Error message when running Random Effects Models in MARK.

[Diego.Pavon]

Hi all,

I am interesting on running random effects models in MARK but I am having some "technical troubles".
I have CMR-based data on Ural and tawny owl (1986 until 2010).
When I get to the "Variance Component Estimation" window I select "random effects", "graphical output", "numerical output" and "intercept only". I always get the error message: "error -- matrix D was found to be not positive definite. Sigma 2 = 2437.8700 (3369.0477 to 1507.2612)"

I only get this error message when retrieving (and using) the Phi(t)p(t) model. When I use any other model, I do not get this message but I cannot get the graphical output. In addition, when I add the model to the browser, the AIC is very large, and the estimates are 0 or 1.

Does anybody know how to solve this problem or have any alternative idea to run these models?

Thank you very much in advance.

Diego

[Diego.Pavon]

Hi again,

I do not know how to edit previous messages. Anyway, I managed to run the random effect from the Phi(t)p(t). The problem was that I was taking all the parameters and I should have left the last parameter out of the analysis because the last Phi can be confounded with the first p (appendix D.4.3).

However, the model does not work properly. The deltaAIC is >300000 larger than the time-dependent model.

phi(t)p(t) AIC=2463 par=46
phi(t)p(t)random effects intercept only TG=14.2 AIC=308467

Any thoughts would be appreciate.

Thank you

[abreton]

Did you select "all of the [p and phi] parameters" except the last [p and phi] or "all of the [phi] parameters" except the last [phi]? I suspect that you did the latter. But If the former, then that's at least part of the problem -- choose only the phi parameters except the last phi. If you're modeling both species, as groups, in the same model, then you'll want to (I suspect) select one species at a time otherwise you're suggesting that survival parameters for the two species came from the same distribution. Given the notation you provided in your post, seems that you're analyzing each species data separately. Another thought, have you looked at the estimates from the phi(t)p(t) model? Especially the survival parameters? Your data may be insufficient for estimating these parameters, have a look at these for problems (e.g., large standard errors), poor estimates from this model may be the cause of the error you're getting in the random effects procedure (explains why the phi(.)p(t) model doesn't throw the error). -- andre

[cooch]

Andre is entirely correct (no surprise there). I'll add a few other comments.

Did you select "all of the [p and phi] parameters" except the last [p and phi] or "all of the [phi] parameters" except the last [phi]? I suspect that you did the latter. But If the former, then that's at least part of the problem -- choose only the phi parameters except the last phi.

Yes, this could very much lead to the problem. There are several examples in Appendix D where if you do this, you'll run into the problem as described. The basic advice is, do not include confounded parameters in RE models. Of course, this is not an issue if you're using a data type for which the fully time-dependent model doesn't have intrinsically confounded parameters.

If you're modeling both species, as groups, in the same model, then you'll want to (I suspect) select one species at a time otherwise you're suggesting that survival parameters for the two species came from the same distribution.

Correct. You can handle modeling each species separately using the DM approach.

Your data may be insufficient for estimating these parameters, have a look at these for problems (e.g., large standard errors), poor estimates from this model may be the cause of the error you're getting in the random effects procedure (explains why the phi(.)p(t) model doesn't throw the error). -- andre

For other issues, see example D.4.2 in Appendix D. Some problems with 'convergence' can be solved either by using simulated annealing, or 'better starting values'.