The impact of context and model choice on the determinants ... · exponential random graph models (ERGMs), a recent methodological advance that allows the explicit modeling of underlying

Strategic Management JournalStrat. Mgmt. J., 37: 2204–2221 (2016)

Published online EarlyView in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/smj.2570Received 5 October 2014; Final revision received 18 February 2016

THE IMPACT OF CONTEXT AND MODEL CHOICEON THE DETERMINANTS OF STRATEGIC ALLIANCEFORMATION: EVIDENCE FROM A STAGEDREPLICATION STUDY

ANINDYA GHOSH,1,2† RAM RANGANATHAN,3*† and LORI ROSENKOPF4†

1 Strategy Department, Indian School of Business, Hyderabad, India2 Department of Entrepreneurship, IESE Business School, University of Navarra,Barcelona, Spain3 Department of Management, McCombs School of Business, University of Texas,Austin, Texas, U.S.A.4 Department of Management, The Wharton School, University of Pennsylvania,Philadelphia, Pennsylvania, U.S.A.

Research summary: Endogenous characteristics of alliance network structure have repeatedlybeen shown to predict future alliance ties in the strategic management literature. Specifically,the concepts and measures of relational, structural, and positional embeddedness (per Gulatiand Gargiulo, 1999), as well as interdependence, are foundational for many studies. We explorethese determinants of alliance formation by replicating the baseline analyses of Ahuja, Polidoro,and Mitchell’s, 2009 SMJ article. We examine the impact of empirical choices with respect totime period, underlying data generating model, and industry by isolating each effect in turn. Wedemonstrate that while geographic similarity and product-market similarity each robustly predictthe interdependence effect, the effects of both technological similarity as well as the embeddednesspredictors are sensitive to context and/or method.

Managerial summary: Our examination of alliance formation in the chemical and semiconductorindustries during the 1990s demonstrates how new alliances may be predicted by both thetechnical, geographic, and product-market fit of potential partners as well as characteristics ofeach partner’s previous network participation. Comparing our results to an earlier study, we findthat geographic and product-market similarity predict alliance formation across both industriesand time frames while prior ties between partners predict alliance formation only when theseindustries are less mature. Other network participation indicators generate nuanced effects, whichunderscore the importance of quasi-replications of alliance formation across industries and timeperiods in building evidence-based management theories. Copyright © 2016 John Wiley & Sons,Ltd.

INTRODUCTION

The approach of predicting dyadic alliance tiesbetween firms by examining alliance network

Keywords: alliances; networks; endogeneity; replication;ERGM*Correspondence to: Ram Ranganathan, Management Depart-ment The University of Texas at Austin McCombs School of Busi-ness 2110 Speedway Stop B6300 Austin, TX 78712-1282.E-mail: [email protected]†Authors contributed equally.

Copyright © 2016 John Wiley & Sons, Ltd.

structure is deeply entrenched in the strategicmanagement literature (e.g., Ahuja, 2000; Ahujaet al., 2009; Ahuja, Soda, and Zaheer, 2012;Chung, Singh, and Lee, 2000; Garcia-Pont andNohria, 2002; Gulati, 1999; Li and Rowley, 2002;Rosenkopf, Metiu, and George, 2001; Rothaermeland Boeker, 2008; Wang and Zajac, 2007). All thesestudies manifest “an endogenous dynamic” (Gulatiand Gargiulo, 1999: 1453) in that the pattern of prioralliances in the network, termed “network embed-dedness,” predicts subsequent alliance formation.

The Determinants of Strategic Alliance Formation 2205

According to Gulati and Gargiulo (1999), thisembeddedness can be relational (a function of theprior history of ties between the firms in a dyad),structural (a function of a small subset of the net-work, such as transitivity), or positional (a functionof the full network, such as joint centrality). Ofcourse, in these studies, alliance formation is alsopredicted by characteristics that can be assessedindependently from the actual presence of thealliance ties themselves (e.g., Chung et al., 2000;Rothaermel and Boeker, 2008; Wang and Zajac,2007); constructs such as interdependence1 derivefrom the characteristics of each firm in the dyad.

Despite the widespread use of this approach, sev-eral empirical concerns exist. First, network effectsmay be sensitive to the industry environment firmsinhabit (Rowley, Behrens, and Krackhardt, 2000) aswell as the time period within an industry (Mad-havan, Koka, and Prescott, 1998). Second, scholarshave called attention to the issue of endogeneityin modeling network ties (e.g., Ahuja et al., 2012;Stuart and Sorenson, 2007). Traditional regressionapproaches to predicting alliance formation treat theexisting alliance network as exogenous, unrealisti-cally assuming that firms do not act strategically toachieve their current network positions (Stuart andSorenson, 2007). Similarly, such approaches alsoignore the real interdependencies between alliancechoices across firms such as Gimeno’s (2004)demonstration that firms that are competitors inthe product-market domain are influenced by eachother’s alliance choices. Finally, composite networkstatistics used to represent network embeddednessmechanisms, such as joint centrality, can mask mul-tiple underlying network generating processes, asthe measure is aggregated across multiple networklevels.

To examine the theoretical and empirical implica-tions of these issues, we replicated the baseline anal-yses from Ahuja, Polidoro, and Mitchell (APM),published in SMJ in 2009. APM demonstratedeffects of positional embeddedness via combinedcentrality; strategic interdependence via techni-cal, geographic, and product-market similarity; andrelational embeddedness via previous alliance ties,using the preferred network modeling techniqueavailable to strategy scholars at that time. APM’seconometric specification is largely comparable to

1Sometimes termed variously as “homophily,” “heterophily,”“similarity,” or “complementarity” in the alliance literature(Ahuja et al., 2012; Rothaermel and Boeker, 2008).

the seminal Gulati and Gargiulo (GG) work,2 buttheir work is more amenable to replication by itsfocus on a clearly bounded industry.3

Our replication proceeds in three stages. First,we reproduce the APM estimation strategy onpublicly available data in their same industry(chemical), only varying the time period (ourdata span 1991–2000 while their proprietarydata span the prior decade). Next, holding theindustry and time period constant, we addressmodel specification and estimation concerns usingexponential random graph models (ERGMs), arecent methodological advance that allows theexplicit modeling of underlying network formationprocesses and dependencies in the alliance networkdata. Finally, we examine the role of industry con-text by repeating the ERGM analysis for alliancedata in the semiconductor industry over the sametime period.

Our results generate two important implica-tions for future alliance formation research. First,we demonstrate both robustness and sensitivity ofresults across industry and time period differences.With regard to robustness, we find that geographicand product-market similarity each consistentlypredicts alliance formation in both our contexts. Incontrast, we find that the maturity of an industrycan substantially alter the effect of both techno-logical similarity and previous ties, suggesting theneed for future research to incorporate such con-tingencies and boundary conditions. Second, wedemonstrate that traditional measures of positionalembeddedness, such as joint centrality, conflatemultiple network generating processes. By usingrecent advances in ERGM techniques to decomposethese processes empirically, we are able to suggest

2The main difference between the specification of APM and thatof GG is that APM does not include a Structural embeddednessmeasure. We examine this issue when we replicate using analternative method.3While the GG study represents the most comprehensive artic-ulation of endogenous determinants of alliance formations, twointerrelated challenges precluded an effective replication. First,their sample was largely proprietary and used a combination ofthe Cooperative Agreements and Technology Indicators (CATI)database and several other hand-collected sources. Second, theirdata set combines data from three different industrial settings(automotive products, new materials, and industrial automation),which do not correspond to well-defined standard industry classi-fications for which other established and more commonly avail-able data sets could be searched via SIC codes. In comparison,while APM’s study also used proprietary data, it was set in theglobal chemical industry, constituting a relatively unequivocalcomposition of firms.

Copyright © 2016 John Wiley & Sons, Ltd. Strat. Mgmt. J., 37: 2204–2221 (2016)DOI: 10.1002/smj

2206 A. Ghosh, R. Ranganathan, and L. Rosenkopf

more precise network predictors to refine conceptslike positional embeddedness by grounding them inactual tie-forming mechanisms such as preferentialattachment and transitivity.

REPLICATION STAGE I: ISOLATINGTIME DIFFERENCES

Sample, data sources, and measures

We collected data for all alliances between firmsin the global chemical industry for the years1991−2000 using SDC Platinum, the largestcross-industry data set of strategic alliances. AsSchilling (2009) notes, the SDC Platinum data setprior to 1990 is quite sparse, so we chose 1991as the first year of our alliance observations.4

We chose a 10-year period to closely matchAPM’s breadth (9 years, 1983–1991). To buildthe industry-based alliance network, we selectedonly those alliances that had both partner firmsbelonging to the focal industry (and at leasttwo partners in the focal industry for multipartyalliances). We matched the firms participatingin these alliances to company background andfinancial data available in Compustat—UnitedStates, Compustat Global Fundamentals, BureauVan Dijk (BvD), and OSIRIS, and supplementedthis with industry-specific data from Chemicaland Engineering News (CEN) and DataQuest.We matched these data using Committee on Uni-form Security Identification Procedures (CUSIP)identifiers, Global Company Key (gvkey), stockmarket tickers, and company names.5 We furtherwinnowed the set of firms to the largest 150 firmsin each industry in each year by revenue.6,7 We

4Like APM, we collected additional presample data (in ourcase, from 1986 to 1990) to build a “Prior Ties” measure forour replication, but we did not use this period to measure thedependent variable.5In instances in which a subsidiary firm was involved in thealliance and the subsidiary firm’s financial data were not reportedseparately, we matched the data to the parent firm’s financialsusing the same techniques. To account for industry mergers, weused the acquiring firm or the merged entity’s financials whereseparate financials were not reported for the constituent firms inthe year of the alliances.6APM’s sample in comparison consists of the largest 97 firms inthe chemical industry.7For firms whose financials were reported in an internationalcurrency, we converted to USD using foreign exchange ratesavailable from the Federal Reserve Bank data set (accessedthrough WRDS).

collected patent data for these firms from the NBERpatent data set using the approach specified in theBronwyn Hall Patent Name Matching project tomatch company names to patent assignees (Hall,Jaffe, and Trajtenberg, 2001). Further, like APM,we include those alliances that involved at leasttwo of the aforementioned 150 firms leading to afinal replication sample of 202 strategic alliancesbetween 139 firms. This level of alliance intensityis lower than that reported for the APM sample,which had 97 firms engaging in 338 alliancesduring their earlier time frame.

We replicated APM’s baseline model (see theirModel 2, p. 953). APM’s model includes the vari-ables Combined centrality and Combined central-ity squared to measure “positional embeddedness”(calculated using the geometric mean of the eigen-vector centrality scores of the two member firmsin a dyad), the variables Previous alliances andPrevious alliances squared to measure “relationalembeddedness” (calculated using the number ofalliances the two member firms in a dyad had in thepast) and the variables Technical similarity, Techni-cal similarity squared (calculated using similarityof patents), Geographic similarity (calculated usingsimilarity of international country presence), andProduct market similarity (calculated using sim-ilarity of industries firms participate in) to mea-sure “interdependence.” APM also include dyadiccontrols for Size, Liquidity, Debt-equity, Patentsand R&D. Our independent variables for the net-work embeddedness and technical similarity mea-sures were created in identical fashion to APM’svariables. For the geographic and product marketsimilarity, as we had less granular data than APM,we computed a direct similarity (homophily) vari-able based on country location and primary indus-try participation. Among the controls, we createdidentical measures to APMs the only exception wasthat we used a Debt-to-Asset ratio instead of aDebt-to-Equity ratio due to data limitations. LikeAPM, we created a binary dependent variable set to1 if the two firms in the dyad formed an alliance in ayear, and 0 otherwise. We used a five-year movingwindow (years “t–5” through “t–1” for alliance for-mation year “t”) to construct our alliance formationnetwork variables.

Table 1 displays the descriptive statistics and cor-relations for the APM chemical replication sample.Although our time period is later than APM’s, oursample is comparable in terms of descriptive statis-tics (see APM, Table 1, p. 951). For instance, the



Tabl

e1.

Des

crip

tive

stat

istic

s(c

hem

ical

)

Var

iabl

esM

ean

S.D

.M

inM

ax1

23

45

67

89

1011

1213

1415

16

1A

llian

cefo

rmed

0.00

70.

080

11

2Jo

intc

entr

ality

0.02

40.

070

0.96

0.04

13

Join

tcen

tral

itysq

uare

d0.

006

0.03

00.

910.

020.

881

4C

omm

onal

lianc

epa

rtne

rs0.

100

0.4

05

0.03

0.75

0.64

15

Tech

nica

lsis

mila

rity

1.41

00.

520

20.

010.

140.

110.

131

6Te

chni

cals

imila

rity

squa

red

2.26

01.

230

40.

010.

160.

120.

140.

971

7G

eogr

aphi

csi

mila

rity

0.22

40.

420

10.

040.

10.

080.

090.

080.

081

8Pr

oduc

t-m

arke

tsim

ilari

ty0.

187

0.39

01

0.04

0.1

0.09

0.07

00

0.09

19

Prev

ious

allia

nces

0.01

70.

170

60

0.35

0.46

0.3

0.05

0.06

0.08

0.07

110

Prev

ious

allia

nces

squa

red

0.03

00.

620

360

0.29

0.49

0.29

0.03

0.03

0.06

0.06

0.82

111

Size

0.41

40.

290

10.

030.

020.

020.

02−

0.04

−0.

030.

120.

060.

040.

041

12Pe

rfor

man

ce0.

040

0.05

00.

62−

0.01

−0.

05−

0.05

−0.

030.

030.

03−

0.08

−0.

03−

0.02

−0.

02−

0.04

113

Liq

uidi

ty0.

710

0.19

0.1

10.

010.

090.

070.

06−

0.01

00.

090.

060.

040.

030.

08−

0.09

114

Deb

t-eq

uity

0.40

60.

30

10

0.08

0.07

0.05

−0.

02−

0.01

0.02

0.1

0.05

0.04

0.16

−0.

050.

091

15Pa

tent

s0.

197

0.29

01

00

00

−0.

10.

04−

0.03

−0.

030

00.

030

0.01

01

16R

&D

0.29

70.

290

10.

020.

040.

040.

04−

0.04

−0.

030.

060.

080.

050.

050.

51−

0.03

0.15

0.12

0.06

1 dependent variable Alliance formed has a mean of0.007 in our sample versus 0.01 in APM’s samplewhile the standard deviation is 0.08 versus 0.10.

Results using identical model

Following APM, we model the data as longitudinalpanels, creating a record of the dependent and inde-pendent variables for each unique dyad in the sam-ple for each year of the replication period. We poolthe observations and use probit regression (probit inStata) with year dummies and robust standard errorsclustered on the dyad to estimate the coefficients.8

To facilitate comparison, we include APM’s base-line results in the first column of Table 2.

Model 1 displays our replication results forAPM’s pooled probit method. Model 2 addsthe structural embeddedness measure (Commonalliance partners) set forth in the seminal GGmodel but not included in APM.9 The last columnof Table 2 compares APM’s baseline model toour replication results from Model 1. We foundan identical positive (marginal effect10 = 0.004in both APM and our replication) and significant(p-value< 0.001) effect for Geographic similarityand a comparable positive (marginal effect= 0.005in our replication versus 0.010 in APM) and signif-icant (p-value< 0.001) effect for Product-marketsimilarity. We also obtained a comparable positive(marginal effect= 0.036 in our replication versus0.029 in APM) and significant (p-value= 0.004)base effect of positional embeddedness (Combinedcentrality). Although we obtained a comparablecoefficient for the second-order term Combinedcentrality squared (marginal effect= -0.04 in ourreplication versus -0.03 in APM), it was not statisti-cally significant (p-value= 0.165 in our replicationversus p-value< 0.01 in APM). For Technical simi-larity, we obtained a weaker effect size (marginaleffect= -0.003) relative to APM (marginaleffect= -0.043 in APM) and the coefficient wasnot statistically significant (p-value= 0.545). For

8Following APM, we also confirmed that our results were consis-tent with random-effects models.9Note that the inclusion of structural embeddedness (which is notsignificant) fully preserves our Model 1 results, and is included toallows us to directly compare the probit models with subsequentERGMs, which include better specified measures for all the threeembeddedness mechanisms.10Calculated as change in the probability of observing an allianceif the variable is increased by one unit, holding all other variablesat their respective sample means.



Table 2. Replication of APM: chemical industry, probit regressions, 1991−2000a

APM replication (pooled probit)

Models variables APM original (pooled probit) 1 2

Alliance formation mechanismsPositional embeddednessCombined centrality 1.60 2.49 2.04

(0.32) (0.87) (1.07)Combined centrality squared −1.45 −2.83 −2.74

(0.51) (2.04) (2.09)Structural embeddednessCommon alliance partners 0.10

(0.10)Strategic interdependenceTechnical similarity −2.40 −0.21 −0.21

(1.13) (0.35) (0.35)Technical similarity squared 1.04 0.11 0.10

(0.39) (0.15) (0.15)Geographic similarity 0.25 0.28 0.28

(0.07) (0.08) (0.08)Product-market similarity 0.58 0.33 0.33

(0.13) (0.08) (0.08)Relational embeddednessPrevious alliances 0.44 0.74 0.70

(0.06) (0.77) (0.78)Previous alliances squared −0.03 −1.23 −1.20

(0.02) (0.39) (0.39)Dyad level controlsSize −0.01 0.37 0.36

(0.09) (0.14) (0.14)Performance −2.92 0.40 0.43

(1.02) (0.63) (0.63)Liquidity 0.21 −0.12 −0.12

(0.12) (0.19) (0.19)Debt-equity 0.00 −0.26 −0.26

(0.09) (0.14) (0.14)R&D 0.04 0.10 0.10

(0.09) (0.13) (0.13)Patents 0.17 −0.04 −0.03

(0.07) (0.16) (0.16)Constant −2.71 −2.67 −2.66

(0.88) (0.30) (0.30)Year dummies Included Included IncludedLog likelihood −1810.68 −650.16 −649.89

a Robust standard errors in parentheses.

relational embeddedness (Previous alliances andPrevious alliances squared), while the coefficientswe obtained are in the same direction as those ofAPM’s, we did not obtain significance for the maineffect (p-value= 0.336) and the coefficient for thesecond-order effect (Previous alliances squared) isnegative and significant (marginal effect= -0.018in our replication versus -0.001 in APM). Figure 1compares the overall effect from the first andsecond-order terms for relational embeddednessbetween our replication and APM’s original results.

Since the variable Previous alliances can take ononly positive integer values, we can conclude fromthe graph that the net effect of Previous alliances onalliance formation likelihood is positive for APM,but negative for our replication. This is driven by themagnitude of the negative coefficient for Previousalliances squared (-1.23) in our replication relativeto the positive coefficient for Previous alliances(0.74). In contrast, APM’s coefficients for Previousalliances and Previous alliances squared were 0.44and -0.03, respectively.



0

0.001

0.002

0.003

0.004

0.005

0.006

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 1 2 3 4 5 6 7

APM Replication

Figure 1. Effect of previous alliances (relational embed-dedness) on alliance formation. (Note: In the above graphX-axis measures Previous alliances, Y-axis (left) measuresLikelihood of alliance formation for APM’s study, Y-axis(right) measures Likelihood of alliance formation for ourprobit replication. All other variables are held at their mean

values for the respective samples)

In sum, while our probit replication results in thechemical industry in a later time period are broadlyconsistent with those of APM, providing confir-matory replications for many of their findings, weare unable to reproduce the effects for interdepen-dence in the technological domain, and we obtain adifferent effect for relational embeddedness in thealliance network. We will examine how the evolu-tion of the chemical industry may have driven someof the differences we observe in our Discussionsection.

Another explanation for replication differences:model choice

While the stage of the industry may account forsome differences, we also know that the applica-tion of traditional dyadic regression methods suchas probit to network data suffers from several short-comings (Stuart and Sorenson, 2007). First andmost well known, these methods assume indepen-dence across alliance ties, which many studies,including GG (see p. 1482), acknowledge as anissue (Robins et al., 2007; Stuart, 1998).11

11For instance, an alliance between one firm in a focal dyadand a competitor of the other firm in the focal dyad may createcompetitive pressures that result in a tie forming in the dyad.Dependencies may arise at distant points in the network aswell—an alliance between competitors of the two focal firmsin a dyad may precipitate an alliance forming between them.In the presence of such interdependencies, traditional methods

Second, predictors derived from one observednetwork inherently lack an adequate stochasticcomponent because they do not have a correspond-ing probability distribution function defined at thenetwork level. For example, on observing one dis-tribution of centrality measures across firms froman actual alliance network composed of 150 chem-ical firms, we are unable to assess whether such adistribution is typical across all possible structuresof 150-firm networks that can emerge, given whatwe know about the underlying firm, dyadic and net-work level mechanisms of alliance formation, andafter accounting for the possibility of random varia-tion in the network generating process (Holland andLeinhardt, 1970, 1981; Pattison et al., 2000).

Third, the use of endogenous network variables topredict alliance formation further introduces depen-dencies over time among the observed networkvariables. The models in Table 2 predict allianceformation using measures that are derived fromsnapshots of the same network—in other words,independent variable measures such as centrality,are derived from a cumulative set of past alliancechoices. Despite the theoretical rationale that thecurrent structure of a network shapes its futureevolution, the use of such measures in traditionalregression models assumes that this structure isindependent of the prior network structure.

Finally, it is difficult to isolate different theo-retical mechanisms from composite network mea-sures like Combined centrality. Centrality is firstcalculated as a function of the full network foreach individual firm, and when the eigenvector cen-trality score is utilized by researchers, its intentis to proxy for theoretical mechanisms such aspower, visibility, or information control (Ahujaet al., 2009; Gulati and Gargiulo, 1999). Yet, at best,this approach conflates multiple underlying mech-anisms. For example, are higher centrality scoresdriven by preferential attachment, where firms withexisting alliance ties will attract even more ties(e.g., Powell et al., 2005)? Or by transitivity (Gulati,1995)? Likely both, and the practice of combiningthe centrality scores of each firm in the dyad furtherobscures these underlying mechanisms.12

of predicting alliance formation can lead to incorrect inferences.While various methodologies such as clustering of standard errorsor correcting for autocorrelation have been used in prior literature,such statistical methodologies are limited to the firm or to the dyadlevel and cannot handle complex multilevel tie dependencies thatinfluence alliance formation.12Note that APM (in Model 3) also include dummies distinguish-ing low-centrality and socially-asymmetric dyads (where one firm



Underlying strategic alliance formation

process

Observed network measures

Observed strategic alliance ties

-Univariate Dependent Variable

CentralityIndirect tiesNetwork density

Observed strategic alliance ties

-Multivariate Dependent Variable

Underlying strategic alliance formation

processLocal network patterns

Node degree distribution – preferential attachment2-paths – structural equivalenceTriangles - transitivity

Assumed todetermine

Modeled to tie &firm (node)dependencies

Logisticregression

ERGM

Alliances assumed to be independent

Predicting the adjacency matrix –no independence assumption

Network Model Specification

Regression Model Specification

Figure 2. Comparing traditional regression and ERGMs in modeling alliance formation

REPLICATION STAGE II: ISOLATINGMODEL DIFFERENCES

To isolate how modeling issues may affect results,we tested a theoretically comparable allianceformation specification using exponential randomgraph models (ERGMs), a recent advance in socialnetwork methodology that overcomes several ofthese limitations (see Cranmer and Desmarais,2011; Holland and Leinhardt, 1981; Lusher, Koski-nen, and Robins, 2012; Robins et al., 2007; Snijderset al., 2006). Figure 2 illustrates the differencesbetween ERGMs and traditional regression meth-ods. ERGMs13 are parametric models of networks(c.f. Lusher et al., 2012) defined by identifyingrelevant local network structural elements thatreflect underlying tie formation mechanisms.While traditional regression methods imputeendogenous network mechanisms by observingsummary network measures over time, ERGMsclearly specify these endogenous processes thatgenerate the observed network structures. ERGMsalso shift the level of analysis from the dyad to

had high centrality while the other had low). These results, whilesignificant when utilized in place of combined centrality, did notachieve significance when combined centrality was simultane-ously included in the model. Similarly, GG also utilized centralityratio as an explanatory variables, but it was not significant.13Also known as p* models.

the network by predicting the entire adjacencymatrix of alliances as the dependent variable.14

This allows the modeling of dependencies betweenpair-wise combinations in the adjacency matrixas well as dependencies beyond dyads, while alsorelaxing the stringent assumption of cross-dyadicindependence in traditional regression models.Finally, ERGMs are stochastic in nature, treatingthe observed network as one instantiation withina distribution of possible networks generated bythe proposed mechanisms, thus accounting for thepossibility of random variation at the network levelthat traditional methods cannot.

Measuring alliance formation mechanismsin ERGMs

ERGMs allow us to separate the conflated mea-sures used in traditional regression models throughprecise local network structural specifications thateach capture mutually exclusive and collectivelyexhaustive tie formation mechanisms. Our first taskwas therefore to select the specific local networkstructural elements that best correspond to the tra-ditional hypothesized alliance formation mecha-nisms, summarized in Figure 3. We discuss each of

14ERGMs generate a joint prediction for all the n*(n – 1)/2 dyadsfor an n-node network.



Figure 3. Local alliance network structural elements in ERGMs

these elements conceptually here and provide addi-tional mathematical details corresponding to theirestimation in the Appendix S1.

The most basic network component, edges, isanalogous to an intercept term in a traditionalregression, and captures the base propensity of anyalliance tie to form as a function of the count of thealliance ties in the network. The remaining compo-nents unpack positional and structural embedded-ness mechanisms and are nested hierarchically.

For positional embeddedness, APM as well asGG suggest that a firm’s network position allows itto benefit from information about alliance opportu-nities in the network beyond its immediate partnersand also provides a signal of ability and prestige inthe collaboration context. APM expected this effectto yield a diminishing advantage with increasingembeddedness; hypothesizing an inverted U-shapedeffect and measuring it using Combined Centrality,the geometric mean of the eigenvector centralitiesof the two firms in the dyad. Since this approachoverlooks dependencies at different levels and con-flates multiple mechanisms, we instead model thismechanism in ERGMs by fitting the degree distribu-tion local network structure which solely models adegree-based preferential attachment mechanism.15

The network statistic for this structure, called gwde-gree (geometrically weighted degree distribution),is modeled using a probability distribution func-tion derived from the curved exponential family.16

15This is measured by the presence of k-stars, that is a central nodeconnected to k others, in the network—for detailed treatment seeHunter (2007) and Hunter and Handcock (2006).16Equation (A1) in the Appendix provides the mathematical basisof the gwdegree statistic.

Put simply, as a firm’s degree (number of allianceties) increases, this statistic decreases exponen-tially (Hunter, 2007; Hunter and Handcock, 2006).Thus, with a positive and significant coefficient, thelog-odds of an alliance tie increases for all dyadiccombinations, but the increase is of smaller mag-nitude for dyads whose constituent firms alreadyhave higher degrees. By its exclusive focus on thedegree of each firm17 and the specification of a prob-ability distribution function that permits inferenceof the degree based attachment mechanism (Hand-cock, 2003), this approach removes the conflationof mechanisms inherent in the composite joint cen-trality measure.

For structural embeddedness, GG hypothesizea transitivity mechanism where firms are morelikely to enter into an alliance when they sharecommon partners. Note that the measure for thistransitivity mechanism in our probit replication(Table 2, Model 2)—a simple count of Commonalliance partners between two firms—is highlycorrelated with the Combined centrality measurefor positional embeddedness (r= 0.75). We over-come this conflation with ERGMs by using triangleconfigurations18 corresponding to the theoreticalmechanism of transitivity, estimated using gwesp

17A firm’s degree is a network statistic that is likely to be amore relevant local measure that affects alliance formation, than aglobal network level centrality measure. For instance, it is easilyconceivable that firms select other firms as partners if they havemany existing alliances (high degree) or vice versa. It is lessobvious that they carry out eigenvector centrality calculationswhile exercising such choices.18Also referred to as triad closure or triangles.



(geometrically weighted edge shared partners),which fits the distribution of the number of trianglesalso using a curved exponential distribution.

Another limitation of using the traditional countmeasure is that we are unable to capture the influ-ence of nested substructures (e.g., equivalence con-siderations within substructures consisting of two,four, or more firms as depicted in Figure 3). ERGMsallow us to model these nested elements usingshared partner distribution structures19 that are esti-mated using a measure called gwdsp (geometricallyweighted dyad shared partners)—a statistic of thedistribution of shared partner firms by unconnectedfirms. Thus, the simultaneous use of gwesp andgwdsp allows us to make stronger inferences of tran-sitivity (Robins, Pattison, and Wang, 2009).20

Finally, ERGMs also enable the modeling ofnode-level and tie-level attributes that additionallymotivate alliance formation. We modeled relationalembeddedness as a tie attribute by a count of prioralliances for each dyad. We also modeled the inter-dependence mechanisms by creating tie characteris-tics that captured similarity on the firms’ geography,product-market, and technology vector (based onpatent classes) attributes. In contrast, we includedall the control measures (Size, Performance, DebtRatio, Liquidity, and Solvency) from the APM base-line model as nodal attributes. Here, ERGMs differslightly from our probit models as they internallycalculate the effect for these nodal attributes basedon the dyadic sum rather than ratios.

ERGM results

Estimation using ERGMs involves the use of a soft-ware module that supports a corresponding imple-mentation of the network structures. The measurescorresponding to the ERGM local network struc-tural elements as well other firm and dyad levelcovariates are specified using a probability func-tion and estimated through maximum likelihood

19While triangle configurations are structural mechanisms thatcorrespond to indirect ties, shared partner distributions model theidea that structural equivalence and multiple connectivity lead toclustered regions in the network.20The benefit of using these geometrically weighted statistics isthat they allow us to parsimoniously describe the network data byreducing the number of parameters. For example, the degree fittingterm only uses two parameters instead of using (n – 1) parameterswhere “n” is the highest degree observed in the network. Similarlyinstead of fitting multiple triangles the gwesp statistic uses onlytwo parameters.

estimation in R (Handcock et al., 2008).21 Weassessed model fits using the Akaike’s InformationCriterion (AIC) (Akaike, 1998) and log-likelihoodstatistics, and we employed graphical tests of good-ness of fit (Goodreau, Kitts, and Morris, 2009)to visualize the match between the predicted andobserved networks.22

Table 3 reports our ERGM results and comparesthem to our probit replication results from Model2 in Table 2.23 With ERGMs, a positive coefficientindicates the higher likelihood of presence of a localnetwork element than one would expect by chance,conditional on the rest of the network, whereas anegative coefficient indicates a lower probability ofthe presence of the structure than expected (Lusheret al., 2012).

A comparison of our ERGM to the probit repli-cation results demonstrates the robustness of thegeographic (marginal effect of 0.0006 versus 0.004in probit) and product market similarity (marginaleffect of 0.001 versus 0.004 in probit) predictors,and reveals three major differences. First, whereasTechnical similarity was insignificant in the priorcase, it is positive and significant in the ERGM(p-value 0.023, and marginal effect of 0.0007 ver-sus -0.008). Thus, after modeling the structure of the

21The estimation uses the MCMC-MLE procedure in the“ERGM” package, a part of the statnet suite of package for R.Substantively, this approach first involves generating a large num-ber of possible networks that might be observed based on theselocal network conditions, and then asking whether the networkof ties in the focal sample is a likely member of the this family.It is important to reiterate that the generated comparison groupof networks is not deterministic but stochastic. Expressed proba-bilistically, we are ultimately able to determine the probability ofobserving the sampled network, given the input specifications ofdifferent local network elements.22Each plot compares the observed data to one hundred randomlygenerated simulated networks obtained from the fitted models.This provides a visual sense of the model fit in terms of somekey properties of the network such as the degree distribution. Allthree structural network statistics fit a curved exponential familymodel that requires the estimation of the decay parameter “𝛼”.This is achieved through an iterative process of fitting modelsfor different values and choosing the one that provides the lowestAIC value. Our models report the decay parameter for the best fitmodel. Additional models and plots are available on request fromthe authors.23The models shown here are those that provide the best fit (bothin terms of the AIC and log-likelihood and visual goodness of fit),therefore we report our results based on it below. We do not reporta model with the gwdsp term because the ERGM with this termwould not converge which is an indication that such structuresare not prevalent in the network. This finding also supports ourassertion that less localized positional measures obscure ratherthan clarify mechanisms.



Table 3. Comparison of probit and ERGM results for chemical industry, 1991−2000a

ERGMs

Models variables Probit replication (Table 2, Model 2) 1 2

Alliance formation mechanismsPositional embeddednessPreferential attachment (combined centrality or gwdegree) 2.04 −1.33 −0.83

(1.07) (0.21) (0.25)Structural embeddednessTransitivity (common alliance partners or gwesp) 0.10 0.37

(0.10) (0.07)Strategic interdependenceTechnical similarity −0.21 0.55 0.50

(0.35) (0.23) (0.23)Geographic similarity 0.28 0.50 0.46

(0.08) (0.15) (0.15)Product-market similarity 0.33 0.95 0.90

(0.08) (0.14) (0.14)Relational embeddednessPrevious alliances 0.70 0.62 0.47

(0.78) (0.38) (0.35)Dyad level controlsSize 0.36 −0.02 −0.02

(0.14) (0.03) (0.03)Performance 0.43 −0.81 −0.67

(0.63) (1.83) (1.89)Liquidity −0.12 −0.09 −0.06

(0.19) (0.24) (0.24)Debt-equity −0.26 0.33 0.22

(0.14) (0.59) (0.57)R&D 0.10 0.53 0.52

(0.13) (0.24) (0.22)Patents −0.03 −0.03 −0.02

(0.16) (0.02) (0.02)Edges NA 293.50 287.41

(70.13) (68.38)Year Yes Yes YesLog-likelihood −649.89 −937.02 −928.81AIC NA 1900.04 1885.62BIC NA 1998.63 1991.8


network more accurately and accounting for inter-dependencies in our ERGM, we see that similar-ity between two firms in the technology domainincreases their alliance propensity. Second, whilethe Common alliance partners measure for struc-tural embeddedness in our probit replication wasnot significant, the gwesp measure in our ERGMis positive and significant (p-value< 0.00001). Itis important to note that while the other variablesof interest did not substantively change with theinclusion of Common alliance partners in the probitreplication, we did observe a drop in significance forCombined centrality (Table 2—compare Models 1and 2), as these two measures are correlated. Thus,

by using a better specified measure in our ERGM,we find support for GG’s original hypothesis thatstructural embeddedness has a positive effect onalliance formation.24

24 This transitivity effect can be interpreted by considering howthe probability of the alliance changes when a pair of connectedfirms increases its number of shared partners by one, ceterisparibus.

log(pafter/pbefore) = 0.376× (1− e ^ (−1.3))k = 0.37× 0.72 kIn other words, it is easiest to complete a triangle when noneexists (k= 0); such a change leads to an increase of 0.37 on thelog-probability scale beyond the effects predicted by other modelterms. However, this increase diminishes for each unit increase ink. Thus, completing a two-triangle when a triangle already existsonly results in an additional increase of 0.37× 0.72; completing



0

3

6

9

-2

-1

0

1

2

3

4

5

6

(A) (B)

0 1 2 3 4 5 6 7 8 9 100

3

6

9

-2

-1

0

1

2

3

4

5

6

0 1 2 3 4 5 6 7 8 9 10

Figure 4. (A) Plot of gwdegree in chemical and (B) plot of gwdegree in semiconductor. (In the above graphs, the verticalaxis measures the change in the log-odds of an alliance forming between two firms and the two axes on the horizontal

plane measure the degrees [number of ties] of those two firms.)

Third, the coefficient for the positional embed-dedness term gwdegree (preferential attachment) isnegative and significant (p-value< 0.00001). Recallthat in ERGM, a negative coefficient for this termmeans that firms with higher degrees (high num-ber of alliance partners) have a higher likelihood ofentering into further collaborative ties compared tofirms with lower degrees, but the increase in thislikelihood diminishes as degree increases, as dis-played in Figure 4(A).25 In the graph, the verticalaxis is a measure of the change in the log-odds ofa tie forming between two firms “i” and “j” if theirrespective degrees increased from [Di, Dj] to [Di+1,Dj+1]. The axes on the horizontal planes are thedegrees of each of the two firms in the dyad. Whileour probit replication did not obtain an effect forthe squared term of Combined centrality, the con-cave curvilinear effect depicted in the graph doessuggests evidence for the positional embeddedness

a three-triangle when a two triangle already exists only gives0.37× 0.53 and so on. Thus, as two firms that are already in analliance, share more and more partners, the propensity to findadditional shared alliance partners decreases.25The model estimates the decay parameter to be 1.1, andthe coefficient obtained is −0.83. The change statisticsfor two nodes with degrees i and j is (1 – e^(-0.1.1))i+(1–e^(-0.1.1))j= 0.67i+ 0.67j. So with a coefficient of -0.83, thelog-odds of an alliance decrease for all degree values of i and j,but this decrease would be of smaller magnitude when i and jhave higher degree.

mechanism originally posited by APM. As thefigure illustrates, the slope is positive, indicatinghigher odds of alliance formation between the twofirms as their combined degree score increases.

All the other alliance formation mechanismsestimated in our ERGM are comparable to ourprobit replication. We continue to find positiveand significant effects for Geographic Similarity(p-value< 0.0001) and Product-market similarity(p-value 0.000078), and are unable to find sig-nificance for the relational embeddedness (Previ-ous alliances) measure (p-value 0.13373, marginaleffect= 0.23).

REPLICATION STAGE III: ISOLATINGINDUSTRY DIFFERENCES

The semiconductor industry as a contrastingcontext

To isolate and analyze the potential effect of indus-try context on our results, we replicated our ERGManalyses using semiconductor industry data fromthe same time frame. The semiconductor industry,with a distributed locus of technological devel-opment across firms, a modular set of products,and continuous innovation pressures, offers arich contrast to the more mature, process-basedchemical industry that is less susceptible to



rapid price-performance improvements based oninnovation (Rosenkopf and Schilling, 2007). Wefollowed an identical sampling strategy to thechemical replication (1991−2000, largest 150firms, within-industry alliances) from the samedata sources. Our final sample consisted of 321strategic alliances.26

ERGM results in semiconductor

We developed an equivalent ERGM for the semi-conductor industry alliance sample. Table 4 showsthe results of the semiconductor ERGM side by sidewith the results from the chemical ERGM.

Several results are consistent; the effects forstructural embeddedness (Transitivity—gwesp[p-value< 0.00001]), Geographic Similarity(p-value= 0.000126) and Product-Market Similar-ity (p-value< 0.00001) continue to persist acrossthe two industries.27 Yet others differ. Technicalsimilarity, which was significant and positivein chemical, is not significant in semiconductor(p-value= 0.553, marginal effect= 0.02). In con-trast, we do find a positive effect for relationalembeddedness in semiconductor (Prior Ties,p-value= 0.000126, marginal effect= 0.75) whichwas insignificant in chemical. Perhaps the moststriking difference is that the effect for positionalembeddedness is reversed in semiconductor(p-value< 0.00001). The corresponding measurein our semiconductor ERGM is positive andsignificant—recall that a positive coefficient forthis term suggests an “anti-preferential attachment”mechanism which is somewhat comparable to anegative effect for a combined centrality measurein traditional regressions. Thus, this demonstrates atendency for alliance formation in low-degree firms,a mechanism that runs counter to that proposed bythe positional embeddedness theory.28 Figure 4(B)

26The means and standard deviations for the measures in thesemiconductor industry are quite similar to those of the chemicalindustry displayed in Table 1 (additional correlation tables for thesemiconductor industry available on request).27The log-odds of two firms to ally given they operate in the samesectors within the industry increases by 0.70, which is a decreasein the marginal effect from 0.0013 to 0.001. For two firms from thesame country, the log-odds are higher by 0.55, which correspondto an increase in the marginal effect from 0.0006 to 0.0007.28In contrast, our examination of alliance formation in the semi-conductor context using APM’s traditional probit methods failedto yield significant effects for Geographic and Product MarketSimilarity. However, they yielded significant results for positionalembeddedness (Combined centrality) and Technical similarity.

Table 4. Comparison between chemical and semicon-ductor ERGM results, 1991−2000a

Models variables

ERGM forchemical(Table 3,Model 2)

ERGM forsemiconductor

Alliance formation mechanismsPositional embeddednessPreferential attachment (gwdegree) −0.83 2.55

(0.25) (0.60)Structural embeddednessTransitivity (gwesp) 0.37 0.56

(0.07) (0.11)Strategic interdependenceTechnical similarity 0.50 0.06

(0.23) (0.25)Geographic similarity 0.46 0.56

(0.15) (0.15)Product-market similarity 0.90 0.73

(0.14) (0.17)Relational embeddednessPrevious alliances 0.47 0.69

(0.35) (0.25)Dyad level controlsSize −0.02 0.08

(0.03) (0.03)Performance −0.67 −0.36

(1.89) (0.27)Liquidity −0.06 0.14

(0.24) (0.20)Debt-equity 0.22 −0.14

(0.57) (1.01)R&D 0.52 0.11

(0.22) (0.04)Patents −0.02 0.15

(0.02) (0.03)Edges 287.41 369.50

(68.38) (76.72)Year Yes YesLog likelihood −928.81 −841.12AIC 1885.62 1712.24BIC 1991.8 1813.53


shows the graphical plot of this mechanism andrelative effect size. When compared to the corre-sponding plot for the chemical industry ERGM(Figure 4(A)), the slope of the curve is negative andsteeper in the semiconductor ERGM, indicating aneffect for positional embeddedness that is strongerand opposite to what we observed in the chemicalindustry ERGM. Among the control variables,in addition to the strong positive effect for R&D(p-value= 0.0084) (as was the case with chemical),there are also positive effects for Patent Count(p-value< 0.00001) and Size (p-value= 0.0127).



DISCUSSION

Our efforts to replicate the baseline model ofalliance formation from APM’s Strategic Man-agement Journal article, and subsequently, isolateeffects of changes in method and context, demon-strate the robustness of predictors like geographicand product-market similarity, while illustrating thenuances of the remaining predictors arising fromdifferent industry settings, time periods, and empir-ical methods. Table 5 summarizes our results forease of comparison and integration. The rows inTable 5 indicate the mutually exclusive theoreticalmechanisms identified in alliance formation litera-ture, broadly categorized as endogenous structuralnetwork drivers and dyad-specific strategic interde-pendence factors. Each column of Table 5 repre-sents a replication across successive shifts in thedimensions of time, choice of method and industrycontext away from the focal APM study. We firstdiscuss the effects of research context shifts (in timeperiod and industry) and then the effect of shiftingto ERGMs. In each case, we develop implicationsboth for empirical research and for theory.

Implications of shifts in research context

Davis and Marquis (2005) argued that theoreti-cal mechanisms underlying firms’ behavior andrelationships between these mechanisms may bereshaped by shifts within industries as well as tran-sitions in the broader social and economic envi-ronments in which they are embedded. As seen bycomparing Columns A and B (time shift) as well asColumns C and D (industry shift), our replicationsdemonstrated that contextual choices of time periodand industry can affect results.

While certain effects persist across the two timeperiods and two industries we compared (see Rows5 and 6 for geographic and product-market sim-ilarity, as well as Row 3 for structural embed-dedness), others vary with the context in whichresearchers situate their empirical investigation. Inparticular, the effect of technical similarity variesdramatically across our replications (see Row 4).Focusing first on the time shift (Columns A andB), while APM derived an inverted-U effect, ourprobit replication in the later time period did notobtain a significant effect. Focusing next on theindustry shift (Columns C and D), our ERGMs alsoyielded different results—a positive effect of tech-nical similarity in chemical, and a nonsignificant Ta

ble

5.C

ompa

riso

nof

pred

icte

dm

echa

nism

sac

ross

repl

icat

ions

isol

atin

gon

edi

men

sion

ofch

ange

Rep

licat

ion

dim

ensi

onsh

ift=

>th

eore

tical

mec

hani

sms

(A)

Ori

gina

lAPM

prob

itre

sult

(’83

-’91

,Tab

le2)

(B)

Tim

e:’9

1-’0

0(c

hem

ical

,pro

bit,

Tabl

e2)

(C)

Mod

el:E

RG

M(’

91-’

00,c

hem

ical

,Tab

le3)

(D)

Indu

stry

:sem

icon

duct

or(’

91-’

00,E

RG

M,T

able

4)

Net

wor

km

echa

nism

s1.

Posi

tiona

lem

bedd

edne

ssPo

sitiv

e+

curv

iline

arPo

sitiv

e+

curv

iline

arPo

sitiv

e+

curv

iline

arN

egat

ive+

curv

iline

ar2.

Rel

atio

nale

mbe

dded

ness

Posi

tive+

curv

iline

arN

egat

ive+

curv

iline

arN

otsi

gnifi

cant

(p=

0.26

)Po

sitiv

e3.

Stru

ctur

alem

bedd

edne

ssPo

sitiv

ePo

sitiv

eSt

rate

gic

inte

rdep

ende

nce

4.Te

chni

cal

Neg

ativ

e+

curv

iline

arN

otsi

gnifi

cant

(p=

0.55

)Po

sitiv

eN

otsi

gnifi

cant

(p=

0.55

)5.

Geo

grap

hic

Posi

tive

Posi

tive

Posi

tive

Posi

tive

6.Pr

oduc

t-m

arke

tPo

sitiv

ePo

sitiv

ePo

sitiv

ePo

sitiv

e

Cur

vilin

ear

(sec

ond-

orde

r)ef

fect

size

com

para

ble

toA

PM,b

utno

tsta

tistic

ally

sign

ifica

nt(p

-val

ue=

0.16

5).

Onl

yse

cond

-ord

eref

fect

isst

atis

tical

lysi

gnifi

cant

.Fir

st-o

rder

effe

ctis

nots

igni

fican

t(p-

valu

e=

0.33

6).



effect in semiconductor. Each of these contrastsmay be consistent with prior research about indus-try and technology life cycles, in that as technicaluncertainty reduces, firms will migrate toward moreexploitation-based alliances (Lavie and Rosenkopf,2006). In other words, firms in less mature indus-tries where technical uncertainty is higher (chemi-cal in the 1980s relative to chemical in the 1990s;or semiconductor in the 1990s relative to chem-ical in the 1990s) are more likely to seek nov-elty through recombining dissimilar technologies(exploration alliances) rather than through con-solidation within the same technological domain(exploitation alliances).

While these contrasts may be instructive, the sub-stantial discrepancies in the effects for Technologysimilarity in our replications raise other empiricalimplications about variation in the measure acrossindustry contexts. For example, following priorresearch, APM used a continuous measure obtainedthrough the use of Euclidean distance between firmsusing patent classes. This measure carries the strongimplicit assumption that all classes are equidistanteven though distance between classes likely variesboth between and within industry.29 Therefore,future research should utilize measures that accountfor the nature of the technology classes within agiven industry. Empirical techniques that conceptu-alize the technology space using citation networksand then apply clustering algorithms to identifysimilar technologies may be useful to obtain betterfit between the concept and measure going forward(Sytch, Tatarynowicz, and Gulati, 2012).

We also note differences in results across timeand industry for the positional and relationalembeddedness predictors (see Rows 1 and 2). Sincethe use of ERGMs fundamentally changes ourmeasures, theorizing, and interpretations, in thissection we will only examine the time period shiftin chemical (probit analyses) and leave interpre-tation of the industry shift (ERGM analyses) forthe subsequent section. Specifically, for positionalembeddedness, while APM obtained a positive,curvilinear result for combined centrality for theirdata, only the main effect obtained significance

29For example, a patent for logic circuitry (e.g., class 326: Elec-tronic digital logic circuitry) will be much more dissimilar (andconceptually more distant) to a patent in materials (e.g., class505: Superconductor technology: apparatus, material, process),than to a patent for photolithography steps (e.g., class 716:Computer-aided design and analysis of circuits and semiconduc-tor masks).

in the chemical industry in our later time perioddata (see Row 1, Columns A and B). In addition,relational embeddedness no longer predicts allianceformation in our data (Row 2, Columns A and B).

A plausible explanation for these shifts is theevolution of the chemical industry which focused,globalized and consolidated during this time period.Some of the biggest restructuring in the chemi-cal industry since the 1920s occurred during the1990s (Alperowicz, 2014). Industry concentrationincreased during the 1991−2000 period with 2,866mergers, divestitures, and asset sales—in compari-son there were only 931 such corporate events in theprior decade.30 These transitions may have shiftedthe drivers of alliance formation. The need to tapinternational markets may have spurred allianceswith new local firms,31 and consequently, the impor-tance of trust (as evidenced through repeated ties)in renewing existing partnerships may have dimin-ished. Increases in ties for leading firms coupledwith consolidation would lessen the diminishingimpact of centrality originally observed by APM.Consolidation may also have resulted in the acqui-sitions of existing alliance partners (e.g., Yang, Lin,and Peng, 2011), thus making relational embed-dedness a predictor of future acquisitions ratherthan alliances. Finally, in a maturing industry, firmsmay have already accrued sufficient information onthe performance of partners from previous partner-ships, thus making the mere fact of having had aprevious alliance a poor predictor of future partner-ships (e.g., Holloway and Parmigiani, 2014).32

All of these findings reinforce the need to eluci-date scope conditions for findings generated in anyresearch context. Researchers should be explicitabout comparing their models and contexts againstpredecessor papers, seeking to isolate changesto one issue and contrasting results. Through

30Source: SDC Platinum, where both acquirer and target firm arein the chemical industry.31This was also the period during which several Asian chemicalfirms from India and China came into prominence and investmentopportunities opened up in Eastern Europe.32Our thesis about industry shifts in this decade in the chemicalindustry was further substantiated when we re-ran our probitregressions dropping successive years from the beginning of thesample (i.e., 1992−2000, 1993−2000… )—we lose effects forvariables such as Combined Centrality immediately. In contrast,when we ran the regressions dropping years from the end of theperiod (i.e., 1991−1999, 1991−1998… ), we retain effects tillour sample shrinks by ∼50 percent. Thus, the industry shift andevolution of alliance drivers likely became more pronounced astime progressed.



incremental, staged variations, such as we havedemonstrated here, researchers can investigate anddevelop mid-range theories with contingenciesrooted in underlying characteristics, such as uncer-tainty and concentration, of particular industries atparticular times.33

Implications of model choice

The use of ERGMs, a methodological advance notavailable earlier to network researchers, to addressshortcomings inherent in dyadic regression analy-ses also enabled us to disentangle traditional net-work embeddedness measures while allowing forbetter specification of network formation mech-anisms. While combined centrality has becomea well-established measure to connote positionalembeddedness in our literature, we have argued thatit conflates different mechanisms such as transitiv-ity, preferential attachment, and homophily. In con-trast, ERGMs more clearly identify and separate thepositional and structural characteristics by model-ing preferential attachment (represented by gwde-gree) and transitivity mechanisms (represented bythe nested local network structure componentsgwesp, and gwdsp) while simultaneously modelingother mechanisms such as interdependence.

Empirically, we observed the effect of shift-ing from a probit model to an ERGM when ouranalyses are isolated to the 1991−2000 chemicalindustry and we vary the modeling technique (com-pare Table 5, Columns B and C). For networkembeddedness predictors, the full specification ofall three measures in our ERGM generates a pos-itive result for structural embeddedness (Row 3)as well as a positive curvilinear result for posi-tional embeddedness (Row 1). Thus, separatingthe traditional Combined centrality construct intoits constituent parts allows for more precise test-ing of specific mechanisms. We demonstrate thatthat the structural mechanisms spurring allianceformation are nodal (degree/visibility) and trian-gular (transitivity/common ties) in the 1991−2000chemical industry. Broader network constructs likegwdsp do not predict alliance formation. This out-come is consonant with recent work suggesting that

33While the practice of pooling multiple industries to obtaingeneralizability is common, we caution against this approach. Inpost-hoc analyses, we found that pooling our data across bothindustries obscured context-specific findings (results availablefrom authors on request).

arguments premised on multistep information flowsbetween organizations are often unrealistic (Ghoshand Rosenkopf, 2015).

Our most important theoretical implicationsfocus on positional and relational embeddedness.Using the better-specified ERGMs allows us toseparate these mechanisms effectively, which varywith industry context (Table 5, Rows 1 and 2,Columns C and D). For positional embedded-ness, alliance formation for firms in the chemicalindustry exhibits preferential attachment, as thecoefficient of gwdegree is negative; in contrast,the coefficient is positive for the semiconductorindustry, suggesting that low degree firms havehigher alliance propensities.34

How can we explain these divergent effects whena cursory examination of the alliance networks inboth industries both present comparable centralitydistributions and visible hubs? We must simulta-neously consider the differing effects of relationalembeddedness, which is positive in semiconductorbut not significant in chemical. Semiconductorfirms that have low degrees (existing number ofalliance partners) have a higher propensity to allywhile in chemical high-degree firms capitalizeon their positional advantage. Taken together,semiconductor firms reproduce preexisting rela-tionships while seeking novel partners, whereaschemical firms seek partnerships with the mostwell-established firms with less regard for priorrelationships.

These differing industry results likely stem fromvariation in alliance formation drivers across indus-tries. The semiconductor industry, characterizedby rapid technological change and distribution oftechnological knowledge, manifests higher techno-logical uncertainty and lower concentration. Here,entrepreneurial (read low degree) firms may be ableto forge alliances based on the potential value oftheir technological capabilities, such as the designcapabilities of fabless semiconductor firms thatcame to the fore during our study’s time frame. Even

34In additional analyses (available on request), we also comparedERGM results for the semiconductor industry with probit regres-sions in the same industry (i.e., varying only the method whileholding the context and the time period constant). While effectsfor Geographic similarity and Previous alliances were significantand consistent, and that for Technical similarity was insignificantand consistent between the two models, effects for the other con-structs differed dramatically. This further underscores the risks ofrelying solely on traditional regression methods to test allianceformation mechanisms.



highly embedded incumbent firms in such con-texts may frequently need to partner with periph-eral or unconnected firms to tap into technologiesthat could become disruptive or to establish stan-dards of products and processes they develop. Incontrast, the chemical setting, characterized by con-solidation and global market expansion, manifestslower technological uncertainty and higher concen-tration, favoring well-established (read high degree)firms.

Of course, while we have identified the benefitsof using ERGM techniques, such an approach isnot without its limitations. For instance, ERGM isknown to be unstable to missing data and modelmisspecification issues (commonly referred to as“model degeneracy” problems). ERGM statisticalpackages are also not widely available acrossplatforms (to the best of our knowledge, “R” is theonly statistical programming toolkit that providesreasonably documented ERGM routines), and areundergoing frequent changes, thus requiring ERGMusers to invest in additional learning. Finally, whileERGM provides a robust method to estimatewhether specific tie generation mechanisms under-lie an observed network at a point in time, modelingevolutionary processes where these mechanismsmay shift over time is less straightforward. How-ever, in balance, as we demonstrate in this article,the benefits of ERGM outweigh both its limitationsand those of traditional network modeling methods.

CONCLUSION

Our effort to replicate the baseline analyses ofAhuja et al. (2009) explored the impact of shifts intime period, modeling approach, and industry con-text on the well-established predictors of allianceformation in the strategy literature. As a result,we identified two critical implications for futureresearch in this domain. First, and perhaps notunexpectedly, context does matter; any choice ofindustry and time period represents a particularstage of an industry lifecycle, and the findings arelikely contingent on underlying characteristics ofthe industry’s technology and organization. Futureresearch must acknowledge these issues and seekcomparability across studies to assess contingencieseffectively.

Second, our use of the newer ERGM technique,not available earlier to network researchers, enabledus to unpack traditional network measures and

demonstrate how to improve operationalizations ofthe network embeddedness constructs. While theoriginal explication of positional, structural, andrelational embeddedness set forth the combinedcentrality measure as the appropriate representationof positional embeddedness, our analyses demon-strated that combined centrality conflates multiplemechanisms. The mutually exclusive and collec-tively exhaustive predictors available in ERGMsallowed us to use standard predictors of priorties, common ties, and degree distribution withoutviolating the assumption of independence acrossdata and measures in traditional dyadic regressions.Such an approach, by retaining multiple levels ofanalysis in the network regression rather than con-solidating all measures to the dyadic level, allowedus to identify divergent mechanisms of alliance for-mation across contexts.

Accordingly, we argue that the ERGM approachshould become an essential part of the standardtoolkit for future research on alliance formation,and that alliance researchers should seek to repli-cate extant findings using this method. In so doing,studies become more comparable not just via tri-angulation across a variety of industries and timeperiods, but also because the design of ERGMsavoids customized and equivocal composite net-work measures. Such comparability benefits ourfield by enabling the development of mid-range the-ory about underlying constructs such as uncertaintyand concentration, which shape alliance networkstructure across a wide swath of industry and tem-poral contexts.

REFERENCES

Ahuja G. 2000. The duality of collaboration: inducementsand opportunities in the formation of interfirm linkages.Strategic Management Journal 21(3): 317–343.

Ahuja G, Polidoro F, Mitchell W. 2009. Structuralhomophily or social asymmetry? The formationof alliances by poorly embedded firms. StrategicManagement Journal 30(9): 941–958.

Ahuja G, Soda G, Zaheer A. 2012. The genesis anddynamics of organizational networks. OrganizationScience 23(2): 434–448.

Akaike H. 1998. Information theory and an extension ofthe maximum likelihood principle. In Selected Papersof Hirotugu Akaike (pp. 199–213). Springer New York.

Alperowicz N. 2014. 1990s: industry restructures, drawsapart from oil and pharma. Available at: http://www.chemweek.com/lab/1990s-Industry-restructures-draws-apart-from-oil-and-pharma_63787.html 2015)(accessed 05 Sep 2014).



Chung SA, Singh H, Lee K. 2000. Complementarity,status similarity and social capital as drivers of allianceformation. Strategic Management Journal 21(1): 1–22.

Cranmer SJ, Desmarais BA. 2011. Inferential networkanalysis with exponential random graph models. Polit-ical Analysis 19(1): 66–86.

Davis GF, Marquis C. 2005. Prospects for organizationtheory in the early twenty-first century: institutionalfields and mechanisms. Organization Science 16(4):332–343.

Garcia-Pont C, Nohria N. 2002. Local versus globalmimetism: the dynamics of alliance formation in theautomobile industry. Strategic Management Journal23(4): 307–321.

Ghosh A, Rosenkopf L. 2015. PERSPECTIVE—shroudedin structure: challenges and opportunities for afriction-based view of network research. OrganizationScience 26(2): 622–631.

Gimeno J. 2004. Competition within and between net-works: the contingent effect of competitive embedded-ness on alliance formation. Academy of ManagementJournal 47(6): 820–842.

Goodreau SM, Kitts JA, Morris M. 2009. Birds of afeather, or friend of a friend? Using exponential randomgraph models to investigate adolescent social networks.Demography 46(1): 103–125.

Gulati R. 1995. Social structure and alliance formation pat-terns: a longitudinal analysis. Administrative ScienceQuarterly 40(4): 619–652.

Gulati R. 1999. Network location and learning: theinfluence of network resources and firm capabilitieson alliance formation. Strategic Management Journal20(5): 397–420.

Gulati R, Gargiulo M. 1999. Where do interorganizationalnetworks come from? American Journal of Sociology104(5): 1439–1493.

Hall BH, Jaffe AB, Trajtenberg M. 2001. The NBER patentcitation data file: lessons, insights and methodologicaltools. National Bureau of Economic Research. No.w8498.

Handcock MS. 2003. Statistical models for social net-works: inference and degeneracy. Dynamic Social Net-work Modeling Analysis 126: 229–252.

Handcock MS, Hunter DR, Butts CT, Goodreau SM, Mor-ris M. 2008. Statnet: software tools for the representa-tion, visualization, analysis and simulation of networkdata. Journal of Statistical Software 24(1): 1548.

Holland PW, Leinhardt S. 1970. A method for detectingstructure in sociometric data. American Journal ofSociology 76(3): 492–513.

Holland PW, Leinhardt S. 1981. An exponential family ofprobability distributions for directed graphs. Journal ofthe American Statistical Association 76(373): 33–50.

Holloway S, Parmigiani A. 2014. Friends and profitsdon’t mix: the performance implications of repeatedpartnerships. Academy of Management Journal 59(2):460–478.

Hunter DR. 2007. Curved exponential family models forsocial networks. Social Networks 29(2): 216–230.

Hunter DR, Handcock MS. 2006. Inference in curvedexponential family models for networks. Journal

of Computational and Graphical Statistics 15(3):565–583.

Lavie D, Rosenkopf L. 2006. Balancing exploration andexploitation in alliance formation. Academy of Manage-ment Journal 49(6): 797–818.

Li SX, Rowley TJ. 2002. Inertia and evaluation mecha-nisms in interorganizational partner selection: syndi-cate formation among US investment banks. Academyof Management Journal 45(6): 1104–1119.

Lusher D, Koskinen J, Robins G. 2012. Exponential Ran-dom Graph Models for Social Networks: Theory, Meth-ods, and Applications. Cambridge University Press.

Madhavan R, Koka BR, Prescott JE. 1998. Networks intransition: how industry events (re) shape interfirmrelationships. Strategic Management Journal 19(5):439–459.

Pattison P, Wasserman S, Robins G, Kanfer AM. 2000.Statistical evaluation of algebraic constraints for socialnetworks. Journal of Mathematical Psychology 44(4):536–568.

Powell WW, White DR, Koput KW, Owen-Smith J. 2005.Network dynamics and field evolution: the growth ofinterorganizational collaboration in the life sciences.American Journal of Sociology 110(4): 1132–1205.

Robins G, Pattison P, Kalish Y, Lusher D. 2007. Anintroduction to exponential random graph (p*) modelsfor social networks. Social Networks 29(2): 173–191.

Robins G, Pattison P, Wang P. 2009. Closure, connectiv-ity and degree distributions: exponential random graph(p*) models for directed social networks. Social Net-works 31(2): 105–117.

Rosenkopf L, Metiu A, George VP. 2001. From the bottomup? Technical committee activity and alliance forma-tion. Administrative Science Quarterly 46(4): 748–772.

Rosenkopf L, Schilling M. 2007. Comparing alliancenetwork structure across industries: observations andexplanations. Strategic Entrepreneurship Journal1(3–4): 191–209.

Rothaermel FT, Boeker W. 2008. Old technology meetsnew technology: complementarities, similarities, andalliance formation. Strategic Management Journal29(1): 47–77.

Rowley T, Behrens D, Krackhardt D. 2000. Redundantgovernance structures: an analysis of structural andrelational embeddedness in the steel and semiconduc-tor industries. Strategic Management Journal 21(3):369–386.

Schilling MA. 2009. Understanding the alliance data.Strategic Management Journal 30(3): 233–260.

Snijders TA, Pattison PE, Robins GL, Handcock MS.2006. New specifications for exponential random graphmodels. Sociological Methodology 36(1): 99–153.

Stuart TE. 1998. Network positions and propensities to col-laborate: an investigation of strategic alliance formationin a high-technology industry. Administrative ScienceQuarterly 43(3): 668–698.

Stuart TE, Sorenson O. 2007. Strategic networks andentrepreneurial ventures. Strategic EntrepreneurshipJournal 1(3–4): 211–227.

Sytch M, Tatarynowicz A, Gulati R. 2012. Toward a theoryof extended contact: the incentives and opportunities



for bridging across network communities. OrganizationScience 23(6): 1658–1681.

Wang L, Zajac EJ. 2007. Alliance or acquisition? Adyadic perspective on interfirm resource combinations.Strategic Management Journal 28(13): 1291–1317.

Yang H, Lin ZJ, Peng MW. 2011. Behind acquisitionsof alliance partners: exploratory learning and networkembeddedness. Academy of Management Journal54(5): 1069–1080.

SUPPORTING INFORMATION

Additional supporting information may be foundin the online version of this article:

Appendix S1. ERGM statistics, specification andestimation.


The impact of context and model choice on the determinants ... · exponential random graph models (ERGMs), a recent methodological advance that allows the explicit modeling of underlying

Documents