|
Abstract:
|
In implementing Bayesian analysis, we face the problem of using data multiple times. We select a model and obtain the posterior distribution using the same set of data. When several analysts model the same data, we wish to combine the models efficiently, which should provide improved predictive performance. This paper tackles these problems via a novel modeling method based on data-splitting. In implementing this method, several data analysts work independently on portions of a dataset, constructing separate models that are updated eventually and combined through Bayesian model averaging. This paper provides theoretical results that characterize general conditions under which data-splitting results in improved estimation. Application of the method to a popular, real dataset shows predictive performance superior to that of many popular automatic modeling techniques.
|
