Propensity score matching and subclassification in observational studies with multi-level treatments

Area of Science:

• Epidemiology
• Biostatistics
• Econometrics

Background:

• Estimating average treatment effects in observational studies is crucial for causal inference.
• Existing propensity score methods are well-established for binary treatments but face challenges with multiple treatment levels.
• Unconfoundedness given pretreatment variables is a key assumption in observational studies.

Purpose of the Study:

• To develop and validate new methods for estimating average treatment effects in observational studies with more than two treatment levels.
• To extend propensity score subclassification and matching methods to the multi-level treatment setting.
• To address biases arising from observed pretreatment variables in multi-level treatment analyses.

Main Methods:

• Development of new methods based on weak unconfoundedness and the generalized propensity score.
• Adaptation of propensity score subclassification and matching for multi-level treatments.
• Application of proposed methods to a real-world case study on fibromyalgia treatment.

Main Results:

• Demonstration that adjusting for a scalar function of pretreatment variables, via the generalized propensity score, eliminates bias.
• Successful application of the novel methods to analyze fibromyalgia treatment effects.
• Simulation study confirms the finite sample performance of the proposed methods.

Conclusions:

• The generalized propensity score provides a unified framework for addressing confounding in multi-level treatment studies.
• The developed methods offer a robust approach to estimating average treatment effects in complex observational settings.
• These advancements have significant implications for causal inference in fields utilizing multi-level interventions.