工商管理英文论文翻译

外文资料翻译

Abstract

This paper introduces the concept of knowledge networks toexplain why some business units are able to benefit from knowledgeresiding in other parts of the c ompany while others arenot. The core premise of this concept is that a proper u nderstandingof effective interunit knowledge sharing in a multiunitfirm requires aj oint consideration of relatedness in knowledgecontent among business units and t he network of lateral interunitrelations that enables task units to access related k nowledge.Results from a study of 120 new product developmentprojects in 41 bu siness units of a large multiunit electronicscompany showed that project teams o btained more existingknowledge from other units and completed their projects fas terto the extent that they had short interunit network paths to unitsthat possessed related knowledge. In contrast, neither networkconnections nor extent of related k nowledge alone explainedthe amount of knowledge obtained and project completi on time.

The results also showed a contingent effect of having directinterunit relations i n knowledge networks: While establisheddirect relations mitigated problems of tra nsferring noncodifiedknowledge, they were harmful when the knowledge to be tra nsferredwas codified, because they were less needed but stillinvolved maintenance costs. These findings suggest that researchon knowledge transfers and synergies i n multiunit firmsshould pursue new perspectives that combine the concepts ofnet

work connections and relatedness in knowledge content.Why are some business u nitsable to benefit from knowledgeresiding in other parts of the company while othersare not? Both strategic management and organization theoryscholars have ex tensively researched this question,but differences in focus between the various ap proacheshave left us with an incomplete understanding of whatcauses knowledge sharing to occur and be beneficialacross business units in multiunit firms. In one line ofresearch, scholars have focused on similarity in knowledgecontent among b usiness units, arguing that a firmand its business units perform better tothe exten t thatunits possess related competencies that can be used bymultiple units (e.g., Rumelt 1974, Markides and Williamson1994, Farjoun 1998). While this knowledg e content viewhas demonstrated the importance of relatedness in skillbase, it doe s not shed much light on the integrative mechanismsthat would allow one busine ss unit to obtainknowledge from another (Ramanujam and Varadarajan1989, Hill 1994). When sharing mechanisms are consideredin this research, it is often assu med that the corporatecenter is able to identify and realize synergies arisingfrom similarity in knowledge content among businessunits, but this assumption is typic ally not tested empiricallyand excludes a consideration of lateral interunit relation s(Chandler 1994, Markides and Williamson 1994,Farjoun 1998).

In other lines of research, in contrast, scholars havedemonstrated the importanc e of havinglateral linkagesamong organization subunits for effective knowledgesha ring to occ ur. Researchhas shown that a subunit’sinformation processing capacity is enhanced by lateralinterunit integration mechanisms (e.g., Galbraith 1973,1994; Egelhoff 1993; Gupta and Govindarajan 2000),product innovation knowledge flow s more efficientlythrough established relationships spanning subunitboundaries (Tu shman 1977, Ghoshal and Bartlett 1988,Nobel and Birkinshaw 1998,Hansen 199 9), and bestpractices are transferred more easily when a positive existingrelations hip exists between the two parties to atransfer (Szulanski 1996). These lines of r esearch on linkageshave, however, not incorporated opportunities forknowledge sh aring based on commonality in knowledgecontent among subunits, but has taken this aspect asgiven.

Yet the existence of both related knowledge in thefirm—i.e., expertise in the f irm’s business units that canbe useful for tasks per formed in a focal business un itand a set of established linkages among business unitsseems necessary for inter unit knowledge sharing to occurand be effective. In this paper, I consider both d imensionsand develop theconcept of task-specific knowledge networks,which comp rise not only those business units thathave related knowledge for a focal task un it, but also theestablished direct and indirect interunit relations connectingthis sub set of business units.

I define establishedinterunit relations as regularly occurring informal contactsbet ween groups of people from different businessunits in a firm, and I assume thatt ask units will be abletouse these relations to search for and access knowledgeresi ding in other business units.

I make two main arguments. First, with respect to indirect relations (i.e., conne ctions throughintermediaries),I argue that task teams in focal business units with shortpath lengths in a knowledge network (i.e., few intermediariesare needed to c onnect with other units) are likelyto obtain more knowledge from other business units andperform better than those with long path lengths becauseof search benef its accruing to business units with shortpath lengths. Long path lengths, in contra st, lead to informationdistortion in the knowledge network, makingsearch for usef ul knowledge more difficult. Second, I arguethat a focal unit’s direct established relations in aknowledge network are a two-edged sword: While theyprovide im mediate access to other business units that possessrelated knowledge, they are als o costly to maintain.

They are, therefore, most effective when they help teamssolve difficult transfer problems, as when the knowledgeto be transferred is noncodified (Szulanski 1996, Hansen1999). Whenthere is no transfer problem, they are likelyto be harmful fort ask-unit effectiveness because of theirmaintenance costs.

This knowledge network model seeks to advance ourunderstanding of knowled ge sharing in multiunit companiesin several ways. First, by integrating the conce ptsof related knowledge and lateral network connections thatenable knowledge sha

ring, the model seeks to extend extantresearch that has addressed only one of th ese aspects.Second, while extant research on knowledge transferstends to focus o n direct relations (i.e., the dyadic linkbetween a recipient and a source unit of k nowledge), Ialso consider the larger organization context of indirect

relations, which are conduits for information about opportunitiesfor knowledge sh aring (cf. Ghoshal and Bartlett1990). This approach enables a richer understandin g ofsearch processes forknowledge use in multiunit firms.Third, while scholars of ten consider the positive effects

of network relations on knowledge sharing, I also considermaintenance costs of n etworks byincorporating thistime commitment in analyzing the impact of interunit networkrelations on knowledge-sharing effectiveness inmultiunit firms.

Knowledge Networks in Multiunit Firms

The joint consideration of related knowledge and lateralinterunit relations of a knowledge network is illustratedin Figure 1 for a new product development team, whichis the unit of analysis in this paper. Diagram 1a illustratesa network of re lations among all business units in a firm,but does not partition these units into those that have relatedknowledge for the focal new product developmentteam, A (i.e., a pure network consideration). Diagram 1b,in contrast, partitions the busines s units in the firm intothose that have related knowledge for the focal productde velopment team (A) and those that have not, but thereis no consideration of then etwork among the units (i.e.,a pure related knowledge consideration). Diagram 1 illustratesa project-specific knowledge network: Businessunits are partitioned into t hose that have related knowledgefor the focal product development team (A), an d thecomplete set of network ofrelations among them are included,including both direct and indirect relations (i.e.,intermediarylinks connecting the focal unit with othersin the knowledge network). Both the indirect and directrelations affect the extent to which a focal product developmentteam is able to obtain knowledge fr om otherbusiness units and use it to perform better.

Effects of Indirect Relations in Knowledge Networks

A product development team’s direct and indirect interunitrelations in its know ledge network affect the effectivenessof its search for useful knowledge by being importantconduits for information about opportunities the existence, whereabouts, a nd relevance of substantiveknowledge residing in other business units. While busi nessunits in the network may not be able to pass onproduct-specific knowledge directly, as such knowledgeoften requires direct interaction with the source to be extracted,a focal team that hears about opportunitiesthrough the network can cont act the source directly toobtain the knowledge. Sucknowledge,as defined here,incl udes product-specific technical know-how, knowledgeabout technologies and mark ets, as well as knowledgeembodied in existing solutions, such as already develop edhardware and software.

Although direct relations in the knowledge networkprovide immediate access a nd hence areespecially usefulfor a focal team inquiring about opportunities, indire ctrelations are beneficialas well, because information aboutopportunities is likely t o be passed on by intermediaryunits and eventually reach the focal team, provide d thatbusiness units in the knowledge networkare reachable.1The idea that interm ediaries pass on messages and thatthey help forge connections has been well sup ported incommunications and social network research. Studies investigatingthe “s mall-world” phenomenon demonstratedthat the path length (i.e., the minimum nu mber of intermediaries)needed to connect two strangers from differentstates in the United Stateswas remarkably short and consistedof about five to seven intermedia ries (Milgram1967, Kochen 1989, Watts 1999). Early work on innovationresearch showed that new product developmentteams benefited from having a gatekeeper o r boundaryspanner, that is, a person who scans and interprets theteam’s environm ent and then passes on information to therest of the tea (Allen 1977, Katz and Tushman 1979).In social network research, Granovetter (1973) showed

that intermediary persons who are weakly tied to a focalperson are uniquely plac ed to pass on information aboutnew job opportunities because they are more like

ly thanstrongly tied connections to possess nonredundant information.

The common thread in these lines of work is thatindirect relations are pervasi ve conduits for information.Intermediaries help forge connections and pass on me ssagesthat bridge two otherwise disconnected actors.

However, indirect interunit relations may not be perfectconduits of informationa bout opportunities. As informationgets passed on across people from different uni ts,there is likely to be some degree of imperfect transmissionof the message abo ut opportunities for knowledgeuse. In particular, when information about opportun itieshas to be passed on through many intermediaries (i.e.,through long paths, cf. Freeman 1979), it is likely to becomedistorted (Bartlett 1932, March and Simon 1958).People who exchange such information are prone to misunderstandingeach other, forgetting details, failing tomention all that they know to others, filtering, or deliberatelywithholding aspects of what they know (Collinsand Guetzkow 1964 Huberand Daft 1987, Gilovich1991). The distortion may be unintentional or delib erate(O’Reilly 1978). Huber (1982) relates a drama tic example,originally provided by Miller (1972), of a mistakemade during the Vietnam War. The chain of mess ageswas as follows: The order from headquarters to the brigadewas “on no occas ion must hamlets be burned down,”the brigade radioed the battalion “do not bur n down anyhamlets unless you are absolutely convinced that the VietCong are in them;” the battalion radioed the infantry companyat the scene “if you think there are any Viet Congin the hamlet, burn it down;” the company commanderordered his troops “burn down that hamlet.” Thus, themore intermediaries needed, the hig her the chances ofsuch distortion, and hence the less precise is the informationth at is passed on (Miller 1972, Huber 1982).

The implication of receiving imprecise information inthis context is that a proj ect team cannot easily focus ona few opportunities that are especially relevant, b ut mustinstead check anumber of imprecise leads to verifywhether they are releva nt for the team, resulting in a moreelaborate interunit search process that takes ti me. For example,a project manager in my study told me that he hadbeen told b y a third party in the company about a groupof engineer in another unit who w

ere supposed to havesome useful technical know-how, but when he was ableto r each them after trying for a while, it turned out thatthe know-how was not relev ant for the project. Such fruitlesssearches not only take time, but also cause dela ys inthe project to the extent that the needed knowledge inputholds up the comp letion of other parts of he project.Because of the problem of information distorti on whenrelying on intermediary units, a focal team is likely tobenefit from short path lengths in the knowledge network(i.e., few intermediaries required to connec t a team in afocal unit with other units). Short path lengths enable theteam to k now about precisely described opportunities involvingrelated knowledge and allow it to discard informationabout irrelevant opportunities. The team can thenfocus on opportunities with a high degree of realizationpotential and can quickly contact p eople in these unitsand begin working with them to extract and incorporatetheir knowledge into the focal project. Thus, less time isspent evaluating and pursuing opportunities, reducing effortsdevoted to problemistic search, including search effo rtsthat establish that no useful opportunities exist(Cyert and March 1992). Teams with short path lengthsare thus more likely than teams with long path lengths to hear about more opportunities that overall yield more usefulknowledge, to the ext ent that opportunities are notredundant to one another. All else equal, this benefi tshould reduce a focal team’s time to complete t he project.The arguments can be summarized in two hypotheses.

HYPOTHESIS 1. The shorter a team’s path lengths inthe knowledge network, the more knowledge obtainedfrom other business units by the team. HYPOTHESIS 2. The shorter a team’s path lengths inthe knowledge network, the shorter th project completiontime.

Effects of Direct Relations in Knowledge Networks

The shortest possible path length is to have an establisheddirect relation to all other business units in a knowledgenetwork. Such a network position does not re quire anyintermediary units and should remove the informationdistortion caused b

y using intermediaries. However, unlikeindirect relations, which are maintained by intermediarybusiness units, direct interunit relations need to bemaintained by peo ple in the focal business unit, possiblyincluding focal team members, and require their own setof activities that take time. In the company I studied, forexample, product developers spent time outside of theirprojects traveling to other business units on a regular basisto discuss technology developments, market opportunities,a nd their respective product development programs.Such interunit network mainten ancecan be adistraction from completing specific project tasks: Timespent on mai ntaining direct contacts is time not spent oncompleting project-related tasks. Although direct interunit relations involve maintenancecosts, they also provide a benefit incertain situations:Established direct relations between a focal team and anotherbusiness unit may be helpful when the team identifiesknowledge that requ ires effort to be moved from thesource unit and incorporated into the project. Fo r example,in a number of projects in my sample, team memberswere frequently able to obtain software code from engineersin other business units, but sometime s the engineerswho wrote the code needed to explain it and help the teamto inc orporate the code into the new project. Receivingsuch help was often much easie r when the team and theengineers providing the code knew each other beforehan d.

This likely positive aspect of direct relations needsto be compared with their maintenance costs.Direct relations are especially helpful when a team isexperienci ng transfer difficulties—i.e., spending significanttime extracting, moving, and inco rporating knowledgefrom other subunits—because the knowledge is noncodified,w hich is defined as knowledge that is difficultto adequately articulate in writing (Zander and Kogut1995, Hansen 1999). Relying on establisheddirect relationsmay ease the difficulties of transferring noncodifiedknowledge, because the team and people in the directlytied unit have most likely worked with each other beforean d have thus established some heuristics for workingtogether, reducing the time itt akes to explainthe knowledgeand understand one another (Uzzi 1997, Hansen199 9). When a focal team experiences significant transferdifficulties because of nonc

odified knowledge, having establisheddirect relations to related business units is li kelyto reduce the amount of time spent transferring knowledge,which may offset the costs of maintaining such relationsand shortening project completion time. In particular,having a number of direct relations in a knowledgenetwork increases th e likelihood that a team will be ableto use one of them in transferring noncodifi ed knowledge.

Thus, while indirect relations are beneficial to the extentthat they serve as inte rmediaries that provide a focal unitwith nonredundant information, direct relations are beneficialto transferring noncodified knowledge, implyingthat the benefit of ha ving intermediaries supplying nonredundantinformation is relative (cf. Burt 1992).I n contrast, this transfer benefit of direct relations isless important when a focal t eam can easily extract andincorporate the knowledge that was identified in anoth ersubunit, as when that knowledge is highly codified. Inthese situations, direct int erunit relations are not usefulfor transfer, but they still carry maintenance costsw hichtake time away from the completion of the project to theextent that team me mbers d not have slack resources thatcan be devoted to maintaining these relatio nships. Themore suchrelations that are maintained by a focal unit,the higher the maintenance costs, and the more time istaken away from completing a project. T he arguments canbe summarized as follows:

HYPOTHESIS 3A. The higher a team’s number of directrelations in the know ledge network, the shorter the projectcompletion time when the knowledge to be transferredis noncodified.

HYPOTHESIS 3B. The higher a team’s number of directrelations in the knowl edge network, the longer the projectcompletion time when the knowledge to be tr ansferred iscodified.

Data and Methods

Setting

I tested the knowledge network model in a large, multidivisionaland multinatio nal electronics company (hereaftercalled “the Company”). I negotiated access to t hecompany through three senior corporate R&D managersand initially visited 14 divisions where I conducted openendedinterviews with 50 project engineers and managersto better understand the context, and todevelop surveyinstruments. The c ompany, which has annual sales ofmore than $5 billion, is involved in developin g, manufacturing,and selling a range of industrial and consumerelectronics produc ts and systems, and is structured into41 fairly autonomous operating divisions tha t are responsiblefor product development, manufacturing, and sales.

By focusing on these divisions, I was able to compareunits that occupy the sa me formal position in the Company,thereby controlling for a potential source of variationin formal structure. They all had the same formalstatus as a business uni t with profit-and-loss responsibility,all had a general manager, and none of the di visionsreported to another division. In additio to interunit relations,there were a f ew other integrative mechanismsacross divisions, notably divisionwide conferences andelectronic knowledge management systems, but initial interviews

revealed that these did not vary much among thedivisions.

Selecting Product Development Projects

I used two surveys: a network survey administered to theR&D managers in th e 41 divisions and a survey for theproject managers of the product development projects included

in this study. In selecting projects, I first createda list of all projects that the di visions had undertaken duringthe three-year period prior to the time of data colle ction.I then excluded very small projects (i.e., those withless than two project en

gineers) and projects that had not

yet moved from the investigation to the developmentphase and were therefore ha rd to track I also excludedidiosyncratic projects that had no meaningful start and end (e.g., special ongoing customer projects). Includingonly successfully complete d projects may lead to an overrepresentationof successful projects, biasing the res ults.

I therefore included both canceled projects and projectsstill in progress. After having removed too-small, premature,and idiosyncratic projects, I ended up with a listof 147 projects. The project managers of 120 of thesereturned their survey, yielding a response rate of 85%. Ofthe 120 projects, 22 were still in progress at the time ofdata collection, four had been canceled, and 54 reporteda significant t ransfer event involving another division.

Specifying Project-Specific Knowledge Networks

Identifying Related Subunits. Together with the threecorporate R&D managers, I developed a list of 22 technicalcompetencies that constituted related knowledgea reas(see Appendix 1 for the list of technical competencies).2 I asked the R&D managers in the divisions to

indicate up to four specific competencies of their divisionson this list and to add any if they thought the listwas incomplete. The three corporate R&D managers r eviewedthe responses to verify whether it made sense togroup those divisions tha t had reported the same competence.The project managers of the 120 projects we rethen asked to indicate what technical competencies thespecific project required and were presented with thesame list that was presented to the divisional R&D managers.

Thus, for a given project, a number of divisionshad a competence that matche d the requirements listedby the project manager (see Appendix 1 for the distribut ionof projects per competence). For example, a projectmanager indicated that his project required technicalcompetencies in three areas: distributed measurement,com

munication system monitoring, and optics. Twelve different divisions had at least one of these technical competenciesand thus constituted theknowledge network fo rthis particular project.Specifying Interunit Relations. A group of engineers ina di vision typically maintained an informal regular contactwith a group of engineers inanother division, and aproject team would use such contacts to access other di visions.These relationships were common knowledge inthat most product develope rs seemed to know about theirexistence and how to use them, and I was told in preliminaryinterviews that a main responsibility of a division’sman agers was to p rovide these contacts for his or herproject teams,should the need arise. I therefor e assumedthat at least one member of a project team woul knowabout the divisi onal-level contacts and that the teammembers could access these contacts if they wanted to.Because of the importance of these interdivisional contactsin the compa ny, I chose to focus on these types ofcontacts.Following previous research, I use d a key informant toobtain information on interdivisional relations (Knokeand Ku klinski 1982, Marsden 1990). I considered the divisionalR&D managers to be the most appropriate informantsbecause they were “in the thick of things” in theR& D department in their division. The R&D manager ineach of the 41 divisions re ceived a questionnaire asking,“Over the past two years, are there any divisions fr omwhom your division regularly sought technical and/ormarket-related input?”3 T he question was followed by alist of the 41 divisions included in the study, allo wingrespondents to indicate whether they had a tie to any onthe list, leading to a complete network where everybodywas asked whether a tie existed with everyb ody else(Marsden 1990). Because I asked everybody to indicatewhether a tie exis ted with each of the other 40 divisions,I avoided a potential bias resulting from having to asksomeone to ascertain whether ties exist among others(Krackhardt an d Kilduff 1999).

To validate the responses, I employed the crossvalidationmethod used by Krac khardt (1990by askingthe R&D managers who comes to them for input. Anactual tie exists when both divisions agree that one comesto the other for input. I then sent an e-mail to all of theR&D managers, asking them about the ones about w

hichthere was no joint agreement. On the basis of their responses,I included som e of these suspect ties and excludedothers.

Merging Network and Project Data. I constructedproject-specific knowledge net works by including all relationsamong divisions possessing related knowledge for

a given project. For example, for the aforementioned projectfor which there were

12 related divisions, I includedall relations among these 12 divisions, and this ne tworkconstituted the project-specific knowledge network. Toconstruct these project -specific networks, I merged theproject data with the divisional network data by ass igninga division’s network relations to its projects. Thus, interdivisionalties bec ame the equivalent of interdivisionalproject ties. It is important to record thevalu es on thenetwork variables prior to the start of a project becausemy theoretical a rguments assume that a project team usesestablished preexisting interunit ties to s earch for andtransfer knowledge. Following the approach of Burt(1992) and Podo lny and Baron (1997), I handled this issueby measuring the interdivisional netwo rk relationsover several years by only assigning network ties thatexisted prior to the start of the project. This proceduregenerated time-varying network data from informationthat the respondents could recall.

The potential bias in this approach is that it may excludesome relations that e xisted prior to a project’s startbut that ceased to exist by the time the R&D ma nagerscompleted the survey. This problem can be partially controlledfor. This pot ential bias should be more of a problemfor projects in divisions in which relatio ns come andgo than in divisions with long-lasting relations. If a division’srelatio ns are long lasting, then it is less likelythat there were some relations that cease d to exist betweenthe time just prior to the project’s start and the time

of surveying. To control for this potential bias, I entereda control variable for th e average age of direct relationsto related subunits (Age relations).

Dependent Variables

Project Completion Time. To assess project task performance,I measured projec

t completion time as thenumber of months from the start of concept developmen tto the time of marketintroduction for a given project (ortime to the end of the study period or cancellation forongoing and canceled projects, respectively). I def inedstarting time as the month when a dedicated personstarted working part or f ull time on the project, whichtypically coincided with the time an account was o penedfor the project. I defined the end date as the date on which

the product was released to shipment, which is a formalmilestone date in this co mpany because it signifies thatthe product is ready to be manufactured and shipp ed ona regular basis. These definitions turned out to be veryclear and provided f ew problems in specifying the start

and completion times, which were 14.8 months on averagefor completed projects. Scholars have proposed two alternative measures ofcompletion time. First, com pletion timecan be measuredas the extent to which the project is finished on sch edule(e.g., Ancona andCaldwell 1992). The assumption in thisschedule measure is that inherent project differences areaccounted for in the original schedule, but als o that everybodysets equally ambitious schedules, which was mostlikely not true in this company, where individual projectmanagers set their own targets. A secon

d approach is togroup projects according to som

e similarity measure andthen tak

e a project’s deviation from the mean completiontime o

f the group (Eisenhardtan d Tabrizi 1995). Theproblem with this approach is that the mean deviationrelies on a clearsimilarity measure that was not easy toattain in this setting. Given that these two alternativemethods seemed problematic, I chose to use the numberof m onths as the dependent variableand then add projectspecificvariables to control for inherent differences betweenthe projects.

Amount Acquired Knowledge. During field interviewsI was told that the most c ommon knowledge that projectteams received from other divisions took the form of technicalsolutions embodied in already developed softwarecode and hardware components. T here were two types of“ware” being used in the projects—standar input to theproducts being made (e.g., components that were used innearly all os cilloscopes being manufactured), and warethat helped solve ad hoc problems that

were unique to agiven project (i.e., technical know-how that had been embodiedi n software code or hardware). While the formerwas typically handled within divi sions, the latter was typicallyobtained through interdivisional network contacts.Bec ause my theoretical analysis focuses on knowledgethat was obtained to solve ad hoc problems for a project,I chose to focus on software and hardware that the f ocalteam obtained from other divisions to solve emergingproblems. With a few e xceptions, most of the ware obtainedfrom other divisions was of this kind.4 Duri ng pretests,project managers thought they could indicate theamount of ware obtai ned from other divisions fairly accurately.

The project manager was asked to indicate thepercentage of all the project’s s oftware andhardware thatcame from other divisions in the company (see Appendi x1 for the specific question). To construct the dependentvariable, I computed thef raction of ware (ranging from

zero to one) that came from other divisions (Amount acquiredknowledge).While e ngineers also obtained other types of knowledgefrom other divisions, such as inf ormal technical advice notembodied in either software or hardware, these were m oredifficult to quantify, and I therefore did not develop a separatedependent varia ble for these types. However, I did askthe project manager to indicate the extent to which theteam had obtained such knowledge, and thismeasure correlated0.7 wi th the chosen dependent variable. It is thuslikely that my measure of amount acq uired knowledge isa proxy for more informal types of knowledge obtainedthroug h the network in this setting.

Independent Variables

Path Lengths in a Knowledge Network. I relied on geodesicsto compute the di stances in the network. A geodesicis the shortest path length (i.e., the one with fewestintermediaries) between a focal division and another divisionin a knowledg e network (Wasserman and Faust1994). However, the measure is complicated bec ause severalof the project-specific knowledge networks were disconnectedin that s

ome divisions did not have a tie withother divisions in the knowledge network. I handled thisproblem by creating a control variable that indicates the

fraction of related divisions that were reachable in aknowledge network (Reach). This variable takes on avalue of zero if no divisions were reachable (i.e., therew ere no paths connecting the divisions) and a value ofone if all divisions in the project-specific knowledge networkwere reachable (the mean value for this variabl e is0.85).I used the measure of closeness centrality to measurepath lengths in the network (Freeman 1979). Closenesswas measured as (Wasserman and Faust 1994) where d(ni, nj) is the geodesics linking divisions ni and nj.Summing over all r eachable related divisions excludingthe focal one (g _ 1), this gives division ni’s total closenessscore. Thimeasure is standardized, so that a divisionhas the shorte st path length (i.e., is closest) to relateddivisions when the index is one and the longest pathlength when the index is near zero (Close related). Thesemeasures w ere computed in UCINET IV (Borgatti et al.1992).

Direct Relations with Divisions in a Knowledge Network.Because direct relation s were asymmetric in thenetwork in the Company, I distinguished between direct relations in which the focal team went to other divisionsfor advice (i.e., advice-s eeking relations) and direct relationin which other divisions went to the focal on e foradvice (i.e., advice-giving relations). Each type of relationsimplies different c osts. Advice-seeking relationsneed to be maintained, while advice-giving relations requiretime helping others. I coded the number of directadvice-seeking relations to related divisions by countingthe number of preexisting divisional ties to divisi ons thathad related knowledge for a project and then assigned thatvalue to the f ocal project (Outdegree related). I thencoded the number of direct advice-giving relations to relateddivisionsby counting the number of preexisting divisionalties in which a related division reportedly wentto the focal division for advice on a reg ular basis (Indegreerelated).

To control for the possibility that these variables aresimply an indication of th e division’soverall number ofdirect relations, I als o included similar measures for directrelations outside a project’s knowledge network. I subtractedthe number of r

elated advice-seeking ties from thetotal number of direct advice-seeking relations for the focaldivision to arrive at the unrelated advice-seeking ties(Outdegree unrel ated). I subtracted the number of relatedadvice-giving ties from the total number of advice-givingties to compute the number of unrelated advice-givingties (Indegr ee unrelated)

.Finally, I included a measure of the strength of relatedadvice-seeking ties. Pre vious research has shown thatweak ties may facilitate search but impede the tran sfer ofcomplex knowledge (Hansen 1999). Although the theoryin this paper does not pertain to the effects of tie weaknesson interunit knowledge transfers, I want ed to control forthe possible effect of tie weakness. Tie weakness wascomputed by asking the R&D divisional managers to indicateon a seven-point scale how fr equently people intheir division talked to people in the other division andhow cl ose their working relationship was (see Appendix1 for the specific questions). I t ook the average frequencyand closeness for related advice-seeking ties to comput ethe measure (Strength related).Noncodified Knowledge. I constructed a three-item scale of noncodification (see Appendix 1 for the specificitems) and asked the pro ject manager to indicate the level

of codification of the knowledge that the project teamreceived from other divisio ns (Noncodified). This variablewas then interacted with the number of relatedadvi ce-seeking relations to test the hypothesis (Noncodified_ outdegree related). Alternative Explanations. I included variables to tesfor the possibility that eithe r short pat lengths or relatedknowledge (but not both) explains the amount of ac quiredknowledge and product development time. First, I includedan overall closen ess centrality measure by usingthe above equations for the closeness centrality m easure,using the entire set of 41 divisions as he relevant network(Close all).

To make this analysis comparable to the restof the analysis, I also included a variable indicating thetotal number of direct advice-seeking relations (Outdegreeal l). If the estimate for the general closeness measureis positive and significant, th en thenetwork argumentabout the importance of close positions (irrespective ofkn owledge relatedness) is plausible. To capture the extentof related knowledge avail

able to a project team, I includeda variable measuring thenumber of related divis ions(No. related units). If this measure of the extent ofrelated knowledge in the Company is positive and significant,then the argument about the importance of re latedknowledge (irrespective of network relations) isplausible.

Control Variables

Betweenness Centrality. Because the closeness centralitymeasures may be correl ated with other centrality measuresthat attempt to capture other causal mechanism s, included a measure of betweenness centrality, which isoften used to measure a focal actor’s brokering positionin the network (Freeman 1979, Brass and Burkhar dt1992, Burt 1992). Divisions with high betweenness maybe in a powerful positi on where they can control the flowof information betweentwo other units, thus u sing thisbenefit to obtain favors from others, such as help in transferringknowled ge. To control for this power-orientedbenefit of central positions, I included a m easure of betweennesscentrality (Wasserman and Faust 1994)where gjk is the nu mber of geodisics linking division j and k, and gjk (ni) is the number of geodisic s linking division j and k that involve the focal division i. The measure is asum of the probabilities that the focal division will fallon the geodesics linking all pa irs of related divisions. Themeasure is standardized as follows:where the denomin ator is the number of pairs of divisionsnot including the focal division i. This m easure rangesfrom zero to one, where one is the maximum related betweennessa mong related divisions (Between related).

Project Attribute Controls. To make the projects comparable,I controlled for se veral project-specific factors. Icontrolled for the extent to which the project useds oftwarecontrols to soe extent for a project team’s motiva tionto conduct searches t hrough the interunit network.The team should be less motivated to the extent tha t itcan use existing ware inside its owndivision. Projectmanagers were asked to i ndicate the percentage of allsoftware and hardware in theproject that they reused orleveraged from their own division (Own existing ware).

I used the log of estimated dollar costs at the start ofthe project to control fo r size and scope differences betweenthe projects (Budget).5 In my field interview s withproject managers, I was also told that estimated costs captureinherent differ ences in technical complexity amongthe projects (the more complex the technolog y, the moreengineering hours billed to the project). I used the budgetfigure to av oid an interaction between final costs and thedependent variable. High final costs may reflect longcompletion time because of more engineering hoursbilled to the project.

I also coded whether a project-specific patent was appliedfor, to measure degre e of innovation (Patent), andwhether the project team developed a product or a system(Product). More innovative projects presumably takelonger to complete. Th e product-systems distinction was

entered as a variable to control for possible differencesbetween these two categor ies with respect to crossdivisionalknowledge use. Each variable was coded as adu mmy variable, where avalue of one indicates a patentand a product, respectively. Finally, because strictly personal relations spanningsubunits may be used by te am membersto obtain knowledge,I entered a control measure that was obtained fr oma third survey that was sent to all engineers on the projectsin the sample (see Hansen et al. 2001). Engineers wereasked to indicate the number of advice-seeki ng relationsthat they personally had to people in other divisions. Ithen summed t hese relations for a team (excluding contactnames mentioned more than once) toa rrive at a teamlevelmeasure of direct interpersonal relations spanningsubunits (Per sonal relations).

Statistical Approach

Because 66 projects did not report any knowledge usefrom other divisions, the dependent variable “amount acquiredknowledge” was set to zero for these projec ts. Becauseof this largenumber of observations with a valueof zero, a least squar es regression model was inappropriate,and I employed a tobit model, using maxi

mum likelihoodestimation (Maddala 1983, Greene 1993).

In addition, the statistical analysis of completion timewas complicated by the f act that 22 of the 120 projectswere still ongoing at the time of data collection. The dataset,therefore, includes right-censored cases (Tuma andHannan 1984). Furt hermore, four projects were canceled.Because the dataset contains right-censored data, ordinaryleast squares regression analysis cannot be employed(Tuma and Han nan 1984), but the problem of right censoringcan be dealtwith by using a hazard rate model. Inthis approach, a project enters the risk set from the timeit was star ted and leaves the risk set when it is completedor canceled. The instantaneoustra nsition rate—the dependentvariable—is a measure of the likelihood of aproject eit her completing or terminating at time t, conditionalon it not having completed or terminated before t. The higher the transition rate, the more likely the projectwill be completed faster. The hzard rate model takesthe following form:where r(t)j i s the completion rate of project j, t is projecttime in the risk set, and r(t)j* is t he completion rate includingthe effects of all of the control variables in themodel. The effects of the independent variables are specifiedin the exponential bracket a is a vector of estimatedcoefficients, and C is a vector of independent variable s.

I used the piecewise exponential specification as implementedin the statistical program TDA, because I didnot want to make any assumption about duration de pendencethat would require a specific parametric distribution.I controlled for durat ion dependence, however, becausethe survivor plot revealed a nonmonotonic curv e(cf. Tuma and Hannan 1984). The plot revealed severaltransition phases occurrin g at 10, 12, 15, 18, and 21months and I therefore enteredsix time-period variabl esthat reflect the time distribution of events. The transitionrate is assumed to be constant within these periods, andcovariates are assumed not to vary across time periods(Blossfeld and Rohwer 1995).

Because multiple projects belong to a division, it ispossible that project-specifi c observations are nonindependentbecause they vary with divisional attributes.I th erefore chose a fixed effect specification and entered26 dummy variables, one for