A. The Econometric Model

We begin with the following panel data model that can nest much of the existing work on the empirics of economic growth, from the “Barro cross-sectional regression” to the static and dynamic panel data techniques:

where Δ y_it is the growth rate of real GDP per capita in country i; and y_it–1 is the logarithm of lagged real GDP per capita. x_it is a vector of explanatory variables; η_t is the time-specific effect; μ_i is the country-specific effect; and “ε_it is the error term.

Much of the empirical growth literature is based on estimations of an equation similar to (1) using a cross-sectional approach, but the drawbacks of this method are well known. Cross-sectional regressions clearly suffer from endogeneity problems as by construction, the initial level of income, y_it–1, is correlated with the error term, ε_it. This endogeneity bias is larger when considering the simultaneous determination of virtually all growth determinants. Furthermore, substantial bias may be induced by the correlation of unobserved country-specific factors and the explanatory variables.

Traditional static panel data estimators such as fixed and random effects are not consistent in the present context either, due to the inclusion of lagged dependent variables in our regressions (e.g. the initial level of GDP per capita). More specifically, the fixed effects estimator is inconsistent because it usually eliminates μ_i by a de-meaning transformation that induces a negative correlation between the transformed error and the lagged dependent variables of order 1/T, which in short panels remains substantial. The assumption of a lack of correlation between μ_i and the explanatory variables required for random effects consistency is also violated as both Δy_it and y_it–1 are functions of μ_i. Furthermore, these estimators will be inconsistent if the errors show either heteroscedasticity or serial correlation.

As it is expressed in equation (1), we specify our growth regression dynamically and include lagged GDP per capita on the right hand side. In this case, the elimination of fixed effects from the equation in any standard OLS-based estimation procedure implies the violation of the orthogonality condition between the error term and explanatory variables. For this reason, we estimate this equation with the system GMM procedure (discussed below in Section B) and contrast it with a cross-sectionally augmented pooled mean group (CPMG) approach. The system GMM procedure accounts for the endogeneity bias induced by reverse causality running from GDP per capita growth to its determinants. The CPMG estimator has a number of methodological and conceptual advantages over the classical panel data approaches (see a detailed discussion below in Section C), and thus strengthens our conclusions.

B. GMM Methodology

To correct for the biases created by lagged endogenous variables and the simultaneity of growth determinants, we use the generalized-method-of-moments (GMM) estimators developed for dynamic panel data models. Following Anderson and Hsiao (1982), and Arellano and Bond (1991), we take first-differences of equation (1) to eliminate the unobserved country fixed effects, μ_i, yielding:

The first-difference of equation (1) gives the transformed error a moving-average structure that is correlated with the differenced lagged dependent variable. Assuming that the error term, ε_it, is not serially correlated and that the explanatory variables x_it are weakly exogenous,³ the difference GMM estimator uses the following moment conditions:

However, in growth regressions where the explanatory variables are persistent over time, lagged levels are often weak instruments for difference equations.⁴ To reduce the potential biases and imprecision associated with the GMM difference estimator, we follow Arellano and Bover (1995) and Blundell and Bond (1998) in employing a system estimator that also includes equation (1) in levels, with the lagged differences of the endogenous variables as instruments (see Levine et al. (2000) and Aghion et al. (2009) among others for applications of this technique). These are appropriate instruments under the assumption that there is no correlation between the differences of the variables and the country-specific effects. Therefore, the additional moment conditions for the regression in levels are:

The moment conditions effectively give us T—1 equations in first differences followed by T equations in levels. The solutions to these equations are then weighted by the inverse of a consistent estimate of the moment condition covariance matrix in a two-step method. To test the validity of the instruments and therefore consistency of the GMM estimator, we consider two specification tests suggested by Arellano and Bond (1991), Arellano and Bover (1995), and Blundell and Bond (1998). The first is a Hansen test of over-identifying restrictions, which tests the overall validity of the instruments, and the second test examines the hypothesis that the error term ε_it is not serially correlated.⁵

The system GMM approach effectively deals with the endogeneity problem and country-specific fixed effects (by allowing the intercepts to differ across cross-sections). However, it restricts all the slope coefficients to be identical across countries; assumes that the time effects are homogenous; and that the errors are cross sectionally independent. If any of these conditions are not satisfied, the GMM method can produce “inconsistent and potentially very misleading” estimates of the average values of parameters; see Pesaran and Smith (1995) for more details. The time-specific heterogeneity is an underestimated but at the same time very important concern in dynamic panel data models. Country-specific time-effects can capture a number of unobservable characteristics in macroeconomic and financial applications like (a) institutional arrangements, (b) the patterns of trade, (c) political developments, and (d) the effect of WTO, to mention a few. The time-specific heterogeneity is induced by oil price shocks and/or other global common factors, which affect all countries but to different degrees. The CPMG methodology explained below in Section C accounts for heterogenous time effects and at the same time deals with cross-sectional dependencies effectively.

C. CPMG Methodology

When panels of data are available, there exist a number of alternative estimation methods that vary on the extent to which they account for parameter heterogeneity. At one extreme is the Mean Group (MG) approach in which separate equations are estimated for each country and the average of estimated coefficients across countries is examined. Pesaran and Smith (1995) show that the MG method produces consistent estimates of the average of the parameters when the time-series dimension of the data is sufficiently large. At the other extreme are the traditional estimators in which dynamics are simply pooled and treated as homogeneous. Early and prominent examples include fixed effects (FE), random effects (RE), and generalized methods of moments (GMM), described in Section III.B. These methods are typically focused on solving the problem of fixed effect heterogeneity in the case of large N and small T panels; whereas they are not designed to correct for the endogeneity induced by the latent heterogeneity. Pesaran and Smith (1995) show that the traditional procedures for the estimation of pooled models can produce inconsistent and potentially misleading estimates of the lagged dependent variable’s parameter in dynamic panel data models if latent heterogeneity is present.

In between the two extremes is the pooled mean group (PMG) estimator of Pesaran et al. (1999) which is an intermediate case between the averaging and pooling methods of estimation, and involves aspects of both. It restricts the long-run coefficients to be homogenous over the cross-sections, but allows for heterogeneity in intercepts, short-run coefficients (including the speed of adjustment) and error variances. The PMG estimator also generates consistent estimates of the mean of short-run coefficients across countries by taking the simple average of individual country coefficients. It can be argued that country heterogeneity is particularly relevant in short-run relationships, given that countries are affected by over lending, borrowing constraints, and financial crises in short-time horizons, albeit to different degrees. On the other hand, there are often good reasons to expect that long-run relationships between variables are homogeneous across countries. Estimators that impose cross-sectional restrictions (PMG) dominate the fully heterogeneous ones (MG) in terms of efficiency if the long-run restrictions are indeed valid. If the constraints are not valid, however, the restricted estimators are inconsistent.⁶

We make use of the PMG estimator because it offers the best available choice in terms of consistency and efficiency in our sample of countries, while it corrects at the same time for the shortcomings of homogeneous panel methods mentioned above. Moreover, we apply the methodology of Pesaran (2006) to the PMG estimator to correct for the cross-sectional dependencies that arise in the error terms from unobserved global factors, since we assume that countries are affected in different ways and to varying degrees by these shocks.⁷

The cross sectionally augmented pooled mean group (CPMG) estimator is based on an Autoregressive Distributive Lag (ARDL) model and thus can be used for long-run analysis. In a series of papers, Pesaran and Smith (1995), Pesaran (1997), and Pesaran and Shin (1999) show that one can use the ARDL approach to produce consistent and efficient estimates of the parameters in a long-run relationship between both integrated and stationary variables, and to conduct inference on these parameters using standard tests. This method avoids the need for pre-testing the order of integration given that they are valid whether the variables of interest are I(0) or I (1). The main requirements for the validity of this methodology are that, first, there exists a long-run relationship among the variables of interest and, second, the dynamic specification of the model is sufficiently augmented so that the regressors become weakly exogenous and the resulting residual is serially uncorrelated.

To explain the CPMG estimator in more details, consider the following ARDL (p, q, q, …, q) model:

where as before i = 1, 2, …, N, t = 1, 2, …, T, x_it is the κ × 1 vector of explanatory variables for group i, μ_i represents the fixed effects; the coefficients of the lagged dependent variables, λ_ij, are scalars and δ_ij are κ × 1 coefficient vectors. We assume that the error term, u_it, has the following multi-factor error structure:

where f_t is a vector of unobserved common shocks, which can be stationary or nonstationary; see Kapetanios et al. (2011). The source of error term dependencies across countries is captured by f_t, whereas the impacts of these factors on each country are governed by the idiosyncratic loadings in γ_i. The individual-specific errors, ε_it, are distributed independently across i and t; they are not correlated with the unobserved common factors or the regressors; and they have zero mean, variance greater than zero, and finite fourth moments. The common factors, or the heterogenous time effects, may be captured/proxied by adding cross sectional averages of the observables to our regressions, see Pesaran (2006) and Binder and Offermanns (2008).

More specifically, combining (3) and (4) and averaging across i yields

where the variables with a bar denote the simple cross section averages of the corresponding variables in year t. Since the error term ε_it is by assumption independently distributed across i and t, its cross-sectional average, ε_t, tends to zero in root mean square error as N becomes large. The common factors can therefore be captured through a linear combination of the cross-sectional averages of the dependent variable and of the regressors:

Using (6) in (4), the error correction representation of (3) becomes:

where

As always T must be large enough so that the model can be estimated for each cross-section. In addition, the roots of equation (7) in y_it must lie outside the unit circle to ensure that Φ_i < 0, and hence that there exists a long-run relationship between y_it and x_it defined by

Finally, the long-run coefficients on x_it, defined by ${\theta }_i=\left(\frac{{\beta }_i}{{\mathrm{\Phi }}_i}\right)$ above, are restricted to be the same across countries, namely:

The CPMG estimator uses a maximum likelihood approach to estimate the model based on the Newton–Raphson algorithm. The lag length for the model can be determined using, for instance, the Schwarz Criterion (SBC) and the null of long-run homogeneity:

can be tested using the Hausman statistic for the coefficient on each of the explanatory variables and for all of them jointly.

A. Commodity Terms of Trade

Our country-specific measure for the CTOT index is from Spatafora and Tytell (2009), and is defined as:

where MUV_t is a manufacturing unit value index used as a deflator, X_ij (M_ij) is the share of exports (imports) of commodity j in country i’s GDP, and P_jt is the individual commodity price.¹¹ By construction, the movements in the CTOT index are due to changes in commodity prices as the export and import shares are taken as fixed and so remain constant over time. For empirical application, we calculate X_ij and M_ij as the average value of these shares between 1970 and 2007. The CTOT index allows countries to be influenced by changes in commodity prices differently, depending on the composition of their export and import baskets. This is in contrast to the ‘standard’ commodity price indices most commonly used in the literature, such as the “All Primary Commodities Index” in International Monetary Fund (2010b), which attaches the same weight to each country in the regression analysis. Equation (10) is used to construct two important variables. The first is a commodity terms of trade growth series, a proxy for resource abundance, calculated as the annual log differences in the CTOT index, and the second is a measure of CTOT volatility; both are explained in more detail below.

Table 2.

Definitions and Sources of Variables Used in Regression Analysis

Table 2.

Definitions and Sources of Variables Used in Regression Analysis

Variable	Definition and Construction	Source
Real GDP per Capita	Ratio of GDP (in 2000 US$) to population.
GDP per Capita Growth	Geometric average growth rate of real GDP per capita.	Authors’ construction using data from the World Bank (2010) World Development Indicators (WDI).
Initial GDP per Capita	Initial value of GDP per capita in the beginning of each five-year period.
TFP	Total factor productivity (TFP).
TFP Growth	Geometric average growth rate of TFP.
Initial TFP	Initial value of TFP in the beginning of each five-year period.
Physical Capital per Capita	each five-year period. Ratio of physical capital to population.	Authors’ construction using data from Heston et al. (2009). See Section IV for more details.
Physical Capital per Capita Growth	Geometric growth rate of physical capital per capita.
Initial Physical Capital Per Capita	Initial value of the ratio of total physical capital to total population in the beginning of each five-year period.
Human Capital per Capita	Ratio of human capital to population.
Human Capital per Capita Growth	Geometric growth rate of human capital per capita.	Authors’ construction using data from Barro and Lee (2010). See Section IV for more details.
Initial Human Capital per Capita	Initial value of the ratio of total human capital to total population in the beginning of each five-year period.
Commodity Terms of Trade Growth	Growth rate of commodity terms of trade index.	Authors’ construction based on Spatafora and Tytell (2009).
Commodity Terms of Trade Volatility	Standard deviation of commodity terms of trade growth in five-year interval.
Export Sophistication Measure	A measure of the productivity level associated with a country’s specialization pattern.	Authors’ construction based on Hausmann et al. (2007) and the World Bank (2010) WDI.
Education	Ratio of total secondary enrollment to the population of the age group that officially corresponds to that level of education.	Authors’ construction using data from UNESCO (2010) UIS.
Trade Openness Government Burden	Ratio of Exports and Imports to GDP. Ratio of government consumption to GDP.	Authors’ construction using data from the World Bank (2010) WDI.
CPI	Consumer price index (2000=100) at the end of the year.
Inflation rate	Annual percentage change in CPI.	Author’s calculations using data from the International Monetary Fund (2010a) World Economic Outlook.
Lack of Price Stability	log(100+inflation rate).

View Table

Table 2.

Definitions and Sources of Variables Used in Regression Analysis

Variable	Definition and Construction	Source
Real GDP per Capita	Ratio of GDP (in 2000 US$) to population.
GDP per Capita Growth	Geometric average growth rate of real GDP per capita.	Authors’ construction using data from the World Bank (2010) World Development Indicators (WDI).
Initial GDP per Capita	Initial value of GDP per capita in the beginning of each five-year period.
TFP	Total factor productivity (TFP).
TFP Growth	Geometric average growth rate of TFP.
Initial TFP	Initial value of TFP in the beginning of each five-year period.
Physical Capital per Capita	each five-year period. Ratio of physical capital to population.	Authors’ construction using data from Heston et al. (2009). See Section IV for more details.
Physical Capital per Capita Growth	Geometric growth rate of physical capital per capita.
Initial Physical Capital Per Capita	Initial value of the ratio of total physical capital to total population in the beginning of each five-year period.
Human Capital per Capita	Ratio of human capital to population.
Human Capital per Capita Growth	Geometric growth rate of human capital per capita.	Authors’ construction using data from Barro and Lee (2010). See Section IV for more details.
Initial Human Capital per Capita	Initial value of the ratio of total human capital to total population in the beginning of each five-year period.
Commodity Terms of Trade Growth	Growth rate of commodity terms of trade index.	Authors’ construction based on Spatafora and Tytell (2009).
Commodity Terms of Trade Volatility	Standard deviation of commodity terms of trade growth in five-year interval.
Export Sophistication Measure	A measure of the productivity level associated with a country’s specialization pattern.	Authors’ construction based on Hausmann et al. (2007) and the World Bank (2010) WDI.
Education	Ratio of total secondary enrollment to the population of the age group that officially corresponds to that level of education.	Authors’ construction using data from UNESCO (2010) UIS.
Trade Openness Government Burden	Ratio of Exports and Imports to GDP. Ratio of government consumption to GDP.	Authors’ construction using data from the World Bank (2010) WDI.
CPI	Consumer price index (2000=100) at the end of the year.
Inflation rate	Annual percentage change in CPI.	Author’s calculations using data from the International Monetary Fund (2010a) World Economic Outlook.
Lack of Price Stability	log(100+inflation rate).

To calculate CTOT growth, we first take the logarithm of (10)

Taking the difference of (11), we obtain the annual growth rate of the CTOT index:

which reflects the changes in the basket of real commodity prices in country i scaled by the importance of each commodity j in that economy’s net exports for that particular good, (X_ij—M_ij).

Resource revenue (or rent), being calculated as the production multiplied by price (minus marginal cost), has been used extensively in a number of recent studies in the resource curse literature as a measure of abundance. Given that production levels do not change much over time and are generally persistent, most changes in resource rents or revenues in the short-run (for instance five-years) are due to price fluctuations. Moreover, the Dutch disease phenomenon focuses on the changes in natural resource prices as the main driver of the eventual drag on TFP and output growth. Therefore, the commodity terms of trade growth considered in this paper, which is a weighted measure of changes in commodity prices, can be seen as a proxy for resource abundance, as well.

In contrast to most studies in the growth literature which employ time-invariant measures of volatility, we construct two time-varying measures. First, we consider the five year non-overlapping standard deviation of g_CTOT;it, annual growth rates of the CTOT index,

where S = 4 as we are working with five-year averages. The volatility of g_CTOT;it, given in equation (13), indicates the extent to which CTOT growth deviates from a given mean at any point in time. Second, as annual data on CTOT volatility is required in the cross-sectionally augmented pooled mean group (CPMG) regressions, we estimate a generalized autoregressive conditional heteroscedasticity (GARCH) model using the logarithm of CTOT_it. This approach estimates the “conditional variance” of the logarithm of the CTOT for each year, independent of other observations. The computed variance series might yield periods with different volatility levels, and therefore a time varying measure. More specifically, we estimate the volatility of the commodity terms of trade from a GARCH(1,1) model on annual observations using a regression of the change in the logarithm of the CTOT variable, g_CTOT;it, on a constant (this formulation is used to avoid prejudging the issue of stationarity) as in Bleaney and Greenaway (2001) and Serven (2003):

where ${\xi }_{it}\sim N(0,{\sigma }_{it}^2),{\xi }_{it-1}^2$ is the squared residuals, ${\sigma }_{CTOT,it}^2$ is the conditional variance of g_CTOOT;it, ${\sigma }_{CTOT,i}^2$ is the unconditional variance, λ₁ is the ARCH parameter, and λ₂ is the GARCH parameter. We calculate CTOT volatility as the square root of ${\sigma }_{CTOT,it}^2$.

The upper graphs in Figure 1 illustrate a simple bivariate relationship between GDP per capita growth and CTOT growth over the entire period 1970–2007, suggesting a mild positive correlation between these two variables for both country groups. Examining the two lower graphs, we observe that while higher CTOT volatility is associated with lower GDP growth in primary commodity exporting countries, this relationship does not hold for the other subsample, which has a more diversified export structure. Overall, the results from Figure 1 represent preliminary evidence that while commodity booms do not reduce output per capita growth (contrary to the resource curse hypothesis), the volatility of CTOT stunts output growth only for primary commodity exporters. This is perhaps not surprising as those countries with a diversified basket of exports, especially manufacturing or service-sector goods, can be expected to grow faster and be better insured against price fluctuations in individual commodities.

In Section V.A, we will add a whole range of control variables and deal with possible endogeneity problems through the system GMM approach, to investigate whether the above results survive for the full sample and the two subsamples, as suggested by Figure 1. We will also investigate the relationship between resource abundance and/or CTOT volatility with that of output growth using annual data and applying a cross-sectionally augmented version of the Pooled Mean Group (CPMG) methodology in Section V.B. Since we also would like to investigate possible mechanisms through which CTOT volatility can harm economic growth, we focus on three channels which have been widely discussed in the literature: (i) TFP growth; (ii) physical capital accumulation; and (iii) human capital acquisition. To do this analysis, we need to construct series for physical and human capital stocks as well as for TFP. In what follows, we briefly describe how these series are constructed.

B. Physical Capital Accumulation

We apply the perpetual inventory method, as in Hall and Jones (1999) for instance, to data from the Penn World Tables (PWT) 6.3; see Heston et al. (2009), to construct the series of the physical capital stock, K_it. We construct the initial stock of capital, K_it0, for country i as:

where δ is the depreciation rate, _gI is the geometric average growth rate of I_it between t₀ and t₀ + 10, and I_it represents gross investment and is defined as:

in which ki_it measures the investment share of real GDP per capita (rgdpch_it) and pop_it is population. Since we have access to data on investment from 1960 for most countries, we set t₀ to this year.¹² Furthermore, we assume a depreciation rate, δ, of six percent and compute the subsequent values of the capital stock as:

C. Human Capital Stock

To calculate the level of human capital stock in country i, we obtain data on the average years of schooling attained (total, primary, secondary, tertiary) in five-year intervals from the Barro and Lee Educational Attainment Dataset 2010. Since annual data is required to retrieve the human capital series, we linearly interpolate the Barro and Lee (2010) dataset. Moreover, we assume that labor is homogeneous within a country and that each unit of labor has s_it years of schooling (education). Therefore, the labor-augmenting human capital is given by:

Following Psacharopoulos (1999),¹³ we specify ψ(s_it) as a piecewise linear function with coefficients (returns to schooling) 0.134 for the first four years of education, 0.101 for the next four years, and 0.068 for any value of s_it > 8.¹⁴

D. Productivity

In constructing the total factor productivity series, we follow Hall and Jones (1999) and assume that output in country i is produced according to the following constant returns to scale production function:

where K_it denotes the stock of physical capital defined in (17), A_it is a labor-augmented productivity factor, H_it is a measure of the average human capital of workers defined in (18), and L_it is labor input in use:

where as before rgdpch_it is real GDP per capita, pop_it is a measure of population and rgdpwok_it is real GDP per worker from the PWT 6.3. The capital share, α, is assumed constant across countries and set equal to 1/3.

Finally, using the data on output per worker, capital, population, and schooling, we can construct the level of total factor productivity as follows:

A. Analysis Using Five-Year Averages

To filter out business cycle fluctuations and to focus on the long-run effects of CTOT growth and volatility, we follow the literature in transforming the annual series into non-overlapping five-year averages. Given the time span of our dataset (from 1970 to 2007), we can construct an unbalanced panel with a maximum of seven five-yearly observations per country covering 1970–2005.

A.1 Volatility and Growth

We propose to use the system GMM estimator described above, but as the two-step standard errors on estimated coefficients will be biased downward in small samples like ours, we make use of Windmeijer (2005) approach to correct for that bias. The following equation is estimated:

$\begin{array}{cc}\hfill g_{y,is}& =(\mathrm{\Phi }-1)y_{is-1}+{\gamma }_1{g}_{CTOT,is}+{\gamma }_2{\sigma }_{CTOT,is}\hfill \\ & +{\gamma }_{3}EXP{Y}_{is}+{\beta }^{\prime }{z}_{is}+{\mu }_i+{\eta }_s+{ϵ}_{is,}\left(22\right)\hfill \end{array}$

where i = 1, 2, …, N, and s = 1, 2, …, S, in which $S=\frac{T}{5},$ with T denoting the years between 1970 and 2005. g_I, is the geometric average growth rate of real GDP per capita between dates s and s–1; y_is–1 is the logarithm real GDP per capita at the beginning of each period; g_CTOT,is is the growth rate of the CTOT index; and σ_{CTOT, is} is its volatility. EXPY_is is a measure of export diversification and z_is is a set of other control variables now standard in the growth literature¹⁵ including education, trade openness, government burden, and lack of price stability. η_s is the time-specific effect; μ_i is the country-specific effect; and ε_it is the error term.

Table 3 presents the estimation results of the impact of commodity terms of trade growth and volatility as well as export diversification on GDP per capita growth. In the first regression using the whole sample of 118 countries, [1.1], we observe that an increase in g_CTOT is both growth enhancing and highly significant. On the other hand, although the coefficient of CTOT volatility is negative, this is in fact insignificant and thus there is no evidence that volatility in commodity prices harms growth for the full sample. As already discussed in Section IV.A, we expect the growth experience of primary commodity exporters to be different from those countries that are not well endowed with a handful of primary products, see Figure 1, and therefore we split the sample into two subsets.

Table 3.

Growth Effects of CTOT Volatility I

Notes: Time and fixed effects are included in all the regressions. Standard errors are presented below the corresponding coefficients in brackets. Symbols ***, **, and * denote significance at 1%, 5%, and at 10% respectively. The impact of CTOT growth and volatility measures the overall average percentage effect of the two CTOT variables on output per capita growth. Source: Authors’ estimations.

Table 3.

Growth Effects of CTOT Volatility I

Estimation Method:		Two-step system GMM with Windmeijer (2005) small sample robust correction.
Unit of Observation, Period: Volatility Measure:		Non-overlapping five-year averages, 1970 - 2005. Five year standard deviation of annual CTOT growth.
		[1.1]	[1.2]	[1.3]
The Dependent Variable is the Growth Rate of Output per Capita		All 118 Countries	62 Commodity Exporters	Other 56 Countries
Initial Output per Capita, in logs		-1.204**	-0.872	-1.738***
		(0.471)	(0.688)	(0.546)
Commodity Terms of Trade Growth		0.240***	0.255***	-0.156
		(0.072)	(0.078)	(0.469)
Commodity Terms of Trade Volatility		-0.105	-0.119**	-0.683
		(0.081)	(0.058)	(0.577)
Export Sophistication Measure, in logs		4.818***	2.787*	3.687**
		(1.830)	(1.638)	(1.465)
Control variables
Education		0.812	1.256	0.054
(secondary enrollment, in logs)		(0.803)	(0.960)	(1.380)
Trade Openness		2.027**	2.587***	2.142**
(trade volume/GDP, in logs)		(1.024)	(0.860)	(0.929)
Government Burden		-2.656**	-4.007***	-0.109
(government consumption/GDP, in logs)		(1.163)	(1.064)	(1.536)
Lack of Price Stability		-6.786***	-6.264**	-11.119***
(log [100 + inflation rate])		(2.412)	(2.485)	(3.773)
Intercept		-1.568	10.872	29.013
		(17.131)	(16.023)	(19.256)
No. Countries/No. Observations		118/664	62/352	56/312
Specification tests (p-values)
(a) Hansen Test		0.121	0.448	0.314
(b) Serial Correlation
	First-order	0.000	0.000	0.003
	Second-order	0.199	0.252	0.674
Impact of CTOT Growth and Volatility		-	-0.312	-

Notes: Time and fixed effects are included in all the regressions. Standard errors are presented below the corresponding coefficients in brackets. Symbols ***, **, and * denote significance at 1%, 5%, and at 10% respectively. The impact of CTOT growth and volatility measures the overall average percentage effect of the two CTOT variables on output per capita growth. Source: Authors’ estimations.

View Table

Table 3.

Growth Effects of CTOT Volatility I

Estimation Method:		Two-step system GMM with Windmeijer (2005) small sample robust correction.
Unit of Observation, Period: Volatility Measure:		Non-overlapping five-year averages, 1970 - 2005. Five year standard deviation of annual CTOT growth.
		[1.1]	[1.2]	[1.3]
The Dependent Variable is the Growth Rate of Output per Capita		All 118 Countries	62 Commodity Exporters	Other 56 Countries
Initial Output per Capita, in logs		-1.204**	-0.872	-1.738***
		(0.471)	(0.688)	(0.546)
Commodity Terms of Trade Growth		0.240***	0.255***	-0.156
		(0.072)	(0.078)	(0.469)
Commodity Terms of Trade Volatility		-0.105	-0.119**	-0.683
		(0.081)	(0.058)	(0.577)
Export Sophistication Measure, in logs		4.818***	2.787*	3.687**
		(1.830)	(1.638)	(1.465)
Control variables
Education		0.812	1.256	0.054
(secondary enrollment, in logs)		(0.803)	(0.960)	(1.380)
Trade Openness		2.027**	2.587***	2.142**
(trade volume/GDP, in logs)		(1.024)	(0.860)	(0.929)
Government Burden		-2.656**	-4.007***	-0.109
(government consumption/GDP, in logs)		(1.163)	(1.064)	(1.536)
Lack of Price Stability		-6.786***	-6.264**	-11.119***
(log [100 + inflation rate])		(2.412)	(2.485)	(3.773)
Intercept		-1.568	10.872	29.013
		(17.131)	(16.023)	(19.256)
No. Countries/No. Observations		118/664	62/352	56/312
Specification tests (p-values)
(a) Hansen Test		0.121	0.448	0.314
(b) Serial Correlation
	First-order	0.000	0.000	0.003
	Second-order	0.199	0.252	0.674
Impact of CTOT Growth and Volatility		-	-0.312	-

Notes: Time and fixed effects are included in all the regressions. Standard errors are presented below the corresponding coefficients in brackets. Symbols ***, **, and * denote significance at 1%, 5%, and at 10% respectively. The impact of CTOT growth and volatility measures the overall average percentage effect of the two CTOT variables on output per capita growth. Source: Authors’ estimations.

Regression [1.2] shows the opposite significant effects of g_CTOT and σ_CTOT on GDP growth for the 62 primary commodity exporting countries in our sample. While commodity price booms significantly increase economic growth, volatility affects it negatively. The positive growth effect of g_CTOT provides evidence against the traditional resource curse hypothesis, which argues that it is the level of resource abundance that affects economic growth negatively. Our findings are in line with results obtained in a number of recent studies in the literature such as Brunnschweiler and Bulte (2008), Cavalcanti et al. (2011a), Cavalcanti et al. (2011b), and van der Ploeg and Poelhekke (2010). The negative relationship between volatility and growth in resource-abundant countries is also documented in van der Ploeg and Poelhekke (2009) and van der Ploeg and Poelhekke (2010), who acknowledge that the source of the resource curse is the volatility of commodity prices as opposed to resource abundance, although their empirical analysis is based on the volatility of unanticipated output growth and not of commodity prices. σ_CTOT is a more appropriate measure to analyze the resource curse paradox as it directly affects a country’s ability to extract from its resource stock and make use of the proceeds. Whereas the volatility of unanticipated output growth is most likely caused by factors that are not directly related to the abundance of natural resources.

To determine the overall growth impact of changes in the CTOT variable and its volatility, we calculate the average percentage effect of the two CTOT variables on output per capita growth using the estimates from regression [1.2]. The overall effect is -0.312 over five years (see Table 3), therefore the negative growth effects of CTOT volatility offset the positive impact of commodity booms, which suggests that volatility, rather than abundance per se, drives the resource curse paradox.¹⁶

These results, however, do not hold for the second subsample that focusses on the remaining 56 non-resource abundant countries; see regression [1.3] of Table 3. For these countries, changes in commodity prices (or their volatility) are not expected to have any major impact on their physical and human capital investment, or their TFP growth rates. Consequently, there should be no significant impact on economic growth stemming from changes in the CTOT variable or its volatility. In addition, these countries generally have highly diversified export and import baskets, implying that the changes in commodity prices should have a lower effect on them as opposed to primary commodity abundant countries. This argument is also supported by observing that the coefficient of export diversification variable, EXPY_it, is significant and positive in all three regressions in Table 3. This finding suggests that diversifying away from exporting only a handful of primary commodities towards technology improving exports can significantly increase the growth rate of an economy. Another related important reason for this asymmetric effect is that most of these economies are financially developed and have access to international capital markets, and so, are well positioned to absorb the shocks from commodity price variations. For instance, 27 out of the 56 countries in the net-commodity importing sample are members of the Organization for Economic Cooperation and Development (OECD).

Note that in all three regressions, the control variables have the expected signs and are all statistically significant except for the education variable in all regressions, and the government burden variable in [1.3]. Overall, while higher level of trade-openness is growth enhancing, price instability and government burden tend to have adverse effects on GDP growth. In addition, there is evidence of income convergence across countries with the coefficient on the lagged-dependent variable being significant for the full sample and the sample consisting of net primary commodity importers. However, this finding should be interpreted with caution as there is a large cross-country heterogeneity in our sample of 118 countries which might render the estimated coefficient on ln y_it–1 biased. Finally, in all regressions, the Hansen and second order serial correlation test statistics, which examine the validity of the instruments used, are well above the conventional significance levels.