RE: st: Residuals in Logistic Regression

[Richard Williams <Richard.A.Williams.5@nd.edu>]

> I never understood the rationale of this whole business. It has to do with 
> whether you focus on individual observations (in which case your degrees 
> of freedom equal your total sample size) or covariate patterns (in which 
> case, the degrees of freedom are dependent on what covariates you include 
> in the model). H&L prefer the latter. I think it leads to certain 
> unintuitive situations. For example, suppose you fit a model with only an 
> intercept. All the residuals will be 0 in this case (there's only a single 
> covariate pattern, ie, the entire sample). Surely that cannot be right? 
> Perfect fit? I don't think so.
Thanks Constantine.  Another intuitive oddity: I took some of the tied 
covariate patterns and added .0001 to the value of a variable in one case 
and subtracted .0001 from the value of a variable in another.  It made no 
difference in the parameter estimates but it greatly changed the 
residuals.  So, trivial differences in the covariate patterns can create 
huge differences in the residuals.

I generally like the way Stata picks the "right" statistic for you, but in 
this case I think it may be being a little too paternalistic; it would be 
nice if it gave you the option of computing residuals on an 
individual-level basis as well as based on the covariate patterns. 
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/