Exercise 9

1. Which type of model has coefficients that vary with i, but are constant with t?
a) pooled model
b) fixed effects
c) random effects
d) none of these

2. Which type of model has coefficients that vary with i and t?

a) pooled model
b) fixed effects
c) random effects
d) none of these

3. In which model are coefficient estimates determined by variation within individuals

rather than variation across individuals?
a) pooled model
b) fixed effects
c) random effects
d) none of these

RAO, YAO ECON312 1

4. Which model is also called an error components model?
a) pooled model
b) fixed effects
c) random effects
d) none of these

5. For a random effects model the least squares estimator is unbiased and consistent.
The errors can be corrected for potential heterogeneity using _________________, but
the estimator with minimum variance is _______________.
a) cluster-robust standard errors, GLS
b) White’s correction, ML
c) 2SLS, fixed effects
d) cluster-robust standard errors, pooled

6. Which of the following is not a reason random effects (RE) results may be preferred
to fixed effects (FE)?
a) RE accounts for the random sampling process that generated the data
b) RE is a GLS estimator so in large samples it has a smaller variance than FE which is
a least squares estimator
c) RE produces a coefficient for race, gender, or other individual characteristics that are
constant over time
d) RE estimates are more robust in the case of endogenous regressors

7. How do you test for endogenous regressors, or correlation between the error term
and any regressor in a random effects model?
a) estimate coefficients with RE and FE, then perform a Hausman test of equality
b) estimate the model capturing estimated residuals, then regress residuals on all
regressors and perform an F-test
c) estimate RE model capturing estimated residuals, then estimate coefficients of
correlation with each regressor

RAO, YAO ECON312 2

d) estimate RE and FE models and perform an F test on each model individually. If the
difference between the F statistics is significant, conclude endogeneity.

8. If you perform a Hausman test on a random effects model and have a test statistic
that exceeds your critical value, what should you conclude?
a) all of the regressors in the RE model are exogenous
b) none of the common RE and FE coefficients are significantly different
c) at least one of the coefficients is significantly different from zero
d) at least one of the regressors in the RE model is endogenous

9. If you perform a Hausman test on a random effects model and have a test statistic
that exceeds your critical value, which of the following is not correct?
a) at least one of the regressors in the RE model is endogenous
b) none of the common coefficient estimates in the RE model will be significantly
different in the FE model
c) FE may be the preferred estimation technique
d) this model may be better estimated using the Hausman-Taylor estimator

10. Suppose you have a long, narrow panel of data and estimate a single equation with
indicator variables and interaction terms for the individuals. In doing this what
assumption from the pooled model have you maintained?
a) coefficients on variables are equal across individuals
b) errors are uncorrelated with any x’s
c) expected value of errors are zero
d) variances of error terms are equal across individuals

11. Suppose you have a long, narrow panel of data and estimate a single equation with
indicator variables and interaction terms for the individuals. In doing this what
assumption from the pooled model have you relaxed?
a) coefficients on variables are equal across individuals
b) errors are uncorrelated with any x’s

RAO, YAO ECON312 3

c) expected value of errors are zero
d) variances of error terms are equal across individuals

12. When an equation is estimated for each individual jointly, taking into account
contemporaneous correlation the resulting model is a(n)
a) Hausman-Taylor model
b) SUR
c) ECM
d) VEC

13. In a random effects model, define the composite error 𝑣𝑖𝑡 = 𝑢𝑖 + 𝑒𝑖𝑡 where 𝑢𝑖 is
uncorrelated with 𝑒𝑖𝑡 and 𝑒𝑖𝑡 have zero mean and constant variance 𝜎𝑒2 and are serially
uncorrelated; The random effects 𝑢𝑖 has zero mean and constant variance 𝜎𝑒2 and serially
uncorrelated.

a) Show that 𝐸 (𝑣𝑖𝑡 ) = 0

b) Find 𝑉𝑎𝑟(𝑣𝑖𝑡 )
c) For 𝑡 ≠ 𝑠, find 𝑐𝑜𝑣(𝑣𝑖𝑡 , 𝑣𝑖𝑠 )

14. Consider a panel data on young women in year 1990 and 1991 with each year contains
716 women who were interviewed. We are interested in the wage equation that relates
the logarithm of WAGE to years of education (EDUC), working experience (EXPER),
its square EXPER2, and dummy variable BLACK, SOUTH, UNION.

The coefficient estimates for the different parts of the questions are given in the following
table with the standard errors in parentheses below the estimated coefficients.

RAO, YAO ECON312 4

 Variable 1990 LS 1991 LS Random Effect Fixed Effect

Intercept 0.2268 0.2216 0.3086 1.5468

(0.1881) (0.2227) (0.1610) (0.2522)

EDUC 0.0762 0.0778 0.0766

(0.0063) (0.0064) (0.0060)

EXPER 0.0875 0.0830 0.0758 0.0575

(0.0265) (0.0292) (0.0205) (0.0330)

EXPERsquare -0.0020 -0.00179 -0.001648 -0.1234

(0.00096) (0.00096) (0.000702) (0.1102)

BLACK -0.1562 -0.1309 -0.1319

(0.0366) (0.0327) (0.0345)

SOUTH -0.1029 -0.1368 -0.1350 -0.3261

(0.0327) (0.0334) (0.0303) (0.1258)

UNION 0.1701 0.1324 0.1170 0.0822

(0.0350) (0.0354) (0.0235) (0.0312)

a) The estimate results (column 2 and 3 in the table) show the least square estimation for
each of the years 1990 and 1991. How do the results compare? For these individual
year estimations, what are you assuming about the regression parameter values across
individuals (heterogeneity)?

b) Allowing heterogeneity across individuals, the wage equation is modeled as

ln( WAGEit )  1i   2 EDUCi   3 EXPERit   4 EXPERit2

  5 BLACK i   6 SOUTH it   7UNION it  eit

Explain any differences in assumptions between this model and the models in part a).
Explain why the variables EDUC and BLACK have the subscripts i rather than i and
t, like the other variables.

c) Estimate the model shown in b) using the fixed effects estimator. The estimated
results are in Column 5 in the Table. Explain why there are no estimated coefficients
for variable EXPER and BLACK.

RAO, YAO ECON312 5

d) Based on the estimation results in column 4 of the table, what is the estimated return
on an additional year of education in the random effects model? Is it statistically
significant? Construct a 95% interval estimate for this parameter.

bFE ,k  bRE ,k
t
[ se(bFE ,k ) 2  se(bRE ,k ) 2 ]1/2
e) Using the t-test statistic , to test (at 5% significance level)
the difference between the fixed effects and random effects estimates of the
coefficients on EXPER, its square EXPER2, SOUTH and UNION. The t values are
as follows:

texp er  0.711 (p  value  0.477)

texp er 2  0.487 ( p  value  0.626)
tsouth  1.565 (p  value  0.118)
tunion  1.692 (p  value  0.091)

Do we reject, or fail to reject the null hypothesis that the difference between the
estimates is zero? Which estimate would be better to use, fixed effects or random
effects estimates?

RAO, YAO ECON312 6