Tree models and other statistical approaches and issues for heterogeneity of treatment effects
With increased emphasis on comparative effectiveness research and patient centered outcomes, substantial attention has focused on estimating and accounting for heterogeneity of treatments effects (HTE). Most standard statistical models, however, instead estimate only average treatment effects (ATE), with the possible addition of a few interactions or subgroup analyses. In contrast, other applications such as predicting a response (to a single treatment), or early detection of disease, frequently utilize modern regression methods, or machine learning techniques to formulate models which implicitly model HTE. A key limitation is translating these methods to comparative effectiveness is that they do not allow for explicitly specifying a primary predictor of interest or report statistics such as significance or false significance levels, or interpretable coefficients and estimates of treatment effects. In this study, we address a very small portion of those challenges, by assessing significance and power for tree models where the treatment variable is forced into the model. Opportunities for larger investigations and applications for funding are also discussed.