Problem calculating relative risk for multi-categorical exposure

[Vivio]

Hi, I am trying to calculate the relative risk of SMM per province and obtain different RR estimates for each province with ON as the reference.

My data looks something like this:

I have tried different methods but none is working. Please help.

Method 1 using binomial does not give me separate RRs. I also saw somewhere about removing the "param" and estimate line but this did not work either, neither does using NLMeans macro.

proc genmod data= smm.uvu descending;
class province1(ref='ON')/ param=ref;
model SMM = province1/d=bin link=log;
estimate 'SMM RR by province' province1 1/exp;
lsmeans province1 / diff exp cl;
title "RR SMM type by province";
run;

Method 2. The proc freq method below is only giving me 2*2 freq without any RR:

proc freq data=uvu2 order=data;
tables province1*SMM/relrisk;
run;

[mkeintz]

It looks like proc freq only does relative risk for 2x2 tables.  So you could create a dataset with a sequence of province pairs,  i.e. all the 'AB' and 'ON' observations (pair_num=1), then all the 'BC' and 'ON' observations (pair_num=2).  Then you could do a sequence of 2x2 tables  using PROC FREQ -- BY PAIR_NUM, as here:

data ON;
  set have (where=(province='ON'));
run;

data test ;
  set have (where=(province^='ON'));
  by province ;
  retain pairnum 0   pairing 'XX vs YY';
  if first.province then do;
    pairnum+1;
    pairing=catx(' ',province,'vs','ON');
  end;
  output;

  if last.province then do ptr=1 to n_on;
    set ON nobs=n_on point=ptr;
    output;
  end;
run;

proc freq data=test;
  by pairnum pairing;
  tables province*smm / nopercent nocol ;
  exact relrisk (column=2 method=noscore);
run;

I used the "method=noscore" option because it is much faster than the default (method=score).  Take a look at the RELRISK <(options)>  section of the EXACT statement documentation.  If you have a lot of pairings to examine, you might want to test this on just a couple to get an idea of how long it will take to run.

And on another note, I don't know how much "confidence" one can have in a set of relrisk confidence limits when each set re-use the same province=ON data.  That's an estimation problem that I never had to face.