When the outcome has more than two categories — a diagnosis, a soil class, a preferred option — a binary model is not enough. Multinomial logistic regression fits the probability of each category against a reference you choose. Magic Stat runs it from your spreadsheet and returns the odds ratios and the classification matrix.
With k outcome categories, multinomial logistic regression fits k−1 equations, each contrasting one category with a reference category. A coefficient is the change in the log-odds of being in that category rather than the reference, per unit of the predictor, and its exponential is an odds ratio. Magic Stat reports every block with its own coefficients and confidence intervals.
The reference category is a choice, not a fact of the data. Changing it does not change the fit — the log-likelihood and the predicted probabilities stay the same — it only changes which comparisons the coefficients describe. The dialog lets you set it, and the default is the first category in order, as in R's nnet::multinom.
This method models nominal categories, where the order carries no meaning. If the outcome is ordinal — mild, moderate, severe — an ordinal model usually fits the ordering better and uses the data more efficiently, and Magic Stat does not fit one.
nnet::multinom.Outcome: a categorical column with three or more levels (two levels belong in the binary logistic dialog).
Reference category: choose which level is the baseline; the default is the first level in order.
Predictors: one or more numeric columns, selected together.
Pick the category that makes the comparisons meaningful — often a control, a baseline or the most common level. Changing the reference never changes the fit; it only re-expresses the coefficients relative to a different baseline.
Each block compares one category with the reference. An odds ratio of 2 in a block means the odds of being in that category rather than the reference are twice as high per one-unit increase in the predictor, holding the others fixed.
With only two categories, use binary logistic regression — the dialog will point you there. With ordered categories, an ordinal model is usually a better fit to the structure than treating the levels as unrelated.
If the categories have a natural order — severity grades, Likert levels — they are ordinal and this nominal method ignores that order. Fitting an ordinal model, such as R's MASS::polr, is the more appropriate route; Magic Stat fits the nominal case.
Each category needs enough observations for its coefficients to be stable; sparse categories give unstable estimates. Magic Stat requires at least 30 complete rows as a floor, but more data per category is the real requirement.
Logistic regression → Generalized linear models → Contingency tables →