LD model with sampling effort varying by individual

[rasrage]

Dear all,
I am attempting to run a live resight/dead recovery model through RMark. I am very new to RMark and maybe my question is trivial, but searching through help pages and pdfs hasn't given me the answer so far.
My data com from monthly telemetry surveys. Most individuals were searched for every month of the study, unless they were eventually found dead. But for some individuals, sampling effort stopped for some reason. So, for example, one individual was located in three consecutive months, then it wasn't seen in month 4, and then no-one ever looked for it again during the rest of the study. Other individuals were looked for, though, so that effort is essentially an individual and time varying categorical covariate, where detection and recovery probability for the second category (when an individual was no longer looked for) have to be fixed to 0.
I managed including effort as a numerical individual and time varying covariate (just to see if I understood the basic mechanism of including such a covariate), but I was unable to include it as a categorical covariate. I tried using a 0/1 dummy variable, making "not sampled" the intercept and fixing that to 0, but that didn't work out either.
My question: is this dummy variable/fixing the intercept to 0 the correct approach and am I just not understanding the process of how to fix the right parameter to 0? Or is there a different approach to doing that?

Here is how I attempted to do the dummy variable/fixing the intercept to 0 approach:
Effort is a matrix, nrow = #individuals, ncol = #primary occasions; entries are 0 if an individual was not looked for on a given occasion and 1 if it was looked for. Column names are "Effort1" through "Effort37".
I added that to the data frame holding the encounter histories using cbind().
I tried defining p as
p.dot=list(formula=~Effort, fixed=list(index=1, value=0))
Thinking that in this model, p would have 2 parameters and the first one would be the intercept. But the model returns nonsensical estimates for all detection parameters.

Thank you in advance for your help and insight,

Rahel

[jlaake]

Rahel-

So for the concept of time-varying individual covariate you can only estimate the beta parameter for the covariate. There is no way to set a beta value in MARK. You can only fix reals. Presumably if you set it up as a 0/1 variable (sampled/not sampled) and let it estimate the parameter it would be a large negative value for the logit link because none will have been seen after that. Note that 1 should be not-sampled. If you make 1, sampled, then your intercept will be a large negative number and everything will look goofy because it is relative of a what is in essence an estimated boundary intercept.

Now, for a couple of other thoughts. Check to see in MARK documentation if LD format allows a negative frequency which is a loss on capture. If it does that would be the easiest thing to do. If it doesn't then another alternative is to put the animals in groups (essentially MARK's categorical covariate). The group should contain individuals in which searching stopped all on the same occasion. Then you can set the real parameter p=0 for that group for the occasions that were not searched. With either of these approaches you don't have the extra beta parameter estimated.

In my marked package I allow different fixed values for each animal-occasion which would enable this to be handled directly. However, I haven't included dead recoveries in any of the models so it is not an option for you at this point. One of the ideas above should work however.

regards --jeff

[rasrage]

Thank you Jeff! I'll explore those options you mentioned.
Best
Rahel

[rasrage]

Double thanks, Jeff, as I *just* found a post where you pretty much already answered that question.