R help - Apr 2008 - lrm -interaction without main effect-error message

Dear all,

this might be not only an R-question but also a statistical.
When I do a logistic regression analysis (species distribution modeling) 
with function lrm (Design package) I get the follwoing error message:

 > tadl1<-lrm(triad~fd+dista+fd2+dista2+fd:dista+dista:geo2, x=T, y=T)

Error in if (!length(fname) || !any(fname == zname)) { :

  missing value where TRUE/FALSE needed


The problem seems to be that geo2 (factor variable with 3 levels) is not 
included as main effect. But when I run the same model with glm it is 
working properly.
However, from an ecological point of view, inclusion of only the 
interaction term makes sense. When running the model with inclusion of 
both main effect and interaction, main effect has no significant 
influence and the interaction only marginaly. And LR-Test underlines 
model simplification.
Does anyone know how to solve the problem? I need the lrm function since 
I have to validate my models via bootstrapping (validate.lrm).

Many thanks!
Eva

-- 
Eva Mosner
Lilienstra?e 9

35043 Marburg

06421-8400886
0151-17867488

Eva Mosner wrote:
> Dear all,
> 
> this might be not only an R-question but also a statistical.
> When I do a logistic regression analysis (species distribution modeling) 
> with function lrm (Design package) I get the follwoing error message:
> 
>  > tadl1<-lrm(triad~fd+dista+fd2+dista2+fd:dista+dista:geo2, x=T,
y=T)
> 
> Error in if (!length(fname) || !any(fname == zname)) { :
> 
>   missing value where TRUE/FALSE needed
> 
> 
> The problem seems to be that geo2 (factor variable with 3 levels) is not 
> included as main effect. But when I run the same model with glm it is 
> working properly.
> However, from an ecological point of view, inclusion of only the 
> interaction term makes sense. When running the model with inclusion of 
> both main effect and interaction, main effect has no significant 
> influence and the interaction only marginaly. And LR-Test underlines 
> model simplification.
> Does anyone know how to solve the problem? I need the lrm function since 
> I have to validate my models via bootstrapping (validate.lrm).
> 
> Many thanks!
> Eva
> 
No!  The test of a 'main effect' that you did is not a valid test and it
invalidates the hierarchy principle.  Don't get lulled into thinking 
that parsimony is a good thing.  Besides getting strange fits you will 
not preserve type I error or confidence interval coverage.  If you were 
doing ols you would be getting an invalid estimate of sigma.

Model simplification is warranted if you tested an appropriate group of 
parameters with a test that has a large number of degrees of freedom. 
For example, you might argue that ALL interaction terms could be dropped 
if the P-value for the combined effects of all interaction parameters is 
0.3.  You might argue that one predictor could be dropped if the 
combined effects of all main effects and interaction effects containing 
the predictor gives a P-value of 0.25.  Both of these tests also respect 
the hierarchy principle.

Frank


-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University