Has the European Monetary Union Increased Trade? Andrew K. Rose UC Berkeley-Haas, CEPR, NBER Heidelberg, Nov 2016 (based on research joint with Reuven Glick, FRBSF)

Currency Unions Bilateral Currency Unions (“Dollarization”) British £: Bahamas (-1965), NZ (-1966), India (-1966), Ireland (-1978) …. US $: Panama, Bahamas (1966-), Ecuador (2000-), El Salvador (2001-), Zimbabwe (2009-) …. Fr Franc: Morocco (-1957), Algeria (-1968) … Multilateral Currency Unions CFA Franc Zones Eastern Caribbean Currency Union Common (Rand) Monetary Area European Economic and Monetary Union, EMU (1999-)

Costs and Benefits of Joining a Monetary Union Key Costs Loss of nominal exchange rate as policy tool Loss of national monetary policy control EMU: these costs are high! Potential (Economic) Benefits Greater transparency of prices encourages greater competition and efficiency Reduced currency risk encourages more trade and investment Is there actually any substantive benefit in the data?

Debate in Literature on Magnitude of Trade Effect of CUs Big, 90-100%. e.g. Glick and Rose (2002), Frankel (2010) Moderate, 40-50% e.g. Eicher and Henn (2011) Small, 0-20% e.g. Micco et al (2003), Bun and Klaasen (2002, 2007), de Nardis and Vicarelli (2003), Flam and Nordstrom (2007), Berger and Nitsch (2008), Camarero et al (2013) Negative? e.g. Baldwin and Taglioni (2007)

Specific Motivation Glick-Rose (2002) used panel approach to investigate effect of currency unions on trade, using data for 1948-1997 before establishment of EMU Found currency unions increase trade by ≈90% Current paper uses data for 1948-2013 and asks Is EMU similar to other currency unions? Is there symmetry between currency union exit and entry? Assumed symmetry before. Couldn’t test because had only 16 entries, 130 exits in 1948-1997 sample Can test now with EMU entries Do advances in methodology matter?

Preview of Findings EMU different from other CUs, increases trade among EMU countries by ~50% Find symmetry Econometric methodology matters a lot Sample matters a lot as well

Measuring Trade Effects “Old” Methodology: Gravity Equation ln(Tradeijt) = CUijt + Zijt + {δt} + ijt Tradeijt = average nominal value of bilateral trade between i and j at time t, CU = 1 if i and j use the same currency at time t and 0 otherwise, Z = gravity control variables, usual suspects: e.g. GDP, distance, common language, border, regional RTA, colonial history, etc. … {δt} = year-specific effects

Methodological Issues in Estimating  Simultaneity Omitted variables Effects of CU between i and j on other countries through “multilateral resistance” effects General equilibrium effects on spending and output for all countries Homogeneity implicit in treating all currency unions alike

Data Set IMF DoTS trade: >200 “countries” 1948-2013 (with gaps) Population, real GDP: WDI > PWT > IFS Country Characteristics: World Factbook Regional Trade Agreements (RTAs): WTO Currency Unions: Glick-Rose updated 1:1 par for extended period of time (not just hard fixes) Transitive: x-y and y-z imply x-z

Gravity  Estimates for Trade  Table 1 EER 2002  New  With non-EMU and EMU CUs dis-aggregated All CUs 1.30 (.13) .92 (.09) All Non-EMU CUs   1.12 (.11) All EMU .02 (.08) Sample period 1948-1997 1948-2013 #Obs. 219,558 426,953 e1.3 ~3.7x e.92 ~2.5x nil Note: Pooled OLS estimates. Other gravity regressors and year dummies included, but not reported. Robust standard errors in parentheses.

Prefer (Within) Fixed Pair Effect Estimator Exploits variation over time, answers the policy question of interest, i.e. the (time series) question “What is the trade effect of a country joining (or leaving) a currency union?” Controls for unobserved pair effects, including potential endogeneity of currency union ln(Tradeijt) = CUijt + Zijt + {δt} + {θij} + ijt

Gravity  Estimates for Trade with Pair Fixed Effects  Table 2 EER 2002  New With non-EMU and EMU CUs dis-aggregated All CUs .65 (.05) .63 (.07) All Non-EMU CUs   .75 (.10) EMU .41 (.05) Sample period #Obs. 1948-1997 219,55 1948-2013 426,953 #Pair FE 11,178 14,801 e.41 -1 ~ 51% Note: Pooled OLS estimates. Other gravity regressors and year dummies included, but not reported. Robust standard errors in parentheses.

How Does this Compare with Literature? Easiest to graph (large) literature

Forest Plot of (45) Literature EMU Estimates

Meta-Estimate Random effects estimator delivers estimate of (exp(.116)-1≈) 12.3% Economically non-trivial Statistically significant Robust to reasonable sub-samples

Meta-Estimates of EMU Trade/Export Effect Estimator Sample Point Est. 95% Confidence Interval P-value, no Hetero. Lower Upper Fixed All (45) .085 .078 .091 .000 Random .116 .084 .147 Export (27) .140 .092 .189 Dyadic (35) .126 .088 .164 Monadic (9) .132 -.027 .291

Publication Bias Over twenty (of 45) papers unpublished Still, can investigate easily with standard techniques Funnel plots of estimate against precision indicates weak right skew Many estimates outside 95% confidence interval! Results in Figure 2 Conclude: little evidence of publication bias But worrying dispersion!

Publication Bias

Why do EMU Estimates Vary Across Studies? Rising with (log) observations Small (positive) effect of years in EMU Positive (big) effect of span in years Positive (big) effect of number countries Histograms, scatterplots, regressions all provided in Figure 3 Note paucity of observations Special note: usually very few countries in sample

EMU Effect and Sample Size

Meta Regressions of EMU Trade Effect Weight Std. Err. Obs-1 GSCites-1 Log Countries .16 (.06) .15 (.05) .20 .11 (.04)   Log Years .14 .13 .09 Obs -.05 -.03 (.03) .01 Time-Vary Country FE -.04 (.07) -.07 -.00 Exports not Trade .07 .04 .05 Dyadic FE .03 .02 -.01 P(value) .63 .81 .78 Adjust R2 .26 .27 .47 .29

Confirmation via Meta-Regression Want to check ocular evidence Strong positive effect of #countries Strong positive effect of #years Other effects? Check via Meta Regression Analysis Check for sensitivity to weighting Check for other determinants

Quick Summary In literature: longer, wider spans of data over both time and countries systematically associated with higher estimate of EMU trade effect Curious … extra data increases γ even if extra observations not directly relevant to EMU! (Explains why these observations – e.g., small/poor countries – often omitted from studies; natural to include only relevant observations when estimating EMU trade effect – encompassing)

Caveat But … only 7 papers in literature use preferred methodology (exports, dyadic and time-varying country fixed effects) … and most papers use few countries (median 22), years (median 20) So, seems wise to check meta-results with actual data, plain-vanilla methodology

What’s Trustworthy? Measuring Trade Effects Newer (Export) Gravity Models Much work on “theory-consistent” gravity estimation Use Least Squares with time-varying country Dummy Variables (LSDV) (+ dyadic FE) to control for multilateral resistance, other general equilibrium effects: ln(Exportsijt) = CUijt + Zijt+ {λit} + {ψjt} + {θij} + ijt Xijt = nominal bilateral exports from i to j at time t, {λit} = set of time-varying exporter dummy variables, {ψjt} = set of time-varying importer dummy variables {θij} = set of time-invariant pair-specific dummy variables

With non-EMU and EMU CUs Gravity  Estimates for Exports with country-year effects for exporter & importer & country pair FE  Table 5 Aggregate With non-EMU and EMU CUs dis-aggregated All CUs .34 (.02) All Non-EMU CUs   .30 (.03) EMU .43 (.02) Sample period 1948-2013 #Obs. 879,794 #Country-year effects 22,438 #Pair FE 33,886 e.43 -1 ~ 54% Note: Other gravity regressors and year dummies included, but not reported. Robust standard errors in parentheses.

Tangent: Allow for Dynamic Effects Add (14) leads and lags around currency union exit/entry i.e. Add ΣkθkCUENTRYijt-k + ΣkφkCUEXITijt-k to gravity equation Distinguish effects between EMU/non-EMU exit and entries Estimate with pair FE Test for Symmetry (post-entry = - post-exit) Find symmetry holds well

Allowing Dynamic Effects, CU exit lowers exports, entry raises exports Figure 2

Symmetry Tests, Exports with country-year and pair FE Table 6 F-stat (p value) After CU Entry = - After CU Exit? .8 (.71) Before CU Entry = - Before CU Exit? .8 (.68) Both 1.0 (.49) After non-EMU CU Entry = After EMU Entry? 1.3 (.17) Before non-EMU CU Entry = Before EMU Entry? 1.4 (.16) 2.8 (.00) After non-EMU CU Exit = - After EMU Entry? .9(.51) Can’t reject Can’t reject Table reports F-test statistic for Ho of identical slopes Σkθk Σkφk for given CU pairs and time periods

Sensitivity Analysis of  Estimates : Dis-aggregating Other CUs   EMU .43** (.02) .43** (.02) Other CUs .30** (.03) -.10 (.06) CFA Franc .58** (.10) ECCU $ 1.64** (.11) Aussie $ .39 (.20) Brit. £ .55** (.03) French Franc .87** (.08) Indian Rupee .52** (.11) US $ -.05 (.06) Note: Other gravity regressors, country-year and pair dummies included, but not reported. 879,794 annual observations, 1948-2013.

Sensitivity Analysis of  EMU Estimates: Varying Country and Sample Period   for EMU 1948-2013 1995-2013 1948-2005 1985-2005 1995-2005 All Countries .43** (.02) [879,794] .47** (.03) [424,230] .18** (.03) [691,074] [386,653] .18** (.04) [235,510] Upper Income Countries (GDP p/c>$12,736) .11** (.03) [75,468] .16** (.03) [45,401] -.02 (.04) [52,103] -.01 (.04) [35,865] -.09* (.04) [22,036] Industrial Countries + Present/future EU -.01 (.02) [73,253] .04 (.02) [26,763] -.09** (.03) [61,939] -.16**(.03) [27,570] -.07 (.04) [15,449] Present/future EU -.27** (.02) [30,731] -.04 (.02) [13,337] -.31** (.04) [25,115] -.29**(.03) [12,230] -.10** (.03) [7,721] Note: dependent variable is log exports. Other gravity regressors, country-year and pair dummies included, but not reported. Robust standard errors in parentheses; no. of obs. in brackets.

Dimensionality Effects Adding more years increases γ! Adding more countries increases γ Consistent with meta-regressions!

Gravity Estimates of EMU Effect Varying end dates and country samples 2001 2003 2005 2007 2009 2011 2013 All .08 (.05) .12 (.04) .17 (.03) .19 .25 (.02) .36 .43 Rich .00 -.01 -.02 -.04 .07 .11 EU -.33 (.06) -.36 -.32 -.28 -.26 (.07) -.25 -.24

Graphical Estimates of EMU Export Effect

Conclusions from Meta-Regression-cum-Regression Analysis Throwing away data easily allows one to estimate small/negative EMU export effect Adding years of data in EMU (relevant!) increases EMU export effect Adding countries outside EMU (seemingly irrelevant!) decreases EMU export effect

Why the Differences? Anderson and van Wincoop (2003, p 176); multilateral trade resistance depends positively on trade barriers with all trading partners Dropping small and/or poor countries (likely to have systematically different trade resistance) leads to biased estimates of multilateral trade resistance; higher multilateral resistance leads to more trade. Downward-biased estimates of multilateral resistance biases γ down. Multilateral trade resistance is a function of all bilateral trade barriers, so all trade partners should be included

Estimates of Multilateral Resistance: Evidence of Bias 2001 2003 2005 2007 2009 2011 2013 All .08 (.05) .12 (.04) .17 (.03) .19 .25 (.02) .36 .43 Rich .00 -.01 -.02 -.04 .07 .11 EU -.33 (.06) -.36 -.32 -.28 -.26 (.07) -.25 -.24 Observations 597,565 642,571 688,519 735,025 782,047 829,708 877,736 42,673 46,851 51,824 57,317 62,764 68,428 75,096 22,887 24,341 25,788 27,350 28,891 30,434 31,982 Average of Exporter/Importer-Year Fixed Effects, {λit}, {ψjt} for rich-country Observations Exporter (All) .94 .95 .97 .98 1.00 Importer (All) 1.18 1.17 1.16 1.15 1.14 1.12 1.10 Average of Exporter/Importer-Year Fixed Effects, {λit}, {ψjt} for other non-rich Observations -.07 -.08 -.09 -.10 Average difference between Full and Rich sample estimates for EMU observations Exporter .65 .64 .62 .66 .67 Importer 1.06 .99 .86 .77 .68

Summary Glick-Rose (2002) concluded Based on: “a pair of countries which joined/left a currency union experienced a near-doubling/halving of bilateral trade.” Based on: Assumption of symmetry between currency union exits and entries Caveat: EMU might be different from other currency unions Our results insensitive to precise econometric methodology Here, re-estimate using variety of models, annual panel >200 countries, 1948-2013, 15 EMU years

Conclusions Symmetry holds between currency union entry and exits Methodology and sample matter Preferred methodology is panel with country-pair fixed effects Preferred sample includes all countries, all periods of time Symmetry holds between currency union entry and exits EMU is different EMU boosts trade by 50% Other currency unions have different effects on trade

Conclusion/Summary: Why do Estimates of EMU Trade Effect Vary so Much? Varying sample sizes by time and (especially) country More Data is Better! Established via meta-analysis and regressions Truncating sample (omitting small/poor countries) biases downward EMU trade effect in a) theory, b) data, and c) literature Including entire post-war sample of countries/years delivers large estimate of EMU export effect of γ≈.43 or (exp(.43-1≈) 54%! Economically large (may grow) Statistically significant (robust t-statistic>20) Quite consistent with Rose-Stanley survey (2005): 47%

Future Research Handling zero and missing trade observations LS estimates may be biased because of: Heteroskedasticity, and/or Discarded observations of zero/missing trade Santos Silva and Tenreyro propose Poisson pseudo-maximum likelihood to handle both But difficult to use in big panels like ours Interaction of effects of joining CUs and other forms of economic integration, such as regional trade arrangements Many countries joined EMU in years prior to joined EMU