different recovery models - alternative optimization?

[mherzog]

I am running some normal recovery data through MARK.

Specifically, I am running a sex*age*time model, and the final result produces estimates (some of which are not estimated), and in the prompt at the end of the run there is something to the extent of convergence "suspect" (can't remember it exactly).

Anyway, I see the MARK provides an alternative optimization method... GOOD! I have now run the data (same model) through this and while it takes about 3-4 times longer, it is happy, no convergence problems, lower deviance, etc. - even calculated parms correctly.... GOOD!

However, now I wanted to get a c-hat for this recovery data, and currently these models are in the brownie parameterization, so I change my model type via the PIM menu to recoveries only, and rerun the model (again - sex*age*time using the alternative optimization checkbox in MARK).

This run, however, produces a different deviance than the brownie run of the same data, optimization method, and model. I didn't think this was supposed to be the case. From my reading, I thought the same model in the brownie vs. recoveries only model would produce the same AIC, deviance, etc. Is this correct?

Mark

[gwhite]

Mark:

The fact that you were having optimization problems tells me that the data are pretty sparse. Likely, because of the sparse data, you have some survival rates that are estimated to be >1 in the Brownie parameterization. If you used the identity link, the estimates will exceed 1. In contrast, the Dead Recoveries parameterization forces survival to be in the 0-1 interval, regardless of the link function. Thus, I suspect that you've got some parameter estimates that are giving you boundary problems.

The 2 parameterizations will give exactly the same deviance for the fully group and time specific model (global model) if no parameter estimates exceed 1. If you constrain all the estimates to 0-1 with the logit or sin links, you should get exactly the same deviance for the global model. If not, then something is amiss in what you've done.

Gary

[cooch]

Mark:

The 2 parameterizations will give exactly the same deviance for the fully group and time specific model (global model) if no parameter estimates exceed 1. If you constrain all the estimates to 0-1 with the logit or sin links, you should get exactly the same deviance for the global model. If not, then something is amiss in what you've done.

Gary

Gary, I've run into this on occasion myself. While I think I've figured out the mechanics of forcing 0-1 with logit or sin links, could you spell out the basic mechanics (or point to the right part of the helpfile)?

Thanks!

[mherzog]

I used the logit link for all of these models.... Also, when you say the "fully group and time effect" model will have the same deviance:

My groups are age and sex, so I was generally using an age*time model as my global model rather than g*t. Are you implying that only the general model g*t will have the same deviance in both parameterizations, and not other models?

I am running this right now, so I'll get my answer shortly with respect to this specific question, but I am curious if you are also stating that the ONLY model that will result in the same deviance between recoveries only and brownie paramterization is the global model? And this "global" model is g*t, and not age*t.

Thanks.

[cooch]

I used the logit link for all of these models.... Also, when you say the "fully group and time effect" model will have the same deviance:

My groups are age and sex, so I was generally using an age*time model as my global model rather than g*t. Are you implying that only the general model g*t will have the same deviance in both parameterizations, and not other models?

I don't think thats what Gary's implying - I think he means that IF you constrain the parameters to be estimated [0,1] bounded, then both parameterizations (Brownie and non-Brownie) should yield the same result.