Problems with "real" covariate parameter estimatio

[NMichel]

OK, you were so helpful before, now I'm back with another question! Now I'm dealing with two different datasets, and with each I'm attempting to model survival as a function of some climate parameters - mostly rainfall estimates.

The first dataset consists of only two models - one with transient Phi and fixed p, the other with a transient*time interaction for Phi and constant p. When I constrained the Phi(trans),p(constant) model with the rainfall covariates in either an additive or multiplicative fashion, the appropriate rainfall and rainfall*trans parameters were estimated. However, when I constrained the Phi(trans*time) model with the rainfall covariates (I tried straight additive as well as interactions with trans &/or time), the rainfall parameter(s) was(were) not estimated, and also many of the other parameters had very high Beta standard errors.

A similar thing occurred with the second dataset. My first set of models included Phi constrained by additive group effects, some of which included additive time effects, and time-dependent p. With these models, rainfall parameters were estimated without any problems. However, I later fixed the recapture probability to 1 or 0 in several years, and when I constrained these partially-fixed-p models with rainfall covariates, the rainfall parameters were unable to be estimated.

Any ideas? I'm guessing it has something to do with the fact that the rainfall parameters are acting as a surrogate for time-dependency and creating some sort of conflicts when you have time-dependency already included and/or have fixed some time-dependent parameters. Am I in the right ball park here? Is it possible to constrain time-dependent models or models with some fixed parameters with real covariates?

Thanks again!

Cheers,
Nicole Michel

[cooch]

Any ideas? I'm guessing it has something to do with the fact that the rainfall parameters are acting as a surrogate for time-dependency and creating some sort of conflicts when you have time-dependency already included and/or have fixed some time-dependent parameters. Am I in the right ball park here? Is it possible to constrain time-dependent models or models with some fixed parameters with real covariates?

Read (or re-read) section 8.1.1 of the same chapter. Example of constraining age models - examples shows a trend, but you could use 'real' covariates in exactly the same manner. Keep aware of the time-scale for the various age classes (as discussed).

The larger issue of constraining with real covariates (and everything else) is covered in some detail in Chapter 7.

If MARK is not yielding estimates, I'd hazardd a guess that your design matrix is incorrect in some fashion.