Losing ground: Latin American growth from 1955 to 1999.

• Author: Grier, Robin

1. Introduction

   Several recent papers document the fact that per capita gross domestic product (GDP) income is not growing very fast in Latin America, even after more than a decade of structural reform. (1) Unfortunately for the region, this lack of convergence is not just a recent phenomenon; relative per capita income in Latin America went from 43% of the rich country average in 1955 to only 24% in 2000. (2) Not only is the region as a whole falling behind, but the fall in relative per capita income is not being driven by any particular outlier country.

   In this paper, I investigate the reasons behind the divergence of average incomes in Latin America and the rich, industrialized countries within the framework of the neoclassical growth model. I first determine whether factor accumulation was high enough in the Latin American countries to have theoretically brought about consistently high GDP growth. Specifically, I estimate an augmented Solow growth model for 22 rich countries for the years 1955-1999. I then combine the coefficients from that estimation with actual factor accumulation in Latin America to forecast how fast a representative rich country would have grown with the same level of accumulation as the Latin American country. (3) If simulated income growth is high (and, in addition, consistently higher than actual growth), then other factors besides factor accumulation are probably a major reason behind a country's low growth.

   I find that actual per capita income growth is consistently well below simulated growth for all of the countries in the sample. Even countries that had long periods of sustained income growth still grew slowly relative to their potential. (4) For instance, average annual per capita income growth in Mexico and Brazil from 1960 to 1979 was 1.97% and 2.85%, respectively. While these rates indicate sustained income growth for two decades, they are still significantly below the model's simulated ones (3.91% for Mexico and 5.33% for Brazil). Their economic growth was relatively high, but given their factor accumulation, these two countries were actually underperforming. I go on to investigate possible causes for the gap between actual and forecasted growth. Specifically, I find that low levels of ethnolinguistic diversity, high import growth, low government consumption spending, and autocracy are all negative and significantly related to the simulation errors. Countries that have tropical climates or are landlocked have higher average simulation error. In addition, I find that the composition of exports matters. The richer countries in the sample that had high manufacturing shares of exports also had lower average simulation errors. Educational quality also helped to explain differences across forecast errors. Countries with lower secondary school class sizes (relative to the rich country average) also had lower simulation errors.

   In section 2, I describe the methodology of the paper in more detail. Section 3 estimates an augmented Solow model of growth for the rich countries and uses the coefficients of that model to simulate growth in Latin America. Section 4 investigates specific reasons why Latin American countries' actual growth was consistently less than simulated growth, while section 5 tests whether these factors help to explain the missing growth gap. Section 6 concludes.

2. Methodology

   Grier (2003), in a departure from the traditional method of calculating productivity via growth accounting, estimates a version of the augmented Solow model for 21 industrialized countries for the 1960-1990 period and uses those coefficients (combined with actual factor accumulation in East Asia) to simulate per capita income growth in East Asia during that time.

   As she points out, industrialized countries are good benchmarks for estimating an augmented Solow model because they are the most likely to be close to their steady-state levels. While no one debates whether Latin American growth is "miraculous" or not, I employ the same method to determine whether Latin American countries are significant underachievers.

   Mankiw, Weil, and Romer (1992) derive the following cross-country growth regression, where [s.sub.k] and [s.sub.h] represent the income shares invested in physical and human capital, d is the depreciation rate, n is the population growth rate, and g is the growth of labor-augmenting technology. [lambda] represents the speed at which a country's output converges to its steady-state level.

   [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (1)

   While the sum of technology growth and depreciation (g + d) is typically assumed to be constant in panel estimations, Maddison (1987) argues that the rate of technological growth in the Organization for Economic Cooperation and Development (OECD) was significantly higher in the pre-1973 era, meaning the assumption of a constant rate of technological growth would be inappropriate. I assume that g + d sum to 0.05, but I include time dummies in the estimation, which should help to account for any unmodeled changes in the rate of technology growth. I also divide the sample into two subsamples (1955-1974 and 1975-1999) and find that I cannot reject the null hypothesis of parameter stability at the 0.10 level.

   Another potential problem in estimating the above regression is the possibility that investment is not exogenous. In the empirical growth literature, there has been more and more recognition that investment is likely to be an endogenous variable in growth regressions. (5) For that reason, I use a generalized method of moments (GMM) estimator with a heteroskedasticity and autocorrelation consistent weighting matrix and one lag of investment to identify the equation. In the next section, I discuss the data and also the results of combining rich country coefficients with Latin American inputs to forecast growth in the region.

3. Results

   Estimating a Baseline Model with Rich Country Data

   In this section, I estimate Equation 1 for a panel of 22 rich countries from 1955 to 1999. The data are averaged into five-year intervals, capturing information from average cross-country differences and temporal fluctuations over the sample period. (6) All economic data are from the Penn World Tables 6.1. (7) The education variable is the average years of secondary schooling in the population older than 25 years of age divided by the total possible numbers of years of secondary schooling. (8) The data for the numerator come from Barro and Lee's (2001) updated education data, while the data on total years of schooling across countries come from UNESCO's Statistical Yearbook on Education (UNESCO 2004).

   I treat investment as an endogenous regressor in the model by using a GMM estimator with a heteroskedasticity and autocorrelation consistent weighting matrix and one lag of investment to identify the equation. (9) I weight the autocovariances with a Bartlett kernel and use the formula of Newey and West (1987) to determine the optimum fixed bandwidth (which turns out to be 4). I find that the augmented Solow model explains over half of the variation of growth rates in the sample, with an [R.sub.2] of 0.56. The coefficient on lagged per capita income is negative and significant at the 0.01 level, supporting the common finding of convergence in industrialized samples. Investment and secondary education are positive and significantly correlated with per capita income growth at the 0.05 and 0.01 level, respectively. The coefficient on n + g + d. however, is insignificantly different from zero. From Equation 1 above, we know that the coefficients of investment, human capital, and n + g + d should add up to zero. I perform a Wald test of the null hypothesis that these coefficients add up to zero and find that I cannot reject the null standard significance levels. Given this, and the fact that the coefficient on n + g + d is very imprecisely estimated, I reestimate the equation imposing the Solow restriction that the three coefficients sum to zero. The coefficients on investment, education, and n + g + d are now 0.0138, 0.00569, and 0.0195, respectively. I will use these coefficients to simulate Latin American growth in the next section.

   Simulated Latin American Growth

   I combine the coefficients from the rich country model with actual factor accumulation in Latin America to calculate how fast a representative rich country would have grown during the 1955-1999 time period with the same levels of accumulation. Following Grier (2003), I look at the size and composition of the simulation errors to determine whether countries are systematically overachieving or underachieving. First, I use the root mean squared error of the forecast as a percentage of average actual growth (hereafter, RMSPE) as a measure of the size of the error. A high RMSPE indicates that the simulated growth is a poor indicator of how fast a country really grew during the time period.

   Second, I use the proportion of the simulation error that is due to bias to determine whether a country's growth forecast is due to systematic differences between the actual and the simulated series. Theil (1961) showed that simulation error can be decomposed into three groups. The first category, which is the error due to a difference in the mean value of the simulated and actual series, is the one that is of most interest for the purposes of this paper. (10) The simulated series may actually replicate the volatility and turning points of the actual series quite closely, but if the means are different, then the forecasted rates will be consistently higher or lower than the actual rates.

   I plot actual and simulated income growth for each country to determine whether the simulated growth rates are higher than actual rates. I will categorize a country as an underachiever if the RMSPE of the forecast is greater than 50%, if 50% or more of the forecast error is due to systematic bias, and if forecasted growth is...