Abstract

ABSTRACT We examine the incentives for firms to vertically integrate through acquisitions and production. We develop a new firm-specific measure of vertical relatedness and integration using 10-K product text. We find that firms in high R&D industries are less likely to become targets in vertical acquisitions or to vertically integrate. These findings are consistent with the idea that firms with unrealized innovation avoid integration to maintain ex ante incentives to invest in intangible assets and to keep residual rights of control. In contrast, firms in high patenting industries with mature product markets are more likely to vertically integrate, consistent with control rights being obtained by firms to facilitate commercialization of already realized innovation. * University of Maryland, University of Southern California, and University of Southern California and National Bureau of Economic Research, respectively. Frésard can be reached at lfresard@rhsmith.umd.edu, Hoberg can be reached at gerard.hoberg@marshall.usc.edu and Phillips can be reached at gordon.phillips@marshall.usc.edu. We thank Yun Ling for excellent research assistance. For helpful comments, we thank Kenneth Ahern, Jean-Noel Barrot, Thomas Bates, Giacinta Cestone, Robert Gibbons, Oliver Hart, Thomas Hellmann, Adrien Matray, Sébastien Michenaud, Steve Tadelis, Ivo Welch and seminar participants at the National Bureau of Economic Research Organizational Economics Meetings, 2015 International Society for New Institutional Economics Meetings, 2014 American Finance Association Meetings, Arizona State University, Carnegie Mellon, Dartmouth College, Humboldt University, IFN Stockholm, Harvard-MIT Organizational Economics joint seminar, Tsinghua University, University of Alberta, University of British Columbia, University of California Los Angeles Universidad de los Andes, University of Maryland, University of Washington, VU Amsterdam, and Wharton. All errors are the authors alone. Copyright c 2015 by Laurent Frésard, Gerard Hoberg and Gordon Phillips. All rights reserved. The scope of firm boundaries and whether to organize transactions within the firm (integration) or by using external purchasing is of major interest in understanding why firms exist. They show that ex ante incentives for a firm to invest in relationship-specific assets are reduced under vertical integration for the firm that gives up its residual rights of control to the other contracting firm. 1 These theories particularly apply to innovation activities as technological investments are typically relationship-specific, not fully contractible and often unverifiable 2 In this paper, we argue and provide evidence that the costs and benefits of vertical integration, and hence firm boundaries, are related to the stage of development of innovation. In particular, we find that the distinction between unrealized innovation in the form of R&D and realized innovation characterized by legally enforceable patents is a key empirical determinant of firms' vertical organization and acquisitions. We capture variations in firm boundaries based on new firm-specific measures of vertical relatedness that we construct using text-based analysis of firm 10-K product descriptions filed with the Securities and Exchange Commission (SEC), and vertically-linked product descriptions from the Bureau of Economic Analysis (BEA) Input-Output tables. This allows us to precisely determine vertical relatedness across firm pairs and identify which mergers and acquisitions are vertically related. This framework also allows us to develop a new measure of vertical inte-1 2 As highlighted by Acemoglu, Aghion, Griffith, and Zilbotti (2010) innovative activities are subjected to the types of problems featured by the transaction costs and property right theories. 1 gration at the firm-level based on whether firms use product vocabulary that spans vertically related markets. Because 10-Ks are updated annually, we relate firm vertical organization to the stage of innovation as these activities dynamically evolve. 3 This more direct approach is not possible using measures of vertical integration based on static methods that rely on SIC or NAICS codes. Using a sample of almost 7,000 publicly-traded firms over the 1996-2008 period, we find strong evidence that firms in R&D intensive industries are less likely to be acquired in vertical transactions. In contrast, firms in patent intensive industries are more likely to be targeted in vertical transactions. In addition, we show that vertical acquisitions tend to occur at times when target firms have accumulated more patents, in contrast to non-vertical acquisitions which are diametrically opposite. The distinction between unrealized and realized innovation also matters for observed firm-level vertical integration. We find strong evidence that firms in R&D intensive industries are less likely to be vertically integrated whereas firms in high patenting industries are more likely to be vertically integrated. These findings are economically large: In our baseline specification, firm vertical integration decreases by 10% following a one-standard deviation increase in R&D intensity, and increases by 7% following a one-standard deviation increase in patenting intensity. These findings are consistent with firms with unrealized innovation avoiding integration to maintain ex ante incentives to invest in intangible assets and maintain residual rights of control as in These results are consistent with a simple model based on incomplete contracting that links R&D and patent intensity to vertical acquisitions. The decision of two firms to integrate vertically depends on the stage of development of the specific asset exchanged in their relationship. The model predicts that when the asset is still in the form of R&D (unrealized innovation), firms optimally will not integrate. 4 To further assess our conclusions, we follow Bloom, Schankerman, and van Reenen (2013a) and exploit variation in R&D tax credits across U.S. states to construct an instrument for industries' R&D intensity. Confirming our baseline results, this instrumental variables framework indicates that an increase in industry R&D significantly lowers the likelihood that a firm will be acquired in a vertical acquisition. Also, higher R&D intensity leads to significantly less firm-level vertical integration. The distinction between high R&D and patents is empirically nontrivial. High R&D does not necessarily lead to high patenting rates, and patent rates vary across industries. As reported by R&D labs, high R&D may not lead to patents due to concerns about appropriability. Their survey points to the ability of others to work around patents using information conveyed by the patent application, causing managers not to patent in many industries. Consistent with property rights for realized innovation being important, we find that the link between patents and vertical integration is only present in industries where patents provide effective coverage. insider indicated that the deal between the two companies would help to bring the relationship-specific assets that are difficult to contract on. 3 "hardware closer to the operating system and achieve a tighter integration." Buying firms to gain control of their realized innovations facilitates commercialization either through reduced ex post hold-up or increased commercialization incentives. 5 An industry that exemplifies our findings regarding dynamic vertical integration is the network equipment industry, which includes Cisco, Broadcom, Citrix, Juniper, Novell, Sycamore, and Utstarcom. During our sample, and using our measures, we find that firms in this industry jointly experienced (A) levels of R&D that peaked and began to decline, (B) levels of patenting activity that rose four to five fold, and (C) levels of vertical integration that also rose four to five fold. The conversion of unrealized innovation into realized patented innovation reduced the incentives for relationship-specific investment, and increased the incentives to vertically integrate in order to transfer control rights to the party commercializing the patents. 6 We also consider the role of supply chain stability and maturity. 4 ex post effects Our findings contribute to a large literature examining the determinants of vertical integration (see 9 Specifically, they show a relative decline in non-production workers in acquired establishments that are vertically related. They also show an increase in products that were made by the acquiring firm previously in the acquired firms' establishments. 5 ing power. Ahern and Harford (2013) The remainder of this paper is organized as follows. Section II develops a simple model of vertical integration. Section III presents the data and develops our measures of vertical relatedness. Section IV examines the effect of innovation activities on vertical acquisitions, and Section V examines the impact on firm vertical integration. Section VI concludes. 10 See http://www.naics.com/info.htm. The Census Department states "NAICS was developed to classify units according to their production function. NAICS results in industries that group units undertaking similar activities using similar resources but does not necessarily group all similar products or outputs." 11 II A Simple Model of Integration To illustrate the contrasting effects of realized and unrealized innovation on firm integration decisions, we develop a simple dynamic incomplete contracting model of vertical acquisition using the framework introduced by The model is meant to illustrate the trade-offs of vertical integration versus separation over time. Consider an upstream supplier and a downstream producer. At each time t, they cooperate to produce a product at a base price P b t . The sale price P t that can charged on consumers further depends on commercialization and product integration investments chosen by the downstream firm as well as R&D investments by the upstream firm that can result in new patentable features. In the spirit of Grossman and Hart (1986) and Aghion and Tirole (1994), we assume that both R&D and commercialization investments are relationship-specific, non-contractible and nonverifiable. At each period, firms can either operate as separate entities or can decide to integrate. 12 Integration implies the acquisition of a firm (or the patent) from a firm by the other firm. The party that sells its assets is called the target and it loses control rights over the assets sold, and thus makes no further relationship-specific investment. For each t, the upstream supplier chooses an x t amount of R&D effort with a cost We assume x t is the non-contractible portion of R&D effort. Thus, if the downstream producer acquires the upstream supplier, x t will be equal to zero. 13 The downstream producer chooses an amount y t of commercialization investment that can also boost the price of the product with a cost m t = c(y t ) = Ry h t . Commercialization investments can include for instance marketing the product, building 12 Our model can be thought of as a model of one firm doing R&D which results in a patent. This patent can be used in the supplier's production process to improve what is sold to the downstream firm. Thus, integration can be viewed as either a bundled sale of all the assets of the target or the sale of a patent that can be separated from the target firm and used by the downstream firm to improve its product. This would come with some cost associated with using for the patent that varies with ownership of the patent or the bundled assets. We discuss these potential ex post costs more later. 13 The contractible portion of R&D effort need not be equal to zero. For simplicity, we focus on the non-contractible portion. 7 a new factory, and hiring sales people. We assume that both g > 1 and h > 1 so that costs are convex. The discount rate is r. The base price P b t takes a value in the set {P 0 , P 1 , . . . P N }, with P s < P s+1 (0 ≤ s ≤ N − 1) and P s+1 − P s < P s − P s−1 (0 ≤ s ≤ N − 1). A success in R&D at time t leads to new features and product enhancements. These product enhancements result in a legally enforceable patent, and boost the base price from P s to P s+1 (0 ≤ s ≤ N − 1). Additional product features and extensions have a positive but decreasing effect on prices. We use X t to denote the result of R&D investment which is realized and observed by both parties at the end of time period t, such that X t = 1 corresponds to a success and X t = 0 to a failure. The probability of success is determined by the R&D investments p(X t = 1) = x t . For simplicity, we assume that the increase in price resulting from commercialization investment is deterministic, and it increases the base price P b t by an amount y t if the firms are separate, and ρ(y t ) if the firms are integrated. Both the level of price impact and the marginal product of commercialization investments are higher under integration, such that ρ(y t ) > y t and ρ (y t ) > 1. 14 The bargaining power of the upstream supplier is α (and the downstream producer 1 − α) in both the ex-ante acquisition negotiations that result in the integration of the two firms, and ex-post renegotiation for splitting total surplus when firms are separate. The model's timing is summarized in 1. The downstream producer decides whether to acquire the upstream supplier, 14 This assumption can arise from the supplier not cooperating fully (withholding some information or selling related products to other firms) with the downstream firm if separate. We do not model the specific reason for the marginal product of commercialization expenditures being higher under integration. In the end, what is crucial is that the marginal product is higher for some types of expenditures if one firm has full control of the assets which can include a patent that is used in the production process. Clearly this is a crucial assumption but one that is likely to be satisfied for production when timely delivery of components are important and when the quality of the engineers or people involved in the production of the components cannot be perfectly observed. It would also be satisfied in situations when it is difficult to contract on all aspects of product quality as in the recent case of Boeing and other firms reintegrating with some of their suppliers given supply chain problems (See: http://www.industryweek.com/companies-ampexecutives/rebalancing-business-model.) 8 and if so, negotiates with the supplier based on each party's bargaining power. 2. R&D investments x t and commercialization investment y t are decided by both parties as ex-ante investments. 4. By the end of the period, the success of R&D investments is realized, so that at the beginning of next period t + 1, both firms observe the value of X t . The realization of R&D and the grant of a patent is key to determining whether firms will integrate or remain separate. Since we make the assumption that X t is realized at the end of each period, the final price charged on consumers is equal to P t = P b t (1 + y t ) under separation and P t = P b t (1 + ρ(y t )) under integration, with the base price P b t = P N if the last period base price is Note that the base price is a contingent variable given the last-period R&D outcome X t−1 . We model the decision of the producer to acquire the supplier and integrate (I) as a real option that, when exercised, is costly to reverse. We denote I = 1 as the situation where firms are integrated, and I = 0 when firms remain separate. In line with The central prediction of the model, shown as Proposition 1 in Appendix 1 is that R&D expenditures are higher when the firms are separate, while commercialization and product integration expenditures are higher when firms are integrated. We 15 We could equivalently consider the case where the upstream firm buys the downstream firm. This would occur if the downstream firm does the R&D and the upstream firm customizes the product features before supplying the product. Note that this is not a crucial assumption. The model can thus be applied in either direction. We focus on the case of the downstream firm buying the upstream firm for simplicity, which is empirically the most frequent case as the previously cited Industry Week article notes. 9 show in Appendix 1 how the integration decision depends on the product price over time. Proposition 2 shows that when P b t = P N (i.e., the maximum price) both firms prefer to be integrated. This result arises because at that price the marginal effect of R&D on the price is zero. 16 Formally, we show that V (P N ) = V (P N ; I = 1) > V (P N ; I = 0) with V denoting the value function that measures total surplus. Moreover, we show as Proposition 3 in Appendix 1 that when P b t is below P N , there exists a state s * such that V (P s ) = V (P s ; I = 1) ≥ V (P s ; I = 0) for any s ≥ s * , and The state s * is the triggering state for integration. This equilibrium is illustrated in Separation optimally allocates residual rights of control to the party whose incentives are more important (the upstream supplier). In contrast, when the asset is more fully developed and its features are protected by a patent (i.e. higher state s resulting from successful R&D), incentives for further R&D by the supplier (x) decline because of the decreasing marginal effect of R&D on the product price. At that time, the incentives for the downstream producer to spend on commercialization to further boost the product price (y) increases. Yet, without legal control rights on the asset (i.e. ownership of the patent), the producer faces hold up risk from the supplier. To encourage commercialization incentives, it is thus optimal for the overall relationship to allocate the residual rights of control to the downstream producer, whose incentives are more important. Hence, integration maximizes total surplus. The model thus delivers the following central prediction: Central Prediction: Firms are likely to remain separate when innovation is unrealized and R&D is important. Firms are more likely to be integrated when the innovation is realized and is protected by patents. 16 What is necessary is that marginal product of the non-contractible R&D declines over time such that the gain from R&D is less than the cost of not-integrating and getting the benefits of commercialization. 10 We test this proposition using new text-based measures of vertical relatedness, and by examining the distinct roles played by R&D and patenting intensity. However, we note that varying the assumptions about contractiblity and how the marginal products of innovation and commercialization evolve will give different predictions. Hence the model is mainly provided to illustrate the economic forces that deliver this central prediction. III Data and Methodology We consider multiple data sources: 10-K business descriptions, Input-Output (IO) tables from the Bureau of Economic Analysis (BEA), COMPUSTAT, SDC Platinum for transactions, and data on announcement returns from CRSP. A Data from 10-K Business Descriptions We start with the Compustat sample of firm-years from 1996 to 2008 with sales of at least $1 million and positive assets. We follow the same procedures as Hoberg and Phillips (2015) to identify, extract, and parse 10-K annual firm business descriptions from the SEC Edgar database. We thus require that firms have machine readable filings of the following types on the SEC Edgar database: "10-K," "10-K405," "10-KSB," or "10-KSB40." These 10-Ks are merged with the Compustat database using using the central index key (CIK) mapping to gvkey provided in the WRDS SEC Analytics package. Item 101 of Regulation S-K requires business descriptions to accurately report (and update each year) the significant products firms offer. We thus obtain 74,379 firm-years in the merged Compustat/Edgar universe. B Data from the Input-Output Tables We use both commodity text and numerical data from the Input-Output (IO) tables from the BEA, which account for the dollar flows between all producers and purchasers in the U.S. economy (including households, the government, and foreign buyers of U.S. exports). The tables are based on two primitives: 'commodity' 11 outputs (any good or service) defined by the Commodity IO Code, and producing In addition to the numerical values in the BEA data, we use an often overlooked resource: the 'Detailed Item Output' table, which verbally describes each commodity and its sub-commodities. The BEA also provides the dollar value of each subcommodity's total production and a commodity's total production is the sum of these sub-commodity figures. 18 Each sub-commodity description uses between 1 to 25 distinct words (the average is 8) that summarizes the nature of the good or service provided. 19 [Insert 18 There are 5,459 sub-commodities and 427 commodities in 2002. The average number of subcommodities per commodity is 12, the minimum is 1 and the maximum is 154. 19 For instance, the commodity 'Footwear Manufacturing' (IO Commodity Code #316100) has 15 sub-commodities including those described as 'rubber and plastics footwear' and 'house slippers'. 12 vertical relation such as 'used in', 'made for' or 'sold to'. Second, we remove any expressions that indicate exceptions (e.g, we drop phrases beginning with 'except' or 'excluding'). Third, we discard uninformative common words from commodity vocabularies. 20 Finally, we remove any words that do not frequently co-appear with the other words in the given commodity vocabulary. This further ensures that horizontal links or asset complementarities are not mislabeled as vertical links. We compute the fraction of times each word in a given IO commodity co-appears with other words in the same IO commodity when the given word appears in a 10-K business description (using all 10-Ks from 1997 to avoid any look ahead bias). We then discard words in the bottom tercile by this measure (the broad words). For example, if there are 21 words in an IO commodity description, we would discard 7 of the 21 words. 21 We are left with 7,735 commodity words that identify vertically related product markets. The 'Detailed Item Output' table also provides the economic importance of each commodity word. We compute the relative economic contribution of a given subcommodity (ω) as the dollar value of its production relative to its commodity's total production. Each word in a sub-commodity's textual description is assigned the same ω. Because a word can appear in several sub-commodities, we sum its ω's within a commodity. A given commodity word is economically more important if this fraction is high. We define the commodity-word correspondence matrix (CW ) as a three-column matrix containing: a commodity, a commodity word, and its economic importance. Because the textual description in the Detailed Item Output table relates to commodities (and not industries), we focus on the intensity of vertical relatedness between pairs of commodities. We construct the sparse square matrix V based on the extent to which a given commodity is vertically linked (upstream or downstream) 20 There are 250 such words including accessories, air, attachment, commercial, component. See the Internet Appendix for a full list. 21 This tercile-based approach is based on Hoberg and Phillips 13 to another commodity. From the Make 22 Note that V is sparse (i.e., most commodities are not vertically related) and is non-symmetric as it features downstream (V c,d ) and upstream (V d,c ) directions. Finally, we create an exit correspondence matrix E to account for production that flows out of the U.S. supply chain. To do so, we use the industries that are present in the Use table but not in the Make table ('final users'). E is a one-column matrix containing the fraction of each commodity that flows to these final users. C Text-based Vertical Relatedness We identify vertical relatedness between firms by jointly using the vocabulary in firm 10-Ks and the vocabulary defining the BEA IO commodities. We link each firm in our Compustat/Edgar universe to the IO commodities by computing the similarity between the given firm's business description and the textual description of each BEA commodity. Because vertical relatedness is observed from BEA at the IO commodity level (see description of the matrix V above), we can score every pair of firms i and j based on the extent to which they are upstream or downstream by 22 Alternatively, we consider in unreported tests the maximum between SU P P and CU ST , and also SU P P , or CU ST alone, to define vertical relatedness. Our results are robust. 14 (1) mapping i's and j's text to the subset of IO commodities it provides, and When computing all textual similarities, we limit attention to words that appear in the Hoberg and Phillips (2015) post-processed universe. We also note that we only use text from 10-Ks to identify the product market each firm operates in (vertical links between vocabularies are then identified using BEA data as discussed above). Although uncommon, a firm will sometimes mention its customers or suppliers in its 10-K. For example, a coal manufacturer might mention in passing that its products are "sold to" the steel industry. To ensure that our firm-product market vectors are not contaminated by such vertical links, we remove any mentions of customers and suppliers using 81 phrases listed in the Internet Appendix. 23 Ultimately, we represent both firm vocabularies and the commodity vocabularies from BEA as vectors with length 60,507, which is the number of nouns and proper nouns appearing in 10-K business descriptions in 15 bounded in [0,1], and a value close to one indicates that firm i's product market vocabulary is a close match to IO commodity c's vocabulary. The matrix B thus indicates which IO commodity a given firm's products is most similar to. We then measure the extent to which firm i is upstream relative to firm j: (1) The triple product (B · V · B ) is an M × M matrix of unadjusted upstream-todownstream links between all firms i to firms j. Note that direction is important, and this matrix is not symmetric. Upstream relatedness of i to j is thus the i'th row and j'th column of this matrix. Firm-pairs receiving the highest scores for vertical relatedness are those having vocabulary that maps most strongly to IO commodities that are vertically related according to the matrix V (constructed only using BEA relatedness data), and those having vocabularies that overlap non-trivially with the vocabularies that are present in the IO commodity dictionary according to the matrix B. Thus, firm i is located upstream from firm j when i's business description is strongly associated with commodities that are used to produce other commodities whose description resembles firm j's product description. Downstream relatedness is simply the mirror image of upstream relatedness, DOW N ij = U P ji . By repeating this procedure for every year in our sample (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008), the matrices U P and DOW N provide a time-varying network of vertical links among individual firms. D NAICS-based Vertical Relatedness Given we are proposing a new way to compute vertical relatedness between firms, we compare the properties of our text-based vertical network to those of the NAICSbased measure used in previous research, which we describe now. One critical difference is that the NAICS-based vertical network is computed using the BEA industry space, and not the BEA commodity space. This is by necessity because the links to NAICS are at the level of BEA industries. Avoiding the need to link to BEA industries is one advantage of the textual vertical network. More generally, the com-16 pounding of imperfections in BEA industries and NAICS industries may result in horizontal contaminations, especially when firms are in markets that do not cleanly map to NAICS industries. In particular, the Census Department states "NAICS was developed to classify units according to their production function. NAICS results in industries that group units undertaking similar activities using similar resources but does not necessarily group all similar products or outputs." To compute the NAICS-based network, we use methods that parallel those discussed above for the BEA commodity space (matrix V ), but we focus on the BEA industry space and construct an analogous matrix Z. We first compute the BEA industry matrix IF LOW as SHARE × U SE, which is the dollar flow from industry i to industry j. We then obtain ISU P P and ICU ST by dividing IF LOW by the total production of industry j and i respectively (using parallel notation as was used to describe the construction of V ). The matrix Z is simply the average between ICU ST and ISU P P . Following common practice in the literature (see for example Fan and Goyal For simplicity, we thus label the two resulting vertical networks as 'NAICS-1%' and 'NAICS-10%', respectively. To ensure our textual networks are comparable, we choose two analogous textual granularity levels: 10% and 1%. These two text-based vertical networks define firm pairs as vertically related when they are among the top 10% and top 1% most vertically related firm-pairs using the textual scores. We label these networks as 'Vertical Text-10%' and 'Vertical Text-1%'. Note that the textual networks generate a set of vertically related peers that is customized to each firm's unique product offerings. 17 These firm level links provide considerably more information than is possible using broad industry links such as those based on NAICS. E Vertical Network Statistics We compare the properties of five key relatedness networks: Vertical Text-10%, Vertical Text-1%, NAICS-10%, NAICS-1%, and the TNIC-3 network developed by [Insert The first row of Panel A in As theories of vertical relatedness and integration often focus on non-financial concepts such as relationship-specific investment and ownership of assets, these results support the use of the text-based network as being more relevant. Although we do not report full details here to conserve space, we conduct two validation tests in the Internet Appendix to this paper. The goal is to compare the ability of the text-based and NAICS-based vertical networks to identify actual instances of vertical relatedness from orthogonal data sources. In the first test, we search all firm 10-Ks to identify direct verbal statements indicating the firm is vertically integrated. We find that the text-based network is roughly four times stronger in predicting these direct 10-K statements than is the NAICS-based network IV Vertical Acquisitions We now use our text-based vertical network to examine innovation activities and vertical organization. We start by studying vertical acquisitions, as these transactions represent a direct way firms can alter their boundaries and modify their degree of integration. To test our main hypothesis and theoretical predictions in the literature (e.g., Grossman and Hart (1986)), we concentrate on targets (the sellers of assets) as they are the party that loses control rights due to the transaction, and for which the trade-off between ex ante investment incentives and ex post hold-up should be important. We thus examine how R&D and patenting intensity are related to the 19 likelihood of being a target in vertical and non-vertical transactions. Comparing vertical to non-vertical transactions is important, as the hypothesized issues of ex-ante incentives, contracting frictions, and potential ex-post hold up are most salient for vertically related firms and our theoretical predictions do not extend to other transactions such as horizontally related acquisitions. 24 A Transactions Sample [Insert 20 have similar granularity levels, it is perhaps surprising that the networks disagree sharply regarding the fraction of transactions that are vertically related. For our primary sample excluding financials, we observe that 39% are vertically related using the Vertical Text-10% network. Using the NAICS-10% network, we observe that just 13% are vertically related. For any network with a granularity of 10%, if transactions are random, we expect to see 10% of transactions belonging to this network. The fact that we find 39% is strong evidence that many transactions occur between vertically related parties. The results also suggest that the accumulated noise associated with NAICS greatly reduces the ability to identify vertically related transactions. We also note that with both networks, vertical deals are almost evenly split between upstream and downstream transactions. We present these results mainly to compare with previous research (based on either SIC or NAICS codes). Confirming existing evidence, the combined returns across all transactions are positive and range from 0.49% to 0.94%. Notably, when vertical transactions are identified using our text-based measure, the combined returns are larger for vertical relative to non-vertical transactions. This supports the idea that vertical deals are value-creating on average as in B Profile of Targets in Vertical Transactions Table IV presents the R&D and patenting profile of targets in vertical and nonvertical deals. We focus on all transactions and we use our text-based network (10%) to identify vertical deals. We consider both industry-(i.e. TNIC-3) and 25 We also find that transactions classified as vertical are followed by an increase in our firmlevel measure of vertical integration (V I). Using the Vertical Text-10% network, acquirers in vertical transactions experience an increase of 6% in V I from one year prior to one year after the acquisition. In contrast, acquirers in non-vertical transactions experience a decrease of 0.70% in V I. When we use the NAICS-10% network to identify vertical mergers, vertical acquirers see a negligible increase of 0.30% in V I. 21 firm-level measures of R&D and patenting activity. We measure R&D intensity as R&D divided by sales, and patenting intensity as the number of patents divided by assets. We describe all variables used in the paper and display summary statistics in Appendix 2. In Panel A, we observe a large difference between targets in vertical and non-vertical deals. When compared to firms that never participate in any acquisitions over the sample period (labeled as non-merging firms), vertical targets exhibit lower levels of R&D and hold more patents. In contrast, targets in nonvertical deals are more R&D intensive with lower patenting intensity. [Insert The results in [Insert C Multivariate Analysis We complement the above univariate tests by estimating probit regressions to examine how R&D and patenting intensity affect the likelihood of becoming a target. The dependent variable is an indicator variable indicating whether a given firm is a target in a vertical or a non-vertical transaction, as noted in the column headers, in a given year. We consider our text-based network when identifying which transactions are vertically related (Vertical Text-10%). Our sample covers the period 1996-2008 and excludes regulated utilities and financial firms. We further require observations to have non-missing values for each variable we use in the estimations. We have 45,198 firm-year observations corresponding to 6,924 distinct firms. For the explanatory variables of interest (in particular R&D and patent intensity), we consider equally-weighted averages across TNIC-3 industries instead of own-firm variables. We note, however, that our conclusion is qualitatively unchanged if we use own-firm R&D and patent variables instead of industry-level variables (see Internet Appendix IA.IV.1). This choice is driven by two considerations. First, focusing on industry lessens endogeneity concerns for both vertical and non-vertical transactions (see Acemoglu, Aghion, Griffith, and Zilbotti (2010)). Indeed, while a firm directly chooses its own degree of vertical integration, it has little choice regarding its industry's level of R&D or patenting activities. Second, the theoretical incentives to vertically integrate should be driven mostly by the characteristics of product markets, which is best captured using industry variables. For instance, as 23 in [Insert (2013) who model and provide evidence that small target firms conduct more R&D when they have a high probability of selling out to larger horizontally related firms. We next focus on the level of patenting activity. Consistent with our hypothesis that ex post successful innovation indicates maturity and lowers the returns from separate R&D investment, we find that vertical targets are more likely to be in high patenting industries. The opposite is true for non-vertical acquisitions, where firms in high patenting industries are less likely to be acquired. The coefficient on this interaction is only significant for non-vertical transactions, which further confirms that these transactions are different. The positive coefficient 26 Note that we cluster standard errors at the industry (using the FIC data from Hoberg and Phillips 27 In our sample, R&D and patenting activity are not perfectly correlated. This correlation is 0.33 across firms, and 0.58 across industries. 24 for industry patenting activity also remains robust for vertical acquisitions. 28 Table V also supports our hypothesis that maturity is an important positive determinant of vertical transactions. For instance, column (1) indicates that firms with lower market-to-book ratios and older firms are more likely to be targets of vertical deals. In contrast, targets in non-vertical deals are more likely to be young and are in less capital intensive industries. These findings are consistent with the following interpretation of the U-shaped relationship between firm maturity and restructuring activity noted in Although our main focus is on transaction targets, we also examine the link between R&D and patenting intensity and the likelihood of becoming a vertical or non-vertical acquirer. We report the results in the Internet Appendix for brevity D State R&D Tax Credit as Instrument Our results so far reveal significant associations between the propensity to be purchased by vertically-related acquirers and the R&D and patenting intensity of the 28 In additional tests that we present in the Internet Appendix for brevity, we show that the results hold when we use lagged values of the independent variables Fully addressing this concern would require instruments that separately affect industry R&D and patenting intensity, without influencing a given firm's propensity to be acquired through other channels. Although such an ideal instrument is not available, we follow Bloom, Schankerman, and van Reenen (2013a) and use taxinduced changes to the user cost of R&D to construct an instrument for industry R&D intensity. State R&D tax credits offer firms credits against state income tax liability based on the amount of qualified research done within the state. 29 In practice, different states have different levels of R&D tax credits, and hence the user cost of R&D is dependent on firm locations and time. As discussed in Bloom, Schankerman, and van Reenen (2013a), the existing literature suggests that the introduction and level of R&D tax credits is quite random.