Multinomial Logistic Regression
Also known as: polytomous logistic regression, softmax regression, multinomial logit, nominal logistic regression
Multinomial logistic regression extends binary logistic regression to outcomes with three or more unordered categories. It models the log-odds of each category relative to a chosen reference category as a linear function of the predictors, and estimates all parameters simultaneously via maximum likelihood. It is the standard choice when the dependent variable is nominal with multiple levels.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use multinomial logistic regression when the outcome is nominal with three or more unordered categories and the predictors are continuous, categorical, or mixed. It is appropriate when no natural ordering exists among the outcome levels (if an ordering exists, prefer ordinal logistic regression). Recommended sample size is at least 10 events per predictor per category contrast. Do not use it when the outcome is ordinal (use proportional-odds or adjacent-category ordinal models), when the outcome is binary (use standard logistic regression), or when the number of categories is very large and sparse (consider regularisation or other classifiers).
Strengths & limitations
- Handles nominal outcomes with any number of unordered categories in a single, unified model.
- Produces calibrated class probabilities via the softmax function, enabling probabilistic prediction and comparison across categories.
- Relative risk ratios are intuitively interpretable: each exp(β) is a multiplicative effect on the odds of one category versus the reference.
- Accommodates continuous, binary, and categorical predictors and their interactions without assuming normality of predictors.
- Maximum-likelihood framework provides likelihood-ratio tests and standard errors for hypothesis testing and confidence intervals.
- Well-supported in all major statistical software (R, Python statsmodels, Stata, SAS, SPSS).
- Requires substantially larger samples than binary logistic regression because K-1 parameter sets are estimated simultaneously; sparse cells severely inflate standard errors.
- The independence of irrelevant alternatives (IIA) assumption is implicit: adding or removing one outcome category should not change the relative odds of the remaining categories, which may not hold in practice.
- Interpretation becomes complex when K is large: there are (K-1) × p coefficients to report and summarise.
- No closed-form solution; convergence can fail with perfect separation or very small category frequencies.
- Ignores natural ordering in the outcome; if an ordering exists, ordinal models are more parsimonious and powerful.
Frequently asked
How does multinomial logistic regression differ from ordinal logistic regression?
Multinomial logistic regression treats the outcome categories as unordered (nominal) and estimates a separate equation for each non-reference category. Ordinal logistic regression exploits a natural ordering among categories and models cumulative log-odds with fewer parameters. When the outcome has a meaningful order, ordinal models are more parsimonious and should be preferred.
What is the independence of irrelevant alternatives (IIA) assumption and how can I test it?
IIA states that the relative probability of choosing category A over B is unchanged by the presence or absence of other categories. It can be tested with the Hausman-McFadden test or the Small-Hsiao test. If IIA is violated, nested logit or mixed logit models that relax the assumption are alternatives.
How do I choose the reference category?
Any category can serve as the reference; model fit is identical regardless of the choice. A natural or meaningful reference (e.g., a control group, the most common category, or the theoretically neutral option) makes coefficients easier to interpret. Changing the reference rescales all coefficients but does not change predicted probabilities or overall fit.
What pseudo-R² should I report for multinomial logistic regression?
McFadden's pseudo-R² = 1 − (log L_model / log L_null) is most commonly reported and comparable across studies. Nagelkerke's R² is also widely used. Neither is equivalent to OLS R²; values above 0.20 generally indicate good fit for McFadden's measure. Always supplement pseudo-R² with the likelihood-ratio chi-square test.
Can I use regularisation with multinomial logistic regression?
Yes. L2 (ridge) and L1 (lasso) penalties can be applied to the multinomial log-likelihood, which is especially useful when predictors are numerous relative to observations or when multicollinearity is present. Regularised multinomial logistic regression is available in scikit-learn, glmnet (R), and related packages.
Sources
- Agresti, A. (2002). Categorical Data Analysis (2nd ed.). Wiley-Interscience. ISBN: 978-0471360933
- Hosmer, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied Logistic Regression (3rd ed.). Wiley. ISBN: 978-0470582473
How to cite this page
ScholarGate. (2026, June 3). Multinomial Logistic Regression. ScholarGate. https://scholargate.app/en/statistics/multinomial-logistic-regression
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Discriminant AnalysisStatistics↔ compare
- Logistic RegressionResearch Statistics↔ compare
- Ordinal Logistic RegressionStatistics↔ compare
- Random ForestMachine learning↔ compare