3.3 Sensitivity analysis
In this section, we investigate the robustness of our results. In the rst set of robustness
checks, we divide our sample in industrialized and developing countries and estimate the
baseline model separately for each of the two groups of countries. In doing so, we allow not
only the coecient on the globalization variable, but also all other coecients to vary between
industrialized and developing countries.
The results are collected in table 3. For each type of educational program, we estimate two
individual models. First, we estimate a pooled OLS model to exploit both the between and
within variation (since pooled OLS models take the between variation into account, they can
be considered as an additional robustness check). Thereafter, we include country xed eects
and re-estimate the model while only using the within-variation. All models also include year
xed eects, and hypothesis tests are conducted with robust standard errors. Total education
expenditures are instrumented with the same instruments as in the baseline models.
First, note that the signs of the coecient are always the same in the pooled OLS and the
xed eects models. The signicance levels, however, tend to be lower in the xed eects
models, but the t-statistics are in general relatively high. Note also that the overidentication
test sometimes performs poorly in the pooled OLS models while it always performs well in
the xed eects models. This indicates that the pooled OLS models probably suer from an
omitted variables bias.
Overall, the results from these robustness checks conrm the baseline results. In developing
countries, globalization apparently leads to higher primary education expenditures while it
seems to have a negative eect on tertiary education expenditures. Conversely, globalization
seems to have no eect on education expenditures in industrialized countries as indicated by
the insignicant coecients in the xed eects models. In other words, the baseline results
are conrmed by this set of robustness checks.
In table 4, we report the results from a second set of robustness checks. In these regressions,
we study whether the results in the baseline regressions are sensitive to (i) outliers, (ii) to
using “budget shares” (primary, secondary, and tertiary education expenditures as share of
total education expenditures) instead of spending per student as dependent variable, and (iii)
to controlling for development aid. The specication is similar to the baseline models in the
sense that we include the interaction of the globalization with that of the developing country
dummy variable and all other control variables from the baseline model.
15