Trade, Finance, and Endogenous Firm Heterogeneity

Trade, Finance, and Endogenous Firm Heterogeneity Abstract We study how financial frictions affect firm-level heterogeneity and trade. We build a model in which productivity differences across monopolistically competitive firms are endogenous and depend on investment decisions at the entry stage. By increasing entry costs, financial frictions lower the exit cutoff and hence the value of investing in bigger projects with more dispersed outcomes. As a result, financial frictions make firms smaller and more homogeneous, and hinder the volume of exports. Export opportunities, instead, shift expected profits to the tail and increase the value of technological heterogeneity. We test these predictions using comparable measures of sales dispersion within 365 manufacturing industries in 119 countries, built from highly disaggregated US import data. Consistent with the model, financial development increases sales dispersion, especially in more financially vulnerable industries; sales dispersion is also increasing in measures of comparative advantage. These results help explaining the effect of financial development and factor endowments on export sales. (JEL: F12, F14) The editor in charge of this paper was Claudio Michelacci. Acknowledgments: We thank the Editor (Claudio Michelacci), four anonymous Referees, Johannes Boehm, Harald Fadinger, Gene Grossman, Elhanan Helpman, Matthias Kehrig, Ferdinando Monte, Fabrizio Zilibotti, and seminar participants at the International Monetary Fund, University of Porto, University of Nottingham, Università della Svizzera Italiana, Queen Mary University of London, Universitat Autònoma de Barcelona, CREI, 2017 CEPR European Research Workshop in International Trade, 2016 “Winter Symposium in Economics and Finance” (Milan), 2016 CEPR Macroeconomics and Growth Workshop (London), MadMac workshop on Growth and Development (Madrid) and the 2016 SED Annual Meeting (Toulouse) for comments. We acknowledge financial support from the Barcelona GSE, the Spanish Ministry of Economy, and Competitiveness (ECO2014-55555-P and ECO2014-59805-P), and the Catalan AGAUR (2014-SGR-546). Bonfiglioli, Crinò and Gancia are Research Fellows at CEPR. Crinò is Research Affiliate at CESifo 1. Introduction Why firms differ so much in sales and productivity, and how these differences vary across industries, countries, and time, are among the most pressing questions across the fields of international trade, macroeconomics, and economic development. Although the literature on firm heterogeneity has exploded since the late 1990s, the existing evidence is often limited to few countries or sectors and theoretical explanations are still scarce.1 One well-established stylized fact is that average firm size increases with per capita income and, according to recent work, so does its dispersion.2 Since financial markets are much less developed in poor countries, a plausible conjecture is that credit frictions may play a role at shaping firm heterogeneity. Financial constraints have also been found to restrict significantly international trade.3 Since export participation is concentrated among the most productive firms, it is then plausible to conjecture that financial frictions may hinder trade by affecting the firm size distribution. The goal of this paper is to shed new light on these hypotheses. We start by introducing financial frictions in a model where productivity differences across firms are endogenous and depend on investment decisions at the entry stage. In most of the literature, credit frictions distort the allocation of resources among existing firms who differ in productivity for exogenous reasons. Instead, we consider the problem of financing an up-front investment, namely, innovation, which affects the variance of the possible realizations of technology. This approach has several advantages. First, financial frictions at the entry stage are highly relevant in practice, especially when financing an investment with uncertain returns. Second, it allows us to highlight some of the economic decisions that shape the equilibrium degree of firm heterogeneity. Next, we take the model to the data. Starting from highly disaggregated product-level US import data, we show how to build comparable measures of sales dispersion across a large set of countries, sectors, and time and use them to test the model. With this uniquely rich dataset, we provide new evidence that financial frictions compress the sales distribution, which in turn has a significant negative effect on export volumes. We now describe more in detail what we do. The first step is to develop a model in which technology differences across firms depend on investment decisions at the entry stage. Our point of departure is a multi-sector and multi-country static version of Melitz (2003), which is the workhorse model of trade with heterogeneous firms. As it is customary, firms draw productivity upon paying an entry cost and exit if they cannot profitably cover a fixed production cost. As in Bonfiglioli, Crinò, and Gancia (2017), however, firms can affect the distribution from which technology is drawn. In particular, investments in bigger innovation projects are associated to more dispersed realizations of productivity. As a result, the ex post degree of heterogeneity in a sector depends on the ex ante choice of the entry investment. In this framework, we introduce frictions, which raise the cost of financing entry in financially vulnerable sectors, and both variable and fixed costs of selling to foreign markets. A key insight of the model is that the possibility to exit insures firms from bad realizations and increases the value of drawing productivity from a more dispersed distribution. This generates two main predictions. First, financial frictions lower the equilibrium degree of heterogeneity in a sector. The intuition for this result is that financial frictions reduce entry, which in turn lowers the minimum productivity needed to survive. But a higher surviving probability lowers the value of drawing from a more dispersed distribution.4 We then show that, by making firms smaller and more homogeneous, financial frictions hinder the volume of exports both along the intensive and the extensive margin, with a stronger effect in sectors that are more financially vulnerable. Second, as in Bonfiglioli et al. (2017), export opportunities, by shifting expected profits to the tail and raising the exit cutoff, increase the value of drawing productivity from a more dispersed distribution thereby generating more heterogeneity. At a first glance, this mechanism seems to capture important real-world phenomena. It is widely documented that entry barriers, financial frictions, and trade costs allow unproductive firms to survive. Limited export opportunities also lower the payoffs of successful products. Our theory suggests that these frictions have additional effects on incentives: they discourage investment in large-scale projects and the use of advanced technologies with high upside potential. As a consequence, in equilibrium firms are small, the resulting distribution of revenue has a low dispersion, and there are few exporters. This picture does not seem far from the reality in many financially underdeveloped countries. Our next step is to test these predictions using highly disaggregated data. To guide the analysis, we use the model to show how the parameter measuring firm heterogeneity at the sector level can be computed from the dispersion of sales across products from any country and industry to a given destination market. We then empirically assess the predictions of the model using extremely detailed data on US imports of roughly 15,000 (HS 10-digit) products from 119 countries and 365 manufacturing industries over 1989–2006. Starting from almost 4 million observations at the country–product–year level, we measure sales dispersion for each country, industry and year as the standard deviation of log exports across products. We thereby obtain a unique dataset, which includes more than 230,000 comparable measures of sales dispersion across countries and manufacturing industries, over a period that spans two decades. The dataset we use has several advantages and some limitations. For our purposes, its most important feature is that it allows us to construct measures of the dispersion of sales to a single market for a large set of countries that differ greatly in the level of financial development and for a large set of sectors that differ greatly in financial vulnerability. This would be hard to do using firm-level data, which are unavailable for many countries and often do not separate sales by destination.5 Moreover, although in the model firms and products coincide, it is not a priori obvious whether its predictions should be tested preferably using firm- or product-level data. In practice, however, measures of heterogeneity across firms or products are highly correlated, as we show using US data. The impossibility to control for firm characteristics is also mitigated by the fact that the mechanism in the model works through an adjustment of the exit cutoff that affects indiscriminately all firms in a sector and by the inclusion of a host of fixed effects. After documenting some interesting statistics on how sales dispersion varies across countries, industries, and time, we study how it depends on financial development and export opportunities. Following a large empirical literature, we identify the effect of financial frictions exploiting cross-country variation in financial development and cross-industry variation in financial vulnerability (Rajan and Zingales 1998; Manova 2013). Our main result is that, consistent with our model, financial development increases sales dispersion, especially in more financially vulnerable industries. Export opportunities, proxied by country-sector measures of comparative advantage as in Romalis (2004), also make the distribution of sales more spread out. These results are robust to controlling for the number of exported products, to the inclusion of country–year and industry–year fixed effects, to the level of industry aggregation, to various changes in the sample such as excluding small exports, to the use of alternative proxies for financial frictions and financial vulnerability, to alternative estimation approaches and measures of sales dispersion, and to instrumenting financial development with historical conditions of countries. We also find that sales dispersion is important for explaining trade flows and the well-known effect of financial frictions on exports (Beck 2002; Manova 2013). Finally, we provide some more direct evidence on the mechanism at work in the model, which operates through changes in the innovation strategies of firms. To this end, we show that our proxies for financial frictions at the country–sector–year level are a significant determinant of major innovations, as measured by the number of utility patents applied for at the US Patent Office, computed separately for each foreign country, industry and application year. In turn, patent applications are positively correlated with sales dispersion, as in our theory. Our model of endogenous firm heterogeneity has been developed in this paper and in Bonfiglioli et al. (2017). In the latter, we abstract from financial frictions and draw implications for wage inequality. We also provide evidence that export opportunities increase firm heterogeneity, innovation, and wage inequality. In the present paper, instead, we introduce financial frictions and extend the model to multiple asymmetric countries. This allows us to derive novel empirical implications. Regarding the evidence, the two papers use completely different data and approaches. In Bonfiglioli et al. (2017) we use US firm-level data; here instead, we use non-US product-level data. Remarkably, the measures of sector-level heterogeneity computed in the different data sets are comparable in magnitude, display similar trends and have similar correlations with export opportunities. Compared to Bonfiglioli et al. (2017), an important advantage of the data used in this paper is also that it enables us to document new empirical patterns. Among these, we extend to a much broader sample the little-known fact that the dispersion of firm size increases with per capita income. For comparison, Poschke (2015) uses survey data from less than 50 countries, and Bartelsman et al. (2009) uses data for 24 countries only. We also document that the dispersion of sales has increased on average by 6% between 1989 and 2006. More importantly, we establish the result that financial frictions have a statistically significant and quantitatively large effect on sales dispersion. For instance, our estimates implies that the average increase in private credit over the sample period could explain 59% of the observed increase in sales dispersion. Besides the evidence in these two papers, our theory accords well with a number of additional observations. For instance, several papers show evidence suggesting that differences in productivity across firms appear to be related to investment in new technologies (e.g., Dunne et al. 2004; Faggio, Salvanes, and Van Reenen 2010; Doraszelski and Jaumandreu 2013). Moreover, the emphasis on the role of entry and product innovation is empirically relevant, given that every year about 25% of consumer goods sold in US markets are new (Broda and Weinstein 2010). Furthermore, as shown for instance in Cabral and Mata (2003), there is already considerable heterogeneity among new firms. The trade-off between large/small innovation projects with more/less variable outcomes seems also to describe well some important aspects of the innovation strategies pursued by different firms. For instance, designing and assembling a new variety of laptop PCs, which mostly requires the use of established technologies, is safer and less costly than developing an entirely new product, such as the iPad. Yet, Apple’s large investment was rewarded with the sale of more than 250 million units over a period of five years only, whereas the sales of manufacturers of traditional computers stagnated. Nevertheless, the choice between innovations differing in the variance of outcomes and the implications for firm heterogeneity has received so far little attention in the literature. An exception is Caggese (2015), who has developed a dynamic model where firms with low profitability invest in radical, high-risk, innovation because they have less to lose in case innovation fails.6 Financial frictions increase the rents of these firms and hence reduce their willingness to take on risk. Our mechanism differs in that it applies to all firms and does not depend on their profit level. Our focus is also entirely different: we study and test the implications for the dispersion of sales and the volume of trade. From a theoretical perspective, the effect of financial frictions on the size distribution of firms is not a priori obvious. Existing models do not focus explicitly on the dispersion of the size distribution and often study how credit constraints distort the allocation of resources across firms.7 Whether these misallocations amplify or dampen the dispersion of sales depends on several factors, such as whether credit constraints bind more for less or more productive firms. For instance, Arellano, Bai, and Zhang (2012) argue that small firms are more likely to be inefficient in scale because they are closer to their borrowing limits, although other results are possible. Hence, the effect of financial frictions on sales dispersion is ultimately an empirical question on which this paper sheds some new light. Our focus on entry is motivated by Midrigan and Xu (2014), who find that financial frictions can distort entry more than the allocation of resources between existing firms. They find that misallocation generated by financial frictions can be small because more efficient producers accumulate internal funds over time and can grow out of their borrowing constraints (see also the survey by Buera, Kaboski, and Shin 2015). Our evidence that financial frictions hinder patenting, which in turn is associated with lower sales dispersion, lends additional support to our modeling assumptions. This paper is closely related to the literature on trade with heterogenous firms. In particular, our findings help understand the role of financial frictions in affecting export decisions. The fact that financial constraints reduce exports disproportionately more than domestic production has been documented in a series of recent contributions (see Chor and Manova 2012; Manova 2013; Paravisini et al. 2015 and all the papers surveyed in Foley and Manova 2015). This literature has provided robust evidence that financial development hinders trade and that this effect is stronger in financially vulnerable sectors. Yet, the exact mechanism through which this happens is still subject to investigation.8 Existing explanations typically assume credit frictions to bind more for exports than for domestic sales. But the foundation for this assumption is not entirely uncontroversial, especially since export volumes are overwhelmingly driven by large firms that are usually less financially constrained (e.g., Beck, Demirgüç-Kunt, and Maksimovic 2005). Our model overcomes these difficulties. Through their effect on the exit cutoff at the sector level, financial frictions affect all firms, including those that are not constrained. Their negative effect on the introduction of new products is also easy to justify, because it is well-known that financing R&D-intensive projects by means of external credit is subject to relevant informational frictions (e.g., Hall and Lerner 2010). Moreover, in our model there is no need to impose any asymmetry on the financial needs of domestic or export activities. Finally, this paper is also part of the broader and growing literature studying the effect of trade on technology choices, such as Bustos (2011).9 We depart from previous works by focusing on the dispersion rather than the level of productivity and studying a mechanism that does not rely on scale effects. Yet, our result that entry can foster the use of better technologies accords well with recent findings that pro-competitive forces appear to have increased firm-level productivity (Khandelwal and Topalova 2011). The remainder of the paper is organized as follows. In Section 2, we build a model in which differences in the variance of firm-level outcomes originate from technological choices at the entry stage and show that financial development and export opportunities generate more heterogeneity in equilibrium. Section 3 derives a number of predictions on how observable measures of within-sector heterogeneity at the country–industry level depend on export opportunities and financial development and how firm heterogeneity affects the margins of trade. Section 4 tests these predictions. Section 5 concludes. 2. The Model We build a multi-sector, multi-country, static model of monopolistic competition between heterogeneous firms along the lines of Melitz and Redding (2014). After paying an entry cost, firms draw their productivity from some distribution and exit if they cannot profitably cover a fixed cost of production. As in Bonfiglioli et al. (2017), we allow the variance of the productivity draws to depend on investment decisions. We then introduce a financial friction between firms, who must borrow to pay the entry investment, and external investors, and study how it affects firm-level heterogeneity. 2.1. Preferences and Demand Country o is populated by a unit measure of risk-neutral households of size Lo. Preferences over consumption of goods produced in I industries are \begin{equation*} U_{o}=\prod _{i=1}^{I}C_{oi}^{\beta _{i}},\ \ \beta _{i}>0,\ \ \sum _{i=1}^{I}\beta _{i}=1. \end{equation*} Each industry i ∈ {1, …, I} produces differentiated varieties and preferences over these varieties take the constant elasticity of substitution form: \begin{equation*} C_{oi}=\left[ \int _{\omega \in \Omega _{oi}}c_{oi}(\omega) ^{\frac{\sigma _{i}-1}{\sigma _{i}}}\text{d}\omega \right] ^{\frac{\sigma _{i}}{\sigma _{i}-1}}, \, \sigma _{i}>1, \end{equation*} where coi(ω) is consumption of variety ω, Ωoi denotes the set of varieties available for consumption in country o in sector i, and σi is the elasticity of substitution between varieties within the industry i. We denote by poi(ω) the price of variety ω in industry i and by Poi the minimum cost of one unit of the consumption basket Coi: \begin{equation*} P_{oi}=\left[ \int _{\omega \in \Omega _{oi}}p_{oi}(\omega) ^{1-\sigma _{i}}\text{d}\omega \right] ^{1/(1-\sigma _{i})}. \end{equation*} Then, demand for a variety can be written as \begin{equation*} c_{oi}(\omega) =\frac{\beta _{i}E_{o}P_{oi}^{\sigma _{i}-1}}{p_{oi}(\omega) ^{\sigma _{i}}}, \end{equation*} where Eo is expenditure available for consumption. 2.2. Industry Equilibrium We now focus on the equilibrium of a single industry i ∈ {1, …, I}. In each industry, every variety ω is produced by monopolistically competitive firms that are heterogeneous in their labor productivity, φ. Since all firms with the same productivity behave symmetrically, we index firms by φ and we identify firms with products. We first describe the technological and financial constraints faced by the typical firm. A firm is run by a manager, who owns the idea needed to produce a given variety. To implement the idea, the manager must choose how much to invest in innovation at the entry stage. Managers have no wealth so that the entry cost, which is borne up-front, must be financed by external capital. Since innovation is not pledgeable and has uncertain returns, it is financed through an equity-like contract according to which investors acquire claims on profit. Once the entry investment is paid, the manager draws productivity from a Pareto distribution, whose shape parameter depends on the size of the investment.10 Next, the firm faces standard production and pricing decisions. There is a fixed cost of selling in a given market and a variable iceberg cost of exporting. Finally, investors need to be paid. We assume that with probability δo the manager returns the profit πi to investors. With probability (1 − δo), instead, the manager can misreport the value of production and repay only a fraction κi < 1 of profit. The parameter κi is an inverse measure of financial vulnerability that, following Rajan and Zingales (1998) and Manova (2013), is assumed to vary across industries for technological reasons. The parameter δo captures instead the strength of financial institutions and is associated to the level of financial development of the country. 2.2.1. Production, Prices, and Profit We solve the problem backward. At the production stage, the manager will choose the price and in which markets to sell (if any) so as to maximize profit. As it is customary, the equilibrium price of a firm with productivity φ serving market d from country o is \begin{equation*} p_{\mathit{doi}}(\varphi) =\frac{\sigma _{i}}{\sigma _{i}-1}\frac{\tau _{\mathit{doi}}w_{o}}{\varphi }, \end{equation*} where wo is the wage in country o and τdoi ≥ 1 is the iceberg cost of shipping from o to d (with τooi = 1) in industry i. Revenues earned from selling to destination d are \begin{equation*} r_{\mathit{doi}}(\varphi )=\beta _{i}E_{d}P_{di}^{\sigma _{i}-1}p_{\mathit{doi}}(\varphi) ^{1-\sigma _{i}}. \end{equation*} Profit earned in destination d is a fraction σi of revenue minus the fixed cost of selling in market d, wofdoi. Hence, \begin{equation} \pi _{\mathit{doi}}(\varphi) =A_{di}\left( \frac{\varphi }{\tau _{\mathit{doi}}w_{o}}\right) ^{\sigma _{i}-1}-w_{o}f_{\mathit{doi}}, \end{equation} (1) where the term $$A_{di}=\beta _{i}E_{d}P_{di}^{\sigma _{i}-1}[( \sigma _{i}) ^{\sigma _{i}}(\sigma _{i}-1)^{1-\sigma _{i}}] ^{-1}$$ captures demand conditions in the destination market. The firm will not find it profitable to serve market d whenever its productivity is below the cutoff \begin{equation} \varphi _{\mathit{doi}}^{\ast }=\tau _{\mathit{doi}}w_{o}\left( \frac{w_{o}f_{\mathit{doi}}}{A_{di}} \right) ^{1/(\sigma _{i}-1)}, \end{equation} (2) corresponding to $$\pi _{\mathit{doi}}\left( \varphi _{\mathit{doi}}^{\ast }\right) =0$$. 2.2.2. Entry Stage We now consider the entry stage. As in Melitz (2003), firms pay a sunk innovation cost to be able to manufacture a new variety with productivity drawn from some distribution with c.d.f. Goi(φ). Hence, combining the pricing and exit decision, we can write ex ante expected profit from market d: \begin{equation} \mathbb {E}[\pi _{\mathit{doi}}] =\int _{0}^{\infty }\pi _{\mathit{doi}}(\varphi) \text{d}G_{oi}(\varphi) =w_{o}f_{\mathit{doi}}\int _{\varphi _{\mathit{doi}}^{\ast }}^{\infty }\left[ \left( \frac{\varphi }{\varphi _{\mathit{doi}}^{\ast }}\right) ^{\sigma _{i}-1}-1\right] \text{d} G_{oi}(\varphi) , \end{equation} (3) where the last equation makes use of (1) and (2). Expected profit from selling in all potential markets is $$\mathbb {E}[\pi _{oi}] =\sum _{d}\mathbb {E}[\pi _{\mathit{doi}}]$$. We depart from the canonical approach by making the distribution Goi(φ) endogenous. To this end, we follow Bonfiglioli et al. (2017) in using a simple model of investment in new products generating a Pareto distribution for φ with mean and variance that depend on firms’ decisions. The model formalizes the idea that firms can choose between smaller projects with less variable returns and larger projects with more spread-out outcomes. In particular, in order to enter, the manager of the firm can choose between a menu of projects of size soi ∈ (0, 1] that allows the firm to manufacture a new variety with productivity drawn from the distribution \begin{equation} G_{oi}(\varphi) =1-\left( \frac{\varphi _{\min }}{\varphi } \right) ^{1/v_{oi}}, \end{equation} (4) where \begin{equation} v_{oi}=\frac{s_{oi}}{\alpha _{i}\sigma _{i}}, \, \alpha _{i}>1. \end{equation} (5) Hence, by choosing the size soi of the project, the firm is selecting to draw φ from a family of Pareto distributions differing in the parameter voi = soi/(αiσi) (i.e., the inverse of the shape parameter).11 The choice of voi affects positively the dispersion of φ. To see this, note that the standard deviation of the log of φ is equal to voi, which can therefore be interpreted as an index of dispersion of the distribution. At the same time, voi also affects the expected value of φ, which is equal to φmin (1 − voi)−1.12 This positive relationship between mean and variance is realistic: Bonfiglioli et al. (2017) find strong evidence of a positive correlation between the average and the dispersion of sales across US firms. Yet, as we show in the Appendix, our main results hold in an alternative model in which firms can choose between distributions that are a mean-preserving spread. How is the initial entry investment determined in equilibrium? To answer this question we turn to the cost of entry. First, we assume that the entry cost, expressed in units of labor, is an increasing and convex function of the investment soi, satisfying the Inada-like condition that the cost tends to infinity as soi approaches the maximum size of one.13 Since voi = soi/(αiσi), the problem of choosing soi can be reformulated as one of choosing voi at the cost woF(voi), with F΄(voi) > 0, F″(voi) > 0, F(0) = 0 and $$\lim _{v_{oi}\rightarrow 1/(\alpha _{i}\sigma _{i})}F(v_{oi})=$$ ∞. Next, recall that woF(voi) must be financed externally and that with probability (1 − δo) managers can hide a fraction (1 − κi) of profit. For simplicity, we normalize the outside option of both managers and investors to zero. Hence, investors expect to be repaid πoi with probability δo and κiπoi, with probability (1 − δo), and competition for funds between managers implies that voi is set so as to maximize the expected returns of investors: \begin{equation} \max _{v_{oi}}\left\lbrace \mathbb {E}[\pi _{oi}] -w_{o}\lambda _{oi}F(v_{oi}) \right\rbrace , \end{equation} (6) where λoi ≡ [δo + (1 − δo)κi]−1 > 1 captures the additional cost of financing the entry investment in the presence of financial frictions (κi < 1 and δo < 1). Finally, free-entry implies that investors must break even, $$\mathbb {E}[\pi _{oi}] =w_{o}\lambda _{oi}F(v_{oi})$$, which is also their (binding) participation constraint.14 To solve (6), we use Goi(φ) to express ex ante expected profits (3) as a function of voi: \begin{equation*} \mathbb {E} [ \pi _{oi}] =\frac{(\sigma _{i}-1)w_{o}}{1/v_{oi}-(\sigma _{i}-1)}\left( \frac{\varphi _{\min }}{\varphi _{ooi}^{\ast }}\right) ^{1/v_{oi}}\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}, \end{equation*} where \begin{equation} \rho _{\mathit{doi}}\equiv \frac{\varphi _{ooi}^{\ast }}{\varphi _{\mathit{doi}}^{\ast }}=\tau _{\mathit{doi}}^{-1}\left( \frac{A_{di}}{f_{\mathit{doi}}}\frac{f_{ooi}}{A_{oi}}\right) ^{1/(\sigma _{i}-1)} \end{equation} (7) is a measure of export opportunities in destination d. In particular, in a given industry i, $$\rho _{\mathit{doi}}^{1/v_{oi}}\in (0,1)$$ is the fraction of country o firms selling to market d. To make sure that the maximand in (6) is concave, the cost function F(voi) must be sufficiently convex. In particular, we define the elasticities of the entry cost and of profit as ηF(voi) ≡ voiF(voi)΄/F(voi) and $$\eta _{\pi }(v_{oi}) \equiv \partial \ln \mathbb {E}[\pi _{oi}] /\partial \ln v_{oi}$$, respectively. We then assume $$\eta _{F}^{\prime }(v_{oi})>\eta _{\pi }^{\prime }(v_{oi})$$. The first order condition for an interior voi is \begin{eqnarray} \frac{\mathbb {E} [ \pi _{oi}] }{v_{oi}}\left[ \frac{1}{1-v_{oi}(\sigma _{i}-1)}+\ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}+\frac{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}\right] \nonumber\\ =w_{o}\lambda _{oi}F^{\prime }(v_{oi}) .\qquad \qquad \qquad \qquad \qquad\qquad \qquad \qquad \end{eqnarray} (8) The left-hand side of (8) is the marginal benefit of increasing voi, whereas the right-hand side is its marginal cost. The terms in brackets, equal to the elasticity of expected profit to voi, capture the fact that a higher v increases expected profits for various reasons. First, it raises the unconditional mean of productivity draws. Second, it increases the probability of drawing a productivity above the cutoff needed to sell to any destination. Third, it increases the relative gains from a high realization of φ when the profit function is convex, that is, when σi > 2 (as can be seen from equation (1)). Yet, both $$\mathbb {E}[\pi]$$ and $$\varphi _{ooi}^{\ast }/\varphi _{\min }$$ are endogenous. To solve for them, we impose free entry, requiring that ex ante expected profit be equal to the entry cost: $$\mathbb {E}[ \pi _{oi}] =w_{o}\lambda _{oi}F(v_{oi})$$. Replacing this into the first-order condition (8), we obtain the following expression: \begin{equation} \frac{1}{1-v_{oi}(\sigma _{i}-1)}+\ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}+\frac{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}=\frac{v_{oi}F^{\prime }(v_{oi}) }{F(v_{oi}) }, \end{equation} (9) where the left-hand side is the elasticity of expected profit, ηπ(voi), whereas the right-hand side is the elasticity of the entry cost, ηF(voi). Under the assumptions that $$\eta _{F}^{\prime }(v_{oi})>\eta _{\pi }^{\prime }(v_{oi})$$ and $$\lim _{v_{oi}\rightarrow 1/\alpha _{i}\sigma _{i}}\eta _{F}(v_{oi})=$$ ∞, there is a unique interior voi satisfying (9). Finally, we need to substitute for the equilibrium exit cutoff for productivity, which is pinned down again by the free-entry condition: \begin{equation} \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}= \frac{\sigma _{i}-1}{1/v_{oi}-(\sigma _{i}-1)}\frac{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}{\lambda _{oi}F(v_{oi}) }. \end{equation} (10) Note that the exit cutoff is decreasing in the cost of financing, λoi: higher financing costs deter entry, thereby reducing the degree of competition and the minimum productivity required to break even. In addition, the exit cutoff is increasing in export opportunities, ρdoi: as it is well-known since Melitz (2003), export opportunities increase profit for more productive firms thereby inducing more entry and making survival more difficult.15 After replacing the cutoff in (9), it can be proved that, for given fixed costs, the left-hand side, that is, the elasticity of expected profit, is increasing in export opportunities and decreasing in the cost of financing. Note also that, in an interior equilibrium, all parameters raising ηπ(voi) also increase the optimal voi. We are then in the position to draw predictions on the equilibrium dispersion of productivity, which is Pareto with minimum $$\varphi _{ooi}^{\ast }$$ and shape parameter 1/voi. Hence, the log of φ is exponential with standard deviation equal to voi.16 Using this result, we can show how the equilibrium dispersion of firm productivity varies across sectors, countries and destination markets as described by Proposition 1. Proposition 1. Assume that the solution to (6) is interior. Then, the equilibrium dispersion of firm productivity in sector i, as measured by the standard deviation of the log of φ, is increasing in export opportunities, ρdoi, and in the financial development of the country of origin, δo, especially in sectors with high financial vulnerability (low κi). \begin{equation*} \frac{\partial v_{oi}}{\partial \rho _{\mathit{doi}}}>0; \frac{\partial v_{oi} }{\partial \delta _{o}}>0; \frac{\partial ^{2}v_{oi}}{\partial \delta _{o}\partial \kappa _{i}}<0. \end{equation*} Proof. See Appendix A key insight to understand the results in Proposition 1 is that the possibility to exit (or, more in general, to discard failed innovations) insures firms from bad realizations and increases the value of drawing productivity from a more dispersed distribution. This generates two main predictions. First, financial frictions lower the equilibrium degree of heterogeneity in a sector. The intuition is as follows. Financial frictions raise the cost of investment and reduce entry, especially in financially vulnerable sectors. This lowers the minimum productivity needed to survive, which in turn reduces the value of drawing productivity from a more dispersed distribution. Second, as in Bonfiglioli et al. (2017), export opportunities, by shifting expected profits to the tail and raising the exit cutoff, increase the value of drawing productivity from a more dispersed distribution thereby generating more heterogeneity. Note that the problem studied so far is simplified by the assumption that all firms entering a given sector in a given country are ex ante identical and therefore end up choosing the same voi. An alternative case would be one in which firms differ in their exposure to financial frictions before the innovation decision is made. Even though more complex, this case is interesting because in reality new products are introduced by firms, some of which have more internal funds (e.g., older and larger incumbents) than others (e.g., startups and small firms).17 To see how this ex ante heterogeneity affects our main results, in the Appendix we sketch a version of the model in which in each industry there is an exogenous mass of entering firms that are not subject to the financial friction (i.e., for them λoi = 1). We then show that, while financially constrained firms behave exactly as in the baseline model, unconstrained firms choose a higher voi. Yet, the choice of voi by any firms is still affected by the exit cutoff as in the baseline model. The additional difficulty is that, with different firms making heterogeneous choices, the overall productivity distribution is not Pareto anymore. Nevertheless, its dispersion can still be characterized analytically using the Theil index, which can be decomposed into within- and between-group components. Factoring in the compositional effects, we then show that the average dispersion in an industry is increasing in the exogenous fraction of financially unconstrained firms, the more so the higher the index of financial vulnerability λoi. Hence, adding ex ante heterogeneity does not alter the main predictions of the model. On the contrary, it suggests that the financial vulnerability of a sector may also be proxied by firm characteristics, such as average size or age, that typically correlate with the presence of financial constraints. 3. Exports, Finance, and Firm Heterogeneity We now derive a number of predictions on how observable measures of within-sector heterogeneity at the country–industry level depend on export opportunities and financial development. We also study how heterogeneity affects the volume of exports at the country–industry level. These predictions will be tested empirically in the next section. 3.1. Sales Dispersion per Destination Market Revenue from market d of firms from country o operating in sector i is a power function of productivity, $$r_{\mathit{doi}}(\varphi )=r_{\mathit{doi}}(\varphi _{\mathit{doi}}^{*})( \varphi /\varphi _{\mathit{doi}}^{\ast }) ^{\sigma _{i}-1}$$. Then, from the properties of the Pareto distribution, it follows that rdoi(φ) is also Pareto distributed with c.d.f. $$G_{r}(r) =1-(r_{\min }/r) ^{1/(v_{oi}(\sigma _{i}-1))}$$, for r > rmin  = σiwofdoi.18 This means that the standard deviation (SD) of the log of sales in industry i is equal to voi(σi − 1), and for given demand elasticity at the sector level, σi, it is determined by voi. Hence, we can apply Proposition 1 to draw results for the determinants of sales dispersion across sectors, countries, and destination markets: Proposition 2. Assume that the solution to (6) is interior. Then, the dispersion of sales from country o to destination d in sector i, as measured by the standard deviation of the log of rdoi, is increasing in export opportunities, ρdoi, and in financial development, δo. The effect of financial development is stronger in sectors with higher financial vulnerability (low κi). \begin{equation*} \frac{\partial SD[\ln r_{\mathit{doi}}] }{\partial \rho _{\mathit{doi}}}>0; \frac{\partial SD [\ln r_{\mathit{doi}}] }{\partial \delta _{o}}>0; \frac{\partial ^{2}SD[\ln r_{\mathit{doi}}]}{\partial \delta _{o}\partial \kappa _{i}}<0. \end{equation*} Proof. This follows from Proposition 1 and from the distribution of revenues, which implies that SD[ln rdoi] = voi(σi − 1). We can also derive testable predictions regarding the effect of export opportunities on equilibrium heterogeneity. Proposition 2 shows that the dispersion of sales is higher in sectors with higher ρdoi. But how can we measure export opportunities in the data? From (7), it can be seen that ρdoi is a negative function of variable trade costs, τdoi. Hence, our results suggest that globalization, by lowering variable trade costs, increases the value of technologies with higher variance and leads to more heterogeneity. Second, there is another important determinant of export opportunities, Adi/Aoi, which captures relative demand conditions. As shown in Bernard, Redding, and Schott (2007), this term in general depends on comparative advantage. In particular, they show that, other things equal, Adi/Aoi will be higher in a country’s comparative advantage industry because profits in the export market are larger relative to profits in the domestic market in comparative advantage industries. It follows that, even if we abstract from microfounding the differences in Adi/Aoi here, we can use existing results to conclude that the exit cutoff, export opportunities, and equilibrium sales dispersion will all be higher in a country’s comparative advantage industries. 3.2. Export Volumes, Firm Heterogeneity, and Finance We now derive predictions for the volume of trade. The total value of exports to destination d from origin o in industry i can be written as \begin{equation*} X_{\mathit{doi}}=\underset{\text{mass of exporters}}{\underbrace{M_{oi}\left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\mathit{doi}}^{\ast }}\right) ^{1/v_{oi}}}}\underset{\text{export per firm}}{\cdot \underbrace{\frac{\sigma _{i}w_{o}f_{\mathit{doi}}}{1-v_{oi}(\sigma _{i}-1)}}}\text{,} \end{equation*} where Moi is the mass of country o firms operating in industry i and $$( \varphi _{ooi}^{\ast }/\varphi _{\mathit{doi}}^{\ast }) ^{1/v_{oi}}$$ is the fraction of firms exporting to destination d. We now study how firm heterogeneity affects various components of the export volume. Consider first the intensive margin. Average sales to market d per firm from country o serving that destination, denoted as xdoi, is \begin{equation*} x_{\mathit{doi}}=\frac{\sigma _{i}w_{o}f_{\mathit{doi}}}{1-v_{oi}(\sigma _{i}-1)}, \end{equation*} which is increasing in voi. The intuition for this result is that a higher voi increases the probability of drawing levels of productivity above any cutoff. Hence, it raises exporters’ average productivity and revenue from any destination market. Interestingly, note also that, for given voi, average export per firm does not depend on the variable trade cost, τdoi, due to a compositional effect. A fall in τdoi induces existing exporters to export more. However, it also induces entry into exporting of less productive firms, which export smaller quantities. The combination of Pareto productivity and constant-elasticity-of-substitution demand functions implies that these two effects cancel out. Although this is certainly a special result, even in more general models these two effects will tend to offset each other. In our model, however, τdoi affects exports per firm through an additional channel: by increasing export opportunities, a lower variable trade cost induces firms to invest in technologies with a higher v, which, as explained previously, imply larger average exports per firm. Consider then the extensive margin of trade. The fraction of country-o firms exporting to market d in industry i can be expressed as \begin{equation*} \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\mathit{doi}}^{\ast }}\right) ^{1/v_{oi}}=\left[ \tau _{\mathit{doi}}\left( \frac{f_{\mathit{doi}}}{f_{ooi}}\frac{A_{oi}}{A_{di}}\right) ^{1/(\sigma _{i}-1)}\right] ^{-1/v_{oi}}, \end{equation*} where, recall, Aoi summarizes demand conditions in market o. To better isolate the effect of voi, consider the case of symmetric countries, that is, Aoi = Adi. Since $$\tau _{\mathit{doi}}( f_{\mathit{doi}}/f_{ooi}) ^{1/(\sigma _{i}-1)}>1$$ (so that not all firms export), it immediately follows that the fraction of exporters is increasing in voi. Intuitively, a higher voi increases the mass in the tail of the distribution and hence the probability that a firm is productive enough to export. In an asymmetric world, the fraction of exporters will also depend on relative demand conditions, Adi/Aoi. For example, in sectors of comparative advantage competition will tend to be tougher in the home market (higher Adi/Aoi) and more firms will export. Finally, the volume of exports from o to d relative to production for the home market is also increasing in voi, \begin{equation*} \frac{X_{\mathit{doi}}}{X_{ooi}}=\tau _{\mathit{doi}}^{-1/v_{oi}}\left( \frac{f_{\mathit{doi}}}{f_{ooi}} \frac{A_{oi}}{A_{di}}\right) ^{-1/\left[ (\sigma _{i}-1)v_{oi}\right] }\frac{f_{\mathit{doi}}}{f_{ooi}}. \end{equation*} Together with Proposition 2, these results imply that financial frictions, by lowering voi, reduce the volume of trade, average sales per exporter and the fraction of exporting firms. 4. Empirical Evidence The main result of the model is that financial development and export opportunities increase the value of investing in bigger innovation projects with more variable outcomes, thereby generating more heterogeneity across firms and a higher volume of trade. In this section, we test these predictions. We start by describing the data and the measure of sales dispersion that we will use, and documenting some new facts about how this measure varies across countries, industries, and years. Next, we study how sales dispersion responds to financial development across industries with different financial vulnerability. We then explore how sales dispersion mediates the effects of financial development and export opportunities on countries’ export flows. Finally, we use patent data to provide some evidence on the mechanism at work in the model, namely, that financial development affects sales dispersion by fostering major innovations. 4.1. Data and Measures of Sales Dispersion Our main measure of dispersion is the standard deviation of log sales in a single destination market. Besides being consistent with Proposition 2, this measure has the convenient property of being scale invariant. To construct it across countries and industries, we use highly detailed product-level data on international trade. In particular, we source data on US imports of roughly 15,000 products—defined at the 10-digit level of the harmonized system (HS) classification—from all countries in the world over 1989–2006 (Feenstra, Romalis, and Schott 2002). These data contain approximately 4 million observations at the country–product–year level.19 We map products into manufacturing industries–defined at the 4-digit level of the 1987 Standard Industry Classification (SIC)–and then construct measures of sales dispersion separately for each country–industry–year triplet. We define sales dispersion as the standard deviation of log exports across the 10-digit products exported to the United States in a given triplet. To ensure that our results are not driven by sample composition, we focus on a consistent sample of 119 countries and 365 industries for which we observe exports to the United States in all years between 1989 and 2006.20 Sales dispersion is observed for triplets that have two or more products exported to the United States. In the remaining triplets, the standard deviation of log exports is unobserved (i.e., it is missing), because either no or a single product is shipped to the American market. Since the United States is the main export destination for most countries in our sample, triplets with two or more exported products are numerous and relatively large.21 Table 1 makes this point by providing details on the structure of our data set in 2006. Note that almost 40% of triplets have at least two products exported to the United States, and that this number rises to 52% when industries are aggregated at the 3-digit level. Moreover, triplets with two or more exported products are large in terms of export value, which equals 85 (178 at the 3-digit level) million dollars on average. At the same time, Table 1 also shows that the measures of sales dispersion are generally based on a large number of products. In particular, the average triplet contains 15 (31 at the 3-digit level) products exported to the United States. Table 1. Sample composition. Country–industry pairs Number of HS-10 products Imports ($ ’000) Number % of Total number Mean Median Min. Max. Mean Median Min. Max. (a) Sample: 119 countries and 365 (4-Digit) industries. Year: 2006 All country–industry pairs 43,435 1.00 6 0 0 608 33,083 0 0 47,181,989 Pairs w/no HS-10 product exported to the United States 21,809 0.50 0 0 0 0 0 0 0 0 Pairs w/1 HS-10 product exported to the United States 4,830 0.11 1 1 1 1 563 11 0.3 201,813 Pairs w/2+ HS-10 products exported to the United States 16,796 0.39 15 7 2 608 85,393 1727 0.5 47,181,989 (b) Sample: 119 countries and 131 (3-Digit) industries. Year: 2006 All country–industry pairs 15,589 1.00 16 2 0 804 92,545 29 0 62,961,319 Pairs w/no HS-10 product exported to the United States 5,876 0.38 0 0 0 0 0 0 0 0 Pairs w/1 HS-10 product exported to the United States 1,614 0.10 1 1 1 1 797 9 0.3 115,469 Pairs w/2+ HS-10 products exported to the United States 8,099 0.52 31 12 2 804 177,972 2484 0.5 62,961,319 Country–industry pairs Number of HS-10 products Imports ($ ’000) Number % of Total number Mean Median Min. Max. Mean Median Min. Max. (a) Sample: 119 countries and 365 (4-Digit) industries. Year: 2006 All country–industry pairs 43,435 1.00 6 0 0 608 33,083 0 0 47,181,989 Pairs w/no HS-10 product exported to the United States 21,809 0.50 0 0 0 0 0 0 0 0 Pairs w/1 HS-10 product exported to the United States 4,830 0.11 1 1 1 1 563 11 0.3 201,813 Pairs w/2+ HS-10 products exported to the United States 16,796 0.39 15 7 2 608 85,393 1727 0.5 47,181,989 (b) Sample: 119 countries and 131 (3-Digit) industries. Year: 2006 All country–industry pairs 15,589 1.00 16 2 0 804 92,545 29 0 62,961,319 Pairs w/no HS-10 product exported to the United States 5,876 0.38 0 0 0 0 0 0 0 0 Pairs w/1 HS-10 product exported to the United States 1,614 0.10 1 1 1 1 797 9 0.3 115,469 Pairs w/2+ HS-10 products exported to the United States 8,099 0.52 31 12 2 804 177,972 2484 0.5 62,961,319 Notes: All statistics use product-level data on exports to the United States at the 10-digit level of the harmonized system (HS) classification (Feenstra et al. 2002). The sample consists of 119 countries that have exported to the United States in at least one industry during all years between 1989–2006. Industries are defined at the 4-digit level of the standard industrial classification (SIC) in panel (a) and at the 3-digit SIC level in panel (b); in each panel, the sample includes industries in which at least one country has exported to the United States during all years between 1989–2006. The standard deviation of log exports (used in subsequent tables) can be defined for country–industry pairs that have at least two HS-10 products exported to the United States; it is instead undefined (i.e., missing) for the other country–industry pairs. View Large Table 1. Sample composition. Country–industry pairs Number of HS-10 products Imports ($ ’000) Number % of Total number Mean Median Min. Max. Mean Median Min. Max. (a) Sample: 119 countries and 365 (4-Digit) industries. Year: 2006 All country–industry pairs 43,435 1.00 6 0 0 608 33,083 0 0 47,181,989 Pairs w/no HS-10 product exported to the United States 21,809 0.50 0 0 0 0 0 0 0 0 Pairs w/1 HS-10 product exported to the United States 4,830 0.11 1 1 1 1 563 11 0.3 201,813 Pairs w/2+ HS-10 products exported to the United States 16,796 0.39 15 7 2 608 85,393 1727 0.5 47,181,989 (b) Sample: 119 countries and 131 (3-Digit) industries. Year: 2006 All country–industry pairs 15,589 1.00 16 2 0 804 92,545 29 0 62,961,319 Pairs w/no HS-10 product exported to the United States 5,876 0.38 0 0 0 0 0 0 0 0 Pairs w/1 HS-10 product exported to the United States 1,614 0.10 1 1 1 1 797 9 0.3 115,469 Pairs w/2+ HS-10 products exported to the United States 8,099 0.52 31 12 2 804 177,972 2484 0.5 62,961,319 Country–industry pairs Number of HS-10 products Imports ($ ’000) Number % of Total number Mean Median Min. Max. Mean Median Min. Max. (a) Sample: 119 countries and 365 (4-Digit) industries. Year: 2006 All country–industry pairs 43,435 1.00 6 0 0 608 33,083 0 0 47,181,989 Pairs w/no HS-10 product exported to the United States 21,809 0.50 0 0 0 0 0 0 0 0 Pairs w/1 HS-10 product exported to the United States 4,830 0.11 1 1 1 1 563 11 0.3 201,813 Pairs w/2+ HS-10 products exported to the United States 16,796 0.39 15 7 2 608 85,393 1727 0.5 47,181,989 (b) Sample: 119 countries and 131 (3-Digit) industries. Year: 2006 All country–industry pairs 15,589 1.00 16 2 0 804 92,545 29 0 62,961,319 Pairs w/no HS-10 product exported to the United States 5,876 0.38 0 0 0 0 0 0 0 0 Pairs w/1 HS-10 product exported to the United States 1,614 0.10 1 1 1 1 797 9 0.3 115,469 Pairs w/2+ HS-10 products exported to the United States 8,099 0.52 31 12 2 804 177,972 2484 0.5 62,961,319 Notes: All statistics use product-level data on exports to the United States at the 10-digit level of the harmonized system (HS) classification (Feenstra et al. 2002). The sample consists of 119 countries that have exported to the United States in at least one industry during all years between 1989–2006. Industries are defined at the 4-digit level of the standard industrial classification (SIC) in panel (a) and at the 3-digit SIC level in panel (b); in each panel, the sample includes industries in which at least one country has exported to the United States during all years between 1989–2006. The standard deviation of log exports (used in subsequent tables) can be defined for country–industry pairs that have at least two HS-10 products exported to the United States; it is instead undefined (i.e., missing) for the other country–industry pairs. View Large The most important and innovative feature of our data set is that it includes approximately 230,000 measures of sales dispersion in a single and large market, across many countries and industries that differ greatly in financial frictions and financial vulnerability. It would be hard to assemble a similar data set using firm-level data. Although in reality the one-to-one correspondence between firms and products postulated in the model does not hold perfectly, it is not a priori obvious whether the predictions should be tested preferably using firm- or product-level data, given that the theory applies more directly to product innovation rather than firm creation. Fortunately, however, this distinction is not too relevant when working with a high level of product disaggregation, as we do. To see this, note that the number of products exported to the United States across countries and industries is typically not far from the number of foreign firms selling in the United States. In particular, although we do not have information on firm-level sales, we were able to obtain data on the number of foreign firms that have exported to the United States in the year 2002, separately for each foreign country and manufacturing industry. This information comes from the PIERS database, and was provided to us by IHS Markit. PIERS covers the universe of US maritime trade transactions, and accounts for 83% of total US imports in 2002. Using this information, we have found that, across all foreign countries and manufacturing industries, the median number of firms exporting to the United States is equal to 8 when defining industries at the 4-digit level, and to 13 when defining industries at the 3-digit level. The corresponding numbers of 10-digit products exported to the United States are 7 and 12, respectively (see Table 1). Hence, the number of units on which our sales dispersion measures are constructed is not far from the one that we would have used had we had access to export data at the firm level. The overall similarity between the number of products and firms is perhaps not too surprising, given that for the average country in our sample it is not very likely that more than one firm exports the same 10-digit product to the United States. Yet, a concern may remain that in large countries the mapping between firms and products may be less accurate. To alleviate this concern, we have checked that the cross-industry variation in sales dispersion obtained from trade data at the 10-digit product level reflects fairly closely the cross-industry variation in sales dispersion obtained from available US firm-level data. To this end, we have computed the standard deviation of log sales using 10-digit product-level data on exports from the United States to the rest of the world (Feenstra et al. 2002) and correlated this measure with the standard deviation of log sales computed with firm-level data from Compustat in 1997 (the midpoint of our sample). Despite important differences between the two data sets, and the fact that firms’ sales do not include only exports, the correlation turned out to be positive, sizable and statistically significant (0.47, p-value 0.03). 4.2. Stylized Facts We now present some new facts about how sales dispersion varies across countries, industries and years. In Table 2, we report descriptive statistics. In each panel, we consider a different sample, and show the mean and standard deviation of sales dispersion for the year 2006, as well as the change in sales dispersion over 1989–2006. We also show statistics on the number of 10-digit products used to construct the measures of sales dispersion in a given panel. In panel (a), we focus on our baseline sample of 119 countries and 365 industries. The mean and standard deviation of sales dispersion, computed across countries and industries, equal 1.94 and 0.88, respectively. Between 1989 and 2006, sales dispersion has increased on average by 6%. Hence, sales dispersion is large, varies greatly both geographically and across sectors, and has risen over the last two decades. In panel (b), we report the same statistics computed on a restricted sample of products, which consist of the 8,548 10-digit codes that are present in HS classification in each year between 1989 and 2006. The numbers are very close to those reported in panel (a), suggesting that our results do not depend on the changes that have occurred over time in the product classification (Schott and Pierce 2012). Table 2. Descriptive statistics on sales dispersion. Mean Std. Dev. Change Mean Std. Dev. Change (a) Consistent countries and industries (b) Consistent countries industries, and products, Sales dispersion 1.94 0.88 0.06 1.92 0.92 0.07 No. of products 15 25 2 11 17 0 (c) Cross-industry (d) Cross-country Sales dispersion 1.62 0.84 0.11 1.95 0.87 0.11 No. of products 9 12 2 12 10 2 Mean Std. Dev. Change Mean Std. Dev. Change (a) Consistent countries and industries (b) Consistent countries industries, and products, Sales dispersion 1.94 0.88 0.06 1.92 0.92 0.07 No. of products 15 25 2 11 17 0 (c) Cross-industry (d) Cross-country Sales dispersion 1.62 0.84 0.11 1.95 0.87 0.11 No. of products 9 12 2 12 10 2 Notes: Sales dispersion is the standard deviation of log exports, computed separately for each exporting country, 4-digit SIC manufacturing industry, and year, using data on exports to the United States at the 10-digit product level. The number of products is the number of 10-digit product codes used to compute the measures of sales dispersion. Mean and standard deviation refer to the year 2006; changes are computed over 1989–2006, and are expressed in percentages for sales dispersion and in units for the number of products. Panel (a) refers to a consistent sample of countries (119) and four-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Panel (b) uses the same sample as in panel (a), but restricts to a consistent set of 10-digit product codes (8548) that are present in the HS classification in all years between 1989 and 2006. The statistics in panels (a) and (b) are computed across all country–industry observations. The statistics in panel (c) are computed across industries within a given country, and are then averaged across the 119 countries. The statistics in panel (d) are computed across countries within a given industry, and are then averaged across the 365 industries. View Large Table 2. Descriptive statistics on sales dispersion. Mean Std. Dev. Change Mean Std. Dev. Change (a) Consistent countries and industries (b) Consistent countries industries, and products, Sales dispersion 1.94 0.88 0.06 1.92 0.92 0.07 No. of products 15 25 2 11 17 0 (c) Cross-industry (d) Cross-country Sales dispersion 1.62 0.84 0.11 1.95 0.87 0.11 No. of products 9 12 2 12 10 2 Mean Std. Dev. Change Mean Std. Dev. Change (a) Consistent countries and industries (b) Consistent countries industries, and products, Sales dispersion 1.94 0.88 0.06 1.92 0.92 0.07 No. of products 15 25 2 11 17 0 (c) Cross-industry (d) Cross-country Sales dispersion 1.62 0.84 0.11 1.95 0.87 0.11 No. of products 9 12 2 12 10 2 Notes: Sales dispersion is the standard deviation of log exports, computed separately for each exporting country, 4-digit SIC manufacturing industry, and year, using data on exports to the United States at the 10-digit product level. The number of products is the number of 10-digit product codes used to compute the measures of sales dispersion. Mean and standard deviation refer to the year 2006; changes are computed over 1989–2006, and are expressed in percentages for sales dispersion and in units for the number of products. Panel (a) refers to a consistent sample of countries (119) and four-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Panel (b) uses the same sample as in panel (a), but restricts to a consistent set of 10-digit product codes (8548) that are present in the HS classification in all years between 1989 and 2006. The statistics in panels (a) and (b) are computed across all country–industry observations. The statistics in panel (c) are computed across industries within a given country, and are then averaged across the 119 countries. The statistics in panel (d) are computed across countries within a given industry, and are then averaged across the 365 industries. View Large Next, we study how sales dispersion varies across countries and industries. In panel c), we focus on cross-industry variation. To this purpose we first compute, separately for each country, the mean and standard deviation of sales dispersion across the 365 industries, as well as the change in sales dispersion over the sample period. Then, we report average statistics across the 119 economies in our sample. In panel d), we focus instead on cross-country variation. To this purpose we first compute, separately for each industry, the mean and standard deviation of sales dispersion across the 119 countries, as well as the change in sales dispersion over the sample period. Then, we report average statistics across the 365 industries in our sample. Note that sales dispersion varies greatly both geographically and across industries, with the cross-country variation being slightly larger than the cross-industry variation. In both cases, sales dispersion has increased over the sample period, by 11% on average. These numbers are comparable to those obtained by Bonfiglioli et al. (2017) using US firm-level data over 1997–2007. Finally, we show that the variation in sales dispersion is not random, but correlates strongly with a number of country characteristics that are relevant for our theory. To this end we first compute, separately for each country, simple averages of sales dispersion across the 365 industries in 2006. Then, we plot this variable against different country characteristics. The results are displayed in Figure 1. The first graph studies how sales dispersion correlates with economic development, as proxied by real per-capita GDP.22 It shows that sales are significantly more dispersed in richer countries. This result confirms, using product-level trade data instead of firm-level data, the evidence from recent work on the firm size distribution, according to which the dispersion in firm size is increasing in countries’ level of development (e.g., Bartelsman et al. 2009; Poschke 2015). The second graph plots average sales dispersion against a standard proxy for countries’ financial development, namely the amount of credit (over GDP) issued by commercial banks and other financial institutions to the private sector. Note that sales dispersion is larger in countries where financial markets are more developed, and the relationship between the two measures is tight. The third graph shows how sales dispersion varies across countries with different levels of regulatory barriers affecting entry costs. In particular, we use an inverse proxy for entry barriers, given by the ranking of countries in terms of an index of doing business: countries occupying a higher position in the ranking have more friendly business regulations.23 Note that sales dispersion is increasing in the index of doing business and, thus, it is higher in countries with lower entry barriers. Finally, in the fourth graph we plot sales dispersion against average exports to the United States per product. The relationship is strong and positive, suggesting that countries with greater sales dispersion export more to a given market. Figure 1. View largeDownload slide Sales dispersion and country characteristics. Sales dispersion is the standard deviation of log exports, computed separately for each exporting country, 4-digit SIC manufacturing industry and year, using data on exports to the United States at the 10-digit product level (Feenstra et al. 2002). Each graph plots average sales dispersion in a given country (across 4-digit industries) against the country characteristic indicated on the horizontal axis. Per-capita GDP is real per-capita GDP from the Penn World Table 8.1. Private credit is the amount of credit issued by commercial banks and other financial institutions to the private sector over GDP, sourced from the Global Financial Development Database. Doing business is the ranking of countries in terms of the corresponding index of business regulation sourced from the World Bank Doing Business Database. Exports to the United States are expressed in millions of US dollars. Standard errors (reported in square brackets) are robust to heteroskedasticity. All graphs refer to the year 2006. Figure 1. View largeDownload slide Sales dispersion and country characteristics. Sales dispersion is the standard deviation of log exports, computed separately for each exporting country, 4-digit SIC manufacturing industry and year, using data on exports to the United States at the 10-digit product level (Feenstra et al. 2002). Each graph plots average sales dispersion in a given country (across 4-digit industries) against the country characteristic indicated on the horizontal axis. Per-capita GDP is real per-capita GDP from the Penn World Table 8.1. Private credit is the amount of credit issued by commercial banks and other financial institutions to the private sector over GDP, sourced from the Global Financial Development Database. Doing business is the ranking of countries in terms of the corresponding index of business regulation sourced from the World Bank Doing Business Database. Exports to the United States are expressed in millions of US dollars. Standard errors (reported in square brackets) are robust to heteroskedasticity. All graphs refer to the year 2006. Overall, Figure 1 shows that sales dispersion is higher in countries that are richer, have better financial markets and exhibit lower entry costs. In turn, greater sales dispersion is associated with greater exports. In the next sections, we exploit highly disaggregated data and variation across countries, industries, and years, to identify the effect of financial development on sales dispersion and the effect of sales dispersion on exports. 4.3. Sales Dispersion and Finance 4.3.1. Empirical Specification and Variables According to Proposition 2, the dispersion of sales from an origin country to a destination market, as measured by the standard deviation of log exports, should be increasing in the country’s level of financial development, especially in industries with higher financial vulnerability. Moreover, better export opportunities should also raise sales dispersion. To test Proposition 2, we estimate variants of the following specification: \begin{eqnarray} SD_{oit} &=&\alpha _{o}+\alpha _{i}+\alpha _{t}+\beta _{1}FD_{ot-1}+\beta _{2}FD_{ot-1}\cdot FV_{i} \nonumber \\ &&+\beta _{3}X_{ot-1}+\beta _{4}X_{ot-1}\cdot Z_{i}+\varepsilon _{oit}, \end{eqnarray} (11) where SDoit is the standard deviation of log exports to the United States from country o in industry i and year t; αo, αi, and αt are country, industry, and year fixed effects, respectively; FDot−1 is a measure of financial development in country o and year t − 1; FVi is a measure of industry i’s financial vulnerability; Xot−1 and Zi are, respectively, vectors of country and industry characteristics that determine comparative advantage, and thus proxy for export opportunities; finally, εoit is an error term.24 Our coefficient of interest is β2, which captures the differential effect of financial development on sales dispersion, across industries characterized by different degrees of financial vulnerability. As discussed in Rajan and Zingales (1998) and Manova (2013), this coefficient is identified by exploiting the asymmetric impact that financial frictions exert on industries, depending on technological characteristics that make industries more or less reliant on the financial system. The advantage of this strategy over a simple cross-country regression is the possibility to control for time-varying country characteristics potentially correlated with financial development.25 We are also interested in the vector of coefficients β4, which measure the impact of export opportunities and are identified similarly. Following, among others, Manova (2013), our preferred proxy for financial development (FDot−1) is private credit, which is a well-measured and internationally comparable indicator of the size of the financial system. In our main specifications, we use two variables for measuring the degree of financial vulnerability of an industry. The first proxy is external finance dependence (EFi), defined as the share of capital expenditure not financed with cash flow from operations (Rajan and Zingales 1998; Manova 2013). This variable is a direct proxy for financial vulnerability, because in sectors where EFi is higher, firms rely more on outside capital to finance their operations. The second proxy is asset tangibility (ATi), defined as the share of net property, plant, and equipment in total assets (Claessens and Laeven 2003; Manova 2013). This variable is an inverse proxy for financial vulnerability, because in sectors where ATi is higher, firms have more tangible assets to pledge as collateral when borrowing. Accordingly, we expect the coefficient β2 in equation (11) to be positive when using EFi and negative when using ATi.26 To construct EFi and ATi, we use US firm-level data, sourced from Compustat for the period 1989–2006.27 Because the United States has one of the most advanced financial systems in the world, using US data makes it more likely that EFi and ATi reflect firms’ actual credit needs and availability of tangible assets (Rajan and Zingales 1998; Claessens and Laeven 2003). At the same time, the ranking of industries in terms of EFi and ATi obtained with US data is likely to be preserved across countries and time periods, because financial vulnerability mostly depends on technological factors—such as the cash harvest period or the type of production process–that are common across economies and largely stable over time (Rajan and Zingales 1998).28 Finally, following Romalis (2004), Levchenko (2007), Nunn (2007), and Chor (2010), we proxy for export opportunities using different country–industry proxies for comparative advantage. These are the interactions between a country’s skill endowment, capital endowment, and institutional quality (Xot−1) with an industry’s skill intensity, capital intensity, and contract intensity (Zi), respectively.29 4.3.2. Baseline Estimates The baseline estimates of equation (11) are reported in Table 3. Standard errors are corrected for two-way clustering by country–industry and industry–year, in order to accommodate both autocorrelated shocks for the same country–industry pair and industry-specific shocks correlated across countries. In column (1), we start with a parsimonious specification that only includes the financial variables and full sets of fixed effects for origin countries (αo), industries (αi), and years (αt). These fixed effects absorb all time-invariant determinants of sales dispersion at the country and industry level, as well as general time trends common to all countries and sectors.30 Consistent with Proposition 2, the results show that sales dispersion is increasing in financial development, especially in financially vulnerable industries, where firms are more dependent on external finance or have fewer tangible assets. In column (2), we add the proxies for export opportunities.31 We find skill endowment, capital endowment and institutional quality to raise sales dispersion relatively more in industries that are skill and capital intensive, or dependent on relationship-specific investments. Hence, sales dispersion is also greater in the presence of better export opportunities, consistent with Proposition 2. Table 3. Sales dispersion and finance—Baseline estimates. (1) (2) (3) (4) (5) Financial development 0.042* 0.061** [0.024] [0.024] Fin. Dev. × External Finance Dependence 0.075*** 0.058*** 0.056*** 0.040*** 0.037*** [0.012] [0.011] [0.012] [0.013] [0.013] Fin. Dev. × Asset Tangibility − 0.150** − 0.219*** − 0.259*** − 0.398*** − 0.411*** [0.076] [0.078] [0.079] [0.085] [0.085] Skill Endowment 0.692*** [0.100] Capital Endowment − 0.247*** [0.033] Skill End. × Skill Intensity 0.350*** 0.384*** 0.256*** 0.246*** [0.037] [0.039] [0.044] [0.044] Cap. End. × Capital Intensity 0.067*** 0.067*** 0.062*** 0.059*** [0.006] [0.006] [0.009] [0.009] Institutional Quality × Contract Intensity 0.172* 0.107 0.164 0.109 [0.104] [0.105] [0.139] [0.139] No. of products 0.003*** [0.000] Observations 234,112 229,128 229,128 229,128 227,583 R-squared 0.19 0.20 0.23 0.25 0.25 Country FE Yes Yes No No No Industry FE Yes Yes No No No Year FE Yes Yes No No No Country–Year FE No No Yes Yes Yes Industry–Year FE No No Yes Yes Yes Price indexes × Industry FE No No No Yes Yes (1) (2) (3) (4) (5) Financial development 0.042* 0.061** [0.024] [0.024] Fin. Dev. × External Finance Dependence 0.075*** 0.058*** 0.056*** 0.040*** 0.037*** [0.012] [0.011] [0.012] [0.013] [0.013] Fin. Dev. × Asset Tangibility − 0.150** − 0.219*** − 0.259*** − 0.398*** − 0.411*** [0.076] [0.078] [0.079] [0.085] [0.085] Skill Endowment 0.692*** [0.100] Capital Endowment − 0.247*** [0.033] Skill End. × Skill Intensity 0.350*** 0.384*** 0.256*** 0.246*** [0.037] [0.039] [0.044] [0.044] Cap. End. × Capital Intensity 0.067*** 0.067*** 0.062*** 0.059*** [0.006] [0.006] [0.009] [0.009] Institutional Quality × Contract Intensity 0.172* 0.107 0.164 0.109 [0.104] [0.105] [0.139] [0.139] No. of products 0.003*** [0.000] Observations 234,112 229,128 229,128 229,128 227,583 R-squared 0.19 0.20 0.23 0.25 0.25 Country FE Yes Yes No No No Industry FE Yes Yes No No No Year FE Yes Yes No No No Country–Year FE No No Yes Yes Yes Industry–Year FE No No Yes Yes Yes Price indexes × Industry FE No No No Yes Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. Financial development is proxied by private credit as a share of GDP. External finance dependence and asset tangibility are, respectively, the share of capital expenditure not financed with cash flow from operations and the share of net property, plant, and equipment in total assets (industry-level averages over 1989–2006). Skill endowment is the log index of human capital per person. Capital endowment is log real capital stock per person engaged. Skill intensity is the log average ratio of nonproduction to production worker employment over 1989–2006. Capital intensity is the log average ratio of real capital stock per worker over 1989–2006. Institutional quality is average rule of law over 1996–2006. Contract intensity is an indicator for the importance of relationship-specific investments in each industry. The number of products is the number of 10-digit product codes that are exported by a given country to the United States in a given industry and year. All time-varying regressors are lagged one period. All regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 3. Sales dispersion and finance—Baseline estimates. (1) (2) (3) (4) (5) Financial development 0.042* 0.061** [0.024] [0.024] Fin. Dev. × External Finance Dependence 0.075*** 0.058*** 0.056*** 0.040*** 0.037*** [0.012] [0.011] [0.012] [0.013] [0.013] Fin. Dev. × Asset Tangibility − 0.150** − 0.219*** − 0.259*** − 0.398*** − 0.411*** [0.076] [0.078] [0.079] [0.085] [0.085] Skill Endowment 0.692*** [0.100] Capital Endowment − 0.247*** [0.033] Skill End. × Skill Intensity 0.350*** 0.384*** 0.256*** 0.246*** [0.037] [0.039] [0.044] [0.044] Cap. End. × Capital Intensity 0.067*** 0.067*** 0.062*** 0.059*** [0.006] [0.006] [0.009] [0.009] Institutional Quality × Contract Intensity 0.172* 0.107 0.164 0.109 [0.104] [0.105] [0.139] [0.139] No. of products 0.003*** [0.000] Observations 234,112 229,128 229,128 229,128 227,583 R-squared 0.19 0.20 0.23 0.25 0.25 Country FE Yes Yes No No No Industry FE Yes Yes No No No Year FE Yes Yes No No No Country–Year FE No No Yes Yes Yes Industry–Year FE No No Yes Yes Yes Price indexes × Industry FE No No No Yes Yes (1) (2) (3) (4) (5) Financial development 0.042* 0.061** [0.024] [0.024] Fin. Dev. × External Finance Dependence 0.075*** 0.058*** 0.056*** 0.040*** 0.037*** [0.012] [0.011] [0.012] [0.013] [0.013] Fin. Dev. × Asset Tangibility − 0.150** − 0.219*** − 0.259*** − 0.398*** − 0.411*** [0.076] [0.078] [0.079] [0.085] [0.085] Skill Endowment 0.692*** [0.100] Capital Endowment − 0.247*** [0.033] Skill End. × Skill Intensity 0.350*** 0.384*** 0.256*** 0.246*** [0.037] [0.039] [0.044] [0.044] Cap. End. × Capital Intensity 0.067*** 0.067*** 0.062*** 0.059*** [0.006] [0.006] [0.009] [0.009] Institutional Quality × Contract Intensity 0.172* 0.107 0.164 0.109 [0.104] [0.105] [0.139] [0.139] No. of products 0.003*** [0.000] Observations 234,112 229,128 229,128 229,128 227,583 R-squared 0.19 0.20 0.23 0.25 0.25 Country FE Yes Yes No No No Industry FE Yes Yes No No No Year FE Yes Yes No No No Country–Year FE No No Yes Yes Yes Industry–Year FE No No Yes Yes Yes Price indexes × Industry FE No No No Yes Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. Financial development is proxied by private credit as a share of GDP. External finance dependence and asset tangibility are, respectively, the share of capital expenditure not financed with cash flow from operations and the share of net property, plant, and equipment in total assets (industry-level averages over 1989–2006). Skill endowment is the log index of human capital per person. Capital endowment is log real capital stock per person engaged. Skill intensity is the log average ratio of nonproduction to production worker employment over 1989–2006. Capital intensity is the log average ratio of real capital stock per worker over 1989–2006. Institutional quality is average rule of law over 1996–2006. Contract intensity is an indicator for the importance of relationship-specific investments in each industry. The number of products is the number of 10-digit product codes that are exported by a given country to the United States in a given industry and year. All time-varying regressors are lagged one period. All regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large In column (3), we replace the country, industry, and year fixed effects with country–year (αot) and industry–year (αit) fixed effects. The latter soak up all shocks hitting a given country or sector in a year.32 Hence, to identify the coefficients, in this specification we exploit the combination of cross-country variation in financial development and endowments within a year, and cross-industry variation in financial vulnerability and factor intensities. Reassuringly, the interaction coefficients are largely unchanged. In column (4), we augment the previous specification by including a full set of interactions between countries’ Consumer Price Indexes and industry dummies. These variables are meant to control for country–industry specific changes in the price indexes (see, e.g., Manova 2013). Our main evidence is unaffected. Finally, in column (5) we control for the number of products exported to the United States within each country–industry–year triplet. This variable has a positive but very small coefficient, and its inclusion does not make any noteworthy change in our main results. This suggests that sales dispersion is not mechanically driven by the number of products on which it is constructed. Furthermore, to make sure that the effect of financial development is not confounded by any correlation with the number of exported products, from now on we control for the latter variable in most of the specifications. 4.3.3. Robustness Checks In this section, we submit the baseline estimates to a large number of robustness checks. We focus on the richest specification reported in column (5) of Table 3. Alternative Samples. In Table 4, we address a number of potential concerns with the composition of the estimation sample. We start by showing that our evidence is not driven by the sample of 10-digit HS products used to construct the measures of sales dispersion. In particular, in column (1) we find similar results when excluding country–industry–year triplets with only two products exported to the United States. In column (2), we instead confirm the main evidence by re-computing SDoit after excluding products with limited exports, that is, products that fall in the bottom 25% of exports within each country–industry–year triplet.33 Table 4. Sales dispersion and finance—Alternative samples. At least 3 products No small products Probit Heckman correction No small countries 3-Digit industries Consistent products (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.031** 0.035*** 0.060*** 0.041*** 0.037*** 0.059*** 0.054*** [0.013] [0.012] [0.005] [0.013] [0.013] [0.017] [0.020] Fin. Dev. × Ass. Tang. − 0.381*** − 0.185** − 0.693*** − 0.384*** − 0.393*** − 0.208* − 0.282** [0.084] [0.079] [0.028] [0.085] [0.086] [0.111] [0.130] Skill End. × Skill Int. 0.240*** 0.129*** 0.288*** 0.266*** 0.258*** 0.156*** 0.131* [0.047] [0.039] [0.007] [0.044] [0.047] [0.056] [0.068] Cap. End. × Cap. Int. 0.060*** 0.040*** 0.017*** 0.064*** 0.061*** 0.051*** 0.036** [0.009] [0.007] [0.001] [0.008] [0.009] [0.013] [0.016] Inst. Qual. × Contr. Int. − 0.126 0.056 2.380*** 0.242* 0.063 0.442** 0.379* [0.143] [0.127] [0.025] [0.140] [0.146] [0.185] [0.213] No. of products 0.003*** 0.001*** 0.003*** 0.002*** 0.002*** [0.000] [0.000] [0.000] [0.000] [0.000] Inverse Mills ratio 0.101*** [0.021] Observations 189,522 227,583 566,020 229,128 197,095 110,346 95,502 R-squared 0.28 0.16 0.25 0.26 0.32 0.28 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes At least 3 products No small products Probit Heckman correction No small countries 3-Digit industries Consistent products (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.031** 0.035*** 0.060*** 0.041*** 0.037*** 0.059*** 0.054*** [0.013] [0.012] [0.005] [0.013] [0.013] [0.017] [0.020] Fin. Dev. × Ass. Tang. − 0.381*** − 0.185** − 0.693*** − 0.384*** − 0.393*** − 0.208* − 0.282** [0.084] [0.079] [0.028] [0.085] [0.086] [0.111] [0.130] Skill End. × Skill Int. 0.240*** 0.129*** 0.288*** 0.266*** 0.258*** 0.156*** 0.131* [0.047] [0.039] [0.007] [0.044] [0.047] [0.056] [0.068] Cap. End. × Cap. Int. 0.060*** 0.040*** 0.017*** 0.064*** 0.061*** 0.051*** 0.036** [0.009] [0.007] [0.001] [0.008] [0.009] [0.013] [0.016] Inst. Qual. × Contr. Int. − 0.126 0.056 2.380*** 0.242* 0.063 0.442** 0.379* [0.143] [0.127] [0.025] [0.140] [0.146] [0.185] [0.213] No. of products 0.003*** 0.001*** 0.003*** 0.002*** 0.002*** [0.000] [0.000] [0.000] [0.000] [0.000] Inverse Mills ratio 0.101*** [0.021] Observations 189,522 227,583 566,020 229,128 197,095 110,346 95,502 R-squared 0.28 0.16 0.25 0.26 0.32 0.28 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Notes: Except for column (3), the dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. In column (3), the dependent variable is instead a dummy, which takes the value of 1 for country–industry–year triplets with two or more products exported to the United States (i.e., triplets for which sales dispersion is defined) and the value of 0 for the remaining triplets (for which sales dispersion is not defined). Column (1) uses country–industry–year observations for which sales dispersion is based on at least three products exported to the United States. In column (2), sales dispersion is computed after excluding the bottom 25% of products (with the smallest value of exports) in each country–industry–year triplet. In column (4), the inverse Mills ratio is constructed as in Heckman (1979), using predicted values from the first-stage probit regression reported in column (3). Column (5) excludes countries with less than 5 million people in 2006. Column (6) defines industries at the 3-digit (instead of 4-digit) level. Column (7) further constructs sales dispersion using a consistent set of 10-digit product codes (8548) that are present in the HS classification in all years between 1989 and 2006. All time-varying regressors are lagged one period. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year, except in column (3), where they are corrected for clustering at the industry–year level. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 4. Sales dispersion and finance—Alternative samples. At least 3 products No small products Probit Heckman correction No small countries 3-Digit industries Consistent products (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.031** 0.035*** 0.060*** 0.041*** 0.037*** 0.059*** 0.054*** [0.013] [0.012] [0.005] [0.013] [0.013] [0.017] [0.020] Fin. Dev. × Ass. Tang. − 0.381*** − 0.185** − 0.693*** − 0.384*** − 0.393*** − 0.208* − 0.282** [0.084] [0.079] [0.028] [0.085] [0.086] [0.111] [0.130] Skill End. × Skill Int. 0.240*** 0.129*** 0.288*** 0.266*** 0.258*** 0.156*** 0.131* [0.047] [0.039] [0.007] [0.044] [0.047] [0.056] [0.068] Cap. End. × Cap. Int. 0.060*** 0.040*** 0.017*** 0.064*** 0.061*** 0.051*** 0.036** [0.009] [0.007] [0.001] [0.008] [0.009] [0.013] [0.016] Inst. Qual. × Contr. Int. − 0.126 0.056 2.380*** 0.242* 0.063 0.442** 0.379* [0.143] [0.127] [0.025] [0.140] [0.146] [0.185] [0.213] No. of products 0.003*** 0.001*** 0.003*** 0.002*** 0.002*** [0.000] [0.000] [0.000] [0.000] [0.000] Inverse Mills ratio 0.101*** [0.021] Observations 189,522 227,583 566,020 229,128 197,095 110,346 95,502 R-squared 0.28 0.16 0.25 0.26 0.32 0.28 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes At least 3 products No small products Probit Heckman correction No small countries 3-Digit industries Consistent products (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.031** 0.035*** 0.060*** 0.041*** 0.037*** 0.059*** 0.054*** [0.013] [0.012] [0.005] [0.013] [0.013] [0.017] [0.020] Fin. Dev. × Ass. Tang. − 0.381*** − 0.185** − 0.693*** − 0.384*** − 0.393*** − 0.208* − 0.282** [0.084] [0.079] [0.028] [0.085] [0.086] [0.111] [0.130] Skill End. × Skill Int. 0.240*** 0.129*** 0.288*** 0.266*** 0.258*** 0.156*** 0.131* [0.047] [0.039] [0.007] [0.044] [0.047] [0.056] [0.068] Cap. End. × Cap. Int. 0.060*** 0.040*** 0.017*** 0.064*** 0.061*** 0.051*** 0.036** [0.009] [0.007] [0.001] [0.008] [0.009] [0.013] [0.016] Inst. Qual. × Contr. Int. − 0.126 0.056 2.380*** 0.242* 0.063 0.442** 0.379* [0.143] [0.127] [0.025] [0.140] [0.146] [0.185] [0.213] No. of products 0.003*** 0.001*** 0.003*** 0.002*** 0.002*** [0.000] [0.000] [0.000] [0.000] [0.000] Inverse Mills ratio 0.101*** [0.021] Observations 189,522 227,583 566,020 229,128 197,095 110,346 95,502 R-squared 0.28 0.16 0.25 0.26 0.32 0.28 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Notes: Except for column (3), the dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. In column (3), the dependent variable is instead a dummy, which takes the value of 1 for country–industry–year triplets with two or more products exported to the United States (i.e., triplets for which sales dispersion is defined) and the value of 0 for the remaining triplets (for which sales dispersion is not defined). Column (1) uses country–industry–year observations for which sales dispersion is based on at least three products exported to the United States. In column (2), sales dispersion is computed after excluding the bottom 25% of products (with the smallest value of exports) in each country–industry–year triplet. In column (4), the inverse Mills ratio is constructed as in Heckman (1979), using predicted values from the first-stage probit regression reported in column (3). Column (5) excludes countries with less than 5 million people in 2006. Column (6) defines industries at the 3-digit (instead of 4-digit) level. Column (7) further constructs sales dispersion using a consistent set of 10-digit product codes (8548) that are present in the HS classification in all years between 1989 and 2006. All time-varying regressors are lagged one period. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year, except in column (3), where they are corrected for clustering at the industry–year level. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large In columns (3)–(6), we use different approaches for accommodating observations with missing sales dispersion, which correspond to triplets with either zero or one product only. A possible concern is that, if the missing values are not random, our evidence might be driven by sample selection bias. We start by addressing this issue with a two-step model à la Heckman (1979). In particular, in column (3) we estimate a Probit model for the probability of observing a triplet with nonmissing sales dispersion. The results show that sales dispersion is more likely to be observed in financially developed countries, especially in industries with greater financial vulnerability.34 Then, using predicted values from column (3), we construct the inverse Mills ratio and include it as an additional control in the main equation (column (4)).35 The coefficient on the inverse Mills ratio is positive and precisely estimated, indicating that the errors of the two equations are correlated, but it is also small in size. Accordingly, correcting the estimates for sample selection yields coefficients that are practically identical to the baseline ones reported in column (4) of Table 3.36 In column (5), we instead exclude small countries (those with less than 5 million people in 2006) and concentrate on large exporters, for which we observe sales dispersion in the vast majority of industries and years. Alternatively, in column (6) we redefine industries at the 3-digit level, since triplets with missing sales dispersion are less numerous when industries are more aggregated, as shown in Table 1. Despite the drop in sample size, our evidence is unchanged also in these specifications. Finally, in column (7) we further restrict the sample to a consistent set of 8,548 products that are present in the HS classification during all years between 1989 and 2006, and we reconstruct the measures of sales dispersion using these products only. Although the HS classification has been partly restructured over the sample period, the main results are unchanged, suggesting that they are not driven by the modifications occurred over time in the product classification. Alternative Proxies. In Table 5, we use alternative measures of financial development and financial vulnerability. We start by replacing private credit with other common proxies for the size of the financial system, namely, deposit money bank assets, liquid liabilities and domestic credit as a share of GDP (columns (1)–(3)).37 The results always show that financial development increases sales dispersion especially in financially vulnerable industries. In column (4) we use instead the log lending rate, which measures the cost incurred by firms for obtaining credit, and is therefore an inverse proxy for the size and efficiency of the financial system.38 Consistent with this interpretation, we find the interactions involving the lending rate to have the opposite signs as those involving private credit or other proxies for size. Table 5. Sales dispersion and finance—Alternative proxies. Bank assets Liquid liabilities Domestic credit Lending rate Lagged financial vulnerability Rankings of financial vulnerability Firm age (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.043*** 0.026** 0.031*** − 0.042*** 0.049*** 0.111*** [0.013] [0.013] [0.012] [0.007] [0.010] [0.032] Fin. Dev. × Ass. Tang. − 0.527*** − 0.639*** − 0.401*** 0.171*** − 0.296*** − 0.176*** [0.080] [0.081] [0.077] [0.048] [0.098] [0.035] Fin. Dev. × Firm Age − 0.118*** [0.030] Skill End. × Skill Int. 0.250*** 0.240*** 0.253*** 0.185*** 0.247*** 0.247*** 0.252*** [0.044] [0.044] [0.044] [0.047] [0.044] [0.044] [0.044] Cap. End. × Cap. Int. 0.061*** 0.061*** 0.058*** 0.035*** 0.060*** 0.060*** 0.059*** [0.009] [0.009] [0.009] [0.010] [0.009] [0.009] [0.009] Inst. Qual. × Contr. Int. 0.064 0.093 0.121 0.325** 0.142 0.090 0.270** [0.139] [0.137] [0.139] [0.146] [0.139] [0.140] [0.136] No. of products 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Observations 226,881 229,112 230,843 216,037 227,296 227,583 228,192 R-squared 0.25 0.25 0.25 0.26 0.25 0.25 0.25 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Bank assets Liquid liabilities Domestic credit Lending rate Lagged financial vulnerability Rankings of financial vulnerability Firm age (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.043*** 0.026** 0.031*** − 0.042*** 0.049*** 0.111*** [0.013] [0.013] [0.012] [0.007] [0.010] [0.032] Fin. Dev. × Ass. Tang. − 0.527*** − 0.639*** − 0.401*** 0.171*** − 0.296*** − 0.176*** [0.080] [0.081] [0.077] [0.048] [0.098] [0.035] Fin. Dev. × Firm Age − 0.118*** [0.030] Skill End. × Skill Int. 0.250*** 0.240*** 0.253*** 0.185*** 0.247*** 0.247*** 0.252*** [0.044] [0.044] [0.044] [0.047] [0.044] [0.044] [0.044] Cap. End. × Cap. Int. 0.061*** 0.061*** 0.058*** 0.035*** 0.060*** 0.060*** 0.059*** [0.009] [0.009] [0.009] [0.010] [0.009] [0.009] [0.009] Inst. Qual. × Contr. Int. 0.064 0.093 0.121 0.325** 0.142 0.090 0.270** [0.139] [0.137] [0.139] [0.146] [0.139] [0.140] [0.136] No. of products 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Observations 226,881 229,112 230,843 216,037 227,296 227,583 228,192 R-squared 0.25 0.25 0.25 0.26 0.25 0.25 0.25 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. Financial development is proxied by deposit money bank assets as a share of GDP in column (1), liquid liabilities as a share of GDP in column (2), domestic credit to the private sector as a share of GDP in column (3), and the log lending rate in column (4). In column (5), external finance dependence and asset tangibility are computed as averages over the presample period, 1979–1988. In column (6), the actual values of external finance dependence and asset tangibility are replaced by the rankings of industries in terms of these variables; the rankings are based on data for 1989–2006 and are normalized between 0 and 1. In column (7), firm age is the log median number of years in which firms in an industry are listed in the US stock market, based on data from Compustat. All time-varying regressors are lagged one period. All regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 5. Sales dispersion and finance—Alternative proxies. Bank assets Liquid liabilities Domestic credit Lending rate Lagged financial vulnerability Rankings of financial vulnerability Firm age (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.043*** 0.026** 0.031*** − 0.042*** 0.049*** 0.111*** [0.013] [0.013] [0.012] [0.007] [0.010] [0.032] Fin. Dev. × Ass. Tang. − 0.527*** − 0.639*** − 0.401*** 0.171*** − 0.296*** − 0.176*** [0.080] [0.081] [0.077] [0.048] [0.098] [0.035] Fin. Dev. × Firm Age − 0.118*** [0.030] Skill End. × Skill Int. 0.250*** 0.240*** 0.253*** 0.185*** 0.247*** 0.247*** 0.252*** [0.044] [0.044] [0.044] [0.047] [0.044] [0.044] [0.044] Cap. End. × Cap. Int. 0.061*** 0.061*** 0.058*** 0.035*** 0.060*** 0.060*** 0.059*** [0.009] [0.009] [0.009] [0.010] [0.009] [0.009] [0.009] Inst. Qual. × Contr. Int. 0.064 0.093 0.121 0.325** 0.142 0.090 0.270** [0.139] [0.137] [0.139] [0.146] [0.139] [0.140] [0.136] No. of products 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Observations 226,881 229,112 230,843 216,037 227,296 227,583 228,192 R-squared 0.25 0.25 0.25 0.26 0.25 0.25 0.25 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Bank assets Liquid liabilities Domestic credit Lending rate Lagged financial vulnerability Rankings of financial vulnerability Firm age (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.043*** 0.026** 0.031*** − 0.042*** 0.049*** 0.111*** [0.013] [0.013] [0.012] [0.007] [0.010] [0.032] Fin. Dev. × Ass. Tang. − 0.527*** − 0.639*** − 0.401*** 0.171*** − 0.296*** − 0.176*** [0.080] [0.081] [0.077] [0.048] [0.098] [0.035] Fin. Dev. × Firm Age − 0.118*** [0.030] Skill End. × Skill Int. 0.250*** 0.240*** 0.253*** 0.185*** 0.247*** 0.247*** 0.252*** [0.044] [0.044] [0.044] [0.047] [0.044] [0.044] [0.044] Cap. End. × Cap. Int. 0.061*** 0.061*** 0.058*** 0.035*** 0.060*** 0.060*** 0.059*** [0.009] [0.009] [0.009] [0.010] [0.009] [0.009] [0.009] Inst. Qual. × Contr. Int. 0.064 0.093 0.121 0.325** 0.142 0.090 0.270** [0.139] [0.137] [0.139] [0.146] [0.139] [0.140] [0.136] No. of products 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Observations 226,881 229,112 230,843 216,037 227,296 227,583 228,192 R-squared 0.25 0.25 0.25 0.26 0.25 0.25 0.25 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. Financial development is proxied by deposit money bank assets as a share of GDP in column (1), liquid liabilities as a share of GDP in column (2), domestic credit to the private sector as a share of GDP in column (3), and the log lending rate in column (4). In column (5), external finance dependence and asset tangibility are computed as averages over the presample period, 1979–1988. In column (6), the actual values of external finance dependence and asset tangibility are replaced by the rankings of industries in terms of these variables; the rankings are based on data for 1989–2006 and are normalized between 0 and 1. In column (7), firm age is the log median number of years in which firms in an industry are listed in the US stock market, based on data from Compustat. All time-varying regressors are lagged one period. All regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Next, we perform robustness checks using alternative proxies for financial vulnerability. In column (5), we replace our main measures with equivalent indicators based on data for the pre-sample decade (1979–1988). In column (6), we instead replace the actual values of EFi and ATi with the rankings of industries in terms of these two variables.39 The results are similar to the baseline estimates, consistent with the idea that cross-industry differences in financial vulnerability are mostly driven by technological factors, which tend to persist both across countries and over time. Finally, in column (7) we use the age of firms in an industry as an alternative proxy for financial vulnerability, as younger firms are typically more constrained than more established ones. We proxy for firm age using the log median number of years in which firms in an industry are listed in the US stock market, based on data from Compustat. We find the interaction of financial development with firm age to be negative and very precisely estimated, implying that financial frictions have stronger effects on sales dispersion in industries that are populated by younger firms. Hence, our evidence is remarkably robust across different proxies for financial vulnerability. Alternative Estimation Approaches. We now show that our main evidence holds when using alternative ways of measuring sales dispersion and alternative strategies for estimating the baseline specification. The results are reported in Table 6. We start by running weighted regressions, which give more weight to triplets with a larger number of products, for which sales dispersion may be measured more precisely. In particular, in column (1) we weight the regression with the log number of products; taking logs avoids giving excessive weight to a few, exceptionally large, triplets. In column (2), we instead weight the regression using industries’ shares in the total number of products exported by a given country to the United States in each year; using share accommodate differences in the number of products sold by different countries in the United States. In both cases, the coefficients on the financial variables are close to our baseline estimates. Table 6. Sales dispersion and finance—Alternative estimation approaches. Weighted regression Weighted regression Pareto shape parameter Pareto shape parameter Pareto shape parameter (1) (2) (3) (4) (5) Fin. Dev. × Ext. Fin. Dep. 0.026** 0.037** 0.033*** 0.025*** [0.012] [0.017] [0.007] [0.006] Fin. Dev. × Ass. Tang. − 0.397*** − 0.591*** − 0.506*** − 0.480*** [0.079] [0.135] [0.037] [0.034] Fin. Dev. × Firm Age − 0.133*** [0.005] Skill End. × Skill Int. 0.260*** 0.055 0.257*** 0.196*** 0.201*** [0.045] [0.065] [0.008] [0.008] [0.009] Cap. End. × Cap. Int. 0.063*** 0.034*** 0.077*** 0.066*** 0.065*** [0.009] [0.010] [0.002] [0.001] [0.002] Inst. Qual. × Contr. Int. − 0.299** − 0.455** − 0.275*** − 0.353*** − 0.154*** [0.136] [0.195] [0.031] [0.028] [0.029] No. of products 0.002*** 0.001*** 0.003*** 0.001*** 0.001*** [0.000] [0.000] [0.000] [0.000] [0.000] Observations 227,583 227,583 189,522 189,522 189,912 R-squared 0.29 0.41 0.25 0.44 0.44 Country–Year FE Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Weighted regression Weighted regression Pareto shape parameter Pareto shape parameter Pareto shape parameter (1) (2) (3) (4) (5) Fin. Dev. × Ext. Fin. Dep. 0.026** 0.037** 0.033*** 0.025*** [0.012] [0.017] [0.007] [0.006] Fin. Dev. × Ass. Tang. − 0.397*** − 0.591*** − 0.506*** − 0.480*** [0.079] [0.135] [0.037] [0.034] Fin. Dev. × Firm Age − 0.133*** [0.005] Skill End. × Skill Int. 0.260*** 0.055 0.257*** 0.196*** 0.201*** [0.045] [0.065] [0.008] [0.008] [0.009] Cap. End. × Cap. Int. 0.063*** 0.034*** 0.077*** 0.066*** 0.065*** [0.009] [0.010] [0.002] [0.001] [0.002] Inst. Qual. × Contr. Int. − 0.299** − 0.455** − 0.275*** − 0.353*** − 0.154*** [0.136] [0.195] [0.031] [0.028] [0.029] No. of products 0.002*** 0.001*** 0.003*** 0.001*** 0.001*** [0.000] [0.000] [0.000] [0.000] [0.000] Observations 227,583 227,583 189,522 189,522 189,912 R-squared 0.29 0.41 0.25 0.44 0.44 Country–Year FE Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Notes: The dependent variable is sales dispersion. In columns (1) and (2), it is defined as the standard deviation of log exports, computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. In columns (3)–(5), sales dispersion is instead constructed as the inverse of the shape parameter of the Pareto distribution. To estimate the shape parameter for each exporting country, industry, and year, a regression of log sales rank on log sales is run for each triplet, using data on exports to the United States at the 10-digit product level; only triplets with at least three products are considered. Sales rank is adjusted by subtracting 0.5 as in Gabaix and Ibragimov (2011). The shape parameters are the absolute values of the coefficients on log sales obtained from these regressions. The regression in column (1) is weighted using the log number of 10-digit products that are exported to the United States in each country–industry–year triplet. The regression in column (2) is weighted using each industry’s share in the total number of 10-digit products that are exported to the United States by each country in each year. The regressions in columns (4) and (5) are weighted using the inverse of the standard errors of the Pareto shape parameters. All time-varying regressors are lagged one period. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year (in columns (1) and (2)) or bootstrapped (100 replications, with observations sampled within country–industry pairs, in columns (3)–(5)). See also notes to previous tables. **Significant at 5%; ***significant at 1%. View Large Table 6. Sales dispersion and finance—Alternative estimation approaches. Weighted regression Weighted regression Pareto shape parameter Pareto shape parameter Pareto shape parameter (1) (2) (3) (4) (5) Fin. Dev. × Ext. Fin. Dep. 0.026** 0.037** 0.033*** 0.025*** [0.012] [0.017] [0.007] [0.006] Fin. Dev. × Ass. Tang. − 0.397*** − 0.591*** − 0.506*** − 0.480*** [0.079] [0.135] [0.037] [0.034] Fin. Dev. × Firm Age − 0.133*** [0.005] Skill End. × Skill Int. 0.260*** 0.055 0.257*** 0.196*** 0.201*** [0.045] [0.065] [0.008] [0.008] [0.009] Cap. End. × Cap. Int. 0.063*** 0.034*** 0.077*** 0.066*** 0.065*** [0.009] [0.010] [0.002] [0.001] [0.002] Inst. Qual. × Contr. Int. − 0.299** − 0.455** − 0.275*** − 0.353*** − 0.154*** [0.136] [0.195] [0.031] [0.028] [0.029] No. of products 0.002*** 0.001*** 0.003*** 0.001*** 0.001*** [0.000] [0.000] [0.000] [0.000] [0.000] Observations 227,583 227,583 189,522 189,522 189,912 R-squared 0.29 0.41 0.25 0.44 0.44 Country–Year FE Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Weighted regression Weighted regression Pareto shape parameter Pareto shape parameter Pareto shape parameter (1) (2) (3) (4) (5) Fin. Dev. × Ext. Fin. Dep. 0.026** 0.037** 0.033*** 0.025*** [0.012] [0.017] [0.007] [0.006] Fin. Dev. × Ass. Tang. − 0.397*** − 0.591*** − 0.506*** − 0.480*** [0.079] [0.135] [0.037] [0.034] Fin. Dev. × Firm Age − 0.133*** [0.005] Skill End. × Skill Int. 0.260*** 0.055 0.257*** 0.196*** 0.201*** [0.045] [0.065] [0.008] [0.008] [0.009] Cap. End. × Cap. Int. 0.063*** 0.034*** 0.077*** 0.066*** 0.065*** [0.009] [0.010] [0.002] [0.001] [0.002] Inst. Qual. × Contr. Int. − 0.299** − 0.455** − 0.275*** − 0.353*** − 0.154*** [0.136] [0.195] [0.031] [0.028] [0.029] No. of products 0.002*** 0.001*** 0.003*** 0.001*** 0.001*** [0.000] [0.000] [0.000] [0.000] [0.000] Observations 227,583 227,583 189,522 189,522 189,912 R-squared 0.29 0.41 0.25 0.44 0.44 Country–Year FE Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Notes: The dependent variable is sales dispersion. In columns (1) and (2), it is defined as the standard deviation of log exports, computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. In columns (3)–(5), sales dispersion is instead constructed as the inverse of the shape parameter of the Pareto distribution. To estimate the shape parameter for each exporting country, industry, and year, a regression of log sales rank on log sales is run for each triplet, using data on exports to the United States at the 10-digit product level; only triplets with at least three products are considered. Sales rank is adjusted by subtracting 0.5 as in Gabaix and Ibragimov (2011). The shape parameters are the absolute values of the coefficients on log sales obtained from these regressions. The regression in column (1) is weighted using the log number of 10-digit products that are exported to the United States in each country–industry–year triplet. The regression in column (2) is weighted using each industry’s share in the total number of 10-digit products that are exported to the United States by each country in each year. The regressions in columns (4) and (5) are weighted using the inverse of the standard errors of the Pareto shape parameters. All time-varying regressors are lagged one period. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year (in columns (1) and (2)) or bootstrapped (100 replications, with observations sampled within country–industry pairs, in columns (3)–(5)). See also notes to previous tables. **Significant at 5%; ***significant at 1%. View Large Next, we use an alternative estimate of sales dispersion. As a baseline measure, we have chosen the standard deviation of log sales both because it is easy to build and interpret, and because it is consistent with the theoretical model. However, under the assumption that sales are Pareto distributed, as in our model, the same measure of dispersion can be estimated as the inverse of the shape parameter. To check that the results are indeed robust to these alternative measures of dispersion, we estimate a separate shape parameter for each country–industry–year triplet, by running regressions of log sales rank on log sales across 10-digit products; the shape parameters are the absolute values of the coefficients on log sales obtained from these regressions.40 It is reassuring that the correlation between the two measures of sales dispersion is extremely high (0.97). In column (3), we use the new measures of sales dispersion in place of the standard deviation of log sales. We bootstrap the standard errors by resampling observations within country–industry pairs, to account for the estimation of the shape parameters in the first stage. Using the inverse of the Pareto shape parameter we obtain coefficients that are very close to our baseline estimates for the standard deviation of log exports. An additional advantage of the Pareto shape parameters is that, since they are estimated, they come with a measure of fit. We exploit this information in column (4), where we repeat the previous specification, but we now weight the observations with the inverse of the standard errors of the Pareto shape parameters. This allows us to give less weight to triplets for which sales dispersion is estimated less precisely. We find no noteworthy change in the main coefficients. Finally, in column (5) we re-estimate the weighted regression using firm age as an alternative proxy for financial vulnerability. We continue to find strong evidence that financial development raises sales dispersion more in financially vulnerable industries. Additional Controls. A possible concern with our baseline results is that the coefficients on financial development may pick up the effects of omitted variables, which are correlated with financial frictions and may also influence sales dispersion. Our identification strategy partly allays this concern. Indeed, our specifications control for country–year and industry–year fixed effects, so the estimated coefficients do not reflect shocks hitting specific countries and sectors in a given year. Hence, in this section we focus on factors that vary both across countries and over time, and that may have differential effects on sales dispersion across sectors. The results are reported in Table 7. In column (1), we include the interactions between real per-capita GDP and the two proxies for financial vulnerability, in order to account for the fact that richer countries are more financially developed. The coefficients on the new interactions are small and not very precisely estimated, suggesting that the effect of economic development on sales dispersion is not heterogeneous across industries. At the same time, our coefficients of interest are largely unchanged, suggesting that the baseline estimates are not contaminated by the correlation of financial development with per-capita income. Table 7. Sales dispersion and finance—Additional controls. Per-capita GDP Import penetration and export intensity Real exchange rate Foreign direct investment Number of HS codes All controls All controls (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.035*** 0.032** 0.037*** 0.038*** 0.039*** 0.037*** [0.013] [0.013] [0.013] [0.014] [0.013] [0.014] Fin. Dev. × Ass. Tang. − 0.484*** − 0.355*** − 0.416*** − 0.378*** − 0.351*** − 0.389*** [0.089] [0.086] [0.085] [0.088] [0.084] [0.091] Fin. Dev. × Firm Age − 0.101*** [0.030] Skill End. × Skill Int. 0.261*** 0.249*** 0.247*** 0.257*** 0.227*** 0.253*** 0.248*** [0.045] [0.044] [0.044] [0.045] [0.044] [0.046] [0.046] Cap. End. × Cap. Int. 0.053*** 0.057*** 0.059*** 0.061*** 0.060*** 0.054*** 0.057*** [0.009] [0.009] [0.009] [0.009] [0.008] [0.009] [0.009] Inst. Qual. × Contr. Int. 0.185 0.057 0.113 0.123 0.090 0.146 0.222 [0.145] [0.140] [0.140] [0.140] [0.138] [0.145] [0.144] No. of products 0.003*** 0.003*** 0.003*** 0.003*** 0.004*** 0.004*** 0.004*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] GDP × Ext. Fin. Dep. 0.003 − 0.002 0.002 [0.011] [0.012] [0.012] GDP × Ass. Tang. 0.126* 0.108 0.050 [0.065] [0.068] [0.066] Imp. Pen. × Ext. Fin. Dep. − 0.205** − 0.229*** − 0.230*** [0.089] [0.086] [0.086] Imp. Pen. × Ass. Tang. − 3.311*** − 3.329*** − 3.293*** [0.550] [0.561] [0.561] Exp. Int. × Ext. Fin. Dep. 0.225** 0.262*** 0.261*** [0.087] [0.085] [0.085] Exp. Int. × Ass. Tang. 2.656*** 2.598*** 2.548*** [0.545] [0.559] [0.559] Exch. Rate × Ext. Fin. Dep. 0.007 0.017 0.022 [0.024] [0.024] [0.024] Exch. Rate × Ass. Tang. 0.257 0.127 0.081 [0.160] [0.162] [0.164] FDI × Ext. Fin. Dep. 0.000 − 0.021 − 0.010 [0.016] [0.017] [0.017] FDI × Ass. Tang. − 0.184* 0.133 0.029 [0.103] [0.115] [0.113] Fin. Dev. × Numb. HS − 0.001*** − 0.001*** − 0.001*** [0.000] [0.000] [0.000] Observations 227,583 227,194 227,583 223,381 227,583 222,992 222,992 R-squared 0.25 0.25 0.25 0.25 0.25 0.25 0.25 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Per-capita GDP Import penetration and export intensity Real exchange rate Foreign direct investment Number of HS codes All controls All controls (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.035*** 0.032** 0.037*** 0.038*** 0.039*** 0.037*** [0.013] [0.013] [0.013] [0.014] [0.013] [0.014] Fin. Dev. × Ass. Tang. − 0.484*** − 0.355*** − 0.416*** − 0.378*** − 0.351*** − 0.389*** [0.089] [0.086] [0.085] [0.088] [0.084] [0.091] Fin. Dev. × Firm Age − 0.101*** [0.030] Skill End. × Skill Int. 0.261*** 0.249*** 0.247*** 0.257*** 0.227*** 0.253*** 0.248*** [0.045] [0.044] [0.044] [0.045] [0.044] [0.046] [0.046] Cap. End. × Cap. Int. 0.053*** 0.057*** 0.059*** 0.061*** 0.060*** 0.054*** 0.057*** [0.009] [0.009] [0.009] [0.009] [0.008] [0.009] [0.009] Inst. Qual. × Contr. Int. 0.185 0.057 0.113 0.123 0.090 0.146 0.222 [0.145] [0.140] [0.140] [0.140] [0.138] [0.145] [0.144] No. of products 0.003*** 0.003*** 0.003*** 0.003*** 0.004*** 0.004*** 0.004*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] GDP × Ext. Fin. Dep. 0.003 − 0.002 0.002 [0.011] [0.012] [0.012] GDP × Ass. Tang. 0.126* 0.108 0.050 [0.065] [0.068] [0.066] Imp. Pen. × Ext. Fin. Dep. − 0.205** − 0.229*** − 0.230*** [0.089] [0.086] [0.086] Imp. Pen. × Ass. Tang. − 3.311*** − 3.329*** − 3.293*** [0.550] [0.561] [0.561] Exp. Int. × Ext. Fin. Dep. 0.225** 0.262*** 0.261*** [0.087] [0.085] [0.085] Exp. Int. × Ass. Tang. 2.656*** 2.598*** 2.548*** [0.545] [0.559] [0.559] Exch. Rate × Ext. Fin. Dep. 0.007 0.017 0.022 [0.024] [0.024] [0.024] Exch. Rate × Ass. Tang. 0.257 0.127 0.081 [0.160] [0.162] [0.164] FDI × Ext. Fin. Dep. 0.000 − 0.021 − 0.010 [0.016] [0.017] [0.017] FDI × Ass. Tang. − 0.184* 0.133 0.029 [0.103] [0.115] [0.113] Fin. Dev. × Numb. HS − 0.001*** − 0.001*** − 0.001*** [0.000] [0.000] [0.000] Observations 227,583 227,194 227,583 223,381 227,583 222,992 222,992 R-squared 0.25 0.25 0.25 0.25 0.25 0.25 0.25 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. GDP is the real per-capita GDP of each country in each year. Import penetration and export intensity are the ratios of imports over apparent consumption (production plus imports minus exports) and of exports over GDP, respectively, in each country and year. The exchange rate is the PPP real exchange rate of each country, relative to the US dollar, in each year. FDI is the ratio of outward FDI over GDP in each country and year. The number of HS codes is the total number of 10-digit codes that belong to each 4-digit SIC industry according to the HS classification in each year. All time-varying regressors are lagged one period. All regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 7. Sales dispersion and finance—Additional controls. Per-capita GDP Import penetration and export intensity Real exchange rate Foreign direct investment Number of HS codes All controls All controls (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.035*** 0.032** 0.037*** 0.038*** 0.039*** 0.037*** [0.013] [0.013] [0.013] [0.014] [0.013] [0.014] Fin. Dev. × Ass. Tang. − 0.484*** − 0.355*** − 0.416*** − 0.378*** − 0.351*** − 0.389*** [0.089] [0.086] [0.085] [0.088] [0.084] [0.091] Fin. Dev. × Firm Age − 0.101*** [0.030] Skill End. × Skill Int. 0.261*** 0.249*** 0.247*** 0.257*** 0.227*** 0.253*** 0.248*** [0.045] [0.044] [0.044] [0.045] [0.044] [0.046] [0.046] Cap. End. × Cap. Int. 0.053*** 0.057*** 0.059*** 0.061*** 0.060*** 0.054*** 0.057*** [0.009] [0.009] [0.009] [0.009] [0.008] [0.009] [0.009] Inst. Qual. × Contr. Int. 0.185 0.057 0.113 0.123 0.090 0.146 0.222 [0.145] [0.140] [0.140] [0.140] [0.138] [0.145] [0.144] No. of products 0.003*** 0.003*** 0.003*** 0.003*** 0.004*** 0.004*** 0.004*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] GDP × Ext. Fin. Dep. 0.003 − 0.002 0.002 [0.011] [0.012] [0.012] GDP × Ass. Tang. 0.126* 0.108 0.050 [0.065] [0.068] [0.066] Imp. Pen. × Ext. Fin. Dep. − 0.205** − 0.229*** − 0.230*** [0.089] [0.086] [0.086] Imp. Pen. × Ass. Tang. − 3.311*** − 3.329*** − 3.293*** [0.550] [0.561] [0.561] Exp. Int. × Ext. Fin. Dep. 0.225** 0.262*** 0.261*** [0.087] [0.085] [0.085] Exp. Int. × Ass. Tang. 2.656*** 2.598*** 2.548*** [0.545] [0.559] [0.559] Exch. Rate × Ext. Fin. Dep. 0.007 0.017 0.022 [0.024] [0.024] [0.024] Exch. Rate × Ass. Tang. 0.257 0.127 0.081 [0.160] [0.162] [0.164] FDI × Ext. Fin. Dep. 0.000 − 0.021 − 0.010 [0.016] [0.017] [0.017] FDI × Ass. Tang. − 0.184* 0.133 0.029 [0.103] [0.115] [0.113] Fin. Dev. × Numb. HS − 0.001*** − 0.001*** − 0.001*** [0.000] [0.000] [0.000] Observations 227,583 227,194 227,583 223,381 227,583 222,992 222,992 R-squared 0.25 0.25 0.25 0.25 0.25 0.25 0.25 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Per-capita GDP Import penetration and export intensity Real exchange rate Foreign direct investment Number of HS codes All controls All controls (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.035*** 0.032** 0.037*** 0.038*** 0.039*** 0.037*** [0.013] [0.013] [0.013] [0.014] [0.013] [0.014] Fin. Dev. × Ass. Tang. − 0.484*** − 0.355*** − 0.416*** − 0.378*** − 0.351*** − 0.389*** [0.089] [0.086] [0.085] [0.088] [0.084] [0.091] Fin. Dev. × Firm Age − 0.101*** [0.030] Skill End. × Skill Int. 0.261*** 0.249*** 0.247*** 0.257*** 0.227*** 0.253*** 0.248*** [0.045] [0.044] [0.044] [0.045] [0.044] [0.046] [0.046] Cap. End. × Cap. Int. 0.053*** 0.057*** 0.059*** 0.061*** 0.060*** 0.054*** 0.057*** [0.009] [0.009] [0.009] [0.009] [0.008] [0.009] [0.009] Inst. Qual. × Contr. Int. 0.185 0.057 0.113 0.123 0.090 0.146 0.222 [0.145] [0.140] [0.140] [0.140] [0.138] [0.145] [0.144] No. of products 0.003*** 0.003*** 0.003*** 0.003*** 0.004*** 0.004*** 0.004*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] GDP × Ext. Fin. Dep. 0.003 − 0.002 0.002 [0.011] [0.012] [0.012] GDP × Ass. Tang. 0.126* 0.108 0.050 [0.065] [0.068] [0.066] Imp. Pen. × Ext. Fin. Dep. − 0.205** − 0.229*** − 0.230*** [0.089] [0.086] [0.086] Imp. Pen. × Ass. Tang. − 3.311*** − 3.329*** − 3.293*** [0.550] [0.561] [0.561] Exp. Int. × Ext. Fin. Dep. 0.225** 0.262*** 0.261*** [0.087] [0.085] [0.085] Exp. Int. × Ass. Tang. 2.656*** 2.598*** 2.548*** [0.545] [0.559] [0.559] Exch. Rate × Ext. Fin. Dep. 0.007 0.017 0.022 [0.024] [0.024] [0.024] Exch. Rate × Ass. Tang. 0.257 0.127 0.081 [0.160] [0.162] [0.164] FDI × Ext. Fin. Dep. 0.000 − 0.021 − 0.010 [0.016] [0.017] [0.017] FDI × Ass. Tang. − 0.184* 0.133 0.029 [0.103] [0.115] [0.113] Fin. Dev. × Numb. HS − 0.001*** − 0.001*** − 0.001*** [0.000] [0.000] [0.000] Observations 227,583 227,194 227,583 223,381 227,583 222,992 222,992 R-squared 0.25 0.25 0.25 0.25 0.25 0.25 0.25 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. GDP is the real per-capita GDP of each country in each year. Import penetration and export intensity are the ratios of imports over apparent consumption (production plus imports minus exports) and of exports over GDP, respectively, in each country and year. The exchange rate is the PPP real exchange rate of each country, relative to the US dollar, in each year. FDI is the ratio of outward FDI over GDP in each country and year. The number of HS codes is the total number of 10-digit codes that belong to each 4-digit SIC industry according to the HS classification in each year. All time-varying regressors are lagged one period. All regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large In columns (2)–(4), we add interactions between the measures of financial vulnerability and variables reflecting the degree of international integration and exposure to foreign competition of a country: import penetration and export intensity (column (2)); the real exchange rate (column (3)); and the ratio of outward FDI to GDP (column (4)).41 Including these variables does not make any noteworthy change in the main coefficients, suggesting that our estimates are not picking up the effects of different forms of international integration. In column (5), we interact financial development with the total number of HS codes that belong to a 4-digit SIC industry in a given year. One may worry that this number, which is determined by an administrative convention, may mechanically drive the measures of sales dispersion. Yet, including the new interaction leaves our main results unaffected. In column (6) we include all these controls in the same specification. Our main evidence is unchanged also in this demanding exercise. Finally, in column (7) we re-estimate the last specification using firm age to proxy for financial vulnerability. Our conclusions continue to hold. Other Issues. The previous sections suggest that our results are unlikely to reflect time-varying shocks occurring in a given country or industry, or the effects of many confounders that vary at the country–industry level. In this section, we discuss other potential identification issues. The first concern is that even the large set of controls used in Table 7 might fail to fully account for time-varying shocks hitting specific country–industry pairs. Although we cannot control for country–industry–year effects, in column (1) of Table 8 we introduce a full set of fixed effects for triplets of broad geographical areas, 3-digit industries and years.42 These fixed effects soak up all time-varying shocks hitting a certain 3-digit sector within a region. As a result, identification now only comes from the remaining variation in financial development across nearby countries, as well as from the remaining variation in financial vulnerability across narrow industries with similar technological content. Reassuringly, the coefficients remain similar to the baseline estimates also in this case. Table 8. Sales dispersion and finance—Other issues. Contemporaneous shocks Underlying trends Country–industry fixed effects Area-SIC3-year effects Based on initial dispersion Based on initial exports Based on initial no. of products All countries Countries with banking crises (1) (2) (3) (4) (5) (6) Fin. Dev. 0.040 0.045 [0.032] [0.042] Fin. Dev. × Ext. Fin. Dep. 0.030** 0.027** 0.036*** 0.038*** 0.026* 0.042* [0.014] [0.012] [0.013] [0.013] [0.015] [0.022] Fin. Dev. × Ass. Tang. − 0.307*** − 0.369*** − 0.393*** − 0.413*** − 0.259** − 0.257* [0.100] [0.076] [0.084] [0.085] [0.121] [0.155] Skill End. 0.281* 0.134 [0.152] [0.223] Cap. End. − 0.136 − 0.087 [0.108] [0.117] Skill End. × Skill Int. 0.216*** 0.199*** 0.251*** 0.242*** − 0.045 − 0.129 [0.048] [0.040] [0.044] [0.044] [0.135] [0.196] Cap. End. × Cap. Int. 0.053*** 0.048*** 0.058*** 0.059*** 0.039 0.022 [0.009] [0.007] [0.009] [0.009] [0.025] [0.028] Inst. Qual. × Contr. Int. − 0.654*** 0.133 0.113 0.107 [0.191] [0.125] [0.139] [0.139] No. of products 0.003*** 0.002*** 0.003*** 0.004*** 0.003*** 0.003*** [0.000] [0.000] [0.000] [0.001] [0.000] [0.001] Observations 227,583 227,583 227,583 227,583 227,583 148,940 R-squared 0.31 0.32 0.25 0.25 0.57 0.59 Country–Year FE Yes Yes Yes Yes No No Industry–Year FE Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Country–Industry trends No Yes Yes Yes No No Area-SIC3-Year FE Yes No No No No No Country–Industry FE No No No No Yes Yes Contemporaneous shocks Underlying trends Country–industry fixed effects Area-SIC3-year effects Based on initial dispersion Based on initial exports Based on initial no. of products All countries Countries with banking crises (1) (2) (3) (4) (5) (6) Fin. Dev. 0.040 0.045 [0.032] [0.042] Fin. Dev. × Ext. Fin. Dep. 0.030** 0.027** 0.036*** 0.038*** 0.026* 0.042* [0.014] [0.012] [0.013] [0.013] [0.015] [0.022] Fin. Dev. × Ass. Tang. − 0.307*** − 0.369*** − 0.393*** − 0.413*** − 0.259** − 0.257* [0.100] [0.076] [0.084] [0.085] [0.121] [0.155] Skill End. 0.281* 0.134 [0.152] [0.223] Cap. End. − 0.136 − 0.087 [0.108] [0.117] Skill End. × Skill Int. 0.216*** 0.199*** 0.251*** 0.242*** − 0.045 − 0.129 [0.048] [0.040] [0.044] [0.044] [0.135] [0.196] Cap. End. × Cap. Int. 0.053*** 0.048*** 0.058*** 0.059*** 0.039 0.022 [0.009] [0.007] [0.009] [0.009] [0.025] [0.028] Inst. Qual. × Contr. Int. − 0.654*** 0.133 0.113 0.107 [0.191] [0.125] [0.139] [0.139] No. of products 0.003*** 0.002*** 0.003*** 0.004*** 0.003*** 0.003*** [0.000] [0.000] [0.000] [0.001] [0.000] [0.001] Observations 227,583 227,583 227,583 227,583 227,583 148,940 R-squared 0.31 0.32 0.25 0.25 0.57 0.59 Country–Year FE Yes Yes Yes Yes No No Industry–Year FE Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Country–Industry trends No Yes Yes Yes No No Area-SIC3-Year FE Yes No No No No No Country–Industry FE No No No No Yes Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. Column (1) controls for contemporaneous shocks. To this purpose, it includes a full set of interactions between the year dummies, dummies for 3-digit SIC industries, and seven dummies for geographical areas, as defined by the World Bank: East Asia and Pacific; Europe and Central Asia; Latin America and the Caribbean; Middle East and North Africa; North America; South Asia; and Sub-Saharan Africa. Columns (2)–(4) control for underlying trends based on pre-existing characteristics of each country–industry pair. To this purpose, each column includes a full set of interactions between the year dummies and the initial (first year) value of the characteristic indicated in the column’s heading. Columns (5) and (6) control for time-invariant country–industry characteristics. To this purpose, each column includes country–industry fixed effects: column (5) uses the whole sample of countries, whereas column (6) uses the subsample of countries that have experienced at least one banking crisis over 1989–2006. All time-varying regressors are lagged one period. Except for column (6), the regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 8. Sales dispersion and finance—Other issues. Contemporaneous shocks Underlying trends Country–industry fixed effects Area-SIC3-year effects Based on initial dispersion Based on initial exports Based on initial no. of products All countries Countries with banking crises (1) (2) (3) (4) (5) (6) Fin. Dev. 0.040 0.045 [0.032] [0.042] Fin. Dev. × Ext. Fin. Dep. 0.030** 0.027** 0.036*** 0.038*** 0.026* 0.042* [0.014] [0.012] [0.013] [0.013] [0.015] [0.022] Fin. Dev. × Ass. Tang. − 0.307*** − 0.369*** − 0.393*** − 0.413*** − 0.259** − 0.257* [0.100] [0.076] [0.084] [0.085] [0.121] [0.155] Skill End. 0.281* 0.134 [0.152] [0.223] Cap. End. − 0.136 − 0.087 [0.108] [0.117] Skill End. × Skill Int. 0.216*** 0.199*** 0.251*** 0.242*** − 0.045 − 0.129 [0.048] [0.040] [0.044] [0.044] [0.135] [0.196] Cap. End. × Cap. Int. 0.053*** 0.048*** 0.058*** 0.059*** 0.039 0.022 [0.009] [0.007] [0.009] [0.009] [0.025] [0.028] Inst. Qual. × Contr. Int. − 0.654*** 0.133 0.113 0.107 [0.191] [0.125] [0.139] [0.139] No. of products 0.003*** 0.002*** 0.003*** 0.004*** 0.003*** 0.003*** [0.000] [0.000] [0.000] [0.001] [0.000] [0.001] Observations 227,583 227,583 227,583 227,583 227,583 148,940 R-squared 0.31 0.32 0.25 0.25 0.57 0.59 Country–Year FE Yes Yes Yes Yes No No Industry–Year FE Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Country–Industry trends No Yes Yes Yes No No Area-SIC3-Year FE Yes No No No No No Country–Industry FE No No No No Yes Yes Contemporaneous shocks Underlying trends Country–industry fixed effects Area-SIC3-year effects Based on initial dispersion Based on initial exports Based on initial no. of products All countries Countries with banking crises (1) (2) (3) (4) (5) (6) Fin. Dev. 0.040 0.045 [0.032] [0.042] Fin. Dev. × Ext. Fin. Dep. 0.030** 0.027** 0.036*** 0.038*** 0.026* 0.042* [0.014] [0.012] [0.013] [0.013] [0.015] [0.022] Fin. Dev. × Ass. Tang. − 0.307*** − 0.369*** − 0.393*** − 0.413*** − 0.259** − 0.257* [0.100] [0.076] [0.084] [0.085] [0.121] [0.155] Skill End. 0.281* 0.134 [0.152] [0.223] Cap. End. − 0.136 − 0.087 [0.108] [0.117] Skill End. × Skill Int. 0.216*** 0.199*** 0.251*** 0.242*** − 0.045 − 0.129 [0.048] [0.040] [0.044] [0.044] [0.135] [0.196] Cap. End. × Cap. Int. 0.053*** 0.048*** 0.058*** 0.059*** 0.039 0.022 [0.009] [0.007] [0.009] [0.009] [0.025] [0.028] Inst. Qual. × Contr. Int. − 0.654*** 0.133 0.113 0.107 [0.191] [0.125] [0.139] [0.139] No. of products 0.003*** 0.002*** 0.003*** 0.004*** 0.003*** 0.003*** [0.000] [0.000] [0.000] [0.001] [0.000] [0.001] Observations 227,583 227,583 227,583 227,583 227,583 148,940 R-squared 0.31 0.32 0.25 0.25 0.57 0.59 Country–Year FE Yes Yes Yes Yes No No Industry–Year FE Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Country–Industry trends No Yes Yes Yes No No Area-SIC3-Year FE Yes No No No No No Country–Industry FE No No No No Yes Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. Column (1) controls for contemporaneous shocks. To this purpose, it includes a full set of interactions between the year dummies, dummies for 3-digit SIC industries, and seven dummies for geographical areas, as defined by the World Bank: East Asia and Pacific; Europe and Central Asia; Latin America and the Caribbean; Middle East and North Africa; North America; South Asia; and Sub-Saharan Africa. Columns (2)–(4) control for underlying trends based on pre-existing characteristics of each country–industry pair. To this purpose, each column includes a full set of interactions between the year dummies and the initial (first year) value of the characteristic indicated in the column’s heading. Columns (5) and (6) control for time-invariant country–industry characteristics. To this purpose, each column includes country–industry fixed effects: column (5) uses the whole sample of countries, whereas column (6) uses the subsample of countries that have experienced at least one banking crisis over 1989–2006. All time-varying regressors are lagged one period. Except for column (6), the regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large The second concern is that our estimates may be driven by differential trends across country–industry pairs. In columns (2)–(4), we therefore control for underlying trends based on pre-existing characteristics of each pair. To this purpose, we interact the time dummies with the first-year value of the characteristic indicated in each column. The coefficients are stable across the board. The third concern is that our results may be contaminated by unobserved, time-invariant, heterogeneity across country–industry pairs. In columns (5) and (6), we address this concern by exploiting the panel structure of the data and including country–industry fixed effects in place of the country–year effects. Compared to previous specifications, we therefore exploit a different source of variation, which is provided by changes in financial development and factor endowments over time within a country, rather than by differences in these variables across countries. Accordingly, this approach is not well-suited to study the effects of export opportunities, because a proper test of comparative advantage requires comparing different countries, as we do in our main specifications. On the contrary, this alternative approach is still well-suited to test the effect of financial frictions, as our theoretical mechanism predicts that sales dispersion should increase after an improvement in financial conditions within a country. We report results for both the whole sample of countries (column (5)) and the subsample of economies that have experienced a banking crisis during the sample period (column (6)).43 For the latter countries, changes in private credit have been larger, thereby providing us with greater time variation for identification. Reassuringly, our evidence is unchanged also in these very demanding specifications. Cross-Sectional and IV Estimates. Finally, we present a set of cross-sectional results, which are obtained by replacing all time-varying variables with their long-run mean over 1989–2006. These regressions further ensure that our main coefficients are not contaminated by temporary shocks. The results are reported in Table 9. In spite of a dramatic loss of observations, the coefficients shown in column (1) are similar to the baseline panel estimates. In column (2), we compare the results based on private credit with those obtained using an index for the quality of institutions that affect credit access. In particular, we use an index for the effectiveness of the legal system at resolving insolvencies.44 This index is time invariant, and can thus be meaningfully used only in a cross-sectional set-up. The results confirm our baseline evidence. In column (3), we re-run the regression reported in column (1), but we exclude the country fixed effects. Unlike the previous specifications, this one allows us to identify the linear term in financial development, and can thus be used to quantify the overall effect of financial frictions on sales dispersion, besides their differential effect across industries (see the next section). The results are close to those reported in column (1). Table 9. Sales dispersion and finance—Cross-sectional results. OLS IV Private credit Resolving insolvencies Private credit Private credit Resolving insolvencies Private credit (1) (2) (3) (4) (5) (6) Fin. Dev. 0.342*** 0.567*** [0.052] [0.092] Fin. Dev. × Ext. Fin. Dep. 0.071*** 0.101*** 0.066*** 0.089*** 0.188*** 0.078*** [0.015] [0.034] [0.016] [0.030] [0.050] [0.030] Fin. Dev. × Ass. Tang. − 0.320*** − 0.368* − 0.345** − 0.589** − 0.862* − 0.835*** [0.120] [0.211] [0.145] [0.251] [0.497] [0.263] Skill End. 0.453*** 0.425*** [0.065] [0.067] Cap. End. − 0.103*** − 0.126*** [0.033] [0.033] Inst. Qual. − 0.413*** − 0.414*** [0.105] [0.117] Skill End. × Skill Int. 0.262*** 0.241*** 0.216*** 0.245*** 0.203*** 0.189*** [0.052] [0.052] [0.051] [0.055] [0.057] [0.053] Cap. End. × Cap. Int. 0.055*** 0.051*** 0.039*** 0.057*** 0.053*** 0.043*** [0.008] [0.009] [0.008] [0.008] [0.009] [0.008] Inst. Qual. × Contr. Int. 0.436** 0.493*** 0.365** 0.333* 0.369** 0.188 [0.172] [0.169] [0.166] [0.180] [0.183] [0.173] No. of products 0.003*** 0.003*** 0.010*** 0.003*** 0.003*** 0.009*** [0.001] [0.001] [0.002] [0.001] [0.001] [0.002] Observations 20,716 20,952 20,716 20,716 20,952 20,716 R-squared 0.36 0.36 0.22 0.29 0.29 0.14 Country FE Yes Yes No Yes Yes No Industry FE Yes Yes Yes Yes Yes Yes First-stage results Kleibergen–Paap F-Statistic − − − 467.2 194.4 555.3 OLS IV Private credit Resolving insolvencies Private credit Private credit Resolving insolvencies Private credit (1) (2) (3) (4) (5) (6) Fin. Dev. 0.342*** 0.567*** [0.052] [0.092] Fin. Dev. × Ext. Fin. Dep. 0.071*** 0.101*** 0.066*** 0.089*** 0.188*** 0.078*** [0.015] [0.034] [0.016] [0.030] [0.050] [0.030] Fin. Dev. × Ass. Tang. − 0.320*** − 0.368* − 0.345** − 0.589** − 0.862* − 0.835*** [0.120] [0.211] [0.145] [0.251] [0.497] [0.263] Skill End. 0.453*** 0.425*** [0.065] [0.067] Cap. End. − 0.103*** − 0.126*** [0.033] [0.033] Inst. Qual. − 0.413*** − 0.414*** [0.105] [0.117] Skill End. × Skill Int. 0.262*** 0.241*** 0.216*** 0.245*** 0.203*** 0.189*** [0.052] [0.052] [0.051] [0.055] [0.057] [0.053] Cap. End. × Cap. Int. 0.055*** 0.051*** 0.039*** 0.057*** 0.053*** 0.043*** [0.008] [0.009] [0.008] [0.008] [0.009] [0.008] Inst. Qual. × Contr. Int. 0.436** 0.493*** 0.365** 0.333* 0.369** 0.188 [0.172] [0.169] [0.166] [0.180] [0.183] [0.173] No. of products 0.003*** 0.003*** 0.010*** 0.003*** 0.003*** 0.009*** [0.001] [0.001] [0.002] [0.001] [0.001] [0.002] Observations 20,716 20,952 20,716 20,716 20,952 20,716 R-squared 0.36 0.36 0.22 0.29 0.29 0.14 Country FE Yes Yes No Yes Yes No Industry FE Yes Yes Yes Yes Yes Yes First-stage results Kleibergen–Paap F-Statistic − − − 467.2 194.4 555.3 Notes: The dependent variable is sales dispersion (the standard deviation of log exports) for each exporting country and industry, computed with data on exports to the United States at the 10-digit product level, and averaged over 1989–2006. Financial development is proxied by private credit in columns (1), (3), (4), and (6), and by an index of insolvencies resolutions in columns (2) and (5). Private credit, factor endowments, and the number of products are averaged over 1989–2006. The index of insolvencies resolutions is normalized between 0 and 1, and takes higher values for countries occupying higher positions in the ranking. In columns (4)–(6), financial development is instrumented using dummies for whether countries’ legal systems are of civil law (French, German, or Scandinavian origins). All regressions are based on a consistent sample of countries (119) and 4-digit SIC industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for clustering by industry. The F-statistics are reported for the Kleibergen–Paap test for weak identification. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 9. Sales dispersion and finance—Cross-sectional results. OLS IV Private credit Resolving insolvencies Private credit Private credit Resolving insolvencies Private credit (1) (2) (3) (4) (5) (6) Fin. Dev. 0.342*** 0.567*** [0.052] [0.092] Fin. Dev. × Ext. Fin. Dep. 0.071*** 0.101*** 0.066*** 0.089*** 0.188*** 0.078*** [0.015] [0.034] [0.016] [0.030] [0.050] [0.030] Fin. Dev. × Ass. Tang. − 0.320*** − 0.368* − 0.345** − 0.589** − 0.862* − 0.835*** [0.120] [0.211] [0.145] [0.251] [0.497] [0.263] Skill End. 0.453*** 0.425*** [0.065] [0.067] Cap. End. − 0.103*** − 0.126*** [0.033] [0.033] Inst. Qual. − 0.413*** − 0.414*** [0.105] [0.117] Skill End. × Skill Int. 0.262*** 0.241*** 0.216*** 0.245*** 0.203*** 0.189*** [0.052] [0.052] [0.051] [0.055] [0.057] [0.053] Cap. End. × Cap. Int. 0.055*** 0.051*** 0.039*** 0.057*** 0.053*** 0.043*** [0.008] [0.009] [0.008] [0.008] [0.009] [0.008] Inst. Qual. × Contr. Int. 0.436** 0.493*** 0.365** 0.333* 0.369** 0.188 [0.172] [0.169] [0.166] [0.180] [0.183] [0.173] No. of products 0.003*** 0.003*** 0.010*** 0.003*** 0.003*** 0.009*** [0.001] [0.001] [0.002] [0.001] [0.001] [0.002] Observations 20,716 20,952 20,716 20,716 20,952 20,716 R-squared 0.36 0.36 0.22 0.29 0.29 0.14 Country FE Yes Yes No Yes Yes No Industry FE Yes Yes Yes Yes Yes Yes First-stage results Kleibergen–Paap F-Statistic − − − 467.2 194.4 555.3 OLS IV Private credit Resolving insolvencies Private credit Private credit Resolving insolvencies Private credit (1) (2) (3) (4) (5) (6) Fin. Dev. 0.342*** 0.567*** [0.052] [0.092] Fin. Dev. × Ext. Fin. Dep. 0.071*** 0.101*** 0.066*** 0.089*** 0.188*** 0.078*** [0.015] [0.034] [0.016] [0.030] [0.050] [0.030] Fin. Dev. × Ass. Tang. − 0.320*** − 0.368* − 0.345** − 0.589** − 0.862* − 0.835*** [0.120] [0.211] [0.145] [0.251] [0.497] [0.263] Skill End. 0.453*** 0.425*** [0.065] [0.067] Cap. End. − 0.103*** − 0.126*** [0.033] [0.033] Inst. Qual. − 0.413*** − 0.414*** [0.105] [0.117] Skill End. × Skill Int. 0.262*** 0.241*** 0.216*** 0.245*** 0.203*** 0.189*** [0.052] [0.052] [0.051] [0.055] [0.057] [0.053] Cap. End. × Cap. Int. 0.055*** 0.051*** 0.039*** 0.057*** 0.053*** 0.043*** [0.008] [0.009] [0.008] [0.008] [0.009] [0.008] Inst. Qual. × Contr. Int. 0.436** 0.493*** 0.365** 0.333* 0.369** 0.188 [0.172] [0.169] [0.166] [0.180] [0.183] [0.173] No. of products 0.003*** 0.003*** 0.010*** 0.003*** 0.003*** 0.009*** [0.001] [0.001] [0.002] [0.001] [0.001] [0.002] Observations 20,716 20,952 20,716 20,716 20,952 20,716 R-squared 0.36 0.36 0.22 0.29 0.29 0.14 Country FE Yes Yes No Yes Yes No Industry FE Yes Yes Yes Yes Yes Yes First-stage results Kleibergen–Paap F-Statistic − − − 467.2 194.4 555.3 Notes: The dependent variable is sales dispersion (the standard deviation of log exports) for each exporting country and industry, computed with data on exports to the United States at the 10-digit product level, and averaged over 1989–2006. Financial development is proxied by private credit in columns (1), (3), (4), and (6), and by an index of insolvencies resolutions in columns (2) and (5). Private credit, factor endowments, and the number of products are averaged over 1989–2006. The index of insolvencies resolutions is normalized between 0 and 1, and takes higher values for countries occupying higher positions in the ranking. In columns (4)–(6), financial development is instrumented using dummies for whether countries’ legal systems are of civil law (French, German, or Scandinavian origins). All regressions are based on a consistent sample of countries (119) and 4-digit SIC industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for clustering by industry. The F-statistics are reported for the Kleibergen–Paap test for weak identification. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Finally, we discuss possibly remaining concerns with endogeneity. As previously shown, our coefficients are robust to controlling for a wide range of factors, suggesting that our evidence is unlikely to reflect simultaneity bias due to omitted variables. Other features of the empirical set-up help allay concerns with reverse causality. The latter would occur if sales dispersion increased in a given country and industry for reasons unrelated to financial development, and if this, in turn, affected the financial variables in a way that could explain the specific pattern of our coefficients. Note, however, that the financial vulnerability measures are based on US data and kept constant over time. Thus, these measures are unlikely to respond to changes in sales dispersion occurring in specific countries and industries. Second, we have shown that our results are unchanged across alternative financial vulnerability measures, and when using proxies based on data for the previous decade. It is unlikely that changes in sales dispersion over the sample period could drive the variation in all of these alternative indicators. Third, our results are robust across a battery of proxies for financial development; we believe it is unlikely that an omitted shock could move all these variables equally and simultaneously. Finally, our results hold when using long-run averages of private credit and a time-invariant index for the quality of financial institutions, which are unlikely to respond to changes in sales dispersion in a given year. Yet, we now show that our evidence is also preserved when using instrumental variables (IV). The latter allow us to isolate the variation in financial development due to countries’ historical conditions, whereas cleaning up the variation due to current economic conditions potentially correlated with sales dispersion. The results are reported in columns (4)–(6) of Table 9. Following La Porta, Lopez-de-Silanes, and Shleifer (2008), we instrument the proxies for financial development using dummies for whether countries’ legal systems are of civil law (French, German, or Scandinavian origins). Consistent with La Porta et al. (2008), we find the nature of countries’ legal systems to be a strong predictor of financial development, suggesting that differences in financial frictions across countries to a large extent reflect historical differences in countries’ legal origins. More importantly, our main evidence is preserved also in these specifications. 4.3.4. Economic Magnitude We now quantify the effect of financial development on sales dispersion. To this purpose, we use the estimates reported in column (3) of Table 9 and study by how much sales dispersion would change following a certain increase in private credit. We start from the average effect, that is, the effect on the industry with the average levels of financial vulnerability. Our estimates imply that an increase in private credit from the 25th percentile of the distribution (17%, roughly the level of Peru) to the 75th percentile (68%, approximately the level of South Korea) would raise sales dispersion by 12.9% on average. For comparison, a commensurate increase in skill (capital) endowment would raise sales dispersion by 8% (15%) in the average industry. The effects of financial development are therefore in the same ballpark as those of export opportunities. These estimates also imply that the observed increase in private credit over the sample period (15 p.p.) could explain 59% of the increase in sales dispersion between 1989 and 2006. Next, we turn to the differential effect of financial development across industries with different levels of financial vulnerability. Our estimates imply that an increase in private credit from the 25th to the 75th percentile would raise sales dispersion by 11.7% in the industry at the first quartile of the distribution by external finance, and by 13.5% in the industry at the third quartile. The same increase in private credit would raise sales dispersion by 11.4% in the industry at the third quartile of the distribution by asset tangibility, and by 14.5% in the industry at the first quartile. 4.4. Trade, Finance, and Sales Dispersion The previous sections have shown that financial development increases sales dispersion especially in financially vulnerable industries. In turn, according to our model, higher sales dispersion should raise both the number of exported products (extensive margin) and exports per product (intensive margin), thereby increasing overall exports. It follows that sales dispersion provides a mechanisms through which financial development could affect export flows across countries and industries. We now provide some evidence on this mechanism. The results are reported in Table 10. In columns (1)–(3), we start by studying how sales dispersion correlates with overall exports and the two margins of trade. To this purpose, we regress log total exports, log number of exported products and log exports per product, respectively, on sales dispersion, controlling for country–year and industry–year effects, as well as for the interactions between countries’ CPI and industry dummies. All coefficients are positive and very precisely estimated. Consistent with our model, greater sales dispersion in a given country and industry is associated with larger exports to the United States, more exported products and greater exports per product. In columns (4)–(6) we replace sales dispersion with its main determinants according to our model and previous empirical results; namely, with the interaction between financial development and financial vulnerability, as well as with export opportunities. The results confirm the well-known fact that financial development increases exports relatively more in financially vulnerable sectors (Beck 2002; Manova 2013), as well as the standard view that countries with larger endowments of skilled labor and capital, or with better institutional quality, export relatively more in industries that are skill and capital intensive, or dependent on relationship-specific investments (Romalis 2004; Levchenko 2007; Nunn 2007; Chor 2010). Finally, in columns (7)–(9) we include all variables simultaneously. The coefficients on sales dispersion remain unchanged, whereas those on financial development and export opportunities drop in size, suggesting that part of the effect of these variables on exports works through the dispersion of sales. Table 10. Trade, finance, and sales dispersion. Total exports No. of products Exports per product Total exports No. of products Exports per product Total exports No. of products Exports per product (1) (2) (3) (4) (5) (6) (7) (8) (9) Sales dispersion 1.635*** 0.185*** 1.450*** 1.632*** 0.183*** 1.449*** [0.014] [0.004] [0.012] [0.015] [0.004] [0.012] Fin. Dev. × Ext. Fin. Dep. 0.141*** 0.040*** 0.101** 0.076** 0.033*** 0.043 [0.049] [0.012] [0.041] [0.035] [0.011] [0.027] Fin. Dev. × Ass. Tang. − 3.153*** − 0.368*** − 2.786*** − 2.504*** − 0.295*** − 2.209*** [0.348] [0.098] [0.290] [0.265] [0.093] [0.213] Skill End. × Skill Int. 1.460*** 0.442*** 1.018*** 1.043*** 0.395*** 0.648*** [0.177] [0.049] [0.145] [0.126] [0.046] [0.098] Cap. End. × Cap. Int. 0.235*** 0.035*** 0.201*** 0.135*** 0.024** 0.111*** [0.035] [0.010] [0.028] [0.026] [0.009] [0.019] Inst. Qual. × Contr. Int. 1.847*** 0.858*** 0.989** 1.579*** 0.828*** 0.751** [0.536] [0.146] [0.439] [0.388] [0.138] [0.301] Observations 259,309 259,309 259,309 229,128 229,128 229,128 229,128 229,128 229,128 R-squared 0.72 0.75 0.69 0.55 0.73 0.48 0.73 0.75 0.70 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Total exports No. of products Exports per product Total exports No. of products Exports per product Total exports No. of products Exports per product (1) (2) (3) (4) (5) (6) (7) (8) (9) Sales dispersion 1.635*** 0.185*** 1.450*** 1.632*** 0.183*** 1.449*** [0.014] [0.004] [0.012] [0.015] [0.004] [0.012] Fin. Dev. × Ext. Fin. Dep. 0.141*** 0.040*** 0.101** 0.076** 0.033*** 0.043 [0.049] [0.012] [0.041] [0.035] [0.011] [0.027] Fin. Dev. × Ass. Tang. − 3.153*** − 0.368*** − 2.786*** − 2.504*** − 0.295*** − 2.209*** [0.348] [0.098] [0.290] [0.265] [0.093] [0.213] Skill End. × Skill Int. 1.460*** 0.442*** 1.018*** 1.043*** 0.395*** 0.648*** [0.177] [0.049] [0.145] [0.126] [0.046] [0.098] Cap. End. × Cap. Int. 0.235*** 0.035*** 0.201*** 0.135*** 0.024** 0.111*** [0.035] [0.010] [0.028] [0.026] [0.009] [0.019] Inst. Qual. × Contr. Int. 1.847*** 0.858*** 0.989** 1.579*** 0.828*** 0.751** [0.536] [0.146] [0.439] [0.388] [0.138] [0.301] Observations 259,309 259,309 259,309 229,128 229,128 229,128 229,128 229,128 229,128 R-squared 0.72 0.75 0.69 0.55 0.73 0.48 0.73 0.75 0.70 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Notes: The dependent variables are indicated in columns’ headings and are all expressed in logs. Sales dispersion is the standard deviation of log exports, computed separately for each exporting country, industry and year, using data on exports to the United States at the 10-digit product level. All time-varying regressors are lagged one period. All regressions are based on a consistent sample of countries (119) and 4-digit SIC industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables.**Significant at 5%; ***significant at 1%. View Large Table 10. Trade, finance, and sales dispersion. Total exports No. of products Exports per product Total exports No. of products Exports per product Total exports No. of products Exports per product (1) (2) (3) (4) (5) (6) (7) (8) (9) Sales dispersion 1.635*** 0.185*** 1.450*** 1.632*** 0.183*** 1.449*** [0.014] [0.004] [0.012] [0.015] [0.004] [0.012] Fin. Dev. × Ext. Fin. Dep. 0.141*** 0.040*** 0.101** 0.076** 0.033*** 0.043 [0.049] [0.012] [0.041] [0.035] [0.011] [0.027] Fin. Dev. × Ass. Tang. − 3.153*** − 0.368*** − 2.786*** − 2.504*** − 0.295*** − 2.209*** [0.348] [0.098] [0.290] [0.265] [0.093] [0.213] Skill End. × Skill Int. 1.460*** 0.442*** 1.018*** 1.043*** 0.395*** 0.648*** [0.177] [0.049] [0.145] [0.126] [0.046] [0.098] Cap. End. × Cap. Int. 0.235*** 0.035*** 0.201*** 0.135*** 0.024** 0.111*** [0.035] [0.010] [0.028] [0.026] [0.009] [0.019] Inst. Qual. × Contr. Int. 1.847*** 0.858*** 0.989** 1.579*** 0.828*** 0.751** [0.536] [0.146] [0.439] [0.388] [0.138] [0.301] Observations 259,309 259,309 259,309 229,128 229,128 229,128 229,128 229,128 229,128 R-squared 0.72 0.75 0.69 0.55 0.73 0.48 0.73 0.75 0.70 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Total exports No. of products Exports per product Total exports No. of products Exports per product Total exports No. of products Exports per product (1) (2) (3) (4) (5) (6) (7) (8) (9) Sales dispersion 1.635*** 0.185*** 1.450*** 1.632*** 0.183*** 1.449*** [0.014] [0.004] [0.012] [0.015] [0.004] [0.012] Fin. Dev. × Ext. Fin. Dep. 0.141*** 0.040*** 0.101** 0.076** 0.033*** 0.043 [0.049] [0.012] [0.041] [0.035] [0.011] [0.027] Fin. Dev. × Ass. Tang. − 3.153*** − 0.368*** − 2.786*** − 2.504*** − 0.295*** − 2.209*** [0.348] [0.098] [0.290] [0.265] [0.093] [0.213] Skill End. × Skill Int. 1.460*** 0.442*** 1.018*** 1.043*** 0.395*** 0.648*** [0.177] [0.049] [0.145] [0.126] [0.046] [0.098] Cap. End. × Cap. Int. 0.235*** 0.035*** 0.201*** 0.135*** 0.024** 0.111*** [0.035] [0.010] [0.028] [0.026] [0.009] [0.019] Inst. Qual. × Contr. Int. 1.847*** 0.858*** 0.989** 1.579*** 0.828*** 0.751** [0.536] [0.146] [0.439] [0.388] [0.138] [0.301] Observations 259,309 259,309 259,309 229,128 229,128 229,128 229,128 229,128 229,128 R-squared 0.72 0.75 0.69 0.55 0.73 0.48 0.73 0.75 0.70 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Notes: The dependent variables are indicated in columns’ headings and are all expressed in logs. Sales dispersion is the standard deviation of log exports, computed separately for each exporting country, industry and year, using data on exports to the United States at the 10-digit product level. All time-varying regressors are lagged one period. All regressions are based on a consistent sample of countries (119) and 4-digit SIC industries (365) with positive exports to the United States in all years between 1989 and 2006. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables.**Significant at 5%; ***significant at 1%. View Large 4.5. Sales Dispersion, Finance, and Innovation In this final section, we provide some evidence on the mechanism through which financial development affects sales dispersion. In the model, firm heterogeneity depends on the innovation strategies chosen by firms. Financial development induces firms to invest in bigger projects with more dispersed outcomes. This translates into a larger share of revenue invested in innovation and a higher incidence of “major” innovations: for any cutoff x, $$\Pr (\varphi >x)$$ increases with v. Are these predictions consistent with the data? To answer this question, we need comparable measures of investment in major innovations across countries, sectors and time, which are not easy to come by. Once again, however, we can overcome the challenge relying on high-quality US data. In particular, we use the number of utility patents applied for at the US Patent Office (USPTO), computed separately for each foreign country, industry and application year. We source the raw patent data from the NBER Patent Data Project. Between 1989 and 2006, a total of 898,589 patents were applied for by foreign entities at the USPTO. These patents belong to 2,183 technology classes, defined according to the International Patent Classification. We map these technology classes into SIC industries using a correspondence table developed by Silverman (1999). Patenting is a relatively rare activity, which is typically concentrated in few countries. For instance, only 49 of the 119 countries in our sample have applied for patents between 1989 and 2006. As a consequence, approximately 80% of the country–industry–year triplets in our sample have zero patent count.45 On the other hand, a unique feature of the USPTO data is that they provide a measure of innovation that is easy to compute and comparable across countries and industries. Another advantage of this measure is that, since only significant innovations are patented in the United States, foreign patent applications can be taken as a reasonable proxy for major innovations. We start by showing that sales dispersion is positively correlated with innovation, as predicted by the model. To this purpose, we regress sales dispersion on patent count across country–industry–year triplets, controlling for country–year fixed effects, industry–year fixed effects and the interactions between countries’ CPI and industry dummies. The results are reported in Table 11. In column (1), we use the whole sample, whereas in column (2) we restrict to the subsample of observations with positive patent count. Finally, in column (3) we replace the country–year effects with country–industry effects, in order to exploit time variation within country–industry pairs for identification. In all cases, the coefficient on patent count is positive and precisely estimated. Although we cannot make any claim regarding causality, this evidence is nevertheless consistent with the model. Table 11. Sales dispersion and innovation. (1) (2) (3) Patent count 0.516*** 0.409*** 0.201** [0.104] [0.099] [0.080] Observations 259,309 54,884 259,309 R-squared 0.25 0.38 0.56 Country–Year FE Yes Yes No Industry–Year FE Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Country–Industry FE No No Yes (1) (2) (3) Patent count 0.516*** 0.409*** 0.201** [0.104] [0.099] [0.080] Observations 259,309 54,884 259,309 R-squared 0.25 0.38 0.56 Country–Year FE Yes Yes No Industry–Year FE Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Country–Industry FE No No Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. Patent count is the number of patents registered at the USPTO in thousands, computed separately for each country, industry and application year. The regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Columns (1) and (3) use the whole sample of observations, whereas column (2) restricts to the subsample of observations with positive patent count. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables. **Significant at 5%; ***significant at 1%. View Large Table 11. Sales dispersion and innovation. (1) (2) (3) Patent count 0.516*** 0.409*** 0.201** [0.104] [0.099] [0.080] Observations 259,309 54,884 259,309 R-squared 0.25 0.38 0.56 Country–Year FE Yes Yes No Industry–Year FE Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Country–Industry FE No No Yes (1) (2) (3) Patent count 0.516*** 0.409*** 0.201** [0.104] [0.099] [0.080] Observations 259,309 54,884 259,309 R-squared 0.25 0.38 0.56 Country–Year FE Yes Yes No Industry–Year FE Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Country–Industry FE No No Yes Notes: The dependent variable is sales dispersion (the standard deviation of log exports), computed separately for each exporting country, industry, and year, using data on exports to the United States at the 10-digit product level. Patent count is the number of patents registered at the USPTO in thousands, computed separately for each country, industry and application year. The regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Columns (1) and (3) use the whole sample of observations, whereas column (2) restricts to the subsample of observations with positive patent count. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year. See also notes to previous tables. **Significant at 5%; ***significant at 1%. View Large Next and more importantly, we study how financial frictions affect innovation. To this purpose, we estimate the baseline specification (see column (5) of Table 3) using patent count instead of sales dispersion as the dependent variable. The results are reported in Table 12. In column (1), we use the whole sample of observations. Consistent with the model, we find that financial development increases innovation relatively more in financially vulnerable industries. In column (2), we restrict to the subsample of observations with positive patent count. The coefficients have the same sign as in column (1), and are now even larger. Although one coefficient is marginally insignificant (p-value 0.135), this reflects the reduced sample size. Indeed, when the two interactions are included individually rather than jointly (columns (3) and (4)), both coefficients regain significance and maintain their size. In column (5), we alternatively deal with the presence of zeros in the patent count variable by using a zero-inflated Poisson model. The coefficients have the same sign as before and are both highly significant. In columns (6) and (7), we use firm age as a proxy for financial vulnerability, focusing on the whole sample and on the subsample of observations with positive patent count, respectively. The results confirm that financial development raises innovation relatively more in financially vulnerable industries. Table 12. Determinants of innovation—Panel regressions. Baseline Zero-inflated Poisson Firm age Whole sample Positive patent count. Positive patent count Positive patent count Whole sample Whole sample Positive patent count (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.006*** 0.007 0.013** 0.189*** [0.002] [0.005] [0.005] [0.008] Fin. Dev. × Ass. Tang. − 0.054*** − 0.283*** − 0.305*** − 0.690*** [0.011] [0.064] [0.067] [0.110] Fin. Dev. × Firm Age − 0.011** − 0.059*** [0.004] [0.023] Skill End. × Skill Int. 0.007*** 0.069*** 0.082*** 0.068*** 0.585*** 0.008*** 0.080*** [0.002] [0.013] [0.015] [0.013] [0.025] [0.002] [0.020] Cap. End. × Cap. Int. 0.000 − 0.001 0.001 − 0.002 0.070*** 0.000 0.000 [0.000] [0.002] [0.002] [0.002] [0.004] [0.000] [0.009] Inst. Qual. × Contr. Int. − 0.034*** − 0.319*** − 0.331*** − 0.316*** 1.675*** − 0.012*** − 0.326*** [0.008] [0.056] [0.060] [0.056] [0.146] [0.004] [0.059] No. of products 0.000*** 0.000*** 0.000** 0.000*** 0.005*** 0.000*** 0.000 [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.006] Observations 227,583 46,864 46,864 46,864 227,583 228,192 46,951 R-squared 0.28 0.43 0.41 0.43 0.26 0.41 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Baseline Zero-inflated Poisson Firm age Whole sample Positive patent count. Positive patent count Positive patent count Whole sample Whole sample Positive patent count (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.006*** 0.007 0.013** 0.189*** [0.002] [0.005] [0.005] [0.008] Fin. Dev. × Ass. Tang. − 0.054*** − 0.283*** − 0.305*** − 0.690*** [0.011] [0.064] [0.067] [0.110] Fin. Dev. × Firm Age − 0.011** − 0.059*** [0.004] [0.023] Skill End. × Skill Int. 0.007*** 0.069*** 0.082*** 0.068*** 0.585*** 0.008*** 0.080*** [0.002] [0.013] [0.015] [0.013] [0.025] [0.002] [0.020] Cap. End. × Cap. Int. 0.000 − 0.001 0.001 − 0.002 0.070*** 0.000 0.000 [0.000] [0.002] [0.002] [0.002] [0.004] [0.000] [0.009] Inst. Qual. × Contr. Int. − 0.034*** − 0.319*** − 0.331*** − 0.316*** 1.675*** − 0.012*** − 0.326*** [0.008] [0.056] [0.060] [0.056] [0.146] [0.004] [0.059] No. of products 0.000*** 0.000*** 0.000** 0.000*** 0.005*** 0.000*** 0.000 [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.006] Observations 227,583 46,864 46,864 46,864 227,583 228,192 46,951 R-squared 0.28 0.43 0.41 0.43 0.26 0.41 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Notes: The dependent variable is the number of patents registered at the USPTO in thousands, computed separately for each country, industry and application year. All time-varying regressors are lagged one period. The regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Columns (1), (5), and (6) use the whole sample of observations, whereas columns (2)–(4) and (7) restrict to the subsample of observations with positive patent count. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year, except in column (5), where they are clustered by industry–year. See also notes to previous tables.*Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 12. Determinants of innovation—Panel regressions. Baseline Zero-inflated Poisson Firm age Whole sample Positive patent count. Positive patent count Positive patent count Whole sample Whole sample Positive patent count (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.006*** 0.007 0.013** 0.189*** [0.002] [0.005] [0.005] [0.008] Fin. Dev. × Ass. Tang. − 0.054*** − 0.283*** − 0.305*** − 0.690*** [0.011] [0.064] [0.067] [0.110] Fin. Dev. × Firm Age − 0.011** − 0.059*** [0.004] [0.023] Skill End. × Skill Int. 0.007*** 0.069*** 0.082*** 0.068*** 0.585*** 0.008*** 0.080*** [0.002] [0.013] [0.015] [0.013] [0.025] [0.002] [0.020] Cap. End. × Cap. Int. 0.000 − 0.001 0.001 − 0.002 0.070*** 0.000 0.000 [0.000] [0.002] [0.002] [0.002] [0.004] [0.000] [0.009] Inst. Qual. × Contr. Int. − 0.034*** − 0.319*** − 0.331*** − 0.316*** 1.675*** − 0.012*** − 0.326*** [0.008] [0.056] [0.060] [0.056] [0.146] [0.004] [0.059] No. of products 0.000*** 0.000*** 0.000** 0.000*** 0.005*** 0.000*** 0.000 [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.006] Observations 227,583 46,864 46,864 46,864 227,583 228,192 46,951 R-squared 0.28 0.43 0.41 0.43 0.26 0.41 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Baseline Zero-inflated Poisson Firm age Whole sample Positive patent count. Positive patent count Positive patent count Whole sample Whole sample Positive patent count (1) (2) (3) (4) (5) (6) (7) Fin. Dev. × Ext. Fin. Dep. 0.006*** 0.007 0.013** 0.189*** [0.002] [0.005] [0.005] [0.008] Fin. Dev. × Ass. Tang. − 0.054*** − 0.283*** − 0.305*** − 0.690*** [0.011] [0.064] [0.067] [0.110] Fin. Dev. × Firm Age − 0.011** − 0.059*** [0.004] [0.023] Skill End. × Skill Int. 0.007*** 0.069*** 0.082*** 0.068*** 0.585*** 0.008*** 0.080*** [0.002] [0.013] [0.015] [0.013] [0.025] [0.002] [0.020] Cap. End. × Cap. Int. 0.000 − 0.001 0.001 − 0.002 0.070*** 0.000 0.000 [0.000] [0.002] [0.002] [0.002] [0.004] [0.000] [0.009] Inst. Qual. × Contr. Int. − 0.034*** − 0.319*** − 0.331*** − 0.316*** 1.675*** − 0.012*** − 0.326*** [0.008] [0.056] [0.060] [0.056] [0.146] [0.004] [0.059] No. of products 0.000*** 0.000*** 0.000** 0.000*** 0.005*** 0.000*** 0.000 [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.006] Observations 227,583 46,864 46,864 46,864 227,583 228,192 46,951 R-squared 0.28 0.43 0.41 0.43 0.26 0.41 Country–Year FE Yes Yes Yes Yes Yes Yes Yes Industry–Year FE Yes Yes Yes Yes Yes Yes Yes Price indexes × Industry FE Yes Yes Yes Yes Yes Yes Yes Notes: The dependent variable is the number of patents registered at the USPTO in thousands, computed separately for each country, industry and application year. All time-varying regressors are lagged one period. The regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Columns (1), (5), and (6) use the whole sample of observations, whereas columns (2)–(4) and (7) restrict to the subsample of observations with positive patent count. Standard errors (reported in square brackets) are corrected for two-way clustering by country–industry and industry–year, except in column (5), where they are clustered by industry–year. See also notes to previous tables.*Significant at 10%; **significant at 5%; ***significant at 1%. View Large Finally, in Table 13 we re-estimate the previous specifications by replacing time-varying variables with their long-run averages. This reduces the incidence of zero patent counts. We also report results for IV specifications, estimated on the whole sample (column (8)) or on the subsample of observations with positive patent count (column (9)). The main results are preserved, and our coefficients of interest are similar in size to those of the panel regressions. Table 13. Determinants of innovation—Cross-sectional regressions. Baseline Zero-inflated Poisson Firm age IV Whole sample Positive patent count Positive patent count Positive patent count Whole sample Whole sample Positive patent count Whole sample Positive patent count (1) (2) (3) (4) (5) (6) (7) (8) (9) Fin. Dev. × Ext. Fin. Dep. 0.006*** 0.006 0.010** 0.315*** 0.006*** 0.012* [0.002] [0.005] [0.005] [0.029] [0.002] [0.007] Fin. Dev. × Ass. Tang. − 0.044*** − 0.197*** − 0.212*** − 1.222*** − 0.053*** − 0.241*** [0.010] [0.041] [0.044] [0.352] [0.013] [0.049] Fin. Dev. × Firm Age − 0.009** − 0.035** [0.004] [0.014] Skill End. × Skill Int. 0.009*** 0.033*** 0.042*** 0.036*** 1.097*** 0.011*** 0.043*** 0.008*** 0.026** [0.002] [0.011] [0.012] [0.010] [0.129] [0.002] [0.010] [0.002] [0.010] Cap. End. × Cap. Int. 0.001*** 0.003** 0.001 0.003** 0.162*** 0.001*** 0.001 0.001*** 0.003*** [0.000] [0.001] [0.001] [0.001] [0.013] [0.000] [0.001] [0.000] [0.001] Inst. Qual. × Contr. Int. 0.002 − 0.057*** − 0.019* − 0.063*** 3.478*** 0.019*** − 0.025** − 0.001 − 0.063*** [0.004] [0.016] [0.011] [0.016] [0.367] [0.004] [0.010] [0.005] [0.018] No. of products 0.000** 0.000*** 0.000** 0.000** 0.005*** 0.000** 0.000** 0.000*** 0.000*** [0.000] [0.000] [0.000] [0.000] [0.001] [0.000] [0.000] [0.000] [0.000] Observations 20,716 5,008 5,008 5,008 20,716 20,771 5,016 20,716 5,008 R-squared 0.23 0.33 0.31 0.33 0.22 0.30 0.21 0.29 Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes First-stage results Kleibergen–Paap F-statistic – – – – – – – 467.2 409.3 Baseline Zero-inflated Poisson Firm age IV Whole sample Positive patent count Positive patent count Positive patent count Whole sample Whole sample Positive patent count Whole sample Positive patent count (1) (2) (3) (4) (5) (6) (7) (8) (9) Fin. Dev. × Ext. Fin. Dep. 0.006*** 0.006 0.010** 0.315*** 0.006*** 0.012* [0.002] [0.005] [0.005] [0.029] [0.002] [0.007] Fin. Dev. × Ass. Tang. − 0.044*** − 0.197*** − 0.212*** − 1.222*** − 0.053*** − 0.241*** [0.010] [0.041] [0.044] [0.352] [0.013] [0.049] Fin. Dev. × Firm Age − 0.009** − 0.035** [0.004] [0.014] Skill End. × Skill Int. 0.009*** 0.033*** 0.042*** 0.036*** 1.097*** 0.011*** 0.043*** 0.008*** 0.026** [0.002] [0.011] [0.012] [0.010] [0.129] [0.002] [0.010] [0.002] [0.010] Cap. End. × Cap. Int. 0.001*** 0.003** 0.001 0.003** 0.162*** 0.001*** 0.001 0.001*** 0.003*** [0.000] [0.001] [0.001] [0.001] [0.013] [0.000] [0.001] [0.000] [0.001] Inst. Qual. × Contr. Int. 0.002 − 0.057*** − 0.019* − 0.063*** 3.478*** 0.019*** − 0.025** − 0.001 − 0.063*** [0.004] [0.016] [0.011] [0.016] [0.367] [0.004] [0.010] [0.005] [0.018] No. of products 0.000** 0.000*** 0.000** 0.000** 0.005*** 0.000** 0.000** 0.000*** 0.000*** [0.000] [0.000] [0.000] [0.000] [0.001] [0.000] [0.000] [0.000] [0.000] Observations 20,716 5,008 5,008 5,008 20,716 20,771 5,016 20,716 5,008 R-squared 0.23 0.33 0.31 0.33 0.22 0.30 0.21 0.29 Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes First-stage results Kleibergen–Paap F-statistic – – – – – – – 467.2 409.3 Notes: The dependent variable is the number of patents registered at the USPTO in thousands, computed separately for each country and industry, and averaged over 1989–2006. Private credit, factor endowments, and the number of products are also averaged over 1989–2006. The regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Columns (1), (5), (6), and (8) use the whole sample of observations, whereas columns (2)–(4), (7), and (9) restrict to the subsample of observations with positive patent count. In columns (8) and (9), financial development is instrumented using dummies for whether countries’ legal systems are of civil law (French, German, or Scandinavian origins). Standard errors (reported in square brackets) are corrected for clustering by industry. The F-statistics are reported for the Kleibergen–Paap test for weak identification. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 13. Determinants of innovation—Cross-sectional regressions. Baseline Zero-inflated Poisson Firm age IV Whole sample Positive patent count Positive patent count Positive patent count Whole sample Whole sample Positive patent count Whole sample Positive patent count (1) (2) (3) (4) (5) (6) (7) (8) (9) Fin. Dev. × Ext. Fin. Dep. 0.006*** 0.006 0.010** 0.315*** 0.006*** 0.012* [0.002] [0.005] [0.005] [0.029] [0.002] [0.007] Fin. Dev. × Ass. Tang. − 0.044*** − 0.197*** − 0.212*** − 1.222*** − 0.053*** − 0.241*** [0.010] [0.041] [0.044] [0.352] [0.013] [0.049] Fin. Dev. × Firm Age − 0.009** − 0.035** [0.004] [0.014] Skill End. × Skill Int. 0.009*** 0.033*** 0.042*** 0.036*** 1.097*** 0.011*** 0.043*** 0.008*** 0.026** [0.002] [0.011] [0.012] [0.010] [0.129] [0.002] [0.010] [0.002] [0.010] Cap. End. × Cap. Int. 0.001*** 0.003** 0.001 0.003** 0.162*** 0.001*** 0.001 0.001*** 0.003*** [0.000] [0.001] [0.001] [0.001] [0.013] [0.000] [0.001] [0.000] [0.001] Inst. Qual. × Contr. Int. 0.002 − 0.057*** − 0.019* − 0.063*** 3.478*** 0.019*** − 0.025** − 0.001 − 0.063*** [0.004] [0.016] [0.011] [0.016] [0.367] [0.004] [0.010] [0.005] [0.018] No. of products 0.000** 0.000*** 0.000** 0.000** 0.005*** 0.000** 0.000** 0.000*** 0.000*** [0.000] [0.000] [0.000] [0.000] [0.001] [0.000] [0.000] [0.000] [0.000] Observations 20,716 5,008 5,008 5,008 20,716 20,771 5,016 20,716 5,008 R-squared 0.23 0.33 0.31 0.33 0.22 0.30 0.21 0.29 Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes First-stage results Kleibergen–Paap F-statistic – – – – – – – 467.2 409.3 Baseline Zero-inflated Poisson Firm age IV Whole sample Positive patent count Positive patent count Positive patent count Whole sample Whole sample Positive patent count Whole sample Positive patent count (1) (2) (3) (4) (5) (6) (7) (8) (9) Fin. Dev. × Ext. Fin. Dep. 0.006*** 0.006 0.010** 0.315*** 0.006*** 0.012* [0.002] [0.005] [0.005] [0.029] [0.002] [0.007] Fin. Dev. × Ass. Tang. − 0.044*** − 0.197*** − 0.212*** − 1.222*** − 0.053*** − 0.241*** [0.010] [0.041] [0.044] [0.352] [0.013] [0.049] Fin. Dev. × Firm Age − 0.009** − 0.035** [0.004] [0.014] Skill End. × Skill Int. 0.009*** 0.033*** 0.042*** 0.036*** 1.097*** 0.011*** 0.043*** 0.008*** 0.026** [0.002] [0.011] [0.012] [0.010] [0.129] [0.002] [0.010] [0.002] [0.010] Cap. End. × Cap. Int. 0.001*** 0.003** 0.001 0.003** 0.162*** 0.001*** 0.001 0.001*** 0.003*** [0.000] [0.001] [0.001] [0.001] [0.013] [0.000] [0.001] [0.000] [0.001] Inst. Qual. × Contr. Int. 0.002 − 0.057*** − 0.019* − 0.063*** 3.478*** 0.019*** − 0.025** − 0.001 − 0.063*** [0.004] [0.016] [0.011] [0.016] [0.367] [0.004] [0.010] [0.005] [0.018] No. of products 0.000** 0.000*** 0.000** 0.000** 0.005*** 0.000** 0.000** 0.000*** 0.000*** [0.000] [0.000] [0.000] [0.000] [0.001] [0.000] [0.000] [0.000] [0.000] Observations 20,716 5,008 5,008 5,008 20,716 20,771 5,016 20,716 5,008 R-squared 0.23 0.33 0.31 0.33 0.22 0.30 0.21 0.29 Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes First-stage results Kleibergen–Paap F-statistic – – – – – – – 467.2 409.3 Notes: The dependent variable is the number of patents registered at the USPTO in thousands, computed separately for each country and industry, and averaged over 1989–2006. Private credit, factor endowments, and the number of products are also averaged over 1989–2006. The regressions are based on a consistent sample of countries (119) and 4-digit industries (365) with positive exports to the United States in all years between 1989 and 2006. Columns (1), (5), (6), and (8) use the whole sample of observations, whereas columns (2)–(4), (7), and (9) restrict to the subsample of observations with positive patent count. In columns (8) and (9), financial development is instrumented using dummies for whether countries’ legal systems are of civil law (French, German, or Scandinavian origins). Standard errors (reported in square brackets) are corrected for clustering by industry. The F-statistics are reported for the Kleibergen–Paap test for weak identification. See also notes to previous tables. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large 5. Conclusions In this paper we have studied how financial development affects firm-level heterogeneity and trade in a model where productivity differences across monopolistically competitive firms are endogenous and depend on investment decisions at the entry stage. By increasing entry costs, financial frictions allow less productive firms to survive and hence lower the value of investing in bigger innovation projects with more dispersed outcomes. As a result, financial frictions make firms more homogeneous and hinder the volume of exports both along the intensive and the extensive margin. Export opportunities, instead, shift expected profits to the tail and increase the value of technological heterogeneity. We have tested these predictions using comparable measures of sales dispersion within 365 manufacturing industries in 119 countries built from highly disaggregated US import data. Consistent with the model, financial development increases sales dispersion, especially in more financially vulnerable industries; sales dispersion is also increasing in measures of comparative advantage. Moreover, sales dispersion is important for explaining the effects of financial development and factor endowments on export sales. Finally, we have also provided some evidence consistent with our mechanism working through innovation. The results in this paper have important implications. First, they help explaining why financial frictions restrain trade more than domestic production. To rationalize this finding, existing models typically assume that credit is relatively more important for financing foreign than domestic activities. The origin of this asymmetry is however not entirely clear. Existing explanations also face the challenge that export volumes are dominated by large firms, and large firms are typically less financially constrained. Our model overcomes both shortcomings. Second, this paper sheds new light on the relationship between trade volumes and finance. In particular, our empirical results help identifying the mechanism through which financial development increases the volume of exports especially in financially vulnerable sectors, suggesting that part of the overall effect works through the dispersion of sales. Third, our results also contribute to understanding why firms are smaller and relatively more homogeneous in less developed countries. Finally, since more productive firms also pay higher wages, this paper also hints to an overlooked channel through which financial development may affect wage inequality.46 Exploring more in detail this hypothesis seems an interesting avenue for future research. Appendix A: Proof of Proposition 1 To prove that the equilibrium voi is increasing in export opportunities and financial development, especially in more financially vulnerable sectors, we first use (9) to define \begin{eqnarray*} W &\equiv & \frac{1}{1-v_{oi}(\sigma _{i}-1)}\ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}+\frac{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}-\frac{v_{oi}F^{\prime }(v_{oi})}{F(v_{oi})}\\ &=& \eta _{\pi }(v_{oi}) -\eta _{F}(v_{oi}) , \end{eqnarray*} and apply the implicit function theorem to obtain the generic expression for the derivative of voi with respect to variable y: \begin{eqnarray*} \frac{\partial v_{oi}}{\partial y}=-\frac{\partial W}{\partial y}\Big/\frac{\partial W}{\partial v_{oi}}. \end{eqnarray*} Under our assumption that $$\eta _{F}^{\prime }(v_{oi}) >\eta _{\pi }^{\prime }(v_{oi})$$, the denominator is negative. Next, we prove that ∂voi/∂ρdoi > 0 by computing \begin{eqnarray*} \frac{\partial W}{\partial \rho _{\mathit{doi}}} &=&\frac{\partial \ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}}{\partial \rho _{\mathit{doi}}}+\frac{\partial \left( \frac{\sum {d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}\right) }{\partial \rho _{\mathit{doi}}} \\ &=& \frac{\sum _{d\ne o}\frac{f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}{\rho _{\mathit{doi}}v_{oi}}}{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}\\ &&-\,\frac{\left( \sum _{d\ne o}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}\right) \left( \sum _{d\ne o} \frac{f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}{\rho _{\mathit{doi}}v_{oi}}\right) }{\left( \sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\right) ^{2}}, \end{eqnarray*} and showing that it is positive. To this end, we set the following condition: \begin{eqnarray*} \frac{\sum _{d\ne o}\frac{1}{\rho _{\mathit{doi}}}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}{\sum _{d\ne o}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}>\frac{\sum _{d\ne o}\frac{1}{\rho _{\mathit{doi}}} f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}, \end{eqnarray*} take the terms for d = o (with fooi and ρooi = 1) out of the summations, and obtain \begin{equation*} \frac{\sum _{d\ne o}\frac{1}{\rho _{\mathit{doi}}}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}{\sum _{d\ne o}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}>\frac{\sum _{d\ne o}\frac{1}{\rho _{\mathit{doi}}} f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}{f_{ooi}+\sum _{d\ne o}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}, \end{equation*} which holds for any ρdoi > 1. We then prove that ∂voi/∂δo > 0 by computing ∂w/∂δo = (∂W/∂λoi)(∂λoi/∂δo), which is positive since \begin{equation*} \frac{\partial W}{\partial \lambda _{oi}}=\frac{\partial \ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}}{\partial \lambda _{oi}}=\frac{\partial \ln \left( \frac{1}{\lambda _{oi}}\right) }{\partial \lambda _{oi}}=-\frac{1}{\lambda _{oi}}\text{ and }\frac{\partial \lambda _{oi}}{\partial \delta _{o}}=-\lambda _{oi}^{2}( 1-\kappa _{i}). \end{equation*} Finally, to prove that ∂2voi/(∂δo∂κi) < 0, we first obtain \begin{equation*} \frac{\partial ^{2}v_{oi}}{\partial \delta _{o}\partial \kappa _{i}}=\frac{\partial \left( -\frac{dW}{d\delta _{o}}/\frac{dW}{dv_{oi}}\right) }{\partial \kappa _{i}}=\frac{-\frac{\partial ^{2}W}{\partial \delta _{o}\partial \kappa _{i}}\frac{\partial W}{\partial v_{oi}}-\frac{\partial ^{2}W}{\partial v_{oi}\partial \kappa _{i}}\frac{\partial W}{\partial \delta _{o}}}{\left( \frac{\partial W}{\partial v_{oi}}\right) ^{2}}, \end{equation*} where the denominator is positive, −(∂W/∂voi) > 0, and −(∂W/∂δo) > 0. We prove the numerator to be negative by computing \begin{equation*} \frac{\partial ^{2}W}{\partial \delta _{o}\partial \kappa _{i}}=\frac{\partial (\lambda _{oi}(1-\kappa _{i}))}{\partial \kappa _{i}}=\lambda _{oi}[\delta _{o}(1-\kappa _{i}) -1] <0, \end{equation*} since both δo and κi take values between 0 and 1, and \begin{equation*} \frac{\partial ^{2}W}{\partial v_{oi}d\kappa _{i}}=\frac{\partial \frac{\partial \ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}}{\partial v_{oi}}}{\partial \kappa _{i}}=\frac{1}{v_{oi}}\frac{\partial \frac{\partial \ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}}{\partial \ln v_{oi}}}{\partial \kappa _{i}}= \frac{\delta _{o}}{v_{oi}}>0, \end{equation*} where the elasticity of $$(\varphi _{ooi}^{\ast }/\varphi _{\min}) ^{1/v_{oi}}$$ with respect to voi is calculated imposing the equilibrium first order condition (9).47 Hence, ∂2voi/(∂δo∂κi) < 0. Appendix B: Mean-Preserving Spreads We now consider the case in which $$\varphi _{\min }=\bar{\varphi}(1-v_{oi})$$ so that the mean $$\mathbb {E}[ \pi _{oi}] =\bar{\varphi }$$ is constant, whereas an increase in voi is still associated to a higher variance, SD[ln φ] = voi. Thus, an increase in v corresponds to a mean-preserving spread. Although the evidence in Bonfiglioli et al. (2017) suggests that the mean and the variance of productivity are likely to be linked, we nevertheless want to show that the main results in the paper still hold if firms can only choose the dispersion of the productivity draw. Assuming $$\varphi _{\min }= (1-v_{oi}) \bar{\varphi }$$, ex ante expected profits become \begin{equation*} \mathbb {E}[ \pi _{oi}] =\frac{(\sigma _{i}-1)w_{o}}{1/v_{oi}-(\sigma _{i}-1)}\left[ \frac{(1-v_{oi}) \bar{\varphi }}{\varphi _{ooi}^{\ast }}\right] ^{1/v_{oi}}\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}. \end{equation*} The first order condition for an interior voi is \begin{eqnarray} w_{o}\lambda _{oi}F^{\prime }(v_{oi}) &=&\frac{\mathbb {E}[ \pi _{oi}] }{v_{oi}}\left[ \frac{1}{1-v_{oi}(\sigma _{i}-1)}-\frac{1}{1-v_{oi}}\right. \\ &&\left. +\ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}}+\frac{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}}{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}}\right] . \nonumber \end{eqnarray} (B.1) Clearly, the fact that the mean of φ is constant lowers the marginal benefit of voi, as captured by the new term −1/(1 − voi) in the left-hand side. Notice that $$\varphi _{ooi}^{\ast }/\varphi _{\min }$$ is still pinned down by the exit condition (10) as in the baseline model. Comparing (B.1) to (9) it is easy to see that the comparative statics for voi are qualitatively unchanged, provided that σ is high enough (σ > 2 is a sufficient condition). Yet, the degree of dispersion chosen in equilibrium is lower in the case of mean-preserving spreads. We now show that the implications for the distribution of revenues, conditional on voi, are also identical. Since revenue from market d of firms from country o operating in sector i is $$r_{\mathit{doi}}(\varphi )=r_{\mathit{doi}}(\varphi _{\mathit{doi}}^{\ast })\left( \varphi /\varphi _{\mathit{doi}}^{\ast }\right) ^{(\sigma _{i}-1)}$$, it follows that rdoi(φ) is Pareto distributed with c.d.f. \begin{equation*} G_{r}(r) =1-(r_{\min }/r) ^{1/[ v_{oi}(\sigma _{i}-1)] },\quad \text{for}\quad r>r_{\min }=\sigma _{i}w_{o}f_{\mathit{doi}}. \end{equation*} Note that revenue of the marginal firm is independent of the productivity distribution because it is pinned down by the exit condition. It then follows that the formulas for the volume of trade are also unchanged. Even if the unconditional average of the productivity distribution does not change with dispersion, since the level of sales of the marginal firm is constant, average sales of operating firms still increase with dispersion. This does not mean that the volume of trade is the same in the two versions of the model. The volume of export is lower relative to the baseline case because the equilibrium voi is lower, but the way in which it varies with voi is unchanged. Appendix C: Adding Financially Unconstrained Firms We now sketch a version of the model in which in each industry there is an exogenous mass of entering firms that are not subject to the financial friction, that is, for them λoi = 1. We denote these firms with the superscript u for “unconstrained” and assume that their measure is fixed exogenously. When entering, these firms will choose $$v_{oi}^{u}$$ so as to maximize: \begin{equation*} \max _{v_{oi}^{u}}\left\lbrace \mathbb {E}\left[ \pi _{oi}^{u}\right] -w_{o}F\left( v_{oi}^{u}\right) \right\rbrace . \end{equation*} Besides the entry stage, all firms with a given productivity are however identical. Hence, the first-order condition of unconstrained firms is \begin{eqnarray*} \frac{\mathbb {E}\left[ \pi _{oi}^{u}\right] }{v_{oi}^{u}}\left[ \frac{1}{1-v_{oi}^{u}(\sigma _{i}-1)}+\ln \left( \frac{\varphi _{ooi}^{\ast }}{\varphi _{\min }}\right) ^{1/v_{oi}^{u}}+\frac{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}^{u}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}^{u}}}{\sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}^{u}}}\right]\qquad\\ = w_{o}F^{\prime }\left( v_{oi}^{u}\right) .\qquad \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad \end{eqnarray*} Note that the left-hand side is identical for all firms. This is so because, after the entry cost is paid, all firms with a given productivity are identical. For the same reason, the exit cutoff is the same for all firms. Thus, the value of drawing productivity from any distribution does not depend on whether the firm is constrained or not. The right-hand side is however different: unconstrained firms face a lower cost of financing the entry investment. Then, the assumption that F is sufficiently convex to make the maximand concave immediately implies that unconstrained firms choose a larger entry investment: $$v_{oi}^{u}>v_{oi}$$. Given that unconstrained firms face a lower entry cost, they have a strictly stronger incentive to enter. However, the number of potential unconstrained entrants is fixed (we assume that a firm is a technology, so that an unconstrained firm with an unsuccessful product cannot re-enter). We then focus on the most interesting case in which unconstrained firms are so few that, after they have all drawn their productivity, entry is still profitable for financially constrained firms.48 Under this assumption, constrained firms, denoted by a superscript c, will continue entering until the free-entry condition $$\mathbb {E}[ \pi _{oi}^{c}] =w_{o}\lambda _{oi}F( v_{oi}^{c})$$ is satisfied for them. This implies that the exit cutoff $$\varphi _{ooi}^{\ast }/\varphi _{\min }$$ is determined as in the baseline model. The choice of $$v_{oi}^{c}$$ is also identical to the baseline model. One key difference now is that in equilibrium there are two types of firms, with different distributions of revenues. On average, financially unconstrained firms are larger and make positive profits. The revenue of unconstrained firms selling to market d from country o in sector i is distributed as a Pareto with c.d.f. \begin{equation*} G_{r}^{u}(r) =1-(r_{\min }/r) ^{1/[ v_{oi}^{u}(\sigma _{i}-1)] },\quad \text{for} \quad r>r_{\min }=\sigma _{i}w_{o}f_{\mathit{doi}}. \end{equation*} The distribution of revenues of constrained firm is also Pareto, it has the same minimum, rmin , but a different shape parameter: $$v_{oi}^{c}(\sigma _{i}-1)<v_{oi}^{u}(\sigma _{i}-1)$$. The overall distribution is not Pareto anymore. However, its dispersion can still be characterized analytically using as a measure the Theil index, which has the advantage of being a weighted average of inequality within subgroups, plus inequality between those subgroups. In particular, denote T(rdoi) as the Theil index of overall inequality of revenues in the destination country d in industry i for firms selling from the country of origin o, and denote the groups of constrained and unconstrained firms with the superscript k ∈ {u, c}. Then, \begin{eqnarray*} T\left( r_{\mathit{doi}}\right) &=& \int _{0}^{\infty }\frac{r_{\mathit{doi}}}{\bar{r}_{\mathit{doi}}}\ln \left( \frac{r_{\mathit{doi}}}{\bar{r}_{\mathit{doi}}}\right) \text{d}\Phi (r_{\mathit{doi}}) \\ &=& \sum \nolimits _{k}\theta _{\mathit{doi}}^{k}T\left(r_{\mathit{doi}}^{k}\right)+\sum \nolimits _{k}\theta _{\mathit{doi}}^{k}\ln \frac{\bar{r} _{\mathit{doi}}^{k}}{\bar{r}_{\mathit{doi}}}, k\in \left\lbrace u,c\right\rbrace , \end{eqnarray*} where $$\bar{r}_{\mathit{doi}}$$ is average revenue, Φ(rdoi) is the cumulative revenue distribution, $$\bar{r}_{\mathit{doi}}^{k}$$ is average revenue in group k, $$T\left(r_{\mathit{doi}}^{k}\right)$$ is the Theil index of dispersion within group k, and $$\theta _{\mathit{doi}}^{k}$$ is the revenue share of group k firms. Given that within each group revenues follow a Pareto distribution we have \begin{equation*} T\left(r_{\mathit{doi}}^{k}\right)=\ln \left(1-v_{\mathit{doi}}^{k}\right)+\frac{v_{\mathit{doi}}^{k}}{1-v_{\mathit{doi}}^{k}}. \end{equation*} It is easy to show that this within-group Theil index is increasing in the dispersion of the Pareto distribution as measured by the parameter $$v_{\mathit{doi}}^{k}$$: \begin{equation*} \frac{\partial T\left(r_{\mathit{doi}}^{k}\right)}{\partial v_{\mathit{doi}}^{k}}>0. \end{equation*} Since $$T\left(r_{\mathit{doi}}^{u}\right)>T\left(r_{\mathit{doi}}^{c}\right)$$ and $$\bar{r}_{\mathit{doi}}^{u}>\bar{r} _{\mathit{doi}}^{c}$$, it follows that the overall Theil index is increasing in the share of financially unconstrained firms: \begin{equation*} \frac{\partial T (r_{\mathit{doi}})}{\partial \theta _{\mathit{doi}}^{u}}=\left[ T\left(r_{\mathit{doi}}^{u}\right)-T\left(r_{\mathit{doi}}^{c}\right)\right] +\ln \frac{\bar{r}_{\mathit{doi}}^{u}}{\bar{r} _{\mathit{doi}}^{c}}>0. \end{equation*} Moreover, since the difference between $$v_{\mathit{doi}}^{u}$$ and $$v_{\mathit{doi}}^{c}$$ is increasing in the level of financial frictions, λoi, we also have \begin{equation*} \frac{\partial ^{2}T(r_{\mathit{doi}}) }{\partial \theta _{\mathit{doi}}^{u}\partial \lambda _{oi}}>0. \end{equation*} In sum, revenue is more dispersed the higher the share of financially unconstrained firms, and the effect is stronger in countries or sectors in which firm-level financial frictions are more severe. Footnotes 1 See for instance Syverson (2011). 2 See Poschke (2015) and Bartelsman, Haltiwanger, and Scarpetta (2009). 3 See for instance Manova (2013), Beck (2002), and Svaleryd and Vlachos (2005). 4 Note that in our model risk is completely diversified. However, expected returns depend on the variance of productivity draws. In a more general model, financial frictions may deter entry also by lowering diversification opportunities as in Michelacci and Schivardi (2013). 5 For instance, Berman and Hericourt (2010) in their study on finance and trade use a sample of only nine countries and around 5,000 firms overall. 6 Gabler and Poschke (2013) study instead how policy distortions affect experimentation by firms. 7 A growing literature studies distortions generating dispersion in the marginal product of factors, but this is a very different question. 8 For instance, in Manova (2013) and Chaney (2016) firms only face liquidity constraints for accessing foreign markets; in Kohn, Leibovici, and Szkup (2016) exporters face relatively higher working capital needs than nonexporters; in Caggese and Cuñat (2013) exporting increases volatility and hence the risk of a costly bankruptcy, which is higher for more productive firms. The results can also be sensitive to assumptions on the credit frictions. For instance, Brooks and Dovis (2015) find that when debt limits of firms respond to profit opportunities, which they argue is the empirically relevant case, credit frictions do not hamper trade-induced reallocations. 9 Foellmi and Oechslin (2016) also study how trade may affect technology adoption in a model with financial frictions. 10 The Pareto distribution is widely used in the literature and has been shown to approximate well observed firm-level characteristics, especially among exporters (e.g., Helpman, Melitz, and Yeaple 2004). As in Chaney (2008), its convenient properties allow us to derive closed-form solutions useful for mapping the model to the data. 11 A simple microfoundation can be built on the idea that the realization of productivity depends both on quality of the project q, which is unknown and uncertain, and the size soi of the investment. In particular, assume that ln φ = soiq + ln φmin , which implies that quality and resources are complements. Then, if quality, q, is exponentially distributed with $$\Pr [q>z] =\exp (-\alpha _{i}\sigma _{i}z),$$ it follows that φ is Pareto distributed with minimum φmin  and shape parameter αiσi/soi. 12 The assumptions αi > 1, soi ∈ (0, 1], and voi = soi/(αiσi) imply that voi < 1/σi < 1, which guarantees that productivity is drawn from a distribution with a finite mean and that $$\mathbb {E}[ \pi _{oi}]$$ converges to a finite value. The condition voi < 1/σi can be relaxed if the number of firms is finite or if there is an upper bound to the support of the Pareto distribution for φ. Yet, the assumption that productivity is less dispersed in industries producing more homogeneous varieties is consistent with Syverson (2004). 13 Equivalently, we could have modified (5) so that the dispersion parameter is a concave function of the entry investment. 14 Consistent with the corporate finance literature (e.g., Tirole 2005), innovation is financed through an equity-like contract and the effect of the financial friction is to increase the cost of R&D capital. Besides the specific microfoundation, the existing evidence shows indeed that small and new innovative firms experience high costs of R&D capital, and that the problem is especially severe in countries where financial markets are less developed (e.g., Hall and Lerner 2010). 15 We assume that fooi is sufficiently high to make sure that $$\varphi _{ooi}^{\ast }/\varphi _{\min }>1$$ in equilibrium. 16 The standard deviation of the log of φ is a common measure of dispersion that has the convenient property of being scale invariant. If φ is Pareto, this measure is also invariant to truncation from below. 17 Ex ante heterogeneity is a natural feature of models with innovation by incumbents. Although our framework can accommodate such asymmetries across firms, adding innovation by incumbents would require a major departure from the current setup and is left to future research. 18 If φ follows a Pareto(φ*, z), then x ≡ ln (φ/φ*) is distributed as an exponential with parameter z. Then, any power function of φ of the type AφB, with A and B constant, is distributed as a Pareto(A(φ*)B, z/B), since AφB = A(φ*)BeBx with Bx ∼ Exp(z/B), by the properties of the exponential distribution. 19 These are the most disaggregated trade data available at the moment. For instance, in other data sources, trade data are reported at the 6-digit (UN Comtrade) or 8-digit (Eurostat Comext) level of product disaggregation. 20 In particular, each of the 119 countries has exported to the United States in at least one industry during all years between 1989 and 2006. By analogy, in each of the 365 industries at least one country has exported to the United States over the same period. Focusing on this consistent sample ensures that our econometric results are not contaminated by the creation of new countries (e.g., the former members of the Soviet Union) and by the presence of small exporters that trade with the United States only occasionally. Nevertheless, the results obtained for the entire sample of countries (171) and industries (377) are very similar (available upon request). 21 For these reasons, we find in what follows that our results are essentially unchanged when using different approaches for accommodating the presence of triplets with missing observations on sales dispersion (see Section 4.3.3 for details). 22 We use data on real per-capita GDP from the Penn World Table 8.1. 23 We source the index of doing business from the World Bank Doing Business Database. 24 We lag all time-varying controls by one period because the effects of financial development on sales dispersion need not fully unfold within a year. Our main results are however robust to using contemporaneous values (available upon request). 25 We discuss these controls and other endogeneity concerns in what follows and in Section 4.3.3. 26 In Section 4.3.3, we show that our results are robust to the use of alternative measures of financial development and financial vulnerability. 27 Following the conventional approach, we take the median value of asset tangibility and average external finance dependence across all firms in an industry over 1989–2006. For 4-digit industries with no firms in Compustat, we use the value of a given variable in the corresponding 3-digit or 2-digit sector. 28 Consistently, in some robustness checks we show that our results are unchanged when using lagged values of EFi and ATi (computed over the decade before the beginning of our sample) or the rankings of industries in terms of these two variables. 29 Skill and capital endowments are the log index of human capital per person and the log real capital stock per person engaged, respectively. Both variables are sourced from the Penn World Table 8.1. Skill and capital intensity are the log ratio of nonproduction to production workers’ employment and the log real capital stock per worker, respectively. Both variables are sourced from the NBER Manufacturing Industry Productivity Database and averaged over 1989–2006. Institutional quality is average rule of law over 1996–2006, sourced from the Worldwide Governance Indicator Database. Contract intensity is the indicator for the importance of relationship-specific investment in each industry, sourced from Nunn (2007). 30 The industry fixed effects also subsume the linear terms in financial vulnerability and factor intensities. 31 Because rule of law does not vary over time, its linear term is captured by the country fixed effects. 32 The country–time and industry–time effects also absorb all country- and industry-specific determinants of sales dispersion. These include the elasticity of substitution, as well as the country and industry components of variable trade costs (e.g., distance and bulkiness). 33 In unreported specifications, we have also estimated the baseline regression after excluding countries with extreme values of private credit (Japan an Sierra Leone) and industries with extreme values of financial vulnerability (SIC 2111, 2836, 3844, and 2421). The coefficients (available upon request) were very close to the baseline estimates, suggesting that our main results are not driven by outliers. 34 Helpman, Melitz, and Rubinstein (2008) and Manova (2013) use a similar two-step model for correcting the estimates of gravity equations from sample selection bias. Consistently, the Probit results in column (3) are similar to those in Manova (2013), who finds the probability of observing a trade flow to be increasing in the exporter’s financial development, the more so in financially vulnerable industries. 35 We omit the number of products from columns (3) and (4), because this variable creates convergence problems when estimating the Probit model. The reason is that the number of products is zero for most of the triplets in which the dependent dummy variable is also zero (see Table 1 for details). This creates nonconcavities in the log-likelihood function, and prevents convergence. The estimates in column (4) should thus be compared with those reported in column (4) of Table 3, which excludes as well the number of products. 36 The coefficients reported in columns (3) and (4) of Table 4 are identified through the implicit assumption that the errors of the two equations are jointly normal. In untabulated regressions (available upon request), we have estimated the Probit model using the lagged dependent variable as an additional regressor, which is excluded from the main equation in column (4) (see Johnson 2012). This variable has strong predicted power, consistent with the existence of fixed export costs. At the same time, our coefficients of interest were very close to those reported in column (4). One caveat with this specification is that past participation in trade may be correlated with some unobserved determinant of sales dispersion. 37 Bank assets are total assets held by commercial banks. As such, they also include credit to the public sector and assets other than credit. This feature makes bank assets a more comprehensive, but less precise, proxy for the size of the financial sector. Liquid liabilities include all liabilities of banks and other financial intermediaries. Thus, this variable may also include liabilities backed by credit to the public sector. Finally, domestic credit also includes credit issued by, and granted to, the public sector, and thus is a broader, but perhaps less precise, measure of the size of the financial system. See Crinò and Ogliari (2017) for more details. 38 The lending rate is the rate charged by banks for loans to private firms. As such, it is a standard proxy for the cost of borrowing in a country (see, e.g., Chor and Manova 2012). We source this variable from the IMF International Financial Statistics and the OECD. 39 To ease the interpretation of the coefficients, we normalize the rankings between 0 and 1. 40 We exclude triplets with only two products, as for these triplets there are fewer observations than parameters to be estimated. Moreover, following Gabaix and Ibragimov (2011), we adjust sales rank by subtracting 0.5, in order to correct for possible small-sample biases. 41 Import penetration and export intensity are, respectively, the ratio of imports over apparent consumption (GDP plus imports minus exports) and the export share of GDP; both variables are constructed with data from the World Development Indicators. The real exchange rate and the FDI share of GDP are sourced from the Penn World Table 8.1 and UNCTAD FDI Statistics, respectively. 42 Geographical areas are seven regions defined by the World Bank: East Asia and Pacific; Europe and Central Asia; Latin America and the Caribbean; Middle East and North Africa; North America; South Asia; and Sub-Saharan Africa. 43 We use information on systemic banking crises from Laeven and Valencia (2012). 44 We source this index from the World Bank Doing Business Database; we normalize it to range between 0 and 1, and so that higher values correspond to countries with a higher position in the ranking (i.e., better institutions). 45 We consider various ways for dealing with the zeros in the regression analysis in what follows. 46 See Michelacci and Quadrini (2009), Philippon and Reshef (2012), and Bonfiglioli (2012) for papers studying how financial development can affect wage and income inequality. 47 In particular, \begin{eqnarray*} && d\ln \big(\varphi _{ooi}^{\ast }/\varphi _{\min }\big) ^{1/v_{oi}}/d\ln v_{oi} \\ &&= [1-v_{oi}(\sigma _{i}-1)] ^{-1}+ \left( \sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\ln \rho _{\mathit{doi}}^{-1/v_{oi}}\right) \Bigg/\left( \sum _{d}f_{\mathit{doi}}\rho _{\mathit{doi}}^{1/v_{oi}}\right) - v_{oi}F^{\prime }(v_{oi}) /F(v_{oi}), \end{eqnarray*} which under (9) is equal to $$-\ln ( \varphi _{ooi}^{\ast }/\varphi _{\min }) ^{1/v_{oi}}$$. 48 The other case is trivial, in that it coincides with the equilibrium of an industry not subject to any financial friction. References Arellano Cristina , Bai Yan , Zhang Jing ( 2012 ). “ Firm Dynamics and Financial Development .” Journal of Monetary Economics , 59 , 533 – 549 . Google Scholar CrossRef Search ADS Bartelsman Eric , Haltiwanger John , Scarpetta Stefano ( 2009 ). “ Measuring and Analyzing Cross Country Differences in Firm Dynamics .” In Producer Dynamics: New Evidence from Micro Data , edited by Timothy Dunne , Bradford Jensen , Mark Roberts . University of Chicago Press , pp. 15 – 76 . Google Scholar CrossRef Search ADS Beck Thorsten ( 2002 ). “ Financial Development and International Trade. Is There a Link? ” Journal of International Economics , 57 , 107 – 131 . Google Scholar CrossRef Search ADS Beck Thorsten , Demirgüç-Kunt Asli , Maksimovic Vojislav ( 2005 ). “ Financial and Legal Constraints to Growth: Does Firm Size Matter? ” The Journal of Finance , 60 , 137 – 177 . Google Scholar CrossRef Search ADS Berman Nicolas , Hericourt Jérôme ( 2010 ). “ Financial Factors and the Margins of Trade: Evidence from Cross-Country Firm-Level Data .” Journal of Development Economics , 93 , 206 – 217 . Google Scholar CrossRef Search ADS Bernard Andrew , Redding Stephen , Schott Peter ( 2007 ). “ Comparative Advantage and Heterogeneous Firms .” Review of Economic Studies , 74 , 31 – 66 . Google Scholar CrossRef Search ADS Bonfiglioli Alessandra ( 2012 ). “ Investor Protection and Income Inequality: Risk Sharing vs Risk Taking .” Journal of Development Economics , 99 , 92 – 104 . Google Scholar CrossRef Search ADS Bonfiglioli Alessandra , Crinò Rosario , Gancia Gino ( 2017 ). “ Betting on Exports: Trade and Endogenous Heterogeneity .” Economic Journal , forthcoming . Broda Christian , Weinstein David ( 2010 ). “ Product Creation and Destruction: Evidence and Price Implications .” American Economic Review , 100 (3) , 691 – 723 . Google Scholar CrossRef Search ADS Brooks Wyatt , Dovis Alessandro ( 2015 ). “ Trade Liberalization with Endogenous Borrowing Constraints .” Working paper , University of Minnesota . Buera Francisco , Kaboski Joseph , Shin Yongseok ( 2015 ). “ Entrepreneurship and Financial Frictions: A Macro-Development Perspective .” Annual Review of Economics , 7 , 409 – 436 . Google Scholar CrossRef Search ADS Bustos Paula ( 2011 ). “ Trade Liberalization, Exports, and Technology Upgrading: Evidence on the Impact of MERCOSUR on Argentinian Firms .” American Economic Review , 101 (1) , 304 – 340 . Google Scholar CrossRef Search ADS Cabral Luís , Mata José ( 2003 ). “ On the Evolution of the Firm Size Distribution: Facts and Theory .” American Economic Review , 93 (4) , 1075 – 1090 . Google Scholar CrossRef Search ADS Caggese Andrea ( 2015 ). “ Financing Constraints, Radical versus Incremental Innovation, and Aggregate Productivity .” Working paper , Universitat Pompeu Fabra . Caggese Andrea , Cuñat Vicente ( 2013 ). “ Financing Constraints, Firm Dynamics, Export Decisions, and Aggregate Productivity .” Review of Economic Dynamics , 16 , 177 – 193 . Google Scholar CrossRef Search ADS Chaney Thomas ( 2008 ). “ Distorted Gravity: The Intensive and Extensive Margins of International Trade .” American Economic Review , 98 (4) , 1707 – 1721 . Google Scholar CrossRef Search ADS Chaney Thomas ( 2016 ). “ Liquidity Constrained Exporters .” Journal of Economic Dynamics and Control , 72 , 141 – 154 . Google Scholar CrossRef Search ADS Chor Davin ( 2010 ). “ Unpacking Sources of Comparative Advantage: A Quantitative Approach .” Journal of International Economics , 82 , 152 – 167 . Google Scholar CrossRef Search ADS Chor Davin , Manova Kalina ( 2012 ). “ Off the Cliff and Back: Credit Conditions and International Trade during the Global Financial Crisis .” Journal of International Economics , 87 , 117 – 133 . Google Scholar CrossRef Search ADS Claessens Stijn , Laeven Luc ( 2003 ). “ Financial Development, Property Rights, and Growth .” Journal of Finance , 58 , 2401 – 2436 . Google Scholar CrossRef Search ADS Crinò Rosario , Ogliari Laura ( 2017 ). “ Financial Imperfections, Product Quality, and International Trade .” Journal of International Economics , 104 , 63 – 84 . Google Scholar CrossRef Search ADS Doraszelski Ulrich , Jaumandreu Jordi ( 2013 ). “ R&D and Productivity: Estimating Endogenous Productivity .” Review of Economic Studies , 80 , 1338-1383 . Google Scholar CrossRef Search ADS Dunne Timothy , Foster Lucia , Haltiwanger John , Troske Kenneth ( 2004 ). “ Wage and Productivity Dispersion in US Manufacturing: The Role of Computer Investments .” Journal of Labor Economics , 22 , 397 – 429 . Google Scholar CrossRef Search ADS Faggio Giulia , Salvanes Kjell , Van Reenen John ( 2010 ). “ The Evolution of Inequality in Productivity and Wages: Panel Data Evidence .” Industrial and Corporate Change , 19 , 1919 – 1951 . Google Scholar CrossRef Search ADS Feenstra Robert , Romalis John , Schott Peter ( 2002 ). “ U.S. Imports, Exports and Tariff Data, 1989-2001 .” NBER Working Paper 9387 , NBER , Cambridge, MA . Foellmi Reto , Oechslin Manuel ( 2016 ). Harmful Pro-Competitive Effects of Trade in Presence of Credit Market Frictions . Manuscript , University of St. Gallen . Foley Fritz , Manova Kalina ( 2015 ). “ International Trade, Multinational Activity, and Corporate Finance .” Annual Review of Economics , 7 , 119 – 146 . Google Scholar CrossRef Search ADS Gabaix Xavier , Ibragimov Rustam ( 2011 ). “ Rank-1/2: A Simple Way to Improve the OLS Estimation of Tail Exponents .” Journal of Business Economics and Statistics , 29 , 24 – 39 . Google Scholar CrossRef Search ADS Gabler Alain , Poschke Markus ( 2013 ). “ Experimentation by Firms, Distortions, and Aggregate Productivity .” Review of Economic Dynamics , 16 , 26 – 38 . Google Scholar CrossRef Search ADS Hall Bronwyn , Lerner Josh ( 2010 ). “ Financing R&D and Innovation .” In Handbook of the Economics of Innovation , edited by Bronwyn Hall , Nathan Rosenberg . Elsevier , pp. 609 – 639 . Heckman James ( 1979 ). “ Sample Selection Bias as a Specification Error .” Econometrica , 47 , 153 – 161 . Google Scholar CrossRef Search ADS Helpman Elhanan , Melitz Marc , Rubinstein Yona ( 2008 ). “ Estimating Trade Flows: Trading Partners and Trading Volumes .” Quarterly Journal of Economics , 123 , 441 – 487 . Google Scholar CrossRef Search ADS Helpman Elhanan , Melitz Marc , Yeaple Stephen ( 2004 ). “ Export Versus FDI with Heterogeneous Firms .” American Economic Review , 94 (1) , 300 – 316 . Google Scholar CrossRef Search ADS Johnson Robert ( 2012 ). “ Trade and Prices with Heterogeneous Firms .” Journal of International Economics , 86 , 43 – 56 . Google Scholar CrossRef Search ADS Khandelwal Amit , Topalova Petia ( 2011 ). “ Trade Liberalization and Firm Productivity: The Case of India .” Review of Economics and Statistics , 93 , 995 – 1009 . Google Scholar CrossRef Search ADS Kohn David , Leibovici Fernando , Szkup Michal ( 2016 ). “ Financial Frictions and New Exporter Dynamics .” International Economic Review , 57 , 453 – 486 . Google Scholar CrossRef Search ADS Laeven Luc , Valencia Fabian ( 2012 ). “ Systemic Banking Crises Database: An Update .” IMF Working Paper 12/163 , IMF , Washington, DC . Google Scholar CrossRef Search ADS La Porta Rafael , Lopez-de-Silanes Florencio , Shleifer Andrei ( 2008 ). “ The Economic Consequences of Legal Origins .” Journal of Economic Literature , 46 , 285-332 . Google Scholar CrossRef Search ADS Levchenko Andrei ( 2007 ). “ Institutional Quality and International Trade .” Review of Economic Studies , 74 , 791 – 819 . Google Scholar CrossRef Search ADS Manova Kalina ( 2013 ). “ Credit Constraints, Heterogeneous Firms, and International Trade .” Review of Economic Studies , 80 , 711 – 744 . Google Scholar CrossRef Search ADS Melitz Marc ( 2003 ). “ The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry Productivity .” Econometrica , 71 , 1695 – 1725 . Google Scholar CrossRef Search ADS Melitz Marc , Redding Stephen ( 2014 ). “ Heterogeneous Firms and Trade .” In Handbook of International Economics , edited by Gita Gopinath , Elhanan Helpman , Kenneth Rogoff . Elsevier , pp. 1 – 54 . Michelacci Claudio , Quadrini Vincenzo ( 2009 ). “ Financial Markets and Wages .” Review of Economic Studies , 76 , 795 – 827 . Google Scholar CrossRef Search ADS Michelacci Claudio , Schivardi Fabiano ( 2013 ). “ Does Idiosyncratic Business Risk Matter for Growth? ” Journal of the European Economic Association , 11 , 343 – 368 . Google Scholar CrossRef Search ADS Midrigan Virgiliu , Xu Daniel ( 2014 ). “ Finance and Misallocation: Evidence from Plant-Level Data .” American Economic Review , 104 (2) , 422 – 458 . Google Scholar CrossRef Search ADS Nunn Nathan ( 2007 ). “ Relationship-Specificity, Incomplete Contracts, and the Pattern of Trade .” Quarterly Journal of Economics , 122 , 569 – 600 . Google Scholar CrossRef Search ADS Paravisini Daniel , Rappoport Veronica , Schnabl Philipp , Wolfenzon Daniel ( 2015 ). “ Dissecting the Effect of Credit Supply On Trade: Evidence From Matched Credit-Export Data .” Review of Economic Studies , 82 , 333 – 359 . Google Scholar CrossRef Search ADS Philippon Thomas , Reshef Ariell ( 2012 ). “ Wages and Human Capital in the U.S. Financial Industry: 1909–2006 .” Quarterly Journal of Economics , 127 , 1551 – 1609 . Google Scholar CrossRef Search ADS Poschke Markus ( 2015 ). “ The Firm Size Distribution across Countries and Skill-Biased Change in Entrepreneurial Technology .” IZA Discussion Paper 7991 , IZA , Bonn . Rajan Raghuram , Zingales Luigi ( 1998 ). “ Financial Dependence and Growth .” American Economic Review , 88 (3) , 559 – 586 . Romalis John ( 2004 ). “ Factor Proportions and the Structure of Commodity Trade .” American Economic Review , 94 (1) , 67 – 97 . Google Scholar CrossRef Search ADS Schott Peter , Pierce Justin ( 2012 ). “ A Concordance Between Ten-Digit U.S. Harmonized System Codes and SIC/NAICS Product Classes and Industries .” Journal of Economic and Social Measurement , 37 , 61 – 96 . Silverman Brian ( 1999 ), “ Technological Resources and the Direction of Corporate Diversification: Toward an Integration of the Resource-Based View and Transaction Cost Economics .” Management Science , 45 , 1109 – 1124 . Google Scholar CrossRef Search ADS Svaleryd Helena , Vlachos Jonas ( 2005 ). “ Financial Markets, the Pattern of Industrial Specialization and Comparative Advantage: Evidence from OECD Countries .” European Economic Review , 49 , 113 – 144 . Google Scholar CrossRef Search ADS Syverson Chad ( 2004 ). “ Product Substitutability and Productivity Dispersion .” Review of Economics and Statistics , 86 , 534 – 550 . Google Scholar CrossRef Search ADS Syverson Chad ( 2011 ). “ What Determines Productivity .” Journal of Economic Literature , 49 , 326 – 365 . Google Scholar CrossRef Search ADS Tirole Jean ( 2005 ). The Theory of Corporate Finance . Princeton University Press . SUPPLEMENTARY DATA Supplementary data are available at JEEA online. © The Author(s) 2018. Published by Oxford University Press on behalf of European Economic Association. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of the European Economic Association Oxford University Press

Trade, Finance, and Endogenous Firm Heterogeneity