Preliminary – Do not quote

Comments welcome

Shuo.jan.dave

The Link between Trade and Income: Export Effect, Import Effect, or Both?

Shuo Zhang

Jan Ondrich

J. David Richardson[1]

Department of Economics

Syracuse University

March 2003

Abstract

A new framework is developed to evaluate how cross-country differences in export openness and import openness in 1990 affected the level of real per capita income. Familiar and novel instruments are used to extract the exogenous components of total trade (exports plus imports) and of net exports (exports minus imports), which in turn imply distinct export and import effects. We build on an existing literature (Frankel-Romer and others) that uses aspects of a country’s geography as instrumental variables for total trade openness. We build on a country’s demography and net wealth abroad to develop a novel instrument for net export openness. Our new estimates reveal that export openness alone correlates with income cross-sectionally, not import openness.

I. Introduction

The doctrine of mercantilism views trade, at least in part, as a zero-sum game. Trade is considered favorable if exports bring in more money than what is paid out for imports. Adam Smith and David Ricardo challenged the mercantilism theory by arguing that overall openness is what matters – nations grow more prosperous through the process of specialization and trade through imports as well as exports. Many developing countries have replaced their strategy of government-protected import substitution industrialization with a market-focused export oriented strategy. They seek to promote economic growth by exporting more and more of their products; this has the feel of neo-mercantilism.

This paper re-poses the question whether trade “raises” a country’s income per person (or per worker); and if it does, what are the channels through which trade affects income, exports or imports? Numerous theoretical and empirical studies have attempted to answer these questions. Theoretical support for the positive link between different sorts of openness and economic growth were provided by Romer (1986) and Lucas (1988). Barro and Sala-i-Martin (1995) and Romer (1992) emphasized how countries that are more open have a greater ability to absorb (import) technological advances, which ultimately leads to higher real per capita income. Krugman (1974), Rodrik (1995), and Rodriguez and Rodrik (2000) on the other hand, cast doubt on the impact of openness on growth. Warner (2003) casts doubt on their doubts and views our central question as a frontier issue.[2]

Empirical studies construct measures of openness variable based on exports[3], imports[4], or the sum of the exports and imports[5]. Levine and Renelt (1992) argue that “all findings using the share of exports in GDP could be obtained almost identically using the total trade or import share. Thus, studies that use export indicators should not be interpreted as studying the relationship between growth and exports per se but rather as studying the relationship between growth and trade defined more broadly.” (p.959) Grossman and Helpman (1991) stated that technological spillovers could come via imports as easily as exports. Lawrence and Weinstein (2001) argue that imports, not exports, contribute importantly to the productivity growth of Japan and Korea.

Frankel and Romer (1999), on which chapter one of this dissertation is based (Zhang 2002a), alleviates many of the conceptual and econometric barriers to these issues by showing how geographical characteristics provide an arguably good instrument for a country’s intrinsic openness. Yet they remark that their trade and income investigation cannot separate the import effect and export effect. Wei (with Wu 2001, 2002a, 2002b) consciously follows Frankel’s and Romer’s lead in a study of how globalization affects Chinese city-level growth and inequality. Wei conflates export and import influences by selecting Chinese cities’ distanced to two major Chinese ports as his instrument for a city’s natural openness.

No one to our knowledge has yet figured a way to do what seems initially the most natural thing. That is to construct a measure of export openness, then an arguably independent measure of import openness, and investigate whether one has a different effect on income than the other, ceteris paribus.

That is the principal objective of this paper -- to identify the separate influences of export openness and import openness on income levels after controlling for endogeneity. In particular, we develop a framework that is slightly different from the Frankel and Romer (1999) income determination model by considering an additional “net trade effect” on income levels. When combined with Frankel and Romer’s “total-trade” effect, the two effects together imply separable export and import effects. The extended income model is intended to contribute to three empirical challenges related to trade and income. The first is the Frankel-Romer determination of instruments for total trade. The second is the determination of instruments for net trade. Thirdly, the model and data employed permit the identification of the impact of exports and imports on income separately.

Our concern about endogeneity in this chapter differs slightly from our concern in the previous chapter. When considering exports and imports separately, the expected bias from ignoring endogeneity is, respectively, caused by the unobserved income determinants that are correlated with exports and imports.

This paper (Zhang 2002b) is one part of my three dissertation papers (Zhang 2002a, Zhang 2002b, Zhang 2003). Zhang (2002a) explores the sensitivity of the Frankel and Romer (1999) empirical relationship between country’s total-trade openness and income level to heteroscedasticity and sample selection in their first-stage bilateral instrumenting regressions. The results support their hypothesis that trade has a significant and positive, yet relatively small impact on income. The present chapter tries to differentiate Frankel’s and Romer’s export openness from import openness. Export openness plays the dominant role. Unfortunately, the unique instrumenting techniques for “net trade” in this chapter cannot be implemented bilaterally, so there remains some doubt about the exact onformity of our conclusions to Frankel and Romer’s. Zhang (2003) employs established and new panel econometrics techniques to further examine the robustness of the findings presented in Frankel and Romer (1999), Zhang (2002a), Zhang (2002b) and other literature.

The rest of the paper is as follows. Section II describes the models. Section III provides data definitions and sources. Section IV reports the empirical results and Section VI contains the conclusions of the paper.

II. Empirical Models

Extensive research has been devoted to investigating factors that influence the per-capita income levels of countries. Recently, an enormous literature has developed on the influence of trade openness.[6] The major shortcoming of many of these empirical studies is their inability to separate the impact of exports and imports. Some focus on one to the neglect of the other; others focus on openness measures that force them to have equal weight. In particular, by regressing income on total trade, many studies embody an underlying assumption that exports and imports contribute equally to income growth. In this section, we develop a simple model based on one of them, the Frankel-Romer study, by separating the total trade share into export share and import share in order to distinguish between export effect and import effect.

(1)

where represents per capita income. and are exports and imports scaled by real GDP.[7] stands for other control variables.

As written, the “true” effects of export openness and import openness, respectively, are and . Yet in practice, researchers usually omit one of the openness terms or force the coefficients to be identical on the two measures of openness (calling the sum of the two “overall” or total “openness”). In principle, this could cause omitted-variable bias in the remaining coefficients of interest or specification bias, due to constraining two coefficients to have the same value. Yet the natural temptation to “just run equation (1)” usually creates misleading inferences because export openness and import openness are indeed highly correlated, as any general-equilibrium thinker knows and as Levine and Renelt observe (above), and because both are acknowledged to be endogenously related income per capita, the focus variable on which they are thought to “operate.”[8] Confronting these challenges, what’s a researcher to do?

Both exports and imports contribute to income growth. Exports promote specialization and exploit economies of scale. An increase in demand for country’s output raises real income. Imports allow countries to take advantage of other countries’ technology embodied in imported inputs. Through intensive involvement in international competition, countries become more productive. But income also has a causal effect on exports and imports. More supply of output leads potentially to more exports. Higher income facilitates consumption of both domestic and foreign products which ultimately raises imports. Therefore, we cannot directly regress income on exports and imports as shown in equation (1). Standard regression is inconsistent in the presence of mutual endogeneity.

Instrumental variable estimation provides a theoretically appealing way to handle the endogeneity problem. The important practical question posed in this paper is how to find two sets of instrumenting variables that can not only capture the exogenous components of exports and imports but also distinguish the export effect from the import effect. Country A’s exports to country B is country B’s imports from country A. There is a tendency both within countries and across them for exports and imports to co-vary for general-equilibrium reasons. Thus it is difficult in practice for instrumental-variable groups to ideally capture the distinction.

This paper proposes an alternative method to distinguishing the effects of imports and exports. The challenge of endogeneity in standard models is maintained by algebraically re-arranging equation (1). Letting T be the total trade (exports plus imports) divided by real GDP and E be the net trade (exports less imports) divided by real GDP, to control for scale effects, equation (1) transforms to

(2)

We argue that finding good instruments for T and E in order to estimate equation (2) is much easier than finding good instruments for X and M in order to estimate equation (1). In Figure 1, the average export share X from 1970 to 1998 is plotted against the average import share M from 1970 to 1998. The World Development Indicator 2000 data cover 174 countries.[9] The figure exhibits a strong positive relationship between export share and import share (with the correlation of 0.85).

In Figure 2, we plot the sum of the export share and the import share along the vertical axis. Along the horizontal axis, we plot the difference between the export share and the import share. The chart shows no evident relationship between the two variables in the long run (the correlation is equal to -0.13.)

Frankel and Romer (1999) have already provided good instrument candidates for the total trade share. The major task of the next section is to present the Frankel-Romer model with the heteroscedasticity correction, propose instruments for the net trade share variable by capturing cross-country differences in borrowing and lending behavior, and investigate the impact of total trade and net trade on income after instrumenting for both endogenous independent variables. Since and , the estimation will shed implicit light on the distinctive export effect and the distinctive import effect.

A. Bilateral Trade Regression

Trade promotes growth through increased specialization, efficient resource allocation, diffusion of international knowledge, and heightened domestic competition (Sachs and Warner 1995). On the other hand, countries that produce more output tend to trade more with the rest of the world. To correct for the simultaneity bias, Frankel and Romer (1999) proposed the geographical characteristics of countries as instruments for total trade. They argued that geography is a powerful determinant of bilateral trade as well as overall trade. Furthermore, countries’ geographical features are not affected by their incomes, or by government policies and other factors that influence income.

This section applies the extended version of the Frankel-Romer bilateral trade model as appeared in Frankel and Rose (2002). We also employ the updated bilateral trade data set which “is estimated to cover at least 98 percent of all trade.” (Frankel and Rose 2002, pp. 462) The first stage regression is based on the international trade extension of the gravity model, namely, that trade volume between two trading partners rises with an increase of national incomes or a reduction of the distance between them:

(3)

.

In equation (3), denotes the bilateral trade between countries i and j (measured as exports plus imports), a is the vector of coefficients, Xij is the vector of the covariates, distij measures the great circle distance between the principle cities of countries i and j, pop and area represent population and area respectively, languageij is a dummy variable which takes the value of 1 if people in country i and country j speak the same language, landlockedij is the number of landlocked countries within the country pair, and borderij is the dummy variable for a common border between two countries. Countries far away from each other tend to trade less. Since country size is inversely related to proximity, the impact of area is expected to be negative. Larger trading partner’s population implies higher demand for the domestic country’s export. Thus, the sign for popj should be positive. Countries are expected to trade more with each other if they have the same official language or share a border. If countries have access to the ocean, shipping costs are significantly reduced which makes it less difficult to trade.

Due to the possibility of heteroscedasticity in large cross-country analysis, this paper conducts the Breusch-Pagan test on equation (3). The test statistic rejects the homoscedasticity assumption with a high level of confidence. To counteract the problem of heteroscedasticity, weighted least square (WLS) estimation is used based on the procedure proposed by Harvey (1976).[10] Assuming the variance of the disturbance term in equation (3) has the following hypothesized specification:

(4)

The estimator of is:

(5)

based on the estimation

(6)

where is the residual resulting from the OLS regression of equation (3) and . The values of both the dependent and independent variables are divided by the square root of the predicted variance of the disturbance term, which corresponds to the pattern of the residuals. This normalizes the residuals so that they are homoscedastic.

The next step is to obtain the predicted value of the dependent variable in levels. Simply exponentiating underestimates the expected value of because