Bayesian Causal Forests (BCF) estimate heterogeneous treatment effects using a separate ensemble of trees for the treatment effect. A key modeling choice is the prior on the leaf values of these treatment trees, which determines how sampling variability is attributed to treatment effects. The authors note that independent shrinkage priors, commonly used in this setting, do not control how a tree distributes variability across its leaves, and their total prior variance grows as more leaves are added.
To address this, the paper introduces the Dirichlet-Tree Bayesian Causal Forest, which places a Dirichlet-Tree prior on the variance shares of each treatment tree's leaves. This prior is generated by stick-breaking along the tree topology, allowing joint regularization of leaf values. The authors characterize the spike-and-tail behavior of the marginal leaf prior and show that the prior adaptively bounds the number of regions where effect heterogeneity appears. They also establish posterior consistency for the forest.
In simulations, the proposed model outperforms standard BCF when heterogeneity is confined to a few regions and reduces spurious heterogeneity elsewhere. Unlike independent shrinkage priors, its accuracy does not deteriorate as trees grow. The method is also evaluated on real data, including a pancreatic cancer survival study, though the abstract does not report specific numerical results for that application.