[R] Model simplification using anova()

[Dieter Menne]

ChCh <jmo101 <at> student.canterbury.ac.nz> writes:

> 
> 
> Hello all,
> 
> I've become confused by the output produced by a call to
> anova(model1,model2). First a brief background.  My model used to predict
> final tree height is summarised here:
> 
>                             Df Sum Sq Mean Sq F value    Pr(>F)    
> Treatment              2   748.35   374.17 21.3096   7.123e-06 ***
> HeightInitial            1   0.31      0.31     0.0178    0.89519    
> DiameterInitial         1   0.52      0.52     0.0298    0.86460    
> Frost                     1   38.29    38.29    2.1807    0.15392    
> HeightInitial:Frost    1   85.83     85.83   4.8882    0.03774 *  
> DiameterInitial:Frost 1   97.90    97.90    5.5754    0.02749 *  
> Residuals                22 386.30   17.56                      
> ---
> 
> Based on this, I should not remove either of the interaction terms, so I
> turned my attention to the main factors. Based on p-values, I removed
> HeightInitial and used a call to anova(model1,model2) to see if this
> resulted in a weaker model.  Here is the output:
...

Though should never remove main factors in the presence of interactions. Check
Bill Venables penitential sermon which quite on top of google when you enter
"exegeses".

http://www.stats.ox.ac.uk/pub/MASS3/Exegeses.pdf

The usual way to simplify models in R is based on the AIC, not on p-values. See
stepAIC and the chapter in MASS.

Dieter