[R] Re: [S] lme vs. aov with Error term

[array chip]

Jose,

Thank you very much for the explanation! 

what about if I specify using "ML" instead of "REML"
in the lme? I found that I still got different answers
even I use "ML" in the lme call. And in any case,
should I trust "lme" more than "aov", or vice versa?

Thanks again


--- jose.pinheiro at pharma.novartis.com wrote:
> > I have a question about using "lme" and "aov" for
> the
> > following dataset. If I understand correctly,
> using
> > "aov" with an Error term in the formula is
> equivalent
> > to using "lme" with default settings, i.e. both
> assume
> > compound symmetry correlation structure.
> 
> Not necessarily, only when the maximum of the REML
> objective
> function falls within the parameter space. This
> obscure observation
> translates, in your case,  into saying that, when
> the "Subject"
> mean square error is larger than the "Within" MSE
> (in the aov
> notation), then the two fits should coincide. This
> is the case for
> your y response:
> 
> Error: subject
>           Df   Sum Sq  Mean Sq F value Pr(>F)
> . . .
> Residuals  4 0.285268 0.071317 
> 
> Error: Within
>                Df   Sum Sq  Mean Sq F value Pr(>F)
> . . .
> Residuals       8 0.231480 0.028935 
> 
> However, when the Subject MSE is smaller than the
> Within MSE, the
> maximum of the REML objective occurs at a negative
> value of the
> between-subject variance. Because the lme algorithm
> forces the
> estimated parameters to remain within the parameter
> space, the
> solution you got probably corresponds to a boundary
> of the parameter
> space (the Subject Std. Dev must be fairly small).
> This is what
> happens for your "x" response. 
> 
> Error: subject
>           Df   Sum Sq  Mean Sq F value  Pr(>F) 
> . . .
> Residuals  4 0.122304 0.030576 
> ---
> 
> 
> Error: Within
>                Df  Sum Sq Mean Sq F value Pr(>F)
> . . .
> Residuals       8 0.37469 0.04684 
> 
> Hope this helps.
> 
> Best,
> --José
> 
>
---------------------------------------------------------
> José Pinheiro
> Biostatistics, Novartis Pharmaceuticals 
> One Health Plaza, Room 419/2115, East Hanover, NJ
> 07936  
>
---------------------------------------------------------
>