Log-odds (logit)

A probability of 0.5 is log-odds 0, 0.75 is about 1.10 and 0.25 is about -1.10. Logistic regression models the log-odds as a weighted sum of inputs and maps it back with the sigmoid 1 / (1 + e^-z). Raw model scores before the sigmoid are often called logits; they are not probabilities and can be any real number. A coefficient b means the odds are multiplied by e^b for one unit of the input.