Scrive Birgit Mair <mair@slcmsr.org>:
> Hallo s-news,
>
> I used lme to fit a linear mixed effects model as follows:
> lme(y~ time + x2 + x3 + x4, random= list(center=~1, pid=~time, data=work)
>
> My qqnorm-plot is heavy tailed, because of a lot of outliers, so the
> residuals are not normally distributed. In Pinheiro and Bates (2000) I
> read that this would lead to conservative tests. One of my fixed effects
> has an p-value of 0.059 in the summary t-test, thus I have to exclude it
> from the model. Does anybody know, if there is a better test with "true"
> p-values or how to deal with this data?
> Thank you in advance
>
> Birgit
Birgit,
I do not know what kind of independent variables you are using, but log
transforming them (or some other transformation) may help to attain normality
of error terms...
Following Pinhero and Bates, the summary t-tests are not the most robust way to
test for the significance of fixed terms. You should rather try F-tests, which
can be easily done with the anova.lme object
hope this helps!
good luck,
Achaz von Hardenberg
Alpine Wildlife Research Centre
Parco Nazionale Gran Paradiso
via della Rocca 47
10123 Torino, Italy
achaz.von.hardenberg@usherbrooke.ca
Tel. (office): +39.011.8606212
Tel. (mobile): +39.328.8736291
Fax: +39.011.8121305
Open access to all papers published in the Journal of Mountain Ecology:
http://www.mountainecology.org
GSE-AIESG (Gruppo Stambecco Europa - Alpine Ibex European Specialist Group):
http://www.gseonline.org
|