schooling is even stronger than in the full sample regressions because of the fall in
standard errors. We proceed similarly with the other specifications of Table 9 and find
the same results as there (we do not report these results). Namely, the regressions a
laBenhabib and Spiegel(1994) andPritchett(2001) do not produce significant coeffi-
cients for the schooling variable. Overall, these results suggest that the significance of
schooling in these regressions is not caused by unrepresentative observations.