I am attempting to estimate abundance of Largemouth Bass sampled by shoreline electrofishing on multiple occasions. Closed-population model selection indicates that capture probabilities varied among occasions (model Mt); however, the closure assumption is questionable due to potential inshore-offshore movement. Therefore, I am also considering POPAN models. There is no evidence of substantial lack of fit of the global POPAN model Phi(t) p(t) pent(t). Models assuming constant Phi or pent seem reasonable because intervals between occasions are fairly consistent and the data are too sparse for the fully time-dependent model to be well supported. In addition to the global model, I am fitting model Phi(.) p(t) pent(.), a death-only model Phi(.) p(t) pent(0), a birth-only model Phi(1) p(t) pent(.), and a closed model Phi(1) p(t) pent(0). I specify constant pent by modifying the design matrix as explained in the MARK book and as I have done successfully with other data sets. The global model runs with no apparent problems. However, when I run model Phi(.) p(t) pent(.) with the default 2ndPart variance estimation, I receive the following message:

* * WARNING * * Divide by zero occurred during variance-covariance calculation of this model.

IEEE flag status at end of variance-covariance calculation:

overflow F

divide by zero T

invalid T

underflow F

inexact T

I have fitted models to comparable data sets for various lakes and years, and this result occurs consistently. Estimated Phi and pent are usually near 1 and 0, respectively, so I wonder if this has something to do with it. I don't receive this message when I run the death-only, birth-only, or closed POPAN models, and the parameter estimates from those models look good. The Phi(.) p(t) pent(.) model runs without error messages if I use the Hessian variance estimation, but then the variance appears to be underestimated. I have searched the MARK book, POPAN literature, and this forum and cannot figure out what's wrong. Am I missing something obvious? Thanks.