The repeated design of our current study produced startle responses that were nested within a three-level structure of longitudinal data (i.e., stimulus valence nested within sessions nested within participants). Therefore, we used the PROC MIXED procedure in SAS (PROC MIXED, SAS Institute Inc, Cary, NC, version 9.1, Littell, Milliken, Stroup, Wolfinger, & Schabenberfer, 2006) to conduct linear multilevel modeling (LMM) to estimate the effects of Group, Session, and Stimulus Valence on startle responses. Analyses of an intercept-only model (i.e., a model with no predictors) found that an unstructured covariance structure provided the best fit with the correlation structure in the data set. A standard approach to model building was followed in which three-level LMMs that included stimulus valence as a level 1 predictor, session as a level 2 predictor, group as a level 3 predictor and cross-level interactions were conducted, in a step-wise fashion, to predict startle responses. All predictors were initially entered as fixed-effect predictors. To assess whether the slopes characterizing the relationship between lower-level predictors (e.g., stimulus valence and session) and startle responses varied among sessions and/or among
