Outlier Identification in Model-Based Cluster Analysis

In model-based clustering based on normal-mixture models, a few outlying observations can influence the cluster structure and number. This paper develops a method to identify these, however it does not attempt to identify clusters amidst a large field of noisy observations. We identify outliers as those observations in a cluster with minimal membership proportion or for which the cluster-specific variance with and without the observation is very different. Results from a simulation study demonstrate the ability of our method to detect true outliers without falsely identifying many non-outliers and improved performance over other approaches, under most scenarios. We use the contributed R package MCLUST for model-based clustering, but propose a modified prior for the cluster-specific variance which avoids degeneracies in estimation procedures. We also compare results from our outlier method to published results on National Hockey League data.