Biometric models were fit using “Open Mx” package, version 3.2.2, for R software [42], and estimated using full information maximum likelihood [35] and a threshold-liability model, which models ordinal categories as arising from estimated thresholds on an underlying normal distribution [41]. For this purpose, the second and third categories of PD criteria were collapsed to avoid empty or rarely endorsed cells in the three-way tables; similarly, criterion counts of 3 or more were collapsed when studying liabilities for composite PDs. We then estimated associations among a wave 1 PD-related variable, wave 1 AUD (5-year recency), and wave 2 AUD (5-year recency) for all the covariance components, A, C, and E using Cholesky decompositions (see Figure 1a for a path diagram of one component) [41]. It is often found that the C component is not statistically significant, leading to an “AE” model. We selected between ACE and AE models using Akaike’s Information Criterion [43]. For individual biometric parameter estimates, likelihood-ratio tests and likelihood-based 95% Confidence Intervals (CIs) were assessed [44].
