Comparing Clustering Algorithms and Transformation Methods for Categorical Data

• Activity Number: 552
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Tingting Zhang* and Jenhao Cheng
• Companies: Press Ganey Associates and Press Ganey Associates
• Keywords: Clustering ; Categorical data ; Health care

Abstract

K-means and centroid linkage are two popular clustering algorithms used to separate target population into several homogeneous groups based on continuous variables. However, very few studies focus on how well these algorithms are performed on categorical data and how to transform categorical data. Here we cluster 547 Press Ganey's Quality Performer clients into several peer groups for fair performance comparison based on their hospital characteristics: beds, geographic region, ownership type, metropolitan area and teaching status. These categorical attributes are transformed into dichotomous dummy variables and continuous principal components. Both K-means and centroid linkage are performed on the two sets of variables with a predefined number of clusters. The clustering results are evaluated base on internal clustering criteria, such as Dunn index, Silhouette coefficient. We find that the combination of principle component analysis and centroid linkage provide the best clusters. The hospitals are classified into 3 peer groups. These hospitals can leverage the clustering results to compare the performance with their peers, and further improve quality of healthcare service.