WPS6279 Policy Research Working Paper 6279 Networks, Firms, and Trade Paulo Bastos Joana Silva The World Bank Development Research Group Trade and Integration Team November 2012 Policy Research Working Paper 6279 Abstract Fixed costs associated with learning about demand They find that larger stocks of emigrants in a given and setting up distribution networks are expected to destination increase export participation and intensity. be lower when there are more potential contacts in the In addition, they show that the former of these effects destination market, suggesting a greater probability of tends to be more pronounced among firms that are more market entry and larger export revenues. The authors likely to have close ties with the emigrants. These results match historically-determined emigration stocks with are consistent with a multiple-destination version of the detailed firm-level data from Portugal to examine the Melitz (2003) model featuring market-specific entry costs effect of migrant networks on these export outcomes. and idiosyncratic firm-destination demand shocks. This paper is a product of the Trade and Integration Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at pbastos@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Networks, …rms, and trade Paulo Bastosy Joana Silvaz Keywords : Networks; gravity; export participation and intensity. JEL classi…cation : F1; F2; L1; L2 We thank two anonymous referees, Jonathan Eaton, Ana Margarida Fernandes, Caroline Freund, Bernard Hoekman, William Maloney, Caglar Ozden and participants at the Midwest International Eco- nomics Group Spring Meetings in Iowa, and the LACEA annual meetings in Buenos Aires for valuable comments and discussions. Joana Silva thanks Instituto Nacional de Estatística for providing access to the micro data employed in this paper.The views expressed in this paper are those of the authors and should not be attributed to the World Bank. y World Bank. z World Bank. 1 1 Introduction Recent research on …rm-level exporting suggests that the idiosyncratic interplay between …rms and markets plays an important role in shaping trade patterns. Examining detailed data on the exports of French …rms, Eaton et al. (2011) show that while a multiple- destination Melitz (2003)-type model does a good job at explaining several empirical reg- ularities, it fails to come to terms with some important features of the data: (1) …rms do not enter markets according to an exact hierarchy; and (2) their sales where they do enter deviate from the exact correlations that the basic model would insist upon.1 To account for these facts, they introduce market and …rm-speci…c heterogeneity in entry costs and demand. But while the modi…ed framework is more consistent with observed trade pat- terns, there is still little evidence on how speci…c potential sources of such heterogeneity in entry costs and demand a¤ect the export performance of individual producers. In this paper, we match historically-determined emigration stocks with detailed …rm- level data from Portugal to examine the e¤ect of migrant networks on export participation and intensity across markets and …rms. Portugal o¤ers unique features for this investiga- tion. First, the motives and timing of the sizable emigration ‡ows observed in the country during the Estado Novo authoritarian regime, along with their steep fall in the aftermath of the democratic revolution of 1974, mitigate the concern that emigration stocks might be endogenous to current trading relationships.2 Second, emigration ‡ows came predomi- nantly from the Northern region of Portugal, suggesting that …rms located there are more likely to have close ties with the emigrants. These historical features allow us to investigate not only whether migrant networks matter for export outcomes, but also if they constitute an important source of market and …rm-speci…c heterogeneity in entry costs and demand. While accounting for the e¤ect of …rm productivity heterogeneity and the customary gravity-type regressors, we …nd that larger stocks of emigrants in a given destination increase the likelihood of export participation. This e¤ect is signi…cantly more pronounced among …rms that are more likely to have close ties with the emigrants – that is, among producers currently located in North of Portugal; and among …rms that were already born when the bulk of emigration ‡ows took place. Conditional on a …rm serving a market, the presence of emigration stocks appears to be an important driver of how much it sells there: using two alternative methods to account for self-selection into export markets – a Tobit model and a Heckman selection model – we …nd that export revenues tend to increase systematically with the number of emigrants in the destination. Taken together, these …ndings suggest that migrant networks are an important source of market and …rm- 1 See also Lawless (2009), who provides related evidence for Ireland. 2 In particular, the circumstances in which emigration ‡ ows took place alleviate concerns of reverse causality. 2 speci…c heterogeneity in entry costs and demand. In addition to the work cited above, this paper is related to recent research using aggre- gated …rm-level data to provide some evidence on how colonial ties and linguistic similarity in‡uence trade ‡ows. Crozet and Koenig (2010) decompose industry-level bilateral exports in France into the number of exporters and the average exports per …rm, and employ each of these variables as the dependent variable in a gravity model. Their estimates suggest that a common language and colonial ties increase both the number of exporters and the average exports per …rm.3 Anderson (2007) uses data from Sweden to investigate the link between bilateral trade and familiarity (captured by dummy variables for Nordic, Baltic and English-speaking nations). He reports a positive relation, due primarily to the exten- sive margin.4 Our paper di¤ers from this line of research in several important respects, however. First, in addition to studying the role of colonial ties/common language, we provide evidence on whether and how migrant networks promote trade. Second, in do- ing so we use theory-grounded measures of export participation and intensity at the level of the individual producer (linked to information on …rms’ labor productivity and size). Third, and most importantly, we exploit the unique historical background underlying the generating process of our data to provide evidence that migrant networks constitute an important source market and …rm-speci…c heterogeneity in entry costs and demand. This paper is also related to recent work using country and product-level data to re- examine the microfoundations of the gravity equation in the context of heterogeneous-…rm models, including Helpman et al. (2008) and Baldwin and Harrigan (2011). Relative to this strand of work, a key distinguishing feature of our empirical analysis is the ability to link information on emigration stocks with detailed data on …rm-level export outcomes and other …rm characteristics. In doing so, we also contribute to a rich literature on the e¤ects of social networks on aggregate trade ‡ows (Gould, 1994; Rauch, 1999; Rauch, 2001; Rauch and Trindade, 2002; Melitz, 2008; Guiso et al, 2009). The remainder of the paper is organized as follows. Section 2 outlines a multi-country heterogeneous-…rm trade model, and uses it to discuss the speci…c channels whereby mi- grant networks may impact on …rms’bilateral exports. Section 3 provides historical back- ground and discusses the timing and drivers of Portuguese emigration ‡ows. Section 4 describes the data employed, before section 5 presents preliminary evidence based on country-level data. Section 6 examines the e¤ect of networks on export participation and intensity at the level of the …rm, emphasizing the role of …rm age and location in shaping 3 Importantly, Crozet and Koenig (2010) also use …rm-level data to estimate key parameters of the model of Chaney (2008), but in doing so restrict the focus to the e¤ect of geographic distance on trade. 4 An important caveat, however, is that the measure of familiarity employed by Anderson (2007) exhibits high colinearity with geographic distance. 3 the former of these relationships. Section 7 concludes. 2 Networks and trade with heterogeneous …rms To guide our empirical analysis, we present a multiple-destination variant of the Melitz (2003) model featuring market-speci…c …xed entry costs and idiosyncratic …rm-destination demand shocks. The version we use draws heavily on Chaney (2008), Lawless (2009), Eaton et al. (2011) and Crozet et al. (2011). Consider a world composed of multiple asymmetric nations. In each country c, con- sumer preferences for a di¤erentiated good are characterized by a Dixit-Stiglitz sub-utility function of the form Z 1 1 Uc = [ac (j )q (j )] dj (1) j c where q (j ) denotes quantity of variety j demanded in country c, c denotes the set of varieties j , and > 1 the constant elasticity of substitution between any pair of varieties. The ac (j ) are …rm-destination demand shocks, a feature that we borrow from Eaton et al. (2011) and Crozet et al. (2011). We assume that ac (j ) represent …rm j ’s network of connections with consumers in country c. Each …rm produces a single horizontally-di¤erentiated variety j by means of a Ricar- w dian technology with unit cost '(j ) , where w is the wage level common to all …rms and '(j ) is a …rm-speci…c productivity parameter, randomly drawn from a distribution with cumulative density function G('). To export to market c producers must incur both a …xed cost, Fc , and a variable cost, c. The latter is modeled as an "iceberg" transportation 1 cost, where if one unit of a good is shipped to country c only a fraction c arrives to the …nal destination. Fixed costs of exporting can stem from a wide variety of activities, for example establishing business contacts and distribution networks, undertaking marketing e¤orts, or overcoming bureaucratic barriers. The presence of information networks pro- vided by emigrants is likely to reduce several of these entry costs (Rauch and Trindade, 2002). For simplicity, we assume that Fc varies between markets, but not across …rms in the same market. Hence ac (j ) are the sole parameters capturing the presence of …rm- speci…c social networks in a given destination, while Fc captures networks e¤ects that are common to all …rms in a market. A valid interpretation is that, by incurring Fc , a …rm with more contacts in country c that is, a …rm with a larger ac (j ) is able to reach a larger number of potential buyers there. The sub-utility (1) is assumed to enter a Cobb-Douglas full utility, which implies that consumers of country c spend an exogenous total amount, Xc , on the di¤erentiated good. If …rm j serves market c, its market share in that country is given by: 4 xc (j ) (pc (j )=(ac (j ))1 =Z (2) Xc (pc (i)=(ac (i))1 di i c where xc (j ) and pc (j ) denote export revenues and prices, respectively, both of which are inclusive of trade costs. Firm j maximizes a destination-speci…c pro…t function given by: w c (j ) = (pc (j ) c )qc (j ) Fc (3) '(j ) As is standard in Dixit-Stiglitz monopolistic competition models, …rms obtain a con- stant mark-up =( 1) and charge the CIF price: w pc (j ) = c (4) 1 '(j ) Using (4) and (2), we derive export revenue from market c conditional on entry: 1 '(j ) 1 1 1 xc (j ) = ac (j ) Xc c Pc (5) w where Pc is the price index in country c. If …rm j serves country c, the net contribution of this destination to pro…ts is given by: c (j ) = xc (j )= Fc (6) Heterogeneity in …rm e¢ ciency and …rm-destination demand shocks imply that not all …rms will …nd it pro…table to export to country c. From (2) and (6), the probability of exporting is given by: 1 '(j ) 1 1 1 P[Ec (j ) = 1] = P[ ac (j ) Xc c Pc > Fc ] w which may equivalently be expressed as: '(j ) P[Ec (j ) = 1] = P[( 1) ln( )+( 1) ln ac (j )+ln Xc +( 1) ln Pc ( 1) ln c ln( Fc ) > 0] w (7) Using (5), we can write the FOB export revenue in market c as: xc (j ) '(j ) ln xfob c (j ) ln( )=( 1) ln( )+( 1) ln ac (j )+ln Xc +( 1) ln Pc ln c (8) c w Equations (7) and (8) deliver a number of familiar predictions about the e¤ect of …rm e¢ ciency and destination-market attributes on export participation and intensity. All else equal, more productive …rms are more likely to enter a given destination and to sell more there. For given productivity, export participation and intensity are expected to increase with the destination’s size and price level, and decrease with variable trade costs. 5 Our main interest lies on the e¤ect of migrant networks on export participation and intensity. From (7) we see that by reducing Fc networks are expected to increase the likelihood of export participation. We also see that, for a given level of e¢ ciency and …xed entry costs, …rms with stronger networks in a given destination – i.e. …rms with larger ac (j ) – are more likely to sell there. Finally, (8) reveals that export intensity is independent of Fc , but may nevertheless rise with migrant networks due to idiosyncratic …rm-destination demand shocks, ac (j ). 3 Historical background Several historical features of Portugal facilitate the analysis of the impact of social net- works on …rm-level export outcomes. We provide a brief summary of such a background, emphasizing the aspects that are more relevant for the empirical analysis. Current stocks of Portuguese emigrants are largely rooted in Estado Novo, the author- itarian regime that ruled the country between 1933 and 1974. The limits on individual rights and freedoms imposed by the regime, together with large income di¤erentials rela- tive to developed countries, created strong pressures for emigration. Yet, as emphasized s by Baganha (2003), for decades such pressures were …ercely controlled by the regime’ highly restrictive emigration policy. In the early 1960s, however, Estado Novo ’s attitudes towards emigration took a surprising turn. Following a shift in industrial policy favoring the development of more capital intensive industries in the Lisbon area, the regime was faced with an excess of unskilled labor supply in the Northern region and began to favor emigration of redundant unskilled workers (Baganha, 2003). By 1965, several obstacles to emigration had been removed and emigration was no longer a crime punishable by law. Largely in response to this policy shift, between 1960 and 1974 over 1.5 million, predomi- nantly low-skilled, Portuguese left the country in search of better economic and political conditions in more developed nations (Figure 1). Following the Carnation Revolution in 1974, which overthrew Estado Novo, emigration ‡ows fell sharply re‡ecting the combined e¤ect of the introduction of democracy and, initially, more restrictive policies towards immigration in several recipient countries. As a result, and also in light of limited return migration from more developed countries, stocks of emigrants have remained largely sta- ble over the past three decades.5 The motives and timing of Portuguese emigration ‡ows alleviate concerns of reverse causality in the relationship between migrant networks and export outcomes. In addition, the concentration of emigration ‡ows in time and space 5 Despite the limited emigration ‡ows observed in Portugal since 1975, the Global Migrant Origin Database (described below) recorded a total of 1.95 million Portuguese emigrants in the year 2000. This amounts to nearly 19% of the population living in Portugal, which has remained fairly stable in recent years. 6 allow us to investigate if migrant networks constitute an important source of market and …rm-speci…c heterogeneity in entry costs and demand. Portugal further provides an interesting setting for re-examining the e¤ect of common language/colonial ties on export participation and intensity. Portuguese-speaking nations belong to three di¤erent continents and di¤er markedly in terms of size. They are also at di¤erent stages of economic development. Brazil was the …rst country to become in- dependent in 1822. The other nations became independent in the wake of the Carnation Revolution of 1974. Table 1 reports the independence dates, o¢ cial languages, and num- ber of Portuguese migrants in each of these countries. Table A.2 in the Appendix displays the distribution of Portuguese migrants across the full set of export destinations. 4 Data Our empirical analysis draws on a rich …rm-destination data set that combines information from several sources: Export ‡ows. Our main data source is the Foreign Trade Statistics (FTS) of Portugal for 2005. This is the country’s o¢ cial information source on international trade statistics, gathering the shipments of virtually all exporting …rms to each destination market. The FTS data are collected in two di¤erent ways. Data on trade with countries outside the EU (external trade) are collected via the customs clearance system, which covers the universe of external trade transactions. Data on the transactions with other EU member states (internal trade) are obtained via the Intrastat system, where the information providers are companies engaged in internal trade and registered in the VAT system whose value of annual shipments exceeds a legally binding threshold (85,000 Euros in 2005). Export values in these data are "free on board", thus excluding any duties or shipping charges. The 2005 FTS data set comprises information on 16,541 exporting …rms and 220 destination markets. Despite the above-mentioned constraint, the export transactions included in these data aggregate to 97 percent of the total value of merchandise exports reported in the o¢ cial national accounts.6 Firm characteristics. Data on additional characteristics of exporting …rms, namely labor productivity (gross value added per worker), employment, age and geographical location, come from 2005 Enterprise Integrated Accounts System (EIAS). This is a census of …rms operating in Portugal run by the National Statistics Institute.7 These data are 6 In the FTS data, Serbia, Montenegro, and Serbia and Montenegro are considered as three di¤erent destinations. We treat them as such in the analysis, but none of our results is in‡uenced by this decision. 7 In both the FTS and the EIAS, …rms are uniquely identi…ed by their tax identi…cation number (NPC). Hence the mapping of the two data sets was straightforward. After merging the two datasets, the Statistics Institute applied a transformation of the NPC to the data that were made available to researchers in order 7 unavailable for a small subset of exporting …rms, which are therefore excluded from the …nal sample used in the empirical analysis. Importing countries. We have supplemented the …rm data with information on each importing country, namely its real GDP (measured at PPP), GDP per worker, and the distance between its most populated city and Lisbon (measured in Kms). The source of the information on distance is CEPII. The remaining variables come from the World Development Indicators (WDI) of the World Bank. Whenever WDI data were reported missing, we have used instead information from the CIA factbook. These data are available for 199 destination markets, including the Portuguese-speaking countries listed in Table 1. Portuguese emigrants. We have further added information on the number of Por- tuguese emigrants in each importing country. These data come from the Global Migrant Origin Database (GMOD) from the the University of Sussex’s Development Research Cen- tre on Migration, Globalization and Poverty. The stocks refer to the year 2000, and are available for 193 of the 199 countries mentioned above. The full GMOD data set consists of a 226x226 matrix of origin-destination stocks. The data are obtained by disaggregating the information on migrant stocks in each destination country as given in the 2000 round of its population census. The …nal data set used in the estimation gathers information on 14,782 exporting …rms and 193 destination markets. Table A1 provides summary statistics on these data.8 5 Networks and trade at the country level Table 2 reports descriptive statistics on the distribution of exporters across destinations. As can be seen, on average each …rm exported to 3.4 countries. However, the mean hides signi…cant …rm heterogeneity. More than one-half of all exporters sell to only one destination (54.2%), but they tend to be relatively small exporters, accounting for only 6.8% of the total export value. By contrast, only 7% of …rms export to more than 10 countries, but they account for 60.2% of the total export value. We proceed by presenting preliminary evidence based on aggregate measures of the extensive and intensive margins of bilateral exports. To do so we decompose the shipments of Portugal to each importing country into two di¤erent terms: P exports c exports c = …rms jc P (9) j …rms jc j to preserve con…dentiality. 8 The …nal data set used in this paper is an extended version of that described in Bastos and Silva (2010), supplemented with information on Portuguese emigrants and a wider set of …rm characteristics. 8 where the …rst term denotes the number of …rms exporting to country c, and the second the average exports per …rm. Figure 2 plots each of these components against the share of each trading partner in total exports. A clear pattern emerges from the data: the number of …rms exporting to Portuguese-speaking countries (identi…ed by a triangle) is relatively high when compared with the share of those countries in Portugal’s total exports. These simple descriptive statistics point, therefore, to an important role of the extensive margin in shaping the association between colonial ties/common language and exports. To investigate the relationship between migrant networks and exports, we estimate a gravity equation of the form: ln exports c = Xc + ln emigrants c + c (10) where the dependent variable is the log of the exports value to country c and Xc is a vector of standard gravity regressors, namely country c’s real GDP and GDP per worker, the geographic distance between Lisbon and the most populated city of the importing country, and binary variables indicating whether the country is a member of the European Union, uses the euro as its currency, and is landlocked.9 Our main interest lies in the coe¢ cients associated with emigrants c , the number of Portuguese emigrants in country c.10 We proceed by regressing the number of exporting …rms and the average exports per …rm against the same set of explanatory variables. Since OLS is a linear operator, these regressions additively decompose the margins whereby each regressor impacts on bilateral exports (Hummels and Klenow, 2005). Table 3 reports the results. Not surprisingly, the estimates reported in column (1) indicate that bilateral exports are signi…cantly higher in the presence of a common lan- guage and sizable emigrant stocks. Columns (2) and (3) report the relative contributions of the number of exporters and average exports per …rm. The point estimates suggest that the increase in the number of exporters accounts for most of the positive e¤ect of social networks on aggregate exports: in column (2) the coe¢ cients of interest are signi…cant at the 1% level and its magnitude is just slightly below the estimate presented in column (1), while in column (3) the coe¢ cient is much smaller and insigni…cant. 9 Bernard et al. (2007) adopt a similar speci…cation to analyse US bilateral exports, but do not focus on the role of social networks. 10 For 7 countries in our data, the number of Portuguese emigrants is zero. Since the log of zero is unde…ned, this issue was dealt with by transforming all zero values to 0.00001, then taking the log. We have veri…ed that all results in the paper prevail when excluding these countries from the estimation sample. 9 6 Networks and trade at the …rm level While their simplicity is attractive, the aggregated measures we employed in the previous section su¤er from important limitations for examining the e¤ect of migrant networks on export participation and intensity. In fact, a key insight of the model presented in section 2 is that the e¤ects on export participation of trade costs and demand factors depend on …rm attributes. We are unable to examine the role of …rm heterogeneity in determining export participation if we measure the extensive margin as the number of exporters. Another concern about this measure stems from the existence of a minimum statistical threshold for the collection of data on internal trade, which implies that the number of exporters is measured with error. On the other hand, the use of average shipments per exporter for measuring the e¤ects of migrant networks on export intensity is also problematic. In the context of a Melitz- type framework, Lawless (2010) shows that the e¤ect of variable trade costs on average sales per exporter is ambiguous. This is because a fall in variable costs increases the sales of every …rm serving a country, but may also lead to entry of lower-productivity …rms (and thus lower-sales …rms) in that market. A similar reasoning applies to the e¤ects of …rm-destination networks, ac (j ). In this section, we therefore exploit the …rm-destination detail of the data to estimate the e¤ect of migrant networks on export participation and intensity at the …rm-level. 6.1 Networks and export participation We begin by focusing on the e¤ects of migrant networks on export participation. Based on (7), we estimate a speci…cation of the form: P(E jc = 1) = P( ln 'j + Xc + ln emigrants c + jc ) (11) where: E jc = 1 if …rm j exports to country c and P is a probability distribution function; 'j is a proxy for …rm e¢ ciency; and jc is the …rm-destination error term. The country- level regressors have the meaning de…ned above. Table 4 reports the estimates yielded by a linear probability model (LPM). The esti- mated coe¢ cients give the marginal e¤ects of each regressor on the probability of a …rm exporting to a given market. In line with the model, the estimates in column (1) reveal that, conditional on the attributes of the export destination, more productive …rms are more likely to sell there. The coe¢ cients associated with the customary gravity variables also present the expected sign: …rms are more likely to export to larger and geographically closer nations, that are non-landlocked, and have the euro as the o¢ cial currency. 10 Of most interest to our analysis, the results in column (1) suggest that migrant net- works have a positive and signi…cant e¤ect on the probability of exporting to a given destination. Doubling the number of Portuguese emigrants in a potential destination in- creases the probability of export participation by about 0.14 of a percentage point. This e¤ect is certainly non-negligible, taking into account that the distribution of emigrant stocks across export destinations is very wide (see Table A2 in the Appendix). The esti- mates further suggest that the probability of exporting is 8.4 percentage points higher if Portuguese is an o¢ cial language of the importing country. While our measure of …rm labor productivity is closely linked to the concept of …rm e¢ ciency highlighted in the theory, it may not be the ideal proxy for e¢ ciency in the more complex real world where …rms employ multiple factors of production. To address this concern, we conduct a similar exercise using the log of …rm size (total employment) as a measure of …rm heterogeneity.11 In heterogeneous-…rm models with multiple factors of production, …rm size tends to be positively correlated with measures of total factor productivity (Verhoogen, 2008; Bartelsman et al. 2008; Hsieh and Klenow, 2009). Reas- suringly the estimates reported in column (2) show that the coe¢ cients of interest remain unchanged when this alternative proxy is used. For robustness, we re-estimate (11) using country-speci…c dummy variables for each of the seven Portuguese-speaking countries in our estimation sample. Columns (3) and (4) present the results. The magnitude of the coe¢ cients varies somewhat across countries, but the estimated coe¢ cients are positive and signi…cant for all countries. More importantly, the coe¢ cient capturing the e¤ects of migrant networks on export participation remains little changed when country-speci…c slopes for these destinations are allowed for. In columns (5) and (6), we use a linear probability model with …rm …xed-e¤ects (LPM- FE) to estimate (11). Rather than relying on a proxy to account for 'j , the LPM-FE esti- mator exploits solely the within-…rm variation in export participation across destinations, thereby accounting for both observed and unobserved …rm heterogeneity. As noted by Baldwin and Harrigan (2011) and Bustos (2011) the LPM-FE estimator is more appropri- ate than the …xed-e¤ects Probit and Logit estimators, as the latter are inconsistent when the number of e¤ects is large (incidental parameters problem). This is clearly the case here. In addition, they note that the LPM-FE is also preferable to random-e¤ects Logit models as the latter embody the (unsuitable) assumption that …rm-e¤ects are orthogonal to country characteristics. Reassuringly, the LMP-FE estimator yields very similar coef- …cients for the e¤ects of emigrant stocks and common language on export participation. As a further robustness check, Table 5 reports marginal e¤ects yielded from a number 11 As a robustness check, we have also used the log of …rm total sales and obtained similar results (available upon request). 11 of alternative estimators: Probit with proxies for …rm e¢ ciency, Probit with …rm random- e¤ects, and Conditional Logit with …rm …xed-e¤ects. Inspecting this table we see that, while the magnitude of the marginal e¤ect of interest varies somewhat across estimators, the key …ndings remains qualitatively unchanged. 6.1.1 Networks and …rm age The motives and timing of Portuguese emigration ‡ows provide an interesting setting for testing for the presence of …rm-destination networks. In particular, …rms that were already born when the historical bulge in Portuguese emigration ‡ows occurred would be expected to have a greater number of contacts among the emigrants, and hence be more responsive to their presence in potential export destinations. To investigate this hypothesis, we exploit information on …rm age and examine the extent to which the e¤ects of migrant networks on export participation vary across …rms from three di¤erent cohorts: born before 1976; in 1976-1990; and after 1990. Table 6 displays the results. Column (1) reports the coe¢ cients on the interaction between the regressors capturing the presence of social networks and binary variables for each cohort. Column (2) reports F-tests on the equality of these coe¢ cients using the 1976-1990 cohort as reference group. The estimates suggest that the e¤ects of migrant networks on export participation are signi…cantly larger for the older cohort. Notice also that a similar conclusion can be drawn from the coe¢ cient on common language/colonial ties, and recall that most Portuguese-speaking countries declared independence around 1975. An important concern about these results is that the di¤erential e¤ects of migrant networks across cohorts might be driven by confounding factors, notably the well-known association between …rm age and productivity (and size) (Cabral and Mata, 2003; Angelini and Generale, 2008). To verify the extent to which underlying di¤erences in …rm e¢ ciency might be driving our estimates, we compare the distribution of …rm labor productivity (FPD) across the three cohorts, using an empirical strategy analogous to Cabral and Mata (2003). Figure 3 shows that the FPD of …rms born after 1990 is more skewed to the right than that of their older counterparts, with the null hypothesis of equality of FPDs being rejected at the 5% level. This pattern is in line with Cabral and Mata (2003) who show that the distribution of …rm size of Portuguese …rms is very skewed to the right at the time of birth, and evolves gradually toward a log normal distribution. We also see, however, that the FPDs of the two older cohorts are visually and statistically indistinguishable from each other. This …nding is also in accordance with the observation by Cabral and Mata (2003) that …rm size distributions tend to evolve gradually to a more symmetric distribution, notably a log-normal distribution. Hence di¤erences in FPD appear unable 12 to explain the di¤erential e¤ects of migrant networks and common language/colonial ties on the export participation of …rms born in 1976-1990 relative to …rms born before 1976. As a further robustness check, we examine the di¤erential e¤ect of social networks across the three cohorts, for each tercile of the …rm productivity distribution.12 Reassur- ingly, the results reported in columns (3) to (8) of Table 7 show that in each productivity tercile the e¤ect of emigrant stocks on export participation is stronger among …rms born before 1976. Overall, the empirical results are therefore suggestive of the presence of …rm-destination social networks. 6.1.2 Networks and …rm location As noted in section 3, the bulk of emigration ‡ows came from the North of Portugal, where agriculture and traditional labor-intensive industries ceased to be among the priorities of s industrial policy, and the excess of labour supply that emerged from this the regime’ policy shift created strong pressures for emigration. This regional heterogeneity in …rm exposure to migrant networks allows us to provide further evidence on the importance of …rm-destination networks in shaping export participation. Column (1) in Table 7 reports the results for the full estimation sample. They suggest that the e¤ects of migrant networks on export participation are indeed more pronounced among …rms currently located in the North of the country; the coe¢ cient of interest is clearly larger, and this di¤erence is statistically signi…cant at the 5% level. Like in the previous sub-section, we worry that the di¤erential e¤ect of migrant net- works across regions might be driven by regional heterogeneity in …rm e¢ ciency. Com- paring the FPDs across the two groups of …rms, we see that the distribution of labor productivity of …rms located in the North of Portugal is relatively more skewed to the right, particularly in the middle of the distribution (Figure 4). For robustness, we there- fore verify if the regional heterogeneity of the e¤ect of networks on export participation is also present in each productivity tercile. The results are reported in columns (3) to (8) of Table 7 and suggest that this is indeed the case: in each productivity tercile, the e¤ect of migrant networks on export participation tends to be signi…cantly larger among Northern …rms.13 Interestingly, the estimates also indicate that the e¤ect of colonial ties/common language on export participation is signi…cantly less pronounced among Northern …rms. This may re‡ect the fact that Lisbon, the capital city and former center of the Portuguese 12 We use the log of …rm productivity relative to the industry to which the …rm belongs, where the latter is de…ned using data on the main product exported by the …rm (at the 5-digit level). This procedure is analogous to that adopted by Bustos (2011). 13 As a further robustness check, we have conducted a similar analysis excluding the …rms born before 1976. The results, not shown but available upon request, are qualitatively similar. 13 colonial empire, is located in the South of the country. Overall, therefore, our estimates suggest that the e¤ects of migrant networks on export participation tend to be more pronounced among the …rms located in the region where such networks are likely to be stronger, in line with the model presented in section 2. 6.2 Networks and export intensity We now turn to the e¤ect of migrant networks on the export intensity of each …rm, conditional on exporting to that market. From (8), we specify the revenue estimating- equation as: fob ln xjc = ln 'j + Xc + ln emigrants c + jc (12) fob is the revenue obtained by …rm j in country c, and the remaining regressors where xjc have the meaning de…ned above. The theory we adopt and the empirical analysis of the previous section suggest that …rms self-select into export markets. An important di¢ culty in estimating (12) is therefore that non-zero export revenues are only observed for the subset of destinations that …rm j serves. Due to sample selection, OLS estimation on non-zero export revenues may deliver biased estimates on the e¤ect of migrant networks on export intensity. We will employ two alternative methods to account for potential selection bias in the estimation of (12). First, we will adopt the Tobit procedure proposed by Crozet et al. (2011). In the context of a similar theoretical framework, Crozet et al. (2011) show that the minimum value of positive …rm-level exports observed in country c can be used fob in Tobit estimation. Replacing x (j ) with as a theory-grounded censoring point for xjc c fob (j ) xc fob (j ) > F = . in (6), we see that …rm j …nds it pro…table to serve country c if xc c c c fob (j ) that is consistent which For each destination, there is therefore a minimum value of xc non-negative pro…ts. Under the assumption of a log-normal distribution for ac (j ), this fob (j ) in country c serves as a yields a Tobit structure where the minimum value of xc theoretically-consistent censoring point. The main advantage of the Tobit estimator is that the correction procedure relies solely on the link between theory and observed trade ‡ow data. For robustness, we will also use a Heckman correction approach similar to Helpman et al. (2008). To avoid identi…cation based on functional form, the Heckman selection model requires identifying at least one variable that a¤ects export participation but not export intensity. From (7) and (8), we see that this requirement is ful…lled by a variable that in‡uences solely …xed costs of exporting. In line with Helpman et al. (2008), we will use measures of entry costs based on World Bank data that may plausibly satisfy this requirement. As a benchmark, we will also present the results of OLS estimates of (12) on non-zero export revenues. 14 Table 8 reports the OLS and Tobit estimates. The OLS estimates are solely based on non-zero export ‡ows and are reported in columns (1) to (3). The speci…cations in columns (1) and (2) use …rm labor productivity and employment, respectively, as proxy for …rm e¢ ciency, while that in column (3) uses a …rm …xed-e¤ects estimator to account for …rm heterogeneity. In all cases, the estimates point to a positive and statistically signi…cant relationship between export volumes and migrant networks. Columns (4) and (5) report the Tobit estimates. In this case, the estimation sample includes both zero and positive …rm-level bilateral trade ‡ows and the dependent variable fob to account for sample selection. The corre- is left-censored by the minimum value of ln xjc sponding marginal e¤ects con…rm the theoretical prediction that, conditional on selection into export markets, more e¢ cient …rms obtain larger revenues. Of most interest to the present analysis, the marginal e¤ects of migrant networks on export intensity are positive and statistically signi…cant at the 1% level, but are now larger in magnitude. Therefore, conditional on a …rm serving a market, the presence of emigration stocks appears to be an important determinant of how much it sells there. A potential concern with the Tobit estimation is that the censoring point might be measured with error. To assess if the estimates are sensitive to the proxy for the entry threshold, in columns (6) and (7) we report marginal e¤ects from Tobit estimation, but with the dependent variable left-censored at zero. Reassuringly, the estimates remain very similar. For robustness, Table 9 reports the results of a Heckman sample selection model. In line with Helpman et al. (2008), we use World Bank data on country-speci…c regulation costs of …rm entry to proxy for …xed costs of exporting. These indicators consist of the number of procedures and the number of days for an entrepreneur to legally start operating a business.14 Columns (1) to (3) report the marginal e¤ects of the Probit selection equation. The results in column (1) show that export participation is signi…cantly less likely in markets characterized by a larger number of procedures to form a business, as would be expected. Column (2) shows that the number of days required to form a business also enters negatively the selection equation, but this e¤ect is not statistically signi…cant. In column (3), we specify a binary indicator for both of these costs, which takes the value of one when the sum of the number of days and procedures is above the median. Once again, the e¤ect of this indicator on the probability of exporting is negative, although statistically insigni…cant. Columns (4) to (6) report the corresponding estimates of the Heckman selection model for the export revenue equation. The estimation can be conducted using full maximum 14 Data refer to 2005, come from the WDI of the World Bank, and are available for 159 countries of our sample. 15 likelihood or the two-step method proposed by Heckman (1979). We report results from the former method only, but have veri…ed that the latter yields very similar estimates. As before, the results point to a positive and statistically signi…cant relationship between export revenue and migrant networks. Results from additional estimations (not reported but available upon request) con…rm that these estimates remain very similar when log employment is used as a proxy of …rm e¢ ciency. Overall, thus, these results suggest that, conditional on a …rm serving a market, the presence of emigration stocks appears to be an important determinant of how much it sells there. This …nding is in line with the theory presented in section 2, which suggests that migrant networks are a source of …rm-destination demand shocks. 7 Concluding remarks We have used detailed …rm-level data from Portugal to examine the e¤ect of migrant networks on export participation and intensity. In doing so, we have exploited unique features from Portugal –notably stable, historically-determined stocks of emigrants with whom some …rms are more likely to have close ties. While accounting for the role of …rm productivity heterogeneity and the customary gravity-type regressors, we have found that larger stocks of Portuguese migrants in a given destination increase the likelihood of export participation. This e¤ect tends to be signi…cantly larger for …rms that are more likely to have a larger number of contacts among the emigrants –i.e., …rms currently located in the North of the country; and …rms that were already born when the bulk of emigration ‡ows took place. Conditional on a …rm serving a market, the presence of migrant networks appears to be an important determinant of how much it sells there. Taken together, these …ndings suggest that migrant networks are an important source of market and …rm-speci…c heterogeneity in entry costs and demand, thereby complementing and extending recent work by Eaton et al. (2011). By way of conclusion, we would like to note that although our empirical …ndings are consistent with a relatively parsimonious static model, a framework highlighting dynamic interactions between the formation of social networks, entry in export markets and …rm growth would likely provide important additional insights. We regard the development of such a framework as an interesting avenue for future research. References [1] Angelini, P. and Generale, A. (2008). "On the evolution of …rm size distributions." American Economic Review, 98(1): 426-438. 16 [2] Andersson, M. (2007). "Entry costs and adjustments on the extensive margin - an analysis of how familiarity breeds exports." Royal Institute of Technology of Stock- holm, CESIS Working Paper 81. [3] Baganha, M. (2003). "From closed to open doors: Portuguese emigration under the corporatist regime." e-Journal of Portuguese History, 1(1): 1-16. [4] Baldwin, R. and Harrigan, J. (2011). "Zeros, quality and space: Trade theory and trade evidence." American Economic Journal: Microeconomics, 3(2): 60-88. [5] Bartelsman, E., Haltiwanger, J. and Scarpetta, S. (2008). "Cross country di¤erences in productivity: The role of allocative e¢ ciency." University of Maryland, mimeo. [6] Bastos, P. and Silva, J. (2010). "The quality of a …rm’s exports: Where you export to matters." Journal of International Economics, 89(2): 99-111. [7] Bernard, A., Jensen, J.B., Redding, S. and Schott, P. (2007). "Firms in international trade." Journal of Economic Perspectives, 21(3): 105-130. [8] Bustos, P. (2011). "Trade liberalization, exports and technology upgrading: Evidence from the impact of MERCOSUR on Argentinean …rms." American Economic Review, 101 (1): 304-340. [9] Cabral, L. and Mata, J. (2003). "On the evolution of the …rm size distribution: Facts and theory." American Economic Review, 93(4): 1075-1090. [10] Chaney, T. (2008). "Distorted gravity: The intensive and extensive margins of inter- national trade." American Economic Review, 98(4): 1707-1721. [11] Crozet, M. and Koenig, P. (2010). "Structural gravity equation with extensive and intensive margins." Canadian Journal of Economics, 43(1): 41-62. [12] Crozet, M., Head, K. and Mayer, T. (2011). "Quality sorting and trade: Firm-level evidence from French wine." Review of Economic Studies, forthcoming. [13] Eaton, J., Kortum, S. and Kramarz, F. (2011). "An anatomy of international trade: Evidence from French …rms", Econometrica, 79(5), 1453-1498. [14] Guiso, L., Sapienza, P. and Zingales, L. (2009). "Cultural biases in economic ex- change?" Quarterly Journal of Economics, 124(3), 1095-1131. [15] Gould, D. (1994). "Immigrant links to the home country: Empirical implications for U.S. bilateral trade ‡ows." Review of Economics and Statistics, 76(2): 302-316. 17 [16] Heckman, J. (1979). "Sample selection bias as a speci…cation error.", Econometrica, 47(1), 153-161. [17] Helpman, E., Melitz, M. and Rubinstein, Y. (2008). "Estimating trade ‡ows: Trading partners and trading volumes." Quarterly Journal of Economics, 123(2): 441-487. [18] Hsieh, C. and Klenow, P. (2009). "Misallocation and manufacturing TFP in China and India." Quarterly Journal of Economics, 124, 1403-1448. [19] Hummels, D. and Klenow, P. (2005). "The variety and quality of a nation’s exports." American Economic Review, 95(3): 704-723. [20] Lawless, M. (2009). "Firm export dynamics and the geography of trade." Journal of International Economics, 77(2): 245-254. [21] Lawless, M. (2010). "Deconstructing gravity: Trade costs and the extensive and in- tensive margins." Canadian Journal of Economics, 43(4): 1149-1172. [22] Melitz, M. (2003). "The impact of trade on intra-industry reallocation and aggregate industry productivity." Econometrica, 71(6): 1695-1725. [23] Melitz, J. (2008). "Language and foreign trade." European Economic Review, 52(4): 667-699. [24] Rauch, J. (1999). "Networks versus markets in international trade." Journal of In- ternational Economics, 48(1): 7-35. [25] Rauch, J. (2001). "Business and social networks in international trade" Journal of Economic Literature, 39, 1177-1203. [26] Rauch, J. and Trindade, V. (2002). "Ethnic Chinese networks in international trade." Review of Economics and Statistics, 84(1): 116–130. [27] Verhoogen, E. (2008). "Trade, quality upgrading and wage inequality in the Mexican manufacturing sector." Quarterly Journal of Economics, 123, 489-530. 18 Figure 1: Portuguese emigration flows, 1930-2003 200,000 180,000 160,000 140,000 120,000 100,000 80,000 60,000 40,000 20,000 End of 2nd World War Carnation revolution 0 1930 1932 1934 1936 1938 1940 1942 1944 1946 1948 1950 1952 1954 1956 1958 1960 1962 1964 1966 1968 1970 1972 1974 1976 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 Legal emigrants Total emigrants Permanent migrants Notes: Data on legal emigrants and total emigrants until 1988 come from the Portuguese Ministry of Foreign Affairs. Data on permanent emigration from 1992 onwards come from the National Statistics Institute. Data on Portuguese emigration for the period 1989-1991 are not available. 1 Figure 2: Decomposing bilateral exports AGO ESP FRA 8 CPV GBR USADEU CHE ITA NLD BEL CAN SWE MOZ BRA DNK STP JPN NOR IRL GNB AUT GRC FIN MAR AUSCHN TUR POL ISR ZAF HKG 6 IND LUX CZE ARE TUN ROMMEX number of firms (ln) MAC HUN RUS NZL KOR TWN CYP SAU SGP EST LBN BGR SVK CHL DZA ZAR MLT ISL THA VEN ARG EGY MYS COG HRV LTU SVN COL PHL SENLVANGA UKR AND PAK IDN SYR KWT PERGIN PAN JOR DOM ATG BHR GHA QAT IRN 4 MRT ECU MUS VNM ANT CIV CRI URY TMPCMR CUB OMN GTM LKA GIB BMU SRB BEN SLV LBY KEN GAB GEO MDG TTO GMB MKD BGD NAM AZE PYF TZA MDA BOL BHS YUG NCL KAZBFABIH SDNSMR HND PRY LBR GNQ MLI ETH BLR UGA ZWE ALBYEM JAMTGO HTI SWZ FRO AFGKHM 2 MHL SUR NER TKM BDIIRQ TCD GUY AIA SYC LCA MWI ARM NICPRK UZB SLE BRBDJI ZMBSOM FJI ABW MNE KGZ BRNCOK MDV RWANPL CYM CAF VCT COMPNG BWA TJK COKTUV UMI MSRERI DMA MNP KNA TON LSO GRD SLB SPM WSM VGB 0 -20 -15 -10 -5 0 export share (ln) 15 TCD SGP DEUESP MNP SMR FRA GBR BWA ITA BEL NLD USA KWTMYS SDNGIB TUR VGB FRO NGA CHL DZA POL HUN FIN SWE average exports per firm (ln) MDA IRN CHN AUT LBY SVN GEO ARG EGYSAU MEX RUS HKG CZE DNK IRL SVK GRC MAR BRA BIH JOR LVA UKR ROM AUS ARE ZAF TUN AGO ERI PRK IRQHND BLR SLV LTU VNM CIV THA ISR SEN JPN CAN NOR CHE COK BDI GNQ BFA KEN SYR GIN ANT BGR VEN LUX TWN KOR SPM NICYEM ETHGABCUB LKA PAN QAT PAK IDN CYP SOM MLI GTM GHA CMR PHL MLTINDMOZ HRV ISL CPV MWI NER KAZ ZWE NAM TZA PER LBN WSM MSR KHM GRD SWZ ALB TGO HTI AZE NCLSRB AND ATG ECU CRI COL EST NZL BRN PNG DJI ARM ZMB TKM GMB MKD PYFTMP BEN BHR DOM MUS MAC GNB SLE BGD PRY URY MRT COG STP 10 SYC AIAJAM TTO BHSOMN ZAR DMA UZB MHL LBR UGA BMU COM MNEAFG BRB SUR LCA BOL YUG LSO UMI ABW KGZ VCT CAF TON CYM SLB MDG NPL GUY RWA FJI TUV COK KNA MDV 5 TJK -20 -15 -10 -5 0 export share (ln) Notes: Portuguese-speaking countries are identified by a triangle. 2 Figure 3: Distribution of log firm labor productivity by age .4 (a) born before 1975 (b) born 1976-1990 (c) born 1991-2005 .3 Density .2 .1 Kolmogorov-Smirnov tests for equality of distributions: (a)=(b): p-value 0.69 (b)=(c): p-value 0.00 0 0 5 10 15 20 ln labor productivity Notes: The curves are based on a normal kernel density smoother with a bandwidth of 0.7. 3 Figure 4: Distribution of log firm labor productivity by region .4 (a) north (b) other .3 Density .2 .1 Kolmogorov-Smirnov tests for equality of distributions: (a)=(b): p-value 0.00 0 0 5 10 15 20 ln labor productivity Notes: The curves are based on a normal kernel density smoother with a bandwidth of 0.7. 4 Table 1: Portuguese-speaking countries Independence Portuguese Country Continent Official languages from Portugal migrants Angola Africa Portuguese 1975 1555 Brazil South America Portuguese 1822 170210 Cape Verde Africa Portuguese 1975 1656 East Timor Asia Tetum, Portuguese 1975 n.a. Guinea Bissau Africa Portuguese 1973 766 Macau-China Asia Chinese, Portuguese 1999 445 Mozambique Africa Portuguese 1975 55520 Sao Tome and Principe Africa Portuguese 1975 814 Table 2: Export destinations per firm Number of destinations served % of firms % of revenue 1 54.2 6.8 2 15.0 6.2 3 7.7 4.7 4 5.0 3.2 5 3.3 3.7 6 2.4 3.9 7 1.8 2.5 8 1.5 3.6 9 1.1 2.6 10 1.0 2.6 More than 10 7.0 60.2 Average number of destinations per firm 3.4 Maximum number of destinations per firm 84 1 Table 3: Networks and the extensive and intensive margins of exports ln exports ln number ln exports Dependent variable: of firms per firm (1) (2) (3) ln GDP 0.7185*** 0.4661*** 0.2524*** (0.0666) (0.0366) (0.0479) ln GDP per capita 0.2369* 0.1932** 0.0437 (0.1344) (0.0765) (0.0792) ln distance -1.5067*** -0.8026*** -0.7042*** (0.1910) (0.1204) (0.1441) EU member 1.2831*** 1.0289*** 0.2542 (0.3033) (0.1983) (0.2281) common currency 0.0954 -0.0575 0.1529 (0.2883) (0.2215) (0.2228) landlocked -1.4478*** -0.9793*** -0.4685* (0.3497) (0.1718) (0.2726) common language 3.7571*** 3.6676*** 0.0895 (0.6136) (0.5326) (0.1996) ln emigrants 0.0736** 0.0609** 0.0127 (0.0308) (0.0235) (0.0213) Observations 193 193 193 R2 0.76 0.82 0.41 Notes: The estimation method is OLS. Robust standard errors are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. 2 Table 4: Networks and export participation Dependent variable: export participation (1) (2) (3) (4) (5) (6) ln firm labor productivity 0.0031*** 0.0031*** (0.0003) (0.0003) ln firm employment 0.0083*** 0.0083*** (0.0012) (0.0012) ln GDP 0.0038*** 0.0038*** 0.0038*** 0.0036*** 0.0036*** 0.0036*** (0.0009) (0.0009) (0.0008) (0.0009) (0.0008) (0.0008) ln GDP per capita 0.0016 0.0016 0.0021* 0.0016 0.0021* 0.0021* (0.0013) (0.0013) (0.0012) (0.0013) (0.0012) (0.0012) ln distance -0.0117** -0.0117** -0.0117** -0.0106** -0.0106** -0.0106** (0.0047) (0.0047) (0.0046) (0.0047) (0.0046) (0.0046) EU member 0.0091 0.0091 0.0099 0.00990.0091 0.0099 (0.0096) (0.0096) (0.0096) (0.0096) (0.0096) (0.0096) common currency 0.0571** 0.0571** 0.0571** 0.0567** 0.0567** 0.0567** (0.0227) (0.0227) (0.0228) (0.0228) (0.0228) (0.0229) landlocked -0.0095** -0.0095** -0.0095* -0.0090* -0.0090* -0.0090* (0.0048) (0.0048) (0.0048) (0.0048) (0.0048) (0.0049) common language 0.0835** 0.0835** 0.0835** (0.0348) (0.0348) (0.0349) ln emigrants 0.0014** 0.0014** 0.0016*** 0.0016*** 0.0014** 0.0016*** (0.0006) (0.0006) (0.0006) (0.0006) (0.0006) (0.0006) Angola 0.2850*** 0.2850*** 0.2850*** (0.0026) (0.0026) (0.0026) Brazil 0.0189*** 0.0189*** 0.0189*** (0.0063) (0.0063) (0.0063) Cape Verde 0.1328*** 0.1328*** 0.1328*** (0.0043) (0.0043) (0.0043) Guinea Bissau 0.0334*** 0.0334*** 0.0334*** (0.0042) (0.0042) (0.0042) Macau 0.0100*** 0.0100*** 0.0100*** (0.0034) (0.0034) (0.0034) Mozambique 0.0511*** 0.0511*** 0.0511*** (0.0039) (0.0039) (0.0039) Sao Tome and Principe 0.0512*** 0.0512*** 0.0512*** (0.0034) (0.0034) (0.0034) Firm fixed-effects No No No No Yes Yes Observations 2852926 2852926 2852926 2852926 2852926 2852926 Firms 14782 14782 14782 14782 14782 14782 Destinations 193 193 193 193 193 193 R2 0.06 0.07 0.08 0.08 0.11 0.12 Notes: The estimation method is OLS. Robust standard errors clustered by importing country are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. 3 Table 5: Networks and export participation, alternative estimators Dependent variable: export participation (1) (2) (3) (4) ln firm labor productivity 0.0014*** (0.0002) ln firm employment 0.0022*** (0.0003) ln GDP 0.0020*** 0.0015*** 0.0007*** 0.0203*** (0.0003) (0.0003) (0.0000) (0.0023) ln GDP per capita 0.0011** 0.0008** 0.0004*** 0.0124*** (0.0004) (0.0003) (0.0000) (0.0015) ln distance -0.0049*** -0.0037*** -0.0020*** -0.0566*** (0.0010) (0.0008) (0.0000) (0.0074) EU member 0.0036** 0.0025* 0.0011*** 0.0274*** (0.0018) (0.0014) (0.0000) (0.0035) common currency 0.0009 0.0007 0.0004 0.0066*** (0.0011) (0.0009) (0.0000) (0.0013) landlocked -0.0024* -0.0018* -0.0009*** -0.0353*** (0.0014) (0.0010) (0.0000) (0.0048) common language 0.0853* 0.0793* 0.0699*** 0.0682*** (0.0464) (0.0462) (0.0016) (0.0091) ln emigrants 0.0006*** 0.0005*** 0.0002*** 0.0067*** (0.0002) (0.0001) (0.0000) (0.0009) Estimator Probit Probit RE Probit Cond. Logit Firm effects No No Random Fixed Observations 2852926 2852926 2852926 2852926 Firms 14782 14782 14782 14782 Destinations 193 193 193 193 Notes: Marginal effects reported. In columns (1) and (2), robust standard errors clustered by importing country are in parentheses. In columns (3) and (4), robust standard errors are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. 4 Table 6: Networks, firm age and export participation Dependent variable: export participation full sample low productivity medium productivity high productivity (1) (2) (3) (4) (5) (6) (7) (8) F-stat F-stat F-stat F-stat Coeff. Coeff. Coeff. Coeff. (P-value) (P-value) (P-value) (P-value) common language * born before 1976 0.1129*** 51.42 0.0893*** 17.68 0.1231*** 44.46 0.1361*** 26.87 (0.0376) (0.00) (0.0328) (0.00) (0.0406) (0.00) (0.0415) (0.00) common language * born in 1976-1990 0.0841** 0.0772** 0.0807** 0.0981** (0.0357) (0.0328) (0.0352) (0.0405) common language * born after 1990 0.0738** 8.60 0.0744** 0.50 0.0735** 3.65 0.0734** 8.69 (0.0337) (0.00) (0.0340) (0.48) (0.0347) (0.06) (0.0323) (0.00) ln emigrants * born before 1976 0.0023*** 8.29 0.0018** 6.29 0.0024** 6.79 0.0031*** 9.41 (0.0009) (0.00) (0.0007) (0.01) (0.0010) (0.01) (0.0011) (0.00) ln emigrants * born in 1976-1990 0.0016** 0.0013** 0.0017** 0.0018** (0.0007) (0.0006) (0.0007) (0.0007) ln emigrants * born after 1990 0.0011** 6.13 0.0008** 5.77 0.0011** 5.93 0.0014** 4.94 (0.0005) (0.01) (0.0004) (0.02) (0.0005) (0.02) (0.0006) (0.02) Firm fixed-effects Yes Yes Yes Yes Observations 2852926 1107627 929874 815425 Firms 14782 5739 4818 4225 Destinations 193 193 193 193 R2 0.11 0.09 0.11 0.13 Notes: The estimation method is OLS. Columns (1), (3), (5) and (7) report estimated coefficients from regressions that additionally include interactions between the other gravity variables and the age-cohort dummies. Robust standard errors clustered by importing country are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. Columns (2), (4), (6) and (8) report F-tests of equality of coefficients across the three age-cohorts, where the reference group is the 1976-1990 cohort. 5 Table 7: Networks, firm location and export participation Dependent variable: export participation all firms low productivity medium productivity high productivity (1) (2) (3) (4) (5) (6) (7) (8) Coeff. F-stat. Coeff. F-stat. Coeff. F-stat. Coeff. F-stat. (P-value) (P-value) (P-value) (P-value) common language * North 0.0488** 0.0463* 0.0487** 0.0524** (0.0238) 8.49 (0.0236) 8.62 (0.0247) 9.22 (0.0231) 7.26 common language * Other 0.1083** (0.00) 0.1008** (0.00) 0.1112** (0.00) 0.1150** (0.01) (0.0431) (0.0408) (0.0443) (0.0450) ln emigrants * North 0.0020** 0.0015** 0.0021** 0.0025*** (0.0008) 5.56 (0.0006) 5.55 (0.0009) 4.83 (0.0010) 6.34 ln emigrants * Other 0.0010** (0.02) 0.0008* (0.02) 0.0011** (0.03) 0.0013** (0.01) (0.0005) (0.0004) (0.0005) (0.0005) Firm fixed-effects Yes Yes Yes Yes Observations 2852926 1107627 929874 815425 Firms 14782 5739 4818 4225 Destinations 193 193 193 193 R2 0.11 0.09 0.12 0.13 Notes: The estimation method is OLS. Columns (1), (3), (5) and (7) report estimated coefficients from regressions that additionally include interactions between the other gravity variables and region dummies. Robust standard errors clustered by importing country are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. Columns (2), (4), (6) and (8) report F-tests of equality of coefficients across regions. 6 Table 8: Networks and export intensity, OLS and Tobit Dependent variable: ln export revenue (1) (2) (3) (4) (5) (6) (7) ln firm labor productivity 0.3808*** 1.9873*** 1.9951*** (0.0387) (0.2012) (0.2009) ln firm employment 0.3310*** 4.1915*** 4.1879*** (0.0231) (0.3596) (0.3573) ln GDP 0.1310*** 0.1414*** 0.1914*** 2.8406*** 2.7758*** 2.8132*** 2.7494*** (0.0347) (0.0377) (0.0459) (0.5817) (0.5958) (0.5733) (0.5875) ln GDP per capita -0.0727* -0.0745* 0.0046 1.5500*** 1.5132*** 1.5341*** 1.4977*** (0.0381) (0.0405) (0.0395) (0.5922) (0.5851) (0.5869) (0.5801) ln distance -0.5246*** -0.5943*** -0.6477*** -6.9170*** -6.9301*** -6.8611*** -6.8748*** (0.0871) (0.0996) (0.0891) (0.9276) (0.9195) (0.9244) (0.9167) EU member 0.9288*** 0.769*** 0.7166*** 4.3223*** 3.9470** 4.3025*** 3.9276** (0.1379) (0.1459) (0.1445) (1.6598) (1.6444) (1.6505) (1.6357) common currency 0.0896 0.1278 0.0171 1.1312 1.1824 1.1789 1.2301 (0.1364) (0.1401) (0.1351) (1.2395) (1.2415) (1.2363) (1.2389) landlocked -0.8273*** -0.8543*** -0.7335*** -4.0660 -4.0295 -4.0425 -4.0078 (0.1124) (0.1171) (0.1200) (2.5010) (2.5175) (2.4721) (2.4891) common language 0.0119 0.3332 0.9917** 25.4061*** 25.6086*** 25.2408*** 25.4474*** (0.2204) (0.256) (0.3998) (5.5898) (5.5996) (5.5485) (5.5589) ln emigrants 0.0454** 0.0517*** 0.0802*** 0.8546*** 0.8464*** 0.8512*** 0.8429*** (0.0178) (0.0188) (0.0205) (0.2456) (0.2499) (0.2436) (0.2478) Estimator OLS OLS OLS Tobit Tobit Tobit Tobit Firm fixed-effects No No Yes No No No No Observations 52053 52053 52053 2852926 2852926 2852926 2852926 Firms 14782 14782 14782 14782 14782 14782 14782 R2 0.18 0.2 0.51 0.15 0.18 0.15 0.18 Destinations 193 193 193 193 193 193 193 Notes: In columns (1) to (3), the estimation method is OLS and only non-zero export flows are used in the estimation. In columns (3) and (4) the estimation method is Tobit and the dependent variable is left-censored at the minimum firm-level export value observed in each destination. In columns (6) and (7) the estimation method is Tobit and the dependent variable is left-censored at zero. Marginal effects reported. Robust standard errors clustered by importing country are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. 7 Table 9: Networks and export intensity, Heckman selection model Dependent variable: export participation ln export revenue (1) (2) (3) (4) (5) (6) ln firm labor productivity 0.0016*** 0.0016*** 0.0016*** 0.4025*** 0.4029*** 0.4241*** (0.0002) (0.0002) (0.0002) (0.0553) (0.0558) (0.0294) ln GDP 0.0023*** 0.0023*** 0.0019*** 0.1590*** 0.1594*** 0.1847*** (0.0006) (0.0006) (0.0006) (0.0579) (0.0584) (0.0511) ln GDP per capita 0.0026*** 0.0025*** 0.0037*** -0.0375 -0.0358 0.0143 (0.0008) (0.0008) (0.0006) (0.1238) (0.1240) (0.0916) ln distance -0.0062*** -0.0062*** -0.0058*** -0.6126*** -0.6132*** -0.6902*** (0.0013) (0.0013) (0.0013) (0.1502) (0.1510) (0.0838) EU member 0.0004 0.0006 0.0016 1.0212*** 1.0219*** 1.0445*** (0.0013) (0.0016) (0.0015) (0.1390) (0.1388) (0.1470) common currency 0.0025 0.0025 0.0014 0.0643 0.0645 0.0821 (0.0017) (0.0016) (0.0013) (0.1270) (0.1276) (0.1367) landlocked -0.0017 -0.0016 -0.0023 -0.8521*** -0.8538*** -0.8807*** (0.0015) (0.0015) (0.0016) (0.1177) (0.1172) (0.1007) common language 0.1508* 0.1547** 0.1243* 0.3722 0.3798 0.6990 (0.0779) (0.0751) (0.0717) (0.5530) (0.5581) (0.4425) ln emigrants 0.0005** 0.0005** 0.0007*** 0.0645** 0.0647** 0.0736*** (0.0002) (0.0002) (0.0002) (0.0257) (0.0255) (0.0227) ln number of procedures -0.0042** -0.0038** (0.0018) (0.0017) ln number of days -0.0004 (0.0009) reg. costs (proc. & days) -0.0005 (0.0014) Observations 2350338 2350338 2350338 2350338 2350338 2350338 Firms 14782 14782 14782 14782 14782 14782 Destinations 159 159 159 159 159 159 R2 0.25 0.25 0.25 Log likelihood -297017.8 -297005 -297929.5 Notes: Columns (1) to (3) report results from the Probit selection equation. Marginal effects and pseudo R2 reported. Columns (4) to (6) report the corresponding maximum likelihood estimates from a Heckman selection model. Robust standard errors clustered by importing country are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. 8 Table A1: Summary statistics, country and firm-level covariates Obs. Mean SD Country-level covariates ln GDP 193 24.00 2.49 ln GDP per capita 193 9.63 9.63 log distance 193 8.55 0.69 EU member 193 0.12 0.12 common currency 193 0.06 0.24 landlocked 193 0.19 0.39 common language 193 0.04 0.19 ln emigrants 193 4.75 3.95 Obs. Mean SD Firm-level covariates ln firm labor productivity 14782 9.92 0.91 ln firm employment 14782 2.68 1.46 firm age 14782 18.49 36.11 % of total Regional distribution North 41.62 Center 21.06 Lisbon 31.72 Alentejo 3.56 Algarve 1.18 Azores 0.35 Madeira 0.5 9 Table A2: Portuguese emigrants, 2000 France 619,847 Philippines 1,753 Macao, China 445 Germany 234,840 Bermuda 1,750 Chile 413 United States 212,318 Cape Verde 1,656 Indonesia 404 Brazil 170,210 Angola 1,555 Morocco 399 Canada 155,984 Austria 1,473 Tanzania 382 Switzerland 104,159 Ghana 1,427 Kyrgyz Republic 373 Spain 56,359 Nigeria 1,338 Moldova 373 Mozambique 55,520 Cote d'Ivoire 1,262 Suriname 371 Venezuela, RB 54,414 Malaysia 1,251 Japan 368 Luxembourg 41,722 Guinea 1,220 Ethiopia 322 United Kingdom 37,910 Hong Kong, China 1,170 Croatia 322 Belgium 21,371 Netherlands Antilles 1,005 Cameroon 315 Pakistan 21,302 Kenya 887 San Marino 314 Zimbabwe 19,729 Sao Tome and Principe 814 Greece 302 Australia 15,441 Congo, Rep. 776 Yemen, Rep. 298 Kuwait 10,411 Norway 769 Costa Rica 297 Netherlands 10,218 Cuba 768 Mexico 270 Andorra 8,873 Guinea-Bissau 766 Turkmenistan 235 South Africa 8,037 Ecuador 759 Turkey 225 Russian Federation 6,451 Lebanon 747 Singapore 213 Italy 5,901 Algeria 713 Bahrain 212 Argentina 5,840 Denmark 686 Syrian Arab Republic 196 Uzbekistan 5,059 Uruguay 680 India 187 Jordan 4,806 China 652 Iraq 178 Israel 3,986 Ireland 601 Estonia 176 Ukraine 3,656 Colombia 589 Togo 170 Nepal 2,876 Zambia 571 Aruba 164 Sweden 2,533 Namibia 566 Faeroe Islands 157 Malawi 2,446 Romania 508 Burundi 155 Libya 1,945 New Caledonia 506 Belarus 149 Burkina Faso 1,846 Taiwan 503 New Zealand 148 United Arab Emirates 1,841 Tajikistan 450 Thailand 143 10 Table A2: Portuguese emigrants, 2000 (continued) Finland 141 Lesotho 45 St. Vincent and the Grenadines 12 Panama 137 Somalia 43 Bosnia and Herzegovina 11 Kazakhstan 127 Cambodia 42 Jamaica 11 Seychelles 123 Paraguay 42 Korea, Rep. 10 Niger 122 Rwanda 40 British Virgin Islands 8 Iceland 116 Czech Republic 39 Bahamas, The 7 Madagascar 115 Grenada 39 Dominica 7 Swaziland 113 Bangladesh 32 Anguilla 6 Georgia 102 Gabon 31 Albania 6 Gambia, The 99 Benin 29 Maldives 6 Uganda 99 Bolivia 28 Solomon Islands 6 Senegal 98 Hungary 28 Northern Mariana Islands 5 Bulgaria 96 Eritrea 27 Botswana 4 Oman 92 Tonga 22 Latvia 4 Mali 91 St. Kitts and Nevis 21 Macedonia, FYR 4 Haiti 86 Korea, Dem. Rep. 21 Slovak Republic 4 Armenia 83 El Salvador 20 Samoa 4 Trinidad and Tobago 83 Tunisia 20 Antigua and Barbuda 3 Gibraltar 81 Comoros 19 Lithuania 3 Dominican Republic 75 Cyprus 19 Mauritania 3 Afghanistan 71 Mauritius 18 Cayman Islands 2 Poland 70 Brunei Darussalam 17 Equatorial Guinea 1 Iran, Islamic Rep. 67 Guyana 17 St. Lucia 1 Liberia 67 Sri Lanka 16 Monserat 1 Malta 63 Marshall Islands 16 St Hellen 1 French Polynesia 63 Barbados 15 Azerbaijan 0 Peru 57 Cook Islands 15 Central African Republic 0 Egypt, Arab Rep. 55 Fiji 15 Federal States of Micronesia 0 Djibouti 53 Guatemala 15 Qatar 0 Chad 52 Sierra Leone 14 Sudan 0 Saudi Arabia 51 Honduras 12 Slovenia 0 Papua New Guinea 48 Nicaragua 12 Tuvalu 0 Vietnam 48 Note: This table reports the number of Portuguese emigrants in 193 countries. The data come from the Global Migrant Origin Database and refer to the year 2000. 11