[R-sig-ME] lme() vs aov()

[Federico Calboli]

Peter Dalgaard wrote:
> The gut reaction is that you shouldn't trust lme() in low-df cases, but 
> in this particular case the issue is different:
> 
>  > summary(mod.aov)
> 
> Error: source
>          Df Sum Sq Mean Sq F value   Pr(>F)  drug       2 61.167  
> 30.583  61.167 0.003703 **
> Residuals  3  1.500   0.500                   ---
> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
> 
> Error: Within
>          Df Sum Sq Mean Sq F value Pr(>F)
> Residuals  6    9.0     1.5             
> Notice that the Residuals Mean Sq is larger in the Within stratum than 
> in the source stratum. In terms of a mixed-effects model, this implies a 
> negative estimate for the variance of the source effect. lme() will have 
> nothing of that and sets it to zero instead. If you drop the 
> Error(source) you get the same F as in lme() although the df differ.
> 
> (The  "negative variance" can be interpreted as negative within-source 
> correlation, but that only works properly for balanced designs. Long 
> story...)

Thank you Peter for the explanation. I'm perfectly happy about this particular 
model, but I'd like to ask you (and everyone else who'd like to chime in), what 
do you mean with "you shouldn't trust lme() in low-df cases"? Why?

(I ask because I often have low-df analyses to do).

Regards,

Federico



-- 
Federico C. F. Calboli
Department of Epidemiology and Public Health
Imperial College, St Mary's Campus
Norfolk Place, London W2 1PG

Tel  +44 (0)20 7594 1602     Fax (+44) 020 7594 3193

f.calboli [.a.t] imperial.ac.uk
f.calboli [.a.t] gmail.com