Skip to content
Prev 164 / 20628 Next

negative variances

On 4/12/07, Tu Yu-Kang <yukangtu at hotmail.com> wrote:
Both of those optimization paths show demonstrate convergence to a
singular variance-covariance matrix.  The first two parameters being
optimized are on the scale of variances (lmer) or standard deviations
(lmer2) and are constrained to be non-negative.  In the case of lmer
they need to be constrained to be slightly positive (>= 1e-10).  You
can see that both case have converged on the boundary although lmer2
achieved a better minimum (lower deviance).

The first parameter represents the relative variance (or relative
standard deviation for lmer2) of the intercept random effect.  The
second parameter is a relative variance (standard deviation) of a
linear combination of the random effects.  The third parameter
determines the linear combination.

To me this indicates model failure.  There are finite variances for
both the slope and intercept random effects but they  are perfectly
correlated.