Sameer B. Srivastava University of California Berkeleyhaas.berkeley.edu/faculty/papers/srivastava_restructuring.pdf · 2 Organizational Restructuring and Social Capital Activation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Organizational Restructuring and Social Capital Activation1
Sameer B. Srivastava
University of California – Berkeley
May, 2013
Keywords: organizational restructuring; social capital; uncertainty;
coping; organizational change; network dynamics
1Direct all correspondence to Sameer B. Srivastava, University of California – Berkeley, Haas School of Business,
545 Student Services, #1900, Berkeley, CA 94720-1900; [email protected]; 510-643-5922. I thank
Jason Beckfield, Chris Muller, Frank Dobbin, Roberto Fernandez, Heather Haveman, Ed Lawler, Ming Leung, Peter
Marsden, Erin Reid, Misiek Piskorski, Eliot Sherman, Toby Stuart, András Tilcsik, Cat Turco and participants of the
MIT-Harvard Economic Sociology Seminar, the Dobbin Research Group, the MIT Economic Sociology Working
Group, and Harvard’s Work, Organizations, and Markets Seminar for helpful comments and suggestions on prior
H2: An increase in job-related uncertainty stemming from organizational
restructuring leads people to activate fewer ties to colleagues who are in their own
subunit.
METHODS
Research Setting
A major information services company, hereafter referred to as InfoCo, served as the research
site for the study. Declining financial performance led InfoCo’s management team to undertake a
major restructuring. The restructuring involved the creation of new subunits, such as a global
marketing function; the combination of existing subunits, such as “solution lines” that integrated
product development and marketing; and the elimination of certain other subunits and job roles.
11
It also resulted in changes in the work group structure. The broad thrust of the changes –
centralization of certain functions, regionalization of others, and downsizing to reduce costs –
was largely consistent with the forms of restructuring experienced by workers across a wide
range of US corporations (Capelli, 2008; Cappelli et al., 1997; Osterman, 2000).
Study Subjects
The study included all 114 US-based members of the InfoCo’s extended leadership group.
Because members of this group all had pre-existing ties to one another, they were well-suited to
a study of social capital activation. The group was mostly male (67.5%) and white (84.2%). A
significant portion (58%) worked in one of two main regional offices; the rest were distributed
among smaller sites. They spanned three salary grades (in ascending rank): 7.5% were
“operational leaders,” 80.3% were “tactical leaders,” and 12.2% were “executive leaders.”
The company granted access to data on the extended leadership group but did not permit
data collection on lower level employees (given concerns about how rank-and-file employees
might react if they learned that their email traffic – even if stripped of identifying information
and content – was being tracked for research purposes). The focus on a relatively senior
employee population raises questions about the generalizability of findings from this study. In
this particular case, however, the restructuring was implemented in a manner that mitigates this
concern. By all accounts, the CEO kept details of the restructuring close to his vest until the
changes were announced. Thus, even this senior group of employees experienced a discrete
period of heightened job-related uncertainty. The qualitative evidence suggests that most
understood the strategic rationale for the change. Thus, most did not face significant strategic
uncertainty during this time. In the period before the new organizational structure was
announced, many seemed to know that some form of restructuring was imminent but remarkably
12
few knew the details (e.g., which subunits would be created, merged, or dissolved; who would
report to whom) or what it would mean for them personally. So in the period before the
announcement, they faced considerable structural uncertainty but limited job-related uncertainty.
Once the new structure was announced, the structural uncertainty was mostly resolved, but there
was a spike in job-related uncertainty as people scrambled to understand what the new structure
would mean for them. Virtually everyone experienced this increase in job-related uncertainty,
and it is this form of uncertainty that is most likely to produce the coping responses theorized
above (Bordia et al., 2004). Indeed, by the time restructuring concluded, many study participants
had been significantly affected: 43 (37.7%) had a change in supervisor, 15 (13.6%) moved to a
different InfoCo division, and 13 (11.4%) exited the company.4 (Some experienced more than
one of these changes.) Those who experienced these changes did not learn of them until after the
initial announcement, and those who were not affected did not know so until considerably later.5
Semi-structured interviews (see details below) provided further evidence that these
individuals experienced an increase job-related uncertainty. As one marketing director reported,
“The announcement happens, and then I get a call from HR and the guy who was my boss at the
time. They say, ‘We’re eliminating your role, and you’re not going to get job you thought you
were going to get.’ Then they offered me another job that I really didn’t want. I was stunned.”
Similarly, a marketing support director, who participated in the process of identifying which
people were let go and who was selected for what open position, described the period as follows:
It was just a terrible, terrible time….All of the leaders were given a certain
number of slots to fill. We had to go through a process of assessing and ranking
4There were no statistically significant differences between those who exited and those who stayed on observable
characteristics such as gender, tenure, or salary band. 5During the time that archival data were being collected, this group was also unaware that it was involved in a
research study. Knowledge of the study was kept to a small group (e.g., CEO, head of HR) to minimize distraction.
13
people – for example, eleven people might be ranked for a job role with ten open
slots. The eleventh person was laid off. If the job role was redefined, we had to
tell all incumbents that they were laid off and had to interview to get their job
back. Everyone was feeling insecure. In my area, news leaked that 40% of the
staff would be let go.
In sum, the individuals included in this study were well-suited to the study of social capital
activation as a coping response to job-related uncertainty.
Data Collection
Four kinds of archival data were collected for the study: (1) internal communication memos,
which were used to construct the timeline of restructuring events; (2) email logs (spanning a
period of 40 weeks) of the extended leadership group6; (3) extracts of InfoCo’s email distribution
lists, which were used to identify shared work groups among employees (based on list co-
membership in a given week); and (4) extracts from InfoCo’s human resources system.
Internal communication memos (and semi-structured interviews) indicated that the period
of greatest job-related uncertainty commenced in Week 9, when the CEO released the first of
several communications that provided details of the new organizational structure. Additional
memos – announcing the formation of new subunits, the consolidation of other units, and the
appointment and departure of personnel – were sent intermittently until Week 18. All changes to
the organizational structure had been made, key positions had been filled, and departing
employees had all exited by Week 18. Thus, Weeks 9 to 18 represented the period of heightened
job-related uncertainty (see Figure 1).
– Figure 1 about here –
6Prior research indicates that this time period is appropriate for the study of employee reactions to restructuring
(Brockner, Tyler, and Cooper-Schneider 1992; Shah 2000).
14
Email logs represented a second key information source. Analyses of email
communication are becoming increasingly common in organizational research (Allatta & Singh,
2011; Hinds & Kiesler, 1995; Kossinets & Watts, 2006). Consistent with the ethical standards
used in prior studies (Borgatti & Molina, 2005; Kadushin, 2005), identifying information (e.g.,
email addresses) was encrypted using an irreversible algorithm, email logs did not contain
message content, and only messages internal to InfoCo were collected. Although these choices
helped protect the privacy of study subjects and the confidentiality of company data, they also
restricted the ability to analyze the meaning embedded in email content.
Email data of this kind have several advantages over traditional network surveys. First,
they can be collected unobtrusively, which can be useful in observing network dynamics during
a politically sensitive time such as an organizational restructuring. Next, they provide a window
into peripheral ties, which network surveys typically do not seek to measure. They can also yield
more reliable indicators of interaction than surveys, which can suffer from various forms of
recall and self-report bias (Marsden, 2011). For longitudinal network analysis, they have the
added benefit of allowing for consistent data collection over time – for example, by avoiding
measurement error that can arise from variability in interviewer techniques and eliminating the
sample attrition that can occur in repeated surveys. At InfoCo, interviewees reported that they
routinely used company email even for personal communication, including messages sent via
personal digital assistants. At the time of the restructuring (early 2008), it was uncommon for
InfoCo employees to use instant messaging or personal email services at work.
These benefits are counterbalanced by certain limitations. First, the trace of email
communication does not always signify purposive interaction. For example, email messages can
sometimes be automatically generated, sent to pre-determined distribution lists, or mindlessly
15
copied to peripheral actors. To address these shortcomings, emails including the phrase “Out of
Office” in the subject line and mass emails (those sent to more than one recipient) were
excluded.7 Next, email logs contain only a subset of communications. At InfoCo, the email
system was linked to electronic calendars and therefore included a record of all formally
scheduled meetings. Nevertheless, given that people likely communicated through a mix of
scheduled meetings, email exchanges, and informal, unscheduled communication, email logs
provide an incomplete window into the communications that took place in this period. Because
more sensitive communication was likely to take place in informal face-to-face and phone
communication, email logs likely represent a conservative indicator of social capital activation in
response to job-related uncertainty.
In addition to email logs, I collected email distribution lists. Just as the choice to use
email data involves tradeoffs, so too does the decision to use email distribution lists to locate
individuals in the space of cross-unit work groups. Widely used across organizations, distribution
lists encapsulate the myriad collective units that exist within an organization but that are often
poorly documented or kept updated. Because list names were encrypted in the data available
from InfoCo, it was not possible to distinguish among different list types. Semi-structured
interviews suggested that InfoCo’s lists primarily corresponded to work groups rather than
subunits. Moreover, in a typical week during the observation period, there were over 2,300 active
distribution lists – far more than the number of subunits to which subjects belonged. Although
the lists obtained from InfoCo likely reflected a mix of cross-unit work groups, work groups
nested within subunits, and subunits themselves, findings from other studies that have used email
7The results reported below were robust to different mass email thresholds – for example, including messages sent to
up to five recipients.
16
distribution lists and had access to list names suggest that the vast majority represent work
groups (Liu, Srivastava, & Stuart, 2013).
To derive from these data a measure of distance in the space of cross-unit work groups,
two additional steps were taken (see below for details on the measure): weighting lists by size
(assuming that small lists are more likely to reflect meaningful work groups in which strong,
trustworthy ties form while large lists are more likely to indicate diffuse work groups or
subunits) and weighting lists by the diversity of subunits represented on the list.
Finally, repeated extracts from InfoCo’s human resource systems were used to construct
time-varying measures of positions in subunits and cross-unit work groups and to identify
sociodemographic characteristics.
Measures
The response variable was a count of the number of one-to-one email messages exchanged in a
given week, t, between a dyadic pair, i and j. Explanatory variables included Same Departmentt
(set to 1 if i and j were in the same department in week t and to 0 otherwise), Both Men, Both
Women, and Distance in Cross-Unit Work Group Spacet.8 In principle, work groups can overlap
significantly with subunits, for example if they are entirely nested within subunits. The measure
of distance in cross-unit work group space was therefore adjusted by the level of subunit
diversity represented on each list (Blau, 1977) and by list size.9 The resulting measure is a
variant of Jaccard’s distance, a widely used distance measure that has a theoretical range from 0
to 1 (Sneath & Sokal, 1973)10
:
8The restructuring affected both subunit and cross-unit work group structure – for example, 6% of dyads
experienced a change in reporting relationship and about 5% of dyads experienced more than a half standard
deviation change in distance in cross-unit work group space. 9The diversity measure was based on the eleven divisional groups (collections of departments) that were represented
on lists. More diverse lists included people from a broader range of subunits than less diverse lists. 10
All continuous covariates are mean-centered in regression analyses. The results reported below were robust to the
use of alternative distance measures, such as one based on Dice’s coefficient (Dice 1945).
17
∑
∑
∑
∑
Where: i, j index members of the dyad
k indexes distribution lists
si,j = 1 if i and j belong to list k; 0 otherwise
mi = 1 if i belongs to list k; 0 otherwise
mj = 1 if j belongs to list k; 0 otherwise
dk = 1 – sum of squared proportions of divisional groups represented on list k
Nk = size of list k
Only 5.2% of dyads that were below the median of this measure were also in the same
department, suggesting that this measure is more or less orthogonal to formal structure.
To identify increases or decreases in social capital activation during the period of
heightened job-related uncertainty, an indicator, Uncertaintyt, set to 1 for Weeks 9 through 18
(from the time structural changes were first announced to when they were fully implemented),
and the following interaction terms were used: Uncertaintyt x Same Departmentt and
Uncertaintyt x Distance in Cross-Unit Work Group Spacet. Hypotheses 1 and 2 predict
significant and negative coefficients, respectively, for the linear combinations: (1) Uncertaintyt
+Uncertaintyt x Distance in Cross-Unit Work Group Spacet and (2) Uncertaintyt +Uncertaintyt x
Same Departmentt.
Given the well-documented tendency in social networks toward various forms of
homophily and propinquity (McPherson, Smith-Lovin, & Cook, 2001), the following controls
were also included: (1) Same Locationt, set to 1 for dyads in the same building and floor (e.g.,
Allen, 1977); (2) Same Salary Gradet, set to 1 for dyads at the same hierarchical rank (e.g., Han,
1996); (3) Same Cohort, set to 1 for dyads hired within one year of each other (e.g., Wagner,
Pfeffer, & O'Reilly, 1984); (4) Same Age, set to 1 for dyads with an age difference of less than
18
four years (e.g., Burt, 2000)11
; (5) Same Ethnicity, set to 1 for dyads listed in the human resource
system as having the same ethnicity – White, Asian, African American, Latino / Hispanic, or
Other (e.g., Mehra, Kilduff, & Brass, 1998); and same gender (Ibarra, 1992a), which was
decomposed into (6) Both Women and (7) Both Men.
Estimation
A dyad-level panel data set of email messages exchanged between i and j in week, t, was
constructed. Analyses of such data must contend with the clustering (i.e., non-independence) of
observations. Error terms in regression analyses will be correlated across observations, a problem
referred to as network autocorrelation. The failure to control for clustering can lead to under-
estimated standard errors and over-rejection of hypothesis tests. To address this issue, a variance
estimator that enables cluster-robust inference when there is multi-way clustering was
implemented (Cameron, Gelbach, & Miller, 2011). This situation arises when – as in this study –
there is clustering at both the cross-sectional and temporal levels. In the case of two-way
clustering, the technique produces three different variance matrices: for the first dimension, for
the second dimension, and for the intersection of the two. The first two matrices are added
together and third subtracted. In the case of three-way clustering, the analogous technique results
in the creation and combination of seven one-way cluster robust variance matrices.12
Thus, I
estimated Poisson regressions with standard errors clustered by sender, by receiver, and by week.
11
A difference of less than four years was also used to define same age in Burt (2000). The results reported below
did not change materially when other age difference cut-offs were used instead. 12
Each of the first three matrices clusters in one dimension. Because some observation pairs are in the same two-
dimensional cluster, considering only these three matrices would result in double counting. So matrices that cluster
on the three combinations of two dimensions are then subtracted. This eliminates double counting but does not
account for pairs that share the same cluster in all three dimensions. So the seventh matrix, which clusters on pairs
sharing the same cluster in all dimensions, is added back (see Cameron, Gelbach, and Miller [2011: 10-11]). This
technique, which also controls for potential over- or under-dispersion in the data, was implemented in STATA using
the “clus_nway” script (Kleinbaum, Stuart, and Tushman 2013).
19
This technique is appropriate for the analysis of dyadic network data, including panel
data (Cameron et al., 2011; Kleinbaum, Stuart, & Tushman, 2013). In simulation studies
(Lindgren, 2010), it performs at least as well as an alternative approach: Multiple Regression
Quadratic Assignment Procedure (MRQAP) with double semi-partialing (DSP) (Dekker,
Krackhardt, & Snijders, 2007). Moreover, it is computationally faster than MRQAP with DSP in
dealing with large data sets (Kleinbaum et al., 2013). Unlike stochastic actor-based models (e.g.,
those estimated using SIENA), this approach does not account for higher-order dependence
structures (e.g., transitive triplets) that may exist in the data. However, stochastic actor-based
models assume a dichotomous response variable and are appropriate when the number of
observations of a network is small – usually less than ten (Snijders, van de Bunt, & Steglich,
2010). Thus, they are not appropriate for this data set, in which the response variable is a count
of email messages exchanged over 40 weeks.
As a robustness check, an alternative approach – estimating Poisson regressions with
fixed effects for every sender and every receiver in the study – was also implemented (Mizruchi,
1989; Reagans & McEvily, 2003). This approach shifts the potentially autocorrelated
disturbances out of the residuals and yields consistent and efficient estimates (Mizruchi, 1989:
421). It also accounts for all time-invariant, unobserved differences among study participants.
The results reported below were materially unchanged with this alternative approach.
RESULTS
Quantitative Analysis
Table 1 reports descriptive statistics and a correlation matrix. As expected, there is a positive
correlation between messages exchanged and various measures of similarity between dyads (e.g.,
20
Same Departmentt and Same Locationt) and a negative correlation between messages exchanged
and Distance in Cross-Unit Work Group Spacet.
– Table 1 about here –
Table 2 provides a comparison of aggregate communication patterns between the periods
of uncertainty and relative stability. Although there was a slight increase in aggregate
communication volume during the weeks of uncertainty, this change was not statistically
significant. Consistent with Hypothesis 1, the correlation between Distance in Cross-Unit Work
Group Spacet and messages exchanged was -0.101 in the period of relative stability and -0.122 in
the period of uncertainty. Similarly, consistent with Hypothesis 2, the proportion of messages
sent between colleagues in the same department was 0.568 in the period of relative stability and
0.516 in the period of uncertainty (p<.001).
– Table 2 about here –
Table 3 reports the results of the regression analyses used for hypothesis testing. Model 1
depicts results from the baseline model. Same Locationt, Same Departmentt, and Both Women
have positive and significant coefficients, while Distance in Cross-Unit Work Group Spacet has
a negative and significant coefficient. The coefficients for Same Departmentt and Distance in
Cross-Unit Work Group Spacet are consistent with prior research indicating a tendency for the
formal and quasi-formal structure to influence communications in the workplace (Allen, 1977;
Hinds & Kiesler, 1995; Srivastava & Banaji, 2011). The positive coefficient for Both Women is
consistent with prior research on homophily in the workplace (Ibarra, 1992a; Kleinbaum et al.,
2013; Lincoln & Miller, 1979), though there is no evidence in this setting for homophily among
men. Unlike prior research (Han, 1996), Same Salary Gradet in this setting has a significant and
21
negative coefficient, perhaps because the senior-most leaders among study subjects were
working more within their divisions across vertical levels than with their peers in other divisions.
Models 2 and 3 test for the significance of the relevant interaction terms: Uncertaintyt x
Distance in Cross-Unit Work Group Spacet and Uncertaintyt x Same Departmentt. In Model 2,
Uncertaintyt represents the effects of job-related uncertainty on colleagues at the mean distance
in cross-unit work group space. It has a slightly negative but not significant coefficient. The
interaction term Uncertaintyt x Distance in Cross-Unit Work Group Spacet, has a negative and
significant coefficient (beta=-0.488; p<.05). In Model 3, Uncertaintyt represents the effects of
job-related uncertainty on colleagues in different departments. It has a slightly positive but not
significant coefficient. The interaction term, Uncertaintyt x Same Departmentt, has a negative
and significant coefficient (beta=-0.216; p<.05). Taken together, Models 2 and 3 indicate
significant uncertainty interaction effects that are consistent with Hypotheses 1 and 2.
Model 4 represents the fully specified model used to conduct more specific hypothesis
tests. The relevant interaction terms are significant and of the expected sign: Uncertaintyt x
Distance in Cross-Unit Work Group Spacet (beta=-0.866; p<.001) and Uncertaintyt x Same
Departmentt (beta=-0.295; p<.01). The linear combination, Uncertaintyt + Uncertaintyt x
Distance in Cross-Unit Work Group Spacet, is negative and significant (beta=-0.789; p<.01).
Thus, there is support for Hypothesis 1. That is, when job-related uncertainty is heightened,
increasing distance in cross-unit work group space tends to suppress the number of messages
exchanged between colleagues. Conversely, colleagues who are more proximate in work group
space are apt to exchange more messages with one another when they face an increase in job-
related uncertainty. Similarly, because the linear combination, Uncertaintyt + Uncertaintyt x
Same Departmentt, is also negative and significant (beta=-0.218, p<.05), there is support for
22
Hypothesis 2. That is, when job-related uncertainty increases, colleagues in the same department
tend to exchange fewer messages with one another.
– Table 3 about here –
Considering that changes in email communication probably represent a conservative
indicator of shifts in social capital activation, these effects were sizable: In the period of
heightened job-related uncertainty relative to stability, there was a 6% increase in the predicted
number of messages among dyads at the 5th
percentile of Distance in Cross-Unit Work Group
Spacet, a 9% decrease in the predicted number of messages among dyads at the 50th
percentile,
and an 11% decrease in the predicted number of messages among dyads at the 95th
percentile.
Turning to subunits, there was a 14% decline in the predicted number of messages exchanged
between colleagues in the same department and a 7% increase in the predicted number of
messages exchanged between colleagues in different departments.
Qualitative Analysis
To help address the limitations of the archival data sources described above, I conducted
supplemental semi-structured interviews with 23 InfoCo employees, who were selected in
consultation with human resource professionals. They were chosen from sub-samples believed to
have experienced higher and lower levels of uncertainty during the restructuring (based on their
job roles) but who remained employed at the firm. Legal concerns kept the company from
granting me access to those who had exited. The interviews, which occurred several months after
the restructuring concluded, lasted between 30 and 45 minutes and were recorded and
transcribed. The 23 semi-structured interviews included 16 with a subset of the individuals
whose email data were analyzed and 7 with human resource professionals who helped to
implement the restructuring. Interviews lasted between 30 and 45 minutes and were recorded and
23
transcribed. The purpose of the interviews was to validate the timeline of events, assess the
nature of uncertainty people experienced, understand how and why they activated their networks
during the restructuring, and determine how they used electronic communication media.
Each set of interviews was coded to identify whether or not the respondent reported: (1)
feeling uncertain during the restructuring; (2) experiencing strategic uncertainty; (3)
experiencing structural uncertainty; (4) experiencing job-related uncertainty; (5) coping with
uncertainty through social capital activation; (6) activating strong tie contacts; (7) activating
weak tie contacts; (8) activating ties within the respondent’s subunit; and (9) activating ties
outside the respondent’s subunit. These categories were not mutually exclusive – for example, a
person could report activating both strong tie and weak tie contacts. Table 4 below provides
illustrative quotations that were coded as belonging to each of these categories.
– Table 4 about here –
Of the 16 individuals whose responses were coded in this manner, 9 were men and 6
were women. 15 of the 16 reported experiencing some form of uncertainty as a result of