66 Chapter 3 because staff and patient mutually agreed that treatment goals were achieved, this was considered as completion of treatment instead of dropout. Power and Sample Calculation Initial power analysis suggested that a sample of 34 per group (68 total) would provide a power of .80 to detect a medium effect. A total of 74 participants (33 to CBT+ and 41 to DBT-BED) were randomized. At the conclusion of the trial, it was discovered that the initial power analysis was incorrect. The actual power to detect a medium effect based upon an alpha of .05 is only .58. Statistical Analysis All analyses were conducted using SPSS Version 25 (IBM Corp, 2017). Significance tests were based on a two-tailed alpha of 0.05. Primary measures of outcome used to evaluate efficacy included OBE episodes and EDE-Q Global scores. Secondary measures of outcome included DEBQ Emotional Eating, EDI-3 Emotional Dysregulation, SCL-90 total score, BDI-II total score, and EDI-3 Low Self-Esteem. Distribution diagnostics for primary and secondary outcome measures suggested that all outcome measures except OBE episodes were symmetrically distributed and appropriate for normal assumption models. CBT+ and DBT-BED treatment groups were compared separately at end of treatment and follow-up controlling for baseline assessment using a generalized linear model with a negative binomial distribution for OBE episodes and a general linear model for all other outcome variables. Given that treatment for both CBT and DBT was delivered in group settings, preliminary models were run nesting participants within therapeutic groups. As no significant variation attributable to therapeutic group was found, subsequent analyses were conducted without nesting. Final models included a main effect for treatment group, and a fixed covariate for baseline assessment. Effect sizes between treatments were calculated using both Cohen’s (Cohen, 1988) d and the success rate difference (SRD; Kraemer & Cupfer, 2006). Cohen’s d values were calculated from covariate-adjusted estimated marginal means; Cohen uses values of 0.2, 0.5, and 0.8 to characterize “small”, “medium”, and “large” differences between groups, respectively. SRD values, which can range from -1 to +1, represent the probability that a randomly selected case from one treatment will have an outcome preferable to a randomly selected case from another treatment. Outcome analyses were based upon the intention-to-treat principle (McCoy, 2017). Multiple imputation was used to impute missing data using fully conditional Markov chain Monte Carlo (MCMC; Schafer, 1987) modeling. The final analyses were based upon the pooled results of 20 separate imputation sets. Sensitivity analyses were conducted using both maximum likelihood imputation and available data analyses to evaluate the consistency of results across differing methods for handling missing data.
RkJQdWJsaXNoZXIy MjY0ODMw