Propensity scores: from naive enthusiasm to intuitive understanding

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Estimating causal effects of binary exposures with confounding is crucial in medical research.
• Outcome regression modeling is a common approach, but propensity score methodology offers an alternative.

Purpose of the Study:

• To provide an intuitive, non-technical discussion of propensity score methodology.
• To motivate the use of propensity scores by analogy with randomized studies.
• To describe four main implementations of propensity score methods and the parameters they estimate.

Main Methods:

• Discussion of propensity score methodology.
• Analogy with randomized studies.
• Description of four implementation methods: matching, stratification, inverse probability of treatment weighting, and covariate adjustment.

Main Results:

• Propensity score methods can estimate causal effects under the assumption of no unmeasured confounders.
• Illustrative example using data on breastfeeding and infant neurodevelopment.
• Discussion of population parameters estimated by different propensity score methods.

Conclusions:

• Propensity score methods are a valuable tool for causal inference in observational studies.
• These methods remain controversial, with ongoing debate regarding their use compared to outcome regression.
• Further research is needed to clarify the relative advantages and disadvantages of propensity score methods.