97 Social Avoidance and Testosterone Enhanced Exposure Efficacy in Women with Social Anxiety Disorder were included as random slopes for participant. Response direction, Time and Group (and their interactions) as random slopes for Stimulus model.1 Per registration, see also parent trial Hutschemaekers et al. (2021), we included endogenous baseline testosterone as an additional control variable (mean saliva sample 1 and 2) in all models. We used the Lme4 package in R (Bates et al., 2013) and p-values were calculated using the likelihood ratio tests in the Afex package (Singmann, 2013). The confidence intervals were determined using Lme4’s confint function using Bootstrapping (1000 simulations). Continuous predictor variables were centered, and sum-to-zero contrasts used. Consistent with the recommendations for mixed models (Pek & Flora, 2018), we report unstandardized effect sizes (estimates). Results Sample characteristics The data of 54 participants were analyzed (Mage = 23.31, SD = 5.64, range = 18–43) since one participant receiving placebo dropped out before the first exposure due to illness. Another participant in the same group dropped out during the first session (3.6%). All other participants completed both sessions and the follow-up. A full overview of the sample characteristics have been described elsewhere (Hutschemaekers et al., 2021). There were no baseline differences between the placebo and testosterone group on any of the AAT reaction times, all p-values > .257 (see Table 5.1). 1 We aimed to test a maximum random effect structure (picture type, response direction and time and their interactions as random slopes for the random intercept of participant and random slopes of response direction, time and group for the random intercept of stimulus model) but this model did not converge due to model estimation problems. Therefore, we ran simpler random effects models by dropping random slopes step by step and comparing the AIC after each step. As is common with mixed models, some of the simplified models also resulted in convergence warnings, but these warnings are more often false positives. In line with the recommendations by Bolker (2022), we used different optimizers (allFit function) and compared the estimates which all showed the same results and highly similar estimates. As such we decided to report the results of the model with the best fit (i.e. lowest AIC) and most extended random effect structure that was modeled. 5
RkJQdWJsaXNoZXIy MTk4NDMw