Modeling averaging with known fate models and no mortality

[jbauder]

I have a telemetry data set for rattlesnakes that I using to obtain annual survival estimates for use in a population viability analysis. I am using known fate models and broke my data into five monthly intervals (summer tracking) and one "winter" interval consisting of the seven months the snakes were underground and not monitored. I am trying to obtain "summer" survival estimates across the five months of active tracking and "winter" survival estimates across the one winter interval. However, some of my summer and winter intervals have no mortality events so survival is 1.0. I would like to use model averaging to obtain model averaged summer and winter survival estimates.
Can I do model averaging when some of the values used in the averaging process have estimates of 1.0 and standard errors of 0.0?

[bacollier]

I have a telemetry data set for rattlesnakes that I using to obtain annual survival estimates for use in a population viability analysis. I am using known fate models and broke my data into five monthly intervals (summer tracking) and one "winter" interval consisting of the seven months the snakes were underground and not monitored. I am trying to obtain "summer" survival estimates across the five months of active tracking and "winter" survival estimates across the one winter interval. However, some of my summer and winter intervals have no mortality events so survival is 1.0. I would like to use model averaging to obtain model averaged summer and winter survival estimates.
Can I do model averaging when some of the values used in the averaging process have estimates of 1.0 and standard errors of 0.0?

jbauder,
A pretty central tenet to survival analysis is that you should have at least 1 'event' within each sampling occasion, that way you don't get survival estimates of 1. Understanding your situation having tinkered with some rattlesnake known fate data before, can you adjust your frequency of encounter histories such that you have events in each occasion. If not, you might try and use profile CI to get lower bounds on your survival estimates for those periods where .

As for your model averaging question, I am not sure, I would think that having a zero SE would cause that 0 to be averaged into the unconditional SE, but I am not sure as I have never had to deal with it.

Bret

[jbauder]

Thanks for your reply Bret.
I re-ran my models with the profile likelihood confidence intervals on the interval that doesn't have any events (i.e., no mortalities). The estimate is still 1.0 and the standard error is 0.1207224E-018 but the profile likelihood confidence interval is 0.62-1.00 so it seems like that solution worked.
My data set consists of two summers and one winter of data. We tracked six snakes during the first summer and through the winter and we confirmed they all survived the winter. We removed those six snakes from the study the second spring and added 16 new snakes to study and tracked them to the end of the summer. Only one snake died during the second summer (at the very last month, September, of tracking) so we only have one event during the study and no events during the winter. So I don't think I can adjust my encounter history frequencies to have an event in each sampling period.
I am only running two models in this analysis: a S(.) model to obtain a single annual survival estimate and a S(season) model to estimate winter and summer survival. The winter survival estimate is the one that is estimated at 1.00.

[bacollier]

Thanks for your reply Bret.
I re-ran my models with the profile likelihood confidence intervals on the interval that doesn't have any events (i.e., no mortalities). The estimate is still 1.0 and the standard error is 0.1207224E-018 but the profile likelihood confidence interval is 0.62-1.00 so it seems like that solution worked.
My data set consists of two summers and one winter of data. We tracked six snakes during the first summer and through the winter and we confirmed they all survived the winter. We removed those six snakes from the study the second spring and added 16 new snakes to study and tracked them to the end of the summer. Only one snake died during the second summer (at the very last month, September, of tracking) so we only have one event during the study and no events during the winter. So I don't think I can adjust my encounter history frequencies to have an event in each sampling period.
I am only running two models in this analysis: a S(.) model to obtain a single annual survival estimate and a S(season) model to estimate winter and summer survival. The winter survival estimate is the one that is estimated at 1.00.

Ok, well, I think you may be using a sledgehammer (MARK) with only 1 event over 2 seasons, estimation of period survival is really not that useful and the profile-likelihood CI's are probably not that reasonable either. You should probably just drop the MARK analysis/model averaging, etc. and provide a simple percentage of snakes that survived the tracking season, it will save you headaches later.

Bret

[dhewitt]

If you run model-averaging in RMark it can be slightly different from MARK. We get boundary estimates sometimes and when MARK does model-averaging somehow Gary avoids the zero SE problem that Bret mentions. Not sure how, and also not sure whether Jeff has changed RMark to make it match MARK. But, beware.