[R-sig-ME] Extracting variances of the estimated variance components in lme4

[Joshua Wiley]

On Thu, May 3, 2012 at 1:22 PM, Ben Bolker <bbolker at gmail.com> wrote:
> Freedom Gumedze <Freedom.Gumedze at ...> writes:
>
>> Douglas and Thierry,
>>
>> Many thanks Douglas for the advice. I will look at the suggestion by
>> Douglas when the URL is visible.
>> The omission of the option for the standard errors of the estimated
>> variances (or std deviations) is understandable to avoid their 'abuse
>> e.g. in significance testing'. However, they should be available (if
>> needed) as they can be obtained from the inverse of information matrix
>> for the var. components.
>
>  It's not quite that easy, because the variance components are not
> estimated on the scale of variances or standard deviations, but on the
> Cholesky scale, so (depending on the model) the information matrix of
> the 'theta' parameter vector (a concatenated vector of the lower triangles
> of the Cholesky factors) may not be easy to translate to the
> information matrix of the standard deviations or variances.  I posted
> some code earlier in response to a query of Josh Wiley's, based on the
> development version of lme4 (sorry), that extracts the deviance function
> and wraps it in a function that transforms standard deviations to
> the Cholesky-factor parameterization -- combining this with finite-difference
> approximations of second derivatives (e.g. from the numDeriv package)
> will give the standard errors of the estimated parameters, if you want
> them.
>
>  I have the intention of including this stuff in a skull-and-crossbones-marked
> section of an "lme4-extras" vignette (if Doug lets me).  The vignette is
> in progress, I can send it on request.

I would certainly be interested in this (more to understand than
anything else).  I agree with not using the standard errors for
significance tests, besides, how cool are those profile plots in Doug
Bates slides?  Nice!  Cannot wait to play with that.


>  Ben Bolker
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models



-- 
Joshua Wiley
Ph.D. Student, Health Psychology
Programmer Analyst II, Statistical Consulting Group
University of California, Los Angeles
https://joshuawiley.com/