Testing Linearity of Logit

Question

Hi Team,

I learnt that Box Tidwell test is not available in JMP at present. Is there any mechanism that JMP provides us to check this assumption that is (I think) essential for Logistic regression ? (that all Continuous independent variables need to be linearly related to the logit of the dependent variable).

This is a request from a professor who is replacing SPSS in the class

thanks

Murali

SamGardner · Answer

Using the "Ingots.JMP" sample data table: you first need to create the transformed variables log[heat] and log[soak] and then include them as interaction terms with their main effect.

the heat variableshows evidence of non-linearity, that is heat*log[heat] is statistically significant.

SamGardner · Answer

@Muralidhara we have received this request before, see Wishlist Item.  My initial investigation into this made me think the Box-Tidwell test is not such a good approach as a specific test statistic.

I think the way you do this is to include the interactions between the continuous predictors and their log transform, and if any of the interactions effects that include a continuous factor is significant, then the continuous factor is deemed to "not be linearly dependent on the logit".  I suppose you could also include higher order effects also (quadratic for instance).  See https://sites.education.miami.edu/statsu/wp-content/uploads/sites/4/2020/07/Logistic-Regression-Webinar.pdf