Two basic statistical strategies of conducting causal inference in real-world studies

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Randomized controlled trials (RCTs) are ideal but costly and limited in generalizability.
• Real-world studies (RWS) offer alternatives but are susceptible to confounding bias.
• Existing causal inference methods for RWS have complex statistical properties and assumptions.

Purpose of the Study:

• To provide a tutorial on two fundamental statistical strategies for causal inference in real-world studies.
• To clarify the underlying model assumptions and statistical properties of these strategies.
• To demonstrate their robustness and the need for sensitivity analysis.

Main Methods:

• Investigated the weighting strategy using propensity-score models (treatment assignment ~ confounders).
• Investigated the standardization strategy using outcome-regression models (outcome variable ~ treatment assignment + confounders).
• Compared the two strategies under identifiability conditions.

Main Results:

• Both weighting and standardization strategies are robust for estimating treatment effects.
• Identifiability conditions for both strategies are the same.
• Sensitivity analysis is equally necessary for both strategies to assess assumption violations.

Conclusions:

• Weighting and standardization are two core, robust strategies for causal inference in real-world studies.
• Despite different initial approaches, both methods rely on the same identifiability conditions.
• Consistent sensitivity analysis is crucial for validating findings from both strategies.