Model Averaging for Improving Inference from Causal Diagrams.

Area of Science:

• Epidemiologic research
• Causal inference
• Statistical modeling

Background:

• Model selection is a critical yet debated aspect of epidemiologic research.
• Lack of consensus on best model identification can lead to "wish bias", where researchers favor models supporting preferred outcomes.
• Directed acyclic graphs (DAGs) help identify adjustment sets for causal effect estimation but can yield multiple valid sets.

Purpose of the Study:

• To propose and evaluate model averaging techniques for obtaining causal estimates from multiple theoretically unbiased models.
• To mitigate the impact of "wish bias" in model selection within epidemiologic studies.
• To integrate uncertainty in the selection of candidate causal models.

Main Methods:

• Combining multiple adjustment sets identified via DAGs.
• Applying three model averaging techniques: information criteria weighting, inverse variance weighting, and bootstrapping.
• Illustrating the approach using data from the Pregnancy, Infection, and Nutrition (PIN) study.

Main Results:

• All three model averaging techniques produced similar, averaged causal estimates.
• The proposed a priori model averaging strategy effectively integrates uncertainty in model selection.
• The approach helps prevent the reporting of selectively attractive estimates from equally valid models.

Conclusions:

• Model averaging offers a robust method for causal effect estimation in the presence of multiple valid adjustment sets.
• This strategy enhances the objectivity of epidemiologic research by reducing "wish bias".
• Integrating model averaging provides a more reliable approach to causal inference.