Bayesian model selection for unit root testing with multiple structural breaks

Abstract

A fully Bayesian approach to unit root testing with multiple structural breaks is presented. For this purpose the number of breaks, the corresponding break dates as well as the number of autoregressive lags are treated as model indicators, whose posterior distributions are explored using a hybrid Markov chain Monte Carlo sampling strategy. The performance of the sampling algorithm is demonstrated on the basis of several Monte Carlo experiments. In a next step the most likely model is used to test for a unit root with possible multiple breaks by computing the posterior probability of this point hypothesis under different prior distributions. The sensitivity of the test results with regard to the assumed prior distribution is analyzed and the Bayes test is compared with some classical unit root tests by means of power functions. Finally, in an empirical application the yearly unemployment rates of 17 OECD countries are analyzed to answer the question if there is persistence after a labor market shock.

Introduction

There has been a growing literature to unit root testing in economic time series over the last three decades starting with the seminal papers of Dickey and Fuller (1979) and Nelson and Plosser (1982). As stressed by many authors the misspecification of the considered test regressions can lead to substantially biased inferences in the class of autoregressive integrated moving average (ARIMA) models, see Banerjee et al. (1993), Stock (1994) and Maddala and Kim (1998) for discussions. The specification regards on the one hand the structure of the stochastic component, i.e. the autoregressive and/or moving average lag orders (see Hall, 1994, Ng and Perron, 2001) and on the other hand the specification of the deterministic components like the inclusion of time trends, the number of possible structural breaks and the timing of these breaks, see Perron (1989), Christiano (1992) and Vogelsang and Perron (1992), inter alia. In the Bayesian unit root literature a heated controversy was devoted to the adequate prior use, model specification issues and the proper modeling of initial conditions, see e.g. Sims (1988), Phillips (1991) and Uhlig (1994), inter alia, and Bauwens et al. (1999) for an overview. In contrast to the classical literature there are only a few approaches to account for structural breaks when testing for stochastic trends, see e.g. Zivot and Phillips (1994), Koop and Steel (1994), DeJong (1996) and Marriott and Newbold (2000). Most of these works treat the model order, in particular the lag order and/or the number of breaks as fixed quantities. In the context of nonlinear models Chen et al. (2013) consider a threshold autoregressive (TAR) model, where the innovation variance is assumed to follow a GARCH process and the mean equation is modeled by a TAR(1) process with a fixed number of regimes. In general, the process of model selection induces uncertainty with respect to any subsequent analysis and thus should be captured in order to improve statistical inference. Although there exist many classical approaches to model selection in dynamic models with structural breaks, which are mainly based on information criteria, a Bayesian framework is chosen here.

In the following a stochastic model selection approach is presented, which can be used to determine the optimal specification for unit root testing in the case of multiple breaks. In a nutshell, the proposed sampling scheme can be regarded as an extension of the approach presented in Zivot and Wang (2000) by estimating the number of breaks, the associated break dates, as well as the number of autoregressive lags simultaneously with all other unknown model parameters. This is accomplished by the introduction of two discrete valued state variables which indicate certain model combinations in the space of candidate models. By estimating the lag order and the number of breaks simultaneously we get an approximation of the joint posterior distribution of these model indicators. Therefore the underlying Bayesian approach allows a probabilistic representation of the model uncertainty through the shape (e.g. multimodality, platykurtosis) of the model posterior distribution. In particular with regard to unit root testing model selection is a crucial part in order to get reliable test decisions. For this reason it is important to capture any uncertainty related to this step. Since the joint distribution of all unknown parameters is of varying dimension, i.e. depends on the specific model complexity, usual sampling techniques to generate random draws from this distribution, like Gibbs sampling, cannot be applied without further modifications. For this purpose a flexible Markov chain Monte Carlo (MCMC) approach is introduced, which enables to jump between parameter spaces of differing dimensionality. The performance of this method is demonstrated on the basis of Monte Carlo (MC) experiments which indicate great reliability in finding the true values of the data generating process (DGP). Using Bayesian methods for model selection in structural break models has the advantage that most of these methods are technically simpler than their classical counterparts, allow for finite-sample inferences that are optimal given the framework, and also allow for nonnested model comparisons (see Zivot and Wang, 2000). Furthermore, with regard to unit root testing a Bayesian model framework is appealing, because unlike in classical approaches inference stays the same here for trending and nontrending data (see Sims and Uhlig, 1991). So far many approaches to model selection in time series models have been proposed in the Bayesian literature. Among the many works that focus on lag order determination in ARMA models are Barnett et al., 1996, Barnett et al., 1997, Huerta and West (1999), Chen (1999), Gerlach et al. (2000), Vermaak et al. (2004), Ehlers and Brooks (2004) and Philippe (2006), inter alia. Many of the existing works that deal with the detection of change points treat the selection of the number of breaks as a successive problem, which is solved by using information criteria or Bayes factors, but do not treat the number of change points together with the number of lags as additional unknown parameters in their sampling schemes, see for example Chib (1998), Zivot and Wang (2000) and Koop and Potter (2004).

Therefore the present work aims to provide contributions in the following directions: a stochastic model selection approach for multiple structural breaks models is proposed, where the autoregressive lag order, the break dates and also the number of breaks can be estimated simultaneously with all other model parameters. As a result the joint posterior distribution of these model indicators is obtained, which can be used for further inference. Most importantly, the model selection approach presented below focuses on the application in unit root testing problems and the use of Augmented-Dickey–Fuller (ADF) regressions (see Said and Dickey, 1984) and thus provides an alternative to classical model selection strategies used in that context. In the following structural breaks are allowed to occur in the deterministic trend function only, because for the subsequent unit root testing a closed-form expression of the required marginal likelihoods will be required. However with regard to the empirical application of Section  7, using unemployment data, for example, allowing for multiple variance breaks is of less importance as it would be in the case of high frequency financial data. Besides model determination a second focus lies on testing for a (zero frequency) unit root when there are multiple breaks in the trend function. The proposed Bayesian unit root test is then compared with several classical unit root tests by means of simulated power functions. MC experiments indicate a superiority of the Bayes test in terms of power especially for moderate and small sample sizes.

In an empirical application, the unemployment rates of 17 OECD countries for the years 1960–2010 are analyzed to answer the question if there is persistence after a labor market shock. The majority of empirical works to test for persistence effects in European unemployment rates have been done by using classical methods. Most of these works apply univariate tests without structural breaks and cannot reject the unit root null hypothesis, see Mitchell (1993), Roed (1996) also Hassler and Wolters (2009). For the US the results are mostly reverse and therefore no high degree of persistence has been found in the unemployment rates, see Nelson and Plosser (1982), Blanchard et al. (1992) also Roed (1996). It is by now well recognized that in case of structural breaks in the data, not allowing for breaks in the test regression can bias the results towards a unit root. Therefore, a second group of studies uses methods that allow for (multiple) structural breaks. Not surprisingly the results of these studies show a clearer tendency against a unit root, see Arestis and Biefang-Frissancho Mariscal (1999), Papell et al. (2000), Papell and Prodan (2004) and also Pascalau (2007). So far only few authors used Bayesian methods to analyze the trend characteristics of unemployment rates, exceptions are Summers (2004), Mikhail et al. (2006) and Berger and Everaert (2008). Since the time span of the data covers the first financial crisis of the year 2008 it is also possible to capture the impact of this event. Thereby, besides the main statistical focus, the present work also aims to provide new empirical evidence to the question if some OECD countries are more likely to recover to their natural rate of unemployment after an exogenous shock or if such an event has a permanent impact on a country’s long run unemployment rate.

The paper is organized as follows: in Section  2 the statistical model is presented and in Section  3 the MCMC sampling algorithm is introduced and its performance is analyzed in Section  4. Then in Section  5 the Bayesian testing approach for a unit root with multiple structural breaks is presented. Further, in Section  6, the sensitivity of the test results with regard to the assumed prior distributions is analyzed and the Bayesian unit root test is compared with some classical unit root tests. In Section  7 the presented methods are applied to annual OECD unemployment rates and Section  8 concludes.