Nonlinearity, Structural Breaks or Outliers in Economic Time Series?

This paper has its motivation from discussions at the E C 2 conference in 1995. y Financial support from the Social Sciences and Humanities Research Council of Canada is gratefully acknowledged , as are helpful discussions with Sid Chib and Les Oxley. z Financial support from the Academic Senate UCLA and Center for Computable Economics, UCLA is gratefully acknowledged. In recent years there has been an increasing interest in nonlinear models as an alternative to the linear speciications which h a ve dominated the applied macroeconomics literature. For many series empirical evidence for nonlinearity exists. However, it is possible that this apparent nonlinearity could be due to structural breaks or outliers. Hence, this paper develops methods for comparing linear, nonlinear, structural break and outlier models. We adopt a Bayesian approach which allows for the easy comparison of non-nested models and surmounts the problems associated with nuisance parameters which are unidentiied under the null which plague classical tests. The computational diiculties associated with the Bayesian approach are surmounted by w orking with autoregressive s w i t c hing models for which analytical posterior results for most of the parameters are available. After motivating and deriving Bayesian methods for such models, an empirical section analyzes the behaviour of the growth of US real GDP and British industrial production.