Date:         Wed, 10 Dec 1997 23:57:10 GMT
Reply-To:     "Steven K. Smith" <sksmith@FACSTAFF.WISC.EDU>
Sender:       "SPSSX(r) Discussion" <SPSSX-L@UGA.CC.UGA.EDU>
Comments:     RFC822 error: <W> Incorrect or incomplete address field found and
              ignored.
From:         "Steven K. Smith" <sksmith@FACSTAFF.WISC.EDU>
Organization: Univ of Wisconsin - Madison
Subject:      nonparametric estimation of class-conditional probability density
              functions
Comments: To: SAS-L@VTVM1.CC.VT.EDU, STAT-L@VM1.MCGILL.CA

I am trying to identify characteristics associated with group membership. Class
variable is binary (1,0). Characteristics are defined by both categorical and
continuous variables.
 
Using SAS I have already estimated a logistic model. Now I would like to try
discriminant analysis, but SAS will only do parametric discriminant analysis
when some independent variables are categorical (CATMOD).
 
Is there someone out there who can suggest a stats package that is more
flexible when it comes to direct estimation of class-conditional probability
density functions when the measure space is defined by both categorical and
continuous variables? I'm looking for something that could do kernal,
kth-nearest-neighbor in addition to parametric (assuming normal mixture).
 
In addition, can someone recommend a text that is useful reference for applied
work? I am using D.J. Hand "Classification and Discrimination". It contains a
nice theoretical discussion, but offers little in the way of practical advice.
I'm not looking for a how-to text, but rather for a reference that can offer
practical decision making criteria on applied methods (e.g., choosing between
logistic and discriminant analysis, criteria for choosing parametric vs
nonparametric methods, testing predictive power of the empirical model, etc.).
 
Thank you.
Janet Marie Eisenhauer
eisenhauer@aae.wisc.edu
Associate Researcher
Land Tenure Center
Univ of Wisconsin - Madison