Logistic Regression using Proc GLIMMIX with random effects

nsns — Tue, 18 Jun 2019 15:19:34 GMT

I am trying to combine response data (i.e. proportions) from multiple studies where each study has only one arm of interest i.e. arm A or arm B. Because in each study I don’t have both arms I cannot perform a classic meta-analysis. The data are percentage of responders. I would like to estimate the percent responders for each arm, the difference and the relative risk between the groups with 95% confidence intervals for each. I thought of using a logistic regression with a random effect (i.e. the individual studies) using Proc GLIMMIX. The confidence intervals are important because I will use these to define a non-inferiority margin. Questions are 1) how can I get confidence intervals for the difference between the arms and 2) how do I get the relative risk?

Sample data and code looks something like this:

data a;

input studyno arm $ n_total n_responders;

cards;

1 B 107 42

2 B 73 41

3 B 75 41

4 B 77 49

5 B 199 123

6 B 201 122

7 B 221 123

8 A 16 4

9 A 47 16

10 A 767 107

11 A 170 20

12 A 51 13

13 A 128 20

14 A 19 5

15 A 14 1

16 A 47 13

17 A 118 23

18 A 58 11

;

run;

proc glimmix data=a;

class arm studyno;

model N_responders/n_total = arm /solution cl dist=binomial link=logit;

random intercept/subject=studyno;

lsmeans arm / cl ilink;

estimate 'A vs B' arm 1 -1 / cl ilink exp;

run;

Thanks in advance for any suggestions.

Re: Logistic Regression using Proc GLIMMIX with random effects

StatDave — Mon, 24 Jun 2019 15:25:13 GMT

Assuming that the studies are independent, then you can just fit an ordinary logistic model and use the NLMEANS macro to estimate the relative risk (or difference in proportions if desired) and its standard error as shown in this note.

topic Logistic Regression using Proc GLIMMIX with random effects in Statistical Procedures

Logistic Regression using Proc GLIMMIX with random effects

Re: Logistic Regression using Proc GLIMMIX with random effects