Random Effects Modeling with Estimation Problems

[JeffHostetler]

I am trying to estimate temporal process variance in survival in a Burnham modeling framework with 13 time steps. Unfortunately, the time-specific survival models run in MARK (with the RMark interface) fail to estimate all parameters; the “estimates” of the last two time steps’ survivals on the real scale are essentially 0 with 0 standard error and their betas on the logit scale are around -24, with either 0 or very high standard errors. I was wondering if this precludes using either the variance components or the MCMC approach in MARK to estimate process variance.

My gut and what playing around I’ve done with both suggest that the variance components approach may not be appropriate but that the MCMC approach might be. But my gut is often wrong! There would be some advantages to using the variance components approach for me if it’s appropriate here (including those from White et al.’s 2009 simulation chapter in Modeling Demographic Processes in Marked Populations). But getting reasonably reasonable estimates of process variance is the important thing.

Thanks,

Jeff

[cooch]

I am trying to estimate temporal process variance in survival in a Burnham modeling framework with 13 time steps. Unfortunately, the time-specific survival models run in MARK (with the RMark interface) fail to estimate all parameters; the “estimates” of the last two time steps’ survivals on the real scale are essentially 0 with 0 standard error and their betas on the logit scale are around -24, with either 0 or very high standard errors. I was wondering if this precludes using either the variance components or the MCMC approach in MARK to estimate process variance.

My gut and what playing around I’ve done with both suggest that the variance components approach may not be appropriate but that the MCMC approach might be. But my gut is often wrong! There would be some advantages to using the variance components approach for me if it’s appropriate here (including those from White et al.’s 2009 simulation chapter in Modeling Demographic Processes in Marked Populations). But getting reasonably reasonable estimates of process variance is the important thing.

Thanks,

Jeff

MCMC is unlikely to help you much, since it yields estimate of sigma which are basically identical to those from the RE approach (after appropriate transformation). More likely, you'll need to drop the last parameters from inclusion in the RE. For example, if you have 20 survival estimates (S(1) -> S(20), in a fully time-dependent model, the last estimate is confounded, so you'd need to constrain the RE for S(1)-> S(19) only.

Also, be sure to try good starting parameters, and run the SA optimzation routine.