CORRUPTION AND GROWTH UNDER WEAK IDENTIFICATION

Abstract

I. INTRODUCTION The purpose of this paper is to revisit Mauro (1995) and re-examine his results as a consequence of recent research that cautions against making traditional types of inference in the presence of weak instruments. Early work done by Nelson and Startz (1990a, 1990b) demonstrates that the distribution of the two-stage least squares (2SLS) estimator and the t-statistic is poorly approximated by the asymptotic representation when instruments are weak. More formally, Staiger and Stock (1997) model the coefficients on the instruments in the first-stage equation in the [n.sup.-1/2] neighborhood of zero in which they show that the 2SLS estimator loses consistency resulting in a nonstandard distribution for the t-statistic. In fact, they show that under the local-to-zero framework, the distribution of the t-statistic depends on the level of endogeneity, strength, and the number of instruments. Thus, the true size of the t-statistic can differ substantially from its perceived size when instruments are weak. Therefore, 95% confidence intervals based on asymptotic approximations, such as [??] [+ or -] 1.96se([??]) are no longer reliable. As a rough rule of thumb, Staiger and Stock (1997) suggest a first-stage F-stat of 10 on the excluded instruments as an indication of acceptable instruments. In terms of point estimation, Bound, Jaeger, and Baker (1995) show that the 2SLS estimator is biased in the same direction as the ordinary least squares (OLS) when using irrelevant instruments. Hahn and Hausman (2002) illustrate a detailed expression for the 2SLS bias showing that it is monotonically decreasing in the instrument strength while monotonically increasing in the degree of endogeneity. Cruz and Moreira (2005) verify these results through Monte Carlo simulations also showing that the classic bias corrected 2SLS estimator of Nagar (1959) provides a poor correction when instruments are very weak. Given that the theory and techniques mentioned above were discovered and developed after Mauro wrote his QJE paper, we attempt to reproduce his results and analyze them in light of these recent findings. From the early work of Left (1964) and Huntington (1968), corruption in government has been of great interest to economists. These studies argue that corruption could be positively correlated with economic performance in the presence of a thick and cumbersome bureaucracy. Corruption in this view greases the wheels of bureaucracy, thus increasing the efficiency in which transactions occur, leading to a positive effect on the economic performance of a country. On the other hand, Rose-Ackerman (1978), Murphy, Shleifer, and Vishny (1991, 1993), and Shleifer and Vishny (1993) provide theoretical arguments that corruption deteriorates economic growth through the misallocation of talent and other resources. Mauro (1995), being the first to introduce instrumental variables (IVs) in the cross-country growth literature, contributes to this debate by examining empirically the relationship between two measures of corruption and investment and economic growth. His results suggest that corruption has a negative impact on investment and economic growth. The significance of his results varies with model specifications but the results appear to be robust when correcting for the endogeneity of corruption. Using the ethnolinguistic fractionalization (ELF) index as an IV, he employs 2SLS estimation to correct for endogeneity created by the two measures of corruption. Mauro (1995)'s work has been of great value in the corruption literature and has had profound influence on the direction of policy making. The paper is one of the most cited papers not just in the literature on corruption but in economics in general. (1) Given the impact that the paper has undoubtedly had on the direction of research, we find it necessary and relevant to revisit the paper in light of the recent developments in the field of econometrics. …