Optimally combining propensity score subclasses.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Propensity score methods, like subclassification, are vital for controlling confounding in non-randomized studies.
• Estimating causal effects requires careful handling of potential biases.

Purpose of the Study:

• To compare two methods of combining propensity score subclass estimates: weighting by proportion and weighting by inverse variance.
• To evaluate their performance under various conditions, including the number of subclasses, propensity score overlap, survey weighting, and treatment effect heterogeneity.

Main Methods:

• Simulation studies were conducted to assess method performance.
• Two combining strategies were compared: weighting by proportion of observations in subclass and weighting by inverse variance.
• Factors varied included number of subclasses, propensity score overlap (positivity), survey weighting, and treatment effect heterogeneity.

Main Results:

• Both methods performed well without positivity violations and with constant treatment effects, inverse variance weighting being slightly superior.
• Weighting by proportion was better when treatment effects varied across subclasses.
• In an example with practical positivity violations but no effect heterogeneity, inverse variance weighting yielded more efficient estimates.

Conclusions:

• Inverse variance weighting offers greater efficiency for combining propensity score subclass estimates, particularly with constant treatment effects.
• The choice of weighting method depends on the presence of treatment effect heterogeneity.
• These findings have implications for causal inference in observational studies, especially in public health research.