[R-lang] Re: p-values from pvals.fnc

[Jonathan Baron]

I have a question about this, which has been bugging me for a while.
It is very probably very naive, but here is a chance to ask it.

pvals.fnc() does not deal with random slopes because it complains
about "random correlation parameters".  I'm not quite sure what this
means, but I think it may mean that it wants to take into account the
correlation of the slopes and intercepts, but it can't.  Consider the
following:

l1 <- lmer(Y ~ X + (1+X|S)
l2 <- lmer(Y ~ X + (1|S) + (0+X|S)
l3 <- lmer(Y' ~ X' + (1|S) + (0+X'|S)

pvals.fnc() does not work with l1.  It will work with l2, but I think
l2 imposes the assumption of no correlation of slopes and intercepts.
This may be approximately true, and maybe you can tell by looking at
the output of l1, where it gives correlations in the random effects
part.  (It also gives them for the fixed effects, but I think that is
not the issue, although I'm not sure.)

If the assumption of no correlation isn't true, can't you transform X
and Y into X' and Y' (or maybe just transform one of them) by adding
or subtracting a constant?  I have tried various ways of doing this,
but none works for me except trial and error.  In many cases,
transformation by a constant would not change the substantive
inferences that we want to draw.

Jon

On 07/30/11 20:53, Jakke Tamminen wrote:
> Roger: Thanks for the information, I guess I have a lot of reading to do!
> 
> Alex and Roger: Looks like my version of lme4 is pretty old, 0.99875-6. If
> the more recent versions don't give you p-values for models with random
> slopes, should I be looking at them (the p-values) at all, or rely on the
> t-statistic (and probably update my packages!)?
> 
> Jakke
> 
> On 30 July 2011 18:45, Alex Fine <afine@bcs.rochester.edu> wrote:
> 
> > Jakke,
> >
> > I'm probably missing something, so I'm not replying-all.  How do you even
> > get pvals.fnc() to work with a model that has random slopes?  I have the
> > most up-to-date version of the languageR package and it won't take models
> > that have anything other than random intercepts.
> >
> > thanks,
> > Alex

-- 
Jonathan Baron, Professor of Psychology, University of Pennsylvania
Home page: http://www.sas.upenn.edu/~baron