Fitting robust-design open multi-state w state uncertainty

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

Fitting robust-design open multi-state w state uncertainty

Postby SGEnglish » Sun Oct 10, 2021 7:50 pm


I am having quite a bit of difficulty with model fitting and interpretation of a multi-state open robust-design model with state uncertainty in RMark.

The data I am trying model is comprised of ~4 years of continuous remotely sensed data for a few thousand individuals. I have collapsed the continuous sensing data into capture histories with 4 primary periods (years) each with 4 secondary periods (seasons). Marking occurred at irregular intervals throughout the 4 years. The population comprises residents and transients which are indistinguishable upon initial capture.

I want to model seasonal movement patterns, so I have been reading the Ruiz-Gutierrez et al. (2016) paper for the state uncertainty application to transience and residency, as well as the Kendall et al. (2019) paper which compares demographic parameters among seasons. I am confused by the apparent simplicity of the models described in the paper when I compare them to the models implemented in RMark, which I would think use the appropriate numerical processes for parameter estimation. Let me explain first the Ruiz-Gutierrez confusion:

This MSORD-SU model is said to have 9 parameters though I only understand 7 of them being detailed. Meanwhile, the RMark model, Robust Design Multi-state Open with Mis-classification (RDMSOpenMisClass) has only 8 parameters. In the manuscript, they drop 2 annual parameters (S, Psi) to focus on within-season estimates. They utilize state structure (omega), probability of entry (pent), persistence (Phi), probability of detection (rho), and probability of correct classification (delta). This leaves me two short from their 9 or 1 parameter short of the 8 in the RMark model (Pi).

Next, in the Kendall et al. (2019) paper, I would have expected that they would use an extension of the above model like the Robust Design Multi-state Open with State Uncertainty and Seasonal Effects (RDMSOpenMCSeas) in RMark, but they used instead the Open Robust Design Multi-state which has 5 parameters rather than the 10. I imagine that the ‘data-hungry’ characteristic of MSORD models described by Boys et al. (2019) is the reason behind this choice but I also struggle to interpret these additional parameters: alpha, c, and d.

Based on the information I have provided, it makes sense that I use the season effects model (RDMSOpenMCSeas) in RMark, yes? The help files with RMark including ‘parameters.txt’ and ‘models.txt’, and reading chapters 10 and 15 didn’t help me very much unfortunately.

Thanks very much to anyone who is able to help me clear up some of this confusion.

Ruiz-Gutierrez, V., Kendall, W. L., Saracco, J. F., & White, G. C. (2016). Overwintering strategies of migratory birds: a novel approach for estimating seasonal movement patterns of residents and transients. Journal of Applied Ecology, 53(4), 1035–1045.

Boys, R. M., Oliveira, C., Pérez-Jorge, S., Prieto, R., Steiner, L., & Silva, M. A. (2019). Multi-state open robust design applied to opportunistic data reveals dynamics of wide-ranging taxa: the sperm whale case. Ecosphere, 10(3), e02610.

Kendall, W. L., Stapleton, S., White, G. C., Richardson, J. I., Pearson, K. N., & Mason, P. (2019). A multistate open robust design: population dynamics, reproductive effort, and phenology of sea turtles from tagging data. Ecological Monographs, 89(1), e01329.
Last edited by SGEnglish on Mon Oct 11, 2021 6:44 pm, edited 2 times in total.
Posts: 3
Joined: Fri Oct 08, 2021 3:59 pm

Re: Fitting robust-design open multi-state w state uncertain

Postby Bill Kendall » Mon Oct 11, 2021 4:38 pm

Some of the aspects of your sampling are not clear, and therefore could use further clarification, but from the basics, with 12 total sampling periods, including 4 primary, that implies you have 3 secondary sampling periods per primary, and therefore the bare minimum needed for an open robust design model. Given you say you have continuous data, keep in mind that the ‘open’ aspect of the open robust design models has to do with staggered arrival and departure times, not mortality or true ingress and not state transitions within those primary periods.

Another issue that is not clear is, for your system, is whether there is an issue of state uncertainty for those that are observed, and what are your states and the nature of the uncertainty.

With respect to the particulars of the models you’ve been looking at, a lot of your confusion is understandable since we don’t have Mark book chapters for any of them except the basic open robust design (my fault, not Evan’s). If you have no state uncertainty, then the MSORD model described in the 2019 paper is appropriate, and includes five parameter types (survival, state transitions, arrival probabilities, persistence probabilities, and detection probabilities), although there is an alternative parameterization (through Change Data Type) to include state structure for a given primary period.

If you have uncertainty about the state of a detected individual, then the MSORD-SU model outlined in the 2016 paper makes sense. With this model there are two additional parameters types: pi(state | unknown) = probability a newly detected individual is in each state, given its state is unknown; and delta(state | state) = probability a recaptured individual’s state is identified.

The last model you mention, with seasonal effects, is an extension of the above model, where the ability to assign a state is itself seasonal (e.g., where there is a breeding state and a non-breeding state and the young have not yet been born when resighting begins, and some may be weaned or dead by the time the resighting ends). So there are two additional parameter types: one for the initiation of the ability to identify the state, and one for the cessation of that ability.

With more information about exactly what your situation is, I could perhaps provide more guidance.
Bill Kendall
Posts: 91
Joined: Wed Jun 04, 2003 8:58 am

Re: Fitting robust-design open multi-state w state uncertain

Postby jlaake » Mon Oct 11, 2021 5:06 pm

I'm going to chime in here with a comment about posting to the RMark sub-forum. Bill's answer makes it fairly clear that this post had absolutely nothing to do with RMark and it should have been posted elsewhere. Unless the subject is specific to some task or problem with RMark then it shouldn't go in the RMark section. This post is a general analysis question about robust design models and analyzing a particularly dataset. RMark is only an interface to MARK and it is MARK that has models and not RMark. And as I have pointed out before, I can compose an interface to these MARK models without knowing very much about the model or how to analyze data with a particular model. I do know some of the model particulars because I have analyzed data with them but with age and retirement that knowledge is decreasing every day.
Posts: 1261
Joined: Fri May 12, 2006 12:50 pm
Location: Escondido, CA

Re: Fitting robust-design open multi-state w state uncertain

Postby SGEnglish » Mon Oct 11, 2021 7:06 pm

First, thank you very much for your reply. I will try to better delineate the R component of my question from the analysis component next time. But for the sake of this thread, ultimately I do aim to fit an RMark model and I hope that this is clear enough. I am not sure if the thread should be moved or not—looks like that is at your discretion.

I have edited my original post to make it more clear. Briefly, my data have 4 primary periods (years) and 4 secondary periods (seasons). Given the relaxed assumption for staggered entry and departure, I think I can roughly assume that the population is ‘closed’ for a period of 1 year. The state-uncertainty is that the population comprises transients and residents, indistinguishable on first encounter.

Next, and what I am really interested in, are the proportions of transients and residents, their seasonal movement (i.e. whether re-detections are less probable for some at certain times of year while they are likely to return the following year/season/never). So in principle the application I thought should be very similar to the 2016 Ruiz-Gutierrez et al. paper: “[…] use time- and state-dependent probabilities of site entry and persistence as indirect measures of movement.” Only that I want to use the model for inter-seasonal movement with my data and not so much intra-seasonal movement. Still, I will need to consider transient and resident state uncertainty because both could bias estimates of persistence or probability of entry into the territory.

Although I haven't got it all worked out just yet, I'll post what I have understood, hopefully to the benefit of someone else toiling away with RMark.

To fit an ‘RDMSOpenMisClass’ model in RMark, several parameters must be constrained in some way or another. The 2016 publication gives a good framework for implementing the all the necessary restrictions with this model in MARK. In RMark, constraining the pi parameter (probability of being in either state given that state is unknown) is simple enough:
Code: Select all
ddl$pi$fix <- 0.5

if there are 2 states. Alternatively, if there are 3 states, I guess:
Code: Select all
ddl$pi$fix <- 0.33

Next, not all parameters are utilized in the model. I believe this means that a parameter is fixed at a value of 1, although I would very much like to hear if somebody knows better.
Posts: 3
Joined: Fri Oct 08, 2021 3:59 pm

Return to RMark

Who is online

Users browsing this forum: No registered users and 1 guest