When the outcome is yes or no, a straight line will predict probabilities below zero and above one. Logistic regression models the log-odds instead. Magic Stat fits it from your spreadsheet and returns odds ratios, a ROC curve with its AUC, and a classification table, in one dialog.
Logistic regression predicts the probability of a binary event through the log-odds, which keeps predictions between 0 and 1 and lets each predictor's effect be written as an odds ratio. An odds ratio of 2 means the odds of the event double per one-unit increase in the predictor, holding the others fixed. Magic Stat reports each coefficient, its odds ratio and their confidence intervals.
How the outcome is encoded matters. A column of 0/1 is used as is; a two-category text column is figured out automatically, with the second level in order treated as the event. The dialog states which level is the event, so the odds ratios cannot be read backwards.
Logistic regression has a specific failure mode: perfect separation, when a predictor splits the two outcome groups completely. Then the maximum-likelihood estimates do not exist, and software that returns huge coefficients with tiny standard errors is reporting an artefact. Magic Stat refuses to present such estimates and names the predictor responsible.
Outcome: a column with two categories or 0/1; the event level is stated explicitly.
Predictors: one or more numeric columns, selected together.
Plots: the predicted-probability S-curve and the ROC curve open in their own editable windows.
A 0/1 column is used directly, with 1 as the event. For a two-level text column, the levels are ordered and the second is treated as the event. The dialog prints the reference and the event level so the odds ratios are unambiguous.
An odds ratio above 1 means the odds of the event rise as the predictor rises; below 1, the odds fall. It is a multiplicative effect on the odds, not on the probability — for a common outcome the two differ substantially.
Almost always perfect separation or a constant predictor. When a predictor splits the outcome groups completely, the maximum-likelihood estimates do not exist and cannot be reported. Magic Stat identifies the offending predictor so you can remove it or reconsider the model.
Logistic regression needs enough events, not just enough rows — a handful of events against many predictors gives unstable estimates. Magic Stat requires at least 20 complete rows as a floor, but that is a minimum, not a target.
No. The fit is a maximum-likelihood logistic model, and the AUC confidence interval uses the DeLong method. If you need penalised or mixed-effects logistic models, those live in other tools and in Magic Stat's other analyses.
Multinomial logistic regression → Generalized linear models → Medical statistics →