Switching PCA for modeling changes in the underlying structure of multivariate time series data
Kim De Roover, Eva Ceulemans, Marieke E. Timmerman, Patrick Onghena · Lirias · 2011
Behavioral researchers are often interested in the underlying structure of multivariate time series data. For instance, given multiple measurements of a set of variables for one subject, one may wonder whether some of the variables covary across time or whether they all fluctuate independently of one another, and whether the underlying structure is the same across all measurements or varies over time. For example, when something happens between two measurements, the amount or nature of covariation in the data can change. To explore such structural differences across time, Switching PCA is developed. Using Switching PCA, clusters of sequential (i.e., a time contiguity constraint is imposed on the clustering, implying that each cluster consists of consecutive measurements only) measurements are induced, according to the underlying structure, and the data within each cluster are modeled by a separate PCA. An algorithm for fitting Switching PCA models is presented. The value of the model for empirical research is demonstrated.