A summary of the responses I received regarding the lack
of estimate of standard error within SAS when the
estimate of Covariance was zero within a Mixed Model.
Thanks to those who helped and forgive me for any
misinterpretation of your responses:
Main interpretation of such a result is that the covariance
estimate is zero and can fairly confidently be interpreted
thus.
The reason for there being no estimate of the standard
error given by SAS is to do with the variance being
bounded below by zero. As the estimate is on its bound,
the "gradient of the likelihood is nonzero and therefore the
standard error can not be estimated".
There are 2 caveats:
" 1. When the study was designed, blocks were defined
in the expectation
that this might yield a more powerful design. I would take
care if
there was prior evidence which suggested that the block
effect would be
important.
2. If the 'between-blocks' MS was substantially less than
the
'within-blocks' MS then this would suggest a problem with
the analysis.
This can happen if there is an important predictive
varaible which is
not included in the analysis and which is well balanced
over the
blocks. Ie it is evidence that you may have blocked on
the wrong
variable."
A potentially safer way of testing for the block effect is to
look at the -2 REML Log Likelihood value when the model
includes and does not include the block term. Look at the
difference in the 2 values and this has a Chi-squared
distribution.
Thanks again,
Roger
Roger Humphry
Epidemiology Unit,
SAC, VSD,
Stratherrick Rd,
Drummondhill,
Inverness,
IV2 4JZ
UK
Tel: 01463 243 030
Fax: 01463 711 103
|