Bayesian additive regression trees (BART) induce a highly regarded prior distribution over functions. BART has proven particularly well-suited to estimating causal effects, originally by simply appending a treatment assignment indicator as a covariate and computing relevant counterfactual predictions. Recently Hahn, Murray and Carvalho (2017) demonstrated that parameterizing a causal model directly in terms of the heterogeneous treatment effect function can be desirable, as it allows for prior elicitation and specification directly on the parameter of interest. That paper focused on binary treatments. In this talk we argue that in regimes with continuous treatments and moderate to strong confounding, this parameterization is even more essential as it provides direct control over the manner and degree of the extrapolation necessary to make causal inference. We describe some difficulties in moving from binary to multi-valued and continuous treatments, and how one might mitigate these in the context of flexible regression models, and BART-based methods in particular.