Multivariate analyses were restricted to well-being measures that couldnt have been observed prior to the year (2004) in which work and income were observed. These included measures that ascertained whether a child had ever experienced or done the behavior in question. For example, a child could have ever been suspended or expelled from school in 2003 or even several years before 2004. Thus, for a variable measured this way, it was not logically possible that the work and income of the family in 2004 affected the likelihood of a child being suspended or expelled unless the child was suspended or expelled in 2004. In contrast, the survey asks about television rules in 2004 the same year that it asks about family work and income. Thus, this variable could be included in the multivariate analyses.
For dichotomous variables, we employed logistical regression. For indexes, we employed ordinary least squares regression.
Our first step was to estimate models in which only the work-poverty variables were included. For these models there were three dichotomous variables (1) income below the poverty line and hours worked below the work standard; (2) income between 100 percent and 200 percent of the poverty line and hours worked at or above the work standard; and (3) income over 200 percent of the poverty line and hours worked at or above the work standard. The omitted group was working poor i.e., income below the poverty line and hours worked at or above the work standard. This approach was expected to yield results similar to the bivariate analyses already discussed. Next, we added the basic demographic controls. Finally, we added the work pattern dichotomous variables: (1) full-time work is the parents dominant pattern over the course of a year i.e., at least 35 hours per week; and (2) parent works at least 50 weeks over the course of a year.