Skip to content
Prev 158 / 20628 Next

negative variances

Negative variances, unless they can be explained as statistical error,
surely indicate that the model that is formulated expecting variances
to be positive is inappropriate.  My preference is to do what MLwiN
does, and allow the estimates to go negative, just so that the user is
alerted to problems with the model, or at least with its formulation.

The model that has negative "variance" components may from the
point of view of getting the variance-covariance structure correct be
OK, just formulated using parameters that, notwithstanding a fervent
wish to interpret them as variances, cannot be so interpreted.

I once heard, from people providing a data analysis service, of
receiving randomized block treatment comparison data where it
turned out that the blocks had been chosen, e.g. with their plots
at increasing distances from a stream.  A negative between block
component of "variance" was, given the design, to be expected.
I did myself once encounter a scientist who thought that blocks
should be chosen, as far as possible, so that each individual block
spanned as large a part of the variation as possible.

A problem with SEMs is that this stuff typically hides under the hood.
It is not obvious how the user might get to look under the hood
(at the layout of the blocks, maybe) and comment, "Ah, I might have
guessed as much."

Graphical models (the name is not as revealing as one might like
of the nature of these models) can be a huge advance because they
do try to pull apart what goes into SEMs, to look under the hood.
See the gR task view for R, that you can get to from a CRAN mirror.

John Maindonald             email: john.maindonald at anu.edu.au
phone : +61 2 (6125)3473    fax  : +61 2(6125)5549
Centre for Mathematics & Its Applications, Room 1194,
John Dedman Mathematical Sciences Building (Building 27)
Australian National University, Canberra ACT 0200.
On 12 Apr 2007, at 6:54 PM, Tu Yu-Kang wrote: