Problem Set 5

ECON 231W

Due Friday November 6th at 12:30pm.

Type it nicely (Latex or Word with equation editor).

Upload the Word or pdf file in blackboard.
Scanned handwritten problem sets are not allowed and will not be graded.

1. You are using data collected from 545 men who worked every year between 1980 and 1987.
The included variables are as follows
variable description

id person identifier
year 1980 to 1987
lwage log(wage)
le log(labor mkt experience)
black Dummy, =1 if black
hisp Dummy, =1 if Hispanic
married Dummy, =1 if married
educ years of schooling
union Dummy, =1 if worked belongs to a union
d81 Dummy, =1 if year = 1981
d82 etc
...
d87
In what follows we will assume that the error terms in our regressions are homoskedastic
(recall that the Hausman test that we saw in class is valid only when we the homoskedasticity
assumption hold). Here we want to study the relation between log(wage), education, and
experience, controlling for some other variables. We assume that the true regression is
87
X
ln (wageit ) = β0 + β1 educit + β2 leit + γY dY + other controls + αi + uit (1)
Y =81

where αi is the unobserved fixed effect. Note that the model includes year-specific dummies.

(a) The sample includes data from 1980 to 1987, but equation (1) includes only dummies for
years 1981 to 1987. Why?
(b) Before dealing with panel data, we have typically assumed that observations are i.i.d..
You know that independent observations are uncorrelated, so, if you can show that two
observations are correlated, you showed that they cannot be independent, and then they
cannot be i.i.d.. Here the error term for worker i in year t is vit = αi + uit , while in year
s the error is vis = αi + uis . Assume that both the fixed effect αi and the other error
component uis have zero expected value, and you can assume that cov (uit , ujs ) = 0 for
every i, j, t, s (unless, of course, i = j and t = s !!). Assume also that the fixed effects are
uncorrelated with all the u’s. Calculate cov (vit , vis ). Is it equal to zero? Do you think
that in a panel data the assumption that observations are i.i.d. is a good one?

1
 You estimate equation (1) using RE (Random Effects), and FE (Fixed Effects) and you
get the following results (standard errors are shown to the right of the corresponding
coefficient):
(1) (2)
Random Effect Fixed Effect

Variable Coeff. s.e. Coeff. s.e.

educ 0.104 0.024

le 0.301 0.092 0.176 0.116
black -0.125 0.139
hisp 0.104 0.162
married 0.067 0.038 0.003 0.042
union 0.050 0.041 0.015 0.044
d81 0.039 0.058 0.088 0.063
d82 -0.020 0.074 0.068 0.086
d83 -0.031 0.089 0.084 0.106
d84 -0.055 0.102 0.083 0.124
d85 -0.019 0.113 0.140 0.139
d86 -0.010 0.124 0.168 0.153
d87 0.081 0.133 0.276 0.165
constant -0.123 0.316

(c) Using the RE results in column (1), interpret the coefficient related to the variables educ
and le.
(d) Using again RE, you run a test for the joint significance of all the year-specific dummies,
and the F -test is equal to 1.576. What do you conclude?
(e) Now you turn to the estimates obtained estimating equation (1) using FE (Fixed Effects).
How do you justify the fact that the FE estimator did not estimate the coefficients for
educ, black, and hisp?
(f) The estimated coefficients using FE and RE look quite different, so you suspect that the
fixed effect might be correlated with one or more included regressors. Therefore, you
decide to run a Hausman test and the result is 17.2 Do you reject the null using a 5%
significance level? And what about using a 10% significance level? What does the result
of the test suggest about the presence of correlation between the fixed effect αi and the
included regressors? Based on the result of the Hausman test, is there evidence that the
RE estimators are not consistent? (Note for the calculation of the number of degrees of
freedom: this test is based on the comparison of coefficients estimated by using FE and
RE, so you can only compare coefficients that are estimated in both models!).

2
2. Empirical exercise to be solved with Stata. You want to study the relation between
murder rates and death penalty. You can find the dataset murder.xls on the Course Website.
The dataset contains data from all US States for 1987, 1990, and 1993. Again, you should
assume that the error terms in these regressions are homoskedastic. Make sure to include
your Stata output with your homework!

(a) Estimate the following model using random effects:

mrdrteit = β0 + β1 execit + γ1 D90t + γ2 D93t + αi + uit . (1)

where mit is the murder rate (murders per 10,000 people) in year t and state i, execit is
the number of executions in year t and during the two previous years, D90t is a dummy
equal to one if t = 1990 (and D93t is analogously defined), and αi is an unobserved state-
specific, time-invariant fixed effect (the variable id in the dataset is a numerical state ID,
which you will want to use instead of state when you estimate this RE model in Stata).
(b) The model does not contain a dummy equal to one if t = 1987. Why?
(c) You observe that, in the data, the unemployment rate in the state is positively correlated
with the number of executions, and you suspect that unemployment may be related to
murder rates. Re-estimate the model using again random effects, but including also a
measure of unemployment. The new model is then the following:

mrdrteit = β0 + β1 execit + β2 unempit + γ1 D90t + γ2 D93t + αi + uit (2)

where unempit is the unemployment rate in state i at time t. How does β1 change when
you include unempit in the regression? Explain the change in terms of omitted variable
bias.
(d) In the two regressions, can you reject the null hypothesis that the predicted effect on the
murder rate of the number of executions is zero, using a 10% significance level?
(e) You suspect that the fixed effect αi may be correlated with the regressors. Re-estimate
model (2) using fixed effects.
(f) By using a Hausman test, you test formally the null hypothesis that the fixed effect is
uncorrelated with the regressors. Perform this test. Can you reject the null, using a 1, 5,
and 10% significance level? Given the result of the Hausman test, which estimator seems
to be consistent, in this empirical application, RE, FE, or both?
(g) Using FE results, calculate a 95% confidence interval for the expected effect on the murder
rate of increasing the number of executions by 30.
(h) Using FE results, test the null hypothesis that the year fixed effects are both equal to
zero. Can you reject the null using a 1 and a 5 significance level? What does the result
tell you about the change in the murder rate in the US over time, during the period
1987-93?
(i) Now you want to test the null hypothesis that the two year fixed effects are identical.
You know that Cov (γ̂1 , γ̂2 ) = 0.24. Calculate the p-value of the test, and briefly interpret
the result (using, again, FE results).