We examined attrition between the 1987 SOH and the full 1992 HDS sample through logistic regressions predicting loss to follow-up. In addition, we conducted a form of propensity score analysis (Rosenbaum & Rubin, 1983) to check whether differential attrition for concordant versus discordant pairs was biasing our conditional odds ratio estimates. A logistic regression equation was used to predict individual non-participation versus participation in the 1992 HDS as a function of variables assessed in the 1987 SOH. Predicted probability of participation (range: 0.08–0.78, median 0.59) was then categorized into quintiles, and this categorical variable, defined for the first twin from each pair, was then included as a moderator in conditional logistic analyses (i.e.., testing for an interaction of participation probability with the drug risk statuses of twin and cotwin when predicting discordance for educational attainment).
