Multinomial Extension of Propensity Score Trimming Methods: A Simulation Study.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Propensity score (PS) trimming methods by Crump et al., Stürmer et al., and Walker et al. were proposed to enhance efficiency and reduce confounding.
• Existing binary PS trimming methods may not adequately address complex scenarios with multiple treatment groups.

Purpose of the Study:

• To generalize existing propensity score trimming methods to accommodate multinomial treatment groups.
• To evaluate the performance of these novel multinomial trimming methods in reducing bias and variance, particularly in the presence of unmeasured confounders.

Main Methods:

• Generalized propensity score trimming definitions for multinomial treatments were developed and validated against binary definitions.
• Simulations were conducted using three treatment groups with varying sizes, employing inverse probability of treatment weights, matching weights, and overlap weights.
• Performance was assessed based on bias reduction and variance reduction under different weighting schemes.

Main Results:

• The proposed multinomial trimming methods successfully reduced bias across various weighting strategies, especially when treatment group sizes were disparate.
• Multinomial Stürmer and Walker trimming demonstrated superior bias reduction in scenarios with highly unequal group sizes.
• Variance reduction was more effective with multinomial Crump and Stürmer trimming when using inverse probability of treatment weights.

Conclusions:

• The generalized multinomial propensity score trimming methods offer significant benefits in controlling for measured and unmeasured confounders in multi-treatment studies.
• These methods provide a valuable extension to existing propensity score techniques for complex causal inference problems.