[R] Variance Component/ICC Confidence Intervals via Bootstrap or Jackknife

[Rick Bilonick]

On Sun, 2006-10-29 at 11:06 -0800, Spencer Graves wrote:
>       I can think of two ways to get confidence intervals on intraclass 
> correlations (ICCs) and more accurate intervals for variance 
> components:  (1) modifying 'simulate.lme' to store the estimated 
> variance components as well as "logLik" and (2) using 'lmer' and 
> 'mcmcsamp' in library(lme4). 
> 
>       The difficulty with (1) is that you have to make a local copy of 
> 'simulate.lme', then figure out where and how to modify it.  I've just 
> looked at the code, and it looks like the required modifications should 
> be fairly straightforward.  The problem with (2) is that you would have 
> to learn the syntax for a nested model for 'lmer'.  It's different from 
> that for 'lme' but not difficult.  The 'lmer' function is newer and 
> better in many ways but is not as well documented and does not have as 
> many helper functions yet.  The best documentation available for it may 
> be the 'MlmSoftRev' vignette in the 'mlmRev' package plus the "R News 
> 5/1" article from May 2005.  If you are not familiar with vignettes, 
> RSiteSearch("graves vignette") produced 74 hits for me just now.  Find 
> 'vignette' in the first hit led me to an earlier description of vignettes. 
> 
>       If it were my problem, I'd probably try the second, though I might 
> try both and compare.  If you try them both, I'd be interested in the 
> comparison.  If the answers were substantially different, I'd worry. 
> 
>       Hope this helps. 
>       Spencer Graves
> 

I'm familiar with making nested models in lmer but I'm not familiar with
mcmcsamp but will check it out. Thanks.

I used nlme/lme because it has the intervals function while (at least
the last time I checked), lme4/lmer did not.

The way I've done the bootstrapping (sampling at each level) sounds the
same as using a simulation. But articles and references I've found
indicate that only the highest level (a if c is nested in b and b is
nested a) should be sampled.

Rick B.



-- 
Richard A. Bilonick, Ph.D.
Assistant Professor
University of Pittsburgh School of Medicine
Department of Ophthalmology
412 648 9138
BST S 207