Population heterogeneity and causal inference.

Area of Science:

• Social Science Research
• Causal Inference
• Observational Data Analysis

Background:

• Population heterogeneity is a fundamental challenge in social science.
• Causal inference from observational data requires strong assumptions due to unobserved confounders and heterogeneous treatment effects.
• Existing research focuses on bias from unobserved pretreatment factors and varying treatment effects.

Purpose of the Study:

• To investigate the evolution of "composition bias" over time.
• To demonstrate how population heterogeneity interacts with treatment effect heterogeneity to create aggregate-level bias.
• To analyze bias in causal inference when treatment propensity is linked to treatment effect variation.

Main Methods:

• Theoretical analysis of composition bias.
• Modeling the dynamic emergence of bias at the aggregate level.
• Examining the interplay between treatment propensity and heterogeneous treatment effects.

Main Results:

• Composition bias evolves dynamically over time due to population heterogeneity.
• Systematic association between treatment propensity and heterogeneous treatment effects generates aggregate bias.
• This dynamic bias can arise even when the ignorability assumption holds at the microlevel.

Conclusions:

• Population heterogeneity significantly impacts causal inference in social sciences.
• Composition bias is a critical, dynamically evolving source of error in observational studies.
• Understanding this bias is crucial for accurate causal claims from social science data.