[R] aov contrasts residual error calculation
spencer.graves at pdf.com
Fri Apr 21 18:32:16 CEST 2006
I'm on the opposite extreme: I don't know aov, and given that it is
largely obsolete, I'm not too eager to learn it.
Steven Lacey wrote:
> Thanks for the reply. I am not using lme because I don’t have the time to
> understand how it works and I have a balanced design, so typcial linear
> modelling in aov should be sufficient for my purposes. Down the road I plan
> to learn lme, but I'm not there yet. So any suggestions with respect to aov
> would be greatly appreciated.
> -----Original Message-----
> From: Jacques Veslot [mailto:jacques.veslot at good.ibl.fr]
> Sent: Friday, April 21, 2006 11:58 AM
> To: Steven Lacey
> Cc: r-help at stat.math.ethz.ch
> Subject: Re: [R] aov contrasts residual error calculation
> why not using lme() ?
> first, you need transform data:
> dat2 <- as.data.frame(lapply(subset(dat, sel=-c(A,B,C)), rep, 3))
> dat2$y <- unlist(subset(dat, sel=c(A,B,C)), F, F)
> dat2$cond <- factor(rep(c("A","B","C"), each=nrow(dat)))
> dat2$inter <- factor(dat2$map):factor(dat2$cond)
> lme1 <- lme(fixed = y ~ mapping + cond + inter + other fixed effects,
> random = ~ 1 |subj, data=dat2,
> Steven Lacey a écrit :
>>I am using aov with an Error component to model some repeated measures
>>data. By repeated measures I mean the data look something like this...
>>subj A B C
>>1 4 11 15
>>2 3 12 17
>>3 5 9 14
>>4 6 10 18
>>For each subject I have 3 observations, one in each of three
>>conditions (A, B, C). I want to test the following contrast (1, 0,
>>-1). One solution is to apply the contrast weights at the subject
>>level explicitly and then call t.test on the difference scores.
>>However, I am looking for a more robust solution as I my actual design
>>has more within-subjects factors and one or more between subjects
>>A better solution is to specify the contrast in an argument to aov.
>>The estimated difference of the contrast is the same as that in the
>>paired t-test, but the residual df are double. While not what I
>>expected, it follows from the documentation, which explicitly states
>>that these contrasts are not to be used for any error term. Even
>>though I specify 1 contrast, there are 2 df for a 3 level factor, and
>>I suspect internally the error term is calculated by pooling across
>>While very useful, I am wondering if there is way to get aov to
>>calculate the residual error term only based on the specified
>>contrasts (i.e., not assume homogeneity of variance and sphericity)
>>for that strata?
>>If not, I could calculate them directly using model.matrix, but I've
>>never done that. If that is the preferred solution, I'd also
>>appreciate coding suggestions to do it efficiently.
>>How would I do the same thing with a two factor anova where one factor
>>is within-subjects and one is between...
>>Mapping Subject A B C
>>1 1 4 11 15
>>Mapping is a between-subject factor. Condition is a within-subject
>>factor. There are 5 levels of mapping, 8 subjects nested in each level
>>of mapping. For each of the 40 combinations of mapping and subject
>>there are 3 observations, one in each level of the condition factor.
>>I want to estimate the pooled error associated with the following set
>>of 4 orthogonal contrasts:
>>What is the best way to do this? One way is to estimate the linear
>>contrast for condition for each subject, create a 40 row matrix where
>>the measure for each combination of mapping and subject is the linear
>>contrast on condition. If I pass this dataframe to aov, the mse it
>>returns is the value I am looking for.
>>If possible, I would like to obtain the estimate without collapsing
>>the dataframe, but am not sure how to proceed. Suggestions?
>>R-help at stat.math.ethz.ch mailing list
>>PLEASE do read the posting guide!
More information about the R-help