A Model of Robust Position in Social Networks
A Model of Robust Position in Social Networks
A Model of Robust Position in Social Networks
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
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 [email protected].
The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to
American Journal of Sociology.
http://www.jstor.org
Matthew S. Bothner
Cornell University
Harrison C. White
Columbia University
What are features of robust rather than fragile social positions? Various
approaches have long equated robust positions with diversification and
fragile, vulnerable ones with concentration. For instance, the early treat-
1
We have benefited from the advice of Peter Bearman, Ron Breiger, Ron Burt, Tim
Conley, James Evans, Jon Kleinberg, John Levi Martin, Mark Mizruchi, and John
Padgett, as well as seminar participants at the University of Chicago, Columbia Uni-
versity, the Tuck School of Business, and the Yale School of Management. Financial
support from the University of Chicago Booth School of Business and from the Ewing
Marion Kauffman Foundation is gratefully acknowledged. We thank Val Burris for
generously making data available on the sociology PhD exchange network. Direct
correspondence to Matthew Bothner, Department of Sociology, 364 Uris Hall, Cornell
University, Ithaca, New York 14853-7601. E-mail: [email protected]
944
THEORETICAL ANTECEDENTS
Our starting point is Abbott’s (2001) observation that an individual’s
status may turn more significantly on the range of his or her current
options than on a history of affiliations with high-status institutions. Ac-
cording to this view, “It should not be assumed that the ultimate elite
career is one that hooks together all the perfect elite trajectories. . . . The
945
most elite individual may be that person who maintains the largest num-
ber of possible future trajectories that s/he could jump onto” (Abbott 2001,
p. 247).
This insight carries several interrelated implications. One is that, al-
though many sociologists have long viewed status largely as the product
of stable traits (Berger et al. 1977), status may instead rise and fall in
response to shifts in the durability of an actor’s location in social structure.
Another implication is that much of the growth and decline in status we
observe in social life may well have its most fundamental antecedents in
audience members’ taste for robustness and distaste for fragility. Abbott’s
claim points directly to the appealing theoretical possibility that possession
of durable social support invites prestige-conferring reactions, whereas
occupants of fragile social positions are seen as illegitimate and therefore
attract responses from others that attenuate their social standing. We
believe that deference accruing to durably situated actors, as well as
disdain for those in fragile roles (evident in diatribes like “he’s only hang-
ing on by a thread—cut him loose”) should figure centrally in our por-
trayals of status-based processes.
Cases of fragility’s corrosive effects also appear as far back as Weber’s
(1978) comparison of two distinct types of charismatic leaders. The first
(most familiar) type of charismatic leader is the extraordinary figure whose
military exploits, intellectual powers, or miraculous works elicit the def-
erence of a broad community of followers. This individual, standing on
a wide base of support, inhabits what for us is a robust position. Con-
versely, for Weber, the second type is the person whom the first type
eventually (with age) consecrates as his or her successor. The second type’s
social standing therefore rests perniciously on the sponsorship of the orig-
inal charismatic leader. Unlike the “pure” charisma of the predecessor,
mere “hereditary charisma” or “lineage charisma” (Weber 1978, pp. 248,
1135–41) marks the successor’s role—a role not nearly as legitimate as
that of the first type. This second type is fragilely situated, having been
chosen by the original, now-deceased leader, and frequently evokes sus-
picion and conflict among prospective followers—those whose allegiance
the first type garnered with relative ease. Weber’s distinction is important
because it brings into focus the unfavorable perceptions that encircle and
damage occupants of fragile social roles. With fragility often come re-
actions that erode legitimacy, such as Quakers’ epithetical descriptions of
successors as “hirelings” (Weber 1978, p. 249).
Whyte’s (1993) celebrated ethnography of Boston’s North End, espe-
cially his analysis of the Nortons street gang, offers further examples. In
particular, two components of Whyte’s qualitative account—his descrip-
tion of the Nortons’ bowling match and its competitive aftermath and
his portrayal of the realignment of the gang after its leader fell from
946
947
Fig. 1.—Social structure of the Nortons street gang, as hand illustrated by Whyte (1993,
p. 13).
In the future, Doc was not nearly as able to prop up Long John, how-
ever. And this fact takes us to our second point—namely, that an indi-
vidual’s position becomes more fragile if that individual’s source of social
support becomes less robust. Doc’s fall—and, due to their coupling in the
network, Long John’s heightened fragility—resulted from his failed foray
into the North End’s political ambit.2 The Nortons grew disillusioned
with Doc and underwent sharp social-structural changes after the summer
of 1938 (Whyte 1993, pp. 42–51). Although there were several exits, entries,
and shifts in stayers’ positions, the most jugular realignments, particularly
for Long John, followed Doc’s repositioning. When Doc’s position became
less robust—Doc had lost much of the support of the Nortons and had
grown dependent on Spongi, a new player, not included in figure 1—
Long John became nearly an isolate.
In his summary of Long John’s dilemma, Whyte (1993, p. 45) high-
lighted a central component of the model of fragility we develop, specif-
ically, that the fragility of an actor’s position is exacerbated when that
actor’s source of social support—in this case, Danny, Mike, and especially
Doc—grows more vulnerable:
2
Although this second point concerns our conceptualization, rather than the expected
consequences, of fragility, it does imply a corollary prediction—namely, that Alec would
have been even more likely to challenge Long John in the bowling alleys had the social
positions occupied by Long John’s supporters grown less robust.
948
Long John divided his time between Spongi’s and the Norton Street corner.
The realignment left him in a vulnerable position. There were two groups
that hung around Spongi’s “joint”: the inner circle and the hangers-on. Spongi
included his brother, Danny, Doc, and two others in the inner circle. When
he went for “coffee-ands,” for a drive, or to the movies, he would invite them
to accompany him. He did not include Long John in his invitations, so Long
John was excluded from the inner circle. Without the support of Doc, Danny,
and Mike, he had no standing among the boys who remained on Norton
Street, and he did not know where to go. (Whyte 1993, p. 45)
MEASUREMENT
The distinctiveness of social positions resembling the one occupied by
Long John brings us to a more formal description of our measurement
of fragility and robustness. We start with the Herfindahl index to account
for concentration and then bring in the recursiveness of Bonacich’s (1987)
method to account for coupling.
The Herfindahl index has long been used to assess industry concen-
tration, where it is written as the sum of firms’ squared market shares
(Schmalensee 1989, pp. 966–67; McLean and Padgett 1997). It has also
been used to quantify religious homogeneity within a geographical area
(Ellison, Burr, and McCall 1997). We take it as our point of departure
for measuring fragility because it aptly captures the degree to which one
node in the network depends on a narrow base of alters rather than facing
diversification across others. More specifically, if we consider a hypo-
thetical relational matrix X , where X p [xij ] , xii p 0, and xij denotes the
flow of respect (endorsements, recognition, esteem, deference, liking, sup-
port, precedence, or honor) directed from individual j to i, then we can
express the Herfindahl index for members of the social network sum-
marized by X very simply as
冘
n⫺1
Hi p dij , (1)
jp1
[冘 ]
2
xij
dij p n⫺1 . (2)
jp1 xij
When applied to relational matrices like X , the Herfindahl index thus
involves these straightforward steps: divide each entry through by its row
sum, square the resulting proportions, and then collect the row sums for
the new matrix of squared proportions. If Hi equals one, i then receives
all of his or her respect or social support from just one other actor in the
network. Under this scenario, i is completely dependent. If, by contrast,
949
the index for a given actor equals 1/ (n ⫺ 1), that actor is then minimally
dependent, and his or her position is consequently far less fragile. The
Herfindahl index captures the concentration of sources of incoming flows
and therefore serves as a constructive baseline for measuring the fragility
of social positions.
Used by itself, however, the Herfindahl index has limitations. When
applied to a given social network, it considers only the distribution of an
actor’s immediate ties, making no allowance for the identities and net-
works of that actor’s contacts to exert influence. The use of Herfindahl’s
measure thus assumes complete decoupling by construction because it
contains no provision for network spillovers. Although full decoupling—
that is, the complete absence of network spillovers—may accurately de-
scribe some social structures, it likely mischaracterizes most networks of
interest to sociologists. Consequently, to build coupling into our measure
of fragility, we turn to a salient feature of Bonacich’s (1987) status mea-
sure, whose structure we now summarize as the backcloth of our ap-
proach.
Using again a hypothetical relational matrix X as a reference point,
Bonacich (1987, pp. 1172–75) recursively defines the status of i as
Si (a, b) p 冘(j
a ⫹ bSj) xij , (3)
950
冘(
follows, where Fi represents the fragility of actor i:
Fi (a, b) p a ⫹ bFj) dij . (4)
j
The core elements in equation (4) are dij and b. Recall that, in discussing
Herfindahl’s index, we set dij p [xij / 冘 jp1 xij ] 2 in equation (2), so that dij
n⫺1
冘
⬁
For equation (5) to converge, b ! l⫺1 must hold, where l equals the largest
normed eigenvalue of D.3 Given this constraint, we write the relationship
between b and l as follows, where c denotes a coupling coefficient:
b p cl⫺1. (6)
Thus, for c p 0, full decoupling is taken to characterize the relevant
network, and fragility therefore correlates perfectly with the Herfindahl
index. Under c p 0, in other words, fragility reduces to concentration.
Conversely, as c r 1, each actor’s fragility score Fi is increasingly affected
by the fragility score Fj of those on whom i depends.
The substantive interpretation of c will of course vary across different
kinds of networks. When analyzing a network through which intangible
resources circulate, such as the Nortons street gang, a researcher would
likely select a high value of c when the members of the network attach
significance to instances in which individuals are “propped up” by others
3
The solution (readily computed with standard packages) to eq. (5) in matrix form is
F (a, b) p a (I ⫺ bD)⫺1 D1, where I is an identity matrix (cf. Bonacich 1987, eq. [4]).
951
R (a, b) p 1 ⫺ a 冘 ⬁
kp0
b kD k⫹11. (7)
4
Although it may be more accurate to refer to our measure as “structural robustness”
or “relational robustness” to distinguish it from the separate ideational component of
robustness (see Moody and White’s [2003, esp. p. 104] discussion), we have opted for
the simplest label. In addition, we choose the scaling constant a so that the sum of
squared fragility scores equals the number of actors in the matrix (cf. Bonacich 1987,
p. 1173). We could also of course substitute in place of 1, as the minuend, a column
vector of constants equal to the largest fragility score plus the smallest fragility score,
in order to linearly convert fragility scores into robustness scores. Using the maximum
plus the minimum has the benefit of making the robustness distribution a mirror image
of the fragility distribution, with the same endpoints and variance.
952
953
5
However, in panel models that supplement those we subsequently discuss, we also
explored results using five distinct levels of c —in particular, c 苸 {0, .25, .5, .75, .99}. We
took this approach to evaluate our assertion that coupling is a central component of
fragility, while also assessing our anticipation of a negative effect of fraternity members’
fragility levels.
6
Although our panel models predict the growth (rather than level) of future status as
a function of current fragility net of current status, we also simulated versions of
Newcomb’s week 1 matrix to ensure that far weaker cross-sectional correlations be-
tween status and fragility are plausible. Using these alternative matrices (again with
17 nodes, each ranking one another from 1 to 16 followed by reverse coding), we not
only observed correlations that were virtually at zero, but we also viewed individuals
who were much more sharply positioned on the off diagonal than 13 and 16 in fig. 2
(and who offer leads for future research on fragility and status). More specifically, our
simulations (not reported but available on request) brought relief to “drones” (i.e.,
individuals receiving lukewarm support from many, while lighting up no one) as in-
habitants of low-status, low-fragility roles, while also highlighting “polarizers” as oc-
cupants of high-status, high-fragility roles. We believe that a promising line of future
research would involve investigating contextual conditions that give rise to particularly
low correlations between fragility and status. One such context may be deeply contested
settings where factions emerge and where polarizers deploy drones as their trusted
between-group emissaries.
954
TABLE 1
A Relational Matrix X1 Recording Likability Levels in Week 1
of Newcomb’s Study
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17)
(1) 0 9 4 4 3 10 2 8 11 15 5 2 16 3 1 9 8
(2) 10 0 7 16 7 4 13 9 1 1 10 6 2 12 8 6 2
(3) 5 1 0 2 6 6 6 1 9 8 13 15 1 9 13 2 7
(4) 6 16 10 0 10 14 14 10 3 3 9 11 10 11 9 14 15
(5) 7 6 9 3 0 2 1 7 4 6 11 12 13 4 16 4 13
(6) 13 5 6 13 1 0 9 16 6 13 3 3 15 8 4 1 6
(7) 4 15 8 14 5 7 0 3 13 14 8 10 5 15 6 3 12
(8) 3 3 2 1 13 15 11 0 2 7 1 4 3 1 5 5 5
(9) 2 7 11 5 12 13 8 6 0 10 14 7 4 16 11 16 14
(10) 1 4 12 10 11 1 7 14 10 0 4 13 9 14 15 8 10
(11) 14 2 15 11 15 3 12 15 16 2 0 14 11 5 14 15 9
(12) 8 11 16 8 14 12 15 12 15 9 15 0 6 10 12 11 16
(13) 16 10 1 9 4 16 3 13 8 5 7 1 0 2 7 7 11
(14) 12 8 5 6 2 5 5 2 12 4 2 9 7 0 2 10 1
(15) 9 12 13 7 9 8 4 5 5 16 6 8 14 13 0 12 3
(16) 11 14 3 12 8 9 10 4 7 11 12 5 8 6 3 0 4
(17) 15 13 14 15 16 11 16 11 14 12 16 16 12 7 10 13 0
base of support (as did Nutsy—Long John’s close counterpart in fig. 1).
More generally, differences in fragility among status equals is not unlike
a case of two towers of equal height that nevertheless stand on markedly
dissimilar foundations. We further explore differences in social founda-
tions in figures 3 and 4—where we depict the positions of individuals 13
and 16, respectively—to highlight distinctive features of our approach.7
7
An alternative, although broadly related, theoretical approach is brought forward in
Mizruchi et al.’s (1986) disaggregation of status scores into derived (generated by alters)
and reflected (generated by ego) components. Although their concept of derived status
bears some resemblance to our concept of fragility, our approach differs from theirs
in that we construct a separate measure rather than present a decomposition of status.
955
Fig. 2.—Status versus fragility in week 1 of Newcomb’s study. Status and fragility scores
were computed from the relational matrix X1 shown in table 1, the first of 15 weekly matrices
Xt used to compute network measures in our panel models. Numbers of individuals’ rows
in Newcomb’s matrices label each observation. Lines correspond to the status and fragility
scores of individuals 13 and 16.
956
8
Code for generating figs. 3 and 4 is available on request. We made line thickness a
957
linear function of entries in D1 and for added emphasis nonlinearly transformed alters’
fragility levels when selecting the radii of their surrounding circles.
958
959
ability, social skill (Leifer 1988), and all other stable traits that plausibly
channel sorting into fragile social positions (cf. Gould 2002, pp. 1143–44).
Using a fixed-effects specification thus eliminates all between-person var-
iation, so that the coefficients reflect within-person effects of fragility on
growth in status. Additionally, as denoted by the indicators tt⫹1 , we adjust
for all types of overall temporal heterogeneity, such as aggregate levels
of reciprocity and transitivity, the eigenvalue from which fragility is com-
puted, and all other global properties of the network (Doreian et al. 1996;
Moody et al. 2005). Since tt⫹1 enter jointly with m i, the effect of age is
also fully accounted for. Correlations and descriptive statistics for vari-
ables in our analyses are included in table 2.9
Table 3 presents six models predicting future status growth in New-
comb’s fraternity. We present one-tailed tests for all continuous covariates,
given the clarity of our expectations about the directions of the effects.
We also present standardized regression coefficients because of their use-
fulness in conveying effect magnitudes in models containing network-
related covariates. Model 1 contains only lagged status and fixed effects
for individuals and periods as a baseline. Inspection of the individual
fixed effects (not reported to conserve space) confirms that variation in
fraternity members’ average levels of status growth is clearly pronounced,
reinforcing the merits of the within-person estimator. Consistent with our
expectations about the importance of coupling, we find in model 2 that
the effect of fragility (as Herfindahl’s measure, where complete decoupling
is assumed) is not discernibly different from zero.
In model 3, we enter FitFcp0 and FitFcp.99 jointly. Whereas FitFcp0 reduces
to the Herfindahl index, FitFcp.99 assumes coupling, with adjacent individ-
uals’ positions influencing the fragility of a given individual’s position in
social structure. In model 3, the coefficient on FitFcp0 stays insignificant,
whereas FitFcp.99 exerts a discernibly negative effect. Through regressing
status growth on FitFcp0 and FitFcp.99 jointly, the estimate on FitFcp.99 reflects
the variation that is orthogonal to FitFcp0. Thus, the persistence of the effect
of FitFcp.99, where coupling is taken into account, offers preliminary evi-
dence to suggest that, at least in the present empirical setting, network
spillovers may matter more than immediate dependencies. When esti-
mating other models (available upon request), with c equal to .25, .5, and
.75, we also found that fragility is significant only at the highest possible
level of coupling, c p .99. This set of outcomes is important because it
suggests that taking coupling into account is a necessary part of under-
standing the link between positional fragility and dynamics of prestige.
Model 4 confirms that the negative effect of fragility measured with cou-
9
We report within-person correlations because our estimator focuses only on within-
person variance.
961
TABLE 2
Correlations and Descriptive Statistics for Variables in the Analysis of
Newcomb’s Fraternity
pling stays significant when its counterpart is excluded from the set of
covariates.
In models 5 and 6, we take two additional steps to assess further our
main finding that fragility negatively affects status growth, before turning
to an application of our fragility measure in a network of economic flows
unrelated to status. First, we guard against the possibility that behavioral
correlates of fragility drive the effect we observe. It is easy to imagine
that (at least some) occupants of fragile roles tend to act sycophantically,
whereas occupants of robust positions are more prone to feather displays
and, in general, carry themselves with considerable aplomb—much to
their advantage, especially in a fraternity-style setting. We have argued,
building on prior sociological theory, that fragility undermines legitimacy
and is therefore status eroding, whereas robustness solidifies legitimacy
and is consequently status enhancing. We nonetheless address the coun-
terpossibility that locations on an axis marked by fragility and robustness
are in fact undetectable by others in a network, making them by definition
inconsequential, in a direct sense, for processes of growth and decline in
status. If this alternative view is descriptive, then an easily tracked be-
havioral correlate of fragility, such as sycophantic conduct (cf. Burt 1976,
pp. 104–9) may well fuel the effect we observe. Although the fixed effects
m i absorb time-constant intrinsic propensities to behave subserviently,
conduct of this type may have a time-changing component as well.
We constructed a weighted asymmetry measure to proxy individuals’
shifting propensities to act obsequiously by working from the premise that
a focal actor is sycophantic to the degree that he sends out more positive
affect in a given exchange than he receives. Our measure thus captures
imbalances in the flows that mark a chosen actor’s set of implicit trades,
so that the individual meting out lots of favorable rankings to those who
send back unfavorable rankings gets a large value. Using our reverse-
962
TABLE 3
Standardized Regression Coefficients for Fixed-Effects Models of Status
Growth in Newcomb’s Fraternity
1 2 3 4 5 6
Status . . . . . . . . . ⫺1.7583 ⫺1.8841 ⫺1.8630 ⫺1.9317 ⫺2.6064 ⫺2.7008
(.1406)** (.1798)** (.1790)** (.1722)** (.3641)** (.3941)**
Fragility
(c p 0) . . . . . ⫺.1286 .4359
(.1147) (.3186)
Fragility
(c p .99) . . . ⫺.5417 ⫺.1770 ⫺.2167 ⫺.3190
(.2855)* (.1023)* (.1033)* (.1312)**
Sycophant . . . . ⫺.6984 ⫺.6112
(.3329)* (.3518)*
N ............. 238 238 238 238 238 224
R2 . . . . . . . . . . . . . .4321 .4355 .4453 .4402 .4520 .4018
Note.—Standard errors are in parentheses. Individual and time fixed effects are included
in each model. R2 values reflect within-person variations.
* Significant at 5%.
** Significant at 1% (one-tailed tests).
Ait p 冘16
jp1
(xijt ⫺ xjit)2
2
# jijt , (9)
where the indicator jijt equals 1 if j gave i a less favorable ranking than
i gave j, ⫺1 otherwise. We expect Ait to diminish future growth in status.
Correspondingly, when we enter Ait in model 5, we observe a strong
negative effect and also find that fragility remains discernibly negative.
Having accounted for fraternity members’ time-varying levels of asym-
metric exchange, we can conclude with greater confidence that fragilely
situated individuals face status-related penalties because of the unique
features of the positions they occupy.
Second, we evaluate our primary result further by ensuring that the
pattern of results does not hinge on any particular subject’s series. We
do so both because of the relatively small size of Newcomb’s panel and
due to the fact that one member of the fraternity persistently garnered
poor ratings. White et al. (1976, p. 759) drew attention to “a scapegoat
. . . (man 10), who received one of the bottom three choices of each of
the other 16 persons.” Correspondingly, the intercept for the tenth indi-
vidual is strongly and conspicuously negative across models 1–5 in table
3. We therefore estimated a version of model 5 omitting the tenth indi-
963
10
We also estimated a version of our final model with status scores computed with
b p 0, rather than the standard three-fourths multiplier, to ensure that our results did
not hinge on a particular choice of b. Although this is in our view a theoretically less
attractive way to proceed (here, status is measured just as counts of reverse-coded
rankings rather than weighting by evaluators’ status), the effect of fragility measured
at c p .99 was still discernible with a coefficient of ⫺.209 and a ⫺2.03 t-test.
11
These supplementary models are available upon request. Significance tests for fra-
gility all exceed in absolute value the critical value of 1.645 for a one-tailed test, ranging
from ⫺1.71 when the eighth individual is deleted to ⫺2.43 when the tenth individual
is omitted, as discussed above.
964
12
We address the possible effects of large diagonals (the amount of intracategory sales,
as when automakers sell to automakers) with a separate adjustment. Whereas our
measure of fragility deals with dependency on other nodes in the network, the ad-
justment we introduce captures dependency on oneself at an aggregate level.
965
TABLE 4 (Continued)
13
Two of the 67 categories—federal general government and state and local general
government—necessarily have undefined sell fragility scores because their row sums
967
Fig. 6.—Chain of sales-based ties leading to a high sell fragility score for chemical prod-
ucts. Chemical products sells primarily to plastics and rubber products, which sells mainly
to construction, which in turn sells mainly to real estate. Number of industries’ rows in the
Bureau of Economic Analysis’s use matrices label each node in the chain.
in It are zero. These sectors are, of course, important buyers in the columns of It,
however, and so are included in the computation of fragility scores. After making the
original matrix It row stochastic and squaring each entry, the undefined entries in their
rows of the resulting matrix were reset to zero. We exclude both governmental cate-
gories from our panel models because we are interested primarily in within effects of
fragility. When between or overall effects of fragility are of interest, we suggest as-
signing a fragility score of zero to actors with zero row sums in the relational matrix
and capturing the consequence of their isolate position with a separate dummy variable.
14
Two additional categories of the original 67—noncomparable imports as well as
scrap, used, and secondhand goods—did not receive inputs from any of the other
categories and therefore had row sums equal to zero in ITt . We also exclude them from
our panel models, given that they allow no within-category variation. We include them,
however, in the columns of ITt when calculating fragility scores because of the inputs
they provide.
968
Fig. 7.—Chain of procurement-based ties leading to a high buy fragility score for truck
transportation. Truck transportation buys mainly from petroleum and coal products, which
buys disproportionately from oil and gas extraction, which then relies mainly on rental and
leasing services and lessors of intangible assets. Number of industries’ rows in the Bureau
of Economic Analysis’s use matrices label each node in the chain.
transportation buys mainly from petroleum and coal products, which buys
disproportionately from oil and gas extraction, which then relies princi-
pally on rental and leasing services and lessors of intangible assets.
To investigate associations between industries’ time-varying levels of
fragility and total value added, we turn to panel models of the form
969
970
them to affect total value added positively, adjusting for total industry
input.
Shifting to X it in equation (10), we include additional covariates that
address two particular features of input-output tables. First, we devised
the dummy variable negativeflowit in light of the small number of cases
in which negative flows of dollars appeared in an industry’s row or column
of the input-output matrices. For example, in 2000, the dollar value of
the input from insurance to farms was ⫺$181.7 million. This is because
the insurance industry weathered losses as the farming industry’s claims
exceeded benefits. Since we recoded negative entries to zeros in I t in order
to calculate fragility and centrality scores, we include the indicator
negativeflowit as a cautionary adjustment.
Second, because our measure of fragility concerns the nature of an
actor’s ties to others in a network, as noted previously we set the diagonal
equal to zero in the matrices I t . Zeros along the diagonal is the rule rather
than the exception in most analyses of social networks. Conversely, given
the broad level of aggregation at which industries are defined for input-
output tables, diagonals measuring intraindustry sales (as when computer
manufacturers sell components to one another) are frequently nonzero.
We therefore enter two further adjustments: intrarowit , the ratio of in-
dustry i’s diagonal to its row sum in I t, and intracolumn it, its diagonal
over its column sum in I t. We use these measures to capture a focal
industry’s time-varying levels of internal exchange.
Finally, as denoted by m i and tt⫹1, we enter fixed effects for industries
and for years, respectively. Using industry-specific fixed effects m i accounts
for a number of factors, including cross-industry heterogeneity in demand,
average distance from the end consumer, and other time-constant traits.
With year indicators tt⫹1, we adjust for various aggregate macroeconomic
fluctuations likely to influence industry performance. Correlations and
descriptive statistics for variables in our analyses are shown in table 5.
Table 6 presents results of five models predicting total value added. We
begin with very simple specifications and then proceed to models that
more stringently adjust for consequential sources of variation. Model 1
contains only total industry input and sell fragility measured with max-
imal coupling. We find, keeping with our expectations, that industries that
are less robustly positioned in their selling patterns add less value. Model
1 thus offers preliminary evidence in support of the supposition that
positional robustness in an economic network elevates performance. Turn-
ing to model 2, we observe a negative effect as well for our measure of
buy fragility.
In model 3, we move to a two-way fixed-effects specification that con-
servatively absorbs between-industry, as well as temporal, heterogeneity.
Given this approach, the available variation is within industry and un-
971
1 2 3 4 5 6 7 8 9 10 11
1. ln (total value added) . . . . . . . .
2. ln (total industry input) . . . .6875
3. Sell fragility (c p .99) . . . . ⫺.2625 ⫺.115
4. Buy fragility (c p .99) . . . . .1774 .1063 .0466
5. Sell centrality . . . . . . . . . . . . . . .1644 .0961 ⫺.0028 ⫺.0661
6. Buy centrality . . . . . . . . . . . . . ⫺.0268 .1121 .1276 .0147 .389
7. Sell fragility (c p 0) . . . . . . .0512 .0806 ⫺.0891 ⫺.2732 .0148 ⫺.0268
8. Buy fragility (c p 0) . . . . . . .2092 .081 .0678 .729 ⫺.0174 .0061 ⫺.2988
972
9. Intrarow . . . . . . . . . . . . . . . . . . . . ⫺.0349 .1066 .0948 .0776 .149 .4178 ⫺.1032 .0113
10. Intracolumn . . . . . . . . . . . . . . .1003 .0085 ⫺.024 ⫺.0032 .4356 .0971 ⫺.1175 ⫺.0351 .5209
11. Negative flow . . . . . . . . . . . . ⫺.0768 ⫺.0107 .027 .047 ⫺.1613 ⫺.0482 .09 ⫺.0082 .0187 .0386 . . .
Mean . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3431 11.3445 .31242 .77544 .66943 .66276 .68663 .7888 .1486 .16088 .04444
SD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11939 .12078 .07849 .12639 .04804 .05283 .04811 .09485 .01328 .011 .11838
1 2 3 4 5
ln (total industry input) . . . . . . . . . . . . . . . . . . . . .903 (.035)** .888 (.033)** .381 (.057)** .398 (.062)** .316 (.064)**
Sell fragility (c p .99) . . . . . . . . . . . . . . . . . . . . . ⫺.089 (.035)* ⫺.230 (.057)** ⫺.241 (.055)**
Buy fragility (c p .99) . . . . . . . . . . . . . . . . . . . . . ⫺.374 (.047)** ⫺.011 (.053) .007 (.049)
Sell centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .394 (.094)** .298 (.115)*
Buy centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .045 (.095) ⫺.010 (.102)
Sell fragility (c p .00) . . . . . . . . . . . . . . . . . . . . . ⫺.031 (.092) .181 (.094)
Buy fragility (c p .00) . . . . . . . . . . . . . . . . . . . . . .218 (.072)** .261 (.068)**
2002 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .023 (.014) .024 (.014) .016 (.013)
2003 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .057 (.014)** .047 (.015)** .043 (.014)**
2004 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .101 (.015)** .100 (.016)** .104 (.016)**
973
2005 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .120 (.018)** .120 (.020)** .130 (.020)**
Intrarow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .084 (.481)
Intracolumn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.048 (.544)
Negative flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ⫺.045 (.036)
Constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.134 (.405)** 1.571 (.377)** 6.804 (.631)** 6.592 (.674)** 7.081 (.676)**
974
15
We are grateful to an AJS reviewer for suggesting that we apply our measure to
Burris’s (2004) data.
975
16
The correlation between Burris’s logged measure of social capital based on Bonacich
(1972) and the measure of centrality based on Bonacich (1987), which we use through-
out the current article, is .90. In addition, results are virtually identical if we replace
the former measure with the latter in our subsequent regression models. We use the
former measure in our reanalysis, for sake of consistency with Burris’s original study.
976
TABLE 7
Models Predicting Departmental Prestige in Burris’s
PhD Exchange Network
1 2 3 4
Social capital . . . . . . . . . . . . . . . . . . . . . 1.118 1.105 .747 1.068
(.069)** (.067)** (.045)** (.078)**
Fragility (c p 0) . . . . . . . . . . . . . . . . . ⫺.085
(.093)
Fragility (c p .99) . . . . . . . . . . . . . . . ⫺.101 ⫺.093 ⫺.096
(.039)* (.036)* (.039)*
Article publications . . . . . . . . . . . . . . .072 .067 .070 .064
(.051) (.049) (.052) (.049)
Citations . . . . . . . . . . . . . . . . . . . . . . . . . . .005 .005 .026 .005
(.010) (.010) (.053) (.010)
Research grants . . . . . . . . . . . . . . . . . . ⫺.000 ⫺.001 ⫺.008 ⫺.000
(.003) (.003) (.045) (.003)
Weighted article publications . . . .180 .176 .112 .165
(.082)* (.079)* (.051)* (.080)*
Book publications . . . . . . . . . . . . . . . . .245 .217 .114 .232
(.090)** (.088)* (.046)* (.089)*
Constant . . . . . . . . . . . . . . . . . . . . . . . . . . ⫺.401 ⫺.294 ⫺.000 ⫺.138
(.150)** (.151) (.034) (.227)
N .................................. 94 94 94 94
R2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .89 .90 .90 .90
Note.—Standard errors are in parentheses. Model 3 reports standardized regression co-
efficients and corresponding standard errors.
* Significant at 5%.
** Significant at 1%.
977
978
Limits on Generality
Before sketching implications for other lines of sociological theory, we
emphasize three conditions that restrict the generality of the perspective
we have presented. First, there are clearly institutional settings in which
extralocal processes do not matter and where decoupling therefore defines
the network. In these settings, fragility is best represented just as a Her-
findahl measure of concentration, without attention to contiguous actors’
identities and networks. Consider in this connection the two-by-two table
979
in figure 8, whose vertical axis extends from fragile to robust and whose
horizontal axis goes from decoupling to coupling. If in fact decoupling
marks the network, then fragility should be measured in a way that is
consistent with the bottom-left cell in figure 8, where A relies completely
on D but where D’s identity and network are inconsequential for A. This
quadrant is therefore quite different from its counterpart on the bottom
right, where A is fragile because of its position at the end of a chain, and
from the upper-right cell, where A is robust by virtue of its wide range
of connections to others who are diversified in their sources of social
support.
The extent to which a social structure is “readable” (White 1992, pp.
106–11) contributes significantly to the degree of coupling depicted on the
horizontal axis and thus is a factor determining the best measurement
strategy. Although many social systems allow incumbents (as well as ob-
servers) to map out who is connected to whom and in what way—consider,
as extremes, a rank-order tournament or, less commonly, a matrix orga-
nization whose publicly displayed organizational chart accurately reflects
its informal structure—other systems are far more opaque, even for in-
siders. There, the concrete patterning of social ties is fuzzy and uncertain
for all involved—incumbents and observers have difficulty discerning, or
even inchoately sensing, nonlocal patterns—and network spillovers are
therefore rare. Consequently, an important condition for our approach is
a reasonable level of perception within the network. Coupled identities,
where spillovers occur, necessarily require opportunities for detecting or
discerning social ties.
We therefore expect extralocal ties to matter appreciably in the esti-
mation of fragility effects within transparent settings like investment
banking, where firms’ locations in public pecking orders are widely mon-
itored (Podolny 1993). In such settings, we recommend computing fragility
with a large b weight. Conversely, for less penetrable settings that would
tend toward the left column in figure 8—such as semiconductors (Podolny,
Stuart, and Hannan 1996), where consumers likely have difficulty dis-
cerning the circuitous patterns of interfirm patenting ties—we anticipate
little in the way of network spillovers and thus suspect that firms’ levels
of fragility will be best calculated more restrictively (i.e., with a smaller
b weight or even with b set equal to zero).
Second, our depiction of A’s position as fragile due to chaining in the
bottom right of figure 8 leads us to reinforce the importance of examining
global-network structure when deciding whether to implement our mea-
sure. As we discuss in more technical detail in our appendix, a central
consideration in determining if our measure will yield information over
and above Herfindahl’s index is whether the structure of the global net-
work permits chains of reasonable length among its nodes or is instead
980
Fig. 8.—Fragility and robustness for decoupled and coupled networks. Solid lines denote
first-order ties of social support. Dashed lines represent second-order ties of social support.
Diameters of circles around nodes reflect levels of robustness. Nodes are circled if their
identity and network affect the robustness of the focal actor A.
981
ployees (Baron 2004). More generally, there are certainly settings in which
fragility is ideologically construed as positive.
To adjudicate among ways of discussing the likely effects of our mea-
sure, we believe it is essential as a starting point to consider differences
in where concentration comes from—from ego or from alter. Consider a
case from the computer industry as an example. Dell’s choice historically
to focus exclusively on the direct sales channel is markedly different from
a fraternity brother who gets a disproportionate share of his social support
from one other member of the house—although Herfindahl’s index will
be high for both. Dell’s strategy has largely proven to be status enhancing,
but in Newcomb’s fraternity we found that concentration is status erod-
ing. Where these cases differ importantly is not in levels of analysis but
in the provenance of concentration and therefore in terms of control.
Whereas a computer maker can deliberately decide to specialize within
a specific segment, the fraternity brother has far less discretion over
whether he receives the lion’s share of his social support from a particular
peer. When this does occur, those other than the focal individual are
consequently the ones imputing (unfavorable) meaning to his concentrated
position.17
Thus, in keeping with lines of research differing from our own, the
measure of fragility we have introduced might be recast as a measure of
focus in other contexts. In such contexts, actors receive rewards for es-
tablishing exclusive ties to those with narrow allegiances, and the fragile–
robust axis labels for the ordinate in figure 8 could therefore be replaced
respectively with focused–disordered. When the cost of forging such ties
varies inversely with actors’ quality levels—as at times they do in ap-
prenticeship programs, for instance—equation (4) will measure focus,
rather than unwanted complexity or disorder, as a viable market signal
in Spence’s (1974) sense.18
To mitigate concerns about generality, in addition to describing how to
substantively reframe our measure to accommodate different settings, we
stress the pervasiveness of the social structural conditions that are con-
sistent with the viewpoint we have advanced: our approach posits that
17
Although we believe that attention to the origins of fragility—with ego or with alter—
serves as a useful starting point, it is insufficient. Like a laser steadily aimed at the
wrong target, ego can of course deliberately narrow its focus on the wrong audience.
18
This point also raises the issue of lineages, if actors in a system span cohorts or
generations. Under this scenario, eq. (4) could measure purity of lineage, as when an
artist or academic comes from the right (and narrow) line of teachers. Purity is then
a by-product of having collaborated exclusively with a master who was him- or herself
sharply discriminating as a disciple. For instance, this is almost certainly the legiti-
mating process the Apostle Paul had in mind when, in his defense, he cited Gamaliel
as his teacher (Acts 22).
982
Future Directions
In addition to replicating our results in other settings, especially in those
in which status in known to precede coveted outcomes, we see three
particularly promising ways to proceed further. First, although our mea-
sure of fragility operates at the level of individual nodes, it is easy to
imagine opportunities to link up with important group-level themes high-
lighted by Moody, White, Harary, and colleagues in their work on struc-
tural cohesion (see, e.g., White and Harary 2001; Moody and White 2003).
Moody and White (2003) discuss robustness at the group level in terms
of a focal group’s capacity to persist in spite of node removal and thus
picture cohesive groups as those that have “a status beyond any individual
group member” (p. 122). They also suggest that future research may benefit
from focusing on the hazards of removal faced by particular nodes. Along
this line, we believe that our measure of fragility may be particularly
valuable.
Consider, as a potential empirical site, work in biology on metabolic
and other cellular networks (Jeong et al. 2000) that focus on individual
substrates, connected by biochemical pathways, constituting a metabolic
network. In this domain, a highly fragile substrate is one that is connected
to one or a few narrowly focused alternate substrates. In addition, not
unlike many social networks, biological networks are often characterized
by long-tailed distributions where a few substrates account for much of
the network connectivity. Although this can result in stability for the
network as a whole—in the sense that networks stay largely intact in the
face of random error—researchers are now interested in examining cross-
organism differences in the identities and durability of individual sub-
strates (Jeong et al. 2000). Arraying nodes in these and similar kinds of
networks according to their fragility may be useful for locating otherwise
hidden fault lines in global structures and thus for further understanding
group-level robustness as envisioned by Moody and White (2003).
983
Two additional next steps are to link notions of fragility and robustness
more tightly to etiologies of status and also to harness the approach we
have developed to further pinpoint the structural foundations of robust
action—or action that preserves options (Leifer 1988; Padgett and Ansell
1993). We conclude with very preliminary sketches of how theories of
status and of robust action might be extended in light of our approach.
We found in our analysis of Newcomb’s data that fraternity members
enjoyed increments in status as their positions within the social structure
became more robust. This finding is consistent with Abbott’s (2001) ob-
servation that status is a by-product of possessing a wide range of future
options. We interpreted this finding on a purely structural level as evidence
consistent with the premise that audience members esteem those in robust
positions. In arguing that structure directly affects perceptions, we sought
to adjust as stringently as possible for fixed and time-changing behavioral
differences in our empirical models. Nevertheless, we believe that future
work on the determinants of status may benefit from examining some of
the possible behavioral concomitants of robustness. These are actions that
emerge from the durability of actors’ positions and in turn contour their
levels of prestige. We refer here in particular to actions that are permissible
and sustainable exclusively for inhabitants of robust positions. Examples
include the “attractive arrogance” Merton (1968, p. 61) ascribed to elite
scientists or excessive acts of generosity directed toward others who lack
the resources to reciprocate. Through tracing out the links among ro-
bustness, modes of conduct, and harvests of future deference, we antic-
ipate that our conceptions of where status comes from are likely to
sharpen.
Turning in conclusion to further implications for theory, we suspect that
robust positions, as we have defined them, render robust action more
sustainable. Work on robust action has immediate precedents in Leifer’s
(1991) research on chess masters. This study found that the very best chess
players are not those endowed with uncommon forecasting abilities but
rather those who can sustain flexibility with respect to (or remain robust
with respect to) opponents’ moves. More generally, robust action refers
to strategic conduct that is difficult for a competitor to interpret and that
therefore keeps its practitioner from getting absorbed in an undesirable
role. For example, by acting robustly in a romantic relation, a suitor veils
actual intentions and therefore circumvents appearing overly committed
and becoming the less powerful, or by acting robustly in international
relations, one nation wishing to conquer another nation adroitly provokes
its target to start a war and then successfully brands it as the instigator
(Leifer 1988, pp. 867–69). Correspondingly, for a political figure, this strat-
egy equates to maneuvering in ways that are indecipherable to opponents
and therefore congenial to continued flexibility (Padgett and Ansell 1993).
984
APPENDIX
985
empirical illustration). Moving from the dyadic to the global level, the
mean path length—the average shortest distance between all connected
pairs—for a given social structure is an index of its overall reachability
or efficiency (Wasserman and Faust 1994). To the extent that mean path
length is low, members of the social structure can get to one another
quickly. In contrast, a high mean path length reflects the presence of long
winding strands of social ties.
We investigate how mean path length conditions fragility scores in order
to assess a direct implication of our approach—namely, that fragility mea-
sured with a high coupling coefficient is most likely to offer unique in-
formation in networks marked by extended chains of dependence. We
believe that pursuing this implication is important on theoretical and
empirical grounds. On a theoretical level, it begins to illuminate the con-
textual underpinnings of fragility by bringing relief to global-level dy-
namics likely to increase variation along a fragility-robustness axis. On
an empirical level, it has the potential to alert researchers to the types of
observable social structures in which attending to context (over and above
the immediate composition of the ego network) in the computation of
fragility yields the greatest payoffs.
We proceeded by generating 1,000 simulated networks of various sizes
and levels of connectivity. Using these networks, we could then assess the
effect of mean path length on our correlation-based outcome of interest,
cor (Fcp.99 , Fcp0 ), which we refer to as cormeasures. The square matrices
from which we constructed these networks ranged in size n from 50 to
150 rows. To induce variation in the total number of nonzero entries across
our various matrices, we worked with an array of “cutpoints” in each
matrix size category.
More precisely, after generating a square matrix of a given size n from
n2 random draws from the standard normal distribution, we set all entries
above the chosen cutpoint equal to one, brought other entries to zero,
and constrained the main diagonal to equal zero regardless of the value
of the initial random draw. We used six cutpoints in total: 0, .5, 1, 1.5, 2,
and 2.25. With equal frequency, we applied the first four of these cutpoints
to 160 matrices of size 50; we applied all but the largest to a total of 600
matrices of sizes 75, 100, and 125; and we used all six cutpoints on 240
matrices of size 150. Thus, for each size-cutpoint pair, we simulated 40
random matrices, yielding a total of 1,000 directed nonvalued networks.
We then computed cormeasures as the correlation between fragility for
c p .99 and fragility for c p 0, as well as mean path length, for each of
these networks. Although our results are virtually identical if we set iso-
lates’ fragility scores equal to zero, we eliminated isolates from our anal-
ysis, both because fragility is undefined for completely disconnected nodes
986
987
988
REFERENCES
Abbott, Andrew. 1981. “Status and Status Strain in the Professions.” American Journal
of Sociology 86:819–35.
———. 1995. “Sequence Analysis: New Methods for Old Ideas.” Annual Review of
Sociology 21:93–113.
———. 2001. Time Matters: On Theory and Method. Chicago: University of Chicago
Press.
Abbott, Andrew, and Alexandra Hrycak. 1990. “Measuring Resemblance in Sequence
Data: An Optimal Matching Analysis of Musicians’ Careers.” American Journal of
Sociology 96:144–85.
Abbott, Andrew, and Angela Tsay. 2000. “Sequence Analysis and Optimal Matching
Methods in Sociology.” Sociological Methods and Research 29:3–33.
Ballester, Coralio, Antoni Calvó-Armengol, and Yves Zenou. 2005. “Who’s Who in
Networks. Wanted: The Key Player.” Econometrica 74:1403–17.
Banfield, Jeffrey D., and Adrian E. Raftery. 1993. “Model-Based Gaussian and Non-
Gaussian Clustering.” Biometrics 49:803–21.
Baron, James N. 2004. “Employing Identities in Organizational Ecology.” Industrial
and Corporate Change 13:3–32.
Berger, Joseph M., Hamit Fişek, Robert Z. Norman, and Morris Zelditch, Jr. 1977.
Status Characteristics and Social Interactions: An Expectation States Approach.
New York: Elsevier.
Blau, Peter. 1964. Exchange and Power in Social Life. New York: Wiley.
Bonacich, Phillip. 1972. “Factoring and Weighting Approaches to Status Scores and
Clique Identification.” Journal of Mathematical Sociology 2:113–20.
———. 1987. “Power and Centrality: A Family of Measures.” American Journal of
Sociology 92:1170–82.
Borgatti, S., M. G. Everett, and L. C. Freeman. 1999. UCINET V for Windows: Software
for Social Network Analysis, ver. 5.2.0.1. Natick, Mass.: Analytic Technologies.
Bothner, Matthew S. 2003. “Competition and Social Influence: The Diffusion of the
Sixth-Generation Processor in the Global Computer Industry.” American Journal of
Sociology 108:1175–1210.
Bothner, Matthew, Richard Haynes, Wonjae Lee, and Edward Bishop Smith. 2010.
“When Do Matthew Effects Occur?” Journal of Mathematical Sociology 34:80–114.
Burris, Val. 2004. “The Academic Caste System: Prestige Hierarchies in PhD Exchange
Networks.” American Sociological Review 69:239–64.
Burt, Ronald S. 1976. “Positions in Networks.” Social Forces 55:93–122.
———. 1982. Toward a Structural Theory of Social Action. New York: Academic
Press.
———. 2005. Brokerage and Closure. New York: Oxford University Press.
———. 2007. “Closure and Stability: Persistent Reputation and Enduring Relation-
ships among Bankers and Analysts.” Pp. 100–143 in The Missing Links: Formation
and Decay in Economic Networks, edited by James Rauch: Russell Sage.
Carroll, Glenn, and Michael T. Hannan. 2000. The Demography of Organizations and
Industries. Princeton, N.J.: Princeton University Press.
Clark, Tim W. 1987. “Black-Footed Ferret Recovery: A Progress Report.” Conservation
Biology 1:8–11.
Doreian, Patrick, Roman Kapuscinski, David Krackhardt, and Janusz Szczypula. 1996.
“A Brief History of Balance through Time.” Journal of Mathematical Sociology 21:
113–31.
Ellison, Christopher G., Jeffrey A. Burr, and Patricia L. McCall. 1997. “Religious
Homogeneity and Metropolitan Suicide Rates.” Social Forces 76:273–99.
Emerson, Richard M. 1962. “Power-Dependence Relations.” American Sociological
Review 27:31–41.
990
991
992