Strategic Management Journal

Strat. Mgmt. J., 33: 611–632 (2012)

Published online EarlyView in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/smj.1969
Received 15 January 2010; Final revision received 10 June 2011

ORGANIZATIONAL STRUCTURE AS A DETERMINANT

OF PERFORMANCE: EVIDENCE FROM MUTUAL
FUNDS
FELIPE A. CSASZAR*
Stephen M. Ross School of Business, University of Michigan, Ann Arbor, Michigan,
U.S.A.

This article develops and tests a model of how organizational structure influences organizational
performance. Organizational structure, conceptualized as the decision-making structure among
a group of individuals, is shown to affect the number of initiatives pursued by organizations and
the omission and commission errors (Type I and II errors, respectively) made by organizations.
The empirical setting is more than 150,000 stock-picking decisions made by 609 mutual funds.
Mutual funds offer an ideal and rare setting to test the theory, since there are detailed records on
the projects they face, the decisions they make, and the outcomes of these decisions. The study’s
independent variable, organizational structure, is coded based on fund management descriptions
made by Morningstar, and estimates of the omission and commission errors are computed
by a novel technique that uses bootstrapping to create measures that are comparable across
funds. The findings suggest that organizational structure has relevant and predictable effects
on a wide range of organizations. In particular, the article shows empirically that increasing
the consensus threshold required by a committee in charge of selecting projects leads to more
omission errors, fewer commission errors, and fewer approved projects. Applications include
designing organizations that achieve a given mix of exploration and exploitation, as well as
predicting the consequences of centralization and decentralization. This work constitutes the
first large-sample empirical test of the model by Sah and Stiglitz (1986). Copyright  2012 John
Wiley & Sons, Ltd.

INTRODUCTION screening biases are there in an organization?’

Yet with few exceptions, questions of this sort
There is a long-standing concern that the strategy remain largely unexplored in the strategy literature
literature needs a better understanding of the rela- (Rumelt, Schendel, and Teece, 1994: 42). The lack
tionship between organizational structure and per- of knowledge about how decision-making structure
formance. This concern goes back at least to Cyert affects organizational performance surfaces repeat-
and March (1963: 21), who posed the following edly in different areas of management. For exam-
questions when motivating their theoretical enter- ple, in the context of ambidextrous organizations,
prise: ‘what happens to information as it is pro- Raisch and Birkinshaw (2008: 380) note that ‘far
cessed through the organization? What predictable less research has traditionally been devoted to how
organizations achieve organizational ambidexter-
Keywords: organization design; omission and commis- ity,’ and in the context of R&D organization,
sion errors; exploration/exploitation; centralization and Argyres and Silverman (2004: 929) show surprise
decentralization; decision making ‘that so little research has addressed the issue of
*Correspondence to: Felipe A. Csaszar, Stephen M. Ross School
of Business, University of Michigan, 701 Tappan Ave. R4336, how internal R&D organization affects the direc-
Ann Arbor, MI 48 109, U.S.A. E-mail: fcsaszar@umich.edu tions and impact of technological innovation by

Copyright  2012 John Wiley & Sons, Ltd.

612 F. A. Csaszar

multidivisional firms.’ These observations are con- committees are used in many settings relevant to
gruent with the view that organization design—the strategic decision making, such as boards of direc-
field specifically devoted to studying the links tors, top management teams, finance committees,
between environment, organizational structure, and and investment teams.
organizational outcomes—is, in many respects, an From a theoretical standpoint, confirming the
emerging field despite its long history (Daft and Sah and Stiglitz model is of special interest to
Lewin, 1993; Zenger and Hesterly, 1997; Foss, organization design, as their theory provides a
2003). parsimonious mechanism to explain how micro
This article contributes to a better understanding decisions (individual choices) are aggregated by
of the relationship between organizational struc- an organizational architecture into macro behav-
ture and organizational performance by providing iors (organization-level performance). In fact, by
the first large sample empirical test of the the- separating performance into omission and commis-
ory developed by Sah and Stiglitz (1986). Among sion errors and then linking structural choices to
other predictions, this theory establishes a causal the occurrence of these errors, Sah and Stiglitz’s
link between the structure of a decision-making model sheds light on such organization design
committee and the number of omission and com- issues as the implications of centralization and
mission errors the committee will make. Errors decentralization and how organizations can pursue
of omission and commission (which are equiv- exploration and exploitation.
alent, respectively, to Type I and II errors in It is especially important to test the Sah and
statistics) correspond to missing good choices Stiglitz model because it has spawned a large num-
(omissions) and pursuing bad choices (commis- ber of descendants.1 Thus, all these works rely
sions). Specifically, Sah and Stiglitz’s (1986) critically on the untested validity of its prede-
theory predicts that committees with a high con- cessor. Despite there being many ways in which
sensus level (e.g., requiring unanimous approval) the work of Sah and Stiglitz (1986) can illumi-
will make relatively few commission errors but nate managerial phenomena, few of the references
many omission errors. In contrast, committees with have come from the management field. The occa-
a low consensus level (e.g., requiring the approval sional exceptions include work on mergers and
of just one of its members) will exhibit the oppo- acquisitions (Gulati and Higgins, 2003; Puranam,
site behavior: they will make few omission errors Powell, and Singh, 2006), venture capital (Lerner,
but many commission errors. Additionally, the the- 1994), technological choices (Garud, Nayyar, and
ory predicts that the higher the consensus level, the Shapira, 1997), the implications of alternative eval-
fewer projects that will be pursued by the commit- uation on search (Knudsen and Levinthal, 2007),
tee. This article finds empirical support for these and analyzing the errors of more complex organi-
three predictions of the Sah and Stiglitz model. zational structures (Csaszar, 2009; Christensen and
The Sah and Stiglitz model makes predictions Knudsen, 2010). Perhaps the lack of empirical val-
regarding organizational performance (e.g., ex- idation explains why few of the references to the
pected omission and commission errors of a given work of Sah and Stiglitz (1986) have come from
organization design). Yet in many cases, by using a the, largely empirical, management field.
contingency or ‘fit’-type of logic, one may extend Sah and Stiglitz’s theory is so simple—it is
the reach of their model to infer predictions regard- based solely on the probabilities that different vot-
ing competitive performance (e.g., profitability in ing rules have of vetoing projects—that one may
the face of competition). For instance, all other be tempted to suppose their predictions are obvi-
things being equal, if in a given competitive situ- ously true in organizations resembling those in
ation omission errors are costlier than commission the theory. But without empirical validation, it
errors, then a firm whose organizational design is not obvious whether Sah and Stiglitz’s (1986)
produces fewer omission errors will be more prof- terse description of organizations has predictive
itable than one whose design produces more omis- value. For example, it could be that their idea
sion errors. of looking at organizations as veto mechanisms
From a practical standpoint, empirically validat-
ing Sah and Stiglitz’s theory is relevant because
it has performance implications for committees, a 1
At the time of this writing, more than 180 citations according
widespread decision-making structure. Moreover, to ISI Web of Science.

Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 613

greatly oversimplifies the communication capac- findings reported here may occur elsewhere is that
ity of individuals or the information aggregation mutual funds are among the most stringent settings
rules actually used by organizations. Alternatively, imaginable: stock returns are eminently random
it could be that the process described by Sah (Fama, 1970), so it is hard to imagine that orga-
and Stiglitz (1986) does operate in organizations, nizational structure could affect any outcome of
but that its effects are minuscule when compared stock picking. Finding empirical support in this
with other concurrent effects, such as groupthink setting, therefore, suggests empirical support in
(Janis, 1972), herding (Bikhchandani, Hirshleifer, other, less stringent settings.
and Welch, 1992), power (Pfeffer, 1992), decision The next section of this article describes the
biases (Kahneman and Tversky, 1979), or mas- mechanism underlying the Sah and Stiglitz model;
sive inaccuracies in managerial decision making the following section connects the theory of Sah
(Starbuck and Mezias, 1996). In sum, determining and Stiglitz to the management literature. Then
the falsifiability of Sah and Stiglitz’s theory is an the tested hypotheses are presented, the empirical
empirical question. Science progresses by theory setting is described, and the results are presented.
building and theory testing, but Sah and Stiglitz’s Finally, the broader theoretical and managerial
theory-building effort has, until now, lacked an implications of this research are discussed.
accompanying theory-testing effort.
This article tests Sah and Stiglitz’s model by
using data on the decision-making structure and AN OVERVIEW OF THE SAH
on the omission and commission errors of mutual AND STIGLITZ MODEL FROM
funds. One could argue that Sah and Stiglitz’s the- AN ORGANIZATION DESIGN
ory has not yet been tested empirically because PERSPECTIVE
collecting information on organizational structure
and errors (particularly omissions) is difficult.
Luckily, mutual funds offer a rare opportunity to To understand the scope and applicability of Sah
observe all the required information: mutual funds’ and Stiglitz’s model, it is useful to start by describ-
decision-making structure (which squarely maps ing how it fits within the three fundamental themes
into the committees described by Sah and Stiglitz) of organization design: (1) organizational search or
is observable from analysts’ reports; and omission alternative generation (e.g., Rivkin and Siggelkow,
and commission errors can be measured by looking 2003; Ethiraj and Levinthal, 2004); (2) alternative
at the investment universe of each fund, the assets evaluation (e.g., Gavetti and Levinthal, 2000;
that each fund decides to buy and not to buy from Knudsen and Levinthal, 2007); and (3) execution
that investment universe, and the ex post return of or implementation (e.g., Hrebiniak and Joyce,
each asset. 1984; Galbraith and Kazanjian, 1986). In light of
The current study explores a particular aspect this characterization, the work of Sah and Stiglitz
of organization design using a particular setting. (1986) falls precisely in the category of alternative
Hence, two questions regarding generalizability evaluation.
emerge: how prevalent is the mechanism stud- Sah and Stiglitz (1986, 1988)2 model how com-
ied, and how representative is the setting. The mittees screen projects: that is, how effective
mechanism studied—voting—is certainly preva- they are at separating good projects from bad
lent, although it is unlikely to occur in such a ones. Although originally developed to compare
stylized way as described by Sah and Stiglitz’s the performance of central planning with that
model. In the real world it is probable that other of free markets, Sah and Stiglitz’s model can
phenomena (e.g., power, politics, herding) would shed light on a broader set of organizational
co-occur. It is interesting that, despite the pos- issues because many organizations use committee-
sibility that any number of factors influence the like structures. Examples include banks choosing
relationship between structure and organizational which loans to approve, venture capital firms and
performance, the results reported here are consis- mutual funds picking investments, movie studios
tent with Sah and Stiglitz’s parsimonious char-
acterization. Regarding the second question, on 2
The paper by Sah and Stiglitz (1988) generalizes Sah and
how representative is the mutual funds setting, one Stiglitz (1986) to committees with arbitrary size and consensus
characteristic of this setting that suggests that the levels. For succinctness, only the earlier paper is cited hereafter.

Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
614 F. A. Csaszar

judging scripts, hiring committees selecting candi- of the three individuals likes the project in order
dates, and top management teams deciding which for it to be approved.
strategic projects to pursue. If the number of decision makers on the com-
Sah and Stiglitz model committees as compris- mittee is fixed, the main predictions of Sah and
ing N decision makers, of which C must approve Stiglitz’s model are that, on average, lowering
a project for it to be approved by the commit- the consensus level leads to: (1) more approved
tee (C stands for ‘consensus level’). For example, projects, (2) fewer omission errors, and (3) more
an organization of three members that approves commission errors.
a project when any member decides to approve These three predictions can be explained using
it is represented by N = 3 and C = 1, or simply basic probability theory. The next example illus-
3/1; likewise, 2/2 denotes a two-member organiza- trates the effect of structure on the number of
tion that only approves projects for which there is projects approved—that is, Prediction 1 of the
consensus. An organization consisting of a single model. Imagine two decision makers, each with
individual is denoted 1/1. a 50 percent chance of approving a project, fac-
The model assumes that individuals are fallible ing 100 projects to be screened. If unanimity is
(i.e., individuals perceive reality in a noisy fash- required (i.e., if both decision makers must agree
ion), homogeneous, and uncorrelated (i.e., noises that a project is good for it to be approved),
in perception are independent and identically dis- then they would approve on average 25 projects
tributed); it also assumes that projects are described (= 100 × 0.5 × 0.5). If instead the approval of just
by a single number (i.e., a project quality, which is one decision maker is required, then the orga-
imperfectly perceived). Like all models, this styl- nization would approve on average 75 projects
ized description of organizations leaves many phe- (conversely to the previous case, now each project
is accepted unless both decision makers reject
nomena outside of its scope, such as organizations
it: 100 − 25 = 75). Although the examples so far
whose task is different from screening projects,
have assumed that decision makers are homoge-
heterogeneity in ability, group dynamics such as
neous (i.e., in the example the two managers had
herding (Bikhchandani et al., 1992) or groupthink
the same 50% probability of approving a project),
(Janis, 1972), and, more generally, organizational
the model can be extended to accommodate het-
structures different from those describable in terms
erogeneous decision makers. But homogeneity is a
of N and C. Nonetheless, the model does permit
reasonable assumption in some settings, as when
one to focus on some basic mechanisms that are managers have similar training, and also more gen-
pervasive within organizations: how centralized or erally because, on average, heterogeneous settings
decentralized the decision process of an organiza- behave like homogeneous settings. (For example,
tion is and how many individuals are involved in it. a committee whose members have a probability of
The following examples illustrate how the model approval uniformly distributed between 40% and
captures these organizational characteristics. 60% behaves, on average, like a committee whose
For instance, a 3/3 could represent the decision- members have a 50% probability of approval.)
making process within a venture capital firm in The effect of structure on omission and com-
which the three partners must agree on any invest- mission errors (Predictions 2 and 3) can be simi-
ment; it could also represent a three-level hierarchy larly illustrated. Imagine that only 50 of the 100
in which projects received by a low-level employee projects are good and that each decision maker
must escalate up to the CEO for approval. In both has equal probability of accepting and reject-
examples, all three individuals must concur on the ing good and bad projects (i.e., there is a 25%
project’s viability before it is approved by the orga- chance of each decision maker either accepting
nization. In contrast, a 3/1 could represent either of a good project, accepting a bad project, reject-
the following decentralized structures: a firm with ing a good project, or rejecting a bad project).
three research engineers, any one of whom may Then, the unanimous committee (which Sah and
independently decide to pursue further research on Stiglitz (1986) call a hierarchy) would make on
a new technology; or it could represent a mutual average 6.25 commission errors (i.e., #projects
fund with three autonomous fund managers, any × probability that both decision makers accept a
one of whom may authorize the purchase of a secu- bad project = 100 × 0.25 × 0.25) and 43.75 omis-
rity. In these last two examples, it suffices that one sion errors (i.e., #projects × probability that any
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 615

of the decision makers reject a good project = the fundamental questions in the fields of strat-
100 × [0.25 + 0.25–0.252 ]). A committee requir- egy (Rumelt et al., 1994: 42) and organization
ing the approval of only one decision maker (which theory (Thompson, 1967), so it is no surprise
Sah and Stiglitz (1986) call a polyarchy) exhibits that it has been addressed extensively from sev-
the converse behavior: on average, it would make eral perspectives since old, even biblical (Van
6.25 omission errors (i.e., #projects × probability Fleet and Bedeian, 1977: 357), times. Therefore,
that both decision makers reject a good project) rather than attempting the impossible task of sum-
and 43.75 commission errors (i.e., #projects × marizing these literatures, this section presents a
probability that any of the decision makers accept broad overview with an emphasis on highlighting
a bad project).3 the main differences and similarities between cur-
Although not developed for this purpose, the rent and previous approaches. The review focuses
Sah and Stiglitz model can be used to analyze mainly on organizational structures whose build-
the effects of centralization and decentralization ing blocks are individuals (as in Cyert and March,
in organizations. The applicability of their model 1963), not business divisions (as in Chandler,
to these effects stems from the facts that ‘decision 1962). The focus on this kind of structures is
makers generally base their actions on estimates consistent with the Carnegie tradition understand-
formulated at other points in the organization’ ing of organizational structure as ‘the pattern of
(Cyert and March, 1963: 85) and that, in cen- communications and relations among a group of
tralized organizations, these estimates must ‘flow human beings, including the processes for making
up’ through more decision makers (before reach- and implementing decisions’ (Simon, 1947/1997:
ing the final decision maker) than in decentralized 18–19).
organizations (Robbins, 1990: 6). Thus, the infor- This section is organized in terms of three main
mation flow in centralized organizations resembles disciplines that share an interest in how organi-
that of hierarchies, whereas the information flow zational structure affects performance: organiza-
in decentralized organizations resembles that of tion design, organizational economics, and signal
polyarchies. In sum, the Sah and Stiglitz frame- detection theory.
work captures the dynamic of information passing
through more filters in centralized than in decen- Organization design
tralized organizations.
The appendix generalizes the examples given Important early attempts at understanding the rela-
here by developing a model that makes the same tionship between structure and performance are
qualitative predictions but under more general present in the work of Chandler (1962) and Barnard
assumptions (regarding number of decision mak- (1938). Their work, like most ensuing efforts in
ers, consensus level, individual screening abilities, organization design, took an information process-
and types of incoming projects). The model in the ing perspective. For instance, Chandler (1962:
appendix does not present new theory, but it does 69–70) cites memos from Du Pont’s reorganiza-
serve as a concise summary of the work of Sah tion in 1919 that were explicit about the role of
and Stiglitz (1986) that is useful for the purposes information processing: ‘the most efficient results
of this article. are obtained at least expense when we coordi-
nate related effort and segregate unrelated effort.’
Similarly, Barnard (1938: 215) mentions that ‘the
function of executives is to serve as channels of
THEORETICAL MOTIVATION
communication so far as communication must pass
though central positions.’
What are the effects of organizational structure Influenced by Barnard, Simon (1947/1997)
on organizational performance? This is among developed a more formal understanding of orga-
nizations as information processing devices com-
3
Calculations similar to these—that is, conjunctive (AND) and posed of boundedly rational individuals. Under
disjunctive (OR) probabilities—are commonly used in reliability this view, organizational structure plays a cen-
theory (e.g., Rausand and Høyland, 2004). Note that Sah (1991: tral role, as it defines how information flows
68) mentions a classic work on reliability (Moore and Shannon,
1956/1993) as an antecedent of the Sah and Stiglitz (1986) and is aggregated inside organizations, allowing
model. organizations to accomplish goals that would be
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
616 F. A. Csaszar

otherwise unattainable by any of its individual Zandt, 1999); the optimal organization of produc-
members. Led by Simon (1947/1997), Cyert and tion as a function of environmental uncertainty
March (1963) gave organizational structure a cen- (Cremer, 1980); the acquisition of knowledge by
tral place in the Carnegie tradition. However, with hierarchies with heterogeneous agents (Geanakop-
one exception (Cohen, March, and Olsen, 1972), los and Milgrom, 1991; Garicano, 2000); the extent
this tradition devoted most of its energies to deci- to which hierarchies can accommodate coordina-
sion making in the absence of concerns about tion and specialization (Hart and Moore, 2005);
organizational structure. In fact, organizational and the relative performance of such common
structure has recently been called a ‘forgotten pil- organizational forms as the M-form and the U-
lar’ of this tradition (Gavetti, Levinthal, and Oca- form (Harris and Raviv, 2002; Qian, Roland, and
sio, 2007: 525). Xu, 2006).
The work of Sah and Stiglitz (1986) is consis- Sah and Stiglitz (1986) contribute to orga-
tent with the information processing perspective nizational economics by introducing two new
that permeates the organization design literature: elements: modeling communication patterns as
it views the role of organizational structure as sequential or parallel circuits and measuring per-
a means to aggregate the information coming formance as omission and commission errors. The
from boundedly rational, fallible individuals. Yet literature that has descended from the work of
in addition to being consistent with organiza- Sah and Stiglitz (1986) has been primarily the-
tion design, it extends the information process- oretical and focused on voting (see the intro-
ing approach by offering a new set of predictions duction of Christensen and Knudsen (2010) for
about structure and types of projects approved. By a review), so the application of their model to
providing empirical support for Sah and Stiglitz’s organization design issues remains largely unex-
model, this article aims to pave the way for these plored.
predictions to be used fruitfully in organization
design. Signal detection theory
Almost without connection to the previous litera-
Organizational economics tures, a rich body of work that addresses many of
the same questions has been developed in social
Several models in organizational economics have psychology, under the label of signal detection
studied the effect of structure on performance. theory, and in the closely related theory of social
In broad terms, the organizational economics lit- decision schemes.
erature on structure can be divided into two Signal detection theory (Peterson, Birdsall, and
strands: incentives4 and information processing. Fox, 1954; Green and Swets, 1966; Macmillan and
The current article is directly related to this latter Creelman, 2004) provides a mathematical frame-
strand. work to analyze perception and decision making
The information processing strand of the organi- by fallible individuals. This theory conceptual-
zational literature has dealt mainly with the selec- izes decision makers as trying to detect a signal
tion of projects and the efficiency aspects of project in a noisy environment, and it provides a set of
implementation. Early works along these lines models, measures, and experiments to assess how
include Williamson (1967) on optimal hierarchy good decision makers are at detecting those sig-
size and Marschak and Radner (1972) on opti- nals under different settings. This theory was first
mal decision making by teams. More recent work used to measure the sensory acuity of military per-
has studied efficiency measures of hierarchies (e.g., sonnel (i.e., radar operators) and was later adapted
Radner, 1992; Bolton and Dewatripont, 1994; Van to study myriad discrimination problems in cog-
nitive processes (e.g., medical diagnosis, weather
4
The incentives strand has dealt mainly with project imple- forecasting, quality control).5 Given the general-
mentation or execution. Topics studied include the relationship ity of signal detection theory, many problems of
between the manager’s incentives and the range of projects
implemented by the firm (Rotemberg and Saloner, 1994); the
5
interplay between organizational structure and formal authority The updated bibliography on the 1988 reprint edition of the
(Aghion and Tirole, 1997); and the types of projects that are classic book on the subject (Green and Swets, 1966) lists more
better served by managers who persuade employees instead of than 1,000 studies published on the subject during the period
resorting to authority (Van den Steen, 2009). from 1967 to 1988 alone.

Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 617

choice can be seen as special cases of it. Of par- HYPOTHESES

ticular interest to this article is that some problems
of managerial decision making have been ana- The independent variable of the study is organi-
lyzed from a signal detection perspective: Puranam zational structure which, from the data set used,
et al. (2006) use the theory to study M&As and can reliably be coded into three nonoverlapping
Garud et al. (1997) use it to study technological categories: (1) mutual funds managed by one indi-
innovation. vidual (in terms of the model outlined earlier,
One of the contributions of signal detection the- this corresponds to N = 1 and C = 1); (2) mutual
ory is representing the omission and commission funds managed by more than one individual, with
errors of different decision makers via a ‘receiver each individual operating independently from the
operating characteristic curve’ (Figure 1 is akin others (in terms of the model, N > 1 and C = 1);
to that representation). Although signal detection and (3) mutual funds managed by more than one
theory and the Sah and Stiglitz model both pay individual and where decisions are made unani-
explicit attention to omission and commission mously (in terms of the model, N > 1 and C =
errors and to their costs, one difference between N). For brevity, these structures will be referred
these approaches is that, whereas Sah and Stiglitz to (respectively) as individual, decentralized, and
(1986) study the performance of committees, sig- centralized. Because of data limitations to be dis-
nal detection theory studies the performance of cussed later, the data set does not contain struc-
individuals. tures featuring intermediate levels of consensus
The theory of social decision schemes (Davis, (i.e., N/C with 1 < C < N, such as N = 7 and
1973, 1992) is similar to the work of Sah and C = 6).
Stiglitz on committees in that it presents mathemat- The dependent variables of the study are the
ical models of group decision making (although three outcomes predicted by the model: number of
performance is not measured as omission and approved projects, omission errors, and commis-
sion errors. Because the model predicts that these
commission errors). This literature has produced
three outcomes are most different for centralized
empirical results (Stoner, 1961; Stasser and Titus,
versus decentralized structures, the hypotheses are
1985; Hinsz, Tindale, and Vollrath, 1997); how-
stated as comparisons between those two struc-
ever, because the experiments have primarily been
tures. It would be possible to formulate six fur-
conducted in the laboratory using small groups
ther hypotheses comparing individual managers to
that meet for brief times, reported results may not
centralized structures and individual managers to
be generalizable to more complex organizations decentralized structures; but, to avoid a litany of
(Argote and Greve, 2007: 344). hypotheses, these comparisons are discussed in the
In sum, an empirically validated version of the results section without being formally enumerated
Sah and Stiglitz model can serve to combine the here.
strengths of both signal detection theory (i.e., its The three hypotheses are equivalent to the three
focus on omission and commission errors) and properties of Sah and Stiglitz’s model described
social decision schemes (its focus on groups), earlier. These three hypotheses encompass all the
thereby bringing closer to organization design the mechanisms exposed in Sah and Stiglitz (1986,
rich bodies of work on group decision making that 1988) regarding committees with exogenously
are based in social psychology. given screening functions.
Although the reviewed literatures have provided
many important insights regarding the impact Hypothesis 1: Decentralized organizations
of structure on performance, the field of orga- accept more projects than centralized
nizations lacks an empirically validated theory organizations.
that—starting from structure at the level of indi-
viduals—is able to predict organization-level Hypothesis 2: Decentralized organizations make
measures of performance that are relevant to firm fewer omission errors than centralized
organizations.
strategy. If empirically validated, Sah and Stiglitz’s
model would offer such a theory. The contribution Hypothesis 3: Decentralized organizations make
of this article is to test this theory empirically for more commission errors than centralized
the first time. organizations.
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
618 F. A. Csaszar

EMPIRICAL SETTING AND APPROACH rejected (stocks that were not bought); and (4)
organizational structure is observable from descrip-
Empirical challenge tions of the fund management prepared by Morn-
ingstar. Additionally, there are thousands of mutual
Before delving into the specifics of the data set and
funds, and the typical fund makes dozens of deci-
statistical methods, it is important to understand
sions per quarter. All these considerations make
the structure of the empirical problem. To test the
mutual funds an exceptional vehicle for studying
hypotheses, all of the following must be observed:
the effects of organization design on organizational
(1) organizations making decisions about projects,
performance; indeed, mutual funds would make
(2) a measure of the quality of each project decided
a good aspirant for the ‘fruit fly’ of organization
upon, (3) the decision that each organization made
design.
with respect to every project it faced, and (4) the
Despite these virtues of mutual funds as an
organizational structure of each organization. Item
empirical setting, there is a strong tradition in
1 exists in many settings (e.g., firms deciding
the finance literature that maintains organizational
whom to hire, where to expand, what to sell). Item
structure should not be a determinant of fund
2 is also readily available in settings where the
performance. In a nutshell, the efficient market
ex post value of the project is visible and can proxy
hypothesis (EMH) (Fama, 1970) purports that all
for the project’s true quality. In the venture capital
available information is already reflected in asset
context, for example, it could be a function of the
prices, rendering future returns unpredictable. If
IPO value of a start-up in which a venture capitalist
that is true, organizational structure should not
considered investing; in the R&D context, it could
predict mutual fund performance. However, two
be the number of citations accrued by a patent after
caveats apply. First, the EMH’s performance mea-
a firm had the opportunity to buy it.
sure is financial return, not omission and commis-
Yet items 3 and 4 are serious hurdles for the
sion errors.7 Second, the EMH is no longer viewed
empirical researcher. First, there is typically no
as invulnerable, since a vast literature on market
track record of the projects an organization consid-
anomalies (e.g., Goetzmann and Ibbotson, 1994;
ered but decided not to pursue (e.g., all the firms
Chevalier and Ellison, 1999; Cohen, Frazzini, and
a venture capitalist screened but did not invest
Malloy, 2008) has emerged in the last 20 years.
in). Second, organizational structure is not avail-
Because the variance explained by market
able from public databases. Organizational charts
anomalies is small (e.g., the typical R 2 of an
are sometimes available, but they give no indica-
anomaly is less than 1%), any variance explained
tion of whether a given decision-making process is
by organizational structure is not expected to be
centralized or decentralized (e.g., by looking at an
large. A further implication is that any explanatory
organizational chart, it is not possible to know the
power the model has will likely increase in settings
decision process used to set the direction of R&D,
where the link between cause and effect is more
perform M&As, or decide on IT investments).
deterministic. Since stock picking is arguably one
Mutual funds offer a rare window into the impli-
of the most random task environments possible, it
cations of organization design on organizational
follows that mutual funds make for a stringent test-
performance because, in this setting, the four nec-
ing arena and that the results of this article serve
essary ingredients are observable: (1) managing
as conservative estimates.
a mutual fund is essentially about making deci-
sions (i.e., deciding what to buy and what to
sell); (2) the ex post return of each investment is Independent variable: mutual fund
a good measure of the quality of each decision;6 organizational structure
(3) regulations require funds to disclose their hold- A mutual fund is a type of investment that pools
ings periodically, which allows the researcher money from many investors to buy a portfolio of
to distinguish between accepted ‘projects’ (i.e., different securities such as stocks, bonds, money
stocks that were bought) and those that the fund
7
A surprising result of this article is that the organizational
6
Stock returns are exogenously given (in most plausible cases structure of a fund can affect its omission and commission errors
they do not depend on anything a fund manager can do) and, in such a way that financial return is not affected—a result that
thus, provide a good match for Sah and Stiglitz’s model, which supports the apparently contradictory predictions of both Sah
treats quality as exogenously given. and Stiglitz (1986) and Fama (1970).

Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 619
Table 1. Examples of how organizational structure is coded from Morningstar’s fund descriptions (the ticker symbol
of each fund is given in parentheses)

Structure (N/C) Excerpts from Morningstar’s mutual fund description

1/1 ‘Ron Baron has been at the helm since the fund’s inception. . . He’s the driving force behind
this portfolio. . . buys companies he thinks can. . .’ (BPTRX)
2/1 ‘Managers Scott Glasser and Peter Hable each run 50 percent of the portfolio. . .’ (CSGWX)
3/1 ‘Three management firms select 10 stocks apiece for this fund’s portfolio.’ (SFVAX)
5/1 ‘(The fund) divvies up assets among five subadvisors, and each picks eight to 15 stocks
according to his own investing style.’ (MSSFX)
2/2 ‘Teresa McRoberts and Patrick Kelly became comanagers of this fund in late September 2004. . .
They don’t pay too much attention to traditional valuation metrics such as. . .’ (ACAAX)
7/7 ‘All investment decisions are vetted by the entire seven-person team. . . Management populates
the fund with 30–50 stocks. . .’ (CBMDX)

market instruments, or other securities. Mutual or recently promoted but retaining his/her ana-
funds in the United States are regulated by the lyst tasks).
U.S. Securities and Exchange Commission (SEC); 2. If the description is explicit about the number
among other requirements, the SEC requires funds of ‘sleeves’ or subadvisors, or if it describes
to report their portfolio holdings at the end of the how managers split their portfolios, N is set to
last trading day of every quarter (Form 13F) and the number of divisions of the portfolio and C
periodically identify their fund managers (Form is set to 1 (since this is a decentralized fund).
487). Mutual funds are heavily scrutinized not only 3. If two or more managers are mentioned but
by the SEC but also by institutional investors and nothing is said about how they coordinate
investment research firms. (e.g., they are addressed as a plurality, as in
Morningstar, one of the leading investment re- ‘they invest in. . .’), it is assumed that the fund
search firms, offers information about mutual funds employs a consensus (N = C) decision proce-
to investors and financial advisors. Using public dure. This is reasonable, as this is the default
sources and periodically meeting with fund man- structure of comanaged funds, and because if
agers, Morningstar’s analysts produce a one-page managers work separately, they have no incen-
report—densely packed with statistics and anal- tive to being reported as working in tandem
ysis—for each fund they track. For the present (managers want to create their own reputation).
study, the most important element of these profiles 4. If no specific manager names are mentioned
is a section entitled ‘governance and management,’ (e.g., the description mentions only a generic
which presents a short biography of the managers ‘the management’) or if the description states
and describes how they manage the portfolio. This that the fund is run by an algorithm (some funds
section of the report contains enough information that track indices operate like this), the fund is
to code organizational structure as modeled in this left unclassified.
article (in terms of number of managers, N, and
level of consensus required, C). To understand how
the coding was done, consider the excerpts shown Less than 4 percent of the funds fell in the
in Table 1, which illustrate typical descriptions. To unclassified group and less than 1 percent of the
increase consistency, four rules were followed for funds had a consensus level other than 1 or N.
the coding: These two classes of funds were eliminated from
the data set.
1. If the description mentions managers’ names, Because fund descriptions do not include such
N is set to the number of people mentioned as nuances as the relative sizes of each sleeve of
manager or comanager, with the exception of a decentralized fund, the organizational structure
people who are described explicitly as having a of the subadvisor of each sleeve, or the share of
secondary role (e.g., if a manager is described power each manager has in a centralized fund,
as subordinate, performing administrative tasks, the funds were aggregated into three broader cat-
not participating in the day-to-day management, egories: 1/1 (managed by an individual), N/1
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
620 F. A. Csaszar

(decentralized), and N/N (centralized). This deci- good assets that the fund bought during period t
sion ensures against overinterpreting the results. is denoted A+ = {a|a ∈ A and r(a) ≥ b} and its
All the funds were coded both by the author and cardinality is denoted n+ . Similarly, the bad assets
one research assistant. The percentage of agree- bought consist of A− = {a|a ∈ A and r(a) < b}
ment between both categorizations was 96 percent. with cardinality n− .
The results presented here use the author’s catego- At first sight, several measures might capture
rization, but all the results are robust to using the the commission error of a fund. Two possibilities
−
other categorization. are the number of bad assets bought,
n , and the
total negative return, TNR = − {a∈A− } r(a) (the
Dependent variables: omission and commission initial minus sign makes the measure increase in
errors the proper direction). Yet a problem now arises in
that—because different funds invest in a differ-
The main intuition behind the measures of omis- ent number of assets and in different investment
sion and commission error developed in this article universes—these raw metrics are not comparable
is as follows: in hindsight, a commission error across funds and, thus, are unsuitable for the pur-
occurred whenever a fund bought an asset that poses of this study.
turned out to have a poor performance (i.e., whose One way of solving the comparability problem
ex post return fell below a given benchmark); sim- is to convert these raw error measures into prob-
ilarly, an omission error occurred whenever a fund abilistic measures that account for the specifics of
failed to buy an asset that turned out to have a good each situation. An example will help clarify this
performance.8 To observe these errors, two types point. Imagine you want to find out who is better
of data are required: the list of assets that a fund at games of chance—someone who flipped a coin
did and did not buy, and the returns of these assets. 100 times and got 60 heads or someone who threw
Good data sources exist for both elements. a die 200 times and got 40 sixes. If a probability
In order to make the discussion more precise, distribution is placed on the outcomes (Pr{Head} =
some notation is useful. For a given mutual fund 1/2 and Pr{Six} = 1/6), it doesn’t matter that each
F at time t, let A = {a1 , a2 , . . . , an } be the set person played a different game; in both cases it is
of assets that F bought during time period t possible to compute a statistic (in this case, a chi-
(subscript t is omitted for convenience). The best squared) and then compare the players in terms of
available information on mutual fund holdings is how unlikely their results were.
reported quarterly, so hereafter the unit of time A first approach to creating probability-adjusted
is one quarter. Let U = {u1 , u2 , . . . , uM } represent measures of a fund’s errors is to use the hypergeo-
the assets in which F can invest, or F ’s investment metric distribution. This distribution, whose prob- 
universe at time t. The number of assets bought by R
ability mass function is f (r; M, R, m) =
F at period t is n, and the number of assets in its     r
investment universe at time t is M. By definition, M −R  M
, is typically illustrated in terms
the assets bought by a fund are a subset of the m−r m
fund’s investment universe, A ⊆ U . of the probability of getting exactly r red marbles
Asset returns are measured as holding period after drawing m marbles (without replacement)
returns; that is, r(a) represents the total return of from an urn containing M marbles of which R
asset a from the end of period t to the end of period are red. Thus, replacing ‘marble’ with ‘stock’ and
t + 1 (this measure accounts for changes in price ‘red’ with ‘bad’ yields a function that computes
as well as any income from dividends). The study the probability of getting a given number of bad
uses a per fund benchmark, defined as the average stocks; the function is already adjusted for port-
return of the assets in the fund’s investment uni- folio size, universe size, and the number of bad
1
r(u ). An asset is
M
verse at time t; thus, b = M stocks in the investment universe. A nice feature
i
i=1 of this approach is that it removes the effect of the
cataloged as ‘good’ if its return in a given period environment from the error measures. For exam-
equals or exceeds the benchmark b. The subset of ple, an economy-wide shock that has a positive
effect on one type of fund but a negative effect on
8
Omission and commission errors can also be measured with another would not distort the measures of error.
respect to sell decisions. This case is discussed later. In other words, that the negatively affected fund
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 621

draws stocks from a ‘tougher’ urn than does the 2.6), a statistic computed via bootstrap converges
positively affected fund is controlled for by the to the real statistic as the number of random draws
hypergeometric distribution. increases. For the data used in this article, each
fund ‘competes’ against 100,000 simulated port-
Using the bootstrap to compute a better folios and the standard error introduced by the
measure of omission and commission errors bootstrap procedure is less than 0.003.
Once the population of portfolios comparable to
The hypergeometric approach just illustrated can a fund F is created, the measure of commission
be further improved by using the bootstrap, an error is simply a measure of the deviance of F ’s
important development in statistics (Efron, 1979; error with respect to the commission errors of
Efron and Tibshirani, 1993). One limitation of the that population. Given the central limit theorem
hypergeometric approach is that it weighs all bad and the large number of simulations, the normal
decisions equally, regardless of the size of the distribution is a good approximation for the TNRs
errors (i.e., a stock that slightly underperformed of the population. Therefore, errors are reported in
the benchmark is counted the same as a stock terms of standardized scores; the higher the score,
whose price collapsed). The bootstrap allows one the higher the error.
to model a probability distribution that takes into The omission error can be defined analogously
account the size of the errors.9 to the commission error. Instead of measuring
The bootstrap consists of creating an arbitrar- TNR, in this case one measures the total unbought
ily good approximation of a population via Monte positive returns (TUPR)—that is, the sum of the
Carlo simulations and using this new population good assets that belong to the investment universe
to compute the exact value of a statistic. In this of fund F but were not bought in
the current
case, the population to be estimated is the set of period. Mathematically, TUPR = {a∈U and a∈A} /
all possible portfolios of a given size that can r(a). Following the previous example, the TUPR
be drawn from a given investment universe. An of fund F is 4 percent (= 1% + 3%). As before,
example will clarify how the bootstrap can be the bootstrap is then used to compute a probability-
used to measure commission errors. Suppose the adjusted measure that is expressed as a standard-
returns of the assets in the investment universe ized score.
of fund F are {−5%, −2%, −1%, 1%, 3%, 4%},
the benchmark is b = 0, and the fund bought the
Data preparation and limitations of the data
three assets that ended up returning {−2%, −1%,
set
4%}. Hence F ’s total negative return is 3 percent
(TNR = −[−2% + −1%]). To assess how large or The content and format of Morningstar’s one-
small this number is, it must be compared to the page mutual fund reports have changed repeatedly
TNRs of the population of funds that can draw over the years, and in 2007 it started including a
 
three stocks from the same investment universe ‘governance and management’ section with en-
6 ough information to code organizational structure
as F . In this example, 20 (= ) other portfo-
3 for a large sample of funds. This implies that
lios could have been bought, but in realistic cases
the data on organizational structure for Decem-
the space of possible portfolios cannot be explored
ber 2007 are available only as a snapshot. There-
exhaustively;10 hence the method relies on ran-
fore, whereas the dependent variables are com-
domly sampling the space of possible portfolios.
puted using errors from 2004Q4 to 2007Q1, funds
With the exception of some well-known patho-
that changed their organizational structure after
logical cases (Davison and Hinkley, 1997, Sec.
2004Q4 but before December 2007 are partially
misclassified in the analysis. Fortunately, changes
9
Distinguishing between large and small errors is a natural in the organizational structure of funds are rare.
property to expect from a measure of errors—especially since
models that conceptualize perception as signal plus noise (such
There are no official statistics, but a good estimate
as Sah and Stiglitz’s or signal detection theory) would predict of change in the organizational structure of mutual
that blunders are less likely to occur than slight errors. funds can be gathered from Morningstar (2008). In
10
The average fund in the data set buys 16 stocks
 from a uni- addition to 500 fund reports, Morningstar (2008:
195
verse of 195, which creates a space of ≈ 1023 possible 29) also includes a brief description of all the man-
16
portfolios. agement changes that took place in these funds
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
622 F. A. Csaszar

during 2007. Of the 500 reported funds, 32 expe- unobserved and observed trades differed in a way
rienced some sort of management change (the most that depended on organizational structure. There
typical change is replacement of a manager) and are no reasons to believe a priori that this might
only four funds experienced a change in organi- be the case.
zational structure as coded in this article. This The returns used to determine whether an invest-
amounts to a 0.8 percent yearly probability of such ment was a good or a bad one were the quarterly
change. returns of each stock from the end of quarter t to
In December 2007, Morningstar kept organiza- the end of quarter t + 1; these returns were gath-
tional descriptions for 1,687 funds. To increase ered from the CRSP data set ‘Monthly Stocks’
comparability, only funds that were primarily using the field ‘Holding Period Return,’ which
devoted to stocks (not other asset classes, such adjusts for stock splits and dividends. Because the
as bonds or options) were selected. Thus, funds exact date at which assets are bought is unknown
were chosen if their asset composition (accord- (i.e., the holdings database has quarterly resolu-
ing to the CRSP data set ‘Mutual Fund Profiles tion), a further intrinsic limitation of the data set
and Monthly Asset Data’) was at least 60 percent is that it fails to account for the return accrued
stocks in the time period under study. This nar- since a stock is bought until the end of that quar-
rowed the list down to 1,087 funds. The CRSP data ter. Yet this lack of data should affect the results
sets ‘Portfolio Holding Information’ and ‘Monthly of the study in a conservative way. The reason is
Stocks’ were then used to choose only those funds that if managers are able to minimize the errors
for which CRSP reported the returns of the indi- they make, this ability should be more notice-
vidual stocks owned by the fund for at least able soon after the decision than later, when more
50 percent of its portfolio value. This reduced the unpredictable events may affect the price of their
list to 642 funds. This drop is primarily explained purchase.
by CRSP’s tracking only the returns of stocks The investment universe of a fund at time t
traded on NYSE, NASDAQ, and AMEX (while was defined as all the stocks available for pur-
many funds invest in international stocks) and to chase at time t from the union of all the hold-
a lesser extent by observations missing from the ings reported by the fund in a trailing window
CRSP portfolio holdings data set. Finally, funds of seven quarters, including the current quarter
for which the Morningstar description did not (i.e., using the last seven Forms 13F reported by
allow an organizational structure to be inferred the fund). There are at least three other ways to
were dropped, leaving the final count at 609 funds define the investment universe, but they present
owned by 154 different parent firms. Collectively, conceptual and practical problems that make them
for the 10 quarters from 2004Q4 to 2007Q1, less preferable than the trailing-period definition.
these funds invested in 5,833 distinct stocks (as The first alternative is to use the investment
identified by their CUSIP number), made 153,457 objective, typically reported by each fund in its
buy decisions, and had $1.6 trillion under manage- prospectus; however, this information is
ment at the end of the period. The range of dates imprecise11 and not always available, so using it to
used is due to data limitations: before 2004Q4 the define the investment universe would have a sub-
CRSP holdings database is sparse; and by Decem- jective quality. A second alternative is to include
ber 2007, CRSP had not yet uploaded the holdings all the 5,833 stocks ever bought by all the funds.
information for the quarters after 2007Q1. This approach was discarded because it is unfair
The stocks that a fund bought during the quarter to count the failure to buy stocks that would never
ending at date t were determined by looking at be bought by a fund as ‘omissions’ (e.g., a utili-
the stocks added to the portfolio since the last ties fund does not buy high-tech stocks). A third
reported quarterly holdings. The quarterly holdings alternative is to use the union of all the stocks
were gathered from the CRSP data set ‘Portfolio
Holdings Information,’ itself a compilation of the 11
For example, a fund may say that it attempts to track a broad
Forms 13F that mutual funds submit to the SEC. index like the S&P500, but this does not imply that it invests
An intrinsic limitation of the data is that, if a stock only in stocks that are listed in the index; many of its investments
is bought and sold during the same quarter, that may fall outside it. Another fund may say that it invests in ‘small
caps,’ a broad category with thousands of stocks, though its
buy decision is unobserved. However, this would investments consistently fall within a group of fewer than 100
pose a problem only if the error measures of the stocks.

Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 623

ever bought by the funds that share the same The controls used, which are in line with those
Morningstar investment category. Like the previ- used in the mutual fund literature (e.g., Chen
ous alternative, this method creates loose invest- et al., 2004), were: (1) Beta i , the risk profile of
ment universes that lead to a similar (albeit less the fund as measured by its Beta with respect to
serious) unfairness problem. In short, letting the the S&P500; (2) log(ParentSize i ), the size of the
deeds of the fund speak for themselves seemed parent firm (the firm owning the fund) as mea-
the most appropriate choice. Robustness checks sured by the logarithm of the number of mutual
showed that the third alternative definition pro- funds that the parent firm owns (within the uni-
duced results qualitatively similar to those reported verse of 1,087 stock mutual funds tracked by
here using the trailing-period definition. Morningstar); (3) log(FundSize i ), the size of the
fund as measured by the logarithm of the net
assets managed by the fund (in millions of dol-
RESULTS lars); and (4) CatXX i , seven investment category
dummies as coded by Morningstar (Large Growth,
Although mutual funds offer a unique window Large Blend, Large Value, Mid-Cap Growth, Small
into the effect of structure on performance, actu- Growth, Small Blend, and Mid-Cap Blend).
ally measuring that effect is relatively challeng- Roughly 80 percent of the funds fell into one of
ing. An ideal test would consist of comparing the these seven categories; the rest were consolidated
performance of mutual funds making investments in an ‘Other’ class that grouped 13 smaller cate-
in the same sector, with the same managers, at gories and was used as the omitted dummy in the
the same time, and differing only in organiza- regressions.
tional structure. This situation is unattainable, so The regressions were run on a pooled cross-
the challenge consists of statistically controlling section (and not in a panel) because there is
for differences other than organizational structure. essentially no variation in the structure of the
Fortunately, the bootstrap (explained earlier) and mutual funds during the period analyzed. The num-
the standard control variables in the mutual funds ber of observations in the regressions is 6,090,
literature (explained later) can control for these since there are 609 funds and each one has
differences.12 data for 10 periods. The omission and commis-
Each of the three hypotheses was tested using a sion errors were computed as probabilities that
regression of the form are independent of the specific environment the
fund was facing (i.e., actual stocks in the invest-
dependent variablei = α0 + α1 Decentralizedi ment universe and its returns), which makes the
pooled cross-section specification an appropriate
+ α2 Individuali + α3 Betai + α4 log(ParentSizei ) choice.
+ α5 log(FundSizei ) + α6 CatLGi + α7 CatLBi To avoid a possible source of endogeneity,
all the controls were measured at the beginning
+ α8 CatLVi + α9 CatMCGi + α10 CatSGi of the period used to compute the dependent
+ α11 CatSBi + α12 CatMCBi + εi , variables (beginning of 2004Q4). To counter the
effects of heteroskedasticity—and because obser-
where the dependent variable is the logarithm vations coming from funds that belong to the same
of the number of stocks bought per quarter to parent firm may not be independent—the stan-
test Hypothesis 1, omission error to test Hypothe- dard errors were computed using cluster-robust
sis 2, and commission error to test Hypothesis 3. estimation (Williams, 2000) with clusters defined
The independent variable of the study, organi- according to the parent firm. All reported p-values
zational structure, was coded as two dummies correspond to two-tailed tests; this is a conserva-
representing the decentralized and the individual tive decision because the hypotheses tested are of
structure (the centralized structure is the omitted the form a < b, which calls only for running one-
dummy). tailed tests.
Table 2 displays summary statistics and correla-
12
tions. The correlations show no evidence of mul-
For an example of the use of the bootstrap in the mutual fund
literature, see Kosowski et al. (2006); for a tutorial presentation, ticollinearity, which is reaffirmed by the variance
see Burns (2004). inflation factors—none of which was larger than
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
624 F. A. Csaszar

Table 2. Descriptive statistics and correlations (for the 609 funds in the sample)

mean sd min max 1 2 3 4

1 Omission error −0.16 0.48 −2.77 1.09 1.00

2 Commission error 0.14 0.47 −1.03 2.35 0.33 1.00
3 Beta 1.15 0.27 −0.09 2.77 0.01 0.05 1.00
4 log(ParentSize) 2.14 1.10 0.00 4.68 −0.11 0.06 −0.04 1.00
5 log(FundSize) 6.17 1.68 −0.24 11.25 0.06 0.08 −0.14 0.32

Table 3. Results of regression analysis of number of stocks bought

Dependent variable: log(#stocks bought)

A1 A2 A3 A4 A5

Decentralized (structure N/1) 0.647∗∗∗ 0.630∗∗∗ 0.582∗∗∗ 0.580∗∗∗ 0.544∗∗∗

(0.141) (0.138) (0.167) (0.169) (0.162)
Individual (structure 1/1) 0.119 0.091 0.041 0.041 0.047
(0.100) (0.100) (0.090) (0.090) (0.091)
Beta 0.848∗∗∗ 0.880∗∗∗ 0.883∗∗∗ 0.419+
(0.236) (0.245) (0.253) (0.250)
log(ParentSize) 0.187∗∗ 0.186∗∗ 0.210∗∗∗
(0.060) (0.062) (0.060)
log(FundSize) 0.003 0.013
(0.027) (0.025)
∗∗∗
Category effects (joint test)
∗∗∗ ∗∗∗ ∗∗ ∗
Constant 2.215 1.259 0.851 0.832 0.982∗
(0.097) (0.273) (0.325) (0.399) (0.414)
Observations 6090 6090 6090 6090 6090
Adjusted R2 0.030 0.090 0.137 0.135 0.213

Note: Robust standard errors between parentheses.

+
p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 (two-tailed tests).

1.6, a number well below the customary thresh- row 1 of Table 4), its logarithm was used as a
old of 10. Of the 609 funds in the data set, the dependent variable. In all the models, the decen-
most common structure is the individual manager tralized structure (N/1) was associated with buying
(324 funds), followed by the centralized structure significantly more stocks than the centralized struc-
(233 funds) and the decentralized structure (52 ture (the effect size corresponds to a 30 to 50%
funds). increase, depending on the model and the value of
the controls). No significant relationship is present
Number of projects accepted for the structure 1/1, yet the sign of its associated
coefficients has the predicted direction in all the
To test Hypothesis 1, the number of stocks bought models.
by each fund was analyzed.13 In order to determine
The coefficients associated with the controls tell
whether the relationship between organizational
stories that are interesting in themselves. Models
structure and number of stocks bought is statisti-
A3 to A5 show that funds belonging to larger par-
cally significant, five models were tested (Table 3).
ent firms buy more stocks—even after the size
Given that the distribution of the number of stocks
of the mutual fund and investment category are
bought is highly skewed (see, e.g., the relation-
ship between the average and the maximum in controlled for. One possible interpretation is that
larger parent firms have better support structures,
13
allowing managers to track more stocks. The
As a robustness check, the number of stocks sold by each
fund was also analyzed. The results were qualitatively the same regressions also show that the more net assets man-
as those presented here for the number of stocks bought. aged by a fund, the more stocks it will invest in;
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 625
Table 4. Descriptive statistics—number of stocks per organizational structure

N/N 1/1 N/1 Total

Centralized Individual Decentralized (609 obs.)
(233 obs.) (324 obs.) (52 obs.)
avg (sd) avg (sd) avg (sd) avg (sd)
[min,max] [min,max] [min,max] [min,max]

1) #stocks bought per quarter 15.9 (27.9) 15.3 (16.7) 26.1 (27.2) 16.5 (22.7)
[1.2, 280.0] [1.3, 131.2] [3.4, 148.2] [1.2, 280.0]
2) #stocks in portfolio 91.5 (105.9) 138.6 (293.7) 171.1 (168.3) 123.3 (230.6)
[18.6, 1220.3] [20.3, 3455.2] [25.6, 990.9] [18.6, 3455.2]

this may reflect that large funds are more likely N/1 is 131.3; and this is not statistically differ-
to run into the liquidity limits of the underlying ent from the portfolio size of structure 1/1, which
stocks. The fact that Beta has a positive effect can is 138.6.) In other words, if structure is mea-
also be explained in terms of liquidity, as higher- sured in the wrong way, the relationship between
Beta stocks, on average, correspond to smaller structure and portfolio size is rendered invisi-
firms. Finally, Model A5 shows that there is a ble even though that relation is actually quite
significant category effect, which gives additional strong.
support to the liquidity explanation because the
categories with the largest positive coefficients are
Omission and commission errors
those involving small companies (only the cate-
gories Small Growth and Small Blend were statis- Figure 1 displays the average omission and
tically significant, with respective coefficients of commission error made by each organizational
0.55 and 0.91). structure. The axes of the figure correspond to the
Models A1 to A5 were rerun using portfolio standardized measures previously described (com-
size instead of number of stocks bought per quar- puted via bootstrap). The figure looks exactly as
ter, and all the results were qualitatively the same. Sah and Stiglitz’s model would predict, with the
This increases confidence in the results by show- centralized fund at the lower right (minimizing
ing that what is true for a flow variable (number commission errors), the decentralized fund at the
of stocks bought) is also true for its corresponding upper left (minimizing omission errors), and the
stock variable (portfolio size). In all, the large and individual manager in between.
significant coefficients accompanying the decen-
tralized structure provide ample evidence that
decentralized funds accept more projects than do
centralized funds (Hypothesis 1).
It is remarkable that this statistically signifi-
cant relationship between structure and portfolio
size has not been reported in the finance litera-
ture—probably because researchers in that field
have conceptualized organizational structure sim-
ply as number of managers (e.g., Chen et al.,
2004). Measuring structure as number of managers
is roughly equivalent to comparing the average
portfolio size of structure N/N and N/1 (since
both structures have N managers) against the port-
folio size of structure 1/1. Under that averaging,
no difference is noticeable. (In particular, from
the numbers in Table 4, one can easily calcu- Figure 1. Average (centroids) omission and commission
late that the average portfolio size of N/N and errors of the three organizational structures
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
626 F. A. Csaszar

Table 5. Results of regression analysis of omission error

Dependent variable: omission error

B1 B2 B3 B4 B5

Decentralized (structure N/1) −0.162∗ −0.162∗ −0.150∗ −0.172∗ −0.161∗

(0.077) (0.077) (0.069) (0.074) (0.078)
Individual (structure 1/1) −0.054 −0.054 −0.042 −0.044 −0.043
(0.046) (0.046) (0.046) (0.046) (0.047)
Beta 0.019 0.011 0.037 0.034
(0.071) (0.071) (0.071) (0.078)
log(ParentSize) −0.047∗∗ −0.062∗∗∗ −0.064∗∗∗
(0.016) (0.017) (0.018)
log(FundSize) 0.033∗∗ 0.035∗∗
(0.012) (0.013)
Category effects (joint test) not sig.
Constant −0.119∗∗ −0.140 −0.039 −0.235+ −0.217+
(0.039) (0.089) (0.100) (0.127) (0.130)
Observations 6090 6090 6090 6090 6090
Adjusted R 2 0.005 0.004 0.014 0.024 0.022

Note: Robust standard errors between parentheses.

+
p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 (two-tailed tests).

All five models in Table 5 support Hypothe- funds having little incentive to exploit small, yet
sis 2 by showing that a decentralized fund makes profitable investment opportunities because their
significantly fewer omissions than a centralized relative contribution to the fund’s overall prof-
one. The magnitude of the coefficients associ- itability would be tiny.
ated with the decentralized structure is sizable, All the models in Table 6 support Hypothesis 3
as it can be shown that decreasing an error by by showing that a decentralized fund makes signifi-
0.15 points of the standardized score is associated cantly more commission errors than does a central-
with a 13 percent increase in annual performance ized one. As before, the coefficients for the indi-
(relative to current performance; e.g., a 10% annual vidual manager have the predicted sign but are not
return would become 11.3%).14 As in the previous significant. Parent size and net assets, which were
set of regressions, the coefficients accompanying significant controls in the regressions of omission
the individual manager have the correct sign but error, are not significant predictors of commis-
are not statistically significant. sion error; this may mean that small funds devote
Among the controls, parent size and net assets comparatively more resources to minimizing com-
appear to be significant determinants of omission mission (rather than omission) errors.
errors. The fact that funds owned by larger firms Two controls that are typically significant in
make fewer omissions indicates that the ability studies of investment performance—the fund’s
to avoid missing investment opportunities may Beta and its investment category—are not signif-
reside partly in routines that are more likely to icant predictors of either omission or commission
exist in larger firms, such as research support ser- errors. The reason is that the bootstrap mechanism
vices, fund manager training, or knowledge sharing used to compute the errors already controls for
among managers of different funds. Conversely, these parameters: each fund is compared in stan-
the finding that funds managing more assets make dardized terms against a large number of funds
more omission errors may be due, in part, to large that draw stocks from the same investment uni-
verse and so, on average, have the same Beta and
14
To compute the effect on a fund’s annual return, a simulation investment category as the focal fund.
was run using parameters representative of the average fund.
This fund buys 16 stocks from a universe of 195 stocks each
quarter; the stock’s returns are drawn from a N(0.0339, 0.2042) Ruling out alternative hypotheses
distribution; the portfolio turnover is one year; and the effect due
to superior stock picking is only effective (this is a conservative Mutual funds offer the best available setting to test
assumption) in the quarter after the stock was bought. the Sah and Stiglitz model’s predictions because
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 627
Table 6. Results of regression analysis of commission error

Dependent variable: commission error

C1 C2 C3 C4 C5

Decentralized (structure N/1) 0.184∗∗ 0.183∗∗ 0.177∗ 0.164∗ 0.146∗

(0.068) (0.067) (0.071) (0.075) (0.073)
Individual (structure 1/1) 0.054 0.051 0.045 0.044 0.045
(0.044) (0.044) (0.042) (0.041) (0.041)
Beta 0.079 0.083 0.098 0.082
(0.079) (0.080) (0.088) (0.103)
log(ParentSize) 0.023 0.014 0.014
(0.019) (0.018) (0.018)
log(FundSize) 0.019 0.018
(0.014) (0.014)
Category effects (joint test) not sig.
Constant 0.097∗∗ 0.008 −0.042 −0.156 −0.183
(0.033) (0.103) (0.118) (0.178) (0.207)
Observations 6090 6090 6090 6090 6090
Adjusted R 2 0.008 0.008 0.010 0.012 0.013

Note: Robust standard errors between parentheses.

+
p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 (two-tailed tests).

this setting offers observability of structure and equally to the performance of the fund vis-à-vis
omissions, a large number of observations, and the benchmark. Omission and commission errors
cross-sectional variance in structure. At the same are equally costly for mutual funds. After all, fail-
time, the setting is not perfect because there is ing to buy a stock that would have contributed
essentially no variation of structure over time. $1 to returns is no more or less costly than
This means that when making causal interpreta- buying a stock that subtracted $1 from returns;
tions, close attention must be paid to possible in both cases there is a loss of $1 with respect
endogeneity and unobserved heterogeneity issues. to a competing fund that did not make the same
The following logic shows how difficult it is to error.
construct alternative hypotheses that might bias the Second, most imaginable unobserved character-
results in the same direction as predicted by the istics should affect both buy and sell decisions.
Sah and Stiglitz model. Yet it turns out that structure affects only the buy
The controls used in the regressions (for parent decisions. (In regressions for the omission and
and firm size, risk profile, and the investment cate- commission errors on the sell decisions, available
gory dummies) serve to rule out simple alternative from the author on request, none of the structure
hypotheses such as relating a type of error to an coefficients was significant.) This is consistent with
investment strategy. More importantly, the boot- a mechanism that surfaced during informal inter-
strap (described earlier)—by controlling for the views with fund managers. They revealed that,
specific environment faced by each fund—takes although the purchase of stocks is quite delibera-
care of a large set of possible issues, such as the tive, the sale of stocks is a semiautomatic process
effect that economy-wide shocks could have on that is often guided by stop-loss orders or tax and
different funds. liquidity considerations.
Unobserved heterogeneity seems unlikely for Another issue that may bias the results would
two reasons. First, any unobserved heterogene- be incorrect imputation of organizational structure.
ity explanation due to a deliberate preference It could be argued that some of the funds that
of managers for omission or commission errors are coded as being managed by one individual
seems implausible because mutual fund managers are really managed by either a hierarchy (N/N)
are primarily concerned with surpassing a bench- or a polyarchy (N/1), but that these details do
mark, not with how this benchmark is surpassed, not appear in the Morningstar report from which
and omission and commission errors contribute structure is coded. Yet if that were the case, it
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
628 F. A. Csaszar

would bias the conclusions against the hypoth- structure is not different from the overall error of
esized results—namely, it would be harder for picking stocks at random: the unpredictability of
funds coded as hierarchy and polyarchy to be sta- returns stated by the efficient market hypothesis
tistically different from the mass of funds coded as holds when looking at the overall error, even
individually managed. In other words, Morningstar if each error measured independently is partly
reports that were imprecise in this way would bias predictable. This equivalency in the cost of errors
the results in a conservative fashion. is also beneficial (and perhaps essential) for the
purposes of the empirical test carried out in this
article, for otherwise it is likely that mutual funds
DISCUSSION would all flock to the structure minimizing the
relevant error and thereby drain the data set of
This study has used mutual funds as a rich variation in the independent variable.
data source to explore how organizational struc- Thus, the mutual fund setting has a particu-
ture affects organizational performance. In perfect lar characteristic that does not generalize to other
accordance with the predictions of the Sah and domains: in the structure-errors-performance chain
Stiglitz model of fallible decision making, decen- of causation, structure affects errors (i.e., organi-
tralized structures accept more projects (Hypothe- zational performance) but errors (as long as they
sis 1), make fewer omission errors (Hypothesis 2), cost the same and add up to the same overall
and make more commission errors (Hypothesis 3) error) do not affect competitive performance. In
than do centralized structures. This section places most other settings (in which the two errors have
these results in perspective. different costs), errors should affect competitive
performance. Generalizing results to such other
domains is discussed next.
Mutual funds and organizational structure
Two questions come to mind regarding the orga-
Generalizability to other domains
nization design of mutual funds: is there an opti-
mal organizational structure for mutual funds? And This article’s proof of the hypotheses derived from
why is the individual manager the most common Sah and Stiglitz (1986) serves as a foundation on
structure? (Note that 53.2% of the funds in the data which to guide structure recommendations. The
set used this structure.) key observation in this context is realizing that dif-
As mentioned previously, omission and commis- ferent organizations face different costs for omis-
sion errors are equally costly for a mutual fund sion and commission errors. For example, juries
concerned only with maximizing returns. Hence, are more concerned with commission errors (i.e.,
the structure this hypothetical fund should choose to avoid convicting the innocent); the typical IT
is the one that minimizes the sum of both errors. department also is presumably more concerned
Strikingly, the sum of the omission and commis- with minimizing commission errors (e.g., not leak-
sion errors (measured as standardized scores) for ing sensitive information) than with minimizing
each of the three structures is statistically indis- omission errors (e.g., implementing every good
tinguishable from zero (i.e., if the coordinates IT innovation); and a well-funded R&D lab in an
of the points on Figure 1 are added, the results industry characterized by first-mover advantages is
are −0.28 + 0.28 = 0.00, −0.17 + 0.15 = −0.02, more likely to be concerned with avoiding omis-
and −0.12 + 0.10 = −0.02 for structures N/1, sion errors. Thus, this article supports the follow-
1/1, and N/N, respectively). Given this equiva- ing recommendations for organizations aiming to
lency in overall errors, it seems natural that most choose the best structure given the environment it
funds choose the least expensive structure. The faces: if the omission error is the costlier error,
existence of funds with structures different from the organization is better served by a polyarchical
1/1 may speak to other concerns, such as secur- (N/1) structure; if the commission error is costlier,
ing continuity against manager turnover, offering the organization is better served by a hierarchical
promotion opportunities to junior employees, or (N/N) structure.15
creating a differentiated product.
There is a special beauty to the fact that in 15
It is possible to make finer-grained structure recommendations
the mutual fund setting, the overall error of each (i.e., recommending a specific N and C) by feeding into the

Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 629

This research also speaks to the unexplored clear mechanisms for both directions of the causal-
question of what are the processes that link orga- ity arrow. If strategy is understood as the pool of
nizational structure to exploration and exploita- all the projects pursued by the organization (akin
tion (Siggelkow and Levinthal, 2003: 650; Argyres to Mintzberg’s (1978) concept of emergent strat-
and Silverman, 2004: 929; Raisch and Birkinshaw, egy), then structure influences the types of project
2008: 380). A relevant observation in address- that end up in that pool (e.g., hierarchy decreases
ing this question is that omission and commission commission errors) and, thus, structure influences
errors are another way of looking at exploration strategy. On the other hand, if strategy is under-
and exploitation (Garud et al., 1997: 33; Garicano stood as a deliberate plan (i.e., in a Chandlerian
and Posner, 2005: 157). The logic of this argu- way), the role of organization becomes subordi-
ment is that, on the one hand, firms in unstable or nate to strategy (e.g., a firm that wants to reduce
fermenting environments must try to avoid omis- commission errors decides to use a hierarchical
sions because these curtail the extent of explo- structure) and so strategy influences structure.
ration of new high-fitness positions. Illustrations
of this behavior are Bill Gates saying that ‘the
real sin is if we (Microsoft’s R&D) miss some- Further work
thing’ (Hawn, 2004: 70) and Andy Grove’s quip, Further research could employ alternative settings
‘miss the moment (for change in a high-tech firm (or perhaps alternative experiments) to explore the
such as Intel) and you start to decline’ (Strat- predictions of the model that cannot be tested
ford, 1993: 58). On the other hand, firms fac- using the current data set. Some questions open
ing stable or incrementally changing environments to empirical examination involve the omission and
try to avoid commission errors, since these could commission errors associated with structures other
disrupt their currently efficient exploitative opera- than those studied here, as well as how commit-
tions. Examples of these phenomena include Proc- tee decision making interacts with other organiza-
ter & Gamble, where new product proposals are tional dynamics, such as power. Another line of
often reviewed more than 40 times before reaching inquiry, very much in the spirit of contingency
the CEO (Herbold, 2002), and IBM’s mainframe theory, could explore whether firms that exhibit
era inspired ‘nonconcur policy,’ which enabled any better structure–environment fit achieve higher
department to veto projects initiated anywhere in performance or exhibit higher survival rates. For
the firm (Gerstner, 2003). Hence, given that (1) it example, in industries requiring more conserva-
has been shown here how organizational structure tive decision making (i.e., where commissions are
can influence the omission and commission errors costlier than omissions), one would expect the per-
made by organizations and (2) previous research formance of firms using higher consensus levels to
has shown that these errors control the degree surpass those using lower levels.
to which organizations can explore and exploit, In more general terms, this article also sug-
this article exposes a mechanism by which orga- gests that decomposing performance into omission
nizational structure can influence exploration and and commission errors can reveal phenomena oth-
exploitation. erwise unobservable when using standard per-
A core debate in strategy and organization formance measures. Hence, future research on
design has concerned the direction of causality in organizations may benefit from including omission
the relationship between strategy and structure. On and commission errors as alternative measures of
the one hand, there is Chandler’s (1962) famous performance.
dictum that ‘structure follows strategy.’ On the
other hand, several authors have argued in favor of
a reverse, complementary logic—in other words, Conclusions
that structure may also influence strategy (see, e.g.,
From a theoretical point of view, this research
Burton and Kuhn, 1980: 4; Burgelman, 1983: 61;
presents a mechanism by which micro decisions
Pettigrew, 1987: 665). The current article describes
are aggregated into macro behaviors and links
to important questions of strategy research—for
model information about the situation in question (e.g., cost of example, ‘do organizations have predictable
each of the errors, cost of each decision maker). biases?’ (Cyert and March, 1963: 21), ‘what do
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
630 F. A. Csaszar

we know about the relationships between organi- (2009), the Organization Science Winter Confer-
zational size (or other stable characteristics) and ence (2010), the MIT organizational economics
behavior?’ (Rumelt et al., 1994: 42), and ‘what is lunch, and the SMJ Conference on Strategy and
the relationship between decision making and deci- Organizational Design at HBS. Research funding
sion outcomes?’ (Zajac and Bazerman, 1991: 37). from the INSEAD Alumni Fund and The Ackoff
From a practical standpoint, this research sheds Fund of the Wharton Risk Management and Deci-
light on how the structure of organizations can sion Processes Center is gratefully acknowledged.
be modified to compensate for the shortcomings Errors remain the author’s own.
of individuals, and it allows several managerial
concerns to be addressed: what type of organi-
zation is required in order to avoid exceeding a REFERENCES
given error level? Is it true that hierarchy ham-
pers innovation? What organizational structures Aghion P, Tirole J. 1997. Formal and real authority in
organizations. Journal of Political Economy 105(1):
can lead to more innovation? In regard to this 1–29.
last question, an important area of application is Argote L, Greve HR. 2007. A behavioral theory of
enabling established organizations to exhibit traits the firm—40 years and counting: introduction and
usually associated with entrepreneurial ventures. impact. Organization Science 18(3): 337–349.
The 9/11 Commission Report contains an eloquent Argyres NS, Silverman BS. 2004. R&D, organization
structure, and the development of corporate tech-
call for this sort of transformation: ‘imagination is nological knowledge. Strategic Management Journal
not a gift usually associated with bureaucracies . . . 25(8/9): 929–958.
it is therefore crucial to find a way of routinizing, Barnard CI. 1938. Functions of the Executive. Harvard
even bureaucratizing, the exercise of imagination’ University Press: Cambridge, MA.
(National Commission on Terrorist Attacks upon Bikhchandani S, Hirshleifer D, Welch I. 1992. A theory
of fads, fashion, custom, and cultural change as
the United States, 2004: 344). informational cascades. Journal of Political Economy
Maritan and Schendel (1997: 259) observe that 100(5): 992–1026.
‘there has been surprisingly little work that has Bolton P, Dewatripont M. 1994. The firm as a com-
explicitly examined the link between the pro- munication network. Quarterly Journal of Economics
109(4): 809–839.
cesses by which strategic decisions are made and Burgelman RA. 1983. A model of the interaction of
their influence on strategy.’ This article aims to strategic behavior, corporate context, and the concept
illuminate that topic by advancing a small step of strategy. The Academy of Management Review 8(1):
toward understanding how organizational struc- 61–70.
ture aggregates individual decisions into strategic Burns P. 2004. Performance measurement via ran-
dom portfolios. Available at: http://www.burns-
outcomes. stat.com/pages/Working/perfmeasrandport.pdf (ac-
cessed 1 October 2010).
Burton RM, Kuhn AJ. 1980. Strategy follows structure:
the missing link of their intertwined relation. Working
ACKNOWLEDGEMENTS paper, Fuqua School of Business, Duke University.
Chandler AD. 1962. Strategy and Structure: Chapters in
I would like to give special thanks to Dan the History of American Industrial Enterprise. The
MIT Press: Cambridge, MA.
Levinthal, Nicolaj Siggelkow, Jitendra Singh, and Chen J, Hong H, Huang M, Kubik JD. 2004. Does fund
Sid Winter for their insights throughout this size erode mutual fund performance? The role of
project. For their valuable comments, I also thank liquidity and organization. American Economic Review
the seminar participants at Columbia, Harvard 94(5): 1276–1302.
Business School, IESE, INSEAD, London Busi- Chevalier J, Ellison G. 1999. Are some mutual fund
managers better than others? Cross-sectional patterns
ness School, The Ohio State University, Stan- in behavior and performance. Journal of Finance
ford Graduate School of Business, Tuck School 54(3): 875–899.
of Business, University of Minnesota, University Christensen M, Knudsen T. 2010. Design of decision
of North Carolina, University of Southern Cali- making-organizations. Management Science 56(1):
fornia, University of Southern Denmark, Univer- 71–89.
Cohen L, Frazzini A, Malloy CJ. 2008. The small
sity of Toronto, and UCLA as well as the Whar- world of investing: board connections and mutual
ton PhD seminar, the 15th CCC doctoral con- fund returns. Journal of Political Economy 116(5):
sortium, the Academy of Management Meeting 951–979.
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
 Organizational Structure as a Determinant of Performance 631
Cohen MD, March JG, Olsen JP. 1972. A garbage can Geanakoplos J, Milgrom P. 1991. A theory of hierarchies
model of organizational choice. Administrative Science based on limited managerial attention. Journal of the
Quarterly 17: 1–25. Japanese and International Economies 5(3): 205–225.
Cremer J. 1980. A partial theory of the optimal Gerstner LV. 2003. Who Says Elephants Can’t Dance?
organization of a bureaucracy. Bell Journal of Leading a Great Enterprise Through Dramatic
Economics 11(2): 683–693. Change. HarperBusiness: New York.
Csaszar FA. 2009. An efficient frontier in organization Goetzmann WN, Ibbotson RG. 1994. Do winners repeat?
design. Available at: http://dx.doi.org/10.2139/ssrn. Journal of Portfolio Management 20(2): 9–18.
1097118 (accessed 15 January 2012). Green DM, Swets JA. 1966. Signal Detection Theory and
Cyert RM, March JG. 1963. A Behavioral Theory of the Psychophysics. Wiley: New York.
Firm. Prentice Hall: Englewood Cliffs, NJ. Gulati R, Higgins MC. 2003. Which ties matter when?
Daft RL, Lewin AY. 1993. Where are the theories for The contingent effects of interorganizational partner-
the new organizational forms: an editorial essay. ships on IPO success. Strategic Management Journal
Organization Science 4(4): i–vi. 24(2): 127–144.
Davis JH. 1973. Group decision and social interaction: Harris M, Raviv A. 2002. Organization design. Manage-
theory of social decision schemes. Psychological ment Science 48(7): 852–865.
Review 80(2): 97–125. Hart O, Moore J. 2005. On the design of hierarchies:
Davis JH. 1992. Some compelling intuitions about coordination versus specialization. Journal of Political
group consensus decisions, theoretical and empiri- Economy 113(4): 675–702.
cal research, and interpersonal aggregation phenom- Hawn C. 2004. Microsoft skills. Fast Company 89: 70.
ena: selected examples, 1950–1990. Organizational Herbold RJ. 2002. Inside Microsoft: balancing creativity
Behavior and Human Decision Processes 52(1): 3–38. and discipline. Harvard Business Review 80(1):
Davison AC, Hinkley DV. 1997. Bootstrap Methods and 73–79.
Their Application. Cambridge University Press: New Hinsz VB, Tindale RS, Vollrath DA. 1997. The emerging
York. conceptualization of groups as information processors.
Efron B. 1979. Bootstrap methods: another look at the Psychological Bulletin 121(1): 43–64.
jackknife. Annals of Statistics 7(1): 1–26. Hrebiniak LG, Joyce WF. 1984. Implementing Strategy.
Efron B, Tibshirani RJ. 1993. An Introduction to the Macmillan: New York.
Bootstrap. Chapman & Hall: New York. Janis IL. 1972. Victims of Groupthink: A Psychological
Ethiraj SK, Levinthal DA. 2004. Bounded rationality Study of Foreign-Policy Decisions and Fiascoes.
and the search for organizational architecture: Houghton Mifflin: Boston, MA.
an evolutionary perspective on the design of Kahneman D, Tversky A. 1979. Prospect theory: analysis
organizations and their evolvability. Administrative of decision under risk. Econometrica 47(2): 263–291.
Science Quarterly 49(3): 404–437. Knudsen T, Levinthal DA. 2007. Two faces of search:
Fama EF. 1970. Efficient capital markets: review of alternative generation and alternative evaluation.
theory and empirical work. Journal of Finance 25(2): Organization Science 18(1): 39–54.
383–423. Kosowski R, Timmermann A, Wermers R, White H.
Foss NJ. 2003. Selective intervention and internal 2006. Can mutual fund ‘stars’ really pick stocks? New
hybrids: interpreting and learning from the rise evidence from a bootstrap analysis. Journal of Finance
and decline of the Oticon spaghetti organization. 61(6): 2551–2595.
Organization Science 14(3): 331–349. Langlois RN. 1997. Cognition and capabilities: opportu-
Galbraith JR, Kazanjian RK. 1986. Strategy Implementa- nities seized and missed in the history of the computer
tion: Structure, Systems, and Process (2nd edn). West industry. In Technological Innovation: Oversights and
Publishing: St. Paul, MN. Foresights, Garud R, Nayyar PR, Shapira Z (eds).
Garicano L. 2000. Hierarchies and the organization Cambridge University Press: New York; 71–94.
of knowledge in production. Journal of Political Lerner J. 1994. The syndication of venture capital
Economy 108(5): 874–904. investments. Financial Management 23(3): 16–27.
Garicano L, Posner RA. 2005. Intelligence failures: an Macmillan NA, Creelman CD. 2004. Detection Theory:
organizational economics perspective. Journal of A User’s Guide (2nd edn). Lawrence Erlbaum
Economic Perspectives 19(4): 151–170. Associates: Mahwah, NJ.
Garud R, Nayyar PR, Shapira Z. 1997. Technological Maritan CA, Schendel DE. 1997. Strategy and decision
choices and the inevitability of errors. In Technolog- processes: what is the linkage? In Strategic Decisions,
ical Innovation: Oversights and Foresights, Garud R, Papadakis V, Barwise P (eds). Kluwer Academic
Nayyar PR, Shapira Z (eds). Cambridge University Publishers: London, U.K.
Press: New York; 20–40. Marschak J, Radner R. 1972. Economic Theory of Teams.
Gavetti G, Levinthal DA. 2000. Looking forward and Yale University Press: New Haven, CT.
looking backward: cognitive and experiential search. Mintzberg H. 1978. Patterns in strategy formation.
Administrative Science Quarterly 45(1): 113–137. Management Science 24(9): 934–948.
Gavetti G, Levinthal DA, Ocasio W. 2007. Neo- Moore EF, Shannon CE. 1956/1993. Reliable circuits
Carnegie: the Carnegie School’s past, present, and using less reliable relays (I and II). In Claude Elwood
reconstructing for the future. Organization Science Shannon: Collected Papers, Sloane NJA, Wyner AD
18(3): 523–536. (eds). Wiley-IEEE Press: New York; 796–830.
Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj
632 F. A. Csaszar

Morningstar. 2008. Morningstar Funds 500 . Wiley: Simon HA. 1947/1997. Administrative Behavior (4th
Hoboken, NJ. edn). Free Press: New York.
National Commission on Terrorist Attacks upon the Starbuck WH, Mezias JM. 1996. Opening Pandora’s
United States. 2004. The 9/11 Commission Report: box: studying the accuracy of managers’ perceptions.
Final Report of the National Commission on Journal of Organizational Behavior 17(2): 99–117.
Terrorist Attacks Upon the United States. Norton: Stasser G, Titus W. 1985. Pooling of unshared informa-
New York. tion in group decision making: biased information
Peterson W, Birdsall T, Fox W. 1954. The theory sampling during discussion. Journal of Personality
of signal detectability. IRE Professional Group on and Social Psychology 48(6): 1467–1478.
Information Theory 4(4): 171–212. Stoner JAF. 1961. A comparison of individual and group
Pettigrew AM. 1987. Context and action in the decisions involving risk. Master’s thesis, School of
transformation of the firm. Journal of Management Industrial Management, Massachusetts Institute of
Studies 24(6): 649–670. Technology.
Pfeffer J. 1992. Managing with Power: Politics and Stratford S. 1993. Andy Grove: how Intel makes spending
Influence in Organizations. Harvard Business School pay off. Fortune, 22 February: 58.
Press: Boston, MA. Thompson J. 1967. Organizations in Action: Social
Puranam P, Powell BC, Singh H. 2006. Due diligence Science Bases in Administrative Theory. McGraw-
failure as a signal detection problem. Strategic Hill: New York.
Organization 4(4): 319–348. Van den Steen E. 2009. Authority versus persuasion.
Qian YY, Roland G, Xu CG. 2006. Coordination and American Economic Review 99(2): 448–453.
experimentation in M-form and U-form organizations. Van Fleet DD, Bedeian AG. 1977. A history of the
Journal of Political Economy 114(2): 366–402. span of management. Academy of Management Review
Radner R. 1992. Hierarchy: the economics of managing. 2(3): 356–372.
Journal of Economic Literature 30(3): 1382–1415. Van Zandt T. 1999. Real-time decentralized information
Raisch S, Birkinshaw J. 2008. Organizational ambidex- processing as a model of organizations with boundedly
terity: antecedents, outcomes, and moderators. Journal rational agents. Review of Economic Studies 66(3):
of Management 34(3): 375–409. 633–658.
Rausand R, Høyland A. 2004. System Reliability Theory: Williams RL. 2000. A note on robust variance estimation
Models, Statistical Methods, and Applications (2nd for cluster-correlated data. Biometrics 56: 645–646.
edn). Wiley-Interscience: Hoboken, NJ. Williamson OE. 1967. Hierarchical control and optimum
Rivkin JW, Siggelkow N. 2003. Balancing search and firm size. Journal of Political Economy 75(2):
stability: interdependencies among elements of 123–138.
organizational design. Management Science 49(3): Zajac EJ, Bazerman MH. 1991. Blind spots in industry
290–311. and competitor analysis: implications of interfirm
Robbins SP. 1990. Organization Theory: Structure, (mis)perceptions for strategic decisions. Academy of
Design, and Applications (3rd edn). Prentice Hall: Management Review 16(1): 37–56.
Englewood Cliffs, NJ. Zenger TR, Hesterly WS. 1997. The disaggregation of
Rotemberg JJ, Saloner G. 1994. Benefits of narrow corporations: selective intervention, high-powered
business strategies. American Economic Review 84(5): incentives, and molecular units. Organization Science
1330–1349. 8(3): 209–222.
Rumelt RP, Schendel DE, Teece DJ. 1994. Fundamental
Issues in Strategy: A Research Agenda. Harvard
Business School Press: Boston, MA.
Sah RK. 1991. Fallibility in human organizations and SUPPORTING INFORMATION
political systems. Journal of Economic Perspectives
5(2): 67–88. Additional supporting information may be found
Sah RK, Stiglitz JE. 1986. The architecture of economic in the online version of this article:
systems: hierarchies and polyarchies. American
Economic Review 76(4): 716–727. APPENDIX: Model
Sah RK, Stiglitz JE. 1988. Committees, hierarchies and Please note: Wiley-Blackwell is not responsible
polyarchies. Economic Journal 98(391): 451–470. for the content of functionality of any supporting
Siggelkow N, Levinthal DA. 2003. Temporarily divide to
conquer: centralized, decentralized, and reintegrated materials supplied by the authors. Any queries
organizational approaches to exploration and adapta- (other than missing material) should be directed
tion. Organization Science 14(6): 650–669. to the corresponding author for the article.

Copyright  2012 John Wiley & Sons, Ltd. Strat. Mgmt. J., 33: 611–632 (2012)
DOI: 10.1002/smj