Researchers has long been focusing on causal inference with binary or categorical treatments, where causal estimands are well understood and inference tools are rich. However, causal problems involving continuous treatments are common in practice, yet a formal framework is scarce in the literature. We extend classic Fisher's randomization test to address an initial question: does a treatment have effect on the outcome of interest conditional on a set of covariates. Our theory starts from randomized experiments and generalizes to observational studies. Inference tools are developed to establish causal relationships and to verify underlying assumptions.
翻译:暂无翻译