# Categorising continous data in logistic regression

If a predictor in a logistic regression is not discriminating between 'goods' and 'bads' for most of the range and is discriminating only at higher or lower values (monotonic), should the variable be used as continuous or as a dummy indicating the range in which it is predictive?

-
Categorizing doesn't always damage the information used for the fit in a (generalized) linear model. For example, if the relationship between $x$ and $y$ is highly non-linear, then you can use the data much more effectively by categorizing $x$ and then estimating it's effect as you would any factor in regression. This might more adequately capture the non-linear aspects of the relationship (if sufficiently many roughly homogeneous categories are made) than calculating a linear fit to the original continuous measurement. –  Macro Sep 8 '11 at 12:18