Tools & techniques--statistics: propensity score techniques.

Area of Science:

• Epidemiology
• Biostatistics

Background:

• Conventional multivariable adjustment is challenging with many pretreatment variables and few outcome events.
• Propensity score (PS) methods offer an alternative for complex observational data analysis.

Purpose of the Study:

• To outline the utility and application of propensity score (PS) techniques in statistical analysis.
• To highlight best practices for deriving and validating PS models.

Main Methods:

• Deriving PS using only relevant pretreatment characteristics associated with the outcome.
• Visually inspecting PS distributions to identify areas of minimal overlap and potential residual confounding.
• Employing data trimming based on PS distribution to reduce residual confounding.
• Utilizing standardized differences in pretreatment characteristics to assess the success of PS methods.

Main Results:

• PS techniques are particularly beneficial when the number of potential confounding pretreatment variables is large and outcome events are small.
• Visual inspection of PS identifies areas of no or minimal overlap, indicating residual confounding.
• Data trimming based on PS distribution effectively minimizes residual confounding.
• Standardized differences serve as a crucial check for the validity of the PS approach.

Conclusions:

• Propensity score (PS) methods provide a feasible approach for complex confounding adjustment in observational studies.
• Careful application, including visual inspection and trimming, is essential for minimizing residual confounding.
• PS methods, like multivariable adjustment, cannot overcome unmeasured or imperfectly measured confounders and do not replace randomized controlled trials.