topic Re: Questions about interaction term in SAS Procedures

Questions about interaction term

ammarhm — Wed, 04 Mar 2015 02:36:32 GMT

I know this forum is mainly for help with the SAS code itself and I have received some fantastic answers here, I would really appreciate if anyone could please help me with this.

I am doing a regression analysis, looking at the effect of 2 drugs (A or B) in a population where we stratified individuals by ethnic group:

Estimates for European Population

proc logistic data=Studydatabase;

class sex drug study;

model Efficacy (event='Good')= drug age sex weight egfr;

where population='European';

run;

Effect	OddsRatioEst	LowerCL	UpperCL	ProbChiSq
Age	1.005	0.973	1.037	0.7729
Drug A vs B	1.862	1.229	2.821	0.0033
Sex Female vs Male	1.469	0.925	2.334	0.1033
Weight	1.015	1.001	1.028	0.0346
eGFR	1.006	0.995	1.016	0.2967

________________________________________

Estimates for Asian Population

proc logistic data=Studydatabase;

class sex drug study;

model Efficacy (event='Good')= drug age sex weight egfr;

where population='Asian';

run;

Obs	Effect	OddsRatioEst	LowerCL	UpperCL	ProbChiSq
1	Age	1.047	0.994	1.103	0.0839
2	Drug A vs B	0.719	0.217	2.382	0.5892
3	Sex Female vs Male	1.932	0.445	8.384	0.3795
4	Weight	0.987	0.954	1.020	0.4357
5	eGFR	1.014	0.991	1.036	0.2300

As you can notice, the influence of drug is significant in the European, but not in the Asian population. One could interpret this as the effect of the drug being modified according to the study subpopulation, and hence expect the p value for interaction to be significant, but it is not at all!

proc logistic data=Studydatabase;

class sex drug study;

model effectnum (event='Good')= study drug age sex weight egfr population*drug;

run;

Interactin term P value

ProbChiSq

0.1874

I know this forum is mainly for questions regarding the SAS code itself, but I would appreciate any help from all the SAS and statistics experts here

Best wishes

Any explanation please? Am I missing something

Re: Questions about interaction term

SteveDenham — Wed, 04 Mar 2015 19:55:29 GMT

Looking at the overlap in the confidence bounds on the two odds ratios, I am not surprised that the test for the point interaction is not significant.

European (1.229, 2.821)

Asian (0.219, 2.382)

Note also, that there is a significant weight effect in the European dataset that is not found in the Asian dataset. This may be a confounded variable. What happens if you add population*weight to the model?

Steve Denham

Re: Questions about interaction term

ammarhm — Wed, 04 Mar 2015 22:48:08 GMT

Dear Steve

Thank you very much, I have also noticed the overlap in the confidence intervals.

Could you please how to add the *weight term in the code? I mean if I do this I would get an error as I don't have a variable (weight):

proc logistic data=Studydatabase;

class sex drug study;

model effectnum (event='Good')= study drug age sex weight egfr population*weight;

run;

Thanks again

Re: Questions about interaction term

ammarhm — Wed, 04 Mar 2015 23:01:37 GMT

Dear Steve

Thank you very much, I have also noticed the overlap in the confidence intervals.

Could you please how to add the *weight term in the code? I mean if I do this I would get an error as I don't have a variable (weight):

proc logistic data=Studydatabase;

class sex drug study;

model effectnum (event='Good')= study drug age sex weight egfr population*weight;

run;

Thanks again

Re: Questions about interaction term

SteveDenham — Thu, 05 Mar 2015 20:25:12 GMT

You had the following code:

proc logistic data=Studydatabase;

class sex drug study;

model effectnum (event='Good')= study drug age sex weight egfr population*drug;

run;

Now one thing I would do before continuing is to include 'population' in the class statement, which would give:

proc logistic data=Studydatabase;

class sex drug study population;

model effectnum (event='Good')= study drug age sex weight egfr population*drug;

run;

Now in both of these model statements, there is a variable called 'weight', right there before egfr. It is not in the class statement, so I assume that it is a continuous covariate, and the Odds Ratio estimate of 0.987 says that the odds of a good event decrease slightly for each one unit increase in weight.

I suggested adding the term weight*population to the model, as I believe the event occurrence relationship to weight may differ in the two populations, and this needs to be removed as a possible confounder of the drug effect. This would result in the following code:

proc logistic data=Studydatabase;

class sex drug study population;

model effectnum (event='Good')= study drug age sex weight egfr population*weight population*drug;

run;

Steve Denham