While I am a big fan of the AR+RE method proposed by Littell, and have recommended it a lot, I suspect that is what is causing your problem. The first thing to try is to remove slice as a random effect. This is the equivalent of assuming that all subjects have a common intercept in the mixed model equations, which does not seem unreasonable from the smattering of data provided.
If that doesn't solve things, you may need to look at more flexible covariance structures than AR(1)--you may have data that makes that assumption about the errors run into difficulties. With that many time points, a spline structure may be able to give you what you need.
Steve Denham
... View more