behavior*time*sex DM help

[scoster]

Hello again,

I'm a novice MARK user, and have learned a lot in the past two months as I've been reading into the gentle introduction and other texts. I don't have anyone on campus I can ask these basic questions, and I recognize that sometimes my questions are unclear (hence no response here on the forum). Through lots of thought and re-reading, I've come to understand several of my questions, but one has still be nagging me. And that is how to construct the DM (using a Huggins closed capture model with 6 trapping sessions) for a model that tests behavior*time*sex. Below is the DM I've constructed, but it does not have the same deviance as when I construct this model using the PIMs. I've been staring at it for days trying to see an error, and I haven't had any luck. Can anyone please spare a moment to help me out with this? I would really appreciate any feedback.

Many thanks,
Stephanie

Int B S T1 T2 T3 T4 T5 BT1 BT2 BT3 BT4 BT5 ST1 ST2 ST3 ST4 ST5 BST1 BST2 BST3 BST4 BST5
P 1 1 1 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0
P 1 1 1 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0
P 1 1 1 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0
P 1 1 1 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0
P 1 1 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1
P 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
P 1 1 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
P 1 1 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
P 1 1 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
P 1 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
P 1 1 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
P 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C 1 0 1 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
C 1 0 1 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
C 1 0 1 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
C 1 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
C 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C 1 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

[cooch]

The model you're interested in (sex*time*behaviour) is not fully identifiable - you can build it with PIMs, but many of the estimates will be 'nonsense' - look at the estimates from the model built using PIMs, and I suspect you'll see this. So, the reason your model deviance differs between models built with PIMs and your attempt at a DM has more to do with this issue, and more how different link functions (sin for PIMs, logit for DM) will handle the non-identifiability issue.

For the moment, ignore the sex effect, and think about being interested in the model time*behaviour. Then, re-read section 14.5.1 - pay close attention to the DM on p. 14- 16. This DM corresponds to time*behaviour, but as noted in *many* places in the text, it is not identifiable without constraints.

[scoster]

Thanks for your response, that is helpful. I recognize the behavior*sex*time model is not fully identifiable, I was hoping to create a DM that matched the deviance of the PIM and then delete the interaction terms to create the more constrained additive model. But I became concerned when the deviance between the DM and the PIM didn't match. Would the following DM therefore be correct for the additive model?

Int B S T1 T2 T3 T4 T5
P 1 1 1 1 0 0 0 0
P 1 1 1 0 1 0 0 0
P 1 1 1 0 0 1 0 0
P 1 1 1 0 0 0 1 0
P 1 1 1 0 0 0 0 1
P 1 1 1 0 0 0 0 0
P 1 1 0 1 0 0 0 0
P 1 1 0 0 1 0 0 0
P 1 1 0 0 0 1 0 0
P 1 1 0 0 0 0 1 0
P 1 1 0 0 0 0 0 1
P 1 1 0 0 0 0 0 0
C 1 0 1 0 1 0 0 0
C 1 0 1 0 0 1 0 0
C 1 0 1 0 0 0 1 0
C 1 0 1 0 0 0 0 1
C 1 0 1 0 0 0 0 0
C 1 0 0 0 1 0 0 0
C 1 0 0 0 0 1 0 0
C 1 0 0 0 0 0 1 0
C 1 0 0 0 0 0 0 1
C 1 0 0 0 0 0 0 0


I also have another related question. If I am choosing the candidate models apriori and in analysis a model turns out to have inestimable parameters, is it appropriate to delete that model from the candidate model set at that time, or should the model stay in the candidate set even though I know the model is not of any value?

Thanks a lot, I look forward to the response.

Stephanie

[dhewitt]

Hi Stephanie,

There is a difference between parameters being inestimable and parameters not being identifiable. First, some identifiable parameters may be inestimable because of data issues (sparsity). The model still has value even if some estimates (hopefully a rather small proportion of them!) are of no value. Second, there are different degrees of problems with parameter identifiability. The one you are encountering here is rather serious, and you need to understand the identifiability issues and then not run models with such problems (or, if you run them "on accident", throw them out and don't think much about them). Some identifiability problems are less serious, like the last p and Phi in CJS models. There the solution is easy -- just don't look at those estimates.

- Dave

[scoster]

Dave,

Thank you for taking the time to respond to my question. Your answer has provided some clarity, and I appreciate that. I wonder when you say:

"Second, there are different degrees of problems with parameter identifiability. The one you are encountering here is rather serious, and you need to understand the identifiability issues and then not run models with such problems"

are you referring to the problems with parameter identifiability in the b*t*s model? I do recognize that those are not identifiable, and should not be run. Or are you referring to the b+t+s model, because I was under the impression that those should be identifiable, because it is constrained (and if I am getting inestimatable parameters, perhaps that is due to data issues as you suggested). Can you help clarify this for me?

Many thanks,
Stephanie

[dhewitt]

I was referring to the fully interactive model. As you note, running that model is not a good idea.