How can I write a macro to execute iterative regressions?

[jmdecu]

Hi,

I have a group of 10 exposure variables and a group of 10 outcome variables.  I would like to write a macro to regress each outcome variable on each exposure variable (a total of 100 regressions) using GEE with logistic regression, but I am unsure of how to code this.  Any help would be much appreciated!

Thanks!

[Reeza]

See about halfway down the page here, though the whole page is useful if you're just starting out with macros.

https://www.ats.ucla.edu/stat/sas/seminars/sas_macros_introduction/default.htm

You can search for:

A macro program for repeating a procedure multiple times

[jmdecu]

Thanks so much for this link!  After reading the page, I was able to create a macro to regress several outcomes on a single predictor, but I would like to regress all of the outcomes on all of the predictors as well.  Are there any more resources you could suggest? 

[Linlin]

Hi,

post your basic code here, we will help you to create a macro.

[jmdecu]

Here is the code I am working with so far:

proc genmod data=xxx descending;

    class exposure pharmid studyarm (ref='0') / param=ref;

    model outcome = exposure studyarm / dist=bin link=logit type3 wald;

    repeated subject = pharmid / type=cs;

run;

I would like to create a macro that will regress each of 10 outcome variables on each of 10 exposure variables using this code.  As I mentioned above, I was able to write a macro that would regress each outcome on a given group of exposure variables, but I haven't figured out how to loop the macro through all of the exposure variables individually.  Here is what I have so far:

%macro preliminary_gee(all_deps, indvars =);

  %let k=1;

  %let dep = %scan(&all_deps, &k);

  %do %while(&dep NE);

    proc genmod data=xxx descending;

    class &indvars pharmid studyarm;

    model &dep = &indvars studyarm / dist=bin link=logit type3 wald;

    repeated subject = pharmid / type=cs;

    run;

    %let k = %eval(&k + 1);

    %let dep = %scan(&all_deps, &k);

  %end;

%mend preliminary_gee;

In addition, it would be great to suppress the output and to just create a dataset containing the OR estimates from each regression.  Again, any ideas you have would be much appreciated.

Thanks!

[Reeza]

Nice! But why loop over only the dependents, why not add another do loop and nest over your independents as well?

If you got this far, I think you can get it, but I'm sure ppl here will help regardless. 

[data_null__]

I don't think JMDECU should do it that way at all.  I would "transform" the data such that the independents and dependents are levels of BY variables and use ONE call to PROC GENMOD.

I don't know exactly which output you want from GENMOD but you can figure that out.

data sample;
   array y[3];
   array x[3];
   do pharmid = 1 to 10;
      studyarm = rantbl(12345,.5);
      do _n_ = 1 to dim(y);
         y[_n_] = rantbl(0,.5);
         x[_n_] = rannor(0);
         end;
      output;
      end;
   run;
proc print;
   run;
proc transpose data=sample out=sample2(rename=(col1=Y)) name=dep;
   by pharmid studyarm X:;
   var y:;
   run;
proc transpose data=sample2 out=sample3(rename=(col1=X)) name=ind;
   by pharmid studyarm dep y;
   run;
proc sort data=sample3;
   by dep ind;
   run;
proc print;
   run;
ods select none;
ods output GEEEmpPEst=GEEEmpPEst;
proc genmod data=sample3 descending;
   by dep ind;
   class pharmid studyarm;
   model y = x studyarm / dist=bin link=logit type3 wald;
   repeated subject = pharmid / type=cs;
   run;
ods select all;
proc print data=GEEEmpPEst;
   run;

Message was edited by: data _null_

[Linlin · Accepted solution]

Hi,

first put your variables in a dataset.  example:

data have;
input (var1 var2) ($1);
cards;
a b
c d
e f
g h
i j
k l
m n
o p
q r
s t
;

data _null_;
set have;
call symputx(cats('dep',_n_),var1);
call symputx(cats('indvars',_n_),var2);
run;

%macro preliminary_gee();
%do i=1 %to 10;
    %do j=1 %to 10;
    proc genmod data=xxx descending;
    class &&indvars&j pharmid studyarm;
    model &&dep&i = &&indvars&j studyarm / dist=bin link=logit type3 wald;
    repeated subject = pharmid / type=cs;
    run;
   %end; %end;
%mend preliminary_gee;

%preliminary_gee()

[jmdecu]

Hi Linlin,

Thanks so much for sending me this code.  It worked!  Just one last question - is it possible to create a single output dataset containing the odds ratio estimates generated by each regression?  I know it is possible to do this when using proc logistic, but I am not sure about proc genmod.

Thanks again!  

[Howles]

Two ways.

1. If you persist with the macro loop approach, include a step before the loop to initialize a zero-row target data set. After the PROC GENMOD step but inside the inner loop, include a PROC APPEND.

2. If you adopt data_null_'s suggestion to ditch the macro looping and employ BY processing instead, you should see the stats automatically accumulate in one data set.

jmdecu wrote:

Hi Linlin,
Thanks so much for sending me this code.  It worked!  Just one last question - is it possible to create a single output dataset containing the odds ratio estimates generated by each regression?  I know it is possible to do this when using proc logistic, but I am not sure about proc genmod.

Thanks again!