On model selection and model misspecification in causal inference

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Standard variable selection procedures are commonly used in observational studies for outcome prediction.
• Their application to assess exposure effects can lead to biased estimators for both effects and their uncertainty.

Purpose of the Study:

• To highlight the sub-optimal nature of standard variable selection in exposure effect assessment.
• To propose a novel confounder-selection procedure that directly improves exposure effect estimation.
• To identify causal inference strategies robust to confounder selection issues.

Main Methods:

• Comparative analysis of standard variable selection versus a proposed targeted procedure.
• Evaluation of causal inference strategies for their robustness and validity of confidence intervals.
• Assessment of bias in exposure effect and uncertainty estimators.

Main Results:

• Standard variable selection procedures are prone to introducing bias in observational studies.
• The proposed procedure directly targets the quality of the exposure effect estimator.
• Certain causal inference strategies provide valid confidence intervals and robustness against misspecification, even when confounder selection is ignored.

Conclusions:

• Rethinking variable selection is crucial for accurate exposure effect estimation in observational research.
• The proposed method offers a more direct approach to improving causal inference quality.
• Specific causal inference techniques demonstrate resilience to confounder selection complexities.