topic A question on the BINWIDTH option in PROC LOGISTIC MODEL Statement. in Statistical Procedures

A question on the BINWIDTH option in PROC LOGISTIC MODEL Statement.

jbilenas — Tue, 01 Feb 2022 21:03:16 GMT

The SAS documentation state that for SAS/STAT releases before SAS/STAT 14.1 the BINWIDTH is 0.002 which I assume results in 500 bins. I typically build Credit Risk logistic models without binning continuous independent variables.

For SAS/STAT versions starting at 14.1 or higher and there are fewer than 5 million observations then the default BINWIDTH=0 which means that no binning is done.

Am I interpreting correctly?

Re: A question on the BINWIDTH option in PROC LOGISTIC MODEL Statement.

ballardw — Tue, 01 Feb 2022 21:19:50 GMT

If you are referring to

if you have a binary response and fewer than 5,000,000 observations, the default is BINWIDTH=0; this means that no binning is performed and the exact values of the statistics are computed using the trapezoidal area described in the section ROC Computations. This method is a bit slower and might require more memory than the binning approach.

Then yes, no binning is done UNLESS you request it. The Default is 0. You can set a bin width to force binning if needed.

Part of the note is there to provide documentation for those folks that run an old program on existing data and get a different result that could occur from the change of the DEFAULT Bidwidth setting, i.e. not specified in the model.

Re: A question on the BINWIDTH option in PROC LOGISTIC MODEL Statement.

Rick_SAS — Tue, 01 Feb 2022 23:18:14 GMT

> I typically build Credit Risk logistic models without binning continuous independent variables.

PROC LOGISTIC does not automatically bin continuous independent variables, and it never has. The doc is talking about the estimation of the association statistics. The binning takes place in the interval of predicted probabilities and is used to quickly estimate some statistics that have to do with goodness of fit.

Using the BINWIDTH= option only affects the concordance and related statistics in the Association table. It doesn't affect parameter estimates and other model-related estimates.