I'm trying to troubleshoot some unexpected behavior in my models. I'm running some multi-state models with a binary time-varying covariate (0 and 1 for two size classes). I understand for a time-varying individual covariate I need a value for every animal on every occasion, but I assume that is only applicable from the first capture forward. Is that correct for multi-state models?

For occasions prior to the first capture, I simply have provided a padded value because MARK does not want a dot or NA in that place. So for example, if I used 9 as a pad, the size covariate matrix for an individual first captured on the 2nd occasion of 4 might look something like this: 9 0 0 1 and that could have a corresponding capture history of 0101 or 0111 or 0100.

To make sure my padded values actually don't make any difference, I started playing around. Here's where things get weird. If I pad the covariate values with 0, 1, 9 and 99 to compare the differences, I get the same results for all examples using the standard optimization routine. However, if I use simulated annealing, I get different results, but only for data padded with 9 or 99. Are those values sneaking into the likelihood? I cannot figure out what is going on. What am I missing here?

https://photos.smugmug.com/Site-Pages/S ... 142450.png

Edited to add: I have checked this same dataset with a basic CJS model and I'm encountering the same problem in the same way. I have triple-checked my data also and cannot find any instances where I inadvertently have an observation in the capture history prior to a size estimate. I'm going to see if there's a published or example dataset I can test this on.