Factoring vs linear modeling in rate estimation: a simulation study of relative accuracy

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Modeling

Background:

• Dose-response modeling in epidemiology often involves factoring ordered variables into dichotomous indicators.
• Factoring can influence estimation variance, bias, and overall accuracy in statistical models.

Purpose of the Study:

• To investigate the impact of variable factoring on the accuracy of rate estimation in epidemiological studies.
• To evaluate how factoring affects model accuracy across diverse population model forms and effect sizes.

Main Methods:

• A simulation study was conducted using 37,500 randomly generated population model forms.
• Poisson regression models, both factored and unfactored, were fitted to simulated follow-up study datasets.
• Models ranged from subadditive to supramultiplicative to capture varied exposure-response relationships.

Main Results:

• Factoring variables sometimes substantially improved, sometimes substantially decreased, and sometimes had little effect on accuracy.
• Accuracy differences between factored and unfactored models were complexly dependent on true vs. fitted model forms, effect strengths, and study size.
• Predicting whether factoring improves or degrades accuracy in practice can be challenging.

Conclusions:

• The impact of factoring on rate estimation accuracy is variable and context-dependent.
• Relying solely on factoring may not consistently yield optimal results.
• Supplementing variable factoring with alternative dose-response modeling strategies is recommended for robust analysis.