Trade and economic development: global causality and development- and openness-related heterogeneity

Abstract
This paper reconsiders the classic relationship between trade and economic development. We examine the short-term and long-run dynamics between trade and income for 167 countries over the period 1970–2011 and assume that the effect is not homogenous for all countries but rather varies according to the development stage and the degree of trade openness. We apply panel cointegration, Granger causality and panel error correction in combination with Dynamic Ordinary Least Squares and General Method of Moments estimation to explore the causal relationship between these two variables. The results suggest a statistically significant positive short-run and long-run global relationship between trade and income. However, when splitting the panel into different income and trade openness groups, a long-run relationship is observed only for high-income countries and countries with a relatively high degree of trade openness.


Introduction
The relationship between international integration and economic development has long been a subject of interest and even controversy. 1 With regard to the theoretical relationship between trade and development, most studies support the proposition that openness to international trade positively affects income. Romer (1993), Helpman (1990, 1991), and Barro and Sala-i-Martin (1995), among others, argue that countries that are more open to trade are better able to catch up to the leading technologies. International trade promotes the efficient allocation of resources through comparative advantages and allows the dissemination of knowledge and technological progress through high-tech imports (Chang et al., 2009;Baldwin et al., 2005). Additionally, trade openness encourages competition in domestic and international markets and allows to better reap the potential benefits of increasing returns to scale and economies of specialization by increasing the market size (Melitz and Ottaviano, 2008;Alesina et al., 2005).
However, the empirical evidence on the relationship between trade and income varies a lot, depending on the econometric technique used, the sample studied, and the exact definition of the variables used. 2 Further, recent developments suggest that trade openness is not always beneficial to economic growth and that the relationship between trade and development may vary with the level of development and trade.
First, due to technological or financial constraints, a lack of human capital, brain drain and weak institutions, less developed countries may fall behind in adopting and using new technologies and thus not be able to fully benefit in the advantages of trade. Due to this limitation in technology and knowledge transfer the countries stay on a technology and income trap where they continue to produce basic goods and services that do not encourage innovation and enable structural transformation and growth. Empirical evidence suggests that the relationship between trade and growth may vary with economic development. Analyzing the relationship between trade and development Kim and Lin (2009) and Kim et al. (2011) identify an income threshold above which greater trade openness has beneficial effects on economic growth and below which heightened trade has harmful consequences. In line with this, Sakyi et al. (2011) and Herzer (2013) find that the effect of trade on income tends to be greater in countries with higher income levels. Thus, the growth effect of trade may differ according to the level of economic development.
In our study we reexamine the relationship between trade and development and contribute to the literature by empirically analyzing if the relationship differs between countries depending on both, their level of income and their degree of trade openness. This fragmented examination adds to the existing literature as it enables us to identify differentiated development-and openness related heterogeneity in the relationship based on a wide range of countries and a long time dimension. Specifically, we shed light on whether benefits are contingent upon the development and openness level. Further, our empirical evidence is based on short-term and long-run causality methods that allow the identification of and differentiation between long-term causality and short-term adjustment. In particular, we apply panel based cointegration tests, error correction and Granger causality models using data for 167 countries over the period 1970-2011 and employ Dynamic Ordinary Least Squares and General Method of Moments estimation techniques, which are appropriate in the presence of endogeneity concerns.
We identify a positive long-run relationship between trade and development, however, our findings support the view that the relation between trade globalization and economic growth is heterogeneous for different stages of development and openness.
The paper proceeds as follows. The next section presents the empirical investigation including the data, the methodology, and the results of the panel unit root tests, the panel cointegration tests, the error correction model and the Granger causality tests for the global relationship and the income-and trade share-related subpanels. Section three concludes.

Empirical analysis
This section aims at empirically investigating the causal relationship between trade and economic development. However, a loop of causality between these variables can lead to endogeneity problems. Endogeneity bias can lead to inconsistent estimates and incorrect inferences. Current literature has identified this problem and suggests two common solutions. One solution to address the endogeneity problem is to use instrumental variables (IV) techniques. For example, several studies use geographic attributes as instruments to identify the effects of trade on income. 3 However, there is also criticism that geography-based instruments are too weak to identify the true effect and are invalid, since they are correlated with other geographic variables that affect income through non-trade channels and so reflect only the geography effect on income. Further, their use can lead to multicollinearity problems between geographical controls, institutional proxies, and the trade instruments, which yields statistically insignificant results (Rodríguez and Rodrik, 2001;Herzer, 2013).
Another solution to address the problem of endogeneity is the use of an appropriate estimation technique. Here the Dynamic Ordinary Least Squares (DOLS) estimation accounts for possible serial correlation and endogeneity of the regressors and thus yields unbiased estimates of the long-run relationship. 4 Further, Generalized Method of Moments (GMM) techniques developed by Arellano and Bond (1991) and Blundell and Bond (1998) use an internally generated set of instruments to alleviate endogeneity concerns and control for the dynamic nature of the relationship. 5 Since the availability of time-varying trade instruments on panel level is limited and weak, and invalid instruments lead to spurious results, we abstract from instrumentalization and approach the problem of endogeneity by focusing on DOLS and GMM estimation techniques.

Data
We use a balanced panel data set containing 167 countries over the period 1970-2011. The bivariate analysis is based on two variables, GDP per capita and trade share in GDP.
GDP per capita corresponds to real GDP per person. It is calculated as real GDP at constant 2005 US$ divided by population. Both, the GDP and the population data are drawn from the Penn World Table 8.0 as provided by Feenstra et al. (2013).
In the literature, trade volume measures and trade restriction measures are used to examine the effect of trade openness and development. Since our research question is aimed at investigating whether additional trade encourages additional development we use trade share as the most common measure of the class of trade volume measures. Trade share is measured by the sum of exports and imports as a percentage of GDP at 2005 constant prices and corresponds to the trade openness variable extracted from the Penn World Tables 8.0.
In addition to the overall panel, we segment the data set into eight subpanels; four according to per-capita income and four according to trade share. For the income classification, we use the World Bank country classification and distinguish between low-income (LI) economies ($1045 or less), lower-middle-income (LMI) economies ($1046 to $4125), upper-middle-income (UMI) economies ($4126 to $12745) and high-income (HI) economies ($12,746 or more). Further, we classify the list of countries into countries with low trade openness (LTO), lower-middle trade openness (LMTO), upper-middle trade openness (UMTO) and high trade openness (HTO) by computing the quartiles of the trade share variable. The limits for the quartiles are <0.44 for the first quartile, 0.44-0.66 for the second quartile, 0.67-0.99 for the third quartile and > 0.99 for the fourth quartile. Table 1 shows the descriptive statistics for the global sample and the income-and trade-related subpanels. The descriptive statistics show that from a diversified perspective there is no clear evidence of a positive relationship between trade and development, rather the causality should be assumed to be heterogeneous. First, for the income 4 DOLS are usually preferred to the OLS estimator because they take care of small sample bias and endogeneity bias by taking the leads and lags of the first-differenced repressors. 5 Dynamic panel data models include lagged levels of the dependent variable as regressors. This violates strict exogeneity, because the lagged dependent variable is necessarily correlated with the idiosyncratic error. When the strict exogeneity assumption is violated, commonly used static panel data techniques are inconsistent, because these estimators require strict exogeneity. The System-GMM estimator uses lags of the dependent variable as instruments for differenced lags of the dependent variable (which are endogenous) to solve the endogeneity, heteroscedasticity and serial correlation problems. subpanels the trade share increases with income, but hardly differs in the groups of UMI and HI. Second, the trade share subpanels show that income does not increase simultaneously with the trade share. Here the UMTO group shows the highest average income; the other groups almost do not differ from each other. The descriptive statistics thus show that a causal analysis of the relationship should consider development and openness related differences.

Global causality between trade and economic development 2.3 General methodology
To explore the short-run and long-run dynamics between GDP p.c. and openness we perform panel unit root and panel cointegration tests. The cointegration test shows us whether there is a long-term relationship or not. Therefore, depending on the result of the cointegration test we perform long-run DOLS estimation and panel vector error correction (for the case of cointegration) or short-term Granger-causality tests (for the case of no cointegration). More specifically, our point of departure is a bivariate autoregressive-distributed lag (ADL) model where index i = 1...N refers to the country and t = 1...T to the period. Following Granger (1969) there is Granger causality from x to y if past values of x improve the prediction of y given the past values of y. 6 With respect to the model x Granger causes y if not all δ j are zero. By the same token Granger causality from y to x occurs if not all γ j are equal to zero. However, Engle and Granger (1987) have shown that, if the series x and y are cointegrated, the standard Granger causality test is misspecified. In this case an error correction model should be used instead. Therefore, in a first step we apply unit root and cointegration tests. Based on the results we determine whether to use the Granger causality framework or an error correction technique to test causality. 7 6 Note that model selection criteria, such as the Bayesian Information Criterion (BIC, (Schwarz, 1978)) or the Akaike Information Criterion (AIC, (Akaike, 1974)), can be used to determine the appropriate model order p. 7 Causality in the sense of Granger implies the ability to predict the behavior of a variable using lagged values of another variable. Granger models can only identify short-term relationships.
In contrast error correction models belongs to a category of multiple time series models most commonly used for data where the underlying variables are cointegrated and thus have a long-run stochastic trend. ECMs are a theoretically-driven approach useful for estimating both short-term and long-term effects of one time series on another. The term error-correction relates to the fact that last-period's deviation from a long-run equilibrium, the error, influences its short-run dynamics. Thus, ECMs directly estimate the speed at which a dependent variable returns to equilibrium after a change in other variables.

Panel unit root test
The Granger causality test requires the variables to be stationary. We check their stationarity using a battery of panel unit root tests. We apply the IPS test by Im et al. (2003) and the Fisher-type unit root test developed by  and Choi (2001) which allow for heterogeneous unit root processes among cross sections but assume that the cross sections are independent. However, for macroeconomic variables like GDP per capita and trade cross-sectional dependence is most likely, e.g., through regional linkages, local spillover effects between countries or globally common shocks. The results for the cross-sectional independence test developed by Pesaran (2004) reveal that the hypothesis of cross-sectional independence is rejected at the 1% significance level for both variables. 8 Therefore, we apply also a secondgeneration unit root test developed by Pesaran (2007) allowing for cross-sectional dependence. The results (presented in Table 10 in the appendix) indicate that the null hypothesis of a unit root cannot be rejected for the levels, however the differences are stationary.

Panel cointegration test
Since the panel unit root tests indicate that the variables are integrated of order one I(1), we test for cointegration using the panel cointegration tests developed by Pedroni (2004) as well as Westerlund (2007) and Persyn and Westerlund (2008). The former test allows for heterogeneity in the panel by permitting unit-specific short-run dynamics, heterogenous trend and slope parameters, fixed effects, and individual specific deterministic trends, while the later additionally allows for cross-sectional dependence. Table 2 presents the results and shows that predominantly the null of no cointegration is rejected for both directions of the cointegration relation, indicating that the variables exhibit a cointegration relationship, regardless of the direction of the cointegration equation.

Error correction model
Engle and Granger (1987) have shown that when the series x and y are cointegrated, a standard Granger causality test as presented in the eqs. (1) and (2) is misspecified, because it does not allow for the distinction between the short-run and the long-run effect. Following Stock and Watson (1993) and Pedroni (2001) the long-run relationship between the cointegrated variables can be estimated using the dynamic OLS estimator for heterogeneous cointegrated panels. The estimator accounts for the endogeneity bias, heteroscedasticity and autocorrelation problems that typically occur when estimating dynamic relationships using the traditional OLS estimator and thus provides consistent and efficient estimates. Considering the short-run dynamics, we additionally estimate error correction models (ECM). These are linear transformations of the above ADL models and provide a link between the short-run and the long-run effect.
The standard error correction procedure is a two-step method where first the error correction terms ECT y and ECT x are obtained by saving residuals of separate estimation of the long-run equilibrium of y i,t on x i,t and x i,t on y i,t . In a second step the ECM with the included error correction term can be estimated. While the coefficients α j and δ j as well as β j and γ j capture the short-run effects, the coefficients λ and κ of the error correction terms give the adjustment rates at which short-run dynamics converge to the long-run equilibrium relationship. If λ and κ are negative and significant a relationship between x and y exist in the long run. Table 3 shows the results of the long-run relationships presenting the Pedroni (2001) DOLS estimator. The results reveal that all coefficients have a positive sign and are statistically significant at the 1% level. We find a positive causal relationship between GDP per capita and trade for both directions for the global panel.
In order to infer the causal short-run relationship between GDP per capita and trade openness a panel vector error correction model is specified. The results of the corresponding error correction regressions are summarized in Table 4. They include the summation of the short-run coefficients with the corresponding F-statistic and the ECT coefficient with the corresponding t -statistic. Since the results of the short-run relationship are sensible to the lag length, the Akaike information criterion (AIC) is used to determine the optimal lag length. The first column explores the dynamics of trade on Note: The Akaike information criterion (AIC) is used to determine the optimal lag length. * p < 0.10, ** p < 0.05, *** p < 0.01 GDP per capita and contains the results with reference to eq. (3), while the second column investigates the other direction of causality and is consequently based on eq. (4).
To avoid the problem of biased estimates through possible endogeneity caused by a simultaneity of GDP per capita and trade, we use the System-GMM estimator developed by Blundell and Bond (1998). This estimator effectively deals with the problem caused by endogenous explanatory variables by using their lagged levels as instruments for the difference equation and lagged differences as instruments for the level Note: Standard errors are reported in parentheses. The Akaike information criterion (AIC) is used to determine the optimal lag length. * p < 0.10, ** p < 0.05, *** p < 0.01 Observations 6181 6181 Note: The Akaike information criterion (AIC) is used to determine the optimal lag length. The short-run effect presents the summation of the short-run coefficients; the corresponding F-statistics is presented in squared parentheses. The long-run effect presents the coefficient for the error correction term with the corresponding standard errors in round parentheses. AR2 presents the p value of the Arelano-Bond (1991) test for autocorrelation of second order. Hansen test presents the p value of the J-test for overidentifying restrictions. * p < 0.10, ** p < 0.05, *** p < 0.01 equations. Therefore, we report several test statistics to evaluate instrument validity. In line with Roodman (2009), we also keep the number of used instruments always below the number of cross-sections by, collapsing'the instrument matrix, to avoid the problem of instrument proliferation, which may render GMM estimates invalid. With regard to the first error correction model presented in the first columns, we can identify both, a significant long-run and short-run effect of trade on income. The coefficient of the error correction term gives the adjustment rate at which short-run dynamics converge to the long-run equilibrium relationship. With respect to our specifications, this is the adjustment rate at which the gap between trade share and GDP per capita is closed. Furthermore, the significant error correction coefficient provides evidence of the already tested cointegration relationship between the variables and implies that short-run adjustments in trade will re-establish the long-run equilibrium when there are deviations from long-run equilibrium. The absolute value of the term provides the speed of the short-run adjustment process, indicating that about 5 % (−0.049) of the discrepancy is corrected in each period. Aside of the speed of adjustment, the results indicate the magnitude of the short-run effect. It is measured by the sum of the included lags of the independent variable and tested via an F-test. The coefficient indicates a positive significant causal effect running from trade to GDP per capita.
As for the other direction of the causal relationship presented in the second column of Table 4, the results also reveal a significant short-run and long-run effect running from GDP per capita to trade. Again, the error correction term is negative and significant. This intensifies the long-run relationship with a short-run adjustment to equilibrium that we already found in the first model. At about 48% (−0.481) the speed of adjustment is higher than in the reversed model. The short-run effect of income on trade is also positive, however at a higher level as the effect of the other direction, indicating that the short-run openness response to a temporary income shock is of a higher magnitude as the income response of a temporary openness shock.

Income-and trade-related differences
The question arises whether this globally identified positive connection applies to all countries, regardless of their development stage and their openness to international trade. Following the argument that least developed countries have a lower adoption capability of global know-how and technical knowledge due to e.g. weak institutions, a lack of investment in human capital and of a well-functioning financial system, developing countries may struggle to employ technologies with the same degree of intensity and versatility as developed countries and thus may not be able to benefit equally from opening up to international trade. In this context Blalock and Gertler (2009) find that firms with investments in research and development and such with highly educated employees adopt more technology from foreign entrants than others and Levchenko (2007) argues that development differences captured by differences in institutional quality imply, among other things, that less developed countries may not gain from trade.
Further, following the reasoning that the relationship between trade and economic development may differ according to the level of trade openness, e.g. due to specialization patterns (Hausmann et al., 2007), different trade structures (Madsen, 2007;Saggi, 2002) and diversification patterns (Hesse, 2006;Naudé and Rossouw, 2011;Herzer and Nowak-Lehnmann, 2006), more open countries may benefit to a different extent from the advantages of trade than less open countries. The empirical evidence also points at a possible asymmetry in the effect and shows that below a specific openness threshold greater trade openness has beneficial effects on economic growth and above the trade effect on growth declines (Zahonogo, 2016).
Hence, the literature suggests heterogeneity in the relationship between trade and development with regard to both, the stage of development and the degree of openness. To account for these possible development-and openness-related effects, we subdivide the global panel into income-and openness-related subpanels and carry out the cointegration and causality analysis for each panel individually.
The results for the unit root tests for the income subpanels are presented in Tables 11 and 12 in the appendix. Predominantly, the statistics indicate that -consistent with the results of the global panel -the variables are non-stationary in levels (Table 11), while the differenced data shows stationarity (Table 12). Since the variables are integrated of order one, in a next step we test for cointegration within the subpanels. Table 5 shows the results for the cointegration tests. It becomes apparent that the subpanel results are not homogeneous and differ according to the income level and the degree of openness. Starting with the difference according to the stage of development, first, high-income countries exhibit a cointegration relationship. These groups of countries seem to have a long-run relationship between percapita income and trade share for both directions of the cointegration equation. Second, the vast majority of the test statistics for the lower income subpanels (low-income, lower-middle income and upper-middle income) do not show a significant cointegration relation, with the exception of the trade-to-income direction for the low-middle income subpanel.
Considering the differences according to the degree of openness the picture is similar. We can identify a cointegration relationship for countries with a high degree of openness, while the test statistics (especially when cross-sectional dependence is taken into consideration) are insignificant for the lower quartiles.
Depending on the result of the cointegration test we now apply DOLS estimation and ECM for those subpanels that show a cointegration relationship and Granger causality test for those subpanels without a long-run relationship. The DOLS estimator allows to account for the magnitude and sign of the long-run relationship. The results are presented in Table 6. The coefficients indicate a highly significant long-run relationship running from trade to GDP p.c. and vice    Note: The Akaike information criterion (AIC) is used to determine the optimal lag length. * p < 0.10, ** p < 0.05, *** p < 0.01 versa for all subpanels where we could identify a cointegration relationship. Further, the beneficial effect of trade on income is much higher for high-income countries than for highly open countries. This is also the case for the positive effect of development on trade, where we again can identify the highest effect for the group of high-income countries. In comparison, the effect for lower-middle income countries is only about half as large. For the openness subpanels the effects are relatively close, but the countries with the highest openness benefit less than those in the upper-middle openness group.
The results of the short-run dynamics are presented in Table 7 for the effect of trade on GDP p.c. and in Table 8 for the opposite direction. Here the short-run effects are calculated within error correction models for the case of a cointegration relationship, and within Granger causality models for the case of no cointegration. Both model classes also control for the inclusion of a lagged dependent variable. The negative significant error correction terms confirm the previously identified long-run relationship. As for short-run effects, only the countries with low levels of income and a lower-middle degree of trade openness show a positive significant short-run effect of trade on income. The group of low-income countries also shows a significant short-run effect for the direction from income to trade where also the coefficient of the upper-middle trade openness panel is significant.
In summary, the long-and short-run results suggest that the relationship between trade and income is not that straightforward. We can identify heterogeneity in the relationship depending on both, the stage of development and the degree of openness.
While trade does not seem to drive economic development for the low-and middle-income countries, where only shot-run benefits can be identified, highincome countries seem to gain income advantages in the long run. The beneficial effects of opening up to international trade thus seem to increase as economies develop, confirming the arguments about the adaptation capabilities of a country in determining technology transfer and knowledge diffusion (Kim and Lin, 2009). This result is consistent with existing empirical evidence, which shows considerable differences in the effects of trade on income across countries (Herzer, 2013)  Note: Standard errors are reported in parentheses. The Akaike information criterion (AIC) is used to determine the optimal lag length. * p < 0.10, ** p < 0.05, *** p < 0.01 and higher trade benefits for developed countries (Kim and Lin, 2009;Kim, 2011). Further, the link between trade and income also appears to be dependent on the degree of trade openness. Countries that are only open to international trade on a medium and small scale are unable to tap into the trade advantages in the form of economies of scale, specialization or technology adoption. The effect heterogeneity may reflect the development-and openness-related trade structure characteristics of countries at different stages of development and global integration (Hausmann et al., 2007;Rodrik, 2006). Contrary to the result of Zahonogo (2016), who identifies higher benefits below a certain threshold of openness, our result suggests that a certain degree of openness is required to reap the benefits. The differences in the results may appear due to the fact that his analysis is based on Sub Saharan African countries only, representing different conditions in terms of development and trade.

Conclusion
In this contribution we reconsider the link between openness and development and examine the causal relationship between trade and income for 167 countries over the period 1970-2011. After reviewing recent research regarding the link between trade and income, we apply panel cointegration, Granger causality and panel error correction Notes: The Akaike information criterion (AIC) is used to determine the optimal lag length. The short-run effect presents the summation of the short-run coefficients and the corresponding F-statistics in squared parentheses. The long-run effect presents the coefficient for the error correction term with the corresponding standard errors in round parentheses. AR2 presents the p value of the Arelano-Bond (1991) test for autocorrelation of second order. Hansen test presents the p value of the J-test for overidentifying restrictions * p < 0.10, ** p < 0.05, *** p < 0.01 in combination with Dynamic Ordinary Least Squares and General Method of Moments estimation to explore the causal relationship between these two variables.
Besides a global analysis, we account for potential development-and opennessrelated effect differences by subdividing the global panel into subpanels and examining causality differences with regard to different stages of development and different degrees of openness.
In line with recent literature, the global analysis reveals a statistically significant positive short-run and long-run relationship between trade and income. The subpanel analysis, however, shows that the identified relationship is not universally valid, it is rather heterogeneous and varies according to development and openness levels. A bidirectional positive long-run relationship is evident only for high-income and highly open countries, while countries with a low level of income or a low degree of openness are not able to benefit.
The heterogeneity of the results indicates that the opening process plays an important role. Countries at a low stage of development and a low degree of global integration where the opening process has so far been insufficient do not participate in the benefits of trade. From a certain development and openness stage -which is accompanied by a link to leading technologies and the transfer of knowledge -trade and economic development show a positive causal relationship. Notes: The Akaike information criterion (AIC) is used to determine the optimal lag length. The short-run effect presents the summation of the short-run coefficients and the corresponding F-statistics in squared parentheses. The long-run effect presents the coefficient for the error correction term with the corresponding standard errors in round parentheses. AR2 presents the p value of the Arelano-Bond (1991) test for autocorrelation of second order. Hansen test presents the p value of the J-test for overidentifying restrictions * p < 0.10, ** p < 0.05, *** p < 0.01  (2007) Z t 4.03 −1.56* −9.83*** −15.03*** Note: The Akaike informatiomn criterion (AIC) is used to determine the optimal lag length. * p < 0.10, ** p < 0.05, *** p < 0.01 Note: Under the null hypothesis of cross-section independence CD~N(0,1). ρ refers to the correlation coefficients between the time-series for each panel member Note: The Akaike information criterion (AIC) is used to determine the optimal lag length. * p < 0.10, ** p < 0.05, *** p < 0.01 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.