Knowledge Item: CA-Technical Issues-05
Alternate Ways of Calculating Data Partitioning Trees (or CHAID Analyses)

CHAID (CHi-square Automatic Interaction Detector) is a data modeling method that has been used extensively in this Knowledge Base. This Knowledge Item (Extended Results) looks at some alternate solutions for a few data sets that would have been obtained if an alternate data modeling method had been used. While we believe that the Exhaustive CHAID algorithm used is the most appropriate data mining method, it is instructive to see alternate results that would have been obtained with a different method.

While our studies of the algorithms in these modeling methods and tests with various datasets led us to choose the Exhaustive CHAID method as the most potentially "accurate" using criteria of recovery of known datasets and theoretical meaningfulness of the obtained results, the C&RT (Classification & Regression Trees, also called C&RT) modeling method is also widely used in the research literature, in part because of its computation speed (an issue which is not as relevant in 2000 as it was 10 years ago). In general our comparison of alternate models suggests that in terms of generally substantive results (as opposed to the minor variations always found between competing statistical modeling methods with slightly different assumptions), the two methods yield generally comparable findings for the types of applications we have studied in this Knowledge Base.

The Extended Results section shows a number of examples where we calculated the trees using alternate methods. The Additional Statistics section shows what happens when we make alternate assumptions about the level of variable scaling (nominal versus ordered categories) in some of our models.

Because of the tests illustrated here, and a much more extensive set of model runs as well as our reading of the literature, we have adapted the modeling strategy discussed more extensively in Knowledge Item: CA-Technical Issues-07.

Exhaustive CHAID

Click graphic to expand. (IE 6 users may also have to click the graphic expansion button in the new window.)

Classification and Regression Trees (C&RT; C&RT)

Knowledge Item Citation: Huba, G. J., Panter, A. T., Melchior, L. A., and the HRSA/HAB SPNS Cooperative Agreement Steering Committee (1998-2001). Knowledge Item: CA-Technical Issues-05 from HRSA/HAB's SPNS Cooperative Agreements on Innovative Models of Care, The Measurement Group Knowledge Base on HIV/AIDS Care, Online at www.TheMeasurementGroup.com.