[R-sig-ME] lmer nonconvergent: care to run and explain?

[Douglas Bates]

I did do the follow-up fit omitting the (1|Mind) term and got the same
results and the same warning.  Neither of those warnings make sense to me.
They are based on evaluations of an approximate Hessian formed (I think) by
finite differences of gradients evaluated by finite difference and those
are very noisy estimates.  In this case the estimated Hessian is diagonal
for the full model and 1 by 1 for the reduced model.  It should not be
problematic to have a large eigenvalue.  All that means is that small
differences in the second parameter (the only parameter in the reduced
model) cause large changes in the objective function, which is okay.

	[[alternative HTML version deleted]]