13239: Noel: FW: [hss] Haitian Scientific Seminar Seminar September 28, 2002 (REMINDER) (fwd)

[Bob Corbett <corbetre@webster.edu>]

From: Alfred Noel <alfred.noel@umb.edu>

REMINDER!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

Alfred G. Noël
The University of Massachusetts Boston
Department of Mathematics
Science Building, Floor 03, Room 00175
100 Morrissey Boulevard
Boston, MA 02125-3393

alfred.noel@umb.edu
http://www.math.umb.edu/~anoel


----------
From: Alfred Noël <alfred.noel@umb.edu>
Date: Wed, 11 Sep 2002 22:46:59 -0400
To: <hss@cs.umb.edu>
Subject: [hss] Haitian Scientific Seminar Seminar September 28, 2002

Dear Friends,
            Below you will find information concerning the next HSS seminar.
You may want to visit http://www.math.umb.edu/hss/seminars/2002/ also if you
have access to the Internet. For further information I can be reached at:

The University of Massachusetts Boston
Department of Mathematics
Science Building, Floor 03, Room 00175
100 Morrissey Boulevard
Boston, MA 02125-3393

alfred.noel@umb.edu
http://www.math.umb.edu/~anoel

----------------------------------------------------------------------------
Haitian Scientific Society
Seminar series

Northeastern University Mathematics Seminar Room 544 Nightingale
September 28, 2002 11:45 AM Ð1:30 PM

Monir Hossain
Massachusetts VA Epidemiology Research
& Information Center, VA Boston Health Care System, MA

A Comparison of Predictive Performance of Statistical Models Classification
and Regression Tree (CART), and Discriminant Analysis (DA) to Identify
whether Veterans were Hospitalized in VA or Medicare


Abstract

The predictive performances of the two statistical models classification and
regression tree (CART) and discriminant analysis (DA) were compared to
predict where veterans were hospitalized. A cohort of 23,362 veterans who
turned 65 years of age in 1994 and who enrolled in Medicare was used to
predict post-65 hospitalizations in VA and Medicare. All 23,362 veterans
were hospitalized at least once in either system. The outcome variable had
three levels: 1) hospitalized in VA only (47.3%), 2) hospitalized in
Medicare only (39.6%), and 3) hospitalized in both (13.1%). The independent
variables used to classify these patients were: gender, 2-levels distance to
nearest VAMC , 3-levels VA eligibility category, 2-levels Medicare
enrollment eligibility category, 4-levels region of residence, 3-levels
Charlson co-morbidity index, and # of hospitals in the county of residence.
The overall correct classifications were similar for CART (45.8%) versus DA
(45.1%) . However, the correct classifications for the 3-levels differed
between CART and DA. The optimal tree model built by CART with 65 terminal
nodes correctly classified the 3-levels as 33.3%, 60.8%,and 45.5%,
respectively. The DA model correctly classified the 3-levels as 28.6%,
68.7%, and 33.7%, respectively. Both models used equal prior probabilities
of the groups. The order of the importance of the independent variables were
found to be the same by both the models. Although overall correct
classifications rates were similar, CART model performed much better in the
smallest (hospitalized in both) and the largest (hospitalized in VA) groups.
If the purpose of modeling is to improve correct classifications of the
smallest and the largest groups then CART model may be more Appropriate than
DA model.
This is joint work with Steven Wright.


-------------------------------------------
Alfred G. Noël
The University of Massachusetts Boston
Department of Mathematics
Science Building, Floor 03, Room 00175
100 Morrissey Boulevard
Boston, MA 02125-3393

alfred.noel@umb.edu
http://www.math.umb.edu/~anoel






-- 
Alfred G. Noël
The University of Massachusetts Boston
Department of Mathematics
Science Building, Floor 03, Room 00175
100 Morrissey Boulevard
Boston, MA 02125-3393

alfred.noel@umb.edu
http://www.math.umb.edu/~anoel