Dealing with transience: TSM or age classes?

posts related to the RMark library, which may not be of general interest to users of 'classic' MARK

Dealing with transience: TSM or age classes?

Postby sharrod » Tue Aug 13, 2019 6:06 pm


I'm hoping to get some clarification on the following issue. I have 16-year CMR data set of birds marked as young (L) and adults (SY and ASY). I'm trying to estimate population sizes of each age class for each year of the study using a POPAN model. I tested the GOF of the most general model and found significant transience among young birds (not surprising since many die after fledging or disperse). Even after manually removing young that were never recaptured, I still have a transience issue. I also tried suppressing the first encounter, which took care of the transience issue but led to a trap-dependence issue.

I know building time since marking (TSM) models (i.e., Φa2*t, pt) is recommended in this instance, but I'm wondering if incorporating the above age classes in each model is enough to take care of it. I created my tsm variable based off this post (viewtopic.php?f=21&t=2591), and since I know age has an effect on survival, I've already incorporated age class into several models. Right now the Phi models I'm working with are

which I'm thinking are overly complicated (and maybe redundant)?

Simply put, do I need to include the tsm variable in all my models, or is including the age classes in each model adequate to deal with transience? Any clarification or suggestions would be greatly appreciated.

Thank you.
Posts: 10
Joined: Thu Aug 24, 2017 12:05 pm

Re: Dealing with transience: TSM or age classes?

Postby B.K. Sandercock » Wed Aug 14, 2019 3:29 am

A time-since-marking (tsm) model and a 2-age class model are the same thing, they both separate the interval after first marking (phi1) from later intervals (phi2+). A difference between the two parameters (ph1 < phi2+) could be due to differences in survival due to age or handling effects, or to differences in site fidelity due to transience, heterogeneity of capture, or again handling effects.

Model selection stats and parameter estimates should be identical for these two models:
Phi.tsm=list(formula=~ tsm)
Phi.2ac=list(formula=~ age_class)

You might set up your models with 3 age classes for young (L -> SY -> ASY), 2 age classes for SY (SY -> ASY) and one age class for ASY (ASY-> ASY).
If you are concerned about transience, use 2+ classes for all: 3 age classes for young (L -> SY -> ASY), 2 age classes for SY (SY -> ASY) and one age class for ASY (ASY1-> ASY2+). Tsm and age effects have the same structure for L and SY birds.
In practice, most young just disappear in vertebrate populations because of some combination of low survival and natal dispersal. Often better to just treat them separately since the recaptures are often low. If you have a small number of philopatric young that do come back, you could just censor their records as young and just use the adult part of their encounter history along with all other adults (e.g., L0SA0A becomes 00SA0A).

Good luck with your project.
B.K. Sandercock
Posts: 48
Joined: Mon Jun 02, 2003 4:18 pm
Location: Norwegian Institute of Nature Research

Re: Dealing with transience: TSM or age classes?

Postby jlaake » Thu Aug 15, 2019 2:44 pm

Thanks for addressing this Brett. I want to address a related issue here.

One of the dangers of RMark is that you can specify a model without having a clue what it specifies or does. That is not a good thing and it is one place where Evan Cooch and I are in complete agreement about RMark. I'm not targeting you specifically because others have done or do the same thing. RMark was created to avoid the drudgery of creating design matrices manually. Unfortunately the side-effect is that you can specify a model with a formula and have no clue what the model is doing. That is not good. You seem to recognize that many of your models were likely over-parameterized. They are! In fact I can't make sense of most of them. Please don't ever specify a model that you don't understand and could explain to some one in words or preferably write down a formula for the model.

I'll take one of the models as an example. ~time*tsm. I'll assume tsm is a numeric variable. This formula specifies that the logit of survival is a linear function like a+b*tsm where a and b differ for each time. So if you had 10 time intervals it would specify 10 intercepts (a) and 10 slopes (b). Now for the first interval, you only have 1 value of tsm=0 because they comprise the first animals that were tagged. You need two values of tsm for a valid regression. But now think about what the model is saying. It is saying that the effect of tsm is changing (different slope) over time. My guess is that you didn't really understand that is what it was doing. These are linear models on the logit (or other link) scale and most people can understand linear equations. If you don't have a good understanding, go back and re-visit an algebra text. Or you can work through the LineaModels.pdf in Anyone who is using RMark should read that pdf if they don't have a complete understanding of linear models with capture-recapture. There is also lots of material about this same subject in Evan's book.
Posts: 1378
Joined: Fri May 12, 2006 12:50 pm
Location: Escondido, CA

Re: Dealing with transience: TSM or age classes?

Postby cooch » Thu Aug 15, 2019 3:11 pm

I don't want to 'pile on', but I will add one short comment...

jlaake wrote:One of the dangers of RMark is that you can specify a model without having a clue what it specifies or does. That is not a good thing and it is one place where Evan Cooch and I are in complete agreement about RMark.

Having to construct the DM for a model 'by hand' in classic (GUI-based) MARK doesn't inherently reduce the chance of error (a lot of people assume that simply because the model runs with an 'error message', that the answer is correct -- one reason as a learning heuristic I force my suffering students to do a lot of things 'by hand') -- but the fact that a hand-coded DM is slower to construct than a model built using RMark syntax might add enough of a 'lag' in the process to force some extra thinking. Or put another way, the elegance and efficiency that RMark affords lets you do silly things faster.

But, on the upside, in the hands of someone who knows what they're doing, RMark lets you build a lot of models easily and quickly, which of course is on reason I urge people who've reached a significant skill level with 'classic MARK' to consider moving to RMark.

p.s. Brett is entirely correct (no surprise) in his comments on TSM and age models. In fact, the TSM model is just a 2-point finite mixture model (Chapter 14) in disguise, and is one of a handful of ways to accomodate unobservable capture heterogeneity by tweaking the structure of the model.
Posts: 1613
Joined: Thu May 15, 2003 4:11 pm
Location: Cornell University

Return to RMark

Who is online

Users browsing this forum: No registered users and 1 guest