three essays on organizational politics, search, and

THREE ESSAYS ON ORGANIZATIONAL

POLITICS, SEARCH, AND LEARNING

A DISSERTATION

SUBMITTED TO THE GRADUATE SCHOOL OF BUSINESS

AND THE COMMITTEE ON GRADUATE STUDIES

OF STANFORD UNIVERSITY

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

FOR THE DEGREE OF

DOCTOR OF PHILOSOPHY

Scott Cohn Ganz

June 2016

This dissertation is online at: http://purl.stanford.edu/dg252nd9441

© 2016 by Scott Cohn Ganz. All Rights Reserved.

Re-distributed by Stanford University under license with the author.

ii

http://purl.stanford.edu/dg252nd9441

I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.

John-Paul Ferguson, Primary Adviser


William Barnett, Co-Adviser


Steven Callander

Approved for the Stanford University Committee on Graduate Studies.

Patricia J. Gumport, Vice Provost for Graduate Education

This signature page was generated electronically upon submission of this dissertation in electronic format. An original signed hard copy of the signature page is on file inUniversity Archives.

iii

Abstract

My dissertation consists of three papers that develop new theory at the nexus of

organizational politics and organization learning. Each paper attempts to integrate

the disparate approaches to group decision-making in organizational theory and po-

litical economics. The first paper examines how preference conflict in organizations

impacts the likelihood that organizations will learn about the performance of existing

policies prior to making changes to organizational strategy. The second paper looks

at how career concerns of managers inside of organizations impact the character and

informativeness of offline experimentation. The third paper uses existing theory on

delegation and informational influence in political economics to better understand

how social movements utilize policy expertise to get access to the political process.

iv

Acknowledgments

This work is a celebration of all of my family and friends who have supported me

and sacrificed on my behalf to get me to where I am today.

I dedicate this work to my father, Steve, whose fight with pancreatic cancer

ended in the fall. He would have cherished this moment, just as he cherished all of

my achievements, large and small. I love him. And I miss him.

I thank my wife, Amy, who uprooted herself from D.C., moved across the country,

and made it work the best she could in Palo Alto so that I could attend Stanford.

She makes my world a loving, happy place. I would be lost without her.

I acknowledge my mother, Audrey, whose creativity, zest for life, and irrepressible

spirit are a constant source of wonder and inspiration.

I am deeply grateful to JP, whose constant reminders that my research is impor-

tant (so I should do it well) motivate me to keep doing better work.

Lastly, I remember Muppet. His playful spirit will remain with me always.

v

Contents

Abstract iv

Acknowledgments v

1 Introduction 1

2 What You Don’t Know Can’t Hurt You 8

3 Organizational Experimenting With Managers 48

4 Social Movements and Legislative Expertise 96

A Appendix to Chapter 2 120

B Appendix to Chapter 3 128

vi

List of Tables

4.1 Negative Binomial Regression Predicting the Count of EMOs

Testifying at a Hearing . . . . . . . . . . . . . . . . . . . . . . . . . . 113

4.2 Negative Binomial Regression Predicting the Count of Non-environmental

Interest Organizations Testifying at a Hearing . . . . . . . . . . . . . 116

4.3 Seemingly Unrelated Regression Predicting the Count of EMOs and

Non-environmental IOs Testifying at a Hearing . . . . . . . . . . . . 117

vii

List of Figures

2.1 Example 1: No Learning . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.2 Example 1: Learning; Low Outcome . . . . . . . . . . . . . . . . . . 30

2.3 Example 1: Learning, Moderate Outcome . . . . . . . . . . . . . . . . 31

2.4 Example 1: Learning, High Outcome . . . . . . . . . . . . . . . . . . 33

2.5 Learning vs. Not Learning, Low Uncertainty . . . . . . . . . . . . . . 41

2.6 Learning vs. Not Learning, Moderate Uncertainty . . . . . . . . . . . 42

2.7 Learning vs. Not Learning, High Uncertainty . . . . . . . . . . . . . . 43

2.8 Learning vs. Not Learning, High Uncertainty Reduction . . . . . . . 44

2.9 Learning vs. Not Learning, Moderate Uncertainty Reduction . . . . . 45

2.10 Equilibrium Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.1 Brownian Path with Known (x0,y0) = (0,2) . . . . . . . . . . . . . . 64

3.2 Brownian Path with Known (x0,y0) = (0,2) and (xp,yp(m)) = (3,�0.25) 66

3.3 x

⇤1(m) conditional of (xp,yp(m)) for (x0,y0) = (0,2) . . . . . . . . . 75

3.4 x

⇤1(m) conditional of (xp,yp(m)) for (x0,y0) = (0, 14) . . . . . . . . . 76

3.5 Expected Utilities under r = GT . . . . . . . . . . . . . . . . . . . . 82

3.6 x⇤p under r = GT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.7 Expected Utilities under r = RS . . . . . . . . . . . . . . . . . . . . . 85

3.8 x⇤p under r = RS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

viii

3.9 Lower Cutpoint under r = FR . . . . . . . . . . . . . . . . . . . . . . 87

3.10 x⇤p under r = FR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

3.11 Expected Utilities under r = OB . . . . . . . . . . . . . . . . . . . . 90

3.12 Lower Bound under r = OB . . . . . . . . . . . . . . . . . . . . . . . 92

3.13 Replacement Probabilities and EUO under Different Replacement Rules 93

3.14 Comparing Guaranteed Tenure and Outcome-based Replacement . . . 94

4.1 Comparison of Means for EMOs

and Non-environmental Interest Organizations . . . . . . . . . . . . . 115

ix

Chapter 1

Introduction

Over the past half-century, the intellectual descendants of Simon, March, Cyert, and

Lindblom in the fields of organizational theory and political science have traveled di-

vergent paths. Organizational research in the Carnegie School tradition focused on

understanding psychological biases of individuals, modeling chaotic decision-making

processes within organizations, and grappling with interdependencies between orga-

nizational strategy and competitive selection (Gavetti et al., 2012, 2007). For the

intellectual descendants of the Carnegie School, agent-based simulations and evo-

lutionary logic became the standard tools for formalizing theoretical insights and

generating novel predictions (Cohen et al., 1960, 1972; Cyert and March, 1963; Har-

rison et al., 2007). Carnegie School scholars emphasized the importance of under-

standing how individuals in organizations learn, how learning facilitates the creation

of organizational strategy, and how strategies evolve in response to changes in the

organizational environment.

Lost along the way, however, was the initial emphasis on the politics of organi-

zational life and, by extension, the politics of organizational learning. The concept

1

CHAPTER 1. INTRODUCTION 2

of the dominant coalition and an academic focus on coercive institutions moved

political coalitions and preference conflict to the periphery of organizational anal-

ysis (Scott, 1987; Simon, 1976; Thompson, 1967). In many organizational theory

models, the interests of organization members are assumed to be aligned with orga-

nizational performance or survival. In other models, it is not clear that organization

members have preferences at all (Bendor, 2010; Bendor et al., 2001; Ethiraj and

Levinthal, 2009; Schilling and Fang, 2014, are important counter-examples). To the

extent that preference conflict is considered in formal modeling efforts, internal dis-

sent arises from differences in experiences, rather than from roles, desires, expertise,

or incentives (Denrell, 2003; Denrell and March, 2001; March, 2006, 2010). While

inter-organizational conflict and politics did remain a part of qualitative work in or-

ganizational behavior (e.g. Eisenhardt and Bourgeois, 1988; Morrill, 1991, 1995; Zald,

1970) and has been re-introduced in recent empirical work at the intersection of so-

cial movements and organizational theory (e.g. Ingram and Simons, 2000; Rao et al.,

2000; Soule, 2012; Weber et al., 2009), the importance of politics in organizational

learning has largely been ignored.

In contrast, over the past half-century, positive political theory and political

economics have placed increasing emphasis on the role of preference conflict in the

study of collective decision-making in government (Diermeier and Krehbiel, 2003;

Moe, 2012). Theoretical approaches to political decision-making became dominated

by non-cooperative game theory and, in particular, games of incomplete information.

Standard models of collective behavior in the legislature were built on the spatial

models of competition that were initially developed in industrial organization eco-

nomics (Austen-Smith and Banks, 1999; Black et al., 1958; Downs, 1957; Krehbiel,

1988). Modern studies of government decision-making characterize legislatures and

bureaucracies as consisting of self-interested individual actors: “a ‘they,’ not an ‘it’.”


(Shepsle, 1992)

Policymaking uncertainty remained an important element of political economy

models. Bounded rationality, incomplete information, and risk aversion became the

theoretical mechanisms that explained why rational individual legislators would agree

to participate in coercive political institutions and delegate decision-making author-

ity to bureaucrats (Bendor and Meirowitz, 2004; Callander, 2008; Gilligan and Kre-

hbiel, 1987, 1990; Krehbiel, 1991). However, rich models of decision-making complex-

ity were pushed aside in favor of those that permitted greater analytic tractability.

Formal models in political science became more interested in the institutions that

facilitate successful organizational learning and superior collective decision-making

than the cognitive processes that underlie organizational learning and the social

forces that promote collective action.

In my dissertation, I reconnect the disparate approaches to group decision-making

in organizational theory and political economics. My dissertation work imports tools,

analytical models, and concepts from political economics into the study of the pol-

itics of organizational learning. The formality and analytical simplicity offered by

spatial models of intergroup conflict provide a tractable way for organizational the-

ory to model the impact of intra-organizational politics on organizational outcomes.

Understanding legislative institutions as a product of risk-averse, boundedly rational

legislators can help social movement theory make sense of the strategic interactions

between private organizations and legislatures. Furthermore, recent attempts to add

richer depictions of policy uncertainty to the formal analysis of political experimen-

tation make it easier for learning theorists to integrate non-cooperative game theory

and endogenous search into realistic models of learning in uncertain environments.

My dissertation consists of three papers that develop new theory at the nexus of

organizational politics and organization learning. The first examines how preference


conflict in organizations impacts the likelihood that organizations will learn about the

performance of existing policies prior to making changes to organizational strategy.

The second looks at how career concerns of managers inside of organizations impact

the character and informativeness of offline experimentation. The third uses existing

theory on delegation and informational influence in political economics to better

understand how social movements utilize policy expertise to get access to the political

process.

Short summaries of each of the three papers follow:

Paper 1: What You Don’t Know Can’t Hurt You

The first paper explores how preference conflict within organizations influences

whether or not organizations choose to learn. The paper emphasizes that failures in

organizational learning need not be the result of cognitive biases, costly information,

or faulty information-aggregation routines. Organizations are political coalitions

that face internal contestation over organizational strategies and goals. The decision

whether to collect information impacts both the goals that organization members

try to meet and the organization’s capacity to meet them. The key insight of the

paper is that even in organizations where actors are rational and where there are no

difficulties in aggregating information, organization members may still choose not to

learn, because members can attain better organizational outcomes (for themselves) if

the organization does not learn than if it does. Furthermore, the paper demonstrates

that individual incentives for strategic ignorance increase when existing policies are

less desirable and when there is greater uncertainty in the relationship between poli-

cies and outcomes. As a result, organizations characterized by political conflict will

not learn when they are likely to be able to build a consensus about the desirability

of organizational change and will learn when they are unlikely to be able to build


such a consensus.

Paper 2: Organizations Learning About Managers Learning About Strategies

The second paper examines the challenges associated with offline experimentation

in organizations made up of managers who are primarily concerned with their own

careers and secondarily concerned with organizational performance. Offline experi-

ments and prototyping are central to the design thinking approach to organizational

innovation and are also a core practice in evidence-based management. However,

unlike online search, for which tendencies towards exploitative and risk-averse inno-

vation are well documented, offline experimentation in organizations is often assumed

to be unproblematic in the organizational literature on search and innovation. My

paper demonstrates how managers who are fearful of being replaced in response to a

poor experimental outcome will pervert the experiments they choose to undertake,

sapping offline experiments of their informational value in the process.

In particular, the paper shows that when organizations evaluate a manager’s

offline prototype relative to the existing status quo, managers tend to be overly risk-

averse in their experimentation behavior. In contrast, if organizations evaluate man-

agers based on whether they are more desirable than an unexperienced replacement,

managers will tend to be overly exploratory in their experimentation behavior. The

paper also examines two policies that organizations can implement in order to try to

overcome this agency problem. Guaranteeing tenure alleviates the career concerns

of managers by ensuring that they will keep their job, notwithstanding the desirabil-

ity of their experiment’s outcome. However, a tenure guarantee also requires that

organizations sometimes keep managers who are likely to be a poor fit. An outcome-

based replacement rule can induce managers to design an experimental prototype in

the way that the organization finds desirable, but also requires that organizations


sometimes replace managers who are likely to produce high quality organizational

outcomes. The paper shows that the freedom afforded by the tenure guarantee leads

to better organizational performance in high complexity task environments. The or-

ganizational control exerted by the outcome-based rule, in contrast, is more desirable

in low uncertainty environments.

Paper 3: Social Movements as Purveyors of Legislative Expertise

The third paper, co-authored with Sarah Soule1, examines the role of informa-

tional influence in the impact of social movements and social movement organizations

(SMOs) on the legislative process. This empirical paper builds on theoretical models

of legislative delegation in which a rational legislature facing an unknown relation-

ship between policies and policy outcomes has the opportunity to delegate legislative

decision-making to an expert organization or agency. We argue that the existing lit-

erature on the role of social movements in the policy process focuses too narrowly on

the importance of mobilizing constituents and providing signals about constituent

preferences to legislators. Modern social movement organizations are also often in-

volved in the policy process by offering legislative expertise to legislators on technical

issues. The environmental movement, in particular, has effectively leveraged its ac-

cess to the environmental science community to become influential in the design of

environmental protection legislation and regulation.

We use a new dataset of organizational testimony at congressional hearings on

environmental protection to examine this channel of influence. We find that envi-

ronmental SMOs are invited to testify in greater numbers in hearings that consider1Scott C. Ganz is the first author on this paper, as he was the primary writer, oversaw data col-

lection, and conducted all statistical analyses. The paper is currently being prepared for submissionto journals.


a specific piece of proposed legislation than in hearings that are exploratory or in-

vestigatory in nature. We also find that this increase in representation is unique

to social movement organizations with an environmental focus; non–environmental

interest organizations do not experience a similar improvement in legislative access.

These findings suggest that, as a result of their scientific expertise, environmental

SMOs are receiving privileged access to the policy process relative to other interest

organizations affected by environmental regulation.

Chapter 2

What You Don’t Know Can’t Hurt

You

Introduction

Two of the most enduring legacies of the Carnegie School of organizational theory

are the importance of organizational learning and the image of the firm as a political

coalition (Cyert and March, 1963; March and Simon, 1958). Despite their common

roots, these two concepts are rarely integrated (Gavetti et al., 2012, 2007; Gibbons,

2003). Research in organizational learning moved quickly into formal modeling and

computer simulations that conceived of the organization as an actor facing a con-

strained optimization problem. Such models implicitly assume that all organizational

decision-makers have the same preferences. Attempts to create models of organiza-

tions characterized by individuals with different preferences, which is the essence

of organizational politics, required confronting the problems of chaotic collective

preferences and unstable collective goals that plagued social choice theorists at the

time (see Austen-Smith and Banks 1999). As a result, most organizational learning

8

CHAPTER 2. WHAT YOU DON’T KNOW CAN’T HURT YOU 9

models assume that organizational preferences over outcomes are well-ordered and

disregard the process through which individual preferences are aggregated (Bendor

et al., 2001; Ethiraj and Levinthal, 2009; Schilling and Fang, 2014, are important

counter-examples).

Ignoring politics had a cost: modern theory solely ascribes organizational fail-

ures to learn to imperfections in information collection and decision-making biases.

In his 2010 book, The Ambiguities of Experience, James March begins his inquiry

with what appears to be an uncontroversial assumption: “Organizations pursue in-

telligence” (March, 2010, p. 1). The remainder of his book—and the lion’s share

of research on organizational learning—describes all of the ways that organizations

fail in this pursuit: organization members imperfectly collect information, which is

badly aggregated by managers, who then make biased decisions based on the infor-

mation they gather (Csaszar and Eggers, 2013; Denrell and March, 2001; Levinthal

and March, 1993; Levitt and March, 1988; March and Simon, 1958). March acknowl-

edges that his treatment of organizational learning ignores the role that conflicts of

interest within organizations play in “influenc[ing] not only the pursuit of intelli-

gence but also its definition” (March, 2010, p. 6). But readers searching elsewhere

for models that describe how internal politics impact the pursuit of organizational

intelligence will be left wanting.

My paper examines conditions under which rational organizations characterized

by political conflict would choose not to pursue knowledge. Organizations are politi-

cal coalitions that face internal contestation over organizational strategies and goals.

The decision whether to collect information impacts both the goals that organization

members try to meet and the organization’s capacity to meet them. More informa-

tion can make it possible for organization members to overcome internal opposition

and implement controversial new policies. But, more information can also provide


internal opponents with ammunition to fight back against proposals to implement

organizational change. For example, in an internal negotiation between management

and the workforce over salary and benefits, an in-depth study of the compensation

practices of close competitors could show that the current scheme is more generous

than existing market norms, which would support proposals by management for ben-

efit cuts. But, the survey could instead uncover that the current regime lags the rest

of the market substantially. Labor leaders could then use this information in support

of more generous compensation. Thus, under some conditions, rational organization

members might choose not to gather more information, if they suspect that greater

clarity will help their opponents.

My paper demonstrates that organizations navigating in a fog need not be full of

irrational members and inoperative information sharing systems. Conflicting pref-

erences among rational members is sufficient in some cases to convince organization

members not to try to learn. Indeed, in organizations characterized by preference

conflict, strategic uncertainty can help organization members overcome political con-

straints on action.

I develop a model that brings political conflict back into organizational learning.

I utilize the analytical machinery for preference aggregation developed in political

economics research (Diermeier and Krehbiel, 2003). That machinery, which is im-

plemented through the “spatial model,” has typically been applied to government

decision-making and majority voting. I show that the intuition underpinning spatial

preferences can be applied to a variety of organizational settings. My paper has a

key insight: even in organizations where actors are rational and where there are no

difficulties in aggregating information, organization members may still choose not to

learn, because members can attain better organizational outcomes (for themselves)

if the organization does not learn than if it does. Further, the model demonstrates


that individual incentives for strategic ignorance increase when existing policies are

less desirable and when there is greater uncertainty in the relationship between exist-

ing policies and their outcomes. As a result, organizations characterized by political

conflict will not learn when they are likely to be able to build a consensus about the

desirability of organizational change and will learn when they are unlikely to be able

to build such a consensus.

Background

Institutional Analysis

I examine the relationship between politics, uncertainty, and learning through

institutional analysis (Diermeier and Krehbiel, 2003). Institutional analysis consists

of four parts. First, I define a series of behavioral postulates for each actor. These

behavioral postulates define the actors’ utility functions and the behavioral rules that

govern individual decision-making. Second, I identify the key institutions of collec-

tive choice. These institutions determine the order in which decisions are made, the

set of behaviors that each actor can take, and the information available to each actor

when it is their turn to act. The third part of institutional analysis logically derives

the behaviors of each actor and describes the collective outcomes resulting from their

choices. The fourth part then uses these predicted behaviors to generate hypotheses

that can be applied to data.1

1Institutional analysis bears strong resemblance to the theory of conflict systems in March (1962),the political model of choice described in Pfeffer (1981), and the political economy approach toorganizational analysis in Zald (1970). March argues that analyses of conflict systems require twonecessary conditions: “(1) that the elementary decision processes be susceptible to treatment asconsistent basic units, and (2) that analytic procedures be available to explore the properties of themodel,” where “consistent basic units...can be defined as having a consistent preference orderingover the possible states of the system” (March 1962: 663). Pfeffer (1981) adds two additionalnecessary conditions in order to “understand organizational choices using a political model”: (1)“what determines each actor’s relative power” and (2) “how the decision process arrives at a decision.”


Behavioral Postulates: Spatial Modeling, Euclidean Preferences, and Outcome

Uncertainty

I define the individual preferences for each actor using spatial modeling. Spatial

models map possible organizational outcomes into Euclidean space and identify an

ideal outcome for each actor. Utility is a strictly decreasing function of the Euclidean

distance between an actor’s ideal outcome and the realized outcome.

Spatial modeling defines actor preferences in a way that accords with theories

of preference conflict in organizations. Most important decisions in organizations

involve tradeoffs about organizational outcomes (Pondy, 1992). Common tradeoffs

include short-term goals versus long-term goals, risk vs. return, and social impact vs.

financial performance. Although all organization members may agree that creating

positive social impact is better than causing social harm, for example, they are likely

to disagree about the ideal balance of social impact and corporate profits. Organiza-

tion members in charge of corporate social responsibility (CSR) would be willing to

forego corporate profits in exchange for positive social impact; shareholders would

prefer high profits at the expense of social welfare goals; managers are somewhere in

the middle.

The spatial model takes this intuition about preferences and projects it into

Euclidean space. We can illustrate the tradeoff between social impact and corporate

profits through an analogy to the set of real numbers, where low values indicate

negative social impact (and more profits) and high values indicate positive social

impact (and less profits). The numerical ideal outcome of the shareholder is less

than the ideal outcome of the manager, which is less than the ideal outcome of

the employee working in CSR. Lower policy outcomes are most desirable for the

shareholder; middling outcomes are most desirable for the manager; high outcomes


are most desirable for the worker in CSR.

Preferences over strategies or policies can be defined spatially in terms of the

outcomes they are expected to produce. For example, in an investment bank there

is likely to be outcome disagreement over the relative levels of risk and expected

return in an investment portfolio. Risk managers prefer lower volatility and more

modest returns. Traders are more risk-seeking. Internal conflict will arise when risk

managers try to implement policies that substitute bureaucratic reporting rules and

restrictive risk limits for trader discretion, because more rules and tighter oversight

generate outcomes more desirable for risk managers than for traders.

However, in many organizational contexts, environmental turbulence, implemen-

tation error, and unpredictable competitor behavior generate stochasticity in the

mapping between the strategies an organization implements and the outcomes that

emerge as a result. Therefore, overt internal conflict in organizations is rarely over

outcomes per se. Instead, political disagreement usually takes place when organiza-

tions set strategy or implement policy. Policies are under the control of organization

members; outcomes are not.

Uncertainty about current policies, however, is qualitatively different from uncer-

tainty about policy changes. The mapping between a status quo policy and status

quo outcome is potentially learnable. A new policy that results from an organiza-

tional change, in contrast, brings with it idiosyncratic uncertainty that is unknowable

prior to its implementation. Lessons learned prior to a policy change may have lim-

ited applicability afterward.

Institutions of Collective Choice: An Agenda-Setting Proposer and an Organiza-

tional Veto

Decisions in organizations frequently involve one actor with proposal power (or


agenda-setting power) and another that has the authority to reject, or veto, pro-

posals. As a result of different interests, financial incentives, operational roles, or

psychological perspectives, the preferences of the proposer and the preferences of the

veto are often not entirely aligned. As a result, the proposer must take the prefer-

ences of the veto into account if they desire to initiate a change to the organization.

This is the decision-making architecture that I examine in my model.

This decision-making architecture has applications to broad organizational con-

texts. It closely resembles top-down decision-making processes in which a manager

proposes a strategic change to a work unit; the work unit, in turn, has the op-

portunity to integrate the change or fight against it (Dalton, 1959; Freeland, 2001;

Granovetter, 2005; Granovetter and Swedberg, 2011). The subordinate’s veto reflects

a classic tension in hierarchical organizations: the managers responsible for setting

the organization’s strategy are less expert than the work units whom they seek to

control (Barnard, 1938; Simon, 1976; Weber, 1978). In Richard Freeland’s study of

the implementation of the “M-form” at General Motors, he concludes: “When top

management makes decisions in a way that appears to eschew expert opinion, subor-

dinates are likely to view the resulting decisions as unjustified and either will put forth

less than consummate effort in carrying them out or will engage in outright resistance

to their implementation” (Freeland, 2001, p. 30). However, the described decision-

making architecture also can be applied to bottom-up decision-making processes in

which work units are in charge of surveying the environment and then proposing

policy changes to management, who can either “rubber stamp” or reject the pro-

posals (Csaszar, 2013; Knudsen and Levinthal, 2007; Sah and Stiglitz, 1986). The

proposal-veto collective choice structure is also relevant to political decision-making

in a government with a legislature responsible for proposing legislation and an exec-

utive with veto authority (Cameron and McCarty, 2004; Krehbiel, 1997; Matthews,


1989; Shepsle and Weingast, 1981).

The Model

Summary

Two organization members, an agenda-setting proposer, P , and a member with

veto power, V , interact to determine an outcome, o, in R. P and V face an uncer-

tain status quo outcome produced by a known status quo policy. There also exist

a continuum of potential alternative policies that map to alternative outcomes with

uncertainty. P and V first decide whether to learn the status quo policy. Then, P

and V negotiate over whether to replace the status quo with an alternative policy.

Preferences and Information

P and V have utility functions, up and uv, that are defined spatially in R, where

o = 0 is P ’s ideal outcome and o = 1 is V ’s ideal outcome. P and V are risk-neutral.

Their utilities are defined by the negative Euclidean distance between their ideal

outcomes and o.

up(o) = �|o|

uv(o) = �|1� o|

There exists an exogenous status quo policy, p 2 R, which maps stochastically to

a policy outcome, p(p), where p(p) = p+Xp. Xp is a random variable distributed

uniformly on [�↵,↵], where ↵ 2 (0, 12 ]. Higher values of ↵ indicate greater uncer-

tainty in the relationship between the status quo policy and its associated outcome.

There also exist a continuum of alternative policies, a 2 R, which map stochas-

tically to alternative outcomes, a(a). The level of stochasticity in the outcomes


resulting from alternative policies is endogenously determined by the behavior of P

and V in the game. a(a) = a + 1I · XIa + (1 � 1I) · X¬I

a , where 1I evaluates to 1

if P and V choose to learn p(p) at the outset of the interaction and evaluates to

0 otherwise. If p(p) is learned, then a(a) = a +XIa . If p(p) is not learned, then

a(a) = a+X¬Ia .

The uncertainty associated with a proposed alternative if p(p) is not learned,

X¬Ia , is distributed uniformly on [�↵,↵]. If p(p) is learned, then the uncertainty

in the mapping between alternatives and their resulting outcomes may be reduced.

The uncertainty associated with a proposed alternative if p(p) is learned, XIa , is

distributed uniformly on [�↵�, ↵�], where � � 1. Higher values of � are associated

with greater uncertainty reduction in the relationship between a and a(a) if p(p)

is learned.2

p, ↵, and � are known by both P and V prior to the game.

Collective Choice Process

The interaction consists of two subgames: the learning subgame and the policy

subgame. In both subgames, P acts as agenda-setter and V has the opportunity to

exercise a veto.

In the learning subgame, P and V decide whether to learn p(p). Then, in the

policy subgame, P can propose an alternative, a, to V . If in the policy subgame, P

proposes an alternative and V accepts the proposal, then the outcome, o, is a(a).

Otherwise, the outcome is p(p).2Note that the mechanism through which learning p(p) generates reduced uncertainty in the

model is through improved “shock absorption” (see Bendor and Meirowitz (2004)) rather thanthrough a correlation between Xp, XI

a and X¬Ia . However, an alternate formulation in which Xp

and Xa are distributed bivariate normal with common variance and a correlation coefficient betweenzero and one generates substantively similar, although less analytically tractable, results.


In the learning subgame, P has the option to propose to V that p(p) be learned.

V then has the opportunity to exercise a veto. The decision to learn p(p) in the

model is akin to the decision by an organization to initiate a program evaluation

(Weiss 1998). If the organization undertakes the evaluation, the organization not

only learns information about existing policies, but also may reduce the uncertainty

surrounding the outcomes resulting from a policy change.

The proposal strategy in the learning subgame is defined by the binary choice

LP 2 {¬L,L}

where LP= ¬L indicates the decision by P not to propose to learn p(p) and

LP= L indicates the decision by P to propose to learn p(p).

The veto strategy in the learning subgame is defined by the binary choice

LV 2 {¬A,A}

where LV= ¬A indicates that V has exercised its veto and LV

= A indicates

that V has accepted the proposal to learn p(p).

If LP= L and LV

= A, then P and V learn p(p). Otherwise, P and V enter

the policy subgame knowing p, but not p(p).

Following the learning subgame, P and V begin the policy subgame. P has the

option to propose a policy alternative, a, to V . The proposal strategy is a choice

PPi 2 {p, a 2 R}

PPi = p indicates that P has not proposed a policy alternative. As a result, p(p)

becomes the outcome, o.


PPi = a indicates that P has proposed a policy alternative to V . If accepted,

a(a) becomes the outcome, o.

The subscript, i 2 {¬I, I} indicates whether P and V are informed about p(p)

prior to beginning the policy subgame. If p(p) has been learned, then i = I.

Otherwise, i = ¬I.

If PPi = a, then V has the choice to accept the alternative or to veto the proposal.

If V vetoes, then p(p) becomes the outcome, o. If V accepts, then a(a) becomes

the outcome, o. The veto strategy for V is a binary choice

PVi 2 {p, a}

where PVi = p indicates that V has exercised its veto and PV

i = a indicates that

V has accepted P ’s proposal. i is defined as above.

At the conclusion of the interaction, P and V receive utility defined by the neg-

ative Euclidean distance of the outcome, o, from their individual ideal outcomes.

Strategies and Solution Concept

Both participants have the ability to perfectly predict the actions of their alter.

As a result, they make decisions that are best responses to all the predicted future

decisions in the game. I employ a subgame perfect Nash equilibrium (SPNE) concept

for the model, which is common in rational actor models of this sort. In the subgame

perfect refinement of Nash equilibrium, actors make expected utility maximizing

decisions conditional on reaching every stage in the game. Unlike Nash equilibrium,

which permits irrational strategies in parts of the game that will never be reached in

equilibrium, SPNE requires full sequential rationality of actor behavior (Fudenberg

and Tirole, 1991).


The key results from the model are whether the subgame perfect strategy for the

proposer includes learning p(p) and whether the subgame perfect strategy for the

veto includes accepting a proposal to learn p(p). If learning is part of the subgame

perfect strategy for P , then the expected utility for the proposer if p(p) is learned

exceeds the expected utility for the proposer if p(p) is not learned. If accepting a

proposal to learn is part of the subgame perfect strategy for V , then the expected

utility for the veto if p(p) is learned exceeds the expected utility for the veto if p(p)

is not learned.

The strategy for P , SP , is a triple of decisions, (LP ,PP¬I ,PP

I ). Define the sub-

game perfect equilibrium strategy for P , S⇤P , as (LP⇤,PP⇤

¬I ,PP⇤I ), which is the set of

decisions that result in the highest expected utility for P conditional on available

information and knowledge of the best responses of V .

The strategy for V , SV , is a triple of decisions, (LV ,PV¬I ,PV

I ). Define the subgame

perfect equilibrium strategy for V , S⇤V , as (LV ⇤,PV ⇤

¬I ,PV ⇤I ).

I introduce a series of tie-breaking assumptions in order to simplify the formal

analysis. If the proposer is indifferent between learning and not learning, the proposer

chooses to learn. If the proposer is indifferent between proposing an alternative or

retaining the status quo policy, the proposer retains the status quo. Finally, if the

veto is indifferent between accepting or vetoing a proposal from the proposer in either

subgame, the veto accepts the proposal.

I illustrate through the model that, depending on the status quo policy (p), the

level of status quo policy uncertainty (↵), and the amount of uncertainty associated

with policy alternatives that is reduced if p(p) is learned (�), the proposer will

sometimes choose to propose to learn p(p) and sometimes will prefer to keep the

organization from learning p(p). Similarly, depending on p, ↵, and �, the veto will

sometimes accept a proposal to learn p(p) and sometimes use its veto in order to


keep the organization from learning p(p).

Equilibrium Results

First, I demonstrate equilibrium existence.

Proposition 1: A unique, pure-strategy SPNE exists

The game is finite and all players have perfect information. Therefore, the game

can be solved by backwards induction and has a pure-strategy Nash equilibrium.

Because all players have strict preferences over possible outcomes (and ties are broken

deterministically), there exists a unique SPNE (Fudenberg and Tirole, 1991).

I derive the solution to the game using backwards induction. First, I examine the

equilibrium behavior in the proposal subgame conditional on p(p) not being learned.

Proposition 2: In the proposal subgame, conditional on p(p) not being learned, the

following defines the Nash equilibrium strategies for P and V:

PP⇤¬I =

8>>>>>>>><

>>>>>>>>:

a = 0 if p 0

p if 0 < p 1

a = 2� p if 1 < p 2

a = 0 if p > 2

PV ⇤¬I = a

If p(p) is not learned, then P and V have no additional information about the

distribution of outcomes resulting from the set of alternatives. Therefore, V decides


whether or not to accept an alternative by comparing the distance of p from its ideal

point with the distance of a from its ideal point. P proposes the alternative, a, that

is as close to 0 as possible that is at least as close to 1 as p.

For p 0 and p > 2, the Euclidean distance of p from V ’s ideal outcome is

sufficiently large that P can propose a = 0 (the best possible alternative from P ’s

perspective) and V will accept.

For p 2 [0, 1], there are no alternatives that P can propose that will also make

V weakly better off. As a result, P does not make a proposal and retains the status

quo policy, p.

For p 2 (1, 2], P is partially constrained by V ’s veto power. The distance of p

from V ’s ideal point is p � 1. As a result, P proposes a equidistant from 1 as p.

a = 2� p.

Because all of the alternatives proposed make V weakly better off, V accepts in

equilibrium.3

Next, I examine the equilibrium behavior in the proposal subgame conditional

on p(p) being learned. Importantly, if p(p) is learned, then the uncertainty in the

mapping between a and a(a) may be reduced.

3For complete proofs, please see Appendix A.


Proposition 3: In the proposal subgame, conditional on p(p) being learned, the fol-

lowing defines the Nash equilibrium strategies for P and V:

PP⇤I =

8>>>>>>>>>>>><

>>>>>>>>>>>>:

a = 0 if p(p) � ↵2�

p if p(p) 2⇣� ↵

2� , 1 +↵2�

⌘

a = 1�r

( p(p)� 1) 2

↵��⇣↵�

⌘2if p(p) 2

⇣1 +

↵2� , 1 +

↵�

i

a = 2� p(p) if p(p) 2⇣1 +

↵�, 2i

a = 0 if p(p) > 2

PV ⇤I = a

The rationale for P ’s proposal strategy if p(p) is learned is the same as had

p(p) not been learned. P aims to propose the alternative closest to 0 that makes V

weakly better off than p(p).

For p(p) sufficiently less than zero or greater than 2, P and V both prefer an

alternative policy at 0 to the known status quo policy outcome.

For p(p) that is negative but near zero, P prefers a known p(p) to any uncertain

alternative policy. The cutoff between P proposing a = 0 and retaining the status

quo policy, p, is one-quarter the range of uncertainty associated with an alternative

policy, or ↵2� .

Similarly, for p(p) sufficiently near 1, there are no alternatives that V finds an

improvement over the status quo. This range is from 1� ↵2� to 1+

↵2� . Therefore, for

all p(p) between � ↵2� and 1 +

↵2� , P does not propose an alternative policy.


Finally, for p(p) 2⇣1 +

↵2� , 2

⌘, P proposes the alternative for which V is indif-

ferent between retaining the status quo and accepting a.

Now, I examine the equilibrium behavior for P and V in the learning subgame,

conditional on the best responses in the proposal subgame.

Proposition 4: In the learning subgame, the following defines the Nash equlibrium

strategies for P and V.

LP⇤=

8><

>:

L if Eup(LP= L) � Eup(LP

= ¬L)

¬L otherwise

LV ⇤=

8>>>>><

>>>>>:

A if p �↵� ↵2�

¬A if p 2⇣�↵� ↵

2� ,↵2� � ↵

⌘

A otherwise

In the learning subgame, the equilibrium behavior for V is more straightforward

than for P .

In order to demonstrate the values of p for which V will veto a proposal to

learn p(p), I partition R into six segments, Pi, where i 2 {1, 2, 3, 4, 5, 6}. P1 =

(�1,�↵� ↵2� ], P2 = (�↵� ↵

2� , 0], P3 = (0, 2� ↵], P4 = (2� ↵, 2], P5 = (2, 2 + ↵],

and P6 = (2 + ↵,1).

For p 2 P1, for all possible values of p(p), P will propose a = 0 if p(p) is

learned. If p(p) is not learned, then P will propose a = 0 as well. Therefore, V is

indifferent between learning p(p) and not. As a result, V will accept a proposal to


learn p(p).

For p 2 P6, the same holds. For all possible values of p(p), P will propose a = 0

if p(p) is learned. If p(p) is not learned, then P will propose a = 0 as well.

For p 2 P2, if p(p) is not learned, then P will propose a = 0, V will accept,

and Euv = �1. Therefore, by learning p(p), V trades off the possibility that

p(p) 2 (� ↵2� , 0), in which case PP⇤

I is p < 0 and Euv < �1 against the possibility

that p(p) � 0, PP⇤I is p � 0, and Euv � �1. For p 2 (�↵� ↵

2� ,↵2� �↵), Pr( p(p) 2

(� ↵2� , 0)) > Pr( p(p) > 0) and V prefers not to learn p(p). For p 2 [

↵2� � ↵, 0], the

opposite is true and V prefers to learn p(p).

For p 2 P3, if p(p) is learned and P proposes an alternative, then V is indifferent

between the alternative and p(p). Similarly, if p(p) is not learned and P proposes

an alternative, V is made exactly as well off as if the status quo policy were to

remain. Therefore, V is indifferent between learning p(p) and not.

For p 2 P4, if p(p) is not learned, then P will propose an alternative that makes

V indifferent between accepting the alternative and retaining p. In contrast, if p(p)

is revealed to be greater than 2, P will propose a = 0, which makes V strictly better

off than p(p). Therefore, V prefers that p(p) be learned.

For p 2 P5, if p(p) is not learned, P will propose a = 0 and Euv(a) = �1. If

p(p) is revealed to be greater than 2, then P will also propose a = 0. But, if p(p)

is revealed to be less than 2, then P will propose a > 0 and Euv(a) > �1. As a

result, V is strictly better off if p(p) is learned.

P ’s decision whether or not to learn is considerably more involved. Implicitly, if

the expected utility from learning exceeds the expected utility from not learning, P

will propose to learn p(p). I demonstrate the equilibrium behavior for P through a

series of numerical examples in order to generate intuition for the conditions under


which not learning dominates learning and vice versa. The first two examples illus-

trate cases where internal politics create incentives for the proposer not to learn. The

third demonstrates how political considerations can make learning more desirable,

rather than less.

Example 1: Price-setting in a Manufacturing Firm

Background

Zbaracki and Bergen (2010) describe the internal politics of price-setting in a

manufacturing firm that sells products to warehouse distributors. They emphasize

the conflict between the marketing team and the sales team. The organization’s

goals included increasing market share and maximizing profit margins. Of these two

goals, the marketing team placed greater emphasis on the market share goal and

the sales team placed greater emphasis on the profit margin goal. These emphases

reflected the roles of each team in the company. Marketers were in charge of setting

pricing strategies that would enlarge the company’s client base. Members of the sales

team were tasked with developing long-term relationships with distributors, which

entailed significant discretion over discounting.

The disagreement over organizational outcomes created conflict over pricing strate-

gies. The marketing team desired lower prices across the board and less individual

discretion for the sales team. The sales team preferred higher prices and greater

discretion to discount. The marketing team justified their position by arguing that

prospective customers paid significant attention to list prices and that there was no

evidence that selective discounting increased revenue for the company. The sales

team believed that large customers did not take list prices seriously. Lower prices

merely constrained the company’s ability to price discriminate.

The conflict remained latent until the firm’s management decided to make a large


capital investment and to purchase a product line from a close competitor, generating

a sizable decline in production costs. The organizational change had a direct impact

on the firm’s existing pricing policy. If list prices were kept at the same levels, then

the sales team would have even more discretion than before. However, the marketing

team had institutional control over the published price lists. Marketing viewed the

decreased production costs as an opportunity to slash prices across the board.

Political conflict erupted between the marketing team and the sales team. Mar-

keting criticized sales for being “champions of high price” (Zbaracki and Bergen,

2010, p. 965). Sales responded that marketing had “very minimal experience in this

industry” and titled the MBA-laden pricing team “mahogany row” (Zbaracki and

Bergen, 2010, p. 965). According to one analyst, meetings between marketing and

sales to negotiate new pricing policies became so contentious that they approached

physical violence.

Notwithstanding the overt disagreement over organizational strategy, the mar-

keting team made no serious attempt to collect information or to model alternative

pricing schemes. Marketing admitted that they could not predict customer response

to a large cut in prices. Interestingly, however, they also did very little to collect in-

formation about the status quo pricing policy. They refused to incorporate the data

on differential pricing across market regions and product lines compiled by the sales

team into their pricing simulations. As the sales team lacked the information and

expertise to develop pricing simulations on their own, the political conflict erupted

in the absence of any hard data.

In the end, firm executives agreed to the recommendations of the marketing team

for the newer, high volume products, but made a specific exception for older products

in order to make the price changes revenue neutral. The director of sales acknowl-

edged that his team had lost in this political conflict: “You’ve got to move on. You


can’t keep fighting it. It is pointless. You have to move on and make it work”

(Zbaracki and Bergen, 2010, p. 966).

Strategic Analysis

I illustrate how it could be in the best interests of the marketing team to refuse

to collect information on the existing pricing scheme.

Define the outcome space by the relative importance of market share and profit

margins, where high market share and low profit margins map to low values in R.

Define preferences over policies by their expected outcomes, where low values indicate

lower prices and high values indicate more discretion.

Let the marketing team be P , which is consistent with marketing’s control over

the price lists and their ability to collect and analyze market data. Let sales be V .

Let p =

54 , indicating that the status quo policy (following the firm’s investment)

was more heavily weighted towards discretion than what was seen as optimal from the

perspectives of both marketing and sales. However, the status quo is substantially

less desirable from the perspective of the marketing team.

Let ↵ =

14 and � = 1, indicating a moderate level of uncertainty about the

mapping from the status quo policy to the status quo outcome. Because the status

quo policy was implemented prior to a significant organizational change, learning

p(p) provides no additional information about the mapping between a proposed

alternative and the associated policy outcome. Note that the sales team, V , will

accept a proposal from the marketing team, P , to learn p(p) (LV ⇤= A).

The strategic analysis involves calculating the expected utility for P if p(p) is

learned and if p(p) is not learned:

If the marketing team does not propose to learn (LP= ¬L), then the best

alternative proposal is 34 (PP⇤

¬L is a =

34), which is the nearest proposal to their ideal


outcome, 0, that makes the sales team just as well off. Because the range of outcomes

associated with the policy are distributed uniformly from 12 to 1, the expected utility

for the marketing team is �34 (Eup(LP

= ¬L) = �34). Note that for ease of notation,

I will denote Eup(LP= ¬L) and Eup(LP

= L) as Eup(¬L) and Eup(L).

The game outcome if the marketing team does not propose to learn (LP= ¬L)

is illustrated in Figure 2.1.

If the marketing team proposes to learn (LP= L), the sales team agrees (LV

=

L), and the organization learns that the status quo policy outcome, p(p), is between

1 and 98 , there is no alternative that the marketing team, P , can propose that makes

the sales team, V , better off. As a result, the marketing team does not propose an

alternative (PP⇤L = p). The probability that the status quo policy outcome lies in

this range is 14 ((Pr( p(p) <

98) =

14). The expected utility for the marketing team,

P , given that p(p) lies in this range is �1716 .

The game outcome if the marketing team proposes to learn (LP= L) and p(p) =

1716 is illustrated in Figure 2.2.


=

L), and the organization learns that the status quo policy outcome, p(p), is between98 and 5

4 , the marketing team will propose an alternative with an expected value

between 34 and 1 that makes the sales team, V , indifferent between the proposed

alternative and the status quo outcome. The probability that p(p) lies in this

interval is 14 (Pr(98 p(p) <

54) =

14). The expected utility for the marketing team,

P , given that p(p) lies in this interval is strictly less than than �34 .




=

L), and the organization learns that the status quo policy outcome, p(p), is between


Figure 2.1: p =

54 . ↵ =

14 . � = 1. LP

= ¬L.

(a) The status quo policy, p = 54 .

0 0.5 1 1.5 2Lower prices/Less discretion

Higher prices/More discretion

Marketing (P )ideal outcome

Sales (V )ideal outcome

Status quo policy, p =

54

Distribution of status quo outcomes, p+Xp ⇠ U [1, 32 ]

(b) The best alternative policy that the marketing team can propose that the sales team will acceptis a = 3

4 .

Alternative propoal, a

NA

Euv(a)

-0.5 0 0.5 0.75 1 1.5 2 2.5

-1.5

-1-0

.50

P ideal outcome V ideal outcome

V ’s utilitycurve

Euv(p) = �14If a is in the

darkened range:V will accept

a =

34

(c) The expected utility for P , Eup = � 34 .


Proposed alternative, a =

34

Distribution of alternative outcomes, a+X¬Ia ⇠ U [

12 , 1]

-0.5 0.0 0.5 1.0 1.5 2.0 2.5


Figure 2.2: p =

54 . ↵ =

14 . � = 1. LP

= L. LV= A. p(p) =

1716 .

(a) Assume that the marketing team proposes to learn, the sales team agrees, and the organizationlearns that the status quo outcome, p(p) =

1716 .





Revealed status quo outcome, p(p) =1716


(b) There are no alternatives that will make both marketing and sales better off. Therefore, Pdoes not propose an alternative policy. The outcome is p(p) = 17

16 . The expected utility for P ,Eup( p(p)) = � 17

16 .

Alternative proposal, a

NA

Euv(a)

-0.5 0 0.5 1 1.5 2 2.5

-1.5

-1-0

.50


V ’s UtilityCurve

Euv( p(p)) = � 116

There exist noalternative proposals that

V will accept


Figure 2.3: p =

54 . ↵ =

14 . � = 1. LP

= L. LV= A. p(p) =

1916 .


1916 .







(b) The best alternative policy that the marketing team can propose that the sales team will acceptis a = 0.82.


NA

Euv(a)

-0.5 0 0.5 0.82 1 1.5 2 2.5

-1.5

-1-0

.50


V ’s UtilityCurve

EUv( p(p)) = � 316

If a is inthe darkened range:

V will accept

a = 0.82

(c) The expected utility for P , Eup = �0.82.


Proposed alternative, a = 0.82

Distribution of alternative outcomes, a+XIa ⇠ U [0.57, 1.07]

-0.5 0.0 0.5 1.0 1.5 2.0 2.5


54 and 3

2 , then the marketing team will propose an alternative policy with an expected

value between 12 and 3

4 such that the sales team is indifferent between the proposed

alternative and the status quo outcome. The probability that p(p) lies in this

interval is 12 (Pr( p(p) � 5

4) =

12). The expected utility for the marketing team

given that p(p) lies in this interval is �58 .



If the marketing team proposes to learn (LP= L), the marketing team has a 25

percent chance of expected utility of �1716 , a 25 percent chance of expected utility less

than �34 , and a 50 percent chance of expected utility of �5

8 . The expected utility for

the marketing team if they propose to learn, therefore, is lower than the expected

utility if they do not (Eup(L) < Eup(¬L)).

Eup(L) = Pr( p(p) <9

8)(Eup| p(p) <

9

8) +

Pr(9

8 p(p) <

5

4)(Eup|

9

8 p(p) <

5

4) + Pr( p(p) �

5

4)(Eup| p(p) �

5

4)

Eup(L) <1

4·�17

16+

1

4·�3

4+

1

2·�5

8

Eup(L) < �3

4

Eup(¬L) = �3

4

Eup(¬L) > Eup(L)

Thus, the marketing team chooses not to propose to learn (LP⇤= ¬L).

Conclusion

Because learning about the uncertain status quo policy increases the likelihood

that the sales team would refuse any potential proposed alternative, the marketing

team prefers to make a proposal without learning the mapping of the status quo


Figure 2.4: p =

54 . ↵ =

14 . � = 1. LP

= L. LV= A. p(p) =

118 .


118 .







(b) The best alternative policy that the marketing team can propose that the sales team will acceptis a = 5

8 .


NA

EUv(a)

-0.5 0 0.5 1 1.5 2 2.5

-1.5

-1-0

.50


V ’s UtilityCurve

EUv( p(p)) = �38

If a is inthe darkened range:

V will accept

a =

58

(c) The expected utility for P , Eup = � 58


Proposed alternative, a =

58

Distribution of alternative outcomes, a+XIa ⇠ U [

38 ,

78 ]

-0.5 0.0 0.5 1.0 1.5 2.0 2.5


policy to its associated outcome.

Example 2: U.S. Congressional Investigations Following the Financial Cri-

sis of 2008

Background

In the aftermath of the 2008 financial crisis and, in particular, following the emer-

gency passage of legislation approving the Troubled Asset Relief Program, Democrat

and Republican legislators in the U.S. Congress agreed that the American financial

regulatory regime was badly broken. Traditionally, Democrats in Congress supported

a more stable financial system at the cost of sacrificed economic growth. Republi-

cans, in contrast, were willing to risk greater financial instability for faster economic

growth. These disagreements over economic values tended to play out in negotia-

tions over financial regulatory policy. Democrats supported more stringent banking

regulation. Republicans tended to favor deregulation.

After the financial crisis, however, Democrats and Republicans agreed that the

needle had swung too far towards deregulation. The debate was over how much

additional regulation was necessary. Congressional Democrats supported replacing

the existing regulatory apparatus with a regime similar to the one in place following

the Great Depression. Congressional Republicans supported more targeted reforms.

It is common for the U.S. legislature to establish investigatory commissions when

national emergencies occur that warrant policy intervention. Investigatory commis-

sions of this type were created to great fanfare following the Great Depression, the

Challenger Disaster, and the terrorist attack on the World Trade Center. The in-

vestigative reports published by these commissions were influential in directing the

subsequent policy debate. Even today, the Pecora Commission, the Rogers Commis-

sion, and the 9/11 Commission remain a part of public discourse.


However, Democrats in the Congress were unenthusiastic about creating a com-

mission following the financial crisis. Chairman of the House Financial Services Com-

mittee Barney Frank called proposals to form an investigatory commission “a silly

idea whose time has come” (Kaiser, 2013). Unable to keep a commission from being

created, Frank instead convinced congressional leadership in the House of Represen-

tatives to stipulate that the commission’s findings be published after the financial

reform effort was complete.

Sweeping financial reforms were signed by the President on July 21, 2010. Major

changes included the creation of new regulatory agencies, significant expansion of

existing regulatory authority, and the re-introduction of restrictions on investment

behavior that had been slowly eroded through deregulation.

The results of the commission’s investigation were published six months later.

Strategic Analysis

I demonstrate one reason why Congressional Democrats were opposed to the

creation of an investigatory commission.

Define the outcome space by the relative importance of economic stability and

economic growth, where high stability and low growth map to low values in R. Define

preferences over policies by their expected outcomes, where low values indicate more

market regulation and high values indicate less market regulation.

Let Democrats be P , which is consistent with the partisan control of the Congress

and the Executive when the commission was proposed. Let Republicans be V .

Let p = 1.9, indicating that the status quo policy (following the financial crisis)

was more heavily weighted towards instability than what was seen as optimal from

the perspectives of both Democrats and Republicans.

Let ↵ =

12 and � =

54 . There is high uncertainty about the mapping between


the status quo policy and status quo outcome, although there is agreement that all

possible status quo outcomes are associated with too little regulation. Learning about

the mapping between the status quo policy and policy outcome does provide some

additional information about the distribution of outcomes resulting from proposed

alternatives. This differs from the previous example, in which learning about the

mapping from status quo policy to status quo outcome did not decrease the range of

possible outcomes resulting from proposed alternatives.

LV ⇤= A, indicating that V will accept a proposal to learn p(p).

If LP= ¬L, then PP ⇤

¬L is a = 0.1, which is the nearest point to 0 which makes V

just as well off. Eup(¬L) = �0.26.

If LP= L and p(p) 2 [1

25 , 1

35), P will propose a = 2 � p(p). The probability

that p(p) 2 [1

25 , 1

35) is 1

5 . The expected utility for P given p(p) 2 [1

25 , 1

35) is �1

2 .


35 , 2), P will propose a 2 (0, 25) . The probability that

p(p) 2 [1

35 , 2) is 2


35 , 2) is strictly less

than �15 .

If LP= L and p(p) 2 [2, 22

5 ]. P will propose a = 0. The probability that

p(p) 2 [2, 225 ] is 2

5 . The expected utility for P given p(p) 2 [2, 225 ] is �1

5 .

As a result, if LP= L, P has a 20 percent chance of expected utility of �1

2 , a 40

percent chance of expected utility of �15 , and a 40 percent chance of expected utility

of strictly less than �15 , which is less than Eup(¬L) = �0.26.

Thus, LP⇤= ¬L.

Conclusion

Congressional Democrats would do fairly well without learning, because the status

quo policy is thought to be so extreme. If they investigate, Congressional Republicans

could learn that the current regulatory system was stronger than had previously been


thought, in which case Democrats would be unable to get as desirable a new policy

accepted by Republicans. Although, if they investigate, Congressional Democrats

could learn that the current regime was even less regulated than legislators previ-

ously thought, the political value of this information is bounded by the Democrats’

ideal outcome.

Example 3: Drug Testing, Major League Baseball, and the Players’ Union

Background

The late 1990s were a boon to Major League Baseball. The assaults on Roger

Maris’ storied home run record by sluggers Mark McGwire, Ken Griffey, Jr., Sammy

Sosa, and Barry Bonds brought fans back to baseball after a nasty labor dispute

canceled the 1994 World Series. However, the power surge also stimulated suspicion

among baseball fans about the possibility of widespread performance-enhancing drug

use. Following a series of investigations into specific players and even some high

profile admissions of guilt, the Office of the Commissioner came to believe that

steroid use was hurting the long-term popularity of the sport.

The Major League Baseball Players Association agreed that steroid use was a

problem, but was deeply uncomfortable with subjecting its members to the steroid

testing regime proposed by the owners. The union was concerned that the owners

would use steroid tests as a tool to discriminate against disfavored players, that

violators could be open to federal prosecution, and that the testing regime could be

a first step towards more draconian oversight over the private lives of players. The

union was willing to make steroid testing a part of the collective bargaining process,

but expected concessions from ownership in order to accept widespread randomized

testing and stiff penalties for violators.

In 2006, Major League Baseball commissioned a report on the prevalence of


steroid use in baseball, which was conducted by former U.S. Senator George Mitchell.

The report, which was completed in the fall of 2007, found conclusive evidence of

performance-enhancing drug use by 89 current and former players (Mitchell, 2007).

Included in the report were also interviews with current and former players and

coaches that insinuated that the problem of performance-enhancing drugs was far

more widespread than had been expected.

The resulting change in player perceptions about the prevalence of performance-

enhancing drugs in baseball combined with the bad public optics of appearing to be

in favor of illegal drugs have led the union to take a more conciliatory stance towards

increased testing. In the 2011 negotiations between the commissioner’s office and the

union, the players accepted significantly more stringent testing standards, including

agreeing to the collection of blood samples to test for human growth hormone. In

2013, the players’ union also agreed to increased penalties for players caught using

performance-enhancing drugs, random blood tests, and substantially increased su-

pervision for past offenders.

Strategic Analysis

I illustrate how the decision to commission a report on steroids in baseball helped

the commissioner’s office in negotiations with the Major League Baseball Players

Association.

Define the outcome space by the relative importance of clean players and per-

sonal privacy, where clean players and low privacy map to low values in R. Define

preferences over policies by their expected outcomes, where low values indicate more

drug testing and high values indicate less drug testing.

Let the Office of the Commissioner be P , which is consistent with power that

owners have over the players and the considerably greater resources at the disposal


of the commissioner than the head of the players’ association. Let the union be V .

Let p = 1, indicating that the status quo policy is optimal from the perspective

of the union, but undesirable from the perspective of the owners.

Let ↵ =

310 and � = 1, indicating a moderate level of uncertainty about the

prevalence of the use of steroids in the current regime. Also, as there is no formal

drug testing policy in place, learning the current rate of steroid usage does not reduce

the uncertainty surrounding an alternative performance-enhancing drug policy.

LV ⇤= A, indicating that V will accept a proposal to learn p(p).

If LP= ¬L, then PP⇤

¬L = p, because there are no alternative policies than can

make the union better off than the status quo. Eup(¬L) = �1.


320 , 1

310 ], P will propose a 2 [

710 , 1]. The probability that

p(p) 2 [1

320 , 1

310 ] is 1


320 , 1

310 ] is strictly

greater than �1.

If LP= L and p 2 [

710 , 1

320), P will not propose an alternative. The probability

that p(p) 2 [

710 , 1

320) is 3

4 . The expected utility for P given p(p) 2 [

710 , 1

320) is �1.

As a result, if LP= L, P has a 75 percent chance of expected utility of �1, but

a 25 percent chance of expected utility greater than �1.

Thus, LP⇤= L.

Conclusion

Absent the report, players would believe that the modest amount of performance-

enhancing drugs in baseball outstripped the necessary sacrifice of personal privacy

associated with more stringent controls. If the report found there was less drug usage

than expected, then the Commissioner’s office would be no worse off. In contrast, if

the report found greater drug usage than expected, then the players would be more

willing to agree to a tighter drug testing regime.


The Learning Decision

These three examples illustrate how conflicting preferences can complicate incen-

tives for organizational learning. I now generalize from the examples and explore

the organization’s decision whether or not to learn for a range of parameter values.

The analysis highlights a few important insights about equilibrium behavior in the

model. First, the strength of the incentives both to learn p(p) and not to learn

p(p) grow with ↵. Second, as � increases, the incentives to learn for both P and

V increase as well. However, even for high values of �, there remain values of p for

which P does not propose to learn p(p) and for which V vetoes a proposal to learn

p(p). Finally, for all values of p such that the organization does not learn p(p), p

is either less than 0 or greater than 1. Therefore, the organization only chooses not

to learn p(p) when the existing policy is undesirable for both P and V .

Figures 2.5, 2.6, and 2.7 illustrate the expected utility functions for the proposer

and the veto conditional on p and whether p(p) is learned for ↵ 2 { 110 ,

14 ,

12} and

� = 1 (on the left) and the difference between the utility functions if p(p) is learned

and if p(p) is not learned (on the right). The top left panels demonstrate the

proposer’s expected utility if p(p) is learned and the proposer’s expected utility

if p(p) is not learned. The bottom left panels demonstrate the veto’s expected

utility if p(p) is learned and the veto’s expected utility if p(p) is not learned. The

right panels illustrates the differences between the two expected utilities for P and

V . Higher values on the right panel indicate a larger benefit to learning relative to

ignorance.

Figure 2.5 compares Eup and Euv when p(p) is learned and when p(p) is not

learned for ↵ =

110 and � = 1. Eup,v(I) represents the expected utility for P or V if

p(p) is learned. Eup,v(¬I) represents the expected utility for P or V if p(p) is not


Figure 2.5: ↵ =

110 . � = 1.

Eup

-1 0 1 2 3

-1.0

-0.6

-0.2

Eup(I)Eup(¬I)

Euv

Status Quo Policy, p

-1 0 1 2 3

-1.0

-0.6

-0.2 Euv(I)

Euv(¬I)

P idealoutcome

V idealoutcome


-1 0 1 2 3

-0.1

0.0

0.1

0.2

Exp

ecte

dU

tility

Gai

nfr

omLe

arni

ng

Eup(I) - Eup(¬I)Euv(I) - Euv(¬I)

P idealoutcome

V idealoutcome


Comparing Eup,v(I) to Eup,v(¬I) Eup,v(I)� Eup,v(¬I)

learned.

Figure 2.5 indicates that when the level of uncertainty in the mapping from

policies to outcomes is small, the strategic choice to learn p(p) is less important.

However, even with low uncertainty, for both P and V the incentives to keep the

organization from learning p(p) are largest when p is believed to be undesirable.

Figures 2.6 and 2.7 demonstrate how these results change as the level of uncer-

tainty increases. Figure 2.6 compares Eup,v(I) and Eup,v(¬I) when ↵ =

14 and � = 1.

Figure 2.7 compares Eup,v(I) and Eup,v(¬I) when ↵ =

12 and � = 1.

There are three major differences between Figures 2.5, 2.6, and 2.7. As uncer-

tainty increases, the size of the expected utility difference between learning and not

learning increases for both P and V . Also, both the range of values for which learn-

ing dominates not learning and the range of values for which not learning dominates


Figure 2.6: ↵ =

14 . � = 1.

Eup

-1 0 1 2 3

-1.0

-0.6

-0.2

Eup(I)Eup(¬I)

Euv


-1 0 1 2 3

-1.0

-0.6

-0.2 Euv(I)

Euv(¬I)

P idealoutcome

V idealoutcome


-1 0 1 2 3

-0.1

0.0

0.1

0.2

Exp

ecte

dU

tility

Gai

nfr

omLe

arni

ng


P idealoutcome

V idealoutcome



learning grow considerably for both P and V . Finally, for both P and V , the strength

of the incentive to keep the organization from learning p(p) conditional on preferring

that p(p) not be learned grows with ↵.

There are many organizational contexts in which learning about the status quo

provides information that is relevant to potential policy changes, as well. Figures 2.8

and 2.9 demonstrate how the equilibrium outcomes change when organizations also

learn about the distribution of outcomes resulting from an alternative policy when

they learn about the status quo, i.e. � > 1. In the examples, ↵ =

12 and � 2 {2, 3}.

Figure 2.8 compares Eup,v(I) and Eup,v(¬I) when a great deal of uncertainty is

reduced as a result of learning: ↵ = 0.5 and � = 3. Figure 2.9 compares Eup,v(I)

and Eup,v(¬I) when a moderate amount of uncertainty is reduced as a result of

learning: ↵ = 0.5 and � = 2.


Figure 2.7: ↵ =

12 . � = 1.

Eup

-1 0 1 2 3

-1.0

-0.6

-0.2

Eup(I)Eup(¬I)

Euv


-1 0 1 2 3

-1.0

-0.6

-0.2 Euv(I)

Euv(¬I)

P idealoutcome

V idealoutcome


-1 0 1 2 3

-0.1

0.0

0.1

0.2

Exp

ecte

dU

tility

Gai

nfr

omLe

arni

ng Eup(I) - Eup(¬I)Euv(I) - Euv(¬I)

P idealoutcome

V idealoutcome



Importantly, even when significant uncertainty is reduced as a result of organi-

zational learning, there remains a range of status quo values for which not learning

dominates learning for both P and V .

Figure 2.10 summarizes the equilibrium outcomes of the learning subgame con-

ditional on p for ↵ 2 { 110 ,

14 ,

12} and � 2 [1, 5]. The areas of the graphs colored black

indicate the sets parameter values for which P will not propose to learn p(p), or

LP ⇤= ¬L. The areas of the graph colored grey indicated the sets of parameter

values for which V will not accept a proposal to learn p(p), or LV ⇤= ¬A.

An important insight from the model that is demonstrated in Figure 2.10 is that

the parameter values for which the organization will not learn are likely to be as-

sociated with Pareto inefficient outcomes. These are precisely the conditions under

which there are likely to exist mutually beneficial alternative policies. Furthermore,


Figure 2.8: ↵ =

12 . � = 3.

Eup

-1 0 1 2 3

-1.0

-0.6

-0.2

Eup(I)Eup(¬I)

Euv


-1 0 1 2 3

-1.0

-0.6

-0.2 Euv(I)

Euv(¬I)

P idealoutcome

V idealoutcome


-1 0 1 2 3

-0.1

0.0

0.1

0.2

Exp

ecte

dU

tility

Gai

nfr

omLe

arni

ng


P idealoutcome

V idealoutcome



the regions of the parameter space for which the organization does not learn grow

with the uncertainty in the mapping between status quo policies and their associated

outcomes. As a result, the model predicts that internal politics make organizations

less likely to learn about existing policies prior to making organizational changes,

especially when there is a great deal of uncertainty surrounding existing policies and

even when learning reduces the uncertainty associated with a proposed alternative.

Conversely, when the organization is unlikely to implement a strategic change, the

organization is more likely to learn.

Discussion and Conclusion

The importance of politics and preference conflict in organizations is as impor-

tant today as when it was first emphasized by organizational scholars at the Carnegie


Figure 2.9: ↵ =

12 . � = 2.

Eup

-1 0 1 2 3

-1.0

-0.6

-0.2

Eup(I)Eup(¬I)

Euv


-1 0 1 2 3

-1.0

-0.6

-0.2 Euv(I)

Euv(¬I)

P idealoutcome

V idealoutcome


-1 0 1 2 3

-0.1

0.0

0.1

0.2

Exp

ecte

dU

tility

Gai

nfr

omLe

arni

ng


P idealoutcome

V idealoutcome



School. However, the implications of internal politics on organizational learning have

gone underexplored in the organizational strategy literature. My paper develops a

model that illustrates how fairly simple political negotiations between two organiza-

tion members over organizational goals create complex incentives related to learning.

Learning is a strategic decision. Under certain conditions, the benefits from learning

are outweighed by the political constraints imposed by more information. As a result,

failures of organizations to learn do not need to be the result of irrationality or sub-

optimal design. Entirely rational organization members facing political opposition

will sometimes prefer ambiguity to certainty.

The idea that organization members can benefit from purposive ignorance is

an old one (e.g. see Feldman and March, 1981; Pfeffer, 1981; Vancil, 1979) but

efforts to formalize the conditions under which organization members are expected


Figure 2.10: Learning Subgame Equilibrium Behavior. ↵ 2 { 110 ,

14 ,

12}. � 2 [1, 5].

-1 0 1 2 3

12

34

5

↵ =

110

Status Quo Policy, p-1 0 1 2 3

12

34

5

↵ =

14

Status Quo Policy, p-1 0 1 2 3

12

34

5

↵ =

12


Leve

lofU

ncer

tain

tyR

educ

tion,�

LP ⇤= ¬L

LV ⇤= ¬A

to constrain organizations from gathering new information are fairly novel. This

paper builds on recent work on Bayesian persuasion by Brocas and Carillo (2007) and

Kamenica and Gentzkow (2011), which examine behavior in games where one actor

can strategically experiment in order to attempt to convince an alter about the state

of the world. The model is also related to research in organizational economics that

demonstrates the challenges of information transmission in principal-agent models

characterized by actors with partially aligned preferences (Austen-Smith and Banks,

2000; Crawford and Sobel, 1982; Gibbons et al., 2012; Milgrom and Roberts, 1988).

However, to my knowledge, this is the first paper to formally show how proposer-

veto political dynamics in organizations create incentives for organization members

to forego learning.

The model has normative implications for managers as well. Managers often face


steady institutional pressure to justify new proposals with objective data and analysis

(Feldman and March, 1981). Naïve managers see these requests for information as

a part of an organizational effort to achieve commonly held goals and, as a result,

always acquiesce. Strategic managers realize that requests for more information by

organization members are a part of a political process. As a result, strategic managers

may refuse to collect information. In fact, strategic managers will sometimes even

be willing to pay a personal cost to impede the capacity of political rivals to learn.

Lastly, the model helps organizational researchers better understand two frus-

trating observations about organizational learning and organizational change. One

reason why organizations frequently fail to learn about existing strategies or policies

prior to implementing organizational changes is because there are groups within or-

ganizations who fear that the results of such an investigation will disadvantage them

politically. These fears, however, do not exist when existing strategies are Pareto

efficient. Thus, organization members are likely to be willing to learn about existing

policies when it is unlikely that there is sufficient support in favor of an alternative to

replace the existing strategy. The ironic conclusion, therefore, is that organizations

will not learn when a policy change is likely to occur and will learn when a policy

change is unlikely to be enacted.

Chapter 3

Organizations Experimenting With

Managers Experimenting With

Strategies

Introduction

A hallmark of successful organizations is their ability to continually produce

fundamental innovations that fuel growth in the long term (Anderson and Tush-

man, 1990; O’Reilly and Tushman, 2013; Tushman and Anderson, 1986; Tushman

and O’Reilly, 1996; Utterback, 1994). However, despite the payoffs for organiza-

tions that sustain the type of wide-ranging exploration required to promote durable

competitive advantage, organizations consistently find it challenging to induce their

members to take big risks with uncertain payoffs. Due to inertial organizational rou-

tines, difficulties with cross-unit collaboration, political conflict, and managerial risk

aversion, innovation in organizations is often overly incremental (Cyert and March,

48

CHAPTER 3. ORGANIZATIONAL EXPERIMENTING WITH MANAGERS 49

1963; DiMaggio and Powell, 1983; Gibbons and Roberts, 2012; Hannan and Free-

man, 1984; Levitt and March, 1988; March, 1991; March and Shapira, 1987; Shapira,

1995). Managers seek smaller, exploitative improvements to existing products, ser-

vices, and technologies instead of the large, bold innovations that are required to

maintain long-term performance.

One way that organizations try to induce managers to overcome inertia, risk

aversion, and incrementalism is by encouraging managers to experiment offline. As

opposed to online search, in which organizations fully implement strategic changes

and then observe whether or not the changes worked, offline experimentation allows

organizations to observe the results of small-scale tests of alternative strategies prior

to committing to a new initiative. Clinical trials by drug companies, A/B testing

by internet search engines, and weapons prototypes in the defense industry are all

examples of offline experiments. By reducing the stakes associated with a failed

experiment, organizations attempt to induce members to take bigger risks. The

knowledge gleaned from offline experiments is then used to inform the strategies

that organizations choose to take online (Csaszar and Levinthal, 2016; Gavetti and

Levinthal, 2000; Kelley and Littman, 2001; Lippman and McCall, 1976; Pfeffer and

Sutton, 2006b).

However, experimenting offline fails to resolve a classic agency problem associated

with risky experimentation in organizations (Aghion and Jackson, 2014; Gibbons and

Roberts, 2012). Offline experiments not only offer managers the opportunity to test

alternative strategies, but also offer organizations the opportunity to evaluate the

competence of managers at their assigned task. It can be difficult for a firm to

distinguish between a good experimenter who had bad luck with an experiment and

a poorly matched or incompetent experimenter. For example, a graduate student in

the biological sciences whose dissertation research finds null results will often have a


difficult time finding a post-doctoral research position, because prospective employers

cannot tell whether the student’s experimental work utilized proper research methods

or whether the experiments the student conducted were fundamentally flawed. As

a result, when a strategic manager faces with the choice between (1) designing an

offline experiment with a goal in mind that the organization finds desirable and

believes to be attainable and (2) designing an experiment with a targeted goal that

is undesirably incremental or impractically blue sky, the manager may prefer the

second. A failure to achieve an attainable goal reveals the manager as possibly

incompetent or poorly matched. Successfully reaching an incremental goal or failing

to achieve an unattainable one reveals less about a manager’s competence or fit.

Managers who seek to avoid revealing information about their quality do not nec-

essarily undertake experiments that are overly incremental from the perspective of

the organization. In fact, some of the same factors that lead managers to be overly

incremental when searching online can lead managers to be overzealous in their offline

search behavior. When pushed, managers will select goals that go far beyond what

can be successfully integrated by an organization in the hope of demonstrating their

capability to be innovative to superiors. Rather than coloring outside the lines, these

managers will color “on an entirely different page” (Kelley and Littman, 2001). IDEO

general manager Tom Kelley describes the “joyful failure” of the Enorme phone, a

product co-designed with an iconic Italian designer that sought to bring Italian de-

sign aesthetics together with Silicon Valley telephone engineering. The effort was

undoubtedly innovative: an example of the phone remains in the permanent collec-

tion of the Museum of Modern Art in New York. It was also an abject failure as a

consumer product. At $150 apiece at a time when customers could easily get a phone

for a few dollars, there was no consumer demand for the “playful cubist” telephone

(Kelley and Littman, 2001). These blue sky efforts are no better for organizations


than unambitious, incremental innovation efforts. Further, given the extensive fo-

cus by researchers on how organizations can overcome inefficient incrementalism and

tendencies toward exploitation in the context of online search, concerns about over-

exploration in offline search pose a novel problem for organizational researchers. The

challenge for organizational leaders is not merely to get managers to take bigger risks.

Instead, the problem that leaders face is to get managers to take risks that are at

once ambitious enough to support long-term goals but not so exploratory as to be

unusable or impractical.

My paper develops a formal model of the behavioral dynamics of offline experi-

mentation in organizations in which managers are primarily concerned about their

own careers and secondarily concerned with organizational performance. It examines

common pitfalls that organizations face when motivating their managers to exper-

iment offline. And, it evaluates potential mechanisms organizations can use to get

managers to undertake more informative offline experiments. Specifically, the paper

envisions a manager tasked with experimenting offline on their organization’s behalf

prior to the organization implementing a novel strategy online. The organization

seeks to maximize the performance of the implemented strategy. The manager fears

being replaced by another manager in the organization.

I utilize a modeling approach for search and experimentation behavior derived

from recent work in political economics. The model demonstrates how career con-

cerns inflamed by the fear of failed experiments limit the amount that organizations

learn from offline experimentation. Importantly, a manager’s fear of being replaced

can lead the manager to both strategically under-experiment and strategically over-

experiment, as both strategies decrease the likelihood that the manager will be re-

vealed to be a poor fit for the project to which they have been assigned.

The model also evaluates mechanisms that an organization can use to maximize


the value of the information gleaned from an offline experiment in the face of this

principal-agent dynamic. On the one hand, organizations can define the range of pos-

sible experimental outcomes that are acceptable ex ante and replace those managers

whose experiments fail to achieve the predetermined goals. On the other, organi-

zations can guarantee that they will not replace an experimenter, notwithstanding

the experimental outcome. I demonstrate that the former—an outcome-based re-

placement rule—dominates the latter—guaranteed tenure—in low uncertainty task

environments and that the latter dominates the former in high uncertainty task envi-

ronments. The freedom to experiment without career risk (combined with stringent

selection processes and socialization in order to prevent shirking) is a common char-

acteristic of strategic human resources policies in a number of famously exploratory

firms that inhabit highly uncertain innovation environments, including Bell Labs,

IDEO, Google, and W.L. Gore (Gertner, 2012; Hamel and Spence, 2010; Kelley and

Littman, 2001; Lashinsky, 2006).

Background and Related Work

Experimentation, Organizational Learning, and Career Concerns

Organizational research on innovation highlights a classic tension between the

need for organizations to undertake risky search and the difficulties involved in pro-

moting risk taking among its members. Sociological approaches emphasize that

strong inertial forces within organizations and dominating institutional pressures to

conform to existing norms generate status quo bias (DiMaggio and Powell, 1983;

Hannan and Freeman, 1984). Psychological and economic treatments focus on the

ways that organizational institutions promote managerial risk aversion (Gibbons

and Roberts, 2012; Holmstrom and Ricart i Costa, 1986; March and Shapira, 1987;

Shapira, 1995). These varied mechanisms help to explain why organizations tend to


be better at exploitative, incremental search than exploratory, uncertain search.

One recipe for promoting greater risk tolerance in search is an organizational com-

mitment to offline experimentation. Offline experiments allow organization members

to “look before they leap” and to develop a more detailed understanding of the forces

that shape the innovation landscape. Prototyping, in particular, has been at the

core of the design thinking approach to innovation and change. By trying new ideas,

seeing what works, and rapidly iterating, organizations can reduce the complexity

of a problem and uncover novel solutions (Kelley and Littman, 2001). For example,

a Bay Area organization charged with supporting families with children with devel-

opmental disabilities successfully used prototyping to improve its service delivery by

streamlining the process through which new clients were evaluated. The organiza-

tion experimented with a Winnebago that took the various evaluators to applicants’

neighborhoods. A process that usually took two to three months was reduced by

ten weeks, on average. Although outfitting the organization with a fleet of RVs was

a non-starter, the results of the experiment convinced the organization to create a

dedicated assessment team in order to reduce the complexity of the initial assessment

process and to make social workers more mobile so that they can meet with families

closer to their homes (Sutton and Hoyt, 2016).

An emphasis on offline experimentation is also central to the literature on evidence-

based management. By strategically selecting alternatives, executing small-scale

market tests, and carefully measuring outcomes, organizations can overcome decision-

making biases and political opposition in order to change existing practices for the

better (Beshears and Gino, 2014; Pfeffer and Sutton, 2006a,b). Pfeffer and Sutton

(2006b) describe how Harrah’s Las Vegas Hotel & Casino overturned industry-wide

conventional wisdom about the best way to get customers to gamble more through

the use of low stakes experiments. In one experiment, Harrah’s offered one group


of customers a standard promotional package that included $30 in chips and $95

in meal and room vouchers and another group an experimental package that only

included $60 in chips. Contrary to the widely held belief that gamblers are enticed

to the tables by fancy meals and entertainment, the second package induced more

gambling revenue at a lower cost.

The economics literature on career concerns, however, paints a less rosy picture of

the capacity for organizations to motivate their employees to undertake exploratory

experiments, even if the outcomes of the experiments have low stakes for the organi-

zation. Managers who are unsure about their competence with a task would prefer

not to reveal information about their quality. Fear of being revealed as incapable

leads managers to forego risk taking, generating inertia, herding behavior, and pan-

dering (Aghion and Jackson, 2014; Canes-Wrone et al., 2001; Gibbons and Roberts,

2012). Canes-Wrone et al. (2001) describe how the re-election concerns of California

Governor and future Chief Justice of the United States Supreme Court Earl Warren

led him to pander to conservative fears about the threat of communism. Although

he initially objected indignantly to calls that UC-Berkeley faculty take an oath dis-

claiming membership in the Communist party, he later reversed course by making

such an oath a condition of all public employment. Warren was fearful that the

revelation of communist sympathies among state employees prior to his upcoming

election against staunch anti-Communist James Roosevelt could irreparably damage

his chances of winning. Requiring state employees to take an oath made him appear

strong against the communist threat.

A key insight of my paper is that career concerns can also lead managers to

be overly exploratory in their offline search behavior. Rather than implementing

an experimental strategy that a manager believes will work well, a manager may

instead experiment with a strategy that they know will likely go too far. Lindbergh


and Miller (2016), two statistically-inclined baseball bloggers who helped to manage

an Independent League baseball team for a summer, recount the challenges of getting

their Sonoma Stompers to buy into the non-traditional ways of arranging fielders on a

baseball diamond that are supported by baseball statisticians.1 Lindbergh and Miller

experimented with these untraditional ideas during a practice game, acknowledging

that “if it works, we’re confident they’ll embrace [our vision]. If it doesn’t, we’re not

confident they’ll have the courage to keep trying. And if it’s somewhere in between,

it probably counts as a loss for us.” Therefore, Lindbergh and Miller required that

the Stompers’ infield position itself unconventionally for every single at bat during

the game. “The point isn’t necessarily for it to work,” they admitted. “The point is

to see where it fails.” By requiring that their team be overly experimental, the team

evaluated Lindbergh and Miller based on whether the team learned something from

the experience, rather than on the team’s performance during the practice game.

Strategic over-experimentation, therefore, helped Lindbergh and Miller solidify their

influence over the team and its coaches.

The paper also examines mechanisms that organizations can use to better align

managerial behavior with the organization’s goals. I show that organizations that

seek to induce effective offline experimentation by replacing managers that fail to

improve on the status quo, a “retrospective satisficing” heuristic, tend to induce

overly incremental experimentation behavior. This is consistent with prior research

that finds that organizations using prior performance as a benchmark tend to under-

take incremental, exploitative search (Cyert and March, 1963; March, 1991; Simon,

1955). I also explore two potential policies that the organization can implement in1They sought to go even further than even “the shift” in which three infielders are moved to one

side of the field, currently en vogue among many Major League teams. They also wanted the teamto implement five-man infields and “the wall,” a configuration with all four infielders on one side ofthe field.


order to improve managerial experimentation behavior: guaranteed tenure and an

outcome-based replacement rule. The first mechanism aligns managerial preferences

through greater freedom; the second achieves the same end through tighter control.

These dual approaches to resolving agency problems are well studied in strategic

human resource management (Adler and Borys, 1996; Baron et al., 1999; Bernstein,

2012; Blau, 1955; Dalton, 1959). Greater managerial discretion generates higher lev-

els of commitment and intrinsic motivation, places decisions with the people who

know operational details best, and promotes greater variability in organizational

decision-making. However, discretion also limits the ability of organizations to over-

see managerial behavior, which can promote shirking, difficulty with coordination,

and political infighting. My model adds to this literature by demonstrating how

tighter control better alleviates the principal-agent dynamics associated with offline

experimentation when the variability in the task environment is low. However, when

uncertainty in the mapping between novel strategies and outcomes is high, aligning

the incentives of the manager and the organization through guaranteed tenure gen-

erates more informative offline experiments and better organizational outcomes.

Models of Learning and Conflict in Organizations

A tradeoff exists in theoretical scholarship in organizational learning between the

researcher’s desire to capture the complexity inherent to the process of discovery

and the researcher’s need to make clear predictions about how changes within orga-

nizations and in the external environment impact performance. To this end, orga-

nizational theory and economics have emphasized different aspects of the challenges

associated with learning in organizations. Organizational theory has focused on the

relationship between the complexity of search and organizational performance (Davis

et al., 2007; Levinthal, 1997; Rivkin, 2000). Organizational and political economics


has been concerned with problems associated with collecting and transmitting infor-

mation within organizations characterized by individuals with differing preferences

and goals (Callander, 2011a,b; Gibbons et al., 2012). My paper seeks to build on a

growing literature that spans these two views.

Within organizational theory, there is an emphasis on understanding how or-

ganizations react to randomness in the fitness landscape. These stochastic models

explore how uncertainty associated with the implementation of novel configurations

of organizational inputs combine with various search strategies to impact long-term

performance and rates of survival. The basic models in this tradition envision the

organization as undertaking random, local search and, through a process of trial-

and-error, eventually finding collections of inputs that dominate all other local com-

binations (Kauffman, 1993; Levinthal, 1997). However, more recent work has also

integrated cognitive representations of the fitness landscape that serve as maps for

evolutionary search. These cognitive maps take into account the roles of science and

abstract knowledge in the innovation process. Organizations cannot know precisely

how a change to a strategic input will impact performance, but organizations may

have general or abstract knowledge undergirding the innovation process that makes

it more likely that a particular change will be performance enhancing. (Csaszar and

Levinthal, 2016; Fleming and Sorenson, 2004; Gavetti and Levinthal, 2000). For

example, Fleming (2002) describes how the design of Hewlett-Packard’s first ink jet

printer was inspired by recent work in high temperature physics on “micro jets,” but

also required a lengthy trial-and-error experimental approach in order to find a resis-

tor that could both generate extreme heat and withstand long term wear-and-tear.

Other theoretical models of organizational search have also sought to integrate

more detailed descriptions of the decision-making process within organizations into


the evolutionary approach. These models highlight how incentive conflict, hierarchi-

cal decision-making, and information-sharing challenges impact the capacity of orga-

nizations to learn about uncertain alternatives (Csaszar and Eggers, 2013; Csaszar

and Levinthal, 2016; Rivkin and Siggelkow, 2003; Schilling and Fang, 2014). How-

ever, bringing these intra-organizational dynamics into the evolutionary framework

requires layering an additional set of exogenous assumptions about the behavior of

individuals inside organizations on top of assumptions about organizational search

behavior, environmental selection, and the structure of the stochastic landscape.

Critics of this approach highlight the incompatibility between the strategic motiva-

tions of individuals within organizations and the non-strategic search that drives evo-

lutionary competition in the organizational population (Callander and Matouschek,

2014; Roberts et al., 2012).

Organizational and political economics tends to prioritize detailed examinations

of the decision-making process inside of organizations at the cost of realistic de-

pictions of randomness in the external environment. Economic models emphasize

the roles that preference conflict, worker incentives, and power asymmetries play in

bringing about suboptimal organizational outcomes. Agency problems, in particu-

lar, are often prominent in economic models of decision-making and learning under

uncertainty (Gibbons et al., 2012). Unlike organizational theory, which relies heavily

on evolutionary logic and simulation modeling, models of learning in economics are

solved in an analytic manner using non-cooperative game theory. Search behavior is

driven endogenously by the behavioral assumptions of individual actors in the model.

However, these rational choice decision-making models are limited in that they often

rely on sparse representations of the complexity of learning, often characterized by

a limited set of strategic alternatives or a single random parameter that, if revealed,


provides a global map of outcomes associated with potential changes to inputs (Ben-

dor and Meirowitz, 2004; Callander, 2008; Crawford and Sobel, 1982; Gilligan and

Krehbiel, 1987, 1989). Economic models in which organizations learn about manage-

rial competence use a similarly simplified structure in which managers are endowed

as either competent or incompetent and the organization seeks to reveal a manager’s

type (Aghion and Jackson, 2014; Canes-Wrone et al., 2001).

More recent work in political economics has sought to integrate richer represen-

tations of the fitness landscape while maintaining the tractability required to make

inferences about how changes to internal institutions impact performance. In partic-

ular, a recent methodological innovation in political and economic models of search

involves representing the relationship between organizational inputs and outcomes

as produced by a unidimensional stochastic process (Callander, 2008, 2011a,b). As

a shorthand, I refer to these one-dimensional stochastic maps between inputs and

outcomes as “fitness functions.” In these models, knowledge of the parameters that

generate the stochastic process create well-defined probability distributions of the

possible outcomes associated with alternative strategies. When combined with a

quadratic-loss utility function, the fitness landscape that the organization faces in

these models becomes smooth and strategically tractable in expectation, while still

rugged and multi-peaked in its realization.

The model I develop here combines the strengths of these two modeling ap-

proaches. The managers in my model face real uncertainty about how organiza-

tional strategies map to organizational outcomes, like in organization theory models.

Yet unlike in evolutionary search models, managerial expectations about unknown

strategies in my model have well-defined probability distributions. As a result, the

manager’s search behavior is endogenous to their preferences, as in economic models.


Further, the fitness function method I use here improves on prior work in organi-

zational economics by exploring the link between managerial incentives surrounding

tenure, managerial search behavior, and organizational performance. Rather than

a manager choosing between a risky and a safe option, which has been the norm

in economic models of career concerns, my model involves selecting amongst a con-

tinuum of interdependent experimental strategies. As a result, I make predictions

about how career concerns impact the character of experimentation that managers

undertake and how different experimentation strategies interact with the uncertainty

of the task environment to affect organizational performance.

The Model

The model features an organization that seeks to improve its performance by

delegating a new project to a tenure-motivated manager. The manager has the op-

portunity to prototype offline prior to the organization going online with the project.

After observing the prototype and the prototype’s performance, the organization’s

leadership has the opportunity to replace the manager. If the organization does not

replace the manager, then the same manager who experimented offline also takes the

project online. If the organization replaces the manager, a new manager is brought

in to take the project online.

Each manager in the organization is endowed with a unique fitness function, which

is represented by a realized path of Brownian motion. The parameters that generate

the stochastic Brownian path are publicly known and provide a baseline cognitive

representation for the relationship between untried strategies and outcomes. As a

manager experiments with new strategies and the associated outcomes are revealed,

the manager updates their cognitive map, in the process decreasing the variance in

the relationship between untried strategies and their outcomes. Importantly, when a


manager experiments, the information gleaned is only relevant for the manager who

undertook the experiment. A manager’s particular implementation of a potential

alternative strategy requires tacit knowledge; the same strategy implemented by two

different managers may generate very different organizational outcomes (Csaszar and

Levinthal, 2016; Kogut and Zander, 1992; Nelson and Winter, 1982; Rivkin, 2000).

From the organization’s perspective, more desirable managers have realized fitness

functions that closely resemble their cognitive map. Organizations seek to optimize

performance by selecting managers who are endowed with fitness functions that are

likely to produce high quality outcomes.

Formally, an organization, O, currently implements a strategy, x0 2 R, which

produces an outcome, y0 2 R. The organization desires an ideal outcome, y⇤, and

seeks to change its strategy in order to achieve it. O delegates authority over the

strategic change to a manager, m, who is selected at random from a group of potential

managers, M . For ease of exposition, let y0 > y⇤ = 0 and assume x0 = 0.

Managers transform strategies, x 2 R, into outcomes, y(m) 2 R, through a fitness

function, m, where m(x) = y(m). Each m 2 M is associated with a different m,

which is selected by Nature. As in Callander (2011a,b), m is defined by a realization

of a path of Brownian motion with known drift, µ, and variance, �2, that is “tied

down” to the known status quo strategy-outcome pair: (x0, y0). As m is the product

of a stochastic process, the realized fitness function m is initially unknown to both

O and m.

m has the following properties: (1) m is a many-to-one function; m : x !

y(m). Although each manager’s strategy, x, only produces one outcome, y(m),

many strategies may produce the same outcome. (2) Larger changes in x produce

larger changes in y(m), in expectation. (3) Ex ante, m is unknown to both O and

m. However, certain properties of m are common knowledge: O and m know the


expected change in y(m) given a one unit change in x, defined as µ. O and m also

know the variance of the distribution of possible outcomes associated with a one-unit

change in x, defined as �2. (4) With each observed pair of x and y(m), O and m

learn more about m and, as a result, predict how m maps from x to y(m) with

greater certainty.

µ and �2 are constant and m(x0) = y0 for all m 2 M . All that varies between

managers, therefore, is the stochastic realization of m. Ex ante, O is indifferent

between all m 2 M .

m has one opportunity to experiment offline with a prototype strategy, xp, prior

to the organization going online with an implemented strategy, x1. The experimen-

tation and implementation proceed as follows: In the first period, m selects xp. O

and m observe xp and yp(m). Then, O has the opportunity to replace m with m0,

who is another manager selected at random from M . m0 transforms strategies to

outcomes through the function m0 . In the second period, m or m0 implements x1

and, subsequently, y1(m) (or y1(m0)) is revealed.2 Importantly, m0 cannot use the

realization of (xp, yp(m)) to inform its choice of strategy. m0 has the same knowledge

of its own fitness function as m had prior to first period.

The utility for O is quadratic-loss: UO = �(y1 � y⇤)2 = �y21. Outcomes closer to

O’s ideal outcome, y⇤, are more desirable than those farther away. The quadratic-loss

utility function also makes O risk-averse such that EUO = �E(y1)2 � var(y1).

m and m0 have lexicographic preferences. Foremost, m is tenure-motivated. Um =

u if m is replaced and u if m is not replaced, where u < u. Secondarily, m and m0

have preferences that align with O’s. (As m0 only has one period of tenure, m0 has

preferences, therefore, that fully align with O’s and will select x1 to maximize EUO.)2From this point forward, both y1(m) and y1(m0) will be referred to as y1 unless it is necessary

to be explicit about whether m or m0 is in charge of implementation in the second period.


The third metric for comparison is that, all else equal, m and m0 have a preference

for certainty over uncertainty. When facing the choice between two strategies that

produce the same expected utility for O, m and m0 will select the one which produces

a known y1. Finally, m and m0 have a status quo bias. When selecting between the

status quo strategy and a prototyped strategy, m and m0 prefer the status quo.

A replacement rule, r, is the rule that O uses to determine whether or not to re-

place m conditional on the observed (xp, yp(m)). I assume that organizations credibly

commit to replacement rules prior to the first period. I examine four possible re-

placement rules: (1) Guaranteed Tenure (GT): O never replaces m. (2) Retrospective

Satisficing (RS): O replaces m if and only if |yp(m) � y⇤| � |y0 � y⇤|. (3) Forward-

looking Rational (FR): O replaces m if and only if EUO(m|(xp, yp(m))) EUO(m0).

(4) Outcome-based Replacement (OB): O replaces m if and only if yp(m) /2 (yp⇤, yp

⇤)

where both yp⇤ and yp

⇤ are at least as undesirable as y0 and are determined by O in

order to maximize EUO.3

Properties of Experimentation on Brownian Paths

Although m and m0 are unknown to m and O, beliefs about untried strategies

are defined by the properties of Brownian motion. In the first period, beliefs about

strategies are tied down by the status quo strategy-outcome pair: (x0, y0). Therefore,

beliefs about the outcomes resulting from prototype strategies to the right and left of

x0 are “open-ended.” Specifically, beliefs about yp(m) are distributed normally with

expected value:

E(yp(m)|xp) = y0 + (xp � x0)µ

3Given the nature of m’s preferences, if O replaces m half of the time if (and only if) theconditions set forth by O are fulfilled, the prototyping strategy for m will be equivalent.


Figure 3.1: Brownian Path with Known (x0,y0) = (0,2)

-2 -1 0 1 2 3 4

-4-2

02

4

Prototype Strategy, xp

Pro

toty

peO

utco

me,

y p(m)

(x0, y0) = (0, 2)µ = �1

�2= 1

and variance:

V ar(yp(m)|xp) = |xp � x0|�2

Figure 3.1 displays one possible realization of m (the solid line) along with

the expected values of yp(m) given xp (the dashed line) and regions defined by one

standard deviation from the expected value of yp(m) (the dotted lines). The status

quo strategy-outcome pair is (0, 2), the drift, µ, is �1, and the variance, �2, is 1.

Following the observation of (xp, yp(m)), m’s beliefs about outcomes resulting

from x1 are updated to take into account the revealed information. Let xl be the

lesser of x0 and xp, and xr be the greater of x0 and xp. Let yl(m) [yr(m)] be the

outcome associated with xl [xr]. For x1 < xl and x1 > xr, the beliefs over untried


strategies remain open ended, tied down on the right by xl in the former and on the

left by xr in the latter. For xl < x1 < xr, however, beliefs about untried policies

are tied down on both ends. Rather than engaging in open-ended search, m is now

exploring on a Brownian bridge. For x1 on a Brownian bridge, beliefs about yp(m)

are distributed normally with expected value:

E(y1(m)|x1, (xl, yl), (xr, yr)) =yr � ylxr � xl

and variance:

V ar(y1(m)|x1, (xl, yl), (xr, yr)) =(xr � x1)(x1 � xl)

xr � xl

�2

Figure 3.2 displays one possible realization of m (solid line) following a prototype-

outcome pair revealed at (3,�0.25), along the with expected values of y1(m) given

x1 (dashed line) and the regions defined by one standard deviation from the expected

value of y1(m) (the dotted lines). As before, the status quo strategy-outcome pair

is (0, 2), the drift, µ = �1, and the variance, �2= 1. Note that the slope of the

expected values of m on the bridge between (x0, y0) and (xp, yp(m)) differs from the

slopes of the expected values on the flanks to the left of xl and to the right of xr.

Further, for untried strategies a given distance from a known strategy-outcome pair,

the variance of outcomes associated with strategies on the bridge is smaller than the

variance associated with outcomes on the left and right flanks.

Equilibrium Definition and Methodology


Figure 3.2: Brownian Path with Known (x0,y0) = (0,2) and(xp,yp(m)) = (3,�0.25)

-1 0 1 2 3 4

-10

12

34

Implemented Strategy, x1

Fina

lOut

com

e,y 1(m)

(x0, y0) = (0, 2)(xp, yp(m)) = (3,�0.25)µ = �1

�2= 1


I utilize a subgame perfect equilibrium concept. As the game proceeds sequen-

tially and there is perfect information, there exists a unique subgame perfect equi-

librium.

I solve the game through backwards induction. The second period equilibrium

strategies for m and m0, x⇤1(m) and x⇤

1(m0), are defined as in Callander (2011a,

2011b). The first period equilibrium strategy for m, defined as x⇤p(r), is the xp that

maximizes EUm conditional on replacement rule r, where r 2 {GT,RS, FR,OR}.

The equilibrium replacement rules for O define conditions under which O will

replace m. For O 2 {GT,RS}, the replacement rules for O are self-explanatory.

For r = FR, the prototype-outcome pairs for which O replaces m, indicated by

(xp,yp)⇤FR, are defined such that if (xp, yp) 2 (xp,yp)

⇤FR, then EUO(m|x⇤

1(m), (xp, yp))

EUO(m0|x⇤

1(m0)). For r = OB, the boundaries assigned in equilibrium by O, (yp⇤, yp⇤),

are defined in order to maximize EUO subject to the best response by m, x⇤p(OB).

In order to find x⇤p, I rely on a combination of analytical methods and compu-

tation. I find the analytic solution for x⇤1(m) conditional on µ, �2, (x0, y0), and

(xp, yp(m)). I also find the analytic solution for x⇤1(m

0) conditional on µ, �2, and

(x0, y0). The nature of the replacement regimes and the properties of Brownian

motion preclude explicit analytic solutions for x⇤p(r). As a result, I define implicit

solutions for x⇤p conditional on µ, �2, r, x⇤

1(m), and x⇤1(m

0). I then solve for x⇤

p(r)

using numerical optimization methods.

I demonstrate x⇤p and EUO(x

⇤p) for a range of y0 > 0 and �2 under different

replacement rules. For analytic simplicity, I fix µ = �1. y0 represents the extent

to which O is dissatisfied with the status quo. �2 represents the complexity of the

problem that the managers face.4

4Callander (2011a; 2011b) demonstrates that search behavior in models of this sort is determinedby the ratio of the increase in uncertainty given a one unit change in x and the change in y given


Analytical Results

In order to solve the model by backwards induction, I first find x⇤1(m) and x⇤

1(m0)

given (x0, y0), (xp, yp(m)), µ, and �. For m and m0, preferences are aligned with O

in the second period. As a result, m and m0 face a problem (nearly) identical to the

one analyzed in Callander (2011a; 2011b). Therefore, Propositions 1 and 2a-2g and

Lemmas 1a-1d are direct applications of Callander (2011a; 2011b).

Second Period Behavior

I first examine x⇤1(m

0). As m has been replaced and the information revealed

during prototyping is of no value to m0, m0 selects x1 based on their beliefs about

m0 in order to maximize EUm0 . This is equivalent to maximizing EUO = �(y1�y⇤)2.

There are three possible strategic changes available to m0 in the second period. m0

can choose not to make a revision to x0, which I describe as a stable strategy. m0 can

choose to make a revision to x0 such that x1 > x0, which I describe as a rightward

change. m0 can also choose the make a revision to x0 such that x1 < x0, a left-

ward change. However, given that y0 > 0 and µ < 0, a leftward change, which moves

y1(m0) further away from y⇤ in expectation, is strictly dominated by a stable strategy.

Lemma 1a: For y0 > �2

2 , the optimal x1 for m0, x⇤1(m

0), produces an expected

outcome of �2

2 . For y0 �2

2 , x⇤1(m) is stable at x0.

Risk aversion leads m0 to aim for an outcome that is between the status quo

outcome and their ideal outcome of 0. The boundary at which the utility sacrificed

from the additional variance associated with more exploration exceeds the utility

benefit from a more desirable outcome in expectation is half the complexity of the

a one unit change in x. Therefore, generality is maintained by assuming µ = �1.


search environment, �2

2 . For status quo outcomes outside of the boundary, m0 will

search rightward. For status quo outcomes interior to the boundary, m0 will prefer

the status quo outcome to the risk associated with rightward search.

Proposition 1: The optimal x1 for m0, x⇤1(m) is:

• If y0 > �2

2 , x⇤1(m

0) = y0 � �2

2 . EUO = �(y0 � �2

4 )�2.

• If y0 �2

2 , x⇤1(m

0) = x0. EUO = �y20.

For y0 that is sufficiently distant from y⇤, m0 explores to the right of x0 in order

to seek a more desirable outcome. For y0 sufficiently close to y⇤, m0 prefers a stable

strategy in which it retains the status quo, x0, than to explore and bear the associ-

ated risk of a bad outcome.

Next, I examine the optimal x1 for m, x⇤1(m), conditional on the realization of

(xp, yp(m)). The strategy depends crucially on whether yp(m) has moved backwards

and away from y0, moved towards y0 but still short of y⇤, or moved past y⇤.

There are five possible strategic changes available to m in the second period,

conditional on (xp, yp(m)). m can retain xp, which I describe as a stable strategy. m

can choose to return to x0, which I describe as backsliding. m can choose x1 > xr,

where xr = max(x0, xp), a rightward change. m can xl < x1 < xr, which I de-

scribe as a change on the Brownian bridge. Finally, m can choose x1 < xl, where

xl = min(x0, xp), a leftward change.

First, I explore the dominant second period strategy for m if m chooses to retain

the status quo and forego the opportunity to prototype. In such a case, the strategic


decision for m in the second period is equivalent to the strategic decision for m0 in

the second period.

Proposition 2a: If xp = x0, then x⇤1(m) = x⇤

1(m0).

Now, I move on to the (more interesting) cases where xp 6= x0. I first ex-

amine the conditions under which backsliding is the dominant strategy for m for

yp(m) > y⇤ = 0.

Lemma 1b: For yp(m) > 0, a leftward change is strictly dominated by a stable

strategy if yp(m) y0; or a backsliding strategy if yp(m) > y0.

As a leftward strategy moves y1 away from y⇤ for yp(m) > 0, m prefers to retain

the known strategy-outcome pair that is closest to y⇤ than to revise its strategy left-

ward.

Lemma 1c: For yp(m) > �2

2 and xp > 0, rightward exploration dominates retaining

the prototype strategy.

After observing yp(m), m chooses whether to continue to search rightward, to

keep the prototype strategy, or to backslide to x0. Lemma 1b notes that, when

choosing between the first two options, m prefers to retain the prototype xp if xp

is sufficiently close to y⇤. Otherwise, m prefers to explore rightward than to retain xp.

Proposition 2b: For values of �2 and y0 and realizations of (xp, yp(m)) that satisfy

the following conditions, the optimal x1 for m, x⇤1(m), involves backsliding to x0.

• If y0 > �2

2 and yp(m) � y20+�2/4

�2 and xp > x0, x⇤1(m) = x0. EUO = �y20.


• If y0 �2

2 and yp(m) � y0, x⇤1(m) = x0. EUO = �y20.

For observed yp(m) sufficiently far away from y⇤ and xp > x0, m prefers to back-

slide to x0 rather than to continue to explore to the right from xp and take on the

associated uncertainty. For all yp(m) > y0 and for y0 sufficiently close to y⇤ (i.e.

y0 �2

2 ), m prefers to backslide to x0 than to either retain xp or search rightward.

For y0 sufficiently far from the ideal outcome (i.e. y0 > �2

2 ) and yp(m) > y0, m

weighs the expected utility associated with exploring rightward with the expected

utility from backsliding to x0. For yp(m) > y0 and yp(m) >> y⇤, m backslides.

For yp(m) closer to y⇤ but not so close as to be desirable in and of itself, a back-

sliding strategy is dominated by a rightward change.

Proposition 2c: For values of �2 and y0 and realizations of (xp, yp(m)) that satisfy

the following conditions, x⇤1(m) involves a rightward revision.

• If y0 > �2

2 and yp(m) <y20+�

2/4�2 < �2

2 and xp > x0, x⇤1(m) = yp(m) � �2

2 .

EUO = �(yp(m)� �2

4 )�2.

• If y0 > �2

2 and yp(m) >q

(y0 � �2

4 )�2 and xp < x0, x⇤

1(m) = y0 � �2

2 . EUO =

�(y0 � �2

4 )�2.

For yp(m) far enough away from y⇤ as to make continued search desirable (but

not so far as to make rightward search less desirable than backsliding) and xp > x0,

m continues to search to the right from xp. For xp < x0, unless xp is sufficiently close

to y⇤ as to make it desirable to be retained, m searches to the right from x0 in period 2.

For yp(m) sufficiently close to y⇤, m retains yp(m) rather than risk the uncer-

tainty associated with revising its strategy. I first consider offline prototypes for


which yp(m) is still short of y⇤ and m chooses to retain yp(m).

Proposition 2d: For values of �2 and y0 and realizations of (xp, yp(m)) that satisfy

the following conditions, x⇤1(m) is stable at xp.

• If �2

2 � yp(m) � 0 and xp > x0, x⇤1(m) = xp. EUO = �yp(m)

2.

• Ifq

(y0 � �2

4 )�2 � xp � 0 and xp < x0, x⇤

1(m) = xp. EUO = �yp(m)

2.

If the prototype outcome is sufficiently close to y⇤ and involved a rightward re-

vision from x0, then m retains xp. In contrast, if the prototype involved a leftward

revision, then m compares the known utility from remaining at xp to the expected

utility from a rightward change from x0.

I now move on to yp(m) < y0 and xp > 0.5 For yp(m) < y0, m has the opportunity

to experiment on the Brownian bridge and reap the gains associated with the de-

creased variance. However, for strategy-outcome pairs for which yp(m) is sufficiently

close to y⇤, m still prefers a stable strategy.

Lemma 1d: For yp(m) < 0 and xp < x0, a rightward change is dominated by xp.

Because the prototype has overshot its target, further exploration to the right

only brings m further from the organization’s goal, in expectation.

Proposition 2e: For values of �2 and y0 and realizations of (xp, yp(m)) that satisfy

the following conditions, x⇤1(m) is stable at xp.

5Note that I do not characterize x⇤1 for yp(m) < 0 and xp < x0. Solving for x⇤

1 in these cases isnot necessary to characterize x⇤

p(r) for any of the replacement rules as first-period outcomes in thisregion are fairly unlikely. Further, the optimal behavior of m in these cases is (perhaps surprisingly)complex to define explicitly.


• If 0 > yp(m) � �y0 and yp(m) � � �2

2|µ| where µ =

yp

(m)�y0xp

, x⇤1(m) = xp.

EUO = �yp(m)

2.

Note that the boundary of yp(m) for which m prefers the known xp to exper-

imenting on the bridge grows more extreme as the expected slope of m on the

bridge becomes shallower. Bigger experiments require larger differences in outcomes

in order to warrant further experimentation in the second period.

However, for sufficiently large first period experiments (i.e., xp >> 0), yp can be

further from y⇤ than y0, yet experimenting on the bridge is still undesirable. As a re-

sult, rather than remaining stable at xp, m backslides to x0, as EUm(x0) > EUm(xp).

Proposition 2f: For values of �2 and y0 and realizations of (xp, yp(m)) that satisfy

the following conditions, x⇤1(m) backslides to x0.

• If yp(m) �y0 and y0 �2

2|µ| , x⇤1(m) = x0, EUO = �y20.

For more extreme values of yp(m), the expected slope is sufficiently large to in-

duce experimentation on the bridge. x⇤1(m) lies between x0 and xp. If y0 is more

desirable than yp, x⇤1(m) < x

p

2 and E(y1) > 0. Otherwise, x⇤1(m) > x

p

2 and E(y1) < 0.

Proposition 2g: For values of �2 and y0 and realizations of (xp, yp(m)) that satisfy

the following conditions, x⇤1(m) is on the bridge spanning (x0, y0) and (xp, yp(m)).

• If yp(m) < � �2

2|µ| , or yp(m) > y0 and y0 >�2

2|µ| :

x⇤1(m) =

2µy0 + �2

2

⇣�2

xp

� µ2⌘

EUO = �(y0 + µx⇤1(m))

2 � (xp � x⇤1(m))x⇤

1(m)

xp

�2


I graphically depict x⇤1(m) given (xp, yp(m)) and �2

= 1. In Figure 3.3, y0 is

sufficiently undesirable as to induce rightward search for some yp > 0. In contrast,

in Figure 3.4, y0 is fairly close to y⇤, which makes rightward search undesirable for

yp > 0.

Replacement Rules and Prototyping Strategies

I now examine the prototyping behavior of m and replacement decisions of O in

greater detail.

Under Guaranteed Tenure, r = GT , the replacement rule fully aligns the pref-

erences of m and O. As a result, m selects x⇤p(GT ) in order to maximize EUO in

period 2 under the assumption that m will not be replaced. Guaranteed Tenure

produces the same behavior for m in equilibrium as if O implemented a replacement

rule in which O assigns m a prototype strategy x⇤⇤p and replaces m if and only if

xp(m) 6= x⇤⇤p .

Note, however, that if there were no principal-agent problems (i.e. m were not

tenure motivated) and m were replaced after period 1 if EUO/m(m) < EUO/m(m0),

x⇤p(GT ) would not be the utility-maximizing level of xp. If m were revealed to be

a poor fit for the task, m would prefer to be replaced with m0 rather than being

retained. This inefficiency associated with being forced to retain an undesirable m

under Guaranteed Tenure leads to an conservative bias in x⇤p(GT ) relative to the

optimal experiment under fully aligned preferences. However, x⇤p(GT ) is equivalent

to the utility-maximizing xp for O when m0 is incapable of improving on the status

quo, which is the case when y0 < �2

2 , or when the fitness functions of potential re-

placements are identical to m’s, i.e. m(x) = m0(x) for all m0 2 M and for all x 2 R.

Proposition 3: x⇤p(GT ) is defined implicitly by the following equation:


Figure 3.3: x

⇤1(m) conditional of (xp,yp(m)) for (x0,y0) = (0,2)

-10 -5 0 5 10 15 20 25

-4-2

02

46

xp

y p(m)

Right

Backslide

Right

StableStable

Stable

Brownian Bridge

Back-slide

(x0, y0) = (0, 2)


Figure 3.4: x

⇤1(m) conditional of (xp,yp(m)) for (x0,y0) = (0, 14)

-0.5 0.0 0.5 1.0

-1.0

-0.5

0.0

0.5

1.0

xp

y p(m)

Backslide Backslide

Stable Stable

Stable

BackslideBrownianBridge

(x0, y0) = (0, 14)


max

xp

EUO(y1|x⇤1(m), x1(m

0) = x0)

Under the Retrospective Satisficing replacement rule, r = RS, O will replace

m if yp is not closer to y⇤ than y0. x⇤p(RS) is, therefore, the prototyping strategy

that maximizes the probability that y0 > |yp|. It is important that m be replaced if

|yp| = y0 in order to induce m to experiment at all.

Proposition 4: x⇤p(RS) is unique and defined implicitly by the following equation:

max

xp

"�

xpp�2xp

!� �

xp � 2y0p

�2xp

!#

One property of x⇤p(RS) is that it exhibits risk aversion. As larger experiments

generate more uncertainty, m will choose a prototype strategy which produces an

outcome that falls short of y⇤, in expectation.

Lemma 2a: x⇤p(RS) > 0 and E( m(x

⇤p(RS))) < y⇤.

Uniqueness and risk aversion, in fact, reflects a more general property of replace-

ment rules in which yp must fall between y0 and a lower bound, yp, in order for m to

be retained. Under a replacement rule of this sort, EUm(xp) is quasi-concave in xp

for xp > 0. Furthermore, E( m(x⇤p(RS))) always falls short of the midpoint between

the upper and lower boundary.

Lemma 2b: For any replacement regime in which m is replaced if yp /2 (yp, y0),


x⇤p(RS) is unique and in (0,

y0�yp

2 ).

Under the Forward-Looking Rational replacement rule, r = FR, O will replace m

when EUO(m|x⇤1(m), (xp, yp(m))) EUO(m

0|x⇤1(m

0)). There are two regions where

this will be the case. First, if yp(m) > y0, then m is less attractive than m0 because

m has to search farther to the right than m0 to reach y⇤. Second, in the regions

where xp > 0 and yp(m) < 0 in which m backslides to x0, m0 can do at least as well

as m by keeping x0 (or improve on x0 by experimenting to the right for y0 >�2

2 ).

In equilibrium, m, therefore, searches to the right until there exist values of yp(m)

given xp that will lead O to replace m. This strategy minimizes the probability that

yp(m) > y0 without adding an additional risk of replacement.

Proposition 5: If y0 > �2

2 , then x⇤p(FR) � 4y0��2. If y0 �2

2 , then x⇤p(FR) =

4y20�2 .

Lemma 3: If the first derivative of the values of (xp, yp) that satisfy EUO(m, (xp, yp)) =

EUO(m0) is less than �1 for y0 >

�2

2 , then x⇤p(FR) = 4y0 � �2.

An Outcome-based replacement rule is similar to Retrospective Satisficing, in

that if yp fails to fall between an upper and lower bound, m will be replaced by m0.

However, Outcome-based replacement allows O to set the upper and lower bound to

be more extreme than y0 and �y0 in order to induce a more desirable first period

experiment than x⇤p(RS). However, the rule has the downside of requiring that O

replace m when yp yp, even when the expected utility of retaining m exceeds the

expected utility of replacing m.

Under Outcome-based replacement, the upper bound, yp⇤ is equal to y0 in equilib-

rium. If it were greater than y0, then m would select xp = x0 and never be replaced.


Therefore, O adjusts yp⇤ in order to induce the desired x⇤

p(OB).

Proposition 6a: yp⇤= y0. yp

⇤ is defined implicitly by the following equation:

max

yp

EUO(x⇤p(OB), x⇤

1(m), x⇤1(m

0))

m selects xp in a similar fashion to the Retrospective Satisficing replacement rule,

although with a lower bound of yp rather than �y0.

Proposition 6b: x⇤p(OB) is defined implicitly by the following equation, where yp

indicates the lower bound for acceptable outcomes defined by O. For a given yp,

there is a unique induced x⇤p(OB) that maximizes the equation.

max

xp

"�

xpp�2xp

!� �

xp + yp � y0p�2xp

!#

Whereas Guaranteed Tenure sometimes forces O to keep m when it would prefer

to replace m, Outcome-based replacement requires that O sometimes replace m when

O would prefer to keep m. Therefore, when choosing between these two replacement

regimes, O trades off the cost of keeping an undesirable m with replacing a desirable

one.

Proposition 7: O weakly prefers r = GT to r = OB when EUO(x⇤p(GT )) �

EUO(x⇤p(OB)). O weakly prefers r = OB to r = GT when EUO(x

⇤p(GT ))

EUO(x⇤p(OB)). O strictly prefers r = GT to r = OB if y0 �2

2 .


Computational Approach

I briefly describe how I computationally derive equilibrium outcomes for x⇤p(r).

All global optimizations use either golden-section search, simulated annealing or

genetic optimization, implemented using the stats, GenSA and rgenoud packages

in R-3.2.1. Appendix B offers an annotated version of the scripts used to calculate

all of the examples in the paper.

I use golden-section optimization or simulated annealing when I can demonstrate

analytically that the function EUm(xp) is quasi-concave in the relevant region. I

use simulated annealing when the bounds on x⇤p(r) are analytically defined but the

function to be optimized is not shown to be quasi-concave, as simulated annealing

tends to be very efficient for unidimensional global optimization subject to a box

constraint (Mullen, 2014). I use genetic optimization when the bounds are not ana-

lytically defined, as genetic optimization allows for the algorithm to explore beyond

predetermined bounds. For more detail on the simulated annealing algorithm, see

Xiang et al. (2013). For more detail on the genetic optimization algorithm, see

Mebane Jr. and Sekhon (2011).

Computational Results

I evaluate the first-period experimentation strategy subject to the various re-

placement rules.

Figure 3.5 demonstrates EUm and EUO conditional on xp under Guaranteed

Tenure for �2 2 {1, 2, 4} and y0 2 {12 , 1, 2, 4} for xp 2 [0, 2y0]. The gray lines indicate

x⇤p. As the preferences of m are aligned with O, the expected utilities of m and O

are equal. Further, EUm appears quasi-concave with respect to xp for the values

depicted. Finally, the utility curves are proportional when the ratios of the distance

from the status quo outcome to the optimal outcome (i.e. y0) and the complexity


of the fitness function (i.e. �2) are equal. (See, for example, �2= 1, y0 = 1 vs.

�2= 2, y0 = 2 vs. �2

= 4, y0 = 4.) This property will hold for all of the replacement

rules.

Figure 3.6 graphs x⇤p conditional on y0 as a function of �2. Most striking is the

non-monotonic relationship between optimal experimentation behavior and the com-

plexity of the task environment. For tasks of low to moderate complexity, optimal

experimentation involves shooting past the ideal outcome (i.e. E( m(xp)) < 0) and

then triangulating on a Brownian bridge in the second period. This makes intuitive

sense. In low complexity task environments, the modest increased uncertainty asso-

ciated with the bigger first period experiment is outstripped by the gains associated

with lower uncertainty in the implementation phase. However, for more complex

environments, the uncertainty associated with a big first period experiment swamps

the benefits from possibly triangulating on a Brownian bridge in the second period.

Further, the optimal experiment is always strictly positive, but approaches zero

as complexity increases. It is always beneficial for m to undertake some experi-

mentation in the first period, because m can backslide in the second period if the

experimental outcome is undesirable. However, for environments characterized by

explosive uncertainty, m makes a very incremental experiment in the hope of moving

marginally closer to the ideal outcome than the status quo. Note that, in expectation,

this sort of experimentation behavior will lead to outcomes that are incrementally

more desirable than the status quo. Approximately half of the time the organization

ends up backsliding to the status quo after observing the experimental outcome.

Finally, as in Figure 3.5, the shapes of the x⇤p functions are proportional. When

the range of �2 is scaled by y0, the four functions are identical in shape, although

varied in the magnitude of x⇤p. This property will also hold for all of the replacement

rules.


Figure 3.5: Expected Utilities under r = GT

x

EUm

-0.17-0.16-0.15-0.14-0.13-0.12

�2= 1, y0 = 0.5

x

EUO

-0.17-0.16-0.15-0.14-0.13-0.12

0.2 0.4 0.6 0.8 1.0

x⇤p(GT )

x

EUm

-0.69-0.62-0.55-0.47-0.4

-0.33�2

= 1, y0 = 1

x

EUO

-0.69-0.62-0.55-0.47-0.4

-0.33

0.0 0.5 1.0 1.5 2.0

x⇤p(GT )

x

EUm

-1.7-1.48-1.26-1.05-0.83-0.61

�2= 1, y0 = 2

xEUO

-1.7-1.48-1.26-1.05-0.83-0.61

0 1 2 3 4

x⇤p(GT )

x

EUm

-3.7-3.15-2.61-2.06-1.52-0.97

�2= 1, y0 = 4

x

EUO

-3.7-3.15-2.61-2.06-1.52-0.97

0 2 4 6 8

x⇤p(GT )

x

EUm

-0.2-0.19-0.18-0.16-0.15-0.14

�2= 2, y0 = 0.5

x

EUO

-0.2-0.19-0.18-0.16-0.15-0.14

0.2 0.4 0.6 0.8 1.0

x⇤p(GT )

x

EUm

-0.76-0.7

-0.64-0.59-0.53-0.47

�2= 2, y0 = 1

x

EUO

-0.76-0.7

-0.64-0.59-0.53-0.47

0.0 0.5 1.0 1.5 2.0

x⇤p(GT )

x

EUm

-2.89-2.57-2.26-1.94-1.63-1.31

�2= 2, y0 = 2

x

EUO

-2.89-2.57-2.26-1.94-1.63-1.31

0 1 2 3 4

x⇤p(GT )

xEUm

-6.9-6.01-5.11-4.22-3.32-2.43

�2= 2, y0 = 4

x

EUO

-6.9-6.01-5.11-4.22-3.32-2.43

0 2 4 6 8

x⇤p(GT )

x

EUm

-0.22

-0.21

-0.19

-0.18

-0.16

-0.15�2

= 4, y0 = 0.5

x

EUO

-0.22

-0.21

-0.19

-0.18

-0.16

-0.15

0.2 0.4 0.6 0.8 1.0

x⇤p(GT )

x

EUm

-0.8-0.75-0.7

-0.66-0.61-0.56

�2= 4, y0 = 1

x

EUO

-0.8-0.75-0.7

-0.66-0.61-0.56

0.0 0.5 1.0 1.5 2.0

x⇤p(GT )

x

EUm

-3.3-3.02-2.74-2.45-2.17-1.89

�2= 4, y0 = 2

x

EUO

-3.3-3.02-2.74-2.45-2.17-1.89

0 1 2 3 4

x⇤p(GT )

x

EUm

-11.79-10.48-9.16-7.85-6.53-5.22

�2= 4, y0 = 4

x

EUO

-11.79-10.48-9.16-7.85-6.53-5.22

0 2 4 6 8

x⇤p(GT )

xp


Figure 3.6: x⇤p under r = GT

s2

x⇤ p(GT)

0.1

0.2

0.3

0.4

0.5

0.6 �2 2 (0.1, 6), y0 = 0.5

s2

x⇤ p(GT)

0.2

0.4

0.6

0.8

1.0

1.2 �2 2 (0.1, 6), y0 = 1

s2

x⇤ p(GT)

1.0

1.4

1.8

2.2

�2 2 (0.1, 6), y0 = 2

s2

x⇤ p(GT)

3.6

4.0

4.4

�2 2 (0.1, 6), y0 = 4

0 1 2 3 4 5 6�2


Figure 3.7 depicts EUm and EUO under Retrospective Satisficing for the same

parameter values as Figure 3.5. Unlike in Figure 3.5, however, EUm is represented

by the probability of m being replaced after the experimental outcome is revealed.

As demonstrated in Lemma 2a, the probability of replacement is quasi-convex with

respect to xp for xp > 0. Note that x⇤p is consistently slightly less extreme than what

is optimal from the perspective of O.

Figure 3.8 shows x⇤p conditional of y0 as a function of �2. Most striking is the

risk averse behavior induced by the Retrospective Satisficing replacement rule. As

noted in Lemma 2a, E( m(x⇤p)) is always short of y⇤, even for low uncertainty task

environments. The figure demonstrates that x⇤p approaches zero as the complexity

of the task environment increases. This incremental search behavior is a product of

the lower boundary of acceptable outcomes being equal to �y0, a constraint that is

weakened under Outcome-based replacement.

Figure 3.9 seeks to demonstrate the truth of the antecedent in Lemma 3. The

graphs demonstrate that the lower cutpoint of the region in which yp < 0 and

m is replaced (i.e. the lower replacement region) decreases more quickly than the

E( m(xp)) as xp grows larger. As a result, x⇤p is equal to the most extreme value of

xp such that m will never be replaced if y0 < 0. For y0 < �2

2 , E( m(xp)) actually

falls within the lower replacement region. As a result, as xp ! 1, the probability of

replacement approaches 1. If xp >�2

2 , then the the lower bound of the replacement

region approaches E( m(xp)) as xp ! 1. As a result, if m were to explore further to

the right, the small benefit associated with the decreased likelihood that m(xp) � y0

would be swamped by the increased likelihood that m(xp) would fall in the lower

replacement region.

Figure 3.10 shows the optimal experimentation behavior for m under a Forward-

looking Rational replacement rule. The fear of replacement, here, induces m to


Figure 3.7: Expected Utilities under r = RS

x

Pr(

repl

ace)

0.360.420.480.540.6

0.66�2

= 1, y0 = 0.5

x

EUO

-0.21-0.19-0.18

-0.16-0.15-0.13

0.0 0.2 0.4 0.6 0.8 1.0

x⇤p(RS)

x

Pr(

repl

ace)

0.250.320.380.450.510.58

�2= 1, y0 = 1

x

EUO

-0.7-0.63-0.56-0.48-0.41-0.34

0.0 0.5 1.0 1.5 2.0

x⇤p(RS)

x

Pr(

repl

ace)

0.130.210.290.360.440.52

�2= 1, y0 = 2

xEUO

-1.7-1.49-1.28-1.07-0.86-0.65

0 1 2 3 4

x⇤p(RS)

x

Pr(

repl

ace)

0.040.130.220.320.410.5

�2= 1, y0 = 4

x

EUO

-3.7-3.17-2.64-2.1

-1.57-1.04

0 2 4 6 8

x⇤p(RS)

x

Pr(

repl

ace)

0.430.490.550.620.680.74

�2= 2, y0 = 0.5

x

EUO

-0.2

-0.19

-0.17

-0.16

-0.15

0.0 0.2 0.4 0.6 0.8 1.0

x⇤p(RS)

x

Pr(

repl

ace)

0.360.420.480.540.6

0.66�2

= 2, y0 = 1

x

EUO

-0.89-0.82-0.75-0.67-0.6

-0.53

0.0 0.5 1.0 1.5 2.0

x⇤p(RS)

x

Pr(

repl

ace)

0.250.320.380.450.510.58

�2= 2, y0 = 2

x

EUO

-2.88-2.58-2.28-1.98-1.68-1.38

0 1 2 3 4

x⇤p(RS)

xPr(

repl

ace)

0.130.210.290.360.440.52

�2= 2, y0 = 4

x

EUO

-6.88-6.02-5.17-4.31-3.46-2.6

0 2 4 6 8

x⇤p(RS)

x

Pr(

repl

ace)

0.470.540.610.670.740.81

�2= 4, y0 = 0.5

x

EUO

-0.22-0.21-0.2

-0.18-0.17-0.16

0.0 0.2 0.4 0.6 0.8 1.0

x⇤p(RS)

x

Pr(

repl

ace)

0.430.490.550.620.680.74

�2= 4, y0 = 1

x

EUO

-0.85-0.8

-0.75-0.69-0.64-0.59

0.0 0.5 1.0 1.5 2.0

x⇤p(RS)

x

Pr(

repl

ace)

0.360.420.480.540.6

0.66�2

= 4, y0 = 2

x

EUO

-3.68-3.37-3.06-2.74-2.43-2.12

0 1 2 3 4

x⇤p(RS)

x

Pr(

repl

ace)

0.250.320.380.450.510.58

�2= 4, y0 = 4

x

EUO

-11.66-10.43

-9.2-7.97-6.74-5.51

0 2 4 6 8

x⇤p(RS)

xp


Figure 3.8: x⇤p under r = RS

s2

x⇤ p(RS)

0.1

0.2

0.3

0.4

�2 2 (0.1, 6), y0 = 0.5

s2

x⇤ p(RS)

0.2

0.4

0.6

0.8

1.0 �2 2 (0.1, 6), y0 = 1

s2

x⇤ p(RS)

0.5

1.0

1.5

2.0 �2 2 (0.1, 6), y0 = 2

s2

x⇤ p(RS)

2.0

2.5

3.0

3.5

4.0 �2 2 (0.1, 6), y0 = 4

0 1 2 3 4 5 6�2


Figure 3.9: Lower Cutpoint under r = FR

x

y p

-10.5-8.5-6.5-4.5-2.5-0.5 cutpoint

y0 � xp

�2= 1, y0 = 0.5

Diff

eren

ce

0

2 4 6 8 10

x

y p-12-9.8-7.6-5.4-3.2

-1 cutpointy0 � xp

�2= 1, y0 = 1

Diff

eren

ce

0.50.60.70.80.9

1

4 6 8 10 12

x

y p

-15-12.4-9.8-7.2-4.6


�2= 1, y0 = 2

Diff

eren

ce

1.72

2.22.52.7

3

8 10 12 14 16

x

y p

-21-17.6-14.2-10.8-7.4


�2= 1, y0 = 4

Diff

eren

ce

4.34.85.45.96.5

7

16 18 20 22 24

x

y p

-19.5-15.6-11.7-7.8-3.9

0 cutpointy0 � xp

�2= 2, y0 = 0.5

Diff

eren

ce

-10-8.1-6.2-4.3-2.4-0.5

2 4 6 8 10

x

y p

-11-9-7-5-3-1 cutpoint

y0 � xp

�2= 2, y0 = 1

Diff

eren

ce

0

2 4 6 8 10 12

x

y p

-14-11.6-9.2-6.8-4.4


�2= 2, y0 = 2

Diff

eren

ce

1.11.31.51.61.8

2

6 8 10 12 14 16

xy p

-20-16.8-13.6-10.4-7.2


�2= 2, y0 = 4

Diff

eren

ce

3.64.14.65.15.5

6

14 16 18 20 22 24

x

y p

-31.5-25.1-18.8-12.5-6.10.2 cutpoint

y0 � xp

�2= 4, y0 = 0.5

Diff

eren

ce

-24-19.4-14.7

-10-5.4-0.8

0 2 4 6 8

x

y p

-19-15.2-11.4-7.6-3.8

0 cutpointy0 � xp

�2= 4, y0 = 1

Diff

eren

ce

-10-8.2-6.4-4.6-2.8

-1

2 4 6 8 10

x

y p

-12-10-8-6-4-2 cutpoint

y0 � xp

�2= 4, y0 = 2

Diff

eren

ce

0

4 6 8 10 12 14

x

y p

-18-15.2-12.4-9.6-6.8


�2= 4, y0 = 4

Diff

eren

ce

2.42.7

33.43.7

4

12 14 16 18 20 22

xp


undertake experiments that are much, much more extreme than under the previ-

ous two replacement rules. Unlike Retrospective Satisficing, which produces overly

incremental experimentation behavior, a Forward-looking Rational rule generates

over-experimentation. By setting an experimental outcome goal that is far past the

optimal outcome, m nearly guarantees that the experiment gleans information that

will be useful in the implementation stage. However, rather than maximizing the

informativeness of the experiment, m maximizes the probability that it will learn

something and, in the process, become marginally more desirable than a potential

replacement, m0.

The last replacement rule that I examine is Outcome-based replacement. Figure

3.11 examines how the behavior of m and the expected utility for O change as a

function of the lower bound of the region of acceptable experimental outcomes, yp.

The upper panels demonstrate how the probability of replacement for m changes as O

adjusts this lower bound, assuming that m rationally chooses xp in order to minimize

the likelihood of being replaced. The gray line indicates the utility-maximizing yp for

O. The lower panels indicate how EUO changes in response to changes in yp, given

the best response of m. Under an Outcome-based rule, the preferences of m and O

are not closely aligned. m would prefer as large a region of acceptable outcomes as

possible. O limits the region in order to induce a desirable x⇤p.

Figure 3.12 illustrates y⇤p as a function of �2. The bound initially declines quickly

in response to the low uncertainty environment in order to induce m to experiment

aggressively. The bound then tightens as the complexity of the task environment

increases. As the uncertainty becomes too explosive, however, the bound loosens

once again. In order to induce sufficiently aggressive experimentation, O gives m

greater freedom when the complexity of the task grows. Note that, for all values of

�2, the region of acceptable experimental outcomes is larger under an Outcome-based


Figure 3.10: x⇤p under r = FR

s2

x⇤ p(FR)

0.5

1.0

1.5

2.0 �2 2 (0.1, 6), y0 = 0.5

s2

x⇤ p(FR)

1.0

2.0

3.0

4.0 �2 2 (0.1, 6), y0 = 1

s2

x⇤ p(FR)

34

56

78 �2 2 (0.1, 6), y0 = 2

s2

x⇤ p(FR)

1011

1213

1415

16

�2 2 (0.1, 6), y0 = 4

0 1 2 3 4 5 6�2


Figure 3.11: Expected Utilities under r = OB

x

Pr(

repl

ace|x⇤ p)

0.210.240.270.3

0.330.36

�2= 1, y0 = 0.5

x

EUO

-0.144

-0.138

-0.132

-0.126

-0.12

-2.0 -1.5 -1.0 -0.5

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.10.130.160.190.220.25

�2= 1, y0 = 1

x

EUO

-0.39-0.374-0.358-0.342-0.326-0.31

-4.0 -3.0 -2.0 -1.0

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.020.040.060.090.110.13

�2= 1, y0 = 2

x

EUO

-0.9

-0.836

-0.772

-0.708

-0.644

-0.58

-8 -6 -4 -2

y⇤p

x

Pr(

repl

ace|x⇤ p)

0

0.01

0.02

0.02

0.03

0.04�2

= 1, y0 = 4

x

EUO

-2.05-1.834-1.618-1.402-1.186-0.97

-16 -12 -8 -4

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.330.350.370.390.410.43

�2= 2, y0 = 0.5

x

EUO

-0.17-0.164-0.158-0.152-0.146-0.14

-2.0 -1.5 -1.0 -0.5

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.210.240.270.3

0.330.36

�2= 2, y0 = 1

x

EUO

-0.59-0.57-0.55-0.53-0.51-0.49

-4.0 -3.0 -2.0 -1.0

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.10.130.160.190.220.25

�2= 2, y0 = 2

x

EUO

-1.55-1.486-1.422-1.358-1.294-1.23

-8 -6 -4 -2

y⇤p

xPr(

repl

ace|x⇤ p)

0.020.040.060.090.110.13

�2= 2, y0 = 4

x

EUO

-3.58-3.332-3.084-2.836-2.588-2.34

-16 -12 -8 -4

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.42

0.43

0.45

0.46

0.47�2

= 4, y0 = 0.5

x

EUO

-0.18

-0.174

-0.168

-0.162

-0.156

-0.15

-2.0 -1.5 -1.0 -0.5

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.330.350.370.390.410.43

�2= 4, y0 = 1

x

EUO

-0.69-0.666-0.642-0.618-0.594-0.57

-4.0 -3.0 -2.0 -1.0

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.210.240.270.3

0.330.36

�2= 4, y0 = 2

x

EUO

-2.37-2.29-2.21-2.13-2.05-1.97

-8 -6 -4 -2

y⇤p

x

Pr(

repl

ace|x⇤ p)

0.10.130.160.190.220.25

�2= 4, y0 = 4

x

EUO

-6.21-5.956-5.702-5.448-5.194-4.94

-16 -12 -8 -4

y⇤p

Lower Bound, yp


replacement rule than under Retrospective Satisficing.

Figure 3.13 examines the expected utilities of m and O given optimal experimen-

tation behavior under the four replacement rules. The panels on the left indicate the

probability of replacement as a function of task complexity. Obviously, Guaranteed

Tenure is most desirable for m. The next most desirable is the Forward-looking Ra-

tional rule, in which m is retained so long as the first period experiment makes m

more desirable than a potential replacement. The two worst replacement rules for m

are the Outcome-based rule and the Retrospective Satisficing rule, the former more

desirable than the latter due to the wider region of acceptable outcomes.

The panels on the right indicate the expected utility for O under the four rules

given an optimal response by m. The Forward-looking Rational rule is least de-

sirable, as the extreme induced experimentation behavior leads to very little useful

information for the implementation stage. An Outcome-based rule is strictly better

than a Retrospective Satisficing rule, due to the additional flexibility it affords to O.

The two replacement rules that generate the best outcomes for O are Guaranteed

Tenure and Outcome-based replacement.

Figure 3.14 compares the desirability of these two replacement rules as a function

of task complexity. For low complexity tasks, an Outcome-based rule dominates, as

O retains the ability to replace undesirable managers without taking on too great

a risk of being forced to replace a desirable one. In contrast, in high uncertainty

environments, the requirement that O replace m if y0 yp becomes more burden-

some. Finally, once the complexity grows sufficiently large, i.e. �2 � 2y0, Guaranteed

Tenure induces the utility-maximizing xp for O if m were solely motivated by orga-

nizational performance.

Discussion


Figure 3.12: Lower Bound under r = OB

s2

y p⇤

-1.0

0-0

.90

-0.8

0-0

.70 �2 2 (0.1, 6), y0 = 0.5

s2

y p⇤

-2.0

-1.8

-1.6

-1.4

-1.2 �2 2 (0.1, 6), y0 = 1

s2

y p⇤

-4.0

-3.5

-3.0

-2.5

�2 2 (0.1, 6), y0 = 2

s2

y p⇤

-8-7

-6-5

-4 �2 2 (0.1, 6), y0 = 4

0 1 2 3 4 5 6�2


Figure 3.13: Replacement Probabilities and EUO under Different Replacement Rules

s2

Pr(replace|x

⇤ p)

0.0

0.1

0.2

0.3

0.4

0.5 �2 2 (0.1, 6), y0 = 0.5

s2

Pr(replace|x

⇤ p)

0.0

0.1

0.2

0.3

0.4

�2 2 (0.1, 6), y0 = 1

s2

Pr(replace|x

⇤ p)

0.0

0.1

0.2

0.3

0.4 �2 2 (0.1, 6), y0 = 2

s2

Pr(replace|x

⇤ p)

0.00

0.10

0.20

0.30

�2 2 (0.1, 6), y0 = 4

s2

EUO(x⇤ p)

-0.1

5-0

.10

-0.0

50.

00 �2 2 (0.1, 6), y0 = 0.5

s2

EUO(x⇤ p)

-0.1

5-0

.10

-0.0

50.

00 �2 2 (0.1, 6), y0 = 0.5

s2

EUO(x⇤ p)

-0.1

5-0

.10

-0.0

50.

00 �2 2 (0.1, 6), y0 = 0.5

s2

EUO(x⇤ p)

-0.1

5-0

.10

-0.0

50.

00 �2 2 (0.1, 6), y0 = 0.5

r = GT r = RS r = FR r = OB


Figure 3.14: Comparing Guaranteed Tenure and Outcome-based Replacement

s2

EUO(GT)�

EUO(OB)

-0.0

040.

000

0.00

4�2 2 (0.1, 6), y0 = 0.5

s2

EUO(GT)�EUO(OB)

-0.0

20.

000.

010.

02

�2 2 (0.1, 6), y0 = 1

s2

EUO(GT)�

EUO(OB)

-0.0

50.

000.

05

�2 2 (0.1, 6), y0 = 2

s2

EUO(GT)�

EUO(OB)

-0.3

-0.2

-0.1

0.0

0.1 �2 2 (0.1, 6), y0 = 4

0 1 2 3 4 5 6�2


Recent management scholarship emphasizes the desirability of experimental pro-

totyping and evidence-based management as a way to improve organizational per-

formance. While it is true that offline experimentation can help organizations avoid

making strategic errors and also encourage exploratory search, the conflicting moti-

vations of managers can sap offline experiments of their informational value. Career

concerns of managers, in particular, may lead to overly exploitative prototyping

behavior and may lead to over-exploratory prototyping behavior, depending on how

organizations evaluate their managers. When managerial competence or effectiveness

with a task is measured relative to an organization’s prior performance, managers

tend to be overly incremental in their experimentation behavior. In contrast, when

managers are judged based on their capacity to implement a more desirable online

change than an unexperienced replacement, managers tend to be overly exploratory.

I evaluate two mechanisms that organizations can use to ameliorate this agency

problem. Guaranteed tenure alleviates the career concerns of managers by ensuring

that they will keep their jobs notwithstanding the desirability of their experimen-

tal prototype. Guaranteed tenure also requires that organizations sometimes keep

managers who the organization believes to be a poor fit. An outcome-based re-

placement rule induces managers to create the prototype that the organization finds

most desirable, but requires that organizations sometimes replace managers who are

likely to produce high quality organizational outcomes. The paper shows that the

freedom afforded by guaranteed tenure leads to better organizational performance

in high complexity task environments. The organizational control exerted by the

outcome-based rule, in contrast, is more desirable in low uncertainty environments.

Chapter 4

Social Movements as Purveyors of

Legislative Expertise

Introduction

For the last two decades, a major emphasis in sociological research on social

movements has been demonstrating the impact of social movements on the policy

process. Scholars have found empirical support for the power of movements in a

variety of policy domains, including civil rights, women’s rights, environmental pro-

tection, medicine, lesbian and gay rights, and antiwar (Amenta et al., 2010). In

recent years, scholars have added nuance to this basic finding by demonstrating the

stages of the policy process wherein movements matter most (e.g. Soule and King,

2006) and by demonstrating that the effect of movements on the policy process is

often contingent on facets of the political environment (e.g. Soule and Olzak, 2004).

At the core of this literature is the idea that social movements matter to the policy

process because they transmit valuable information to legislators (Gillion, 2013).

For example, large public protests send a signal to lawmakers about the interests

96

CHAPTER 4. SOCIAL MOVEMENTS AND LEGISLATIVE EXPERTISE 97

and level of commitment of constituents. Similarly, letter-writing campaigns can

demonstrate the breadth of support for a previously overlooked legislative matter

(e.g. Gillion, 2013; Lohmann, 1993). By providing informative signals to lawmakers

about constituent preferences and priorities, the activities of social movements can

move certain issues up the legislative agenda and create opportunities for novel issues

to gain legislative attention (Soule and King, 2006). Baumgartner and Jones (2015)

describe this type of information as entropic information, because it helps to bring

order to the chaotic set of problems and interests that make up the legislative process.

Baumgartner and Jones (2015) also identify a second class of information that

is equally important to the legislative process, but which has received considerably

less attention in the social movements literature: expert information. While entropic

information helps lawmakers decide what issues are important to their constituents,

expert information helps legislators design policies that achieve desired goals. Once

a problem has been identified and the legislature sets out to write new legislation,

lawmakers rely on organizations and individuals with expertise to help them craft an

appropriate solution (Bendor and Meirowitz, 2004; Callander, 2008; Esterling, 2004).

External experts also lend credibility to the policy process by making legislative

outcomes appear to be the product of rational deliberation (Epstein, 1996; Nelkin,

1975).

Social movements and, in particular, social movement organizations (SMOs), have

increasingly become sources of policy expertise, particularly for legislative issues re-

lated to the environment and public health (Andrews, 2006; Bosso, 2005; Burstein

and Hirsh, 2007; Epstein, 1996; Shaiko, 1999). Their expertise gains them access

to and influence over the policy process. Congressional committees often design

consequential, highly technical legislation with short timelines for consideration and

limited internal subject-matter expertise (Hall and Deardorff, 2006; Weiss, 1992).


As a result, SMOs, think tanks, and other interest organizations are invited into the

policy process in order to help lawmakers with legislative research, policy communi-

cation, and even drafting legislative language (Baumgartner and Leech, 1998; Berry,

1977; Heinz et al., 1993; Morales, 2013; Schlozman and Tierney, 1986). In addition

to extensive work behind-the-scenes, SMOs and other interest organizations are also

frequently invited to publicly testify before Congress on pending pieces of legislation

(Andrews and Edwards, 2004; Burstein and Hirsh, 2007).

Entropic information and expert information are relevant at different moments in

the legislative process (Baumgartner and Jones, 2015). Entropic information is most

important when the legislature is searching for new items to place on the agenda

or prioritizing the pressing matters before the Congress. Organizations that can

convincingly signal the breadth and strength of constituent preferences, therefore,

should have outsized access to the policy process early in the agenda-setting phase

(Baumgartner and Jones, 2009; Kingdon, 1984). Expert information, in contrast,

is valuable once the Congress has focused on a particular legislative matter and is

in the process of selecting a policy instrument. For legislative issues nearing votes

in committee or on the floor of the House or Senate, therefore, organizations with

expertise should have privileged access to the policy process.

In our paper, we examine patterns of access to the legislative process for en-

vironmental social movement organizations (EMOs). In addition to engaging in

outsider activism, the environmental movement has actively sought to legitimate its

policy claims and influence skeptical policymakers through the development of scien-

tific knowledge (Andrews and Edwards, 2004; Andrews, 2006; Hadden, 2015; Rome,

2013). Further, the uncertainty and complexity associated with environmental poli-

cymaking creates opportunities for EMOs to use their internal expertise in order to

exercise informational influence on the policy process (Allen et al., 2001; Eden, 1998;


Farber, 2003; Stone, 1993). As a result, the environmental movement is a valuable

place to begin an examination of how SMOs utilize scientific expertise in order to

gain access to the policy process.

In our study, we use invitations to testify before Congress as an indication that

an EMO has been given access to the policy process. We then use the characteristics

of the hearings to which the organizations are invited in order to learn when in the

policy process EMOs are most influential. Finally, we compare the observed patterns

of EMO testimony against what we observe for other interest organizations that lack

environmental expertise.

We find that EMOs have privileged access to the legislative process when com-

mittees are considering specific pieces of environmental legislation. EMOs and non-

environmental interest groups are invited to testify before Congress in roughly equal

numbers during exploratory, investigatory, or oversight hearings. However, EMO

representation at congressional hearings increases once a congressional committee is

in deliberations over legislative language, and by a significantly greater amount than

non-environmental interest organizations.

We conclude that the role of SMOs as providers of expert information to the pol-

icy process is important, but under-examined in the social movements literature. The

provision of expert information is a key avenue through which SMOs can generate

access to elite actors, which has previously been identified as important to successful

policy outcomes (Amenta et al., 1992, 1994; Cress and Snow, 2000; Soule and King,

2006). Further, we demonstrate how SMOs with expertise have the opportunity to

influence legislation at a fairly advanced stage in the policy process (King et al.,

2007, 2005; Olzak and Soule, 2009; Soule and King, 2006).

Theory and Hypotheses


The existing literature identifies two primary mechanisms for how social move-

ments impact the policy process (see e.g. Andrews and Edwards, 2004). The first

points to the signals that protests send to legislators and other important interests.

The second centers on the way social movement organizations gain access to and

allies within the legislature. Our paper focuses primarily on the second of these

mechanisms, but we briefly describe the first.

Protest Signals and Legislative Response

Many researchers argue that social movements impact legislators because they

send important informational signals about constituents’ views on issues (e.g. An-

drews and Edwards, 2004; Gillion, 2012, 2013; Lohmann, 1993). As Lohmann (1993,

p. 320) notes, protest “informs the political leader’s decisions and thus affects policy

outcomes.” This idea is embedded in the democratic theory of politics, which asserts

that electoral incentives align the legislative decisions of officeholders with the pref-

erences of the majority of their constituency (e.g. Downs, 1957; Erikson et al., 2002;

Page and Shapiro, 1983; Soule and Olzak, 2004). The activities of social movements,

much like public opinion polls, serve as a gauge of constituents’ preferences on a

given issue.

Much of the sociological work on signaling focuses on public protest, rather than

other activities that social movements engage in, because public protest is one of the

most visible activities social movements undertake (Kollman, 1998; Lipsky, 1968).

This existing work examines several facets of protest that make it more worthy of

congressional attention, such as the frequency or the size of protest events (e.g.

Agnone, 2007; Burstein and Freudenburg, 1978; DeNardo, 1985; King et al., 2007;

Lohmann, 1993; McAdam and Su, 2002; Olzak and Soule, 2009; Santoro, 2002, 2008).

To the extent that the signal sent by protest is perceived to reflect the interests of


the voting public, legislators will react with policy decisions that are in line with

the protest’s demands (Gillion, 2013). In addition to the direct impact of protest

on legislators, the literature also argues that protest sends signals to bystander con-

stituents, who may either join the movement or pressure legislators in other ways

(Andrews and Edwards, 2004; Giugni, 2004; Jenkins, 1987; Lipsky, 1968). Protest is

powerful because it can threaten legislators directly, or do so indirectly by its impact

on the opinions, values, and beliefs of other important constituencies.

Social Movement Organizations and Legislative Response

In addition to the signaling mechanism of protest, some movement scholars argue

that another important mechanism of movement legislative influence is via social

movement organizations’ ability to gain access to the legislative process or mobilize

allies in Congress (Amenta et al., 2010; Andrews and Edwards, 2004; Kane, 2003;

King et al., 2007; McCammon et al., 2001; Olzak and Soule, 2009; Soule et al., 1999;

Soule, 2004; Soule and Olzak, 2004). Elite allies, in turn, can act as a mouthpiece for

a social movement within the government, provide an avenue for social movement

organizations to get access to the legislative process, and ensure that the issues pro-

moted by the social movement receive favorable placement on the legislative agenda

(Amenta et al., 1992, 1994; Cress and Snow, 2000; Soule and King, 2006).

The empirical literature on the impact of movement organizations on legislators

has established a correlation between the density (i.e., number) of social movement

organizations in a given location (i.e., city, state, or nation) and the passage of leg-

islation favorable to the movement in question. For example, Soule et al. (1999)

find a correlation between the number of feminist SMOs and roll call votes and con-

gressional hearings on women’s issues (for similar findings, see Johnson et al., 2010;


Kane, 2003; Olzak and Soule, 2009; Soule, 2004). This literature has also estab-

lished a correlation between the use of certain tactics mainly associated with social

movement organizations and policymaking around the movement in question. For

example, Soule et al. (1999) find a correlation between the mention of letter-writing,

petitions, and lobbying in news accounts of feminist protest and congressional hear-

ings and roll call votes on women’s issues (for similar findings, see McCammon et al.,

2001; Olzak and Soule, 2009).

This empirical work is important because it suggests that social movement or-

ganizations likely influence legislation in some way. However, the literature thus far

lacks a nuanced understanding of what exactly it is that movement organizations do

to produce these responses from policymakers. We argue that providing information

and expertise to legislators is an important but largely underexamined mechanism

of movement influence on policy (Burstein and Hirsh, 2007, is an important excep-

tion). Like other actors in the “policy community or subsystem” of government, such

as bureaucrats, researchers, and other interest groups (Sabatier, 1991), social move-

ment organizations help to draft legislation, testify at hearings, and present research

(e.g. Berry, 1977; Heinz et al., 1993; Schlozman and Tierney, 1986). We develop this

argument in the next section.

Expert Information and the Legislative Process

There are two kinds of information that are relevant to the impact of social

movements on policy. The first, entropic information, is critical to the search stage

of policy, when legislators are deciding the issues upon which to focus their attention

(Baumgartner and Jones, 2015). The second kind of information relevant to the

impact of movements on policy, and the focus of this paper, is expert information.

The policy outcomes of proposed bills are often difficult to predict beforehand,


and may interact with existing legislation in unexpected ways. Thus, legislators

rely on expert information in order to ensure that the policy outcomes of a piece of

legislation reflect the intended policy goals (Baumgartner and Jones, 2015; Callander,

2008; Esterling, 2004). Furthermore, gaining public support for legislative initiatives

requires convincing the public that a new bill will benefit their interests. Legislation

that is well justified at the outset can calm the concerns of a mistrustful electorate

(Cobb and Kuklinski, 1997; Esterling, 2004). Finally, a norm of scientific rationalism

pervades legislative debate. Legislators and the public often insist that a policy has

a scientifically supported basis for implementation (Ezrahi, 1990; Habermas, 1971;

Nelkin, 1975).

In order to write legislation that reflects the policy goals of informed lawmakers,

the Congress created the committee system in order to ensure that Members develop

expertise on issues of interest. Each committee is endowed with exclusive jurisdic-

tion over a piece of the legislative agenda (King, 1994, 1997; Krehbiel, 1991; Smith

and Deering, 1997). Committee members are expected to become resident experts

in their assigned areas of jurisdiction (Cooper, 1970; Gilligan and Krehbiel, 1987,

1990; Krehbiel, 1991, 2004). As a result, the task of writing important legislation is

commonly delegated to committee leadership in conjunction with other committee

members. Further, if a new piece of legislation originates outside of the committee

with jurisdiction, the congressional leadership refers it to the relevant committee.

Committees then have the option to edit the legislation and report it back to the

full legislative body for consideration. Committee specialization gives the legislature

confidence that the lawmakers who are perceived to be most knowledgeable have first

reviewed all the bills that receive consideration by the full House or Senate (Krehbiel,

1991; Shepsle and Weingast, 1987; Smith and Deering, 1997).


Committee members and their staffs rely on other entities for information on is-

sues that require significant scientific or technical expertise. Despite the considerable

resources that exist in the executive agencies, the legislative branch often chooses in-

stead to seek expertise elsewhere in order to protect its independence and circumvent

potential political conflicts between the legislative and executive branches (Brudnick,

2011; Weiss, 1992). For some policy questions, committees can find sufficient exper-

tise within the legislative branch, in particular through the Congressional Research

Service, which acts as an internal think-tank for legislators and their staffs. How-

ever, committees work on short and unpredictable time schedules. The 600-person

Congressional Research Service is often overtaxed with requests and lacks necessary

specialized knowledge to answer many questions (Kosar, 2015). As a result, commit-

tees often look beyond the legislative branch for policy expertise (Hall and Deardorff,

2006; Hansen, 1991; Wright, 1995).

Interest organizations, think tanks, and advocacy groups make themselves avail-

able to assist Members of Congress and their staffs in need of answers to important

legislative questions (Wright, 1995). These extra-governmental organizations aspire

to become trusted resources for legislative committees in order to gain influence over

the design of policy proposals. Legislators know that the information offered by in-

terest organizations comes from a biased perspective. But, for interest organizations

with policy goals that are sufficiently aligned with the lawmaker’s and for issues

around which there is significant policy uncertainty, legislators are willing to rely on

this information in order to ameliorate the risk of unintended outcomes. This legisla-

tive bias-variance trade off leads to policy outcomes that are better aligned with the

preferences of the interest organization than would be predicted by the preferences

of the median legislator (Bendor and Meirowitz, 2004; Callander, 2008; Gilligan and

Krehbiel, 1987; Krehbiel, 1991; Wright, 1995).


Social movement organizations are well suited to provide expert information and

take advantage of informational influence. Social movements have long relied on sci-

entific knowledge as a tool to support their organizational goals (Brown et al., 2004;

Epstein, 1996; Ezrahi, 1990; Habermas, 1971; Johnson, 2006; Rome, 2013). Social

movements often have mature networks of experts willing to aid in their causes.

The environmental movement has been particularly effective at engaging environ-

mental scientists in order to justify legislative efforts on species protection, forest

preservation, and emissions reduction (Bosso, 2005; Shaiko, 1999). Recent move-

ments have also demonstrated the ability to work within the scientific community

to generate new scientific knowledge in order to justify claims or refute opponents’

claims (Brown et al., 2004; Epstein, 1996; Jamison, 2001; Rome, 2013). Further,

the public views social movement organizations differently than corporate lobbyists

and business groups. Social movement organizations are perceived as more trust-

worthy advocates for the public interest than corporations or, for that matter, the

government (Edelman, 2015; Lyon and Maxwell, 2008; Soule, 2009). Therefore, by

partnering with a social movement organization, legislators may be able to secure

broader political support for a legislative initiative.

Congressional Committees and Information Collection

Empirical examination of which organizations provide expert information to pol-

icymakers and committee staffs is complicated by the informal nature of the rela-

tionships between the Congress and interest organizations. Private meetings, phone

calls, and e-mail exchanges between Members, their staffs, and representatives of in-

terested parties occur daily. Because the Congress is exempt from keeping workplace

records and Freedom of Information Act requests, systematic data on expert access

is challenging to collect.


However, interviews by the authors indicate that examining which organizations

testify at congressional hearings provides an accurate picture of the organizations

that are most influential on a legislative topic. Staff members often invite testi-

mony from the organizations that they have worked with in private. Furthermore,

staff preparation for a hearing often involves lengthy conversations with prospective

witnesses. According to an employee at a prominent environmental advocacy orga-

nization: “If you’ve been working closely with [a committee] on legislation and then

[the committee has] a hearing related to it, they’re going to want to bring the people

that know the most about the issue they’re working on.”

Hearings provide an institutionalized, public avenue for committees to seek expert

information. The Congress holds thousands of public hearings every year. A hearing

can include as little as a single witness and as many as a few dozen witnesses. In

addition to movement organizations, other interest organizations are commonly in-

vited to testify (e.g., labor unions, trade associations and industry groups). Affected

parties may be called in order to offer a real-world perspective or to demonstrate how

a proposed legislative change would impact a key constituency. For technical or com-

plex policy questions, academic researchers may be called to testify, as well. Finally,

it is not unusual for other government officials to testify at hearings, particularly

when the topic under consideration cuts across jurisdictional lines (Baumgartner

and Jones, 2015; Burstein and Hirsh, 2007; Leyden, 1995; Talbert et al., 1995).

Hearings are one part theatrical puffery and one part legislative substance. Hear-

ings serve as an opportunity for lawmakers to get media attention and to demonstrate

issue knowledge to constituents. Hearings are carefully designed to elicit testimony

to support the interests of the majority party. Question-and-answer sessions are of-

ten scripted, with each Member assigned particular questions to elicit predetermined

responses (Leyden, 1995; Oleszek, 1989). Yet, the ceremonial aspects of hearings


should not belie their legislative role. The process of hearing preparation provides

deadlines for staff and substantive education for lawmakers. Organizations invited

to testify often create detailed reports and white papers to support submitted testi-

mony. Finally, although some hearings receive a great deal of media attention, and a

few are even televised live, even lightly publicized hearings are often important events

for affected constituencies (Baumgartner and Jones, 2015; Burstein and Hirsh, 2007;

Diermeier and Feddersen, 2000; Oleszek, 1989).

The content of a particular hearing provides information about the stage in the

legislative process when an organization is given access. There are two broad classes

of congressional hearings: non-legislative and legislative. Non-legislative hearings are

not tied to a specific bill and may focus on virtually any topic deemed important by

a committee (Talbert et al., 1995). Two common types of non-legislative hearings

are investigatory and oversight hearings, which usually examine the efficacy of a

federal program or potential wrongdoing by a private company or government official.

Non-legislative hearings can also be called for the purpose of raising the profile of

a legislative issue, to demonstrate responsiveness to important constituents, or to

publicly claim jurisdiction over an issue or set of issues. The decision to hold a non-

legislative hearing indicates that an issue is in a preliminary stage in the legislative

process.

Legislative hearings, in contrast, are tied to a specific bill that may become law.

Legislative hearings are called after a bill has been introduced by a Member and

referred to a committee (Baumgartner and Jones, 2015; Talbert et al., 1995). The

decision to hold a legislative hearing indicates that a committee is seriously consid-

ering referring a bill to the full legislative body. A committee will hold legislative

hearings just prior to the bill “markup.” During markup, the committee members

submit amendments to a bill under consideration and then vote on whether to send


the amended bill back to the full legislature (Koempel and Schneider, 2007; Sachs,

2003; Smith and Deering, 1997).

Hypotheses

Given the above discussion, we develop two core hypotheses about EMO access

to the legislative process. First, we predict that EMOs will have greater access to

the legislative process as a bill reaches a more advanced stage within the committee.

During the search phase of the policy process, the committee seeks information from

a broad set of interests with diverse areas of knowledge. In contrast, once a com-

mittee has decided to put forth new legislation to the full Congress, the committee

seeks specialists with subject area expertise. Therefore, we expect that the number

of EMOs invited to testify at hearings that are explicitly tied to the design of a piece

of environmental protection legislation will exceed the number of EMOs in hearings

related to environmental protection that are non-legislative.

H1: More EMOs are invited to testify at legislative hearings than non-legislative hear-

ings.

We would not predict this same pattern to hold for interest organizations that

lack environmental expertise. During the search phase, we expect the legislature to

invite other interest organizations with strong preferences on environmental protec-

tion to testify alongside EMOs. Once a committee’s agenda is determined and a

particular piece of legislation is under consideration, we expect that the committee

will be more discerning. Only those organizations with specific expertise and, likely,

with whom the committee has a prior relationship, will be invited to testify. As a

result, we expect that the increase in representation for EMOs will exceed the change


in representation for non-environmental interest organizations.

H2: The increase in EMO representation will exceed the increase in representation of

non-environmental interest organizations across non-legislative and legislative hear-

ings.

Empirical Setting: The Environmental Movement in the United States

We examine the participation of environmental movement organizations in envi-

ronmental protection hearings. We collect data on congressional hearings related to

species and forestry protection from 1988 to 2007. We choose species and forestry

issues because these are traditionally at the nexus of environmental activism and

environmental science.

Data

We collect data on every congressional hearing held on “species and forest pro-

tection” in the U.S. House of Representatives over 10 Congresses from 1988 (100th)

through 2007 (109th). We use the Policy Agendas Project coding scheme to identify

the relevant hearings. From these data on hearings, we code several relevant pieces of

information. First, we utilize the classification scheme created by the Policy Agendas

Project to determine whether a hearing was legislative or non-legislative in nature.

The Policy Agendas Project classifies hearings as legislative if they mention a bill

number or name explicitly in the title of the hearing. The hearing is classified as

non-legislative if there is no bill number or bill name indicated (Talbert et al., 1995).

In all, there were 276 hearings during this period. 122 were legislative and 154 were

non-legislative.

Next, we examine the organizational affiliations of each of the witnesses invited


to testify at each of these 276 hearings. We collected and transcribed the organiza-

tional affiliation of each witness using the ProQuest Congressional Hearing Digital

Collection. In all, we obtained a list of 2,611 witnesses at the 276 environmental

hearings in this time period. These witnesses were affiliated with 1,019 different

organizations. Once we had this list of 1,019 organizations, we coded each by orga-

nizational type. We researched each organization using the organization’s website

and/or in other public materials about the organization (e.g., newspaper articles,

research articles, etc.). For defunct organizations, we also used the Way Back Ma-

chine (http://archive.org/web/) to find archived websites. We found information on

950 of the organizations that testified. We created three organizational categories:

(1) environmental movement organization (n=153); (2) non-environmental interest

organization (n=182); (3) not interest organization (n=615).

We coded an organization as an EMO if on its website or other public materials

it was stated that the organization’s primary purpose is environmental protection

and indicates that the organization publicly agitates for environmental protection.

We did not distinguish between organizations based on particular policy positions.

Most supported environmental conservation for its own end (e.g., Wilderness Society,

Sierra Club). Some, however, supported preservation for the purposes of recreation,

hunting, or fishing (e.g., U.S. Sportsmen’s Alliance, Safari Club International). We

categorized regional chapters of organizations by their national organizing group

(e.g., the Hawaii Audubon Society was grouped with the National Audubon Society).

We identified 153 of the organizations as EMOs.

We coded an organization as a non-environmental interest organization if the or-

ganization’s goals were not primarily environmental and a stated method for achiev-

ing its organizational goals was government influence (see Burstein, 1998; Walker,


1991; Wilson, 1973). Labor unions, professional associations, and local civic organi-

zations were coded as non-environmental interest organizations. Although the prefer-

ences of a non-environmental interest organization may align with the environmental

movement on a particular legislative issue, we attempted to categorize organizations

based on their stated organizational mission. Therefore, all organizations that ex-

plicitly represent commercial hunters, commercial fishermen, forest industries, and

companies that manufacture or sell recreational products were categorized in this

group. Organizations that publicly supported environmental protection, but whose

membership consisted exclusively of industry interests were grouped in this category

as well. We identified 182 of the organizations as non-environmental interest groups.

All other organizations were classified as “not interest organization.” These in-

clude government and quasi-governmental organizations, think tanks, for-profit cor-

porations, and universities. Zoos, aquariums, and circuses were coded in this cat-

egory, as well. (However, organizations involved in animal conservation that also

manage zoos or aquariums, such as the Wildlife Conservation Society, were coded as

EMOs). The remaining 615 organizations were classified to be in this category.

We also collect a series of variables about hearings that we use as controls in our

regression analyses. These include the committee that is holding the hearing, the

two-year Congress in which the hearing takes place, and the number of sessions in

the hearing. Consistent with the Policy Agendas Project, we also note whether the

hearing was related to appropriations, the creation of a new government agency, or

legislation proposed by the President. We control for the length of the hearing in case

legislative hearings and non-legislative hearings tend to have a different number of

witnesses. We control for the committee holding the hearing, the Congress in which

the hearing takes place, and the type of hearing to ensure that our estimates are not

being driven by a single committee, single Congress, or single type of hearing. Note


that by controlling for the two-year Congress, we implicitly also control for partisan

control of the Congress.

Analysis and Results

We begin by examining whether more EMOs are invited to testify in legislative

hearings than non-legislative hearings. We find that legislative hearings do tend

to include more EMOs than non-legislative hearings. The average non-legislative

hearing includes testimony from 1.21 EMOs, compared with 1.95 EMO participants

at legislative hearings. A Welch Two-Sample T-Test indicates that this difference is

significant at the 99% level.

We test this relationship using a series of regression analyses, which allow us to

control for potential alternative explanations. We use negative binomial regression,

which is common for modeling count data in the presence of overdispersion. We

estimate the model using the “glm.nb” command in R 3.1.0.

The results are presented in Table 4.1. The first model predicts the number of

EMOs invited to testify at each hearing, after controlling for the two-year Congress,

the length of the hearing, the congressional committee that called the hearing,

whether the hearing was related to appropriations, the creation of a new agency,

or in response to a Presidential proposal. We then add an indicator variable for

whether the hearing was legislative in nature. The coefficient on this variable is

positive and significant, which is consistent with our theory.


Table 4.1: Negative Binomial Regression Predicting the Count of EMOsTestifying at a Hearing

Dependent variable:EMO Testimony Count(1) (2)

Intercept �0.460 �0.710⇤(0.362) (0.364)

Legislative Hearing 0.391⇤⇤⇤(0.121)

Number of Sessions 0.910⇤⇤⇤ 0.880⇤⇤⇤(0.121) (0.120)

Agriculture Committee �0.087 �0.073(0.264) (0.258)

Resources Committee 0.124 0.116(0.201) (0.196)

Merchant Marines and Fisheries Committee 0.012 0.033(0.228) (0.222)

Appropriations Hearing �0.481 �0.516⇤(0.316) (0.313)

Agency Creation Hearing 0.554⇤ 0.381(0.300) (0.296)

Presidential Proposal Hearing �0.012 0.196(0.648) (0.642)

Congress Dummies Yes Yes

Observations 276 276

Note: ⇤p<0.1; ⇤⇤p<0.05; ⇤⇤⇤p<0.01


Our analysis supports the contention that SMOs are an important source of

expert information for congressional committees. The decision to invite EMOs in

larger numbers to legislative hearings indicates that EMO access increases once a

committee has determined its priorities and begins to craft a piece of legislation.

We then test whether the increased representation at legislative hearings is unique

to EMOs. Our null hypothesis is that all interest groups receive greater representa-

tion at legislative hearings. 1.27 non-environmental interest organizations are invited

to testify at the average legislative hearing, compared with 1.20 at a non-legislative

hearing. This is not a statistically significant difference. Figure 4.1 illustrates the

means and 95 percent confidence intervals for EMO and non-environmental interest

organization testimony count across legislative and non-legislative hearings.

We examine the same models for the non-environmental interest organizations as

we did for the EMOs. As before, we estimate a negative binomial regression model

with Congress and hearing controls. The results are presented in Table 4.2. We

do not find a significant difference in representation for non-environmental interest

organizations across non-legislative and legislative hearings.

In order to test the hypothesis that the difference between EMO participation at

legislative and non-legislative hearings exceeds the difference for non-environmental

interest organizations, we compare the difference between the differences using a

Wald test and find that the difference is significant at the 1 percent level. We

also examine the relationship using Seemingly Unrelated Regression, which allows

us to efficiently compare the effect size of nested models with common independent

variables, but different dependent variables. We estimate the model is using the

‘systemfit’ package in R 3.1.0. The results are presented in Table 4.3.


Figure 4.1: Comparison of Means for EMOsand Non-environmental Interest Organizations

Cou

ntof

Test

imon

ype

rH

earin

gby

Org

Typ

e

0.5

1.0

1.5

2.0

2.5

Means and 95% Confidence Intervals

EMOs Non-EMO IOs

Non-LegislativeLegislative Non-

LegislativeLegislative

(a)

(b)

(c)(d)

Pr(a=b) < 0.001 Pr(c=d) = 0.734


Table 4.2: Negative Binomial Regression Predicting the Count of Non-environmentalInterest Organizations Testifying at a Hearing

Dependent variable:Non-environmental IO Testimony Count

(1) (2)

Intercept �1.987⇤⇤⇤ �2.017⇤⇤⇤(0.502) (0.512)

Legislative Hearing 0.049(0.154)


Agriculture Committee 0.793⇤⇤ 0.796⇤⇤(0.336) (0.336)


Merchant Marines and Fisheries Committee 0.955⇤⇤⇤ 0.957⇤⇤⇤(0.309) (0.309)

Appropriations Hearing �0.229 �0.241(0.363) (0.365)

Agency Creation Hearing �0.129 �0.155(0.450) (0.456)

Presidential Proposal Hearing �1.288 �1.261(1.089) (1.092)



Note: ⇤p<0.1; ⇤⇤p<0.05; ⇤⇤⇤p<0.01


Table 4.3: Seemingly Unrelated Regression Predicting the Count of EMOs andNon-environmental IOs Testifying at a Hearing

Dependent variable:EMO Count Non-EMO IO Count

(1) (2)

Intercept 0.236 �1.092(0.561) (0.565)

Legislative Hearing 0.530 �0.030(0.197)⇤⇤ (0.199)


Agriculture Committee �0.224 0.938⇤(0.439) (0.443)


Merchant Marines and Fisheries Committee �0.263 0.915⇤(0.360) (0.363)

Appropriations Hearing �0.363 0.098(0.378) (0.382)

Agency Creation Hearing 0.866 �0.205(0.567) (0.572)

Presidential Proposal Hearing 0.363 �0.701(0.781) (0.787)



Note: ⇤p<0.1; ⇤⇤p<0.05; ⇤⇤⇤p<0.01


We then test the linear restriction that the coefficients on “Legislative Hearing”

are different across the nested models using a chi-squared test. We use the ‘linearHy-

pothesis’ command in the ‘car’ package in R 3.1.0. We find that the difference in

differences is significant at the 5 percent level.

These results provide additional evidence that SMOs are different from other

types of interest groups. Although EMOs and non-environmental interest organiza-

tions are invited in roughly equal numbers to non-legislative hearings, the increased

reliance on EMOs for expertise during legislative hearings is unique.

Conclusion and Areas for Further Research

Social movements attempting to impact the policy process seek access to elite

government actors. One way to get lawmakers to pay attention to social movement

goals is by transmitting entropic information, which is the type of information that

helps lawmakers determine their legislative priorities during each term in office. By

convincing elected officials that the movement speaks for a large and dedicated group

of voters, social protests and other forms of outsider activism are effective strategies

for utilizing the power of entropic information to increase the political influence of a

social movement.

However, social movements and SMOs also gain access to the policy process

through their organizational expertise. Lawmakers trade off a desire to achieve policy

goals with a fear of unintended consequences. Interest organizations with policy

expertise, as a result, are given privileged access to the policy process when legislation

is being crafted and debated in the legislature. Interest organizations with policy

expertise can then use this access to influence legislative outcomes. Due to their

high level of public trust and varied sources of technical knowledge, SMOs working

in scientific domains are particularly well suited to exploit expertise-based influence


on the policy process.

In our paper, we examine the patterns of EMO access to the policy process. Con-

sistent with theories of expertise-based influence, we find that EMOs get privileged

access to the legislature when congressional committees are debating specific pieces

of legislation. Further, this enhanced level of access is unique to EMOs; we do not

observe the same patterns for other interest organizations with strong opinions on

environmental regulation.

However, a crucial limitation of our paper is that we can only demonstrate leg-

islative access, not policy influence. In future work, we seek to tie expertise-based

access to concrete changes in legislative outcomes, which has been the strength of

prior work on the information-based political strategies of SMOs. We hope that

our analysis acts as a springboard for future studies that show the impact of SMO

expertise on policy outcomes, test the comparative effectiveness of expertise-based

strategies, and examine how social movements balance entropy-based and expertise-

based informational strategies to achieve their policy goals.

Appendix A

Appendix to Chapter 2

Proof of Proposition 2

I find the solution to the model using backwards induction. First, I determine

the optimal strategies for P and V conditional on reaching the proposal subgame.

Then, I determine the optimal strategies in the learning subgame, given subsequent

equilibrium strategies.

Proposal Subgame Given p(p) Not Learned

I first find the outcome of the proposal subgame, assuming that in the learning

subgame either LP= ¬L, or LP

= L and LV= ¬A.

Define the acceptance set for V , Av(p), as the set of alternatives which would be

accepted by V conditional on p. Spatially, Av(p), is the set of points in R which are

closer to 1 than p:

Av(p) =

8><

>:

[p, 2� p] if p 1

[2� p, p] if p > 1

120

APPENDIX A. APPENDIX TO CHAPTER 2 121

Define a as the alternative in Av(p) that is optimal for P to propose. If Eup(a) �

Eup(p), then PP⇤¬L = a. Otherwise, PP⇤

¬L = p.

PP⇤¬L =

8>>>>>>>><

>>>>>>>>:

a = 0 if p 0

p if 0 < p 1

a = 2� p if 1 < p 2

a = 0 if p > 2

Because the proposals are all contained in the acceptance set, Av(p), V will accept

a in equilibrium.

PV ⇤¬L = a

The expected utility for P , conditional on p, ↵, and �, and either LP= ¬L, or

LP= L and LV

= ¬A:

If PP⇤¬L is a = 0:

Eup = � [Pr( a(a) 0) |E( a(a)| a(a) 0)|+ Pr( a(a) > 0) |E( a(a)| a(a) > 0)|] =

�h12 ·

↵2� +

12 ·

↵2�

i= � ↵

2�

If PP⇤¬L = p and p 2 (0,↵]:

Eup = � [Pr( p(p) 0) |E( p(p)| p(p) 0)|+ Pr( p(p) > 0) |E( p(p)| p(p) > 0)|] =

�h⇣

�(p�↵)2↵ · p�↵

2

⌘+

�p+↵2↵ · p+↵

2

�i= �p2+↵2

2↵

If PP⇤¬L = p and p 2 (↵, 1]: Eup = �p.

If PP⇤¬L is a 2 (↵, 1): Eup = �a.

If PP⇤¬L is a 2 (0,↵]: Eup = �a+↵

2↵ .

Proposal Subgame Given p(p) Learned


Next, I find the solution to the proposal subgame, assuming that LP= L and

LV= A:

V defines its acceptance set with respect to p(p) in the same way as it defined

Av(p). However, now the acceptance set is defined with respect to the learned

p(p). As a result, for p(p) near 1, there do not exist any values of a such that

Euv(a) > Euv( p(p)). There are also two sets of intermediate values, where V ’s

preference for a known p(p) over an uncertain a makes the range of Av( p(p))

smaller than the range of Av(p).

For sufficiently low p(p), V is indifferent between p(p) and a = p(p), because

p+ ↵�< 1. I find the maximum value for which V is indifferent:

p(p) +↵

�= 1

p(p) = 1� ↵

�

Therefore, for all p(p) 1� ↵�, V will accept p(p) � a � 1. By the symmetry

of V ’s utility curve, V will also accept 1 < a 2� p(p). Therefore, if p(p) 1� ↵�:

Av( p(p)) = [ p(p), 2� p(p)].

For 1 � ↵�< p(p) < 1, V prefers a known p(p) to an uncertain a. In order to

define the indifference set, I need to find the minimum and maximum a such that V


is indifferent. First, I find the minimum a that satisfies:

Euv( p(p)) = Euv(a)

�(1� p(p)) = � [Pr( a(a) 1)E( a(a)| a(a) 1) + Pr(a > 1) · E( a(a)| a(a) > 1)]

1� p(p) =

1� (a� ↵

�)

2

↵�

·✓1�

1 + a� ↵�

2

◆!+

a+ ↵

�� 1

2

↵�

·✓a+ ↵

�+ 1

2

� 1

◆!

=

(a� 1)

2+ (

↵�)

2

2

↵�

a = 1�r

2(1� p(p))↵

�� (

↵

�)

2

By the same logic, the maximum a that satisfies the equality is 1+q2(1� p(p))

↵�� (

↵�)

2.

I find the minimum of the range of values of p(p) for which there does not exist

an a that V will accept by finding the values of p(p) for which the set defined above

is empty:

Find p(p) that satisfies: 1�q

2(1� p(p))↵�� (

↵�)

2= 1+

q2(1� p(p))

↵�� (

↵�)

2.q2(1� p(p))

↵�� (

↵�)

2= 0. p(p) = 1� ↵

2� .

Therefore, if p(p) 2 (1� ↵�, 1� ↵

2� ]:

Av( p(p)) = [1�q

2(1� p(p))↵�� (

↵�)

2, 1 +q2(1� p(p))

↵�� (

↵�)

2].

Further, if p(p) 2 (1� ↵2� , 1]: Av( p(p)) = ;.

By the symmetry of V ’s utility curve, I define the acceptance sets for other


possible values of p(p).

Av( p(p)) =

8>>>>>>>>>>>><

>>>>>>>>>>>>:

[ p(p), 2� p(p)] if p(p) 1� ↵�

h1�

q(1� p(p))2

↵�� (

↵�)

2, 1 +q

(1� p(p))2↵�� (

↵�)

2i

if 1� ↵�< p(p) 1� ↵

2�

; if 1� ↵2� < p(p) 1 +

↵2�

h1�

q( p(p)� 1)2

↵�� (

↵�)

2, 1 +q

( p(p)� 1)2

↵�� (

↵�)

2i

if 1 + ↵2� < p(p) 1 +

↵2�

[2� p(p), p(p)] if p(p) > 1 +

↵�

As before, define a as the a 2 Av( p(p)) that is optimal for P to propose. If

Eup(a) � Eup( p(p)), then PP⇤L = a. Otherwise, PP⇤

L = p.

PP⇤L =

8>>>>>>>>>>>><

>>>>>>>>>>>>:

a = 0 if p(p) � ↵2�

p if � ↵2� < p(p) < 1 +

↵2�

a = 1�q

( p(p)� 1)2

↵�� (

↵�)

2 if 1 + ↵2� < p(p) 1 +

↵�

a = 2� p(p) if 1 + ↵�< p(p) 2

a = 0 if p(p) > 2

PV ⇤L = a

Because the proposals are all contained in the acceptance set, Av( p(p)), V will

accept in equilibrium.

The expected utility for P , conditional on p(p), ↵ and �, and LP= L and

LV= A:


Eup( p(p)) =

8>>>>>>>>>>>><

>>>>>>>>>>>>:

� ↵2� if p(p) � ↵

2�

� p(p) if � ↵2� < p(p) < 1 +

↵2�

�a if 1 + ↵2� < p(p) 2� ↵

�

�a+↵

�

2↵

�

if 1 + ↵�< p(p) 2

� ↵2� if p(p) > 2

Learning Subgame

In order to determine the optimal strategies for P in the learning subgame, I

first calculate the expected utility of P conditional on learning prior to p(p) being

learned. Let Eup(p, L) denote the expected utility of P given p conditional on p(p)

being learned and Eup(p,¬L) denote the expected utility of P given p conditional

on p(p) not being learned. Euv(p, L) and Euv(p,¬L) are similarly defined.

I partition the possible values for p 2 R into six segments, ⇢j(R), j 2 {1, 2, 3, 4, 5, 6},

• In ⇢1, p(p) � ↵2� . PP⇤

L is a = 0.

• In ⇢2, �↵�< p(p) 1 +

↵2� . PP⇤

L = p(p).

• In ⇢3, 1 +

↵2� < p(p) 1 +

↵�. PP⇤

L is a = 1 �q

( p(p)� 1)2

↵�� (

↵�)

2.

a 2 [1� ↵�, 1� ↵

2� ).

• In ⇢4, 1 + ↵�< p(p) 2� ↵

�. PP⇤

L is a = 2� p(p). a 2 [

↵�, 1� ↵

�).

• In ⇢5, 2� ↵�< p(p) 2. PP⇤

L is a = 2� p(p). a 2 [0, ↵�).

• In ⇢6, p(p) > 2. PP⇤L is a = 0.


The following describes Eup(p) if p(p) is learned. Conceptually, the solution is the

sum of the product of the probabilities that p(p) lies in a partition and the expected

utility from a given that p(p) lies in the partition:

Define ⇢ as the partition in which p � ↵ lies. If ⇢j = ⇢ and ⇢j 6= ⇢, then

Pr( p(p) 2 ⇢j) =max(⇢

j

)�(p�↵)2↵ , E( p(p)| p(p) 2 ⇢j) =

max(⇢j

)+(p�↵)2 .

Define ⇢ as the partition in which p + ↵ lies. If ⇢j =⇢ and ⇢j 6= ⇢, then

Pr( p(p)| p(p) 2 ⇢j) =p+↵�min(⇢

j

)2↵ , E( p(p)| p(p) 2 ⇢j) =

p

(p)+↵+min(⇢j

)2 .

If ⇢ = ⇢, then Pr( p(p)| p(p) 2 ⇢j) = 1, E( p(p)| p(p) 2 ⇢j) = p.

Define b⇢ as the set of partitions which lie between ⇢ and ⇢.

If ⇢j 2 b⇢, then Pr( p(p)| p(p) 2 ⇢j) =

max(⇢j

)�min(⇢j

)2↵ , E( p(p)| p(p) 2 ⇢j) =

max(⇢j

)+min(⇢j

)2 .

If ⇢j /2 ⇢ [ ⇢ [ b⇢, then Pr( p(p) 2 ⇢j) = 0.

Eup(p, L) = �X

2

6666666664

(Pr( p(p) 2 ⇢1) + Pr( p(p) 2 ⇢6))↵2�

Pr( p(p) 2 ⇢2)E( p(p)| p(p) 2 ⇢2)

Pr( p(p) 2 ⇢3) ·

Pr( p(p) 2 ⇢4) · (2� E( p(p)| p(p) 2 ⇢4))

Pr( p(p) 2 ⇢5) · E⇣

(2� p

(p))2+(↵�

)2

2↵

�

| p(p) 2 ⇢5

⌘

3

7777777775

=

8>>>>>>>>><

>>>>>>>>>:

1�R 1+↵

�

p�↵(

2( p

(p)�1)↵�

�(↵�

)2)

12

1+↵

�

�(p�↵) d p(p) if ⇢3 2 ⇢ and ⇢3 /2 ⇢

1�R p+↵

1+ ↵

2�

(

2( p

(p)�1)↵�

�(↵�

)2)

12

p+↵�(1+ ↵

2� )d p(p) if ⇢3 2 ⇢ and ⇢3 /2 ⇢

1�R p+↵

p�↵(

2( p

(p)�1)↵�

�(↵�

)2)

12

2↵ d p(p) if ⇢3 2 ⇢ and ⇢3 2 ⇢

1�R 1+↵

�

1+ ↵

2�

(

2( p

(p)�1)↵�

�(↵�

)2)

12

↵

2�d p(p) if ⇢3 2 b⇢

For P , LP⇤ is defined implicitly. If Eup(p, L) � Eup(p,¬L), then LP⇤= L.

Otherwise, LP⇤= ¬L.


Now, I consider the values of p for which V accepts P ’s proposal to learn. As

demonstrated in the text of the paper, if PP⇤L is a, then, for all values of p(p), V is

made at least as well off as if p(p) had not been learned. As a result, LV ⇤= A.

Therefore, I limit attention to those values of p for which PP⇤L = p. For p +

↵ = � ↵2� + ✏, where ✏ is small, V prefers that p(p) not be learned, because if

p(p) 2⇣� ↵

2� ,�↵2� + ✏

⌘, then PP⇤

L = p and uv < �1. In contrast, if p(p) is not

learned, PP⇤¬L is a = 0 and uv = �1. Therefore, p = �↵ � ↵

2� defines the lower

boundary of the region where PV ⇤= ¬A.

There exists a cutpoint at which V is indifferent between p(p) being learned and

not that provides the upper boundary for this region. At that cutpoint, Euv(p,¬L) =

Euv(p, L). This cutpoint is defined by p =

↵2� � ↵. At p =

↵2� � ↵, Euv(p, L) =

�1, because for all possible values of p(p) such that PP⇤L = p, Pr( p(p)) > 0 =

Pr( p(p)) < 0. Therefore, for p 2 (�↵� ↵2� ,

↵2��↵), L

V ⇤= ¬A. Otherwise, LV ⇤

= A.

This fully characterizes the equilibrium behavior described in Proposition 2.

Appendix B

Appendix to Chapter 3

Introduction

In this Appendix, I provide proofs of the analytical results in the paper and pro-

vide annotated code that details the optimization routines used to compute values

of x⇤p(r) and yp

⇤.

Analytical Results

Propositions 4 and 6b and Lemma 2a and 2b

Begin with a general case where there exists a convex region of acceptable yp(m)

such that m is retained iff yp(m) 2 (yp, y0), where y0 > �y0 � yp. The probability

that m is retained is:

�

xpp�2xp

!� �

xp + yp � y0p�2xp

!

which I abbreviate as �(f(xp))� �(g(xp)).

Define x⇤p as the value of xp that maximizes �(f(xp))� �(g(xp)).

128

APPENDIX B. APPENDIX TO CHAPTER 3 129

For xp > 0:

�(f(xp))0= �(f(xp))f

0(xp)

�(f(xp))00= �f(xp)�(f(xp))(f

0(xp))

2+ �(f(xp))f

00(xp)

f(xp) > 0

�(f(xp)) > 0

f 0(xp) =

1

2

p�2xp

> 0

f 00(xp) = � �2

4(�2xp)32

< 0

�(f(xp))00 < 0

Therefore, �(f(xp))0 > 0 and �(f(xp))

00 < 0 and �(f(xp)) is concave.

Also:

�(g(xp))0= �(g(xp))g

0(xp)

�(g(xp))00= �g(xp)�(g(xp))(g

0(xp))

2+ �(g(xp))g

00(xp)

g0(xp) = f 0(xp) +

y0 � yp

2�x3/2p

> f 0(xp)

g00(xp) =

3(yp � y0)� xp

4x2p

p�2xp

For xp �y0�y

p

2 :

�(g(xp)) � �(f(xp))

d

dxp

[�(f(xp))� �(g(xp))] < 0


Therefore, the probability of retention is strictly decreasing in x⇤p �

y0�yp

2 .

For xp 2 (0,y0�y

p

2 ]:

�(f(xp)) > �(g(xp))

g00(xp) > 0

�(g(xp))00 > 0

As �(g(xp)) is convex, ��(g(xp)) is concave. As the sum of two concave functions

is concave, �(f(xp)) � �(g(xp)) is concave as well in the region. Thus, for xp > 0,

the function is quasi-concave. As the derivative of �(f(xp))��(g(xp)) with respect

to xp is negative at y0�yp

2 , x⇤p <

y0�yp

2 .

Finally, I examine xp 0. For all xp < 0, the probability of replacement strictly

exceeds 12 . For xp > 0 and very small, d

dxp

⇡ �(f(xp))0 > 0 because �(g(x)) ⇡ 0.

As �(f(xp))0 > 0, the probability of replacement for xp very small is strictly less

than 12 and thus dominates all xp < 0. (xp = 0 is dominated by any xp 6= 0, as the

probability of replacement is 1 if m chooses xp = 0.)

As xp 0 is strictly dominated and the probability of being retained is quasi-

concave in xp > 0, uniqueness has been demonstrated for r = OB and r = RS.

Further, x⇤p must be greater than 0 and less than y0�y

p

2 .

Proposition 5 and Lemma 3

For all xp, the value of yp that minimizes EUO is �y0 and EUO((xp,�y0)) = �xp

�2

4

(Callander 2011a; 2011b). Setting EUO(m0) equal to the former, I find that under

Forward-looking Replacement if xp < 4y0 � �2 for y0 > �2

2 or xp <4y20�2 for y0 �2

2 ,


the only replacement region for m is if yp > y0.

If y0 > �2

2 and xp � 4y0 � �2 or if y0 �2

2 and xp � 4y20�2 , then there exist two

regions in which m will be replaced. The “upper region” is all yp(m) � y0. The “lower

region” consists of values of yp(m) < 0 such that µ is sufficiently shallow to warrant

replacement.

The “lower bound” of the upper region is y0. (There is no “upper bound” to the

upper region.)

The “upper bound” and “lower bound” of the lower region are defined by the

values of (xp, yp(m)) such that EUO(m) EUO(m0). If m would choose to backslide

to y0, m0 weakly dominates m. Therefore, the upper bound of the lower region is

weakly greater than the upper bound of the region below 0 in which m backslides,

defined in Proposition 2e. The lower bound of the lower region is weakly less than

the lower bound of the region below 0 for which m backslides, defined in Proposition

2f.

First, I examine y0 �2

2 and xp <4y20�2 . m will be replaced for any pair of (xp, yp)

that leads m to backslide to y0. I focus on the lower region. The upper boundary of

the lower region is �y0, as yp(m) > �y0 makes a stable strategy for m preferable to

replacement with m0. Rearranging terms from Proposition 2f, the lower boundary of

the lower region is y0� xp

�2

2y0. As �2 > 2y0, this means that the lower bound decreases

more quickly than E( m(xp)). Further, at xp =

4y20�2 , the lower bound is equal to

E( m(xp)).

By these properties:


� =�2

2y0> 1

Pr(replace) =

"1� �

xpp�2xp

!#+

"�

xp � 2y0p

�2xp

!� �

(1� �)xpp

�2xp

!#

As xp > xp � 2y0 > 0, 0 < ddx

p

�

✓xpp�2x

p

◆< d

dxp

�

✓xp

�2y0p�2x

p

◆. As d

dxp

(1��)xpp

�2xp

< 0,

ddx

p

�

✓(1��)x

pp�2x

p

◆< 0. Therefore, I have shown EUm(xp =

4y20�2 ) > EUm(xp >

4y20�2 ).

For any xp <4y20�2 , the probability of replacement is 1� �

✓xpp�2x

p

◆, which has a

negative first derivative with respect to xp. Therefore, EUm(xp =4y20�2 ) > EUm(xp <

4y20�2 ). As a result, for y0 �2

2 , x⇤p =

4y20�2 .

Now, I turn to y0 >�2

2 and xp � 4y0 � �2. The upper bound of the lower region

must be greater than �y0, as �y0 provides the worst outcome of any yp(m) for O.

It also must be weakly diverging from y0 as xp increases, as for any x, x0 such that

x < x0 and for any 0 > y > �y0, EU(x⇤1, (x, y)) > EU(x⇤

1, (x0, y) due to the flatter

slope on the Brownian bridge.

Formally, let x, x0 > 0 and let 0 < y < �y0. Let x+ be the value of x1 that

maximizes EUO(x0, y). Let xo be the value of x1 such that E( (x+, (x0, y))) =

E( (xo, (x, y))). As x⇤

x0 =

xo

x, V ar( (x+

)) > V ar( (xo)) and EUO(x

+, (x0, y)) <

EUO(xo, (x, y)).

If the derivative of the lower bound of the lower region (LB) with respect to xp

is strictly less than ddx

p

E( m(xp)) = �1, then the first derivative of LB�(y0�xp

)p�2x

p

is

negative as well. Note that if yp(m) E( m(xp)), m is necessarily more desirable

than m0 as µ < µ and the variance of search on the Brownian bridge is less than

the variance associated with open-ended search. Therefore, �✓

LB�(y0�xp

)p�2x

p

◆> 1

2 and


ddx

p

�

✓LB�(y0�x

p

)p�2x

p

◆< 0.

By these properties and noting that xp > xp � 2y0:

Pr(replace) >

"1� �

xpp�2xp

!#+

"�

xp � 2y0p

�2xp

!� �

LB � (y0 � xp)p

�2xp

!#

As the upper bound of the lower region is increasing more quickly than the lower

bound of the upper region with respect to xp and the lower bound of the lower region

is strictly decreasing with respect to xp, the probability of replacement is strictly in-

creasing with respect to xp if ddx

p

LB < �1.

Computational Results

Finding x

⇤p(GT)

Under Guaranteed Tenure, m selects x⇤p in order to maximize EUO (or minimize

the squared distance between y1 and y⇤ = 0). For a given xp, the pdf of yp, f(yp) ⇠

N (y0 � xp,p

xp�2) and EUO(xp) =

REUO(yp)f(yp)dyp.

In the second period, there are five possible changes that m can make to xp: (1)

m can backslide to x0; (2) m can remain stable at xp; (3) m can explore to the

right; (4) m can search on the Brownian bridge; and (5) m can explore to the left.

Propositions 2a through 2g define the conditions under which m does each and the

associated expected utility of O.

First, I determine a series of cutpoints across which the relationship between xp

and x⇤1(m) changes:

• The first cutpoint, c1 =y20+�

4/4�2 , defines the minimum value of yp for which m

backslides to x0 for xp � 0 and y0 >�2

2 .


• The second cutpoint, c2 = y0, defines the minimum value of yp for which m

backslides to x0 for y0 �2

2 .

• The third cutpoint, c3 =

q(y0 � �2

4 )�2, defines the minimum value of yp for

which m explores to the right from x0 for xp < 0 and y0 >�2

2 .

• The fourth cutpoint, c4 =

�2

2 , defines the minimum value of yp for which m

explores to the right from xp for xp � 0.

• The fifth cutpoint, c5 =

y0�p

y20+2xp

�2

2 , defines the minimum value of yp for

which m remains stable at xp for 0 xp 4y20�2 .

• The sixth cutpoint, c6 = �y0, defines the minimum value of yp for which m

remains stable at xp for xp >4y20�2 .

• The seventh cutpoint, c7 = y0� xp

�2

2y0, defines the minimum value of yp for which

m backslides to x0 for xp >4y20�2 .

Using these cutpoints, the following equation defines EUO(yp|xp) under r = GT :

• If xp < 0 and y0 >�2

2

EUO < � 1� �

xp + c3 � y0p

|xp|�2

!!✓y0 �

�2

4

◆�2�

Z c3

0

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!y2pdyp


• If xp < 0 and y0 �2

2

EUO < � 1� �

xp + c2 � y0p

|xp|�2

!!y20�

Z c2

0

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!y2pdyp

• If 0 xp 4y20�2 and y0 >

�2

2

µ =

yp � y0xp

x⇤1(m) =

2µy0 + �2

2

⇣�2

xp

� µ2⌘

EUO = � 1� �

xp + c1 � y0p

|xp|�2

!!y20�

Z c1

c4

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!✓yp �

�2

4

◆�2dyp�

Z c4

c5

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!y2pdyp�

Z c5

�1

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!✓(y0 + µx⇤

1(m))

2+

(xp � x⇤1(m))x⇤

1(m)

xp

�2

◆dyp


• If 0 xp 4y20�2 and y0 <

�2

2

EUO = � 1� �

xp + c2 � y0p

|xp|�2

!!y20�

Z c2

c5

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!y2pdyp�

Z c5

�1

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!✓(y0 + µx⇤

1(m))

2+

(xp � x⇤1(m))x⇤

1(m)

xp

�2

◆dyp

• If xp >4y20�2 and y0 >

�2

2

EUO = � 1� �

xp + c1 � y0p

|xp|�2

!!y20�

Z c1

c4

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!✓yp �

�2

4

◆�2dyp�

Z c4

c6

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!y2pdyp�

�

xp + c6 � y0p

|xp|�2

!� �

xp + c7 � y0p

xp�2

!!y20�

Z c7

�1

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!✓(y0 + µx⇤

1(m))

2+

(xp � x⇤1(m))x⇤

1(m)

xp

�2

◆dyp


• If xp >4y20�2 and y0 �2

2

EUO = � 1� �

xp + c2 � y0p

|xp|�2

!!y20�

Z c2

c6

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!✓yp �

�2

4

◆�2dyp�

�

xp + c6 � y0p

|xp|�2

!� �

xp + c7 � y0p

|xp|�2

!!y20�

Z c7

�1

1p|xp|�2

�

xp + yp � y0p

|xp|�2

!✓(y0 + µx⇤

1(m))

2+

(xp � x⇤1(m))x⇤

1(m)

xp

�2

◆dyp

The function “CompleteIntegral.GT” calculates EUO(yp|xp) under r = GT . It

utilizes a series of helper functions that calculate pieces of the partitioned density of

yp resulting from xp. The function “integrate_sg” corrects the output of the integral

for very high variance density functions:

integrand.1 = function(y_p, x_p = x_p, s2 = s2, y_0 = y_0) {

dnorm(y_p, y_0 - x_p, sqrt(abs(x_p) * s2)) * -(y_p - s2/4) *

s2

}


EU_O <- -y_p^2

dnorm(y_p, y_0 - x_p, sqrt(abs(x_p) * s2)) * EU_O

}



mu_hat <- (y_p - y_0)/x_p

p2_star <- (2 * mu_hat * y_0 + s2)/(2 * (s2/x_p - mu_hat^2))

out <- -(y_0 + mu_hat * p2_star)^2 - ((x_p - p2_star) * p2_star *

s2)/x_p

dnorm(y_p, y_0 - x_p, sqrt(abs(x_p) * s2)) * out

}

CompleteIntegral.GT = function(x_p, s2, y_0) {

if (x_p > 0) {

c_1 <- (y_0^2 + s2^2/4)/s2

c_2 <- y_0

c_3 <- ifelse(y_0 >= s2/4, sqrt((y_0 - s2/4) * s2), NA)

c_4 <- s2/2

c_5 <- (y_0 - sqrt(y_0^2 + 2 * x_p * s2))/2

c_6 <- -y_0

c_7 <- y_0 - (x_p * s2)/(2 * y_0)

}

if (x_p < 0) {

c_1 <- (y_0^2 + s2/4)/s2

c_2 <- y_0

c_3 <- ifelse(y_0 >= s2/4, sqrt((y_0 - s2/4) * s2), NA)

}


if (x_p == 0) {

return(-(y_0 - s2/4) * s2)

}

if (x_p < 0 & y_0 > s2/2) {

part1 <- -(1 - pnorm((x_p + c_3 - y_0)/sqrt(abs(x_p) *

s2))) * (y_0 - s2/4) * s2

part2 <- integrate(integrand.2, lower = 0, upper = c_3,

x_p = x_p, s2 = s2, y_0 = y_0)$value

return(part1 + part2)

}

if (x_p < 0 & y_0 <= s2/2) {


s2))) * y_0^2

part2 <- integrate(integrand.2, lower = 0, upper = c_2,

x_p = x_p, s2 = s2, y_0 = y_0)$value

return(part1 + part2)

}

if (x_p > 0 & x_p <= 4 * y_0^2/s2 & y_0 > s2/2) {


s2))) * y_0^2

part2 <- integrate_sg(integrand.1, lower = c_4, upper = c_1,

x_p = x_p, s2 = s2, y_0 = y_0)$value


part3 <- integrate(integrand.2, lower = c_5, upper = c_4,

x_p = x_p, s2 = s2, y_0 = y_0)$value

part4 <- integrate(integrand.3, lower = -Inf, upper = c_5,

x_p = x_p, s2 = s2, y_0 = y_0)$value

return(part1 + part2 + part3 + part4)

}

if (x_p > 0 & x_p <= 4 * y_0^2/s2 & y_0 <= s2/2) {


s2))) * y_0^2


x_p = x_p, s2 = s2, y_0 = y_0)$value

part3 <- integrate_sg(integrand.3, lower = -Inf, upper = c_5,

x_p = x_p, s2 = s2, y_0 = y_0)$value

return(part1 + part2 + part3)

}

if (x_p > 4 * y_0^2/s2 & y_0 > s2/2) {


s2))) * y_0^2


x_p = x_p, s2 = s2, y_0 = y_0)$value


x_p = x_p, s2 = s2, y_0 = y_0)$value

part4a <- pnorm((x_p + c_6 - y_0)/sqrt(abs(x_p) * s2)) -


pnorm((x_p + c_7 - y_0)/sqrt(abs(x_p) * s2))

part4b <- -y_0^2


x_p = x_p, s2 = s2, y_0 = y_0)$value

return(part1 + part2 + part3 + part4a * part4b + part5)

}

if (x_p > 4 * y_0^2/s2 & y_0 <= s2/2) {


s2))) * y_0^2


x_p = x_p, s2 = s2, y_0 = y_0)$value

part3a <- pnorm((x_p + c_6 - y_0)/sqrt(abs(x_p) * s2)) -

pnorm((x_p + c_7 - y_0)/sqrt(abs(x_p) * s2))

part3b <- -y_0^2


x_p = x_p, s2 = s2, y_0 = y_0)$value

return(part1 + part2 + part3a * part3b + part4)

}

}

I now find xp that maximizes EUO(yp|xp) under GT. I maximize CompleteInte-

gral.GT using the function “optimise.GT”. The function uses a genetic optimization

routine to search for a global maximum.


optimise.GT = function(s2, y_0) genoud(CompleteIntegral.GT, nvars = 1,

Domains = rbind(c(0.05, 4 * y_0)), boundary.enforcement = 0,

s2 = s2, y_0 = y_0, print.level = 0, max = T)

Finding x

⇤p(RS)

Under a Retrospective Satisficing replacement rule, m selects x⇤p in order to

maximize EUm(yp|xp), which is the probability that |yp| < y0. Pr(|yp| < y0) =

�

✓xpp

xp

�2

◆��

✓xp

�2y0pxp

�2

◆, which is calculated by the function “CompleteIntegral.RS”.

CompleteIntegral.RS = function(x_p, s2, y_0) {

-(pnorm(x_p/sqrt(abs(x_p) * s2)) - pnorm((x_p - 2 * y_0)/sqrt(abs(x_p) *

s2)))

}

I now find xp that maximizes EUm under RS. I maximize CompleteIntegral.GT

subject to the box constraint that 0 < xp y0 using Simulated Annuealing in the

function “optimise.RS”.

optimise.RS = function(s2, y_0) GenSA(y_0/2, CompleteIntegral.RS,

10^-8, y_0, control = list(verbose = FALSE, max.call = 1000),

s2 = s2, y_0 = y_0)

Finding yp⇤

Under a Outcome-based Replacement rule, O sets yp in order to maximize EUO.

Given yp, m then chooses x⇤p(OB) in order to minimize the probability of being

replaced. EUO(x⇤p(OB)|yp) under Outcome-based Replacement will be the calculated

in the same fashion as EUO(x⇤p(GT )) under Guaranteed Tenure, except that for

yp � y0 and yp yp, EUO = �⇣y0 � �2

4

⌘�2.


The function “CompleteIntegral.OB” calculates EUO(yp) under r = OB:

EU_m.OB = function(x_p, LB, s2, y_0) {

UB <- y_0 - 10^(-5)

-((pnorm(UB, y_0 - x_p, sqrt(x_p * s2)) - pnorm(LB, y_0 -

x_p, sqrt(x_p * s2))))

}

x_p.star.OB = function(LB, s2, y_0) {

(optimise(EU_m.OB, lower = 10^-5, upper = y_0 - (LB + y_0)/2,

LB = LB, s2 = s2, y_0 = y_0, maximum = F)$minimum)

}

CompleteIntegral.OB = function(LB, s2, y_0) {

if (LB == -Inf)

return(-y_0^2)

if (LB >= y_0 - 2 * 10^-5) {

return(ifelse(y_0 <= s2/2, -y_0^2 - (LB - y_0 + 10^-5),

-(y_0 - s2/4) * s2 - (LB - y_0 + 10^-5)))

}

x_p <- x_p.star.OB(LB, s2, y_0)

c_1 <- y_0

c_4 <- s2/2

c_5 <- (y_0 - sqrt(y_0^2 + 2 * x_p * s2))/2

c_6 <- -y_0

c_7 <- y_0 - (x_p * s2)/(2 * y_0)


if (y_0 > s2/2) {

top <- -(1 - pnorm((x_p + c_1 - y_0)/sqrt(abs(x_p) *

s2))) * (y_0 - s2/4) * s2

bottom <- -pnorm((x_p + LB - y_0)/sqrt(abs(x_p) * s2)) *

(y_0 - s2/4) * s2

} else {

top <- -(1 - pnorm((x_p + c_1 - y_0)/sqrt(abs(x_p) *

s2))) * y_0^2

bottom <- -pnorm((x_p + LB - y_0)/sqrt(abs(x_p) * s2)) *

y_0^2

}

if (x_p >= 0 & x_p <= 4 * y_0^2/s2 & y_0 > s2/2) {

if (LB > c_4) {

part1 <- integrate(integrand.1, lower = LB, upper = c_1,

x_p = x_p, s2 = s2, y_0 = y_0)$value

part2 <- part3 <- 0

} else if (LB > c_5) {


x_p = x_p, s2 = s2, y_0 = y_0)$value


x_p = x_p, s2 = s2, y_0 = y_0)$value

part3 <- 0

} else {


x_p = x_p, s2 = s2, y_0 = y_0)$value



x_p = x_p, s2 = s2, y_0 = y_0)$value


x_p = x_p, s2 = s2, y_0 = y_0)$value

}

return(part1 + part2 + part3 + top + bottom)

} else if (x_p >= 0 & x_p <= 4 * y_0^2/s2 & y_0 <= s2/2) {

if (LB > c_5) {


x_p = x_p, s2 = s2, y_0 = y_0)$value

part2 <- 0

} else {


x_p = x_p, s2 = s2, y_0 = y_0)$value


x_p = x_p, s2 = s2, y_0 = y_0)$value

}

return(part1 + part2 + top + bottom)

} else if (x_p > 4 * y_0^2/s2 & y_0 > s2/2) {

if (LB > c_4) {

part1 <- integrate_sg(integrand.1, lower = LB, upper = c_1,

x_p = x_p, s2 = s2, y_0 = y_0)$value

part2 <- part3 <- part4 <- 0




x_p = x_p, s2 = s2, y_0 = y_0)$value


x_p = x_p, s2 = s2, y_0 = y_0)$value

part3 <- part4 <- 0



x_p = x_p, s2 = s2, y_0 = y_0)$value


x_p = x_p, s2 = s2, y_0 = y_0)$value

part3 <- (pnorm((x_p + c_6 - y_0)/sqrt(abs(x_p) *

s2)) - pnorm((x_p + LB - y_0)/sqrt(abs(x_p) *

s2))) * -y_0^2

part4 <- 0

} else {


x_p = x_p, s2 = s2, y_0 = y_0)$value


x_p = x_p, s2 = s2, y_0 = y_0)$value


s2)) - pnorm((x_p + c_7 - y_0)/sqrt(abs(x_p) *

s2))) * -y_0^2


x_p = x_p, s2 = s2, y_0 = y_0)$value

}

total <- part1 + part2 + part3 + part4 + top + bottom


return(ifelse(round(total, 5) == -y_0^2, total + LB,

total))

} else if (x_p > 4 * y_0^2/s2 & y_0 <= s2/2) {

if (LB > c_6) {


x_p = x_p, s2 = s2, y_0 = y_0)$value

part2 <- part3 <- 0



x_p = x_p, s2 = s2, y_0 = y_0)$value


s2)) - pnorm((x_p + LB - y_0)/sqrt(abs(x_p) *

s2))) * -y_0^2

part3 <- 0

} else {


x_p = x_p, s2 = s2, y_0 = y_0)$value


s2)) - pnorm((x_p + c_7 - y_0)/sqrt(abs(x_p) *

s2))) * -y_0^2


x_p = x_p, s2 = s2, y_0 = y_0)$value

}

total <- part1 + part2 + part3 + top + bottom

}


}

I now find yp that maximizes EUO(yp) under OB. I maximize CompleteInte-

gral.OB using the function “optimise.OB”. The function uses a genetic optimization

routine to search for a global maximum.

optimise.OB = function(s2, y_0) genoud(CompleteIntegral.OB, nvars = 1,

Domains = rbind(c(-20 * y_0, y_0 - 10^-5)), boundary.enforcement = 0,

s2 = s2, y_0 = y_0, print.level = 0, max = T)

Bibliography

Adler, P. S. and Borys, B. (1996). Two Types of Bureaucracy: Enabling and Coercive.

Administrative Science Quarterly, 41(1):61–89.

Aghion, P. and Jackson, M. (2014). Inducing Leaders to Take Risky Decisions: Dis-

missal, Tenure, and Term Limits. National Bureau of Economic Research Working

Paper Series, No. 20301.

Agnone, J. (2007). Amplifying Public Opinion: The Policy Impact of the US Envi-

ronmental Movement. Social Forces, 85(4):1593–1620.

Allen, M., Raper, S., and Mitchell, J. (2001). Uncertainty in the IPCC’s Third

Assessment Report. Science, 293(5529):430–433.

Amenta, E., Caren, N., Chiarello, E., and Su, Y. (2010). The Political Consequences

of Social Movements. Annual Review of Sociology, 36(1):287–307.

Amenta, E., Carruthers, B. G., and Zylan, Y. (1992). A Hero for the Aged? The

Townsend Movement, the Political Mediation Model, and US Old-Age Policy,

1934-1950. American Journal of Sociology, 98(2):308–339.

Amenta, E., Dunleavy, K., and Bernstein, M. (1994). Stolen Thunder? Huey Long’s

149

BIBLIOGRAPHY 150

“Share Our Wealth,” Political Mediation, and the Second New Deal. American

Sociological Review, 59(5):678–702.

Anderson, P. and Tushman, M. L. (1990). Technological Discontinuities and Domi-

nant Designs: A Cyclical Model of Technological Change. Administrative Science

Quarterly, pages 604–633.

Andrews, K. T. and Edwards, B. (2004). Advocacy Organizations in the U.S. Political

Process. Annual Review of Sociology, 30(1):479–506.

Andrews, R. N. L. (2006). Managing the Environment, Managing Ourselves: A

History of American Environmental Policy. Yale University Press, New Haven,

2nd edition.

Austen-Smith, D. and Banks, J. S. (1999). Positive Political Theory I: Collective

Preference. Michigan Studies in Political Analysis. University of Michigan Press,

Ann Arbor.

Austen-Smith, D. and Banks, J. S. (2000). Cheap Talk and Burned Money. Journal

of Economic Theory, 91(1):1–16.

Barnard, C. I. (1938). The Functions of the Executive. Harvard University Press,

Cambridge.

Baron, J. N., Hannan, M. T., and Burton, M. D. (1999). Building the Iron Cage: De-

terminants of Managerial Intensity in the Early Years of Organizations. American

Sociological Review, 64(4):527–547.

Baumgartner, F. R. and Jones, B. D. (2009). Agendas and Instability in American

Politics. University of Chicago Press, Chicago, 2nd edition.

BIBLIOGRAPHY 151

Baumgartner, F. R. and Jones, B. D. (2015). The Politics of Information: Problem

Definition and the Course of Public Policy in America. The University of Chicago

Press, Chicago.

Baumgartner, F. R. and Leech, B. L. (1998). Basic Interests: The Importance of

Groups in Politics and in Political Science. Princeton University Press, Princeton,

N.J.

Bendor, J. (2010). Bounded Rationality and Politics. University of California Press,

Berkeley.

Bendor, J. and Meirowitz, A. (2004). Spatial Models of Delegation. American

Political Science Review, 98(2):293–310.

Bendor, J., Moe, T. M., and Shotts, K. W. (2001). Recycling the Garbage Can:

An Assessment of the Research Program. American Political Science Review,

95(1):169–190.

Bernstein, E. S. (2012). The Transparency Paradox a Role for Privacy in Orga-

nizational Learning and Operational Control. Administrative Science Quarterly,

57(2):181–216.

Berry, J. M. (1977). Lobbying for the People: The Political Behaviour of Public

Interest Groups. Princeton University Press, Princeton, N.J.

Beshears, J. and Gino, F. (2014). Experiment with Organizational Change Before

Going All In. HBR.org.

Black, D., Newing, R. A., McLean, I., McMillan, A., and Monroe, B. L. (1958). The

Theory of Committees and Elections. Cambridge University Press, New York.

BIBLIOGRAPHY 152

Blau, P. M. (1955). The Dynamics of Bureaucracy: A Study of Interpersonal Rela-

tions in Two Government Agencies. The University of Chicago Press, Chicago.

Bosso, C. J. (2005). Environment, Inc.: From Grassroots to Beltway. University

Press of Kansas, Lawrence.

Brocas, I. and Carillo, J. D. (2007). Influence Through Ignorance. RAND Journal

of Economics, 38(4):931–947.

Brown, P., Zavestoski, S., McCormick, S., Mayer, B., Morello-Frosch, R., and

Gasior Altman, R. (2004). Embodied Health Movements: New Approaches to

Social Movements in Health. Sociology of health & illness, 26(1):50–80.

Brudnick, I. A. (2011). The Congressional Research Service and the American Leg-

islative Process. Technical Report 7-5700, Congressional Research Service, Wash-

ington, D.C.

Burstein, P. (1998). Interest Organizations, Political Parties, and the Study of Demo-

cratic Politics. In Costain, A. N. and McFarland, A. S., editors, Social Movements

and American Political Institutions, pages 39–56. Rowman & Littlefield Publish-

ers, Oxford.

Burstein, P. and Freudenburg, W. (1978). Changing Public Policy: The Impact of

Public Opinion, Antiwar Demonstrations, and War Costs on Senate Voting on

Vietnam War Motions. American Journal of Sociology, 84(1):99–122.

Burstein, P. and Hirsh, C. E. (2007). Interest Organizations, Information, and Policy

Innovation in the U.S. Congress. Sociological Forum, 22(2):174–199.

Callander, S. (2008). A Theory of Policy Expertise. Quarterly Journal of Political

Science, 3(2):123–140.

BIBLIOGRAPHY 153

Callander, S. (2011a). Searching and Learning by Trial and Error. American Eco-

nomic Review, 101:2277–2308.

Callander, S. (2011b). Searching for Good Policies. American Political Science

Review, 105(4):643–662.

Callander, S. and Matouschek, N. (2014). Managing on Rugged Landscapes. Working

Paper.

Cameron, C. and McCarty, N. (2004). Models of Veto and Veto Bargaining. Annual

Reviews of Political Science, 7:409–435.

Canes-Wrone, B., Herron, M. C., and Shotts, K. W. (2001). Leadership and Pander-

ing: A Theory of Executive Policymaking. American Journal of Political Science,

45(3):532–550.

Cobb, M. D. and Kuklinski, J. H. (1997). Changing Minds: Political Arguments and

Political Persuasion. American Journal of Political Science, 41(1):88–121.

Cohen, K. J., Cyert, R. M., Dill, W. R., Kuehn, A. A., Miller, M. H., Van Wormer,

T. A., and Winters, P. R. (1960). The Carnegie Tech Management Game. The

Journal of Business, 33(4):303–321.

Cohen, M. D., March, J. G., and Olsen, J. P. (1972). A Garbage Can Model of

Organizational Choice. Administrative Science Quarterly, pages 1–25.

Cooper, J. (1970). The Origins of Standing Committees and the Development of the

Modern House. Rice University Studies, Houston.

Crawford, V. P. and Sobel, J. (1982). Strategic Information Transmission. Econo-

metrica: Journal of the Econometric Society, 50(6):1431–1451.

BIBLIOGRAPHY 154

Cress, D. M. and Snow, D. A. (2000). The Outcomes of Homeless Mobilization: The

Influence of Organization, Disruption, Political Mediation, and Framing. American

Journal of Sociology, 105(4):1063–1104.

Csaszar, F. A. (2013). An Efficient Frontier in Organization Design: Organizational

Structure as a Determinant of Exploration and Exploitation. Organization Science,

24(4):1083–1101.

Csaszar, F. A. and Eggers, J. P. (2013). Organizational Decision Making: An Infor-

mation Aggregation View. Management Science, 59(10):2257–2277.

Csaszar, F. A. and Levinthal, D. A. (2016). Mental Representation and the Discovery

of New Strategies. Strategic Management Journal.

Cyert, R. M. and March, J. G. (1963). A Behavioral Theory of the Firm. Blackwell,

Cambridge, 2nd edition.

Dalton, M. (1959). Men Who Manage: Fusions of Feeling and Theory in Adminis-

tration. John Wiley & Sons, New York.

Davis, J. P., Eisenhardt, K. M., and Bingham, C. B. (2007). Developing Theory

through Simulation Methods. Academy of Management Review, 32(2):480–499.

DeNardo, J. (1985). Power in Numbers. Princeton University Press, Princeton, N.J.

Denrell, J. (2003). Vicarious Learning, Undersampling of Failure, and the Myths of

Management. Organization Science, 14(3):227–243.

Denrell, J. and March, J. G. (2001). Adaptation as Information Restriction: The

Hot Stove Effect. Organization Science, 12(5):523–538.

BIBLIOGRAPHY 155

Diermeier, D. and Feddersen, T. J. (2000). Information and Congressional Hearings.

American Journal of Political Science, 44(1):51–65.

Diermeier, D. and Krehbiel, K. (2003). Institutionalism as a Methodology. Journal

of Theoretical Politics, 15(2):123–144.

DiMaggio, P. J. and Powell, W. W. (1983). The Iron Cage Revisited: Institutional

Isomorphism and Collective Rationality in Organizational Fields. American Soci-

ological Review, 48(2):147–160.

Downs, A. (1957). An Economic Theory of Political Action in a Democracy. Journal

of Political Economy, pages 135–150.

Edelman (2015). 2015 Edelman Trust Barometer Results. Technical report.

Eden, S. (1998). Environmental Issues: Knowledge, Uncertainty and the Environ-

ment. Progress in Human Geography, 22(3):425–432.

Eisenhardt, K. M. and Bourgeois, L. J. (1988). Politics of Strategic Decision Making

in High-Velocity Environments: Toward a Midrange Theory. Academy of Manage-

ment Journal, 31(4):737–770.

Epstein, S. (1996). Impure Science: AIDS, Activism, and the Politics of Knowledge.

Univ of California Press.

Erikson, R. S., MacKuen, M. B., and Stimson, J. A. (2002). The Macro Polity.

Cambridge University Press, New York.

Esterling, K. M. (2004). The Political Economy of Expertise. The University of

Michigan Press, Ann Arbor.

BIBLIOGRAPHY 156

Ethiraj, S. K. and Levinthal, D. (2009). Hoping for A to Z While Rewarding Only

A: Complex Organizations and Multiple Goals. Organization Science, 20(1):4–21.

Ezrahi, Y. (1990). The Descent of Icarus: Science and the Transformation of Con-

temporary Democracy. Harvard University Press, Cambridge.

Farber, D. A. (2003). Probabilities Behaving Badly: Complexity Theory and Envi-

ronmental Uncertainty. UC Davis L. Rev., 37:145.

Feldman, M. S. and March, J. G. (1981). Information in Organizations as Signal and

Symbol. Administrative Science Quarterly, 26(2):171–186.

Fleming, L. (2002). Finding the Organizational Sources of Technological Break-

throughs: The Story of Hewlett Packard’s Thermal Ink Jet. Industrial and Cor-

porate Change, 11(5):1059–1084.

Fleming, L. and Sorenson, O. (2004). Science as a Map in Technological Search.

Strategic Management Journal, 25:909–928.

Freeland, R. (2001). The Struggle for Control of the Modern Corporation: Organi-

zational Change at General Motors, 1924-1970. Cambridge University Press, New

York.

Fudenberg, D. and Tirole, J. (1991). Game Theory. MIT Press, Cambridge.

Gavetti, G., Greve, H. R., Levinthal, D. A., and Ocasio, W. (2012). The Behavioral

Theory of the Firm: Assessment and Prospects. Academy of Managament Annals,

6(1):1–40.

Gavetti, G. and Levinthal, D. (2000). Looking Forward and Looking Backward:

BIBLIOGRAPHY 157

Cognitive and Experiential Search. Administrative Science Quarterly, 45(1):113–

137.

Gavetti, G., Levinthal, D., and Ocasio, W. (2007). Perspective–Neo-Carnegie: The

Carnegie School’s Past, Present, and Reconstructing for the Future. Organization

Science, 18(3):523–536.

Gertner, J. (2012). The Idea Factory: Bell Labs and the Great Age of American

Innovation. The Penguin Press, New York.

Gibbons, R. (2003). Team Theory, Garbage Cans, and Real Organizations: Some

History and Prospects of Economic Research on Decision-Making in Organizations.

Industrial and Corporate Change, 12(4):753–787.

Gibbons, R., Matouschek, N., and Roberts, J. (2012). Decisions in Organizations. In

Gibbons, R. and Roberts, J., editors, The Handbook of Organizational Economics,

pages 373–431. Princeton University Press, Princeton.

Gibbons, R. and Roberts, J. (2012). Economic Theories of Incentives in Organiza-

tions. In Gibbons, R. and Roberts, J., editors, The Handbook of Organizational

Economics, pages 56–99. Princeton University Press, Princeton.

Gilligan, T. W. and Krehbiel, K. (1987). Collective Decisionmaking and Standing

Committees: An Informational Rationale for Restrictive Amendment Procedures.

Journal of Law, Economics, & Organization, 3(2):287–335.

Gilligan, T. W. and Krehbiel, K. (1989). Asymmetric Information and Legislative

Rules with a Heterogeneous Committee. American Journal of Political Science,

33(2):459–490.

BIBLIOGRAPHY 158

Gilligan, T. W. and Krehbiel, K. (1990). Organization of Informative Committees

by a Rational Legislature. American Journal of Political Science, 34(2):531–564.

Gillion, D. Q. (2012). Protest and Congressional Behavior: Assessing Racial and

Ethnic Minority Protests in the District. Journal of Politics, 74(04):950–962.

Gillion, D. Q. (2013). The Political Power of Protest: Minority Activism and Shifts

in Public Policy. Cambridge University Press, New York.

Giugni, M. (2004). Social Protest and Policy Change: Ecology, Antinuclear, and

Peace Movements in Comparative Perspective. Rowman & Littlefield, New York.

Granovetter, M. (2005). The Impact of Social Structure on Economic Outcomes.

Journal of Economic Perspectives, 19(1):33–50.

Granovetter, M. and Swedberg, R. (2011). The Sociology of Economic Life. Westview

Press, Boulder, third edition.

Habermas, J. (1971). Toward a Rational Society: Student Protest, Science, and

Politics, volume 404. Beacon Press.

Hadden, J. (2015). Networks in Contention. Cambridge University Press, New York.

Hall, R. L. and Deardorff, A. V. (2006). Lobbying as Legislative Subsidy. American


Hamel, G. and Spence, S. (2010). Innovation Democracy: W.L. Gore’s Original

Management Model. Management Innovation Exchange.

Hannan, M. T. and Freeman, J. (1984). Structural Inertia and Organizational

Change. American Sociological Review, 49(2):149–164.

BIBLIOGRAPHY 159

Hansen, J. M. (1991). Gaining Access: Congress and the Farm Lobby, 1919-1981.

University of Chicago Press, Chicago.

Harrison, J. R., Lin, Z., Carroll, G. R., and Carley, K. M. (2007). Simulation

Modeling in Organizational and Management Research. Academy of Management

Review, 32(4):1229–1245.

Heinz, J. P., Laumann, E., Nelson, R., and Salisbury, R. (1993). The Hollow Core:

Private Interests in National Policy Making. Harvard University Press, Cambridge.

Holmstrom, B. and Ricart i Costa, J. (1986). Managerial Incentives and Capital

Management. Quarterly Journal of Economics, 101(4):835–860.

Ingram, P. and Simons, T. (2000). State Formation, Ideological Competition, and

the Ecology of Israeli Workers’ Cooperatives, 1920-1992. Administrative Science

Quarterly, 45(1):25–53.

Jamison, A. (2001). The Making of Green Knowledge: Environmental Politics and

Cultural Transformation. Cambridge University Press, New York.

Jenkins, J. C. (1987). Non-Profit Organizations and Policy Advocacy. In Powell, W.,

editor, Non-Profit Organization: A Handbook. Yale University Press, New Haven.

Johnson, E. W., Agnone, J., and McCarthy, J. D. (2010). Movement Organizations,

Synergistic Tactics and Environmental Public Policy. Social Forces, 88(5):2267–

2292.

Johnson, S. (2006). The Ghost Map: The Story of London’s Most Terrifying

Epidemic–and How It Changed Science, Cities, and the Modern World. River-

head Books, New York.

BIBLIOGRAPHY 160

Kaiser, R. (2013). Act of Congress: How America’s Essential Institution Works, and

How It Doesn’t. Alfred A. Knopf, New York.

Kamenica, E. and Gentzkow, M. (2011). Bayesian Persuasion. American Economic

Review, 101(6):2590–2615.

Kane, M. (2003). Social Movement Policy Success: Decriminalizing State Sodomy

Laws, 1969–1998. Mobilization: An International Quarterly, 8(3):313–334.

Kauffman, S. A. (1993). The Origins of Order: Self Organization and Selection in

Evolution. Oxford University Press, New York.

Kelley, T. and Littman, J. (2001). The Art of Innovation: Lessons in Creativity from

IDEO, America’s Leading Design Firm. Currency/Doubleday, New York.

King, B. G., Bentele, K. G., and Soule, S. A. (2007). Protest and Policymaking:

Explaining Fluctuation in Congressional Attention to Rights Issues, 1960–1986.

Social Forces, 86(1):137–163.

King, B. G., Cornwall, M., and Dahlin, E. C. (2005). Winning Woman Suffrage One

Step at a Time: Social Movements and the Logic of the Legislative Process. Social

Forces, 83(3):1211–1234.

King, D. C. (1994). The Nature of Congressional Committee Jurisdictions. American


King, D. C. (1997). Turf Wars: How Congressional Committees Claim Jurisdiction.


Kingdon, J. W. (1984). Agendas, Alternatives, and Public Policies. Little, Brown,

Boston.

BIBLIOGRAPHY 161

Knudsen, T. and Levinthal, D. A. (2007). Two Faces of Search: Alternative Gener-

ation and Alternative Evaluation. Organization Science, 18(1):39–54.

Koempel, M. L. and Schneider, J. (2007). Congressional Deskbook: The Practical

and Comprehensive Guide to Congress. TheCapitol.Net.

Kogut, B. and Zander, U. (1992). Knowledge of the Firm, Combinative Capabilities,

and the Replication of Technology. Organization Science, 3(3):383–397.

Kollman, K. (1998). Outside Lobbying: Public Opinion and Interest Group Strategies.

Princeton University Press, Princeton.

Kosar, K. R. (2015). Why I Quit the Congressional Research Service. Washington

Monthly.

Krehbiel, K. (1988). Spatial Models of Legislative Choice. Legislative Studies Quar-

terly, 13(3):259–319.

Krehbiel, K. (1991). Information and Legislative Organization. University of Michi-

gan Press, Ann Arbor.

Krehbiel, K. (1997). Pivotal Politics: A Theory of U.S. Lawmaking. University of

Chicago Press, Chicago.

Krehbiel, K. (2004). Legislative Organization. Journal of Economic Perspectives,

18(1):113–128.

Lashinsky, A. (2006). Chaos By Design. Fortune.

Levinthal, D. A. (1997). Adaptation on Rugged Landscapes. Management Science,

43(7):934–950.

BIBLIOGRAPHY 162

Levinthal, D. A. and March, J. G. (1993). The Myopia of Learning. Strategic

Management Journal, 14:95–112.

Levitt, B. and March, J. G. (1988). Organizational Learning. Annual Reviews of

Sociology, 14:319–40.

Leyden, K. M. (1995). Interest Group Resources and Testimony at Congressional

Hearings. Legislative Studies Quarterly, 20(3):431–439.

Lindbergh, B. and Miller, S. (2016). The Only Rule Is It Has To Work. Henry Holt

and Co., New York. Apple E-book.

Lippman, S. A. and McCall, J. J. (1976). The Economics of Job Search: A Survey.

Economic Inquiry, 14(2):155–189.

Lipsky, M. (1968). Protest as a Political Resource. American Political Science

Review, 62(04):1144–1158.

Lohmann, S. (1993). A Signaling Model of Informative and Manipulative Political

Action. American Political Science Review, 87(02):319–333.

Lyon, T. P. and Maxwell, J. W. (2008). Corporate Social Responsibility and the

Environment: A Theoretical Perspective. Review of Environmental Economics

and Policy, 2(2):240–260.

March, J. G. (1962). The Business Firm as a Political Coalition. Journal of Politics,

24(4):662–678.

March, J. G. (1991). Exploration and Exploitation in Organizational Learning. Or-

ganization Science, 2(1):71–87.

BIBLIOGRAPHY 163

March, J. G. (2006). Rationality, Foolishness, and Adaptive Intelligence. Strategic

Management Journal, 27(3):201–214.

March, J. G. (2010). The Ambiguities of Experience. Cornell University Press, Ithaca.

March, J. G. and Shapira, Z. (1987). Managerial Perspectives on Risk and Risk

Taking. Management Science, 33(11):1404–1418.

March, J. G. and Simon, H. A. (1958). Organizations. Wiley, New York.

Matthews, S. A. (1989). Veto Threats: Rhetoric in a Bargaining Game. Quarterly

Journal of Economics, 104(2):347–369.

McAdam, D. and Su, Y. (2002). The War at Home: Antiwar Protests and Congres-

sional Voting, 1965 to 1973. American Sociological Review, pages 696–721.

McCammon, H. J., Campbell, K. E., Granberg, E. M., and Mowery, C. (2001). How

Movements Win: Gendered Opportunity Structures and U.S. Women’s Suffrage

Movements, 1866 to 1919. American Sociological Review, 66(1):49–70.

Mebane Jr., W. R. and Sekhon, J. S. (2011). Genetic Optimization Using Derivatives:

The Rgenoud Package for R. Journal of Statistical Software, 42(11):1–26.

Milgrom, P. and Roberts, J. (1988). An Economic Approach to Influence Activities

in Organizations. American Journal of Sociology, 94(Supplement: Organizations

and Institutions: Sociological and Economic Approaches to the Analysis of Social

Structure):S154–S179.

Mitchell, G. J. (2007). Report to the Commission of Baseball of an Independent

Investigation into the Illegal Use of Steroids and Other Performance Enhancing

BIBLIOGRAPHY 164

Substances by Players in Major League Baseball. Technical report, DLA Piper US

LLP.

Moe, T. M. (2012). Delegation, Control, and the Study of Public Bureaucracy. In

The Forum, volume 10.

Morales, D. M. (2013). Policymaking as Gift Exchange: An Ethnographic Study of

Lobbyists. Dissertation, Stanford University, Stanford, CA.

Morrill, C. (1991). Conflict Management and Organizational Change. American

Journal of Sociology, 97(3):585–621.

Morrill, C. (1995). The Executive Way: Conflict Management in Organizations.


Mullen, K. M. (2014). Continuous Global Optimization in R. Journal of Statistical

Software, 60(6):1–45.

Nelkin, D. (1975). The Political Impact of Technical Expertise. Social Studies of

Science, 5(1):35–54.

Nelson, R. R. and Winter, S. G. (1982). An Evolutionary Theory of Economic

Change. The Belknap Press of Harvard University Press, Cambridge.

Oleszek, W. J. (1989). Congressional Procedures and the Policy Process. Congres-

sional Quarterly Press, Washington, 3rd edition.

Olzak, S. and Soule, S. A. (2009). Cross-Cutting Influences of Environmental Protest

and Legislation. Social Forces, 88(1):201–225.

O’Reilly, C. A. and Tushman, M. L. (2013). Organizational Ambidexterity: Past,

Present, and Future. Academy of Management Perspectives, 27(4):324–338.

BIBLIOGRAPHY 165

Page, B. I. and Shapiro, R. Y. (1983). Effects of Public Opinion on Policy. American


Pfeffer, J. (1981). Power in Organizations. Ballinger Pub. Co., Cambridge.

Pfeffer, J. and Sutton, R. I. (2006a). Evidence-based Management. Harvard Business

Review.

Pfeffer, J. and Sutton, R. I. (2006b). Hard Facts, Dangerous Half-Truths, and Total

Nonsense. Harvard Business School Press, Boston.

Pondy, L. R. (1992). Reflections on Organizational Conflict. Journal of Organiza-

tional Behavior, 13:257–261.

Rao, H., Morrill, C., and Zald, M. N. (2000). Power Plays: How Social Movements

and Collective Action Create New Organizational Forms. Research in Organiza-

tional Behavior, 22:239–82.

Rivkin, J. W. (2000). Imitation of Complex Strategies. Management Science,

46(6):824–844.

Rivkin, J. W. and Siggelkow, N. (2003). Balancing Search and Stability: Inter-

dependencies Among Elements of Organizational Design. Management Science,

49(3):290–311.

Roberts, J., Saloner, G., and Gibbons, R. (2012). Strategy in Organizations. In

Gibbons, R. and Roberts, J., editors, The Handbook of Organizational Economics,

pages 373–431. Princeton University Press, Princeton.

Rome, A. (2013). The Genius of Earth Day: How a 1970 Teach-in Unexpectedly

Made the First Green Generation. Hill and Wang, New York.

BIBLIOGRAPHY 166

Sabatier, P. A. (1991). Toward Better Theories of the Policy Process. PS: Political

Science & Politics, 24(2):147–156.

Sachs, R. C. (2003). A Guide to Congressional Hearings. Novinka Books, New York.

Sah, R. K. and Stiglitz, J. E. (1986). The Architecture of Economic Systems: Hier-

archies and Polyarchies. American Economic Review, 76(4):716–727.

Santoro, W. A. (2002). The Civil Rights Movement’s Struggle for Fair Employment:

A “Dramatic Events–Conventional Politics” Model. Social Forces, 81(1):177–206.

Santoro, W. A. (2008). The Civil Rights Movement and the Right to Vote: Black

Protest, Segregationist Violence and the Audience. Social Forces, 86(4):1391–1414.

Schilling, M. A. and Fang, C. (2014). When Hubs Forget, Lie, and Play Favorites: In-

terpersonal Network Structure, Information Distortion, and Organizational Learn-

ing. Strategic Management Journal, 35(7):974–994.

Schlozman, K. L. and Tierney, J. T. (1986). Organized Interests and American

Democracy. Harper and Row, New York.

Scott, W. R. (1987). Organizations. Prentice Hall, Englewood Cliffs, NJ.

Shaiko, R. G. (1999). Voices and Echoes for the Environment: Public Interest Rep-

resentation in the 1990s and Beyond. Columbia University Press, New York.

Shapira, Z. (1995). Risk Taking: A Managerial Perspective. Russell Sage Foundation.

Shepsle, K. A. (1992). Congress Is a “They,” Not an “It”: Legislative Intent as

Oxymoron. International Review of Law and Economics, 12(2):239–256.

BIBLIOGRAPHY 167

Shepsle, K. A. and Weingast, B. R. (1981). Structure Induced Equilibrium and

Legislative Choice. Public Choice, 37:503–519.

Shepsle, K. A. and Weingast, B. R. (1987). The Institutional Foundations of Com-

mittee Power. American Political Science Review, 81(1):85–104.

Simon, H. A. (1955). A Behavioral Model of Rational Choice. Quarterly Journal of

Economics, 69(1):99–118.

Simon, H. A. (1976). Administrative Behavior: A Study of Decision-making Processes

in Administrative Organization. Free Press, New York, 3rd edition edition.

Smith, S. S. and Deering, C. J. (1997). Committees in Congress. CQ Press, Wash-

ington, D.C., 3rd edition.

Soule, S. and King, B. (2006). The Stages of the Policy Process and the Equal Rights

Amendment, 1972–1982. American Journal of Sociology, 111(6):1871–1909.

Soule, S., McAdam, D., McCarthy, J., and Su, Y. (1999). Protest Events: Cause

or Consequence of State Action? The U.S. Women’s Movement and Federal Con-

gressional Activities, 1956–1979. Mobilization, 4(2):239.

Soule, S. A. (2004). Going to the Chapel? Same-sex Marriage Bans in the United

States, 1973-2000. Social Problems, 51(4):453–477.

Soule, S. A. (2009). Contention and Corporate Social Responsibility. Cambridge

University Press, New York.

Soule, S. A. (2012). Social Movements and Markets, Industries, and Firms. Organi-

zation Studies, 33(12):1715–1733.

BIBLIOGRAPHY 168

Soule, S. A. and Olzak, S. (2004). When Do Movements Matter? The Politics of

Contingency and the Equal Rights Amendment. American Sociological Review,

69(4):473–497.

Stone, C. D. (1993). The Gnat Is Older Than Man: Global Environment and Human

Agenda. Princeton University Press, Princeton, N.J.

Sutton, R. I. and Hoyt, D. (2016). Better Service, Faster: A Design Thinking Case

Study. Harvard Business Review.

Talbert, J. C., Jones, B. D., and Baumgartner, F. R. (1995). Nonlegislative Hearings

and Policy Change in Congress. American Journal of Political Science, 39(2):383–

405.

Thompson, J. D. (1967). Organizations in Action: Social Science Bases of Admin-

istrative Theory. Transaction Publishers, New Brunswick.

Tushman, M. L. and Anderson, P. (1986). Technological Discontinuities and Orga-

nizational Environments. Administrative Science Quarterly, 31(3):439–465.

Tushman, M. L. and O’Reilly, C. A. (1996). The Ambidextrous Organizations: Man-

aging Evolutionary and Revolutionary Change. California Management Review,

38(4):8–30.

Utterback, J. (1994). Mastering the Dynamics of Innovation: How Companies Can

Seize Opportunities in the Face of Technological Change. Harvard Business School

Press, Boston.

Vancil, R. F. (1979). Decentralization: Managerial Ambiguity by Design. Dow Jones-

Irwin, Homewood, Ill.

BIBLIOGRAPHY 169

Walker, J. L. (1991). Mobilizing Interest Groups in America: Patrons, Professions,

and Social Movements. University of Michigan Press, Ann Arbor.

Weber, K., Rao, H., and Thomas, L. G. (2009). From Streets to Suites: How the Anti-

Biotech Movement Affected German Pharmaceutical Firms. American Sociological

Review, 74(1):106–127.

Weber, M. (1978). Economy and Society: An Outline of Interpretive Sociology.

University of California Press, Berkeley.

Weiss, C. H. (1992). Helping Government Think: Functions and Consequences of

Policy Analysis Organizations. In Organizations for Policy Analysis. Sage Publi-

cations, Newbury Park, CA.

Wilson, J. Q. (1973). Political Organizations. New York, Basic Books,.

Wright, J. R. (1995). Interest Groups and Congress: Lobbying, Contributions, and

Influence. Allyn & Bacon, Boston.

Xiang, Y., Gubian, S., Suomela, B., and Hoeng, J. (2013). Generalized Simulated

Annealing for Global Optimization: The GenSA Package. R Journal, 5(1):13–28.

Zald, M. N. (1970). Organizational Change: The Political Economy of the YMCA.


Zbaracki, M. J. and Bergen, M. (2010). When Truces Collapse: A Longitudinal

Study of Price-adjustment Routines. Organization Science, 21(5):955–972.

three essays on organizational politics, search, and

Documents