Reproduce SAS Genmod in Stata

[Jon Heintz]

I am trying to reproduce a GLM in Stata that I originally created in SAS. Although my estimates are exactly the same, the standard errors are not, and I'm not quite sure why. I tested all of the vce options in Stata, but the errors still do not match. Note that cost is continuous while treatment, var1, and var2 are all categorical. Here is both codes:

SAS

proc genmod data=data;
 class clinician treatment var1 var2 / param=ref;
 model cost = treatment var1 var2/ link=LOG dist=gamma;
repeated SUBJECT=clinician/ PRINTMLE TYPE=ind;
run;

Stata

glm cost i.treatment i.var1 i.var2, f(gam) l(log) cluster(clinician)

Appreciate the help.

[Hemanshu Kumar]

I can't really answer your question except to suggest perhaps looking at the code underlying the glm command in Stata:

viewsource glm.ado

[Joseph Coveney]

I am trying to reproduce a GLM in Stata that I originally created in SAS. Although my estimates are exactly the same, the standard errors are not, and I'm not quite sure why. . . .

proc genmod  . . .;
. . .;
repeated SUBJECT=. . .;

I take it that you've already looked at this xtgee-related FAQ on StataCorp's website, and perhaps this post on the old Statlist.

[Jon Heintz]

Joseph Coveney, thank you, and yes I looked at both of those. I understand the estimation algorithm is slightly different; however, the errors are different enough to change the significance and I'm not which to trust.

[Weiwen Ng]

It turns out that the intro to PROC GENMOD, which does fit GLMs, says this:

The GENMOD procedure can fit models to correlated responses by the GEE method. You can use PROC GENMOD to fit models with most of the correlation structures from Liang and Zeger (1986) by using GEEs. For more details on GEEs, see Hardin and Hilbe (2003); Diggle, Liang, and Zeger (1994); Lipsitz et al. (1994).

and the REPEATED statement does seem to cause it to fit a GEE. Yours specifies type = independent, I.e. independence working correlation.

[Jon Heintz]

Figured it out. The answer is twofold.

1) As Joseph Coveney said, Genmod and Stata estimate the scale parameter and correlation structure slightly differently. Details are given on xtgee-related FAQ.

2) Genmod by default produces score tests while Stata produces Wald tests. Wald tests can be requested for SAS, but score tests cannot be produced by Stata.