New Causal Inference Method Addresses Selection Bias in Observational Studies
A novel method has been introduced by researchers to detect causal effects in the presence of selection bias, a frequent challenge in observational studies that often lack a representative sample of the entire population. The findings, shared on arXiv, outline both necessary and sufficient conditions for determining the average treatment effect (ATE) amid selection bias. This technique employs minimal assumptions regarding probability classes to define propensity scores and selection probabilities, enhancing existing graphical criteria. The research tackles problems such as 'healthy volunteer bias' in extensive biobanks, where participants tend to be healthier and wealthier than the broader population. This study provides a deeper insight into causal effect identification with significantly weaker assumptions compared to prior research.
Key facts
- Selection bias is pervasive in observational studies.
- Large scale biobanks can exhibit 'healthy volunteer bias'.
- Recovering causal effects from sub-populations is important in causal inference.
- Estimating ATE from selected populations can result in biased estimates.
- The paper investigates identifiability of ATE under selection bias.
- Provides necessary and sufficient conditions for ATE identifiability.
- Leverages weak assumptions on probability classes.
- Extends existing graphical identifiability criteria.
Entities
Institutions
- arXiv