Appendix: Determining Whether Countries with Stable Exchange Rates and a Common Currency Trade More

Peter Clark, Shang-Jin Wei, Natalia Tamirisa, Azim Sadikov, and Li Zeng
Published Date:
September 2004
  • ShareShare
Show Summary Details

While there is some evidence of a negative effect of exchange rate volatility on trade, it is not robust to variations in specification. This is true both for aggregate trade as well as for trade in homogeneous and differentiated products separately. Therefore, the overall message from the empirical analysis is that, if exchange rate volatility depresses trade, the effect is unlikely to be quantitatively large. The basis for these conclusions is discussed in more detail, below.

The Gravity Model: An Empirical Analysis

The empirical analysis in this study is based on the standard gravity framework, whereby trade between two countries is modeled as a function of incomes (economic mass) of these countries and the distance between them. The framework has proved to be robust and successful in a wide variety of empirical applications. Moreover, the gravity model has strong foundations in international trade theories, from those based on country differences in factor endowments or technology to models of increasing returns to scale and monopolistic competition. Entering incomes in the product form is well established theoretically in the trade literature.22

Besides the economic mass and distance, the empirical specifications of the gravity model typically control for other factors augmenting or reducing trade, such as land areas, cultural similarity, geographical position, historical links, and preferential trading arrangements, all of which tend to affect the transaction costs relevant for bilateral trade and have been found to be statistically significant determinants of trade in various empirical applications. The model also controls typically for the level of economic development, which is expected to have a positive effect on trade because more-developed countries tend to specialize and trade more. With all of these by now fairly standard explanatory variables included in the gravity equation as controls, the focus of interest here is on the introduction of alternative measures of exchange rate variability to see to what extent this particular variable may affect transaction costs and thereby affect the level of bilateral trade between two trading partners.

To control for remoteness or multilateral resistance effects, country-specific fixed effects are included in the model. The concept of multilateral resistance was proposed by Anderson and Van Wincoop (2003) and is defined as a function of unobservable equilibrium price indices that depend on all bilateral trade barriers and income shares of the trading partners. The gravity equation can then be interpreted as indicating that bilateral trade depends on the bilateral trade barriers between the two countries in question, relative to the product of their multilateral resistance indices. For a given bilateral trade barrier between the two countries, higher barriers between them and their other trading partners would reduce the relative price of goods traded between them, thus raising bilateral trade. In empirical applications, the multilateral resistance indices can be conveniently proxied by country-specific fixed effects. The model also includes time effects to control for time-specific factors, such as global business cycles, oil price shocks, etc., so that the intercept in the model is allowed to change both across countries and over time. In addition, as an experiment, time-varying, country-fixed effects have been included, which is more general than including time dummies and country-fixed effects separately. The time-varying, country-fixed effects are arguably more consistent with the notion of a time-varying multilateral resistance emphasized in recent trade theories. At the same time, part of the forces underlying bilateral exchange rate volatility is time varying and country specific. The inclusion of the time-varying, country-fixed effects could also overcorrect. For example, an unexpected increase in one country’s money supply could raise all the bilateral exchange rate volatility involving that country. Even if this increase in volatility depresses all bilateral trade involving that country, a specification that controls for that country’s time-varying fixed effects would not be able to capture a negative effect of exchange rate volatility on trade. This qualification must be kept in mind in interpreting the result.

It is useful to note that up until very recently the literature that fits a gravity model to trade data seldom included any type of country-fixed effects. It is still rarer to include time-varying, country-fixed effects. Augmenting the empirical trade equation with various kinds of fixed effects may be considered one of the value-added aspects of this paper from a methodological point of view.

Aggregate Trade

The benchmark panel specification for the analysis of aggregate trade is similar to that used by Rose (2002). The model using ordinary least squares with robust standard errors is estimated based on the log-linear transformation:

where ltradeijt denotes the logarithm of the real value of aggregate bilateral trade between country i and j at time t; lrgdpijt is the logarithm of the product of real GDPs of countries i and j at time t; lrgdppcijt is the logarithm of the product of real GDP per capita of countries i and j at time t; lareapij is the logarithm of the product of the land areas of countries i and j; ldistij is the logarithm of distance between i and j; lrealijt is the long-run real IFS-based measure of volatility in the bilateral exchange rate of countries i and j at time t; and custrictijt is a dummy variable taking the value of 1, if countries i and j share a common currency at time t, and zero otherwise. The coefficients of interest are those on the measure of exchange rate volatility, lrealijt, and the currency union dummy, custrictijt.

Other variables control for various cultural, geographical, and historical factors: comlangij is a dummy taking the value of 1 if i and j have a common language; islandij is the number of islands and landlij is the number of landlocked countries in the country pair; borderij is a dummy taking a value of 1 when i and j share a common border; comcolij is a dummy taking a value of 1 if after 1945 i and j were colonies with the same colonizer; curcolijt is a dummy taking the value of 1 if i was a colony of j at time t, or vice versa; colonyijt is a dummy taking a value of 1 if i ever colonized j, or vice versa; and comctryijt is a dummy taking a value of 1 if i and j belong to the same nation.

There are also several controls for trade policy factors:23ftaijt is a dummy variable if i and j are members in the same regional trading arrangement; gspijt is a dummy taking the value of 1 if i was a Generalized System of Preferences beneficiary of j or vice versa at time t; and oneinijt and bothinijt are dummies taking a value of 1 if either i or j, or both, were members of GATT/WTO at time t. Finally, the vectors fe and te denote country- and year-specific dummies. The error term εijt is assumed to be well behaved.

Disaggregated Trade

For the analysis of disaggregated trade, two equations are considered separately for trade in differentiated products and in homogenous products, which are estimated by the Seemingly Unrelated Regressions (SUR) technique. This specification allows the parameters on the same variables to be different for different types of trade, while the error terms for a given country pair are correlated in the two equations.

A few comments are needed to clarify the less familiar analysis of disaggregated trade. Higher exchange rate volatility can be viewed as an increase in a type of transaction cost in international trade. More concretely, it may add noise to the price signal, and hence make it more difficult and more costly for buyers and sellers in the international market to find the right match for trading goods. A given increase in search costs, however, could play a different role in the overall transaction costs for trade in homogeneous products versus differentiated products. For homogeneous products, such as wheat, an importer is not concerned with who the producer is because the products are easily comparable, and price is the primary decision factor. On the other hand, heterogeneous products, such as digital cameras or tennis shoes, tend to be branded because there are additional characteristics other than price that would affect an importer’s purchase decision. For even more differentiated products, such as machine tools, again price would not necessarily be the key factor affecting the purchase decision.

Noting the difference in search costs in international trade for these two types of goods, Rauch (1999) presented some evidence suggesting that a given increase in transaction costs has a bigger negative effect on the volume of trade in differentiated products than in homogeneous products. He did not look into the effect of exchange rate volatility on trade, however. Extending his logic, one might hypothesize that a given increment in exchange rate volatility would also dampen trade in differentiated products more than trade in homogenous products. A recent paper by Broda and Romalis (2003) contains a theoretical model that assumes (as opposed to derives) this difference in the effects of exchange rate volatility. The authors also report some empirical evidence demonstrating that exchange rate volatility deters trade in differentiated products more than trade in homogenous products. Their regression specification, however, does not include as control variables most of the usual country-pair characteristics described above. Given that many developing countries are striving to move toward producing and exporting more differentiated products, it is interesting to test this hypothesis using a regression specification similar to that used for the analysis of aggregate trade.

Data and Sources

Aggregate Trade

Estimating the aggregate trade model requires data on total bilateral trade, incomes, population, distance, as well as geographical, cultural, and historical information. The study uses a panel data set that covers 178 IMF member countries every fifth year from 1975 to 2000.24 Summary statistics and correlations for the data set are presented in Appendix Table A16. The list of countries in the sample is presented in Appendix Table A1.

The data set is an updated version of Rose’s data set. To extend the data to 2000, the bilateral trade data series are constructed exactly following Rose’s study: bilateral merchandise trade data are from the IMF’s Direction of Trade Statistics. Bilateral trade is measured in U.S. dollars as total trade (exports plus imports) between the two countries in question, deflated by the U.S. Consumer Price Index (1982–83 prices) for urban areas (available from Real GDP and population data come exclusively from the World Bank’s World Development Indicators.25

The benchmark model focuses on the long-run measure of IFS-based real exchange rate volatility calculated as the standard deviation of the first-difference of the monthly natural logarithm of the bilateral real exchange rate in the five years preceding year t. The monthly bilateral exchange rates are obtained from the IFS. To obtain exchange rates for each European Monetary Union member currency for the years 1999 and 2000, the euro exchange rates were converted using the irrevocably fixed conversion rates obtained from the official European Central Bank website. Real exchange rates are constructed by using consumer prices from the IFS.

To check the robustness of results, alternative, yet analogously calculated, measures of exchange rate volatility were examined: long-run, IFS-based nominal exchange rate volatility; short-run, contemporaneous IFS-based real and nominal exchange rate volatility; and short-run and long-run volatility of real parallel market rates, data for which come from Reinhart and Rogoff (2002). For more details on these measures of exchange rate volatility, see Section III of this study.

As part of the robustness analysis, the conditional volatilities of the exchange rates estimated using a GARCH (1, 1) were considered.26 The underlying equation for the model is an ARIMA (0, 1, 0) process of the exchange rates (in the logarithmic form), which implies that the log difference of the exchange rates is a random walk with drift. This model yields an estimate of volatility that is the standard deviation of the error term in the underlying equation, conditional upon historical information from all previous months in the five-year period. The last estimated conditional standard deviation of each country pair is used as the approximation of the conditional volatility at the beginning of next period. For example, the conditional volatility of 1975 equals the estimated conditional standard deviation for December 1974 in the GARCH regressions.

The GARCH regressions on the monthly exchange rates are run for six five-year panels, with the first one being 1970–74 and the last one being 1995–99. In each five-year panel, exchange rate data are further grouped into three categories—those of developed country pairs, of developing country pairs, and of country pairs between developed and developing countries—which bring the total number of GARCH regressions to 15. To ensure that the estimated coefficients satisfy the stationarity conditions,27 we exclude countries with hyperinflation episodes and countries with extreme exchange rate fluctuations, defined as a change in the log exchange rate in absolute value in any month exceeding a threshold of 1, or |[d log(exrt)]t| > 1. The threshold amounts to a monthly appreciation of over 170 percent or a monthly depreciation of over 60 percent.28 In addition, in each panel the series length of the exchange rate for each country pair is required to be greater than or equal to 30. This process produces 124 countries in total for the estimation sample, which are used throughout the study.29

Disaggregated Trade

For disaggregated trade, data on the value of bilateral imports for 98 industries are obtained from the United Nations’ COMTRADE database and cover 39 countries (see Appendix Table A17) during the period 1975–2000. Import data are disaggregated at the Standard International Trade Classification (SITC) four-digit level, rev. 1, and are deflated by the U.S. urban Consumer Price Index (1982–84 prices).

The classification of products into differentiated and homogenous varieties follows the strategy in Rauch (1999). Conceptually, Rauch identifies first two types of homogenous products, those traded on an organized exchange (commodities) and those whose prices are reported regularly in a professional trade publication (referenced-price products). All other products are then defined as differentiated products. Rauch implemented the classification on SITC, rev. 2, industries.

There were instances when the classification for a given product was ambiguous. Hence, Rauch produced two separate classification systems: one, conservative aggregation, attributed all ambiguous products to the homogeneous category, and the other, liberal aggregation, attributed all ambiguous products to the differentiated category. Rauch provides an appendix that lists the classification results of all SITC, rev. 2, industries at the four-digit level. This study uses a concordance from SITC, rev. 1, to SITC, rev. 2 (available from, and then applies the Rauch classification to the data. To minimize the impact of misclassified products on the conclusions, all products whose classifications are ambiguous are excluded, and only those products whose degrees of classification are relatively clear are used. That leaves 81 industries for which the classification is relatively unambiguous; 22 of these are classified as homogeneous, and the remaining 59 are classified as differentiated products. The classification lists are presented in Appendix Table A18.

The series for bilateral imports of homogeneous products are obtained by summing sectoral import data across all sectors classified as homogeneous for a given country pair in a given year. Bilateral imports of differentiated products are constructed similarly. GDP and GDP per capita data are from the World Bank Indicators database. All other variables are from the aggregate trade data set described above.

Appendix Table A16 presents some summary statistics on the two types of products in the sample over the years. As can be seen, the total value of trade in differentiated products has been more than twice that of trade in homogeneous products in the sample. There is a modest increase in the share of trade in differentiated products in total trade in the sample, to 83 percent in 2000 from 75 percent in 1975.

Key Findings

The gravity model performs well empirically, yielding precise and generally reasonable estimates (Appendix Table A9) that are broadly consistent with the results of other papers employing a gravity model using trade data. The coefficient on distance is negative and statistically significant, ranging around −1.50 across the different variations of the model. The coefficient on the economic mass is positive and in general statistically significant, ranging from 0.83 in the model with time-varying fixed effects estimated on a sample excluding high-inflation countries, to 0.06 in the same model including such countries. The high sensitivity of the coefficient for the economic mass to the inclusion of high-inflation countries suggests that high-inflation episodes tend to distort economic relations between trade and other behavioral and policy variables, indirectly justifying the exclusion of such countries from the sample on which baseline regressions are estimated in this study.

Other control variables are also significant for the most part and have the expected signs. For example, a common language, Free Trade Agreement membership, Generalized System of Preference relationship, being a colonizer and a colony, and colonization by the same country all have a positive and statistically significant effect on trade. The role of some controls, however, is sensitive to the specification of the model. For example, the level of economic development, as measured by the real GDP per capita, has a positive and statistically significant effect on trade only in the model that includes high-inflation countries and time-varying country effects (Appendix Table A9, column 5). World Trade Organization membership is positive and statistically significant in most specifications, on balance suggesting that it has a trade-enhancing effect over and above other factors.

Main Results of the Effect of Exchange Rate Volatility on Trade

Some benchmark results of using the gravity equation to estimate the effect of exchange rate variability on aggregate trade are given in Appendix Table A9. These equations use the standard measure of volatility, i.e., that of the long-run real IFS exchange rate, the coefficient of which is shown in the first row of the table.30 As shown in column 1, which uses both time-and country-fixed effects, there is a statistically significant negative impact on the level of trade.31 This impact can be computed as the effect of increasing volatility by one standard deviation around its mean, which implies a reduction in trade flows of almost 7 percent.32 Employing this same specification but using the full sample of countries, as shown in column 4, the estimated reduction in trade generated by a one standard deviation increase in volatility generates a reduction in trade of somewhat over 9 percent.33 These estimates are comparable to those found by other authors using the same methodology, e.g., Rose (2000), who estimates a reduction of 13 percent, and Tenreyro (2003), who makes estimates ranging from 4–8 percent.

An alternative specification is used in the results reported in column 2 of Appendix Table A9, where individual country-fixed effects are replaced with country pair–fixed effects. The main advantage of this approach is that it allows one to control for unobserved cultural, economic, historical, geographical, and other factors that are specific to a given pair of countries.34 Omitting such factors may bias the estimation results if they are correlated with other regressors in the model. An F-test indicates that the estimated coefficients for the country-pair fixed effects are jointly significant. It turns out that there is very little effect on the coefficient of exchange rate volatility, which is essentially the same as in column 1 of the same table.35

The finding of a negative impact of exchange rate volatility, however, is not evident in a more general specification in which country- and time-fixed effects are replaced with time-varying fixed effects. Allowing for time variation in country-fixed effects is more consistent with the theoretical concept of multilateral resistance proposed by Anderson and van Wincoop (2003) because such multilateral resistance indices are likely to vary over time. Moreover, an F-test comparing the two specifications indicates that the latter is preferred on statistical grounds. As shown in Appendix Table A9, column 3, this modification of the model results in a positive estimated impact of exchange rate volatility on trade (but not significant at the 90 percent confidence level). Using the time-varying country effects approach for the full sample, as reported in column 5, the estimated effect of volatility is negative and the same size as in column 4, but not statistically different from zero.

What might account for the difference in the results? One possible explanation runs as follows: time-varying country-fixed effects in principle control for all unidentified time-varying factors that are country specific, including the effective, i.e., overall exchange rate volatility for each of the trading partners in question. Indeed, when one includes the measure of effective volatility (at the country level, as opposed to bilateral level) in the basic model with time-invariant country effects,36 the coefficient on this measure of effective volatility is negative and statistically significant, while the coefficient on the bilateral measure of exchange rate volatility becomes positive and similar in magnitude to that in the model with time-varying fixed effects. This shows that the negative effect of bilateral volatility on trade is not robust to controlling for broader aspects of exchange rate volatility and, more generally, for all aspects of multilateral resistance.

These benchmark results show that there is evidence of a negative effect of exchange rate volatility on the level of trade, but the magnitude of the estimated impact appears to be small. This finding is not robust to the choice of estimation technique, however. In particular, the negative effect disappears in a general model that controls for time-varying factors that are country specific and is in line with the most recent theoretical work on the gravity model of trade. This disparity in the findings characterizes not only the above results using the benchmark measure of exchange rate volatility for aggregate trade, but also the results reported below that use alternative measures of volatility, look at different country groupings and different types of traded goods, and use alternative estimation techniques that attempt to control for the possibility that exchange rate volatility is not exogenous. Thus, to anticipate the overall conclusion, while there is evidence that increased exchange rate variability reduces the volume of international trade, this finding depends on the particular estimation technique employed, so that it cannot be considered an overwhelmingly robust empirical result.

Alternative Measures of Volatility

The estimated impact of alternative measures of volatility on trade, using time- and individual country-fixed effects, is reported in Appendix Table A11. When one includes in column 1 a measure of the short-run real exchange rate volatility in the model as the difference from the long-run real volatility,37 short-run volatility can be seen as having an additional dampening effect on trade over and above the negative effect arising from the long-run volatility. The magnitude of this additional effect is about half of the long-run volatility effect. This finding could be interpreted as indicating that trading firms form their expectations of the future exchange rate volatility based on both historical and contemporaneous volatility.

The volatility in the parallel market exchange rate has a broadly similar effect on trade as the volatility in the official IFS-based exchange rate. Appendix Table A11, columns 2 and 3, reports regressions where the volatility measures based on the parallel market rates are included as differences from the IFS-based volatility measures, in addition to the official volatility measures. The coefficient on the long-run parallel market volatility is negative and statistically significant, about the same as the coefficient on the official exchange rate volatility in column 2, but about one-third the size in column 3. In the short run, however, the volatility of the parallel market exchange rate does not appear to affect trade.38 These results suggest that parallel market rates are also relevant for trade transactions in addition to the official exchange rates.39

Given that nominal and real exchange rates are highly correlated, it is probably not too surprising that their volatilities have similar effects on trade. In Appendix Table A11, column 4, the coefficient on the nominal exchange rate volatility (–2.60) is close to that on the real exchange rate volatility (–2.37) in Appendix Table A9, column 1.

So far, this study has considered a simple statistical measure of exchange rate volatility. This is now replaced with a conditional measure of exchange rate volatility, as estimated from a GARCH model (Appendix Table A11, column 5). The coefficient of the conditional volatility measure is virtually identical to that of the unconditional measure (−2.20 versus −2.37). Irrespective of whether one assumes that trading firms condition their expectations on the available historical information (GARCH) or that they project volatility using a simple statistical approach, exchange rate volatility has a statistically significant negative effect on trade of a broadly similar magnitude. In terms of the impact of exchange rate volatility on the level of trade, the estimates are comparable to those discussed above in connection with the benchmark results. They range from a low of about 5 percent for the GARCH estimate of volatility (column 5), to 25 percent for the combined effect of short- and long-run real official volatilities.

These results using various measures of exchange rate volatility are broadly robust to an alternative model specification where country fixed effects are replaced with country-pair effects (Appendix Table A12). The estimated coefficient of volatility is consistently negative, in nearly all cases statistically significant, and tends to be somewhat higher.40 As a consequence, the impact of higher exchange rate volatility on trade is also larger, ranging from a reduction of 8 percent for the long-run nominal exchange rate to a decline of 26 percent for the combined impact of short- and long-run real official rates.

When country-fixed effects and time effects are replaced with time-varying country effects, however (Appendix Table A13), this modification of the model reverses the impact of the study’s standard measures of bilateral exchange rate volatility on trade in the long and short run—it becomes positive and statistically significant—and the effect of parallel market volatility becomes insignificant. This lack of robustness to alternative specifications is in line with that discussed above in connection with the results for the benchmark measure of volatility. With this particular model, increased exchange rate variability now has an estimated positive impact on trade, ranging from 10 percent (column 4) to 34 percent (column 3).

It should be noted, however, that the equation results reported in column 4 of Appendix Table A13 include the effective or overall exchange rate volatility of a country that is used in Section III. Because bilateral trade flows are the dependent variable, the sum of the effective volatilities of the country pairs is used as the regressor. The idea in this specification is to examine the effect of bilateral exchange rate volatility relative to the aggregate measure of volatility, which is a component of the multilateral resistance to trade mentioned above. One would expect that an increase in the variability of the bilateral exchange rate between two countries would have a negative effect on their bilateral trade, and that an increase in the variability of all other exchange rates would tend to raise trade between the two countries in question; as such trade would become relatively less risky. In fact, the empirical results are counter to this expectation. Nonetheless, the net impact of a one standard deviation increase in volatility is a reduction in trade of about 13 percent because the negative effect of the higher effective volatility more than offsets the positive effect of the rise in bilateral volatility.41

Controlling for Endogeneity of Exchange Rate Volatility

So far, exchange rate volatility has been assumed to be exogenous to trade. This assumption, however, may not be warranted: to the extent that countries implement policies aimed to lower exchange rate volatility in order to increase their trade, the baseline equation would suffer from endogeneity bias. Two instrumental variable approaches were used to control for this possibility:42 that proposed by Frankel and Wei (1993), whereby the volatility in the relative quantity of money is an instrumental variable for exchange rate volatility, and that proposed by Tenreyro (2003), which relates exchange rate volatility to the incidence and the propensity of countries’ to share a common anchor. Neither of these instruments is perfect, but each has its advantages: the Frankel-Wei approach is simple and easy to implement, while that of Tenreyro’s instrumental variable appeals to the modern optimal currency framework of Alesina, Barro, and Tenreyro (2002).

Controlling for endogeneity using the Frankel-Wei instrumental variable approach (see first panel of Appendix Table A10) modifies the basic results of the role of exchange rate volatility. While the coefficient on exchange rate volatility remains negative in every specification, it is statistically significant only in the equation with country and time effects for both real and nominal exchange rates. In these two cases, the estimated coefficients are much larger than those reported above. The negative trade effects are also considerably larger—about 90 percent for the real rate and 125 percent for the nominal exchange rate—which seems implausible compared to the findings described above.

The results using the Tenreyro instrumental variable approach are reported in the second and third panels of Appendix Table A10. When the instrumental variable is the dummy for a common anchor, the coefficients on the exchange rate volatility measure become statistically insignificant across all specifications. When the propensity to share a common anchor is used as an instrumental variable, these coefficients are negative in all specifications and statistically significant in the specifications with both country- and time-fixed effects, and with time-varying fixed effects. In both cases, the estimated coefficients are extremely large, implying reductions in trade ranging from 115 percent to 265 percent (column 1) for a one standard deviation increase in volatility. These are far beyond any other estimates in this paper or in the literature, and, as such, should probably be viewed as outliers.

Several reasons may account for differences in the magnitude and sign of these coefficients on the volatility measure and those obtained by Tenreyro. The most important one is that in her regressions volatility appears as log (1 + standard deviation of the exchange rate), whereas the regressions of this study include just the volatility or the standard deviation of the exchange rate for consistency with the benchmark OLS estimations and instrumental variable regressions of Frankel and Wei (1993). The specification of this study differs from Tenreyro in other respects as well: common language and border dummies considered significant are included in the logit regressions estimating the propensity to adopt a common anchor; the study controls for whether or not the trading partners are related as a colony and a colonizer, as well as WTO membership, common colonizer status, and whether or not the trading partners are island economies; and trade between the two countries, rather than just bilateral exports, are used as the left hand–side variable.

Differentiation Across Country Groups

Findings indicate that across countries the impact of exchange rate volatility on trade is not uniform. In particular, volatile exchange rates appear to be more damaging for trade among developing countries than for trade among advanced economies. As shown in Appendix Table A15, column 1, the coefficients on the exchange rate volatility measures interacted with dummies for trade among advanced and developing countries (denoted by NS) and among developing countries (denoted by SS) and are negative and statistically significant; their net magnitude is −2.23 and −3.22, respectively. This is consistent with the possibility that developing countries are less able to manage currency risk. Foreign exchange markets are typically nascent and less liquid in developing countries, limiting firms’ opportunities for hedging foreign exchange risk. With time-varying fixed effects, shown in column 2, however, there is essentially no impact of volatility on trade flows among NS and SS.

Do Members of Currency Unions Trade More?

The main results of this study confirm the remarkable finding of Rose (2000) that common currency arrangements triple trade (Appendix Table A9, column 1), as the coefficient on the currency union dummy is comparable to that found in his paper. The trade-enhancing benefits of currency unions appear to exceed by far the gains from a substantial reduction in exchange rate volatility, although, as discussed above, the instrumental variable (IV) estimation results also indicate a very large benefit in trade gains arising from a decline in volatility.

While the trade-enhancing effect of a common currency is robust to controlling for time-varying fixed effects (Appendix Table A13, columns 1 and 5), it breaks down in a model with country-pair fixed effects (Appendix Table A9, column 2, and Appendix Table A13, columns 1, 4, and 5), in line with the findings in Pakko and Wall (2001). The statistical insignificance of the currency union dummy in the model with country-pair effects—which Rose (2002) did not use in his analysis—suggests that the trade-enhancing effect of a common currency found in specifications omitting country-pair effects reflects an estimation bias because the omitted factors apparently are correlated with trade volume and with the likelihood that countries use a common currency (for example, common history or institutional and regulatory similarities among countries that are members in a currency union). Of course, currency unions have country-pair characteristics that evolve slowly over time. It would thus appear that the power of a test for the effect of currency unions on trade becomes much weaker when country-pair fixed effects are included.

Moreover, the beneficial effect of a common currency on trade is not uniform across country groupings because it appears to be limited to currency unions among developing countries. When the currency union dummy is interacted with the dummy for the developing country pairs, the coefficient on this product term is positive and significant, while the general currency union dummy becomes negative and statistically significant, suggesting that currency unions other than those among developing countries impair trade between them (Appendix Table A15). This result suggests that currency unions may have a large positive impact on trade only in cases where transaction costs are high, where policy credibility problems are acute, or where hedging opportunities are limited.

Disaggregated Trade: Does Exchange Rate Volatility Have a Different Effect on Trade in Differentiated Versus Homogeneous Products?

The finding that the negative effect of exchange rate volatility on trade is not robust also carries over when looking at disaggregated trade. As discussed in the text, recent developments in the economics of trade suggest that a given increase in transaction costs (of which exchange rate volatility is a component) could have a larger negative effect on trade in differentiated products than on trade in homogenous products. The evidence for this possibility is examined in this subsection.

In the first specification (reported in the first two columns of Appendix Table A14), the study estimates a system of two equations by the SUR technique with time- and country-fixed effects. In this case, the coefficients on exchange rate volatility are negative in both equations, and the volatility effect is statistically significant only in the equation on trade in differentiated products. In other words, consistent with the conjecture above, exchange rate volatility has a negative effect on trade in differentiated products, but not on trade in homogenous products. This conclusion, however, is not robust. In the last two columns of Appendix Table A14, when time-varying, country-fixed effects are included (which are more general than the inclusion of a combination of time- and country-fixed effects and are dictated by the recent theory underlying the gravity specification used here), the conclusion is overturned. More precisely, the coefficients on exchange rate volatility are not statistically different from zero for trade both in differentiated and in homogenous products. As extensions, the study also examines the effects of short-run exchange rate volatility and volatility of the parallel market exchange rate: the results are qualitatively the same as described above.

Thus, the overall conclusion on disaggregated trade is the same as that for aggregated trade: namely, the evidence is not overwhelmingly robust to indicate that exchange rate volatility has a negative effect on trade. Using an array of alternative formulations involving different measures of exchange rate volatility, estimation techniques, different country groupings, and dis-aggregation by type of product, one does find fairly systematic evidence of a negative effect of volatility on trade. Once one takes into account other factors that would affect trade in a more general model involving the time-varying multilateral resistance as emphasized by recent trade theory, however, this negative effect disappears. Thus, whether or not one finds evidence that exchange rate volatility depresses the volume of trade and has a larger negative effect on differentiated than on homogeneous products, depends on the particular methodology used in the estimation.


Bilateral tariff and nontariff barriers are excluded from the model due to unavailability of data.


A key regressor in the analysis is long-run exchange rate volatility, which is constructed from five-year intervals. When trade is sampled every fifth year, exchange rate volatility can then be constructed from non-overlapping five-year periods.


In contrast, Rose uses several sources: World Development Indicators, Penn World Tables, and International Financial Statistics. WTO and Free Trade Agreement dummies for 2000 are extended based on the information available from the WTO official website (


As an alternative to estimating volatility using parametric models such as GARCH, Andersen and others (2001) propose examining the realized volatility directly, which has the advantage of being model independent. This approach, however, is very data intensive and thus cannot be implemented in this study.


The estimated coefficients of the regressions ensure that all time-varying variance (c2 t) processes are stable. Furthermore, the results of nine regressions satisfy the sufficient conditions that will guarantee the GARCH processes to be covariance stationary (see Greene, 2000, p. 802).


The following countries are excluded from all regressions: Angola, Argentina, Armenia, Azerbaijan, Bolivia, Brazil, Bulgaria, Chile, Democratic Republic of the Congo, Republic of Congo, Dominican Republic, Ghana, Honduras, Iran, Israel, Lithuania, Mexico, Nicaragua, Nigeria, Peru, Romania, Sudan, Suriname, Tajikistan, Turkmenistan, Uganda, Ukraine, Yugoslavia, Zambia. For regressions using 1995–99 data, five more countries are excluded: Belarus, Indonesia, Sierra Leone, Sri Lanka, and República Bolivariana de Venezuela.


Owing to missing data, the number of countries in the estimation samples for different years is less than 124.


Appendix Table A11 reports the effects of alternative measures of exchange rate variability.


This equation was also estimated without country-fixed effects, and an F-test confirmed that the inclusion of such effects is warranted.


This impact is computed as the estimated coefficient in the regression equation multiplied by one standard deviation of the volatility measure, multiplied by 100 to convert to percent.


While the coefficient of volatility in this regression is about one-half that in column 1, the standard deviation of volatility is over twice as large, as shown in Appendix Table A16, with the overall result being a somewhat bigger trade impact. Given that the larger sample includes countries with substantial exchange rate changes, this is not at all surprising.


In this specification, distance, land area, and other time-invariant bilateral variables become redundant in regressions with country pair–fixed effects and therefore are excluded from the regression.


As discussed below, however, this is not true for all of the other estimated coefficients and, in particular, for the dummy variable for a common currency union.


See Appendix Table A13, column 4.


Short-run volatility is not strongly correlated with long-run volatility in the sample: the correlation coefficient is 0.38.


When the alternative measures of exchange rate volatility are included as separate regressors in the gravity equation, the estimated coefficients are similar to those reported above and are statistically significant, except for the volatility of the short-run parallel rate.


It is also worth noting that the IFS-based exchange rates used in the benchmark regressions are not only official rates but also include market and principal rates. The exchange rates in IFS are classified into three broad categories, reflecting the role of the authorities in the determination of exchange rates and/or the multiplicity of exchange rates in a country. The market rate is used to describe exchange rates determined largely by market forces: the official rate is used to describe an exchange rate determined by the authorities, sometimes in a flexible manner. For countries maintaining multiple exchange arrangements, the rates are labeled principal rate, secondary rate, and tertiary rate. The official rate is included in the series only if neither the market nor the principal rate is available. The IFS-based measures are thus reasonably well correlated with parallel market rates, with the coefficient of correlation of 0.65.


Again, when the alternative measures of volatility are entered in the equation by themselves, the coefficients are similar to those reported in Appendix Table A12 and are statistically significant.


In estimating the equation in column 4 of Appendix Table A13, country-fixed effects are included in the model instead of time-varying country effects, on the assumption that the impact of effective volatility would be largely absorbed by the time-varying country dummies. When this same equation was estimated with time-varying country effects instead of the time- and country-fixed effects used in column 4, however, the results were very similar, with a significant negative impact on trade somewhat higher at 19 percent.


An alternative third instrumental variable approach, based on Devereux and Lane (2003), combines the factors underlying the optimal currency area theory with the factors underlying financial links to explain the volatility of the bilateral exchange rate for trading partners. While appealing conceptually, this approach is highly data intensive and was not implemented due to the unavailability of data.

    Other Resources Citing This Publication