Model Selection in CAPTURE

[jaxzwolf]

I am using CAPTURE to obtain population estimates for a mark-recapture study I completed this summer. Data sets for two of my study sites are sparse, with very few captures and no recaptures. The problem I'm encountering is that CAPTURE is unable to select an appropriate model for data sets in which no recaptures have occurred. It assumes the best population estimator for those data is M(tbh), an analysis which CAPTURE is unable to run. I have read that MARK can run an M(tbh) model, but I have no familiarity whatsoever with the program.

I have been able to obtain estimates running some models individually, including jackknife, zippen, removal, and mt-chao. Estimates obtained using jackknife seem unrealistically high. Is there a way for CAPTURE to select an appropriate population estimator with no recaptures? If not, how do I know which population estimate is the most realistic? In other words, if CAPTURE cannot select a model for me, how do I choose the most appropriate estimator?

[cooch]

I am using CAPTURE to obtain population estimates for a mark-recapture study I completed this summer. Data sets for two of my study sites are sparse, with very few captures and no recaptures. The problem I'm encountering is that CAPTURE is unable to select an appropriate model for data sets in which no recaptures have occurred. It assumes the best population estimator for those data is M(tbh), an analysis which CAPTURE is unable to run. I have read that MARK can run an M(tbh) model, but I have no familiarity whatsoever with the program.

I have been able to obtain estimates running some models individually, including jackknife, zippen, removal, and mt-chao. Estimates obtained using jackknife seem unrealistically high. Is there a way for CAPTURE to select an appropriate population estimator with no recaptures? If not, how do I know which population estimate is the most realistic? In other words, if CAPTURE cannot select a model for me, how do I choose the most appropriate estimator?

Well, the answer depends on several things, the most important of which is how long you want to spend on the problem. If you have a bit of time, you really should learn to use MARK, which is the de facto replacement for CAPTURE. Your best starting point is to point your browser at

http://www.phidot.org/software/mark/docs/book/

and go from there. There is a chapter in the book on closed population estimates, but you'd not want to jump into the deep-end with that first, without developing some basic skills.

If you insist on using CAPTURE, consider using the 'Estimator Appropriate' option. It will do rudimentary model selection for you, but ignores model selection uncertainty (i.e., won't give you model average estimates of abundance, if I recall correctly).

[pollock]

I just wanted to check that you said you had no recptures at all? That is a huge problem!
Think about the LP estimator in this case.One obtains an infinite estimate!
ken Pollock

[cooch]

I just wanted to check that you said you had no recptures at all? That is a huge problem!
Think about the LP estimator in this case.One obtains an infinite estimate!
ken Pollock

Whoops - didn't read the original note that carefully (I saw CAPTURE, model selection, and assumed that this was the basis of the question). Indeed, as Ken points out, if there are no recaptures, then there isn't anything you can estimate - the simple LP (Lincoln-Peterson) estimator would tell you why. If you have only a few recaptures, things become estimable, but the precision will typically be exceedingly poor.

[gwhite]

Petersen, NOT Peterson!

[cooch]

Petersen, NOT Peterson!

Heh - yup. My spell-checker changed it automatically before I hit the submit button.

Guess I'll stick with L-P next time.

[murray.efford]

Some other advice
- model selection in CAPTURE was always a bit of a joke, and certainly leads to screwy results over time if you re-select the 'appropriate' model for each session (dataset).
- you don't say what sort of data you have. If they are from a trapping grid or similar spatial array then you should be using spatial methods that take care of the spatial component of individual heterogeneity, and estimate population density - see e.g. Borchers & Efford Biometrics 64:377-385 and www.otago.ac.nz/density
- the way to deal with sparse data (occasional zeros) is to pool estimates of detection parameters across sessions. Gary has written on this somewhere, and it can be done for nonspatial models in MARK. For spatial models try DENSITY 'between-session' ML SECR models (I admit this is not well documented).
Murray

[jaxzwolf]

I'm sorry if I'm being naieve, here, but please bear with me. This is my first foray into population analyses.

I had ten trapping occasions for each study site. At two of my locations I had some captures, but no recaptures. I have read in some studies that in these instances the total population is taken to be the number of captures. I am trying to see if there is another (better) way to estimate population size for these data sets.

I had only one trapping session at each site; thus, I am unable to pool my data.

Do you think that using a program like DENSITY would be better in this instance than trying to get population estimates with CAPTURE? Could I utilize DENSITY if my trapping grid is unevenly spaced (i.e. traps set 82 m apart on the x-axis and 45 m on the y-axis with alternating rows of 4 and 5 traps)?

[murray.efford]

Hi
DENSITY works with any trap configuration.
You can and should pool detection parameters across grids (these can be treated as 'sessions' in DENSITY even though they are trapped at the same time).
Murray

[jaxzwolf]

Unfortunately, I won't be able to pool my data across grids because I am trying to assess whether population/density differs based on differing forest structural characteristics of each site. Essentially, my data sets are very small, but there is no feasible way for me to combine them and still address my research questions.

Is there any method of obtaining population estimates or density estimates using a single data set with 10 trap days in which no recaptures occurred?

I was able to obtain multiple population estimates for these sites running models individually in CAPTURE. I just don't know how accurate each of these estimates are, or if I can "choose" one as the most realistic/most robust estimate to use in my analyses.