4.3  Results excluding extreme outliers and after adding a quadratic

Our analysis so far has included all Dunedin Study participants for whom we have positive incomes for themselves and for their parents. However, overseas studies have sometimes excluded cases with extremely low incomes, often with the aim of testing the robustness of the results (Couch and Lillard, 1998, pp. 314, 328; Raaum, et al., 2007, pp. 15, 17; Solon, 1992, pp. 401-402). Our dataset also contains a small number of very low incomes for parents that analysis suggested could have a disproportionately large effect on the results. In three cases, currency conversions also resulted in participants’ incomes being much lower than could be expected on the basis of other data.

Table 3 shows the results after excluding cases with extremely low parental incomes and participants' incomes that seem distorted by currency conversions. Excluding these cases tends to increase the intergenerational income elasticity point estimates, particularly for women.[25] For models two and five, the p value for the F statistic fell to 0.035 and 0.030 respectively, indicating that the explanatory variables now have a statistically significant effect on the incomes of women. If we probed further into people's circumstances we could probably justify excluding additional cases, such as those who were temporarily out of the workforce looking after children.

While our Table 1 results would seem to be more internationally comparable than the Table 3 results, comparing these two tables shows our intergenerational income elasticity results are sensitive to small changes in the data used. This point has also been made by overseas economists researching intergenerational income mobility (Couch and Lillard, 1998, p. 328; Minicozzi, 2003, p. 309).

We also experimented with adding a polynomial term to see if the relationship between parental income and the income of their children was non-linear. However, this term was not statistically significant, and we therefore cannot reject the null hypothesis that the relationship is linear. This topic is discussed further in Section 8.1.5.

Column entries are unstandardised linear regression coefficients. Values are for log income. Standard errors are in brackets. *=p<.10, **=p<.05, ***=p<.01. Those whose income is missing or declared zero income (eg homemakers) have been excluded. Income is an extremely sensitive topic and missing values, usually for the fathers of participants, have reduced the number of cases.

This section has investigated intergenerational income mobility using unique Dunedin Study data. Although our confidence intervals are narrower than for the only other study of intergenerational income mobility in New Zealand, they are still quite wide. Parents' incomes also explained only a very small proportion of the variance in the incomes of men and women who were born in Dunedin, confirming that other factors have a large effect on their incomes.

Our research has a number of limitations. For instance, we used two years of self-reported income data for parents and one year of data for their children to model people's long-term economic situations. Because people's incomes tend to vary from year to year, the data imperfectly measures the economic circumstances of some participants and their parents (Corak, 2006, p. 6). The results are also affected by measurement error, resulting from the use of income bands rather than exact amounts. In addition, at age 32 many of the participants in the Dunedin Study probably had not yet reached their peak earning years. People in jobs with high life-time earnings tend to reach their peak earning years later in life than people in jobs with low life-time earnings, and using data from early in people’s careers may result in their economic situation being inaccurately measured (Haider and Solon, 2006, p. 1310). Data from the age 38 Dunedin Study assessments (which are taking place in 2010-11) and from other future assessments may therefore produce different results, particularly for women (Bohlmark and Lindquist, 2006).