Cross-Validation and Bootstrap to Evaluate the Number of Subgroups in Finite Mixture Models

• In Event: Marriott Roundtable Session 29
  In Roundtable Session: Advances in Clustering and Classification

Abstract

Practitioners of finite mixture models often have difficulty selecting the correct number of classes due to conflicting fit indices. Recent empirical and simulation studies advised cross validation as an alternative way to measure the accuracy of class enumeration with growth mixture models, because it iteratively tests models while controlling for Type I error rate. Such studies are limited to hold-out and k-fold cross validation and conversely suggests that the k-fold approach only performs well when classes have large separation. Our current project expands this area of research to four types class enumeration methods: k-fold, leave-out-one, hold out, and bootstrap. We also compare the performance of each type of cross validation on two mixture models, latent class and latent profile analyses.

Author

• Zachary Kendall Collier, The University of Delaware