Alternate link functions?

[Tim]

I'm analyzing telemetry data using the known fate framework. I've got 4 encounter periods and 25 groups representing combinations of 5 stages and 5 years of study. Survival is high to begin with, thus its not surprising that MARK is having trouble estimating many of the group-specific S values in the full group*encounter period and group (stage*year) models probably because these parameters are close to the upper boundary of the logit link function. Simpler models work fine.

My question: What's the advice for trying alternate link functions? Is it simply a trial and error affair or are some link functions known to be better alternatives to the logit when survival is high?

Tim

[cooch]

I'm analyzing telemetry data using the known fate framework. I've got 4 encounter periods and 25 groups representing combinations of 5 stages and 5 years of study. Survival is high to begin with, thus its not surprising that MARK is having trouble estimating many of the group-specific S values in the full group*encounter period and group (stage*year) models probably because these parameters are close to the upper boundary of the logit link function. Simpler models work fine.

My question: What's the advice for trying alternate link functions? Is it simply a trial and error affair or are some link functions known to be better alternatives to the logit when survival is high?

Tim

As discussed in several places in 'the book', the sin function has better properties than the logit for boundary parameters. You might also need to consider (i) structure of the DM, and (ii) optimization method as part of the overall strategy to maximize identifiability, but start with looking at the various discussions of sin versus logit link.