Nested DM Functions – Nest survival stage model

[birdman]

I am trying to use the functions suggested in the nest survival chapter of the Gentle Intro to generate a stage specific model of nest survival (2 stages, egg and nestling). I’m getting bizarre results and would like some input. I searched the forum and found some potentially helpful things, but so far, I’m still stuck.

For each nest record, “i” is the initiation of egg laying. I have an Age covariate denoting the negative age of each nest on the first day of the nesting season which allows me to model age related effects using the Add(age,0)… Add(age,n) function. Because we know hatching date very accurately and there is some variation around the 17-18 day mean incubation period length, I included a calculated covariate (HtchDay) for the age of the nest at hatching (generally 17 or 18, but including a few shorter and longer intervals). I then used the combined function ge(Add(age,n),HtchDay) to create a dummy variable where 1 = nestling stage. My nest season is 146 days long from first initiation to last fledge, so my DM includes 145 parameters.

I ran the model S = B0 + B1xstage and my results included B0 = 18.809 and B1 = -15.781, but no standard errors for either. So I get survival of 1 during egg stage, survival of 0.954 in nestling stage, and no error.

Can anyone enlighten me as to what (probably simplistic)mistake I’m making?
Thanks
Pru

[birdman]

I've continued to work on this issue a bit and still can't find a reason that I'm getting daily survival of 1 for the egg stage and no standard errors. I'd really appreciate if anyone who has run nest stage models as described in the Gentle Intro could weigh in here with what might be going wrong. Given that my other models (age, T, etc.) run and produce reasonable results with appropriate error estimates, I feel it is a model structure problem and not a data structure issue.

I also realize I posted this around the time school sessions were finishing up so maybe it was largely overlooked.

Cheers,
Pru

[bacollier]

I've continued to work on this issue a bit and still can't find a reason that I'm getting daily survival of 1 for the egg stage and no standard errors. I'd really appreciate if anyone who has run nest stage models as described in the Gentle Intro could weigh in here with what might be going wrong. Given that my other models (age, T, etc.) run and produce reasonable results with appropriate error estimates, I feel it is a model structure problem and not a data structure issue.

I also realize I posted this around the time school sessions were finishing up so maybe it was largely overlooked.

Cheers,
Pru

Pru,
Yeah, or no one knew the answer right off the top of their head (which does happen). Shoot me your input datafile and the mark.dbf and .fpt files offlist and I will try to take a look at it this weekend (bret@tamu.edu).

Bret

[bacollier]

Pru,

Will reply to you off list as well tomorrow, but I wanted to post this so that it would be archived.

I messed around with your nest survival stage only model DM a bit tonight and I can get it to give me 'reasonable' results (in terms of the parameter estimates and their associated standard errors not being basically unestimable). Based on your code, I think I debugged the issue to the "HtchDay" covariate value in the DM line GE(add(Age, value), HtchDay). I tested a few different options on how to do this and what seems to work is when I replace that covariate value "HtchDay" with a constant numerical value (I used 17 days based on what you had said earlier). When I do this, the estimates of the beta parameters seem estimable and reasonable (see bottom) of message.

So, I think that the use of the value "HtchDay" is causing the issue. Offhand the only reason I think it might be messing up is that the length of the 'stage' needs to be a fixed value across all nests for the analysis in MARK, and yours was variable in your dataset-but I am guessing so maybe Gary or someone else more familiar with the internals on that can advise. I did not delve overly deeply into your data, but I think this may get you closer.

here is what my testing ended up with beta estimate-wise.

LOGIT Link Function Parameters of {Test}
95% Confidence Interval
Parameter Beta Standard Error Lower Upper
------------------------- -------------- -------------- -------------- --------------
1:int 4.0757566 0.0308664 4.0152585 4.1362547
2:Bretcolumn -0.5632560 0.0407542 -0.6431341 -0.4833779

[birdman]

Bret,

Thanks much for having a look at this. I too suspected the HtchDay variable, and ran a model setting a constant incubation interval (18 days I think) by using GE(add(Age, value), 17) in the DM. At least on my initial run, this didn't work out any better, but it is possible I simply coded something incorrectly in my haste to test that idea. I'll play around with this some and see if I can get the same kind of results you found, and then if I can get other, more relevant models (stage+age for example) to fall into step.

If Gary, Evan, or others have comments on whether you can (theoretically) use variables, as opposed to fixed values, in the GE function I'd be interested to hear.

Pru

[birdman]

I promised Bret after getting some help from him off-list that I'd post a final resolution to the issue discussed above for others who may come across it. As Bret surmised, there was a problem with using the HtchDay variable.

Briefly, I was trying to build a stage model for assessing differences in daily survival by egg or nestling stage. Our study species is fairly consistent regarding incubation length, but there is some variation (deriving from clutch size), and I wanted each nest to be properly categorized as either egg or nestling stage. So, my variable [HtchDay] = (#Days from 1st egg laid to 1st hatch). My plan was to use this as an individual variable combined with the nest age [Age] in a Greater Than or Equal To function as:

GE(Add(Age,<incrementing #>),HtchDay)

Use of the age variable to generate age and stage models is discussed in the Gentle Intro. Bret discusses using a fixed value (as outlined in the G.I.) above. I continued to play around with this and finally figured out what was actually going on (I think) with my nest-based incubation length variable.

Many nests failed before hatching. Those nests were given a “0” value in the HtchDay variable. Plugging the 0 into the GE equation above results in such nests never having an egg stage. As soon as they are initiated they are nestling stage (via the calculation). Exactly why this would confound estimation is beyond my desire to get my brain around at the moment, but when I substitute “21” (the average incubation period length) for “0” for the GE breakpoint for incubation stage failed nests, the model converges very nicely and estimates Betas similar to what Bret generated in his run using 17 days for all nests.

LOGIT Link Function Parameters of {S(Stage) NestBased}
95% Confidence Interval
Parameter Beta Standard Error Lower Upper
------------------------- -------------- -------------- -------------- --------------
1:int 4.0493177 0.0281058 3.9942304 4.1044050
2:Stage -0.6203024 0.0401688 -0.6990332 -0.5415715