mixed modeling in MARK?

[rso]

Hi All,
I would like to specify time as a random effect on PSI and S in a multi-state closed robust design with misclassification, keeping other covariates fixed. I have read Appendix D of "A Gentle Intro" and phidot postings on random effects in MARK, but still am having trouble creating the model output that I would like: single beta for the random time covariate, and individual betas for all the fixed covariates; is this possible to construct in MARK, or will I need to switch to RMark to utilize the mixed modeling commands?

The dataset consists of two groups, an individual time-varying covariate, two temporal covariates (coded in fixed effects DM), and two states (missclassification in one direction only).

For RMark, is there an example available for constructing MSCRD with misclassification? (I am not sure how to specify the unknown state).

thanks,

Ricka

[jlaake]

RMark is an interface to MARK and all of the modelling is done in MARK. The only mixed-modelling in MARK that I know of is the CJS Random model which allows individual heterogeneity in Phi and p. But Gary is always adding new models. Alternatively, you can fit a temporal model and then use the variance components output to look at temporal variation. In RMark that can be done with var.components and var.components.reml.

With regard to your question about MSCRD, I'm not sure what you are referring to. MSCRD does not have miss-classification. It can have an unobserved state which by definition is not observed and is not in the capture history. As in the ?crdms example, its value is specified in strata.labels.

The models with misclassification are RDMSMisClass and RDMSOpenMisClass in which during secondary periods model is closed and open respectively. There are no examples in RMark currently for these models.

There is a recent Dec 2012 article in Methods in Ecology and Evolution by Ford et al that discusses using ADMBre to include invididual random effects in a multi-state model. You may also want to look at some work I've been doing https://github.com/jlaake/ADMB-Examples and specifically https://github.com/jlaake/ADMB-Examples ... eneity.pdf

regards --jeff

[rso]

Thank you Jeff. I appreciate your quick reply and will look at the links. For MSCRD with missclassification, I meant RDMSMisClass (I think these are the same models? based on Kendall et al. 2012. Ecology 93(4): 913-920).

thanks,

Rikca

[jlaake]

Yes, RDMSMissClass is a closed RD multi-state model with misclassification although as Bill will say misclassification should really be "state uncertainty". In MARK these models use "u" (a lower case u) as the uncertain state designation.

--jeff

[rso]

One follow up question - I'm realizing that I may be setting up the fixed model incorrectly in the first place: is it acceptable to include year-specific covariates with a temporally-varying covariate in the same model?

thanks,
Ricka

[jlaake]

I assume you mean something like

~time + Effort

where Effort has a numeric value for each time and time is a factor variable. Answer is no! It is over-parameterized. Now if the formula is for a secondary occasion in a robust design and it is something like

~session + Effort

and Effort varies within secondary occasion that would be perfectly acceptable because there is variation in effort across secondary occasions. ~session would be the intercept for the session and the values for a secondary occasion would be intercept + b*effort_j where effort_j is the effort on the jth secondary occasion. You could even have ~session*Effort - a different intercept and slope for the session.

--jeff

[rso]

Okay - thanks again Jeff!

[cooch]

I assume you mean something like

~time + Effort

where Effort has a numeric value for each time and time is a factor variable. Answer is no!

Indeed - See the Mark help menu:

Help | Ten Commandments -> very first commandment.

"Thou shalt not use a temporal covariate with a fully time-varying model..."

Jeff (acting as Moses, here) explains why.