Keywords: Dirichlet Process, Indirect Comparisons, Benefit/risk
In clinical development programs where component clinical trials compare different sets of non-unique doses of a compound, characterizing which dose has the optimal benefit-risk profile becomes challenging. For example, component trials in the program may enroll different types of patients, e.g., naïve patients, patients who can tolerate background therapy, patients who are contra-indicated to background therapy. For making indirect comparisons, combining trials with variable study design or methodological diversity in a network meta-analysis increases the risk of bias. Hence, care must be implemented particularly because baselines for these component trials are heterogeneous and could not possibly be explained by a normal distribution even if they are modelled separately from the relative effects. In this talk, I will explore the use Dirichlet Process priors to explain heterogeneity as it accommodates both between-study variability and non-normality (multi-modality and skewness) in the distribution of random effects. I will also discuss different types of models that relaxes the normality assumption.