1. The final logistic regression specifications whose results were reported below were generated by a simple regression pruning procedure based on three sets of variables. ? Pure Time Effects: We have eleven years of data, so a fully specified model would include a constant and ten year effects. To allow pruning of the specification, we specify the year effects as a linear time trend, an offset for the unvalidated years, a linear time trend specific to the unvalidated years, and year dummies for 1992;Huber/linearization approximation. Estimation is done in SAS using a PROC IML routine provided by Dan McCaffrey of the RAND Statistics Group,1990
2. with survey year) over the eleven years of data in our analysis file. Given this list of potential covariates, we proceed in four steps: 1. We begin with all of the pure time effects and the time invariance covariate effects included. 2. We then delete all variables (from the time effects and from the time invariant covariate effects) that do not have a p-value less than 0.10. Note that we only do this once. We do not iterate;? Linear Covariate x Time Interactions: We allow the effect of each of the covariates to vary linearly
3. Under-Reporting of Medicaid and Welfare in the Current Population Survey