Determinants and costs of current account reversals under heterogeneity and serial correlation

Abstract

Using large panel data sets for analysis of determinants and costs of reversals asks for controls of latent heterogeneity among countries. This article performs a Bayesian analysis, which allows for a parsimonious yet flexible handling of country specific heterogeneity via random coefficients and serially correlated errors. Consideration of persistence within the employed macroeconomic data is important to gauge the impact of explaining variables suggested by theory correctly. Bayesian specification tests provide evidence in favour of models incorporating heterogeneity and serial correlation. The results suggest that costs of reversals are overestimated, when country specific heterogeneity is neglected and stress the importance of external variables in explaining current account reversals. Results are checked for robustness against the underlying reversal definition.

Keywords:

• current account reversals
• panel treatment model
• random parameters
• serial correlation
• Bayesian analysis
• Markov Chain Monte Carlo estimation

JEL Classification:

• C30
• F32
• F43
• C33
• C35

Acknowledgements

For comments and suggestions received during the referee process and helpful discussions on earlier versions of this article the author thanks the editor and participants of the 2007 European Meeting of the Econometric Society in Budapest.

Notes

¹ Geweke et al. (1997) note that the numerical accuracy of Gibbs sampling in the context of a multinomial multiperiod probit is superior to other approaches based on simulated maximum likelihood or simulated moment conditions in the presence of strong serial correlation. This provides another argument in favour of using a Bayesian approach.

² The following list of countries is analysed: Argentina, Bangladesh, Benin, Bolivia, Botswana, Brazil, Burkina Faso, Burundi, Cameroon, Central African Republic, Chile, China, Colombia, Republic of the Congo, Costa Rica, Cote d’Ivoire, Dominican Republic, Ecuador, Egypt United Arab Republic, El Salvador, Gabon, Gambia, The Ghana, Guatemala, Guinea-Bissau, Haiti, Honduras, Hungary, India, Indonesia, Jordan, Kenya, Lesotho, Madagascar, Malawi, Malaysia, Mali, Mauritania, Mexico, Morocco, Niger, Nigeria, Pakistan, Panama, Paraguay, Peru, Philippines, Rwanda, Senegal, Seychelles, Sierra Leone, Sri Lanka, Swaziland, Thailand, Togo, Tunisia, Turkey, Uruguay, Venezuela, RB, Zimbabwe.

³ The definitions follow Alesina and Perotti (1997) applying similar definitions in the context of fiscal stabilization.

⁴ Starting point of the derivation is to decompose the log marginal likelihood of all data S into

As this identity holds for each point θ within the parameter space of the model, it is calculated at a point within the highest density region where θ* is the posterior mean. The first component gives the log likelihood. For the pooled treatment model it has a closed form. For the specifications allowing for serial correlation and heterogeneity, the likelihood is computed using the Geweke–Hajivassiliou–Keane (GHK)-simulator with 200 replications ensuring numerical accuracy (see Geweke et al. (1994) for details). The second component is the log prior of all model parameters evaluated at the estimated parameter values. The last component of the marginal likelihood is the full posterior distribution of the model parameters, which is obtained iterative runs of shortened Gibb samplers.

⁵ If B < 0 no evidence for the specification under H ₀ is found, for 0 < B < 1.15 very slight evidence in favour of H ₀ is found, with 1.15 < B < 2.3 the evidence is slight, strong evidence is found for 2.3 < B < 4.6 and very strong evidence is found for B > 4.6.

⁶ Convergence has been checked based the diagnostic convergence statistics as introduced by Geweke (1992). A summary of convergence diagnostics is provided by the author upon request.

⁷A maximum likelihood analysis with heteroscedastic variance modelled as σ_it  = exp{γX_it } points in the same direction.

⁸ Priors of mean parameters are not subject to sensitivity analysis, since they are chosen to be generally uninformative with a variance of 1000. In general, conjugate priors are chosen in order to facilitate closed form sampling. Further details are provided upon request by the author.

⁹ An analytical solution to the corresponding integration problem is not available although the error of the probit and growth equation are jointly normally distributed, which would allow to compute the expectations and conditional on the random coefficients defining the range of integration. However, the succeeding integral over the random coefficients has then no closed form solution.

¹⁰ Note that this approach in dealing with the regressors also incorporates in an ad hoc manner reactions of the weak exogenous regressors (e.g. the reserve variable) on a current account reversal.

¹¹ Sampling of errors is performed based in two steps. The joint distribution of errors is decomposed in the marginal distributions of errors within the initial period t ₀ and the corresponding conditional distributions of all other considered errors.

¹² Note also that for all other parameters the cross validation experiment indicates the stability of estimates.