problems with log(-log) link

[Karim]

I am trying to fit models for mortality hazards and am thus drawn to the log(-log) link. But when I specify this link (actually I do parm-specific and use log(-log) for survival and strata transition but logit for resighting) MARK returns a run-time error M6202: MATH -log:SING error. When I hit okay it appends the model results and appears to have fit the model, calculated the deviance and number of parameters, etc. but the beta values are so wacky that it cannot reconstitute the real parameter estimates. The deviance (and thus AICc) is often very close to the the deviance of the equivalent model fit with a logit link, which appears to work fine. Survival estimated with the logit link is quite high (>0.97) so am not sure whether being so close to the boundary is causing the numerical problems. Does anybody have any thoughts on this? Is there something obvious that I'm missing? Any suggestions for things to change to get the log(-log) link to work? For me, a linear age trend or additive effect of strata on the hazard really is more meaningful than a similar effect on logit(survival). Thanks