This question is more general than RMark-specific, but I'll leave it here for now.
TheCarmeleon wrote:Hi everyone, I'm currently running some CJS models in RMark using capture-mark-recapture data, over a time period of 6 years. There would be more but the pandemic has put a stop to things over the past couple of years.
I anticipate this will be a common issue moving forward. I'm personally dealing with lack/cessation of data collection for several very long-term data sets (>40 years minimum). If/when things start up again, there will (I suspect) be a bunch of questions of 'what to do if you have large-ish' gaps in the time series. I digress...
I've run GOF tests with the data, and everything passed with flying colours (which was a relief!). Despite this, I'm slightly concerned as the top three models (based on the AIC table) are telling me that the survival estimates (Phi) are 1. When using less data (during the time when I was first learning how to use RMARK over the summer), I had some more sensible outputs, that made sense knowing what I do about our study species.
What would you suggest I can do to help troubleshoot this problem? It is a problem?
I'd start by seeing if the parameter(s) that are being estimated close to the boundary (in your - typical - case, 1.0) are actually 1.0, or if the numerical optimization simply doesn't have enough 'information' (data) to be able to differentiate one estimate (say, 0.985) from another (say, 1.0). MARK has the ability to differentiate between 'intrinsically' or 'extrinsically' nonidentiable parameters, using 'data cloning' -- Appendix F. [You can't do data cloning automatically using RMark -- it is a function (one of several) that is built in to the classic MARK GUI. If you read Appendix F, you could probably figure out how to tackle it manually -- starting by creating a copy of the .inp file where the number of encounters is multiplied by some factor -- say 100 -- but that might be more work than you want.]
Then, if you end up concluding that - yup - some parameters really are 1.0, then there is an argument you could make that these are not estimated parameters, and you could manually reduce the parameter count(s) accordingly, which will in turn adjust the AIC.
So, thats a starting point.
Having said that, 9/10 times when you have parameters estimated at ~1.0 (99% of the time for a time-dependent model for some/all of the model parameters), the root cause is 'extrinsic nonidentifiability' (meaning, insufficient data).
Good luck...