The Importance of Industry Links in Merger Waves KENNETH R. AHERN and JARRAD HARFORD⋆ ABSTRACT We represent the economy as a network of industries connected through customer and supplier trade flows. Using this network topology, we find that stronger product market connections lead to a greater incidence of cross-industry mergers. Second, mergers propagate in waves across the network through customer-supplier links. Merger activity transmits to close industries quickly and to distant industries with a delay. Finally, economy-wide merger waves are driven by merger activity in industries that are centrally located in the product market network. Overall, we show that the network of real economic transactions helps to explain the formation and propagation of merger waves. Journal of Finance, forthcoming ⋆Kenneth Ahern is at the University of Southern California, Marshall School of Business. Jarrad Harford is at the University of Washington, Foster School of Business. We thank two anony- mous referees, an anonymous associate editor, Cam Harvey (editor), Sugato Bhattacharyya, Hans Degryse, Ran Duchin, Gerard Hoberg, Jonathan Karpoff, Han Kim, Vojislav Maksimovic, David McLean, Sara Moeller, Gordon Phillips, Ed Rice, Matthew Rhodes-Kropf, David Robinson, Shawn Thomas, Karin Thorburn and seminar participants at the 2010 Texas Finance Festival, European Summer Symposium in Financial Markets (2010) First European Center for Corporate Control Studies Workshop (2010), 2010 Frontiers in Finance Conference, 2011 Washington University Con- ferenceon CorporateFinance, 2011 AFAMeetings, 2011 UBC Winter Finance Conference, Georgia State University, University of Illinois, University of Maryland, University of Michigan, University of North Carolina – Chapel Hill, University of Pittsburgh, and University of Wisconsin for helpful suggestions. We also thank Jared Stanfield for excellent research assistance. This research was partially completed while Kenneth Ahern was at the University of Michigan. 1 A growing body of evidence shows that industry characteristics affect many firm decisions, includ- ing financial policy (MacKay and Phillips, 2005), internal capital markets (Lamont, 1997), and corporate governance (Giroud and Mueller, 2010). This line of research emphasizes that strategic interactions between firms and their industry rivals have important implications for fundamental questions in financial economics. We broaden this analysis by making a simple, though consequen- tial, observation: Industries do not exist in isolation, but rather are connected through a complex network of customer-supplier relationships. This implies that whole industries may be affected by shocks that are transmitted through the customer-supplier network. In this paper, we inves- tigate how inter-industry relations affect the timing and incidence of one of the most important phenomena in corporate finance: merger waves. Theindustrynetwork modelof an economy has at least threenew implications for merger waves. First, industry-level economic shocks could lead to cross-industry vertical merger waves. Though it is well documented that merger waves occur within industries (Mitchell and Mulherin, 1996; Maksimovic and Phillips, 2001; Rhodes-Kropf, Robinson, and Viswanathan, 2005), vertical merger waves may be just as common. Second, merger waves could propagate through customer and supplier links without direct vertical integration. For instance, the reorganization of a supplier industry could cause a customer industry to reorganize in response. Third, the structure of the industry network could determine how industry-level M&A activity aggregates into an economy- widemergerwave. Theseimplicationsareimportantforunderstandinghoweconomicfundamentals at the industry-level influence economy-wide outcomes. To test these three implications, we empirically model the product market network using input- output data from the Bureau of Economic Analysis. These data provide trade flows between 471 industriesaccounting for allsectors in theeconomy. Usingthese industrydefinitions, wealso create a network representing cross-industry mergers over the period 1986 to 2010, where the strength of the connection between two industries is proportional to the level of their cross-industry merger activity. Thus, for a comprehensive set of industries, we define two different types of inter-industry connections: input-output trade flows and cross-industry mergers. We first characterize the product market and merger networks. We find that both networks are sparse, but highly inter-connected through a relatively small set of centralized ‘hub’ industries. 2 To illustrate, more than 95% of industry-pairs in the product market network have almost no customer-supplier relations. Similarly, all cross-industry mergers in our sample occur in just 6% of all possible industry-pairs. This means that the average industry engages in mergers with a small set of local industries that are closely related through customer-supplier links. We also find that the product market and merger networks both exhibit small-world properties, where the average industryisseparatedfrommostotherindustriesbyonlytwoorthreedirectconnections,evenacross 471 different industries. In addition, we find that an industry’s centrality in the product market network iscorrelated with its centrality in themergernetwork, as are other network characteristics, such as clustering andaverage distance. Thus, thestructureof themerger network is highly similar to the structure of the product market network. This characterization of the product market and merger networks supports our first finding that vertical mergers are common and highly clustered in a relatively small set of directly-linked industry-pairs. Ofthe51,002 mergersinthesample,61% areinter-industrymergers. Priorresearch identifiesmanyreasonsforverticalmergers.1 Neoclassicaltheoryproposesthatverticalmergersmay eliminate an existing inefficiency, such as double price markups in successive monopolies (Spengler, 1950; Perry, 1978b) or input substitution (Vernon and Graham, 1971; Schmalensee, 1973; Warren- Boulton, 1974). Another neoclassical motive for vertical mergers is to prevent resale of an input in downstream industries in order to allow price discrimination across different price elasticities of demand (Perry, 1978a; Katz, 1987). An alternative to the neoclassical theory, transaction costs may lead to vertical integration if the net benefits of internal transactions are larger than those of transacting in a market (Coase, 1937; Williamson, 1979). The costs of market transactions and the correspondingholdupproblemsincreasewithuncertaintyandwithrelationship-specificinvestments (Klein, Crawford,andAlchian,1978). Thus,firmswithcomplementaryassets maymergewitheach other to overcome incomplete contracts (Rhodes-Kropf and Robinson, 2008). Wefindevidenceconsistentwithtransactioncosttheories,whileshowingthatinput-outputtrade flows predict cross-industry mergers. We estimate exponential random graph models (ERGM), which are multivariate maximum likelihood regressions developed specifically to allow for simulta- neous dependence relations between all nodes in a network. The ERGM results show that there 1Comprehensive surveysof themotives for vertical integration can be found in Tirole (1988) and Perry (1989). 3 are more inter-industry mergers between two industries when they have stronger customer-supplier relations, controllingforindustryvaluation, scope, size, returns,concentration, andmacroeconomic shocks. We also findthat cross-industry mergers are more likely when industries have greater R&D expenditures and that R&D magnifies the effects of productmarket links. To the degree that R&D proxies for incomplete contracts, these results are consistent with holdupproblems. Inaddition, we find evidence that cross-industry mergers are positively related to asset complementarity, following Hoberg and Phillips (2010b). We are careful to note that we do not claim to separately identify each motivation for vertical mergers, but instead, we provide evidence that shows that product market trade flows have a first-order effect on the incidence of cross-industry mergers. The relations between product market links and mergers are economically significant. Industry pairs without a meaningful economic connection have, on average, 0.11 mergers between them over the sample period. Those with a strong connection have an average of 12.5 mergers. This effect is present in every year from 1986 to 2010 and is stronger duringmarket boomsand aggregate merger waves. Thisimpliesthateconomicfundamentalsaremore,notless,importantduringmergerwaves. The second implication of the industry network model is that merger waves could propagate through customer and supplier links without direct vertical integration. Galbraith (1952) predicts thatindustryconsolidationinanupstreamindustryleadstoconsolidationinadownstreamindustry tocounteractthemonopolypowercreatedthroughtheinitialconsolidation. Morerecenttheoretical industrial organization models predict that changes in the substitutability of products or changes to the cost structure of one industry affect the incentives to merge for firms in vertically related industries (Horn and Wolinsky, 1988; Inderst and Wey, 2003). Thus, merger activity could be transmitted through economic links between industries, even without vertical integration. Consistent with this, we find evidence that mergers propagate across the industry network fol- lowing a wave-like pattern. We measure each industry’s exposure to merger activity in related industries, not including mergers with the industry itself. We use graph theory techniques to iden- tifywhichindustriesareclose andwhicharedistantintheproductmarketnetwork. Accountingfor a number of controls, including industry fixed effects and an industry’s own lagged merger activity, we find that mergers in close industries have a strong positive effect on an industry’s own merger activity after a one-year delay, while merger activity in distant industries has a positive impact 4 after a delay of two or three years. Thus, merger waves travel across customer-supplier links, even without direct vertical integration. We also find that the impact of mergers in supplier industries is largerandtravels fasteracrossthenetwork, thandoestheimpactofmergersincustomerindustries. This likely reflects the fact that the supplier network is more densely connected. In the last section of the paper, we investigate the third implication of the network perspective: the structure of the industry network could determine how industry-level M&A activity aggregates into an economy-wide merger wave. In vector autoregressions, we find that the industries that experience merger waves during the height of overall economy-wide merger activity are the most central industries in the productmarket network. This is a direct consequence of the highly skewed distributionof inter-industry connections. Asmerger activity transmits across thenetwork towards more central industries, many overlapping industry waves occur, which produces an aggregate merger wave. This evidence contradicts the idea that industry merger activity caused by random shocks do not cluster in time, and therefore cannot explain economy-wide aggregate merger waves (Shleifer and Vishny, 2003). Our evidence suggests that even if the initial industry shocks are random, aggregate merger waves occur, in part, because of the structure of the industry network. This paper makes two primary contributions to the literature. First, this paper is related to recent research that investigates the role of industry relations in corporate finance. Bhattacharyya and Nain (2011) study the price effects on suppliers and customers following horizontal mergers. Becker and Thomas (2010) examine how changes in concentration in downstream industries affect concentration in upstream industries. Fee and Thomas (2004) and Shahrur (2005) use vertical relations to test the effects of horizontal mergers on market power, buildingfrom Eckbo (1983) and Stillman (1983). Hertzel, Li, Officer, and Rodgers (2008) find that suppliers to firms that file for bankruptcy suffer negative and significant wealth effects. Our paper is the first to focus on the role of input-output connections for cross-industry mergers. Although it is generally accepted that some mergers are motivated by vertical integration, very little about vertical mergers has actually been documented. Fan and Goyal (2006) report that prior to their paper, even basic facts such as the proportion of mergers that are vertical were unknown. Our paper is unique in that we study the determinants of the incidence and timing of inter-industry mergers across all industries, rather than the value implications of the mergers that do occur. Our paper is also related to a strain 5 of recent research on merger waves, including Maksimovic, Phillips, and Yang (2010), Duchin and Schmidt (2012), Garfinkel and Hankins (2011), and Ovtchinnikov (2010). The second contribution of this paper is to model the economy as a network of customer and supplier relations. This approach is related to Hoberg and Phillips (2010a, 2010b), who use network techniques to group firms based on textual product market descriptions. In our paper, we exploit the input-output trade flows to model network ties based on exogenous real economic trade flows between industries. The network approach provides key benefits over the analysis of single connections between suppliers and customers. In particular, by considering all industries, we alleviate selection bias caused by only considering industry-pairs directly involved in mergers. Second, thenetwork approachexplicitly accounts fordependenciesbetween allindustries,including higher-order connections and allows for tests of the propagation of industry-level shocks from one industrytoanother, acrosstheentireeconomy. Webelievethatthisapproachwillhavefar-reaching applications for understandingthe interaction of corporate financeand industrialorganization. For the sake of brevity, we present only a fraction of the description of the product market network in the paper, but provide a comprehensive report in the Internet Appendix, which may be useful for future research. The rest of this paper is organized as follows. Section I presents the industry and merger data and describes the construction of the networks we analyze in the paper. Section II presents tests that compare the industry input-output network to the merger network in a static setting. In Section III, we present tests of the propagation of merger waves across the industry network over time. Section IV presents tests of aggregate merger waves and network centrality. Section V concludes. I. Data Sources and Methods A. Customer-Supplier Trade Network Data Since 1947, the Bureau of Economic Analysis (BEA) has provided Input-Output (IO) accounts of dollar flows between all producers and purchasers in the U.S. economy. Producers include all industrial and service sectors as well as household production. Purchasers include industrial 6 sectors, households, and government entities. Thus, these data cover the entire economy, not just manufacturing industries. The IO tables are based primarily on data from the Economic Census and are updated every five years with a five-year lag. Since our merger data (described below) cover the period 1986 to 2010, we use the IO tables from the years 1982, 1987, 1992, 1997, and 2002, the most recent report as of July 2012. The BEA defines industries at two levels of aggregation, detailed and summary. The number of detailed industries, excluding households and government sectors, ranges between 411 and 478 in the different reports. This is slightly more narrow than the 416 three-digit 1987 SIC codes, but substantially more coarse than the 1,005 four-digit SIC codes. The detailed IO industries are also closer to the number of four-digit NAICS codes in 1997 (313), than to the number of five-digit NAICS codes (721), or six-digit NAICS codes (1,179). The number of summary-level IO industries ranges between 77 and 126, which is similar to two-digit SIC codes (83) and three-digit NAICS codes (96).2 Thus, the coarseness of the IO industry definitions are roughly equivalent to two and three-digit SIC codes, which have been used extensively in prior research. In each report, the BEA updates the classifications used in the IO tables to reflect changes in the economy. The classifications are designed to group firms into industries that best measure customer and supplier relations, using the most recent standardized industry classifications. Prior to 1997, the IO industries were defined based on 1977 and 1987 SIC codes. In 1997 and 2002, the BEA based the IO industries on 1997 and 2002 NAICS codes, following the policy of most U.S. government agencies to switch from SIC to NAICS codes. Concordance tables between NAICS and SIC codes and IO industry codes are provided by the BEA. Since our unit of observation is an industry-pair, to maintain consistency over the years in our sample we cannot combine data from different BEA reports in the same analysis. Therefore, in the main analysis, we present results using the 1997 detail-level IO definitions. We choose the 1997 report because 1997 splits our merger data into two approximately equal time periods. The 1997 report is also concurrent with the largest aggregate merger activity in our sample period. We choose to focus on the detail-level industries in the main analysis, because it allows for a more granular representation of the economy. Therefore, unless otherwise noted, the results presented 2InternetAppendixTableIreportsthenumberofindustriesacrossSIC,NAICS,andBEAIOdefinitionsforvarious years. 7 in the paper refer to the detail-level industries in 1997. However, for robustness, in the Internet Appendix, we run our tests using both detailed and summary-level IO relations from the 1982, 1987, 1992, and 2002 reports. We will direct the reader to specific tables in the Internet Appendix for each robustness check. EachIOreportdefines‘commodity’outputsandproducing‘industries.’ Acommodity, asdefined by the BEA, is any good or service that is produced. An industry may produce more than one commodity, which means that more than one industry may produce the same good or service. However, the output of an industry is typically dominated by one commodity. The ‘Make’ table of the IO report records the dollar value of each commodity produced by the producing industry. In the 1997 report, there are 480 commodities and 491 industries in the Make table. The ‘Use’ table defines the dollar value of each commodity that is purchased by each industry or final user. There are 486 commodities in the Use table purchased by 504 industries or final users.3 Costs are reported in both purchaser and producer costs (the differences are due to retail and wholesale markups, taxes, and other transaction costs). Throughout the paper we use producers’ prices, but using purchasers’ prices makes little difference. From the Use and Make tables, we create matrices that record flows of inputs and outputs between industries. Following Becker and Thomas (2010) we calculate SHARE, an I ×C matrix (Industry × Commodity) that records the percentage of commodity c produced by industry i. The USE matrix is a C×I matrix that records the dollar value of industry i’s purchases of commodity c as an input. The REVSHARE matrix is SHARE×USE and is the I×I matrix of dollar flows fromthecustomerindustryoncolumn j tosupplierindustryonrow i. Finally, theCUST matrix is REVSHARE divided by the sum of all sales for an industry. The SUPP matrix is REVSHARE divided by the sum of all purchases, by industry. The CUST matrix records the percentage of industry i’s sales that are purchased by industry j. The SUPP matrix records the percentage of industry j’s inputs that are purchased from industry i. These two matrices describe the relative trade flows between all industries in the economy. 3The six additional commodities that are in the Use table but not in the Make table are, noncomparable imports, usedandsecondhandgoods, restofworld adjustmenttofinaluses,compensation ofemployees,indirect businesstax and nontax liability, and other value added. The thirteen industries or final users in the Use table that are not in the Make table include personal consumption expenditures, private fixed investment, change in private inventories, exports and imports, and federal and state government expenditures. 8 The IO tables treat compensation of employees as a commodity input in production. However, thereisnocorrespondingindustrythatproducescompensation. Because of this, employee compen- sation gets dropped from the industry matrices. Therefore, we create an artificial labor industry to make sure that we account for labor as an input in the industry matrices. Without including labor costs, other inputs in labor-intensive industries will appear to be a larger component of total inputsthanthey actually are, relative tocapital-intensive industries. Theadditional laborindustry is used only to account for inputs, and we do not include labor as an industry or commodity in our finalsample. After excludinghousehold and government industries, as well as exports and imports, and making a few minor adjustments, we are left with 471 industries. A detailed description of the data is reported in Section I of the Internet Appendix. Oneoftheimportantfeaturesoftheinput-outputmatrix isthatitislargely exogenous tomerger activity. Thisisbecausethebasicinputrequirements intheproductionof any goodaredetermined mainly by thegood’s productionfunction, notby theownership structureof the firmsthat produce the inputs.4 The exogeneity of the product market network, with respect to ownership, mitigates concerns about reverse causality, where merger waves cause product market relations to change. In addition, by using the 1982 IO reports in robustness tests, we ensure that the IO relations are exogenous to merger activity from 1986 to 2010. B. Merger Network Data Merger data are from SDC Thomson Platinum database. We collect data for all mergers that meet the following criteria: 1) Announcement dates between 1/1/1986 and 12/31/2010; 2) Both target and acquirer are U.S. firms; 3) The acquirer buys 20% or more of the target’s shares; 4) The acquirer owns 51% or more of the target’s shares after the deal; 5) Only completed mergers; and 6) Transaction values of at least $1 million. Since the focus of this study is merger activity, rather than wealth effects, we do not restrict the legal form of organization of the target or acquirer. This produces a sample of 51,002 observations. By not restricting our sample to public firms, we have a much more complete sample than is typically used in existing merger research. 4Itispossiblethatverticallyintegratedfirmscouldusesubstituteinputsbasedontheirownershipofcertainsupplier segments. However, if input substitution leads to inefficient production, these firms are unlikely to survive, or, alternatively,theinputsubstitutionisnotimportant. Forthistoaffectourresults,theinputsubstitutionwouldneed to occur at an industry-level,rather than at a firm-level. 9 For each observation, we record the value of the deal, the date, and the NAICS codes of the acquirer and target. Because SDC records 2007 NAICS codes we convert all NAICS codes from SDC to 1997 NAICS codes to match to the IO data. Then for each deal we map the 1997 NAICS to the appropriate 1997 IO industry. In the robustness tests that use IO reports from years other than 1997, we match SIC codes from SDC. This means, for example, that in the tests that use the 1982 IO reports, we first convert 1987 SIC codes reported in SDC to 1977 SIC codes to match to the IO definitions. Section I of the Internet Appendix provides more details on the mapping between industry classifications. Next, we record merger activity both yearly and cross-sectionally for each directed IO industry- pair of acquirer and target industries. This produces 4712 = 221,841 unique pairs. Directed industry pairs means that we differentiate between acquirer and target industries. For each time window (yearly and cross-sectionally) we record the number and dollar value of mergers where the acquirer is in industry i and the target is in industry j. This means we have separate observations for deals involving acquirers in industry i that are buying targets in industry j and deals involving acquirers in industry j that are buying targets in industry i. Since in inter-industry mergers, it is likely that the acquirer could be in either industry, we also record the data in a non-directed way between two industries. This yields 1 ×471×(471−1) = 110,685 unique industry pairs per 2 window of observation. Inthemainanalysis, wematch firmstoIOindustriesusingtheir primaryNAICScode. However, this does not account for diversified firms. We address this concern in a few ways. First, as mentioned previously, the IO industries are roughly as coarse as three-digit SIC codes. Firms with multiple, but related, segments, will tend to get assigned to the same IO industry, regardless of which segment’s 6-digit NAICS code is used. However, this does not account for firms with multiple unrelated segments that would be assigned to different IO industries, depending upon which industry is listed as its primary segment. Therefore, we use three alternative methods to assign firms to IO industries. In the firstalternative method, we use all industry codes reported in SDC to identify a full set of IO industries per firm. We then assign equal weight to merger counts and dollar volumes for each of these IO industry codes. For instance, if an acquirer is in industries 1 and 2 and a target is in