Skip to content
Prev 9133 / 20628 Next

glmmADMB random interaction effects and likelihood estimates

Seth W. Bigelow <seth at ...> writes:
[snip]
[snip]
My guess here is that the log-likelihood differs only by a tiny
amount, because the estimated variances for time2 and time3 are
about 7-8 orders of magnitude less than the intercept variance.
The basic issue is that in a data set of this size (not tiny, but
not huge and quite noisy), you just don't have much power to detect
differences in slope among plots.

  You got a bit lucky getting the models to fit at all; on my platform
(Ubuntu 10.04 32-bit) the models with a time:plot interaction (with
time either continuous or a factor) failed because the hessian was
non-positive-definite.

  You could compare logLik() for the two models to see if they differ
by some tiny amount, but it's also possible (because of the way
AD Model Builder returns its results) that the log-likelihoods will
have been rounded sufficiently that they appear identical.

  Ben Bolker