Analyses were carried out using R version 2.15.0 [47], and all mixed-effects models included both individual and study site as nested random intercept factors, as well as time as a random slope factor. A first-order autoregressive error structure suitable for longitudinal data was used [46]. Mixed-effects models utilized an intent-to-study approach based on intent-to-treat principles that are commonly employed in longitudinal clinical trials by including all eligible individuals who entered the study, regardless of whether they completed all study assessment periods [48]. This approach was taken because examining only those participants who complete the entire study is well-known to bias parameter estimates in longitudinal studies with attrition [48–50]. Rather than discard partial study completers and potentially bias the final sample of individuals analyzed, the expectation-maximization approach was used to handle missing data during maximum likelihood estimation at the time of analysis [51–52]. This approach to missing data is not imputation-based, but relies on estimating model parameters (e.g., individual trajectories) using all available data (e.g., existing measurement occasions for the individual, overall sample parameter estimates, model covariates), and is currently the
