EcoSta 2022: Start Registration
View Submission - EcoSta2022
A0277
Title: Design consistent Bayesian tree models Authors:  Daniell Toth - US Bureau of Labor Statistics (United States) [presenting]
Scott Holan - University of Missouri (United States)
Diya Bhaduri - University of Missouri (United States)
Abstract: Tree models provide a method for analyzing survey data because of the easy way they can handle a large number of variables with many interactions often found in this type of data. However, until recently design consistent tree modeling algorithms have not been available for use on data collected from a complex sample design. Design consistent algorithms are very desirable due to the many potential applications of these methods to survey data. As these applications have become more complex, interest in modeling the conditional distribution at each node using more sophisticated models has grown. Bayesian tree modeling approaches with a prior distribution on the set of all possible tree models and then selecting the optimal model using a stochastic search have been developed for independent data, but there are no methods for incorporating survey weights to produce design consistent models. Since the Bayesian framework allows for easily incorporating more complex models, we propose extending the Bayesian tree algorithm research to obtain a design consistent Bayesian tree model. The methods are illustrated through empirical simulation and an application to the Consumer Expenditure Survey.