4 Results

View the most recent version.

4.1 The degree and nature of intergenerational education mobility

Tables 6 and 7 offer results from the estimation of Equation (1), using both the grouped data estimator from the census, and individual level information from the Ethnic Diversity Survey (EDS) for men and women 25 to 37 years of age. This is done using father's years of education as the regressor in the first panel, mother's education in the second, and both at the same time as reported in the final panel of each of the tables. For men 25 to 37 years of age, every additional year of education their fathers have is associated with 0.13 years of more education. This estimate is virtually the same, regardless of whether the census estimate or the EDS estimate is relied upon. This suggests that the grouping estimator does not suffer from undo problems associated with the use of potential as opposed to actual fathers, and that there is likely little measurement error in this information. However, another interpretation is also possible. The grouped data estimator based on census data is a sum of the parental influence at the individual level, and the influence of the average level of education in the community. As such, the similarity in the results might be interpreted as suggesting that there is no influence of so-called 'social capital,' in the sense that Borjas (1992) uses that term. For women the point estimates are different at 0.10 and 0.16, but the standard error is 0.03, suggesting that the confidence intervals overlap. Further, all of these estimates appear to be about the same—within one standard error— if mother's education rather than father's is used as the right-hand side variable.

The second result from these tables is that at 0.13 and 0.16 the estimates are lower than those for third-generation Canadians of the same age cohort. The educational attainment of men and women whose parents were born in Canada is much more strongly tied to that of their fathers and their mothers than it is for second-generation Canadians. For every additional year of parental education the child's education is in the neighbourhood of 0.37 to 0.4 years higher, triple the estimate for Canadian-born children whose parents were immigrants. These results are also robust to using mother's education as the regressor. This contrasts with the finding in Aydemir, Chen and Corak (forthcoming), showing that the intergenerational elasticity of annual earnings, estimated to be about 0.2, is the same among second-generation Canadians as it is among the population as a whole.

Finally, the last panel of the table, by including both paternal and maternal years of education in the equation, makes explicit—when the focus is on the EDS—that for the second-generation sample the mother's and father's education have roughly the same association with the son's education. For every additional year of paternal education, the education of second-generation Canadian men is 0.08 years higher; and for every additional year of maternal education, it is 0.1 years higher; the standard error of these estimates being 0.05. Paternal education seems to be more important in the case of women, as there is no statistically significant association with maternal education. The education of third-generation men is more tightly associated with paternal years of schooling, but there is no difference between parental effects for women.

However, the findings from the census are different than anything else observed. The coefficient on paternal education is much higher at 0.74 for sons and 0.58 for daughters, while that for maternal education is equally as great in magnitude but opposite in sign. In fact, there is near collinearity between the variables in these data. Using father's education as the regressand and mother's education as the regressor, least squares leads to a coefficient of 1.05 and an R-squared of 0.97. This likely suggests that the large change in the parameter estimates are a function of the high correlation in father's and mother's education at the community level. Further, the results at the individual level from the EDS do not show these patterns. There is no way in which we can improve on this by, for example, adding more observations. As it is, we are using a census and maximizing the number of groups that can reasonably be observed. Accordingly, we proceed by dropping one of the variables, mother's education, in our analysis since the results based upon the EDS offer the slight suggestion that paternal education is more often statistically significant.

Traditionally in the child development literature, maternal education is seen as the prime influence on child attainments, as for example in the discussion by Haveman and Wolfe (1994: 99–101). But recent research has brought this into some question because of the lack of controls for paternal education in many of these studies. If there is assortative mating so that the education levels of parents are similar, the use of only maternal education could be misleading (Behrman and Rosenzweig 2002). Indeed, Sen and Clemente (unpublished) offer an analysis of intergenerational educational attainments using the Canadian General Social Survey and obtain results similar to ours. Their results are for the entire population and are best compared with those reported in Tables 6 and 7 under the heading 'entire population.' They also find that the probability of postsecondary education is positively related to that of both parents, but somewhat more strongly to fathers.⁷ These findings can also be used to motivate the focus in the remainder of our analysis on the relationship between child outcomes and paternal education.⁸

⁷ The results they report in the second columns of their Tables 2 and 3 are not, however, directly comparable in magnitude with our findings because they deal with the probability of any postsecondary education or any university education rather than years of education. They are also not restricted to the age cohort upon which we focus. Their linear probability model of any postsecondary education leads to coefficients of 0.28 and 0.24 for indicators of whether the father attended postsecondary and whether the mother attended postsecondary. They also control for age, gender, marital status and province. The coefficients are estimated to be 0.27 and 0.18 when the probability of any university education is being examined. de Haan and Plug (2007, Table 2) also report a similar result from the Wisconsin Longitudinal Study.

⁸ We also used quantile regressions to amplify slightly the findings from the Ethnic Diversity Survey in order to highlight which part of the distribution contributes to the difference in the intergenerational covariance of years of education. The results were not strong and unambiguous. The least squares estimate of 0.134 for second- generation men is driven more by those sons at and below the median than those above, but that the estimate of 0.4 for third-generation men is driven by those in the top half of the distribution. That is, the link between parent and child education is stronger for high-achieving sons among the native population, but stronger for low- achieving sons for the second-generation population. But these tendencies were slight, and overall there were no really strong differences. The second-generation estimates are always much lower than those for the third generation throughout the entire distribution of child attainments. These general conclusions also held for women.

4.2 Parental education and earnings

Table 8 offers census-based least squares results examining the association of both paternal education and income with child education attainment. The results reported in the first column repeat, for the sake of reference, the results from the first columns of Tables 6 and 7, indicating the small positive association between father-child years of schooling. These coefficients are statistically significant at any marginal significance level, being three times as great as the standard error, and explaining about one fifth to one third of the total variance depending upon whether the focus is on men or women. This is in sharp contrast with the findings in Column 2, which are based on only the logarithm of paternal weekly earnings as the regressor. The coefficient is not statistically different from zero, neither for men nor for women, explaining none of the variation in the data. Finally, and not surprisingly, when both paternal years of education and earnings are used in the model education dominates, it actually turns out that earnings are negatively associated with the child's years of schooling—being on the margin of statistical significance at the 95% level—and the coefficient on education becomes larger in magnitude.

The suggestion in all of this is that, on average, paternal earnings on their own have no strong association with the education outcomes of children, either sons or daughters. There isn't a straightforward interpretation to give to these results. They are certainly not causal, but at the same time they don't simply reflect a near collinearity in the variables. The correlations between parental education and earnings are 0.6192 for fathers and 0.3244 for mothers. An unobserved effect may be at play. For example, it is possible that children of low-income parents have had more altruistic parents that have invested more heavily in non-monetary aspects of human capital than their higher earning counterparts. At the least, these patterns suggest that the education outcomes of second-generation children is much more closely correlated with the education of their parents, and relatedly to the institutional structure of an education system that does not appear to limit access according to income.

4.3 Changes in the intergenerational association of education

Tables 9 and 10 offer an expanded version of the EDS results presented in Tables 6 and 7 by fully interacting Equation (1) with birth cohort effects. The base case is the cohort 25 to 34 years of age, and separate intercepts and slopes are added for those 35 to 44 years, 45 to 64 years, and finally those 65 and older.⁹ Three results follow from this exercise for both men and women.

First, for both the second-generation and the third-generation populations the slope coefficients seem to be the same across all birth cohorts. Rarely are the estimated coefficients for the interaction terms with paternal education greater than one standard error, and they are never greater than two standard errors. Individually these coefficients are not statistically significant from zero, and F-tests do not reject the null hypothesis that collectively they equal zero.

Second, the estimates of the constant term make clear that second-generation Canadians obtain more years of schooling than those born in the country with Canadian-born parents. To be precise, for those 25 to 34 years of age the difference in years of schooling for men is four years in favour of second-generation Canadians; for women it is almost two and one-half years.¹⁰

Third, the separate intercepts for each birth cohort suggest that only in the case of the very oldest cohort, those older than 65 years in 2001, are the years of schooling different. This cohort obtained from two and three-quarters to four and two-thirds years fewer schooling than all younger cohorts. This could reasonably be attributed to changes in school leaving legislation as these individuals would have been 15 years of age at some point before 1950 (Oreopoulos 2006). It is the statistically significant result for this single cohort that drives the results of F-tests to a point that we cannot reasonably reject the null that all intercepts are collectively equal to zero.

With the possibility of this last exception, the results from this model show that for both men and women the intergenerational association in educational attainment, including overall average attainment, has been stable across all birth cohorts. None of the findings associated with Tables 6 and 7 need be modified: the Canadian-born sons of immigrants obtain about 0.10 years more schooling for every additional year their fathers have, and the daughters about the same at 0.16; this is significantly lower than the tie between the Canadian-born children of Canadian-born parents who obtain an additional 0.3 to 0.4 years of schooling for each additional year. In particular, the degree of intergenerational mobility among most recent second-generation Canadians is no stronger or no weaker than it has always been, and has not changed relative to third-generation Canadians.

It should be noted that the youngest second-generation cohort in our analysis, those 25 to 34 years of age in 2000, were born on average in 1970 and no earlier than 1966. In other words, this cohort was born just after the implementation of important policy changes that led to the removal of the national origin quota system as a means of selecting immigrants. Therefore, their parents likely entered the country before this system was replaced by a points-based policy geared to labour market integration. As such the extent to which these findings can be extrapolated into the future is an open question. The results may differ if the analysis were to be replicated in the future with more recent cohorts of immigrants and their children, those who were selected under the new policy regime, and who accordingly were much more diverse in their national origins.

⁹ Our original inclination was to use 10-year age cohorts, but the group 55 to 64 represented about 7% of the samples, and we decided to aggregate it with 45 to 54 year olds after preliminary regressions revealed no statistically significant results.

¹⁰ It should again be noted that these results pertain to the reference case of those living in Ontario.