R help - Mar 2006 - Odd anova(lm()) order phenomenon, looking for an explanation

Hi everyone,

I'm witnessing an odd modelling phenomenon that I can't explain.  If
anyone has seen this before, or can explain what's going on would let
me know, I'd be very grateful!  Especially if I'm just being dim.

I'm fitting a pair of continuous variates and their interaction to
some residuals from another model.  The sequential anova statement
changes with the term order; that's fine.  But each term explains a
much larger Sum Sq when it is listed second than when it is listed
first.


> anova(lm(residuals(delta.point.lm.0) ~ canopy.h + canopy.d,
+          data=snow))
Analysis of Variance Table

Response: residuals(delta.point.lm.0)
           Df Sum Sq Mean Sq F value    Pr(>F)    
canopy.h    1  156.2   156.2  11.118 0.0009613 ***
canopy.d    1  198.0   198.0  14.098 0.0002080 ***
Residuals 303 4256.6    14.0                      
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1
> anova(lm(residuals(delta.point.lm.0) ~ canopy.d + canopy.h,
+          data=snow))
Analysis of Variance Table

Response: residuals(delta.point.lm.0)
           Df Sum Sq Mean Sq F value    Pr(>F)    
canopy.d    1    0.4     0.4  0.0284    0.8664    
canopy.h    1  353.8   353.8 25.1871 8.887e-07 ***
Residuals 303 4256.6    14.0                      
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1 
> 


I would have expected any term to explain less Sum Sq if listed second
than if listed first.  Is my intuition awry?  Does anyone have any
modelling insight to help me interpret what I'm seeing?

Cheers

Andrew
-- 
Andrew Robinson  
Department of Mathematics and Statistics            Tel: +61-3-8344-9763
University of Melbourne, VIC 3010 Australia         Fax: +61-3-8344-4599
Email: a.robinson at ms.unimelb.edu.au         http://www.ms.unimelb.edu.au

G'day Andrew,
>>>>> "AR" == Andrew Robinson <A.Robinson at
ms.unimelb.edu.au> writes:
    AR> I would have expected any term to explain less Sum Sq if
    AR> listed second than if listed first.  Is my intuition awry?
Yes.  :-)

I would not expect that *any* term explains less Sum Sq if listed
second, then life and (linear) modelling would be simple.  The problem
with multiple regression is that a covariate might look unimportant if
used first (i.e. has small Sum Sq associated with it in the anova
table), but if we first correct for other regressor, then this
covariate becomes important all of a sudden (i.e. has large Sum Sq
associated with it in the anova table).

What surprised me, was that you observed this phenomenon with respect
to both regressors.  If only one had displayed this behaviour, I would
have readily explained it as above, but that both display it, I found
surprising too.

    AR> Does anyone have any modelling insight to help me interpret
    AR> what I'm seeing?
Don't know if the following example, which shows the same behaviour,
leads to any insight. 
> n <- 100
> x1 <- runif(n, -1,1)
> x2 <- runif(n, -1,1)
> y <- x1*x1*x2 + rnorm(n, sd=0.05)
> y <- y - mean(y)
> anova(lm(y~x1+x2))
Analysis of Variance Table

Response: y
          Df Sum Sq Mean Sq  F value Pr(>F)    
x1         1 0.0055  0.0055   0.1485 0.7008    
x2         1 5.0071  5.0071 134.3499 <2e-16 ***
Residuals 97 3.6151  0.0373                    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1 
> anova(lm(y~x2+x1))
Analysis of Variance Table

Response: y
          Df Sum Sq Mean Sq  F value Pr(>F)    
x2         1 4.9930  4.9930 133.9723 <2e-16 ***
x1         1 0.0196  0.0196   0.5261   0.47    
Residuals 97 3.6151  0.0373                    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1

Cheers,

        Berwin

========================== Full address ===========================Berwin A
Turlach                      Tel.: +61 (8) 6488 3338 (secr)
School of Mathematics and Statistics        +61 (8) 6488 3383 (self)      
The University of Western Australia   FAX : +61 (8) 6488 1028
35 Stirling Highway                   
Crawley WA 6009                e-mail: berwin at maths.uwa.edu.au
Australia                        http://www.maths.uwa.edu.au/~berwin