Title: Identifying influential observations in complex Bayesian mediation models with ordinal outcome
Authors: Sarka Rusa - Charles University (Czech Republic) [presenting]
Arnost Komarek - Charles University Prague (Czech Republic)
Emmanuel Lesaffre - University of Leuven (Belgium)
Luk Bruyneel - University of Leuven (Belgium)
Abstract: Although increasingly complicated (moderated) mediation models are being employed in practice, most of existing mediation literature has not dealt with model diagnostics. We propose a Bayesian approach to the detection of influential observations (or sets of observations). Importance sampling with weights which take advantage of the dependence structure in mediation models is utilized in order to estimate the case-deleted posterior means of the parameters. The method is applied to the ordinal measurements of patients' willingness to recommend hospitals collected on patients in a large European study to answer the research question whether the outcome depends on recorded system-level features in the organization of nursing care, and whether the related effect is mediated by two measurements of nursing care left undone and possibly moderated by nurse education.