Lexicographic Biases in International Trade

Transcription

1 Biases in International Hua Cheng 1 Cui Hu 2 Ben G. Li 3 1 Peking University, China 2 Central University of Finance and Economics, China 3 Boston College, United States June /36

2 Names of traders should not matter if the market is efficient 2/36

3 At least, names should not matter lexicographically But, there are many such examples. Given alphabetically earlier names: Stocks are more traded (Jacobs and Hillert, 2016; Itzkowitz, Itzkowitz, and Rothbort, 2016) Scholars receive more citations and are more successful (Einav and Yariv, 2006; Huang, 2014; Ray and Robson, 2018) Alumni are more likely to denote (Meer and Rosen, 2011). 3/36

4 We investigate if this exists in int l trade Speaking of reality: most neoclassical huge product space sufficiently heterogeneous traders least personal transactions Speaking of theory: most GE-ish decentralized decisions efficiency of price mechanism the necessity of market clearance Speaking of identification: bilateral countries use different languages trader (firm) names are ranked differently in different languages thus, diff-in-diff 4/36

5 Our laboratory: China China: not Indo-European Chinese: language has two forms character form romanization form Chinese exports largest number of firms serve a wide range of countries 5/36

6 Literature 1 bias (cited earlier): not limited to English/Anglo-cultures not limited to machine-driven contexts 2 Language & trade (Egger and Lassmann, 2012, 2015; Egger and Toubal, 2016; Melitz and Toubal, 2014) A new dimension where language plays a role Name and language interaction 3 Information & trade: they focus on how information matters Profitability updating (Rauch and Casella, 2003; Albornoz et al., 2012) Foreign marketing (Arkolakis, 2010) Costly searching (Allen, 2014) Networking (Rauch, 1999, 2001; Rauch and Trindade, 2002; Rauch and Watson, 2004; Chaney, 2014) We focus on what information matters 6/36

7 1: Chinese language The basic unit of the Chinese language is a logogram Figure 1: A Chinese Logogram This Chinese logogram (romanized as zhong) means middle. Its character order is 4, while its romanization order is 26 (as z in the English alphabet). 7/36

8 1, cont d Every Chinese character has two lexicographic orders: Character order: ascending in stroke count Romanization order: using the modern English alphabet, A to Z (1 to 26) Words are ordered by their initials 8/36

9 Example: lexicographic orders Char. order 4 Rom. order 26(Z) Romanized: Meaning China (Z) 8(H) Zhongxing Telecom. Equipment (ZTE) Corporation Huawei Technologies Co. 9/36

10 Distribution of initials (Char. order) Compared against modern Chinese use (data on the latter from Da (2004)) 10/36

11 Distribution of initials (Rom. order) Compared against modern Chinese use (data on the latter from Da (2004)) 11/36

12 2: Language proximity For Char. order: language proximity with Chinese For Rom. order: language proximity with English (Chinese romanization uses the modern English alphabet) Latin, Greek, Phoenician, Hebrew, Arabic. 12/36

13 Our measure of language proximity Melitz and Toubal (2014): aggregate language-to-language proximity to country-to-country language proximity Table 2: Language Proximity (Descending Order) With Chinese With (American) English Japan 0.57 USA 0.42 France 0.33 Russia 0.32 Germany 0.24 Germany 2.59 France 1.42 Russia 1.30 Japan 0.55 China /36

14 ln T fct = β 1 Char Order f Chn Proximity c + β 2 Roma Order f Eng Proximity c + δ ft + δ ct + ɛ fct, f firm, c country, t year δ ft : firm-year FE δ ct : country-year FE Can be considered as (firm-level) gravity equation Firm-level cleverness? δ ft Country-level immigration change? δ ct 14/36

15 Some findings to start with (1/3) Without interaction terms: 1 Do lexicographic orders matter unilaterally? 2 To add firm FEs, we use relative orders within a year (the same absolute order ranks differently across years) Measure of order Absolute Relative Character order (0.0013) (3.6703) Romanization order (0.0003) (0.8228) Control variables Yes Yes Fixed effects None Firm Dependent varialbe is log(export volume). Robust standard errors in parentheses, clustering at the firm-country level. *** p<0.01, ** p<0.05, * p< /36

16 Some findings to start with (2/3) Measure of order Relative Relative Character order *** (0.0483) Romanization order *** (0.0551) Chinese proximity *** * (0.0073) (0.0076) English proximity *** *** (0.0023) (0.0024) Character order X Chinese proximity *** *** (0.0126) (0.0131) Romanization order X English proximity ** ** (0.0043) (0.0045) Fixed effects Firm and Year Firm-year pair to be continued next slide /36

17 Some findings to start with (2/3): control variables (Continued) ln(distance) *** *** (0.0018) (0.0019) ln(gdp of destination country) *** *** (0.0005) (0.0006) Dummy: Common legal system (before transi *** *** (0.0031) (0.0032) Dummy: Common legal system (after transiti *** *** (0.0032) (0.0033) Dummy: Common religion *** *** (0.0993) (0.1030) Dummy: Free trade agreement *** *** (0.0025) (0.0026) Dummy: Common borders *** *** (0.0037) (0.0038) Fixed effects Firm and Year Firm-year pair 17/36

18 Some findings to start with (3/3): some eyeballing (Median export volume to HIGH-LP proximity countries) MINUS (Median export export volume to LOW-LP proximity countries) 18/36

19 results Table 4: Results (1) (2) (3) Panel A: Full sample Character order X Chinese proximity *** *** (0.0011) (0.0011) Romanization order X English proximity *** *** (0.0002) (0.0002) Firm-year and country-year fixed effects Y Y Y Dependent varialbe is log(export volume). Robust standard errors in parentheses, clustering at the firm-country level. *** p<0.01, ** p<0.05, * p< /36

20 I: Every single year ( ) Character order X Chinese proximity *** *** *** *** *** (0.0033) (0.0030) (0.0027) (0.0023) (0.0025) Romanization order X English proximity * * * *** *** (0.0005) (0.0004) (0.0004) (0.0003) (0.0004) Firm-year and country-year fixed effects Y Y Y Y Y Observations 699, ,084 1,083,908 1,535,582 1,334, Character order X Chinese proximity *** *** *** *** *** (0.0022) (0.0020) (0.0019) (0.0021) (0.0021) Romanization order X English proximity *** *** *** *** ** (0.0003) (0.0003) (0.0003) (0.0003) (0.0003) Firm-year and country-year fixed effects Y Y Y Y Y Observations 1,587,000 1,855,434 2,047,199 1,717,734 1,740,466 Dependent varialbe is log(export volume). Robust standard errors in parentheses, clustering at the firm-country level. *** p<0.01, ** p<0.05, * p< /36

21 II: Put zeros back Margin: Sample Top 50 destination countries (by firm number) Intensive with zeros Top 50 destination countries (by export volume) Top 8,000 firms (by destination country number) Character order X *** *** *** Chinese proximity ( ) ( ) ( ) Romanization order X *** ** *** English proximity ( ) ( ) ( ) Firm-year and country-year fixed effects Y Y Y Observations 91,812,950 91,812,950 4,485,814 Dependent varialbe is log(export volume). Robust standard errors in parentheses, clustering at the firm-country level. *** p<0.01, ** p<0.05, * p< /36

22 III: English only (1) (2) (3) (4) (5) Character order X *** *** *** *** *** Chinese proximity (0.0011) (0.0011) (0.0011) (0.0011) (0.0011) Romanization order X *** English as official language (0.0003) (0.0004) Romanization order X *** *** English as spoken language (0.0004) (0.0006) Romanization order X *** ** English as native language (0.0006) (0.0008) Romanization order X *** *** Alternative English proximity (0.0001) (0.0001) Firm-year and country-year fixed effects Y Y Y Y Y Observations 14,516,125 14,516,125 14,516,125 14,516,125 14,516,125 Dependent varialbe is log(export volume). Robust standard errors in parentheses, clustering at the firm-country level. *** p<0.01, ** p<0.05, * p< /36

23 I: By volume orders matter more for obscure firms. But how to measure visibility? Panel (a) Panel (b) Firm visibility Firm-year visibility High Low High Low Character order X *** *** *** *** Chinese proximity (0.0012) (0.0025) (0.0012) (0.0018) Romanization order X *** *** *** *** English proximity (0.0002) (0.0003) (0.0002) (0.0003) Panel (c) Panel (d) Firm-year-country visibility Firm-year-country-product visibility High Low High Low Character order X *** *** Chinese proximity (0.0010) (0.0008) (0.0010) (0.0012) Romanization order X *** *** *** *** English proximity (0.0001) (0.0001) (0.0001) (0.0002) 23/36

24 II: By product Panel A: Organized-exchange traded products Character order X Chinese proximity *** *** (0.0013) (0.0013) Romanization order X English proximity *** *** (0.0002) (0.0002) Firm-year and country-year fixed effects Y Y Y VS. Panel C: Differentiated products Character order X Chinese proximity (0.0057) (0.0057) Romanization order X English proximity *** *** (0.0010) (0.0010) Firm-year and country-year fixed effects Y Y Y 24/36

25 III: By exporter type Panel A: Different export regimes Export regime Ordinary trade Processing trade Character order X *** *** *** *** Chinese proximity (0.0011) (0.0011) (0.0023) (0.0023) Romanization order X *** *** *** *** English proximity (0.0002) (0.0002) (0.0003) (0.0003) Firm-year and country-year FE Y Y Y Y Y Y Panel B: Different exporter types Exporter role Intermediaries Non-intermediaries Character order X *** *** *** *** Chinese proximity (0.0017) (0.0017) (0.0015) (0.0015) Romanization order X *** *** *** *** English proximity (0.0003) (0.0003) (0.0002) (0.0002) Firm-year and country-year FE Y Y Y Y Y Y 25/36

26 Magnitudes of the bias For the Character-Chinese combination: ˆβ Char Order Chn Proximity (1) For the Romanization-English combination: ˆβ Roma Order Eng Proximity (2) where is interquartile (from 25th percentile to 75th percentile) across firms (f ) or countries (n) 26/36

27 Quantitative Implications (Reduced-form Approach) Interquartile imputations (export volume): Character-Chinese based: Romanization-English based: Pro Simple and clear Con Comparative statics Also, the mechanism remains unclear 27/36

28 Structural approach, a la Allen (2014) Basic idea: Importers search exporters in lexicographic group l with probability s l As in Allen s model, the search is for better prices We estimate ŝ l To obtain the lexicographic bias in search probability Then, with other parameters and data, we can estimate the lexicographic bias in export volume Merits: (of Allen s approach) Aggregated data: lexicographic group-country-product-year (lngt) (of our setup) Can use reduced form to check 28/36

29 Model (3-min version) Importers from country n conduct undirected sequential search in China for product g in order for prices lower than those outside China with capacity φ, Pareto-distributed (θ n ) cost is f ng per search Upon paying f ng, she draws a Chinese exporter (name) and receives a price quote. Her probability of drawing from lexicographic group l is s lng. To decide whether to accept the price quote or search again: ( ) p max ng V ng (p; φ) = min φp, V ng (p ; φ)df ng (p ) + f ng, (3) p min ng where F ng ( ) is the cdf of prices in China in that ng-importer s belief. 29/36

30 Estimation (1-min version) Share of imports of product g by country n from the lexicographic group l is where K ng (p) λ lng = s lng K ng (p ng ) 1 θn png p p min ng is the value of search (p p)df ng (p ) = p lng K ng (p) θn 2 dp, (4) p p min ng (p p )df ng (p ) (5) Remember the first goal is to estimate 30/36

31 A reduced-form check: ŝ lng! Countries import from Full the lexi group for at least 4 years Panel A1: the Character-Chinese combination Character order X ** *** Chinese proximity ( ) ( ) Observations 83,165 79,432 Product year and country FEs Panel A2: the Romanization-English combination Romanization order X * * English proximity ( ) ( ) Observations 63,031 60,329 Product year and country FEs /36

32 Counterfactuals vs. factuals 1 Information frictions come from {s} 2 Use the {ŝ} to make predictions 3 Turn off {s} 4 Make predictions again 32/36

33 Counterfactuals vs. factuals (cont d) Under the hood: construct Q lng by excluding a reference l: png ŝ Q predicted lng ng dp lng = png ŝ lng Kng θn 2 dp. (6) p lng p lng K θn 2 33/36

34 Ultimately, what interests us is Q g ( prox75 Q lexi75 /Qprox75 lexi25 Q prox25 lexi75 /Qprox25 lexi25 ) g = p prox75,g ŝ lexi75,prox75,g p prox75,lexi75,g K θ prox75 2 prox75,g dp p prox75,g ŝ lexi25,prox75,g p prox75,lexi25,g K θ prox75 2 prox75,g dp p prox25,g ŝ lexi75,prox25,g p prox25,lexi75,g K θ prox25 2 prox25,g dp p prox25,g ŝ lexi25,prox25,g p prox25,lexi75,g K θ prox25 2 prox25,g dp Here is the structural interquartile-based quantitative implications: ln Q g = (ln Q prox25 lexi75 ln Qprox25 lexi25 }{{} ) interquartile comparison in lexicographic order (ln Q prox75 lexi75 Qprox75 lexi25 }{{} ) interquartile comparison in lexicographic order, (7) (8) 34/36

35 Quantitative Implications (Structural) Findings (interquartile range double ratio means) Character-Chinese based: Romanization-English based: To summarize the quantitative implications, Magnitudes of the lexicographic bias: at least 1 percent, under both combinations, using both approaches checks: e.g. structural approach sample + reduced form approach 35/36

36 Take away order is a comparative advantage At least one percent 36/36