Next, propensity score analysis (PSA) was conducted to compare intact and separated families by predicted probability of separation or propensity to separate. (Although more elaborate matching methods exist, we know of none formulated for use with clustered observations; see Rosenbaum, 2010). PSA was limited to data available by mother report to reduce bias associated with paternal reports (where missingness is much higher for separated families; Waldron et al., 2013), resulting in a reduced sample of 1645 twin-families (201 AA, 1444 EA); comparisons of early substance use by PSA were further limited to 2038 individual twins (263 AA, 1775 EA) aged 15 and older. Predicted probabilities of parental separation were derived from logistic regression models predicting separation from variables presumed to predate relationship dissolution, including mother self-report alcohol dependence symptoms, mother report of paternal alcohol problems, maternal age at twins’ birth, paternal educational attainment, maternal smoking, and two-way interactions between parental alcohol problems and (i) maternal age at twins’ birth and (ii) paternal education. Although maternal education was examined as a covariate in Cox analyses, paternal education was more strongly predictive
