A machine learning compatible method for ordinal propensity score stratification and matching.

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning in Health Research

Background:

• Machine learning for propensity scores in multivalued treatment studies has advanced, but adjustment techniques lag.
• Current machine learning propensity models lack a single balancing score for stratification and matching.
• Existing methods struggle with the complexity of multivalued treatments in observational studies.

Purpose of the Study:

• To develop a flexible ordinal propensity scoring methodology for multivalued treatments.
• To create a scalar balancing score from machine learning propensity models for improved causal inference.
• To address the limitations of current propensity score adjustment techniques in complex observational studies.

Main Methods:

• Developed the Generalized Propensity Score - Cumulative Distribution Function (GPS-CDF) method.
• Fits a one-parameter power function to the CDF of the GPS vector from machine learning models.
• Utilizes the resulting scalar parameter ($$) as a balancing score for stratification or matching.

Main Results:

• Simulation studies demonstrate improved covariate balance and minimal bias in average treatment effect (ATE) estimates.
• The GPS-CDF method maintains coverage probability in statistical analyses.
• The method effectively produces unbiased ATE estimates by grouping similar subjects.

Conclusions:

• The GPS-CDF method offers a flexible and robust approach to propensity score adjustment in multivalued treatment studies.
• This methodology enhances causal inference from observational data by providing a reliable balancing score.
• The approach is applicable to real-world studies, such as the MATCh study, to investigate treatment effects.