Recursive Partitioning Approaches to Modeling Genetics as a Source of Treatment Effect Heterogeneity

Pharmacogenomics has held the promise of delivering, “personalized medicine,” by identifying genetic determinants of variations in treatment response. While pharmacogenomics has successfully identified a handful of functional polymorphisms (relatively common variations in a single gene that affect the activity of a drug), the promise of personalized medicine has gone largely unfulfilled. Genetics researchers have come to realize that more complex interactions of genes, or of genes and environment, are more likely to underlie treatment effect heterogeneity than individual genes alone. The classical biostatistics approach to clinical trial data analysis requires that subgroups be defined ex ante. But because of the sheer number of genetic and environmental predictors and the potential for multi-level interactions, a curse of dimensionality becomes readily apparent in experimental design. Even post hoc subgroup analysis will suffer from this curse. Recursive partitioning methods (including classification and regression trees and random forests) repeatedly split observations into groups with maximally dissimilar outcomes according to the value of predictor variables. Their recursive nature allows for the exploration of multi-way interactions. However, existing, “greedy,” tree-building algorithms such as CART may fail to find highly important interactions when the direct effects of individual genes or environmental variables are small. While ensemble-based approaches such as bagging or random forests may more often correctly classify individuals, the required aggregation or averaging of classifiers makes interpretation of the models difficult. In this paper, an alternative approach is described that constructs and evaluates sets of random trees, scoring repeatedly-encountered interactions in proportion to their predictive ability and to the performance of the trees in which the interactions appear. Monte Carlo experiments are used to estimate the value of information generated by including subgroup effects identified by the algorithm in a simulated cost-effectiveness analysis of clinical trial data