[R-sig-ME] lmer residual variance estimate with prior weights [SEC=UNCLASSIFIED]

[Ben Bolker]

Steve Candy <Steve.Candy at ...> writes:

> 
> Hi mixed-modellers
 
> I am fitting a simple linear mixed model to some abundance data
> which are mean densities on the log10 scale for a set of spatial
> cells with sample sizes used as prior weights. I define a random
> cell intercept model and fit a linear year trend. I get very similar
> estimates of the intercept and slope when using each of lmer(.) and
> asreml(.) but get much smaller estimates of the SEs of these
> parameters for lmer(.). This is due to a much smaller estimate of
> the residual standard error with estimate of 0.534 for lmer and
> 3.016 for asreml with corresponding estimates of the cell-level
> standard deviation of 0.0014 for lmer and 0.2151 (=0.04626^0.5) for
> asreml. Comparing these to a simple lm(.) fit gives a similar but
> slightly higher estimate of the residual standard error compared to
> the asreml estimate with the a value of 3.116. The lmer estimate
> appears to be orders of magnitude out. Am I interpreting these
> results correctly? Has it something to do with how the weighting is
> done!

[context snipped because gmane hates me]

I have a sneaking suspicion that lmer doesn't scale the sum of
the weights to 1 before doing calculations, as it arguably should.
What happens if you manually scale the weights to 1 (which seems
appropriate in this case)?

  Ben Bolker