Individuals with less than 16 years of education in 1987 were significantly less likely to participate in 1992 (OR=0.84, 95% CI 0.75–0.94), as were minorities (Black OR=0.63, 0.52–0.75 “Other” ethnic group OR=0.46, 0.24–0.86). Those who served in Southeast Asia were more likely to participate in 1992 (OR=1.12, 1.02–1.22). When we tested for interactions between substance use risk factor and participation probability quintile, we found one marginally significant interaction (p=.037, for ever use of cannabis) and two strongly significant interactions (p=.007 for daily smoking, and p=.005 for nicotine dependence), with a negative Beta coefficient for the interaction term in all three cases, implying higher conditional odds ratios in quintiles associated with low response probability. Thus, selective attrition may have caused us to underestimate the true conditional odds ratios for these variables. Further tests using conditional logistic models indicated that inclusion of the interaction term altered the significance of the substance risk factor only for nicotine dependence in the adjusted model (the OR was significant with the inclusion of the quintile interaction term (p<.001), but just failed to reach significance in the
