topic Re: PROC QLIM and PROC Logistic results not matching in Statistical Procedures

PROC QLIM and PROC Logistic results not matching

tbell — Mon, 08 Mar 2021 17:24:17 GMT

Morning,

We do not have much experience with PROC QLIM but needed to use it to get a Mills ratio. Our dependent variable has 2 values (0,1) and here is our code.

proc qlim data= &indata. ;
model MA_ENROLL = <list of variables> / discrete(d=normal);
output out=in.prob_enroll_&indata. marginal mills prob xbeta;
run;

The results of this model are:

We did the same using PROC Logistic -- for which we are very familiar. We got the same results for the dependent variable = 1. But got 1-p results for dependent variable = 0. And these results are more expected given the data.

So it appears the the above table from QLIM (peach color cell) is the probability of the dependent variable being 0 for observations when the dependent variable is 0. And not the probability of the dependent variable being 1 for observations when the dependent variable is 0.

Thanks in advance for help.

Re: PROC QLIM and PROC Logistic results not matching

tbell — Mon, 08 Mar 2021 17:31:47 GMT

I forgot to include the logistic results for comparison.

Re: PROC QLIM and PROC Logistic results not matching

ballardw — Mon, 08 Mar 2021 17:54:44 GMT

It might not hurt to provide the Proc Logistic Code as well.

You might test your hypothesis by changing the reference value in proc logistic.

Re: PROC QLIM and PROC Logistic results not matching

tbell — Mon, 08 Mar 2021 18:01:08 GMT

Thanks. Here is the form of the logistic code:

proc logistic DESCENDING data = newdata;
model y = x1 /link=probit;
output out=predict2 pred=phat;
run;

Re: PROC QLIM and PROC Logistic results not matching

StatDave — Mon, 08 Mar 2021 18:27:19 GMT

PROC LOGISTIC models the probability of the first sorted response level by default, which means 0 for a 0/1 response variable. I believe QLIM expects a 0/1 response variable and always models the probability of 1. If you can to model the probability of 1 in PROC LOGISTIC, specify the EVENT="1" response option: model y(event="1") = ...

Re: PROC QLIM and PROC Logistic results not matching

tbell — Mon, 08 Mar 2021 18:46:20 GMT

Thanks. We used the DESCENDING option in Logistic so I think we are getting probability of 1. And that is what we want. But how do we get the same in QLIM?

Re: PROC QLIM and PROC Logistic results not matching

StatDave — Mon, 08 Mar 2021 19:10:18 GMT

Try the PROBALL option in the OUTPUT statement instead of the PROB option. One of the two resulting columns should match the predicted probabilities output by PROC LOGISTIC.

Re: PROC QLIM and PROC Logistic results not matching

tbell — Mon, 08 Mar 2021 19:19:27 GMT

Thanks Dave,

Will try that -- currently the system with the data is down at the moment, but will try that asap. And will "accept as solution" if this works. 🙂