fixing parameter values for individual covariates

[benaug]

Hopefully this is an easy question. Is there a way in Rmark to fix the parameter values of individual covariates? As far as I can tell, I can only fix elements of the m-arrays.

Thanks!

[jlaake]

You can only fix real parameters and not betas if that is what you are asking. If you are asking whether you can fix a real parameter at a specific value of the covariate, I can't think of any way to do it unless you split the histories into groups based on the value of the covariate and fixed the real parameters for groups with specific covariate values. But if this is an individual numeric covariate, there would be no way to guarantee continuity. But I may be misunderstanding what you are asking.

[benaug]

Thanks, Jeff. You're probably wondering "why would you ever want to do such a thing?". I'm working with someone who has mark recapture data where some individuals were radio tagged during some occasions, but not all occasions for any one individual. These individuals have p=1 on those occasions and we want do fit the CJS model. I could describe this as "time-varying groups" or maybe a "partially known fate" analysis. I'm sure I could code this myself, but I was looking for a way to trick MARK into doing it. All my ideas involved fixing individual covariates since I can't do time-varying groups in MARK. Any ideas?

[jlaake]

Yes I was wondering and that explanation helped. Depending on how many of these individuals you have and how much they vary across time, you could create groups where each individual in the group was radio-tagged on the same set of occasions. Then in the design data for p for the times for each group you can set fix=1 for those times and all other values to NA. Alternatively you can create a 0/1 individual time-varying covariate that is only one when the critter has a collar. Then include that variable in your model and the beta will "converge" to some large positive value such that p=1. Then adjust your parameter counts by -1. Note that a time-varying individual covariate is in your data set and the names of them are basettt where base is the name of the variable you'll use in the formula and ttt is the values of time at each occasion. See documentation.

As a completely different alternative you might try my marked package on CRAN. This is one of the places where its flexibility is useful. It creates a design dataframe with a row for each occasion for each animal. Then you can set the variable fix in the p design dataframe to 1 for just those occasions per animal where it is radiocollared and NA elsewhere. If you go down this route, use the version of marked on my github site and google drive. I recently made some updates in regard to fixing parameters and have not yet posted it to CRAN.

--jeff

[benaug]

Thanks, Jeff. I went with individual time-varying covariates in MARK. It doesn't count that parameter anyway since it is on a boundary. Doing this gives unbiased estimates of phi with nominal coverage in the scenarios I considered. The only problem that would require further attention is model averaging capture probability, but this is a nuisance parameter.

[jlaake]

Note the default argument is adjust=TRUE so RMark assumes the DM is full rank and will adjust the parameter count from what MARK provides. So in this case set adjust=FALSE so that parameter is not counted. Just remember that MARK does that for any parameter at a boundary (particularly for logit link) and some are truly estimated and should be counted.

--jeff

[benaug]

Right. Any ideas on how to estimate c-hat with the individual radio covariates?

[cooch]

Right. Any ideas on how to estimate c-hat with the individual radio covariates?

There are no GOF tests for models (i) with individual covariates, or (ii) known-fate models.

[looking4lions]

Hi benaug,

I am currently having the same issue. I have data on both collared and non-collared lions and wanted to adopt the same approach using time-varying individual covariates. Have you produced a paper that I can reference for this approach, or know of a paper that has adopted this approach?

Thank you!