[R] Trouble with mixed model anova (incl. random block)
Philippe.Hensel at noaa.gov
Wed Nov 17 23:28:29 CET 2010
Dear R community,
I have slightly unbalanced survey/land elevation data (over 7,000 data
points, grouped according to 60 transects, 30 in each of two
treatments). Observations in both treatments are taken simultaneously,
and the surveying of all 60 transects took 22 days. Comparisons among
treatments have to be restricted to the same days the two treatments
I think the following model would be appropriate (at least at a first cut):
Day (Random block)
Day x Treatment (Random, error for Treatment)
Transect (Day, Treatment) (Random)
Point (Transect, Day, Treatment) (Random - residual error)
What are good candidates for algorithms to test this nested error
anova? I have been trying lme() and lmer(), but I am confused as to
how I should code for the nested error structure. I have tried:
lme(Elevation ~ Treatment, random = ~Day | Transect, data)
lmer(Elevation ~ Treatment + (1|Day / Transect), data)
I am not yet very familiar with the R output, so it is hard for me to
figure out if I have written the model statements correctly.
How do I code for Day x Transect to make sure that it is being used as
the error term for evaluating the fixed Treatment effect?
How do I code for Transect(Day, Treatment)?
In the anova output for the lme function, I get a highly significant
intercept, but a non-significant Treatment effect (p=0.34).
In the anova output for the lmer function, I get a significant Treatment
effect (p = 0.03). How can I reconcile these two?
The log-likelihoods are similar: 15183 (lme) and 15189 (lmer). Values
of AIC and BIC are also similar between the two models.
Before I go any further and compare to an intercept-only model (no
Treatment effects), I wanted to make sure I chose the correct path and
Thank you in advance for any assistance.
Philippe Hensel, PhD
NOAA National Geodetic Survey
1315 East-West Hwy.
Silver Spring MD 20910
(301) 713 3198 ext. 137
More information about the R-help