In <53rcep$4nj@GSB-Kwanza.Stanford.EDU>, doncram@leland.Stanford.EDU (Donald Peter Cram) writes: >Is it possible to test equality of coefficients, and perform other tests, >in PROC LOGISTIC like one can in PROC REG? How can I do hypothesis >testing in LOGISTIC? > >Below is a sample program that doesn't work. > > >data sodas; >title "Analysis of Soda Preferences"; >input soda $ cola sweet sub1-sub5; >cards; >coke 1 0 2 5 1 3 4 >pepsi 1 1 1 6 2 4 3 >tab 1 0 6 1 3 6 2 >dietpepsi 1 0 5 2 4 5 1 >sevenup 0 1 3 3 5 2 5 >sprite1 0 1 4 4 6 1 6 >sprite2 0 0 4 4 6 1 6 >sprite3 0 0 4 4 6 1 6 >; > > >proc reg data=sodas; > model cola = sweet sub1 sub2; > test sub1 = sub2; > test sub1 = 0; > >proc logistic descending data=sodas; > model cola = sweet sub1 sub2; > test sub1 = sub2; > test sub1 = 0; > >The PROC REG works fine but the test statement in PROC LOGISTIC >yields: >ERROR 180-322: Statement is not valid or it is used out of proper order. >

Don,

Instead of using PROC LOGISTIC, try PROC GENMOD. GENMOD has a contrast statement which can be employed to provide tests for linear combinations of the predictor variables. In order to fit a logistic regression in GENMOD, you must specify the model in events/trials syntax. In your dataset, you have only the events part (since you have a binary response instead of a binomial response). Thus, you must first create the trials variable. In this example, trials would be a constant=1. The syntax for GENMOD would then be:

proc genmod data=sodas; model cola/denom = sweet sub1 sub2 / error=binomial; contrast "Sub1=?Sub2" sub1 1 sub2 -1 / wald; contrast "Sub1=?0" sub1 1 / wald; run;

If you do not specify the wald option, then GENMOD will compute a likelihood ratio test under the contrast constraint. By specifying WALD, your second test is identical to the p-value which is returned in the table of parameter estimates.

(Note that for the data you presented, GENMOD and LOGISTIC return different tables of parameter estimates because there is "quasi-complete separation of data points". I presume that you have more data that you would actually employ in your regression.)

Dale McLerran Fred Hutchinson Cancer Research Center Ph: (206) 667-2926 1124 Columbia Street Fax: (206) 667-5977 Seattle, WA 98104