🚀 go-pugleaf

Deanna Wild <dkw4@NIOSHE2.EM.CDC.GOV> writes...

> Anyone familiar with a FORMAL 'test for (linear in the log odds) trend'
> in logistic regression?
>
> We tend to use an informal approach here - i.e.; checking to see if quadratic
> terms or the like are needed beyond the linear term.  Or, we categorize the
> continuous independent variable and check to see how consistent the change in
> the resulting indicator variables' coefficients is.
>
.... **stuff deleted** ....

-----------------------------------------------------------------
Deanna,

##  To my knowledge there exists ** NO ** formal test to check
    the linearity assumptions is a logistic model. As you point
    out, one generally introduces a quadratic and/or cubic
    terms in the model and tests for the adequacy of the
    linearity of a continous covariate (based on the likelihood
    ratio test).

##  An alternate approach is to fit models using restricted
    cubic splines for the continous covariate(s). Also, a plot
    of the spline transformation versus the continous
    variable will most often help the analyst to assess if the
    variable behaves linearly in the logit or not.

##  Frank Harrell (Div of Biometry at Duke Univ), among others,
    has authored some excellent papers on using the spline
    approach. Their group has also written some SAS MACROS for
    fitting models using cubic splines. A very interesting
    approach indeed !

##  I would definitely like to hear more about the existence
    of "formal" tests, if they exist.

--  Cande V Ananth (ananth@UNCMVS.OIT.UNC.EDU)
    Dept of Boistatistics, SPH
    Univ of North Carolina at Chapel Hill