How to perform prespecified subgroup analyses when using propensity score methods in the case of imbalanced subgroups

Area of Science:

• Biostatistics
• Epidemiology
• Observational Data Analysis

Background:

• Investigating treatment-by-subset interactions using propensity-score (PS) modeling on observational data with right-censored outcomes is crucial.
• Challenges arise in implementation, particularly with imbalanced subsets regarding prognostic features and treatment prevalence.

Purpose of the Study:

• To compare two propensity-score (PS) estimation strategies: 'across subset' (whole sample) and 'within subsets' (separately).
• To evaluate various PS models and estimands for treatment-by-subset interaction analysis.
• To illustrate these approaches using a real-world example of parotid cancer treatment.

Main Methods:

• Conducted a simulation study comparing 'across subset' and 'within subsets' PS estimation strategies.
• Investigated multiple PS models and estimands.
• Applied the strategies to observational data on facial nerve resection for parotid cancer, stratified by pretreatment facial palsy.

Main Results:

• Both PS estimation strategies yielded similar bias and variance for estimated treatment effects.
• The 'across subsets' strategy showed a slight advantage in very small samples when interaction terms were included.
• Propensity-score matching without replacement led to biased estimates in imbalanced subsets and should be avoided.

Conclusions:

• For analyzing treatment effect heterogeneity in small samples with imbalanced data, the 'across subsets' PS estimation strategy is recommended.
• Utilize PS matching with replacement or weighting methods for estimating average treatment effects in the treated or overlap populations.
• Propensity-score matching without replacement is not suitable for imbalanced subsets in this context.