numerical convergence problem with multistrata model

[justyn]

I'm working on a longterm study of the cooperatively breeding Acorn Woodpecker. MARK fails to reach numerical convergence with even the simplest model (all parameters constant). I have 33 occasions of two groups (males and females) and three strata (juvenile, helper, breeder) and I'm trying to estimate sex and strata specific differences in survival and transitional probabilities. Juveniles can never be recaptured (because at the end of the first occasion they must either disperse/die, or transition to either helper or breeder). Additionally, its impossible to transition from Breeder or Helper back to juvenile. I feel like this is what is causing MARK problems, but I'm not sure.

Any suggestions?

U-care also crashes when trying to estimate GOF.

Cheers,
Justyn Stahl

[justyn]

PS I feel like the MLogit approach would help, but I'm never given the opportunity to specify link values...

[cooch]

I'm working on a longterm study of the cooperatively breeding Acorn Woodpecker. MARK fails to reach numerical convergence with even the simplest model (all parameters constant). I have 33 occasions of two groups (males and females) and three strata (juvenile, helper, breeder) and I'm trying to estimate sex and strata specific differences in survival and transitional probabilities. Juveniles can never be recaptured (because at the end of the first occasion they must either disperse/die, or transition to either helper or breeder). Additionally, its impossible to transition from Breeder or Helper back to juvenile. I feel like this is what is causing MARK problems, but I'm not sure.

Any suggestions?

With that many logical constraints, I'm guessing you're not fixing some parameters at the time you run the numerical estimation. You should.

[cooch]

PS I feel like the MLogit approach would help, but I'm never given the opportunity to specify link values...

Have you read Chapter 9? The MLogit link is discussed on pp. 19-20. Also, discussion of fixing parameters is presented in the same chapter.

[gwhite]

Justin:
First, you should set up the transitions to be the ones you want to estimate or fix. To do this, open the results browser window, and then open the PIM | Change PIM definitions menu. In that dialog, select the transitions that you want to obtain by subtraction, i.e., which transitions are not estimated, but obtained by subtracting from 1 the sum of the estimated transitions. You need to do this change each time your retrieve a model from the results browser because I haven't been cleaver enough to get MARK to retrieve these setting from a previously run model.
Second, when you run the model, select the parameter specific link function option. Then, you will be asked to assign a link function to each of the real parameters.
Finally, if you are still having trouble with convergence, and are sure that you have specified the model correctly, use the simulated annealing optimization algorithm (the alternate optimization algorithm). Combined with the mlogit links, I've never had a problem getting a solution with that approach.

Gary

[justyn]

Thanks guys! I was unable to fix Psi1,1 because it was being estimated by subtraction. This was the one part I overlooked. Thanks again.

[cooch]

Thanks guys! I was unable to fix Psi1,1 because it was being estimated by subtraction. This was the one part I overlooked. Thanks again.

As per Gary's note, you can specify which parameters are being estimated in the MS model - allowing you to fix whichever ones you want/need to. While the 'switching' is covered in the latest incarnation of the MS chapter, the utility of this function for fixing logically constrained parameters (which Gary noted) is not - I'll have to add it to the next revision.

[justyn]

As I understand it, specifying which transition parameters are being estimated in the MS model only applies when using PIMs. Is there a way to modify the design matrix so that Psi(1,1) is present (so I can fix it to zero)? I have a lot of group covariates (weather, food, etc.) that may be the underlying cause of the temporal variation in survival/transistion probabilities, and obviously need to move beyond PIMs.

[cooch]

As I understand it, specifying which transition parameters are being estimated in the MS model only applies when using PIMs. Is there a way to modify the design matrix so that Psi(1,1) is present (so I can fix it to zero)? I have a lot of group covariates (weather, food, etc.) that may be the underlying cause of the temporal variation in survival/transistion probabilities, and obviously need to move beyond PIMs.

Actually, I think you misunderstand how the DM and the PIMs interact. The PIMs define the underlying parameter structure, which you then constrain using the DM. This general idea is covered in some detail in Chapter 7.

So, in your case, specify the transitions you want, then constrain them using the DM.

[justyn]

As I understand it, specifying which transition parameters are being estimated in the MS model only applies when using PIMs. Is there a way to modify the design matrix so that Psi(1,1) is present (so I can fix it to zero)? I have a lot of group covariates (weather, food, etc.) that may be the underlying cause of the temporal variation in survival/transistion probabilities, and obviously need to move beyond PIMs.

Actually, I think you misunderstand how the DM and the PIMs interact. The PIMs define the underlying parameter structure, which you then constrain using the DM. This general idea is covered in some detail in Chapter 7.

So, in your case, specify the transitions you want, then constrain them using the DM.

I've assigned Psi1,2 to be obtained by subtraction, and therefore Psi1,1 and 1,3 appear in the Design Matrix. I fix 1,1 to zero, and then constrain 1,3 according to covariates. Obviously, Psi1,2 does not appear in the DM, but it is still being "constrained" in the same manner as Psi1,3?