Predicted Probabilities and Inference with Multinomial Logit

Multinomial logit (MNL) differs from many other econometric methods because it estimates the effects of variables upon nominal, not ordered outcomes. One consequence of this is that the estimated coefficients vary depending upon a researcher’s decision about the choice of a reference, or “baseline,” outcome. Most researchers realize this in principle, but many focus upon the statistical significance of MNL coefficients for inference in the same way that they use the coefficients from models with ordered dependent variables. In some instances, this leads researchers to report statistics that do not reflect the correct quantities of interest and reach flawed conclusions. In this note, I argue that researchers need to orient their approach to analyzing both the substantive and statistical significance of predicted probabilities of interest that match their research questions.