[R] imposing constraints on the covariance matrix of random effects in lme4?

[Douglas Bates]

On 7/12/07, JVVerkuilen <jvverkuilen at gmail.com> wrote:
> Hello all,

> I am using lme4 to fit some mixed logistic regressions. I need to
> impose an identification constraint of the following form:

> (1        sig12)
> (sig12  sig22)

> and have not figured out how to do it, i.e., sig11 = 1 but the rest of
> the parameters are free to vary. Is this possible and, if so, how?

> I've been looking through the archive and help to no avail, but
> perhaps I'm just missing something.

Current versions of the lme4 package do not allow such a constraint to
be imposed.  It would not be overwhelmingly difficult to do so because
the value that you wish to constrain appears explicitly as one of the
parameters in the optimization of (the approximation to) the
log-likelihood.  However, doing this would involve writing writing
custom versions of some of the functions coded in C. If you are
willing to undertake this contact me off-list and I will describe
exactly what would need to be done.