Data Analysis Plan
Linear mixed models were used to evaluate whether (a) BT would have worse SIP and social skills over time compared to ST and (b) whether social skills and accuracy in recognizing facial expressions and others’ social intentions were positively associated with peer relationship quality. For longitudinal designs with two time points and data missing at random, maximum likelihood estimation used in mixed models produces less biased estimates than least squares approaches [19]. Thus, we used a random intercept model with restricted maximum likelihood estimation to test our hypotheses. All continuous predictors were grand-mean centered to aid the interpretation of coefficients.
In the model for hypothesis (a), DANVA-2, Irony & Empathy, and SSiS were entered as outcome variables in three separate models, while time, group and their interaction were entered as predictors. Three models were run for hypothesis (b) that differed only in terms of the specific predictor variable tested. The three predictors included baseline values of SSiS, DANVA-2, and Irony & Empathy. For example, for the model evaluating the influence of baseline social skills, PROMIS Pediatric Peer Relationships was the outcome variable, while time, baseline SSiS, group and a 3-way interaction between them were entered as predictors. We also conducted two post hoc analyses in the BT group, where we evaluated the moderating effect of baseline time-since-diagnosis on Irony & Empathy and social skills. Examination of diagnostic plots at levels 1 and 2 showed that models met assumptions of normality, linearity, independence and constant variance. The nlme package from the R Statistical Computing Language was used to run the linear mixed models and the sjPlot R package was used to produce tables and graphs of simple slopes.
RESULTS