tnr2013 ron burt, network advantage on how the network was built
DESCRIPTION
TRANSCRIPT
To download a copy of the working paper,
go to http://faculty.chicagobooth.edu/ronald.burt/research.
Path Dependent Network Advantage
A certain Ron Burt Jen Merluzzi
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
Bob Bob Bob Bob Bob
February OctoberJune AugustApril
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
Deb Deb Deb Deb Deb
1
3
11
58
Bob
NonRedundantContacts(thinsolidline)
Network Density& Constraint
(bold lineis constraint)
Bob IsAlways aBroker
Deb Buildsvia SerialClosure
(Metricsoscillatethrough
reversals)
December Network Survey
1
3
11
58
Deb
Figure 5. Broker Network Can Result from Always
Being a Broker, or from Serial Closure
Path Dependent Network Advantage
Puzzle: Individual Differences
in Returns to Network Advantage
Volatility Treated as Noise
Volatility Treated as Signal
Conclusions & Serial Closure
Stra
tegi
c Lea
ders
hip
Crea
ting
Valu
e: T
he S
ocial
Cap
ital o
f Bro
kera
ge (p
age 8
)
James
Robert
1
23
54
6
7
A
B
C & D 25
1
0
100
29
Group A
Group B
Group C
Group D
Density Table
0
85
5
0
0
person 3: .402 = [.25+0]2 + [.25+.084]2 + [.25+.091]2 + [.25+.084]2
Robert: .148 = [.077+0]2 + [.154+0]2 + [.154+0]2 + [.154+0]2 + [.154+0]2 + [.154+0]2 + [.154+0]2
Network Constraint(C = Σj cij = Σj [pij + Σq piqpqj]2, i,j ≠ q)
Networkindicates
distributionof sticky
information, which defines
advantage.
From Figure 1.1 in Brokerage and Closure. For an HBR treatment of the network distinction between Robert and James, see in the course packet Kotter's classic distinction between "leaders" versus "managers." Robert ideally corresponds to the image of a "T-shaped manager," nicely articulated in Hansen's HBR paper in the course packet.
Small World of Organizations & Markets
Stra
tegi
c Lea
ders
hip
Crea
ting
Valu
e: T
he S
ocial
Cap
ital o
f Bro
kera
ge (p
age 2
3)
Network Constraintmany ——— Structural Holes ——— few
Z-Sc
ore
Res
idua
l Ach
ieve
men
t(e
valu
atio
n, c
ompe
nsat
ion,
pro
mot
ion)
Graph (A) shows achievement decreasing with less access to structural holes. Circles are z-score residual achievement for 1,986 observations averaged within five-point intervals of network constraint in each of six management populations (analysts, bankers, and managers in Asia, Europe, and North America; heteroscedasticity is negligible, c2 = 2.97, 1 d.f., P ~ .08; Burt, 2010:26, cf. Burt, 2005:56). Bold line is the vertical axis predicted by the natural logarithm of network constraint. Graph (B) shows the raw data averaged in Graph (A). Vertical axis is wider to accommodate wider achievement differences. Heteroscedasticity is high given wide achievement differences between brokers (c2 = 269.5, 1 d.f., P < .001), but returns to brokerage remain statistically significant when adjusted for hteroscedasticity (Huber-White, t = -8.49). Figure 1 in Burt, "Network-relevant personality and the agency question" (2012 AJS)
A. Achievement Scores Higher than Peers on Averager = -.58, t = -6.78, n = 85
(heteroscedasticity c2 = 2.97, 1 d.f., P < .08)
B. Vary Widely between Individualsr = -.24, t = -9.98, n = 1,989
(heteroscedasticity c2 = 269.5, 1 d.f., P < .001)
In Sum: Individuals Differ Greatly in Returns to Brokerage
Stra
tegi
c Lea
ders
hip
Crea
ting
Valu
e: T
he S
ocial
Cap
ital o
f Bro
kera
ge (p
age 1
3)
Most Senior Job Ranks(29.5 mean network constraint)
Next-Lower,Middle Ranks(56.4 mean constraint)
Next-Lower,Senior Ranks(41.9 mean constraint)
r2 = .61
Net
wor
k St
atus
(S)
(Si =
Σj z
ji Sj,
divi
ded
by m
ean
so a
vera
ge is
1.0
)
Perc
ent o
f Peo
ple
with
in E
ach
Leve
l of J
ob R
anks
18%
1%
Network Constraintmany ——— Structural Holes ——— few
Network Constraintmany ——— Structural Holes ——— few
Network Brokers Tend To Be the LeadersConstraint and status are computed from work discussion networks around twelve hundred managers in four organizations.
A. In the formalorganization
B. And in the informal organization
Figure 2.4 in Burt, "Network structure of advantage" (2013 manuscript)
Path Dependent Network Advantage
Puzzle: Individual Differences
in Returns to Network Advantage
Volatility Treated as Noise
Volatility Treated as Signal
Conclusions & Serial Closure
Current Network
Advantage Subsequent Previous
Network Volatility — Noise or Signal?
Illustra(ng the privilege accorded stability in network analysis, here is Laumann & Pappi (1976:213) on community elites: "Despite differences in nuance associated with 'structure,' the root meaning refers to a persis(ng order or paKern of rela(onships among units of sociological analysis, be they individual actors, classes of actors, or behavioral paKerns.” Nadel (1957:9) is cited as precedent: "We iden(fy the mutual ways of ac(ng of individuals as 'rela(onships' only when the former exhibit some consistency and constancy, wince without these aKributes they would merely be single or disjointed acts.” Of course, implicit in Laumann & Pappi’s emphasis on stability is Nadel’s (1957:8) conceptual dis(nc(on between stable structures and variable parts: "structure indicates an ordered arrangement of parts, which can be treated as transposable, being rela(vely invariant, while the parts themselves are variable.”
Data 1
Four annual panels of compensation and network data on 346 investment bankers employed by the study organization in each of the four panels.
Network data are from annual 360 evaluation process (zji = 1 if j cited i or i cited j as a colleague with whom citer had frequent and substantive business during previous year, 0 otherwise; summary colleague evaluation for banker i is average across colleagues j of zji on 4-point scale).
There are also some control variables (banker’s job rank, colleague evaluation, years in organization, minority, works at headquarters). Measures of network advantage:
Network eigenvector measures centrality/status (si = ∑j pjisj, where pji is proportion of i’s relations that are with j) Network constraint measures lack of access to structural holes (ci = ∑j cij, where cji = (pij + ∑k pikpkj)2, k ≠ i, j, and pij is the proportion of i’s relations that are with j).
Legend: Color indicates banker job rank: top (gold), senior (gray), or other (white). Shape indicates location: US (circle) or elsewhere (square). Lines in sociogram A connect bankers linked by a citation in any of the four years. Lines in sociogram B connect bankers linked by a citation in all four years.
Figure 1. Enduring Banker Relations Better Reveal Social Clusters
A B
Network Status Network Constraint Job Rank 2 Job Rank 3 Job Rank 4 Colleague Evaluation Years with the Organization Minority (gender or race) US Headquarters Intercept Multiple Correlation Squared Number of Observations
.41 —
.20 .48
1.48 .17
.004 -.07 -.11
-.91 .71 346
(.05) ** (.08) * (.09) ** (.10) ** (.04) ** (.01) (.07) (.06)
NOTE — Unstandardized OLS regression coefficients are presented with standard errors in parentheses. Compensation is measured as a z-score. Network status is an eigenvector score normalized to the average banker. Network constraint is the log of constraint. See Appendix B for control variables, means, standard deviations, and correlations. Models I and II predict compensation summed across years from network indices computed from relations pooled over time (relation is 1 if it occurs in only one year, 2 if two years, etc.). Network status is correlated -.86 with log network constraint. Models III and IV predict annual compensation averaged across years from network indices computed for each year then averaged across years. Network status is again correlated -.86 with log network constraint. Models V and VI predict compensation next year from network indices this year (with standard errors adjusted for autocorrelation between repeated observations of the same bankers using the “cluster” option in STATA). * p < .05, ** p ≤ .001
Table 2.
Compensation Returns to Network Advantage
I
.47 —
.20 .50
1.37 .13
.001 -.05 -.14
-.91 .74 346
(.05) ** (.08) * (.08) ** (.10) ** (.04) ** (.01) (.07) (.06) *
II
— -.31
.20 .51
1.64 .18
.008 -.08 -.06
.21 .68 346
(.07) ** (.09) * (.09) ** (.11) ** (.04) ** (.01) (.08) (.06)
IV
— -.41
.21 .52
1.58 .17
.006 -.07 -.07
.78 .70 346
(.08) ** (.09) * (.09) ** (.11) ** (.04) ** (.01) (.07) (.06)
V
All Years Between Years
.33 —
.23 .59
1.55 .12
-.003 -.07 -.09
-.81 .71
1038
(.04) ** (.03) ** (.06) ** (.10) ** (.02) ** (.01) (.04) (.05)
III
— -.27
.24 .62
1.74 .14
.002 -.09 -.04
.31 .66
1038
(.04) ** (.03) ** (.06) ** (.11) ** (.03) ** (.01) (.05) (.05)
VI
Within Years
Path Dependent Network Advantage
Puzzle: Individual Differences
in Returns to Network Advantage
Volatility Treated as Noise
Volatility Treated as Signal
Conclusions & Serial Closure
Data 2
Four annual panels of compensation and network data on 346 investment bankers employed by the study organization in each of the four panels. Network data are from annual 360 evaluation process (zji = 1 if j cited i or i cited j as a colleague with whom citer had frequent and substantive business during previous year, 0 otherwise; colleague evaluation is average of zji on 4-point scale). There are also some control variables (banker’s job rank, colleague evaluation, years in organization, minority, works at headquarters).
Measures of network advantage:
Network eigenvector measuring centrality/status (si = ∑j zjisj) Network constraint measuring lack of access to structural holes (ci = ∑j cij, where cji = (pij + ∑k pikpkj)2, k ≠ i, j and pij is the proportion of i’s relations that are with j).
Measures of network volatility: Positive churn dummy (1 if middle third of percent change in contacts in 4 years) Wide variation dummy (1 if above-median standard deviation in network score over time) Trend up dummy (1 if positive slope on banker’s over time) Negative trend dummy (1 if negative slope on banker’s over time) Reversals dummy (1 if slope of change in banker’s scores reverses during the 4 years)
Figure 4.Illustrative Variation, Trend, and Reversal
Reversal refers to change this year contradicted next year, as when banker status decreases this year after increasing last year (e.g., bankers C and D). No reversals means that banker status was stable from year to year (e.g., banker E), or changed in one direction (e.g., banker A’s increasing status).
0.0
0.5
1.0
1.5
2.0
2.5
3.0 A. Trending-Positive Banker, initially on periphery (1.86 mean over time, 1.19 SD)
B. One-Reversal Banker (1.98 mean status over time, .48 SD)
C. Two-Reversal Banker (1.20 mean over time, .73 SD)
D. Two-Reversal Banker (1.97 mean over time, .64 SD)
E. No-Change Banker (.74 mean over time, .07 SD)
F. One-Reversal Banker (.62 mean over time, .20 SD)
G. Trending-Negative Banker, initially central (.84 mean over time, .78 SD)
Ban
ker S
tatu
s w
ithin
Ann
ual N
etw
ork
YearOne
YearTwo
YearFour
YearThree
Ave
rage
Ban
ker
Sta
tus
acro
ss th
e A
nnua
l Net
wor
ks
A
E
D
C
G
F
B
Table 4.
Compensation Returns to Network Advantage and Volatility
Volatility Level Adjustments (at median network advantage) Positive Churn in Contacts Wide Variation in Advantage Positive Trend in Advantage Negative Trend in Advantage Reversal in Advantage Network Advantage Slopes (at low volatility) Network Status Network Constraint Volatility Slope Adjustments (Positive Churn)*(Network-Median) (Wide Variation)*(Network-Median) (Positive Trend)*(Network-Median) (Negative Trend)*(Network-Median) (Reversal)*(Network-Median) Intercept R2
NOTE — OLS regression coefficients are presented with standard errors in parentheses. Z-score compensation, network status, and network constraint are average annual scores as in Models III and IV in Table 2. Means, standard deviations, and correlations for Models VII and VIII are given in Appendix B, Tables B2 and B3 respectively. All models include the seven Table 2 control variables for job rank, colleague evaluation, years with organization, minority, and US headquarters. Volatility variable “Positive Churn” equals one for bankers outside the shaded cells in Table 3 indicating stability traps. Level adjustments show change in compensation associated with the five binary volatility measures defined in the text. Slope adjustments are interaction terms between volatility variables and network advantage as a deviation from its median. * p < .05, ** p ≤ .01
III
-.00 .07 .02
-.07 -.00
.06 —
-.02 .15 .13 .22 .38
-.67 .76
(.01) (.08) (.08) (.10) (.08) (.14) — (.16) (.12) (.18) (.15) (.14)**
.47 —
-.91 .74
(.05)** —
VII Network Status Predictions Network Constraint Predictions
-.00
.25 —
.31
-.79 .76
(.06) (.07)** — (.08)**
VIII IV
.10
.10 -.03 -.19 .07
— -.25
-.16 .11
-.10 .20
-.31
.33
.71
(.10) (.10) (.12) (.11) (.08) — (.13)* (.18) (.18) (.22) (.19) (.15)*
— -.41
.78
.70
— (.08)**
IX
.11
— -.28
-.33
.33
.71
(.06) — (.09)** (.12)**
X
Path Dependent Network Advantage
Puzzle: Individual Differences
in Returns to Network Advantage
Volatility Treated as Noise
Volatility Treated as Signal
Conclusions & Serial Closure
Three Conclusions and an Inference
(1) Network volatility does not affect performance directly. Compensation is neither higher nor lower for bankers in volatile networks — holding constant level of network advantage and allowing for slope adjustments. (2) Volatility has its effect by enhancing a person’s returns to level of network advantage. Adding volatility to the predictions did not strengthen the network effect. It disaggregated the effect into a portion due to level of advantage and a portion due to volatility. Just holding constant the corrosive effect of stability traps, the split is about equal between network advantage level versus volatility. (3) The network volatility that protects against stability traps is a specific kind. It is not making new contacts in place of old. Churn should be moderate, about what is typical for the median person. It is not a matter of trend. Trend has no association with performance beyond level of network advantage. The variation associated with compensation is reversal: a pattern in which advantage is lost then regained, or gained then lost. Bankers who go through network reversals enjoy significantly higher returns to their level of network advantage. In fact, compensation has no association with level of network advantage for the bankers who failed to experience a reversal during the four years. Our inference from the analysis is that a substantial portion of network advantage is path dependent — advantage is about level of advantage, but it is also about how one’s level of advantage developed. In contrast to the positive image of continuous access to structural holes, the “serial closure” hypothesis is that advantage depends on discontinuous access.
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
Bob Bob Bob Bob Bob
February OctoberJune AugustApril
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
1
3
2
10
1112
65
498
7
Deb Deb Deb Deb Deb
1
3
11
58
Bob
NonRedundantContacts(thinsolidline)
Network Density& Constraint
(bold lineis constraint)
Bob IsAlways aBroker
Deb Buildsvia SerialClosure
(Metricsoscillatethrough
reversals)
December Network Survey
1
3
11
58
Deb
Figure 5. Broker Network Can Result from Always
Being a Broker, or from Serial Closure
Table 6. Returns to Brokerage Are Higher for Bankers Who Experienced Network Reversals
NOTE — Rows distinguish bankers by reversals in network status and constraint using the two dummy “reversal” variables in Table 4. Columns distinguish the bankers by network constraint. The third of bankers with the lowest average constraint scores over time are in the “High” access column. The third with the highest average constraint scores are in the “Low” access column. Cell entries are mean z-score compensation adjusted for the seven control variables in Table 2 (Model IV). Number of bankers is given in parentheses. Asterisks indicate compensation averages in the nine table cells that are significantly high or low (log network constraint in Model IV in Table 2 was replaced by 8 dummy variables distinguishing the nine cells using the middle cell as a reference category).
* P < .01 ** P < .001
Network Reversals
Access to Structural Holes
High (Brokers) Average Low
Yes (banker reversals in status AND
access to structural holes, n = 76)
1.11** (25)
-.16 (27)
-.25 (24)
Probably (banker reversal in status OR
access to structural holes, n = 143)
.34* (47)
-.13 (48)
-.11 (48)
No (no reversals, n = 127)
-.39 (39)
-.02 (41)
-.11 (47)
All Bankers .26 (111)
-.10 (116)
-.14 (119)