topic Re: Robust Sandwich Estimate from PROC Logistic in Statistical Procedures

Robust Sandwich Estimate from PROC Logistic

jdserbon — Tue, 26 Feb 2013 13:38:32 GMT

I have the following pseduo-code in which I need to calculate robust sandwich standard error estimates. I do not believe you can perform this using PROC LOGISTIC, correct? If not, which PROC would you recommend? Which options to specify to substitute for the strata statement? Thanks!

proc logistic data = data;

class x5 /param = ref ref = first;

model y (event = '1')= x1 x2 x3 x4 x5 x6;

strata x7;

where (x9 = 1 or x10 = 1) & x5 ne '' & x7 ne '' & y_denom = 1 &

x8 in ('name1','name2','name3','name4','name5');

by x8;

run;

Re: Robust Sandwich Estimate from PROC Logistic

SteveDenham — Wed, 27 Feb 2013 12:32:09 GMT

Try this:

proc glimmix data=data empirical=classical; /* For a classical sandwich estimation */

class x5(param=ref ref=first) x7;

model y(event='1')=x1 x2 x3 x4 x5 x6/dist=binomial solution;

random intercept/subject=x7; /* To mimic the strata */

where (x9 = 1 or x10 = 1) & x5 ne '' & x7 ne '' & y_denom = 1 &

x8 in ('name1','name2','name3','name4','name5');

by x8;

run;

The code is untested, so there will probably be some bugs along the way, but GLIMMIX is the only way I can think of to get sandwich estimators for generalized linear models.

Steve Denham