[R-lang] question about model interpretation using lmer

[David Reitter]

On 23 Apr 2008, at 20:53, Schack Tang, Katie wrote:
>
> 1) Does the fact that voiced (vce) fails to reach significance on  
> its own indicate that its effect on F0 is not significantly  
> different from the effect of an implosive on F0?
>
Yes.  Your example suggests you're using the default contrasts.
> 2) What can I conclude about the sonorant (son) and voiceless  
> (vceless)?  Can I just conclude that they both raise F0 compared to  
> implosives?
>
Yes.
>   Or can I also conclude that voiceless raises F0 more than sonorant  
> does?
>
No.
> If the former, how can I test the latter--is it possible to specify  
> which level of a factor is withheld in lmer?
>

Normally, the first factor is withheld; manipulating the contrasts  
matrix will give you the right results.

Usually, calculating effect sizes and confidence intervals is very  
informative.  First, you will need to manipulate the contrasts:  F0 ~  
1/PrecSegment instead of  F0 ~ 1 + PrecSegment  (the "1" intercept is  
implicit) should give you a model without explicit intercept, but  
separate estimates for all PrecSegment levels.

To estimate confidence intervals, I've had comparatively good results  
with the pvals.fnc function (languageR package) and its Markov Chain  
Monte-Carlo sampling strategy.  I had to slightly patch this function  
to accept "glmer" rather than just "lmer" type models, and I also  
understand (p.c.) that recent versions do not yet work well with more  
than one random effect.  There are occasional problems with minimal  
(i.e. zero-sized) confidence intervals; manipulating the number of  
simulations (a parameter to the pvals.fnc function) can help.

As far as I know, the confidence intervals that are estimated this way  
are non-simultaneous.  Thus, fishing for effects without defining  
explicit hypotheses will have the usual consequences.  (If you have  
corpus data, the use of separate samples for each hypothesis should  
give you reliable p's and confidence intervals in that respect.)


--
David Reitter
ICCS/HCRC, Informatics, University of Edinburgh
http://www.david-reitter.com




-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 2450 bytes
Desc: not available
Url : http://pidgin.ucsd.edu/pipermail/r-lang/attachments/20080424/515db72a/attachment.bin