Missing samples with live recaptures

[geoffwah]

Dear All

From my readings of 'the book' and past discussions on this forum, I was under the impression that MARK handles missing samples when analysing live recapture data no probs. However, I'm having some troubles in this regard. I am interested in survival rates of frogs, and conducted a mark-recap study at various ponds over two years. However, not all ponds were sampled in both years. Thus, I have 18 occasions in total (8 in year 1, 10 in year 2), but roughly half of the sites were only sampled in the first year.

Thus, I have:
Frog 1: 000011010000001000
etc
Frog 28: 10100110..........
etc
Frog 50: 000011010101001000
etc

I can run live recapture models on these data, but the Deviance estimate is always 0. However, when I run the models using data from only one season, no problems.

If anyone has some ideas on where I may be going wrong here I would grealty appreciate it! Is there an inherent problem I'm not aware of, or is there possible problems with my data (e.g. too sparse??)

Thanks in advance

Geoff

[abreton]

I searched (briefly) the MARK help file and Chapter 2 in 'the book' and wasn't able to locate any information about using a "." to code for missing data in the CJS and closed captures encounter histories - so perhaps you're not aware that this is an option? Point your browser at http://www.phidot.org/forum/viewtopic.p ... ssing+data
...this is a archived post from the phidot website. I used the key words "missing" and "data" and this was about the third hit. See comments from Gary and others.

Two other thoughts come to mind after reading your post. Based on the structure of your encounter histories, I assume your trying to estimate survival from one month to the next with perhaps one interval (winter?) being a few months longer than the standard 1 month interval. Keep in mind that if survival is extremely high (>~95%) from month to month in these frogs then you're likely to have survival estimates on a boundary which can cause convergence problems when the logit link function is deployed. Convergence failure from this source could cause the behavior your detecting. Sparse data on each sampling occasion at each pond could also be the issue. Try simpler models and/or pooling data.

Also, you may be sampling 'all month' each month. If so, then your sampling periods would be far from 'instantaneous', an assumption of the CJS and other CMR model types. I suggest having a look at text on page 423 starting "The assumption of instantaneous..." in Williams et al. 2002 (Analysis and Management of Animal Populations). You may have to pool data to approximately meet this assumption. Note, this issue wouldn't cause convergence failure. Good luck.

[cooch]

I searched (briefly) the MARK help file and Chapter 2 in 'the book' and wasn't able to locate any information about using a "." to code for missing data in the CJS and closed captures encounter histories - so perhaps you're not aware that this is an option? Point your browser at http://www.phidot.org/forum/viewtopic.p ... ssing+data
...this is a archived post from the phidot website. I used the key words "missing" and "data" and this was about the third hit. See comments from Gary and others.

Not in the book because it isn't fully implemented, and isn't quite ready for prime-time in general (wide) use.

[Eric Janney]

I am having similar issues with groups not being sampled on all occasions. I have 13 years of CJS data using PIT tags on fish. In year 11 we switched to a new and improved PIT tag that has a much better read range. So, on occasion 11 we started tagging newly captured individuals with the newer tag type; however, we do not "retag" fish recaptured with the old tags. Because we use remote underwater PIT detection systems, I hypothesized that fish tagged with newer tags will have higher p's in occasion 12 and 13 than fish tagged in occasions 1-10, but no differences in survival because the tag size/shape is the same. So, I now have four grougs: males with old tags, males with new tags, females with old tags, and females with new tags. It has been a bit of a nightmare trying to code this in the design matrix and I'm getting different parameter counts (which I expected) than when I run the same model using the PIMS. I am also getting different deviances (which I didn't expect), but I'm sure it's due to sparse data in some years.

[geoffwah]

Hi All

Much thanks for these comments. Yes, the discussion previously re missing data and the use of "." for CJS models was primarily why I thought this was do-able. And yes, I do have a much large time interval between the first and second seasons, however, survival rates shouldn't be that high in these frogs, and was using the Sin link so hopefully this means this shouldn't be the problem. I sampled one night each fortnight so these data meet the assumtpion of instantaneous sampling too.

Perhaps my data are just too sparse - recapture rates are quite low for example (with the single season data, p was in the order of 0.14).

I have just discussed the issue with a friend who informs me that constraining the p's to zero for frogs in those ponds not sampled in the second year might be the way to go. I have subsequently found a post from Marc Kery in this regard: http://www.phidot.org/forum/viewtopic.p ... +occasions

I'll give this a go, but any further thoughts would be greatly appreciated!

Regards

Geoff