Two essays in applied econometrics on duration analysis and model selection

Abstract

The thesis consists of two essays in applied econometrics. Essay 1 adopts duration analysis to investigate what factors affect the length of an "interest rate spell" - the period during which the interest rate instrument remains unchanged under inflation targeting. Nowadays, an increasing number of countries have transitioned to inflation-targeting regime, and most of them have achieved stable economic growth and low rates of inflation. Not surprisingly, inflation targeting has attracted growing interest among economists in recent years. Shih and Giles (2009) first applied duration analysis to model the length of an interest rate spell under inflation targeting, and their analysis was directed to the experiences of Canada. This essay extends Shih and Giles (2009) work through a cross-national analysis of eight inflation-targeting countries (or areas). Both parametric and nonparametric methods are employed for the analysis. The conclusion is consistent with that of Shih and Giles (2009), that is, the length of an interest rate spell is affected by both the rate of inflation and the rate of economic growth; the influence of exchange and unemployment rates proved to be insignificant. Moreover, empirical results support that inflation-targeting central banks usually design their monetary policies based on the Taylor Rule. Essay 2 compares the performances of four model selection criteria: Akaike’s information criterion (AIC), the Schwarz Information criterion (SIC), the smallsample bias corrected AIC (AICc) and the small-sample bias corrected SIC (SICc) in a linear regression model with heteroskedasticitic or mutually dependent errors. Monte Carlo simulation was used to conduct the analysis. AIC and SIC are widelyused information criteria, but they have a tendency to overfitting in small samples. The AICc and SICc can reduce this small-sample bias. Most literature compares the model selection criteria in the situation when the error terms are identically and independently distributed. Ohtani (2003) first examined their small-sample performances in a linear regression model with first-order autocorrelated errors. Stipulated by Ohtani (2003), this essay considers other types of error terms, including multiplicative heteroskedasticity, the first-order moving average (MA (1)) and the first-order autoregressive conditional heteroskedasticity (ARCH (1)). It was found that AIC and SIC tend to overfit in small samples, and AICc and SICc outperform AIC and SIC respectively. As the sample size increases, AIC and SIC have a decreasing tendency to overfitting and respectively converge to AICc and SICc. In addition, SIC and SICc perform better than AIC and AICc respectively.