Accounting for Missing Covariates in Heterogeneous Treatment Estimation

For example, if the initial study was an RCT, it may have failed to measure practically important Many applications of causal inference require covariates [Kahan et al., 2014] such as social using treatment effects estimated on a study determinants of health [Huang et al., 2024]. Since the population to make decisions in a separate intervention has not previously been used by the health target population. We consider the challenging system, no outcome data linked to these new covariates setting where there are covariates that are is available. However, treatment decisions would observed in the target population that were ideally reflect whether the intervention is likely to be not seen in the original study. Our goal is to beneficial to a patient conditional on all information estimate the tightest possible bounds on heterogeneous available, not just covariates that happened to be in the treatment effects conditioned on original study. This paper studies the question: how such newly observed covariates. We introduce precisely can we identify treatment effects conditional a novel partial identification strategy based on such new covariates? If precise estimates are available, on ideas from ecological inference; the main the decision maker can proceed confidently with idea is that estimates of conditional treatment deployment. Conversely, if considerable uncertainty remains effects for the full covariate set must about an important subgroup, a decision maker marginalize correctly when restricted to only may exercise more caution or invest more resources in the covariates observed in both populations.