Nest Survival - Different phenologies

[Cristiana]

Hi all!
I’m working on the survival from egg to first instar of Colias palaeno (a wonderful butterfly!).
I’ve followed more than 300 eggs during 3-7 days interval until their death or successful hatching for a temporal interval of 83 days.
I would analyse my data with the Nest Survival model in Rmark, but I have one problem.
My data come from different study sites located at different altitude, in which consequently developmental time and phenology are not the same.
To be clearer: in site A I have eggs from day 1 to 53 (than all eggs are death or already first instar); in site B I have eggs from day 7 to 83 (developmental time is longer and oviposition is less synchronised).
I would compare survival rate among sites, inserted as a group in the input file, and using other individual covariates.
When I test for a time trend, Rmark gave me daily survival rate for site A even for the days between 54 and 83, that shouldn’t be realistic because I have no more specimens in this period.
How can I resolve this problem? The only way is to analyse the data for each site separately? But I would like to analyse them all together, to examine the ‘site effect’.
Should I set to 1 daily survival rate for site A for the days between 54 and 83?
Any suggestion is really welcome!
Thanks a lot and sorry for the silly question!
Cristiana

[jlaake]

Cristiana-

Hopefully those familiar with nest survival models will offer their ideas. You didn't tell us much about what model you fit to the data. For example, if you fit ~NestAge then it will fit a smooth trend in survival over age that is constant across sites. If you fit ~site*NestAge then it would fit a decline in age that was specific to the site. However, regardless of the model MARK will produce estimates for each sampling occasion with a model like those because it is a smoothed estimate. I think your intuition is for a model in which survival was estimated for each time/age interval - like a t (time as factor) versus T (time as linear trend) model. But even if you fit it as a factor variable, you would get estimates for those times if you assumed it was constant over site or additive over site. It sounds like you are likely to have an interaction of time/age with site, so you may not gain much with a pooled analysis but by analyzing the data together it lets you do the model selection to show what is the most appropriate model. I would NOT do separate analyses by site.

Now it may be that something like site*NestAge may be the most appropriate model and it may produce estimates for times in which there were no surviving critters. That doesn't mean that there is something wrong with the model necessarily. In particular, if the cumulative survival to that time was essentially 0, then those estimates beyond that time are not meaningful and have little impact on the model fit. By analogy, think about fitting a linear model for some portion of a growth curve. While it may be linear over the range you measured, it might make nonsensical predictions outside of the range of data. Doesn't mean that it isn't a good model to describe the data you have.

Hope this helps. --jeff

[Cristiana]

Yes, thanks.
It helped me.
I was testing a model with Time*Area; I followed your suggestions and now results are clear to me.
It is just necessary not to graph or to consider daily survival estimates for days in which I had no more samples.
I've tried deleting them from the design data list and the results are still the same!

Thanks a lot,
Cristiana