Anda di halaman 1dari 52

Social Structure from Multiple Networks. I.

Blockmodels of Roles and Positions


Author(s): Harrison C. White, Scott A. Boorman, Ronald L. Breiger
Source: The American Journal of Sociology, Vol. 81, No. 4 (Jan., 1976), pp. 730-780
Published by: The University of Chicago Press
Stable URL: http://www.jstor.org/stable/2777596 .
Accessed: 02/03/2011 09:43

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=ucpress. .

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to The
American Journal of Sociology.

http://www.jstor.org
Social Structure from Multiple Networks.
I. Blockmodels of Roles and Positions'
HarrisonC. White
Harvard University

Scott A. Boorman
Universityof Pennsylvania

Ronald L. Breiger
Harvard University

Networksof severaldistincttypesof social tie are aggregatedby


a dual modelthatpartitionsa populationwhilesimultaneously identi-
fyingpatternsof relations.Conceptsand algorithmsare demon-
stratedin fivecase studiesinvolvingup to 100 personsand up to
eighttypesof tie,overas manyas 15 timeperiods.In each case the
modelidentifiesa concretesocial structure.Role and positioncon-
cepts are then identifiedand interpretedin termsof these new
modelsof concretesocial structure.
Part II, to be publishedin the
May issue of this Journal (Boormanand White 1976), will show
how the operationalmeaningof role structures in smallpopulations
can be generatedfromthesociometric blockmodels of Part I.

During the past decade, the networkmetaphorhas becomeincreasingly


popular with social scientists;2 it has even penetratedthe conservative

1 Support from the National Science Foundation under grant GS-2689 is gratefully
acknowledged.In addition to Phipps Arabie, GregoryH. Heil, Paul R. Levitt, and
Francois Lorrain (who have coauthored related papers with us), Paul Bernard and
Joseph E. Schwartz had substantial,specificimpact on the work. The generosityof
Belver C. Griffith, Nicholas C. Mullins, and S. Frank Sampson in supplying and
interpretingdata is deeply appreciated,as were A. P. M. Coxon's detailed comments
on earlier drafts.The editorialadvice of Carolyn J. Mullins led to notable improve-
ments in the exposition.Thanks are due the Mathematical Social Science Board for
supportingtwo small conferenceson models of role networks,at which early versions
of this work were discussed. Access to computerfacilitieswas kindly given by the
CambridgeProject and its director,Dr. Douwe Yntema. The senior author wrote a
draft of this paper while holding a GuggenheimFellowship.
2 Network metaphorsdate back at least to Simmel (1950, 1955; firstpublished in
1908) and the so-called formalschool of German sociologists.Simmel emphasized the
ubiquity of social networksbased on "the actual similarityof [individuals'] talents,
inclinations,activities,and so on" (1955, p. 128) and which cross-cutthe categorical
attributesof persons. Von Wiese, stronglyinfluencedby Simmel, stressedthe multi-
plicity of types of social ties and the analytic desirability of reducing network
structures.If the "constantlyflowingstream of interhumanactivity" were halted in
its course for one moment,von Wiese (1941, pp. 29-30) suggested,we would observe

730 AJS Volume81 Number4


Social StructurefromMultipleNetworks.I
precinctsof economics(Boorman 1975; Marschak and Radner 1972;
Schelling1971; see also Leijonhufvud1968). Sociologists'and anthro-
pologists'attemptsto developthemetaphorintooperationalconceptshave
taken two directions.One has emphasizedthe paths or "threads"in a
singlenetwork:the mannerin whichlong chains of contactwind their
way through largesocialsystems(Milgram1967; Pool and Kochen 1958;
Rapoport 1963; Coleman 1964; Hunter and Shotland 1974; White
1970a, 1970b; Lee 1969; Granovetter1973, 1974). The secondhas em-
phasizedthe "knittedness" of interconnectionswithina networkand the
overlaps between multiple (many-stranded)types of networksfor a
given population(typicallysmall; see TheoreticalBackgroundsection,
below). Our operationalconceptsfollow the second traditionbut are
consistentwith the first.
Afterdemonstrating the utilityof theseconceptsas appliedto fivecase
studies,we redefinethe classicconceptsof role and positionso thatthey
apply to concrete,observableinteractions, orderedby a new framework.
We take as given the incidenceof each of several distincttypesof tie
across all pairs in a population(see for examplefigs. 1 and 3 below).
Ties of each giventypeare treatedas a separateentity(a matrix).Each
is a separatenetworkto be contrastedwithothersuch networks,rather
than mergedwith themto forma complexbond betweeneach pair of
actors.This analyticsegregation of networktypesis basic to our frame-
work. From it, aggregationemergesas a concept with dual aspects:
actors are partitionedinto structurally equivalentsets withineach net-
work; simultaneously, though,networksare mappedinto a set of images
that can be specificallyinterpretedforspecificpopulations.The resulting
"blockmodel"is a viewof social structure obtaineddirectlyfromaggrega-
tion of the relationaldata withoutimposingany a prioricategoriesor
attributesfor actors. Our fundamentalargumentis that the enormous
varietyof concretesocial structures is reflected
in the varietyof possible
blockmodels;furthermore, blockmodelsprovide tools for orderingthis
diversity.
The essentialphenomenon portrayedin networkimagery,we argue,is
the absenceof connections betweennamedindividuals.The logical sym-
metrybetweenties that are "present"and ties that are "absent" (i.e.,
all others) has encouragedproponentsof graph theoryto overlookthe

"an apparentlyimpenetrablenetworkof lines between men. There is not only a line


connectingA with B, and B with C, etc., but C is directlyconnected with A, and,
moreover,A, B, and C are enclosed within a circle. Not only is there one line con-
nectingA with B, and not only one circlein which they are both enclosed,but there
are many connectinglines. . . . A static analysis of the sphere of the interhuman
will . . . consistin the dismemberment
and reconstruction of this systemof relations.
Outside this network,above and below it, there can be nothingthat is social, unless
we leave the plane of empirical observation."

731
AmericanJournalof Sociology
social asymmetry that existsbetweensocial action and its complement
(Harary,Norman,and Cartwright 1965; cf. Simmel1950,pp. 311-16).
This paperand its forthcomingcompanion, Part II, presentno modelsof
processesover time; thereare neitherpredictionsof otherbehaviornor
explicationsof a stochasticprocessof tie formation and dissolutionthat
would sustainan observedblockmodel.In this paper the argumentsfor
a blockmodelas a pictureof social structureare specificto the contextof,
and the data available for,each case study.4Yet blockmodelsprovidea
natural frameworkfor discussingvarious types of structuralchange:
numerouschangesin individualties can still be consistentwith an un-
changedstructural pattern;changesin the "circulation"of actorsamong
thestructurally equivalentsetscan stillreflectthe same structuralpattern
fora givennetwork,and changesin networkpatternscan occurand yet
leave sets of actorsunchanged.
The next sectionof this paper examinesthe broad theoreticalunder-
pinningsof our research.The secondmajorsectionpresentsdefinitions and
themethodsof analysis.The thirdsectionexhibitsanalysesbased on five
case studies.The fourthsectionprovidesan interpretation of "role" and
"position."

THEORETICAL BACKGROUND

Insightfulexpositionsof recentworkon networkinterrelations are those


by Mitchell(1969, chap. 1) and Barnes (1972). While we use themas
centralreferences, we want to state one fundamental disagreement. Both
see networkanalysisto date as, at best, an eclecticbag of techniques
(Barnes 1972, p. 3) for studyingthe details of individuals'variability
around some basic orderingby categoriesand concreteorganizations
(Mitchell 1969, p. 10). We would like the reader to entertaininstead
the idea that the presentlyexisting,largelycategoricaldescriptionsof
social structure
have no solid theoretical
grounding;furthermore, network
conceptsmay providethe onlyway to constructa theoryof social struc-
ture.
Perhapsthe major thrustof classical social theorywas its recognition
of the historicaldissolutionof categoricalboundariesforsocial relations,
whetherthe changewas perceivedas a transitionfromstatusto contract
(Maine), fromGemeinschaft to Gesellschaft(T6nnies), frommechanical

3 Recognizingthat the "holes" in a networkmay defineits structurewas a primary


substantivemotivationfor the work reportedhere. There are obvious analogies with
homologytheoryin algebra (Hilton and Wylie 1960), thoughthe relevantmathematics
is quite different.
4 In addition, White (1974b) has calculated probabilitiesfor the occurrencepurely
by chance of the simplestblockmodels.

732
Social StructurefromMultipleNetworks.I
to organicsolidarity(Durkheim),fromtraditional to means-rational orien-
tation (Weber), or fromascribedto achieved status (Linton). In our
view, the major problemwith postclassicalsocial theoryhas been that
its conceptsremainweddedto categoricalimagery.All sociologists'dis-
courserestson primitive terms-"status,""role,""group,""social control,"
"interaction,' and "society"do notbeginto exhaustthelist-whichrequire
an aggregation principlein that theirreferents
are aggregatesof persons,
collectivities,
interrelated "positions,"or "generalizedactors." However,
sociologistshave been largelycontentto aggregatein only two ways:
eitherby positingcategoricalaggregates(e.g., "functionalsubsystems,"
"classes") whoserelationto concretesocial structurehas been tenuous;or
by cross-tabulating individualsaccordingto theirattributes(e.g., lower-
middle-classwhite Protestantswho live in inner city areas and vote
Democrat). Both methodshave "oftenled to the neglectof social struc-
tureand of the relationsamongindividuals"(Coleman 1958).5 In con-
trastto the standardwisdom,thereis a growinglist of empiricalfindings
regardingthe effect(and frequency)of "accidents"and "luck" in the
actual functioning of societies: the transmissionof useful information
among scientists(Menzel 1962), the attainmentof general economic
success (Jenckset al. 1972), and the locationof desirablejobs (Grano-
vetter1974; see also Boorman 1975). These findingsforceus to ask
whetherthe stuffof social actionis, in fact,waitingto be discoveredin
the networkof interstices thatexistoutsidethe normativeconstructs and
the attributebreakdownsof our everydaycategories.

Overall Social Structure


Nadel's The Theoryof Social Structure(1957), one of the fewpieces of
sustainedanalyticalexegesisin sociology,inspiredthe work(White 1963;
Lorrainand White 1971) fromwhichthesepapers grew.His focuswas
the interrelations of roles.In dealingwithrole "frames"and theirinter-
lock,6he confronted theinteractionof culturalsystemsand concretesocial
structure,a topicon whichwe spendlittletime.However,we do develop,
in a limitedcontext,two of Nadel's mostimportantideas. First, social
structureis regularitiesin the patternsof relationsamong concrete
entities;it is not a harmonyamongabstractnormsand values or a classi-

5 There are some exceptionsto these tendencies,e.g., reference-group theory (Merton


1959, pp. 281-86), and Znaniecki's (1940) embedding of "role" concepts in "social
circles"; nevertheless,there is a remarkablelack of attention to aggregationas a
central problem for sociological theory. Leijonhufvud's (1968, chap. 3) critique of
neoclassical economics for avoiding similar questions is relevanthere. See also Green
(1964) for a more orthodox review of economic aggregationconcepts.
6 This topic, of course, entails the attendantcomplexitiesof interrelating
the multiple
perspectivesof actors in actual societies.

733
AmericanJournalof Sociology
ficationof concreteentitiesby theirattributes. Second,to describesocial
structure, we mustaggregatetheseregularities in a fashionconsistentwith
theirinherentnatureas networks.
The culturaland social-psychological meaningsof actual ties are largely
bypassedin the development. We focusinsteadon interpreting the pat-
ternsamongtypesof tie foundin blockmodels.Our sole assumptionhere
is that all ties of a given observedtype share a commonsignification
(whatever their contentmay be). From these patterns,we develop
below (and in Part II) operationalconceptsof role and position.7
In our view "position,"in the concretesense of officein a formal
organization or membership on a committee, is a conceptquiteindependent
from"role." The blockmodelsof thispaper can be said to identifyposi-
tions,but onlyin an elementary sense.In Part II we extendthe analysis
to encompassmultipleegos and, thence,role structures; we hope also
thatthisextension can describe,in thelanguageof Mitchell(1969, pp. 45-
49), the existenceand interrelations of "institutions."
At best,blockmodelscan makeonlya partialcontribution to the analy-
sis of formalorganizations as structures of offices. The networkmetaphor
is unavoidablein developingmodelsof formalorganization, even of the
simplestkind (Williamson1970, chap. 2). However, fundamentally new
developmentsof the metaphorare needed, such as that proposed by
Friedell (1967) and that impliedby the argumentof Cohen and March
(1974).
Analyzingsystemsof formalorganizationswill require still further
developmentsof networkimagery,and these cannot be divorcedfrom
modelsof elites and the ways in which they may controllarge social
systemsthroughthe structure of networkaccess. Recentworkon director
interlocks(e.g., Levine 1972) and on advisorysystems(Mullins 1972),
as well as formally analogousmodelsof interlocks in dual individual-posi-
tion systems(Breiger 1974b; Bonacich 1972), may point in the right
direction.
One practicalreasonfor the cautionof Mitchelland Barnes in using
networkconceptswas the lack of satisfactory methodsfor aggregating
networks amongindividuals.A relatedreasonwas the paucityof research
on networks amongnodesthatrepresented and organizations.
collectivities
Thereare fewsystematic analysesof networks amongsuchnodes (but see
Fortes 1945, Mayer 1960, and Savage and Deutsch 1960); however,
engineering and operationsresearchhave huge descriptiveand normative
7 Our stresson relationships
among patternssuggestedto one of our refereesan analogy
to Levi-Strauss'swork on "meaning."He therebycreditedus with too much and too
little.We use that termwithoutthe richethnographicinsightof Levi-Strauss; however,
we do discuss the falsifiabilityof an ideal-type pattern and (White 1974b) its null
expectation. In our view, the delineation of concrete social structure should be
analyticallydivorced from symbolic and cultural analysis.

734
Social StructurefromMultipleNetworks.I
literatureson flowswithinnetworks(see Ford and Fulkerson1962) that
may provesuggestive(see White 1973).8
Both Mitchelland Barnesemphasized"anchored"networks(networks
seenfromtheperspective of a particularmember),becausetheywantedto
showhow networkconceptsilluminatethe manipulativeactivitiesof con-
crete personsin real situations.Their conceptualapproachdiffersfrom
our observerviewpointin this paper and the multiple-ego viewpointof
Part II. In particular,they mergeddifferent types of tie and inferred
a complex,overallquality fromthe multiplexbond betweenthe anchor
personand each of his contacts.'3 In contrast,we argue for the value
(fromthe observer'sviewpoint)of treatingthe networkbased on each
analyticallyseparable type of tie as a separate entity.Furthermore,
Mitchelland Barnespaid moreattentionto thedifferent facetsof meaning
measuredforeach typeof tie; theyalso stressedthe importanceof rich
observation by a participantobserveras, forexample,in Kapferer'swork
we argue,following
(1972).10 In contrast, Durkheim,thata theoryshould
be developedonlyin termsof the overallstructure thatis the contextfor
particulartransactions.We cite as evidenceBoorman's(1975) modelof
job information exchange,notingthat his resultsregardingstabilityand
optimality wereobtainedfrompostulatesof a verysimple,overallnetwork
structure."1
Mitchell and Barnes treated sociometry,especially balance theory,
withsomedisdain.12Althoughmanypowerfulanalysesof data have used
a varietyof sociometricconcepts(see, e.g., the excellentsurveyarticles
by Glanzerand Glaser [1959, 1961]), manyof the data are fromexperi-
mental"groups"and otherpopulationsaggregatedwithina sociological
vacuum.Moreover,withthe crucialexceptionof the analysesby Davis,
Holland, and Leinhardt(see Part II), balance theoristshave had little

8Both this paper and Part II deal wholly with data on individuals,but our motive
for developingthe methodsreportedhere was partlythat we thinkthey will be fruit-
ful for analyzing data on networksamong collectivities(see Breiger,Boorman, and
Arabie [19751 reanalyzingdata of Levine [19721).
9 But note the lament of Mitchell's colleague Boissevain (1973, p. xi) that "the
problem of handling multiplexor many-strandedrelationshipsremains,in spite of
the increasinglysophisticatedanalytical apparatus provided by network analysis."
10 Kapferer went even further,attempting to develop an exchange theory for
transactionsbetween pairs in a network.
11 We believe that blockmodels,which representstatic structure,will be a useful
frameworkfor developing social exchange theory. Ekeh's (1974) recent review of
"the two traditions"in social exchange theory urged the importanceof interaction
between restrictedexchange (Homans) and generalized exchange (Levi-Strauss).
Blockmodelsseem a natural contextfor such a merger.
12Mitchell (1969, p. 7) venturedso far as to termbalance theory,the most interest-
ing analytic developmentin this tradition (see e.g., Harary et al. [1965, chap. 131),
''a toy of the lecture-roomtheoretician."

735
AmericanJournalof Sociology
stomachforactual data. Yet we thinkMitchelland Barnes too hasty.
Our own approachowes much to sociometry, particularlyits encourage-
mentof systematicdata reportsin contrastto richintuitiveobservation,
and blockmodelsincludevariousformsof balance theoryas special cases
(see Part II).

ContrastswithSociometry
Sociometry's mostcommongoal fora singletypeof tie was the identifica-
tionof cliques (or similarconfigurations) of tightlyclusteredindividuals;
a secondarygoal was chainsof connectivity. The clique conceptembodies
the rootidea of aggregation by relationsratherthanby attributesthatis
indispensableto blockmodels.Sociometry'sothermajor goal (most not-
able in balancetheory)has been to interpret theinterpenetration,or over-
lap, amongdifferent typesof tie.
We now draw fivecontrastsbetweensociometry and blockmodels. The
firsttwo are promptedby the restrictive natureof the clique concept.
First,personsnot in cliques are usuallydisregarded(i.e., treatedas out-
side the effective
sociometric system).In contrast,blockmodeling requires
searchingfora completepartition, such thatsets of personscan be struc-
turallyimportant regardlessof whetherthe sets resemblecliques.Second,
even when (as in MacRae 1960) cliques are definedas we definestruc-
in tiesto thirdpartiesratherthan
turallyequivalentsets (i.e., by similarity
by choicesof one another),the clique imageryis retainedand is often
allowed to limitthe interpretation. In blockmodels,on the otherhand,
partitioning of individualsis onlyone side of a dual problem;the other
is to interpretthe patternformedon the one or morenetworksby the
partition.13
The thirdcontrastis in use of spatial imagery.Most sociometry deals
with only one type of tie, sometimesan overall type constructedfrom
separatekinds of data. Several investigators (e.g., Laumann and Pappi
1973) eschewthe crudityof clique description, preferring insteadto view
the populationas embeddedin some abstractspace (usually Euclidean).
Even ordinalmeasuresof similarity betweenpairs can be convertedinto
quantitativemeasuresof locationand distancethroughsome variantof
multidimensional scaling (Shepard 1962; Kruskal 1964a, 1964b; McFar-
land and Brown1973; Arabieand Boorman1973; Shepard1974). Cliques,
as well as manyothersociometric concepts(e.g., connectivity), can then
be expressedin termsof locationsand distanceswithinthe space. In con-
trast,blockmodelsassume no such spatial embedding.Presumablyeach

13 Basic to our work has been our desire to conceptualizemany ideal-type patterns,
each suggestiveof a differentform of social organization,and to performtests that
reveal which (of all conceivable patterns) actually exist in a population.

736
Social StructurefromMultipleNetworks.I
network,each distinctquality of tie, requiresits own space, whereas
blockmodelsare able to sidestepthis matter.Perhapsa morebasic ques-
tion,though,is whetherany spatial representation is suitablefora net-
work,sincethe essentialfeatureof social networksmay well be the sharp
breaksin patterns-the"holes" in the networks.'4
The fourthcontrastconcernsboundaries.Sociometryusually takes as
givensomesplitbetweenthepopulationstudiedand the restof the world
(i.e., it assumes a clearcutenclave). However,Barnes emphasizedthe
artificiality of this presuppositionand urged instead a distinctionbe-
tweenthe finite"reach" of networkeffectsand the notionof a sharp
boundaryarounda particularset of people (1972, p. 16). In two of the
case studiesbelow (the biomedicaland the Firth-Sterling), we argue that
blockmodelsapply to networksamong people who are embeddedin a
largerworldand who thus comprisean "open" population.
The fifthcontrastinvolves a basic methodologicalissue. Moreno's
originalemphasison concretediagramsof ties among individualswas
sound.As sociometric analysis"advanced,"though,it became more and
moreweddedto approximations by indices (of whichspatial embeddings
and triadinventories[Holland and Leinhardt1976] are amongthe most
sophisticated).Balance theoryinitiallysignaleda reversalof this trend;
however,as soonas it becameclearthatno real data sets were"balanced"
in the classicalsense,researchers began an unrewarding searchforindices
of thedegreeof deviationfromclassicalbalance (see, e.g.,Flament1963).
We argue,instead,that sociologicalanalysisneeds explicitmodelsof the
structures in observedpopulations,not measuresor statisticalindicesof
deviationsfromsome convenientideal structure.Blockmodelswere de-
velopedto meetthisneed.

METHODS: PHENOMENOLOGY AND ALGORITHMS

StructuralEquivalenceand Blockmodels
Considerties of one type, fromone person to another,arrayedas a
square matrix,witha row and the corresponding columnof the matrix
assignedto each personin the population.Create a separatematrixfor
each typeof tie. Figure I presentsthreekindsof tie amongbiomedical
researchersspecializingin the neural controlof food and water intake
Maier, and Miller 1973). An entryof X means that a tie is
(Griffith,
"present"; a blank space, that a tie is "absent." The left-handmatrix
representsall ties of mutual personalcontact.The other two matrices

14 Breigeret al. (1975) compared blockmodelswith multidimensional scaling solutions


in considerabledetail for several sets of data (using both the MDSCAL and INDSCAL
algorithms); they argue that there is agreementbetween results obtained by the
various approaches.

737
-~~~~~~~
k k k k kb k :i
k k k k --------v

1 a a a a a ak a kc k
a a a k
a ><
a ><
a , a csE
(: c
ck k
I k k k k k k
kc sv
k k k k
Caaak k k
k k k k k k k>
~~
~~~ ~~~~~~~~
, kRR aR k
uS~~ k k k k k k ><

ul
z~~
I
I
~~~~
>C

k k >1
C 1
=r
CT
cr k k k k ;a a k k I.
I
D~~~~~~~~~ CC aI
k k >1;$I C
k k ,

a kb k k k k k k k e Cb e

I k Caaa a k
a Nb Nbl <
k k k k k k k k C0=
___________---

X~~~~~~~~~~~~~~
Ok k
_________------ Vkk
1R P a a a | W

3 >< kC >< k,CS < O 0 .

t1J~~ ~ ~~
kb ~ ~~~ k k I c
X

O I
lOC~~~~
k k
~ ~k ~ ~~~~~~~
wa
Na aa
k
aI.
I - E

18 aR k w a a I S
k k
k~~~o
--~~~~~~~~

k k k k~~~~~~b I ,,,

.1 k k k :k h k l k

Ik k
1 k kI U e

I k k: k
kC
-4
ab
k
El

I k k C R1 =

I k v~~~~~~~~~~~
k k V k
kC
I j S O

H~~~~~~~~~~~~~~~~
C> Otu Dba
H H
>nSt
c1 H H
sr
H1 H
:
H
1n_u
rt N CN CN CN CNCh CN CN CNn
DFO < c
Social StructurefromMultipleNetworks.I
represent "unawarenessof man or his work,"distinguishing pairs of indi-
viduals who reciprocate"unawareness"(the right-handmatrix) from
pairs in whichonly one individualindicatedan "unawareness"tie (the
middlematrix).Only an arbitrarily chosensubsetof 28 membersof the
fullsample (N = 107) is included.
Blockmodeling beginswithweakeningand extendingthe algebraiccon-
cept of "structurallyequivalent" actors in a network(Lorrain and
White 1971). A self-consistent search procedureis used to partitiona
populationinto sets of structurally equivalentactors blocks. In each
data matrix,we rearrangethe rowand columnof each individual,so that
the membersof a blockare groupedtogether. We also use the termblock
fora rectangular submatrixin whichtiesof thegiventypefrommembers
of one block to membersof anotherblock are reported.(The context
will specifywhichof the two meaningsis intended.)Attention is focused
particularlyon blocks which have no, or very few,instancesof ties:
theseare termedzeroblocks.
Look ahead to figure3, in whichthe 28 personsin figure1 have been
partitioned into fourblocks. (For example,the firstblockhas fivemem-
bers: individualsnumbered9, 26, 23, 4, 1.) The rowsand columnsfor
individualshave been rearrangedso that each of the threematricescan
be seen as 16 blocksdisplayingties fromone of the foursets (blocks) of
individualsto another.For example,in each matrixof figure3 the upper
leftblockreportsany tiesamongthe firstfiveindividuals;adjoiningit on
the rightis theblockreporting any tiesfromthesefiveto the secondset
[block] of sixindividuals;and so on. Thereare eightzeroblocksin theleft
matrix,fivein themiddleone,and fourin theright-hand one. The pattern
of zeroblocksin thisfigureis interpreted in the next section,wherecase
studiesare discussed.
A blockmodelis a hypothesis about a set of data matrices:it specifies
foreach matrixwhichblockswillbe zeroblockswhensomecommonparti-
tion of the populationis imposedon all the matrices(as in fig.3). A
blockmodelconsistsof a square binarymatrix,called an image,foreach
typeof tie. Each image has a row and a corresponding columnforeach
block (in fig.3, the top panel of three4 X 4 matricesshowsan imagefor
each typeof tie). The orderingof blockswithinthe blockedmatricesis
arbitrary, as is the orderingof memberswithina block.
Five ideas are basic to blockmodels.First, structuralequivalencere-
quires that membersof the populationbe partitionedinto distinctsets,
each treatedhomogeneously not only in its internalrelationsbut also in
its relationsto each othersuch set. Second, the primaryindicatorof a
relationbetweensetsis not theoccurrencebut theabsenceof tiesbetween
individualsin the sets. Third,manydifferent typesof tie are needed to
portraythe social structure of a population.Fourth,the natureof a type

739
AmericanJournalof Sociology
of tie is inferred
fromthepattern,in a givenpopulation,of all tiesof that
type.
The fifthidea embodiesour basic claim about aggregation.A model
of social structurerequiresspecifying,
foreach pair of sets on each type
of tie,whetheror not a zeroblockexists.Thus, a givenset-a "block" in
our terminology-may be debarred,by the model,fromties of one type
withseveralotherblocks,and fromties of a secondtypewitha different
collectionof otherblocks,and so on.

Bonds and Segregation


Suppose the networksfora givenpopulationsatisfya particularblock-
model hypothesis.When these matricesare rearrangedand partitioned
accordingly, thoseblockswhichare not zeroblocksare usuallynot com-
pletelyfilledwithties. They have a speckledappearancebecause choices
frommembersof the row block to membersof the columnblock are
interspersed withblanks showingno ties. There are several reasonsfor
thissituation.First,if the data are sociometricresponses,thenthe num-
ber of responsesper personis usually limitedto an arbitrarynumber
thatis normallyinsufficient to yieldsolidblocksof entries.If an observer
infersties (insteadof askingthe subjectsabout them),he will be unable
to monitorall possiblepair interactions continuously;thus he may miss
briefoccasionalcontactsthatare,in fact,enoughto maintaina tie (if not
to originateit). At any given time,chance fluctuations(who has been
talkingto whom,etc.) may determinewhichparticularties are coded as
present.
Second, aside fromlimitationsin data collection,the persons being
studiedmaynot be motivatedto reportall theirties.At some times,they
mayevenwishto concealsometiesfromothers,or evenfromthemselves-
hencealso fromany investigator. The act of revealinga tie-one's asym-
metriccontribution to a pair connection-isa tacticaldecisionin an on-
goingsituation.'5
Third,thereis no need fora personto maintaineverytie to all indi-
viduals who belongto his own or anotherblock,even thoughthe num-
beroftiesbetween(or within)theseblocksmaybe considerable. Maintain-
ing a tie requirestimeand energy;in addition,it makesa claim on an-
otherperson'sattention.Blockmodeling requiresthat ties of a giventype
fromany personin one block to any personin anotherbe equivalentin
structuralsignificance;however,not everyoneneed choose to mobilize
all such ties all the time (White 1974a).
Finally,the populationneed not be a naturalgroupin mutual face-
15This theme is furtherdeveloped in Schelling (1960). That self-censorshipactually
occurs is furthersuggestedby the Firth-SterlingCorporation data analyzed below.

740
Social StructurefromMultipleNetworks.I
to-facecontact,all of whosemembersare automaticallyacquaintedwith
one another.It maybe a contactnetwork in whicha particularpersonmay
nevereven have heardof half the others,as in figure1. Blockmodelsare
whollyapplicableto suchcases.
The blockmodelhypothesized fora set of matricesis an interrelatedset
of inferencesfromthosedata to an aggregatedpatternof ties amongcer-
tainsetsof persons.The memberships of thesesets (the blocks) are influ-
encedby each otherthroughtheincidenceof tiesof everytypeacrossthe
wholepopulation.Bonds is the term'6assignedto thoseblockswhichare
notzeroblocks, eventhoughmanyor mostof theentriesare blanks.
Sociometricstars and otherconceptsthat try to captureindividuals'
popularityhave no direct analogue in blockmodels.17 Segregationof
choices,as betweenboys and girlsin grammarschool classrooms(Bjer-
stedt 1956), has oftenbeen noticedin sociometricanalyses.This phe-
nomenoncan be describedby zeroblocksin a blockmodel,but has ap-
parentlybeen investigatedonly for a prioricategoriesof persons,such
as male and female.With reference to less extremeformsof segregation,
thinkof zeroblocks(such as those in the mutualcontactmatrixof fig.
3) as markingtheboundariesof choicesby subgroups.Because individual
popularitydepends on the size and compositionof the group or unit
underconsideration, it may be arguedthat thisclass of sociometric con-
ceptsdependslogicallyon blockmodels(or somecloselyrelatedapparatus)
fordelimiting the boundariesof such units.Withinthe top leftblock of
the figure3 mutualcontactmatrix,forexample,it is apparentthat the
firstindividual(#9) is mostpopular (he is chosenby each of his block-
mates) while the second individual(#26) receivesfewerchoices.This
factis masked,however,in the unpermuted data of figure1, whichshows
individual#26 receivingmorechoicesoverallthanindividual#9.

Images fromReciprocityand Reflexivity


Much of sociometryemphasizesthe distinctionbetweena reciprocated tie
and an asymmetric tie (Davis 1970). This distinction
is used in block-
modeling,but usually betweenblocks and not betweenindividuals.On
the otherhand, reflexivity
is merelya technicalquestionin sociometry,
whereasin blockmodelsthe existenceof diagonal entriesfor a given
image involvesa crucial substantivequestion.In sociometrythereare

16 Breigeret al. (1975) used the term "1-block" instead of bond.


17 Connectivityin a sociometricgraph may depend cruciallyon a single tie between
two individualsand is thereforehard to relate to blockmodelideas (compare also the
concept of a sociometric "bridge" suggested in Granovetter 1973). Holland and
Leinhardt (1973) exploredthe sensitivityof sociometricmodels to measurementerror
in sociometry,but they came to unjustifiablypessimisticconclusions.

741
AmericanJournalof Sociology
only fourpossibilitiesforany pair of individualson a giventypeof tie:
reciprocal,null, or one-way(two possibilities).In contrast,blockmodels
permit16 imageson two blocks; these are shown,and fixedlabels as-
signed,in figure2. In sociometry, threeof the fourpossibilities
are struc-
turallydistinct;in blockmodeling, 10 of the 16 imagesare distinct(see
fig.2).
In the blockmodels forfiveof our case studies,imagesE and V (and F
and W) occurredfrequently and are substantively important.V can cap-
ture,in a crude 2 X 2 pattern,the structureof a hierarchyas usually
idealized: tiesof deferenceextendedwithineach blockto one's immediate
superiorsare also observedfromthe lowerto the higherblocks in the
hierarchy.In contrast,image Y can be seen as aggregatedeferenceac-
cordedonlyby personsin the lowerblock onlyto personsin the higher
block. Call E the "hangers-on"pattern.It suggestsdifferential standing,
but in a differentway.Here thesecondblockis notinternally coherentbut
is part of an overalldeferencestructure. The E imagesuggestsa distinc-
tionbetweenthe centerand the periphery.
ImagesP and N represent pure reflexivity
and pure symmetry, respec-
tively.Sociometric balancetheorymay be expressedby thesetwoimages:
positivebindingwithineach of two cliques (reflexivity-P)and hostility
betweenthe two cliques (symmetry-N). Element C singles out one
clique fromthe remainingisolatedindividuals.If the type of tie repre-
sentedhere has negativeconnotations, theneitherC or D showsa con-
centration of hostilitywithina subset.
ImagesH and T are patternsforperceptions held by membersof both
blocks: all ties frombothblocksgo to one block. Images S and G show
peoplesplitintoa passiveblockon theone hand and an activeblock that
does not discriminate betweenitselfand the passivesin ties of the type
understudy.The usual forced-choice proceduresforgathering sociometric
data wouldprecludediscoveryof patternsS and G.

of Blocks
The Aggregation
The imageof a blockmodelmay be portrayedto any desireddegreeof re-
finement by combiningblocks to achieve a "coarser" image (one with
a smallernumberof rowsand columns)or by further splittingthe exist-
ing blocksto achievea "finer"(larger) image.Any blockmodelon three
blocks,forexample,can be collapsedformally into a blockmodelon two
blocksby combiningany two of the blocks.Since the orderof blocksis
arbitrary,thereare (211-1 1) ways to collapse an n-blockblockmodel
intopossibleblockmodelson twoblocks.'8Of course,theruleforcollapsing
18 Each of the original data matrices reports informationonly about ties of that
type for all pairs in the population. Permutingthe rows (and correspondingcolumns)

742
Social StructurefromMultipleNetworks.I
mustrequirethatif any one of the refinedblocksbeingcombinedinto a
coarseblockis a bond, the coarseblock mustalso be a bond. Oftentwo
blocks are not sufficientto captureeven the grosspatternsin a block-
modelfora case study.19
Observethat each bond in the C,F pair of images for two blocks is
also a bondin theE,F pair: evenwhenthenumberof blocksis the same,
one blockmodelmay be a refined(i.e., more demanding)versionof
another.In principle,one can constructan inclusionlattice
21 of block-
models,on a givennumberof typesof tie, beginningwiththoseon two
blocksand thenextendingthe latticeto threeand moreblocks.In prac-
tice,the possibleblockmodelsare far too numerousforthisto be useful.
For example,thereare 104 singleimageswith threeblocks,whichare
distinctunderpermutation of blocks.
Formally,thepairof imagesC,F is simplya moredemandingversionof
V,F, but (as will becomeapparent) the social structures
describedhave
quite differentqualities. Blockmodelsprovidea framework for making
substantivejudgmentsand interpretations; theysupplya set of formal
answers.However,thesolutionmustbe proposed,as wellas validated,on
substantivegrounds.

Two Algorithms
For the cases studiedto date, up to halftheblockshave been zeroblocks.
Intuitively, to findany partitionof rowsand columnsfor
it is surprising
a set of arbitrarymatriceswhichfita blockmodelcontainingmanyzero-
blocks,but in principletherecould be many.The numberof possible
partitionsis astronomical.G. H. Heil has devisedan efficient computer
algorithmforconstructing all assignments(if any) of men to blocksfor
which the rearrangeddata matricesobey the given blockmodel.This
algorithm,called BLOCKER, is describedin detail elsewhere(Heil and
White 1974).
In carryingout BLOCKER, one can identifypersonswhose assignment
to one or more particularblocks in the blockmodeleffectively deter-
mines the placementof many otherpersons.Such individualsmay be

does not change the information.After a partitionis imposed, the blocked matrix
for a given relation contains the same pair data as before. The use of blockmodels
can be urged on the purelymethodologicalgroundsthat they permitflexibleaggrega-
tion which retainsthis permutationinvariance (see White 1974a).
19 True for the biomedical and the monasterycases examined below. In contrast,
the essentialpatterns,and thus qualities,in the Newcomb fraternitycan be described
by a pair of images (V,F).
20Szasz (1963) provides relevant lattice-theoreticbackground. When building such
a lattice,the order of blocks is obviously significantin comparingmatrices.

743
o 0
z
o 0

1 0 0 0
C D
O O 0 1

0 i 0 0
x y
O 0 1 0

1 0
p
0 1

0 1
N
1 0

1 1 0 0
G S
O 0 1 1

1 0 0 1
H T
1 0 0 1

1 1 0 1
E F
1 0 1 1

1 0 1 1
V w
1 1 0 1

1 1
U
1 1
FIG. 2.-The 16 possible 2 X 2 binary matrices. Grouped into 10 rows, one for
each set that is equivalentunderpermutationof the two blocks. The letterlabels used
here and in the text are applied only to these matrices (never to designate other
images or to data matrices).
Social StructurefromMultipleNetworks.I
termedcrystallizers:theyresemblesociometric starsin importance, notbe-
cause of thenumberof choicestheyreceivebut becauseof theirstrategic,
"structural" positionin theoverallmatrixfollowing fromthehypothesized
model.Otherpersonsare allowedmultiple,alternativeassignments by the
BLOCKER algorithm;theseare termedfloaters. They are somewhatanal-
ogousto thesociometric isolate,whoreceivesfewor no choices.21
The numberof different blockmodelhypotheses, even withjust two or
threeblocks and just two typesof tie, is so large that some otherap-
proachis desirableforinitialexploration.Breigerhas developeda hier-
archicalclustering algorithmthatpartitionsmeninto possibleblocksand
thenfindsa blockmodelby inspectingthe data matricesrearrangedac-
cordingto the partition.22 It is called CONCOR; its formalbehavioris
analyzedin Breigeret al. (1975); some mathematical propertiesare de-
scribedin Schwartz(1974).
The difference betweenthe two algorithms is as follows.CONCOR pro-
duces fromraw data an assignment of individualsto blocks,and thence
suggestsa blockmodelhypothesis.BLOCKER demandsa blockmodelhy-
pothesisand derivesfromit any assignment of mento blocksthatsatisfies
the hypothesisforthe givenset of data matrices.Matrix entriesin any
numericalformcan be directinputto CONCOR, whereaseach tie mustbe
coded as either0 or 1 beforeBLOCKER can be used. Substantivejudgment
is requiredin both: in CONCOR, on whento stop the further splittingof
blocks;in BLOCKER, on whatblockmodels constitute appropriatehypotheses.
BLOCKER searchesforpure zeroblocks;an assignment is rejectedfora
blockmodelhypothesis if even one tie therebyappears in any zeroblock.
CONCOR partitions the populationin a way that may be intuitively char-
acterizedas yieldingsharpcontrastsin densitiesof ties betweendifferent
blocks.The image suggestedby CONCOR can be refinedby varyingthe
cutofflevel of tie densitybelow whicha block is coded as a zeroblock.

21 See AppendixA formore specificinformationabout the assignmentof men to blocks.


22 Specifically,given k matrices,each of size n X n and reportingties among a
population of n actors, a two-dimensionalmatrix (MO) with k X n rows and n
columnsis formedby "stacking"each of the k matricesone above the other,taking
care to preservecolumn ordering.(Alternatively,the 2nk X n array of each matrix
and its transposemay be formed.) The n X n correlationmatrix (M,) of product-
moment correlationcoefficients among columns of MO is then formed.This process
is iterated (MJ+1 is the matrix of correlationsamong all pairs of columns of Mj)
until a limit matrixis obtained,which may be permutedto yield a bipartitedivision
of the actors (columns) into exactly two subsets (blocks) of sizes s and n - s. A
refinementto any desired number of blocks may be obtained by creating a new
array Mo with k X n rows and s columns, k X n rows and n - s columns, etc.
In theirdetailed descriptionof the algorithm,Breigeret al. (1975) include extensions
and comparisonswith multidimensionalscaling and hierarchicalclusteringmethods
in the literature.We wish to thank A. Tagg of the Universityof Surrey for calling
our attentionto the anticipationof this algorithmin McQuitty and Clark (1968).

745
AmericanJournalof Sociology
The imagerequiredby BLOCKER can be refinedby following the inclusion
lattice for images.
In the theoryand interpretationof blockmodels,neitherBLOCKER nor
CONCOR is indispensable.It is possible, thoughlaborious,to find and
test blockmodelhypothesesby simplyinspectingmany permutations of
the data matrices.

FIVE CASE STUDIES

This sectionreportsfivecase studieson whichwe testedblockmodels.


Four concernadultsin worksituations;onlyone (the Firth-Sterling Cor-
porationmanagement)includedall of the population'srelevantauthority
figures.Two are panel studiesusingat least fourtimeperiods.Four of
the populationsare primarilyface-to-face groupsof fewerthan 20 mem-
bers,and one is a subsamplefroma largersample (N = 107). Four are
exposedto normalturnover in members. All fiveincludesystematic data on
individualattributes. All thepopulationsare twentieth-century and Ameri-
can. This sectionconcludesby illustrating the use of blockmodelswith
overtimedata for two of the studies.
For fourof our case studies(all but the biomedicalresearchnetwork),
detailedindependent analysesare presentedin the originalstudiesof the
interactions.These discussionshave informedour search forblockmodel
hypotheses used as inputto BLOCKER and are also used to validatesome
of our findings.We cite here fivegeneralfindingsfromthe case studies
whichillustratethis validationof the blockmodelsand also additional
insightsobtained.(1) CONCOR, a mechanicalsearchalgorithmnot depen-
denton ourperceptions of the"meaning"of thedata, producesa partition
of individualsinto blockswhichis equivalentat the three-or four-block
levelof refinement to the identificationof majorgroupings in the original
studyof the interactions, in all fourcases for whichsuch comparisons
maybe made (see resultsreportedbelowand in Breigeret al. 1975). This
strongfindingsuggeststhe validityof our approachto networkaggrega-
tion. (2) Even thoughmany of our hypotheseswhich are tested by
BLOCKER are informed by our readingof the originalstudies,blockmodels
constructed across severaldifferent typesof social relationin each study
are valuablein portraying theoverallsocial structure.(3) BLOCKER parti-
tionsforcoarse (two- and three-block)modelsagree withthe partitions
independently derivedby CONCOR. (4) In each case study,we suggest
additionalinterpretation (based on a finerpartitionof individualsinto
blocksand/ora consideration of the patternsof relationsbroughtout in
the blockmodel)whichgoes beyondthe analysesof the originalaccounts.
(5) The case of the biomedicalresearchnetwork,for whicha detailed
analysisof theinteractions amongthepopulation'smembersdoes notexist,

746
Social StructurefromMultipleNetworks.I
illustrates
theutilityof blockmodels procedurein search-
as an exploratory
ing forstructurewhenthe onlyavailableclues consistof reportsof inter-
actionamongpairsofindividuals.

A BiomedicalResearchNetwork
Data.-Griffithet al. (1973) identified studyingtheneural
173 scientists
controlof hungerand thirst.Of these,107 respondedto Griffith's question-
naire. In more than half the possible instances(as fig. 1 shows) one
respondent was unawareof another,as can be expectedin an openpopula-
tion.
Blockmodel.-In order to apply BLOCKER to these data, each entry
in the matricesmustbe coded in binaryform.Only a reciprocated choice
Onl"mutualcontact"was regardedas strongenoughby itselfto preventa
zeroblock; thus symmetric choices on mutual contactwere coded "X"
and the restcoded "blank." "Unawareof" is not like other,substantive,
typesof tie,so unreciprocated choiceson it weretreatedas a distincttype
of tie, withreciprocated choicesconstitutinga thirdtype.A blockmodel
hypothesis, statedin the top panel of figure3, forthesethreetypesof tie
was developedby systematically exploringintuitivelyplausible block-
models.At the left in figure3 is the partition-theunique assignment
(definedin AppendixA)-of mento blocksthatBLOCKER yieldswhenthe
data are testedagainstthe hypothesis. The bottompanel showsthe data
matriceswithrowsand columnsblockedin conformity withthispartition:
inspection showstheblockmodelis confirmed.
Interpretation.-Interpret the blockmodelin status terms.(Blocks are
orderedfromhigh to low status.) On symmetric "mutual contact,"the
bottomtwoblocksare connectedneitherinternally norto one another,and
thebottomblockhas no connections withany block,includingitself.The
bottomblock belongsto the populationonlyin the culturalsense that it
has no asymmetric unawareness of theblock thatis obviouslythe leading
set of researchers. No block has asymmetric unawarenessties to the top
block; yet thatblock has asymmetric unawarenessentriesto each of the
others;we mightcall thisa snob effect.
Furthertests.-CONCOR was applied (not shown here) to the data
matricesfor "mutual contact"and "unaware." Its firstsplit of the 28
menyieldeda partitionsimilarto BLOCKER'S: the firsttwogroupsin the
latterbecameone group,and the last two a second,exceptfortwo inter-
changes.Aftertwo moresplits,thepartitionwas
(1 4 9 23 2 10 26) (24 19 12 14) (6 7 28 11 15 13 16)
(18 22 3 5 8 17 20 2125 27). [1]
The extremeblocks are close to those of figure3; the middle two are

747
IN~~ N4 N N N I ~NN NN N

z~~~~~ ;qA
N
daaaa
N
a lba tb
N
t t
N
C

oSY iI II 0<~< II N N <


II N NIa
N 1
CD CIN
9N4NN 4 N N
:t I I N N IN NN4W I
_ r4p-Ir4 I
1e IN k s< | 9NI NN N k kI c
oo-4
o o-
~ Z I
I
- I-
I
- - -
I-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~-
I
- I -
N
s
N
I
-
N
N
-

:C I--------_--_-____________ - _
I I I N I N NNNC I s
0 I I I N I Na a1
NN Ct
I I I Ibe I
I I I I N N NN4

I N N N -
- C I I I
:
II N I
CCDI

I NNN I I I
I NN
9-1 9- 9-:v- ~ 1
IeN~N~iNCk4
NN c:
k IIII M
NN
V

?- J |IN NNN Ic N I kN _ I -

I N I I I N I >

0000 1 - _ - ___ - -- ------ 0m

0 I I I N |I

!~~~~~~~~~~~~~~~~~~~~~~~~~~~ 0-
=> ~~ ~ I I

I I- --I - - - - - - -
M~~~
C I) C I
I; I - Ft

CD I) ID I) t I I
40 j I I- I
lft ~~~~~-
I --- ---I I O
I~~~~~~~~~~~~~~~~~
- ---- --- - --
I N IN I I
N I N I I I I~~~~~~~~~~~~~~~~~
U
2 I N I -I - - - I -4
0000
,-I~~~~~00U
0 IN~~~~C-
N0C) tI
I~~C- 1
N4j
q
WC. t0 I.tC V
N C 4I 4C -
GC4
I4
' t
-1V qV
)oC)V4
_ , t4C,
b
Social StructurefromMultipleNetworks.I
moremixed.For all threetypesof tie,we constructed a blockmodelfrom
theCONCORpartitionby defining a zeroblockas any blockwithless thana
specifiedfractionof theaveragedensityof entriesforthegiventypeof tie.
For any cutofffractionbetween1/10 and 1/2, the resultingblockmodel
was verycloseto thatof figure3.
Each of twoadditionalarbitrarily chosensets of 28 persons(not over-
lappingwitheach otheror withthefirst)was thenblockmodeled in exactly
the same way, again usingCONCOR. The two resultingblockmodelswere
close to each otherand to the blockmodelin figure3. Althoughthe full
sampleof 107 had not yet been simultaneously partitioned(see Breiger
1976), we inferredthat in the largerstudyboth the blockmodelpattern
and the block memberships would correspondwith the resultsreported
forthesesubsamples.Everyonein the networkknowsthe top dogs (block
a), but althoughthesetop dogs collaboratewithsomeresearchers in lower
strata,theyappear to remainignorantof most lessermortals.Members
of block b appear to be very active researchers, aware of one another.
Unlikethosein the bottomblock,membersof the thirdblock (c) are not
just on the sidelines; they frequently see at least some researchersin
higherblocks. Clearly,the completeinterconnectedness of a face-to-face
groupor othercommunity is notnecessaryforthe coherenceof thissocial
structure. Neithertheseblocksnor (moreimportantly) theglobalpattern
of relationsover the networkwould emergefromcounts of individual
''popularity"or fromconventional cliqueanalysis.

A Monasteryin Crisis
Data.-Sampson's (1969) detailed account of social relationsin an
isolatedAmericanmonastery shouldbecomea classic.Duringa 12-month
period,muchof it in residenceas an "experimenter on vision,"Sampson
developedan extraordinary varietyof observational,interview,and ex-
perimentalinformation on the monastery'ssocial structure.Toward the
end of his study,a majorblowupin the monastery culminatedin a mass
exodusof membersby expulsionand resignation.
Sampsondefinedfoursortsof relation-Affect, Esteem,Influence,and
Sanction-on whichrespondents wereto givetheirfirstthreechoices,first
on thepositiveside and thenon thenegative.Figure4 showsthesechoices
forhis fourthtimeperiod,beforethe blowupbut aftera new cohortof
noviceshad settledin. For example,novice#3 liked novice#1 best,and
thereforea 3 (representinghighestchoice) is enteredin the intersection
of the #3 rowand the #1 columnof the top leftmatrix(an entryof 2
meanssecondchoiceand a 1 meansthirdchoice). We assigneach monk
a numberfrom1 to 18 (roughlyin theirorderof joiningthe monastery-
the same orderSampsonused). We give each of the eightpositiveand

749
Americanjournal of Sociology
ill 500 110 100
Oil 110 050 050

LIKE ESTEEM INFLUENCE PRAISE

51 1132 1 I 1 1132 1 I 321S I II I I


91I1 312 I II1 3 21 11 I I 1 132 I II I
6 112 31 I II1 32 1 I 112 3 11 I 1132 1 I
4l3 12l I 1 12 13 1 I 112 13 1 I 112 13 1 I I
11 12 31 1 I I 1 23 I I I 32 1 I I I I 3 21 SI I
8 1__123-_ ...... I.. . 1 _221 l.....I II--- 1 - l 1 - -----? I- - I I-- . ..211 1 --? -- I..----
12 1 1 32 1 1 I 1 321 I I1 1 32 1 1 II I I I
i I 11 2 13 1 11 312 I II1 12 3 11 II1 12 3 lii
2 1 123 11 Ill 21 3 I II1 123 ii 1 123 ii I
14I 113 2 I I I 1132 2 I I I 1132 2 I I I I 33 2 11 I
15 I 12 3 1 I I I I 132 I I I I 32 1 I I I I 132 I I
7!1 11 3 21 I I 113 121 I I 11 3 21 I 1 I 3 121 I
.1? I.. .. .1 _ _2_.l......l I I ?.. . 11 3 22_.l----..I I-- -- - .- -2.l1.. ----..I I--- -- 2 --.
13 12 I 1 1 31 13 11 2 1 1 13 1 1 2 1 1 13 2 1 I I
3 I I 3 12 121 I I 3 12 111 I I 3 12 1 I I I 3 12 1 I
17 1 Il 1 231 1 Il 112 3111 Ii 1 231 1 I I I

111 ill 101 ill


110 110 110 i11

ANTAGONISM DISESTEEM NEGINFL BLAME

io! TI I TIT I I IT T T T I
5 I I I I I 1321 I I I 1321 I I I I
I 1 3211 1 I 11 321 1 Ii 1 321 1 I I I
61 1 12132 11 11 3 122 11 I 3 1211 1 I 3 11211
4l 1231 I I I 1 231 I I I 1 231 I I I I 32 11 I
11 I I 13 I 2 1 I I 13 11 2 1 I I 3 I 121 I I 1 I 231
-a I--- -- _ 1 - .-2... --I --- ---I? ?- 1 _..221 ---? -- --? - 3 12 ---..I I-- -- -I--- -- 1- 22 1I
12 I I I I I I I I I I I I I I 2 I 131
ill1 1 2311 II 1 23 11 112 I 11 31 1 2 1 1 311
2 1 3 21 Il II1 3 1 112 11 3 1 12 II I 13 1
14 l 3 21 1ii I 22 1 13 111 I 2 I 13 111 I 32 1 Ii I
i5 1 1 I 132 I I 31 1 1 2111 I 3 I 121 I 1 I 132 11
7 1 23 11 I I I 32 21 1 21 1 I 32 21 1 21 1 I 12 I 13 I
I ? I ?-- --- 112.1. I I -- ? _ --? ?- 1.2211 --- 22 ?1 --- -- 1 - l I---? ?_ ? ?--- 112221I
13 1 2 11 3 I I I 3 I 2 Il I I 2 11 3 I I I I 1 I 231
3 1 23 12 11 II1 23 11 I II1 23 11 I II1 3 121 I I
17 1 31 21 I I I 32 11 I I I 32 11 I I I I I I
18 12 ? 3--2 -- -- -I...... I I .. 22_31 ..... .. ... I I-- ...122. 1- -- I? - I----. I I-- -- -I-- -- -I..---- I

FIG. 4.-Blockmodel for the monastery,time 4: images and data matrices.In the
latter,3 stands for firstchoice, 2 for second, and 1 for third on each type of tie.
Source: Sampson (1969).

negativerelationsa distinctname.In all of our case studieswe use "Like"


and "Antagonism"forpositiveand negativeaffect,respectively.
Blockmodeling.-CONCOR was appliedto all eightmatricesof figure423
This algorithmis not explicitlyconcernedwithlocatingzeroblocks;data
of all threechoicerankingsare processedby it. Aftertwosplitstheparti-
tionintothreeblockswas
(10 5 9 6 4 11 8) (12 1 2 14 15 7 16) (13 3 17 18). [2]

This partitionwas thenimposedon the eightdata matrices(the results


are shownin fig.4). We assumedthatin a populationof 18 persons,only
thetoptwochoiceswerestrongenoughto invalidatea zeroblock(establish
a bond), so each block whichcontainedno entrygreaterthan 1 was

results are reportedin Breiger et al. (1975) because they, like


23 Slightlydifferent
Sampson, summedplus and minus entriesto combine the eight types of tie into four
matricesforCONCOR input.Our choice is perhapspreferablebecause it makes less strong
measurementassumptions; see also n. 24 below.

750
Social StructurefromMultipleNetworks.I
represented as a zeroblock.The resultingblockmodelis shownin the top
paneloffigure 4.
To applyBLOCKER, whichrequiresbinaryinputdata, values of 2 and 3
werecodedas 1, therestas 0. Whentheblockmodelshownin figure4 was
testedon thedata usingBLOCKER, theuniqueassignment foundwas exactly
thatderivedfromCONCOR!Directinspectionof figure4 confirms thisfind-
ing (note thatthirdchoices,codedas 1, shouldbe ignored).
In the image foresteem,the threeblocks are orderedin a complete
linearhierarchy, whichis certainlyplausiblein a monastery. The bottom
blockhas a bond to itselfand to each of the higherblockson everykind
of positiverelation;yet it also has reciprocatedbonds withboth of the
otherblocks on all fournegativerelations.Liking bonds are universal
withone exception:the secondblockdoes not matchits esteembond for
the firstwith a likingbond. The two top blocks exchangeno positive
sanction("praise" in fig.4); however,the firstblock, top on esteem,
concedesinfluence to thesecond.
Refinements and interpretations.-Ifwe returnto the CONCOR approach
and raise the cutoffdensityfor zeroblocksto half the average density,
theresultingblockmodel is

1 0 0 1 0 0 1 1 0 1 0 0
0 1 0 0 1 0 0 1 0 0 1 0
0 1 1 0 1 1 0 1 1 0 0 1
Like Esteem Influence Praise

0 1 1 0 1 1 0 1 1 0 1 1
1 0 1 1 0 1 1 0 1 1 0 1
1 1 0 1 0 0 1 0 0 0 0 1
Antagonism Disesteem Neg. Infl. Blame

The Like image is identicalwith Esteem,and DisesteemwithNegative


Influence.As in the previousblockmodel,the top two blocks exchange
negativebonds of all fourtypes,but thereis only one kind of negative
bond fromthe bottomblock to the second block. The concretesocial
structuresuggestedis muchthe same foreitherversion:a top-esteemed
block unambivalently positivetowarditself,in conflict
withbut conceding
influenceto a second,moreambivalent, block,to whichis attacheda block
oflosers.
For thissmallpopulation,furtherrefinement of thepartitionby mechan-
ical applicationof algorithmsis not justified,but hints in Sampson's
historicalaccountand inspectionof the matricessuggesteda refinement

751
AmericanJournalof Sociology
of the partitioninto fiveblocks,witheach of the two leftblocksin [2]
beingsplitas:
(10 5 9) (6 4 11 8) (12 1 2) (14 15 7 16) (13 3 17 18). [3]
This is theorderingofrowsand columnsused in figure4. BLOCKER, applied
as before,verifiedthat thispartitionprovidedthe unique assignmentfor
a certainblockmodel;the questionis: Does the blockmodelmake sense?
Considerfirstthe esteemimagefromBLOCKER in thisrefinedblockmodel
(thelinesshowdivisionsintotheblocksof [2] ):
a b c d e
a 0 1 0 0 0
b 1 1 0 0 0
c O 1 1 0 0
d 0 0 1 1 0
e 10 1o 1 1
(Recall thatthisimagefitsthe data with2's and 3's coded 1 and all other
entries0.) Name the blocks fromtop to bottomand fromleft to right
a, b, c, d, e. In brief,withintheold top block,a is now a hanger-onto b;
withintheold secondblock,d defersto thecoreblockc. It was c, but not
d, thatesteemedthe initialtop block,and thenit onlyesteemedthe core
(b). Moreover,it was the hanger-on,a, who concededinfluenceto the
initialsecond block,but only to its core, c. (The old bottomblock of
losersremainsthe same; it [e] esteemedonlythe hangers-on[a] of the
old topblock.)
The otherimagesin the refinedblockmodelcan be read fromthe data
matricesin figure4. All fourpositiveimagesconfirm the hangers-onand
deferencestructureswithinthe formerblocks. There are three more
likingbonds than esteembonds,but exceptforthis fact and the special
asymmetry in esteemand influenceamongthe top blocks (already men-
tioned),the fourpositiveimagesare almostidentical.The refinement of
negativebonds is simpler:in each of the fournegativeimagesthereare
reciprocalbonds betweenb and the bottomthreeblocks (c, d, e) but
almostnoneto b's hangers-on(a). And theloser,e, thoughreceivingmany
or all typesof negativebondsfromtheotherfourblocksand reciprocating
to botha and b, sends no negativebond to d; however,e sends all four
typesof negativebondto d's masters(c).
ComparisonwithSampson'sanalysis.-Both the blockmodelsforthree
blocks and that for five blocks can be comparedwith Sampson'sown
analysis.Sampson (1969, p. 370) positeda definiteclique structurefor
the monasteryat time T4, on the basis of sociometricgraphs (drawn
fromthedata shownin fig.4), his ownobservation, and his interpretation
of eventsand personalattributes.His Young Turks are led by monks2,

752
Social StructurefromMultipleNetworks.I
1, and 12 (in descendingorderof leadership),with 14, 15, 7, and 16 as
followers. His Loyal Oppositionis led by 4, with5 a popularmember,6
and 11 as members, and 9 less fullyattached.He saw threeOutcasts: 3,
17, and 18. The otherthreemonks(10, 8, and 13) waveredbetweenthe
two cliques,whichhe describedas beingin intenseconflict.
Both CONCOR and BLOCKER agreedon thesplitintothreeblocks (shown
in fig.4). Sampson'sLoyal Oppositionis whollycontainedin the first
block; the Young Turks are exactlythe men in the second block; the
Outcastsare whollycontainedin the thirdblock. Sampson'sWaverers8
and 10 are in the Loyal Oppositionblock,whereasWaverer13 is in the
Outcastblock.
Our refinedfive-block blockmodelsplitstheYoung Turksexactlyas did
Sampson,with1, 2, and 12 as leaders; however,the firsttwo blockssplit
theLoyal Oppositiondifferently,as wellas enlarging it. Monk 5 is changed
fromSampson's"socio-emotional leader" (p. 360) to membership in the
Loyal Opposition'shangers-onblock. Sampsonearlierobserved(p. 322,
n. 32): "His [monk5's] circumspect aloofnessfrominterpersonal conflicts
servedto preservehis relativelyhighrankingon mostmeasuresthrough-
out the study,but as a consequence,his influenceon otherswas more
that of a detachedrole modelthan a framerof opinionor action." And
Waverer8 is in the leadingblock (b) of the Loyal Oppositionaccording
to theblockmodel.
The patternof relationsgivenin figure4's blockmodel theeightimages
on threeblocks accordswithSampson'sbasic contentionof a fightbe-
tweentheLoyal Oppositionand theYoungTurks.The blockmodelreports
strongambivalencewithinthe Young Turks simultaneouspositiveand
negativebondsof manytypes but nonewithintheenlargedLoyal Opposi-
tion; all of thesereportsaccordwiththedetailedstatements in Sampson's
analysis.The blockmodelmakes threefurtherimportantassertions:the
Young Turksare concededtop positionin a linearhierarchyof the three
blockson influence, whilethe enlargedLoyal Oppositionis concededtop
spotin a hierarchy on esteem.Third,thebottomblockhas but one internal
typeof bond whichis negative;it also receivesLike bonds fromabove,
as wellas thenegativebondsthatits membersreturnin kind: thisimplies
thatthebottomblockis a meaningful social unitin a sensedifferent from
thatofSampson's"pseudo-group."
Bits of evidencein Sampson'sdetailed textualaccount supportthese
further assertions;forexample,monk13 nominates, and is the onlyman
to vote for,3 as chairmanof an importantmeeting(p. 354) .24 Our

24 Monk 13 is placed in the Loyal Opposition


by Breiger et al. (1975) because of
the way they apply the Breiger algorithm; see n. 23 above. Yet when they did a
multidimensionalscaling analysis for comparison,using Kruskal's MDSCAL algorithm,

753
AmericanJournalof Sociology
assertionscontradictsome of Sampson'ssummarystatements,but this
factper se is not as importantas the fact that blockmodeling has per-
mittedus to move beyondthe pictureSampsondrew-beyondthe kinds
of inferences thatare technicallyfeasiblefromsociometricdiagrams.The
blockmodelon fiveblocksnecessarilydiffers fromSampson'sconclusions;
it analyzesunitsfinerthanhis factionsbut ignoresthedistinctivebehavior
of individualsthat he emphasized.Until the matricesfor earlier time
periodsare takenup below,the main supportforthe refinedblockmodel
is theconsistency acrossimagesof thepatternwithineach faction.
One weekaftertheperiodto whichthesedata refer,an explosionstarted.
The Superiorand theNovice Master,togetherwiththe handfulof senior
monks(none of themincludedin the sociometric populationof the 18
monksin training),decidedin theirregularreviewprocessto expelmonks
2, 3, 17, and 18. The reasonsgivenby the seniorstaffwere,forthe latter
three,that theywere "too immature"and had "personalityproblems,"
whilemonk2 was considered"too independent, questioningand arrogant"
(Sampson 1969,p. 373). Only monks1, 2, and 3 of the 18 had been to
collegeand werecandidatesto be fullclericalmonks;theywereolderand
restlessevenunderthedrasticallyreduceddisciplinetheirseniorshad insti-
tuteda yearearlier,beforetheirarrivaland thatof monks10 through18.
Almostat once,monk1 voluntarily departed.Then,withina week,monks
16, 15, 14, and 7 left,in thatorder.A fewdays later,13 and 8 left,also
voluntarily. A monthlater still,monk10 left.Of the six remainingfrom
the 18, notethatfourhad been therein theold days beforethe changeof
discipline,and fivewerein theLoyal Opposition.A puzzle in bothSamp-
son'spictureand theblockmodelis whymonk12 remained.Otherwisethe
blocksfoundin buildingtherefined blockmodelfrompuresociometric data
fit the initialdeparturesperfectly:monk 1 followedhis blockmateim-
mediately,and the nextwave of fourwas preciselythe block assertedto
be theirsubsidiary;onlyafterthemdid monk13 leave, and he preceded
thetwofromtheLoyal Opposition.

Cliques and Strata in the Bank WiringRoom


Data.-Homans's (1950) classic account of the Bank WiringRoom
suggesteda six-blockblockmodel;25 only afterassessingthis hunchwill
in the usual way. The originalmonographic
we apply the two algorithms
treatment(in Roethlisberger
and Dickson 1939; hereafter abbreviatedas

monk 13 was placed substantiallycloser to other Outcasts, a placement consistent


with our presentblocking.
25This is a slight simplificationof an earlier blockmodel on seven blocks, reported
in White (1974a). Inspector 13, a separate block there, has been combined with
the third block (Wiremen W2 and W5).

754
Social StructurefromMultipleNetworks.I
R-D) is itselfa classic.A production
sectionthatwiredswitchboard banks
in a WesternElectricplantwas transferred to a separateroom; fora year
men sharedthis roomwithan observerwho,withotherresearchers, re-
peatedlyinterviewed the 14 men,monitored officialrecords,and so forth.
The observersreportedtheirjudgmentof the incidenceof five types
of tie26(shownin the matricesof fig.5). Two are the familiarLike and

110010 111000 001000


100000 111000 001000
000000 110010 111111
000110 000111 001000
100100 001111 001001
000000 000110 001010
LIKE GAMES ANTAGONISM
W4 I XX X I I I I I I XXIXXIXX I I I I I I I X I I I I
S1 Ix xix I I iX I I ix xix Ixx I I I I I I I X I I I I
W I lxxxi I I I I IXXx I XiXXI I I I I I I I I I I
I I I l X X _
W2 I I I I I I I I_ I xi XI_ I I
W5 1 1 1 1 I I Ixxxix
XXXI I I
II iX I I lxx
X I I XiXXiX
IXXIX lxxi
IJ
W8 1 I I xixXi I I I I I X iX I I xxxi I I

W7 I X I I lxxi I I I I I x lxxi xix I I Ixxxi I iX I

W6 I I I IX I I I I I xxix I I I I I XX i ix I I
S2 1 __1__1__ _ _ _ __1_1X _ _ _

111011 noll0 l
100100 000000
100000 0on00o
000111 1001ll
100110 100111
100111 101110
HELP WINDOWS

i xI xI
SiS I X I
I I I i ixII I II I I x I VxiXx ix I
I I I I X I IX I
wS I1 I I I I I I I I I I I I

W5 I1___1___1_1_
Xi I I I I I 1_1__
lxx I I __l_
---__ I i I
WB I I I I xix ix I I x I I I xIxx I

W7 I___1_1--LX_
W9 I I I I xi I ix I
iI_ I --
_1 ixxi
X_ IX xix
lX I

W6 1 xi i xxix I I xx I I x xxixxi I

FIG. 5.-Blockmodel for the Bank Wiring Room: images and data matrices

26 An additional type of tie is not reportedhere


because it could be received only
by soldermen.

755
AmericanJournalof Sociology
Antagonism, but threeothersreportcontext-specific types:"Games" is the
designationfora kind of affectionate horseplay(including"pinging"the
upper arm); "Help" (with productiontasks) is our name fora second
type; "Windows"reportschronicquarrelsover openingwindows.The
observerssaw these as stable interactionpatternsestablishedwhen the
sectionhad settleddown.In all typesbut Help, each tie is reciprocated. A
man need not "send" any ties of a giventype.When applyingBLOCKER,
each tiereported as presentwas coded"1."
In additionto twoinspectors(called hereafter I1 and 13,afterHomans),
therewerenine wiremen(numberedfromWl to W9 by positionin the
roomfromfrontto back) and threesoldermen(Si, S2, S4). Layout was
fixed:W1-W3 wereassistedby the soldermanat the frontof the room,
Si; W4-W6 by S2; and W7-W9 by S4; I1 and 13 sharedinspectionof
banksfromthemiddleteam.
Interpretingall the kindsof evidence,R-D, followedby Homans,con-
cludedthat the social structure was based on two cliques,locatedmainly
in the frontand the back of the room,respectively. They listedthe mem-
bers in the frontas Wi, W3, W4, Si, and I1, and those in the back
as W7, W8, W9, and S4; but theyalso saw nuances withineach group
and discussedotherindividualsas fringemembers.At otherplaces in their
accounts,theyemphasizedthat individualshave differential standingor
prestigein the informalsocial structure.To us, their account strongly
suggesteda hangers-on patternwithineach clique as well as stratacutting
acrossthecliques.
Developinga blockmodel. Games ties weredescribedas friendly, and
forboththemand Like ties the hangers-on pattern(the E elementwithin
the clique) seemsappropriate.But Gamestiesweremuchmorenumerous
than Like ties,so it seemedlikelythereshouldbe more men in a core
groupon Games thanin a core groupon Like. Like and Games between
themthusshoulddifferentiate each clique into threeblocks.Antagonism
was concentrated on, and within,the set of men whomboth R-D and
Homans judged to be, at best, marginalto the cliques,and who would
appear in the bottomblocksof the two cliques.These threetypesof tie
suggesteda blockmodelforsix blockson Like, Games,and Antagonism;
it was not clear what patternto expecton Help or on Windows,except
that the lattershouldbe concentrated in the back of the room.This ap-
proachsuggestsnot onlythe imagesbut also the memberships of at least
thehigherblocks.
The initialblockmodel was adjustedby inspection of theactual choices;
it is shownin thetoppanel of figure5 as the firstthreeimages.The data,
also shownin figure5, fit this blockmodel:BLOCKER indeed yields the
uniqueassignment shown(see AppendixA). Two important pointsshould

756
Social StructurefromMultipleNetworks.I
be made. This partitionis consistentwith indeed a refinement
of-
Homans'scliques.The partition
is
(W4 S1 W3) (Wi 1i) (W2 W5 13)
(W8 W9) (W7 S4) (W6 S2). [4]
Furthermore,each of the threeimagesis close to our firstguess forit.
Nested hangers-onpatternsover Like and Games, for each clique, are
shown,togetherwithone symmetric bond joiningthe cliques.All Antago-
nismties are withinor withthe bottomblockin each clique; in addition,
the frontclique's marginalmen receivemostof the negativebonds from
both the frontand the back cliques. The bonds in the last two images
in figure5 (Help and Windows)weresimplyread fromthe data matrices
upon whichthe givenpartitionwas imposed.
Testing the blockmodel.-WhenCONCOR is applied to all five data
matrices,the firstsplit is exactlybetweenthe firstthreeblocksand the
last three.Wheneach of thetwosetsof blocksis routinely split,theresult
is
(W4 Si W3 WI I1) (W2 W5 13)
(W8 W9 W7 S4) (W6 S2), [5]
again perfectconformation to boundariesin the fullsix-blockblockmodel!
CONCOR firstdistinguished betweenthe cliques and then,withineach set
positivelyboundtogether, distinguished
strata.A blockmodelon thesefour
blockscan be aggregatedfromtheone on six blocksin figure5 by taking
thelogicalunionof thefirsttworowsand the firsttwocolumns,and then
doinglikewiseforthe fourthand fifthrowsand columns.
One can also emphasizethe differentiationinto strataas the overriding
feature,ratherthan the split betweencliques. Suppose we combinethe
firstand fourthblocksof figure5, thesecondand fifth, and the thirdand
sixth.Whenunionsof the corresponding rowsand columnsare computed,
thefiveimagesbecome
L G A H W
I I 0 1 1 1 0 0 1 1 1 1 1 1 1
1 0 0 1 1 1 0 0 1 1 10 1 1 1
0 0 0 1 1 0 1 1 1 1 1 1 1 1 1
When thisblockmodelwas appliedby BLOCKER to the fivedata matrices,
we obtaineda singlesolution:the top block is the unionof the firstand
fourth blocksoffigure5, and so forth.
The analysesof the Bank WiringRoomin R-D and in Homansare the
basis forthe blockmodeland so can hardlybe cited as independent evi-
dence.

757
AmericanJournalof Sociology
Test based on reasons for work stoppage.-In the original report,
Roethlisberger and Dickson (1939, pp. 428-32) stated that the bank
wiringdepartment allowedthe nine wiremento claim allowancesforun-
usual workstoppages"beyondtheircontrol."Wiremenfrequently claimed
moretimeallowancesthanwerenecessary(contraryto the intentof the
wage-incentive scheme) because theywere willingto trade some loss of
income for some gain in security(expressedas uniformity in output
curves).Each timea wiremanclaimeda timeallowance,he was supposed
to give the reasonfor the delay. R-D coded 12 classes of reasonsand
thencross-tabulated the claims by reasonand wireman.Generalizedat-
titudesneed have littlerelationto specificpositionin a particularpopula-
tion,but context-specific
attitudessuch as thesereasonsshouldbe affected
by one's positionand should,hence,resemblethoseof othersin equivalent
positions.
In thiscross-tabulation,each wiremanhas a columnand each reasona
row.We used CONCOR to splitthewiremenon thebasis of theirrespective
columnsofcounts.27 The resultis
(WI W2 W3 W4 W5) (W6 W7 W8 W9). [6]
It is at once apparentthat the splitis preciselythat betweenthe blocks
of the frontclique and thoseof the back (see [4] or [5]). In particular,
themarginalmembersW2, W5, and W6, whomHomansdid notplace,are
each groupedwiththewiremenfromtheappropriateupperblocks.
In only two of our case studiesdid the populationhave specificjob
assignments, and only the Bank WiringRoom also had fixedlocations.
Let us returnto thesix-blockpartitiongivenin figure5. We alreadyknow
thatjointmembership in a blockfollowsneitherfromhavingthesame one
of the threejobs nor fromhavingdifferent jobs. Sayles (1958, a mono-
graphon industrialworkgroups)criticizedtheearlierliteratureforgiving
too muchindependentimportanceto dynamicsin small groupsas such;
thekeysto social structure and process,he argued,werekindsof job and
the flowof workimposedamongjobs. At firstsight,it seems that our
analysis providesa counterexample;however,close examinationshows
that the split betweencliques is as importanta determinant of the six
blocksas is thedivisionamongstrata,and the cliquesclearlyemergefrom
the layoutof the room.Similarly,forall typesof tie exceptAntagonism,
the patternof bondsbetweenblocksdependson the clique split as much
as on strata.

Newcomb'sSecond Fraternity
Data.-Newcomb (1961) analyzed two experiments in which 17 pre-
lived togetherin a fraternity-
viouslyunacquaintedmale undergraduates
27 See Appendix B for furtherdetails.

758
Social StructurefromMultipleNetworks.I
stylehouse,expensespaid. They weresubjectto observationand required
to supplymany self-reports, includinga completerank orderingof the
other 16 by "favorableness of feeling"duringeach of the 16 weeks of
the experiment. Newcombreportedonly measuresof associationfor the
rank orderings(whichcontributed only indirectlyto his accountof the
social-psychologicaldynamics).Here we take polar typesof tie (Like and
Antagonism),abstractedfromthe rank order for the last week of the
secondexperiment, and suggesta blockmodelforthreeblocks.The indi-
vidual-leveldata weredescribedby Newcomb'sassociate,Nordlie (1958),
who developedan independentinterpretation that is more explicitthan
Newcomb's.Individualsare numberedas in Nordlie'sAppendixA, as are
ranks(from1 formostfavorableto 16 forleast; no tiespermitted).(The
parallelfirstexperiment will be discussedin Part II.)
Developinga blockmodel.-In a populationof this size, the top two
rankswerebelievedto represent strongfriendshipchoices; thebottomtwo,
strongantagonism.There is a scapegoatin this group (man 10), who
receivedone of the bottomthreechoicesof each of the other16 persons.
For applicationof BLOCKER, the top twochoiceswerecoded as Like ties;
the bottomthree,as Antagonismties. There seemedto be a top group
that disdainedthe others,so the V,F blockmodel(for two blocks) was
hypothesized. The resultwas a splitof menintoblocksforwhichLike and
Antagonism satisfythe blockmodel;men 13, 9, 17, 1, 8, 6, and 4 werein
thetopblock.This is theonlysplitthatyieldsa solution,and it stipulates
two floaters(men 2 and 5) who can be placed separatelyor togetherin
eitherblock(see also AppendixA).
An obviousrefinement is a split of the bottomblock into (1) losers
and (2) a stratumnot internallyantagonisticand ambivalently oriented
to thetopblockofseven: theblockmodel is
1 0 0 0 0 1
1 1 0 1 0 1
1 1 1 1 1 1
Like Antagonism

Developing a blockmodel.-BLOCKER, using this model, yielded the


following:a secondblock (men 7, 11, 12, 2) and a bottomblock (14, 3,
10, 16, 5, 15); the top block is unchanged,but the two floatersare
now morerestricted.
CONCOR was thenapplied to the completerankings(each rank treated
as an integer)to yield threeblocks.These wereidenticalwiththe three
blocksBLOCKER yieldedfromthe top two and bottomthreechoices.To
definethe blockmodel,we state a cutoffdensity (the highestaverage
densityof "choices"in a block thatpermitit to be coded as a zeroblock

759
AmericanJournalof Sociology
forthat typeof tie).28 With the top two ranksdefinedas Like choices
and thebottomthreeas Antagonism, forany cutoffdensitybelowone-fifth
the averagedensityon that type of tie, the resultingblockmodelis the
same as the one presentedabove. When the cutoffdensityis raised fur-
ther,the firstbondsto disappearare the bottomblock'spositivebond to
itselfand its negativebond to the secondblock.
Comparisonof the blockmodelwithNordlie'sinterpretation.-Nordlie's
(1958, pp. 67-77) studyof "sugbrouping" yieldsan assignment of mento
"clusters"for week 15 that is entirelyconsistentwith our blockmodel.
For each pair of men,Nordliecomputedthe rank correlationcoefficient
of the choicessent by each to the 15 others,excludingthemselves.This
matrixof "intra-pairagreement"was then clusteredas describedby
Nordlie(1958, pp. 70-71) to producesubgroupsof nonoverlapping mem-
bershipand a residueof men assignedto no cluster.Nordlie discovers
fivesubgroupsforweek 15 of the secondexperiment:(1, 5, 6, 8, 13), (2,
4, 17), (7, 11, 12), (3, 14), (15, 16). Men 9 and 10 are not assigned.
Each of the five subgroupsis containedwithina single block of our
three-block blockmodel,if the two floaters(men 2 and 5) are allowed
eitherof theirassignments.
Nordlie does not distinguishLike and Antagonism, and he does not
go on to examine and interpretconcretepatternsof ties among his
clusters.The blockmodelabove suggestsa combination of balance theory
(in some form)withhierarchyas the forcesat work.

in a Company
ManagementConflict

Data.-This section develops a blockmodelon eight types of tie


specificallyrelevantto the businessactivitiesof the 16 top managers(in
a companywithabout 2,000 employeesproducing
1958) of Firth-Sterling,
specialtyalloyand abrasivesproductsforindustrialuse (fora fulldescrip-
tionof the data, see White 1961). The blocksand patternsfitwell with
independentevidence (White 1960, 1961) of chronicconflictsover re-
searchand development of new products(the primemanagement issue in
a companywhoseproductline turnedover everyfewyearsin the era of
the R&D fever).Managersare numberedas in the originalreport(White
1961, table 1): 1-S are R&D managers;6-8, sales managers;9 and 10,
production managers;11 and 12,executives;and 13-16,topstaff managers;
in addition,6, 9, and 12 are vice-presidents,and 11 is the president(a

28 It would not be satisfactoryto use average rank order in a block in place of


density of choices. The former measure averages Like and Antagonism choices
numerically.Ambivalence would be excluded; in contrast,a blockmodel permits a
given image to contain both a bond on Like and a bond on Antagonism.

760
Social StructurefromMultipleNetworks.I
man of elitesocial and businessstandingbroughtinto the ailingcompany
a fewyearsearlier).
Developinga blockmodel.-Withsuch ties, therewere no precedents
forhypothesizing a blockmodel.CONCOR was applied to the data matrices
for all eight detailed types of relation (the seven reportedby White
[1961,p. 192], withchoiceson thefifthtype-Respect forKnowledgeand
RespectforDecisions-separatedinto two matrices).The partitioninto
threeblocksfound29was
(11 9 13) (2 4 6 7 8 1 15) (5 10 12 14 3 16). [7]
Since each choice had been carefullyconsideredby the respondents,
the cutoffdensitychosen for zeroblockswas zero (i.e., a single choice
in a block preventedits being coded as a zeroblock).The top panel of
figure6 reportsour blockmodelon threeimages: SimilarBusinessPolicy,
Personal Friendship,and Uncomfortableness. The blockmodelwas ob-
tainedby imposingthe above partitiononto the matricesfortheserela-
tions.The bottompanelshowstheresulting blockedmatrices.(Obviously-

110 000 011


110 111 101
101 011 000
SIMILAR POLICY FRIENDSHIP UNCOMFORTABLENESS

11 X I I I I I I xx
XX I
gl I I I I I I I I Ix II
31X I XXI _ I I _ _ II Ix IXx
2 IX I XX I I I XIX I I I I XX I
4 I III I I I X I I I I I
6 1 1 1 1 1 I XX XI I I x I I X I
7 IX I X X I I I I XXx I I I X I I
8 IX I x I I I I xx XI I I x lx I
51 I 1 X I I I I x XX I I IX I IXXX I

5 1x_ I XX- I I I-X I IXI - I-


1 IlXx I I I I X I I I I I I
lo IXXX
I I I I I I I I I I
14 1 1 1 1 1 111 1 111
3 I X II I I I I I I I I
16 I- XxI ? I _ ?_ _ I _ I __ _ ?

FIG. 6.-Blockmodel for Firth-Sterlingmanagement: images and data matrices.


Source: White (1961).

29 Following the usual procedure,all eight 16 X 16 data matriceswere first"stacked"


into a 128 X 16 array. It was then found that two of the 16 columnsin this array-
those for men 8 and 16-consisted solely of zeroes. As the product-momentcorrelation
is undefinedfor vectors with no variance, these two columns were removed and
CONCOR was applied to the remaining14 columns. Men 8 and 16 were then placed
arbitrarilyinto blocks. When BLOCKER was applied to test fig. 6's model against the
data shown there,men 8 and 16 appeared as the only "floaters."We note that no
other columns contain zero variance in any of the data matrices for the five case
studies.

761
AmericanJournalof Sociology
an implicationof the zero cutoffdensity-thispartitioning matchesthat
producedby BLOCKER forthisblockmodel.)
Includedin both the Policyand Uncomfortableness matricesof choices
are guessesabout the question"who . . . mightsingleyou out . . . ?"
(for each relation). Guesses were requestedin order to increase the
scope of choices, especially on uncomfortableness. Indeed, 19 of the
27 guessescoincidedwithdirectchoices,whilefewwerereciprocated(an
indicationthat thesewere surrogatechoicesratherthan realisticpercep-
tionsof others'choices). Everyguess thatdid not coincidewitha choice
fellintoone of the 10 bonds shownin the images; none fellinto any of
the eightzeroblocks.
Interpretation.-Wenow turn to a substantiveexaminationof the
blockmodel, considering firstthedivisionof menintoblocks.The partition
is neitherintodepartments nor into formalranks.When CONCOR was ap-
plied twicemore (to split the two big blocks), thenmen 4 and 15 were
separatedout withintheleft-hand block; and men3, 16, and 5 weresplit
fromtherestof theright-hand block (see [7] ). These fiveblocks,ordered
as in thepermutation shownabove,correspond exactlyto the groupingof
managersby theirattitudesto R&D (inferredfromtheirquestionnaire
responses,interviews, and actionsin specificconflictsover R&D [White
1960, 19611). Indeed, the permutation classifiesmanagersby degreeof
hardheadednesstoward R&D. The Executive Vice-President(12), the
Treasurer (14), and the Director of AbrasivesProduction(10) were
utterlyskeptical.The PersonnelManager (16) and the two R&D man-
agersconcernedwithalloys (3 and 5)--the moreroutineside of produc-
tion,in whichonlyroutinedevelopment was carriedon-were closeseconds.
All threeSales Managers,plus the overallcoordinatorof R&D (1) and
the man (2) who ran the most speculativeresearchproject,were most
optimisticand favorableto R&D. The President'sTrouble Shooter(15,
in chargeof expeditingsome new products)and manager4 (a respected
engineerwithseveralinventionsthat he was tryingto produceby a job
shop operation)were favorablebut not optimisticabout results.In the
middlewere the President(11), the Vice PresidentforProduction(9),
and the Accountant(13), in the observer'sopinion the most realistic
personsof the entiregroup; all threewere acutely aware both of the
poor recordof Firth-Sterling's R&D and of the crucialimportance(for
thecompany'simageand borrowing capacity)of havingan R&D program.
The blockmodelon thethreemainblockswiththe threespecificimages
makesequal sense.In figure6 the block of 11, 9, and 13 is put on top,
consonantwithits obviouspreeminence in the eyes of all. It exchanges
policybondswiththe Sales block (the middleblockin the partition[7])
and receivesa policybond fromthe hardheadedblock; in addition,each
blockhas a policybond to itself.On PersonalFriendship,the top block

762
Social StructurefromMultipleNetworks.I
sendsno bonds,and it is notchosenby thehardheadedblock (a factshow-
ingtherealismof thelatter); whereastheSales block,in themostvulner-
able positionon R&D, claimsbonds to all threeblocks.There is no Un-
comfortablenessbondwithinanyblock.The top blockand theSales block
are uncomfortable witheach other,and each is ill at ease withthe hard-
headedblock.The latterhad repeatedspecificclasheswiththeSales block
(it seemslikely,fromthe interviews,thatthe hardheadsweresimplyun-
willingto revealspecificnegativechoicesforany reason).
We have not reproducedhere the matrixof choiceson "the managers
withwhomyou have the mostdealings."The image for theseobjective
choiceswouldcontainonlybonds,no zeroblocks.Indeed,if thedensitiesof
choicesare computedbyblock,theresultsare
(11 9 13) .50 .14 .11
(2 1 6 7 8 4 15) .57 .17 .12
(3 16 5 10 12 14) .33 .10 .10.
of entrieswithincolumns.Every-
The strikingfeatureis the homogeneity
one claimsmostdealingswiththe President'sblock,next mostwiththe
Sales block,and least withthehardheadedblock; but in thissmallpopu-
lationof top managers,each blockshowssome heavy contactswithman-
agersin eachotherblock.

Blockmodelsover Time
Blockmodelsalso make sense out of data describingsocial structureover
time.The possibilities are numerous.Blockscan be stableover time,with
theblockmodelchanging.On theotherhand,a blockmodelmaybe stable,
withtheblocks'memberships changingas rolesand positionsrotateamong
individuals(of course,we would need independentconfirmation of such
changes).Or therecan be completestability,at least forthe coarseparti-
tionsinto blocks togetherwithassociatedblockmodelimages.Successive
observationsof choicesexistedfortwo of the cases analyzedabove, the
monastery and the fraternity.The resultsare quite similar.
The fraternity data.-The stabilityof both the blockmodeland the
blocks,afterthefirstfewweeksof maneuvering, is the mainresultforthe
fraternity.We imposedthe three-block partitionfoundforthe finalweek
on the data foreach earlierweek.We also computedthe densityof the
top two choicesin each block foreach week (numberof choicesdivided
by numberof cells in whichchoicescould occur); we thencomputedthe
densityforthe bottomthreechoices.Figure 7 showsthe resultsof this
procedureforselectedweeks: 0, 3, 5, 8, 13, 15. The development seems
clear.In theinitialweek,theblocksshowlittlevariationin density;thus,
thereis littlejustification
forasserting
thatthethreeblocksexistas distinct

763
AmericanJournalof Sociology
WEEK LIKE ANTAGONISM

.21 .18 0 .14 .18 .31

0 .14 .25 .05 .29 .08 .12

.10 .20 .10 .24 .12 .17

.25 .07 .04 .02 .21 .33

3 .20 .15 0 .25 0 .20

.20 .08 .06 .21 .12 .20

.28 .04 .02 .02 .07 .43

5 .11 .42 0 .11 0 .37

.21 .08 .03 .14 .04 .37

.29 .04 .02 0 .10 .43

8 .14 .33 0 .07 0 .42

.20 .12 .03 .07 0 .50

.31 .04 .02 .02 .07 .43

13 .10 .42 0 .03 0 .46

.21 .08 .03 .02 .12 .43

.33 0 0 0 0 .50

15 .10 .42 0 .07 0 .42

.20 .12 .03 .05 .04 .50

FIG. 7.-Newcomb's second fraternity:densityof choices in each block, for selected


weeks. Partition of men into blocks: (13 9 17 1 8 6 4) (7 11 12 2) (14 3 10 16
5 15). Blocks are taken fromthe blockmodeltestedand sustainedon data for week 15.

structuralentities.By the thirdweek,though,a patternis clearlydiscern-


ible; the top two blocks clearlyshow no internalantagonism,and the
thirdis clearlyat the bottomof a three-part Then, by at least
hierarchy.
thefifthweek,notonlythefinalblocksbut also thefinalblockmodelhave
emergedwithremarkableclarity.Thereafterthe stabilityis marked.Now
if eitherBLOCKER or CONCOR is applied to the data foran intermediate
week,muchthe same blocksand blockmodelare found,whereasneithera
clear blockmodelnor clear blocksemergedfromtheirapplicationto data
forthefirsttwoweeks.
This conclusionhas littleinterestif menare simplyrepeatingthe same

764
Social StructurefromMultipleNetworks.I
choicesweekafterweek; theargument in our sectionon methodswas that
menshouldbe expectedto changetheirchoicesover time,but withinthe
confinesof blocks that containbonds. For example,take choices made
by the top block.In week0, thesesevenmen made 14 Like choices(top
two ranks); eightwereof differentpersonsfromthosewhomeach chose
in the finalweek.Of theseeightchoices,fivecreatedbondsnot predicted
in thecoarseV,F model(thatis, thebondswerefromthe top blockto the
otherblockof 10 men); threeadded to the existingbond. By chance,one
would expect an outcomelike this: 10/17 in the forbiddenblock and
7/17in theblockallowedaccordingto theLike image.
In orderto completethis example,we repeatedthe comparisonwith
week 15 foreach weekin turn.Definea changeforweeki to be a choice
in weeki thatis not matchedby a choicein week 15. We thensummed
changesforweeks0-4 and comparedthemwithchangessummedforthe
10 weeks5-15.30 Data forthe firstfiveweeksshow as manychangesof
individualLike choices(38) as weremade in the last 10. In the firstfive
weeks,60% of the changesfall in the zeroblock-exactlythe chance ex-
pectation-butin the last 10 only34%oof the (relativelyfewer)changes
are in thezeroblockof theweek15 model.Whenparallelcountsweremade
for Antagonismchoices (the bottom three ranks), a similar picture
emerged:45%oof theearlyweeks'choicesdiffered fromthoseof week 15,
comparedwith22%oforthe late weeks.For the earlyweeks,exactlythe
chanceexpectation(40%) fellin thezeroblockin theF image,whileonly
14%oof the (relativelyinfrequent)changesin late weekslay in this for-
bidden block. There is some indication,then, that individuals'choices
continuechangingevenaftera blockmodelhas stabilized;in addition,after
week 4, the changes-both for positiveand for negativeties-conform
muchbetterto thefinalblockmodel.31
The monasterydata.-Sampson claimed that, during the period to
whichhisdata refer, groupings
emergedwithinthemonastery and polariza-
tiondevelopedamongthem; blockmodeling does not show such clear-cut
changesovertime.At periodTi, beforethenew cohortof novicesarrived,
monks4, 5, and 6 werethecoremen,with9 and 8 moreperipheral;7 was
isolated,but immediately becamefriendlywithmonk16 in thenewgroup.
By T2, accordingto Sampson'sobservation, monks1 and 2 stoodout as
themostrespectedin thewholegroupof 18. Then incidentsmultipliedas
traditionaldisciplinesurfaced(even thoughin much milderform): for
30 No data were collectedfor week 9.
31In furtherwork (not reportedhere) on both fraternityexperiments,we distin-
guished five blocks in the late weeks. The pattern within the top block reportedin
the text, when split, is stable, but the personnel assigned to each half "circulate"
over time. Withinthe split top block, the elementsE and S are the images on Like
and Antagonism:a hangers-onpattern reinforcedby common rejectionof the lower
half.

765
AmericanJournalof Sociology

example,somenewnoviceswereshockedwhenshownthe waxedwhipfor
the Order'straditionalmortifications.
AboutperiodT3, monk2, without
explicitdisapprovalfromthe officials,
instigatedmeetingsof the novices
to discusstheirroutine.In the formalvote forchairmanof the meeting,
he received11 voteswhilemonk1 receivedthree(only monk4 was not
present).
Our analysisof changeis confoundedby the fact that the sociometric
fromT4. Thus the data forperiodsT2
data used here are retrospective
throughT4 showonlymoderatechanges.Even whenthe refinedpartition
into fiveblocksis imposedupon the earlierdata matrices,the countsin
blocks are quite uniformacross time periods. The sums of weighted
choices,by block,forLike are shownin the top panel of figure8.

Like:

T2 T3 T4

2 12 0 3 1 2 10 4 0 2 4 11 2 0 2

11 5 6 2 0 10 12 1 1 0 9 14 0 1 0

3 0 8 5 2 1 3 11 4 0 0 0 11 4 3
o 2 15 9 0 1 0 16 9 0 0 0 16 8 0

2 0 10 2 10 2 0 5 1 16 2 0 5 1 18

Antagonism:

T2 T3 T4

00 0 0 6 0 0 0 0 6 0 0 0 0 6

0 1 2 8 13 0 1 16 0 6 0 0 13 4 9

4 7 0 1 6 3 10 0 0 5 1 5 2 3 1

1 7 0 0 17 3 12 0 0 10 0 15 0 0 9

4 15 4 1 0 0 19 5 0 0 2 17 6 1 0

FIG. 8.-Monastery data: sums for Like (top panel) and Antagonism (bottom
panel) choices over threetime periods. (Individual choices contributeeither +3, +2,
or +1 to the sums, dependingon tie strength.)

Since the same block membershipsare used for each period and the
blocks are of nearlyuniformsize (with eitherthreeor fourmembers),
it was unnecessaryto convertto a densitymeasure.It is clear that the
blocksfromT4 also stand out as distinctiveunitsat the earlierperiods
(i.e., the groupingsexisted fromthe beginning).Certainlythere is a

766
Social StructurefromMultipleNetworks.I
consistentdeclinein Like betweenthe top twoblocks (roughly,Sampson's
Loyal Opposition) and the middle two (the Young Turks); however,
the buildupof likingwithinthe bottomblock (whichSampsondismissed
as the Outcasts) is equally striking.So also is the increasingfocus of
likingties withintheLoyal Opposition(largelyfromties withdrawn from
theotherblocks)upontheirleaders.
The parallel sums forAntagonismare shownin the bottompanel of
figure8. Again,the fiveblocksare discernibleat the earlierperiods.The
only notablepatternchangesare the concentration of antagonismfrom
the leaderswe identifyin the Loyal Oppositionupon the Young Turks'
leaders (theseLoyal Oppositionleaders,in turn,are increasingly disliked
by followerswithintheYoungTurks).
Much the same changesover timeare seen in the otherthreepositive
typesof tie and in the negativetypes.Even the refinedpartitioninto
fiveblocksyieldssharpdiscrimination in the twoearlierperiods.Sampson
gave carefulinstructions for the retrospective choices, remindingeach
respondent of a salienteventmarkingthatperiod.But it is impossibleto
validatethemas faithful representations of perceptionsat T2 and T3.
We notedin the sectionon methodsthatconstraints on data collection,
randomfluctuation in social relations,
and thedifferential
maintenance and
concealmentof ties over time aive rise in a naturalway to "speckled"
blocks (bonds) thatare onlypartlyfilledwithchoices.Our modelallows
forchangesin the novices' ties over time,but our hypothesiswas that
such changeswouldgenerallyconfinethemselves to bonds and would not
affectzeroblocks(as definedby the blockmodelforperiodT4). If choices
forT2 and T3 are reallyjust surrogatesforT4, theyshouldfall within
the bonds specifiedby the blockmodelof T4; the same outcomeshould
also holdif theyare trulyfora different periodbut thesameblockmodelis
hypothesized.Accordingto the "top-two,bottom-two"cutoffsused in
applyingBLOCKER, among the 25 blocks in each image,about half are
zeroblocks.Of the 131 choiceson the fourpositivetypesof tie at T3, a
third(43 choices) do not coincidewithsome choice at T4. Of these43,
11 createdbonds withinzeroblocks.For negativetypesof tie, the cor-
responding figuresare: 128 choices,48 T3 discrepanciesfromT4, and 7
entriesfallingat T3 into blockswhichare zeroblocksat T4. To put this
anotherway,of the 2,183 entriesthatare zero in the eightdata matrices
forT4, the majority(59%o) are in zeroblocks;however,of the cells that
are zero at T4 but not at T3, four-fifths (80%o) are confinedat T3 to
blocksthatare bondsin theT4 blockmodel.If T2 is comparedto T4, the
figuresfor positiveties are: 130 choices,62 discrepancies;20 of the
latterappear in zeroblocksof the T4 blockmodel.For negativeties the
figuresare 121,69,and 17,respectively.
The implications seem simpleand agreein generalwiththe Newcomb

767
AmericanJournalof Sociology
analysisabove: theSampsonblockmodelis largelystableacrosstimewith
respectto bothblocksof menand images,and it emergedquicklyafterthe
populationwas constituted. of theseconclu-
To bringout the significance
sions,it is worthdrawinga contrastwithevolutionin a moreclassicsense.
There is littlesupportherefora naive borrowing of ideas frombiological
evolution,wherenaturalselectionis classicallyviewed as producingits
resultsgraduallyon a timescale of manygenerations (Mayr 1963; Levins
1968). Instead,the pictureconveyed,at least by these data, is one of
"saltationist"evolution,whereadaptationsarise rapidlyand withfinality
(Ford 1955). To take anotherparallel,this findingis in part consistent
withthe economists'view of a systemof actorswho maximizeprofitsor
and instantaneously,
utilityeffortlessly thoughwiththe crucialdifference
thatno explicitmaximizing behaviorhas yetbeen identified.

IMPLICATIONS: CONCRETE SOCIAL STRUCTURE

Sociologistshave long been torturedby theirinabilityto specifyclearly


the meaningof two fundamental terms,"role" and "position" (see the
excellentbriefreviewby Catton [1964, pp. 936-43]). All agree thatno
cogenttheoryof social structurecan dispensewith the conceptsthese
termstryto capture(one reasonMitchell[1969] is so soft-spoken about
networksis that theyseem conceptuallyremotefromrole and position).
Part of the trouble,we submit,is the lack of generallyapplicableopera-
tionalizations:no matterhow cogentthe prose discussionof role and
position(or cognateterms),the sense of insightfades as the writer(or
his reader) triesto apply themto variousconcretesocial structures.32
We now suggestthat the purposelyneutraltermsemployeduntilnow
-block and blockmodel-provideoperationaldefinitions forthe substan-
tive conceptsof role and position.We then suggesta way to interpret
concretesocial structuresin theseterms.We requirethreeprimitive terms:

(1) claim-the generictermforeachinstance of a tie fromonemem-


ber of a populationto another.Claimsmay be made eitherby the
members themselves (as in thebiomedicalscientists'
data) or by an out-
side observer(as in theBankWiringRoomdata).
(2) type of tie (abbreviated toft)-a networkof all claimsof a
specifiedtype.The meaningof "type"is initially(and explicitly)left
32 The study by Gross, Mason, and McEachern (1958) of the school superintendency
is one of the very few systematicapplicationsof such conceptsto concretesituations.
Because they deal with explicit officesin explicit formal organizations,they bypass
the derivationof positions from networksof relations (at the cost of other concep-
tual difficulties).In our terms,they moved directlyto the interpretation of types of
tie, using data on perceptionsof rights and duties among counterpartoffices; in
addition, they were aided by pooled data from differentconcrete populations that
are, by definition,parallel in organization.

768
Social StructurefromMultipleNetworks.I
open,thepostulatebeingthata sharedmeaningis accordedeach type
throughout thepopulation.
(3) population-the set of personsor otheractorswhoseclaimsare
reported.Its membership is the choiceof theinvestigator;
it neednot
be restricted
to a naturalenclaveor a setin whicheachactorknowsmost
of theothers.
The data are a set of networks:foreach toft,the claimsissuedby each
personregardingall othersin the populationare reportedas a binary
matrix.Most commonly the data will be sociometric,
and thenthe claims
cannotbe consideredto be acknowledged by the targets.With respectto
claims,the choice of jurisprudential languageis not accidental: Block-
modelsmay have naturalapplicationsto the analysisof complexlawsuits,
and theyseem particularly adapted to handlingcounterclaims and cross-
claimsin multiplepartylitigation(e.g., Lasa per L'Industriadel Marmo
Societaper Azioniv. SouthernBuilders,Inc., 45 F.R.D. 435 [W.D. Tenn.
1967], reversed,414 F.2d 143 [6th Cir. 1969]).

Blockmodelsas Roles amongPositions

A blockmodelis a hypothesis,a representation proposedfor the social


structurethat existsin the population'sclaims.Three termscan be used
to describea blockmodel:
(4) position-each of thesetsintowhichthepopulation
is partitioned
is a position.The technical
term"block"is a synonymforthissubstan-
tive concept.33
(5) bond-a nonzeroentryfromone positionto anotherin theimage
for a toft.
(6) image-thereport, in theformof a binarymatrix,
of thebonds
on a giventoftamongall positions.
a blockmodelis a simultaneous
By its definition, graphhomomorphism
in mathematicalterms(Heil and White 1974). Mapping the population
into positionsrequiresmappingeach data graph matrixsimultaneously
onto the correspondingimage.By the definition of a homomorphic map-
ping,thereis no bond fromone positionto anotherif and onlyif there
is no claim fromany memberof the firstpositionto any memberof the
otherposition.Thus,in thetermsused earlier,each imageis fullyspecified
fromits zeroblockswhenthepersonsin a data matrixare partitioned into
positions.

33 We agree with Catton (1964, p. 942) that there has been an evolution toward
clarity about these concepts. Our definitionscan be seen as operationalizingthose
proposedby Larsen and Catton (1962) aftertheirthoughtfulanalysisof the literature:
"position . . . location of a person or a categoryof personsin a set of social relation-
ships . . ."; "role . . . a patternof collectivelyheld expectationswhich defineappro-
priate behavior for persons in a given social position."

769
AmericanJournalof Sociology
The theoreticalcontentof a blockmodelhypothesis,then,is a trans-
formation of individualclaimsinto a statementof social bonds (i.e., ties
withinand amongpositions).Reflexivity (the existenceof "self-choices"
or unityentrieson the diagonalof a matrix) now emergesas a major
substantivequestionabout each position,ratherthan (as in sociometry)
a meretechnicality of presentation
fora claims matrix.We argue (and
furtherelaboratein Part II) that a bond has trulysocial content-is
"received"by the targetpositionas well as "sent" by the rowposition-
whetheror not thereis a reciprocalbond on that toftfromthe targetto
thesenderposition.
Not onlyhave we leftopen the actual semanticcontentof the claims
on a toft; we have also not specifiedthe generalnatureof tofts.One or
moreof the toftsmightreferto rights,anotherto duties,anotherto con-
tracts,anotherto evaluations,and so on throughthe long list of general
naturesproposedin variousanalysesof roles (once again, legal applica-
tionsare obvious). It seemsto us worthwhile to definepositionand role
withinas abstractand flexiblea culturalframework as possible,so that
theculturalcontentof thesocialstructure becomesa questionforempirical
researchratherthan a matterof definition. Sustainedanalysisof social
structure, we reiterate,requireskeepingthe two sides of the analysisas
distinctas possible.
Our next definitionis abstractin culturaltermsbut specificto the
populationand itspositionsin theblockmodel:
(7) role set-a pair of vectorsfor all blockmodel
images:one the
orderedset of rowsforthatposition,theotherthecolumnsforthatpo-
sition.34
Thus a role set is simplya statementof bonds sent and bonds received,
by a position,on all the tofts.Observethata position'sbond (or lack of
one) to itselfon each toftis partof its roleset. Thereis no need to report
theroleset foreach positionas a pair of vectorsbecause theset of images
fortheblockmodel is a completeinventory of therolesetsforall positions.

Arrays
Withoutfurther data, a blockmodelcannotbe validated.One simpleform
of validationis to establishstability,
or coherentchange,in theblockmodel
over time,as was done in the previoussection.Anotheris to test the
existenceof bondson a giventoftby meansof directobservations on inter-
34 This definitionis compatiblewith Merton's usage (1959, p. 369). A person's "status
set," in Merton's terms,includeshis positionsin distinctblockmodelsfor the different
populationsand contextsin which he participates.In a given blockmodelan individual
is indistinguishablefromother membersof his position; his individualityrests in his
unique "status set."

770
Social StructurefromMultipleNetworks.I
actionsor flowsof a suitablekind.A thirdinvolvesassessingthe homo-
geneityof personsin a positionwithregardto suitablecharacteristics.
The key wordsare "coherent"and "suitable": validationrequiressub-
stantivejudgmentas wellas further data. Even if,withminimalsubstan-
tive argument,one became convincedthat a blockmodelwas a valid
descriptionof a given concretesocial structure,therewould still be a
feelingof incompleteness. The remainingquestionis whetherthe roles
amongpositionscoherein some intelligibleway in an overall structure.
In Part II we proposemethodsforansweringthisquestionfromthe per-
spectiveof personsin variouspositions:in brief,we proposecriteriafor
rolestructures(cf.Nadel 1957).
One may also search for an overall interpretationof the blockmodel
froman observer's perspective.
Does everytofthave substantive interpreta-
tionsuch thatthe positionscan be seen as an array35intelligible
on each
toft?Perhapsthesimplestexamplewouldbe theblockmodelV,W; thatis,
theimages

1 0 11
and
1 1 0 1

Where the firsttofthas connotationsof deference,and the second of


domination,one interprets the firstpositionas the superiorof the second.
Or,one couldbeginwithno knowledgeof the imputedqualityof the tofts
but withattributedata forthe individualmembers:forexample,the fact
that membersof the second positionshared higherstandingin various
othercontextsand populations.Then one would inferthat the second
tofthad the qualityof deferenceand the firstconnoteddomination.
As we have seenin the case studies,one mustusuallyworkwitha wide
varietyof clues as to possiblestandingsof positionsand qualitiesof tofts.
As more case studies are analyzed,we can hope that regularitieswill
emerge,clarifying what sort of array is to be foundin what context.
"Context"mustincludewhattoftsone can and does elicitas data.
An arraycan take many forms,as the case studiesshow. The linear
hierarchyor partial orderis only one class of possibilities,thoughan
importantone. The array may indicatehierarchythoughall bonds are
reciprocated:see, for example,the Bank WiringRoom blockmodelon
like and antagonism, whichis a refinement of an E,F blockmodel,not a
V,W one. And the hierarchical arraysuggestedforNewcomb'sfraternity
is a refinement of a blockmodelwith both symmetric and asymmetric
bonds,namely,V,F.
35We are temptedto include "status" in this term,since many writersgive it con-
notationswe intendfor"array" (see Zelditch 1968); but the confusionwould be great
because "status" is commonlyrestrictedto the contextof vertical rankings.

771
AmericanJournalof Sociology
Classic balance theorysuggeststhe P,N array as the concretesocial
structureseen by an observer.Here thereis no questionof hierarchy;
instead,we have the familiaridea of rigidpolarizationinto two opposing
parties.
If one findstheblockmodelC,D on a pair of tofts,it suggeststhatthere
is no array(no overallsocialstructure on thatpopulation); thepopulation
fallsintotwounrelatedcliques,each of whichrespondsonlyon its "own"
toft.On the otherhand,the blockmodelX,Y suggestsa highlyintegrated
structure;forexample,let X mean"seeksas husbands"and Y mean"seeks
as wife,"and partitionthepopulationbysex.
There are 57 distinctpairs of imageson two positions;but we do not
speculateabout the significance of each one. The case studiesillustrate
how the investigator is usuallyled towarddiscriminating morethan two
positionsin his blockmodelhypothesis. Taxonomiesof multipositionblock-
models,36 and taxonomiesof tofts,mustbe evolvedas the empiricalbase
is widened;in addition,carefulattentionmustbe givento themanyrele-
vant conceptualanalyses in the literature.37 Some developmentsalong
theselineswillbe a byproduct of Part II.

CONCLUSION

Afterplacingour workin the contextof othernetworkanalysesof social


we definedoperationalconceptsand methodsand thenpresented
structure,
examples.The conceptof a block-a set of personsstructurally equivalent
withrespectto othersuch sets acrossseveraltypesof relation-has led to
the development of blockmodelsforvariedrelationsin fivequite different
populations.The examplesillustratethereliability and robustnessof both
the algorithms and theirresults.The blockmodelfor each examplewas
interpretable as an abstractpatternamong a few aggregateunits that
characterized the moredetailedinteraction amonga largerpopulationof
individuals.Independentevidenceavailable for fourof the populations
supportsthose blockmodels.Furthermore, in each case, the blockmodel
identified
regularities in relationalstructurethat had not been proposed
previously.In the two instancesforwhichdata were available,we dis-
coveredthatchangesover timewereconsistent withthe blockmodels. We
then interpreted familiarconceptsin role theorywithinour conceptual
framework. In Part II we propose ways to test for and describerole

36 Francois Lorrain and Joseph E. Schwartz have prepared taxonomies,fromthe ob-


server's viewpoint,of blockmodelson two positions for pairs of tofts,and on three
positions for one toft.
37 Breiger (1974a) showed how typologies of communitypower structuresmay be
translatedinto blockmodelhypothesesand thus actually become open to disconfirma-
tion.

772
Social StructurefromMultipleNetworks.I
structures, a more complexlevel of overall integrationthat need not
accompanyregularities of theconcretesortdescribedhere.
The approacheswe have describedare applicable to open network
populations(e.g., a contactnetworkcomposedof biomedicalresearchers)
as well as to the small,closed groupsthat are the traditionalpreserveof
sociometric investigators.Blockmodelapplicationsto a thirdclass of popu-
lations-those comprising groupsas varied as informalassociationsand
corporateboards of directors-are reportedelsewhere(Breiger et al.
1975).38This broadapplicability resultsfromtheabilityof blockmodels to
addressthe problematic issue of locatingthe boundariesof social interac-
tion.In our view,large,looselybounded,"open" populationsare particu-
larlyinteresting subjectsforblockmodeling. We suggestedthatour sample
of 28 scientists(fromGriffith's 107 respondents)reflectedthe structural
patternevident in the larger population.This claim deservesserious
scrutiny:ifit is upheldin futurework,networkanalysiscan be generalized
to thestudyofpopulationsfarlargerthanthosestudiedby earliernetwork
techniques.
Even ifevolutionary or discontinuous changesof structurecan be identi-
fied and congruencewith personaland culturalperceptionsestablished,
models of structureare not sufficient unto themselves.Eventuallyone
mustbe able to showhowconcretesocialprocessesand individualmanipu-
lationsshape and are shapedby structure. A naturalnextstep,then,is to
identifyhow flowsof information and othertransactions relateto images
and theirchange. One fundamentalproblemhere is that many social
settingsmay admit not just a singleequilibriumoutcome,but multiple
alternative equilibria,withwhichparticularequilibrium is reacheddepend-
ingin parton accidentsof earlyinteraction (comparetheearliercomments
on the Newcombfraternity data; comparealso Simon's [1957] modelof
Homans'stheoryof smallgroups).In turn,the interesting questionsmay
bear on what externalforcesmay cause a social structureto pass from
one equilibrium configurationto another.A numberof quite distincttradi-
tionsof modelbuildingseemto be converging on thissameset of problems,
includingSchelling'swork on "tipping"phenomena(1971) as well as
modelsof the geneticevolutionof social behavior(Boormanand Levitt
1973; Boorman1974).
Thereis an important limitation in theviewpointurgedthusfar.As we
stated in the very firstparagraph,connectivity propertiesin networks
receiveonly tangentialattention.The contrastbetweenweak and strong
tiesshouldbe a majorfactorin connectivity analysesforlargepopulations
38L. Groeneveld (1974) also reportssuch an analysis. He found blockmodels,for
three successive time periods, among 34 program units (containing both staff and
advisors) of the National Science Foundation. The five tofts used were objective
measures of mobility,authority,and interlockamong these supra-individual"nodes."

773
AmericanJournalof Sociology
(Rapoportand Horvath1961; Granovetter 1973; Boorman1975). Block-
modelsas developedthus far deal chieflywith strongties, whetherin
open-network populationsor in small enclaves.It may be harderto find
nontrivialblockmodels forweak ties (e.g., sociometric choicesin casual or
temporary groupssuchas thoseused in socialpsychological experiments).3
Connectivity has also been a major concernin sociometricstudiesof
closed populations.Overlapsamong cliques oftenhave been used (e.g.,
Coleman1961; Young and Larson 1965) to indicatestructure(although
thereis no agreed-upon or powerfulway to modeloverlaps).Blockmodels
imposea partitionof a populationintodisjointblocksor positions(except
for floaters-see AppendixA). Clique overlap is capturedonly by the
deviceof cross-cuttingpartitions(as in theBank WiringRoom). We focus
insteadon modelingoverlapsbetweenindependentnetworksand on ag-
gregatingmultiplenetworkswhileretainingtheirnetworkcharacter(the
incidenceof "holes" as zeroblocks).The practicaltradeoffis between
developingstatisticalmeasuresof clique overlap (e.g., Alba 1973) on the
one hand, and on the otherhand interpreting aggregatepatternsacross
multiplenetworks. New techniquesmay providenew vistas.Of particular
promiseis a new nonhierarchical clusteringmodelnow in the processof
development by Arabieand Shepard(Arabie and Shepard1973; Shepard
1974). The idea underlying thismodelis one of allowingarbitrarily given
proximitydata to definetheirown lattice of overlappingclustersin a
naturaland unconstrained way.
Finally,we need to enlargethe base of case studies.We particularly
need data fromwhichwe can learn how new recruitsto a population
mergewith,or change,the structurereflectedin a blockmodel.In the
terminology introduced earlier,whenwill a new recruitbecomea "crystal-
lizer,"and whena "floater"?When will his initialties of varioustypes
relegatehimto a blockthatis marginalto thestructure? Is theE pattern,
on specifiedtofts,the normalcontextfor indoctrination or socialization
into an open networkpopulation?Is the V,F array (on positive-and
negative-affectties,respectively)characteristic of clusteringand deference
withinstable small,closed groups?Only with a broaderempiricalbase
can generalregularitiesbe sought.

39 Observation of troops and packs of primates and other vertebratesis beginning


to result in systematic reports of network ties among individual animals (e.g.,
Struhsaker1967; Schaller 1972; see also the analysis of Struhsaker'sdata in Arabie
and Boorman 1973). If we were to find blockmodelsfittingsuch data, they would
suggestto us that theseties are of strength(durability,intensity)comparableto strong
ties among humans and that individualityis as pronouncedamong animals as among
people (see Wilson 1971, p. 459). Since individualityseems much more muted or
even nonexistentamong insectsand other invertebrates, we presumethat the network
models discussed in this paper are less appropriate for analyzing such populations.

774
Social StructurefromMultipleNetworks.I
APPENDIX A

Floatersand Crystallizers
In each case studyreportedin thispaper,BLOCKER yieldsa uniqueassign-
mentof mento blocks,subjectto twolimitations. First,an assignment is
termeduniqueeven thoughone or moremencan each be in two or more
blocks,forthegivenuniqueassignment of therestof thepopulation;how-
ever,these "floaters"mustbe specified.The extremeexampleis a man
withno tiesof any type,whocan therefore be in any blocksimplybecause
he is irrelevant;it wouldbe misleadingto createa differentoverallassign-
mentforeach of his possiblelocations.Second,the minimumnumberof
men any block can containmustbe specified;otherwise,forexample,a
vacuous solutionwithno men in some block would have to be counted.
The acronymforthisparameter,MIN, is used hereafter to designatethe
minimumblock size thoughtappropriatefor a given populationand
blockmodel.For convenience, the two specificationsare reportedtogether
hereforeachcase study:

Floaters: Man i between


StudyAuthor # Blocks MIN or among Blocks J
Maier,and Miller
Griffith, 4 5 23 between 1 and 3
19 and 24 among 2, 3, or 4
Any man in 4 may also be put in 3
Sampson (1) 3 3 None
(2) 5 3 None
Roethlisberger
et al. (1) 6 2 8 among blocks 1, 4, and 6
13 between 1 and 2
(2) 3 4 1, 9, 12, 13 between 1 and 2
8 between 1 and 3
Newcomb 3 4 2 between 1 and 2
5 between 1 and 3
White 3 3 8 and 16 among all threeblocks

For a detailedanalysisof assignments


and floaters,see Heil and White
(1974).

APPENDIX B

Attributesand Blocks in the Bank WiringGroup


In the dual mannerelaboratedby Breiger(1974b), reasonscan also be
regardedas "individuals,"the cross-tabulation transposed,and CONCOR
appliedagain. (First,though,in orderto equalize theinfluence
of different
wiremenon theresult,thecolumnof countsfora wiremanwas converted

775
AmericanJournalof Sociology
to percentages.)The partitionof reasonsinto threesets was (1 4 5 10)
(7 8 9 11) (2 3 6 12), withreasonsnumberedaccordingto the orderin
whichtheyare listedin thesource(table XXIX there).The sumof counts
in the originalmatrixbetweeneach of a set of reasonsand each of the
menon onesideofthesplitaboveis

16 23
3 106
33 43

wherethetop rowis the leftset of reasons,etc. The wiremenin the back


clique usuallycited reasonsfromthe secondset; the wiremenfromthe
front,almostnever.These reasons,7, 8, 9, 11, associatedwiththewiremen
W6, W7, W8, W9, include "Waitingfor Solderman"and "Waitingfor
Inspector,"but not "WaitingforTrucker."Roethlisberger and Dickson
(1939) notedthatthe "waiting"class of reasonsis highlysubjectiveand
"includesreasonswhichplaced the blame fordelays upon people instead
of upon things"(p. 431). This maximaltakingadvantageof discretion,
coupledwithresentment at equals and superiors(soldermenand inspec-
tors,not truckers),fitsin withthe tendencyto fightover windowswithin
the back clique (see fig. 5) and conforms with R-D's findingthat the
back clique "resentedany show of superiority morethan the othersdid
because theywerein the mostsubordinateposition"(p. 523). Reason 9
is somewhata technicality, as wiremanW9 was assignedto makingrepairs
(p. 432). Reason 11 is not discussedin thesource(see p. 432).
If the back clique's reasons tendedto fall into an imputed"resent-
ment"category,the frontclique's reasonsindicated"playingthe game"
or perhapsa "craftsman'sethic." In roughlyone-fifth of all instancesof
timeallowancesforthe frontclique, its membersgave no reasonat all,
whilethecomparablefigureforthe back clique was 87o. Mostly,the rea-
sonsof thefrontclique are objective,as R-D definethatterm(e.g., cable
reversalsweredetectedby inspectors'tests). However,theylabel "defec-
tive wire"and "defectivesolder"as subjectivereasons.It may be argued
thatthesereasonsare different fromthosein the "waiting"class (reasons
6, 7, 8) in thattheyrepresent a "professional"judgment.

REFERENCES

[NOTE.-The unpublished Ph.D. dissertationscited may be obtained in microfilm


copies from UniversityMicrofilms,Ann Arbor, Michigan.]
Alba, R. D. 1973. "A Graph-theoreticDefinitionof a SociometricClique." Journal
of MathematicalSociology 3 (1): 113-26.
Arabie, P., and S. A. Boorman. 1973. "MultidimensionalScaling of Measures of
Distance betweenPartitions."Journal of MathematicalPsychology10 (May): 148-
203.

776
Social StructurefromMultipleNetworks.I
Arabie, P., and R. N. Shepard. 1973. "Representationof Similaritiesas Additive
Combinations of Discrete, Overlapping Properties." Paper read at Mathematical
PsychologyMeetings,Montreal,August 1973.
Barnes, J. A. 1972. "Social Networks." Addison-Wesley Module Publications 26.
Reading, Mass.: Addison-Wesley.
Bjerstedt,A. 1956. Interpretationsof SociometricChoice Status. Lund: Gleerup.
Boissevain,J. 1973. Prefaceto NetworkAnalysis: Studies in Human Interaction,edited
by J. Boissevain and J. C. Mitchell. The Hague: Mouton.
Bonacich,P. 1972. "Techniques forAnalyzingOverlappingMemberships."Pp. 176-85 in
Sociological Methodology 1972, edited by Herbert L. Costner. San Francisco:
Jossey-Bass.
Boorman, S. A. 1974. "Island Models for Takeover by a Social Trait Facing a
Frequency-dependentSelection Barrier in a Mendelian Population." Proceedingsof
the National Academy of Sciences, USA 71 (11): 2103-7.
. 1975. "A Combinatorial Optimization Model for Transmission of Job In-
formationthrough Contact Networks." Bell Journal of Economics 6 (Spring):
216-49.
Boorman, S. A., and P. R. Levitt. 1973. "A Frequency-dependentNatural Selection
Model for the Evolution of Social Cooperation Networks." Proceedings of the
National Academy of Sciences, USA 70 (1): 187-89.
Boorman, S. A., and H. C. White. 1976. "Social StructurefromMultiple Networks.
II. Role Structures."AmericanJournal of Sociology,in press.
Breiger,R. L. 1974a. "Blockmodel Analysis of Elite Structures."Paper presentedat
the M.S.S.B. Conferenceon Elites, Harvard University,October 1974.
. 1974b. "The Duality of Persons and Groups." Social Forces 53 (December):
181-90.
. 1976. "Career Attributesand Network Structure: A Blockmodel Study of a
Biomedical Research Specialty." AmericanSociological Review, vol. 41 (February),
in press.
Breiger,R. L., S. A. Boorman, and P. Arabie. 1975. "An Algorithmfor Clustering
Relational Data with Applicationsto Social Network Analysisand Comparison with
Multidimensional Scaling." Journal of Mathematical Psychology 12 (August):
328-83.
Catton, W. R., Jr. 1964. "The Development of Sociological Thought." Pp. 912-50 in
Handbook of Modern Sociology, edited by Robert E. L. Faris. Chicago: Rand
McNally.
Cohen, M. D., and J. G. March. 1974. Leadership and Ambiguity: The American
College President.New York: McGraw-Hill.
Coleman, J. S. 1958. "Relational Analysis: The Study of Social Organizationswith
Survey Methods." Human Organization17 (Winter): 28-36.
1961. The AdolescentSociety. New York: Free Press.
1964. Introductionto Mathematical Sociology. New York: Free Press.
Davis, J. A. 1970. "Clusteringand Hierarchyin InterpersonalRelations: Testing Two
Graph Theoretical Models on 742 Sociograms." American Sociological Review 35
(October): 843-52.
Ekeh, P. 1974. Social Exchange Theory. Cambridge,Mass.: Harvard UniversityPress.
Flament, C. 1963. Applications of Graph Theory to Group Structure. Englewood
Cliffs,N.J.: Prentice-Hall.
Ford, E. B. 1955. "Rapid Evolution and the Conditions Which Make It Possible."
Cold SpringHarbor Symposia on QuantitativeBiology 20: 230-38.
Ford, L. R., Jr., and D. R. Fulkerson. 1962. Flows in Networks. Princeton,N.J.:
Princeton UniversityPress.
Fortes, M. 1945. The Dynamics of Clanship among the Tallensi. Oxford: Oxford
UniversityPress.
Friedell, M. F. 1967. "Organizationsas Semilattices."American Sociological Review
32 (February): 46-54.
Glanzer,M., and R. Glaser. 1959. "Techniques for the Study of Group Structureand

777
AmericanJournalof Sociology
Behavior: I. Analysisof Structure."PsychologicalBulletin 56 (September): 317-32.
. 1961. "Techniques for the Study of Group Structureand Behavior: II. Em-
pirical Studies of the Effectsof Structurein Small Groups." PsychologicalBulletin
58 (January): 1-27.
Granovetter,M. S. 1973. "The Strengthof Weak Ties." AmericanJournalof Sociology
68 (May): 1360-80.
. 1974. Gettinga Job: A Study of Contacts and Careers. Cambridge,Mass.:
Harvard UniversityPress.
Green,H. A. J. 1964. Aggregationin Economic Analysis. Princeton,N.J.: Princeton
UniversityPress.
Griffith, B. C., V. L. Maier, and A. J. Miller. 1973. "Describing Communication
Networksthroughthe Use of Matrix-basedMeasures." Mimeographed.Philadelphia:
Drexel University,Graduate School of Library Science.
Groeneveld,L. P. 1974. "Bureaucracy and the Organizationof Science: A Structural
Analysis of the National Science Foundation." Ph.D. dissertation,Department of
Sociology, Universityof Indiana.
Gross, N., W. S. Mason, and A. W. McEachern. 1958. Explorationsin Role Analysis.
New York: Wiley.
Harary, F., R. Norman, and D. Cartwright.1965. StructuralModels: An Introduction
to the Theory of Directed Graphs. New York: Wiley.
Heil, G. H., and H. C. White. 1974. "An Algorithmfor ConstructingHomomorphisms
of Multiple Graphs." Unpublished paper, Department of Sociology, Harvard Uni-
versity.
Hilton, P. J., and S. Wylie. 1960. Homology Theory: An Introductionto Algebraic
Topology. New York: Cambridge University Press.
Holland, P. W., and S. Leinhardt. 1973. "The StructuralImplicationsof Measurement
Error in Sociometry."Journal of MathematicalSociology 3 (January): 1-27.
. 1976. "Local Structurein Social Networks."In Sociological Methodology1976,
edited by David R. Heise. San Francisco: Jossey-Bass,in press.
Homans, G. C. 1950. The Human Group. New York: Harcourt, Brace & World.
Hunter,J. E., and R. L. Shotland. 1974. "TreatingData Collectedby the 'Small World'
Method as a Markov Process." Social Forces 52 (March): 321-32.
Jencks, C., et al. 1972. Inequality: A Reassessment of the Effect of Family and
Schooling in America. New York: Harper & Row.
Kapferer, B. 1972. Strategy and Transaction in an African Factory. Manchester:
Manchester UniversityPress.
Kruskal, J. B. 1964a. "MultidimensionalScaling by OptimizingGoodness of Fit to a
Nonmetric Hypothesis." Psychometrika29 (March): 1-28.
. 1964b. "Nonmetric MultidimensionalScaling: A Numerical Method." Psy-
chometrika29 (June): 115-59.
Larsen, 0. N., and W. R. Catton, Jr. 1962. Concept and Sociology. New York: Harper
& Row.
Laumann, E. O., and F. U. Pappi. 1973. "New Directionsin the Study of Community
Elites." American Sociological Review 38 (April): 212-30.
Lee, N. Howell. 1969. The Search for an Abortionist.Chicago: Universityof Chicago
Press.
Leijonhufvud,A. 1968. On Keynesian Economics and the Economics of Keynes: A
Study in Monetary Theory. London: Oxford UniversityPress.
Levine, J. H. 1972. "The Sphere of Influence." American Sociological Review 37
(February): 14-27.
Levins, R. 1968. Evolution in Changing Environments.Princeton, N.J.: Princeton
UniversityPress.
Lorrain, F. P., and H. C. White. 1971. "Structural Equivalence of Individuals in
Social Networks."Journal of MathematicalSociology 1 (January): 49-80.
McFarland, D. D., and D. J. Brown. 1973. "Social Distance as a Metric: A Systematic
Introductionto Smallest Space Analysis." Pp. 213-53 in Bonds of Pluralism: The

778
Social StructurefromMultipleNetworks.I
Form and Substance of Urban Social Networks,edited by Edward 0. Lauimann.
New York: Wiley.
McQuitty,L. L., and J. A. Clark. 1968. "Clusters from Iterative IntercolumnarCor-
relational Analysis." Educational and Psychological Measurement 28 (Summer):
211-38.
MacRae, D. 1960. "Direct Factor Analysis of Sociometric Data." Sociometry 23
(December): 360-71.
Marschak, J., and R. Radner. 1972. Economic Theory of Teams. New Haven, Conn.:
Yale UniversityPress.
Mayer, A. C. 1960. Caste and Kinship in Central India: A Village and Its Region.
Berkeley: Universityof California Press.
Mayr, E. 1963. Animal Species and Evolution. Cambridge,Mass.: Harvard University
Press.
Menzel, H. 1962. "Planned and Unplanned ScientificCommunication."Pp. 417-41 in
The Sociology of Science,edited by Bernard Barber and Walter Hirsch. New York:
Free Press.
Merton,R. K. 1959. Social Theoryand Social Structure.2d ed. Glencoe,Ill.: Free Press.
Milgram,S. 1967. "The Small World Problem." PsychologyToday 1 (May): 61-67.
Mitchell,J. C., ed. 1969. Social Networks and Urban Situations. Manchester: Man-
chesterUniversityPress, for Institute of Social Research, Zambia.
Mullins, N. C. 1972. "The Structureof an Elite: The Public Advisory Structureof
the Public Health Service." Science Studies 2 (January): 3-29.
Nadel, S. F. 1957. The Theory of Social Structure.London: Cohen & West.
Newcomb, T. M. 1961. The Acquaintance Process. New York: Holt, Rinehart &
Winston.
Nordlie, P. G. 1958. "A LongitudinalStudy of InterpersonalAttractionin a Natural
Group Setting." Ph.D. dissertation,Universityof Michigan.
Pool, I. de S., and M. Kochen. 1958. Unpublished manuscripton contact networks.
Cambridge: MassachusettsInstituteof Technology.
Rapoport, A. 1963. "Mathematical Models of Social Interaction." Pp. 493-579 in
Handbook of MathematicalPsychology,Vol. II, edited by R. D. Luce, R. R. Bush,
and E. Galanter. New York: Wiley.
Rapoport, A., and J. W. Horvath. 1961. "A Study of a Large Sociogram." Behavioral
Science 6 (October): 280-85.
Roethlisberger, F. J., and W. J. Dickson. 1939. Management and the Worker. Cam-
bridge,Mass.: Harvard UniversityPress.
Sampson, S. F. 1969. "Crisis in a Cloister."Ph.D. dissertation,Cornell University.
Savage, I. R., and K. A. Deutsch. 1960. "A StatisticalModel of the Gross Analysis of
Transaction Flows." Econometrica 28 (July): 551-72.
Sayles, L. R. 1958. Behavior of IndustrialWork Groups. New York: Wiley.
Schaller,G. B. 1972. The SerengetiLion. Chicago: Universityof Chicago Press.
Schelling,T. C. 1960. The Strategyof Conflict.Cambridge,Mass.: Harvard University
Press.
. 1971. "Dynamic Models of Segregation."Journal of Mathematical Sociology
1 (July): 143-86.
Schwartz, J. E. 1974. "An Examination of Iterative Blocking Algorithms for
Sociometry." Unpublished paper, Department of Sociology, Harvard University.
Shepard, R. N. 1962. "The Analysis of Proximities:MultidimensionalScaling with an
Unknown Distance Function." Psychometrika27 (1): 125-40; 27 (2): 219-46.
. 1974. "Representationof Structurein SimilarityData: Problems and Pros-
pects." Psychometrika39 (December): 373-421.
Simmel,G. 1950. The Sociology of Georg Simmel. Translated and edited by Kurt H.
Wolff.New York: Free Press.
. 1955. Conflict and the Web of Group-Affiliations. Translated by Kurt H.
Wolffand ReinhardtBendix. New York: Free Press.
Simon, H. A. 1957. Models of Man. New York: Wiley.

779
AmericanJournalof Sociology
Struhsaker,T. T. 1967. "Social Structure among Vervet Monkeys (Cercopithecus
aethiops)." Behaviour 29 (1): 83-121.
Suppes, P. 1960. AxiomaticSet Theory. New York: Van Nostrand.
Szasz, G. 1963. Introductionto Lattice Theory. New York: Academic Press.
White, H. C. 1960. "Management Conflictover Research and Development: A Case
Study." Ph.D. dissertation,Department of Sociology, Princeton University.
. 1961. "Management Conflictand SociometricStructure."AmericanJournal of
Sociology 67 (September): 185-99.
1963. An Anatomy of Kinship. Englewood Cliffs,N.J.: Prentice-Hall.
1970a. Chains of Opportunity.Cambridge,Mass.: Harvard UniversityPress.
1970b. "Search Parameters for the Small World Problem." Social Forces 49
(December): 259-64.
. 1973. "Everyday Life in Stochastic Networks." Sociological Inquiry 43
(Spring): 43-49.
. 1974a. "Models for Interrelated Roles from Multiple Networks in Small
Populations." Pp. 52-102 in Proceedings of the Conferenceon the Application of
UndergraduateMathematicsin the Engineering,Life, Managerial and Social Sciences,
edited by P. J. Knopp and G. H. Meyer. Atlanta: Georgia Instituteof Technology.
. 1974b. "Null Probabilitiesfor Blockmodels of Social Networks." Unpublished
paper, Departmentof Sociology,Harvard University.
Wiese, L. von. 1941. Sociology. Edited and annotated by F. H. Mueller. New York:
Oskar Priest.
Williamson,0. E. 1970. Corporate Control and Business Behavior. Englewood Cliffs,
N.J.: Prentice-Hall.
Wilson,E. 0. 1971. The Insect Societies. Cambridge,Mass.: Harvard UniversityPress.
Young, R. C., and 0. F. Larson. 1965. "A New Approach to CommunityStructure."
American Sociological Review 30 (December): 926-34.
Zelditch, M., Jr. 1968. "Status, Social." International Encyclopedia of the Social
Sciences 13:250-56.
Znaniecki,F. 1940. The Social Role of the Man of Knowledge. New York: Columbia
UniversityPress.

780