Large Sample Size Effects on Maximum Likelihood Estimation in Multinomial Logistic Regression
Abstract Multinomial logistic regression is used for outcomes with more than two categories, but evidence on how large sample sizes affect coefficient-estimation performance remains limited for specific model structures. Previous studies have often focused on prediction, sample-size planning, sparse-data methods, or other estimation settings. This study assessed coefficient performance in a five-category multinomial logistic regression model with ten binary explanatory variables, using category 5 as the reference. Monte Carlo simulation was conducted at sample sizes of 1,000, 3,000, 5,000, and 10,000, with 10,000 datasets generated at each size. The model was fitted to each dataset, and the mean estimates, bias, empirical standard errors, and mean squared errors (MSEs) were calculated for 44 coefficients. All 10,000 model fits succeeded at each sample size. Across the 44 coefficients, mean absolute bias decreased from 0.0453 at \(\:n\:=\:\text{1,000}\) to 0.0038 at \(\:n\:=\:\text{10,000}\). Median coefficient-level MSE decreased from 0.0598 to 0.0055, while mean empirical standard error declined from 0.3216 to 0.0922. MSE decreased for every coefficient at each successive sample size. The study fills a focused evidence gap by comparing coefficient-level estimation measures across several large sample sizes within the same five-category model. Under the specified simulation conditions, larger sample sizes were associated with estimates that were, overall, less biased and more precise.
Authors
- A. W. Babayemi
- Abubakar Abdullahi
Institutions
- Federal University Birnin Kebbi (NG)
- Abdullahi Fodio University of Science and Technology (NG)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-16
- DOI
- https://doi.org/10.21203/rs.3.rs-11464640/v1
- Primary Topic
- Statistical Methods in Epidemiology
- Type
- preprint