The Causal Impact of Common Language on International Trade: Evidence from a Spatial Regression Discontinuity Design Peter H. Egger1,2 , Andrea Lassmann1 1 2
ETH Zurich CEPR, CESifo
Workshop Presentation at NTU Singapore, August 14, 2013
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Language and culture Language in the humanities & social sciences Language differences as cultural barriers to human interaction [philosophy (Herder, Wittgenstein); sociology (Gumperz 82); psychology (Matser et al. 10); economics (Falck et al. 12)]
Common culture by way of common language as a robust driver of cross-border activity [migration (Chiswick 92); multinational firm activity (Bergstrand & Egger 07); international trade]
Common language drives bilateral trade [Head & Ries 98; Rauch & Trindade 02; Melitz 08; Firdmuc & Fidrmuc 09; Melitz & Toubal 12]
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Language and cross-border transactions Channels of influence Common language lowers transaction costs (common spoken language, CSL) Common culture entails knowledge of social norms (common native language/cultural traits, CNL) Problem Cultural element of common native language is correlated with vast number of confounding country-specific factors: cultural, economic, institutional, legal, and political There is a quest for identification
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Language and trade Effect of language on pattern of trade Modern gravity models of bilateral trade have this form Mij = e 𝜆langij dij 𝜇i mj uij
(1)
Mij ∼imports; langij ∼binary common language indicator; 𝜆 ∼direct effect of common language on imports; dij ∼non-language trade cost effects (distance, borders, tariffs, trade agreements); mj ∼importer stuff (GDP, consumer price index); 𝜇i ∼exporter stuff (productivity, factor costs, endowments, etc.); uij ∼stochastic term Vast quantitative evidence on overall language effect in cross-country studies (Egger & Lassmann 12) ˆ is very sensitive to in-/exclusion of control variables ⇒ 𝜆 endogenous common culture Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Interpretation of 𝜆
Facets of common language (Melitz & Toubal 12): Common Spoken Language (CSL) Common Native Language (CNL) Common Official language (COL)
𝜆 reflects weighted impact of CSL as mere communication means and CNL as contextual cultural factor (Melitz & Toubal 12) ˆ is difficult: e 𝜆langij ∼ e 𝜆CSL CSLij e 𝜆CNL CNLij Interpretation of 𝜆 Earlier estimates can not be interpreted as evidence of causal effect of common culture on trade
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Roadmap for the talk
1
Identifying assumptions for the causal effect of the common cultural aspect of common native language
2
Data on trade patterns and language use in Switzerland
3
Technique – a fuzzy RDD for estimating the LATE of common native language on trade
4
Main results for various margins of trade
5
Sensitivity analysis and extensions
6
Conclusions
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Identification strategy Imports of j from i and of k from i: Mij Mik
= e 𝜆CNL CNLij +𝜆CSL CSLij dij 𝜇i mj uij , = e
𝜆CNL CNLik +𝜆CSL CSLik
dik 𝜇i mk uik
(2) (3)
Pick countries and zip codes such that CNLij = 1 while CNLik = 0, CSLij ≈ CSLik and dij ≈ dik . Then, uij Mij = e 𝜆CNL Mik uik
(4)
𝜆CNL estimated as a constant to log-transformed relation ship u in (4) requires independence of (CNLij − CNLik ) and ln uikij
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Figure: Languages spoken in Switzerland by majority
Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Causal effect of CNL on trade This paper: causal impact of CNL on international trade from spatial Regression Discontinuity Design (RDD) Consider three official (spoken and understood) languages of Switzerland spoken in adjacent countries, German, French, Italian CNL as cultural proximity rather than CSL Compare to na¨ıve gravity regressions for Swiss regions using equ. (1): 0.99 for import value and 0.81 for import transactions Consider various margins of trade and a number of extensions
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Data Determine CNL shares at Swiss zip code level (2 language borders, 3 languages: German, French, Italian) using 2000 Census data (Swiss Federal Statistical Office) Combine with Swiss transaction-level data on external goods imports transactions (> 100m observations) from border countries of Swiss Federal Customs Administration (averaged over years 2006-2011, aggregated by zip code and destination) Determine nearest road (1482 zip codes) and great circle distance (1700 zip codes) of every zip code centroid to respective internal (≤ 50 km) and international CNL border (1990 Census and spatial data from Swisstopo/“Die Post”) Language borders largely unchanged since centuries
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Figure: Share of German-speaking population in Switzerland
German language share (Equal intervals) 0.00 - 0.20 0.21 - 0.40 0.41 - 0.60 0.61 - 0.80 0.81 - 1.00
Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Figure: Share of French-speaking population in Switzerland
French language share (Equal intervals) 0.00 - 0.20 0.21 - 0.40 0.41 - 0.60 0.61 - 0.80 0.81 - 1.00
Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Figure: Share of Italian-speaking population in Switzerland
Italian language share (Equal intervals) 0.00 - 0.20 0.21 - 0.40 0.41 - 0.59 0.60 - 0.79 0.80 - 0.99
Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
0
Share of language population .2 .6 .8 .4
1
Native language share in %
-50
-40
-30
-20 -10 0 10 20 Distance to language border in km Egger, Lassmann
RDD Language and Trade
30
40
50
Introduction
Data
RDD set-up
Results
Robustness
Figure: Transactions from German-speaking countries in % of total imports
Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Figure: Transactions from French-speaking countries in % of total imports
Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Figure: Transactions from Italian-speaking countries in % of total imports
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Figure: Transactions from ROW in % of total imports
Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Fuzzy RDD
Quasi-experimental design [Thistlewaite & Campbell 60; Imbens & Lemieux 08; Angrist & Pischke 09, Lee & Lemieux 10]
Use discontinuous determination of treatment status through forcing variable (xij ) which otherwise determines outcome (yij ) continuously Flexible parametric/nonparametric control function about xij Sharp RDD (Br¨ ugger, Lalive & Zweimueller 09; Eugster & Parchet 11) versus fuzzy: discontinuity in CNL is about 0.65 ⇒ Fuzzy RDD
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Fuzzy RDD Difference at limit in space (internal distance to CNL border → 0) of outcomes on ”left” (non-CNL with exporter) and ”right” (CNL language with exporter) at internal border Outcome yij : CNL as fixed (values and transactions shares, # products traded) vs variable trade cost (unit value, value per transaction, quantity per transaction) Fuzzy treatment CNLij : language shares by zip-code; grouped data Instrument Ruleij : binary historical language zone by zip-code (∼ 1 if share > 50%, 0 else) Forcing variable xij : distance to language border (< 0 for non-CNL and ≥ 0 for CNL) Covariate zij : demeaned distance to national border Samples: all versus intra-cantonal language borders only (homogeneous institutions) within 50km from borders Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Fuzzy RDD
Define sufficiently smooth polynomial/nonparametric functions: f0 (xij ) at CNLij < 0, f1 (xij ) at CNLij ≥ 0.5, f1∗ (xij ) ≡ f1 (xij ) − f0 (xij ). Conditional expectations of yij : E [yij ∣xij , CNLij < 0.5] = 𝛼 + f0 (xij ) E [yij ∣xij , CNLij =≥ 0.5] = 𝛼 + 𝜆CNL + f1 (xij ) E [yij ∣xij , CNLij ≥ 0.5] − E [yij ∣xij , CNLij < 0.5] = 𝜆CNL + f1∗ (xij )
Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Fuzzy RDD Definition of fuzzy Local Average Treatment Effect (LATE) Estimate by 2SLS: ≡
ATE
=
E [yij ∣xij , 0 ≤ xij ] − E [yij ∣xij , xij < 0] [CNLij ∣0 ≤ xij ] − [CNLij ∣xij < 0] [ ] f1∗ (xij ) 𝜆CNL + E [CNLij ∣0 ≤ xij ] − E [CNLij ∣xij < 0]
(5)
Nonparametric version in small neighborhood around discontinuity: LATE
≡ =
(E [yij ∣0 ≤ xij < Δ] − E [yij ∣ − Δ < xij < 0]) ([CNLij ∣0 ≤ xij < Δ] − [CNLij ∣ − Δ < xij < 0]) 𝜆CNL . lim
Δ→0
Egger, Lassmann
RDD Language and Trade
(6)
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
10
20
Share of import value 40 30
50
60
Import value share in %
-50
-40
-30
-20 -10 0 10 20 Distance to language border in km Egger, Lassmann
RDD Language and Trade
30
40
50
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
10
Share of import transactions 30 40 20
50
Import transaction share in %
-50
-40
-30
-20 -10 0 10 20 Distance to language border in km Egger, Lassmann
RDD Language and Trade
30
40
50
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
0
Number of imported products (HS 8) 200 400 600
800
Number of products
-50
-40
-30
-20 -10 0 10 20 Distance to language border in km Egger, Lassmann
RDD Language and Trade
30
40
50
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
4
Log import unit value 5 4.5
5.5
Log unit value
-50
-40
-30
-20 -10 0 10 20 Distance to language border in km Egger, Lassmann
RDD Language and Trade
30
40
50
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
7.8
Log intensive margin of imports 8 8.2 8.4 8.6
8.8
Log average value per transaction
-50
-40
-30
-20 -10 0 10 20 Distance to language border in km Egger, Lassmann
RDD Language and Trade
30
40
50
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
6
Log import quantity per transaction 6.5 7 7.5
8
Log average quantity per transaction
-50
-40
-30
-20 -10 0 10 20 Distance to language border in km Egger, Lassmann
RDD Language and Trade
30
40
50
Introduction
Data
RDD set-up
Results
Robustness
Results: Great-circle distance (all borders)
Value share Obs. Trans. share Obs. # Products Obs. Value/Quant. Obs. Value/Trans. Obs. Quant./Trans. Obs.
Parametric 1st 2nd ∗∗∗ 0.231 0.173∗∗∗ 3414 3414 0.227∗∗∗ 0.191∗∗∗ 3414 3414 169∗∗∗ 124∗∗ 3414 3414 0.148 0.085 3395 3395 –0.004 –0.060 3395 3395 –0.026 –0.085 3395 3395 Egger, Lassmann
Nonp. 3rd 0.131∗∗∗ 3414 0.163∗∗∗ 3414 118 3414 –0.194 3395 –0.037 3395 –0.310 3395
RDD Language and Trade
0.160∗∗∗ 1304/19 0.159∗∗∗ 1227/18 164∗∗∗ 1249/18 –0.008 1689/24 –0.044 1859/27 -0.083 2731/39
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Results: Cont’d (intra-cantonal borders)
Value share Obs. Trans. share Obs. # Products Obs. Value/Quant. Obs. Value/Trans. Obs. Quant./Trans. Obs.
Parametric 1st 2nd ∗∗∗ 0.202 0.126∗∗∗ 1872 1872 0.209∗∗∗ 0.157∗∗∗ 1872 1872 149∗∗∗ 115∗ 1872 1872 0.092 0.099 1858 1858 –0.120 –0.171 1858 1858 –0.098 –0.279 1858 1858 Egger, Lassmann
Nonp. 3rd 0.096∗ 1872 0.145∗∗∗ 1872 203∗∗ 1872 –0.256 1858 –0.279 1858 –0.256 1858
RDD Language and Trade
0.136∗∗∗ 890/21 0.168∗∗∗ 779/18 178∗∗∗ 1007/24 –0.008 1224/29 –0.181 1420/34 –0.182 1084/25
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Results: Road distance (all borders)
Value share Obs. Trans. share Obs. # Products Obs. Value/Quant. Obs. Value/Trans. Obs. Quant./Trans. Obs.
Parametric 1st 2nd ∗∗∗ 0.194 0.184∗∗∗ 2968 2968 0.209∗∗∗ 0.201∗∗∗ 2968 2968 184∗∗∗ 85 2968 2968 0.088 0.085 2954 2954 –0.067 –0.083 2954 2954 –0.023 –0.260 2954 2954 Egger, Lassmann
Nonp. 3rd 0.134∗∗∗ 2968 0.187∗∗∗ 2968 90 2968 –0.112 2954 –0.227 2954 –0.275 2954
RDD Language and Trade
0.130∗∗∗ 1061/19 0.188∗∗∗ 1133/21 145∗∗∗ 2928/50 –0.229 855/16 –0.150 1436/25 –0.194 1754/30
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Results: Cont’d (intra-cantonal borders)
Value share Obs. Trans. share Obs. # Products Obs. Value/Quant. Obs. Value/Trans. Obs. Quant./Trans. Obs.
Parametric 1st 2nd ∗∗∗ 0.163 0.162∗∗∗ 1644 1644 0.187∗∗∗ 0.186∗∗∗ 1644 1644 186∗∗∗ 101 1644 1644 0.059 0.134 1633 1633 –0.180 –0.174 1633 1633 –0.146 –0.208 1633 1633 Egger, Lassmann
Nonp. 3rd 0.137∗∗∗ 1644 0.215∗∗∗ 1644 136 1644 –0.065 1633 –0.417 1633 –0.154 1633
RDD Language and Trade
0.138∗∗∗ 720/21 0.204∗∗∗ 696/20 150∗∗ 1602/49 –0.108 704/20 –0.329 742/21 –0.138 1033/29
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Sensitivity analysis and extensions Sensitivity of nonparametric estimates to choice of bandwidth No discontinuities for imports from rest-of-world No placebo discontinuities (at median distance of ≈ ±25 km; and at 15 km) No discontinuities for zip code-specific effects (mj )!!!
return
LATE for additional (product intensive) margins of trade confirms discontinuity for extensive margins LATE across transaction size quartiles: increasing up to third quartile, lowest in fourth quartile (cross-border selling) LATE at distance from language border: increases to > 0.2 with exclusion of 5, 10, 15 km from border (cross-border shopping & working) Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
Sensitivity analysis and extensions LATE across product categories (97 HS 2-digit industries, Rauch goods classification, selected products): Dispersed across industries Matters particularly for reference-priced and differentiated goods rather than homogeneous goods Strong discontinuity for selected differentiated goods (suits, cars)
LATE by border (French-German versus Italian-German language border): 0.284 (value share) and 0.300 (transactions share) at German-Italian language border 0.114 (value share) and 0.150 (transactions share) at French-German language border Importance of geographical barriers (St. Gotthard, Bernina)
LATE for specific native languages (not borders!) Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Results
Robustness
Conclusion
Conclusion What is the causal effect of CNL on trade? CNL raises import volume share by 13 pp., transactions share by 19 pp., and number of products by 145. No significant effect on log unit value, log transaction value and log quantity per transaction CNL matters for fixed trade costs rather than variable trade costs (Helpman, Melitz & Rubinstein 08?) Corresponding semi-elasticities of CNL are 0.28 for log value and 0.31 for log number of transactions Much lower than effect of COL found in na¨ıve gravity regressions for Switzerland (0.99 for import value and 0.81 for import transactions) CNL effect is heterogeneous: matters for reference-priced and differentiated goods rather than for homogeneous goods; hump-shaped effect across transaction quartiles; dispersed across industries Egger, Lassmann
RDD Language and Trade
Introduction
Data
RDD set-up
Results
Robustness
The END
Thanks a lot for your attention!
Egger, Lassmann
RDD Language and Trade
Conclusion
Introduction
Data
RDD set-up
Robustness
Conclusion
return
-1
Region fixed effects (normalized) .5 1 -.5 0
1.5
Zip-code fixed effects
Results
-50
-40
-30
-20 -10 0 10 20 Distance to language border in km Egger, Lassmann
30
RDD Language and Trade
40
50