We considered using logistic regression, with PTSD as the outcome, and predisposition, event type and their interaction, as independent variables. However, doing so is akin to adjusting for an intermediate variable (here, exposure to trauma type that might be influenced by predispositions), which can introduce a bias (Robins & Greenland, 1992). To circumvent this potential bias, we modeled PTSD and trauma type jointly by multinomial logistic regressions. The outcome (Y) in each regression has four categories. For example, in the comparisons between sexual assault (the severe event in this analysis) and accidents (a lower-magnitude event), the categories are: Y = 1 if the respondent (R) developed PTSD after a sexual assault; Y = 2 if R experienced sexual assault but did not develop PTSD; Y = 3 if R developed PTSD after an accident; and Y = 4 if R experienced an accident but did not develop PTSD. The model specifies the probability of outcome Y = j (j = 1, 2, 3, 4), as a function of predisposition and potential confounders as follows: πj(xi)=Pr(Yi=j|xi)=exp(xi′βj)∑k=14exp(xi′βk), where xi is a vector
