Growth Empirics: A Panel Data Approach--A Reply

Abstract

Research on convergence has indeed proceeded through several stages. In the first, represented by such works as Baumol [1986], convergence was studied under the implicit assumption that all countries of the sample had the same steady state levels of income. This notion of convergence later came to be known as absolute convergence. In the second stage, represented by papers such as Barro and Sala-i-Martin [1992] and Mankiw, Romer, and Weil [1992], the concept of conditional convergence was put forward and rigorously defined. It was now emphasized that the growth theory did not imply identical steady state levels of income for all countries and that, in studying convergence, these differences needed to be allowed. However, differences in steady state levels were considered under the assumption of parametric homogeneity of the underlying production function. Yet the assumption that all countries have identical production functions and differ only in the value of the variables of this function and not in its parameters was neither appealing nor tenable. In the third stage, represented by such works as Islam [1995], this assumption of strict parametric homogeneity was relaxed, and, through the use of panel data methodology, the aggregate production function was allowed to differ across countries with respect to the productivity shift parameter. However, this introduction of heterogeneity was still within the general framework of allowing only the steady state level of income to differ across countries, and it maintained the assumption that countries shared the same steady state growth rate. It is in this latter respect that Lee, Pesaran, and Smith [1995] and their current Comment [1998] on the basis of that work represent a significant change. They have extended the use of the panel data method to allow the countries to differ not only with regard to the steady state level but also with respect to the steady state growth rate (g). When this is done, the rate of convergence (A), according to their maximum likelihood estimation, jumps to 0.23. Clearly, a different estimation method is not the main reason for this substantial increase because when g is