Firm Boundaries and Delegation Research Chapter 1

1 Introduction Why do firms integrate suppliers? One benefit, of course, is direct control of production: in a world replete with contracting frictions, ownership can facilitate top management’s ability to impose productivity-enhancing decisions, such as re-tooling, human capital investment, or standards conformity, that its suppliers otherwise would not make. Equally important, if less appreciated, integration confers greater control over the firm’s internal organization. Among the residual decision rights bundled with owning an asset is the ability to re-assign its use or control to others. Over the course of a lengthy and uncertain production process, different types of problems are bound to arise, and top managers may wish to re-allocate decision rights among themselves and their suppliers to solve them according to their relative expertise. Within a firm’s boundaries, management can do this relatively seamlessly, choosing to delegate production decisions to its integrated suppliers or to centralize those decisions, depending on which problem arises. This option is hardly available outside the firm, where suppliers retain those control-allocation rights among their prerogatives of ownership, and are less likely to exercise them in management’s interests. Firm boundaries and the allocation of decision-making inside the firm are thus intrinsically linked. The “control over control” that comes with ownership helps guarantee the firm a minimum quality and quantity of inputs, and thereby introduces a novel mechanism of supply assurance as a rationale for integration: the advantage of ownership is not so much that it can be used to force a supplier to provide an input that he might otherwise sell to someone else, but that it allows the firm to deploy control to the party best suited to using it. Despite their evident connection, the interplay between vertical integration and delegation has scarcely been explored. This paper brings these organizational design decisions together, both theoretically and empirically. It first develops a model to jointly study vertical integration and delegation. It then assesses the evidence in light of the model, combining information on vertical integration and delegation for firms in multiple countries and industries. Our analysis is founded on a well-known conceptual distinction between outsourcing (non-integration) and delegation. While both are instruments for re-assigning control rights, the first is defined by formal titles of ownership and requires legal intervention to reverse, via an asset purchase. By contrast, delegation is a non-contractible act of relinquishing control that can in principle be revoked at will by owner fiat.1 One contribution of this paper is to make this distinction operational by generating testable predictions about the differential 1 See for instance Williamson (1991) and Aghion and Tirole (1997). Indeed, non-integration has legal force, while delegation does not. The law not only regulates and registers asset sales, it frequently adjudicates disputes between parties who hold separate titles. Once integrated, the parties largely forego appeal to the law in many of their disputes. 1 responses of delegation and outsourcing to changes in the firm’s environment. In our theoretical model, the primary organizational design problem is the allocation of control over production decisions. We consider a world in which such allocations are not contractible, but are among the prerogatives of ownership. A headquarters (HQ) and one of its input suppliers decide whether to integrate before starting production and learning which of many potential “problems” — new designs, quality control or compatibility issues, supply chain disruptions, new hiring or capital equipment requirements — will have to be tackled. In the course of addressing a problem, a non-contractible production decision, modeled as a choice among horizontally differentiated standards, will also have to be made. All decisions work equally well for generating the input, but preferences over them diverge between the parties, possibly due to differences in training, background, corporate culture, or managerial vision.2 These private costs are the source of interest conflict in the model. Solving problems and choosing standards cannot be separated: both are in the hands of whoever controls the production process. HQ and the supplier differ along two dimensions beyond their conflicting preferences over production standards. First, distinct problems play differently to their comparative advantages: some will be easier to solve for the supplier, others for HQ. Second, contracting frictions imply that the supplier has smaller residual stakes in the enterprise profit than HQ. Under integration, HQ owns the primary asset the supplier uses. After a problem arises, she can therefore decide whether to retain control (centralize), solving the problem and implementing the production decision herself; or to offer control to the supplier (delegate), who then chooses the production standard and addresses the problem. HQ’s decision to delegate comes down to trading off a supplier’s sometimes superior problem solving ability against the private costs he will impose on her through his own choice of production standard. In so doing, she does not internalize the supplier’s private cost incurred because of her own choice of production standard. By contrast, with non-integration, the supplier owns the asset, and therefore the option to delegate to HQ. However, given his small residual stakes, and his own neglect of HQ’s private costs, he never has an interest in doing so. When the integration decision is made before production begins, the parties are deciding between two imperfect control allocation mechanisms: one that eventually offers the supplier too little control (integration), and one that offers him too much (non-integration). The model delivers several testable predictions about input value and uncertainty as ex2 For instance, HQ favors redundant features and user-friendliness in the software bundled with the game consoles it sells, while the software engineers prefer elegant design and low maintenance. Or each party bristles at the personality types of the otherwise capable candidates it finds to fill a post. This “standards” approach to modeling tensions about the way operations should be carried out within firms has been employed by Van den Steen (2005, 2010) and Hart and Holmström (2010), among others. 2 ogenous determinants of integration and delegation decisions. First, the greater the value contribution of an input to the enterprise, the more likely HQ should be to delegate to the integrated supplier who provides it: with greater value, there is a larger opportunity cost for HQ to indulge in the private benefit of keeping control and providing a mediocre problem solution rather than letting the supplier provide a superior one. Second, a supplier of a more valuable input should be more likely to be integrated. This result derives from the difference in the parties’ delegation behavior: integration becomes relatively more efficient as value increases, because HQ is more likely to allocate control to the party with the better problem solution; by contrast, since a non-integrated supplier never delegates, his relative efficiency never increases. Third, first-order shifts in the supplier’s problem-solving ability should lead to more delegation. Finally, under some empirically mild conditions, HQ should be more likely to integrate inputs whose production carries more problem risk. The logic for this result is similar to that concerning the role of uncertainty in option theory (Dixit and Pindyck, 1994), since the flexibility to reassign control in the face of new information turns integration into a real option. To assess the evidence, we combine information from the WorldBase dataset by Dun & Bradstreet and the World Management Survey (WMS), which allows us to measure integration and delegation choices for firms in 20 countries. A first look at these data reveals a positive correlation between delegation and integration: more vertically integrated firms tend to grant more autonomy to their plant managers.3 This result underscores the conceptual distinction between delegation and outsourcing (non-integration): while both involve shifting control from an HQ to a supplier, they are clearly not interchangeable, since they are negatively correlated in the data. In our empirical analysis, we assess the validity of the predictions generated by the model regarding the effects of input value and uncertainty. We measure vertical integration at the firm-input level, coding which inputs the firm integrates within its boundaries. This reflects the choices faced by firms in our theoretical model, in which HQ decides whether or not to integrate each input supplier. Our delegation measure captures the degree of autonomy granted by a firm to the managers of its integrated suppliers when faced with various production decisions (e.g., hiring a new employee, introducing a new product, making a capital investment). To test the model’s predictions about the role of input value, we exploit exogenous vari3 See Section A-2.2 of the Empirical Appendix. A simple application of our theoretical model’s results offer a rationale for the positive correlation between vertical integration and autonomy: both could be driven by a common (unobserved) driver in the form of enterprise value (e.g., HQ productivity or final good price). Higher value enterprises should integrate more inputs and grant more authority to their integrated suppliers. 3 ation in input-output (IO) coefficients using data from the US Bureau of Economic Analysis (BEA). IO coefficients capture exactly the role of input value in our model, since they measure the dollar value of each input used in the production of an output. To capture exogenous variation in uncertainty, we use the mean and dispersion of the distribution of productivity of independent suppliers in an industry. The evidence confirms the role of input value for integration and delegation, in line with the predictions of our model. We find that final good producers are indeed more likely to integrate suppliers of more valuable inputs, as proxied by input-output coefficients. Among integrated suppliers, more autonomy is granted to those producing more valuable inputs. Also consistent with the theoretical model, we find that delegation increases with the mean productivity in the input industry, while riskiness of the input industry has no significant effect on delegation. Moreover, the probability that firms integrate a particular input increases with the riskiness of the input industry. This finding is reminiscent of the literature on supply assurance motives for integration, as discussed in the next section. The empirical results of our baseline regressions hold up in a battery of robustness checks, such as restricting the analysis to different samples (e.g., only multi-plant firms), including input-industry and output-industry fixed effects, and additional controls (e.g., firm and plant characteristics, as well as measures of contracting frictions). In the integration regressions, we can also include firm fixed effects, exploiting variation within firms across input industries (in the riskiness of the industry or the value of the input) to identify the role of input value and uncertainty. The structure of the paper is as follows. Section 2 briefly reviews the related literature. Section 3 presents the theoretical model. Section 4 describes the data and variables used in the empirical analysis. Section 5 presents the empirical results. Section 6 discusses alternative theoretical explanations for these results. Section 7 concludes. 2 Related Literature Our work is related to several streams of literature. First, we build on the vast literature on the determinants of firm boundaries. Many theoretical studies have looked at the technological/contractual determinants of vertical integration (e.g., Coase, 1937; Grossman and Hart, 1986; Hart and Moore, 1990; Holmström and Milgrom, 1991; Hart and Holmström, 2010); another strand has focused on market determinants (e.g., McLaren, 2000; Grossman and Helpman, 2002; Legros and Newman, 2013; Conconi et al., 2012). A number of papers study legal/institutional determinants of integration (e.g., Acemoglu et al., 2009, Macchi- 4 avello, 2012). The view of integration in our model is similar to that of Williamson (1975), and puts it in the “ex-post non-contractible” branch of incomplete-contracts economics (e.g., the 2002 version of Hart and Holmström, 2010; Aghion et al., 2002). The empirical literature on firm boundaries is also very large, and growing; Lafontaine and Slade (2007, 2013) provide excellent overviews. Our empirical analysis of ownership is closest to some recent papers emphasizing supply assurance (e.g., Macchiavello and MiquelFlorensa, 2017) and the relative technological importance of inputs (e.g., Berlingieri et al., 2021; Hansman et al., 2020) as motives for integration. We focus on backward integration, i.e., producers’ decisions to integrate input suppliers. This is in line with most of the empirical literature on vertical integration, including Acemoglu et al. (2009 and 2010), Alfaro et al. (2016 and 2019), and many other papers surveyed in Lafontaine and Slade (2007 and 2013). Second, our paper contributes to the literature on delegation within firms. Theoretical studies in this literature include, among others, Holmström (1984), Aghion and Tirole (1997), Dessein (2002), Hart and Moore (2005), Alonso et al. (2008), Alonso and Matouschek (2008), and Marin and Verdier (2008). 4 Our approach to modeling delegation is related to work on the design of knowledge hierarchies (Garicano, 2000) and referrals (Garicano and Santos, 2004) insofar as we are concerned with allocation of decision making among the organization’s members according to the expertise at solving particular production problems.5 On the empirical side, contributions include Acemoglu et al. (2007), Guadalupe and Wulf (2010), Bloom et al. (2012), McElheran (2014), Graham et al. (2015), Wu (2017), and Katayama et al. (2018). In this paper, we bring these organizational choices together. A number of papers have studied pairwise interactions of organizational design elements from the theoretical point of view. Examples include Holmström and Tirole (1991), Holmström and Milgrom (1991, 1994), and Friebel and Raith (2010). Although some formal studies have emphasized the conceptual difference between integration and delegation (Baker et al., 1999; Hart and Holmström, 2010), there has been little theoretical work to operationalize these differences. And, to the best of our knowledge, there is no systematic empirical work along those lines. 4 Much (but by no means all) of this literature views delegation as a means of achieving better outcomes by assigning decision rights to (ex-ante) better informed parties; often this helps to incentivize delegates to become more informed in the first place. In our simplified model of delegation, the assignment of control is instead a response to (symmetric) information: the (ex-post) sufficiently more capable (or possibly less time constrained) party gets it. The two approaches are complementary — the production decisions could involve the acquisition of further information — and our approach is mainly for tractability. 5 In those papers the allocation of control is decided contingently through contracts rather than managerial authority, and they abstract from incentive problems, which play a key role in the comparative statics of our model. Moreover, they endogenize knowledge acquisition, while we take it as given. A more minor distinction is that those papers treat the quality of problem solutions as discrete rather than continuous. 5 Our paper also contributes to the literature on supply assurance motives for integration (e.g., Carlton, 1979; Bolton and Whinston, 1993). Those papers tend to focus on demand uncertainty and the ability of non-integrated suppliers to sell their inputs to other buyers. The supply assurance in our model derives from uncertainty about problems that may arise during the production process and the ability of the firm’s members to solve them. Our model refines the “value principle” that emphasizes the role of pecuniary variables such as profitability and prices in organizational design, and extends it beyond the integration context where it has already been applied (Legros and Newman, 2013, 2017; Alfaro et al., 2016). While option value has been considered in some studies of integration (e.g., Bradley et al., 1988; Grullon et al., 2012; Bena and Li, 2014), our analysis provides a novel mechanism, based on the notion of asset ownership as a bundle of control rights, rather than just a claim on return streams. Moreover, input value affects not only firm boundaries, but also the internal control structure, through its effect on centralization/delegation decisions. 3 The Model We present a simple theoretical model in which some modeling assumptions are aimed at delivering predictions that we can test with our data rather than at theoretical generality. The interested reader is referred to Legros and Newman (2022) for a broader treatment. 3.1 Environment Consider a production process in which a final good j is produced with I inputs indexed by i. An enterprise is composed of an HQ, who produces the final good, and I suppliers Si , each comprising a manager (who will also be referred to as a supplier) and an indivisible productive asset that can be owned by the supplying manager or by HQ. For empirical purposes, we shall think of each supplier Si as representative of its industry; there may be differences across industries. All players are risk neutral. The expected value of a good produced by the enterprise can be written as Af Pj I X πi,j EVi,j . (1) i=1 Af represent the exogenous productivity of the HQ of the enterprise, capturing her entrepreneurial competence or the profitability of her product. Pj is the price of the final good and πi,j EVi,j is the contribution of each supplier i. The latter is decomposable into an exoge6 nous, technologically dictated value share πi,j and an endogenous quantity EVi,j that will be shaped by production and organizational decisions made by HQ and supplier Si . For now, we focus on the relationship between HQ and a typical supplier, suppressing the index notation and normalizing Af = Pj = 1. Discussion of the effects of these parameters will be taken up in Section 3.2.2. There are two types of uncertainty in this model. The first is the familiar quantity or quality risk, and related information asymmetries, that cloud inference about underlying decision variables, making them non-contractible. The second is more specific to the organizational design concerns in this paper, which is uncertainty over which specific problems will arise in the course of production — and therefore who is best suited to to tackle them — that drives the delegation decision and is the source of the option value of integration. 3.1.1 Technology, Timing and Information We model the production process in the simplest possible way: there is a binary decision d ∈ {h, s} over how to use the asset that can be carried out by either party, depending on the assignment of control. The choices h, s are equally effective from the viewpoint of production efficiency, but impose different private costs on the two parties, e.g., resulting from different corporate cultures or different preferences over design standards. Choice h is HQ’s preferred option, costing her zero, but imposing a positive cost cS per unit of produced input on the supplier; similarly, option s costs the supplier nothing but imposes cost cH > 0 per unit on HQ. To ensure gains from trade, the cost parameters satisfy π > max{cS , cH } (since control is ultimately allocated to one party or the other, only one will bear a cost). The sequence of events is as follows. First S and HQ agree on who owns the primary productive asset S uses, and on a (positive or negative) side-payment T from HQ to S to ensure participation. At this point, as is standard in models of ownership, the two parties are “locked in” to each other for the duration, because some sunk investment (physical, human capital, enduring the legal process of registering title, or simply finding each other) has to be made to get production going. After production commences, a “problem” (or opportunity) arises; this is represented by the random variable y ≥ 0 with c.d.f G(y) and density g(y). The problem can be solved by either party, but to do so, the party needs to be given control of the production decision d. At this point, the parties agree on a payment t(y) from HQ to S, due on receipt of the input batch, which may depend on the problem that arose. Next, the owner of the asset decides who will solve the problem/make the production decision, and the assignee then chooses d. Finally, the inputs are delivered, HQ privately observes the number of usable units in the batch, and appropriates the realized value less the payment. 7 How well the problem is solved depends not only on the problem itself, but also on the party that solves it: the efficacy of the solution under HQ-control is a constant z > 0, while under S-control it is y (thus, we label problems by the competence with which the supplier would solve them). Efficacy here is measured in terms of the expected number of usable inputs that will ultimately be produced; the actual number depends partly on random factors outside the parties’ control, and only HQ ultimately observes how many units are usable.6 HQ is a generalist, equally adept at addressing all problems (hence the constant efficacy z; this can be relaxed) The supplier is a specialist, better than HQ at solving some of problems, worse at others (0 < G(z) < 1). It is convenient, though not strictly necessary, that HQ is on average better at solving problems than the supplier: Ey ≤ z. (A1) HQ’s competence z, as well as the supplier’s effectiveness distribution G(y) are common knowledge: each party knows exactly how well the other would perform in any given situation y; there is only uncertainty as to which situation will arise. 3.1.2 Contracting Contracting is limited to the transfers of ownership and side payments to facilitate those, and to further monetary transfers payable upon delivery of a batch of inputs. What is not contractible are the decision d (making control relevant), and the number of inputs that HQ finds usable (thereby eliminating the supplier’s stake in the continuation value of the relationship). The identity of the actual decision makers/problem solvers is also not contractible. This ensures that allocating control must be informal, determined by the owner, who could always overturn a decision that was supposed to be under the control of the other party, yet deny that in court. Even if the court knows the state y (or infers it from the payment t(y)), that is not enough to establish who made a decision. Finally, as HQ’s aggregate profit depends on many different inputs, profit sharing would add little in terms of meaningful incentives, so we abstract from those. For empirical purposes, we take the payment to be equal to the market price of the input p multiplied by the expected quantity. Thus, t(y) is equal to py if the anticipated party in control is the supplier, and pz if it is HQ. When negotiating ownership, the supplier and HQ take into account the private costs 6 The private observation prevents the payment from depending on realized rather than expected quantity; thus only HQ has an interest in the continuation value once t(y) is negotiated. Private observation could be interpreted as HQ’s subjective assessment of quality, or the compatibility of the input with HQ’s production technology or with the inputs of her other suppliers; none of these are likely to be contractible – if a part looks wrong, or doesn’t work, it would be too costly for a court to determine fault and compensation. 8 that will ensue. We assume that both parties have sufficient cash at the time of ownership contracting to make any side payments needed to settle the distribution of surplus required to strike a deal. Thus we will be considering “efficient” ownership choices in what follows. Note that in contrast to ownership, the availability of side payments does not imply that an efficient control allocation can be achieved, because of the non-verifiability of the decision maker’s identity. By way of summary, the sequence of events is illustrated in Figure 1. Problem y Integration decision, transfer T Delivery of input batch Delegation decision Negotiate delivery payment t(y) Choose d and solve problem y HQ pays t(y), observes usable inputs Figure 1: Timing of Events 3.1.3 A Benchmark: Efficient Allocation of Control A surplus maximizing control allocation would have S deciding whenever the total surplus is greater under S control than under HQ control: y(π − cH ) ≥ z(π − cS ), or π − cS . (2) π − cH In a world where the identity of the deciding party is contractible, ownership and delegation are irrelevant, at least as far as the control allocation is concerned: the optimal allocation can be achieved by contract, and need not be relegated to those blunt mechanisms. A contract could directly implement a control allocation by specifying who decides as a function of y, independent of the compensation structure or of who owns the asset. 7 y ≥ y ∗ (π) := z Having ownership determined after y is realized would also yield efficient control: if y ≥ y ∗ , let S own; if y ≤ y ∗ , let HQ own. We have already ruled this out, as it may forego many of the (here un-modeled) benefits of ownership that pervade the literature, specifically providing incentives for (or direct implementation of) sunk relationship-specific investments. 7 9 3.2 Input Value and Organizational Design We now return to the world in which control is not contractible, and is instead assigned by the owner of the asset. First we consider delegation under each ownership structure, and then, using results of that analysis, discuss the determination of ownership. 3.2.1 Delegation Under Integration and Non-integration Because HQ privately observes how many inputs actually are usable (generate the positive value π), she is the residual claimant of value. By contrast, S receives a transfer that cannot depend on realized value, only on its expectation y. It follows that S will never delegate when he owns the asset (i.e. under non-integration): centralizing pays him t(y), while delegating yields t(y) − zcS . On the other hand, if HQ owns the asset (i.e. under integration), she prefers to delegate when y(π − cH ) ≥ zπ, in other words when π . (3) y ≥ y H (π) := z π − cH Note that like S, HQ solves problems too often compared to the efficient control allocation, since y H (π) > y ∗ (π). But y H (π) is decreasing in π, so the probability 1 − G(y H (π)) that HQ delegates increases with π. As the value of trade increases, its weight in HQ’s calculations grows vis-à-vis her private costs, and she is more willing to delegate, approaching the efficient control allocation as π grows large. The literature has frequently cited consonance in parties’ preferences as a critical determinant of delegation (e.g., Aghion and Tirole, 1997; Dessein, 2002). Here, higher transaction values endogenously increase consonance. We can thus state the following: Result 1. (Value principle for delegation): HQ’s propensity to delegate increases with π. Observe that in contrast to this strong prediction about the relationship between supplier control and input value under integration, efficient control allocations need not display such a property. Instead, the response would be sensitive to the private cost parameters: from (2), efficient supplier control increases with π if cS < cH , and decreases with π if cS > cH . 3.2.2 Comparing Ownership Structures Since S always retains control when he owns the asset, the total expected surplus under nonintegration (i.e. supplier ownership) is simply W S (π) := Ey(π − cH ). 10 (4) The total expected surplus under integration (HQ ownership) is H H S H Z ∞ W (π) := G(y (π))z(π − c ) + (π − c ) yg(y)dy. (5) y H (π) That is, G(y H (π)) of the time HQ centralizes, choosing d = h, and generating the expected usable output z, which has net per-unit value (π − cS ). The rest of the time, she bears the cost cH in proportion to the number of usable units expected under S-control, where d = s. Assuming efficient allocation of ownership, integration occurs whenever W H (π)−W S (π) is non-negative, which can be written Z y H (π) [z(π − cS ) − y(π − cH )]g(y)dy ≥ 0. (6) 0 As π is embedded in this condition, it is natural to ask what the value principle implies for the propensity of HQ to integrate her suppliers (Legros and Newman, 2013; Alfaro et al., 2016).8 Supposing then that a supplier whose input is worth π is integrated, what can be said about a supplier with a more valuable input worth π ′ ? Notice that the left hand side of expression (6) is increasing in π. 9 Thus, the value difference W H (π) − W S (π) has the single-crossing property, so we can state the following monotonicity result: Result 2. (Value principle for integration): If a supplier who generates value π is integrated, so is a supplier who generates value π ′ > π. The result stems from the differences in HQ’s and S’s delegation behavior under their respective ownership. Compared to S-ownership, HQ-ownership becomes relatively more efficient zcS at allocating control as π increases: y H (π) − y ∗ (π) = π−c H decreases with π; meanwhile S always chooses to retain control, so the corresponding efficiency gain is smaller or negative.10 Looking back at expression (1) for the expected value of an enterprise, Results 1 and 2 imply that the suppliers with higher input value are the ones most likely to be integrated and, among those that are, to have higher degrees of delegation. 8 We cannot simply lean on those papers’ results because the mechanisms that allow parties to benefit from integration there differ from the control-over-control mechanism at work here. R yH 9 Its π-derivative is g(y H )yπH z(π − cS ) + G(y H )z − y H g(y H )yπH (π − cH ) − 0 yg(y)dy , which using R yH y H = zπ/(π − cH ), reduces to [zG(y H ) − 0 yg(y)dy] − g(y H )yπH zcS . From (A1), the bracketed term is R yH non-negative (z ≥ Ey ≥ 0 yg(y)dy/G(y H )), and the last term is positive because the derivative yπH < 0. 10 S ownership may become absolutely more efficient with increasing π through its effect on y ∗ (π), but only if cS < cH , in which case, under (A1), HQ integrates for all π > cS anyway. Dispensing with (A1) may instead generate a U-shaped relation between value and integration, but only if cH is larger than cS and the supplier’s average competence is sufficiently higher than HQ’s. 11 Other sources of variation in enterprise value would have similar effects as input value. In particular, we could allow value to vary across firms within an industry, e.g., if their HQs have different productivity (Af in equation (1), which we have so far set equal to 1). Then Af π would replace π in all the formal expressions above. The effects of variation in Af on integration and delegation choices would be identical to those of input value: both the propensity to integrate suppliers within the firm boundaries and the propensity to delegate decisions to integrated suppliers would increase in HQ’s productivity. Similarly, if the output price Pj varies across industries, firms in high-value industries would have greater propensities than firms in low-value industries both to integrate their suppliers and to delegate to them.11 These simple extensions of the model imply that there should be positive covariation between the degree to which firms are vertically integrated and the extent to which they delegate, in line with what we observe in our data (see Figure A-7 in the Appendix): more valuable firms (e.g., those that have more productive HQs) should integrate more suppliers and grant more autonomy to them. Taken together, Results 1 and 2 help to underscore and operationalize the conceptual distinction between the informal organizational decision to delegate and the formal one of outsourcing (non-integration). Although both move decision-making away from the “center,” the degree of delegation increases, while the propensity to outsource decreases, in response to changes in the same variable, namely input (or enterprise) value. 3.3 Uncertainty and Organizational Design As has been noted elsewhere (Hart and Holmström, 2010), and as the present model illustrates, uncertainty plays an essential role for delegation as well as for ownership. In particular, if the parties could perfectly anticipate y at the time of the integration decision, there would never be a strict incentive to integrate and subsequently delegate: either y ≤ y ∗ (π), in which case there will be integration with centralization, or y > y ∗ (π), in which case non-integration performs at least as well as integration (strictly so if y < y H , for then HQ would inefficiently centralize, or if the acquisition of the supplier’s asset has even infinitesimal cost). Contrary to the data, no variation in autonomy among integrated suppliers would be observed, as no HQ would delegate at all. As we have already discussed, the the value of owning assets is affected by the option to centralize, which provides a hedge against risk in y. Without that uncertainty, the sole role of ownership would be to enforce the efficient assignment of control, as in footnote 7, by integrating if and only if y ≤ y ∗ (π). In that case, against the strong prediction in Result 2, 11 Evidence on the integration effect can be found in Alfaro et al. (2016), and McGowan (2017). 12 integration would increase with π only if cS > cH . To be more precise about how uncertainty affects delegation and integration, we consider the comparative statics of (first-order) stochastic dominance and of (Rothschild-Stiglitz) risk in the distribution G(y).12 3.3.1 Delegation The degree of delegation is 1−G(y H (π)), the probability that y exceeds the threshold y H (π). From the definition of stochastic dominance the following is immediate as well as intuitive: Result 3. Stochastic increases in the supplier productivity distribution G(y) lead to a higher probability of delegation. When problems are more likely to land in the supplier’s sphere of competence, it is more likely that he will be the one asked to solve them. However, because the cumulative probability at any given interior y value — y H (π) in particular — can increase or decrease with increases in risk (risk-comparable c.d.f.’s cross at least once), there is no general prediction about the effects of risk on delegation: increasing risk tends to fatten both upper and lower tails, meaning only that there is a reduced likelihood that problems that crop up will be the ones in the supplier’s “domain of mediocrity.” Likely as not, they will play more to his strengths as to his weaknesses, yielding an ambiguous change in HQ’s willingness to delegate. 3.3.2 Integration It is illuminating to rewrite inequality (6) as the following “put-option” condition: E max{πz − (π − cH )y, 0} ≥ cS zG(y H (π)). The integrand max{πz − (π − cH )y, 0} is decreasing and convex in y. So the relative benefit of integration over non-integration (value of the option) decreases with stochastic increases in G(y): the more likely the supplier will be the one better able to solve the problem that arises, the less valuable is the option to centralize, and with that, the lower is the value of integration. Unlike in standard real-options theory, however, the cost of acquiring the asset, which is the supplier’s expected private cost, is systematically tied to the underlying uncertainty, for it 12 A distribution Ĝ is stochastically larger than G if Ĝ(y) ≤ G(y) everywhere, Rwith strict inequality on a Rx x non-null set. Ĝ is riskier than G if the two distributions have the same mean and 0 Ĝ(y)dy ≥ 0 G(y)dy for all x, with strict inequality on a non-null set; equivalently, Ĝ can be derived from G via a sequence of mean-preserving spreads; or EG u(y) ≥ EĜ u(y) for all concave functions u(·). 13 depends on the likelihood of centralization G(y H (π)). This decreases with G(y) as well, so the net effect on integration appears to be difficult to determine without more specific information about the distributions. As would be expected from the theory of options, where more uncertainty increases option value, more risk in the supplier’s problem-solving capacity raises the value of integration. But again, the cost of integration cS zG(y H (π)) could rise or fall with increases in risk. Despite this ambiguity, the effects on integration of changes in the riskiness of G(y) can be signed for certain classes of distributions, in particular the lognormal family, which is salient for our empirical analysis. Within that class, the relative benefit effect dominates the cost effect, even if that happens to be countervailing, so we have: Result 4. If there is integration at lognormal G(y), there is integration at a riskier lognormal Ĝ(y). For the proof of this result, see Section A-1.1 of the Theoretical Appendix. Results 3 and 4 imply that delegation and integration co-vary with different moments of the productivity distribution: roughly speaking, delegation co-varies with the mean, but not with risk; integration co-varies with risk, but not with the mean. These results provide another illustration of how the conceptual distinction between delegation and outsourcing is manifested in terms of measurable quantities. 3.4 From Theory to Data According to the theoretical model (Results 1 and 2), both integration and delegation choices should depend on input value. Suppliers of more valuable inputs should be more likely to be integrated with firms; among the integrated suppliers, those producing more valuable inputs should be delegated more decisions. These results lead to the following testable predictions, which we will bring to the data in Section 5: P.1: Firms should be more likely to delegate production decisions to integrated suppliers that produce more valuable inputs. P.2: Firms should be more likely to integrate suppliers that produce more valuable inputs. In Section 4.3, we describe how we measure delegation, integration, and input value, to assess the validity of the above predictions. 14 The theory also implies that delegation and integration should depend on G(y), the distribution of problems that can arise during production. In our empirical analysis, we will bring to the data Results 3 and 4 by constructing a proxy for G(y) for each input industry. To this end, in what follows, we will assume that the distribution of problems is specific to each input industry and that each supplier has one customer, at least under the period of observation.13 According to the model, a supplier who is not integrated will receive a payment py, no matter how large or small y is. Thus, interpreting y in the model as expected output per unit of resources (e.g., employment), the observable labor productivity of a non-integrated supplier should reflect the problem encountered producing his input. Under the assumption that the problems encountered are i.i.d. among suppliers in the same industry i, the empirical distribution of sales per worker of non-integrated suppliers in industry i can thus be used as a proxy for the distribution Gi (y).14 In particular, if there is a large enough number of suppliers in an industry and the empirical distribution of labor productivity is approximately lognormal, as is the case in our data (see Figure A-3 in the Empirical Appendix), then we can take Gi (y) to be lognormal, validating one of the hypotheses of Result 4. Lognormality also allows for simple parametric measurement of stochastic dominance and risk in terms of the mean and coefficient of variation (see Levy, 1973), which will be employed in the empirical analysis. In particular, controlling for the coefficient of variation, differences in the mean productivity of suppliers in input industries proxy for stochastic differences in Gi (y); controlling for the mean, differences in the coefficient of variation proxy for differences in Rothschild-Stiglitz risk.15 In Section 5, we will assess the validity of the following predictions, which correspond to theoretical Results 3 and 4: P.3: Firms should be more likely to delegate to their integrated suppliers that operate in stochastically more productive input industries. 13 Alternatively, we could allow for multiple buyers for each supplier and assume that the problem drawn is perfectly correlated across his customers. 14 By contrast, the model implies that the observed productivity distribution of the integrated suppliers is not a good proxy for Gi (y). This is because the contribution of an integrated supplier depends on whether or not HQ centralizes, which she does whenever the supplier has low competence. In that case, it is HQ’s own competence z that determines productivity. The observed productivity of integrated suppliers is thus left-censored, clouding the relationship between various orderings of observed distributions from two industries i, î and orderings of the underlying true distributions Gi (y), Gî (y). In the case of non-integrated suppliers, who endogenously do not delegate, there is no censoring problem. Another reason for excluding integrated suppliers when constructing the uncertainty measures is that transfer pricing effects may distort their measured labor productivity. 15 Variation in the distribution of labor productivity across industries reflects not only differences in output, but also in the prices prevailing in those industries. Our proxy for first-order stochastic differences in Gi (y) will thus suffer from measurement error, if we do not control for industry prices. The coefficient of variation of labor productivity is instead independent of prices. 15 P.4: Firms should be more likely to integrate suppliers that operate in riskier input industries. 4 Data and Variables 4.1 Datasets In what follows, we describe the datasets used in our empirical analysis to construct firm-level integration and delegation measures and to assess the role of input value and uncertainty in shaping these decisions. 4.1.1 World Management Survey The World Management Survey (WMS) is a large scale project aimed at collecting high quality data on organizational design of firms around the world and has been used in many studies (e.g., Bloom et al., 2012). The survey is conducted through phone interviews with plant managers. Several features of the survey design help to make the data of high quality. First, the survey is “double blind”, i.e., managers do not know they are being scored and interviewers do not know the plant’s performance.16 This enables scoring to be based on the interviewer’s evaluation of the firm’s actual organizational practices, rather than their aspirations, the manager’s perceptions, or the interviewer’s impressions. Second, each interviewer ran 85 interviews on average. This allows to include interviewer fixed effects, which help to address concerns over the reliability and consistency of the answers. Third, information on the interview process itself (duration, day-of-the-week), the manager (seniority, job tenure) and the location of the CEO of the firm was collected. These survey measures are used as “noise controls” to help reduce measurement error. The main wave of interviews was run in the summer of 2006, followed by smaller waves in 2009 and 2010. The survey achieved a 45% response rate, which is very high for company surveys.17 Overall, the WMS contains around 11,691 plants in 20 countries. The sampling frame was drawn to be representative of medium sized manufacturing firms in each country: median plant employment is 150, mean plant employment is 277, with a standard deviation of 405 (see row “WMS dataset” in Table A-1). 16 The interviewers were given the name and contact details of the firm, but no financial details. The high success rate is due to the fact that (i) the interview did not discuss firm’s finances, (ii) there were written endorsement of many institutions like the Bundesbank, Banque de France, UK Treasury, and World Bank, and (iii) high quality MBA-type students were hired to run the surveys. 17 16 For each firm, one (randomly selected) plant is surveyed in the WMS. Thus, we do not typically observe variation in delegation across plants belonging to the same firm. In some cases, the same plant was interviewed again in the later waves. 4.1.2 WorldBase WorldBase provides coverage of public and private firms in more than 200 countries and territories.18 This dataset has been used extensively in the empirical literature on firm boundaries (e.g., Acemoglu et al., 2009; Alfaro et al., 2019). The unit of observation is the establishment/plant, namely a single physical location where industrial operations or services are performed or business is conducted. Each establishment in WorldBase is identified by a unique nine-digit sequence called Data Universal Numbering System (DUNS) number. For each establishment, WorldBase provides information on its primary industry and up to five secondary industries in which each establishment operates. These are classified based on the 1987 US Standard Industrial Classification (SIC). The activities are recorded at the SIC4 level (935 industries, of which 459 are in manufacturing). Worldbase also provides additional information of the plants (e.g., location, sales, employment). WorldBase allows us to trace ownership linkages between establishments. In particular, we can use DUNS numbers to link plants that have the same domestic or global parent. D&B defines a parent as a corporation that owns more than 50 percent of another corporation. To construct firm-level variables, we link all plants that have the same domestic ultimate owner.19 We use the 2005 WorldBase dataset and focus on the 20 countries that are also included in the WMS. WorldBase contains 17,371,146 plants (corresponding to 16,718,199 parent firms). Median plant employment is 2, the mean is 288, and the standard deviation is 5,428 (see row “WorldBase dataset” in Table A-1). 4.2 Samples In the empirical analysis, we use two samples constructed from the datasets described above. 18 WorldBase is the core database with which D&B populates its commercial data products that provide information about the “activities, decision makers, finances, operations and markets” of the clients’ potential customers, competitors, and suppliers. The dataset is not publicly available but was released to us by Dun and Bradstreet. The sample was restricted to plants for which primary SIC code information and employment were available (due to cost considerations). For more information see: http://www.dnb.com/us/about/db_database/dnbinfoquality.html. 19 A “Domestic Ultimate” is a subsidiary within the global family tree which is the highest ranking member within a specific country and is identified by a “domuduns”’ code. A “Global Ultimate” is the top most responsible entity within the global family tree and is identified by “gluduns” code. The two codes only differ in the case of multinationals firms. In the case of multinational corporations, we follow Alfaro et al. (2016) and split them into several corporations, considering the domestic ultimate in each country as the relevant headquarters. 17 4.2.1 Delegation Sample This is the sample that is used to test the predictions related to delegation. It is constructed by combining information from WorldBase and the WMS. Notice that we are not able to use the full WMS to test the model’s predictions concerning the role of input value and uncertainty on delegation choices. This is because testing predictions P.1 and P.3 requires information on ownership and input-output linkages between the plant surveyed in the WMS and its central headquarters, which we can only get for plants that are also in the WorldBase dataset. For the United States and Canada we have linked plants interviewed in the WMS to plants in WorldBase using a common plant identifier (the DUNS number). For the remaining countries, there is no common plant identifier, so have used a string matching algorithm based on company names and location information to link plants in the WMS to firms in WorldBase. We have then manually checked the results of the matching process. To construct firm-level variables, we have used ownership information from WorldBase to identify the parent of any matched plant. As mentioned above, the WMS is focused on medium-sized manufacturing plants, while WorldBase contains lots of very small plants. The matched sample includes 3,444 plant-year observations located in 20 countries.20 The final sample used in the delegation regressions, which we henceforth call “delegation sample”, is a bit smaller due to missing information on some control variables for some of the plants. It consists of 2,889 plant-year observations, for plants operating in 574 sectors, corresponding to 2,253 firms. In this sample, 1,663 observations correspond to plants that are part of multi-plant firms, while the remaining observations correspond to single-plant firms. Note that, even though in principle only one plant per firm was interviewed in the WMS, some plants were interviewed more than once. This explains why the number of firms is smaller than the number of plant-year observations.21 As shown in Table A-1, this is a representative sample from the WMS: median plant employment is 150, the mean is 254, and the standard deviation is 367. Table A-2 reports the number of observations (at the plant level) by country in the delegation sample. Using information for all the countries in the sample is key to having a sufficient number of observations for our econometric analysis. Even the country with the maximum number of observations (the United States) accounts for only 20 percent of the sample. 20 The remaining plants in WMS could not be matched to a firm in WorldBase, either because they are not in the WorldBase sample, or because they could not be uniquely assigned to a specific firm based on the string matching algorithm. Since the success of the algorithm is random, the matched sample corresponds to a random subsample of the WMS. 21 The WMS includes only three cases in which two plants belonging to the same firm are surveyed. 18 4.2.2 Integration Sample This sample is constructed exclusively from the WorldBase dataset. It is used to test the predictions related to integration. We exclude firms that have less than 20 employees, to correct for differences in the coverage of small firms across countries (see also Klapper et al., 2006). Since the WMS contains exclusively manufacturing plants, we restrict the attention to firms that have a primary SIC code in manufacturing (between SIC 2000 and 3999). To be able to exploit within-firm variation across inputs, we select firms that integrate at least one input different from their primary output j. The final sample used in the integration regressions, which we henceforth call “integration sample”, includes 67,111 plants, corresponding to 66,102 firms, operating in 459 sectors, located in 20 countries. As shown in Table A-1, this sample features more variation in plant size compared to the delegation sample: median plant employment is 42, the mean is 147, and the standard deviation is 3,187. Table A-3 reports the number of observations (at the firm-input level) by country. 4.3 Key Variables In what follows, we define the main variables used in the empirical analysis. Tables A-5 and A-4 in the Empirical Appendix present summary statistics for these variables. 4.3.1 Delegation Our measure of delegation comes from the WMS survey of Bloom et al. (2012). They conducted in-depth interviews with the plant managers of medium-sized manufacturing firms, excluding those where the CEO and the plant manager was the same person (4.9% of their interviews). Plant managers were asked to state the degree of autonomy they have when hiring a new full-time permanent shop floor employee, introducing a new product, or in sales and marketing decisions. These qualitative variables were scaled from a score of 1 (defined as all decisions taken at the corporate headquarters), to a score of 5 (defined as complete autonomy granted to the plant manager). They were also asked how much capital investment they could undertake without prior authorization from the corporate headquarters. This is a continuous variable enumerated in national currency that is converted into dollars using PPPs.22 Since the scaling may vary across questions, we have standardized the scores from the four autonomy 22 In Appendix Figure A-6, we detail the individual questions in the same order as they appear in the survey. 19 questions to z-scores, by normalizing each question to mean zero and standard deviation one.23 The variable Delegationf (j,c,)p(i) measures the overall degree of autonomy that the headquarters of firm f (with primary activity j, located in country c) grants to the senior manager of plant p (with primary activity i).24 It is the average across the four z-scores for plant p belonging to firm f .25 We use information on ownership linkages from the WorldBase dataset to link a plant to its parent firm. Notice that the design of the survey implies that delegation is only measured for integrated suppliers. This is in line with our theoretical model, in which the control of the production of the input can be delegated to a supplier only within firm boundaries.26 Figure A-1 in the Empirical Appendix shows the distribution of Delegationf (j,c,)p(i) in the delegation sample. As pointed out by Bloom et al. (2012), their survey includes different types of firms: many have only one production plant, others are multi-plant firms;27 in most cases, the CEO is located in a different location than the plant manager interviewed; in others cases, it is located on site. In Section 5.2, we examine the effects of input value and supplier uncertainty on delegation choices. As mentioned before, the WMS typically contains information on one plant per firm. This implies that we can only observe the degree of autonomy granted by the CEO to one plant manager. We use the primary SIC4 code of the parent firm to identify the output industry j and the primary SIC4 code of the plant to identify the input industry i and rely on cross-firm variation in Delegationf (j,c,)p(i) to identify the role of input value. 4.3.2 Integration To distinguish between integrated and non-integrated inputs, we build on the methodology developed by Fan and Lang (2000), combining information on firms’ reported activities with 23 The continuous measure of delegation using in the empirical analysis stands in apparent contrast to the binary delegation choice in the theoretical model. However, the two can be reconciled by supposing that the production process is subdivided into a number of tasks, each of which is subject to a problem shock, and can be delegated or centralized. HQ delegates a task whenever the supplier’s productivity on it exceeds the threshold y ∗ (π), and centralizes otherwise. Interpreting the number or fraction of tasks delegated as the degree of delegation yields a measure that has the same properties to the probability of delegation in the baseline model. 24 We do need to keep track of the location of the plant, since in our dataset f and p are always located in the same country (see footnote 19). 25 All subcomponents are highly correlated with overall delegation score. The highest correlation (0.72) is with the z-score for marketing autonomy. 26 In the case of non-integration, there is de-facto decentralization of production of the input: HQ does not have the authority to centralize production decisions, and the supplier has no incentive to give control to HQ. 27 In the case of firms with multiple plants, the random plants surveyed are the best guess at the average degree of decentralization in the firm as a whole. 20 input-output (IO) tables. As explained below, in our empirical analysis, we measure vertical integration at the firm-input level, in line with our theoretical model: for each firm, we identify the set of inputs needed to produce its final good and estimate the probability that the firm produces each input within its boundaries. To measure vertical integration, we need standardized and disaggregated data on input requirements for each output sector. It should be stressed that disaggregated IO tables are only available for a few countries. Moreover, when available, they are usually based on different sector classifications. For these reasons, we follow previous studies (e.g., Acemoglu et al., 2009; Alfaro et al., 2016, 2019) in using US IO tables to capture technological linkages between sectors. These tables should be informative about input flows across industries to the extent that these are determined by technology. As pointed out by Acemoglu et al. (2009), assuming that the US IO structure carries over to other countries also mitigates concerns about the endogeneity of technology.28 The data are from the Benchmark IO tables of the Bureau of Economic Analysis (BEA). We employ the Use of Commodities by Industries after Redefinitions 1992 (Producers’ Prices) tables. The BEA uses six-digit industry codes, while the classification of production activities in WorldBase is based on the SIC 1987 classification. To convert the input-output data at the 4-digit SIC level, we use the concordance guide provided by the BEA.29 This allows us to measure vertical linkages between 935 manufacturing and non-manufacturing SIC4 industries. For each output industry j, IO tables report the dollar value of i used as an input in the production of $1 of j, also known as the direct requirements coefficient, IOi,j . To measure vertical integration, we combine the IO tables with information from WorldBase on the primary and secondary activities of each firm f .30 We proceed in four steps. First, we use the primary SIC4 code of each firm f to identify its output sector j. Second, we use IO tables to identify the set of inputs required to produced j, which we denote by S(j) = {i : IOi,j > 0}. Third, we identify which inputs firm f integrates, using the primary and secondary SIC codes reported by the firm and all its subsidiaries (if any). We define 28 The results are robust to restricting the analysis to high-income members of the Organisation for Economic Co-operation and Development (OECD), which are closer to the United States in terms of technology. 29 The concordance table can be found at http://www.bea.gov/industry/exe/ndn0017.exe. It assigns a unique 6-digit IO industry to each 4-digit SIC code in manufacturing (i.e., between 2000 and 3999). Outside these sectors, each SIC code may be matched to multiple IO codes. When this is the case, we consider all possible matches. In robustness checks, we show that our results are robust to restricting the analysis to manufacturing inputs, for which the concordance between the SIC and IO industry codes is one to-one. 30 As mentioned above, each establishment in WorldBase reports a primary activity and up to five secondary SIC codes that are produced at the location of the production facility. A firm with a headquarters and k establishments can thus in principle report up to 5+6∗k integrated secondary activities. Most firms in our integration sample have a single establishment. However, they still report secondary activities. Right censoring is not a problem: only 0.01 percent of plants in the WorldBase sample report 5 secondary SIC codes. 21 I(f ) ⊆ S(j) to be the set of integrated inputs, which the firm can in principle obtain within its ownership boundaries. The complement set N I(f ) = S(j) \ I(f ) identifies the nonintegrated inputs, i.e., the inputs required in the production of the firm’s output that are not included in I(f ). Finally, having identified the set of integrated and non-integrated inputs for each firm f , we can construct the variable Integrationf (j,c),i . This is a dummy variable equal to 1 if firm f (producing primary output j, located in country c) integrates a supplier in input industry i within its boundaries. In our empirical analysis, we exploit variation in Integrationf (j,c),i within and across firms to study how input value and input risk shape integration choices. To keep the analysis tractable and exclude cases of mechanical integration (almost all sectors use their own output as an input), we limit the sample to firms that integrate at least one input different from their output j and to the top 100 (manufacturing and non-manufacturing) inputs i used by j, as ranked by the IO coefficients (see also Alfaro et al., 2019).31 As mentioned before, US IO tables are highly disaggregated, providing information on vertical linkages between 935 manufacturing and non-manufacturing industries.32 As a result, even when focusing on the top 100 inputs, the average probability that a firm integrates any input is only around 1 percent (see Table A-5). 4.3.3 Input Value To examine how input value affects delegation and integration choices, we use the variable IOi,j described above. This is the direct requirement coefficient for the pair of sectors i and j, which captures the dollar value of input i used in the production of one dollar of j. Thus IO coefficients IOij are a close empirical counterpart to πij in our theoretical model. In robustness checks, we use different quartiles of the IOi,j variable to check for non-linear effects of input value on organizational choices. Figure A-2 in the Empirical Appendix shows the distribution of IOi,j in the WorldBase sample. Not surprisingly, given that the BEA input-output tables are highly disaggregated, the average IOi,j is only 4 cents in the delegation sample (see Table A-4) and 5 cents in the integration sample (see Table A-5). 31 The results are robust to restricting the analysis to the top 10 or top 20 inputs. Outside manufacturing, the correspondence between BEA and SIC industry codes is not always unique and this may induce measurement error in the IO coefficients. In fact, the effect of input value on integration is even stronger for manufacturing inputs, where the correspondence is one to one. 32 22 4.3.4 Uncertainty in Input Industries To test prediction P.3, we use information from the full WorldBase dataset to measure labor productivity (sales per employee) of non-integrated suppliers in an input industry. As discussed in Section 3.4, this is can be used as a proxy for the distribution Gi (y). In particular, we construct the variable Mean Productivityi,c as the arithmetic average of labor productivity of non-integrated suppliers of input i in country c. To minimize measurement error, we consider all plants that report SIC4 code i as their only production activity. Overall, labor productivity is computed for a sample of 12,063,180 non-integrated plants in WorldBase. The distribution of labor productivity of independent input suppliers approximates a lognormal distribution (see Figure A-3 in the Empirical Appendix).33 Figure A-4 in the Empirical Appendix shows the distribution of Mean Productivityi,c in the integration sample. To assess the validity of prediction P.4, we use CV Productivityi,c , the coefficient of variation of productivity of suppliers in the same input industry.34 Controlling for the mean of supplier productivity, this variable can be used as a proxy for riskiness of the input industry in the Rothschild-Stiglitz sense (see Levy, 1973). Figure A-5 in the Empirical Appendix shows the distribution of CV Productivityi,c in the integration sample. In robustness checks, construct alternative measures of uncertainty in input industries, using data on stock market returns of US firms from Bloom et al. (2018). We also verify that all our results are robust to constructing Mean Productivityi,c and CV Productivityi,c after winsorizing labor productivity at the 5th and 95th percentile (results available upon request). 5 Empirical Results 5.1 The Effects of Input Value and Uncertainty on Delegation In this section, we assess our model’s predictions concerning the effects of input value and of uncertainty in input industries on delegation and choices. To this purpose, we use the delegation sample to estimate: Delegationf (j,c),p(i) = β1 IOi,j + β2 Mean Productivityi,c + β3 CV Productivityi,c + +β4 Xp + β5 Xf + δi + δj + δc + ϵf (j,c),p(i) . 33 (7) The figure is constructed by regressing log labor productivity of all independent suppliers in WorldBase on 4-digit-industry × country dummies. It thus shows within-industry-country variation in log labor productivity. 34 Notice that, unlike Mean Productivityi,c , this variable is independent of industry prices. 23 The dependent variable is the degree of autonomy granted by the CEO of the parent firm f (with primary activity j, located in country c) to the senior manager of plant p (with primary activity i). The variable IOi,j is the direct requirement coefficient for the sector pair ij. Mean Productivityi,c and CV Productivityi,c measure respectively the mean and coefficient of variation of labor productivity of independent suppliers in industry i in country c. Some specifications include vectors of plant-level controls (Xp ), firm-level controls (Xf ), inputsector and output-sector fixed effects (δi and δj ), and country fixed effects (δc ). Plant-level controls are the percentage of the plant’s employees with a bachelor’s degree or higher and the plant’s share of the firm’s employment. Firm-level controls are the employment and age of the parent firm. We cluster standard errors at the input-industry i level.35 Prediction P.1 of our theoretical model suggests that a producer of good j should delegate more decisions to integrated suppliers that produce more valuable inputs, implying that the coefficient β1 should be positive and significant. According to prediction P.3 of our model, firms should delegate more decisions to suppliers that operate in more productive industries, so we also expect the coefficient β2 to be positive and significant. The results of estimating (7) are reported in Table 1. We present first a specification that includes the key control variable with input-industry fixed effects (column 1), and then further include country fixed effects (column 2), output-industry fixed effects (column 3), and the plant and firm controls (column 4). Consider first the effects of input value. As mentioned in Section 4, the WMS does not allows us to observe variation in delegation across plants belonging to the same firm. As a result, β1 cannot be identified by comparing the degree of autonomy granted by firm f producing good j (e.g., automobiles) to the managers of two of its plants, one producing input i (e.g., plastics materials and resins), the other producing input i′ (e.g., engines). We can, however, exploit cross-firm variation in the degree of delegation to identify the coefficient of IOi,j . In particular, we can compare the degree of autonomy granted by two firms producing the same output (e.g., automobiles) to the managers of one of their plants, one producing input i (e.g., plastics materials and resins), the other producing input i′ (e.g., engines). We can also compare the degree of autonomy granted to two plants producing the same input i (e.g., plastics materials and resins) who belong to firms making different final goods (e.g., automobiles and fabricated pipe and fittings). 35 The results are robust to clustering standard errors at the input-output level (the level of variation of IOi,j ) and at the industry-country level (the level of variation of Mean Productivityi,c and CV Productivityi,c ). 24