WPS4525 Policy ReseaRch WoRking PaPeR 4525 Credit Chains and Sectoral Comovement: Does the Use of Trade Credit Amplify Sectoral Shocks? Claudio Raddatz The World Bank Development Research Group Macroeconomics and Growth Team February 2008 Policy ReseaRch WoRking PaPeR 4525 Abstract This paper provides evidence of the presence and increase in their correlation. The analysis uses detailed relevance of a credit-chain amplification mechanism data on the correlations and input-output relations of by looking at its implications for the correlation of 378 manufacturing industry-pairs across 44 countries industries. In particular, it tests the hypothesis that with different degrees of use of trade credit. The results an increase in the use of trade-credit along the input- provide strong support for this hypothesis and indicate output chain linking two industries results in an that the mechanism is quantitatively relevant. This paper--a product of the Growth and the Macroeconomics Team, Development Research Group--is part of a larger effort in the department to understand the determinants of macroeconomic volatility. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at craddatz@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Credit Chains and Sectoral Comovement: Does the Use of Trade Credit Amplify Sectoral Shocks? Claudio Raddatz The World Bank E-mail address: craddatz@worldbank.org. I am grateful to Marcela Eslava, Eduardo Engel, Pablo García, and participants at seminars at Universidad de Chile, Universidad Adolfo Ibañez, Central Bank of Chile, LACEA, Latin Finance Network, and the World Bank for valuable comments. All remaining errors are my responsibility. The views expressed in this paper are the author's only and do not necessarily represent those of the World Bank, its Executive Directors, or the countries they represent. 1 Introduction Trade credit is an important source of short-term financing for firms, not only in the U.S., as documented by Petersen and Rajan (1997), but also around the world. For instance, accounts payables are larger than short-term debt in 60 percent of the countries covered by Worldscope. Also, across the world most firms simultaneously receive credit from their suppliers and grant it to customers concentrated on specific industries. These characteristics of trade credit financing have led some authors to propose it as a mechanism for the propagation and amplification of idiosyncratic shocks. The intuition is straightforward; a firm facing a default by its customers may run into liquidity problems and default to its own suppliers. Therefore, a temporary shock to the liquidity of some firms in a network that borrows from each other may cause a chain reaction in which other firms also get in financial di culties resulting in a large and persistent decline in aggregate activity. This idea was first formalized by Kiyotaki and Moore (1997) in a partial equilibrium setting, and recently extended to a general equilibrium environment by Cardoso-Lecourtois (2004) and Boissay (2006) who have also provided evidence of the potential quantitative importance of the mechanism by calibrating their models to the cases of Mexico and the U.S. Anecdotal evidence from surveys of firms suggests that this mechanism is relevant at the mi- croeconomic level. Bradley and Rubach (2002) found that 31 percent of a sample of 131 firms filing for bankruptcy in the U.S. reported the non-payment by customers and businesses as the major cause for bankruptcy, and 66 percent reported that the non-payment of trade creditors had a great adverse e ect on their operations. Also, based on evidence from the U.K., Chittenden and Bragg (1997) argue that firms typically respond to late payment by customers by delaying their payments to their trade creditors. However, so far there is no formal systematic empirical evidence of the presence and importance of the transmission of shocks through credit chains, which has led to some skepticism on the relevance of this mechanism at the macro level.1 This lack of evidence is most likely the result of the absence of data sources containing the information on trade credit relations among firms required for a direct test of the hypothesis that these relations transmit and amplify shocks. This paper partly fills this empirical gap by providing systematic evidence of the presence of the credit-chain amplification mechanism and quantifying its importance. It does so by taking an indirect approach that exploits the implications of the mechanism for the comovement of di erent industries and relies on widely available industry-level data. The idea behind this approach is that although the exact relations among individual firms are typically unknown, existing input-output matrices provide information on the trade relations among industries that combined with firms' balance sheets data can be used to build a measure of the use of trade credit along the chain of 1The reasons for this skepticism is that in some countries the amounts of net trade credit are typically small, as noticed by Coricelli and Masten (2004), which suggests the possibility that defaults may cancel out at the aggregate level if the network is circular. Basically this means that it is theoretically possible that each firm defaults on each other without a ecting actual production decisions. Each firm obtains the liquidity it needs by defaulting on its suppliers 1 industries linking two sectors and test whether according to theory this measure explains their comovement. I build my test using detailed data on the output of 28 manufacturing industries in a sample of 44 countries during 1980-2004 to compute the within-country correlation of the 378 resulting industry-pairs. I also gather information on the use of trade credit at the industry and country level during the same period and on input-output relations among these industries. Under some assumptions, I use these data to construct a measure of the trade credit use along the chain linking a pair of industries that approximately corresponds to a weighted average of the relative use of trade credit across all sectors linking those industries. This measure, which I label the credit-chain distance between two industries in a given country, is derived from an extension that introduces the possibility of credit-chain amplification in reduced form to the multi-sector model of Shea (2002). In constructing the measure of credit-chain distance for industry pairs in di erent countries I make three assumptions regarding the similarity of input-output relations and relative intensity of use of trade credit across countries because of data availability problems. First, since I lack detailed input-output data for most countries in my sample, I assume that input-output linkages across industries are mostly technologically determined, so I can use data from U.S. input-output matrices to proxy for these relations in other countries. Second, in most developing counties standard data sources do not contain enough firms to build industry level measures of the use of trade credit. Thus, in most of the paper I follow the assumption of Fisman and Love (2003) that the relative use of trade credit across industries is partly technologically determined. Finally, since I do not have information on the di erential use of trade credit across supplier industries, I assume that the fraction of goods that firms in a given sector buy on trade credit is similar across suppliers. I discuss these assumptions in detail in the paper, provide evidence of their validity, relax them as much as possible, and also show that they most likely result in attenuation bias that reduces the power of the tests and stacks the cards against finding any significant result. I test my main hypothesis by running a set of regressions of the output correlation of di erent industry pairs across countries against the measure of credit-chain distance and a number of controls that include country and industry-pair fixed-e ects. Consistently with the hypothesis of credit-chain amplification, I find that an increase in the use of trade credit along the chain linking two sectors, both as a fraction of the cost of inputs and relative to bank credit, significantly increases their output correlation. This result is robust to changes in the specific sample, measure of correlation, and the consideration of a set of alternative explanations. In terms of economic magnitude, my point estimates indicate that a one-standard-deviation increase in the measure of credit-chain distance increases the sectoral correlation of value added of an industry pair in four percentage points, which corresponds to 16 percent of the observed standard deviation of sectoral correlations in the data (0.27). At the aggregate level, if the average country in the sample increases its ratio of accounts payables to the cost of goods sold from seven to 33 percent­a movement from the bottom to the top levels observed in the sample, the variance of its manufacturing sector will increase in about 20 percent. These magnitudes are meaningful 2 and, moreover, the econometric procedure will tend to bias downwards the estimated coe cients, so they should be considered as lower bounds of the true economic relevance of the credit-chain amplification mechanism. This paper contributes to several strands of literature. First and most directly, to the theoretical literature on credit-chain amplification of financial distress started by Kiyotaki and Moore (1997) and extended by Cardoso-Lecourtois (2004) and Boissay (2006). It adds to this literature by provid- ing systematic empirical evidence of the presence and relevance of this mechanism. The paper also contributes to the finance literature on trade credit, which has mainly focused on explaining the reasons why firms use trade credit financing (see Petersen and Rajan, 1997; Burkart and Ellingsen, 2004, for good summaries) and has only recently explored the relations between trade credit use and the macroeconomy. Fisman and Love (2003) provide evidence that firms can use trade credit to substitute for formal bank financing and therefore industries where trade credit is more important can grow relatively faster in countries with weaker formal financial systems. Love et al. (2007), also provide evidence that a bank credit crunch a ects the availability of liquidity in the form of trade credit by restricting the access to formal finance of firms that previously acted as financial intermediaries and provided trade credit to other firms. This paper provides evidence of an addi- tional mechanism by which the use of trade credit itself may have macroeconomic consequences by amplifying and transmitting shocks across firms in the economy. Finally, this paper also relates to the macroeconomic literature on sectoral comovement, which has documented that business cycles are characterized by positive comovement across sectors. This has led many authors to propose mechanisms that can account for this comovement without relying on aggregate shocks. One of the proposed mechanisms is the transmission of shocks through input-output linkages (see Long and Plosser, 1983; Horvath, 2000, among others). This paper provides evidence of the relevance of this mechanism by showing that, across a large sample of countries, industries that are closer in an input-output sense are relatively more correlated. The rest of the paper is structured as follows. Section 2 presents a stylized model of the transmission of shocks through credit chains in a multi-sector economy. Section 3 discusses the empirical implementation of the test and presents the di erent specifications that I estimate. Section 4 explains the manner in which the di erent variables included in the empirical specifications are measured and the di erent data sources used for this purpose. Section 5 presents the main results. Section 6 explores the robustness of these results. Section 7 concludes. 2 Theoretical Underpinnings This section presents a simple model to illustrate the role of input-output linkages in financial contagion through credit chains. The purpose of the model is to show as simply as possible that introducing multiple intermediate goods in a stylized model of credit-chain amplification results in a relation between the use of trade credit, input-output linkages and sectoral correlations that can be tested in the data without knowing the individual links across firms in di erent industries. The 3 model does not address the question of why non-financial firms extend credit to other companies. A large recent literature has explored this question in detail and proposed many mechanisms that can account for this fact (see Burkart and Ellingsen, 2004, for a recent review of the literature). Instead, I assume that, for reasons outside the model, firms sell a fraction of their goods on credit and can simultaneously access a financial intermediary to cover down payments to other firms if necessary. This coexistence of trade and bank credit has been widely documented and explained by several models such as Biais and Gollier (1997), and Burkart and Ellingsen (2004). Time lasts two periods, t = {0, 1}. There are N sectors, each one identified with a particular intermediate good and formed by a unit mass of ex-ante identical entrepreneurs/consumers indexed by z [0, 1] who are risk neutral, consume a final good at the end of t = 1, and can save at an interest rate that I assume equal to zero. At t = 0 all entrepreneurs in sector i are endowed with Xi units of type-i intermediate good that they can combine with the other types of intermediate goods to produce Yi(z) units of the final good at t = 1 using the following technology: Yi(z) =Ai(z)(Ki(z)) , (1) Ki(z) =min X1(z),..., XN(z) i i , ai1 aiN where z indexes the entrepreneur, < 1 is common for all sectors, Xj(z) is the amount of inter- i mediate good j used in the production of good i by entrepreneur z, and aij is a technical coe cient that indicates the relative amount of good j that should be optimally used in the production of good i.2 Since the production of final goods requires the combination of various types of intermediates, type i entrepreneurs need to exchange part of their endowments with entrepreneurs from other sectors. To highlight the role of individual relations among entrepreneurs I assume that exchanges are personalized: entrepreneur z in sector i (henceforth zi) gets its supply of type j intermediate good from entrepreneur z in sector j (zj). I also assume that, for reasons outside the model, in this exchange entrepreneur zj extends trade credit to zi for a fraction i j of the value of the purchase expressed in units of t = 1 final good. Entrepreneurs do not decide how much trade credit to extend but take i j as given. Each entrepreneur has also access to funds from a bank to finance down payments if necessary, subject to a borrowing constraint. Intermediate goods markets are competitive and each entrepreneur takes the price of all intermediate goods (including hers) as given. To simplify matters even further, following Kiyotaki and Moore (1997) and Cardoso-Lecourtois (2004) I assume that there is no ex-ante uncertainty, so shocks at t = 1 are zero probability events and all entrepreneurs plan considering Ai as given. Under these conditions, the problem of entrepreneur z in sector i at t = 0 is to select a combi- nation of intermediate inputs to solve 2I use a fixed-proportions technology to highlight the input-output linkages that are the focus of the empirical section but the choice is not crucial for the derivation below. 4 max j i i i(z) = Yi(z) + pi Xi (z) j i jpjXj(z) Bi(z) (2) {Xj}N i j=1 j j s.t. Yi(z) = AiKi(z) , i i Ki(z) = min X1(z),..., XN(z) , ai1 aiN Bi(z) = (1 i i j j j)pjXj(z) (1 i)piXi (z), j j Bi(z) i i p Xi, where pi is the price of intermediate good i in terms of units of the final good at t = 1. The first term in the objective function corresponds to the value of the final good produced. The second term captures the accounts receivable of the entrepreneur, a fraction j i of the total value of the intermediate good supplied to the entrepreneur z in each sector j entrepreneurs at t = 0. Analogously, the third term corresponds to entrepreneur zi's accounts payable. Finally, Bi(z) is the amount borrowed (saved) at t = 0 to finance down-payments. The first two constraints in the maximization problem reproduce the technology available to the entrepreneur, the third is the budget constraint at t = 0, and the last is a borrowing constraint that restricts the amount that the entrepreneur can borrow to a fraction of her net wealth at t = 0. In this simple setting, an entrepreneur's demand for intermediate goods is Xj = aijKi, where Ki is adjusted according to i whether the borrowing constraint binds. If there are no shocks (and no defaults) profits will correspond to the income accrued from the production of the final good and the selling of initial endowments minus the cost of intermediate inputs, but the presence of defaults may lead profits to depart from this value. In this model a default will occur when an entrepreneur su ers a productivity shock that places its total revenue below its total debt. In this situation the entrepreneur is protected by limited liability and uses her available funds to settle debts on a pro-rata basis. The consequences of a default will depend on whether it is caused by an idiosyncratic or a sectoral shock. If an idiosyncratic shock a ects a countable number of entrepreneurs it will have a ect only those entrepreneurs linked to them but will have no aggregate consequences. For instance, consider the case in which only entrepreneur zk defaults and pays a fraction 0 µk(z) 1 of her debt. Because of the nature of the links among entrepreneurs, in any sector i only entrepreneur zi's balance sheet will be directly a ected by the default according to: j k k i i i(z) = Yi(z) + pi ajiKj i + pi a Kkµk(z) i i pj a Ki j j Bi. (3) j=k j Faced with this balance sheet, the entrepreneur may be forced to default on part of her its debt if i(z) < 0. Otherwise, she will honor her debts and µi(z) = 1. When i(z) < 0, the pro-rata 5 assumption implies that µi(z) = A(z)Ki + pi j j j i i a Kjµj(z) . (4) j j pjaijKi pi j(1 i)ajiKj Equations (3) and (4) show the standard transmission and amplification of a shock by the credit chain that has been modeled in the literature. A bad realization of Ak(z) may lead entrepreneur z from sector k to default on its debt, pushing µk(z) below one. This reduces the profits of z entrepreneurs in all other sectors, in particular in sector i. If the impact of the decline of µk(z) leads zi entrepreneurs to default, the fraction µi(z) they repay is increasing in µk(z), which further reduces the profits of entrepreneurs in sector k, and so forth. The transmission of this idiosyncratic shock is, however, confined to those firms directly linked to zk and has no sectoral consequences. This can be easily seen from the expression for sectoral profits i j j i = i(v)dv = Yi(v)dv + pi a Kj µj(v)dv Di(v)dv v v j i i v v = Yi(v)dv + pi j j a Kj Di v j i i where Di(v) is the total debt (payables plus bank debt) of entrepreneur vi. The last equality indicates that aggregate profits with idiosyncratic defaults are equal to those in absence of defaults, and follows from µj(v) = 1 almost everywhere in response to idiosyncratic shocks. This extreme result is due to the modelling individual entrepreneurs as a continuum, but captures the intuition that with a large number of firms the consequences of idiosyncratic shocks may not map into the aggregate and tracking them require knowledge of the specific linkages among firms. The situation is di erent for sectoral shocks, because their consequences for other sectors can be tracked without knowing the individual links among firms. To see this, consider the case in which a fraction k of entrepreneurs are under distress in sector k and repay a fraction µk of their debts. The size of this fraction is a measure of the size of the sectoral shock, and I assume without loss of generality that it corresponds to entrepreneurs 0 z k. The profits of entrepreneur zi are now Yi(z) + pi j j j=k Xi z , i + piXi k k iµk Di if 0 k i(z) = Yi(z) + pi j j j Xi if < z 1. i Di k The consequences of the default on sector i s aggregate profits will depend on the fraction of firms a ected that are also lead to default. Assuming that productivities are distributed independently of whether a firm's clients are in default, aggregate output is N k i= (1 (µk) k)E( i (z)) k (1 (µk))piaki iKk(1 µk), where (µk) is the fraction of the a ected firms that are led to default, which depends on the size of the default µk, and E( N i ) denotes the expected profits with no default. Thus, the direct impact on sector i s profits (and consumption) iof an increase in the fraction of entrepreneurs from sector 6 k that are in distress is i= (µk)E( N k i ) (1 (µk))piaki iKk(1 µk). k This expression shows that, although idiosyncratic firm-level shocks a ect only the individual firms linked to those that default, a sectoral shock have aggregate consequences for all those sectors that extended credit to the defaulting one. Moreover, the size of the impact will be increasing on the strength of the input-output relation aki and the fraction of the purchases that were made on trade credit k i. This relation also applies to the comovement between the profits of sectors i and k because there is a direct relation between the fraction of k firms that defaults kand sector k's aggregate profits: N k i E( ) k (1 (µk))piaki iKk(1 µk). (5) N = (µk) N k E( ) E( ) i i This partial derivative captures only the first round e ects. The total derivative will also depend on the number of sectors that are in distress as a result of the initial default, that is, on the whole credit-chain. Although the model cannot be solved in closed form, the direct e ect derived in equation (5) predicts that stronger input output linkages (higher aki ) and a more extensive use of trade credit (larger k i ) lead to a stronger comovement between sectors after one of them su ers a negative shock This is the basic prediction I test in the data. The intuition for this result is straightforward. When an upstream supplier extends trade credit to downstream firms, it faces default risk. If there is a default, the supplier's financial health will depend on the size of the default with respect to its profits, which depends on the fraction of the total production sold to the defaulting sector (the input output linkages), and the fraction of those purchases made on credit (the use of trade credit). If some suppliers' financial health is significantly a ected they may also partially default on their obligations with banks and suppliers further upstream in the credit chain, generating the credit-chain amplification previously described in the literature and captured in this model. The model also highlights the crucial role that the use of trade credit has on the transmission of both idiosyncratic and sectoral shocks. In this model there would be no transmission if all purchases of intermediate goods were paid upfront using bank credit because banks are assumed to have deep pockets and not default, stopping the spread of the financial contagion.3 The model is too stylized to capture the direct demand e ects transmitted through input-output linkages with perfect credit markets (see Long and Plosser, 1983). The reason is that it considers only two periods to illustrate as simply as possible the transmission of default risk in a credit chain. However, in a dynamic version of the model a decline in sector k's revenues at t = 1 may reduce its demand for all intermediate inputs if firms are financially constrained or if the decline comes from a persistent productivity shock as in Long and Plosser (1983). I do not focus on this aspect of transmission because it complicates the model significantly and it has been extensively analyzed 3The model of Kiyotaki and Moore (1997) also exhibits this property. 7 in the literature, but I will consider it in the empirical approach. Summing up, this model predicts that the correlation between two sectors resulting from negative shocks will depend on the strength of their input-output linkages and the extent of use of trade credit. If the mechanism is empirically relevant, an increase in the use of trade credit should raise relatively more the correlation between sectors with stronger input output linkages. The ability of the model to explain unconditional correlations depends on the relative importance of negative and positive shocks, with the latter biasing tests of this prediction based on unconditional correlations towards zero. 3 Empirical approach To test the empirical predictions of the stylized model outlined above I move to a simple multisectoral model based on Long and Plosser (1983) and Shea (2002) that captures the essence of the mechanism and predicts an empirical relation between a specific measure of the use of trade credit along the whole chain of sectors linking two industries and their comovement. Consider Shea (2002)'s description of the evolution of sectoral output in a multi-sector economy without trade credit and upstream transmission of shocks through input-output linkages: yt = D , t (6) where y = (y1,...,yN) is a vector of sectoral output fluctuations (sectors 1 to N), = ( 1,..., N ) is a vector of sectoral shocks, and D is the DEM matrix derived by Shea (2002), whose dij element is the share of sector j in the total demand for industry i 's goods though direct and indirect linkages. This structural equation can be written in reduced form as yt = Byt + t, (7) where the elements of B = I D 1 measure the share of the total demand for industry i's good directly attributable to sector j. A simple modification to equation (7) introduces the possibility that the use of trade credit may a ect the propagation of sectoral fluctuations. Let pij be the fraction of the direct demand bij supplied through trade credit, so that bij = pijbij + (1 pij)bij. If this fraction had an additional e ect on the transmission of shocks, the coe cient of direct linkages would be bij(1+ pij) instead of bij, with parameterizing the importance of trade credit use. Further assuming for data availability reasons that pij is constant across suppliers (pij = pj i)4 the reduced form and structural relations among sectors would correspond to yt = B(I + P)yt + t, = A( ,B,P) , t (8) 4I discuss in detail the consequences of this assumption in section 4.1 8 where P is a matrix that has sector j's fraction of inputs purchased on credit (pj) in its main diagonal, and A = (I B(I + P)) 1 is a function of B, P, and . If the use of trade credit does not matter for the transmission of shocks ( = 0) this expression corresponds to the structural equation (6). The matrix A embodies the e ect of the total linkages and the credit chain. This can be seen by taking a linear approximation to A around = 0 to obtain A D + , (9) where D is the demand matrix defined above, and = DBPD captures the e ect of trade credit operating through the credit-chain. Assuming for simplicity that sectoral shocks iare i.i.d, equation (8) implies that the correlation between sectors i and k is jaijakj ik= , 1/2 j ij a2 j kj a2 where aij is the (i,j) element of the A matrix defined above. This correlation depends on a complex non-linear manner on the coe cient , but to a first order approximation it corresponds to jdijdkj dijdkj(~ij + c~kj) c ik + , (10) 1/2 j d2 d2 1/2 j ij d2 j kj d2 l il l kl where c~ij = dij ij l dil il, i, j, d2ij l il d2 measures the use of trade credit along the credit-chain linking i and j relative to the chains linking i and other sectors ( ij is the i,j element of ). The first term of equation (10), which I label the input-output distance between sectors i and k, corresponds to the correlation in absence of trade credit amplification and depends only on the strength of the linkages between these sectors, including indirect linkages through other sectors. The second term of the equation, which I label the credit-chain distance between industries i and k and denote by Cik, is a weighted average of the relative use of trade credit across all sectors j linking these industries, where the weights are determined by the product of the direct and indirect linkages between sector j and both i and k. Intuitively, shocks to sector j will increase (decrease) the correlation between industries i and k if the use of trade credit along the chain linking j and these industries is higher (lower) than average and if the linkages between j and both industries are important.5 5The derivation above assumes for simplicity that trade credit use equally amplifies positive and negative shocks. If the trade credit mechanism is asymmetric the exact expression will be such that the term multiplying the credit chain distance will be a function of and specific parameters of the distribution of shocks. For instance, when shocks are normally distributed with mean zero and variance 2 the coe cient multiplying the credit chain distance is /2, which means that the estimated coe cient from a regression is half the true one. 9 Equation (10) suggests that the hypothesis that the use of trade credit along the chain of industries linking two sectors a ects the transmission of shocks (i.e. whether = 0 in equation (8)) can be tested by checking whether the correlation between a pair of sectors i and k depends on their credit-chain distance, Cik. This test can be implemented by estimating the parameters of the following equation ikc = c+ ik+ Cikc + Wikc + ikc, (11) where ik captures the distance e ect and other fixed determinants of the correlation between a pair of industries, c is a country fixed e ect that captures among other things di erences in the relative importance of aggregate shocks, and Wikc includes other determinants of sectoral correlation. The variable of interest is Cikc, the credit-chain distance between industries i and k in country c. If shocks are independent but not identically distributed the approximation in equation (10) would contain an additional term that depends on the variances of the di erent sectors relative to the average, and if there is an aggregate shock the approximation would include a term that is a function of the distances and the variance of this shock. These additional terms cannot be directly computed from the data because the variances of the real shocks are unknown but I will address concerns that their presence may significantly bias the estimation of equation (11) building proxies for them that I will include in Wikc in some specifications. A similar derivation as above can be used to test the hypothesis that the use of trade credit relative to credit from intermediaries reduces the transmission of shocks, as predicted by the theory of "big pockets", by simply assuming that the fraction of purchases not financed by trade credit weaken the transmission and taking a linear approximation around a relative use of trade credit equal to zero to obtain that the coe cient of direct linkages would be bij(1 sij), where sij is the ratio of formal credit to trade credit as sources of financing the purchases of inputs. The rest of the derivation is analogous changing the sign of the coe cient , and leads to a credit chain distance measure based on the relative use of trade credit as a source of financing instead of its absolute use as derived above. Although the main discussion will focus on the measure based on the absolute use, I also build the measure based on relative use and apply it in section (5) to delve deeper into the mechanism of credit chain amplification. In what follows I assume that the ratio of an industry's use of trade credit to the average use in a country (Pic/Pc) is constant across countries, so I can express the elements of P as the product of this ratio, Pi, and a country's use of trade credit: Pic = Pi Pc. This assumption, which I discuss in detail in the next section, corresponds to assuming that some industries tend to rely more on trade credit financing than other for industry-specific reasons that do not vary substantially across countries. Under these conditions, the credit-chain distance between two industries in a given country can be written as the product of their generic credit-chain distance and the country's typical use of trade credit Cikc = Cik Pc. 10 4 Measuring sectoral comovement, input-output linkages, and the use of trade credit 4.1 The use of trade credit The main measure of the intensity of use of trade credit is the Payables to Cost of Goods Sold Ratio, P, (henceforth Payables Financing) which is the inverse of the payables turnover ratio widely used in financial analysis.6 For a given firm and year it corresponds to the average of the accounts payable at the end of years t and t 1 divided by the total cost of goods sold in year t and measures the fraction of input purchases financed with supplier's credit. The model presented in section 2, as well as the model of Kiyotaki and Moore (1997), suggests that the use of financial intermediaries as a source of finance for the purchase of intermediate inputs could attenuate the transmission of shocks through credit chains. To test this prediction, I also measure the relative use of intermediaries versus suppliers as sources of short-term financing by the ratio of short term debt to accounts payable, S. A high value for this ratio indicates that a firm obtains most of its short term financing from the financial system. The empirical analysis requires representative values of these measures for various manufacturing sectors in di erent countries, which I obtain in the following manner (focusing on the payables financing ratio, P, to fix ideas). First, I compute a representative value of P for each country c (Pc) as the median of P across all manufacturing firms in the country with more than five years of data in Worldscope 2006, except for the U.S.where I use data from Compustat. Additionally, for the U.S. I also calculate industry-level values of P, Pi,US, aggregating all firms in a given industry in the same way described above, which then I divide by the US country-level value, PUS, to obtain a relative index of payables for each industry in the U.S., Pi = Pi,US/PUS. Industries in the U.S. where Pi,US is higher (lower) than the overall U.S. median PUS will have a relative value of accounts payables Pi higher (lower) than one. Finally, assuming that these relative values for di erent industries are constant across countries, I estimate the industry i level value of accounts payables in a given country c, Pic, as the product of the US relative value of industry i and country c 's overall median, Pic = Pi Pc. Several aspects of this procedure require further discussion. The assumption of constant relative ratios across countries is equivalent to assuming that some industries tend to rely relatively more on trade credit financing for technological reasons that do not vary substantially across countries. This assumption was previously used by Fisman and Love (2003) who also provide a series of theoretical and empirical justifications for it. My motivation for making this assumption is mainly data availability. The 2006 version of Worldscope contains data on about 10,500 manufacturing firms in 58 countries, but for most of the 36 developing countries included it does not report su cient firms at the 3-digit ISIC industry-level. For instance, more than 40 percent of the countries in the sample have fewer than five manufacturing industries with more than five firms, and almost 6For a recent application of this measure to analyze the determinants of trade credit use, see Demirguc-Kunt and Maksimovic (2001). 11 all countries with more than 10 industries in this situation are developed countries. Nevertheless, there are three reasons to believe this assumption is unlikely to a ect importantly the results. First, available evidence suggests that the relative use of trade credit across industries is to an important degree technologically determined. The ratios computed for industries with more than 20 firms in a country are significantly correlated with the ratios computed for the U.S.; the coe cient of a regression of the available country-level ratios against the US ratios results is 0.92, significant at the 1 percent level. Second, to the extent that this assumption introduces classical measurement error to the measure of the use of trade credit at the country-industry level, it will result on a downward bias on the coe cient of interest that can be substantial (see appendix). Finally, using the available data to estimate the use of trade credit at the country-industry level produces similar results to those obtained under the assumption of constant relative ratios (see section 6). Another potentially important concern is that both data sources (Worldscope and Compustat) contain only information of publicly listed companies. If these companies' use of trade credit is di erent from that of the rest of the firms the estimates of the overall use of trade credit in a country will be biased. This is possible because public firms tend to be larger than private ones, and the use of trade credit is likely to depend on a firm's size. Nevertheless, as long as the variation in the use of trade credit across countries is mainly determined by those countries' structural characteristics the bias induced by the use of public companies will a ect mainly the level of the measures but not their relative position across countries. On a more pragmatic note, to my knowledge there are no better comprehensive data sources available. The measures of Payables Financing (P) and Short Term Debt to Payables (S) for the sample countries are presented in Table 1. The sample includes 43 countries; 22 developed, and 21 devel- oping ones (including 3 low income). The most represented regions are Western Europe, with 15 countries, followed by East Asia and Pacific with 10, and the least represented are Sub-Saharan Africa, South Asia, and North America, with only two countries each. The main constraint to the sample is the availability of data on the use of trade credit; only for ten countries trade credit data is available but not the correlation data described below.7 Column (3) indicates whether the measures of trade credit were built using data only from manufacturing firms (quality 1) or from the whole corporate sector (quality 2). In most cases (36 countries) there were enough manufacturing firms (10 firms) with more than five years of data to build the manufacturing specific measure, but in a few cases with less than this number non-manufacturing firms were also included in the measure. Nevertheless, all the results we discuss below carry to the subsample of high quality countries (see section 6). Firms in the average sample country finance about 15 percent of inputs' cost with trade credit (see the mean value at the bottom of column (1)). The distribution of this ratio across countries is symmetric around this mean and there is a reasonable degree of variation that puts the 25th and 75th percentiles at 12 and 18 percent. The typical country also uses intermediary and supplier's 7These countries are Argentina, Brazil, Czech Republic, New Zealand, Pakistan, Russia, Switzerland, Slovakia, Slovenia, and Thailand. The main reason why these countries are excluded is that they do not count with 15 observations of the growth rate during 1980-2003 to construct the correlation matrices. 12 credit in relatively equal proportions (see the bottom of column (2) where the mean and median of the ratio are close to 1), but there is important degree of variation around the central tendency, with the 25th and 75th percentile values located at 0.6 and 1.6, respectively. As expected, the correlation between the two measures of trade credit use is negative but small (-0.32). The relative measures (i.e. the ratios with respect to the mean across industries) of Payables Financing (Pi) and Short Term Debt to Payables (Si ) for the 28 ISIC industries in the U.S. are reported in Table 2. The median in column (1) is close to 1, which suggests that the distribution of Payables Financing across industries is relatively symmetric because the mean is one by construc- tion. In contrast, the median in column (2) is below one, indicating that the distribution is skewed to the left. As it was with the country level variables, the degree of variation across industries is reasonable but not large. The 25th and 75th percentiles of Pi across industries correspond to 0.87 and 1.21, respectively. Again, this variation is slightly larger for Si, with the corresponding figures at 0.71 and 1.15. 4.2 Sectoral correlations The main measure of sectoral comovement is the correlation of the growth rates of real value added across 28 industries covering the complete manufacturing sector. The data used to build these correlations come from the United Nations Industrial Development Organization (2005), Industrial Statistics Database (henceforth UNIDO). This database contains information on nominal value added, employment, number of establishments, wages and salaries, gross output, index of industrial production, and gross capital formation for 179 countries and 28 three-digit ISIC manufacturing industries during 1963 and 2003 but the actual coverage of the data is smaller because of missing information. I use data for the period 1980-2003 so that the comovement is measured over approx- imately the same period where trade credit data is available, and for reasons explained below keep only those industries with at least 15 years of data during this period. To compute the correlations, I first calculate the annual growth rates of real value added for each of the 28 industries in the di erent countries included in the database during 1980-2003. Following Rajan and Zingales (1998), I deflate the nominal value added of each industry in a country with the overall Producer Price Index for that country (from International Monetary Fund, 2005). Next, in each country I keep only those industries with at least 15 observations of this growth rate, and calculate their within-country, across-years correlation. To fix ideas, let gict denote the growth rate of industry i in country c between years t 1 and t. The estimate of the correlation between industries i and j in country c corresponds to T (gict g¯ic)(gjct g¯jc) Tij1 rij = c t=1 , (12) 1/2 T T t=1((Ti 1)/Ti)(gict g¯ic)2 t=1 ((Tj 1)/Tj)(gjct g¯jc)2 where g¯ic is the time-average of gict, Tij is the number of observations in which there are data for both sectors i, and j, and Ti is the number of observations in sector i. The reason to keep 13 only industries with at least 15 observations for the growth rates is to avoid having a short rank correlation matrix.8 I compute several other measures of comovement for robustness checks. One potential problem with the baseline measure is the use of a common deflator; if there were important heterogeneity in the evolution of prices across industries, the correlations computed with a common deflator may be driven by the correlation of relative inflation rates instead of the correlation of output growth. To address this concern I also compute the correlation of the growth rates of the index of industrial production reported in UNIDO in a similar way as described above. Although results obtained using this measure should not be a ected by the relative price problem, the production index data is of lower quality and smaller coverage than the value added data, so I keep the later as benchmark.9 Nevertheless, this choice will not a ect the results. Finally, I also computed the correlations of the HP filtered series of real value added and industrial production to address possible concerns regarding the de-trending procedure implicitly used when calculating the correlation in growth rates. The average correlations across manufacturing sectors during the period for the 43 countries in the sample, as well as the number of industry pairs with data are reported in Table 3. The average correlation of real value added and industrial production growth across sectors are always positive and around 0.25 on average. The inter-quartile range situates the average correlations between 0.2 and 0.36 for value added growth and 0.22 and 0.43 for the index of industrial production. This is consistent with the extensive evidence of positive comovement across sectors during the business cycle. The summary statistics are similar for the two measures, whose averages, reported in the table, are also positively and significantly correlated (correlation of 0.39 significant at the 2 percent level). As mentioned above, the coverage is better for the measure based in value added, for which there are 13,182 industry pairs compared with 12,548 for the measure based in industrial production. Both measures of correlation are positively but not significantly correlated with the average level of GDP per capita during 1980-2000. The positive correlation may indicate that idiosyncratic sectoral shocks are relatively more important in poorer countries. I will address this issue in the empirical analysis. Finally, Table 4 provides information on the variation of the correlation measures in the two dimensions of the data: industry-pairs and countries. The first two columns report information for the raw correlations, and the last two for the residual correlations after partialling-out country and industry-pair fixed e ects. Each column (which corresponds to one measure of correlation), first reports the overall variation (as measured by the standard deviation) and its decomposition in the within­ and between­countries dimensions. In columns (3) and (4) the between countries variation is zero by construction. The last two rows of each column show the average standard deviations of 8With N sectors and T observations, there are N(N 1)/2 correlation coe cients to be estimated from NT observations. The order condition, therefore, requires that T (N 1)/2 for a full rank matrix. With 28 sectors, this requires 14 observations as a minimum. I allowed for one more than that. 9Yamada (2005), points out that reported indexes are in many cases not consistent over time and across industries in terms of ISIC aggregation. Also, most developing countries tend to report indexes based on the fixed weight Laspeyres formula that are not appropriate for the computation of growth rates (chain-indexes being the correct ones). 14 the correlation measures within industry-pair across-countries, and within-country across industry- pairs. The former is a measure of the typical variation of correlations across industry-pairs and the latter a measure of the typical variation of correlations within a country.10 The table shows that both measures exhibit considerable variations in the within and between dimensions, although the latter is somewhat smaller than the former. The bottom two rows also show that the typical within- industry and within-country variations in correlations are similar, and, only slightly reduced by partialling-out industry-pair and country fixed e ects. This indicates that an important component of the variation of the correlations is idiosyncratic (within industry-pair and country). On one hand, this is good news for the identification of the mechanism because it focuses precisely on the within-country across industry-pairs variation of the data, but on the other hand it could just mean that correlations are noisily estimated and that the standard deviation of the correlations in the sample is a demanding metric to gauge the economic relevance of the estimates. 4.3 Input-Output linkages A central assumption of this paper is that input-output linkages between industries are largely technologically determined, based on which the linkage measures obtained in a country with good available information, like the U.S., can be extrapolated to the rest of the countries in the sample. The main reason for assuming a constant distance across industry pairs is the lack of comparable information on input-output relations for a broad set of countries at a good level of aggregation; OECD data on input-output matrices cover only 20 countries and 20 manufacturing sectors and data from Olarreaga and Nicita (2004) Trade and protection database covers a larger number of countries but divides manufactures in 17 sectors only and lacks some key variables required to construct the distance matrices. Nevertheless, this assumption should be relatively uncontroversial because, by construction input-output matrices reflect technological links across sectors, and, therefore, their variation across countries is likely to be limited. Table 5 and Figure 1 provide some evidence in this regard. Table 5 reports the variance decomposition of the Direct Requirement Matrices, a crucial input on the construction of the distance measures, obtained from Olarreaga and Nicita (2004) for 67 countries. Although the level of aggregation is di erent from the one used in this paper, the table makes apparent that almost all the variation in the direct requirement matrices is within country, across industry-pairs. Additional evidence in support of this assumption comes from Figure 1, which plots the input-output distance built using U.S. data as described below and the same distances estimated using U.K. data. The relation between the two measures is striking (the correlation is 0.98). Under the assumption of technologically determined links, I build the matrix D of contempo- 10 For any variable xic with variation across industry pairs (i) and countries (c), the within country variation corre- sponds to the standard deviation of xic x ¯.c +x.., where the dot represents the average in that particular dimension. ¯ The between-countries variation therefore corresponds to the standard deviation of x.c. The within industry-pairs, ¯ across-countries variation is computed as follows: for each industry-pair, I first calculate the standard deviation of its correlation across the countries in the sample and then compute the average of these standard deviations across all industry-pairs. I compute the within-country, across industry-pairs variation similarly by switching the order of aggregation. 15 raneous transmission described in section 3 (see Eq. [6]) for the 28 three-digit ISIC manufacturing industries in which I have comovement data using information from the 1992 commodity-by-industry (USE) and industry-by-commodity (MAKE) matrices produced by the U.S. Bureau of Economic Analysis (BEA) and my own correspondence between the BEA industry classifications and ISIC. The appendix describes the construction of this matrix in detail. Using this matrix I compute the input-output and (generic) credit-chain distances across industry pairs as described in equation (10) of section 3. Since the sectoral classification consists of 28 industries at three-digit ISIC level, the C matrix of credit-chain distances contains 378 di erent distances (the distance between a sector and itself is normalized to 100). The distribution of the distances is significantly skewed with most of its mass close to zero (around 40 percent) because of the well-known sparsity of the input-output matrices (see Figure 2).11 Table 6 shows the twenty industry pairs with the closest credit-chain distances, which represent about five percent of the total number of pairs. The closest pairs are intuitive; for instance, the two pairs with the smaller distances are the ones formed by the Textiles and Wearing apparels industries and Transport Equipment and Fabricated Metal Products industries. Although pairs that are close according to credit-chain distance also tend to have a small input-output distance, there is some important degree of variation between these measures of similarity; only half of the pairs reported in Table 6 are also among the 20 with the closest input-output distance and the rank correlation between them is only 0.17. The overall correlation between both distance measures is 0.66, but is mainly driven by the di erence between the group of industries with small linkages and the rest; in fact, the correlation drops to 0.5 when looking only to those industry-pairs with input-output distances above the median. Thus, the overall pattern suggests that the measure of credit-chain distances is partly determined by the input-output linkages among sectors, but the relative use of trade credit in the chain linking two sectors introduces non-trivial variation in the measure with respect to the standard input-output distance. 5 Results This section presents the results of the test of the hypothesis that the intensity of use of trade credit along the chain linking two industries increases their correlation. Following the empirical approach outlined in section 3, I test this hypothesis by using the data to estimate the parameters of equation (11) and test whether the coe cient of the interaction of the generic credit chain distance and the use of trade credit in the country (a) is significantly positive. The coe cients obtained by ordinary least squares (OLS) for various versions of equation (11) that include di erent sets of additional controls are presented in Table 7. In Panel A the dependent variable is the correlation of growth rates of real value-added between industry-pairs in di erent countries and in Panel B is 11According to Horvath (1998, 2000) this sparsity is crucial to understand why the law of large numbers does not necessarily apply to models where transmission of shocks is given by input-output relations. 16 the correlation of growth rates of the industrial production index. The estimated coe cients for the parameter of interest are reported in the first row of the table. The main result of the paper can be seen in the baseline regression reported in column (1), which includes only the credit-chain distance (i.e. the interaction of the generic credit chain distance Cik and a country's Payables Financing) and a set of industry-pair and country fixed-e ects. The coef- ficient obtained for the credit-chain distance based on payables is positive and strongly significant. An upstream sector that is linked to a downstream industry through a series of sectors with high use of trade credit su ers a stronger contraction as a result to a negative shock to the downstream industry. This in turn translates in a higher correlation between them. Columns (2) and (3) of each panel show that the basic result survives the addition of variables capturing the size and the first two moments of the growth performance of the di erent sectors across countries. Regressions in column (2) add two measures of the size of each of the industries included in a country pair: the (log) average number of firms in each industry during the period and the industries' average share of total manufacturing value added. Adding the number of firms is important to control for the e ect of diversification on sectoral correlation; assuming that there are both aggregate and idiosyncratic firm level shocks, an increase in the number of firms in all sectors of a country results in a relatively larger role for aggregate shocks and more correlated sectors. Adding the share on total manufacturing value-added of each of the industries forming an industry pair controls for the possibility of aggregate spillovers. As noticed by Shea (2002), in presence of external economies of scale, such as those suggested by Baxter and King (1991), larger industries will generate larger spillovers to the rest of the economy and therefore be more correlated with all other sectors. The results show that controlling for these variables does not a ect the coe cient obtained for the main interaction. As expected, the coe cients obtained for the (log) number of firms in each of the industries are positive and significant, indicating that industries with more firms are more correlated with the rest of the economy and among themselves. Surprisingly, the coe cients obtained for the industry shares are significant but negative, but this is because the regressions are already controlling for the number of firms, which correlates with the industry share. Regressions including only the shares yield positive coe cients. Intuitively, after controlling for the number of firms, the share of an industry is a measure of value added or output per firm. The results therefore indicate that industries whose firms are relatively more productive (in the sense of having a higher level of output per firm) are typically less correlated with the rest of the economy. The regressions in column (3) further add the average and standard deviation of the corresponding growth rate of each of the industries included in a pair to control for the simple possibility that industries with similar trends or shocks can be more correlated within and across countries. The results indicate that industries that grow relatively faster and are more volatile tend to be less correlated with the rest of the sectors, but, again, adding these controls does not substantially a ect the sign, magnitude, or significance of the coe cient for the credit chain distance.12 12Similar results are obtained when adding the di erence in average and standard deviation between the two industries instead of adding the values for each industry separately (not reported). 17 Figure 3 illustrates the di erential e ects that the use of trade credit has across industry- pairs with di erent credit-chain distances. Panel A shows the relation between a country's level of payables financing and the correlation of the Iron and Steel and Transport Equipment industries (ISIC codes 371 and 384, respectively). The positive relation is apparent; for this industry pair an increase in the use of trade credit increases the correlation. Panel B instead shows the same relation for the pair formed by the Tobacco and Footwear industries (ISIC 314 and 323, respectively), which have almost no credit chain weighted linkages. The lack of association between trade credit use and correlation is also evident. These are the di erences that are captured by the interaction variable. Panels C and D show that this e ect is not exclusive to the pairs of industries just described. Panel C plots the relation between payables financing and sectoral correlation for the 20 industry pairs with the closest credit chain distance and panel D does the same for the 20 industry pairs with the highest distance. Again, the panels show that an increase in the use of trade credit increases the correlation of those industry pairs with a high Cik but has no e ect on those that are far in the credit-chain sense.13 In terms of economic magnitude, the point estimates (column (2)) indicate that a one-standard- deviation increase in the overall measure of credit-chain distance Cikc results in a decline in corre- lation of four percentage points or 16 percent the standard deviation of industry-pair correlations within and across countries (0.27). In terms of relative di erences, an increase in payable financing from the 25th to the 75th percentile level would increase the correlation of the industry pair at the 75th percentile of credit chain distance (closeness) in almost two percentage points more than that of the pair at the 25th percentile, or about ten percent of the interquartile range of average sectoral correlations (0.16). These magnitudes are economically meaningful and also likely to be conservative for several reasons. First, the appendix shows that the coe cients likely su er from attenuation bias because of the presence of measurement error in the use of trade credit. The size of the bias will depend on the unknown variance of the measurement error, but the estimates indicate that it can be substantial. For instance, for a standard deviation of the error equal to the standard deviation of Payables Financing across countries the bias may be as large as 60 percent. Second, as explained in section 2, the credit-chain amplification mechanism is mainly associated with increases in the correlation across industries resulting from the impact of negative shocks. Therefore, using the unconditional correlation as dependent variable will also lead to attenuation bias because there is an important component of the observed correlations that the model does not explain. Third, the measures of correlations most likely contain an important amount of noise resulting from the small number of observations used in their estimation. Finally, as shown in Table 4 and briefly discussed above, the measured correlations of growth in value-added and the index of industrial production across industry-pairs and countries exhibit a considerable amount of variation even after accounting for country- and industry-specific components, which makes di cult for any individual mechanism to account for a large fraction of the residual variability. The aggregate implications of the mechanism can also be estimated by computing the impact on 13The results reported in the figure are robust to excluding Italy (ITA) from the sample. 18 aggregate volatility of an increase in a country's overall use of trade credit implied by the coe cients. This can be easily done by noticing that a country's aggregate manufacturing variance ( 2 Y ) can be written as 2 1 1 Y = 2R 2 , where is a vector containing the share of each manufacturing sector on total manufacturing value added, is a matrix that has the variance of each manufacturing sector in the diagonal (and zero elsewhere), and R is the correlation matrix between the di erent manufacturing sectors, such that Ri,j = i,j. Thus, extrapolating from equation (11) 2 Y = 1 1 2C~ 2 , P where C~ is the matrix of generic credit-chain distances, and the partial derivative indicates that the expression corresponds to the impact of trade-credit use through changes in correlations only. Evaluating this expression for a country with the average sectoral shares and variances yields that an increase in the aggregate level of Payables Financing from the 25th to the 75th percentile level (from 12 to 18 percent) reduces the variance of value-added growth in the manufacturing sector of the average country by five percent. A bigger increase from the lowest to the highest level ac- counts payables in the sample (from 7 to 33) reduces this variance in 20 percent.14 Moreover, as discussed above, these magnitudes are likely to be a lower bound of the true e ect on aggregate vari- ance. Overall, the evidence suggests that the credit-chain amplification mechanism is quantitatively relevant. Using the Short Term Debt to Payables ratio of the di erent industries and countries to compute the credit-chain distance Cikc, a di erent albeit complementary measure of the use of trade credit, yields similar results. This can be seen in columns (1) to (3) of both panels of Table 8. Since this is an inverse measure of the use of trade credit, the credit chain mechanism predicts a negative coe cient for the main variable (see the discussion in section 3). Consistently, in all the regressions the coe cient is negative and strongly significant; an increase in the relative importance of bank credit vis-à-vis supplier's credit along the credit chain reduces the transmission of shocks. This is consistent with Kiyotaki and Moore (1997)'s idea that it is the presence of "big pockets" along the chain that can dampen the transmission of shocks. The economic magnitude of the coe cient associated with this measure of trade credit use is smaller than that obtained when using payables financing. A one standard deviation increase in the credit chain distance based on this measure reduces the correlation of real value added growth in two percentage points; half of the reduction obtained with the baseline measure. Despite its smaller magnitude, the attenuation of sectoral comovement resulting from an increase in the use of formal financing is mostly complementary to the one resulting from an overall decrease in the use of trade credit, as shown in columns (4) to (6) of both panels of Table 8. The regressions reported in this table include both measures of credit-chain distance simultaneously and show that 14Seven and 25 percent of the inter-quartile range of variance of aggregate manufacturing value added growth. 19 the coe cients associated with the measure based on payables financing decline only slightly with respect to those previously reported and the coe cients associated with the short term debt to payables are mostly una ected. This indicates that both the use of trade credit relative to other sources of short term financing and as a fraction of input purchases are important for amplification. A reduction in the use of trade credit as a fraction of the cost of inputs that is compensated with an increase in the use of intermediary financing has a much larger impact on reducing sectoral comovement. Evaluated at the median level short term debt to payables, the same increase in payable financing from the 25th to the 75th percentile discussed above would increase the correlation of the industry pair at the 75th percentile of credit chain distance in almost 2.5 percentage points more than that of the pair at the 25th percentile; this is fifty percent larger than the direct e ect without changing the use of short term debt, and about 15 percent of the interquartile range of average sectoral correlations. 6 Robustness This section explores the robustness of the main result to variations in the sample of countries and industry-pairs included, the use of di erent measures of correlation and credit chain distance, and alternative explanations associated with potential omitted variables. 6.1 Sample issues A first concern regarding the baseline results is that they may be driven by special observations. This concern is valid considering the skewness of the distribution of the measures of credit-chain distances and trade credit use, but it turns out to be unimportant. First, no individual country or industry pair is driving the results. This can be seen in the results summarized in both panels of Figure 4. Panel A reports the distribution of the 43 coe cients for the credit-chain distance obtained after re-estimating the baseline regression dropping one country at a time. All coe cients are positive, statistically significant at the one percent level, and tightly distributed around the median coe cient of 1.98, with the smallest coe cient (1.36) obtained when Japan is excluded from the sample. The histogram presented in Panel B addresses the robustness to the exclusion of specific industry pairs in a similar manner; all the 378 coe cients, are significant and similar to the coe cient of the baseline regression. Second, no particular groups of observations are driving the findings. The regressions presented in Table 9 check for various of these possibilities. Column (1) reports the results obtained using only the sample of 39 "high quality" countries where the use of trade credit of the manufacturing sector could be computed. Column (2) presents results obtained after dropping the ten percent of countries with the highest and lowest levels of Payables Financing (i.e. dropping 20 percent of the sample in total). Column (3) shows similar results after dropping the 5 percent of industry pairs with the highest and lowest generic credit chain distances (10 percent of sample), and Column (4) reports the results obtained using a robust estimation method. In all cases and in both panels the coe cient of the credit-chain distance is not importantly a ected. 20 6.2 Measurement The regressions reported so far have used the correlation in the growth rates of real value added and industrial production index as measures of comovement. The exercises presented in Table 10 consider three alternatives. Column (1) addresses potential concerns with the use of first di erences as a de-trending procedure and reports the coe cients obtained when using as dependent variable the correlation of the series of real value added de-trended using the Hodrik-Prescott filter. Column (2) uses a transformation of the correlation of the growth rate of real value added aimed to tackle the potential heteroskedasticity of the residuals resulting from the correlations taking values between -1 and 1. The transformed correlations are r~ik = ln(rik + 1) c c ln(1 rik), which assumes a parametric c non-linear relation between the correlations rik and the variables of the model (see Otto et al., c 2001, who introduced this). In both cases, the coe cient for the credit-chain distance is negatively significant. Finally, Column (3) addresses concerns about the impact of outliers in the measures of correlations by replacing the standard measure described in equation (12) by a robust measure. The results are again largely una ected.15 The last two regressions presented in Table 10 check the robustness of the results to changes in the measure of credit chain distance. Column (4) reports the results obtained after relaxing as much as possible the assumption of fixed relative use of trade credit across industries that introduced in section 4.1 and using the available country-specific measures of sectoral use of trade credit that can be obtained from Worldscope. The measure of credit chain distance applied in this column uses the country-industry value of payables financing, Pic, measured directly from Worldscope for all country-industries with data on more than five firms (395 cases in the sample), and the product of the U.S. relative and the corresponding country median (Pic = Pi,US Pc) in the remaining cases (1305 cases). The results show that using as much country level information as possible does not a ect the main conclusion of the paper; the coe cient obtained for the credit chain distance is similar to that previously reported. Column (5) explores the sensitivity of the main coe cient to the particular choice of the U.S. as baseline country for measuring the input-output relations across industries and uses instead credit-chain distances computed using input-output matrices from the U.K. The rest of the procedure to construct the distance is the same as in the baseline case. Again, the coe cient is of the right sign and significance and are only marginally smaller. 6.3 Are industries with small credit chain distances similar in other dimensions? Industries that are closer in terms of credit-chain distance may also be similar in other dimensions such as technologies, degrees of external financial requirements, liquidity needs, capital intensity, and stage of production. To the extent that these other dimensions of similarity are emphasized by country characteristics related to the use of trade credit, the results could be spuriously related to the omission of these potential determinants of comovement. The regressions presented in Table 11 explore several of these possibilities. 15The specific measure is the weighed correlation between two industries, where the weights are obtained from a robust regression between the two growth rate series. 21 The measure of credit chain distance could be capturing technological similarities across indus- tries that buy (sell) comparable proportions of goods from (to) other industries, shown by Conley and Dupor (2003) to be related to the comovement of productivity in the US. To check this con- cern I add the interaction of the measures of BUY and SELL distances between industries of Conley and Dupor (2003) and a country's payables financing to the baseline specification.16 The main coe cient is not importantly a ected and the results indicate that only industries that sell goods to similar sectors in similar proportions are significantly more correlated in countries with higher payables financing. This suggests that a higher use of trade credit increases the relevance of backward linkages as sources of comovement, consistently with the upward transmission of shocks emphasized by the credit-chain mechanism. Industries with high degrees of external financial requirements or liquidity needs may respond more to shocks in countries with underdeveloped financial systems. Since a country's use of trade credit may be related to the availability of formal financing associated with financial development it is necessary to control for this potential source of comovement. The degree of financial development of a country can also be related to the availability of capital, which can induce comovement across industries with similar capital intensities. All these possibilities are considered in the regression reported in column (2) that adds to the baseline specification the interaction of an industry-pair's similarity in terms of external financing requirements (measured as in Rajan and Zingales, 1998) and liquidity needs (measured as in Raddatz, 2006) and a country's level of financial development, and the product of an industry-pair's similarity in terms of capital intensity (measured as in Raddatz, 2006) and a country's (log) aggregate capital per worker. Along each dimension, the degree of similarity is computed as the absolute value of the di erence between each industry's measure. The degree of financial development is measured as the (log) average private credit to GDP ratio during 1980-2000 from ?, and capital per-worker is the average for the same period from ?(henceforth PWT). The results show that only industries with similar liquidity needs are less correlated in financially developed countries, suggesting that the increased access to financing eases the response of industries with high liquidity needs to negative shocks and reduces the resulting correlation,17 but this channel does not importantly a ect the coe cient associated with credit chain distance. Sectors producing goods with similar degrees of complexity can also be relatively more correlated. For instance, Clark (1999) and Huang (2001) have documented that the response of quantities versus prices in response to a monetary policy shock vary with the stage of production. Therefore, sectors at comparable stages of production will exhibit more coordinated responses in quantities than sectors at di erent stages. I use the Gini coe cient of the distribution of input costs to capture the complexity/stage of production of the goods produced by an industry (see Blanchard and Kremer, 1997; Kremer, 1993) and measure the similarity of two industries in this dimension 16Let (i,j) be the dollar value of compensation to sector i for goods used in industry j obtained directly from the and SELL distances correspond to the Euclidean distance of the~vectorsSELL industries and correspond to BUY (i,j) = input output tables. An industry's fraction of the costs and demand of other sector are obtainedPby normalizing the12 matrix across rows and columns to obtain B(i,j) = ( i,j)/ Pk (k,jand ) and (i,j) = ( i,j)/ k (i,k). The BUY 17For the relation between liquidity needs and vulnerability to financial underdevelopment see Raddatz (2006). ^Pk( 1 B(k,i) B(k,j))2 2 of costs and demand shares between~two (i,j) = ^Pk( (i,k) (j,k))2 . 22 as the absolute value of the di erence in their Gini coe cients (obtained from Cowan and Neut, 2002). The regression presented in column (3) adds to the main specification the interaction of this variable with the overall volatility of a country measured by the standard deviation of the growth rate of real GDP per capita. Although, consistently with this mechanism, sectors producing goods of similar complexity are indeed more correlated in more volatile countries, the result regarding the role of credit-chains remains unaltered. The credit-chain mechanism emphasized throughout the paper focuses mainly on the transmis- sion of shocks through backward linkages but forward linkages may also be important, not only for technological reasons (productivity shocks upstream translating into an increase in production downstream because of an expansion of the supply of intermediates) but also for credit reasons because an upstream shock could a ect the supply of trade credit for downstream industries. The regressions presented in columns (4) to (6) check for this possibility. Column (4) starts by adding the interaction between the input-output distance between two industries derived from the COST matrix of transmission of upstream shocks described in Shea (2002) and a country's payables fi- nancing to check if the main result could be simply related to upstream transmission being more important in countries with higher use of trade credit.18 The results suggest that industries with stronger forward linkages tend to co-move more in countries with higher payables financing, but this channel does not importantly a ect the findings regarding the role of credit chain distance. The regressions reported in columns (5) and (6) go one step forward and ask whether the forward linkages may matter in the transmission of credit default. The regression in column (5) adds to the baseline specification a measure analogous to the credit chain distance built using the COST matrix (instead of D) and payables financing to check for the possibility that upstream shocks a ect relatively more downstream industries that rely more on trade credit. The regression in column (6) instead focuses on the constraints to the supply of credit and adds a measure of distance built using the COST matrix and the upstream industry's level of accounts receivables to sales as a measure of the amount of credit o ered to downstream industries. The results indicate that only the measure based on the use of credit downstream is related to the degree of comovement in countries with higher aggregate use of trade credit, which suggest that upstream shocks benefit relatively more downstream industries that use relatively more trade credit, but do not provide evidence consistent with contractions on trade credit supply. The derivation of equation (10) assumed that sectoral shocks were independent and identically distributed. Relaxing this assumption would result in additional terms on the linear approximation that will also be a function of the elements of the D matrix, raising, therefore, the possibility that the credit chain distance could be picking the e ect of one of these omitted variables. One possible deviation from this assumption comes from di erences in the variance of the shocks across industries. In this case, the linear approximation (10) would include a term that is a function of the coe cients of the D matrix and the relative variances of the sectoral shocks. This term 18The input-output distance based in this matrix is computed analogously to the distance based on the D matrix (see Eq. [10]) 23 cannot be directly computed because these variances are unobserved, so the regression in column (7) adds a proxy built using the relative variances of the output of the various industries instead. Although imperfect, this measure should properly capture the correlation coming from the input- output linkages, which is the source of concern. The results clearly indicate that this concern is not relevant for the main findings. Another possibility is the presence of aggregate shocks. In the model of section 3, sectoral shocks result in aggregate fluctuations because of the input-output linkages, therefore, aggregate shocks are not necessary to generate comovement. Nevertheless, the presence of truly aggregate shocks will also add a term to the linear approximation that will depend on the input-output linkages and the variance of the aggregate shock, but whose exact form will depend on specific assumptions regarding the transmission of this shock. For this reason, the regression in column (8) controls for this potential source of bias in a general manner by adding the interaction of a country's overall volatility and the generic credit chain distance, which maximizes the potential correlation with the component of the transmission of aggregate shocks that is a function of input- output linkages. 6.4 Alternative explanations and further evidence on the mechanism Since the measure of credit chain distance was constructed using U.S. data, one possible interpreta- tion of the findings is that they are simply indicating that input-output linkages can better explain sectoral comovement in countries that are similar to the U.S. This is unlikely because the U.S. is not atypical in its level of trade credit use (payables financing in the U.S. is 0.13, just below the sample median of 0.14). Nevertheless, I check for this possibility by adding to the baseline specification the interaction of the generic credit-chain distance and the (log) average real GDP per capita (from PWT) during 1980-2000. The results, reported in column (1) of Table 12, show that consistently with the measurement hypothesis linkages are more important in developed countries, but this does not significantly a ect the findings regarding credit chains and sectoral comovement. Column (2) shows similar results when also controlling for a country's overall degree of volatility (as measured by the standard deviation of the growth rate of real value added during 1980-2003). The results obtained using the ratio of short term debt to payables (see Table 8) indicate that the actual use of formal financing from financial intermediaries vis-a-vis suppliers credit partly alleviates the transmission of shocks through credit chains. Building on these results, column (3) checks whether the availability of formal financing a ects the importance of the credit-chain mechanism by allowing the parameter to depend on a country's degree of financial development, which amounts to add a triple interaction term. Surprisingly, the results indicate that the credit chain mechanism is stronger in more financially developed countries.19 A possible interpretation of this finding is that financial development also increases the supply of trade credit, at least in this sample, which would be consistent with Demirguc-Kunt and Maksimovic (2001) who maintain that these sources of financing are complements. If this is the case, increasing the availability of formal 19This finding is robust to controlling for the simple interaction between financial development and GDP per capita and generic credit chain distance. 24 credit does not necessarily reduce the importance of trade credit in the actual composition of firm's financing. To check whether the actual composition matters, the regression in column (4) instead allows the parameter to depend on the country level ratio of short term debt to payables and finds that the use of formal financing reduces the importance of the credit-chain mechanism. A firm that sells its intermediate goods to downstream producers in a foreign country is probably not a ected by shocks to local downstream producers. This observation suggests that the relevance of the credit chain mechanism may depend on a country's degree of trade openness. The exercise presented in column (5), which includes the interaction between an industry pair's generic credit chain distance and a country's degree of trade openness (measured as the log ratio of exports plus imports to GDP), shows that indeed input-output linkages are less important in more open countries but that this finding is not behind the main result of the paper. This result is robust to controlling for the overall level of development (not reported). The credit chain distance between two sectors can be decomposed in various forms that shed light on the mechanism and the sources of identification. First, the regression reported in column (6) determines the relative importance of direct and indirect linkages (first and higher round e ects in the transmission of shocks) by computing the credit chain distance resulting from first round e ects only and adding it to the baseline specification. 20 The results show a higher and statistically significant coe cient for the distance based on first round e ects than for the overall distance, which suggests that the marginal impact of these linkages on sectoral correlations is higher. However, their overall contribution to the di erences in correlation is smaller because although their coe cient is almost four times larger than that of the overall distance, their sample variation is an order of magnitude smaller. Indirect linkages are, therefore, qualitatively and quantitatively important for the mechanism. Second, the regression in column (7) disentangles the contribution to sectoral correlations of di erences in the use of trade credit across industries from the contribution of their average use. The credit chain distance in equation (10) can be trivially written as the distance when all industries use the average level of trade credit plus the contribution of the industries' deviations from that average. These two components (that add to the credit chain distance) are separately included in the regression in column (8). The results show that it is the average use of trade credit along the chain of industries linking two sectors that is behind the result of this paper. The di erences across industries do not contribute significantly and, if anything, tend to reduce the correlation. The lack of significance is not surprising considering the limited variation of the relative use across industries compared to the variation in input-output linkages, and the negative sign seems to indicate that industries with stronger linkages with the rest of the economy tend to use relatively less trade credit. 20Iterating on the reduced form equation (7) yields the recursive representation of equation (8), y = (B + B2 + y, which shows that the structural form corresponds to the sum of the first, second,eand ~ B + .. . + B ) + B ~3 ~N ~N +1 higher order rounds of e ects of the shocks. The first order e ect is given by B and the resulting correlation matrix is derived as in section 3. e 25 7 Concluding remarks This paper has provided indirect evidence of the presence and quantitative importance of the credit- chain amplification mechanism first described in Kiyotaki and Moore (1997) by looking at its impli- cations for the comovement of industries within and across countries. The results, which exploit the variation in correlations of industry-pairs and use of trade credit across countries, robustly indicate that, consistently with the presence of a credit-chain amplification mechanism, an increase in the intensity of use of trade credit along the input-output chain linking two industries augments the correlation between them. In terms of economic significance, the evidence indicates that, without being a first order de- terminant of comovement and volatility, the credit-chain amplification mechanism is quantitatively relevant. Moreover, there are several reasons to consider the estimates as conservative. In addition to the possible causes of attenuation already discussed in the paper, there are at least three other dimensions that deserve to be considered in future work. First, one possible cause of the limited variation in the measures of distances across industry pairs is the level of aggregation used in this paper. At a lower level of aggregation, it may be possible to capture more of the local interactions between similar industries that is now part of the "diagonal" terms of the distance matrices. The reason for not working at a lower level of aggregation in this paper is the exponen- tial increase in the number of observations in the time dimension that are required to satisfy the order condition in the computation of the correlation matrices, but this restriction may become less important as the time coverage of existing datasets increases. Second, the data used in this paper do not include the non-manufacturing sectors, such as retailing, the main users of trade credit and likely a major contributor to credit chain amplification. Finding ways of including these sectors is likely to be relevant to quantify the real importance of this mechanism. Finally, for data availability reasons and to avoid endogeneity problems, this paper's analysis focused on the unconditional cor- relations, despite that according to theory the credit-chain amplification mechanism a ects mainly the comovement resulting from negative shocks. As discussed in the paper, the presence of this asymmetry is likely to weaken the tests and bias downwards the estimates of the relevance of the mechanism. Addressing some of the issues should form part of future research in this area. 26 Appendix A Building the Ultimate Demand Requirements Matrix D Following Shea (1990, 2002) I build the D matrix for the 28 three-digit ISIC industries using information from the 1992 commodity-by-industry (USE) and industry-by-commodity (MAKE) matrices produced by the US bureau of economic analysis (BEA) and my own correspondence between the industry classification used by the BEA and ISIC . The first step in the construction of this matrix is the computation of the Direct Cost Share Matrix (DCS) whose i,j element is the amount of industry i input required per dollar of industry j's output. I compute this matrix directly from the U.S. benchmark input-output matrices by multiplying the MAKE and USE matrices to obtain each industry's use (in dollars) of goods produced by other industries, dividing each column by the total value of the output produced by that particular industry to obtain the Direct Requirement Matrix(DRQ), and then correcting this matrix for the presence of non-zero terms in the diagonal resulting from aggregation according to DCS = (I diag(DRQ)) 1 (DRQ diag(DRQ)) where the diag() operator extracts the main diagonal of a matrix. Next I compute the COST matrix, whose i,j element is the ultimate dollar requirement of good j per dollar sold of good i, by calculating the transpose of the Leontie inverse of the DCS matrix (I DCS ) 1, and extracting the 28 rows and columns corresponding to the ISIC manufacturing industries. Finally, I calculate the Ultimate Demand Requirement Matrix (D) as N Dik = COSTkiak/ COSTziaz , z=1 or, in matrix notation, D = diag(COST Diag(a)) 1COST Diag(a) where ak is the steady state share of good k in overall consumption, a = (a1,...,aN) and the Diag() operator takes a vector and places it in the main diagonal of a matrix. I compute the industry k's final demand following in Shea (2002) as the sum of purchases from consumption, government, and other non-manufacturing industries. The matrix for the U.K. used in the robustness section was built in the same manner. 27 B Measurement error B.1 Errors in the measurement of sectoral use of trade credit I now describe the impact that the presence of classical measurement error in the sectoral use of trade credit on the estimated coe cient for credit chain distance. According to equation (9) in the paper, A D + C to a first order approximation, where C = DBPD. Consider the case of classical measurement error in the diagonal matrix P of sectoral use of trade credit P = P~ + , where is a matrix that contains in the diagonal the error terms in the measure of trade credit use of the di erent sectors and zero elsewhere and P~ is the matrix of true sectoral use. Assume that, according with standard measurement error, the matrix is such that E[~i ] = 0, p j i, j 1,..,N. (13) Under these assumptions the true matrix C~ can be written as C~ =DB(P )D =C µ. It is easy to verify that under assumption (13) the components of C~ and (linear combinations of p~ and , respectively) are orthogonal, so that µ corresponds to standard measurement error for the C matrix. On the other hand, the correlation matrix of sectoral output correspond to = diag(AA ) 1 1 2 AA diag(AA ) 2 , (14) where A = (I B(I P)) 1, and, to a first order approximation around = 0 = ( )| d ( ) =0 + . d =0 After some simple but tedious algebra, this approximation can be written as ( ) diag(DD ) 1 1 1 3 2 DD diag(DD ) 2 diag(DD ) 2DD diag(DD ) 2 diag CD + ~ diag(DD ) 1 1 3 1 2 CD + DC~ ~ diag(DD ) 2 diag(DD ) 2 diag(CD )DD diag(DD ) ~ 2 , which is the matrix form of equation (10) in the paper. This expression clearly shows that the first order approximation of the correlation matrix is linear in the true matrix C~, and, therefore, as a 28 function of and C can be written as ( ,C) = ( , C~) + ( ,µ), where it is still true that the first term on the RHS is a linear combination of the elements of P~ and the second term is a linear combination of the elements of µ. Therefore, under assumption (13), the second term corresponds to classical measurement error associated with the first term, which attenuates the estimated coe cient for . Using a similar derivation is possible to show that deviations from the assumption of a firm's constant relative use of trade credit across suppliers (pij = pj) introduce no bias in the estimated coe cients to a first order approximation. I skip the details for reasons of space, but the argument relies on the matrix A being equal to (I B(I + P) B) ~ 1, where under classical measurement error the matrix B (the element by element product of the matrix D and the random deviations ~ from the common usage e ect) is orthogonal to P. This implies that to a first order approximation the correlation matrix equals the one obtained under the common usage assumption ( , C~) plus an orthogonal term. Although the derivation above shows that to a first order approximation measurement error in the sectoral use of trade credit translates into measurement error in the distances, one might still be concerned that the non-linearity of the true relation between correlations and the matrix of sectoral use could change the bias. In other words, one might be concerned about the quality of the first order approximation. To address this concern I performed a series of Montecarlo simulations to explore empirically the direction of the bias introduced by the measurement error. I perform these simulations in the following manner. First, I take my current estimates of the sectoral use of trade credit, measured as Payables Financing as the true values (P~) and generate a series of "observed" values (P) by adding some noise to the current estimates. I try di erent mag- nitudes for the size of this noise around the baseline level corresponding to the standard deviation of the observed values of payables financing across industries and countries. I then assume a value of 2 for (as in the baseline regression) compute the true A matrix as A = (I B(I P~))1, and use this matrix to compute simulated correlations among industries within countries according to equation (14). Next, I compute the credit-chain distance measure as described in the paper using the simulated values of P and run a regression between the simulated correlations and credit-chain distances for the various industries in the sample of fictitious countries, which has the same size as the actual sample, controlling for industry-pair and country fixed e ects. I repeated this exercise 100 times for each value of the standard deviation of the noise between 0.01 and 0.16 (the standard deviation of P in the sample is 0.08). The results are summarized in Figure A1 (Panel A), which, for each value of the standard deviation of the errors shows the empirical distribution of the di erence between the simulated and true coe cients. The continuous line correspond to the mean of this distribution and the shaded area is the 95 percent empirical confidence interval. The x-axis display the corresponding standard deviations of the measurement error as a fraction of the sample standard deviation of P. The figure 29 clearly shows that the introduction of noise in the measure of the sectoral use of trade credit results in attenuation bias and that the size of the bias increases with the magnitude of the measurement error (as captured by its standard deviation). It also suggests that the attenuation bias may be quantitatively important, reaching about 60 percent for a standard deviation equal to the observed variation of payables financing across countries and industries. B.2 Impact of the use of a common measure of Input-Output distances The main body of the paper presented some evidence indicating that input-output linkages among industries are mainly technologically determined and, therefore, that extrapolation of the distance measures built using U.S. data to other countries is unlikely to result in significant biases. Here I formally examine this statement by performing a series of Montecarlo simulations to explore the potential biases arising from making this assumption in a setting analogous to the one used in the paper. The simulations are performed as follows. I first compute a fictitious true D matrix for each country in the sample as the sum of the D matrix computed for the U.S. plus some random noise. Notice that the D matrices are such that their columns add to one, and typically have large values in the diagonal (mean value is 0.78) because most the final demand for a product is typically embodied in final purchases of the product itself.21 I, therefore, introduce noise in a manner that is roughly consistent with these characteristics by first re-scaling the matrix such that its rows add to 100, adding to each element a small error derived from a uniform distribution with positive domain between zero and a varying number set between 0.5 and 2, and then re-normalizing the rows. In this way I preserve the heavy-diagonal property of the simulated matrices but at the same time introduce significant noise.22 After generating the simulated true matrices, I consider country 1 as the benchmark country and pick the matrix simulated for it as the baseline measure, which I impute to all other countries. With the true distances, the actual measures of trade credit use, and a value of the parameter of 2 I simulate the correlations according to equation (14), and with the imputed common distance compute the credit-chain distances between industries in di erent countries as described in the paper. I then run a regression between the simulated correlations and credit-chain distances in the fictitious sample. I do this 100 times for each value of the upper limit of the uniform distribution between 0.5 and 2. The results of the simulation are summarized in the Panel B of Figure A1. The figure is in the same scale as the one in Panel A to facilitate the comparison of the magnitude of the di erent biases, and shows that the assumption of a common matrix also biases the estimated coe cient downwards, but the size of the bias is small, never exceeding 10 percent. The bias slightly increases with the size of the error, but this is unnoticeable using this scale. 21This is a standard characteristic of input-output matrices that should hold across countries. 22For instance, when drawing noise from a uniform distribution between zero and one, the expected value of the draw is 0.5. Considering that there are 27 other sectors, the expected sum of the noise across these sectors is about 14. So, even in the extreme case in which the share of the diagonal was 1, this share is expected to fall to 0.88 after introducing the noise. Proceeding in this way likely introduces more noise than what would be observed in real data. 30 Overall, the Montecarlo experiments confirm the intuition that the bias introduced by the as- sumption of a common distance across countries goes in the right direction (attenuation) and is probably small. References Baxter, Marianne and Robert G. King, "Productive externalities and business cycles," Tech- nical Report 1991. Biais, Bruno and Christian Gollier, "Trade Credit and Credit Rationing," Review of Financial Studies, 1997, 10 (4), 903­37. Blanchard, Olivier and Michael Kremer, "Disorganization," The Quarterly Journal of Eco- nomics, November 1997, 112 (4), 1091­1126. Boissay, Frederic, "Credit chains and the propagation of financial distress," Working Paper Series 573, European Central Bank January 2006. Bradley, D. and M. Rubach, "Trade Credit and Small Business: A cause of Business Failure?," Technical Report August 2002. Mimeo, University of Central Arkansas. Burkart, Mike and Tore Ellingsen, "In-Kind Finance: A Theory of Trade Credit," American Economic Review, June 2004, 94 (3), 569­590. Cardoso-Lecourtois, Miguel, "Chain Reactions, Trade Credit and the Business Cycle," Econo- metric Society 2004 North American Summer Meetings 331, Econometric Society August 2004. Chittenden, F. and R. Bragg, "Trade Credit, Cash Flows, and SMEs in the U.K., Germany, and France," International Small Business Journal, 1997, 16 (1), 22­35. Clark, Todd E., "The Responses of Prices at Di erent Stages of Production to Monetary Policy Shocks," Review of Economics and Statistics, August 1999, 81 (3), 420 ­33. Conley, Timothy and Bill Dupor, "A Spatial Analysis of Sectoral Complementarities," Journal of Political Economy, April 2003, 111 (2), 311­352. Coricelli, Fabrizio and Igor Masten, "Growth and Volatility in Transition Coun- tries: The role of Credit," 2004. Mimeo, International Monetary Fund. Available at https://www.imf.org/external/np/res/seminars/2004/calvo/pdf/corice.pdf. Cowan, Kevin and Alejandro Neut, "Intermediate Goods, Institutions and Output per Worker," Department of Economics, Massachusetts Institute of Technology 2002. Demirguc-Kunt, Asli and Vojislav Maksimovic, "Firms as financial intermediaries - evidence from trade credit data," Policy Research Working Paper Series 2696, The World Bank October 2001. 31 Fisman, Raymond and Inessa Love, "Trade Credit, Financial Intermediary Development, and Industry Growth," Journal of Finance, 02 2003, 58 (1), 353­374. Horvath, Michael, "Cyclicality and sectoral linkages: Aggregate fluctuations from independent sectoral shocks," Review of Economic Dynamics, October 1998, 1 (4), 781­808. , "Sectoral shocks and aggregate fluctuations," Journal of Monetary Economics, February 2000, 45 (1), 69­106. International Monetary Fund, "International Financial Statistics," CD-ROM 2005. Kiyotaki, Nobuhiro and John Moore, "Credit Chains," Technical Report 1997. Kremer, Michael, "The O-Ring Theory of Economic Development," Quarterly Journal of Eco- nomics, August 1993, 108 (3), 551­175. Liu, Zheng Huang Kevin X. D.and, "Production Chains and General Equilibrium Aggregate Dynamics," Journal of Monetary Economics, October 2001, 48 (2), 437 ­62. Long, John and Charles Plosser, "Real Business Cycles," Journal of Political Economy, February 1983, 91 (1), 39­69. Love, Inessa, Lorenzo A. Preve, and Virginia Sarria-Allende, "Trade credit and bank credit: Evidence from recent financial crises," Journal of Financial Economics, February 2007, 83 (2), 453­469. Olarreaga, Marcelo and Alessandro Nicita, "Trade and Production, 1976-99," 2004. The World Bank, Policy Research Working Paper Series: 2701. Otto, Glenn, Graham Voss, and Luke Willard, "Understanding OECD Output Correlations," RBA Research Discussion Papers rdp2001-05, Reserve Bank of Australia September 2001. Petersen, Mitchell A and Raghuram G Rajan, "Trade Credit: Theories and Evidence," Review of Financial Studies, 1997, 10 (3), 661­91. Raddatz, Claudio, "Liquidity Needs and Vulnerability to Financial Underdevelopment," Journal of Financial Economics, 2006, 80, 677­722. Rajan, Raghuram and Luigi Zingales, "Financial Dependence and Growth," American Eco- nomic Review, 1998, 88, 559­586. Shea, John, "Two Essays in Empirical Macroeconomics." PhD dissertation, Massachusetts Institute of Technology 1990. , "Complementarities and comovements," Journal of Money, Credit, and Banking, May 2002, 34 (2), 412­433. 32 United Nations Industrial Development Organization, "Industrial Statistics Database," CD- ROM 2005. Yamada, Tetsuo, "Relevance and Applicability of the UNIDO Industrial Statistics Database for Research Purposes," 2005. Manuscript, Department of Economic and Social A airs Statistics Division, United Nations. 33 Table 1. Trade credit use in sample countries Country Payables Short Term Debt Quality Trade Financing to Payables Credit Data (1) (2) (3) Australia 0.13 0.24 1 Austria 0.11 1.60 1 Belgium 0.15 0.91 1 Canada 0.19 0.25 1 Chile 0.11 1.86 1 China 0.20 2.73 1 Colombia 0.07 2.45 1 Denmark 0.10 1.00 1 Egypt 0.18 3.04 2 Spain 0.20 0.79 1 Finland 0.10 1.20 1 France 0.17 0.55 1 United Kingdom 0.16 0.42 1 Greece 0.14 1.78 1 Hong Kong, China 0.15 1.21 1 Hungary 0.13 0.87 1 Indonesia 0.12 3.26 1 India 0.17 0.89 1 Ireland 0.15 0.56 1 Iceland 0.15 1.85 2 Israel 0.18 0.58 1 Italy 0.33 0.81 1 Jordan 0.13 1.26 2 Japan 0.24 0.76 1 Korea 0.13 2.15 1 Sri Lanka 0.08 3.09 2 Morocco 0.21 0.27 2 Mexico 0.14 1.12 1 Malaysia 0.12 1.94 1 Netherlands 0.12 0.61 1 Norway 0.10 0.31 1 Peru 0.11 2.36 1 Philippines 0.18 1.29 1 Poland 0.14 0.78 1 Portugal 0.12 1.53 1 Singapore 0.15 0.82 1 Sweden 0.11 0.31 1 Turkey 0.13 1.10 1 Taiwan, China 0.12 1.98 1 United States 0.13 0.40 1 Venezuela, RB 0.20 1.25 2 South Africa 0.19 0.21 1 Zimbabwe 0.08 0.92 2 Mean 0.15 1.24 -- Median 0.14 1.00 -- St. Dev. 0.05 0.83 -- Percentile 25 0.12 0.59 -- Percentile 75 0.18 1.82 -- Payables Financing is the ratio of accounts payables to the cost of goods sold; Short term debt to Payables is the ratio of short term debt to accounts payables. The figures reported for each country correspond to the median level of each ratio across all manufacturing firms in the country, except for the countries where quality equals 2 (column (3)), where it corresponds to the ratio across all firms. For each firm, each measure is the median across the years with data during the period 1980 to 2005. Only firms with more than 5 years of data are included in the computation of the country median and only countries with more than 10 of these firms are included in the sample. 34 Table 2. Relative use of Trade credit across U.S. manufacturing industries Relative Payables Relative Short Term ISIC code Industry Financing Debt to Payables (1) (2) 311 Food products 0.82 1.01 313 Beverages 1.37 0.55 314 Tobacco 0.92 0.92 321 Textiles 0.93 0.94 322 Wearing apparel, except footwear 0.93 1.92 323 Leather products 0.69 2.10 324 Footwear, except rubber or plastic 0.73 1.27 331 Wood products, except furniture 0.62 1.16 332 Furniture, except metal 0.77 0.75 341 Paper and products 0.86 0.57 342 Printing and publishing 0.95 0.87 351 Industrial chemicals 1.23 0.56 352 Other chemicals 1.51 0.58 353 Petroleum refineries 1.27 0.43 354 Misc. petroleum and coal products 1.02 0.97 355 Rubber products 0.87 0.85 356 Plastic products 1.01 0.72 361 Pottery, china, earthenware 0.58 0.88 362 Glass and products 0.94 1.35 369 Other non-metallic mineral products 1.06 0.67 371 Iron and steel 0.94 0.70 372 Non-ferrous metals 0.95 1.18 381 Fabricated metal products 0.97 1.15 382 Machinery, except electrical 1.32 0.80 383 Machinery, electric 1.27 1.03 384 Transport equipment 0.90 0.78 385 Professional & scientific equipment 1.36 0.90 390 Other manufactured products 1.20 2.00 Median 0.95 0.89 Percentile 25 0.87 0.71 Percentile 75 1.21 1.15 Correlations Relative Inv. Payables Turnover 1 Relative Short Term Debt to Payables -0.32 1 Column (1) reports the ratio of accounts payable to cost of goods sold in a given manufacturing industry in the U.S. to the overall U.S. mean in manufactures. Similar ratios are reported for short-term debt to payables (column (2)). The values of each measure for a given industry correspond to the median across all U.S. firms in that industry included in Compustat during 1980-1989. The value for the U.S. manufacturing sector as a whole corresponds to the median across industries. For a given firm, the ratios correspond to their median across the years in which the firm reported data. 35 Table 3. Average correlations and number of observations in sample countries Country Average corr. of No. ind. pairs with Average corr. of No. ind. Pairs Average GDP Per VA growth VA growth IIP growth with IIP growth capita 1980-2000 (1) (2) (3) (3) (4) Australia 0.22 190 0.35 190 20,483 Austria 0.13 378 0.23 378 19,220 Belgium 0.07 91 0.15 231 19,290 Canada 0.17 378 0.46 378 21,974 Chile 0.26 378 0.36 378 6,861 China 0.48 231 -- -- 2,133 Colombia 0.15 378 0.29 378 4,897 Denmark 0.29 378 0.23 378 21,853 Egypt 0.03 378 0.14 378 3,226 Spain 0.23 351 0.25 351 14,062 Finland 0.13 378 0.30 378 18,726 France 0.77 325 0.23 325 19,028 United Kingdom 0.32 378 0.54 378 17,888 Greece 0.20 210 0.00 210 12,185 Hong Kong, China -- -- 0.49 322 20,416 Hungary 0.69 325 0.57 351 9,093 Indonesia 0.20 276 0.20 276 2,889 India 0.14 378 0.11 378 1,705 Ireland 0.11 325 0.27 325 14,819 Iceland 0.06 231 -- -- 20,625 Israel 0.11 253 0.36 231 14,005 Italy 0.40 325 0.22 377 18,505 Jordan 0.05 325 0.09 120 3,962 Japan 0.31 378 0.39 378 20,773 Korea 0.38 378 0.46 378 10,010 Sri Lanka 0.32 325 -- -- 2,564 Morocco 0.03 66 0.07 136 3,418 Mexico 0.10 325 0.65 325 7,587 Malaysia 0.16 378 0.19 325 7,001 Netherlands 0.28 276 0.16 275 19,206 Norway 0.14 325 0.17 377 21,372 Peru 0.31 378 0.60 378 4,404 Philippines 0.29 378 0.12 378 3,075 Poland 0.53 45 0.65 215 7,001 Portugal 0.17 351 0.17 378 11,815 Singapore 0.24 276 0.22 276 16,627 Sweden 0.49 378 0.21 378 19,878 Turkey 0.15 378 0.23 378 5,630 Taiwan, China 0.29 378 0.26 378 10,732 United States 0.41 378 0.43 378 26,235 Venezuela, RB 0.31 378 -- -- 7,090 South Africa 0.26 253 0.39 253 7,645 Zimbabwe 0.15 300 0.29 253 2,721 Mean 0.25 314 0.30 322 12,154 Median 0.22 338 0.25 377 11,815 St. Dev. 0.17 88 0.16 75 7,347 Percentile 25 0.14 276 0.18 276 5,264 Percentile 75 0.31 378 0.39 378 19,213 Column (1) reports the average correlation of value added growth among industry pairs (excluding the correlation between an industry and itself). Column (2) displays the number of non-repeated industry pairs in which we have data on the correlation of value added growth per country (a given industry pair is counted only once). Columns (3) and (4) present similar information for the correlation of the growth of the industrial production index. Column (5) shows the average real GDP per capita of each country during the period 1980-2000. 36 Table 4. Analysis of Variance of Correlation Measures Panel A: Raw data Panel B: Partialling out country and industry pairs FE Correlation of Correlation of Correlation of growth rate of Correlation of growth rate of growth rate of index of growth rate of index of real value added industrial real value added industrial produciton produciton (1) (2) (3) (4) Overall 0.33 0.33 0.27 0.25 Between countries 0.16 0.17 -- -- Within countries 0.29 0.28 0.27 0.25 Within industry-pair across countries 0.29 0.28 0.27 0.27 Within country across industry pairs 0.29 0.28 0.27 0.27 The table reports various summary statistics of the variation of the correlation of the growth or real value added (columns (1) and (3)) and index of industrial production (columns (2) and (4)) across industry pairs along several dimensions. Panel A presents information for the raw measures of correlation, and Panel B reports indicators for the residual correlations after partialling-out country and industry pair fixed effects. The row labeled Overall reports the overall standard deviation of each variable across countries and industry pairs. Between countries and Within countries report the standard deviation of the correlation measures along these two dimensions. The last two rows of each column report the average within industry-pair across-countries, and within-country across industry- pairs standard deviations of the correlation measures. The within-industry across countries variation is computed taking, for each industry-pair, the standard deviation of its measured correlations across the countries in the sample, and then calculating the average of these standard deviations across all industry-pairs. The within-country, across industry pairs variation is computed similarly by switching the order of aggregation 37 Table 5. Analysis of Variance of Measures of Direct Requirements Direct Requirements Overall 0.026 Between countries 0.004 Within countries 0.026 Within industry-pair across countries 0.007 Within country across industry pairs 0.024 The table reports various summary statistics of the variation of the components of the matrices of direct requirements across industry pairs and countries along several dimensions. The row labeled Overall reports the overall standard deviation of the measures across countries and industry pairs. Between countries and Within countries report their standard deviation along these two dimensions. The last two rows report the average within industry-pair across-countries, and within-country across industry- pairs standard deviations of the direct requirement measures. The within-industry across countries variation is computed taking, for each industry-pair, the standard deviation of its direct requirement measures across the countries in the sample, and then calculating the average of these standard deviations across all industry-pairs. The within-country, across industry pairs variation is computed similarly by switching the order of aggregation 38 Table 6. Industry pairs with the smallest Credit-Chain Distances ISIC code ISIC code Credit chain industry 1 Industry 1 industry 2 Industry 2 Distance Ranking Wearing apparel, except 321 Textiles 322 footwear 90.67 1 381 Fabricated metal products 384 Transport equipment 63.17 2 351 Industrial chemicals 352 Other chemicals 63.06 3 356 Plastic products 384 Transport equipment 61.31 4 351 Industrial chemicals 384 Transport equipment 60.15 5 362 Glass and products 384 Transport equipment 59.47 6 356 Plastic products 372 Non-ferrous metals 54.15 7 351 Industrial chemicals 372 Non-ferrous metals 53.30 8 Footwear, except rubber or 323 Leather products 324 plastic 49.27 9 362 Glass and products 372 Non-ferrous metals 49.11 10 356 Plastic products 371 Iron and steel 48.90 11 351 Industrial chemicals 371 Iron and steel 47.18 12 355 Rubber products 384 Transport equipment 46.95 13 351 Industrial chemicals 356 Plastic products 44.69 14 341 Paper and products 342 Printing and publishing 43.77 15 362 Glass and products 371 Iron and steel 42.99 16 356 Plastic products 383 Machinery, electric 40.30 17 355 Rubber products 372 Non-ferrous metals 39.77 18 362 Glass and products 383 Machinery, electric 39.28 19 371 Iron and steel 384 Transport equipment 39.24 20 The table shows the input-output distances of the twenty industry pairs with the smallest (generic) credit-chain distances computed using the 1992 U.S. input-output matrices. The first four columns of the table describe the names and ISIC codes of the industries that comprise each pair. The column labeled Credit-Output Distance displays the estimated value of the distance. This measure is such that a higher value implies a smaller distance. The last column indicates the ranking in terms of closeness of each industry pair across the whole set of 378 possible pairs. 39 Table 7. Credit chains and sectoral correlations Panel A Panel B Dependent variable is the correlation of real Dependent variable is the correlation of value added growth growth in industrial production index (1) (2) (3) (1) (2) (3) Credit Chain Distance (Payables Financing)() 1.978*** 2.058*** 1.927*** 1.455*** 1.504*** 1.481*** (0.327) (0.327) (0.327) (0.357) (0.372) (0.371) (log) Number of establishments industry 1 -- 0.024*** 0.019*** -- 0.037*** 0.036*** (0.004) (0.004) (0.004) (0.004) (log) Number of establishments industry 2 -- 0.024*** 0.022*** -- 0.011** 0.012** (0.004) (0.004) (0.005) (0.005) Share of total manufacturing VA industry 1 -- -0.239** -0.305*** -- -0.243** -0.301** (0.098) (0.098) (0.114) (0.118) Share of total manufacturing VA industry 2 -- -0.669*** -0.672*** -- -0.149 -0.089 (0.104) (0.105) (0.112) (0.115) Average growth industry 1 -- -- -0.131* -- -- -0.426*** (0.073) (0.109) Average growth industry 2 -- -- -0.004 -- -- -0.713*** (0.074) (0.093) Standard deviation growth industry 1 -- -- -0.205*** -- -- -0.337*** (0.027) (0.074) Standard deviation growth industry 2 -- -- -0.102*** -- -- -0.237*** (0.022) (0.057) Observations 13182 12683 12683 12548 11146 11146 R-squared 0.34 0.35 0.35 0.41 0.40 0.41 Industry pair FE Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Credit chain distance (Payables Financing) is the measures of the intensity of use of trade credit in the chain linking two industries based on the Payable Financing ratio. (log) Number of establishments industry 1 (2) is the log of the average number of firms in the first (second) industry in the corresponding industry pair. Share of total manufacturing VA industry 1 (2) is the average share of the first (second) industry pair on total manufacturing value added. Average growth industry (1) is the average growth of real value added (Panel A) and index of industrial production (Panel B) in the first (second) industry in the pair. Standard deviation growth industry 1 (2)) is the standard deviation of the growth of real value added (Panel A) and index of industrial production (Panel B) in the first (second) industry in the pair. All averages mentioned above are computed over the period 1980-2000. All regressions include country and industry pair fixed effects. Standard errors are robust to heteroskedasticity. * = Significant at 10 percent level, ** = Significant at 5 percent level, *** = Significant at 1 percent level. 40 Table 8. Credit chains, relative use of formal financing and sectoral correlations Panel A: Dependent variable is the correlation of real value added growth (1) (2) (3) (4) (5) (6) Credit Chain Distance (Short term debt to payables) -0.097*** -0.092*** -0.099*** -0.072*** -0.065*** -0.075*** (0.021) (0.021) (0.023) (0.022) (0.022) (0.024) Credit Chain Distance (Payables Financing) -- -- -- 1.676*** 1.798*** 1.627*** (0.344) (0.343) (0.347) (log) Number of establishments industry 1 -- 0.023*** 0.018*** -- 0.023*** 0.018*** (0.004) (0.004) (0.004) (0.004) (log) Number of establishments industry 2 -- 0.024*** 0.021*** -- 0.024*** 0.021*** (0.004) (0.004) (0.004) (0.004) Share of total manufacturing VA industry 1 -- -0.224** -0.295*** -- -0.245** -0.314*** (0.098) (0.098) (0.098) (0.098) Share of total manufacturing VA industry 2 -- -0.667*** -0.675*** -- -0.684*** -0.690*** (0.105) (0.105) (0.104) (0.105) Average growth industry 1 -- -- -0.141* -- -- -0.130* (0.073) (0.073) Average growth industry 2 -- -- 0.001 -- -- -0.001 (0.073) (0.073) Standard deviation growth industry 1 -- -- -0.212*** -- -- -0.208*** (0.027) (0.027) Standard deviation growth industry 2 -- -- -0.103*** -- -- -0.102*** (0.022) (0.022) Observations 13182 12683 12683 13182 12683 12683 R-squared 0.34 0.35 0.35 0.34 0.35 0.35 Panel B: Dependent variable is the correlation of growth in industrial production index (1) (2) (3) (4) (5) (6) Credit Chain Distance (Short term debt to payables) -0.066*** -0.071*** -0.070*** -0.045* -0.050* -0.050* (0.026) (0.027) (0.026) (0.027) (0.029) (0.028) Credit Chain Distance (Payables Financing) -- -- -- 1.275*** 1.316*** 1.293*** (0.382) (0.395) (0.393) (log) Number of establishments industry 1 -- 0.036*** 0.036*** -- 0.037*** 0.036*** (0.004) (0.004) (0.004) (0.004) (log) Number of establishments industry 2 -- 0.010** 0.011** -- 0.011** 0.012** (0.005) (0.005) (0.005) (0.005) Share of total manufacturing VA industry 1 -- -0.225** -0.282** -- -0.245** -0.303*** (0.114) (0.117) (0.114) (0.118) Share of total manufacturing VA industry 2 -- -0.147 -0.087 -- -0.163 -0.101 (0.112) (0.115) (0.112) (0.115) Average growth industry 1 -- -- -0.443*** -- -- -0.429*** (0.109) (0.109) Average growth industry 2 -- -- -0.711*** -- -- -0.714*** (0.093) (0.093) Standard deviation growth industry 1 -- -- -0.338*** -- -- -0.339*** (0.074) (0.074) Standard deviation growth industry 2 -- -- -0.234*** -- -- -0.233*** (0.057) (0.057) Observations 12548 11146 11146 12548 11146 11146 R-squared 0.41 0.40 0.41 0.41 0.40 0.41 Credit chain distance (Payables Financing) is the measures of the intensity of use of trade credit in the chain linking two industries based on the Payable Financing ratio. Credit chain distance (Short term debt to payables) is the intensity of use of trade credit in the chain linking two industries based on the Short-term debt to payables ratio. (log) Number of establishments industry 1 (2) is the log of the average number of firms in the first (second) industry in the corresponding industry pair. Share of total manufacturing VA industry 1 (2) is the average share of the first (second) industry pair on total manufacturing value added. Average growth industry (1) is the average growth of real value added (Panel A) and index of industrial production (Panel B) in the first (second) industry in the pair. Standard deviation growth industry 1 (2)) is the standard deviation of the growth of real value added (Panel A) and index of industrial production (Panel B) in the first (second) industry in the pair. All averages mentioned above are computed over the period 1980-2000. All regressions include country and industry pair fixed effects. Standard errors are robust to heteroskedasticity. * = Significant at 10 percent level, ** = Significant at 5 percent level, *** = Significant at 1 percent level. 41 Table 9. Variations in the sample High Dropping Dropping Quality extreme Extreme Robust sample Trade Credit distances Estimation (1) (2) (3) (4) Credit Chain Distance ( Payables Financing)() 2.323*** 1.846*** 2.952*** 1.895*** (0.340) (0.660) (0.561) (0.373) (log) Number of establishments industry 1 0.026*** 0.020*** 0.025*** 0.024*** (0.004) (0.005) (0.004) (0.004) (log) Number of establishments industry 2 0.020*** 0.019*** 0.024*** 0.027*** (0.005) (0.005) (0.005) (0.004) Share of total manufacturing VA industry 1 -0.330*** -0.442*** -0.313*** -0.255*** (0.123) (0.111) (0.103) (0.095) Share of total manufacturing VA industry 2 -0.696*** -0.772*** -0.699*** -0.720*** (0.116) (0.113) (0.108) (0.102) Constant -0.074 -0.026 -0.091** -0.296*** (0.047) (0.046) (0.044) (0.074) Observations 10680 9804 11431 12683 R-squared 0.36 0.37 0.35 0.35 Industry pair FE Yes Yes Yes Yes Country FE Yes Yes Yes Yes Credit chain distance is the measures of the intensity of use of trade credit in the chain linking two industries based on the Payable Financing ratio. (log) Number of establishments industry 1 (2) is the log of the average number of firms in the first (second) industry in the corresponding industry pair. Share of total manufacturing VA industry 1 (2) is the average share of the first (second) industry pair on total manufacturing value added. The dependent variable is the correlation of growth rates of real value added between industry-pairs in different countries. The results reported correspond to those obtained in our high quality sample of 39 countries (column (1)), dropping the countries where the corresponding measures of trade credit were below the 5th and above the 95th percentile levels observed in our main sample (column (2)), and dropping those industries where the corresponding measures of the generic credit chain distances were below the 5th and above the 95th percentile levels observed in our set of industry pairs (column (3)). The coefficients reported in Column (4) were obtained in a robust regression. All averages mentioned above were computed over the period 1980-2000. All regressions include country and industry pair fixed effects. Standard errors are robust to heteroskedasticity. * = Significant at 10 percent level, ** = Significant at 5 percent level, *** = Significant at 1 percent level. 42 Table 10. Different measures of correlations and distances Using Using HP Transformed Robust Country Using UK filter Correlation correlation level input-output Information data (1) (2) (3) (4) (5) Credit Chain Distance ( Payables Financing)( ) 1.745*** 4.592*** 1.722*** -- -- (0.368) (0.805) (0.393) Credit Chain Distance Country Basis () -- -- -- 1.871*** -- (0.306) Credit Chain Distance (using UK data)() -- -- -- -- 1.618*** (0.293) (log) Number of establishments industry 1 0.023*** 0.057*** 0.023*** 0.024*** 0.024*** (0.005) (0.009) (0.004) (0.004) (0.004) (log) Number of establishments industry 2 0.001 0.049*** 0.030*** 0.025*** 0.025*** (0.006) (0.011) (0.005) (0.004) (0.004) Share of total manufacturing VA industry 1 -0.230* -0.627*** -0.327*** -0.230** -0.232** (0.122) (0.226) (0.111) (0.098) (0.098) Share of total manufacturing VA industry 2 -0.074 -1.403*** -0.945*** -0.666*** -0.671*** (0.126) (0.243) (0.117) (0.104) (0.104) Constant 0.455*** -0.190** -0.079* -0.085** -0.088** (0.061) (0.095) (0.044) (0.040) (0.040) Observations 10194 12651 12683 12683 12683 R-squared 0.32 0.38 0.33 0.35 0.35 Industry pair FE Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Credit chain distance is the measures of the intensity of use of trade credit in the chain linking two industries based on the Payables Financing ratio. (log) Number of establishments industry 1 (2) is the log of the average number of firms in the first (second) industry in the corresponding industry pair. Share of total manufacturing VA industry 1 (2) is the average share of the first (second) industry pair on total manufacturing value added. In Columns (1) to (3) the dependent variables are the correlation of the series of real value added detrended using the Hodrik-Prescott filter, the Otto et al. (2001) transformation of the correlation of the growth rate of real value, and a robust measure of the correlation of real value added growth between the industries in a pair, respectively. Columns (4) and (5) present results obtained using a measure of credit chain distance that exploits the existing information on industry level use of trade credit across countries and a measure based on the UK input-output matrices, respectively. All averages mentioned above were computed over the period 1980-2000. All regressions include country and industry pair fixed effects. Standard errors are robust to heteroskedasticity. * = Significant at 10 percent level, ** = Significant at 5 percent level, *** = Significant at 1 percent level. 43 Table 11. Other dimensions of similarity across industry pairs (1) (2) (3) (4) (5) (6) (7) (8) Credit Chain Distance (Payables Financing) () 2.177*** 2.037*** 2.036*** 1.782*** 2.020*** 2.057*** 2.063*** 1.900*** (0.345) (0.337) (0.328) (0.372) (0.326) (0.327) (0.328) (0.338) BUY distance X Payables Financing -0.723 -- -- -- -- -- -- -- (0.484) -- -- -- -- -- -- -- SELL distance X Payables Financing 0.563** -- -- -- -- -- -- -- (0.229) -- -- -- -- -- -- -- Dist. external finance X Financial Development -- -0.000 -- -- -- -- -- -- -- (0.014) -- -- -- -- -- -- Dist. liquidity needs X Financial Development -- -0.437*** -- -- -- -- -- -- -- (0.129) -- -- -- -- -- -- Dist. capital per employee x Capital per Worker -- 0.001 -- -- -- -- -- -- -- (0.023) -- -- -- -- -- -- Dist. Gini intermediary shares X Overall Volatility -- -- 21.047* -- -- -- -- -- -- -- (10.763) -- -- -- -- -- Cost distance X Payables Financing -- -- -- 1.950 -- -- -- -- -- -- -- (1.244) -- -- -- -- Cost Based Credit Chain (Payables Financing) -- -- -- -- 2.854*** -- -- -- -- -- -- -- (0.828) -- -- -- Cost Based Credit Chain (Receivables Financing) -- -- -- -- -- 0.139 -- -- -- -- -- -- -- (0.864) -- -- Correlation from differences in shocks volatilities -- -- -- -- -- -- 0.006 -- -- -- -- -- -- -- (0.013) -- Generic Credit Chain Dist. X Overall Volatility -- -- -- -- -- -- -- -2.408* -- -- -- -- -- -- -- (1.447) Observations 12683 10241 12683 12683 12683 12683 12683 12683 R-squared 0.35 0.34 0.35 0.35 0.35 0.35 0.35 0.35 Number of establishments both industries Yes Yes Yes Yes Yes Yes Yes Yes Share of total manufacturing VA both industries Yes Yes Yes Yes Yes Yes Yes Yes Industry pair FE Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes The dependent variable is the correlation of growth rates of real value added between industry-pairs in different countries. Credit chain distance is the measures of the intensity of use of trade credit in the chain linking two industries based on Payable Financing. BUY (SELL) distance X Payables Financing are the interaction between Conley and Dupor (2003)'s measure of similarity between two industries in terms of suppliers (customers) and a country's median level of payables to cost of goods sold. Dist. external finance X Financial Development is the interaction of the absolute value of an industry pair's difference in external financial requirements (Rajan and Zingales, 1998) and a country's level of Financial Development measured as the (log) average ratio of Private Credit to GDP. Dist. liquidity needs X Financial Development, Dist. capital per employee X Capital per Worker, and Dist. Gini intermediary shares X Overall Volatility are computed analogously. Cost Distance x Payables Financing is the interaction between Shea (2002)'s measure of an industry pair's strength of forward linkages and a country's payables financing. Cost based Credit Chain (Payables financing) and Cost based Credit Chain (Receivables financing) are the measures of credit chains built using Shea (2002)'s COST matrix and payables and receivables financing as measures of trade credit use, respectively. Correlation from differences in shocks volatility is the contribution of differences in the volatility of shocks to various industries to an industry pair correlation. Generic Credit Chain Dist. X Overall Volatility is the interaction of an industry pair's generic credit chain distance and a country's standard deviation of real GDP per capita growth. All averages and standard deviations mentioned above are computed over the period 1980-2000. All regressions include country fixed effects and control for the average number of firms and share of total manufacturing value added of each industry in the pair. Standard errors are robust and clustered at the industry-pair level. * = Significant at 10 percent level, ** = Significant at 5 percent level, *** = Significant at 1 percent level. 44 Table 12. Alternative Explanations (1) (2) (3) (4) (5) (6) (7) Credit Chain Distance (Payables Financing) () 1.924*** 1.886*** 2.272*** 2.357*** 1.867*** 2.034*** -- (0.332) (0.341) (0.330) (0.334) (0.331) (0.328) -- Generic Credit Chain Distance X (log) GDP per capita 0.070*** 0.064*** -- -- -- -- -- (0.023) (0.024) -- -- -- -- -- Generic Credit Chain Distance X Growth Volatility -- -0.732 -- -- -- -- -- -- (1.544) -- -- -- -- -- Credit Chain Distance X Financial Development -- -- 0.810*** -- -- -- -- -- -- (0.210) -- -- -- -- Credit Chain Distance X Short term debt to Payables -- -- -- -0.603*** -- -- -- -- -- -- (0.154) -- -- -- Generic Credit Chain Distance X Trade Openness -- -- -- -- -0.086*** -- -- -- -- -- -- (0.030) -- -- Direct Credit Chain Distance (Payables Financing) -- -- -- -- -- 7.427** -- -- -- -- -- -- (2.988) -- Credit Chain Distance (Common use) -- -- -- -- -- -- 2.294*** -- -- -- -- -- -- (0.345) Credit Chain Distance (Differential use) -- -- -- -- -- -- -2.697 -- -- -- -- -- -- (2.465) Observations 12683 12683 12683 12683 12683 12683 12683 R-squared 0.35 0.35 0.35 0.35 0.35 0.35 0.35 Number of establishments both industries Yes Yes Yes Yes Yes Yes Yes Share of total manufacturing VA both industries Yes Yes Yes Yes Yes Yes Yes Industry pair FE Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Credit chain distance (Payables Financing) is the measures of the intensity of use of trade credit in the chain linking two industries based on the ratio of payables to cost of goods sold. Generic Credit Chain Distance x (log) GDP per capita, Generic Credit Chain Distance x Growth Volatility, and Generic Credit Chain Distance X Trade Openness are the interactions between an industry pair's generic credit chain distance and a country's (log) GDP per capita, standard deviation of the growth of real GDP per capita, and the log average ratio of total exports to GDP. Credit Chain Distance X Financial Development and Credit Chain Distance X Short-term debt to payables are the interactions between these two variables. Direct Credit Chain Distance is the credit chain distance computed considering only the direct linkages among industries. Credit Chain Distance (Common Use) is the distance computed using the average payable financing of all industries, and Credit Chain Distance (differential use) is the distance computed using the industries deviations from that average only. These two distances add to the baseline measure of credit chain distance. All averages mentioned above are computed over the period 1980-2000. All regressions include country fixed effects and control for the average number of firms and share of total manufacturing value added of each industry in the pair. Standard errors are robust to heteroskedasticity. * = Significant at 10 percent level, ** = Significant at 5 percent level, *** = Significant at 1 percent level. 45 Figure 1. Input-Output Distances US and UK .8 atad S .6 U ec ants di .4 utptuo t-u p In .2 0 0 .2 .4 .6 .8 1 Input-output distance UK data The figure plots the input-output distances of the 378 industry pairs formed by the combination of our 28 3-digit ISIC manufacturing industries computed using data from UK (x-axis) and US (y-axis) input- output matrices. 46 Figure 2. Distribution of Credit-Chain Distances across industries .4 .3 noitca .2 Fr .1 0 0 .2 .4 .6 .8 Credit Chain Distance The figure shows the histogram of generic credit-chain distances across (non-repeated) industry pairs excluding the pairs corresponding to an industry and itself. The y-axis displays the fraction of the population in each of the bins depicted in the x-axis. 47 Figure 3. Differential effect of trade credit use in sectors with small and large distances Panel A. Payables financing and the correlation Panel B. Payables financing and the correlation between Iron and Steel and Transport Equipment between Tobacco and Footwear industries (large industries (small distance) distance) SWE 1 .5 HUN FRA ISR CAN CHN JPN ITA KOR FRA ZAF AUTUSA LKA SGP VEN .5 ESP 0 PRT PHL PER COL MYSMEX LKA noi PER n SWEHUN TUR GBR IND at FIN iota KOR GRCGBR TWN JOR elrroc SGP EGY IRL rreloc IDN ITA IRL JPN 0 CHL AUT PRT IND 5 EGY PHL -. ISL JOR ZWE TWNAUS DNK CAN ZWE CHL USA CHN DNK FIN NOR VEN ESP TUR 5-. COL MYS NLD -1 -.1 0 .1 .2 -.1 0 .1 .2 Payables Financing Payables Financing coef = .48734521, (robust) se = .8235791, t = .59 coef = 2.9879355, (robust) se = .92204959, t = 3.24 Panel C. Payables financing and the correlation Panel C. Payables financing and the correlation between 20 industry pairs with smaller distance between 20 industry pairs with larger distance 1 1 IRL HUN HUN HUN HUN FRA FRA FRA FRA FRA FRA FRA CHN SWEHUN HUN FRA HUN JPN PER PER HUN SWEHUN SWEHUN FRA FRA CHN CHN LKA FRA FRA DNKSWEHUN BEL FRA JPN VEN SWEPRT SWEUSA NLDHUN NLDKOR FRA FRA FRA FRA PERPRT IDNHUNGRCISL IDNKOR HUN BELGBRFRA HUN .5 LKA DNK SWE HUN SWEHUN HUN USA HUN HUN SGP .5 FRA FRA FRA SWEPRT HUN FRA ITA USA FRA FRA VEN PHL PHLVEN JPN ITA HUN FRAPHL FRA FRAZAF VEN JPN JPN ZWE CHN JPN ZWE LKA FINPERNLDKORBEL LKA NOR SWEHUN PERNLD PERIDNJORPOL FRAPHLVEN PHL SWEHUNMEX SWETWN MEX JOR BEL PHL PHL ITA ITA COL ZWE SWE KORPOL SWE HUN ISR CHLHUN JPN ITA ITA ITA ITA ZWE LKA JPN LKA SWEUSAMEX HUNMEX ITA JPN JPN ZWEDNKCHLHUNMEXSGP SWE SWEUSA CHLHUNMEXGBRFRAISRCAN USA US FRA ESP LKA CHN PERNLDUSA SWEHUNMEX USA KOR PHLCANVEN TWNKOR ISR SWE USAMEXGBR TUR JPN CH ESP JPN ITA ITA LKA DNK CHLPRTKOR SWETUR EGYCHN EGYVEN EGYESP KOR ISL ISRZAFESP JPN JPN ITA DNK PERAUT JOR PERIDN PER MYSGRCGBR KOR KOR EGY ESP ESP JPN ITA COL LKA ZWEDNKCHLTWN ZWEDNKCHLTWN GRCGBR LKA DNK AUTKOR FINSWETWNMEX GBRINDPHLVENESP DNK IDNAUSGRC PER HUN MEX SWE TWN PER JOR JOR IRLSGP ZAFVEN JPN ITA JPN JPN JPN noi ZWENOR TWNMEX LKADNK NLDKOR PER PHLZAFVEN ZAFCHN VEN ESP COL PER USAMEX GBRINDISRCAN ZAFESP SWETWN GRC TWNGRC JOR JPN JPN JPN ITA ITA LKA DNKSWEUSA LKA DNKFINAUTMYSMEX CHL SWETWNMEXSGP SWEIDNAUS SGP MYS INDPHLVEN EGY EGY ZAFESP COL LKA FINSWE LKA FINAUTTWN BELGBR PERNLDKORIRL SWE ZAFESP ESP ITA LKA DNKCHLTWNGRCGBR INDPHLCAN ZWE NOR IDNMYS LKA FINCHLTURAUSSGP PERPRTKOR ISR EGY ESP ITA COL COL DNKCHLMYSAUSIRLGBR PHL DNKAUTUSA BELGBR DNK FINCHLMYS NOR USAMEXGBR PERPRTKOR SGP PER JOR TUR PHLVEN ISR VEN EGYVEN PHLVEN ISRCACHN EGYESP CHN ESP LKA FINAUTKOR PER USA IRL DNK TUR PERIDNKOR SWETUR BEL KOR ISLGBR JPN ITA ITA ITA ITA LKA DNKAUTTURGRCSGPINDPHLCAN LKA FINAUTTWN NORNLDHUNGRCSGP FRAEGYCAN PERPRTTURPOLGBR SWEIDNKOR AUTTWNMEXSGP AUTTUR TUR ISL IRLGBR JPN ITA noi NOR EGYCHN JPN ITA COL KORIRLGBRINDEGY VEN COL FINCHLMYS NOR AUTUSA USA ISL ZAFESP CAN VEN ITA ITA at 0 ZWEDNKCHLTWN PERAUTTUR COL LKA ZWE DNKAUTMYSGRCIRL FINCHLIDNUSA ISL NORIDN CHLAUTKOR SGP INDPHLCAN CHLMYSAUSSGP FINPRTTWNGRCGBR CHLPRTKO CHLUSAAUSIRL AUTJOR NLDTWNGRC PRTTUR ISRCAN ZAFCHN ESP JPN ITA COL LKA KOR AUSISLGBR SGP FRAEGYVEN GBR PHLCANN ZAFVEN ZAFESP ESP ZWE DNKAUTTUR LKA PERNLDKOR JOR ESP ESP at COL COL DNKCHL PERIDNKOR IRLGBR PERAUTUSA SWETWNAUS GBR INDEGYCAN SWENLDKOR ISL PRTTUR IRLGBR VENMAR JPN JPN COL CHLTWNMEXSGP ISRCAN ZAFCHN ESP SWEMYSAUS CHLMYS IDNTUR CHN elrroc 0 ZWE NORPRTUSA ZWEDNKAUTMYSAUS GBRINDEGYVEN FINSWEMYS AUTTWN PHL ESP TU GB ZAFESP JPN ITA LKADNKCHLJORKORIRL LKAFINCHLPRTTUR DNKAUTUSAMEXSGPINDINDPHLVEN FINAUTMYSAUSSGP PERUSAGRCSGP NLDTUR VEN VEN CHN ESP ITA ITA ITA ITA ITA COL SWEJOR COL NORNLDTUR EGY JPN AUTJOR IRL AUTTUR PRT PRT IDNUSAPOL GBR PHLVENMAR EGYCAN JPN COL COL PRTUSAGRC ISR PHLCANESP LKA FINCHLIDNKOR FINCHLIDNKOR FINAUTMYS PERNLDTURGRCGBR IND PERPRT AUS ISL TWN TWN MYS JOR JOR GBRINDPHL CAN ISL SGP INDPHL GBRINDISR ISL SGP LKA ZAF FINNLDJORAUSISLGBRINDPHLEGYVEN CHLMYSKORIRL SWETWN CHL USAMEXSGPINDPHL CAN ESP COL COL DNKAUTTWN PHLCANCHN LKA ZWE ZWE CHLIDNKOR PRTTUR IDNTUR ITA COL AUTMYS TWNAUS IRLGBRINDPHL NLDAUSIRL JOR IRLSGP JOR PHLCAN ISRCAN EGYESP COL FINCHLIDN MEXISL NORPRTJOR PERPRTKOR TWN MYS TUR JOR JOR IRLGBRINDPHLCANMAR USA BEL JOR ISRCAN ISRVEN EGY EGY COLZWEDNKNLDUSA COLZWEFINSWEMYS ZAFCHN FINCHLIDNKOR PERAUTUSA IRLSGPINDPHLZAFESP PERIDNTUR PERPRTKOR COL LKA NORPRT NORNLDKOR ISL PEAUTMYS GRC IRL COL CAN CHN ITA COL PRTTUR IRLGBRIND EGYVEN USA MYS TWN TUR ZAF ITA ZWE FINCHLTUR MEXSGPINDEGYVEN GBR TUR elrroc FINAUTIDN PRT TWN USA TURGRC GBRIND CAN MAR COL ZWEFINPERMYSAUSIRLGBRIND DNK IDNJOR DNKIDNTUR FINAUTMYS CANESP JPN CHLMYS NLDJOR KOR IND CANVEN COL EGYCANMAR EGY ESP ZAF ZAFCHN CAN VEN FIN FIN PERPRT MYSAUSIRL ISL INDPHLCAN USA SGP SGP EGY LKA ESP COL JPN COL NLDTURGRC GBRIND PRT MYS LKA DNKCHLTWN LKA NORNLDTURMEX LKA LKA DNKAUTMYS FINCHLTWN NORPRT NORPRT PRTKORIRL IDN LKA FIN NORAUTJORMEXIRL FIN PERIDNJOR PRT IRL JOR USA ISL IDN MYS TWNMEXISL MYS MYS JOR HUNGRCISL IRL EGYVENMAR VEN COL NOR PRTJOR NORCHLTUR MYS MYSAUSIRL JOR PHLCAN CAN CAN ESP TWN TURGRCSGPINDPHLCAESP TUR IND IND COL FINCHLIDNTWNBELGBRINDISREGYCHN NORAUTUSA MYS BEL IDN AUSIRL TWN TUR ESP ZWEDNKAUTTUR ZWEDNKAUTTWN GRC NORPRTJOR FIN AUTJOR GRC ZAF ZAF CHLTWN TWNMEXISL INDPHLCAN COL PHL CAN JOR INDEGYVEN PHL MAR COL ZWE FIN MYS IRL EGYESP ITA INDISR VEN EGY ISR COL 5 FIN TU IRL 5-. FIN EGYCHN ESP COL NOR PRT TWNMEXSGP INDEGY ESP EGYCHN JPN COL ZWE FIN DNK NOR INDISR EGY ISR ISR EGY FIN NLDMYSAUS TURMEX ISRCAN ESP -. DNKCHLMYSAUSIRL IDNKOR TUR JORMEXISL JO ISR DNK NLDPRT DNKAUTMYS DNKAUTJOR NORPRTJOR PRTJORBEL TWN MYS BEL IRL ISL ISR ZWE FIN ISR ZWE ZWE EGY NLD MYSMEX AUS NLD MEX EGY ZWE NOR DNKAUTJORMEX PER CHL BEL NOR CHL GRCISL IRL IRL EGY PER MEXISL CAN MAR MEXISL ZWE MEX MAR JPN NOR NOR TWNMEX -1 -1 HUN -.1 0 .1 .2 -.1 0 .1 .2 Payables Financing Payables Financing coef = 1.1496666, (robust) se = .41925543, t = 2.74 coef = .046824, (robust) se = .44577899, t = .11 The figures in Panels A to D show the relation between the median level of Payables Financing among manufacturing firms and the correlation of real value added growth in the industry pair formed by the Iron and Steel and Fabricated Metal Product industries, the industry pair formed by the Tobacco and Footwear industries, the 20 industry pairs that are closer in credit chain distance, and the 20 industry pairs that are farther in credit chain distance, respectively. Each figure reports the coefficient (coef) of an OLS regression between the measures of correlation and payables financing (plus an industry-pair fixed effect in panes C and D) for the industry pairs considered in each case. The reported standard errors and t-statistics (se and t) are clustered at the country level. 48 Figure 4. Distribution of main coefficient after dropping countries and industries Panel A. Distribution of main coefficient after dropping one country at a time .8 .6 noitca .4 Fr .2 0 1.4 1.6 1.8 2 2.2 2.4 Coefficient for the credit-chain distance median=1.98, mean=1.98, sd=.14, iqr=.08 Panel B. Distribution of main coefficient after dropping one industry-pair at a time .5 .4 .3 noitca Fr .2 .1 0 1.9 2 2.1 2.2 Coefficient for the credit-chain distance median=1.98, mean=1.98, sd=.02, iqr=.01 The figure in Panel A shows the distribution of the OLS coefficient of credit-chain distance obtained after dropping one country of the sample at a time. Panel B shows a similar histogram obtained after dropping one industry-pair at a time 49 Figure A1. Effect of measurement error on the coefficient estimated for the effect of credit-chain distance on correlations Panel A. Measurement error on the Payables Financing ratio 2 ). efoc 5 1. uert 1 morf ecn .5 reeffid(.f 0 5 -. oec deta -1 mits 5. -1 E -2 0 .5 1 1.5 2 Standard deviation of measurement error (fraction of sample sd.) Empirical 95% conf. interval Empirical Mean Panel B. Measurement error associated with the use of a common measure of Input-Output Linkages 2 ).feoc 5 1. eurt 1 omrf .5 ontiaived(. 0 5-. efoc de -1 matti 5. -1 Es -2 50 100 150 200 Standard deviation of measurement error Empirical 95% conf. interval Empirical Mean The two panels of the figure display the empirical distribution of the difference between the estimated and true coefficient for the effect of credit-chain distance on sectoral correlations (y-axis) obtained from a series of Montecarlo simulations for different variances of the measurement error on the payables financing (Panel A), and Input-Output Linkages (Panel B) (x-axis). In each panel, the line represents the mean estimated coefficient and the shaded area the empirical 95 percent distribution. The value of the true coefficient is all simulations is 2. 50