topic Re: Identify the best predictor of a count outcome in Statistical Procedures

Identify the best predictor of a count outcome

Melk — Fri, 24 Jan 2020 23:58:25 GMT

I am running a negative binomial regression on various predictors of a count outcome and want to be able to make a conclusion on which one is the "best" at predicting the counts. Are there any metrics that can be compared between models to make this conclusion? I was thinking psuedo r-squared, but understand that that can't be interpreted the same as a regular r-square. Any tips?

Re: Identify the best predictor of a count outcome

PGStats — Sat, 25 Jan 2020 04:44:55 GMT

If I needed a single goodness of fit measure, I would choose AICc. It is valid and robust as long as you are comparing models fit to the same set of dependent observations (in your case, the same set of counts), irrespective of the structure of the model. The lowest value is the best.

That said, never accept a model on the basis of a single measure. Always cheks the fit graphically. SAS provides many standard graphs (e.g. residuals vs predicted) to do so.

Re: Identify the best predictor of a count outcome

StatDave — Mon, 27 Jan 2020 16:24:54 GMT

Another possibility is to use the RsquareV macro to either get an R-square measure for the separate models or to assess the relative variable importance for each predictor in a multi-predictor model.