Professional Documents
Culture Documents
North-Holland
John M. BOLLAND
Unroersrty of Alabama **
Although the concept of centrality has been well developed m the social networks hterature, Its
emplncal development has lagged somewhat Thus paper moves a step m that dlrectlon by
assessing the performance of four centrality models under a vanety of known and controlled
situations It begms by exammmg the assumptions underlying each model, as well as Its behavior
m a commumty influence network. It then assesses the robustness and sensltlvlty of each model
under condltlons of random and systematic variation Introduced mto tbs network.
Introduction
* Thus 1s a rewslon of a paper ongmally presented at the 1986 meeting of the Sunbelt Social
Network Conference. It benefitted from comments by Phil Bonaclch, Wayne Baker, Kathleen
Bolland, and anonymous reviewers
l * Instrtute for Social Science Research, Umverslty of Alabama, P.O. Box 587, Tuscaloosa, AL
35487, U S A.
c D(l) = IL,
J
(1)
J M Bolland / The performance of four centrahty models 235
c D(J)
C’D(r) = -
n-l’
n-l
c’ =-
C(r) n
Cd*,
g EJm
b ‘Jm =-
g Jm
2CB(,,
c’B(r) =
?I*-3n+2’
a(m)
'J ’ (7)
Us,, equals z’s degree when m = 1, and the number of paths of length
i m originating at i when m > 1. By setting m to a sufficiently large
value, V = u[i] converges to an equilibrium, although the values of V
tend toward infinity. Bonacich showed that by dividing Equation (7) by
A” (where X is the largest eigenvalue of A ), V converges to the
eigenvector of A. In turn,
?I,
C’CF(r) =
CW
I
) = (9)
Max
i
iG(.)- G,,))
I i
’ I ~111 use the followmg notanon throughout the paper CL,, denotes I’S centrahty based on
betweenness, CW ) denotes network centrahzatlon based on betweenness, and C, denotes the
general betweenness model Smular notation IS used to described closeness, degree, and contmumg
flow. In addltlon, Cc,, denotes I’S centrahty based on some speclfxd model, and C,) denotes
network centrahzatlon speclfvxl by that model
J M Bolland / The performance of four centralrty models 231
with a single hub (where C&.) = 1, C& 1 ) = 1, and C& ) = l), and he l
(11)
s = d-” (15)
238 J M Bolland / The performance offour centralrty models
CD ‘x geodesm/all paths
CC 1 geodesm
CB 0 geodesm
c CF 0 all paths
I found with the simulated data reported in Study II. These correla-
tions suggest that CL,, conveys the most unique information about the
centrality of individual vertices. Several specific comparisons from the
Chillicothe data, while supportmg this conclusion, also suggest that
C’cr(,) offers unique insights into centrality within the network.
(1) C&p in comparison with the other three models, deflates the
centrality of external network members (e.g., those from the county
and parochial school districts: vertices #II, #17, #21, #23,
#66, #70, #84, #lOO).
(2) C&C,,9 m comparison with the other models, mflates the centrality
of vertex #26, an elementary school principal at the time of the
study. A year later, he was appointed Assistant Supermtendent.
(3) CL,,,, in comparison with the other models, deflates the centrality
of vertex #4, a clerk in the Chillicothe School District Office
whose ties to other staff members were based exclusively on
frequent interaction rather than attributions of importance.
(4) CL,,,9 in comparison with the other models, inflates the centrality
of vertices # 70 and # 100, both of whom serve boundary-spanning
functions in the network.
Table 1
Intercorrelatlons among four centrahty models and network centrahzatlon mdlces for the Chdh-
cothe data
Correlations Centrahzatlon
I I
C&J) C’C(J) CW, C&c,, Ongmal Transformed C,,
CatI, 0 74 0.76 0 51 0.173 0217
Ccc,,
, 0 89 0 78 0 242 0 419
Cdl, 0 83 0 222 0 291
C’CR,) 0 919
J M Bolland / The performance offour centralrty models 241
lo-
09-
c
0
R 06-
R
E
L 07-
*
T 06-
0
N OS-
04-
03: , , I I r 031 , , , I r
0 02 05 10 15 20 0 02 05 10 15 20
error error
Fig 2 (a) Mean Pearson correlation between unperturbed centrahty mdlces and those dewed
from randomly perturbed data under Assumption I. (b) Mean Pearson correlation between
unperturbed centrahty mdlces and those denved from randomly perturbed data under Assump-
tlon II
Random vanatlon.
I initially assessed the robustness of each model under each assumption
as the Pearson correlation between the centrality of the vertices in A
and their centrality in each of the replications of A’; Figure 2 shows
these results for each error level. Under Assumption I, the mean
correlation for C’B(rj quickly becomes unsatisfactory with increasing
error, while that for the other three models deteriorates less rapidly.
Under Assumption II, deterioration is less rapid, with C&,, and Cd,,
showing considerable robustness even at high error levels. Since de-
’ When a smgle lmk 1s added, N and the degree of the target vertex constram the number of
rephcatlons. Thus, for vertex #2 (whose degree equals ten), only 21 rephcatlons are possible
wthout redundancy In contrast, vertices #9 and #94 each have a degree equal to one, so 30
rephcatlons are possible wthout redundancy
J.M Bolland / The performance offourcentralrty models 243
where O(,, = I’S centrality in A and E(C,,,) = z’s mean centrality across
the 100 replications of A’ for each error level. The standard error of
C(I, K(,,) eq uals the standard deviation of Cc,, across the 100 repli-
cations of A’ for each error level. Obviously, since the mean of a
distribution of z-scores is zero, the overall bias of any perturbed model
is also zero. But for any given vertex, the model may or may not be
biased.
Table 2 reports Bias(,) and &ccc,,for the four vertices for each model
under the 20% error condition for each assumption. In general, CL,,,
and CL,,, show the greatest bias under Assumption I, while C&,, and
C&,) show the least. However, C&,, shows comparatively high bias for
vertex #94; this suggests that while Cc, may be comparatively unbi-
ased at the network’s center, it is biased at the network’s periphery. In
the same way, C&,, shows considerable bias for vertex #ll, suggest-
ing that it may be biased for network bridges. An analysis of 5occl,
under Assumption I shows little difference in variability for the four
models or for different regions of the network. Under Assumption II,
the performance of all models improves substantially. However, CL,,,
remains the most biased of the models, and C&(rj remains the most
unbiased of the models. Further, CL,,, shows the greatest overall
variability, while C&r(,) shows the least variability. The notable excep-
tion to this generalization, however, is the poor performance of C&,,
for vertex #94.
Systematic variatzon.
This type of variation might occur were an individual to expand his or
3 I omt the complete results for each error level m ths paper to conserve space. These results are
avadable upon request.
244 J M Bolland / The offourcenrralrty
performance models
Table 2
Mean standardued centrahty mdlces, bias, and standard errors of models for the random
vanatlon condltlon
Assumption I
ID Error
2 000 0 95 1 61 2 51 2 39
0 20 1 47 0 52 123 138 - -0.23 0 79 158 - -093 083 1.59 -0 80 0.75
9 000 - 0.89 -041 - 1 00 -053
0.20 -045 044 071 -053 0 12 0.89 -053 047 084 -049 0 04 0 79
11 0.00 0 62 0.48 0 55 -053
0 20 036 -0.26 101 029 - 019 086 -034 0 21 0.90 0 17 0.70 0 89
94 000 -089 -153 -1.00 -087
0 20 -039 050 084 -061 092 096 -056 044 084 -065 022 083
ID Error
2 000 0 95 1 61 2.51 2 39
0 20 147 052 100 137 -024 069 2.26 - -0 25 0.59 231 -008 039
9 000 -089 -041 - 1 00 -0.53
0 20 - 0.68 021 049 -049 -0.08 052 -0 87 013 038 -0.54 -001 027
11 000 0 62 048 0 55 -0 53
0.20 0 36 026 0 74 025 -023 050 047 - -008 053 -035 018 036
94 0.00 -089 -153 -100 -087
0 20 -052 0.37 058 -086 067 0 78 -077 023 039 -084 003 029
Table 3
Local sensrtivtty and configural sensihvlty of network centrahty models m the fwe-lmk systematic
vanatton conditron
Ca Cc Co CCF
Target
#2 ongmal 0 95 1.61 2 51 2 39
5 lmks 4.26 3.31 0 33 264 103 024 3.72 1 21 0 04 292 0.53 003
#9 ongmal -089 - 0.41 -100 -053
5 lmks 108 197 0.64 1.05 146 029 0.84 1.84 002 033 086 033
#11 ongmal 0.62 0.48 0.55 -053
5 hnks 3 21 265 0.40 174 1.26 025 221 166 0.05 0.62 1 15 0.42
#94 ongmal - 0.89 - 1.53 -100 - 0.87
5 lmks 1.49 2.38 0 58 0.99 2 52 0.34 0.84 1 84 0 02 -001 0.86 037
4 Another potentrally useful concept IS global sensttwtty, or the extent to whxh lmks added to a
target vertex affect the centrahty of others m various regrons of the network In an expanded
version of thrs paper, I examine the global sensitwrty of each model, wtth the concluston that the
models show only modest differences on thts cntenon. These results are also avarlable upon
request.
246 J M Bolland / The performance
offourcentralrty models
Table 4
Mean centrahzatlon mdlces, bias, and standard errors of models for the random vanatlon
condltlon
Assumptron I
c0 Blas( ) c0 Blas( )
CB 000 0 173 0.217
0 20 0.056 -0 117 0 019 0 030 -0 187 0004
Cc 0.00 0 242 0419
0 20 0 196 -0046 0.044 0 330 -0089 0 043
CD 000 0 222 0.297
0 20 0210 -0012 0 048 0 216 -0081 0 028
Cc, 000 0919
0.20 0.534 -0385 0 070
Assumptron II
c, ) By, %, c, ) Blas() %,
addition of 5 random links to their ego networks (in the case of vertex
# 11, this occurs with the addition of the first link). C&,, shows this to
occur only for vertex #11 (and just barely at the fifth link), while CL,,,
and C&r) show it not to occur at all. Thus, C&,, portrays the network
as volatile, while the other models (particularly I&,,,) portray it as
more stable.
Table 5
Local sensmvlty and configural sensrtlvlty of network centrahzatlon models m the five-hnk
systematic vanation condltlon
Conclusions
Evuluatlon of C,
way, causing ambiguity about the true meaning of C&,,. Even at the
network’s periphery, however, high local sensitivity is troublesome, and
C’B(Ij fails to pass the “sow’s ear” test. Like the sow’s ear, which cannot
be turned into a silk purse, neither vertex #94 nor vertex # 9 should
become more central than vertex #2 through modest networking
activity (i.e., the addition of five random links to their ego networks).
Yet, for both, C&,, shows this to happen. Finally, C&,, is relatively
sensitive to the exact configuration of the systematic change that is
introduced, suggesting its useful network properties.
Overall, C, seems quite useful for what it says that the other models
do not address. Its sensitivity provides it with a good capability of
capturing subtle network changes, reducing its likelihood of commit-
ting the network-equivalent of a Type II error; but at the same time, it
is very susceptible to Type I errors. Its interpretation therefore must be
cautious, particularly if data are prone to even modest error.
Evaluation of Cc
ment error is most likely, C&,, may be even more unstable than these
results indicate. If the network’s periphery is the focus of inquiry, then
C’cc,) may perform well, allowing the investigator to uncover subtle
aspects of centrality that other models may miss; otherwise, it may
distort interpretation of network structure. In general, Cc falls in the
middle of the four models with respect to most criteria. Therefore, it is
neither clearly good nor bad. This may be useful - and perhaps
accounts for its generally greater use as an intuitive model of centrality
_ for those who wish to avoid risk, but it also limits its usefulness for
those who wish to capture the full complexity of network structure.
Evaluation of CD
Evaluation of Cc,
References
Bavelas, A
1948 “A mathematical model for group structures ” Apphed Anthropology 7 16-30
Bolland, J.M.
1985 “Perceived leadership stahthty and the structure of urban agenda-settmg networks.”
Socral Nehvorks 7 153-172.
Bonactch, P.
1971 “Factoring and wetghtmg approaches to status scores and chque tdentrftcatton ” Journal
of Mathematrcal Sonologv 2 113- 120
1987 “Power and centrahty. a fannly of measures.” Amencan Journal of Socrology 92
1170-1182
Freeman, L C.
1979 “Centrality m networks I. conceptual clanfxatron ” Socral Network I 215-239
Harary , F , R.2 Norman and D. Cartwnght
1965 Structural Models. New York. Wiley.
Holland, P. and S. Lemhart
1973 “The structural rmphcatrons of measurement in soctometry.” Journal of Mathematical
Socmfogv 3. 85-112.
J M Bollond / The performance of four centrolrty models 253