Skip to content

newbie question on contrasts and aov

6 messages · Wolfgang Pauli, Brian Ripley, Salah Mahmud +2 more

#
I try to move from SPSS to R/S and am trying to reproduce the results of SPSS 
in R. I calculated a one-way anova with "spk" as experimental factor and erp 
as depended variable. 
The result of the Anova are the same concearning the mean square, F and p 
values. But I also wanted to caculate the contr.sdif(4) contrast on spk. The 
results are completely different now. I hope anybody can help me.

Thanks, Wolfgang

This is what I get in SPSS:
Tests of Within-Subjects Contrasts
Measure: MEASURE_1 
Source		SPKType III Sum of Squares	df	Mean Square	F	Sig.
SPK		Level 2 vs. Level 1	3,493	1	3,493	2,026	,178
			Level 3 vs. Previous	20,358	1	20,358	10,168	,007
			Level 4 vs. Previous	18,808	1	18,808	15,368	,002
Error(SPK)	Level 2 vs. Level 1	22,414	13	1,724			
			Level 3 vs. Previous	26,030	13	2,002			
			Level 4 vs. Previous	15,911	13	1,224			

This is the result in R:
Error: sub
          Df Sum Sq Mean Sq F value Pr(>F)
Residuals 13 205.79   15.83

Error: Within
          Df Sum Sq Mean Sq F value    Pr(>F)
spk        3 29.425   9.808  9.4467 8.055e-05 ***
spk: p   1  1.747   1.747  1.6821 0.2022649
spk: q   1 13.572  13.572 13.0719 0.0008479 ***
spk: r   1 14.106  14.106 13.5861 0.0006915 ***
Residuals 39 40.493   1.038
---
Signif. codes:  0 `***' 0.001 `**' 0.01 `*' 0.05 `.' 0.1 ` ' 1



Spk.df <- data.frame(sub,spk,erp)
subset(Spk.df, subset=(sub!="14oddball" & sub!="18odd" & sub!="19odd" & 
sub!="20oddball")) -> Spk.selected.df
contrasts(Spk.selected.df$spk) <- contr.sdif(4)
aov(erp ~ spk + Error(sub), data=Spk.selected.df) -> Spk.aov
summary(Spk.aov,data=Spk.selected.df,split=list(spk=list(p=1,q=2,r=3)))

this is the the beginning of the dataframe, which I use:
         sub  spk    erp
1  10oddball spk1  2.587
2  11oddball spk1 -0.335
3  12oddball spk1  5.564
5  15oddball spk1  0.691
6  17oddball spk1 -1.846
10 21oddball spk1  1.825
11 22oddball spk1  0.370
12  2oddball spk1  3.234
13  3oddball spk1  1.462
14  5oddball spk1  2.535
15  6oddball spk1  9.373
16  7oddball spk1  2.132
17  8oddball spk1 -0.518
18  9oddball spk1  2.450
19 10oddball spk2  2.909
20 11oddball spk2  0.708
21 12oddball spk2  4.684
23 15oddball spk2  3.599
...
#
Notice  `SPKType III Sum of Squares'.  I don't believe your contrasts are 
orthogonal, and R's are sequential sum of squares.

Also, are you sure these are the same contrasts?  I presume this is
contr.sdif from MASS (in which case it is churlish not to credit it), and
SPSS's contrasts look more like Helmert contrasts from their labelling.

Since it appears all your treatments are within subjects you do seem to be 
making life difficult for yourself. Although I would have done a simple 
fixed-effects analysis, applying summary.lm to the bottom stratum would 
give you simple t-tests for each contrast, including actual estimates of 
the magnitudes.
On Sun, 11 Jan 2004, Wolfgang Pauli wrote:

            

  
    
1 day later
#
Does anyone know of the existence of R code for estimating the survivor
function and its standard error where survival time (T) is defined as
the time between two interval-censored events, i.e., both events are
only known to occur within an interval? This is the situation that
commonly arises in longitudinal studies of viral infections. For
instance T could be the time of clearing HPV infection when the dates of
acquisition and loss (clearance) of infection are not measured exactly
but known to occur between two consecutive dates (dates of prescheduled
follow-up visits).

Many thanks,

Salah
#
Have you looked at library(survival)?  Unless I misunderstand what you 
want, it should be there.  Further documentation is provided in Venables 
and Ripley (2002) Modern Applied Statistics with W, Therneau and 
Grambsch (2000) Modeling Survival Data, Harrell (2001) Regression 
Modeling Strategies (all Springer), and in materials downloadable from 
www.r-project.org;  see especially search -> "R site search".  Have you 
tried these sources? 

      hope this helps.  spencer graves
Salah Mahmud wrote:

            
#
On Mon, 12 Jan 2004, Spencer Graves wrote:

            
It isn't.

	-thomas
Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle
16 days later
#
In the meantime I figured out that the Difference-contrast is not quite what I 
was looking for. But I still have two questions

1) Why do I get different results for Helmert contrasts in SPSS and R. I guess 
the contrast matrixes of Helmert are about the same in SPSS and R. I probably 
make a mistake as i am a newbie to R. I thought that it might be, because I 
have a repaeted measures design. That's why I put the Error(sub) in the 
formula of aov. 

2)
I tried to make my own contrast matrix, to compute comparisons between 
adjacent factor levels, i.e. 1-2, 2-3 and 3-4. My matrix looks like this:
     [,1] [,2] [,3]
[1,]    1    0    0
[2,]   -1    1    0
[3,]    0   -1    1
[4,]    0    0   -1

But then I get the same result as with contr.helmert(4). What is wrong I 
really don't get it!

Thank you,

Wolfgang Pauli
On Sunday January 11 2004 18:07, you wrote: