These cross-tabulations do not hold anything constant, including the effect of the other adolescent behaviors on each adult outcome. The anomalies observed may reflect some other factor correlated with late initiation or delinquent behavior. To hold constant other factors that influence these adult outcomes, we estimate a set of regressions. These regressions take the general form
y = f(B,F,X)+ g,
where y is one of the ten outcomes of interest, B represents the five adolescent risky behaviors, F represents the three family environment variables, X represents all the other control variables and is a random error term.
The other control variables include sex, race/ethnicity, educational attainment, and Rosenberg's measure of self-esteem (measured in 1980).15 The regressions are estimated unweighted and each regression includes all the behaviors, family environmental variables, and control variables. The regressions allow us to examine the relationship of the adolescent risky behaviors, family structure, parents' education, and parental alcoholism to the adult outcomes while holding each set of variables constant. The regressions also allow us to control for participation in all five adolescent risky behaviors so that the impact of any particular behavior can be identified separately from the others. We might expect, similar to Newcomb and Bentler's (1988) findings, that each adolescent behavior may show a unique pattern of influence on adult outcomes. We now examine sequentially the relationship of the four sets of adolescent factors to the adult outcomes. The full set of regressions with all variables is shown in Appendix B.