tivation by means of proxy variables lowers the estimated coefficient for at-
tendance from 0.09 to 0.073. This result could be interpreted positively, as in
Romer (1993), as a sign that “an important part of the relationship reflects a
genuine effect of attendance”. An alternative, more plausible interpretation
is that, despite the introduction of a set of control variables, the relation-
ship still reflects the impact of omitted factors correlated with regressors:
to the extent that, despite the control factors, there are still unobservable
fixed effects correlated with attendance, βOLS remains biased and inconsis-
tent (likely to be over-estimated). We thus turn to estimators that, under
specific assumptions, are immune from the omitted variable bias.
Table 3 presents the results obtained estimating alternative specifications
of the relationship between attendance and performance by instrumental
variables (2SLS), using travel time, work and web as instruments for atten-
dance. The results indicate that the estimated coefficient for attendance is
very sensitive to the set of controls included in the specification. In particu-
lar, the inclusion of ability proxies (column 3) determines a large drop in the
estimated coefficient (from 0.15 in the basic univariate specification to 0.07 in
the full specification) and a negative impact on its significance. In the com-
plete specification (column 5) the estimated coefficient for attendance falls
to 0.065 and is not statistically significant. The high sensitivity of the results
to the set of controls suggests that the instruments are not valid. This is
confirmed by the results of the test for overidentifying restrictions (presented
in the last two rows of table 3), that reject the hypothesis of instrument va-
lidity for all models (except, marginally, for the full specification in column
5).22
Given that IV estimation does not provide a solution to the omitted vari-
able bias, we now turn to estimates obtained by exploiting the panel structure
of the data set. Table 4 presents estimates of the fixed effect model (column
1) and the random effect model for alternative specifications (columns 2-4).
The coefficient for lectures attended is positive and statistically significant
in all models reported. The random effect estimates are similar to the OLS
estimates, indicating that the weight of the between component in the error
term is small relative to that of the within component. In particular, in the
full specification (column 4) the RE model indicates that attending an extra
22This also implies that the assumptions on which the Davidson-McKinnon test (whose
results do not reject the null hypothesis of exogeneity for attendance) is based are not
met: the instruments are not truly exogenous, so that the IV estimator is not consistent.
13