Industrialization, convergence, and patterns of growth.

• Author: Cho, Dongchul

1. Introduction

   The long-run behavior of national growth rates has long been of great interest because it sheds light on future income disparities across countries as well as the prospective income of individual countries. Recently, the interest has been further stimulated, for the neoclassical growth model [6; 31] and new endogenous growth models [16; 26] yield sharply different predictions: while the neoclassical growth model predicts that countries with similar preferences and technology will converge to similar levels of per capita income, endogenous growth models predict that there will be no such tendency.

   In this regard, a substantial body of empirical study has examined whether the regression of the growth rate on the level of income per capita indeed produces a negative coefficient as predicted by the neoclassical model. Evidence is mixed, however. In particular, the regression results are very sensitive to the selection of countries: typically, the results for relatively developed countries are consistent with the convergence hypothesis [3; 11], but the results for the samples including less developed countries are rather in conflict with the convergence argument [9; 27]. The goals of this paper are (1) to document a stylized nonlinear (humped) pattern in growth, (2) to demonstrate how an explicit recognition of this pattern helps reconcile the conflicting results on convergence (and the conditional convergence result, below), and (3) to suggest a potential explanation for the humped pattern from a view of industrialization.

   Section II shows that there exists an economically and statistically significant hump in the growth rate of postwar cross-country data: on average, middle income countries grew the fastest, high income countries the next, and low income countries the slowest. Thus, a negative correlation between growth rates and income per capita is observed when low income countries are excluded from the sample, but no such correlation is found for a larger class of countries.

   To find the factors that can explain the fast growth of middle income countries, this paper first considers the most widely used three explanatory variables in growth regressions: the investment to GDP ratio, the percentage of age group enrolled in secondary education, and the rate of population growth. Both parametric and nonparametric analyses show that these variables cannot explain the humped pattern: middle income countries grew far faster than could be explained by these variables. In addition, the failure of the regressions to capture the humped pattern results in the conditional convergence results [2; 18]: the regressions tend to under-predict the growth rates for middle income countries and over-predict for high income countries, and the resulting positive/negative residuals for middle/high income countries generate negative correlations between the growth rate and income per capita.

   Section III, therefore, examines another factor of growth--industrialization. It has long been argued that economies can exhibit a spurt in growth during the course of industrialization. Rosenstein-Rodan [28] notes that industrialization of some leading sectors, which needs a large initial set-up cost, can "big-push" the rest of the economy to industrialize. Rostow [29] also argues that economies can "take-off" when some social/economic preconditions (e.g., infrastructure such as railroads) are met. These big-push and take-off ideas have been explored both empirically [7; 10] and theoretically [1; 8; 19; 20].

   In spite of rigorous models on industrialization, there remains a fundamental difficulty in assessing the empirical plausibility of the models--how to measure "industrialization"? This paper uses the increase in the portion of the labor force employed in manufacturing production as a proxy variable for industrialization, although this is admittedly not the perfect proxy. Along with the proxy variables for capital accumulation, this variable appears to explain the pattern in growth suitably: high income countries grew faster than low income countries because high income countries accumulated (both physical and human) capital faster, but middle income countries grew even faster because of drastic industrialization. When the explanatory variables appropriately describe the humped pattern, one can hardly find the conditional convergence result. Further examinations suggest that the paper's result appears neither a sheer coincidence nor the result of the reversed causation between the growth rate and the proxy variable for industrialization.

   Regarding the convergence vs. divergence debate, the observation in this paper seems most relevant to the argument of Baumol and Wolff [4, 1155]: "The results indicate that smaller groups of countries began to converge as early as, perhaps, 1860; that the size of the convergence club has since risen". That is, income per capita of a country in the early stage of industrialization does not converge to the levels of leading countries. Going through industrialization, however, the country can reduce the income differential from leading countries and eventually join the "convergence club".

   Section IV briefly discusses the patterns of growth from time-series data for the United Kingdom, the United States, and Japan over the past 100 years. Section V concludes, and the Appendix lists the sources of the data.

2. Convergence and Humped Pattern in Growth

   Convergence and the Selection of Countries

   The simplest but commonly used convergence test is to check whether growth rates are indeed negatively correlated with initial levels of income per capita. Table II shows the results for 95 countries in the Summers-Heston [32] data during the 1965-80 period. Results reported in this section are virtually identical to the results from the data for more countries and longer sample periods,(1) although this paper uses a little restrictive sample due to the data availability of the variables considered below. The average annual growth rate of GDP per capita (GRO hereafter) for TABULAR DATA OMITTED TABULAR DATA OMITTED each country was regressed on 1965 log GDP per capita ([Y.sub.0] hereafter). The positive coefficient estimate (first regression) indicates that on average high income countries grew faster than low income countries, which is against the convergence hypothesis.

   The result, however, differs substantially across two half-sized subsamples, 48 LDC's and 47 DC's, categorized by [Y.sub.0], the initial GDP per capita (second and third regressions): it is positive for the LDC's, but negative for the DC's. An F-test rejects (at a 5 percent significance level) the null hypothesis that the regression coefficients are the same across the subsamples. In fact, there exists a significant concave pattern: when the quadratic term of [Y.sub.0] is included in the regression, it appears significant (fourth regression). The quadratic regression estimates imply that the growth rate attains the maximum around the countries with $2700 of GDP per capita (in 1980 international prices of the Summers-Heston [32] data).

   This nonlinearity in the pattern of growth is best revealed by nonparametric estimation. (For the argument in this paper, readers who are not familiar with nonparametric estimation may simply regard it as a sophisticated moving average method.) Figure 1A plots GRO against [Y.sub.0] along with the result of the nonparametric kernel estimation (solid line) and the 95 percent confidence band (dotted lines).(2) Despite the wide dispersion, the weighted average of the growth rates (solid line) increases from 1 to 3.2 percent, and then decreases to 2.6 percent thereafter. That is, middle income countries grew the fastest, high income countries the next, and low income countries the slowest.

   The implication of the concave pattern regarding the above linear convergence test is clear: the more low income countries are included in the sample, the less likely the convergence result appears. Figure 1B confirms this proposition. I ranked the 95 countries in the order of [Y.sub.0], and performed the linear regression of GRO on [Y.sub.0] for the sample with 15 highest income countries. Then I ran the same regression for the sample with an additional next highest income country at a time. Figure 1B shows how the coefficient estimate increases with the number of countries included in the sample. Perhaps it is well known that there is strong evidence for the convergence across developed countries such as OECD countries, while there is no such evidence for a wider range of countries. In light of Figure 1B, the convergence result among developed countries may be considered as a local manifestation of the global humped pattern.

   Conditional Convergence and Humped Pattern in Growth

   Even for a wide range of countries, the growth rate appears to be negatively correlated with income per capita if some variables are controlled for to capture heterogeneities across countries. Barro [2], Mankiw, Romer and Weil [18], and many others have found this result, called the conditional convergence.

   Table IIIA, Regression 1, reports the results of the regressions using the most commonly used three control variables: the investment to GDP ratio (INV, a proxy for a physical capital accumulation rate), the percentage of age group enrolled in secondary education (EDU, a proxy TABULAR DATA OMITTED for a human capital accumulation rate), and the average annual rate of population growth (POP). The population growth rate does not appear significant at all in the first regression. This result may not be...