In a causal analysis, a researcher may believe that the effect of treatment is not constant across the outcome distribution, making the common assumption of normally distributed data inappropriate. For example, a treatment may have greater impact at lower outcome values. In these cases, a natural estimand of interest is the quantile treatment effect (QTE) — the effect of treatment at a given quantile of the outcome distribution. One may also compute the QTE across various quantiles to understand how a treatment changes the overall shape of the outcome distribution. To this end, we propose a flexible extension of the nonparametric regression method BART, which we call Density Regression BART. Density Regression BART incorporates a latent variable to model complex structure in the mean and variance functions, going beyond BART’s standard framework of normality. This allows us to accurately estimate QTEs even when the treatment effect is skewed, multimodal, or otherwise irregular. The Bayesian nature of our method additionally allows us to estimate arbitrary functionals of the QTE, while maintaining proper uncertainty in these estimates. We assess the performance of our method against competing methods for QTE estimation in a simulation study and apply Density Regression BART to a real-world example estimating returns to education.