20. An alternative question we could have posed would be whether family environment mitigates the impact of risky behaviors on adult outcomes for early initiators as compared with late initiators. We have chosen not to frame the question this way because we want to examine the effects of family environment given the choice to initiate at early ages. Those who initiate at later ages or never initiate are following a different path. Family environment may involve different processes for these groups. In addition, since our oldest group includes both late initiators and non-initiators, we cannot disentangle the influence of family environment for these two groups.
21. By limiting the sample to early initiators, the model is equivalent to fully inter-acting "early initiator" with all the variables in the regression, but only examining the results for the early initiators. Thus the estimated values for family structure or parents' education are similar to what would be obtained from interacting family structure and parents' education with age of initiation in the regressions in Chapter V. This functional form also allows all the other variables to have effects specifically for early initiators
22. This method of combining age groups to define early initiators makes the functional form only similar to, but not the same as, the fully interactive model described above.
23. Small sample sizes limit our ability to study other family types. Furthermore, this restriction provides the cleanest way to address the issue. In these two family situations, the measures of mothers' and fathers' education specifically relate to the parents with whom the youth has lived and do so exclusively. This would not be true in other family situations. For example, to study a family comprising the child's mother and a stepfather would require knowing the education of the mother and stepfather, and perhaps also the biological father. We would need to know the timing of the divorce (if there was one) and the re-marriage and the degree to which the father has stayed involved financially and directly with the child. Even knowing all of this, we would have to estimate a complicated relationship between the influence of different parents' education, biological and step, on the child. This places too many demands on the data. For single mother households, we face a similar, though not as complex problem. We do not know the financial or time contributions made by the absent father. However, because this family type is so common, we include it. Since we can only look at mothers' education, there may be significant omitted variable bias without a measure of father's education.
24. We also estimated the regressions including father's education to examine whether there were differences. However, some of the values lacked credibility, i.e. they were orders of magnitude different than previously observed and changed much too dramatically from the regression without father's education. At best, these estimates would be difficult to interpret since we do not have measures of the father's financial contribution or level of involvement with the youth.