The log multinomial regression model for nominal outcomes with more than two attributes

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Modeling

Background:

• Log binomial models estimate risk ratios for binary outcomes.
• Estimating relative risks for multi-category outcomes requires advanced statistical methods.

Purpose of the Study:

• To propose and evaluate the log multinomial model for relative risk estimation in nominal outcomes.
• To compare its performance against multinomial logistic regression and Poisson regression methods.

Main Methods:

• Extensive data simulations were used for performance comparison.
• The log multinomial model was compared with expanded data multinomial logistic regression and Poisson regression models.
• Model performance was assessed based on "inadmissible" solutions (out-of-bounds probabilities), relative bias, and mean squared error.

Main Results:

• The log multinomial model yielded "inadmissible" solutions in over 50% of some data settings.
• Alternative methods produced out-of-bounds probabilities for the log multinomial model in up to 27% of successfully fitted samples.
• Log multinomial coefficient estimates generally exhibited less relative bias and mean squared error compared to alternative methods.

Conclusions:

• The log multinomial model provides a practical approach for adjusted risk ratio estimation in multinomial settings.
• The model requires careful implementation and attention to detail to mitigate issues with probability bounds.
• It offers a valuable tool for analyzing complex categorical outcome data in epidemiological research.