Most of time conducting experiment is expensive. Suppose I and my collaborator decided to use a known databank(NHANE, NCI's data) to do some digging with some untested hypothesis given. Most of time any form of (observational/clinical trial) experiment will adjust according to hypothesis for confounders or at least try to balance covariates. A known databank does not necessarily balance covariates at all.
Suppose I can carry out calculation of causal effect defined by Rubin. Even if I balance covariates by propensity/weighting methods, to what extent can I trust my conclusion on significance of causal effect/association?
