These four models successively impose fewer assumptions on the data, thereby making the
estimates more robust while simultaneously decreasing their eciency. If all models produce
qualitatively similar results, however, we should be reasonably condent that our conclusions
are valid.
The rst model for each spending category (entitled FE 1) includes the globalization vari-
able but not its interaction with the developing countries dummy. By not including the
interaction eect in this model, we eectively assume that globalization has the same eect
in both industrialized and developing countries.
The second model (entitled FE 2) includes the interaction eect and thus takes into account
that the marginal eect of globalization might dier between industrialized and developing
countries.
The third model (entitled IV), while controlling for all variables that are considered in
the second model, takes additionally into account that total education expenditures and the
expenditures for primary, secondary, and tertiary education are by construction simultane-
ously determined. We deal with this endogeneity problem by instrumenting total education
expenditures. The IV estimator, while being less ecient than OLS, leads to consistent es-
timates. We consider this as our preferred model because it is probably the most reasonable
compromise between consistency and eciency.
In the fourth model (entitled IV & CL), re-estimate of the third model, but conduct the
hypothesis tests on the basis of clustered standard errors. This model is therefore robust to
arbitrary forms of autocorrelation, but may be particularly inecient.
We use as instruments for total education expenditures in the third and fourth model (i) the
gross enrollment rates and (ii) the population shares of the relevant age groups for the three
educational programs (school age population). To minimize the potential for endogeneity
problems, we use in each set of regressions for one type of educational program the enrollment
and population shares in the other two types. For example, we use as instruments in the
models with secondary education expenditures as dependent variable the gross enrollment
rate in primary and tertiary education, and the population shares of the age groups relevant
for primary and tertiary education. The only exception is the model with primary education
expenditures. Here, the Hansen J test for overidentication rejects the null when the gross
enrollment rate in tertiary education is used as an instrument. We therefore exclude this
variable from our set of instruments in the models with primary education expenditures as
dependent variable.
The theoretical rationale for using the dierent enrollment rates and population shares as
instruments is that they are likely to be correlated with total education expenditures.11 As
11Note that the direction of the correlation is not trivial. While a higher overall gross enrollment rate
should be positively correlated with higher total education expenditures, this is not necessarily true for the
individual gross enrollment rates in the three educational programs. That is, a higher gross enrollment rate in
a lower-level educational program implies that the gross enrollment rate in the next higher program tends to
be smaller, for example because of repeaters. Since education expenditures per student are on average larger
12