Sourcing parts of complex products: evidence on ...directory.umm.ac.id/Data Elmu/jurnal/A/Accounting...Milgrom and Roberts (1992) discuss two important ways in which make-or-buy decisions

Sourcing parts of complex products: evidence on transactionscosts, high-powered incentives and ex-post opportunism

Shannon W. Anderson a,*, David Glenn b, Karen L. Sedatole c

aUniversity of Michigan Business School, Ann Arbor, MI 48109-1234, USAbFaculty of Commerce and Business Administration, University of British Columbia, Vancouver, BC, Canada V6T 1Z2

cCollege of Business Administration, University of Texas, Austin, TX 78712-1172, USA

Abstract

This paper revisits evidence on the correlates of sourcing decisions in the US auto industry to see whether adoptionof new contracting terms and early involvement of suppliers in design activities (e.g. ``relational contracting'') yieldsdi�erent results as compared to previous ®ndings. Previous studies ®nd that US auto ®rms insource complex parts that

require investments in speci®c assets. Absent large di�erences in production costs, the results suggest that transactionscosts associated with external suppliers exceed transactions costs associated with internal suppliers (e.g. loss of highpowered incentives). Using data on 156 sourcing decisions for process tooling (dies) of a new car program we ®nd that

under the new relational contracting regime, transaction cost theory continues to have explanatory power for sourcingdecisions; however, attributes that favored insourcing in previous studies favor outsourcing in this setting. Moreover,more complex subassemblies are associated with fewer distinct suppliers than expected Ð evidence of a tendency to co-locate decision rights to reduce transactions costs related to system interactions. After controlling for transaction

characteristics that are associated with the sourcing decision, we ®nd no evidence that outsourcing is associated withincreased ex post opportunism by the ®rm (e.g. agreement about contract completion); however, outsourced parts aresubmitted by suppliers for evaluation signi®cantly later than insourced parts (e.g. delivery holdup). # 2000 Elsevier

Science Ltd. All rights reserved.

1. Introduction

With the emergence of hybrid organizationalarrangements, the distinction between marketsand hierarchies is blurred and the relation betweentransactions costs (Williamson, 1975) or ®rm cap-abilities (Porter, 1980, 1985; Prahalad & Hamel,1990) and e�cient organizational boundaries isclouded. New approaches to monitoring and

measurement and ``relational contracting'' between®rms and their suppliers introduce elements ofhierarchy to markets; while decentralization andinnovations in compensation contracts within the®rm introduce market elements to hierarchies(Daft & Lewin, 1993; Zenger & Hesterly, 1997).Previous studies ®nd that, in spite of the loss of``high powered incentives'' (pro®t motives) asso-ciated with hierarchies, US auto ®rms insourcecomplex parts that require investments in speci®cassets (human and physical capital) (Maston,Meehan & Snyder, 1989; Monteverde & Teece,1982; Walker & Weber, 1984, 1987). This paperexamines the relation between transaction attri-

0361-3682/00/$ - see front matter # 2000 Elsevier Science Ltd. All rights reserved.

PI I : S0361-3682(00 )00015-5

Accounting, Organizations and Society 25 (2000) 723±749

www.elsevier.com/locate/aos

* Corresponding author. Tel.: +1-734-647-3308; fax: +1-

734-936-8716.

E-mail address: [email protected] (S.W. Anderson).

butes, sourcing decisions and subsequent contractperformance after a US auto ®rm adopts sub-jective evaluation methods in conjunction withinvolving suppliers in early design activities (toge-ther termed ``relational contracting''). Theseapproaches are similar to those of Japanese auto®rms, where di�erent relations between transac-tions attributes and sourcing decisions areobserved (Dyer, 1996).Extensive vertical integration and the arms-

length nature of transactions with external suppliersthat typify US auto ®rms are a marked contrast to®rm boundaries and strategic supply relationshipsobserved in the Japanese auto industry (Cusu-mano, 1985; Dyer, 1996; Womack, Jones & Roos,1990). Faced with evidence that Japanese ®rmsdevelop new vehicles in a fraction of the time, USauto ®rms have focused on reducing a majorcomponent of vehicle development time Ð thetime to develop and build process tooling (dies)associated with metal stampings (Clark & Fuji-moto, 1991). Major changes include making diesourcing decisions earlier, so that suppliers areinvolved in all stages of die development, andmodifying the process for evaluating contractcompletion to emphasize subjective judgement andpart interdependencies. These changes resemblethe co-specialization of human assets and relianceon trust (as opposed to contract law) that Dyerdescribes as emblematic of supplier relations in theJapanese auto industry.In this paper we revisit evidence on the corre-

lates of sourcing decisions in the US auto industryto see whether adoption of new contracting termsand early involvement of suppliers in designactivities (e.g. ``relational contracting'') yields dif-ferent results as compared to ®ndings of previousstudies. Transactions cost economics might predictthat early supplier involvement favors insourcingto avert opportunistic appropriation by the ®rm ofinnovation activities (e.g. ``stealing'' supplier ideasin the development stage) (Walker, 1994). Similarly,subjective evaluation could be predicted to favorinsourcing as measures of contract completionbecome ambiguous (Milgrom & Roberts, 1992).However, studies have shown that decentralizedorganizations that use relative performance mea-sures to evaluate division managers experience

similar problems (e.g. managers haggle over inter-nal transfer prices, Eccles & White, 1988; Poppo,1995; Vancil, 1978). We examine the associationbetween transaction attributes and sourcing deci-sions, but make no directional predictions aboutrelational contracting favoring in(out)sourcingbecause, as Walker and Poppo (1991, p. 66) note,``. . . how the theory should be used as a predictorof shifts in the current boundaries of the corpora-tion is unclear.''We use data from a new car program to inves-

tigate the relation between 156 sourcing decisionsfor process tooling (dies) and proxies for contractuncertainty. A unique feature of the database, thephysical relation between parts and subassembliesof a single vehicle, is exploited to extend previousstudies. Speci®cally, early studies fail to detect``systems e�ects'' in part sourcing. We test thehypothesis that in addition to part-level char-acteristics, there are joint contract uncertaintiesthat a�ect sourcing decisions of parts that aremembers of a common subassembly. The hypothesisis motivated by the transactions cost literature oninterdependencies as well as recent developments inthe product development (e.g. Smith & Eppinger,1997b) and contracting (e.g. Baiman, Fisher &Rajan, 1999) literatures. We ®nd that both part andsystem characteristics in¯uence the probability ofoutsourcing. Part and system complexity are sig-ni®cantly related to outsourcing; however, in theopposite direction found in previous studies Ð thegreater the complexity of the part and its sub-assembly, the greater the likelihood of outsourcing.We also ®nd signi®cant part±system interactions:simple parts that are members of systems withmore design constraints are more likely to be out-sourced than similar parts in unconstrained systems.After including measures of part and subassemblycomplexity, there is still evidence that subassemblymembership accounts for signi®cant unexplainedvariation in part-level sourcing Ð there are systems-e�ects in sourcing. Finally, we provide evidence thatmore severe uncertainty at the subassembly level isassociated with sourcing parts to fewer suppliersthan would be expected by chance alone; that is, atendency to co-locate parts of a subassembly withfewer suppliers. In the extreme, this tendencymanifests itself as sole sourcing.

724 S.W. Anderson et al. / Accounting, Organizations and Society 25 (2000) 723±749

Cynics might argue that although US ®rmsadopt the rhetoric of relational contracting, theproof lies in subsequent incidence of opportunism.Although we do not have access to contract price(or internal cost) data, we investigate two measuresthat are sensitive to opportunistic behavior. The®rst measure proxies for opportunistic behavior ofthe ®rm and the second proxies for opportunisticbehavior of suppliers. We use as a measure ofopportunistic holdup of die suppliers, the frequencywith which die rework is demanded by the ®rm as acondition for contract completion. Controlling fordie complexity, we ®nd that the incidence of dierework requests is higher for insourced parts thanfor outsourced parts. Subassembly membershipexplains variation in the number of reworkrequests Ð in short, problems tend to be sharedby all members of a subassembly. Outsourcedparts that are members of subassemblies with noinsourced parts experience a slightly higher inci-dence of rework requests than outsourced parts inother subassemblies. These results indicate thatthe ®rm does not use rework to penalize out-sourced parts at the expense of insourced parts. Asecond measure of opportunistic behavior that weinvestigate is the incidence of supplier delays insubmitting parts for evaluation of contract com-pletion. We ®nd that on average, outsourced partsare submitted 70 days later than insourced parts.In sum, after controlling for transaction char-acteristics that are associated with the sourcingdecision, we ®nd little evidence that outsourcing isassociated with increased ex-post opportunism bythe ®rm (e.g. agreement about contract comple-tion); however, it appears that external suppliersengage in delivery holdup. We o�er an explana-tion of these ®ndings from our discussion withmanagers at the ®rm and at external suppliers.Milgrom and Roberts (1992) discuss two

important ways in which make-or-buy decisionsmay violate the theory that ®rms minimize thesum of production and transaction costs. First,production and transactions costs may not beseparable. Second, those making the sourcingdecision may not strictly maximize ®rm value iflocal (e.g. departmental) objectives con¯ict with®rm goals. This is particularly likely in ourresearch setting since the US auto industry was

one of the ®rst modern organizational bureau-cracies (Chandler, 1962). We use ®eld research toinvestigate the degree to which these assumptionsare violated in our research setting. We ®nd thatboth assumptions are problematic; however, all ofthe studies to which we compare our results usethe US auto industry as the research setting andthe adoption of relational contracting did notin¯uence intra-®rm boundaries in our researchsite. Thus the absence of strict adherence to ®rmpro®t maximization is unlikely to explain di�er-ences between our results and previous studies orto explain potentially di�erent decisions beingmade before and after the adoption of relationalcontracting. We posit that limitations associatedwith the assumption of separable production andtransaction costs are more likely to explain thesedi�erences. Speci®cally, we argue that previousstudies explain long term investments in produc-tion capacity, but provide limited insights aboutpart-level sourcing decisions given the existence ofcapacity. Although we do not have data thatwould allow us to explore directly changes in sour-cing decisions associated with adopting relationalcontracting, we present an explanation for howrelational contracting might alter the relationbetween production and transaction costs in a man-ner that could lead to di�erent sourcing decisions.Section 2 reviews the transactions cost literature

on boundaries of the ®rm with an emphasis on theunique aspects of sourcing complex, new products.Section 3 describes the research setting and vari-able measures. Section 4 presents the results ofempirical tests on the determinants of sourcingdecisions. Section 5 compares the incidence of ex-post opportunism for di�erently sourced dies.Section 6 discusses the results using informationgained through ®eld research at the ®rm and itssuppliers. Section 7 concludes with the propositionthat strategic supply chain management and thehybrid organizational forms that it portends placenew demands on accountants and accountingresearchers. First, it is less appropriate than everfor accountants who develop data for sourcingdecisions to focus narrowly on production costs.The value chain perspective of strategic cost man-agement with its focus on ``costs of ownership''rather than supplier price is essential. Second,

S.W. Anderson et al. / Accounting, Organizations and Society 25 (2000) 723±749 725

relational contracting and associated performancemeasures (e.g. supplier ratings) that lend hier-archical characteristics to markets indicate anopportunity to augment our understanding oftraditional organizational control systems.

2. Literature overview

2.1. Transactions costs and the boundaries of the ®rm

In his seminal paper, Coase (1937) identi®estransaction costs as the primary determinant ofthe boundaries of the ®rm. Ideally, contractsbetween buyers and sellers provide adaptationstrategies for all possible contingencies. However, thisrequires either certainty regarding the future eco-nomic environment or unbounded rational reasoning(knowing all possible future states). Transactioncosts arise because complete contracting is oftenimpossible and incomplete contracts give rise tosubsequent renegotiations when the balance ofpower between the transacting parties shifts (Wil-liamson, 1979, 1985). Transaction costs includethe costs associated with writing contracts as wellas the costs of opportunistic ``holdup'' at a laterdate. Although internal organization or ``hier-archies'' are posited to o�er lower costs of coor-dination and control and to avert subsequentopportunistic behavior, Alchian and Demsetz(1972) and Vancil (1978) discuss related problems ofdecentralized ®rms. A major concern is the loss of``high powered incentives'' when the pay-for-per-formance link is attenuated by internal production.Firms are hypothesized to choose organizational

boundaries to minimize the sum of productionand transactions costs (Williamson, 1975, 1979).Milgrom and Roberts (1992) identify ®ve attri-butes of a business exchange that are positivelyassociated with transaction costs: (1) the necessityof investments in durable, speci®c assets; (2)infrequency of transacting; (3) task complexityand uncertainty; (4) di�culty in measuring taskperformance; and, (5) interdependencies withother transactions. The necessity of early invest-ments in durable, transactions-speci®c assets (e.g.human or physical capital) shifts the balance ofpower between transaction participants, because

in later renegotiations these costs are sunk costs ofthe party that incurs them. Infrequent transactionsincrease the likelihood of opportunistic behaviorin later periods by reducing the threat of retribu-tion. In situations where broader market reputa-tions are at stake, infrequent transactions may besustainable; however, Klein (1988) points out thateven long-term contracts often do not providesu�cient adaptation mechanisms and in¯exibilitymay actually induce holdup. Task complexity,uncertainty and measurement problems exacer-bate the problem of identifying and contractingfor contingencies. Interdependencies introducecontingencies among transactions that suggest co-location (e.g. system-level sourcing) or that requirehigh level coordination.The ®ve transaction attributes indicate settings

in which opportunistic behavior is likely. If trans-actions costs o�set production cost advantages ofthe external supplier, the ®rm subsumes the activityÐan outcome termed ``vertical integration'' or`ìnsourcing.'' Empirical research indirectly teststransactions cost theory by relating observed sour-cing decisions to transaction attributes that proxyfor transactions costs. Evidence on the relationbetween transactions-speci®c investments (e.g.Joskow, 1987, 1988; Klein, 1988; Masten, 1984;Masten, Meehan & Snyder, 1989; Monteverde &Teece, 1982; Novak & Eppinger, 1998; Ulset,1996), contract duration (Crocker & Masten,1988; Joskow, 1987, 1988), and technologicaluncertainty (Ulset) generally supports the theory.1

The consistency of the empirical results is star-tling in light of two problems with the hypothesisthat ®rms take sourcing decisions to minimize thesum of production and transactions costs (Mil-grom & Roberts, 1992). First, production andtransactions costs are rarely neatly separable. Forexample, the choice of production technology (andsubsequent production costs) is often inextricablylinked with production volume, which in turndepends on whether the ®rm produces some or allproducts internally. Second, decision-makers arelikely to be a�ected by wealth e�ects associatedwith sourcing, and thus are unlikely to take decisions

1 Shelanski and Klein (1995) provide a comprehensive

review of empirical research on transaction costs.


that strictlymaximize ®rmpro®t.Walker andWeber(1994) hypothesize that production and transact-ions costs di�er in ``the way that they are mappedonto the decision-making process. (p. 378)''Because production costs are objectively ``calcu-lated'' by the accounting system, while transactionscosts are assessed subjectively through indirect indi-cators, ``functional managers are likely to di�er inthe importance that they assign to reducing trans-action costs. . . Consequently, the e�ect transactionscosts have on a make-or-buy choice can partlyre¯ect the in¯uence exerted by the purchasing man-ager.'' They ®nd that production cost di�erences aremore in¯uential in sourcing decisions than transactioncost di�erences and that experience of the decision-maker is related to assessments of technologicaluncertainty. Poppo (1995) ®nds that pro®t centermanagers engage in in¯uence activities that increasethe costs of price renegotiations above the level that isobserved in comparable external market transac-tions. Whyte (1994) and Roodhooft and Warlop(1999) provide experimental evidence that man-agers are more reluctant to outsource wheninvestments in speci®c assets are necessary; andcontrary to theory, managers consider previousinternal investments in speci®c assets (e.g. sunkcosts) a reason to insource. Finally, studies thatfocus on cognitive limitations indicate that in certaincircumstances decision-makers systematically mis-estimate (or fail to consider) transaction costs(Drtina, 1994; Lacity, Willcocks & Feeny, 1996).

2.2. The role of transactions costs in sourcingdecisions for complex, new products

This paper considers collaborations betweenproducers and suppliers on development of pro-cess tooling for new products. Although thesecollaborations have the potential to make bothparties better o�, unique di�culties of accountingand contracting for the division of spoils frominnovation activities may preclude collaboration.For example, the likelihood of design changes andrework in new product development creates awindow for opportunistic behavior and, antici-pated by both parties, in¯uences contract terms.Delaying supplier selection until the productdesign is ®nalized may reduce recontracting; how-

ever, it lengthens product development time andoverlooks the fact that development activities areoften the only way to discover design ¯aws (Wheel-wright & Clark, 1992). A di�erent approachinvolves sourcing tools earlier, based on incompletepart speci®cations and ``soft'' estimates of the ®naltool cost, and involving the supplier in developmentof ®nal speci®cations (Birou & Fawcett, 1994; Boz-dogan, Deyst, Hoult & Lucas, 1998; Krishnan,Eppinger &Whitney, 1995a, 1995b; Tottie & Lager,1995; Ward, Liker, Cristiano & Sobek, 1995).Reduced development time may arise as a result of awider range of novel solutions o�ered by the sup-plier or more rapid convergence on the best diedesign (Bidault, Despres & Butler, 1998; Clark,1989; Fitzgerald, 1999; Smith & Eppinger, 1997a).These bene®ts must be weighed against the nearcertainty of recontracting and associated costs.When products are assemblies of complex, rela-

ted components, each potentially sourced from adi�erent supplier, the traditional notion of the®rm as a ``nexus'' of many bilateral supply con-tracts (Jenson & Meckling, 1976) becomes lessappropriate. Suppliers a�ect and are a�ected byother suppliers. The ®rm is one node, albeit cen-tral, in a network of interdependent players and anovel form of joint transactions costs must beconsidered in determining whether parts should beproduced by the ®rm or purchased from externalsuppliers (Armour & Teece, 1980). Design inter-dependencies may give rise to central coordinationor to ``sole sourcing'' (Milgrom & Roberts, 1992).Like individual part speci®cations, part interac-tions are often uncertain and ill-de®ned at the timethat a sourcing decision is taken (Smith & Eppin-ger, 1997b).2 One form of uncertainty is associated

2 An example of an ill-de®ned, uncertain interaction is ``tol-

erance buildup'' between assembled parts (Taguchi, Elsayed &

Hsiang, 1989). Tolerance buildup refers to situations in which

two randomly selected parts cannot be assembled properly or,

when assembled, do not jointly meet design speci®cations for

the assembled product in spite of each part being drawn from a

population of parts that are on average at design nominal and

within design tolerances. In cases of ¯exible materials, uncer-

tain tolerance distributions or uncertain covariance among

parts, designers typically set individual part tolerances using

simple design rules and rely on empirical observation in proto-

typing to reveal areas that require tolerance revision (Boot-

hroyd, Dewhurst & Knight, 1994).


with assigning fault to a particular part when twoor more assembled parts fail to jointly satisfy the®rm's objectives.3 This interdependency creates atype of opportunistic behavior that may inhibitcollaboration with suppliers when part-level inter-actions are poorly understood Ð opportunisticshifting of blame for contractual failures amongsuppliers of interacting parts. The ®rm may bemore motivated to engage in this behavior whenone or more parts of a system are insourced;however, one can also imagine this arising if cer-tain external suppliers (e.g. ``tier-1'' suppliers) arefavored or are subject to di�erent transactionscosts (e.g. uncertainty is resolved more or lessquickly in the course of product development).Forces that temper the risks of opportunism

include the expectation of future transactionsbetween the two parties and the reputational con-sequences in the broader marketplace to blatantopportunism. As more ®rms engage in suppliercollaborations on new product development, therehas been a concomitant increase in the practice ofconcentrating purchases among a small group of``®rst tier'' suppliers.4 By forming partnershipswith a few key suppliers, ®rms increase transactionfrequency. Although purposeful concentration ofpower among a few suppliers seems to controvertbasic arguments of strategic management [e.g. oneof Porter's (1980) ``®ve forces'' of competition issuppliers], Richardson (1993) and Wasti and Liker(1997) describe how Japanese ®rms rely on parallelsourcing, repeat contracts, shared technologicalknowledge and reputational consequences to sta-bilize this seemingly fragile equilibrium. Moreover,it is unclear that decentralized ®rms that separateinnovation activities (e.g. die development) fromthe use of the innovation (e.g. production facilities)

are immune to these problems when bureaucraticine�ciency and shirking are considered.

2.3. Research on sourcing and transactions costs inthe US auto industry

Automobiles are complex assemblies of manymanufactured parts. A large number of complex,interdependent parts, uncertainty in transactionsand the possibility of signi®cant transactions-speci®cinvestments, have made auto manufacturing anatural laboratory for testing transactions costtheory. Monteverde and Teece (1982), Walker andWeber (1984, 1987), Masten, et al. (1989), Dyer(1996) and Novak and Eppinger (1998) exploitthis setting. Our data do not allow us to performwithin-®rm comparative analysis of determinantsof sourcing decisions before and after the adoptionof relational contracting. Consequently, we usethese studies as a benchmark for how sourcingdecisions were made in the US auto industrybefore supply chain management strategies werewidely considered. This section brie¯y describeseach study.Monteverde and Teece (1982) examine average

sourcing outcomes as judged by an industry pro-fessional for 133 categories of parts for two USauto makers in 1976. They ®nd that insourcing ispositively related to engineering design e�ort andthe degree to which part designs are ®rm-speci®c.There is little evidence that system membership isrelated to sourcing, and the authors do not investi-gate whether ``sole sourcing,'' an extreme solutionof co-locating parts that interact, is more likely inthe presence of system e�ects. One explanation ofthe failure to detect system e�ects on individualpart sourcing decisions is that system membershipis de®ned in relation to assembly sequence andautomotive functions (e.g. transmission parts). Itis unclear that these are relevant ``systems'' forsourcing decisions.5 Masten et al. (1989) extendMonteverde and Teece by demonstrating thatinvestments in transactions-speci®c human capital

3 Baiman, Fisher and Rajan (1999) consider design inter-

dependencies in contracting for supplier performance.4 Increased interest in monitoring and measuring supplier

performance is a recurring theme in the literature (Ittner &

Ponemon, 1992). As outcomes are better de®ned and perfor-

mance more easily measured, these measures may be used in

enforceable contracts Ð thereby permitting a broader array of

collaborations to ®t within traditional market exchanges. Zen-

ger & Hesterly (1997) discuss the use of control mechanisms in

markets.

5 The 156 stamped metal parts that we examine are together

only one of the 133 categories of parts (termed ``body sheet

metal'') that Monteverde and Teece (1982) consider. Body

sheet metal is one of 37 part categories that comprise the `àuto

body'' system.


(e.g. engineering design e�orts) have greater in¯u-ence on sourcing decisions than investments intransactions-speci®c physical capital. They do notconsider system e�ects.Walker and Weber (1984) investigate determi-

nants of 60 sourcing decisions taken within acomponent division of a large US auto manu-facturer over a 3 year period. Factors consideredto in¯uence transactions costs were: supplier mar-ket competition and uncertainty associated with®nal product volume and with product designchanges. The study is unique in its considerationof the decision-making process. Aside from trans-actions costs, factors hypothesized to in¯uence thesourcing decision, are buyer experience and com-parative production costs between the supplierand the ®rm. Production cost di�erences arefound to be the strongest predictors of sourcingoutcomes with volume uncertainty and suppliermarket competition having small but signi®cante�ects that are consistent with transaction costtheory. Walker and Weber (1987) revisit theresults to investigate the possibility that competi-tion in supplier markets moderates the relationbetween sourcing decisions and transaction costs.When supplier markets are not discriminated, neithertype of uncertainty in¯uences sourcing decisions.However, distinguishing high and low competitionsettings, volume uncertainty is associated withinsourcing in low competition markets and productdesign uncertainty is associated with outsourcingin high competition markets. In both cases, pro-duction cost di�erences are signi®cant predictorsof sourcing.Dyer (1996) compares a sample of 50 supplier±

®rm relationships for US (3) and Japanese (2) auto®rms. Purchasing managers in each ®rm identi®ed25 suppliers with whom they worked closely and25 suppliers with whom they had arms-lengthrelationships. For each supplier±®rm pair, surveydata was gathered on investments in speci®cassets, mechanisms employed to guard againstopportunistic behavior (e.g. legal contracts, trust,reputation), and information exchange. Consistentwith previous studies, US ®rms use hierarchieswhen asset speci®city is high and arms-lengthmarket relations otherwise. Japanese ®rms makegreater use of hybrid governance forms that share

characteristics of markets and hierarchies. Dyerconcludes that this form yields co-specialization ofhuman assets that results in superior coordinationand information sharing between ®rms withoutlosses associated with hierarchies (e.g. loss of highpowered incentives).Finally, Novak and Eppinger (1998) use the

automobile as the unit of analysis and examine thepercent of components insourced for 134 luxuryvehicles produced by several worldwide auto®rms. They ®nd that architectural complexity, aunion contract with provisions for in-house com-ponent production, and membership of the vehiclein a platform car program that shares parts, areassociated with insourcing. They ®nd no evidenceof systems e�ects related to complexity of sevenauto systems (e.g. auto body).In summary, previous studies have consistently

found that sourcing decisions in US auto ®rmsconform to the predictions of transactions costeconomics. Controlling for di�erences in produc-tion costs between the ®rm and its suppliers, assetspeci®city, task complexity and task uncertaintyare strong predictors of insourcing. Evidence onthe relation between transaction interdependenciesand part-level sourcing decisions Ð so called``systems e�ects'' Ð is inconclusive. This paperinvestigates whether adoption by a US ®rm ofrelational modes of contracting that are similar tothose used by Japanese ®rms, yields di�erent rela-tions between part and system characteristics thatre¯ect transaction costs and sourcing decisions ascompared to these studies.

3. Research setting and variable measures

3.1. Die development and associated transactionscosts

Stamped metal parts are made by passing metal``blanks'' (e.g. pre-cut metal rectangles) throughautomated presses that hold ``dies.'' Dies aremolds with upper and lower parts that whenforced together, impart shape to metal. Typicallyseveral dies are used in sequence to gradually cut,trim and shape the desired part. Dies that producea single part represent one sourcing decision. All


die sourcing decisions for a vehicle are assumed tooccur simultaneously Ð that is, die development isnot sourced in ``rounds'' with new informationpotentially becoming available between decisions.The cost of dies for stamped metal body parts

accounts for approximately half of the capitalinvestment for producing a new automobile inexisting manufacturing facilities. Die developmentactivities account for the majority of the productdevelopment cycle time.6 All US auto manu-facturers have internal die development facilities;however, no ®rm has enough die shop capacity toperform all of its own die development, repair andmaintenance.7 Agreements with the labor unioncreate conditions that favor using existing dieshops to full capacity; however, no die set in oursample was destined for insourcing as a result ofunion agreements. Similarly, no die set was destinedfor outsourcing; the internal die shops are capableof producing all dies, had su�cient capacity toproduce any die set, and could be asked by thepurchasing department to submit cost estimatesfor any job. Thus we investigate variation in diesourcing, conditional on a certain (unknown)internal capacity utilization for die development.8

Although capital investments associated withdie development technology are substantial, theyare not transaction-speci®c. Suppliers may havespecial skills with certain parts (e.g. skill in pro-ducing dies for auto doors); however, their com-petence is not linked to a particular automanufacturer or vehicle. Un®nished, in-process

dies are clearly transactions-speci®c assets, havingno alternative use other than production of partsfor which they are designed. However, this is trueof all the dies that we examine, so transactions-speci®c investments are unlikely to explain varia-tion in die sourcing decisions among stampedparts. Similarly, the group of eligible suppliers isroughly constant for all of the sourcing decisions,so we can rule out di�erences in the intensity ofcompetition in supplier markets as explainingsourcing decisions (Walker & Weber, 1984, 1987).We turn instead to uncertainty in contracting

for dies as the source of transactions costs thatin¯uence die sourcing decisions. Although thereare several approaches for contracting for diedevelopment, the most common, and the one usedfor all of the parts that we study involves themanufacturer seeking bids for die development onthe basis of detailed part speci®cations and pre-liminary plans for the dies. Although the ®rm isbuying dies, contract completion is de®ned by diesthat produce parts that conform to design speci®-cations. While preliminary die designs may helpthe supplier develop an initial price quote, pro-duction of the die design is not a condition forcontract completion.Die development is subject to two sources of

uncertainty.9 The ®rst source of uncertainty isassociated with estimating the resources requiredto produce dies that yield parts that meet speci®-cations. Uncertainty in estimating inputs stemsfrom ``dimensional bias'' Ð the dimensions ofparts produced by a die are rarely the same asthose of the die because material ¯ows underpressure are unpredictable. Relating die dimen-sions and stamping press forces to part dimensionsis an art. Dies are produced, tried and reworked toremove dimensional biases (e.g. grinding highspots) in a costly trial and error process.The second uncertainty in die development is

associated with judgement and subjectivity inascertaining that parts meet speci®cations. In thepast, auto ®rms considered a die development con-tract complete when dies consistently produced

6 Clark and Fujimoto (1991, p. 184) document dramatic dif-

ferences between US automobile manufacturers, who require

14±16 months to develop dies, and Japanese competitors who

accomplish the same task in 6 months. Despite spending longer

in developing and testing the dies, US ®rms experience more

quality problems in production start-up and require more time

to reach stable production and acceptable quality (p. 191).7 Many years ago, the ®rm selected a level of internal die

development capacity that is signi®cantly less than the demand

represented by all die development and all die repairs. Using a

cuto� of 80% internal production to de®ne insourcing, Mon-

teverde and Teece (1982) would have coded die development as

primarily `òutsourced'' work. We seek to explain which jobs are

(in)outsourced given that only a portion of die development can

be done internally.8 We do not have data on the anticipated capacity utilization

of internal die shops at the time that the vehicle program was

sourced.

9 Hammett, Majeske and Baron (1998) and Liu, Hu and

Woo (1996) are engineering discussions of uncertainty in metal

stamping.


parts that conformed to dimensional speci®ca-tions. However, there is considerable subjectivityin evaluating ``consistent performance to speci®-cation.'' The evidence is typically the dimensionalperformance of a small sample of parts run on dieshop presses. Uncertainty in evaluating the evi-dence is caused by run-to-run variation, which canresult from inconsistencies between die set ups,sheet metal batch-to-batch variation, or changesof critical press settings. Questions of adequatesample size and e�ects of transferring the dies tothe production presses at the ®rm's fabricationplant enter into subjective evaluation of whetherthe dies are satisfactory.10 Design speci®cationsthat are excessively tight and somewhat arbitraryfurther complicate subjective evaluation (Ander-son & Sedatole, 1998). For example, designersmay assign a tolerance of plus or minus 1.0 mm asa ``rule of thumb'' rather than because a largerdimension would compromise part ®t or perfor-mance or because less dimensional variationwould be unnecessarily costly. Similarly, designersmay identify a large number of measurement pointsin an e�ort to constrain part shape. However, thinmetal conforms to its supporting structure, and asa result, di�erent measurements of the same partcan di�er depending on how the part is supportedduring measurement. In sum, strict design tolerancesfor metal parts often increase die developmenttime and cost without producing substantiallybetter or di�erent parts.

3.2. A new approach to supplier relations andcontract evaluation

In an e�ort to reduce vehicle development time,in the 1990s US auto ®rms began sourcing dies

earlier, based on incomplete part speci®cations,and adopted processes for evaluating contractualperformance that incorporated subjective evalua-tions of whether parts ``®t'' with other parts(Hammett et al., 1998). These approaches resembleapproaches used by Japanese manufacturers. Inaddition to data on part-level dimensional perfor-mance, Japanese ®rms evaluate ``screw bodies'' Ðparts that are screwed or riveted together to formsubassemblies that are joined to create the autobody. At an early stage in the die developmentprocess, after the dies are capable of producingparts without deformity but before any ®ne tuninghas taken place to achieve speci®ed part tolerances(e.g. mean and variance of part dimensions), partsare assembled. If the assembled parts are jointlysatisfactory, the dies are declared complete andpurchased at full price even absent the ability tomeet individual part speci®cations. Die develop-ment time and cost can be dramatically reduced;however, contract completion for any single partis intertwined with that of other parts in the sub-assembly. If parts do not pass subassembly eva-luation, the supplier is given suggestions for dierework. Fig. 1 depicts the hierarchical con®gura-tion of subassemblies that are evaluated. Eachcircular ``node'' indicates a subassembly that isevaluated to ascertain contract completion fordies. Hierarchically related nodes (``tiers'') re¯ectthe gradual build-up of the auto-body. First tiersubassembly evaluations are targeted for a parti-cular date and occur as soon as all parts are sub-mitted. Higher tier subassembly evaluations havelater target dates.11 If no acceptable subassemblycan be created from parts that do not meet speci-®cations, then the default is the old mode ofcontracting Ð the supplier must build dies that

10 Die shops test dies on try out presses that di�er from

production conditions. Die shops typically do not have a line of

presses (or a transfer press) connected by automatic material

handling equipment. Rather, they mount the ®rst die of a die

set in the tryout press and manually produce the parts. Then

they remove the ®rst die, mount the second die and manually

perform the second operation on the parts. This continues

through the die set until the parts are complete. The alternative

of moving the dies to the actual production line for tryout is

rarely used because it is too costly to remove a full line of

stamping presses from production.

11 Higher tier contracts are not necessarily ``held up'' by lower

tier contracts. Lower tier contracts that are judged in compliance

prior to higher tier evaluations generally are not ``revisited''

under the same contract. If the ®rm believes that problems diag-

nosed at a higher tier can be remedied by changes in lower tier

parts previously judged to be in compliance, then a second con-

tract for an engineering change would be needed. If the dies have

been transferred to the production facility, internal die facilities

may perform the rework.


make parts that meet the dimensional speci®ca-tions.12

3.3. Measurement of variables

The data are product speci®cations, sourcingdecisions, and subsequent evaluations for the diesets of 170 stamped metal body parts for a new carprogram of a major automobile manufacturer.13

Fig. 1 depicts the hierarchical relation between 170stamped parts of the vehicle and the 32 sub-

assemblies in which they are evaluated.14 Numbersin squares indicate parts that ®rst enter each sub-assembly. Missing part speci®cation data (in par-entheses) limits the analysis to 156 sourcing decisionsthat resulted in 23 external and eight internal diesuppliers being awarded die development con-tracts. In contrast to previous studies, our ``sys-tems'' re¯ect terms of the sourcing contract(groups that are evaluated for contract com-pliance) rather than post-hoc groupings thatre¯ect automobile functions or the sequence ofvehicle assembly. We use measures of character-istics of the 156 parts and the 32 subassembliesthat they comprise to test the hypothesis thatsourcing decisions are jointly in¯uenced by partand subassembly complexity. The data alsoinclude information about outcomes of part-levelevaluations within the subassembly of which it is®rst a member, including the date on which partswere submitted by the supplier for evaluation and

Fig. 1. Auto body buildup sequence for part evaluation.

12 If a product design ¯aw is discovered, an engineering

change is made to the part speci®cation and the ®rm and sup-

plier negotiate the price of die rework. In interviews with sup-

pliers, it became clear that the likelihood of engineering

changes is considered when quoting prices for die development.13 The car is a mid-size, mid-price passenger car that was

introduced for the ®rst time during the 1990s. The vehicle was

new at the time of the sourcing decisions. It is not a ``reskin-

ning'' of an existing car line or a ``carryover'' nameplate. This is

important because it limits the degree to which historical rela-

tionships between suppliers and the ®rm may in¯uence speci®c

sourcing decisions. Clearly, however, certain suppliers may be

more likely than others to win contracts as a result of historical

relations with the ®rm on di�erent car programs.

14 One of the 33 subassemblies depicted in Fig. 1 is strictly

evaluated for subassembly ®t, with no new parts included;

consequently, it is not included in the analysis.


whether the supplier was required to perform dierework. We use this information to constructproxy measures for ex-post opportunism.The dependent variable for the part-level sour-

cing analysis, INSOURCE, takes on the value 1 ifthe dies for the part are built by the auto companyand 0 if the dies are built by an external supplier(outsourced). The dependent variable used toinvestigate factors that contribute to sole sourcingof parts from a particular subassembly is the dif-ference between the actual number of di�erentsuppliers and the expected number of di�erentsuppliers for a given subassembly. The expectednumber of suppliers is the number expected whenN suppliers are selected with replacement from apopulation of 31 internal and external suppliers,15

where N is the number of parts in the subassembly.Thus the variable, CONCENTRATION, repre-sents a crude attempt to measure `àbnormal''deviation in the total number of suppliers thatemerge for a subassembly following the pursuit ofany particular sourcing strategy, given the numberof parts that comprise the subassembly and thepopulation of suppliers from which to choose.16

Negative values (using fewer suppliers thanexpected) suggest an overt strategy to co-locateparts of a subassembly. For the sample of 32 sub-assemblies, CONCENTRATION takes a valueless than zero for 22 subassemblies and rangesfrom ÿ7.50 to +0.18 with a mean of ÿ1.87 and astandard deviation of 2.25. On average, fewersuppliers are used than would be suggested bychance alone.We use two dependent variables as proxies for

the extent of opportunistic behavior by the ®rmand by suppliers after the contract is awarded. Onemeasure, REWORK, is an integer that indicates

the frequency with which subassembly evaluationsresulted in the die supplier being asked to performdie rework under terms of the original contract.By demanding rework, the ®rm increases suppli-ers' costs with no recourse for renegotiating theprice (unless the rework is in conjunction with anengineering change). A second measure that iscontrolled by suppliers, DELAY, is the number ofdays between the date on which parts are reques-ted by the ®rm to be available for subassemblyevaluation and the date that parts are actuallysubmitted. By delaying part submission, the sup-plier may decrease the ®rm's propensity todemand rework or increase the likelihood thatanother part within the subassembly (produced bya more compliant supplier) is targeted for rework.Table 1 presents descriptive statistics for thedependent and independent variables at both thepart and subassembly-level.Two measures are used to proxy for contract

uncertainties in die development:17

LNCONPNT: The natural log of the numberof dimensional control points on the partdesign. On part drawings, designers identifycritical points where completed parts must bemeasured to ascertain dimensional perfor-mance (e.g. mean and sample variation). Diedevelopment contract compliance is judged inpart by measuring performance to speci®ca-tion at these critical points. Uncertaintyabout material ¯ow properties causes design-ers to select more control points in areas ofthe part with unusual curvature, in very thinregions and in areas that adjoin other parts,where ``®t and ®nish'' are critical. Thus, thenumber of critical control points is a proxyfor part complexity and uncertainties in con-tracting for die development. A log transformis used because the data are skewed by severalparts with a large number of critical points ascompared to the remaining sample.

15 The analysis was repeated using a population of 24 sup-

pliers (that is, treating the eight internal suppliers as one), with

no qualitative change in the results. The results are also quali-

tatively similar using a simple dichotomous measure set equal

to one for parts with fewer than expected suppliers and zero

otherwise.16 A limitation of our study is that we have no data on

unique supplier capabilities or on which suppliers were con-

sidered for a given part. Consequently we assume that the

population from which the actual supplier was drawn includes

all suppliers that eventually received at least one part.

17 A third indicator for whether the part is an outer body

panel was considered but found to have no relation to sourcing

decisions or subsequent contract outcomes. Consequently, it is

not included in the reported results.


LNTHICK: The natural logarithm of metalthickness (in millimeters) for the part. Therigidity of thicker parts makes them easier tomeasure with precision relative to designspeci®cations and tolerances. As a result,there is less uncertainty in ascertaining whe-ther design speci®cations have been met.Although it is more di�cult to resolve issuesof performance relative to design speci®ca-tions for thin parts, the parts are more easilyaccepted in a less-than-perfect state. Thinparts are often married to rigid parts in the

sub-assembly; consequently part-level imper-fections may be o�set by systems e�ects.

Novak and Eppinger (1998) use similar mea-sures in their composite vehicle ``body complex-ity'' scale, created by summing the number of:body side outer parts, the number of innerparts, the number of di�erent metal thicknessesfor all parts, the number of joints for all partsand the maximum number of die strikes for anypart. The engineering literature on metal stamp-ing also supports these proxies for contract

Table 1

Descriptive statistics

Panel A. Part-level variables

The level one (N=156 parts) dependent variable, INSOURCE, is an indicator variable equal to one for parts that are insourced and

zero for parts that are outsourced. REWORK, is the number of times that a supplier was formally asked to perform die rework as a

result of subassembly evaluations. DELAY is the number of days between the request date and submission by the supplier of parts

for evaluation. The level one predictors are the natural logarithm of part metal thickness (LNTHICK), the natural logarithm of the

total number of critical measurement points (LNCONPNT), and the number of days between product release and the requested date

for part submission to evaluation (LEADTIME).

Pearson and Spearman correlations (above/below diagonal)

Mean S.D. INSOURCE REWORK DELAY LNTHICK LNCONPNT LEADTIME

INSOURCE 0.40 0.49 ÿ0.00 ÿ0.10 0.15� ÿ0.22*** 0.06

REWORK 0.65 1.00 0.01 0.12 ÿ0.04 ÿ0.32*** 0.23***

DELAY 77.46 75.39 ÿ0.12 0.17** ÿ0.20** 0.24*** 0.33***

LNTHICK 0.16 0.43 0.15� ÿ0.04 ÿ0.25*** ÿ0.22*** ÿ0.00LNCONPNT 1.83 0.58 ÿ0.21*** 0.32*** 0.29*** ÿ0.24*** 0.26***

LEADTIME 473.65 62.86 0.07 0.21** 0.35*** 0.03 0.22***

Panel B. Subassembly-level variables

The level two (N=32 subassemblies) predictors are the number of new parts that enter each separately evaluated subassembly

(NUMPARTS), an indicator variable (HI-TIER) equal to one if the subassembly is at the lowest tier of the vehicle build sequence

(see Fig. 1) and zero otherwise, the average number of critical measurement points per part in the subassembly (AVLNCONPNT),

and the average thickness (AVLNTHICK) for the new parts within each the subassembly.

Pearson and Spearman correlations (above/below diagonal)

Mean S.D. NUMPARTS HI-TIER ALNTHICK ALNCONPNT PROPIN

NUMPARTS 6.22 3.62 0.07 0.33� ÿ0.14 ÿ0.34�HI-TIER 0.34 0.48 0.06 0.00 ÿ0.06 ÿ0.11ALNTHICK 0.12 0.28 0.42** ÿ0.00 ÿ0.30� 0.00

ALNCONPNT 1.78 0.26 ÿ0.15 0.23 ÿ0.35** 0.09

PROPIN 0.47 0.33 ÿ0.34� ÿ0.10 0.02 0.02

�,��,��Correlation is signi®cant with P-value 40.10, 0.05 and 0.01 (two-tailed test).


uncertainty (e.g. Hammett et al., 1998; Liu etal., 1996).We hypothesize that four subassembly-level

characteristics in¯uence die sourcing decisions: thenumber of parts in the subassembly (NUMPARTS);an indicator for whether the subassembly is amember of a second tier or higher subassembly(HI-TIER; see Fig. 1); and two measures of theaverage part complexity of the subassembly,ALNCONPNT and ALNTHICK. The latter twovariables are averages of the part-level character-istics, LNCONPNT and LNTHICK, for the partsin a subassembly (treating lower-tier subassembliesas one ``part''). The other variables are describedbelow. Together these four variables proxy forjoint transactions costs that arise among inter-dependent parts.

NUMPARTS: The number of parts ®rstincluded in a subassembly and the number oflower level subassemblies that are evaluatedtogether for contract compliance (e.g. for asecond tier subassembly that is formed byjoining two ®rst tier subassemblies and addingfour parts, NUMPARTS equals 6). AsNUMPARTS increases, there are moreopportunities for parts to interact with oneanother (e.g. more mating surfaces of adjacentparts). The greater the number of adjoiningparts, the more likely are part interactions thatgive rise to joint transactions costs.

HI-TIER: The 32 automobile subassembliesare in a nested, six-level structure. Data limita-tions preclude modeling the fully nested struc-ture.18 Nonetheless, uncertainties associatedwith producing dies for parts in tier-1 sub-assemblies are di�erent from those associatedwith parts in tier-2 and higher subassemblies.Tier-2 and higher parts are married to a pre-existing subassembly that by virtue of assemblyis somewhat rigid. Related to the concepts of

part thickness, parts that are added to highertier subassemblies are easier to evaluatebecause of the subassembly support structure.Another potential di�erence between tier-1and higher tier parts is the information setting.Although it is not necessary, it is possible fordies for tier-1 parts to be judged in contractcompliance before tier-2 and higher parts areevaluated. If this occurs, then tier-2 andhigher parts are somewhat constrained to ®twithin the envelope of the tier-1 subassembly.By judging lower tier parts in compliance, the®rm limits its freedom to later shift blame fornoncompliance to these parts. As an extremeillustration, if all tier-1 parts achieve com-pliance before tier-2 subassemblies are eval-uated, then if noncompliance is observed attier-2 and there is ambiguity about the cause,only those parts that enter at tier-2 are can-didates for blame. While it is true that tier-1compliance is rarely revisited, higher levelevaluations are not detained by lower-levelcompliance. Nonetheless, evaluations occur insequence, and thus it is reasonable to hypothe-size that ®rst tier subassembliesÐ for which noprior evaluations exist, are in a di�erent infor-mation state than higher tier subassemblies.We use an indicator variable, HI-TIER, todistinguish the 11 upper tier subassemblies.

4. Analysis of sourcing decisions

4.1. Part-level sourcing: model speci®cation

Part-level sourcing is modeled as a binary choicebetween insourcing dies associated with part i, insubassembly j, (INSOURCEij=1) or purchasingthem (INSOURCEij=0). The logistic distributionlinks the binary choice to part-level (X) and sub-assembly-level (Y) characteristics that proxy forcontractual uncertainties. Applying a logistictransformation and incorporating the hypothe-sized relation between parts and subassemblies,yields a hierarchical generalized linear model thatcan be estimated (Bryk & Raudenbush, 1992;Goldstein, 1995; Kreft & De Leeuw, 1998).

18 Investigations of a three-level nested structure (in Fig. 1,

parts in tier-1 and subassemblies and parts in tier-2) revealed

no evidence of higher order systems e�ects on part-level sour-

cing decisions. Consequently we do not believe this is a serious

limitation of this study.


Expressed in terms of the part and subassemblylevel of analysis, the estimated model is:

Level 1 : logP

1ÿ P� ��

ij

� �0j �Xk�1�kjXkij � "ij

for part i in subassembly j;where "ij � N�0; �2�and Xk are k part-level characteristics

Level 2 : �0j � 00 �Xl�1 0lYlj � u0j

�kj � k0 �Xl�1 klYlj

for subassembly j;where u0j � N�0; �2� and Yl

are l subassembly-level characteristics:

Expressed in combined form, the estimatedmodel is:

logP

1ÿ P� ��

ij

� 00 �Xl�1 0lYlj � u0j

!

�Xk�1

ko �Xl�1 klYlj

!Xkij � "ij:

Reorganizing yields an equation with ®xed level-1 e�ects, ®xed level-2 e�ects, ®xed interactionsbetween levels, and random e�ects:

� 00 �Xk�1 k0Xkij

!�

Xl�1 0lYlj

!

�Xk�1

Xkij

Xl�1 klYlj

! !� u0j � "ijÿ �

In accounting research, it is more common tosee the second level of the analysis modeled as a®xed e�ect. For example, the ANCOVA (analysisof covariance) model is a regression model thatincludes continuous independent variables as wellas dummy variables that are used to code groupmembership. Compared to the level-2 equationabove, ANCOVA does not consider factors thatexplain system e�ects (e.g. subassembly character-istics) nor does it assume that system e�ects aremeasured with error and drawn from a commondistribution. Kreft and De Leeuw (1998) compare

alternative modeling strategies and discuss condi-tions under which multilevel modeling is a superiorapproach. The most important statistical con-sideration in using the multilevel approach is thepresence of intraclass correlation.19 Multilevelmodeling addresses limitations of disaggregate(e.g. part-level) modeling brought about by intra-class correlation. Equally important, it does notsu�er the ine�ciency of aggregate (e.g. vehicle-level) modeling brought about by discardinginformation about lower level variation. A ®nalconsideration in selecting a multilevel modelingapproach is whether the researcher is fundamen-tally interested in explaining both level-1 andlevel-2 variation. Because theory suggests thatboth part and systems-level characteristics areassociated with transaction costs and sourcingdecisions, multilevel modeling is appropriate forthis research.In the following sections, we discuss the results of

estimating four increasingly complex hierarchicalmodels of sourcing. The strategy of reporting suc-cessively complex models is recommended in theliterature on multilevel modeling as a means ofascertaining model stability between levels (Snij-ders & Bosker, 1999). Evidence of model stabilityenhances our con®dence that the model is prop-erly speci®ed. It is particularly important in thecase of nonlinear models because familiar measuresof model ®t (e.g. likelihood ratio tests andexplained variance) that are derived from linearapproximations are somewhat suspect.

4.2. A two-level model with a random intercept

Model 1 in Table 2 is the simplest test of inter-dependence among sourcing decisions of partswith common subassemblies. The model is similar

19 Barcikowski (1981) provides evidence on the in¯ation of

the implicit alpha-level for testing coe�cients in the presence of

intraclass correlation for di�erent group sizes. For a group size

of 10 with relatively low intraclass correlation (0.05), the alpha

level of 0.05 is in¯ated to 0.11. For intraclass correlation equal

to 0.20, the alpha level is in¯ated to 0.28. The intraclass corre-

lation of our data is approximately 0.15 and group sizes are

typically less than ten parts. Consequently, analysis at the part

level could easily result in overstated signi®cance of regression

coe�cients.


Table 2

The association between part-level sourcing decisions and transactions costs

This table presents results of modeling the relation between part level sourcing decisions and variables that proxy for contractual

uncertainties in die development. The dependent variable takes on the value one for parts that are sourced internally and zero for

parts that are outsourced. A hierarchical linear modeling approach is employed to allow exploration of interdependencies among

parts that are members of a common subassembly. The data are sourcing decisions and related part descriptions for 156 automotive

stamped metal parts from a single new vehicle. The parts are evaluated for contract compliance based in part on how they perform

when assembled in one of 32 subassemblies; thus we also consider how subassembly characteristics a�ect sourcing. A general

expression for the logistic, two-level model of insourcing that is estimated is:

Level 1: log P1ÿP� �

� �ij� �0j �

Pk�1�kjXkij � "ij for part i in subassembly j, where "�N(0,�2) and Xk are k part-level characteristics.

Level 2: �0j � 00 �Pl�1 0lYlj � u0j

�kj � k0 �Pl�1 klYlj for subassembly j, where u�N(0,�) and Yl are l subassembly-level characteristics.

Expressed in combined form, the estimated model is:

log P1ÿP� �

� �ij� 00 �

Pl�1 0lYlj � u0j

� �� P

k�1 ko �

Pl�1 klYlj

� �Xkij � "ij:

Reorganizing yields an equation with ®xed level-1 e�ects, ®xed level-2 e�ects, ®xed interactions between levels and random e�ects:

� 00 �Pk�1 k0Xkij

� �� P

l�1 0lYlj

� �� P

k�1Xkij

Pl�1 klYlj

� �� u0j � "ijÿ �

The results of the estimated models are grouped in this manner and the models are presented in order from least to most complex.

Level-1 variables (X) are group mean centered. Level 2 variables (Y), with the exception of the indicator variable HI-TIER, are

grand mean centereda

Model 1 Model 2 Model 3 Model 4

Level 1 ®xed e�ects

INTERCEPT ( 00) ÿ0.27 ÿ0.29 ÿ0.25 ÿ0.26(0.27) (0.29) (0.37) (0.44)

LNTHICK ( 10) 1.12** 1.17** 2.92***

(0.57) (0.58) (1.00)

LNCONPNT ( 20) ÿ0.79** ÿ0.83** ÿ0.65(0.38) (0.39) (0.58)


NUMPARTS ( 01) ÿ0.17� ÿ0.21�(0.09) (0.10)

HI-TIER ( 02) 0.18 0.17

(0.67) (0.78)

ALNTHICK ( 03) 1.09 1.31

(1.27) (1.50)

ALNCONPT ( 04) ÿ0.16 ÿ0.10(1.28) (1.46)

Interactions

LNTHICK * NUMPARTS ( 11) ÿ0.08(0.23)

LNTHICK * HI-TIER ( 12) ÿ6.61***(2.23)

LNTHICK * ALNTHICK ( 13) ÿ0.57(3.20)

LNCONPT * NUMPARTS ( 21) ÿ0.20(0.15)

LNCONPT * HI-TIER ( 22) 1.10

(1.36)

LNCONPT * ALNCONPT ( 24) ÿ0.67(1.56)

(continued on next page)


to the familiar one-way ANOVA model in whichthe estimated intercept is the ``grand mean;'' how-ever, the degree to which group (subassembly)means di�er from the grand mean is treated as arandom e�ect:

INSOURCEij ��0j � "ij�0j � 00�u0j

where " � N�0; �2� and u � N�0; �2�:

The results indicate that while, on average theintercept, 00, is not signi®cantly di�erent fromzero (translating the log odds back to the prob-ability yields the percent of insourcing in the sample),there is signi®cant subassembly-level variationaround zero as indicated by the signi®cant standarddeviation of INTERCEPT. This is similar to(unreported) ANOVA results of signi®cant (F-statistic of 2.57) between-group variation thataccounts for 39% of variation in the data.Knowledge of group membership and the posteriordistribution of outcomes allow 75% of part sour-cing decisions to be correctly predicted (comparedto 60% correctly predicted knowing only the pos-terior distribution).

4.3. A two-level model with ®xed e�ects and arandom intercept

Model 2 in Table 2 introduces variables thatproxy for part-level transactions costs, whileretaining the simplest form of interdependence inthe sourcing decisions of parts from common

subassemblies. Each level-1 variable is centeredaround its subassembly mean, thus the estimatedintercept is the grand mean of the log odds ofinsourcing for parts that are of average complexityin their own subassembly.20 The estimated relation is:

INSOURCEij ��0j � �1jLNTHICKij��2jLNCONPNTij � "ij�0j � 00�u0j

where " � N�0; �2� and u � N�0; �2�:

As before, although the intercept, 00, is notsigni®cantly di�erent from zero, there is signi®cantsubassembly-level variation around zero. Part-level variables are signi®cantly related to variationin sourcing decisions within each subassembly.Thicker parts with fewer control points are morelikely to be insourced. Stated di�erently, morecomplex parts are sourced with external suppliersin spite of transactions costs associated with con-tract uncertainty. Although part-level variablesare signi®cantly associated with sourcing out-comes, an insigni®cant di�erence between theapproximate likelihood ratio of Model 1 and 2indicates that Model 2 is not a signi®cant

Table 2 (continued)

Random e�ects

S.D. of INTERCEPT (u0) 1.09*** 1.20*** 1.27*** 1.53***

Model ®t and improvement statistics

Approximate likelihood ratio (ÿ2 * log likelihood) 439.4 438.3 438.5 418.5

Di�erence from model 1 likelihood ratio N/A 1.1 0.9 20.9**

Percent correctly predictedb 75.7 66.6 67.3 75.0

�,��,��signi®cant with P-value40.10, 0.05 and 0.01 (two-tailed test).a Group mean centering level 1 variables in conjunction with including grand mean centered variables in the second level model

that are group averages of the level 1 variable (e.g. ALNTHICK) imparts unique meaning to the estimated coe�cients. The estimated

coe�cient for the level 1 independent variable describes variation in sourcing decisions within a level 2 group while the estimated

coe�cient for the level 2 variable describes the variation between groups. Kreft and De Leeuw (1998) discuss alternatives for variable

centering for two-level models.b A naõÈ ve model, in which only knowledge of the posterior distribution is available, correctly predicts 60 % of sourcing decisions.

20 Level 1 variables are centred around the group mean to

lend a meaningful interpretation (between versus within-group

e�ects) to coe�cients of subsequent models (discussed later). It

is retained in this model to enhance comparability across mod-

els.


improvement over Model 1.21 Only 67% of sour-cing decisions are correctly classi®ed.

4.4. Means as outcomes model: a two-level modelwith ®xed and random components

Model 3 introduces to Model 2 variables thatproxy for contractual uncertainties that arise withsubassembly-level interactions among parts. Thismodel is commonly known as a ``means as out-comes'' model because the level-1 intercept, ismodeled as an `òutcome'' of level-2 explanatoryvariables. Again, level-1 variables are centeredaround the group (subassembly) mean. With theexception of the indicator variable, HI-TIER,Level-2 variables are centered around the grandmean. The estimated model is:

INSOURCEij � �0j � �1jLNTHICKij��2jLNCONPNTij � "ij�0j � 00� 01NUMPARTSj � 02HI-TIERj� 03ALNTHICKj � 04ALNCONPNTj � u0j

where " � N�0; �2� and u � N�0; �2�:

The intercept, 00, is not signi®cantly di�erentfrom zero; however, there is signi®cant sub-assembly-level variation around zero. Part-levelvariables continue to explain variation in sourcingdecisions within subassemblies. The number ofparts in the subassembly is signi®cantly related tosourcing variation between subassemblies. How-ever, as the signi®cant random component of theintercept indicates, even after including this vari-able, signi®cant unexplained subassembly-levelvariation remains. Between-subassembly variationin the mean likelihood of insourcing is not relatedto HI-TIER, ALNTHICK or ALNCONPNT.Although part and subassembly-level variables aresigni®cantly associated with sourcing outcomes,

an insigni®cant (Chi-squared test) di�erencebetween the approximate likelihood ratios ofModel 1 and 3 indicates that Model 3 is not a sig-ni®cant improvement over Model 1. As in the caseof Model 2, only 67% of sourcing decisions arecorrectly classi®ed.

4.5. Slopes and means as outcomes model: a two-level interaction model with ®xed and randomcomponents

Model 4 introduces to Model 3 interactionsbetween part and subassembly characteristics.This ``slopes and means as outcomes'' model,allows all level-1 coe�cients (intercepts andslopes) to be `òutcomes'' of level-2 explanatoryvariables; however, only the intercept varies ran-domly:

INSOURCEij � �0j � �1jLNTHICKij��2jLNCONPNTij � "ij�0j � 00� 01NUMPARTSj � 02HI-TIERj� 03ALNTHICKj � 04ALNCONPNTj � u0j

�1j � 10 � 11NUMPARTSj � 12HI-TIERj� 13ALNTHICKj

�2j � 20 � 21NUMPARTSj � 22HI-TIERj� 24ALNCONPNT22

j

where " � N�0; �2� and u � N�0; �2�:

The intercept, 00, varies signi®cantly aroundzero even after including variables that are sig-ni®cantly related to subassembly-level variation.With the exception of LNCONPNT, the part andsubassembly level variables previously found to besigni®cant remain so; however, the interpretation ofthe e�ect of LNTHICK is augmented with evidenceof a signi®cant interaction with HI-TIER. For partsthat are of greater than average thickness, there is anincreased likelihood of outsourcing if the sub-assembly is in the second tier or higher of the21 In the multilevel modeling literature, the likelihood ratio is

commonly referred to as a ``deviance'' statistic. For nonlinear

models this statistic is derived from a linear approximation. As a

result, it is possible for a slightly more complex model to have a

higher likelihood ratio of models 2 and 3. This is a general pro-

blem of determining model ®t for nonlinear models rather than a

problem of multilevel models. See Snijders and Bosker (1999) for a

complete discussion of model ®t for nonlinear, multilevel models.

22 Average subassembly thickness but average subassembly

control points is hypothesised to in¯uence the coe�cient for the

marginal in¯uence of part-level thickness on part sourcing.

Conversely average subassembly control points but not average

part thickness is hypothesised to be related to the marginal

e�ect of part-level control points on part sourcing.


vehicle evaluation sequence. To illustrate, considera part that is average in every respect exceptthickness Ð instead allow a thickness of onestandard deviation above the grand mean logthickness (0.43). If this part is a member of a sub-assembly that is average in every respect and is a®rst tier subassembly, the probability of insour-cing is 0.73=[1+ exp{ÿ(ÿ0.26+(2.92*0.43))}] ÿ1.If the part is a member of a subassembly that is inthe second tier or higher of the evaluation sequence,the probability drops to 0.14=[1+ exp{ÿ(ÿ0.26+(2.92*0.43)+(0.43*1*ÿ6.61) +0.17)}] ÿ1. Stated dif-ferently, parts that are members of upper tier sub-assemblies, the evaluations of which are possiblyconstrained by evaluations of tier-1 subassemblies,are more likely to be outsourced if they are thick.In contrast to previous models, a signi®cant dif-ference between the approximate deviance statis-tics of Models 1 and 4 indicates that Model 4 is asigni®cant improvement over Model 1. Model 4correctly classi®es 75% of sourcing decisions anddoes so using information that is readily availableto the ®rm and its suppliers prior to the pricequote and the sourcing decision.

4.6. Sole sourcing subassemblies as an extremesolution to systems-related transactions costs

The previous models examine sourcing decisionsof individual parts as a function of part and sub-assembly-level characteristics. In this section weuse the identity of suppliers to whom parts aresourced to address the transactions cost hypoth-esis that severe systems interactions lead to opti-mal co-location of interacting parts with a limitednumber of suppliers (either internal or external).Table 3 provides results of estimating the relationbetween the abnormal number of di�erent suppliers(CONCENTRATION) and subassembly char-acteristics that re¯ect part interactions (HITIER,NUMPARTS, ALNTHICK, and ALNCONPNT)for the 32 subassemblies. The results indicate thatas the number of parts in the subassembly increases,sourcing decisions become more concentrated witha few suppliers. In contrast, for HITIER sub-assemblies, where rework alternatives may beconstrained by previous decisions and where joiningrigid subassemblies may limit blame-shifting and

other opportunistic behavior, more suppliers areemployed than expected. The average thickness anddimensionality of parts in the subassembly have noe�ect on sole sourcing tendencies.23 In sum, there isevidence of a propensity to concentrate parts ofcomplex subassemblies with fewer suppliers.

Table 3

Evidence on sole sourcing of subassemblies.

This table presents results of modeling the relation between the

extent to which fewer di�erent suppliers are used than would be

expected by chance alone (CONCENTRATION) and interac-

tions among parts that comprise the 32 subassemblies. The

number of suppliers that would be observed if suppliers were

selected by chance alone is the number of suppliers expected

given that Nj parts are selected with replacement from a popu-

lation of 31 (23 external and eight internala) die suppliers,

where Nj is the number of parts in subassembly j. CON-

CENTRATION is the di�erence between the actual number of

di�erent suppliers selected and the expected number of suppli-

ers, where negative values indicate selecting fewer suppliers

than expected (e.g. tendencies to sole source).

Independent variable Coe�cient estimate (standard error)

Constant 0.578

(1.52)

NUMPARTS ÿ0.527***(0.058)

HI-TIER 0.979**

(0.411)

ALNTHICK ÿ0.577(0.773)

ALNCONPNT 0.319

(0.798)

Adjusted R2 0.76

Model F-statistic 25.79

(p-value) (0.000)

�,��,��signi®cant with P-value 4 0.10, 0.05 and 0.01 (two-

tailed test).a The results are qualitatively similar when a population of

23 In unreported results we substitute for the average thickness

and number of control points of parts in the subassembly, the

minimum thickness and the maximum number of control points

(e.g. maximum part-level complexity) for any single part in the

subassembly. The coe�cients for NUMPARTS and HI-TIER

are qualitatively similar, and the coe�cient on the natural loga-

rithm of the minimum part thickness is negative and signi®cant at

the P40.10 level (one-tailed test). The results are consistent with

the interpretation that concentrating parts wit a limited number

of suppliers is more likely when part interactions are pronounced.


5. Evidence of ex-post opportunism

The ®rm appears to outsource complex parts incomplex subassemblies and insource simple parts

in simple subassemblies Ð precisely opposite of thepattern observed in earlier studies of sourcing in theUS auto industry. Adoption of more relationalmodes of contracting and functional evaluation of

Table 4

Evidence on ex post opportunism

This table presents results of modeling the relation between two measures of ex post opportunism (REWORK and DELAY) and

previous sourcing decisions (OUTSOURCE at the part level and PROPIN at the subassembly level), controlling for part and sub-

assembly complexity. OUTSOURCE is set equal to 1 if the part is outsourced and PROPIN is the proportion of parts that ®rst enter a

subassembly that are insourced. Hierarchical linear modeling is used to explore interdependencies among parts (n=156) that are

members of a common subassembly (N=32). With the exception of the indicator variable, OUTSOURCE, and the control variable,

LEADTIME, the level-1 independent variables are group mean centered. With the exception of PROPIN, the level-2 independent

variables are grand mean centered

Firm opportunism Supplier opportunism

Dependent Variablea REWORKb DELAY


INTERCEPT ( 00) 0.12 ÿ39.60(0.39) (48.81)

LNTHICK ( 10) 0.12 ÿ68.43***(0.46) (22.71)

LNCONPNT ( 20) 0.60� 34.40**

(0.34) (17.52)

OUTSOURCE ( 30) ÿ1.68** 69.44**

(0.75) (35.11)

LNTHICK * OUTSOURCE ( 40) ÿ0.57 46.90**

(0.43) (21.49)

LNCONPNT * OUTSOURCE ( 50) 0.48 ÿ2.99(0.39) (20.77)

LEADTIME ( 60) (control) 0.30***

(0.09)


ALNTHICK ( 01) 1.19� 2.26

(0.66) (35.10)

ALNCONPNT ( 02) 0.68 73.50**

(0.61) (35.72)

NUMPARTS ( 03) 0.02 ÿ2.79(0.04) (2.32)

PROPIN ( 04) ÿ1.19** ÿ19.11(0.57) (31.57)

Random e�ects

S.D. of INTERCEPT (u0) 0.29 29.19***

S.D. of the error term (linear model) 62.27

�,��,��coe�cient is signi®cant with P-value40.10, 0.05 and 0.01 (two-tailed test).a Dependent variable de®nitions: REWORK, number of times that a supplier was formally asked to perform die rework as a result

of subassembly evaluations; DELAY, the number of days between the date that the ®rm requested part submission for subassembly

evaluation and the actual date of part submission by the supplier.b REWORK, an integer with values between zero and four, is assumed to have a Poisson distribution. Results are qualitatively

similar using a logistic function to represent the probability of any incidence of rework. The population average model is reported.


parts within subassemblies seems to be associatedwith a di�erent relation between sourcing andtransactions costs. An obvious question is whetherafter controlling for characteristics that in¯uencesourcing, outsourced parts subsequently experi-ence greater incidence of opportunistic behaviorthan insourced parts. In Table 4 we investigatethis using two measures that are sensitive toopportunistic behavior.

5.1. Opportunistic demands for die rework

The ®rst column of Table 4 examines one mani-festation of opportunistic behavior by the ®rm Ðwhether parts from outsourced dies fail sub-assembly evaluation more often, with require-ments that the supplier perform die rework as acondition for completing the contract. Controllingfor complexity that makes rework likely regardlessof the source, we estimate the relation between theincidence of rework requests (REWORK) and diesourcing for the part (OUTSOURCE) and itssubassembly (PROPIN). As in the sourcing ana-lysis, we consider the possibility that part-levelrework requests are in¯uenced by rework deci-sions for parts of the same subassembly. Theintuition, again, is that there are multiple ways ofbringing nonconforming subassemblies into con-formance. After sourcing decisions are made, the®rm may be motivated to impose rework onexternal suppliers rather than internal suppliers ifthey can do so under the terms of the originalcontract. If this behavior is common, then thecoe�cient on PROPIN should be positively rela-ted to rework.We use hierarchical linear modeling for a Pois-

son distributed variable because REWORK is apositive integer. Measures of part complexity aregroup mean centered and measures of averagesubassembly complexity are grand mean centered;thus the coe�cients on the level-1 variables repre-sent within group e�ects while the coe�cients onthe level-2 variables represent between groupe�ects. The indicator variable, OUTSOURCE isnot centered and takes on the value of one foroutsourced parts. The uncentered level-2 variable,PROPIN, is the proportion of parts that ®rst enterthe subassembly that are insourced. At the part

level we switch from an indicator of insourcing toan indicator of outsourcing to enhance interpreta-tion of the results about the marginal e�ects ofoutsourcing on rework for an average part (sub-assembly unspeci®ed):

. 00+( 04*(0.47)) is the log of the incidenceof rework for insourced parts of averagecomplexity in average subassemblies. Theestimated incidence of rework is 0.64.

. ( 00+( 04*(0.47)+ 30) is the log of the inci-dence of rework for outsourced parts ofaverage complexity in average subassemblies.The estimated incidence of rework is 0.12.

. ( 00+ 30) is the log of the incidence ofrework for outsourced parts of averagecomplexity in subassemblies with noinsourced parts. The estimated incidence ofrework is 0.21.

A comparison of the ®rst two points indicatesthat the ®rm is not using its right to demandrework to penalize external suppliers. Rather,after controlling for part and subassembly com-plexity, insourced parts are more likely to requiredie rework than outsourced parts. A comparisonof the last two points indicates that the ®rm is notexploiting the opportunity to shift blame for sub-assembly failure from internally to externallysourced dies. We ®nd a signi®cant between-groupe�ect in REWORK that is related to the averagepart complexity of the subassembly and the pro-portion of insourced parts. These systems e�ectssupport the hypothesis that problems observed insubassembly evaluation tend to create die reworkfor all members of the subassembly. After includingthe level-2 explanatory variables for the randomintercept, the remaining unexplained variation isnot signi®cantly di�erent from zero.

5.2. Opportunistic delays in submitting parts forevaluation

The second column of Table 4 examines thenumber of days between the date when the part isrequested by the ®rm to be ready for subassemblyevaluation and the date when parts are providedby the supplier to the ®rm for evaluation


(DELAY). Although it is a linear rather than aPoisson model, the included variables are similarto the previous model for REWORK. We intro-duce the variable, LEADTIME, which is thenumber of days between completion of partdesigns and the requested part submission date, tocontrol for submission delays that stem fromdelays that may be primarily the responsibility ofthe ®rm. As before, variable centering lends spe-ci®c interpretations to the coe�cients about themarginal e�ects of outsourcing on submissiondelays for an average part in an average sub-assembly:

. 00+( 60*473.65)+( 04*0.47) is the numberof days that an insourced part of averagecomplexity in an average subassembly isdelayed. The estimated delay is 93.5 days.

. 00+( 60*473.65)+( 04*0.47)+ 30 is thenumber of days that an outsourced part ofaverage complexity in an average sub-assembly is delayed. The estimated delay is163.0 days.

. 00+( 60*473.65)+ 30 is the number of daysthat an outsourced part of average complex-ity in a subassembly with no insourced partsis delayed. The estimated delay is 171.9 days.

After controlling for part complexity and sub-assembly membership, for average parts in aver-age subassemblies external suppliers delaysubmission for evaluation by 70 days as comparedto internal suppliers. As expected, the marginale�ect of part thickness (e.g. reduced complexity) isto reduce delay time; however, external suppliersdo not take full advantage of producing an easierpart, submitting these parts later than internalsuppliers. So, for example, evaluated at one stan-dard deviation above the grand mean thickness,insourced parts are delayed 64 days (as opposed to93.5 days) while outsourced parts are delayed 154days (as opposed to 163 days). There is signi®cantvariation in submission delay between sub-assemblies; however, variation is unrelated to thesourcing status of other parts in the subassembly.We observe a signi®cant between-group e�ect inDELAY that is related to the average part com-plexity in the subassembly but is unrelated to the

sourcing of parts. These systems e�ects are con-sistent with the hypothesis that problems observedin subassembly evaluation tend to be sharedamong members of the subassembly. Afterincluding the level-2 explanatory variables for therandom intercept, there remains signi®cant unex-plained variation between subassemblies.

6. Field investigation of results

In light of ®ndings that di�er substantially fromprevious studies, we conducted interviews withmanagers, cost estimators, and design engineersfrom the ®rm and die suppliers to try to under-stand our results. Two assumptions that Milgromand Roberts (1992) characterize as fundamentalweaknesses of transactions costs theory Ð inse-parable production and transaction costs andwealth e�ects for decision-makers Ð appear to beproblematic in our research setting. As we describebelow, inseparable production and transactionscosts, are more likely to explain di�erencesbetween our results and previous studies, sincewealth e�ects are likely to be common to studiesof the US auto industry. Similarly, inseparablecosts are more likely than wealth e�ects to beassociated with shifts in sourcing decisions fol-lowing the adoption of relational contractingbecause organizational boundaries within the ®rmwere not altered in conjunction with the adoptionof relational contracting. Our data do not permit usto corroborate changes in the pattern of sourcingdecisions as a function of part and subassemblycharacteristics before and after adoption of rela-tional contracting. Nonetheless, as we describebelow, it seems likely that the nature of the con-tractual changes altered the nature of productionand transaction cost interactions by increasing thevariance associated with die making inputs.Our interviews revealed that sourcing decisions

are more aptly described as short term optimizationof capacity utilization, given the level of die buildcapacity. The decision about ®rm boundaries (thatempirical transactions cost studies examine) wastaken earlier, when long term decisions about die-build capacity were taken. Importantly, decisionsabout long-run capacity investment and short-run


utilization are in¯uenced by an important internaluse of die-build capacity Ð die maintenance andrepair. Once dies are in production, the cost of dierepair becomes trivial when compared to theopportunity cost of idle presses. Consequently, theappropriate level and use of new die-build capa-city depends on repair and maintenance demandsthat old dies create. In the extreme, new die-buildactivities could be fully outsourced not becausetransactions costs are inconsequential, but because,in the short run, the opportunity cost of usinginternal assets for die-build rather than die main-tenance is exceedingly high. Internal sourcing maybe a strategic choice to maintain capability in diedevelopment, but it is also a secondary use ofexcess capacity in die development assets. In thissetting, the long term strategic decision is whetherto invest in internal production capability and ifso, at what level. The (relatively) short term part-level sourcing decision becomes, which parts toproduce in the die build capacity that remainsafter accommodating die repair and maintenancedemands. Sourcing decisions are an attempt to``smooth'' utilization of these assets. An explana-tion that is consistent with our results is that con-stant utilization is attained by insourcing dies thatare associated with low uncertainty about resourcedemands Ð those with lower than average trans-actions costs. Consider further the possibility thatthe new approach of involving suppliers earlier andpermitting acceptance of dies that would previouslyhave been termed ``nonconforming,'' reduces theexpected cost of production (to the supplier) butalmost certainly increases the variance of produc-tion costs. If so, this presents a situation in whichthe adoption of relational contracting could causeparts that would previously have been insourced tobe outsourced. These observations are consistentwith economic theory, which would consideralternative uses of assets in a make versus buyanalysis. However, they are inconsistent withempirical research on transactions costs in theauto industry, which relies on the supplier and theproducer having roughly similar production coststo hypothesize that di�erent sourcing decisionsre¯ect the in¯uence of transactions costs.The matter of the decision-making process for

make versus buy decisions is complex. Transactions

cost theory assumes that the sourcing decision ismade by someone seeking to maximize the valueof the ®rm using complete information aboutproduction costs and alternative uses of assets. Inour ®rm the internal die development shops aremembers of a di�erent operating division from thepurchasing organization that makes sourcingdecisions and the manufacturing organization that``buys'' dies from internal or external die shops.24

Although the ®rm has taken steps to enhancecommunications and align incentives between allthree groups, interviews indicate that sourcingdecisions are more appropriately described asselection by a decision-maker with local objectivesof an internal or an external supplier. Self-inter-ested internal and external suppliers provide datato in¯uence the decision, and the decision-makerhas limited opportunities to verify suppliers'claims about production costs. While concernsabout wealth e�ects associated with sourcingdecisions are substantial, it is unlikely that di�er-ences in wealth e�ects explain di�ering results onthe e�ect of transactions costs on sourcing deci-sions between our study and those that precededus because of the common, largely stable industrysetting. It is also unlikely that wealth e�ects chan-ged signi®cantly in conjunction with adoption ofrelational contracting because (for better orworse) decision rights over the sourcing decisionand organizational boundaries were not changedas part of the new contracting approach.Asked why he was given the most complex jobs,

the President of one external supplier remarked,`ìnternal die shops don't want to do the hard jobsbecause they don't want to be on the hook if theycan't make the parts. And the [®rm's] engineersdon't want them to get the work because theydon't think they'll be responsive [to reworkrequests].'' The implication is that internal dieshops submit high cost estimates for complex diesto minimize their exposure to the perceived risk ofpoor performance. Asked about ®ndings thatinsourced parts are more likely to face demandsfor rework while outsourced parts are submittedlater for evaluation, the manager explained the

24 The designers of the parts that the dies must produce are

in yet a fourth division.


results as evidence that suppliers do not fully trustthe ®rm. He claimed that, anticipating that theirparts are likely to be blamed for problems, externalsuppliers do more work to perfect the dies (relativeto part-level speci®cations) than internal suppliers Ðresulting in ``better'' parts (lower rework) but takinglonger to achieve (longer submission delays). Ofcourse the data are also consistent with supplierssimply holding the dies longer as a means ofaltering the likelihood of being required to dorework. In general, managers at the ®rm agreedthat externally produced dies are being `òver-worked'' relative to internal dies and relative totheir preferences. However they acknowledge sup-pliers' fear of ``punitive'' rework requests andhope to counter it over time with behavior to thecontrary.In summary, as uncertainty in die development

has increased with earlier involvement of suppliersand more subjective, interdependent evaluationsof contract completion, it appears that incre-mental costs associated with the loss of high-pow-ered incentives are greater than incrementaltransactions costs. Moreover, although we ®nd noevidence that the ®rm behaves opportunisticallyduring subassembly evaluations, this may be apremature conclusion. It is possible that, fearingopportunistic behavior, external suppliers submitparts from dies that are more developed thanthose of internal suppliers. If so, it is unclearwhether external suppliers would be subject tomore demands for rework than internal suppliersif they did not make these investments in diedevelopment.These observations are not startling to account-

ing researchers who are accustomed to thinkingabout organizations as decentralized groups thatare populated with self-interested people. How-ever, once these organizational aspects of themake versus buy decision are contemplated, it isdi�cult to interpret the empirical literature ineconomics as evidence that transactions costs are aleading determinant of sourcing decisions. Inmodeling average sourcing behavior for broadclasses of auto parts (e.g. Masten et al., 1989;Monteverde & Teece, 1982), early studies focus ondeterminants of whether any internal productioncapability exists rather than on part sourcing

decisions, conditional on the existence of internalcapability. There is little recognition that in theshort run a ®rm may outsource activities in whichit has an absolute cost advantage to free resourcesfor activities in which it has a greater relative costadvantage. Related to this, in testing whetherinsourcing is positively associated with transac-tions costs, previous research relies heavily ontransactions costs favoring insourcing and`ùnderstates the costs of managing interunit rela-tionships within the organization (Walker &Poppo, 1991).'' Together these limitations aresymptomatic of simplifying assumptions commonin the empirical literature. Field research impli-cates problems with these assumptions as oneexplanation for why we observe di�erent sourcingdecisions from those of previous studies when weconsider a setting that uses more relational modesof contracting.

7. Conclusion

Economists often explain manufacturing ®rms'decisions about whether to make or buy parts asthe low cost solution, considering both productionand transactions costs. Management accountingtexts urge managers to consider ``qualitative fac-tors'' associated with the reliability and perma-nence of suppliers' prices in reaching sourcingdecisions. In this paper, we revisit data on therelation between proxies for transactions costs andsourcing decisions in the US auto industry usingdata from a ®rm that has started involving sup-pliers in design activities and has adopted sub-jective, interdependent evaluations of contractcompletion. The results indicate that under thenew contracting regime, factors that previous studiesfound to favor insourcing now favor outsourcing.Speci®cally, complex parts are more likely to beoutsourced, as are parts that are members of morecomplex, interdependent subassemblies. We also®nd that systems interactions among parts areassociated with fewer suppliers being selected thanwould be expected by chance alone Ð consistentwith the hypothesis that sole sourcing mitigatestransactions costs associated with systems uncer-tainties.


In spite of results that di�er signi®cantly fromprevious studies, we ®nd no evidence that the ®rmdemands rework more frequently for outsourcedparts as compared to insourced parts. Moreover,although incidence of opportunistic behavior cov-aries with subassembly membership, there is noevidence that ``problems'' are disproportionatelyattributed to outsourced parts when the sub-assembly contains at least one insourced part. We®nd that suppliers are more likely to delay sub-mission of parts for evaluation as compared tointernal suppliers. An explanation that emergedfrom our ®eld research is that, skeptical about thenew subjective evaluation methods, suppliers domore work on the dies than internal suppliers andmore than the ®rm prefers. A second explanationis that delays re¯ect ``delivery hold ups'' intendedto alter the ®rms' propensity to demand rework.The data do not permit us to discriminate betweenthese explanations which have very di�erentimplications for improving the new contractingapproach.The transactions costs literature and our study

have implications for practicing accountants andaccounting researchers. Accountants are oftencalled upon to assess whether it is more econom-ical for manufacturing ®rms to make or buy partsin light of alternative uses of production assets.Management accounting textbooks (e.g. Atkin-son, Banker, Kaplan & Young, 1997, pp. 364±8)typically focus on the di�culty of correctly identi-fying avoidable production costs and comparingthem with external suppliers' prices. Strategicsupply chain management and the hybrid organi-zational forms that it portends place new demandson accountants. Accounting texts mention moredi�cult to assess ``qualitative factors'' that in¯u-ence the reliability and permanence of supplierprices (p. 367). Our research indicates that it is lessappropriate than ever for accountants chargedwith developing data for sourcing decisions tofocus narrowly on production costs. The valuechain perspective of strategic cost managementwith its focus on ``costs of ownership'' rather thansupplier price is essential, and we can not assume that``costs of ownership'' are minimized with internalproduction. In short, estimates of uncertaintyassociated with the reliability and permanence of

prices are relevant for both internal and externalsuppliers. Further, relational contracting andassociated performance measures (e.g. supplierratings) that lend hierarchical characteristics tomarkets suggest a need to augment our under-standing of organizational control systems.Although our research site provides a unique

organizational perspective on sourcing that isabsent from some previous studies, it is not with-out limitations. The empirical analysis requiresseveral assumptions. First, we do not have data onsourcing decisions before and after the adoptionof relational contracting. Rather, we must rely ona large body of existing research to proxy forsourcing outcomes prior to the change in con-tracting. Although the industry setting is commonacross all of the studies, we must be cautiousabout overstating the comparability of the studies.A second limitation is that in the absence of dataon the date of the sourcing decisions, we assumethat sourcing decisions for all die sets are madesimultaneously. Thus, there is no time dimensionto the decision process that would allow di�erentstates of information to in¯uence di�erent sour-cing decisions. If sourcing is conducted in``rounds'' our analysis would incorrectly ignorecovariance between early and late decisions. Asecond stronger assumption is that sourcing isrelated only to the characteristics of the part thatthe dies must produce and its subassembly. Intruth, sourcing is almost certainly in¯uenced byhistorical supplier performance. This limitationspeaks to covariance over time between the diesourcing decisions for the vehicle that we studyand all prior die sourcing decisions. Futureresearch could address this issue with data onsourcing decisions within and across vehicledevelopment projects. Finally, although our inter-actions with the ®rm indicates that product designand system architecture decisions precede diesourcing decisions Ð indeed product design andprocess design are governed by di�erent organiza-tions in the ®rm Ð it is possible that productdesigners anticipate sourcing decisions in theirproduct designs (e.g. Novak & Eppinger, 1998).Similarly, it is possible that product designersanticipate die design concerns. Although we thinkthat our results re¯ect causal relations between


part and system complexity and subsequent diesourcing decisions, we can not rule out the possi-bility of a more complex, simultaneous optimiza-tion of product and process design and sourcing.In spite of the limitations, this paper di�ers in

several important ways from previous studies.First, rather than study all components of theautomobile, we focus on tools (dies) that producethe metal stampings for the auto body. An attrac-tive feature of studying dies is that the sourcingdecisions are taken in consideration of a commongroup of eligible suppliers who use comparable diebuild technology and are subject to similar pro-duction and transactions costs. Focusing on sour-cing decisions for dies alone also permits us todevelop measures of transactions costs that aretailored to the speci®c supply setting. For exam-ple, we use many of the measures of part com-plexity that Novak and Eppinger (1998) consider.However, we are not forced to adopt ad-hocmethods of summarizing the data to create a scalefor complexity of stamped body parts that can becombined with aggregate scales of complexity inother part systems (e.g. the auto transmission) fora measure of vehicle part complexity. Previousstudies have attributed variation in sourcing out-comes across broad categories of parts to transac-tions-speci®c investments. We focus instead onwithin-category variation.A second strength is analyzing data from actual

sourcing decisions associated with a real auto-mobile. With the exception of Walker and Weber(1984, 1987) and Novak and Eppinger (1998),previous studies examine average part sourcingdecisions of one or two ®rms over all vehiclesproduced for a period of time. Focusing on a sin-gle new automobile provides us a large number ofsourcing decisions that were made at approxi-mately the same time and subject to the sameenvironmental uncertainties and organizationalsourcing strategies. It also allows us to exploit the®rm's structure of subassembly evaluations ratherthan relying on crude part groupings [e.g. Mon-teverde and Teece (1982) treat all stamped bodyparts as a single part and arbitrarily group it withpaint, safety belts and automotive glass becausetogether these are assembled to create the ``body''of the automobile]. Using a grouping structure

that mirrors that of the supply contract allows usto make stronger claims that the systems e�ectsthat we estimate are in fact those that in¯uence thesourcing decision.

Acknowledgements

We are grateful for ®nancial support from theNational Science Foundation's Innovation andOrganizational Change Program (SBR-9712997)and from the Tauber Manufacturing Institute atthe University of Michigan. The University ofMichigan's Transportation Research Institute wasinstrumental in helping us obtain access to thedata. We also thank Stan Baiman, Robin Cooper,Jay Baron, Pat Hammet, Je�rey Hand, GlennHoetker, Chris Ittner, Bill Lanen, Steve Pollockand participants at the 1999 AOS/USC Conferenceon `Àccounting, Organizational transformationand New Organizational Forms.'' Workshop par-ticipants at the Wharton Business School of theUniversity of Pennsylvania and the GoizuetaBusiness School at Emory University also pro-vided helpful comments.

References

Alchian, A., & Demsetz, H. (1972). Production, information

costs, and economic organization. American Economic

Review, 62, 777±795.

Anderson, S. W., & Sedatole, K. (1998). Designing quality into

products: the use of accounting data in new product devel-

opment. Accounting Horizons. 12 (3), 213±233.

Armour, H. O., & Teece, D. J. (1980). Vertical integration and

technological innovation. The Review of Economics and Sta-

tistics, pp. 470±474.

Atkinson, A. A., Banker, R. D., Kaplan, R. S., & Young, S. M.

(1997). Management accounting. NJ: Prentice Hall: Upper

Saddle River.

Baiman, S., Fisher, P. E. & Rajan, M. V. (1999) Performance

measurement and design in supply chains. Working paper.

University of Pennsylvania, The Wharton School.

Barcikowski, R. S. (1981). Statistical power with group mean as

the unit of analysis. Journal of Educational Statistics, 6 (3),

267±285.

Bidault, F., Despres, C., & Butler, C. (1998). New product

development and early supplier involvement (ESI): The dri-

vers of ESI adoption. International Journal of Technology

Management, 15(1/2), 49±69.

Birou, L. M., & Fawcett, S. E. (1994). Supplier involvement in

integrated product development: a comparison of US and


European practices. International Journal of Physical Dis-

tribution and Logistics Management, 24(5), 4±14.

Boothroyd, G., Dewhurst, P., & Knight,W. (1994). Product

design for manufacture and assembly. New York: Marcel

Dekker.

Bozdogan, K., Deyst, J., Hoult, D., & Lucas, M. (1998).

Architectural innovation in product development through

early supplier integration. R & D Management, 28(3), 163.

Bryk, A. S., & Raudenbush, S. W. (1992). Hierarchical linear

models: Applications and data analysis. Newbury Park, CA:

Sage Publications.

Chandler, A. (1962). Strategy and structure. Cambridge, MA:

MIT Press.

Clark, K. (1989). Project scope and project performance: The

e�ect of parts strategy and supplier involvement on product

development. Management Science, 35(10), 1247±1263.

Clark, K., & Fujimoto, T. (1991). Product development perfor-

mance: Strategy, organization and management in the world

auto industry. Boston: MA: Harvard Business School Press.

Coase, R. H. (1937). The nature of the ®rm. Economica, 4,

386.

Crocker, K. J., & Masten, S. E. (1988). Mitigating contractual

hazards: unilateral options and contract length. Rand Journal

of Economics, 19, 327±343.

Cusumano, M. (1985). The Japanese automobile industry: Tech-

nology and management at Nissan and Toyota. Cambridge,

MA: Harvard University Ð The Council on East Asian

Studies.

Daft, R. L., & Lewin, A. Y. (1993). Where are the theories for

the `new' organizational forms? An editorial essay. Organi-

zation Science, 4, i±vi.

Drtina, R. E. (1994) The outsourcing decision. Management

Accounting (USA), March, 56±62.

Dyer, J. (1996). Does governance matter Keiretsu alliances and

asset speci®city as sources of Japanese competitive advan-

tage? Organization Science, 7(6), 649±666.

Eccles, R., & White, H. (1988). Price and authority in inter-pro®t

center transactions. American Journal of Sociology, 94, 17±51.

Fitzgerald, K. R. (1999). Buying in the OEM Ð Part II. Pur-

chasing, 126(3), 50±57.

Goldstein, H. (1995). Multilevel statistical models. New York:

Halsted Press.

Hammett, P. C., Majeske, K. D., & Baron, J. S. (1998). Func-

tional build: integrating automotive body-design and process

development. In IEEE proceedings annual reliability and

maintainability symposium, Anaheim, CA.

Ittner, C., & Ponemon, L. (1992). Measuring the cost of own-

ership, Journal of Cost Management, Fall, 42±51.

Jenson, M., & Meckling, W. (1976). Theory of the ®rm: man-

agerial behavior, agency costs and capital structure. Journal

of Financial Economics, 3, 305±360.

Joskow, P. (1987). Contract duration and relationships-speci®c

investments: empirical evidence from coal markets. The

Americal Economic Review, 37(1), 168±185.

Joskow, P. (1988). Asset speci®city and the structure of vertical

relationships: empirical evidence. Journal of Law Economics

and Organization, 4, 95±117.

Klein, B. (1988). Vertical integration as organizational owner-

ship: the Fisher Body±General Motors relationship revisited.

Journal of Law, Economics and Organization, 4, 199±213.

Kreft, I., & De Leeuw, J. (1998). Introducing multilevel model-

ing. London: Sage Publications.

Krishnan, V., Eppinger, S. D., & Whitney, D. E. (1995). A

model-based framework to overlap product development

activities. Management Science, 43(4), 437±451.

Krishnan, V., Eppinger, S. D., & Whitney, D. E. (1995).

Accelerating product development by the exchange of pre-

liminary product design information. Journal of Mechanical

Design, 117(4), 491±498.

Lacity, M. C., Willcocks, L. P., & Feeny, D. F. (1996). The

value of selective IT outsourcing. Sloan Management Review,

Spring, 13±25.

Liu, S. C., Hu, S. J., & Woo, T. C. (1996). Tolerance analysis

for sheet metal assemblies. ASME Journal of Mechanical

Design, 118, 62±67.

Masten, S. (1984). The organization of production: evidence

from the aerospace industry. Journal of Law and Economics,

27, 403±417.

Masten, S., Meehan, J., & Snyder, E. (1989). Vertical integra-

tion in the US auto industry. Journal of Economic Behavior

and Organization, 12, 265±273.

Milgrom, P., & Roberts, J. (1992). Economics, organization and

management. Englewood Cli�s: NJ: Prentice Hall.

Monteverde, K., & Teece, D. (1982). Supplier switching costs

and vertical integration in the automobile industry. Bell

Journal of Economics, 13, 206±213.

Novak, S., & Eppinger, S. D. (1998, November). Sourcing by

design: Product architecture and the supply chain. Working

paper 4045. Massachusetts Institute of Technology: Sloan

School of Management.

Poppo, L. (1995). In¯uence activities and strategic coordina-

tion: Two distinctions of internal and external markets.

Management Science, 41(12), 1845±1859.

Porter, M. E. (1980). Competitive strategy: Techniques for ana-

lyzing industries and competitors. New York: The Free Press.

Porter, M. E. (1985). Competitive advantage. New York: The

Free Press.

Prahalad, C. K., & Hamel, G. (1990). The core competencies of

the corporation. Harvard Business Review, 86, 79±91.

Richardson, J. (1993). Parallel sourcing and supplier perfor-

mance in the Japanese automobile industry. Strategic Man-

agement Journal, 14(5), 339±350.

Roodhooft, F., & Warlop, L. (1999). On the role of sunk costs

and asset speci®city in outsourcing decisions: a research note.

Accounting, Organizations and Society, 24(4), 363±369.

Shelanski, H. A., & Klein, P. G. (1995). Empirical research in

transaction cost economics: review and assessment. Journal

of Law, Economics and Organization, 11(2), 335±361.

Smith, R. P., & Eppinger, S. D. (1997). Identifying controlling

features of engineering design iteration. Management Sci-

ence, 43(3), 276±293.

Smith, R. P., & Eppinger, S. D. (1997). A predictive model of

sequential iteration in engineering design. Management Sci-

ence, 43(8), 1104±1120.


Snijders, T., & Bosker, R. (1999). Multilevel analysis. London:

Sage Publications.

Taguchi, G., Elsayed, E., & Hsiang, T. (1989). Quality engi-

neering in production systems. New York: McGraw-Hill.

Tottie, M., & Lager, T. (1995). QFD Ð Linking the customer

to the product development process as part of the TQM

concept. R & D Management, 25(3), 257±268.

Ulset, S. (1996). R&D outsourcing and contractual governance:

an empirical study of commercial R&D projects. Journal of

Economic Behavior and Organization, 30(1), 63.

Vancil, R. (1978). Decentralization: Managerial ambiguity by

design. Homewood, IL: Irwin.

Walker, G. (1994). Asset choice and supplier performance in

two organizations Ð US and Japanese. Organization Science,

5(4), 583±593.

Walker, G., & Poppo, L. (1991). Pro®t centers, single-source

suppliers and transactions costs. Administrative Science

Quarterly, 36(1), 66±87.

Walker, G., & Weber, D. (1984). A transaction cost approach

to make-or-buy decisions. Administrative Science Quarterly,

29(3), 373±391.

Walker, G., & Weber, D. (1987). Supplier competition, uncer-

tainty and make or buy decisions. Academy of Management

Journal, 30(3), 589±596.

Ward, A., Liker, J. K., Cristiano, J. J., & Sobek, D. K. (1995).

The second Toyota paradox: How delaying decisions can

make better cars faster.. Sloan Management Review, 36(3),

43±61.

Wasti, S. N., & Liker, J. K. (1997). Risky business or competi-

tive power Supplier invovlement in Japanese product design?

The Journal of Product Innovation Management, 14(5), 337±

355.

Wheelwright, S. C., & Clark, K. B. (1992). Revolutionizing

product development: Quantum leaps in speed, e�ciency and

quality. New York: The Free Press.

Whyte, G. (1994). The role of asset speci®city in the vertical

integration decision. Journal of Economic Behavior and

Organization, 23(3), 287±302.

Williamson, O. E. (1975). Markets and hierarchies: Analysis and

antitrust implications. New York: The Free Press.

Williamson, O. E. (1979). Transaction-cost economics: the

governance of contract relations. Journal of Law and Eco-

nomics, 22, 233±261.

Williamson, O. E. (1985). The economic institutions of capital-

ism: Firms, markets relational contracting. New York: The

Free Press.

Womack, J. P., Jones, D. T., & Roos, D. (1990). The machine

that changed the world. New York: Harper Perennial.

Zenger, T. R., & Hesterly, W. S. (1997). The disaggregation of

corporations: selective intervention, High-powered incentives

and molecular units. Organization Science, 8(3), 209±222.


Documents

Sourcing parts of complex products: evidence on ...directory.umm.ac.id/Data Elmu/jurnal/A/Accounting...Milgrom and Roberts (1992) discuss two important ways in which make-or-buy decisions