Comparison of robustness to outliers between robust poisson models and log-binomial models when estimating relative...

Area of Science:

• Biostatistics
• Statistical Modeling
• Epidemiology

Background:

• Log-binomial regression and robust Poisson models are popular for estimating relative risks of binary outcomes.
• Log-binomial models are thought to be more efficient, while robust Poisson models may handle outliers better.
• Limited evidence exists comparing the robustness of these two models.

Purpose of the Study:

• To evaluate the performance of robust Poisson and log-binomial models in scenarios with outliers.
• To compare their effectiveness in estimating relative risks for common binary outcomes.

Main Methods:

• A simulation study was conducted.
• The performance of robust Poisson and log-binomial models was assessed in various scenarios with simulated outliers.

Main Results:

• When the relationship between outcome and covariate was simple (log-linear), both models showed similar biases and mean square errors.
• Robust Poisson models consistently outperformed log-binomial models when the true relationship involved higher-order terms, even with low outlier levels.

Conclusions:

• Robust Poisson models demonstrate greater robustness to outliers than log-binomial models for estimating relative risks.
• Awareness of model limitations is crucial for appropriate selection in statistical analysis.