Causal Inference
Interventionalist interpretations of studies involving compound treatments Kerollos Wanis* Kerollos Wanis Aaron Sarvet
When studying interventions, investigators will attempt to specify all their relevant characteristics. When a treatment has unspecified outcome relevant features, it has been referred to in the causal inference literature as a compound treatment. The study of compound treatments has been posed as a serious problem in causal inference research, due to violations of the ”Stable Unit Treatment Value Assumption.” But many treatments of interest are, in fact, compound treatments. This is true not only in observational research, but also in experimental studies. Investigators are unlikely to know all the outcome relevant features of an intervention to be studied. Indeed, the act of studying an intervention suggests some ignorance about its outcome causing mechanisms. And, even if known, were every intervention feature specified, the study might be obscure and of little practical relevance because the intervention would not reflect the natural flexibility of undergoing treatment in the real-world. We argue that causal effects of compound treatments are not only well-defined features of a study population, but might in fact be the effects most useful to decision makers in many scenarios. Inspired by the treatment decomposition that characterizes separable effect estimands, we reconsider compound treatments and the effects of interest when two such treatments are compared. This decomposition allows us to consider hypothetical interventions that modify the compound treatment. When the relevant features of a compound treatment are known and their values measured, data from observational or experimental studies can be used to study hypothetical modifications of compound treatments. When the outcome relevant features are not known, a separable effects decomposition can still allow investigators to study hypothetical modifications of compound treatments. We apply these methods to study the effect of donation after cardiac death on the survival of liver transplant recipients.