Sourcing By Design: Product Complexity and the Supply Chain
Sharon Novak • Steven D. Eppinger Kellogg Graduate School of Management, Northwestern University, 2001 Sheridan Road, Evanston, Illinois 60208-2009 Sloan School of Management, Massachusetts Institute of Technology, 50 Memorial Drive, Cambridge, Massachusetts 02142 [email protected] • [email protected]
his paper focuses on the connection between product complexity and vertical integration Tusing original empirical evidence from the auto industry. A rich literature has addressed the choice between internal production and external sourcing ofcomponents in the auto industry. More recent literature has developed the concept ofproduct architecture as another choice variable that may be one ofthe important contributors to product complexity. In this paper, we connect these two important decisions and study them jointly. We use the prop- erty rights approach to argue that complexity in product design and vertical integration of production are complements: that in-house production is more attractive when product com- plexity is high, as firms seek to capture the benefits oftheir investment in the skills needed to coordinate development ofcomplex designs. We test this hypothesis with a simultaneous equations model applied to data from the luxury-performance segment of the auto indus- try. We find a significant and positive relationship between product complexity and vertical integration. This has implications for optimal incentive structures within firms, as well as for interpreting firm performance. (Product Development; Product Complexity; Product Architecture; Property Rights; Transaction Costs; Vertical Integration; Automotive Industry; Supply Chain Management)
1. Introduction to coordination challenges during product develop- This paper focuses on the connection between ment. We use the property rights approach to argue product complexity and vertical integration using that complexity in product design and vertical inte- original empirical evidence from the auto indus- gration ofproduction are complements: that in-house try. Product complexity has three main elements: production is more attractive when product complex- (1) the number ofproduct components to spec- ity is high, as firms seek to capture the benefits of ify and produce, (2) the extent of interactions to their investment in the skills needed to coordinate manage between these components (parts coupling), development and production ofcomplex designs. and (3) the degree ofproduct novelty. Variations We test this hypothesis with a simultaneous equa- in product complexity are driven by a number of tions model applied to original data from the luxury- factors such as choices in performance, technol- performance segment of the auto industry. ogy, and product architecture. The effect of this This research builds on efforts to capture the role product design choice on the outsourcing decision can ofasset specificity in the vertical integration deci- be profound, as greater product complexity gives rise sion. Product complexity creates a variety oftrans-
0025-1909/01/4701/0189$5.00 Management Science © 2001 INFORMS 1526-5501 electronic ISSN Vol. 47, No. 1, January 2001 pp. 189–204 NOVAK AND EPPINGER Product Complexity and the Supply Chain action costs, such as the coordination cost to design mance. New interactions between components may and execute production.1 Consider the following three be discovered and must be explored to optimize the examples taken from our work in the auto industry. new suspension. However, once a particular front- 1. Number of Components. Vehicles can be designed and rear-suspension system has been developed and for electronic control with one to several multiplexing used in a vehicle, the process ofcoordinating devel- switches. Adding more multiplexes, such as modules opment for another vehicle with the same type of sus- in the doors and body, simplifies wiring, which saves pension is much easier. space, reduces weight, and improves electrical perfor- In light ofthe costs associated with product com- mance. However, each additional multiplex requires plexity, both transaction cost theory and the property its own part drawing, part number, complicated elec- rights approach offer similar predictions, namely, that tronic testing and validation, and associated tracking product complexity and vertical integration are com- and design work. Thus, adding more parts adds to the plements. Transaction cost theory suggests that a firm coordination needed to ensure vehicle development. seeking to minimize the coordination costs associated 2. Component Interactions. Vehicles are often with developing a complex system will internalize designed with front-wheel drive to provide better production.2 As formalized by Grossman and Hart handling on slippery roads. However, a front-wheel- (1986), the property rights view is that only physical drive automatic transmission must be much narrower asset ownership affects the incentives of the parties to than a rear-wheel-drive transmission to fit between invest in the skills required for coordination of com- the front wheels. This narrow design requires much plex designs. Vertical integration ofproduction pro- tighter physical adjacency ofcomponents within the vides the manufacturer with residual rights of control transmission system. As a result, making changes to over those assets, and allows the manufacturer to cap- a component is much more likely to require further ture the benefits ofits investment. The manufacturer parts changes, as the components are tightly cou- will integrate production ofcomplex systems to cap- pled. This requires that any part changes must be ture the benefits ofits skill investment. further coordinated with the designs of all the cou- For the purposes ofempirical testing, we adopt a pled parts. Thus, the more interconnected are the model based on the property rights framework for parts in a system, the more difficult it is to coordinate the following reason: The investment in skills needed development. to coordinate the product development process takes 3. Product Novelty. When a product involves a place during the design phase ofproduct develop- new architecture or new technologies, there is not ment, but the benefits ofthis investment are deter- a stable, well-understood set ofinteractions between mined during the production stage. That is, most components. The process ofidentifyingand under- auto components first are designed, then prototypes standing these relationships adds to the difficulty of are produced and tested, after which any necessary coordinating development. For example, in the design changes are made to the initial design. It is not possi- ofthe vehicle suspension system, occasionally a new ble before testing to enumerate the exact amount and configuration will be used for either the front- or nature ofchanges in design that will be required, thus rear-suspension design. When such a new type of product development in the auto industry is a classic suspension is utilized, it affects the entire vehicle’s example ofcontractual incompleteness. Aftertesting, dynamics. Lengthy and difficult development itera- the party that owns the assets at the production stage tions are required to create the desired vehicle perfor- determines the changes that are to be made to the initial design. This is why we use a framework that
1 We refer to product complexity as a proxy for transaction costs. connects skill investment to asset ownership. However, product complexity is neither limited to asset speci- ficity nor does it necessarily capture asset specificity in entirety, 2 In Williamson’s (1979) view, for example, vertical integration as asset specificity is a very abstract concept. We use the more allows bargaining issues to be resolved by fiat; he assumes that concrete assumptions ofthe property rights literature in testing employees (in contrast to suppliers) obey orders. Thus, coordina- our hypothesis. tion within the firm will be better than coordination between firms.
190 Management Science/Vol. 47, No. 1, January 2001 NOVAK AND EPPINGER Product Complexity and the Supply Chain
The usual empirical model offirm structure takes product complexity and vertical integration. In § 5 asset specificity as exogenous, focusing on firm struc- we examine the variety ofarchitecture and sourcing ture as determined by specificity. Our empirical strategies observed in the industry with respect to analysis explicitly accounts for the fact that prod- system performance. The paper concludes with a dis- uct complexity and firm structure are interrelated cussion about the implications ofthese findings for decisions.3 Our data set includes not only the two practice and for further research. simultaneous choices ofvertical integration (in-house production) and product complexity, but also exoge- nous variables, which should affect only one of the 2. Related Literature two choices directly. This research builds on literature from several fields The standard models also focus on component-level ofstudy. Scholars in both economics and manage- analysis ofvertical integration in the auto industry. ment science have addressed the make/buy decision. Our unit ofanalysis is the automotive system, which The importance ofproduct architecture and complex- we believe has greater explanatory power. Complex- ity has been established in the systems engineering ity is a property ofthe development tasks that can literature. We link these disparate strands ofresearch arise as a result ofthe number ofelements, the cou- to create a framework in which to study the joint pling ofthose elements, and the novelty ofthe prod- decisions ofproduct complexity and firm structure. uct. This depends on the complexity ofthe individual We begin with a review ofthe economic frame- component, as well as on the interrelation ofcompo- work posed by transaction cost theory and the prop- nents within the system. The sourcing decision, then, erty rights approach in order to motivate the issues requires careful evaluation of the trade-offs associated confronting the firm in the make/buy decision. We with the system as a whole, as well as the part-by-part review the existing empirical literature on make/buy decision. Our analysis highlights how the relationship and introduce product development concepts used in between product complexity and vertical integration our alternate model ofthe coupling between product might affect the conclusions that can be drawn from complexity and vertical integration. earlier studies and allows for better understanding of the interaction between these decisions. 2.1. Economic Theories of the Firm Our main result is evidence ofcomplementarity Concepts of Transaction Costs. In economics, firm- between product complexity and vertical integration. level and interfirm activities are modeled as contracts This is objectively measured and can be applied (Jensen and Meckling 1976). According to transaction to other complex manufactured products. This find- cost theory, the determinants ofmake versus buy are: ing has important implications for interpreting firm (a) asset specificity, (b) uncertainty, (c) frequency of performance. We examine evidence of clustering transactions, and (d) opportunism. The decision to within the auto industry around high-performance make or to buy a part also depends on the cost associ- combinations. ated with writing and monitoring a contract between The remainder ofthis paper is divided into six sec- the firm and an outside supplier—the costs associated tions. We review the current literature in § 2. We with the transaction.4 Transaction cost theory suggests then describe our methods in § 3. Section 4 contains three main reasons why it is difficult to write contracts. the statistical evidence linking choices about vertical integration to product complexity. We present sum- mary data on the auto systems analyzed in our data 4 To date, the empirical testing with regard to theories ofthe firm set. We present evidence ofcomplementarity between in economics has been restricted to testing oftransaction costs. While our modeling framework is based on the assumptions of the property rights approach (Grossman and Hart 1986) we also 3 Our empirical approach is consistent with the theoretical view of interpret our hypothesis using the more general assumptions of Helper and Levine (1992) that asset specificity is jointly endogenous transaction cost theory in order to motivate our discussion ofthe with firm structure. existing empirical literature.
Management Science/Vol. 47, No. 1, January 2001 191 NOVAK AND EPPINGER Product Complexity and the Supply Chain
First, specifying all contingencies relevant to the agree- nonappropriability oftechnical knowledge, leading ment is costly, ifnot impossible. Second, negotiating to profits that the firm can capture only through the responsibilities ofall parties under all possible integration. contingencies is difficult. Third, when the transaction Masten et al. (1989) separate physical-asset speci- involves a high degree ofasset specificity coupled ficity and human-capital specificity into two mea- with uncertainty and opportunism, writing and mon- sures. They use the Monteverde and Teece (1982) itoring such a contract would be prohibitively expen- measure ofengineering hours forhuman capital and sive (Williamson 1979). This means that the parties also use a measure ofthe extent to which components will write incomplete contracts. That is, the contracts are produced using physical assets specific to the do not anticipate and cover all possible outcomes. company. They find that only engineering hours have Thus, the parties may have to renegotiate the agree- a significant effect on vertical integration, and argue ment should an unforeseen event occur, and this can that this demonstrates that human capital is more be time consuming and costly. This literature predicts important than physical-asset specificity in influenc- integration when transaction costs are high. One kind ing the decision to vertically integrate. oftransaction cost which might be particularly salient We believe that the use ofengineering hours as is coordination cost. a proxy for asset specificity by Monteverde and Klein et al. (1978) argue that coordination prob- Teece (1982) and Masten et al. (1989) potentially con- lems arising from relationship-specific investments founds the amount ofwork required to develop a will be less severe ifthe transaction is internalized. component with the type ofwork. Many hours could The firm is distinct from a market entity in that dis- be spent to create a simplified design, which reduces putes within the firm can be resolved without legal coordination costs. Alternatively, many hours could action or outside monitoring. Therefore, the firm is be spent on a highly complex design to optimize able to resolve unforeseen problems internally and product performance, resulting in designs requiring thus is able to capture the full gains from its invest- greater coordination. Therefore, engineering hours is ment under all possible outcomes. From this notion, not a direct measure ofthe cost ofthe transaction in Klein et al. conclude that vertical integration is the the market. We propose that transaction costs are best likely firm structure for transactions in which asset- represented as a function of product complexity. specific investment is important. In a study ofstrategic business units across six- teen industries, Harrigan (1986) suggests that suc- Measurement of Transaction Costs. As Joskow cessful firms with high product complexity had a (1988) notes, the abstract nature oftransaction cost higher degree ofintegration than firms with less theory makes empirical testing difficult. Testing product complexity. Our research, focusing on com- requires concrete measures ofasset specificity, as well plexity at the system level within one industry, pro- as a means to test when and how specific invest- vides the opportunity to explore such issues in greater ments become important. Comparing the relative per- detail. Masten (1984), Walker and Weber (1987), and formance of firms and markets has proven diffi- Masten et al. (1991) also address the relationship cult. Since 1982, a number ofpapers have tackled between product complexity and vertical integration, the problem ofempirical testing oftransaction costs. with complexity measured at the component level. Monteverde and Teece (1982) analyze a data set com- Our system-level measures capture interactions that prised of133 component-sourcing decisions at Ford may be missed in component-level analysis. and General Motors in 1976. They proxy asset speci- ficity with the number ofengineering hours required The Property Rights Approach. The property to design a particular component. The dependent rights view ofthe firm is a more formal model ofver- variable is the mode ofthe transaction: vertical inte- tical integration. This approach focuses on how dif- gration (in-house component production) versus mar- ferent ownership structures affect relationship-specific ket transaction (outsourced component). They argue investment incentives ofthe contracting parties. In that investing more engineering hours increases the this view, exemplified by Grossman and Hart (1986)
192 Management Science/Vol. 47, No. 1, January 2001 NOVAK AND EPPINGER Product Complexity and the Supply Chain and Hart and Moore (1990), physical-asset ownership the case that outsourced components are always mod- confers residual rights of control over relationship- ular and thereby simpler and faster to develop than specific assets when contracts are incomplete. In par- internally developed components. A part that is very ticular, the ultimate division ofprofits will depend on complicated due to its performance requirements may the relative strengths ofthe parties when they rene- take more time to develop with a supplier than gotiate, not at the point ofthe initial contract. This within a firm due to coordination problems between implies that nonowning parties will underinvest in the firms. Companies that emphasize rapid product project-specific coordination skills, as they will not development may wish to reduce the design iteration be able to capture the full benefits of their invest- required between systems as a result ofa more com- ment. This generates the result that the investing plex design by modifying the design itself, whether or party should own relationship-specific assets. not it is outsourced. The decision to change the part There has not been any empirical testing ofthe design characteristics and the choice ofwhether or assumptions ofthe property rights approach to not to outsource the part are separate, but, we argue date.5 Whinston (1997) compares the Monteverde and tightly linked. Teece (1982) study with the property rights approach, Clark and Fujimoto (1991) and Clark (1989) look at and demonstrates that in order for vertical integration the choice between new and existing components in to be more likely with greater investment in skills, an empirical study ofproduct development projects the returns to investment by the manufacturer must in the global auto industry. They examine the impact exceed the returns to investment by the supplier. We of“project scope”—a measure ofthe uniqueness of believe that this assumption is reasonable for the auto the part (vs. carryover parts) and the extent ofdevel- industry, as the auto manufacturer can gain returns to opment carried out by outside suppliers, on project its investment in the total vehicle by bringing in out- performance (lead time and cost). The authors found side suppliers, but an outside supplier cannot gener- that 67% ofJapanese projects were “black-box,” or ate another vehicle to utilize its system investment as developed by suppliers, as compared with 16% of easily. U.S. vehicles. They argue that the black-box sys- tem is effective because the link between design and 2.2. Theories of Sourcing and Design in manufacturing is strong. They argue that the high Operations Management percentage ofunique parts and high supplier involve- The operations management literature also addresses ment contributes to an observed Japanese advantage the choice between in-house production and out- in project lead time and cost. In this paper, our focus is sourcing ofcomponents. This trade-offisexemplified on how complexity in product design affects produc- by the decision between purchasing standard parts tion, and we do not address outsourcing ofdesign. 6 and developing custom components in house. Stan- The concept ofsystem interfacesused widely in the dard parts—ifthey have clean, or well-understood, system-engineering literature (Suh 1990, Alexander interfaces—can be outsourced more easily, but may 1964), provides an opportunity to enrich the measure- present trade-offs in terms of performance and cost ment ofcoordination costs. Ifan outsourced compo- over custom parts development (Ulrich and Ellison nent can be designed with well-defined interfaces, it forthcoming). may not require much coordination with suppliers Baldwin and Clark (2000) argue that outsourcing, during development. or selecting existing components from suppliers, may Fine and Whitney (1996) argue that a critical capa- allow a company to benefit from competition among bility in product development is the ability to write suppliers. However, we argue that it is not necessarily
6 Ulrich and Ellison (1998) argue that splitting design and produc- 5 We are not empirically testing the assumptions ofthe property tion ofcoupled systems should be avoided, but that both design rights approach in this paper; we are testing our hypothesis that and production ofsuch systems could be outsourced and achieve product complexity and vertical integration are complements. integration.
Management Science/Vol. 47, No. 1, January 2001 193 NOVAK AND EPPINGER Product Complexity and the Supply Chain competent specifications for components and systems the sample—three in Japan, three in Europe, and two and to be sure the specifications are realized. They list in the United States, account for roughly 90% of the three distinct outsourcing motivations: development global luxury-performance market by sales volume. capability, manufacturing competitiveness, and prod- We studied luxury-performance cars, defined by uct technology. The decision to outsource depends on Consumer Reports as vehicles priced above $30,000 whether the firms seek knowledge or capacity and in 1995. Our motivation for choosing the luxury- whether the product is readily decomposable from the performance segment is that more expensive vehicles rest ofthe system. In our audio example, many com- have a wider range ofavailable choices ofproduct panies with different information-processing capabil- architecture. As these are flagship vehicles, we expect ities but choosing the simpler architecture (separate that this segment allows for the most powerful and cellular phone circuitry) are able to outsource phones comparable test ofthe possibilities available to firms effectively. Indeed, outsourcing a modular component in both design and sourcing. A review ofthe data (see may be an effective short-term remedy to a lack of Table 1) indicates that there is a wide range in prod- development capacity within the firm. However, the uct complexity choices and sourcing choices, as well make/buy decision also affects the future capabilities as in system performance. We collected data focused ofthe firm. (Fine 1998) A firm that repeatedly relies on the same components in a single vehicle segment on an outside supplier to develop a system that is in the auto industry in order to remove possible mea- highly complex may not be able to maintain the skills surement problems caused by a data set which com- needed to develop the system internally. Over time, bines information from different vehicle types, such the firm may lose the option to develop the system as that of Clark and Fujimoto (1991), or from different internally as these skills atrophy. The firm can become component types, such as Masten et al. (1989). dependent on the supplier and therefore less able to The unit ofanalysis is the automotive system. share in the surplus generated by the development Within each company, for each luxury vehicle for each ofthe product. In our model, we address how own- time period in the sample, we have collected data on ership structures affect the incentive of the parties to seven key systems: engine, transmission, body, electri- invest in relationship-specific skills, but we do not cal, suspension, steering and brakes. Table 1 presents address the long-term implications ofsuch actions. summary data on the systems, with respect to com- Our research builds upon the work ofthe above- plexity, sourcing and quality.7 mentioned authors. Several researchers and theo- Data Collection. The data were collected through ries on the complexity ofengineering design have on-site interviews at all companies in the study. also focused on coupling and interactions (Alexander Over 1000 people were interviewed, including CEOs, 1964, Rechtin and Maier 1997, Suh 1999), but to chiefengineers, project managers, and system engi- the best ofour knowledge no statistically signifi- neers involved in development ofeach vehicle for cant testing ofthe relationship between this type of each time period in the study. All participants were product complexity and vertical integration has been assured that only aggregate data would be presented, done before. The data set analyzed here provides the and confidentiality agreements were signed with each opportunity to explore these ideas empirically in a company. manner that has not been previously possible. Data were collected in several stages. First, after signing the agreement with each firm, a letter was 3. Methods sent requesting interviews with relevant project man- agers, system engineers, design engineers, purchas- Sample. The analysis in this paper is based on a ing managers, and manufacturing engineers for each study ofproduct architecture and sourcing in the auto vehicle for each time period in the study. The relevant industry. Our newly developed data set covers com- ponents in eight vehicles over five overlapping five- 7 For reasons ofconfidentiality, company-specific product complex- year time periods from 1980–1995. The companies in ity, sourcing, and quality means are not presented.
194 Management Science/Vol. 47, No. 1, January 2001 NOVAK AND EPPINGER Product Complexity and the Supply Chain
parties were identified by the corporate liaison for Figure 1 Hypothesized Relationships each company, and on-site meetings were arranged. In order to ensure data accuracy, all interviewees were given an overview ofthe research project and definitions ofkey terms. All subjects were given a list ofquestions pertaining to the design and sourc- ing ofcomponents within their respective systems. The questions principally focused on objective infor- mation (e.g., number ofparts in the body side) so as to minimize the likelihood ofresponse bias. The inter- views were conducted on-site at each company, rang- ing from three days to three months. All interviewees were given the option ofbeing interviewed in their native languages. United States and European inter- views were conducted in English and Japanese inter- The independent variables, or outside factors affect- 8 views were conducted in Japanese. ing product complexity (CMPLX), are performance goals (PERF), major change (MAJ), worker skills (SKLZ), technology breaks (TECH), sunk cost (SNK), 4. Relationship Between Product and platform requirements (PLAT). The independent Complexity and Sourcing variables affecting vertical integration (VERTINT) are In this section, we begin by describing the factors that sunk cost (SNK), platform requirements (PLAT), plant directly affect the costs and benefits of product com- capacity (CAP), vehicle volume (VOL), and union plexity and vertical integration. Figure 1 illustrates the requirements (UNION). We define these variables hypothesized relationship between these two choices. and their predicted relationship to our dependent variables below. Summary statistics are presented in
8 All interviews were conducted by one ofthe authors (S.N.). Table 1. Professor Kentaro Nobeoka, a scholar with extensive experi- The dependent variable VERTINT is the per- ence in the Japanese auto industry, provided Japanese interview centage ofthe system produced in house, with 1 interpretation.
Table 1 Summary Data on Automotive Systems
Mean Std Dev Mean Std Dev Quality System CMPLX CMPLX VERTINT VERTINT Range∗
Suspension 0.35 0.26 0.50 0.30 1 to 5 (all) Brakes 0.46 0.28 0.21 0.38 1 to 5 (all) Transmission 0.42 0.17 0.42 0.41 1 to 4 Engine 0.47 0.25 0.48 0.25 1 to 5 (all) Steering 0.55 0.230.43 0.40 1 to 5 (all) Body 0.46 0.19 0.26 0.27 1 to 5 (all) Electrical 0.21 0.21 0.33 0.40 1 to 5 (all)
Note. *CMPLX = product complexity, defined from 0 (low) to 1 (high) system complexity. See Appendix A for system-specific measures. *VERTINT = vertical integration, defined as the percent of the system components produced in-house. A score of 1 indicates in-house production of system components. *Quality is defined according to Consumer Reports Reliability Reviews, which are related according to vehicle system. A score of 5 = fewer than 2% problems per system (p.p.s.), the top C.R. score; 4 = 2% pps, 3 = 5% to 9 3% pps; 2 = 9 3%to14 8% pps; 1 = more than 14 8% pps.
Management Science/Vol. 47, No. 1, January 2001 195 NOVAK AND EPPINGER Product Complexity and the Supply Chain indicating in-house production ofall components. 9 PERF is a (0–1) measure which proxies for desired For each component, system, vehicle model, and time performance at the system level. Certain performance period, we have collected data on the make/buy goals necessitate more complex product designs, such decision outcome. The system measure is constructed as more integrated architectures (Ulrich 1995). For by equally weighting the measure ofeach compo- example, a result ofdesigning to meet high top- nent within the system. Parts supplied to firms by speed capability is a body system consisting oftightly wholly owned subsidiaries, such as the Delphi divi- interconnected parts.12 In our data set, performance sion ofGeneral Motors, are treated as in-house. goals were provided by vehicle product managers, Parts produced by partially owned suppliers, such as on a 0–10 scale, with 0 indicating no importance Nippondenso (Toyota group), were treated as outside for product performance goals and 10 indicating that suppliers. Sourcing spanned the entire range from 0 the vehicle competes based on high performance. (outsourced) to 1 (in-house production), with a mean We expect systems for which performance goals are of0.37 and a standard deviation of0.36, as shown in very high to be associated with product complexity, Table 1. and, hence, we expect a positive relationship between The measures ofproduct complexity used in this PERF and CMPLX. paper are based on detailed system design and manu- MAJ is the dummy for vehicle design status, tak- facturing data. For each system, we estimate product ing on a value of1 ifthe vehicle is undergoing a complexity on a spectrum from 0 to 1 (no complex major change. The timing ofmajor changes ranges system interactions to high product complexity) as an from every four years to every seven years (Clark unweighted average ofcharacteristics ofdesign com- and Fujimoto 1991). The firm has an opportunity to plexity.10 For some systems, measures include char- change product complexity in major changes, and we acteristics such as “newness”—the degree to which expect that in performance vehicles these changes a design configuration has been used in the com- should involve greater performance, and therefore pany and in the vehicle. For example, product com- greater product complexity. We expect a positive rela- plexity in the suspension system is calculated as an tionship between MAJ and CMPLX. unweighted average ofthree (0–1) measures: newness SKLZ is a dummy variable reflecting the presence ofthe design, number ofmoving parts in the sus- ofa worker skills/plant location effect.For example, pension, and whether the suspension is active or pas- a body design featuring many complex manual welds sive.11 This measure is then used for all components in cannot be manufactured in an area where workers the system. The dependent variable product complex- are not trained in advanced welding. Vehicle prod- ity (CMPLX) measures the complexity ofthe system, uct managers were asked whether absence ofworker with a score of1 indicating high system complexity. skills played a role in design considerations for each As shown in Table 1, product complexity spanned the system. A score of1 indicates a yes answer, that full range from 0 (no complex system interactions) to skill limitations were a factor in system design. Thus 0.99 (very high product complexity), with a mean of we expect a negative relationship between SKLZ and 0.42 and a standard deviation of0.27. CMPLX. TECH, the dummy for the state of technology, 9 Masten et al. (1989) use this measure ofsourcing at the component takes on a value of1 forthe year in which cer- level. We believe system-level analysis captures more information tain innovations, such as antilock brakes and new about sourcing behavior. This requires weighting all components electronics technology in suspension systems, are equally, as any attempt to capture value ofthe component requires introduced. This variable reflects technological inno- decomposing down to the component level. We discuss the impli- cations ofthis assumption forour model measurement in § 6. vations that have enabled increased product perfor- 10 For each system, measures ofcomplexity were chosen on the mance deliverable via modular components and we basis ofsystem-engineering principles. The complexity measures used are discussed further in the appendix. 12 This is due to the requirements for overall mass reduction in 11 These measures are discussed further in appendix. order to attain high top speeds.
196 Management Science/Vol. 47, No. 1, January 2001 NOVAK AND EPPINGER Product Complexity and the Supply Chain thus expect a negative relationship between TECH tem, like a one-piece body side, exceeds the capacity and CMPLX. ofcurrent plant equipment, it may be outsourced. For PLAT is a dummy variable for platform require- this reason we predict a negative relationship between ments in parts, indicating (with a “1”) whether the CAP and VERTINT. component was designed to be used by more than VOL is the variable for vehicle volume. We calcu- one vehicle. The literature on system design suggests late volume two ways, as absolute company volume, that constraining a component or system to meet the and as the percentage ofthe overall firm devoted to requirements ofmore than one vehicle necessarily luxury-performance cars. We believe both measures limits the performance optimization of that part rela- can influence sourcing decisions. BMW, for exam- tive to the vehicle in question (Ulrich 1995). For exam- ple, is much smaller than Toyota in absolute volume, ple, the Ford Taurus underbody greatly restricted but Toyota’s luxuxry-performance volume is much design complexity on the Lincoln Continental under- smaller than BMW’s. BMW may be able to command body design that was built on the same platform. a larger, not smaller, ordering capacity with suppliers For this reason, we predict that PLAT will have a neg- because ofits much larger luxury-performance mar- ative affect on CMPLX. Platform requirements could ket. Toyota may also be able to use its market domi- support in-house production through economies of nance in other segments to source more effectively in scale achieved through parts sharing. For this reason, luxury performance. For this reason we make no pre- we hypothesize a positive relationship between PLAT and VERTINT.13 diction about the direction ofthe relationship between SNK is a dummy variable for existing sunk VOL and VERTINT. cost/plant investment. Managers were asked whether The dummy variable UNION takes on a value of1 or not existing plant equipment directly affected their ifa component is produced in-house and is covered design choices for the system. Systems are often under a union agreement. Ifa system is produced in designed around investment in process equipment in a plant with a union agreement, it may be very dif- the plants. This may constrain design to a more com- ficult to outsource any ofthe components in the sys- plex company-specific process or to a simpler process. tem because ofthe extreme cost and risks associated Thus, we expect SNK to have a significant effect on with union renegotiation. For this reason we expect a CMPLX but make no prediction on the direction of positive relationship between UNION and VERTINT. the relationship. Managers were also asked whether or not existing plant capabilities directly affected their 4.1. The Statistical Model sourcing decision. A system may be built in-house Our principal concern in this paper is to study the as a result ofexisting plant investment. On the other relationship between product complexity and sourc- hand, systems are often outsourced because of exist- ing. The preceding discussion suggests that some ing in-plant manufacturing problems. For this reason, form of integration is likely to be chosen as product we also test for the relationship between SNK and VERTINT, but we make no prediction on the direction complexity increases. However, we have argued that ofthe relationship. product complexity and sourcing are coupled. Econo- CAP is a dummy variable indicating limited plant metrically, this suggests a model where product com- production capacity or capability. System managers plexity and sourcing are simultaneously determined, were asked ifthe plant had insufficientcapacity to so that our model should treat these two variables as manufacture system designs in-house. If a certain sys- jointly endogenous. Hausman (1983) has shown that using an instru- mental variables approach always leads to consis- 13 Consistent with transaction cost theory, we assume that although tent estimation for an identified model; this approach suppliers may be able to enjoy the same economies ofscale, they will not pass along the full savings of platform sourcing, because is taken in this paper. To test for the relationship ofthe holdup problem discussed in § 2. between product complexity and sourcing, we esti-
Management Science/Vol. 47, No. 1, January 2001 197 NOVAK AND EPPINGER Product Complexity and the Supply Chain mate the following model: affecting this decision through a second-order effect (through the rest ofthe system). The idea behind the CMPLX = 10 + 11 PERF + 12MAJ + 14SKLZ instrumental variables approach is that these vari- ables are uncorrelated with stochastic disturbances in + 15TECH + 16SNK + 17PLAT (1) the dependent variable to be measured, but are corre- + VERTINT + 14 1 1 lated with the jointly endogenous variable.
VERTINT = 20 + 23CAP + 26SNK For example, making a major change to a system such as the engine consists ofchanging the design + VOL + UNION (2) 27 28 ofthe engine, thus directly affectingits product com- 2 15 + 29PLAT + CMPLX + 2 plexity. It may be the case that this change in prod- uct complexity requires a change in the sourcing of Optimal Instrumental Variables. A consistent esti- the engine system components. This effect, however, mator ofthe system described by (2.1) and (2.2) is ofsecond order. That is, major changes do not is optimal instrumental variables, instrumenting for directly create changes in sourcing for a system; in CMPLX in (2.2) and VERTINT in (2.1) with the instru- fact, many companies prefer to use the same sup- ments Z. The instruments used for Equation (2.2) are: ply strategy even when system changes in design PERF, MAJ, SKLZ, and TECH. The instruments used are made. For this reason, we expect sourcing to be for Equation (2.1) are CAP, VOL, and UNION. As indirectly affected by major change through product this system is highly coupled, we expect that there complexity, and we expect to see that sourcing and will be correlation between all of the factors affect- major change are highly correlated, but we do not ing both decisions.16 However, the presence ofa vari- expect that unobserved variation in sourcing will be able in a particular equation indicates that this factor correlated with major change. Similarly, we assume directly affects the decision to be made, rather than that performance goals, worker skills, and technology breaks affect sourcing decisions through their direct 14 Ifone were to run only the Regression (2.2), using a method such effect on product complexity, and thus are correlated as ordinary least squares (OLS), the resulting measure for would 20 with sourcing, but are uncorrelated with unexplained be biased and inconsistent, because ofthe presence of in the 1 variation in sourcing. We assume that capacity lim- variable CMPLX. This is because there may be unobserved factors in sourcing which are correlated with the complexity decision. For itations, company volume, and union are similarly this reason, we use a simultaneous equation. uncorrelated with unexplained variations in product 15 Plus year, company, and system dummies. complexity, but are correlated with product complex- 16 Table 2 presents simple correlations ofthe variables. ity decisions.
Table 2 Correlations
system cmplx perf maj cap sklz tech snk vol union plat vertint system 1 cmplx −0 07 1 perf 0 05 0 25 1 maj 0 01 0 29 0 15 cap 0 27 −0 11 −0 20 1 sklz 0 51 −0 06 −0 20 0 59 1 tech 0 14 −0 09 0 09 0 10 15 0 18 1 snk 0 19 0 02 −0 2 −0 10 30 47 −0 11 vol 0 01 −0 24 −0 1 −0 20 15 −0 1 −0 10 09 1 union 0 01 0 32 −0 2 −0 4 −0 2 −0 −0 10 22 0 77 1 plat 0 35 −0 12 −0 1 −0 10 0 1 −0 20 05 0 0 17 1 vertint −0 17 −0 15 −0 2 −0 2 −0 2 −0 3 −0 20 0 74 0 55 0 06 1
198 Management Science/Vol. 47, No. 1, January 2001 NOVAK AND EPPINGER Product Complexity and the Supply Chain
Table 3 Regression Results for Product Complexity Table 4 Hausman Specification Test
Dependent Variable Coefficient Std. Error zp>z Dependent Variable Coefficient Std. Error zp>z
CMPLX VERTINT Independent Variables Independent Variables PERF −0 46 0 66 −0 69 0 49 CMPLX 1 61 0 59 ∗2 70 007 MAJ 1 14 0 43 ∗2 68 0 007 SNK −1 40 56 ∗ − 2 439 0 02 SKLZ −0 87 0 64 −1 370 17 VOL −43 5 −1 127 0 26 TECH −1 08 0 51 ∗∗ − 2 12 0 03 UNION 4 92 3 ∗2 09 0 04 SNK 1 130 61 1 84 0 07 PLAT 0 64 0 3 ∗2 10 04 PLAT −0 81 0 39 ∗∗ − 2 10 04 Error Term Instrumental Variables VHAT −0 730 87 −0 85 0 4 CAP 0 630 5 −1 20 21 System Dummy Variables VOL −1 27 1 46 −0 87 0 38 SUSP 0 320 46 0 70 5 UNION −0 530 67 −0 80 42 BRKS −0 02 0 5 −0 05 0 96 TRANS −0 92 0 47 ∗∗ − 1 922 0 05 ∗ = significant at the 0.01 level; ∗∗ = significant at the 0.05 level. N = 134, ENGINE −0 25 0 49 −0 50 6 Adjusted R2 = 0 30; Chi2 = 34 95 STEER −0 25 0 49 −0 50 6 Specification Tests. To test the effect of CMPLX on ELECTRICAL 0 11 0 530 22 0 82 VERTINT (Equation 2.2), we estimate Equation (2.1) ∗ = significant at the 0.01 level; ∗∗ = significant at the 0.05 level. N = 134, and formally validate our model using the Hausman Pseudo R2 = 0 37 Specification Test (Hausman and Wise 1978) as fol- confidence.19 As there is no relationship between our lows. We believe that the CMPLX decision is made as a result ofperformance goals, major changes, worker estimated error and VERTINT, this allows us to mea- skills, technology breaks, and platform requirements, sure the effect of CMPLX on VERTINT using an as well as potentially unobserved other factors, which ordered probit regression. In summary, our model accounts for the economet- we label “ 1”. Using instrumental variables, it is pos- sible to estimate CMPLX directly as a function of all ric implications ofthe coupling between the complex- the instruments, generating a predicted value cˆ.17 We ity and sourcing decisions. Our study design permits first ran a probit regression ofall the instruments us to measure the impact ofproduct complexity on CMPLX to obtain cˆ. The results ofthis regression on sourcing, and we have formally validated our are presented in Table 3. The difference between the approach using a test ofthe simultaneity between the observed values CMPLX and the new values cˆ is our complexity and sourcing decisions. Again, consistent with system engineering and transaction cost argu- estimate of 1, or the unobserved factors in CMPLX, which we label vˆ. We then ran an ordered probit ments, we predict that CMPLX will have a positive regression ofEquation (2.2), with the addition of vˆ affect on VERTINT 2 > 0. No predictions are made as a variable. If, in fact, we have removed the rela- concerning company-specific effects. tionship ofVERTINT on CMPLX, there should be no 18 4.2. The Effects of Product Complexity on relationship between 1 and VERTINT. That is, there should be no correlation between unobserved vari- Sourcing ables in CMPLX and the sourcing decision. Table 4 Table 5 presents the results ofan ordered probit presents the results ofthe Hausman Specification Test. regression of Equation (2.2). We tested for the effect of The relationship between vˆ and VERTINT is not sig- years on sourcing; we found that year dummies are nificantly different from zero, at standard levels of not significant, and thus dropped year dummies from
17 This is the same as running the reduced form regression 19 This result assumes errors are normally distributed. However, CMPLX = Z + V. See Hausman (1983). with a Z statistic of −0 845, less than one, a regression with log 18 This is the same as running VERTINT = x + 1CMPLX + 2vˆ. values is unlikely to reverse the result ofrejection ofthe null
H0 2 = 0. See Hausman (1983). hypothesis.
Management Science/Vol. 47, No. 1, January 2001 199 NOVAK AND EPPINGER Product Complexity and the Supply Chain
Table 5 Regression Results for Sourcing renegotiate the union agreement, a prospect that is Dependent Variable Coefficient Std. Error zp>z costly at best, and at worst, can result in a debilitating strike. As a result, U.S. firms face a far greater penalty VERTINT in attempting to outsource existing components, and CMPLX 1 61 0 59 ∗2 70 007 Independent Variables are thus more likely to build even simpler systems SNK −1 40 56 ∗ − 2 439 0 02 in-house. VOL −43 5 −1 127 0 26 Platform requirements (PLAT) also affected sourc- UNION 5 12 4 ∗2 095 0 04 ing positively at a 95% confidence level. This indicates ∗ PLAT 0 62 0 3 2 018 0 04 that constraining components to serve a platform System Dummy Variables SUSP 0 275 0 54 0 507 0 61 ofvehicles increased the likelihood ofin-house pro- BRKS −0 01 0 57 −0 017 0 99 duction. This result, consistent with transaction cost TRANS −1 42 0 74 ∗∗ − 1 92 0 05 theory, suggests that the benefits to increased order ENGINE −0 14 0 61 −0 229 0 82 volume made possible by producing a common STEER −0 130 62 −0 204 0 84 component for several vehicles are better obtained BODY −0 08 0 53 −0 149 0 88 through in-house production. This is because the ∗ = significant at the 0.01 level; ∗∗ = significant at the 0.05 level. N = 134, firm would face additional monitoring costs by 2 2 Adjusted R = 0 37; Chi = 113 28 outsourcing, allowing suppliers to build similar economies ofscale. It may also be the case that the our estimation.20 Company dummies were included platform requirement requires more modular designs, but are not reported for reasons of confidentiality. One which affect sourcing indirectly by reducing the coor- company had a significant, and negative, effect on dination costs associated with outsourcing. sourcing; the rest ofthe companies did not have a sig- Sunk cost (SNK) had a negative and significant nificant relationship with sourcing. System, volume, effect on sourcing, supporting the argument that sys- and capacity dummies did not play a significant role tems may be outsourced because ofexisting in-plant in the sourcing decision. variation. CMPLX, our measure ofproduct complexity, has a In summary, the predictions ofour model—that positive and significant effect on the percentage of the CMPLX, UNION, and PLAT increase the likelihood component produced in-house at a 99% confidence ofvertical integration and that SNK decreases the level. This means that an increase in product com- likelihood ofvertical integration—are supported by plexity is correlated with an increase in in-house pro- our regression results. duction. This is consistent with our prediction that investment in skills to coordinate complex designs is supported by ownership ofthe key assets used in 5. Impact on Product Quality production. Performance As predicted, UNION, which measures the extent Our hypothesis, that product complexity and vertical to which union requirements influence the decision integration are complementary, suggests a model of to vertically integrate, has a positive and significant firm performance as a function of the interaction of impact on in-house production at a 95% confidence these two organizational design choices. Our predic- level. This is consistent with our interview data on tion is that both complex systems produced in-house, the U.S. auto industry, where union agreements cover and simple systems outsourced, will be positively all components currently manufactured within a U.S. correlated with quality performance.21 Obviously, plant. To outsource a component, it is necessary to
21 Athey and Stern (1998) point out that there may be unobserved 20 In this test, we test the null that all year dummy coefficients are factors in the organizational choice equations that affect the per- equal to zero. At F = 0 76, P>F= 0 60, so we are unable to reject formance results. With this in mind, we do not present a formal the null, and thus are able to drop dummy variables. regression.
200 Management Science/Vol. 47, No. 1, January 2001 NOVAK AND EPPINGER Product Complexity and the Supply Chain performance is affected by a variety of other factors in low, and their performance can be seen as consis- addition to complexity and sourcing decisions. With tent with our hypothesis that complex in-house ought this is mind, we present system observations that to perform highly. However, any attempt to correct informally support our hypothesis. for content requires decomposing down to the com- Quality, as evaluated at the system level using ponent level. We recognize that this may understate Consumer Reports reliability data, ranged from 5, a the sourcing percentage at some ofthe firms in our score indicating fewer than 2% of problems reported, study. As this bias reduces the likelihood that we will to a score of1, indicating greater than 14.8% reported observe the predicted relationship between architec- problems. As shown in Table 1, all five possible ture and sourcing, we believe such corrections would scores were reported for all of the systems in the strengthen our results. study. Across all the systems, mean quality perfor- All the companies manufacture the five major mance also fit our predictions. We define CMPLX engine-system components (cylinder head, engine above 0.5 as complex (below 0.5 as simple) and sourc- block, crankshaft, camshaft, and intake manifold). The ing (VERTINT) above 0.5 in-house as in-house (below variation in sourcing centered around more poten- 0.5 as outsourced). Simple outsourced systems had tially decomposable components. The top perform- the highest mean quality of3.7 out of5. Complex ers were complex outsourced, again, because ofthe in-house had the next highest mean, at 3.2. Complex equal weighting ofcomponent sourcing. The worst outsourced had a mean quality of3.1. Simple in-house performer was simple in-house. In the brake system, had a mean quality of2.5. all but one ofthe companies in the study outsource In the suspension system, the top performer in the brakes as a complete system. The lack ofvariance in category, with a reliability score of5, featured a sim- sourcing limits our ability to interpret quality with ple outsourced design, which supports our view that respect to our hypothesized relationship between simpler products can be more effectively outsourced. architecture and sourcing. The top performers were The worst performer in the category, with a score of both complex outsourced and simple outsourced. The 1, featured a complex outsourced design, consistent worst performer was simple in-house. In the electri- with the idea that complex designs cannot be easily cal system, the worst performer was simple in-house decomposed and outsourced. In the body system, the and the top performer was simple outsourced. best performers, with scores of 5, were complex in- The steering system results also reflect another house and simple outsourced. The worst performers, measurement issue we encountered in our data set. with scores of1, were simple in-house and complex The best performer, with a score of 5, was complex outsourced. Again, this is consistent with our hypoth- outsourced. The worst performers were simple in- esized relationship between complexity and sourc- house and complex in-house. Past empirical studies ing. The top performers in the transmission system, ofthe auto industry have treated the relationship with scores of4, were simple outsourced and complex between Japanese manufacturers and their partially outsourced. The worst performers, with scores of 2, owned keiretsu suppliers as comparable to rela- were simple in-house. The performance of the simple tionships between U.S. and European automakers outsourced and simple in-house vehicles is consistent and their suppliers. That is, studies like Clark and with our predictions. Fujimoto (1991) have treated parts developed by The performance of the vehicle featuring the com- keiretsu suppliers as outsourced by the parent firms.22 plex transmission design with outsourced produc- However, in the complex-outsourced steering system, tion oftransmissions raises a measurement issue we as well as in many ofthe systems in the study, encountered with our data set. The system in question was highly complex, and the company in question 22 Clark and Fujimoto (1991) also treat fully owned subsidiaries produced the most integral components ofthe system such as Delphi as suppliers, where most empirical studies in-house, outsourcing only the simplest components. (Monteverde and Teece 1982, Masten et al. 1989, etc.) treat wholly On a content basis, their in-house production was not owned subsidiaries as in-house.
Management Science/Vol. 47, No. 1, January 2001 201 NOVAK AND EPPINGER Product Complexity and the Supply Chain all Japanese companies who outsourced individual strongly support the strategic importance ofthe prod- parts concentrated the more complex ofthose parts uct decision in the make/buy process. Given our in keiretsu suppliers, and outsourced simpler parts to observation that there are benefits to concentrat- financially separate suppliers. In contrast, most U.S. ing production ofcomplex systems in-house and and European companies typically outsource more to outsourcing simpler systems, efficiency arguments complex parts such as entire door systems or dash suggest that profit-maximizing firms should only assemblies to financially separate suppliers. operate according to these approaches. This raises The work ofmany researchers such as Asanuma the question ofwhy we ever observe firms behav- (1989), Helper (1995), Dyer (1996), and Fujimoto ing otherwise. We believe that this is a result ofthe (1989) indicates that the keiretsu relationship permits chronological and organizational separation ofthese richer information exchange between Japanese man- decisions in auto companies. Product design engi- ufacturers and their partially owned keiretsu sup- neers typically determine product architecture and pliers than between financially separate firms. If complexity. Purchasing agents typically make sourc- greater information sharing were possible between ing decisions. While these groups certainly inter- keiretsu firms, then the coordination problem encoun- act, they do not make these decisions jointly. Our tered by the firm in components development results suggest that greater coordination ofthese func- would be lower with keiretsu (versus non-keiretsu) tions within the product development process could suppliers. By component-based coordination cost, improve firm performance. Our findings also raise then, keiretsu sourcing may be closer to in-house pro- theoretical and empirical issues that we believe war- duction, and the complex-outsourced vehicle in ques- rant further examination. We detail two issues of pri- tion is closer to the complex in-house sourcing we mary concern below. hypothesize should be associated with greater qual- A major simplifying assumption of this paper is ity. We have treated keiretsu firms as out-of-house that sourcing can be treated as a binary decision— with respect to the parent firms to be consistent with either to make or to buy a part. This is done to our system-level analysis, as well as with existing be consistent with the simplest economic theory of empirical methodologies. This assumption, however, vertical integration. However, actual sourcing rela- also potentially understates the sourcing measure for tionships are more complex than simply make or Japanese firms in the study. buy. We observed other types ofcontracting arrange- Our evidence regarding quality suggests that there ments such as keiretsu relationships, joint ownership is not an optimal way to configure the firm or the agreements, equipment loans, and arms-length sub- product, but rather that multiple optima exist. This sidiaries, as well as make/buy practices. These prac- suggests that companies should not necessarily seek tices can create very different information structures, to emulate the “Toyota way” ofoutsourcing or BMW- with potential differences in the coordination costs style product development. Rather, our research sug- faced by the firm in a contracting relationship. We gests that a company that optimizes over both the believe that expanding the measures ofsourcing prac- requirements ofits product and the capability ofits tices is an important direction for future work on supply chain will outperform one which focuses only make/buy.23 In addition to the need to enrich the on firm structure or product characteristics. concept ofthe make/buy decision, our results also raise issues with regard to the information structure offirms. 6. Discussion and Conclusion We find that the quality benefits to designing sim- In summary, our model provides evidence ofcomple- pler systems for outsourcing as well as the qual- mentarity between product complexity and vertical ity penalties for attempting to outsource complex integration, as well as evidence ofclustering within the auto industry around high performance combi- 23 Ulrich and Ellison (1998) propose some alternative sourcing mea- nations ofthe two choice variables. These results sures in an empirical study ofbicycle sourcing.
202 Management Science/Vol. 47, No. 1, January 2001 NOVAK AND EPPINGER Product Complexity and the Supply Chain
systems outweighed the quality benefits ofin-house For each system, responses were translated into the 0–1 mea- production ofsimple systems as well as in-house pro- sure by equally weighting the answers (in most cases replies were duction ofcomplex systems. This suggests that a com- ranked from 0–5). For example, in the suspension system, key char- plex design is still difficult to execute in-house, and acteristics were newness, defined as the extent to which the suspen- sion had been used in the company; the number ofmoving parts that developing a simple part in-house does not nec- in the suspension; and degree ofactive suspension. For newness, essarily improve its quality over outsourcing such respondents were asked whether the suspension configuration had a part. been used before in the study vehicle type, in other vehicle types, In his 1988 review ofempirical work in transac- and ifthe front and rear suspension configuration had been used tion cost theory, Paul Joskow raises the question: Why separately in any vehicles. Responses were scored 0 for no expe- rience, 1 for front and rear used but not together, 2 for front and should information sharing among employees within rear used in different vehicle type, 3 for front and rear suspen- a firm be better than information sharing among sion new. This measure was then scaled to 0, 1/3, 2/3, 1 in our interested parties in a transaction? While our findings data set. For moving parts, the configuration with the fewest num- do not directly speak to this question, we believe that ber ofmoving parts, the McPherson strut system, was scored 0, we have identified an appropriate framework, that the SLA system was scored 1, air McPherson was scored 2, double ofcomplementarity between product complexity and wishbone and delta link systems were scored 3, and rescaled sim- ilarly. The active/passive measure was scored 0 for a passive sus- sourcing, through which to further explore this issue. pension, 1 for partially active, and 2 for a fully active and scaled as described above. The three measures were combined to yield a Acknowledgments score from 0 to 1. Support was provided by the International Motor Vehicle Program, For the body system, key characteristics were the number of the Center for Innovation in Product Development, the Interna- parts for the body side outer and inner, the number of sheet metal tional Center for Research on the Management of Technology, and thicknesses used, the type ofjoints used, and the number ofhits the Industrial Performance Center, all at M.I.T. Extremely valu- for the most complex part. For the brake system, key characteris- able comments were provided by Susan Athey, David Ellison, tics were the number ofchannels, number ofsolenoids, whether Charles Fine, Oliver Hart, Jerry Hausman, Rebecca Henderson, the system included a traction controller, and whether the ABS sys- Paul Joskow, and Nelson Repenning. The authors are very grate- tem was electrical or mechanical. Key characteristics for the trans- ful for the cooperation of the engineers and managers at our mission system were rear wheel vs. front wheel drive, gearsets, study companies and for assistance from Takahiro Fujimoto, Johan traction control, and automatic vs. manual. For the engine system, Lilliecreutz, and Kentaro Nobeoka. key characteristics were electronic control, traction control, trans- verse axis, and cam configuration. For the steering system, key Appendix A. Product Complexity Measures characteristics were adjustability, knuckle attachment, and airbag This appendix provides a briefoverview ofthe methods used to integration. For the electrical system, number ofmultiplexes, the evaluate product complexity, defined as all interactions affecting wiring configuration, and system integration (controls) were the the difficulty of coordinating changes during the product develop- key characteristics.24 ment process. We first compiled a list ofsystem-engineering princi- ples for product development using an extensive literature survey. Chiefengineers foreach system at each company were also asked References to list “key characteristics” most representative ofproduct com- Alexander, Christopher. 1964. Notes on the Synthesis of Form. plexity, and the lists were reviewed and combined. From this list, Harvard University Press, Cambridge, MA. we determined a set ofquestions to measure the system character- Asanuma, B., 1989. Manufacturer-supplier relationships in Japan istics. Experts from all of the participating companies were asked and the concept ofrelationship-specific skill. J. Japanese Internat. to review the list ofquestions, and to add or question any item. Econom. 3, 1–30. These reviews helped to limit the potential for bias in the ques- Athey, Susan, Scott Stern. 1998. An empirical framework for test- tions by involving a large number ofexperts. In some cases, as ing theories about complementarity in organizational design. with the body, brakes, and steering systems, differences in product Working paper, National Bureau ofEconomic Research, Inc., architecture, such as parts in the body side, mechanical vs. elec- No. 6600, Washington, D.C. trical ABS, and airbag integration, were seen to directly contribute Baldwin, Carliss, Kim B. Clark. 2000. Design Rules: Volume I, The to differences in coordination. In other systems, such as transmis- Power Of Modularity, MIT Press, Cambridge, MA. sion, factors related to electronic interaction and coupling, such as traction control, were more significant for coordination of develop- ment. The remaining factors reflect the list that was agreed upon 24 Interested readers may contact the authors for more on the ques- by the entire sample. tions asked and the measuring scale used.
Management Science/Vol. 47, No. 1, January 2001 203 NOVAK AND EPPINGER Product Complexity and the Supply Chain
Clark, Kim B. 1989. Project scope and project performance: The Jensen, M., W. Meckling. 1976. Theory ofthe firm: Manage- effect of parts strategy and supplier involvement in product rial behavior, agency costs, and capital structure. J. Financial development. Management Sci. 35 (10) 1247–1263. Econom. 3 305–360. , Takahiro Fujimoto. 1991. Product Development Performance, Joskow, Paul. 1988. Asset specificity and the structure ofverti- Strategy, Organization, and Management in the World Auto Indus- cal relationships: Empirical evidence. J. Law, Econom. Organ. 4 try. Harvard Business School Press, Boston, MA. 95–117. Dyer, Jeffrey H. 1996. Specialized supplier networks as a source of Klein, B., Robert G. Crawford, Armen A. Alchian. 1978. Vertical competitive advantage. Strategic Management J. 17(4) 271–291. integration appropriable rents, and the competitive contracting Eisenhardt, Kathleen M., Behnam N. Tabrizi. 1995. Accelerating process. J. Law Econom. 21 297–336. adaptive processes: Product innovation in the global computer Masten, Scott. 1984. The organization ofproduction: Evidence from industry. Admin. Sci. Quart. 40 84–110. the aerospace industry. J. Law Econom. 27 403–417. Eppinger, Steven D., D. E. Whitney, R. P. Smith, D. A. Gebala. 1994. , J. Meehan, E. Snyder. 1989. Vertical integration in the U.S. A model-based method for organizing tasks in product devel- auto industry. J. Econom. Behavior Organ. 12 265–273. opment. Res. Engrg. Design 6(1) 1–13. , , . 1991. The costs oforganization. J. Econom. Fine, Charles. 1998. Clockspeed: Winning Industry Control in the Age Behavior Organ. 7(1) 1–25. of Temporary Advantage. Perseus Books, Reading, MA. Monteverde, Kirk, David Teece. 1982. Supplier switching costs and , Daniel Whitney. 1996. Is the make versus buy decision vertical integration in the automobile industry. Bell J. Econom. a core competence? Working paper, MIT International Motor 13 206–213. Vehicle Program, Cambridge, MA. Pimmler, Thomas, Steven D. Eppinger. 1994. Integration analysis of Fujimoto, Takahiro. 1997. The Japanese automobile supplier system: product decompositions. Design Eng. 68 343–351. Framework, facts and reinterpretation. Working paper, Tokyo Rechtin, Eberhardt, Mark W. Maier. 1997. The Art of Systems Archi- University, Tokyo, Japan. tecting. CRC Press, Boca Raton, FL. . 1989. Organizations for effective product development: The Simon, Herbert A. 1969. Sciences of the Artificial. M.I.T. Press, case ofthe global automotive industry. Unpublished D. B. A. Cambridge, MA. dissertation, Harvard Business School, Cambridge, MA. Suh, Nam. 1990. The Principles of Design. Oxford University Press, Grossman, S., O. Hart. 1986. The costs and benefits ofownership: New York. A theory ofvertical and lateral integration. J. Political Econom. . 1999. A theory ofcomplexity, periodicity, and the design 94 691–719. axioms. Res. Engr. Design 11 116–131. Harrigan, Kathryn Rudie. 1986. Matching vertical integration strate- Ulrich, Karl. 1995. The role ofproduct architecture in the manufac- gies to competitive conditions. Strategic Management 7 535–555. turing firm. Res. Policy 24 419–440. Hart, Oliver. 1995. Firms, Contracts and Financial Structure. , David Ellison. 1998. Beyond make-buy: Internalization and Clarendon Press, Oxford, U.K. integration ofdesign and production. Working paper, Wharton , J. Moore. 1990. Property rights and the nature ofthe firm. School, University ofPennsylvania, Philadelphia. J. Political Econom. 98 1119–1158. , . 1999. Holistic customer requirements and the Hausman, Jerry. 1983. Specification and estimation ofsimultaneous design-select decision. Management Sci. 45 (5) 641–658. equation models. Z. Giriliches and M. Intriligator, ed. Handbook , Steven D. Eppinger. 1995. Product Design and Development. of Econometrics, Vol. 1, North-Holland Pub., New York, 403–426. McGraw-Hill, New York. , 1994. Valuation of new goods under perfect and imperfect , D. Sartorius, S. Pearson, M. Jakiela. 1993. Including the competition. Working paper 94–121, MIT Dept. ofEconomics, value of time in design-for-manufacturing decision making. Cambridge, MA. Management Sci. 39(4) 429–447. , D. Wise. 1978. A conditional probit model for qualitative Urban, Glen L., John R. Hauser. 1993. Design and Marketing of New choice. Econometrica 46. Products. 2nd ed. Prentice Hall, Englewood Cliffs, NJ. Helper, Susan. 1995. Supplier relations and the adoption ofnew Walker, Gordon, David Weber. 1987. Supplier competition, uncer- technology: Results ofsurvey research in the U.S. auto indus- tainty, and make or buy decisions. Acad. Management J. 30 (3) try. Working paper, National Bureau ofEconomic Research, 589–596. Inc., No. 5278, Washington, D.C. Whinston, Michael D. 1997. On the transaction cost determinants of , David Levine. 1992. Long term supplier relations and vertical integration. Working paper, Department ofEconomics, product-market structure. J. Law. Econom. Organ. 8(3) 561–581. Northwestern University, Evanston, IL. Holmstrom, B., P. Milgrom. 1994. The firm as an incentive system. Williamson, Oliver E. 1979. Transaction-cost economics: The gover- Amer. Econom. Rev. 84 972–991. nance ofcontract relations. J. Law, Econom. Organ. 22 3–61.
Accepted by Karl Ulrich; received November 7, 1998. This paper was with the authors 11 months for 4 revisions.
204 Management Science/Vol. 47, No. 1, January 2001