Date:         Mon, 29 Sep 2008 11:04:15 -0500
Reply-To:     Mary <mlhoward@avalon.net>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Mary <mlhoward@AVALON.NET>
Subject:      Re: quasicomplete separation of data points  -- Follow up
Comments: To: "Presley, Rodney (CDC/CCHP/NCBDDD)" <rpa9@CDC.GOV>
Content-Type: text/plain; charset="iso-8859-1"

I'm not an expert, but did take the Categorical Data course from SI.   Effect coding for three values would create two design variables, but **with three categories each** such as:
Value 1, Low=  dummy1=1, dummy2=0
Value 2, Medium= dummy1=0 dummy2=1
value 3, High= dummy1=-1, dummy2=-1

But effect coding (param=ref ref='3') would still create two design variable, but with **two categories each**, such as:
value1, Low= dummy1=1 dummy2=0;
value2, Medium= dummy1=0 dummy2=1
value3, High= dummy1=0, dummy2=0;

Thus the effect coding has more categories (0, 1, -1) than the reference coding (0, 1), and that could contibute to the sparseness of cells, such as envisioning a three way matrix of dummy1, dummy2, and the response variable, I believe that you'd wind up with a 3 by 3 by 2 matrix for effect coding (18 cells) versus a 2 by 2 by 2 matrix for reference coding (8 cells). 

Peter Flom strongly recommends always using reference coding, but mostly for the interpretability of the responses, but it does seem to me that this difference in sparseness could be a factor in choosing reference coding as well, particularly when you are coding your other variable with reference coding.  Someone please correct me if I haven't stated this correctly.

-Mary
  ----- Original Message ----- 
  From: Presley, Rodney (CDC/CCHP/NCBDDD) 
  To: SAS-L@LISTSERV.UGA.EDU 
  Sent: Monday, September 29, 2008 10:10 AM
  Subject: Re: quasicomplete separation of data points -- Follow up



  It appears from the log below that the parameterization method
  specified, or if omitted defaulted to param=effect, does impact the
  results.   I do not know enough about logistic regression to understand
  why this would be so.  It seems to me that it should not.  But it does.


   


  Rodney J. Presley, PhD 
  Division of Hereditary Blood Disorders *-+-*al 

  Center on Birth Defects and Developmental Disabilities 
  Centers for Disease Control and Prevention 
  1600 Clifton Road, MS E 64 
  Atlanta, GA 30333 
  Work Phone: 404-498-6732
  Main Number: 404-498-6750
  Fax: 404-498-6799
  Email: RPresley@cdc.gov