WPS5657 Policy Research Working Paper 5657 Success and Failure of African Exporters Olivier Cadot Leonardo Iacovone Denisse Pierola Ferdinand Rauch The World Bank Africa Region Finance and Private Sector Development Department & Poverty Reduction and Economic Management Network International Trade Department May 2011 Policy Research Working Paper 5657 Abstract Using a novel dataset with transactions level exports data same product to the same destination from the same from four African countries (Malawi, Mali, Senegal and country, pointing towards the existence of cross-firm Tanzania), this paper uncovers evidence of a high degree synergies. Accordingly the evidence is consistent with of experimentation at the extensive margin associated the hypothesis that those synergies may be driven by with low survival rates, consistent with high and middle information spillovers. More intuitively and consistently income country evidence. Consequently, the authors with multi-product firms models, the analysis also finds focus on the questions of what determines success and that firms more diversified in terms of products, but even survival beyond the first year and find that survival more in terms of markets, are more likely to be successful probability rises with the number of firms exporting the and survive beyond the first year. This paper is a joint product of the Finance and Private Sector Development Department, Africa Region; and the International Trade Department, Poverty Reduction and Economic Management Network. 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 authors may be contacted at Liacovone@worldbank.org, f.g.rauch@lse.ac.uk, ocadot@worldbank.org, mpierola@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 Success and failure of African exporters∗ Olivier Cadot, †Leonardo Iacovone, ‡Denisse Pierola, §Ferdinand Rauch¶ May 10, 2011 Keywords: Africa, Export survival, Externalities, Firms. JEL classiï¬?cation: F10, F14, O55. ∗ This paper is part of a World Bank research project initiated by Paul Brenton and Denisse Pierola. Financial support from the BNPP Trust Fund on Improving the Survival of African Exports and from the World Bank Research Support Budget is gratefully ac- e e knowledged. We are also grateful to the Malawi Revenue Authority, the Direction G´n´rale e e e e des Douanes du Mali, the Direction G´n´rale des Douanes du S´n´gal, and the Tanzania Revenue Authority for their cooperation and willingness to share data. We thank Frances Aidoo, William Baah-Boateng, Sidiki Guindo, Anthony Mveyange and Nelson Nsiku for their assistance in the collection of the customs data used in this paper. † University of Lausanne and Trade Division, World Bank, 1818 H Street, NW; Wash- ington, DC, 20433, USA, CEPR and CEPREMAP. Email:ocadot@worldbank.org. ‡ Private and Financial Sector Department, Africa Region, World Bank, 1818 H Street, NW; Washington, DC, 20433, USA. Email: Liacovone@worldbank.org (Corresponding author). § Development Economics Research Group - Trade team, World Bank, 1818 H Street, NW; Washington, DC, 20433, USA. Email: mpierola@worldbank.org. ¶ CEP, London School of Economics, Houghton Street, WC2A 2AE, London, UK. Email: f.g.rauch@lse.ac.uk 1 Introduction In spite of great strides since the late 1990s, low-income country exports are still marginal in world trade and suffer from various vulnerabilities, ranging from low unit values to volatility, concentration and low survival (see for example Fugazza and Molina 2009). Low survival is not necessarily a sign of welfare loss if it reflects strong experimentation at the extensive margin, but it can be inefficient if sunk costs of entry and exit are substantial, as suggested by the work of Das, Robert and Tybout (2007). In particular, the weak contribution of the extensive margin to overall export growth (Besedes and Prusa 2007) may be explained by high failure rates, the flip side of low survival. Thus, identifying drivers of the ‘sustainability margin’ of exports is important for our understanding of the constraints to low-income country export growth. This is what the present paper sets up to do, using a new transaction-level export dataset obtained from Customs authorities in four African countries. The key contributions of this paper are twofold. First, we document, for a sample of four low income African countries, a set of stylized facts on export survival that is broadly consistent with the emerging ï¬?rm-level literature, so far conï¬?ned to OECD or middle-income countries. Second, we provide novel evidence on a key outcome of interest, namely survival beyond the ï¬?rst year after entry into export markets. In particular, we identify a cross-ï¬?rm “syn- ergy effectâ€? suggestive of external economies in export survival, a potential driver of sustainability that has been hitherto overlooked in the literature. Increasing the number of exporters of similar products to the same destina- tion exerts a positive externality on new entrants. That is, the more similar they are, the higher the survival probability of new entrants. When evaluat- ing the mechanisms behind these results we ï¬?nd evidence consistent with the existence of information spillovers driving them. For a Senegalese exporter, for instance, the probability of surviving past the ï¬?rst year (22% in 2001) would rise to 26% if the number of national competitors selling the same product (product identiï¬?ed at HS 6 digits) on the same destination market were to double from the baseline 22 to 44. Our results may help explain a ï¬?nding highlighted in Easterly, Resheff and Schwenkenberg (2009), namely, that national export success often takes the form of ‘big hits’, with one narrow export item suddenly growing rapidly. 2 If a sufficient number of exporters target one market simultaneously, our results imply that their chances of surviving increase, possibly triggering a virtuous cycle of entry, survival and growth. Like Eaton et al. (2008), we also ï¬?nd that export spells that survive tend to grow. For instance, in Senegal, products that entered a market in 2001 and survived till 2008 had reached, by then, four times their entry volume. Lastly, export scale and scope at the ï¬?rm level, by which we mean respec- tively destinations per product and products per destination, both evaluated at the time the ï¬?rm launches a new product-destination combination, matter for its survival. From a policy perspective, our ï¬?ndings could be construed as contribut- ing to a possible rationale for using public funds to promote national exports abroad. The synergy we identify is akin to external economies, as the pres- ence of competitors from the same countries exporting the same product to the same destination provides potential new entrants with information on the proï¬?tability of these exports ventures, help identifying potential buyers as well as provide information about the consumers’ preferences and there- fore increases their likelihood of surviving. Similarly, this information be- comes available to ï¬?nancial institutions and ease ï¬?nancial constraints of new exporters as shown by the fact that these ‘synergy effects’ are especially important for ï¬?rms in sectors that are more dependent from the ï¬?nancial sector. However, these external economies may not be fully internalized by exporters as incumbents have not sufficient incentives to explicitly assist new entrants, leading to a market failure. Public intervention, in the form of ex- port promotion, through market-product speciï¬?c information and “matching with buyersâ€? services, could possibly help overcome these market failures, al- though the record of publicly-ï¬?nanced export promotion is patchy, especially in developing countries (see Lederman, Olarreaga and Payton 2010), and the effects we identify even if statistically robust are quantitatively small. The paper is organized as follows. Section 2 discusses the recent liter- ature related to this paper. Section 3 presents qualitative evidence based on a recent survey of African exporters conducted by the World Bank that motivates the consequent analysis and presents a brief description of the data. Section 4 discusses the estimation strategy and the result, Section 5 3 concludes. 2 Export survival: What do we know? At the product level, the determinants of export survival have been ex- plored by a small but growing literature. Besedes and Prusa (2006) used two panels of U.S. imports, one spanning 1972-88 with tariff-schedule data, the other spanning the 1989-2001 period with 10-digit data (the Feenstra-NBER dataset). In both cases, they found that half of all trade relationships lasted only one year and three quarters lasted three years or less. Once censoring was taken into account, median duration was two years. Most strikingly, this pattern of short duration was robust to aggregation at HS6, even though one would expect interruptions to be smoothed out by aggregation. They also found negative duration dependence, meaning that the hazard rate fell as export spells grew older. This ï¬?nding, however, has been recently contested by Brenton, Saborowsky and von Uexhull (2010).1 In terms of survival de- terminants, Besedes and Prusa (2006) found that industrial-country exports lasted longer, and so did exports of machinery, a ï¬?nding conï¬?rmed by the analysis of Asian trade flows by Obashi (2010). Besedes and Prusa (2006) explored the determinants of export survival further by testing the implications of a search model proposed by Rauch and Watson (2003) in which importers search for low-cost suppliers and exporters invest optimally in production capacity in the face of moral hazard (risk of non-payment). Such model implies that, in general, smaller initial trans- actions have a lower life expectancy; however, differentiated goods, where moral hazard is highest, involve both smaller initial transactions and longer life expectancy. The model’s predictions are upheld by Cox regressions on U.S. import data using Rauch’s (1999) index of product differentiation as a regressor. That is, the hazard rate is 23% higher for homogenous products than for differentiated ones, although initial transactions are 40% to 350% larger. In related work, Besedes (2008) also ï¬?nds supports for the Rauch- 1 Brenton et al. argue that the assumption of proportional hazards, which is needed for Cox regressions to be valid, typically does not hold in export-duration samples (this can be o veriï¬?ed using a Sch¨nfeld test). Using the alternative Prentice-Gloeckler (1978) estimator, they ï¬?nd no duration dependence. Brenton et al.’s critique applies to the quasi-entirety of the export-survival literature. 4 Watson hypotheses on a restricted sample of Rauch-differentiated products where he proxies search costs by the number of potential suppliers and reli- ability by income levels. Evidence on trade flows from other countries largely conï¬?rmed these early ï¬?ndings. The determinants of export duration were explored by Nitsch (2009) using Cox regressions on a ten-year panel of German imports at the HS8 level. He found that gravity variables (distance, exporter GDP, common language, common border, etc.) influenced the duration of trade flows pretty much the same way they influenced trade volumes. Interestingly, he found that the short duration of trade flows held even when flows below 10’000 euros were excluded. Fugazza and Molina (2009) extended the exploration to a nine-year panel of HS6 bilateral trade flows between 96 countries using, as regressors, gravity variables and time required for export procedures (based on the World Bank’s Doing Business surveys) as proxies for ï¬?xed costs. Be- sides usual ï¬?ndings on the effect of gravity variables and income levels, they also found that ï¬?xed costs reduced survival.2 A similar exercise was carried out on Asian trade flows by Obashi (2010) with largely convergent results. In particular, the 2-to-3 year median survival seems to hold across all samples studied. Obashi also found that vertical trade relationships (involving the sale of semi-ï¬?nished product) have hazard rates one-third lower than those involving the sale of ï¬?nal goods, and that they are less sensitive to trade costs (e.g. distance or exchange-rate fluctuations). A smaller number of recent papers have made use of the growing avail- ability of ï¬?rm-level datasets to shed new light on the determinants of export o survival. For instance, G¨rg et al. (2008) tested the implications of the heterogeneous-ï¬?rm model of Bernard et al. (2006) on a rich panel of 2,043 Hungarian ï¬?rms spanning the transition from centrally-planned to market economy (1992-2003). Their data contained ï¬?rm characteristics and exports at the ï¬?rm-HS6 level. They found large product turnover during the period as ï¬?rms constantly rearranged their product portfolios. They also found longer survival for products located close to the ï¬?rm’s core competencies and to the country’s comparative advantage. These results are consistent with those 2 This is unintuitive: in microeconomics, the shut-down point depends on average vari- able costs, not on ï¬?xed costs. However the ï¬?xed export costs they consider are incurred for each transaction, although they do not depend on transaction size. They are therefore not really ï¬?xed when looking at flows aggregated to the annual level. 5 of Iacovone and Javorcik (2010) who showed the importance of churning at the ï¬?rm level in response to exogenous opportunities provided by increased o globalization. Alvarez and L´pez (2008) used Tobit regressions to study the determinants of industry-level rates of entry and exit into exporting using a 10-year panel of 5’000 Chilean plants. They found that within-industry heterogeneity, measured (inter alia) by the dispersion of ï¬?rm-level produc- tivity levels, played an important role in explaining ï¬?rm turnover in and out of exporting. By contrast, trade costs, factor intensities, and exchange- rate fluctuations were found to have only marginal impacts. Carballo and Volpe (2008) used a 6-year panel of ï¬?rm-level Peruvian exports at the HS10 level to explore how diversiï¬?cation strategies (in terms of products and mar- kets) affected the survival of ï¬?rm-level exporting activity. They found that both geographical and product-wise diversiï¬?cation raised survival, but geo- graphical diversiï¬?cation more so—presumably because it proxies for product quality. 3 Data 3.1 Qualitative evidence from a World Bank survey Preliminary indications on how African exporters venture and survive (or not) on foreign markets can be gleaned from a 2009 survey focusing on ex- port survival conducted by the World Bank in four African countries.3 The survey, which had three sections (basic information on the ï¬?rm, constraints on survival, and opportunities and plans for future expansion), asked ex- porters speciï¬?c questions on their initial entry into and survival on export markets. On the basis of the information provided, respondents were clas- siï¬?ed into three categories: (i) current exporters, (ii) past exporters (who failed), and (iii) intermittent exporters. As shown by Table 7 in the appendix, roughly two thirds of the respon- dents (a bit more among regular exporters) identiï¬?ed their ï¬?rst client through relatives, friends, intermediaries and suppliers. More formal or technology- related channels (e.g. trade fairs or online research) came only second, and 3 The countries are Malawi, Mali, Senegal and Tanzania. See Appendix 1 for more background information on the survey. 6 only a tenth of the initial contacts were made through export promotion agencies or exporters’ associations. This highlights the importance of infor- mal networks and suggest that the “thicknessâ€? of a certain network where there are many ï¬?rms exporting similar countries to similar destination may help to expand the chance of identifying appropriate buyers among through contacts with relatives, friends, intermediaries and suppliers.4 Product experience, whether through domestic or foreign sales, appears as a strong driver of geographical export expansion. A majority of respon- dents reported that their initial export product was one they were already selling domestically, as opposed to starting a new line taylored to the for- eign customer’s needs. This suggests that experience matters and indirectly, it also could imply a natural 3-step expansion strategy: ï¬?rst the domestic market, then regional markets with similar preferences (so domestically sold products can be tried there), ï¬?nally more differentiated markets. This is con- e sistent with results in Cadot, Carr`re and Strauss-Kahn (2009) who showed that the survival of LDC exports was higher when export to OECD markets was preceded by a small number of years of exports to regional markets. Moreover, when asked whether their most recent export product in a given destination was a new one or one that had previously been exported else- where, respondents overwhelmingly indicated the latter. When asked how the opportunity to export a new product came about in the ï¬?rst place, the majority of regular exporters answered that they were approached by an existing buyer asking for a new product, suggesting that export experience matters beyond domestic experience in terms of establish- ing a “networkâ€? of buyers to identify market opportunities. Finally, in an open question about constraints on export (or export expan- sion in the case of the current exporters), a large proportion of respondents (31%) identiï¬?ed access to ï¬?nance as the main factor limiting their operations. Moreover, the percentage was higher (42%) among past (failed) exporters, suggesting that credit constraints are not just a perception, but a reality effectively hurting the survival of exports. 4 The role of networks for trade is a theme largely developed in the writings of Rauch (1999) 7 3.2 Customs data Our export dataset is generated from raw data ï¬?les collected by customs authorities containing export flows at the transaction level. The ï¬?les were provided by the customs authorities of Malawi, Mali, Senegal and Tanza- nia. Each of them contains information on products exported at the highest level of disaggregation of the HS code used by these administrations: 10-digit for Mali and Senegal and 8-digit for Malawi and Tanzania. In addition to product information, each ï¬?le contains information on destination market, FOB shipment value, net weight, port used and date of transaction. Original names and tax IDs identifying the individual ï¬?rm were replaced by ‘dummy’ digital IDs so as to preserve conï¬?dentiality. We aggregated transactions up to annual totals at the 6-digit level, the standard level used in cross-country comparisons. Finally, for consistency, we ï¬?ltered out years with different port coverage. For instance, for Malawi we have information from 2004 onward; however, as fewer ports were covered in 2004 than in other years, we ex- cluded 2004 from our sample for that country. Sample periods are 2005-2008 for Malawi and Mali, 2000-2008 for Senegal, and 2003-2008 for Tanzania. Table 1 presents some basic descriptive statistics. Tanzania has the largest number of exporters (1,359), followed by Malawi (856), Senegal (715), and Mali (280); however, they are less diversiï¬?ed than those of other coun- tries in our sample in terms of markets. Mali’s exporters are, on average, the most diversiï¬?ed in terms of products. Our variables of interests are indexed as follows. Let f be a ï¬?rm, d a destination, p a product (at HS6), t the starting year of an export spell, and c vf pdt the dollar value of exports of product p to destination d in calendar year t by ï¬?rm f from country c. Because there are no multi-country ï¬?rms in our sample, indexing observations by ï¬?rm eliminates the need to index them by origin country. We aggregate transactions to annual (f, p, d, t) quartets, our primary sample unit. 8 Table 1: Descriptive statistics Nr ï¬?rms Nr prod. Nr dest. Nr prod/ï¬?rm Nr dest/ï¬?rm Nr ï¬?rms/prod Nr ï¬?rms/dest Init. value (USD) Mean Median Mean Median Mean Median Mean Median Mean Median Mali 280 575 99 2.54 2 3.89 2 1.89 1 7.18 2 219,694 5,373 Malawi 856 932 102 1.57 1 4.10 2 3.76 1 13.19 3 106,475 571 Senegal 715 1,653 100 3.10 1 6.76 2 2.92 2 22.17 5 47,111 3,446 Tanzania 1,359 1,689 137 2.49 1 3.62 1 2.91 1 24.69 7 83,078 2,858 This table shows, for each source country: The number of ï¬?rms, products, destinations, the number of products per ï¬?rm, number of destination per ï¬?rm, number of ï¬?rms per product, number of ï¬?rms per destination and the value of those ï¬?rms 9 that entered the export market. All values are computed for the year 2006. Before turning to survival analysis (next section), a few observations are important. Following the literature on the intensive and extensive margins (e.g. Evenett and Venables 2003 or Brenton and Newfarmer 2007), we group our primary sample units into new ï¬?rms, new products (for existing ï¬?rms), new destinations (for existing ï¬?rm-products), and continuing ï¬?rm-product- destinations. Items labeled ‘new’ refer to units that are present in the data at time t but not at time t − 1.5 These groupings create four mutually ex- clusive categories. The ‘new-ï¬?rm’ category includes all product-destination combinations served at time t by an exporter appearing in the data in that year (except the ï¬?rst year). The ‘new-product’ category includes all product- destination combinations served at time t by an existing exporter —one that already exported at t − 1— who did not export that product anywhere at t − 1. The ‘new-destination’ category includes all product-destination com- binations served at time t by an existing exporter who did not serve that destination with any product at t − 1. The ‘existing product-destination’ category includes all product-destination combinations served at time t by an exporter who was also serving that product-destination at t − 1. More formally, let vf,t−1 stand for f ’s exports of any product to any destination at t − 1, vf p,t−1 for its exports of product p to any destination, vf d,t−1 for its exports of any product to destination d, vf pd,t−1 for its exports of product p to destination d. Our four categories are NF = {(f, p, d, t)s.t.vf pdt > 0 and vf,t−1 = 0}, NP = {(f, p, d, t)s.t.vf pdt > 0, vf,t−1 > 0 and vf p,t−1 = 0}, ND = {(f, p, d, t)s.t.vf pdt > 0, vf,t−1 > 0 and vf d,t−1 = 0}, EPD = {(f, p, d, t)s.t.vf pdt > 0 and vf pd,t−1 > 0}. The dollar value of export sales in the ï¬?rst three categories can only go from zero at t − 1 to some positive value at t; these variations add up to the ex- tensive margin. Changes in the dollar value of exports in the last category form the intensive margin. Figure 1 decomposes the exports flows into these four categories both in terms of their number, i.e. count of trade flows, and value. 5 Observations in the sample period’s initial year are considered left-censored and not used. 10 Figure 1: Decomposition of exports flows 11 Note: This graph classiï¬?es each of the origin-ï¬?rm-product-destination observations into one of four mutually exclusive groups: New Destination includes units of existing ï¬?rms which export an existing product to a new destination; New Products includes existing ï¬?rms that add a product to their portfolio, New Firms includes all units from ï¬?rms that did not export before, while Continued includes all other units. The ï¬?rst set of graphs displays the share of observations, and the second set the share of total values of each category. Analyzing the export values, existing products sold in existing destina- tions (i.e. observations for which ï¬?rm, destination and HS6 at time t are all the same as they were at time t-1) dominate in dollar value, although not always in the count of observations. For example, in Tanzania, continued ï¬?rm-product-destinations accounted for 90 percent of export value in 2006 but only for 25% of the observation count. This suggests that our coun- tries experiment substantially. This fact is consistent with the ï¬?ndings of Cadot, Carriere and Strauss-Kahn (2010) for low-income countries, Freund and Pierola (2010) for Peru and Iacovone and Javorcik (2010) for Mexico. Continuing ï¬?rm product destinations make up a relatively small number of export transactions, but a large share of export values. This conï¬?rms the ï¬?ndings of Besedes and Prusa (2007) and Brenton and Newfarmer (2007), who also show the importance of the intensive margin in explaining export growth in developing countries (see also Evenett and Venables 2002). Another interesting stylized fact, consistent with existing ï¬?rm-level lit- erature modeling exporters dynamics (Rauch and Watson 2003), conï¬?rms that when a ï¬?rm’s product manages to survive in a given destination market beyond its ï¬?rst year, it will grow signiï¬?cantly over time. Conditional on sur- vival, Senegalese ï¬?rm-product-destinations that appeared in 2001 (we don’t know the initial year of those appearing in 2000, the sample’s initial year, because they are censored) grew by a factor of over four between 2001 and 2008. Similarly, Tanzanian ï¬?rm-product-destinations that appeared in 2005 grew by a factor of over three by 2008. Following Brooks (2006), Table 2 shows the number of ï¬?rms, ï¬?rm-products, and ï¬?rm-product-destinations by a given year of entry and tracks the sur- vival of this cohort over time for each origin country. Naturally, the numbers decrease because of the exit. What is remarkable, however, is how large the attrition is in the ï¬?rst year and how quickly it slows down over time. For instance, in Senegal, of the 206 ï¬?rms that started exporting in 2001, only 84 made it to 2002 (a death rate of 59%); however, of the 24 still around in 2007, only 3 had failed by 2008 (a death rate of “justâ€? 12%). To make this point more clear, the third column of Table 2, calculated from the second one, shows the survival rate with respect to the previous year (i.e. one minus the annual death rate). Survival rates increase over time. For instance, 59 percent of ï¬?rms that entered in 2001 dropped out by 2002, while 13 percent of ï¬?rms that survived until 2007 survive also until 2008. This casual obser- 12 Table 2: Survival cohorts Senegal Tanzania Mali Malawi Entry:2001 Entry:2004 Entry:2005 Entry:2005 Nr Y-Exit Exit Nr Y-Exit Exit Nr Y-Exit Exit Nr Y-Exit Exit Firm 2001 206 2002 84 0.59 0.59 2003 57 0.32 0.72 2004 40 0.30 0.81 420 2005 35 0.13 0.83 194 0.54 0.54 273 670 2006 29 0.17 0.86 118 0.39 0.72 159 0.42 0.42 217 0.68 0.68 2007 24 0.17 0.88 85 0.28 0.80 123 0.23 0.55 154 0.29 0.77 2008 21 0.13 0.90 75 0.12 0.82 103 0.16 0.62 126 0.18 0.81 Product 2001 2055 2002 449 0.78 0.78 2003 192 0.57 0.91 2004 117 0.39 0.94 2656 2005 94 0.20 0.95 497 0.81 0.81 1047 3322 2006 78 0.17 0.96 200 0.60 0.92 305 0.71 0.71 325 0.90 0.90 2007 61 0.22 0.97 106 0.47 0.96 166 0.46 0.84 174 0.46 0.95 2008 54 0.11 0.97 71 0.33 0.97 123 0.26 0.88 127 0.27 0.96 Product destinations 2001 3326 2002 718 0.78 0.78 2003 356 0.50 0.89 2004 245 0.31 0.93 4908 2005 167 0.32 0.95 837 0.83 0.83 1391 3828 2006 129 0.23 0.96 295 0.65 0.94 286 0.79 0.79 509 0.87 0.87 2007 101 0.22 0.97 167 0.43 0.97 122 0.57 0.91 316 0.38 0.92 2008 84 0.17 0.97 113 0.32 0.98 82 0.33 0.94 224 0.29 0.94 Note: In the columns labelled Nr we document for each origin country the number of ï¬?rms products and destinations in the ï¬?rst available year, and follow this cohort of units over time. Column Y-Exit shows the exit rate (ie. the share of units that left) with respect to the previous year, and column Exit the exit rate with respect to the entry year. 13 vation is consistent with Besedes and Prusa’s decreasing-hazard rate ï¬?nding (annual death rates are discrete-time approximations to instantaneous hazard rates) although, as noted, this ï¬?nding must be taken cautiously. Comparing the upper panel (ï¬?rms) with middle and lower ones (products and product- destinations respectively), there is less stability at more disaggregate levels. Additionally, the fourth column shows cumulative death rates relative to the ï¬?rst year. In all cases these rates are high, and above 80% in 2008 in most cases (with the only exception of Mali at the ï¬?rm level). In all four countries, the very high death rates after the ï¬?rst year suggest that a binary coding of survival based on second-year outcomes is a good summary mea- sure of survival. Overall, the results presented in Figure 1 and Table 2 suggest that there is substantial churning in export products and destinations within ï¬?rms; in other words, ï¬?rms continuously experiment with products and destinations. Thus, Hausman and Rodrik’s ‘self-discovery’ process (Hausman and Rodrik 2003) seems to hold not only at the national level, but also—quite naturally— at the ï¬?rm level. This pattern is also consistent with the notion that ï¬?rms face uncertainty about export costs or demand parameters, a notion that is central to the heterogeneous-ï¬?rms literature. In sum, the preliminary evidence presented above conï¬?rms existing ï¬?nd- ings about export growth and survival: a) the intensive margin represents the largest share of export growth in terms of values, however these values are concentrated over a small number of transactions and ï¬?rms; b) there is substantial experimentation in the exporting activity in the form of entry by new ï¬?rms or the introduction of new products or destinations each year; c) one-year survival rates are low; past the ï¬?rst year, death rates signiï¬?cantly slow down and transaction volumes grow. 4 Estimation strategy and results 4.1 Estimation strategy After aggregating the transactions to cumulated annual totals, the primary sample remains a panel, as each ï¬?rm-product-destination (f, p, d) triplet is 14 observed repeatedly over several years. However, as we are interested in the survival past the ï¬?rst year, the data needs to undertake a second transfor- mation. We deï¬?ne a new (f, p, d, t) quartet as one that appears for the ï¬?rst time in the database, and say that this quartet ‘survives’ if it lasts more than one year. The quartet is then associated to a survival dummy (our depen- dent variable) equal to one. If it lasts only one year, the survival dummy is set equal to zero for that quartet. If it has already appeared in the sample or if it is left-censored (i.e. already active the ï¬?rst year of the sample), we drop it. Multiple spells account for only a very small number of observations, since our sample periods are only a few years except for Senegal. Thus, we reduce our panel to a quasi-cross-section, even though each observation has an initial-year tag allowing us to control for calendar time. Doing so allows us to bypass the issue of how long a spell break should be to be considered a ‘death’, an issue that has been discussed at length in the survival liter- ature and that has no clear-cut answer. Two additional reasons make this binary deï¬?nition of survival attractive. First, our panels are too short to carry out a full-fledged survival analysis. Second, as the descriptive analysis above showed, once a ï¬?rm has survived the ï¬?rst year, its survival probabil- ity dramatically increases; so understanding survival beyond the ï¬?rst year is especially important.6 As already noted, ï¬?rm and country indices are redundant, so we use either a country superscript c or a ï¬?rm subscript f , but not both, and run our regressions on a pooled cross-country sample.7 Our dependent variable is 1 ifvf pdt > 0, vf pd,t− = 0 ∀ > 0, andvf pd,t+1 > 0 sf pdt = (1) 0 ifvf pdt > 0, vf pd,t− = 0 ∀ > 0, andvf pd,t+1 = 0. In 1, the expression “∀ > 0â€? means “over the sample periodâ€? as a single spell over the sample period could be a multiple one over an (unobserved) longer sample. The estimating equation is 6 This choice comes with both a cost and a beneï¬?t. On one hand, we lose information, as a two-year spell is treated as equivalent to a 3- or 4-year one; on the other hand, we gain robustness, as the probability of wrongly treating a two-or-more year spell as a one-year one is fairly low. 7 We also ran, for robustness, separate regressions by origin country. The results of these regressions are available upon request. They are qualitatively similar to those of cross-country (pooled) regressions reported here. 15 Pr(sf pdt = 1) = φ (xf pdt β + δi + δcd + δt + uf pdt ) (2) where φ is the probit function and uf pdt is an error term. Our speciï¬?cation in- cludes industry ï¬?xed effects at HS2 (δi ), origin-destination ï¬?xed effects (δcd ), and spell-start year ï¬?xed effects (δt ). The vector of regressors xf pdt includes measures of the ï¬?rm’s scale and scope as well as proxies for agglomeration and market attractiveness. These proxies are counts of (i) nc , the number pdt of ï¬?rms from origin country c exporting product p to destination d; (ii) nf pt , the number of destinations to which ï¬?rm f exports product p; (iii) nf dt , the number of products that ï¬?rm f exports to destination d; (iv) nc , the num- dt ber of (product × ï¬?rm) combinations active in the bilateral trade between origin c and destination d; they also include (v) vf pdt , the initial value of ï¬?rm f ’s export spell (product p to destination d); and (vi) zf p , the share of product p in ï¬?rm f ’s overall export sales. That is, the notation convention is to omit the index of the dimension over which the count is summed. All counts are put in logs, and we use robust standard errors clustered at the product-destination level throughout. In customs data, E.U. countries are entered as separate destinations rather than as a whole. We have kept this convention, so a destination should be taken, as far as the E.U. is concerned, as a member state. This creates an asymmetry in the treatment of destinations between the U.S., which is taken as a whole, and the E.U., which is broken down. However, as African exports tend to be heavily concentrated on E.U. markets, the al- ternative assumption (bundling all E.U. destinations together) would have drastically reduced the number of destinations and potentially obfuscated some geographical diversiï¬?cation issues, as marketing channels are, in spite of the Single Market, still somewhat separate across E.U. member states. We estimate equation 2 by probit, reporting marginal effects. Typically, marginal effects of a probit estimation can be interpreted like the coefficient in a linear probability model, and also in the present case a robustness check reveals that quantitatively the difference between the results from a linear probability model and the probit’s marginal effects at the mean are very small.8 8 Results of a comparison of the linear probability model and Probit estimates are available upon request. 16 4.2 Baseline results Baseline regression results are shown in Table 3. Before turning to their detailed interpretation, it is important to stress that the effects to be dis- cussed are simultaneously present in each regression and so are conditional on each other. Also, it is important to note that these must be interpreted as conditional on starting to export. The probability of survival beyond the ï¬?rst year t can be estimated only for those trade flows that started at t − 1, so we exclude left-censored spells (those already active at the start of the sample) and multiple ones.9 The ï¬?rst column presents the baseline results. The second differs from the ï¬?rst in that all right-hand side (RHS) variables are lagged by one year. The third and the fourth include one additional control each, the share of product p in ï¬?rm f ’s export portfolio in the third and origin country c’s revealed comparative advantage (RCA) in product p in the fourth. The ï¬?fth runs a counterfactual experiment which is discussed below. Consider ï¬?rst the results in Column (1). The ï¬?rst regressor of interest is ln nc , the log of the number of ï¬?rms selling the same product (p) in the same pdt destination (d). That is, if spell (f, p, d, t) is mens’ t-shirts sold in France by a Senegalese ï¬?rm in 2006, ln nc is the log of the number of Senegalese pdt ï¬?rms exporting mens’ t-shirts in France in 2006. The effect is positive and signiï¬?cant at the 1% level in all speciï¬?cations. That is, more companies from the same country selling the same product in the same destination together raise each other’s survival probability. This is a striking network effect, to which we will come back at some length later on. How large is the effect? Let us write the probability of survival as Ï€f pdt = Pr(sf pdt = 1). Recalling that the coefficients reported in Table 3 are marginal effects, using the point estimate of of 0.0566 in the ï¬?rst cell of Column (1), and the average number of Senegalese ï¬?rms selling to each destination (nSEN = 22) we can write pdt 9 The number of multiple spells is very limited and their inclusion does not influence our results. 17 Table 3: Determinants of survival past the ï¬?rst year Regressors (log) (1) (2) (3) (4) (5) nc pdt Firm count 0.0566*** 0.0431*** 0.0544*** 0.0563*** (0.00283) (0.00306) (0.00282) (0.00285) n−c pdt Firm count 0.00449 (0.00727) nf pt Dest. count 0.125*** 0.0820*** 0.125*** 0.125*** 0.116*** (0.00270) (0.00296) (0.00269) (0.0027) (0.00397) nf dt Prod. count 0.0375*** 0.0224*** 0.0478*** 0.0375*** 0.0301*** (0.00163) (0.00152) (0.00184) (0.00163) (0.00218) vf pdt Init. value 0.0304*** 0.0332*** 0.0277*** 0.0304*** 0.0335*** (0.000898) (0.000889) (0.000921) (0.000898) (0.00125) nc dt Prod. × ï¬?rm -0.00477 -0.0213*** -0.00723 -0.00472 -0.00131 (0.00594) (0.00397) (0.00595) (0.00594) (0.0084) zf p Prod. share 0.0771*** (0.00640) RCAcp <0.0001 <0.0001 (<0.0001) (<0.0001) RHS vars lagged No Yes No No No HS2 FE Yes Yes Yes Yes Yes Origin-dest. FE Yes Yes Yes Yes Yes Time effects Yes Yes Yes Yes Yes Obs. 57,063 57,063 57,063 57,063 11,185 Note: Probit estimations, marginal effects reported. Origin-destination, hs2 and year ï¬?xed effects. Robust standard errors clustered by origin-destination-product. 18 ∆πf pdt = 0.0566 [ln(nc + 1) − ln(nc )] pdt pdt = 0.0566 [ln(23) − ln(22)] (3) = 0.0025. Using the illustrative attrition rates in Table 2, a Senegalese ï¬?rm entering in 2001 has a ï¬?rst-year attrition rate of 0.78 (78%) at the product-destination spell level, implying a survival probability of 0.22. We take this number as our baseline probability of survival, Ï€f pdt . Raising it by 0.0025 means a neg- ligible increase of 0.2 percentage points. Doubling the number of national competitors on a given product-destination niche, from the baseline of 22 to 44, would raise the ï¬?rst-year survival probability by 3.9 percentage points, from 0.22 to 0.26 (a proportional increase of 18%). Skipping the second regressor which is a placebo used in Column (5) and discussed later on, the second regressor in Column (1) is ln nf pt , the count of destinations to which product p is exported by ï¬?rm f , a proxy for the scale at which the ï¬?rm exports, and thus probably produces, product p. Scale signiï¬?cantly raises the probability of survival in all regressions. This may reflect either more robust production lines (say, a larger number of machines, meaning that failure of one of them is more easily made up by others), better information about the cross-country drivers of a product’s demand, or, alternatively, higher product quality. How large is the effect? Using a calculation similar to that in (3), an additional destination10 raises the probability of survival by ∆πf pdt = 0.125 [ln(2.55) − ln(1.55)] = 0.062. That is, the baseline ï¬?rst-year probability of survival of a spell goes up by 6.2 percentage points, from 0.22 to 0.28, when the mean Senegalese exporter adds one destination to his portfolio at the product level. If she was to double the number of his destinations for that product, the ï¬?rst-year survival prob- ability would rise by 8.7 percentage points, from 0.22 to 0.31 (a proportional rise of 39%). 10 The average number of destinations per product, for a Senegalese ï¬?rm, is 1.55. This is lower than the number appearing in Table 1 which is the total number of destinations per ï¬?rm, not per ï¬?rm-product. 19 The next regressor, ln nf dt , is the log of the number of products ï¬?rm f exports to destination d, a proxy for its ‘scope’ in that destination. The ef- fect is, again, positive and signiï¬?cant. As for its magnitude, if our Senegalese ï¬?rm adds one product to its average destination d, from a baseline of 3.48 products,11 the usual calculation gives a rise of just 1 percentage point in the survival probability. With a doubling of the number of products, the survival probability rises by 2.5 percentage points, from 0.22 to 0.245, a proportional rise of 11%. Thus, adding one product to a given destination has a smaller effect on spell survival (1 percentage point) than adding an additional destination for that product (1.7 percentage point). This is somewhat natural, as our analysis is at a disaggregated level in terms of products (5,000 products at HS6), so the additional product sold on destination d can be very close to the original; by contrast, destination countries are much fewer, so adding one more shipping destination for product p is a substantial move (although it may involve adding one E.U. member state which would mean expanding within the Single Market space). An alternative explanation goes as follows. Increasing either scope or size raises the ï¬?rm’s visibility and therefore has a positive demand effect. However, there may be supply effects running at cross-purposes. When a ï¬?rm adds one export destination to a given line of products, it expands production, potentially making the value chain more robust to accidental fluctuations. By contrast, when it adds one product to a destination, the ï¬?rm diversiï¬?es production and therefore spreads managerial attention and risk management over a wider range of activities, potentially resulting in more accidents. In that case, the supply effect runs against the demand effect, resulting in a lower net change in spell survival. The next regressor is a control for the export spell’s initial value, vf pdt , which has been shown to correlate with spell survival at the product (multi- ï¬?rm) level. This is conï¬?rmed at the ï¬?rm level, although the effect is, again, small. Using the coefficient in Table 3 (0.0304), a doubling of the initial value of the Senegalese ï¬?rm’s average export spell ($47’111 from Table 1) would raise the probability of spell survival by 0.021, or 2 percentage points, from 0.22 to 0.241. 11 Again, this number differs from the one appearing in Table 1, which is the total number of products per ï¬?rm, not per ï¬?rm-destination. 20 The last regressor, ln nc , is a count of the ï¬?rm-product pairs from coun- dt try c active on destination d. If c is Senegal and one Senegalese ï¬?rm sells two HS6 products in the E.U. and another one sells three, nc = 5 for all dt ï¬?ve observations with c = Senegal and d = E.U. in year t. It is a proxy for the size of the bilateral trade relationship.This variable is never signiï¬?cant except in Column (2). Column (2) of table 3 is very similar to Column (1) except that all the explanatory variables are lagged by one period. Results are essentially un- changed, except for nc whose coefficient becomes negative and signiï¬?cant. dt What that means is that more ï¬?rm-product combinations from a given origin to a given destination are associated with a lower probability of survival past the ï¬?rst year. Without making too much of this result, one can interpret it as follows. Given that we include origin-destination ï¬?xed effects, nc picks dt up only the time-variant component of bilateral shocks, like booms in the destination market. The negative coefficient suggests that a growth expan- sion (a boom) in t − 1 triggers crowding in followed by retrenchment.12 Column (3) introduces an additional regressor. The literature on multi- product ï¬?rms suggests that ï¬?rms have core and marginal products, and that they have a stronger competitive advantage in the former (see for instance Eckel and Neary 2010 for a theoretical model and Iacovone, Rauch and Win- ters 2010 for an empirical test of this hypothesis). For each multiproduct ï¬?rm f and product p, we proxy how close is that product from the ï¬?rm’s ‘core’ by ln zf p , the log of its share in the total ï¬?rm’s export sales. Results suggest that it correlates positively with ï¬?rst-year survival probability even after controling for dollar initial value; that is, the probability of survival for ‘core’ products is substantially higher than for others. For instance, a prod- uct representing 80% of the ï¬?rm’s export sales (all destinations together) would have a ï¬?rst-year survival probability on a given destination higher by 12 Conï¬?rming this interpretation, when we exclude the destination ï¬?xed effects, the coefficient on this variable becomes positive, suggesting that permanently more attractive markets are associated with longer survival, which is consistent with our interpretation. This “crowding-inâ€? result is also consistent with a ï¬?nding by Bussolo, Iacovone, and Molina (2010) who found, using ï¬?rm-level data from the Dominican Republic, that the reduction of tariffs following the signature of CAFTA led to some over-crowding of Dominican exports, followed by retrenchment. 21 10 percentage points than a product representing 20% of the ï¬?rm’s export sales. In column (4) of table 3, we control for a potential omitted variable that could bias our results if country c had a comparative advantage in product p, explaining both that it had more exporters of that product (in destination d or elsewhere) and that product p had a better survival outlook. As a control for this, we use the initial (sample-start) value of Balassa’s revealed comparative advantage (RCA) index deï¬?ned, for product p, as vpc / p vpc RCApc = (4) vpw / p vpw where vpc stands for country c’s exports of product p and xpw for world ex- ports of that good. Balassa’s index measures the ratio of the share of product p in country c’s export basket relative to it share in the world’s export bas- ket. The higher it is, the more that country is revealed to have a comparative advantage in that product. We compute it at HS6 from mean exports for 1999, 2000 and 2001. Results are robust to the inclusion of this control. Finally, Column (5) provides a key test of whether our synergy effect is spurrious by replacing it with a ‘placebo’. Namely, we replace ln nc , the pdt number of ï¬?rms exporting the same product to the same destination from the same country, by ln n−c , the number of ï¬?rms exporting the same product to pdt the same destination from other countries. For instance, consider an export spell of boys’ swimwear (HS611239) to Germany by a Senegalese ï¬?rm. On the right-hand side of the equation, instead of the number of other Sene- galese ï¬?rms exporting HS611239 to Germany, we will now have the number of ï¬?rms exporting HS611239 to Germany from other countries in our sam- ple (Tanzania, Malawi and Mali). This variable may be positive or zero. It may also be missing, as our national samples have some non-overlapping years, so the sample size is substantially lower. It should also be kept in mind that the placebo we are using is neither random nor “matchedâ€?, being dictated by data availability. It is thus not a rigorous counterfactual. Be that as it may, whereas the synergy effect comes out very strongly in all speciï¬?cations, whether pooled across countries (as reported in Table 3) or run separately by country, the placebo effect is never signiï¬?cant. This test contributes to increase our conï¬?dence that are “synergy effectâ€? is not identi- 22 fying some spurious correlation; it also suggests that there is some national element in the synergy we identify (recall that regressions include bilateral origin-destination ï¬?xed effects). 4.3 Interpreting the synergy effect 4.3.1 Extended networks and “institutional production capabili- tiesâ€? We now turn to possible interpretations of the synergy effect that we iden- tiï¬?ed. We ï¬?rst explore if the synergy effect we identiï¬?ed in Table 3 carries over to extended networks of exporters of “similarâ€? products. This has the advantage of reinforcing our attempt to ï¬?lter out omitted-variable bias, as extended networks at the industry level may pick up the effect of compara- tive advantage, infrastructure, and intermediation channels in a more robust way than Balassa indices calculated at the HS6 level do. First, we deï¬?ne a new regressor, which we will call HS4 for simplicity, equal to the number of products other than p exported by ï¬?rm f to des- tination d and belonging to p’s HS4 heading. Table 4 reports results with HS4 added to the main speciï¬?cation. The new variable has a positive and signiï¬?cant effect on survival, but it does not affect the signiï¬?cance or mag- nitude of our synergy effect. In column (2), we interact this variable with nc , the synergy effect. Again, the synergy effect itself remains positive and pdt signiï¬?cant, but the coefficient on the interaction term is negative. What this means is that the more there are ï¬?rms selling “similarâ€? (same HS4 but dif- ferent HS6) products, the less ï¬?rm f is sensitive to the ‘network’ of ï¬?rms selling the exact same product (at HS6)— intuitively networks of identical and “similarâ€? products are somewhat substitutes. As an alternative, in column (3) we deï¬?ne a new variable, HK, equal to the weighted sum of the number of ï¬?rms exporting product p∗ to the same destination d where the weights are equal to the ‘distance’ between p and p∗ in the sense of Hausmann and Klinger (2006).13 This new variable has no signiï¬?cant effect on the probability of survival, but it does not af- 13 Hausmann and Klinger’s measure of proximity is essentially a measure of the proba- bility that two goods are exported simultaneously by a country. 23 Table 4: Extended networks (1) (2) (3) (4) nc pdt 0.0551*** 0.0699*** 0.0607*** 0.0709*** (0.00364) (0.00624) (0.00300) (0.00362) nf pt 0.151*** 0.151*** 0.132*** 0.132*** (0.00374) (0.00374) (0.00289) (0.00289) nf dt 0.0322*** 0.0321*** 0.0305*** 0.0304*** (0.00112) (0.00112) (0.000916) (0.000915) vf pdt 0.0414*** 0.0413*** 0.0375*** 0.0376*** (0.00199) (0.00198) (0.00166) (0.00165) nc dt -0.0170** -0.0178** 0.00101 0.00175 (0.00860) (0.00858) (0.00585) (0.00585) HS4 0.0113*** 0.0170*** (0.00285) (0.00330) HS4 ×npdt -0.00805*** (0.00284) HK 0.00466 0.0179*** (0.00419) (0.00483) HK ×npdt -0.0241*** (0.00553) Observations 38451 38451 52212 52212 Note: Probit estimations, marginal effects reported. Origin-destination, HS2 and year ï¬?xed effects. Robust standard errors clustered by origin-destination- product. fect it; interacted, in Column (4), the effect is, again, negative, suggesting some substitutability between the networks of identical and “closeâ€? products. In conclusion, in this subsection we evaluated the possibility that our results could be driven not by the “synergyâ€? effects due to the presence of companies exporting same HS6 products to same destination but rather by some “broaderâ€? extended networks. In the light of recent work of Hidalgo et al (2007) we could be concerned that a key omitted variable driving our re- sults is indeed the existence of some broader “institutional capabilitiesâ€?. For this reason, we added to our baseline speciï¬?cations two new variable captur- ing these potential “institutional capabilitiesâ€? and found that, while indeed 24 these are important and there seems to be some substitutability between these and the “synergy effectâ€?, nevertheless the inclusion of these variables does not alter our previous results. 4.3.2 Information and access to ï¬?nance We now turn to an exploration of the mechanisms that could explain our re- sults, primarily focusing on the hypothesis that the “synergy effectâ€? is driven by the existence of some “information spilloversâ€?. First, this synnergy effect could indicate the presence of information ex- ternalities. For instance, when technical regulations or buyer policies change in the destination market, exporters may share information about upcoming changes, improving their ability to anticipate and adapt to these changes. Alternatively, buyers may take suppliers from a given country more seriously (and therefore share more information with them or show more flexibility in the face of glitches) when there is a critical mass of them and improve their reliability. If this conjecture is correct, we should expect a stronger synergy effect for products characterized by higher quality heterogeneity for which information asymmetries between buyers and producers are potentially more important. We proxy product p’s quality heterogeneity by Ï?p , the coefficient of variation of its FOB unit value across exporters in 2000 (the initial value in our sample) using COMTRADE data, with a higher Ï?p meaning more heterogeneous quality.14 The results are presented in column (1) of Table 5. The coefficient on the interaction term Ï?p × ln npdt is positive, although sig- niï¬?cant only at the 10% level, suggesting that the synnergy effect is stronger for products with a high unit-value dispersion, where information is more important. Given the importance of ï¬?nance, as shown by the survey discussed in Sec- tion 3, an alternative hypothesis could be that while information is still a key determinant of the synergy effect the mechanism behind it could be instead different. Consider the following scenario. A Senegalese ï¬?rm is approached by a US buyer to provide a small trial order of t-shirts. Upon successful delivery and sale, the buyer is happy and contacts again the Senegalese ï¬?rm for a larger order. Now the Senegalese ï¬?rm has to ramp up capacity and, 14 We explored results on sub-samples split by Rauch’s categories in a table that is available upon request. 25 Table 5: Mechanisms behind the synergy effect (1) (2) (3) (4) (5) nc pdt 0.0552*** 0.0512*** 0.0932*** 0.0455*** 0.0816*** (0.00437) (0.00496) (0.0112) (0.00577) (0.0121) nf pt 0.132*** 0.132*** 0.138*** 0.132*** 0.137*** (0.00290) (0.00289) (0.00343) (0.00289) (0.00343) nf dt 0.0375*** 0.0377*** 0.0369*** 0.0378*** 0.0370*** (0.00166) (0.00165) (0.00183) (0.00165) (0.00183) vf pdt 0.0305*** 0.0307*** 0.0290*** 0.0307*** 0.0291*** (0.000916) (0.000918) (0.00104) (0.000918) (0.00104) nc dt 0.00110 0.00131 -0.00275 0.00142 -0.00234 (0.00585) (0.00585) (0.00635) (0.00585) (0.00635) Ï? -0.00610 -0.00641 -0.0194*** (0.00500) (0.00499) (0.00565) Ï? p × nc pdt 0.00954* 0.00971* 0.0139** (0.00561) (0.00553) (0.00653) rp 0.0114* 0.0115* (0.00640) (0.00640) rp × nc pdt 0.0140** 0.0141** (0.00586) (0.00584) κ 0.168*** 0.174*** (0.0451) (0.0451) κp × nc pdt -0.115*** -0.107*** (0.0346) (0.0347) Obs. 52212 52212 37838 52212 37838 Note: Probit estimations, marginal effects reported. Origin-destination, HS2 and year ï¬?xed effects. Robust standard errors clustered by origin-destination-product. 26 for that, it needs support from ï¬?nancial institutions. But the ï¬?nancial insti- tutions may not take letters of credit from the buyer at face value, because are aware of all sorts of glitches – quality or other – that may emerge down the line. Anecdotal experience suggests that, in Sub-Saharan Africa, the banks’ response will typically be ‘no’ irrespective of the “proofs of proï¬?tabil- ityâ€? that the exporter shows, and the trade relationship with the US buyer will end before it had a chance to bear fruit. However, if several Senegalese ï¬?rms already sell t-shirts on the US market, the same ï¬?nancial institutions may be more easily convinced about the chances of success of this venture and better evaluate the potential risks involved in this transaction. If this scenario is representative, the synergy effect should be stronger for products that are especially dependent on external ï¬?nance than for others as initial ï¬?nancial constraints would be more binding in these sectors. In order to test this conjecture we interact our variable identifying the synergy effect with the measure of dependence from external ï¬?nance proposed by Rajan and Zingales (1998).15 We construct our rp variable at the product level by using concordance tables between ISIC3 and HS6 classiï¬?cation, and assign- ing to each HS product the Rajan-Zingales index of the ISIC code to which that product belong. Column (2) of Table 5 shows that the interaction term rp × ln npdt is positive and signiï¬?cant. As an alternative way of getting a handle on the degree of dependence from ï¬?nance, we use a proxy for ‘asset tangibility’ proposed by Braun (2003).16 The idea that ï¬?rms with more tangible assets presents lower risks as these provides real guarantees for bank loans, and information asymmetries (ad- verse selection or moral hazard) are less important with good collateral, so synnergy effects should play a lesser role. In accordance with this conjecture, in column (3) of Table 5 we show that the interaction of asset tangibility (rp ) 15 Rajan and Zingales’ measure of ï¬?nancial dependence is an industry-level variable calculated for 27 3-digits ISIC industries and nine 4-digits ones using compustat data for the US. Let k be capital expenditure and x operational cash flow at the ï¬?rm level. Rajan and Zingales’ index for industry j, rj , is the median value of (k −x)/k across all compustat ï¬?rms in industry j. Index values, given in Table 1 of Rajan-Zingales (1998), range from -0.45 for tobacco (ISIC 314) to 1.49 for drugs (ISIC 3522). 16 Braun proxies asset tangibility by the ratio of net property, plant and equipment to market value at the ï¬?rm level, using US compustat data. The industry-level variable is constructed, like in Rajan-Zingales, by taking the industry median at the ISIC 3-digit level. Index values, given in Table 1 of Braun (2003), range from 0.09 (leather products) to 0.67 (petroleum reï¬?neries). 27 and the synergy effect has a negative and signiï¬?cant coefficient, implying that ï¬?rms belonging to industries with high asset tangibility (essentially capital- intensive industries) are less sensitive to the synergy effect.17 Given that these interactions have signiï¬?cant explanatory power, we com- bine them to address a potential omitted variable bias, and to compare there coefficients in a joint multivariate regression. In columns (4) and (5) we combine one regressors capturing each of the three hypothesis that these in- teractions try to capture: information, ï¬?nancial constraints or capabilities. We ï¬?nd that typically the same signs, magnitudes and statistical signiï¬?cance levels persist as the ones just discussed, and the interpretations from above are valid when we control for all these effects simultaneously. 5 Concluding remarks In spite of their growing interest for the profession, ï¬?rm-level datasets are still rare for low-income countries, and virtually inexistent for African coun- tries. Our exploration of African customs data on ï¬?rm-level exports revealed a set of stylized facts that are consistent with evidence from previous studies analyzing rich or middle-income countries. We showed that exporters in our set of African countries experiment a lot on export markets, at a low scale and with low survival rates, particularly in the ï¬?rst year. That is, they op- erate in a difficult environment characterized by very high “infant-mortality ratesâ€?. Therefore we investigate more in detail what determines if they sur- vive beyond their ï¬?rst year. The most striking ï¬?nding coming out of our analysis —and which could not be observed on the product-level data used by previous studies of export survival—is that exporters of similar goods to the same destination exert a positive externality on new entrants. That is, the more they are, the higher the survival probability of new entrants—although the effect is relatively small. This ï¬?nding is at ï¬?rst sight surprising, as one might expect that ex- porters of a given product to the same destination may crowd out each other, either through price competition or simply by offering more choice to buy- 17 Similarly as done for the proxy of external dependence borrowed from Rajan and Zingales, we construct the asset-tangibility variable at the product level, κp , by assigning to that product the corresponding ISIC3 value of Braun’s index. 28 ers who could them ‘hop’ from one to the other, reducing survival rates at the individual level. Strikingly, the synergy effect disappears if we measure the network as the number of ï¬?rms exporting the same product from other origin countries from our dataset. That is, the synergy effect is truly national. Various conjectures could explain our result. First, it could be driven by omitted-variable bias (e.g. supportive infrastructure at the national level or comparative advantage). We control for this by including the country’s revealed-comparative advantage index as a regressor, without altering the results. Relatedly, we follow the idea developed in various papers by Haus- mann and Klinger (2006) that product-speciï¬?c capabilities explain success in export markets and investigate if our synergy effect disappears when con- trolling for some proxies of these “production capabilitiesâ€?, which would be more likely to be driven by omitted variables. Again, our results are robust, although we also ï¬?nd that synergy effects and production capabilities appear to be substitutes for each other. Finally, we explore various conjectures drawing on information asymme- tries and access to ï¬?nance. For instance, access to credit may be easier when many exporters of the same product from the same origin simultaneously op- erate in the same destination, as larger numbers may provide signals about proï¬?tability to both new entrants as well as ï¬?nancing institutions. First, our hypothesis is that an exporter may obtain precious information through the network of competitors, potential buyers, relatives or friends involved in the same manufacturing activity and exporting to the same market. Second, our hypothesis is that an isolated exporter might have more difficulties convinc- ing the ï¬?nancial institutions that the risks she faces are manageable given the uncertain environment of export relations. If other ï¬?rms are success- fully in operation, by contrast, the ï¬?nancial institution can use the success of others as a predictor of its client’s potential. We verify these conjectures in different ways. First, we interact the synergy effect with quality hetero- geneity (proxied by the cross-country dispersion in unit values at the product level). Second, we interact it with indicators of dependence on bank ï¬?nance and asset structure (as a measure of the scope for moral hazard). In both cases, interaction terms are positive and strongly signiï¬?cant, suggesting that synergy effects are stronger in sectors where informational asymmetries are higher, and dependence on external ï¬?nance is more intense. 29 Our results are suggestive of a potential market failure if exporters fail to internalize the positive externality that they exert on new entrants. This may be taken as an argument in support of government-sponsored export promo- tion. However, policy implications should be interpreted very cautiously, as the record of export promotion in developing countries is highly uneven. In addition it may well be that exporters could internalize the externality through mutual-support professional organizations. 30 References [1] Alvarez, Roberto, and Ricardo Lopez, 2008, ‘Entry and Exit in Interna- tional Markets: Evidence from Chilean Data’, Review of International Economics 16, 692-708. [2] Albornoz, Facundo, Hector Pardo, Gegory Corcos and Emanuel Ornelas, 2010, ‘Sequential Exporting’; CEPR discussion paper 8103. [3] Bernard, Andrew, Bradford Jensen and Peter Schott, 2006, ‘Trade Costs, Firms and Productivity’, Journal of Monetary Economics 53, 917-937. [4] Besedes, Tibor, 2008, ‘A Search Cost Perspective on Formation and Duration of Trade’, Review of International Economics 16, 853-849. [5] Besedes, Tibor, and Thomas Prusa, 2006, ‘Ins, Outs and the Duration of Trade,â€? Canadian Journal of Economics 104, [6] Besedes, Tibor, and Thomas Prusa, 2006, ‘A Search Cost Perspective on Formation and Ducation of Trade’, Journal of International Economics 70, 339-358. [7] Besedes, Tibor, and Thomas Prusa, 2007, ‘The Role of Extensive and Intensive Margins and Export Growth,’, NBER Working Paper 13628. [8] Braun, Mathias (2003), ‘Financial Contractibility and Asset Hardness’; unpublished dissertation, Harvard University. [9] Brenton, Paul, Christian Saborowski, and Erik von Uexkull, 2009, ‘What determines the low survival rate of developing country exportflows?’, World Bank Policy Research Working Paper WPS 4951. [10] Brenton, Paul, and Richard Newfarmer, 2007, ‘Watching More Than the Discovery Channel: Export Cycles and Diversiï¬?cation in Development’, World Bank Policy Research Working Paper. u [11] Brenton, Paul, Denisse Pierola and Erik von Uexk¨ll, 2009, ‘The Life and Death of Trade Flows: Understanding the Survival Rates of Developing-Country Exporters’, Breaking Into New Markets: Emerg- ing Lessons for Export Diversiï¬?cation, Shaw , Newfarmer, Walkenhorst (Edts.), The World Bank. 31 [12] Brooks, E., 2006, ‘Why don’t Firms Export More? Product Quality and Colombian plants’, Journal of Development Economics 80, 160-178. o [13] Carballo, Jer´nimo, and Christian Volpe Martincus, 2008, ‘Is export promotion effective in developing countries? Firm-level evidence on the intensive and the extensive margins of exports’, Journal of International Economics 76, 89-106. e e [14] Cadot, Olivier, C´line Carr`re and Vanessa Strauss-Kahn, 2007, ‘Trade Diversiï¬?cation, Income, and Growth: What Do We Know?’, mimeo, ILO. [15] —, — and —, 2011, ‘Export Diversiï¬?cation: What’s behind the Hump?’, forthcoming, The Review of Economics and Statistics. [16] Das, Sanghamitra, Mark J. Roberts and James R. Tybout, 2007, ‘Mar- ket Entry Costs, Producer Heterogeneity and Export Dynamics’, Econo- metrica 75, 837-873. [17] Easterly, William, Ariel Reshef and Julia Schwenkenberg, 2009, ‘The Power of Exports’, World Bank Policy Research Working Paper 5081. [18] Eaton, Jonathan, Marcela Eslava, Maurice Kugler and James Tybout, 2008, ‘Export Growth in Colombia: Firm-Level Evidence’, The Orga- nization of Firms in a Global Economy, edited by Elhanan Helpman, Dalia Marin, and Thierry Verdier, Harvard University Press. [19] Eckel, Carsten, and Peter Neary, 2010, ‘Multi-Product Firms and Flexi- ble Manufacturing in the Global Economy’, Review of Economic Studies 77, 188-217. [20] Evenett, Simon J., and Anthony Venables, 2002, ‘Export Growth by De- veloping Economies: Market Entry and Bilateral Trade’ mimeo, London School of Economics. [21] Feenstra, R.C., 1996, ‘US imports, 1972-1994: Data and Concordances’, NBER Working Paper 5515. [22] Frankel, Jeffrey A., and David Romer, 1999, ‘Does Trade Cause Growth?’, The American Economic Review 89, 379-399. 32 [23] Freund, Caroline, and Denisse Pierola, 2010, ‘Export entrepreneurs: ev- idence from Peru’, World Bank Policy Research Working Paper Series 5407. [24] Gorg, Holger, Aoife Hanley and Eric Strobl, 2008, ‘Productivity effects of international outsourcing: Evidence from plant level data’, Canadian Journal of Economics 41, 670-688. [25] Hausmann, Ricardo, 2008, ‘The Other Hand: High Bandwidth Devel- opment Policy’, Harvard University, John F. Kennedy School of Gov- ernment Working Paper rwp08-060. [26] Hausmann, Ricardo, and Bailey Klinger, 2006, ‘Structural Transforma- tion and Patterns of Comparative Advantage in the Product Space’, Harvard University, John F. Kennedy School of Government Working Paper rwp06-041. [27] Hausmann, Ricardo, Jason Hwang and Dani Rodrik, 2007, ‘What you export matters’, Journal of Economic Growth 12, 1-25. [28] Hausmann, Ricardo, and Dani Rodrik, 2003, ‘Economic Development as Self-Discovery’, Journal of Development Economics 72, 603-633. [29] Hidalgo, Cesar, Bailey Klinger, A.L. Barabasi and Ricardo Hausmann, 2007, ‘The Product Space Conditions the Development of Nations’, Sci- ence, 317(5837): 482-487. [30] Iacovone, Leonardo and Beata Javorcik, 2010, ‘Multi-Product Ex- porters: Product Churning, Uncertainty and Export Discoveries’, forth- coming, Economic Journal. [31] Lederman, Daniel, Marcelo Olarreaga and Lucie Payton (2010), ‘Ex- port Promotion Agencies: Do They Work?’; Journal of Development Economics 91, 257-265. [32] Molina, Ana Cristina, and Marco Fugazza, 2009, ‘The determinants of trade survival’, HEI Working Paper 5. [33] —, Maurizio Bussolo and Leonardo Iacovone, 2010, ‘The DR-CAFTA and the extensive margin : a ï¬?rm-level analysis,’ World Bank Policy Research Working Paper Series 5340. 33 [34] Nitsch, Volker, 2009, ‘Die Another Day: Duration in German Import Trade’, CESifo Working Paper Series 2085. [35] Obashi, Ayako, 2010, ‘Stability of Production Networks in East Asia: Duration and Survival of Trade’, Japan and the World Economy 22, 21-30. [36] Rajan, Raghuram, and Luigi Zingales (1998), ‘Financial Dependence and Growth’; American Economic Review 88, 559-586. [37] Rauch, James E., 1999, ‘Networks Versus Markets in International Trade’, Journal of International Economics 48, 7-35. [38] — and Joel Watson, 2003, ‘Starting small in an Unfamiliar Environ- ment’, International Journal of Industrial Organization 21, 1021-1042. [39] World Bank (2009), ‘World Bank Indicators Database’. 34 6 Appendix 1 This survey was conducted over a sample of exporters randomly drawn from the customs data in each country, after applying some pre-established guide- lines that took into account the following criteria: • exporting status of the ï¬?rm, • its size, • its location • the economic sector (at the 2-digit level of the HS Code) In particular, all the exporters in each country were classiï¬?ed in four groups according to the evolution of their exporting status: a) regular ex- porters are those exporters with consecutive exports until 2008 (last year covered by the customs data in all four countries), b) past exporters are the exporters who were exporting consecutively for at least two years and then exited the market before 2008, c) intermittent exporters are those who ex- ported erratically during the period included in the sample and ï¬?nally, d) new exporters are those exporters who appear for the ï¬?rst time in the sample in 2008. Over 200 ï¬?rms were contacted in each country; however, due to low coop- eration and identiï¬?cation problems with some of the ï¬?rms, the ï¬?nal sample by country and exporting group is as follows: Country Intermit New Reg Past Total Mwi 9 9 59 14 91 Mli 10 18 48 22 98 Sen 15 25 43 39 122 Tza 15 7 48 14 84 Total 49 59 198 89 395 35 Table 6: Survey Responses on Importance of Networks (in %) Question 1: First time exporters: How was the contact with the ï¬?rst client made? MLI MWI SEN TZA All Research online 14 11 24 35 21 Third party contact 73 68 77 51 67 Competitors’ network 8 12 24 11 14 Trade Fair 20 12 19 34 21 Export Promotion Agency 12 11 5 13 10 Exporters’ Association 9 7 8 8 8 Another channel 16 24 5 11 14 Question 2: If the company looked for its buyers, how did it approach them? Research online 26 31 29 41 32 Third party contact 74 72 76 57 70 Competitors’ network 19 18 23 21 20 Trade Fair 40 35 28 52 39 Export Promotion Agency 18 19 11 21 17 Exporters’ Association 14 5 6 17 11 Another channel 10 20 15 6 13 Question 3: If the buyers approached the company, how did they approach it? Company’s website 22 30 29 53 33 Old clients of the company 25 28 33 32 30 Third-party contacts 62 75 75 66 69 Competitors’ network 14 28 21 26 22 Trade Fair 34 33 20 55 35 Export Promotion Agency 18 21 7 25 18 Another channel 9 22 15 8 13 Question 4: How did the opportunity to export a new product come about? An existing buyer approached the company 54 46 50 68 54 The company saw saw demand in a buyers’ market 33 46 50 56 46 The company saw successful competitors 17 27 13 32 22 Success with selling the product domestically 38 42 44 68 48 Through a third party 46 23 25 35 32 Any other type of opportunity? 17 19 13 6 14 36