Methods for Extending Inferences From Observational Studies: Considering Causal Structures, Identification...

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Quantitative generalizability and transportability often focus on randomized trials.
• Extending findings from observational studies presents unique challenges due to complex selection mechanisms.

Purpose of the Study:

• To describe identifiability assumptions and methods for extending causal effect estimates from observational studies to target populations.
• To highlight differences in methods compared to extending findings from randomized trials.

Main Methods:

• Described identifiability assumptions and identification using observed data.
• Employed statistical methods including weighting, outcome modeling, and doubly robust approaches.
• Illustrated method performance through a simulation study.

Main Results:

• Estimators must address confounding when selection occurs on exposure and confounders.
• When selection also involves mediators, estimators require methods to account for both selection and confounding using different variable sets.
• Demonstrated the performance of various statistical methods in simulation.

Conclusions:

• The proposed methods are applicable to complex causal structures common in observational studies, particularly life course epidemiology.
• Identified conceptual implications and practical questions for applying these methods in real-world scenarios.
• Emphasized the need for robust methods to handle selection bias and confounding for valid causal inference.