The Trade Facilitation Impact of the Chinese Diaspora

Using an enhanced data set on the population share of overseas Chinese in 1970 and 1990, this paper analyses the impact of the Chinese diaspora on facilitating China's bilateral trade during the period 1973&#8211;2013. Our findings suggest substantial trade creation effects resulting from the presence of ethnic Chinese in the trade partner population. Diaspora impacts on Chinese bilateral imports are in general higher than those found for exports. Coethnic networks play a larger role as long as the partner country does not have an RTA with China in place. Among export sectors, effects found were strongest for food, as well as for machinery and transport equipment. In regard to imports, coethnic networks matter mostly for raw materials, machinery and transport equipment, and chemicals.


| 2413
MARTÍNEZ-ZARZOSO ANd RUdOLF one-side gravity models of Chinese exports and imports, respectively. Analyses are carried out also at the sectoral level to identify heterogeneity of diaspora effects across main product groups.
The present study contributes to the existing literature in the following ways: first, by using a new dataset on ethnic Chinese compared to earlier studies, we are able to expand the number of countries included in the analysis from 63 (R&T and related studies) to 175. This represents a big increase in the trade flows covered. Second, to the best of our knowledge, this is the first study that focuses on the question how coethnic networks affect bilateral trade flows of a major migrant sending country. Earlier single-country studies on the nexus of migration and trade have all used the perspective of the migrant receiving country (Gould (1994) for the United States; Peri and Requena-Silvente (2010) for Spain; Bratti, Benedictis, and Santoni (2014) for Italy, among others). Third, in contrast to earlier studies that used total trade, we estimate trade facilitation effects separately for exports and imports, accounting for the facts that the preference channel should only affect Chinese exports and that coethnic business networks can be of different importance for the sending country's exporters versus importers. We further study the heterogeneity of effects by trading sector. Fourth, we analyse the role of informal coethnic networks as substitutes for formal trade agreements. Lastly, regarding our methodological approach, we control for time-invariant characteristics of countries and global time trends that can influence migration and trade by applying panel data techniques. In contrast to R&T and similar to Felbermayr et al. (2010), we depart from the traditional gravity model (GM) and use a Poisson pseudo-maximum likelihood (PPML) correlated random-effects estimator. 3 Our estimation technique differs from Felbermayr et al. (2010) in that, whereas they use cross-sectional estimations (for single years: 1980, 1990, 2000), we exploit the longitudinal nature of bilateral trade data and we allow for the fact that China does not export to (import from) all countries in our dataset. In particular, we use a (PPML) correlated random-effects estimator with regional fixed effects that allows us to control for the unobserved heterogeneity that is region-specific and time-invariant. We also use a system-GMM estimator (Blundell & Bond, 1998) that addresses the endogeneity of the Chinese network variable. We further use panel fixed-effects estimations as a robustness check.
Our main results indicate that Chinese coethnic networks indeed positively affect China's bilateral trade flows. We find substantial trade creation effects resulting from the presence of ethnic Chinese in the trade partner population. Diaspora impacts on Chinese imports are higher than those found for exports. Coethnic networks play a larger role as long as the partner country does not have a regional trade agreement with the PRC. Sectoral analyses suggest that, among Chinese exports, diaspora effects are strongest for the food sector, as well as machinery and transport equipment. In regard to imports, coethnic networks matter mostly for raw materials, machinery and transport equipment, and chemicals.
The remainder of this paper is organised as follows. Section 2 describes the data and the empirical strategy used in this study. Section 3 presents main results, extensions to main results and sector-specific analyses, while section 4 reports robustness checks. Conclusions are outlined in Section 5.

| Overseas Chinese data
Following earlier studies that use R&T-style ethnographic data, the Chinese diaspora in this study refers to the group of people that were born in or claim ancestry to China but who reside outside the People's Republic of China, Taiwan, Hong Kong and Macau (Poston, Mao, & Mei-Yu, 1994;Poston & Wong, 2016;Priebe & Rudolf, 2015;Rauch & Trindade, 2002). These definitions therefore involve Chinese born in mainland China but living abroad (first-generation or foreign-born migrants) as well as those Chinese that were born outside mainland China and who continue to live outside of China, in certain countries already for multiple generations. Changes in the number of overseas Chinese over time can therefore be attributed to both, fertility levels among Chinese in the host country and outmigration from China to the host country or from one host country to another.
The present study draws on country-level data on the number of overseas Chinese for 1970 and 1990 from the World Christian Encyclopedia (WCE; Barrett, Kurian, & Johnson, 1982, 2001, which provides one of the most comprehensive ethnographic datasets to date and has recently been used by Priebe and Rudolf (2015) for a related analysis. WCE is a highly detailed census of churches and religions around the world which, as a by-product, offers detailed ethnographic data for most countries. WCE's ethnographic data rely on countries' population census data and additional secondary data sources.
World Christian Encyclopedia data have been frequently used in ethnic-related studies. Reynal-Querol (2005a: 298, 2005b) describe it as "one of the most detailed data for ethnolinguistic diversity." In regard to measuring the size and distribution of the Chinese diaspora during historic periods, WCE's main advantage is that it has data available for many more countries compared to OCAC data used in earlier studies on the trade effects of the Chinese diaspora. For example, the 1981 World Christian Encyclopedia provides data on overseas Chinese for 183 countries. In contrast, OCAC data from its Overseas Chinese Economy Yearbook 1983/84 provide records for only 94 countries (OCAC, 1983). In particular, OCAC reports do not indicate whether countries that are not listed have no overseas Chinese population or are missing due to other reasons (political instability, civil war, lack of census data, etc.). The inability to distinguish missing values from zeros is the main shortcoming of OCAC data.
The WCE dataset that we use in this study provides information on the number of overseas Chinese for a total of 165 countries in 1970 and 186 countries in 1990. In the following analysis, our main variable of interest is the proportion of overseas Chinese in the total population of a host countryhereafter referred to as sharechinese. Table 1 shows the distribution of overseas Chinese across world regions and over time using a combination of OCAC and WCE data and our own imputations. The total size of the Chinese diaspora has been constantly growing over time and across all world regions.
It should be noted that our definition of migration differs from many recent trade-migration studies that use data on first-generation migrants only (the so-called "foreign-born concept," see, e.g., Bratti  Felbermayr & Toubal, 2012;Özden, Parsons, Schiff, & Walmsley, 2011;Peri & Requena-Silvente, 2010). We hypothesise that trade effects of migration go well beyond the first generation of migrants, particularly with regard to the trade-cost channel. Anecdotal evidence from the Chinese diaspora suggests that migrants active in trading often take a generation to build a functioning family business in their new country of residence. Thus, this wider definition of migration adds a new perspective on the trade-migration nexus to the literature.

| Trade and gravity data
Bilateral Chinese exports and imports from 1973 to 2013 are taken from UN-COMTRADE. 4 Data on countries' GDP and population are drawn from the World Development Indicators Database (World Bank, 2016). Distances between capitals, as well as trade impeding or promoting factors such as common border or being landlocked, are taken from the CEPII database (Mayer & Zignago, 2005). RTA and WTO dummy variables are-an actualised version-from De Sousa (2012). Table 2 presents summary statistics of the above variables. 5

| Empirical strategy
Over the past two decades, the gravity model of trade has evolved into a sophisticated tool to analyse the broad determinants of bilateral trade flows, among them a number of policy factors such as regional trade agreements (RTA), trade facilitation factors, tariffs, regulations and others (Feenstra, 2016). The gravity model has been broadly used to investigate the role played by specific policy or geographical variables in explaining bilateral trade flows. Consistent with this approach, and in order to investigate the effect of the presence of Chinese networks on Chinese trade flows, we include the variable sharechinese (in the trade partner country) as a "trade facilitator." According to the underlying theory that has been reformulated and extended by Anderson and van Wincoop (2003), our model assumes constant elasticity of substitution and product differentiation by place of origin. In addition, prices differ among locations due to symmetric bilateral trade costs. 6 The reduced form of the model is specified as: where X ijt is bilateral exports from country i to country j in year t, and Y it , Y jt and Y t W are the GDPs in the exporting country, the importing country and the world in year t, respectively. t ijt denotes trade costs between the exporter and the importer in year t, and P it and P jt are price indices that account for the so-called multilateral resistance factors and are a function of the trade cost of a country with respect to all countries in the world. The empirical specification in log-linear form is given by: 4 Data were downloaded from https://comtr ade.un.org/data/ on 13 April, 2017.
5 Following earlier studies, Hong Kong, Taiwan and Macao have been excluded from the analysis. The data that support the findings of this study are available from the corresponding author upon reasonable request. 6 The assumption of symmetric trade costs is not required in the empirical application. (1) Electronic copy available at: https://ssrn.com/abstract=3706982 where ln denotes natural logarithms. The estimation of Equation (2) is not straightforward due to the presence of trade costs and multilateral resistance terms. The trade-cost function is assumed to be a linear function of a number of trade Notes: ln denotes natural logarithm; exports are in thousands of US$. Y_imp and Y_exp denote gross domestic product of exporter and importer country, respectively. Pop denotes population and area denotes the geographical area of the countries. Dist is the distance between capital cities of origin and destination countries. Border (landlocked) is a dummy variable that takes the value of 1 when the trading countries share a border (do not have an exit to the sea), and zero otherwise. WTO (RTA) is a dummy variable that takes the value of 1 when the trading countries belong to the World Trade Organization (to the same regional trade agreement), and zero otherwise.

| 2417
MARTÍNEZ-ZARZOSO ANd RUdOLF barriers, namely, the time-invariant determinants of trade flows, including distance, area, common border, landlocked dummies and the time-varying RTA and WTO variables. In the recent gravity literature, multilateral resistance terms are modelled as time-varying or time-invariant country-specific dummies when a full-gravity is estimated. However, for a single exporter, some specificity applies. In our empirical application, we focus exclusively on exports from (imports to) China over time for all its trading partners for which the relevant data are available. We therefore specify a one-side gravity model to explain bilateral exports and imports, in which trade partners are indexed by j and years by t.
Since we have a single country (China) as exporter (importer) to (from) any other country in the world, and the target variable is time-invariant and country-specific, we opted by including regional time-invariant dummies in the main specification a way to proxy for multilateral resistance. Regionspecific fixed effects are used in order to mitigate potential biases due to time-invariant unobserved heterogeneity due to cultural factors, for example. Asian countries may have communalities that affect bilateral trade between Asian countries in a different way as trade between countries in different regions. We explain below the panel data techniques used to control also for permanent country heterogeneity. We account for global trends in trade by adding common time dummies.
After dropping the i subscript, substitution of the trade-cost function into Equation (2) and addition of regional dummies, time dummies and an idiosyncratic error term suggests estimating: where Dist j denotes geographical distance from China to country j. Landlocked j takes the value of one when country j is landlocked, and zero otherwise. Border j takes the value of one when country j shares a border with China, zero otherwise. Area j is the geographical area of country j in squared km, RTA jt takes the value of one when China and the country j are members of the same regional trade agreement in year t, zero otherwise, and WTO takes the value of one when both countries are members of the WTO. Most important for the sake of this study, sharechinese j denotes the population share of ethnic Chinese in country j in 1970 or in 1990. Moreover, t denotes a set of year dummies that proxy for time-variant common factors (globalisation) that affect Chinese trade flows to all its partners. Finally, r denotes regional fixed effects and u jt is the error term that is assumed to be well-behaved.
It should be reasonable to assume that the variable sharechinese j70 is exogenous with regard to Chinese bilateral exports and imports from 1973 to 2013. As noted in Priebe and Rudolf (2015), most Chinese outmigration took place before 1952 and then again after the Open Door policies by Deng Xiaoping in 1978. The PRC's economic isolation in combination with strict migration controls between 1952 and the end of the 1970s is reason enough to believe that overseas Chinese in 1970 were not able to foresee Chinese bilateral trade during the four decades post-1973. The use of sharechinese j90 , on the other hand, is more controversial. While potentially it allows us to estimate the relationship between overseas Chinese and Chinese bilateral trade with more up-todate population figures and for a larger sample of countries, it could clearly violate the exogeneity assumption of our estimators. To remedy this problem, we use only the trade period 1991-2013 in our estimations with this variable, in combination with other approaches to establish robustness (system-GMM).
As regards the techniques used to estimate the gravity model, the main novelties are reviewed by Head and Mayer (2014). The authors discuss the main trade theories supporting the model and estimation challenges involved for correct identification of trade effects of specific economic and political factors. In our choice of estimation techniques, we have to consider that we are estimating a one-side gravity and that the variables in the model cannot capture all influences on China's trade. The use of (3) ln X jt = 0 + 1 ln Y jt + 2 ln Dist j + 3 Landlocked j + 4 ln Area j + 5 Border j Electronic copy available at: https://ssrn.com/abstract=3706982 panel data techniques enables us to control for permanent unobserved country-specific heterogeneity. In particular, we specify the gravity model as: where α j is an unobserved country-specific effect that represents the permanent cross-country heterogeneity. The model could be estimated using a random-effects approach if α j is assumed to be uncorrelated with the regressors. Since this assumption is difficult to maintain in practice, 7 we adopt a correlated random-effects approach and assume that the country-specific effects are a function of the time averages of the time-variant variables: Substituting Equation (5) into Equation (4) we obtain: Equation (6) can be estimated using random effects. Since the above approach does not allow for zero trade and given that for the regressions with sectoral data an important percentage of the observations are zeros or missing, we also estimate a PPML correlated random-effects model. According to it, the dependent variable is in levels. It has also the added advantage that the model is robust to heteroscedasticity in the error term.
Finally, in order to account for the potential endogeneity of the target variable (sharechinese) and to allow for dynamics in the dependent variable, we estimated a system-GMM model in which the first lag of the dependent variable is added as regressor and internal instruments (further lags of the dependent variable and lags of the other time-variant variables) are considered as instruments of the potentially endogenous variables, that is the lag-dependent variable and the sharechinese variable.

| Chinese diaspora and China's bilateral trade
The main results for Chinese bilateral exports and imports are presented in Tables 3 and 4, respectively. In both tables (columns (1) to (4)), we use sharechinese in trade partner countries in 1970 (sharechinese70) for the trade period 1973-2013. In Tables A1 and A2 in the Appendix, we employ sharechinese measured in 1990 (sharechinese90) for the period 1991-2013 instead.
The OLS estimates applied to the traditional specification of the gravity model with time fixed effects and regional fixed effects (OLS_TFE) are shown in column (1), country random effects are added to this specification (RE_TFE) in column (2), and the correlated random-effects estimates, resulting from adding to the RE_TFE model the averages of the time-variant variables as regressors (CRE_TFE) are shown in column (3). Finally, the results from the correlated (4) ln X jt = j + 1 ln Y jt + 2 ln Dist j + 3 Landlocked j + 4 ln Area j + 5 Border j + 6 sharechinese j70(90) + 7 RTA jt + 8 WTO jt + r + t + u jt , 7 It is also rejected by the regression-based Hausman test that consists on testing for the joint significance of the coefficients of the time averages of the time variant variables in Equation (6).

MARTÍNEZ-ZARZOSO ANd RUdOLF
Sharechinese70 is positive and statistically significant in all model specifications and the estimated effects found are sizeable. For bilateral exports (Table 3), the coefficients indicate that a 1-percentage point increase in sharechinese70 in the trade partner population is associated with an increase in Chinese exports to this trade partner ranging between 1.7 and 3.1% in the case of sharechinese70 (between 1.6 and 3.9% in the case of sharechinese90, see Table A1). We regard CRE_TFE and CRE_ PPML as our preferred specifications, which show the most conservative estimates. Other factors that determine Chinese bilateral exports are trade partners' market size, access to the sea and distance to the PRC, which is largely in line with earlier studies such as Caporale et al. (2015). The models show an overall strong fit (R 2 = .87 in model 1). Most standard control variables show expected signs and levels of statistical significance. Table 4 presents an equivalent analysis of China's bilateral imports. Results again suggest a strong role of overseas Chinese in determining China's bilateral trade. Interestingly, coefficients on the import side (Table 4) are found to be generally larger than those on the export side (Table 3). Focusing on our preferred specifications, coefficients for sharechinese70 range between 3.1 and 4.4 (those for sharechinese90 between 4.6 and 5.7, see Table A2). These coefficients indicate that a 1-percentage point increase in the share of ethnic Chinese in the trade partner population is associated with an increase in Chinese imports from the partner country by approximately three to six per cent. The models show a good fit (R 2 = .74 in model 1), although it is lower compared to export models in Table 3.
Coethnic networks might play a more important role in countries with weak institutional environments. In a similar fashion, to the degree that trade relations are not formalised in bilateral or multilateral treaties, informal coethnic networks might play a more important role in facilitating trade between two countries. Therefore, one might want to test whether the importance of the Chinese diaspora in trade relations diminishes after countries form a regional trade agreement with the PRC. Table 5 examines this hypothesis for both Chinese exports and imports by adding an interaction term between RTA and sharechinese to the gravity model (RTA × sh70 and RTA × sh90). Here, the CRE_PPML model has been applied; however, results are similar for other estimators. Estimation results in Table 5 confirm the above hypothesis. When countries form an RTA with the PRC, the effect of sharechinese diminishes by 27 per cent for exports using the full trade period and sharechinese70 (column (1)) and by 25 per cent when using sharechinese90 in column (2). For imports (columns (3) and (4)), the reduction is around 16 per cent. These estimates suggest that, even though the formalisation of trade relations tends to lower the role of informal networks in trade, the effect of overseas Chinese continues to be sizeable and highly significant after two countries form a trade agreement.
The gravity model can further be specified in a dynamic panel framework. Tables A3 and A4 in the Appendix provide further evidence from system-GMM estimations which add the lagged dependent variable to the model. GMM-style instruments are used for lagged exports (imports) in columns (1) and (2) and additionally also for sharechinese in columns (3) and (4). Estimation results show that the lagged dependent variable is indeed significant in most models, the effect being more sizeable for exports. In contrast to imports, Chinese exports seem to follow more predictable patterns. The significance of our main variable of interest, sharechinese, is also confirmed by GMM estimations. Longrun coefficients of sharechinese for exports are 3.6 and 4.6 (columns (3) and (4) of Table A3, when sharechinese70/sharechinese90 is considered endogenous), while the respective long-run coefficients for imports are 7.8 and 8.3 (columns (3) and (4) of Table A4). 8 The interaction term rta × sharechinese is significant in seven out of eight models of Tables A3 and A4, confirming the earlier hypothesis.
Based on our main estimates, we have calculated trade creation effects and the equivalent ad valorem tariff reduction for Chinese exports and imports. Using the estimates from column (4) of Tables 3 and 4 for the CRE-PPML model, for Chinese exports (Table 3) we obtain a trade creation effect of 2.63% 9 (2.42, using Table A1, column (4) estimates) if sharechinese70 (90) moves from zero to the sample average (0.015). 10 For Chinese imports, the trade creation effect is significantly higher at 4.82% (7.08, using Table A2, column (4) estimates). These trade creation effects are equivalent to a hypothetical tariff reduction of about 0.37 (0.34) percentage points for Chinese exports and of about 9 This figure has been obtained using the formula: [Exp(coef × avsh) − 1] × 100; where coef = coefficient of sharechinese70 in column (4) of Table 3 and avsh = sample mean of sharechinese70 (See summary statistics in Table 2). 10 Sample averages of 1970 sharechinese are 0.013 for the export sample and 0.017 for the import sample. Thus, we take the average of the two figures (0.015) for the calculation of trade creation and AVT reduction effects. Notes: Robust standard errors in brackets, cluster by country-pair. Average variables of the time-variant variables, regional dummies, and time dummies are omitted to save space. Stata command used: xtpoisson, re. ***p < .01. **p < .05. *p < .1.

| 2423
MARTÍNEZ-ZARZOSO ANd RUdOLF 0.67 (0.98) for Chinese imports (using an elasticity of substitution of 8, as in Anderson and van Wincoop (2004)). 11 Putting our results into perspective, the magnitudes of our estimates are consistent with those found by Felbermayr et al. (2010), to the best of our knowledge the only earlier study that estimated the trade creation effects of the Chinese diaspora for the PRC's bilateral trade. Table A5 in the Appendix compares our findings with those of Felbermayr et al., who found a trade creation (TC) effect of 2.4-8.2% equivalent to a reduction in ad-valorem tariffs (AVT) by 0.34-1.15 percentage points. We estimate this relationship over a much larger sample of countries and distinguish between export and import creation effects. We find a TC of 2.4-2.6% for exports and 4.8-7.1% for imports, respectively. That is equivalent to an AVT reduction of 0.34-0.37 (0.67-0.98) for exports (imports).
Assuming that the preference channel is only present in host countries' imports from China (Chinese exports in our study), then more products from China could be imported as the diaspora grows and the demand for Chinese products increase, but over time overseas Chinese could start producing substitutes in the destination country and the preference motive could have a negative effect on Chinese exports that could be replaced by locally produced Chinese products. In addition, one can expect that preferences for "Chinese" products diminish for younger generations that claim Chinese ancestry but were no longer born in China, but in the host country. The cost reduction link should also be present in host countries' imports from China since Chinese exporters will have more potential links in host countries and more knowledge of the corresponding foreign markets which often give rise to significant Chinese distribution and retail networks. However, since Chinese exports mainly go to developed countries, the cost channel could be less important on average than for Chinese imports and the preference channel is probably overcompensated by a substitution channel when we consider trade in future periods as we do in our estimations.
Our results indicate that the trade creation effect is significantly higher in magnitude for Chinese imports (host countries' exports to China) than for Chinese exports (host countries' imports from China) and hence support a strong trade facilitation mechanism in place for the former. The Chinese diaspora can be expected to be better informed about the Chinese bureaucracy apparatus, contract law enforcement issues, or even the importance of bribes, and this may result to be very useful when doing business in China. Additionally, being able to trust key partners in China may prove to be very valuable, as uncertainty is reduced. The diaspora has the knowledge of how to export to China, where they have family, relatives and business networks.

| Sector-specific effects
Besides its overall trade expansion, in particular since its WTO accession in 2001, China has shown extraordinary trade growth in specific sectors. Among exports, machinery and transport equipment and other manufactures were the two leading sectors in 2015. The two major import sectors were primary commodities and machinery and transport equipment. Given sector-specific developments, it appears useful to examine the trade facilitation effect of overseas Chinese through the sectoral lens. Table 6 summarises the effects of sharechinese on Chinese bilateral exports and imports by sector using a number of estimation techniques ranging from a log-log model with time and regional fixed effects (first row of each sector), adding random effects (second row), log-log correlated random effects (third row) and PPML with time and regional fixed effects (fourth row). Several observations can be made. First, positive and statistically significant effects prevail in almost all sectors and independent of the estimator and time period used. Second, confirming earlier results, sharechinese90 shows somewhat stronger effects compared to sharechinese70. Third, with regard to exports, results suggest that coethnic networks play a particularly important role in China's overseas sales of food, and machinery and transport equipment. The strong effect on food exports is not surprising and confirms the importance of the preference channel. In contrast, the evidence of overseas Chinese promoting the PRC's exports of machinery and transport equipment points more towards the trade-cost channel. Lastly, with regard to imports, overseas Chinese facilitate Chinese imports in raw materials, machinery and transport equipment, and chemicals. Given the high dependence of the Chinese economy on raw material imports, the Chinese Diaspora appears to play a vital role in the global sourcing of

| ROBUSTNESS
As a first robustness check, we use OCAC data on sharechinese to compare our main results obtained using WCE data. As mentioned earlier, OCAC data provide information on the number of overseas Chinese for significantly fewer countries as compared to WCE data, particularly for periods before 2000, but has been the standard source for data on overseas Chinese used in past studies on the Chinese diaspora. In practice, we estimate the same gravity models as presented in Tables 3 and 4 for exports and imports and replace sharechinese1970 (WCE) with OCAC data on sharechinese for single years. Results of using the OCAC sharechinese variable for 1963 are presented in the first row of the first and second part of Table A6 in the Appendix. The population share of overseas Chinese in trade partner countries has a strong effect on both Chinese bilateral exports and imports. Effects found are higher than those found earlier in Tables 3 and 4, while at the same time the number of countries covered reduced from 155 (150) using WCE data to 82 (80) using OCAC data in export (import) regressions. Table A6 shows further results from regressions using later OCAC years, and it can be observed that coefficients for sharechinese decrease systematically when using more recent years. Therefore, estimates using earlier OCAC years might have been biased upwards due to the sample selection correlated with time that is inherent in OCAC data. While having substantially fewer countries available, OCAC data have the potential advantage that it is measured regularly and thus one can try to use it in longitudinal form. We merged data for the Chinese diaspora for the years 1963, 1984, 2000, 2003, 2005 Table A7 in the Appendix. Three models have been estimated for comparative purposes. First, results of a correlated random-effects model (CRE-TFE) are presented in columns (1) and (4) for exports and imports, respectively. Second, a panel fixed-effect (FE) model, retaining only the within-variation in columns (2) and (5) and, third, the equivalent PPML FE model in columns (3) and (6). The estimations mostly confirm our main results, showing that an increase in the Chinese diaspora increases both exports and imports. However, when using panel methods, the magnitude of the effect is higher for exports than for imports, contrary to what we obtained in Tables 3 and 4. It is worth noticing that the difference in results could be due to unobserved heterogeneity that was left uncontrolled for in the regressions presented in Tables 3 and  4. However, it could also be due to smaller sample size, sample selection that is correlated with time, and the resulting unbalanced nature of the OCAC data panel. OCAC data are particularly lacking for many countries before the year 2000; thus, identification using within-variation only is likely to be biased towards changes in variables in later years.
We further tried to estimate the gravity model using fixed-effects panel data estimations with WCE data for 1970 and 1990 exploiting the difference between the Chinese share in both periods; however, with a single change in the target variable, the results are instable and there is not enough variability to explain changes in exports/imports.
Another sensitivity check, as suggested by Priebe and Rudolf (2015) in their analysis of diaspora effects on economic growth in host countries, consists of estimating the model only for the sample of countries with sharechinese ≤ 0.05, ≤0.03 and ≤0.01. This exercise helps us to verify that results are not only driven by a few countries with high population shares of overseas Chinese such as Singapore, Malaysia or Thailand. Estimating over the restricted sample using WCE data confirms our main results. In accordance with this exercise, we also checked for nonlinearities in the effect of sharechinese by splitting it into several bins and also by using a quadratic specification. Results indicate that the coefficients of different bins were not statistically different and the quadratic term was not significant; thus, the effects do not appear to be nonlinear.

| CONCLUDING REMARK S
This article evaluates the role of the Chinese diaspora in explaining Chinese bilateral trade flows. In order to achieve this goal, we use a new dataset on the population share of overseas Chinese and estimate one-side gravity models of Chinese exports and imports, respectively.
The present study contributes to the existing literature in the following ways: first, by using a new dataset on ethnic Chinese compared to earlier studies, we were able to expand the number of countries included in the analysis from 63 (R&T and related studies) to 175. Second, to the best of our knowledge, this is the first study that focuses on the question how coethnic networks affect bilateral trade flows of a major migrant sending country. Earlier single-country studies on the nexus of migration and trade have all used the perspective of the migrant receiving country. Third, in contrast to earlier studies which used total trade, we estimated trade facilitation effects separately for exports and imports, accounting for the facts that the preference channel should only affect Chinese exports and that coethnic business networks can be of different importance for the sending country's exporters versus importers. We further studied the heterogeneity of effects by trading sector. Fourth, we analyse the role of informal coethnic networks as substitutes for formal trade agreements.
Our findings suggest substantial trade creation effects resulting from the presence of ethnic Chinese in the trade partner population. Among export sectors, effects found were strongest for food, as well as for machinery and transport equipment. In regard to imports, the largest effects were found for raw materials, machinery and transport equipment, and chemicals. Interestingly, for food products we find a higher effect for exports than for imports, indicating that for this specific sector the preference effect could play an important role. It is also worth noting that for raw materials, machinery and chemicals, the trade creation effects are in general higher for imports than for exports, supporting the importance of the trade-cost channel for these sectors. Putting our results into perspective, trade creation effects found are consistent with those of earlier studies such as Felbermayr et al. (2010). We are able to distinguish between Chinese export creation and import creation effects and find a trade creation of 2.4-2.6% and 4.8-7.1% respectively, equivalent to an AVT reduction of 0.34-0.37 and 0.67-0.98 percentage points.
Diaspora impacts on Chinese imports are in general higher than those found for exports. This result supports a strong trade facilitation mechanism in place particularly for Chinese imports. The Chinese diaspora can be expected to have a better idea of the Chinese bureaucracy apparatus, contract law enforcement issues, or even the importance of bribes and being able to trust key partners in China may prove to be very valuable, as uncertainty is reduced. However, this mechanism could be less relevant for Chinese exports since the bulk of Chinese exports goes to developed countries with sound institutions, and hence, the trade-cost channel is probably less important.
Relatively higher trade creation effects for imports (compared to exports) of the sending country stand in contrast to findings from earlier studies (Genc, Gheasi, Nijkamp, & Poot, 2012) who more often find the opposite to be true. A couple of explanations can be thought of. First, in contrast to the present study, earlier studies were limited by the use of first-generation migrant data. It is likely-and | 2427

MARTÍNEZ-ZARZOSO ANd RUdOLF
in line with anecdotal evidence-that first-generation migrants contribute relatively more to imports from their source country than to exports to the same, indicating that the preference channel plays a larger role for first-generation migrants. Moreover, as it takes time to build up a functioning business and to be able to contribute to exports of the host country, it could be expected that the trade-cost channel takes more time to be effective. If that was the case, migrants could be expected to contribute relatively more to their host countries' net bilateral exports with their country of origin as time passes. Second, overseas Chinese are often characterised by a comparatively strong emphasis on family economic success, including long working hours, and thrifty and dynamic family businesses (Bolt, 1996;Folk & Jomo, 2013;Gomez, Hsiao, & Xiao, 2003). Thus, for the Chinese diaspora the trade-cost channel might play a more important role than the preference channel compared to other diasporas. The fact that the diaspora trade creation effect is lower for Chinese exports does not support the existence of a strong preference effect in aggregate exports. Although we cannot separate the preference effect from the trade-cost channel, a strong preference effect should lead to a higher trade creation effect on Chinese exports in comparison with Chinese imports and we find the opposite.

ACKNOWLEDGMENTS
We would like to thank the editor and two anonymous referees for their useful suggestions. We are also grateful to the participants in the XVIII International Economics Conference hold in Huelva for their helpful comments. World Christian encyclopedia (Barrett et al., 1982(Barrett et al., , 2001 1973-2013 150 World Christian encyclopedia (Barrett et al., 1982(Barrett et al., , 2001