hence, within each subject,
the demeaned variables all have a mean of zero. For time-invariant variables, e.g. gender,
the demeaned variables will have a value of 0 for every case, and since they are constants
they will drop out of any further analysis. This basically gets rid of all between-subject
variability (which may be contaminated by omitted variable bias) and leaves only the
within-subject variability to analyze. This method works for linear regression models but
does not work for things like logistic regression.
• Unconditional maximum likelihood. With UML, dummy variables are created for each
subject (except one) and included in the model. So, for example, if you had 2000 subjects
each of whom was measured at 5 points in time, you would include 1,999 dummy
variables in the model. Needless to say, this can be pretty time consuming, and can
produce a lot of coefficients that you aren’t really interested in! However, Allison argues
that it is better to use nbreg with UML than it is to use Stata’s xtnbreg, fe. The
latter, he claims, uses a flawed approach and does not, in fact control for all stable
predictors. UML can also be used for linear regression but produces biased estimates with
logistic regression.
• Conditional maximum likelihood. This is used for logistic regression and some other
statistical techniques. Quoting Allison (p. 32; αi refers to the fixed effects parameters),

Note that, with the conditional logit model, for all subjects where the dependent variable
is a constant (e.g. at all five time periods the subject has a value of 1 on the dependent variable,
or a value of zero) the case is dropped from the statistical analysis. Basically, there is no
alternative possibility to compare to, e.g. the only way you can have 5 ones is by being a one at
every time period.

Panel Data 4: Fixed Effects vs Random Effects Models Page 5

Before proceeding, we will show examples of UML (the dummy variable for each case
approach). This will show that regress using UML gives the same results as xtreg,fe but
different results when using logit and xtlogit, fe. The data sets used here are also used
in Allison’s book.

. set more off

. use https://www3.nd.edu/~rwilliam/statafiles/nlsy.dta, clear
. des anti* self* pov*

storage display value

variable name type format label variable label
--------------------------------------------------------------------------------------
--------------------------------------------------------------
anti90 byte %8.0g child antisocial behavior in 1990
anti92 byte %8.0g child antisocial behavior in 1992
anti94 byte %8.0g child antisocial behavior in 1994
self90 byte %8.0g child self-esteem in 1990
pov90 byte %8.0g family poverty status in 1990
[some output deleted]

. gen id=_n
. reshape long anti pov self, i(id) j(year)
(note: j = 90 92 94)

Data wide -> long

-----------------------------------------------------------------------------
Number of obs. 581 -> 1743
Number of variables 17 -> 12
j variable (3 values) -> year
xij variables:
anti90 anti92 anti94 -> anti
pov90 pov92 pov94 -> pov
self90 self92 self94 -> self
-----------------------------------------------------------------------------

. xtset id year
panel variable: id (strongly balanced)
time variable: year, 90 to 94, but with gaps
delta: 1 unit

. * UML works fine with linear regression model

. xtreg anti self pov i.year, fe

Fixed-effects (within) regression Number of obs = 1743

Group variable: id Number of groups = 581

R-sq: within = 0.0331 Obs per group: min = 3

between = 0.0418 avg = 3.0
overall = 0.0359 max = 3

F(4,1158) = 9.92
corr(u_i, Xb) = 0.0683 Prob > F = 0.0000

Panel Data 4: Fixed Effects vs Random Effects Models Page 6

------------------------------------------------------------------------------
anti | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
self | -.0551514 .0105258 -5.24 0.000 -.0758031 -.0344997
pov | .1124749 .0934099 1.20 0.229 -.0707967 .2957464
|
year |
92 | .0443934 .058584 0.76 0.449 -.0705493 .159336
94 | .2107366 .0587978 3.58 0.000 .0953744 .3260987
|
_cons | 2.637156 .2173038 12.14 0.000 2.210803 3.06351
-------------+----------------------------------------------------------------
sigma_u | 1.3218868
sigma_e | .99707353
rho | .63737335 (fraction of variance due to u_i)
------------------------------------------------------------------------------
F test that all u_i=0: F(580, 1158) = 5.16 Prob > F = 0.0000

. set matsize 2000

. reg anti self pov i.year i.id

Source | SS df MS Number of obs = 1743

-------------+------------------------------ F(584, 1158) = 5.48
Model | 3181.88311 584 5.44842999 Prob > F = 0.0000
Residual | 1151.23221 1158 .994155619 R-squared = 0.7343
-------------+------------------------------ Adj R-squared = 0.6003
Total | 4333.11532 1742 2.48743704 Root MSE = .99707

------------------------------------------------------------------------------
anti | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
self | -.0551514 .0105258 -5.24 0.000 -.0758031 -.0344997
pov | .1124749 .0934099 1.20 0.229 -.0707967 .2957464
|
year |
92 | .0443934 .058584 0.76 0.449 -.0705493 .159336
94 | .2107366 .0587978 3.58 0.000 .0953744 .3260987
|
id |
2 | -.8875251 .8194485 -1.08 0.279 -2.495295 .7202448
3 | 4.130859 .8194591 5.04 0.000 2.523068 5.738649
[Rest of coefficients for dummy variables for ids are deleted]

. * UML does not work fine with logit -- Need conditional model instead
. xtlogit pov mother spouse school hours i.year, fe nolog
note: multiple positive outcomes within groups encountered.
note: 324 groups (1620 obs) dropped because of all positive or
all negative outcomes.

Conditional fixed-effects logistic regression Number of obs = 4135

Group variable: id Number of groups = 827

Obs per group: min = 5

avg = 5.0
max = 5

LR chi2(8) = 97.28
Log likelihood = -1520.1139 Prob > chi2 = 0.0000

Panel Data 4: Fixed Effects vs Random Effects Models Page 7

------------------------------------------------------------------------------
pov | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mother | .5824322 .1595831 3.65 0.000 .269655 .8952094
spouse | -.7477585 .1753466 -4.26 0.000 -1.091431 -.4040854
school | .2718653 .1127331 2.41 0.016 .0509125 .4928181
hours | -.0196461 .0031504 -6.24 0.000 -.0258208 -.0134714
|
year |
2 | .3317803 .1015628 3.27 0.001 .132721 .5308397
3 | .3349777 .1082496 3.09 0.002 .1228124 .547143
4 | .4327654 .1165144 3.71 0.000 .2044013 .6611295
5 | .4025012 .1275277 3.16 0.002 .1525514 .652451
------------------------------------------------------------------------------

. logit pov mother spouse school hours i.year i.id, nolog

note: 141.id != 0 predicts failure perfectly
141.id dropped and 5 obs not used
note: 298.id != 0 predicts success perfectly
298.id dropped and 5 obs not used

[Other similar warnings deleted – these are the 324 cases where the outcome is the
same at all 5 time periods for the case]

Logistic regression Number of obs = 4135

LR chi2(834) = 998.93
Prob > chi2 = 0.0001
Log likelihood = -2304.2196 Pseudo R2 = 0.1781

------------------------------------------------------------------------------
pov | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mother | .7341873 .179498 4.09 0.000 .3823778 1.085997
spouse | -.9407072 .1971326 -4.77 0.000 -1.32708 -.5543344
school | .3410341 .1264389 2.70 0.007 .0932184 .5888497
hours | -.0246849 .0035439 -6.97 0.000 -.0316308 -.0177391
|
year |
2 | .4196558 .1142231 3.67 0.000 .1957827 .643529
3 | .4218788 .121389 3.48 0.001 .1839608 .6597968
4 | .5452897 .1306011 4.18 0.000 .2893163 .8012631
5 | .5071969 .1427835 3.55 0.000 .2273463 .7870475
|
id |
75 | -.107972 1.592235 -0.07 0.946 -3.228695 3.012751
92 | 1.206116 1.476275 0.82 0.414 -1.68733 4.099562
[Coefficients for other id dummies not shown]

Panel Data 4: Fixed Effects vs Random Effects Models Page 8