Re: Help: Non-positive semi-definite message
From: LSheiner <lewis@c255.ucsf.edu>
Subject: Re: Help: Non-positive semi-definite message
Date: Thu, 16 Aug 2001 09:57:44 -0700
Beware an inter-individual random effect variance estimate near zero:
It is almost never what it appears!
When a random effect variance is estimated to be zero and the model
used has a diagonal OMEGA, it usually means that the model is
over-parameterized in random effects and the data are insufficient to
estimate the number of distinct components of variance taht are
present.
I personally haven't seen this suddenly happen only
upon addition of a covariate, but it could.
A helpful diagnostic is to run the model with a full OMEGA (i.e., if the
number
of etas is 4, then use $OMEGA BLOCK(4)). On output, convert OMEGA to
its correlation
form (divide the off-diagonal values by the square root of the product
of the two corresponding diagonal elements ... thus the corr value
associated
with OMEGA(2,3) is OMEGA(2,3)/(OMEGA(2,2)*OMEGA(3,3))**.5). If one of
the off-diagonal terms in the correlation matrix in the same row/col as
the variance that went to zero on the diagonal OMEGA run is nearly
unity,
then over-parameterization in random effects is probably the reason for
the zero estimate,
not that the covariate has removed all variability ....
LBS.
--
_/ _/ _/_/ _/_/_/ _/_/_/ Lewis B Sheiner, MD (lewis@c255.ucsf.edu)
_/ _/ _/ _/_ _/_/ Professor: Lab. Med., Bioph. Sci., Med.
_/ _/ _/ _/ _/ Box 0626, UCSF, SF, CA, 94143-0626
_/_/ _/_/ _/_/_/ _/ 415-476-1965 (v), 415-476-2796 (fax)