[R-sig-ME] aov p-value on nested models and issue with missing variables in lmer

Douglas Bates bates at stat.wisc.edu
Tue Mar 22 16:52:42 CET 2011

On Mon, Mar 21, 2011 at 10:10 AM, Divya Mehta <mehta at mpipsykl.mpg.de> wrote:
> Greetings to you all,
> I have read the long discussions regarding the lmer and why one is not
> provided with p-values but rather the t-values.
> Also, the discussion regarding the comparison of two nested lmer models
> using anova function.
> Since i am not a statistician, I would be very grateful for some expert
> advice regarding a problem that I am currently facing.
> I have compared two groups using a paired t-test and for the top hits I
> wanted to check if the p-values from the anova on two nested lmer models is
> comparable and I saw that there were not comparable at all (of course these
> are two different statistics!).
> Perhaps I did not frame the model correctly, here is what I did.
> lr1<- lmer(Y ~ timepoint   +(1|person), REML=F, data)
> lr2<- lmer(Y ~   (1|person), REML=F, data)
> anova(lr2,1r2)#
> I am interested in the effects of the fixed "timepoint" variable and
> "person" is the random factor since for each person i have two repeated
> measures in this case for the "Y".
> I removed all NAs and took only 45 paired samples so I assumed this would be
> relatively similar to the paired t-test.
> Also when I ran the analysis with one individual who did not have two
> repeated measures like all others, i got different values for the anova
> outputs and the power seemed lower than when i removed the one individual -
> this again surprised me a bit since i thought the whole idea of mixed models
> was the flexibility of having some missing data points?
> I am just wondering if I have done something silly :)

I don't think so but there is not really enough information to go on
here.   Can you provide a reproducible example, even if it is
simulated data, so that we can see exactly what you did and what the
results were?

My inclination is to suggest that if the data have a structure for
which a paired t-test would be appropriate then I would use that,
simply because it is easier to explain.

More information about the R-sig-mixed-models mailing list