topic Re: Proc genmod qic in Statistical Procedures

Proc genmod qic

dapenDaniel — Thu, 25 Jun 2020 02:05:48 GMT

hello SAS experts,

I am using proc genmod to run a GEE regression. The dependent variable is count data so negative binomial distribution is specified. The code is below.

proc genmod data=have;
class firmid year;
model DV = IV control variables / dist=nb link=log;
repeated subject = firmid / within=year type=ind;
run;

The results show a negative QIC (-158900.5758). How to interpret this negative QIC? Thanks.

Re: Proc genmod qic

Ksharp — Thu, 25 Jun 2020 11:48:26 GMT

I think it like AIC BIC is a goodness of fit statistic ( abs(QIC) ). Check how much the model fit data.
Near zero means the model is better .
@Rick_SAS might explain it more detail .

Re: Proc genmod qic

Rick_SAS — Thu, 25 Jun 2020 12:45:14 GMT

Like other likelihood-based criteria, you shouldn't worry about the actual number or whether it is positive or negative. These "information criteria" are used to compare different models. The model with the smaller QIC is a better fit. For more information about QIC, see

https://support.sas.com/kb/23/109.html

Re: Proc genmod qic

dapenDaniel — Thu, 25 Jun 2020 16:38:52 GMT

Thank you @Rick_SAS

Please allow me to ask a following question. If I have two models as below, is the model with more negative QIC score (Model1) the better one since it is smaller?

Model1 Model2

QIC -15 -10

Re: Proc genmod qic

dapenDaniel — Thu, 25 Jun 2020 16:46:09 GMT

Thank you @Ksharp

If I have two models, both QICs are negative (let's say -15 and -10). Is the one whose QIC is -10 better? If so, when QICs are negative, the one with larger QIC is better. Am I correct?

Re: Proc genmod qic

Rick_SAS — Thu, 25 Jun 2020 16:55:41 GMT

Correct. Because the QIC for Model1 is less than the QIC for Model2, Model1 is better according to that criterion.

Re: Proc genmod qic

dapenDaniel — Thu, 25 Jun 2020 17:02:00 GMT

Thank you very much! @Rick_SAS