Hello mixed-modelers:
I am confused about detection of differences among fixed effect estimates in
mixed models. I am using lme() to compare the effect of the factor Exam (3
levels: P0,P1,P5) on the continuous variable BA. Measurements were done on
64 Plots. This is modeled with the statement m0 <- BA~Exam, data.,
random=~1|plot. Summary(m0) tells me that Exam has a highly significant
effect, and that both P1 and P5 differ highly significantly from P0, but not
whether P1 and P5 differ. If I use a contrast statement ("library(contrast),
contrast(m0, list(Exam='P1'), list(Exam='P5'), the contrast is not
significant (p=0.16). However, if I just use a paired t-test, the
difference is significant at p=0.001. I also get significant results with
lme if I just delete the 'P0' level. I would be grateful if anyone could
point me in the right direction for making a robust (yet sensitive)
determination of whether these factor levels differ. I possess a copy of the
Ur-text for R mixed models (i.e., Pinheiro & Bates) but it has not yet
provided insight. Appreciatively, Seth W. Bigelow
Linear mixed-effects model fit by REML
Data: s2
AIC BIC logLik
1746.913 1763.122 -868.4565
Random effects:
Formula: ~1 | Plot
(Intercept) Residual
StdDev: 25.779 16.18827
Fixed effects: BA ~ Exam
Value Std.Error DF t-value p-value
(Intercept) 72.01422 3.805048 126 18.925968 0e+00
ExamP1 -15.01266 2.861708 126 -5.246047 0e+00
ExamP5 -10.99922 2.861708 126 -3.843585 2e-04
Correlation:
(Intr) ExamP1
ExamP1 -0.376
ExamP5 -0.376 0.500
Standardized Within-Group Residuals:
Min Q1 Med Q3 Max
-1.74478652 -0.54972545 0.02464379 0.41020121 3.24236603
Number of Observations: 192
Number of Groups: 64
CONTRAST statement
lme model parameter contrast
Contrast S.E. Lower Upper t df Pr(>|t|)
1 -4.013438 2.861708 -9.622282 1.595407 -1.4 187 0.1624
[[alternative HTML version deleted]]