In quantitative genetics, the heritability of a continuous trait denotes the proportion of its variance which is attributable to genetic differences. A classical method for assessing heritability contrasts the correlation of the trait observed across pairs of identical twins (= monozygotic twins = MZs) with that observed across pairs of fraternal twins (= dizygotic twins = DZs). In its simplest version, the twin method attributes the greater correlation of MZs entirely to the perfect correlation of their genotypes. Since the genotypes of DZs, like those of ordinary siblings, correlate only about 1/2, the very simplest twin method just doubles the difference between the two observed correlations to estimate heritability. Economists may have become aware of the heritability concept, and of the twin method, via the great IQ debate, in particular via the books of Arthur Jensen (1972, 1973), Richard J. Herrnstein, and Christopher Jencks. A series of articles by Paul Taubman (1976a,b), and Jere Behrman and Taubman, has now brought heritability and twin methods into economics itself. Twin data on schooling, initial occupation, later occupation, and earnings lead to such inferences as Genetics by itself accounts for roughly 30 to 40 of everything except initial occupation, where it accounts for 8 percent (Behrman and Taubman,

The empirical results on dependency rates and savings rates reported by Nathaniel Leff (1969) cannot be correct. For several cross-country samples, Leff estimates pairs of equations of the form (1) ^1 = 00-1- aiXi + CiXi-f (2) ^2 = 60 + bixi + biXi-f 63*3-|- biXi-f- d where yi = ln S/Y = ln domestic savings ratio yi = In S/N = In per capita savings xi = ln Y/N = In per capita income a:2 = | = growth rate of per capita income X3 = ln Di = ln percentage of population aged 14 or less Xi = ln Di = ln percentage of population aged 65 or more, and a and b are least-squares regression co-efl&cients, and e \\ and e ^ are least-squares resid-uals. As noted by Leff, S/N^iS/Y)iY/N). Consequently, yi = yi-\\-Xi. Least-squares re-gression being what it is, a proper com-putation of (2) should produce = flo + (1 + ai)xi, + 0.2*2 +(3) That is, regressing y ^ on the x should give the same coefficients and the same residuals as occur when y ^ is regressed on the *, except for the coefficient of Xi, which should in-crease by exactly 1. Furthermore, if regression coefficients are guaranteed to be equal, their standard errors, and hence their ^-ratios, must be equal. If regression coefficients are guar-* Professor of economics, University of Wisconsin, Madison. anteed to differ by unity, their standard errors must be equal, and hence their t-ratios must be related by bx/si, = (ajAa.)((l-f ai)/ai) But the results Leff reports do not satisfy these arithmetic requirements. For example, consider the upper panel of his Table 1, p. 891, which refers to a sample of 47 under-developed countries. In the present notation we find:

“Does increasing women's schooling raise the schooling of the next generation?” is the question posed by Jere R. Behrman and Mark R. Rosenzweig (2002). Their answer to the question is no. In fact, they conclude that raising women's schooling may lower the schooling of the next generation. We show that Behrman and Rosenzweig's results are not robust to alternative coding schemes and sample selection rules, and argue that their policy inference may be misguided.