Decision Theoretic Subgroup Detection With Bayesian Machine Learning
By: Entejar Alam, Poorbita Kundu, Antonio R. Linero
Potential Business Impact:
Finds best medicine groups for patients.
We consider the problem of identifying promising subpopulations in terms of treatment effectiveness or treatment effect heterogeneity, from a Bayesian decision theoretic perspective. We first show that a straight-forward application of Bayesian decision theory to subgroup detection leads to a counter-intuitive risk-seeking (RS) behavior. Motivated by this observation, we introduce the Bayesian Risk-Aware Inference and Detection of Subgroups (BRAIDS) utility and use it to perform subgroup selection and post selection inference. The BRAIDS utility interpolates between risk-seeking (RS) and risk-averse (RA) identifications of subgroups, with a variant of the virtual twins algorithm as its risk-neutral midpoint. We also argue that effective subgroup estimation and inference requires the use of regularization priors to safeguard inferences from the winner's curse. We provide empirical evidence that posterior credible intervals for subgroup effects can still obtain nominal coverage levels, provided that an appropriate prior distribution is chosen. The proposed framework is illustrated on data from clinical trial assessing the efficacy of canagliflozin as a treatment for type 2 diabetes.
Similar Papers
Causal Clustering for Conditional Average Treatment Effects Estimation and Subgroup Discovery
Machine Learning (Stat)
Finds groups who benefit most from treatments.
Data-driven controlled subgroup selection in clinical trials
Methodology
Finds best patient groups for medicine.
Identifying Treatment Effect Heterogeneity with Bayesian Hierarchical Adjustable Random Partition in Adaptive Enrichment Trials
Methodology
Finds best medicine for different patient groups.