A comparison of mixed effects logistic regression models for binary response data with two nested levels of...

Area of Science:

• Biostatistics
• Toxicology
• Statistical Modeling

Background:

• Developmental toxicity studies generate complex, nested data.
• Rat pup outcomes are nested within litters, creating multiple levels of clustering.
• Ignoring clustering levels can lead to biased statistical inference.

Purpose of the Study:

• To compare mixed effects logistic regression models for nested binary data.
• To evaluate the impact of ignoring clustering levels in developmental toxicity data.
• To identify the most reliable model for analyzing pup and litter effects.

Main Methods:

• Comparison of three mixed effects logistic regression models: M1 (three-level), M2 (two-level, pup-clustered), and M3 (two-level, litter-clustered).
• Simulations and data analyses using developmental toxicity datasets.
• Assessment of bias and confidence interval coverage for fixed effects.

Main Results:

• The M3 model (litter-clustered only) exhibited significant bias for all effect types.
• The M1 model (three-level) achieved nominal confidence interval coverage.
• The M2 model (pup-clustered only) reduced coverage for litter-level effects, while M3 showed poor coverage for both pup and malformation effects.

Conclusions:

• A three-level model (M1) is recommended for accurately analyzing developmental toxicity data with nested structures.
• Ignoring pup or litter clustering (M2, M3) can lead to biased results and unreliable confidence intervals.
• Model choice significantly impacts the validity of conclusions in nested toxicity data analysis.