That seems very odd. Have you submitted this to SAS tech support?
You could try the glimmix procedure for fitting a GEE using ML estimators instead of moment estimators of the working correlation structure. Glimmix code that fits the same model as your genmod code would be:
proc glimmix data=crash empirical=hc0;
class ID year;
model accident = major minor / dist=negbin link=log;
random _residual_ / subject=ID;
run;
The glimmix procedure will produce type3 Wald tests by default, so you do not need to specify any additional options to obtain these statistics. The empirical=hc0 option requests traditional sandwich standard errors. Note that the glimmix procedure offers other methods for computing robust standard errors - improvements on the traditional sandwich estimates. You might want to use empirical=hc3. See the following manuscripts regarding sandwich estimates:
http://weber.ucsd.edu/~mbacci/white/pub_files/hwcv-019.pdf
http://www.jstor.org/stable/pdfplus/2685594.pdf
SAS documentation discussing use of the glimmix procedure for producing a GEE-type analysis can be found at:
http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_glimmix_sect...
HTH