Rethinking the Match: A Simulation-Based Assessment of Congeniality in continuous Prediction Models

Missing data is a common challenge in medical research, and selecting an appropriate imputation method is crucial for accurate predictions. Congeniality refers to the alignment between the assumptions of the imputation model and the substantive prediction model. While this concept is well-understood in the context of parameter estimation, its implications for predictive performance and model calibration remain unclear. We evaluated congenial and uncongenial model combinations across various scenarios, reflecting different relationships between predictors and the outcome. Our analysis focused on predictive accuracy and calibration in settings with continuous predictors and continuous outcomes. To illustrate these findings, we conducted a case study using the MIMIC-III dataset. Across all simulation scenarios, congeniality had no observable impact on the accuracy or calibration of model combinations. However, patterns from both the simulation study and the MIMIC-III case study suggested that interactions between the imputation model and the substantive prediction model can influence overall performance. Accuracy and calibration are not determined by congeniality, while the combination of imputation and substantive prediction model matters. This is a Master thesis written at the University Medical Centre Utrecht and supervised by Maarten van Smeden, Anne de Hond, and Alex Carriero.