we analyze the effects of generosity by estimating models with those selected components of
social expenditures for which appropriate population shares serving as proxies for number of
recipients are available. We employ either the share of elderly in the population or the share of
unemployed in the active population. Indeed, omission of beneficiary measures might have
biased our previous results as the spending estimates might capture population composition
effects. Simple correlations of the spending measures with the number of their specific
beneficiaries are large.20
In Table 7 we present results for models with measures of the number of recipients included.
Taken all together, the results are not sensitive to inclusion of proxies for the number of
beneficiaries. Pension spending and active labor market policy spending still exert a student
performance lowering impact when the share of elderly and the unemployment rate,
respectively, are included in the model (columns (1) and (2)), while the effects of
unemployment and health care spending remain insignificant (columns (3) and (4)). The
similarity of the coefficients on the spending variables with the original models reported in
Table 5 suggests that the bias from using spending measured per GDP (conditional on
population size) in place of per recipient is rather small. Regarding pension spending in
column (1), the significance level is reduced to 5 percent when the share of elderly is
included. While the share of the population above 60 years of age is insignificant, the test of
joint significance clearly suggests that both variables are jointly related to student
achievement. The effect of active labor market policies spending is equally lowered in
significance (now at 5 percent level) when the unemployment rate is included (column (2)),
while this time the test of joint significance clearly suggests that only one of the variables is
related to student achievement.
5.5. Robustness analyses
The student test scores from the 1980s are not average results for jointly conducted
Mathematics and Science tests as those achievement tests in the ‘post-communist period’
(1990s and beyond), but separate tests on the two subjects. Another reason for splitting the
sample into two time-periods is that many argue that test designs and test procedures have
improved over time. Therefore, the dependent variable may incorporate a larger measurement
error in the 1980s than in later periods. Table 8 presents results for regressions on the
subsample for the post-communist period 1990-2003. Columns (1)-(3) in the table show that
the effects of all our measures of government generosity - government consumption spending,
social transfers, and income tax progressivity - are in fact larger in this subsample compared
20 The correlation coefficients between unemployment spending and unemployment rate is 0.51 and between
pension spending and the share of the population above the age of 60 is 0.86. The correlation between active
labor market policy spending and the unemployment rate is only 0.17.
22