Occupancy models in MARK

[Thurston]

Hi, everyone.

I'm presently working on incorporating occupancy models into my research program and hope to use MARK for this purpose. I have a good understanding of the theory and practice of occupancy models, but have little experience actually running the analyses. I've just finished reading the "Gentle Introduction" and feel fairly comfortable with MARK in general, but not in how MARK handles occupancy models. Is there a resource (yet) for using MARK for this purpose, either coming from phidot or elsewhere?

Thanks much!

[cooch]

Hi, everyone.

I'm presently working on incorporating occupancy models into my research program and hope to use MARK for this purpose. I have a good understanding of the theory and practice of occupancy models, but have little experience actually running the analyses. I've just finished reading the "Gentle Introduction" and feel fairly comfortable with MARK in general, but not in how MARK handles occupancy models. Is there a resource (yet) for using MARK for this purpose, either coming from phidot or elsewhere?

Thanks much!

Try searching the forum first (as suggested in the MARK FAQ) - you would have discovered (for example)

http://www.phidot.org/forum/viewtopic.p ... ncy+models

Then, always check the MARK help file - if it hasn't been written up for 'the book' yet, (i) it will, eventually, and (ii) in the meantime, the help file is your best reference.

[Thurston]

Try searching the forum first (as suggested in the MARK FAQ) - you would have discovered (for example)

http://www.phidot.org/forum/viewtopic.p ... ncy+models

Then, always check the MARK help file - if it hasn't been written up for 'the book' yet, (i) it will, eventually, and (ii) in the meantime, the help file is your best reference.

Thanks for the message. Suffice it say, however, that I had already conducted the searches you suggest (I even RTFM and checked the help files ). I also own the new book on occupancy estimation. But none of these sources provide much specific information on actually running occupancy models in MARK. I'll now wait patiently and am looking forward to a chapter on occupancy estimation in the Gentle Introduction (which, incidentally, I've found very useful for a number of other topics).

Thanks to those of you who sent me helpful links via e-mail. They were very useful and will certainly help get me started.

Cheers.

[cooch]

Thanks for the message. Suffice it say, however, that I had already conducted the searches you suggest (I even RTFM and checked the help files ).

Actually, for simple occupancy models, you really don't need much more than the help file. With even some understanding of basic ideas in MARK (i.e., having worked through chapoter 3 -> 8, which I'll assume you've done if you RTFM), you should be able to figure out how to work occupancy models. Until you get to RD versions, occupancy models are pretty straightforward.

[darryl]

Thanks for the message. Suffice it say, however, that I had already conducted the searches you suggest (I even RTFM and checked the help files ).

Actually, for simple occupancy models, you really don't need much more than the help file. With even some understanding of basic ideas in MARK (i.e., having worked through chapoter 3 -> 8, which I'll assume you've done if you RTFM), you should be able to figure out how to work occupancy models. Until you get to RD versions, occupancy models are pretty straightforward.

Actually it's my fault there's no section on this in the MARK book yet, I got too busy writting that other book. The main trick you need to think about is how to deal with survey-specifc covariates (eg air temp) that you might want to use to model detection probabilities. If you have T survey occcasions, then you need to enter the air temperature (say) as T "individual" covariates. Then when setting up your design matrix you have the T covariates stacked on one another in a single column (ie the magnitude of the effect of air temp on p is the same with each survey). I suggest you standardise the covariates yourself before inputing the data to MARK, and make sure you turn off the 'standardise covariates' option when you do your analysis, particularly if you've got some missing observations in there.

Hope this helps
Darryl

[cooch]

...I suggest you standardise the covariates yourself before inputing the data to MARK, and make sure you turn off the 'standardise covariates' option when you do your analysis, particularly if you've got some missing observations in there.

The standardize covariates option in MARK does a standard z-transform on the data (i.e., mean 0, -3 <-> 3 range in data). Main reason to do this is it helps simplify scaling issues when interpreting results.

The other reason to standardize (in some fashion) - numerical estimation issue - has been dealt with in a recent tweak Gary made to MARK, such that the standardization is now done automatically. See

http://www.phidot.org/forum/viewtopic.php?t=384

[darryl]

...I suggest you standardise the covariates yourself before inputing the data to MARK, and make sure you turn off the 'standardise covariates' option when you do your analysis, particularly if you've got some missing observations in there.

The standardize covariates option in MARK does a standard z-transform on the data (i.e., mean 0, -3 <-> 3 range in data). Main reason to do this is it helps simplify scaling issues when interpreting results.

The other reason to standardize (in some fashion) - numerical estimation issue - has been dealt with in a recent tweak Gary made to MARK, such that the standardization is now done automatically. See

http://www.phidot.org/forum/viewtopic.php?t=384

Ok, my main reason for suggesting you standardise was for the numerical reasons, so perhaps that isn't so relevant with the new tweak. Although, and perhaps this is one for Gary, can the internal rescaling handle situations where you have multiple individual covariates associated with a single beta parameter?

[gwhite]

As Evan pointed out, the current version of MARK does internally standardize the covariates, in fact the entire design matrix. So, it is not absolutely necessary to standardize the individual covariates for numerical optimization reasons.

Darryl asks if multiple covariates within in a single column are handled? Yes -- because each column (i.e. all values in the design matrix column, not just the individual covariates independently) is scaled to a value in the interval -1 to 1 for the optimization process, and then the betas and their VC matrix back-scaled to the original values in the design matrix. So, multiple covariates in one column do work fine with this new scaling algorithm. The new scaling algorithm is transparent to the user, although the astute user may notice small differences (hopefully the 5 or more significant digit) in the beta estimates between old runs and new runs that
incorporate the scaling algorithm.

A reason to scale individual covariates to zero is to be able to include missing values. If you compute the mean of the non-missing values of an individual covariate, and then scale the non-missing values to have a mean of zero, the missing values can be included in the analysis as zero values, and will not affect the slope of the estimated beta. I wouldn't advise this trick for a covariate with a large percentage of missing values because you have no power, but this approach does work for a "small" number of missing values.

Gary