Average causal effects from nonrandomized studies: a practical guide and simulated example.

Area of Science:

• Statistics
• Epidemiology
• Social Sciences

Background:

• Random assignment simplifies causal inference.
• Observational studies often suffer from confounding bias, which traditional methods like analysis of covariance may not fully address.

Purpose of the Study:

• To review Rubin's definition of average causal effect (ACE).
• To distinguish ACE from regression coefficients.
• To present and evaluate nine strategies for estimating ACEs in observational data.

Main Methods:

• Review of regression-based, propensity score, and doubly robust methods for ACE estimation.
• Simulation of an observational study on dieting and emotional distress in adolescent girls.
• Assessment of methods based on bias, efficiency, and interval coverage.

Main Results:

• Traditional regression methods may not eliminate bias in observational studies.
• Propensity score and doubly robust methods offer improved approaches for causal inference.
• The study provides practical guidance and formulas for standard errors.

Conclusions:

• Accurate causal inference from observational data requires careful consideration of confounding.
• The reviewed methods offer researchers tools to better estimate average causal effects.
• This work aids researchers in drawing more reliable conclusions from non-randomized studies.