| Description |
This dissertation explores a new class of statistics in change point analysis which we call R´enyi-type statistics. We wish to decide between the null hypothesis claiming that in a sample, a set of parameters of interest is stable (that is, the parameter values do not change in the sample) and the alternative hypothesis claiming that at some unknown point in the sample, the value of the parameters change. We show that R´enyi-type statistics are well-adapted to situations in which the change in parameter values occur very early or late in the sample especially when compared to other statistics deciding between these same hypotheses, having superior power to competing methods in such contexts. We describe the procedure and show how it can detect changes in mean, in regression models, and in models estimated via generalized method of moments estimators, along with other contexts. We demonstrate our procedures on several real data examples. We can detect changes early in linear models examining the relationship between workers productivity and compensation, Fama-French models and CAPM models, and models describing the relationship between oil prices and the strength of the dollar. A software implementation of the procedures is available via the R package CPAT. |
