Comparison of logistic-regression based methods for simple mediation analysis with a dichotomous outcome variable

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Logistic regression is frequently employed for mediation analysis involving dichotomous outcomes.
• Previous research indicates that scale changes in logistic regression models can impact indirect effect and proportion mediated estimates.
• Standardization has been proposed as a method to address these scale-related issues.

Purpose of the Study:

• To compare the performance of unstandardized and standardized estimates for indirect effects and proportion mediated.
• To evaluate these estimates across multiple regression, structural equation modeling (SEM), and potential outcomes frameworks.
• To provide guidance on optimal methods for mediation analysis with dichotomous outcomes.

Main Methods:

• A simulation study was conducted to assess the performance of different estimation methods.
• Two real-life datasets from an observational cohort study (n=360) were analyzed.
• Performance was evaluated based on bias and efficiency of effect estimates.

Main Results:

• Estimates from the potential outcomes framework demonstrated the lowest bias and highest efficiency.
• Crude indirect effect (ab) and proportion mediated (ab/(ab+c’)) based on multiple regression and SEM also showed favorable performance.
• Standardized estimates did not consistently outperform unstandardized estimates.

Conclusions:

• The potential outcomes framework provides reliable estimates for mediation analysis with logistic regression.
• Unstandardized indirect effect (ab) and proportion mediated (ab/(ab+c’)) estimates from multiple regression and SEM are recommended.
• Standardization of coefficients before estimation is not advised as it may not enhance performance.