Page 213 - Demo

General Discussion2117for their dependency. Importantly, the interpretation of coefficients in these models are highly dependent on the model definition, such as whether one level of a variable serves as reference (treatment coding) or whether the coefficient of each level relates to the average effect of all levels (sum coding). To answer specific questions, such as whether there is a significant relation to the outcome in a specific group, additional post-hoc comparisons (e.g., slope comparisons against zero for continuous variables) are necessary. Thus, researchers should be careful in interpreting their model fits, and be aware whether they are actually able to answer their questions, based on the model definition. Moreover, the robustness of observed (small) effects is often questionable and the ideal of normally distributed outcome data, or at least residuals, often does not match reality. Under the guidance of my colleague Tom Roth, I employed Bayesian statistics to address these issues. Namely, in the majority of my studies, I defined Bayesian multi-level models in addition to Frequentist multi-level models, thereby gaining more insight in the robustness of results. Furthermore, in Chapter 6, in which the sample size of the social anxiety group was low, Bayesian statistics allowed me to evaluate whether there was enough evidence for the alternative hypothesis, and also for the null hypothesis. The Bayesian approach further allowed me to model data distributions that were difficult to capture in commonly-used model families in Frequentist analyses. Rating scales, for example, are not always interpreted as continuous, with most people choosing middle values. In my studies, I observed that people sometimes tend to choose extremes over middle values in their confidence ratings (Chapter 5) or that they interpret scales in a more ordinal than continuous way, with distances not being equal between all data points (Chapter 6). Bayesian modeling allowed me to translate these observations in models, using generalized linear mixed models with a zero-one-inflated family (Chapter 5) and sequential mixed models with an s-ratio family (Chapter 6). Exploring the nature of collected data and searching for well-suited approaches to analyze them is a necessary step for researchers to obtain confidence in their results. Practical ImplicationsBefore directly evaluating the results of this thesis in the light of clinical practice, I would like to discuss an important implication which is related to my personal motivation to conduct psychological research. Progressing through my Bachelor%u2019s