[R] When models and anova(model) disagree...

[Greg Snow]

The individual tests on coefficients in logistic regression are generally based on a Wald test statistic.  Unfortunately there is a bit of a paradox possible in this case where the coefficient is highly significant, but due to a flattening of the likelihood the standard error is overestimated and the p-value ends up non-significant.  The anova function uses the likelihood ratio test which is not affected by this and is the more trustworthy statistic to use.

I assume that you are using the lrm function from the rms package, the book that that package goes along with gives more detail (including the name of the paradox which I don't remember at the moment (and my copy of the book is currently 40 miles away)).

-----Original Message-----
From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org] On Behalf Of Rob James
Sent: Wednesday, June 15, 2011 4:18 PM
To: r-help at r-project.org
Subject: [R] When models and anova(model) disagree...


I have a situation where the parameter estimates from lrm identify a 
binary predictor variable ("X") as clearly non-significant (p>0.3), but 
the ANOVA of that same model gives X  a chi^2-df rank of > 200, and 
adjudicates X  and one interaction of X and a continuous measure as 
highly significant.  The N is massive and X has two categories, each 
with > 100,000 observations.  I would expect X to have a significant 
impact on the outcome.

The full model includes a large number of continuous (coded with rcs 
with 3 knots)  and categorical variables, as well as a plethora of 
interactions between the categorical and continuous variables. Only one 
of the interactions between the binary variable and the other 
categorical or continuous variables is statistically significant.

Can anyone offer a suggestion on what might explain this discordance?

______________________________________________
R-help at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.