Home › Stats-Regression

Multinomial logit — Lodestar.Stats.Regression

One entry point, MultinomialLogit.Fit, for a response with more than two unordered categories. Each category but the smallest label gets its own equation against that reference, and the fit reports what statsmodels' MNLogit(...).fit() summary holds: the coefficients per equation with their standard errors, z statistics, p-values and intervals, the log-likelihood, McFadden's pseudo-R², the likelihood-ratio test, AIC and BIC.

Why it is not a GLM family. The fit is Newton-Raphson on the analytic score and Hessian, not IRLS. Its table is a matrix, one row per equation, where GlmSummary's is a vector.

The ordered model is not here. statsmodels' OrderedModel differentiates its likelihood numerically, and does not reproduce its own answer at the corpus tolerance (decision 0004).

Types

Type	What it is
MultinomialLogit	Fits the model by Newton-Raphson and builds the table.
MultinomialLogitSummary	The coefficients per equation, their inference, and the whole-model numbers.
MultinomialLogitOptions	The intercept, the confidence level, and Newton's budget and tolerance.

See also

• Generalized linear models — the binomial logit is this model at two categories.
• statsmodels → .NET — what is delegated and what is not.
• Python → C# equivalence.