The longitudinal model employed was saturated, in that all across-time CPD covariances were simultaneously estimated. Within-family clustering due to twin status was accounted for with standard twin modeling (Neale & Cardon, 1992). This technique decomposes the observed CPD phenotypic variance-covariance matrix into additive genetic variance (A), shared environmental variance (C), and unshared environmental variance (E). Because so little variance is observed for CPD at age 11, our longitudinal models only included measurements at age 14, 17, 20, and 24. All heritability models were fit with OpenMx 1.0.6 (Boker, et al., 2011) of the R Environment version 2.13.1 (R Development Core Team, 2011). Confidence intervals were computed using non-linear constraints and likelihood ratio tests according to standard practice (Miller & Neale, 1995; Neale & Miller, 1997). All models were fit under full information maximum likelihood to account for missing data. Covariates used in these analyses were year of birth, sex, and cohort status. Covariates were incorporated as fixed effects as described in the next section.
