Vicki Allison
2006-Nov-14 21:59 UTC
[R] Repeated measures by lme and aov give different results
I am analyzing data from an experiment with two factors: Carbon (+/-)
and O3 (+/-), with 4 replicates of each treatment, and 4 harvests over a
year. The treatments are assigned in a block design to individual
Rings.
I have approaches this as a repeated measures design. Fixed factors
are Carbon, O3 and Harvest, with Ring assigned as a random variable. I
have performed repeated measures analysis on this data set two different
ways: one utilizing lme (as described in Crawley, 2002), and the second
using aov (based on Baron and Li, 2006). Using lme I get very
conservative p-values, while aov gives me significant p-values,
consistent with those I obtain performing this analysis in SYSTAT. Can
anyone explain how these models differ, and which is more appropriate to
the experimental design I have described? The code I use, and the
output obtained follow:
1 lme model
library(nlme)
M5 <-lme(ln_tot_lgth ~ Carbon*O3*Harv., random = ~-1|Ring)
anova(M5, type="marginal")
# Output
numDF denDF F-value p-value
(Intercept) 1 44 176.59692 <.0001
Carbon 1 12 0.42187 0.5282
O3 1 12 0.06507 0.8030
Harv. 1 44 17.15861 0.0002
Carbon:O3 1 12 0.23747 0.6348
Carbon:Harv. 1 44 0.85829 0.3593
O3:Harv. 1 44 0.04524 0.8325
Carbon:O3:Harv. 1 44 0.05645 0.8133> plot(M5)
2 aov model
M6<-aov(ln_tot_lgth ~ O3*Harv.*Carbon + Error (Ring/Carbon+O3))
summary(M6)
plot(M6)
# Output
Error: Ring
Df Sum Sq Mean Sq F value Pr(>F)
O3 1 1.76999 1.76999 8.2645 0.01396 *
Carbon 1 0.64766 0.64766 3.0241 0.10760
O3:Carbon 1 0.15777 0.15777 0.7366 0.40756
Residuals 12 2.57002 0.21417
Error: Within
Df Sum Sq Mean Sq F value Pr(>F)
Harv. 1 33.541 33.541 84.0109 9.14e-12 ***
O3:Harv. 1 0.001 0.001 0.0036 0.9524
Harv.:Carbon 1 0.414 0.414 1.0362 0.3143
O3:Harv.:Carbon 1 0.020 0.020 0.0508 0.8226
Residuals 44 17.567 0.399
*** Note change of location***
Victoria Allison
Landcare Research
Private Bag 92170
Auckland 1142
New Zealand
Phone: +64 9 574 4164
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
WARNING: This email and any attachments may be confidential ...{{dropped}}
Spencer Graves
2006-Nov-16 21:14 UTC
[R] Repeated measures by lme and aov give different results
RSiteSearch("lme and aov") returned 350 hits for me just now. I'm
sure that many are not relevant to your question, but I believe some
are. Beyond this, there is now and R Wiki, accessible via
www.r-project.org -> Documentation: Wiki (or directly as
http://wiki.r-project.org/rwiki/doku.php?id=guides:lmer-tests&s=lme%20and%20aov).
The first hit in a search there for "lme and aov" "is an edited
transcript of a long thread in R-help starting Sept 7, 2006 from a
comment by Hank Stevens, with Douglas Bates as leading actor."
(http://wiki.r-project.org/rwiki/doku.php?id=guides:lmer-tests&s=lme%20and%20aov).
If that fails to answer your questions on this, please submit
another post. Please realize however that the expected number and
quality of replies is inversely proportional to some large power of the
length and complexity of your question.
Hope this helps.
Spencer Graves
Vicki Allison wrote:> I am analyzing data from an experiment with two factors: Carbon (+/-)
> and O3 (+/-), with 4 replicates of each treatment, and 4 harvests over a
> year. The treatments are assigned in a block design to individual
> Rings.
>
> I have approaches this as a repeated measures design. Fixed factors
> are Carbon, O3 and Harvest, with Ring assigned as a random variable. I
> have performed repeated measures analysis on this data set two different
> ways: one utilizing lme (as described in Crawley, 2002), and the second
> using aov (based on Baron and Li, 2006). Using lme I get very
> conservative p-values, while aov gives me significant p-values,
> consistent with those I obtain performing this analysis in SYSTAT. Can
> anyone explain how these models differ, and which is more appropriate to
> the experimental design I have described? The code I use, and the
> output obtained follow:
>
> 1 lme model
>
> library(nlme)
> M5 <-lme(ln_tot_lgth ~ Carbon*O3*Harv., random = ~-1|Ring)
> anova(M5, type="marginal")
>
> # Output
> numDF denDF F-value p-value
> (Intercept) 1 44 176.59692 <.0001
> Carbon 1 12 0.42187 0.5282
> O3 1 12 0.06507 0.8030
> Harv. 1 44 17.15861 0.0002
> Carbon:O3 1 12 0.23747 0.6348
> Carbon:Harv. 1 44 0.85829 0.3593
> O3:Harv. 1 44 0.04524 0.8325
> Carbon:O3:Harv. 1 44 0.05645 0.8133
>
>> plot(M5)
>>
>
>
> 2 aov model
>
> M6<-aov(ln_tot_lgth ~ O3*Harv.*Carbon + Error (Ring/Carbon+O3))
> summary(M6)
> plot(M6)
>
> # Output
> Error: Ring
> Df Sum Sq Mean Sq F value Pr(>F)
> O3 1 1.76999 1.76999 8.2645 0.01396 *
> Carbon 1 0.64766 0.64766 3.0241 0.10760
> O3:Carbon 1 0.15777 0.15777 0.7366 0.40756
> Residuals 12 2.57002 0.21417
>
> Error: Within
> Df Sum Sq Mean Sq F value Pr(>F)
> Harv. 1 33.541 33.541 84.0109 9.14e-12 ***
> O3:Harv. 1 0.001 0.001 0.0036 0.9524
> Harv.:Carbon 1 0.414 0.414 1.0362 0.3143
> O3:Harv.:Carbon 1 0.020 0.020 0.0508 0.8226
> Residuals 44 17.567 0.399
>
>
> *** Note change of location***
>
> Victoria Allison
> Landcare Research
> Private Bag 92170
> Auckland 1142
> New Zealand
> Phone: +64 9 574 4164
>
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> WARNING: This email and any attachments may be confidential ...{{dropped}}
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
Andrew Robinson
2006-Nov-16 21:43 UTC
[R] Repeated measures by lme and aov give different results
Vicki, take a look at P. 47 of the book by Pinheiro and Bates, if you have a copy (if not, get one!) They show the analysis of a split-plot design using lme there. Here, both aov and lme are estimating two levels of variation. The key difference is that with aov the Ring-level variation is not being included in the test of all the terms, whereas with lme it is being included. If you want to eliminate it from the tests, which will replicate the aov analysis, then you need to include it as a fixed effect. To pinpoint the question: in the simplest case of a randomized complete block design, we need to ask whether or not the block-level variation should be included in the error sum of squares. aov() excludes it, lme() includes it unless Block is included as a fixed effect. Whether or not you want to exclude it depends on what you think about the design and where the variation is coming from. The traditional RCB excludes it, but whether or not you should depends on the circumstances. Cheers Andrew On Wed, November 15, 2006 8:59 am, Vicki Allison wrote:> I am analyzing data from an experiment with two factors: Carbon (+/-) > and O3 (+/-), with 4 replicates of each treatment, and 4 harvests over a > year. The treatments are assigned in a block design to individual > Rings. > > I have approaches this as a repeated measures design. Fixed factors > are Carbon, O3 and Harvest, with Ring assigned as a random variable. I > have performed repeated measures analysis on this data set two different > ways: one utilizing lme (as described in Crawley, 2002), and the second > using aov (based on Baron and Li, 2006). Using lme I get very > conservative p-values, while aov gives me significant p-values, > consistent with those I obtain performing this analysis in SYSTAT. Can > anyone explain how these models differ, and which is more appropriate to > the experimental design I have described? The code I use, and the > output obtained follow: > > 1 lme model > > library(nlme) > M5 <-lme(ln_tot_lgth ~ Carbon*O3*Harv., random = ~-1|Ring) > anova(M5, type="marginal") > > # Output > numDF denDF F-value p-value > (Intercept) 1 44 176.59692 <.0001 > Carbon 1 12 0.42187 0.5282 > O3 1 12 0.06507 0.8030 > Harv. 1 44 17.15861 0.0002 > Carbon:O3 1 12 0.23747 0.6348 > Carbon:Harv. 1 44 0.85829 0.3593 > O3:Harv. 1 44 0.04524 0.8325 > Carbon:O3:Harv. 1 44 0.05645 0.8133 >> plot(M5) > > > 2 aov model > > M6<-aov(ln_tot_lgth ~ O3*Harv.*Carbon + Error (Ring/Carbon+O3)) > summary(M6) > plot(M6) > > # Output > Error: Ring > Df Sum Sq Mean Sq F value Pr(>F) > O3 1 1.76999 1.76999 8.2645 0.01396 * > Carbon 1 0.64766 0.64766 3.0241 0.10760 > O3:Carbon 1 0.15777 0.15777 0.7366 0.40756 > Residuals 12 2.57002 0.21417 > > Error: Within > Df Sum Sq Mean Sq F value Pr(>F) > Harv. 1 33.541 33.541 84.0109 9.14e-12 *** > O3:Harv. 1 0.001 0.001 0.0036 0.9524 > Harv.:Carbon 1 0.414 0.414 1.0362 0.3143 > O3:Harv.:Carbon 1 0.020 0.020 0.0508 0.8226 > Residuals 44 17.567 0.399 > > > *** Note change of location*** > > Victoria Allison > Landcare Research > Private Bag 92170 > Auckland 1142 > New Zealand > Phone: +64 9 574 4164 > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > WARNING: This email and any attachments may be confidential ...{{dropped}} > > ______________________________________________ > R-help at stat.math.ethz.ch mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide > http://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code. >Andrew Robinson Senior Lecturer in Statistics Tel: +61-3-8344-9763 Department of Mathematics and Statistics Fax: +61-3-8344 4599 University of Melbourne, VIC 3010 Australia Email: a.robinson at ms.unimelb.edu.au Website: http://www.ms.unimelb.edu.au