Title: Achieving a Comprehensive Grasp of Selection Bias in Causal Effect Identification

Abstract:

Selection bias is a common challenge in observational research. For instance, data from large-scale biobanks often display a "healthy volunteer bias," where participants are generally healthier and possess higher socio-economic status than the broader population they are intended to reflect. Addressing this issue is critical in causal inference; deriving average treatment effects (ATE) from such selected sub-populations can lead to significantly skewed estimates of the ATE for the general population. This study explores the identifiability of the ATE when selection bias is present. By imposing minimal assumptions on probability classes, we characterize both the propensity score and selection probability to establish necessary and sufficient conditions for ATE identifiability. Our findings build upon prior research by expanding existing graphical identifiability criteria, thereby providing a more thorough framework for identifying causal effects under selection bias, supported by strictly weaker conditions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC