[R] Marginal and Discrete Effects of a Logit

[Frank E Harrell Jr]

Robi Ragan wrote:
> I am trying to compute the marginal effects from a logit.
> 
> dPr(y=1)/dx_k
> 
> and
> 
> delta Pr(y=1|xbar)/ delta x_k
> 
> I have searched the archives and seen the question asked, but never an
> answer given. Is this possible in either lrm or glm?
> 
> Thanks,
> 

Taking the derivative does yield a kind of marginal effect but not one 
that most people use.  Marginal effects on a relative basis are 
traditionally computed via odds ratios (do ?summary.Design).  You could 
also compute effects as probability differences but confidence intervals 
would not be automatic.  I prefer to give confidence intervals for odds 
ratios and a graph translating odds ratios to risk differences as a 
(strong) function of baseline risk (such a graph is in my book 
Regression Modeling Strategies). -Frank

-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University