[R-sig-ME] Random effects variance estimates lme vs lmer

[David Duffy]

On Tue, 30 Nov 2010, Beth Holbrook wrote:

> I've been running some mixed models in lmer and lme and I've found that in some
> models lmer estimates essentially a zero variance for a random effect while lme
> estimates a variance greater than zero.  Everything else (parameter estimates,
> etc.) looks essentially the same.  Can someone help explain this? Is there
> something differently I should be doing in lmer?
>> mode1 <- lme(rxndist ~ lightclass, random=~1|trial, daphnia, method='ML')
>> mode2 <- lmer(rxndist ~ lightclass + (1|trial), daphnia, REML=FALSE)
>        (Intercept) Residual
> StdDev:   0.3537305 1.798685
>
> Groups   Name        Variance   Std.Dev.
> trial    (Intercept) 9.2927e-10 3.0484e-05
> Residual             3.3602e+00 1.8331e+00

You have specified the models correctly.  The lmer random effect 
estimate should be non-zero, so it has got stuck.  However, the MLE is
not greatly different from zero, so it is not practically important (the 
difference in likelihoods between the model with s2=0 and s2=0.125 is
only 0.38).  As with your earlier queries, the development version of
lme4 might have converged more closely.


-- 
| David Duffy (MBBS PhD)                                         ,-_|\
| email: davidD at qimr.edu.au  ph: INT+61+7+3362-0217 fax: -0101  /     *
| Epidemiology Unit, Queensland Institute of Medical Research   \_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029, Australia  GPG 4D0B994A v