R help - Mar 2010 - logistic model diagnostics residuals.lrm {design}, residuals()

I am interested in a model diagnostic for logistic regression which is normally
distributed (much like the residuals in linear regression with are ~
N(0,variance unknown).

My understanding is that most (all?) of the residuals returned by residuals.lrm
{design} either don't have a well defined distribution or are distributed as
Chi-Square.

Have I overlooked a residual measure or would it be possible to transform one of
the residual measures into something reasonably 'normal' while retaining
information from the residual so I could compare between models (obviously I
could blom transform any of the measures, but then I'd always get a standard
normal)?

Cheers,
bimal

Bimal P Chaudhari, MPH
MD Candidate, 2011
Boston University
MS Candidate, 2010
Washington University in St Louis


	[[alternative HTML version deleted]]

Why do you need a diagnostic that has properties from the normal?  Logistic
regression is based on binary (binomial distribution) data, not continuous data.
Any transform that forced normality (even just under a given null hypothesis)
would probably distort any real information that might be gleaned.

What are you really trying to accomplish?  It is probably easier to address that
then to do an artificial 'normalization'.

-- 
Gregory (Greg) L. Snow Ph.D.
Statistical Data Center
Intermountain Healthcare
greg.snow at imail.org
801.408.8111

> -----Original Message-----
> From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-
> project.org] On Behalf Of Chaudhari, Bimal
> Sent: Thursday, March 11, 2010 10:10 AM
> To: r-help at r-project.org
> Subject: [R] logistic model diagnostics residuals.lrm {design},
> residuals()
> 
> I am interested in a model diagnostic for logistic regression which is
> normally distributed (much like the residuals in linear regression with
> are ~ N(0,variance unknown).
> 
> My understanding is that most (all?) of the residuals returned by
> residuals.lrm {design} either don't have a well defined distribution or
> are distributed as Chi-Square.
> 
> Have I overlooked a residual measure or would it be possible to
> transform one of the residual measures into something reasonably
> 'normal' while retaining information from the residual so I could
> compare between models (obviously I could blom transform any of the
> measures, but then I'd always get a standard normal)?
> 
> Cheers,
> bimal
> 
> Bimal P Chaudhari, MPH
> MD Candidate, 2011
> Boston University
> MS Candidate, 2010
> Washington University in St Louis
> 
> 
> 	[[alternative HTML version deleted]]
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-
> guide.html
> and provide commented, minimal, self-contained, reproducible code.

Chaudhari, Bimal wrote:
> I am interested in a model diagnostic for logistic regression which is
normally distributed (much like the residuals in linear regression with are ~
N(0,variance unknown).
> 
> My understanding is that most (all?) of the residuals returned by
residuals.lrm {design} either don't have a well defined distribution or are
distributed as Chi-Square.
> 
> Have I overlooked a residual measure or would it be possible to transform
one of the residual measures into something reasonably 'normal' while
retaining information from the residual so I could compare between models
(obviously I could blom transform any of the measures, but then I'd always
get a standard normal)?
> 
> Cheers,
> bimal
Hi Bimal,

What would make it necessary for the residuals to have a certain 
distribution?  Why would you expect a categorical Y variable to give 
risk to residuals with a nice distributions?

You can do residual diagnostics without worrying about the distribution.

Frank
> 
> Bimal P Chaudhari, MPH
> MD Candidate, 2011
> Boston University
> MS Candidate, 2010
> Washington University in St Louis
> 
> 
> 	[[alternative HTML version deleted]]
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 

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