[R] manova question

[Ranjan Maitra]

Dear John, Peter and others,

So, I now have a query at an even more elementary level and that is
regarding my results from anova.mlm() not matching the car package's
Manova(). Specifically, I have been trying the following out with regard
to a simple one-way MANOVA setup. So, I try out the following using R:

******* R code *******

morel <- read.table(file = "http://www.public.iastate.edu/~maitra/stat501/datasets/morel.dat",  col.names = c("studentgroup", "aptitude", "mathematics", "language", "generalknowledge"))

morel[,1] <- as.factor(morel[,1])
fit <- anova.mlm(as.matrix(morel[,-1]) ~ morel[,1])

summary(fit, test="Wilks")



*** providing the output ***


           Df   Wilks approx F num Df den Df    Pr(>F)    
morel[, 1]  2 0.54345   6.7736      8    152 1.384e-07 ***
Residuals  79               

*** end of output
                 
             

The above is correct, also by doing the calculations "by hand".

Then, I use the car package, following the help function on Anova()
and do the following:


******* R code ********

morel <- read.table(file = "http://www.public.iastate.edu/~maitra/stat501/datasets/morel.dat",  col.names=c("studentgroup", "aptitude", "mathematics", "language", "generalknowledge"))

library(car)
fit1 <- Manova( lm( cbind(aptitude, mathematics, language, generalknowledge) ~ studentgroup , data = morel))
summary(fit1, test = "Wilks")


****** providing the output *****


Type II MANOVA Tests:

Sum of squares and products for error:
                 aptitude mathematics   language generalknowledge
aptitude         78506.68  13976.5677 11041.9434        4330.1304
mathematics      13976.57  16040.3996  3979.9528        -416.4845
language         11041.94   3979.9528  6035.6132        -372.8491
generalknowledge  4330.13   -416.4845  -372.8491        7097.9562

------------------------------------------

Term: studentgroup 

Sum of squares and products for the hypothesis:
                  aptitude mathematics   language generalknowledge
aptitude         1129.7271    996.0542  237.54441        -880.4353
mathematics       996.0542    878.1980  209.43741        -776.2594
language          237.5444    209.4374   49.94777        -185.1266
generalknowledge -880.4353   -776.2594 -185.12655         686.1536

Multivariate Test: studentgroup
             Df test stat approx F num Df den Df   Pr(>F)  
studentgroup  1 0.8620544 3.080378      4     77 0.020805 *
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 


***** end of output.



Which is very different from the previous results. So what am I doing wrong here? Same issues arise with the other tests also (Pillai, Roy, Hotelling-Lawley, etc).

Many thanks and best wishes,
Ranjan












On Sun, 20 Mar 2011 19:29:41 -0500 John Fox <jfox at mcmaster.ca> wrote:

> Dear Peter and Ranjan,
> 
> In addition to Anova(), linearHypothesis() in the car package handles
> multivariate linear models, including those for repeated measures.
> 
> Best,
>  John
> 
> --------------------------------
> John Fox
> Senator William McMaster
>   Professor of Social Statistics
> Department of Sociology
> McMaster University
> Hamilton, Ontario, Canada
> http://socserv.mcmaster.ca/jfox
> 
> 
> 
> 
> > -----Original Message-----
> > From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org]
> > On Behalf Of peter dalgaard
> > Sent: March-20-11 6:50 PM
> > To: Ranjan Maitra
> > Cc: R-help
> > Subject: Re: [R] manova question
> > 
> > 
> > On Mar 20, 2011, at 21:05 , Ranjan Maitra wrote:
> > 
> > > Dear friends,
> > >
> > > Sorry for this somewhat generically titled posting but I had a
> > > question with using contrasts in a manova context. So here is my
> > question:
> > >
> > > Suppose I am interested in doing inference on \beta in the case of the
> > > model given by:
> > >
> > > Y = X %*% \beta + e
> > >
> > > where Y is a n x p matrix of observations, X is a n x m design matrix,
> > > \beta is m x p matrix of parameters, and e is a normally-distributed
> > > random matrix with mean zero and independent rows, each having
> > > dispersion matrix given by \Sigma. Then, I know (I think) how to
> > > perform MANOVA. Specifically, I use:
> > >
> > > fit <- manova(Y ~ X)
> > >
> > > and
> > >
> > > summary(fit) will allow me to perform appropriate inference on beta.
> > >
> > > Now, suppose I am interested in doing inference on C %*% \beta %*% M
> > > (say testing whether this is equal to zero) with C and M being q x m
> > > and p x r matrices, respectively (with q, r both being no more than
> > > p), then can this be done using the manova object from the above? How?
> > > If not, is there an efficient way to do this?
> > 
> > Check out anova.mlm(), it does most of this sort of testing. Not quite
> > the "C %*% ..." bit because the linear model code is not really built to
> > handle linear constraints, but rather compare nested models, each
> > specified using a set of betas. (So you usually test whether a subset of
> > betas is zero).
> > 
> > Also check out the "car" package. Its Anova() function does some similar
> > stuff.
> > 
> > If noone has done so already, I wouldn't think it to be very hard to
> > implement the general case. Most of the bits are there already.
> > 
> > --
> > Peter Dalgaard
> > Center for Statistics, Copenhagen Business School Solbjerg Plads 3, 2000
> > Frederiksberg, Denmark
> > Phone: (+45)38153501
> > Email: pd.mes at cbs.dk  Priv: PDalgd at gmail.com
> > 
> > ______________________________________________
> > R-help at r-project.org 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.
>