Discriminative Ability of the Beta-Binomial Regression Model for Prediction
Patient-reported outcomes (PROs) are widely used as primary endpoints in clinical research, and beta binomial regression (BBR) is suitable for modelling them. Prediction models aid clinical decision-making and often lead to clinical prediction rules estimating an individual’s risk of adverse events. Their discriminative ability is usually assessed through the AUC in logistic regression or the c-index and concordance probability in Cox models. Because prediction models depend on the response type, we introduce a measure to estimate the discriminative ability of the BBR model: the generalized c-index (GEC), and compare it with other concordance-based indices. The proposal is evaluated through simulation and applied to a cohort of patients with chronic obstructive pulmonary disease to predict health related quality of life. The model shows good discriminative performance (GEC = 0.87). We recommend the GEC as a robust estimator of discrimination for BBR in PRO prediction.
Keywords: Beta-binomial regression prediction models discriminative ability