[R] ERROR NaNs produced; when comparing two logistic regression models with the ANOVA CHI test

[Prof Brian Ripley]

You need the models to be NESTED for this test to be valid.  Look up 
`analysis of deviance' in your textbooks.

You get the error because you have two unnested models with the same 
number of parameters, so the F numerator is divided by 0. (But this is 
merely a symptom of incorrect usage.)

On Tue, 22 Feb 2005, Jan Verbesselt wrote:

> Dear R-list,
>
> *When comparing two logistic regression models with the anova CHi test, I
> obtain the following error: (there are no NA's in the time series). How can
> this be solved such that I can compare two models on the same dataset were
> different explanatory variables are used?
>
>
>    l.KBDI <- glm(zna.arson2 ~ zna.KBDI,family = binomial)
>    l.NDWI <- glm(zna.arson2 ~ zna.NDWI,family = binomial)
>
> anova(l.KBDI,l.NDWI, test = "Chi")
>
>
> Analysis of Deviance Table
>
> Model 1: zna.arson2 ~ zna.KBDI
> Model 2: zna.arson2 ~ zna.NDWI
>  Resid. Df Resid. Dev  Df Deviance P(>|Chi|)
> 1       110    123.227
> 2       110    113.593   0    9.635
> Warning message:
> NaNs produced in: pchisq(q, df, lower.tail, log.p)
>
>
> *Is this caused by the non-linearity of the data?
> *How can I obtain a p-value?

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595