Group Title: experimental investigation of managers' escalation errors
Title: An experimental investigation of managers' escalation errors
CITATION THUMBNAILS PAGE IMAGE ZOOMABLE
Full Citation
STANDARD VIEW MARC VIEW
Permanent Link: http://ufdc.ufl.edu/UF00102729/00001
 Material Information
Title: An experimental investigation of managers' escalation errors
Physical Description: Book
Language: English
Creator: Radtke, Robin Rae, 1963-
Copyright Date: 1992
 Record Information
Bibliographic ID: UF00102729
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
Resource Identifier: oclc - 27779164
ltuf - AJM7678

Full Text












AN EXPERIMENTAL INVESTIGATION OF
MANAGERS' ESCALATION ERRORS

















By

ROBIN RAE RADTKE


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA


1992















ACKNOWLEDGEMENTS


I would like to thank the members of my dissertation

committee for all of their help: Professor Bipin Ajinkya

(Chairman), Professor Doug Snowball, Professor Jeffrey Yost

and Professor Ronald Randles. Their guidance and words of

wisdom were greatly appreciated throughout this project.

Additionally, I am grateful to all the professors from the

Fisher School of Accounting and the Department of Finance,

Insurance and Real Estate who supplied me with class time to

recruit subjects. My experimental assistants, Tom Bristow,

Steve Cox, Dominique Marchand, Sean Robb, Sherry Ropp and

Sean Williams deserve praise for helping me administer the

experiment. The student subjects who participated in the

experiment, most of whom were very good natured, are

acknowledged for giving their time and effort. Last but not

least, I would like to thank my family and friends for their

love and support.
















TABLE OF CONTENTS

page

ACKNOWLEDGEMENTS........................... ........... ii

ABSTRACT.............................................. v

CHAPTERS

1 INTRODUCTION AND BACKGROUND................... 1

Introduction................................ 1
Motivation for the Research................. 3
Overview of Research Method and
Organization.............................. 5

2 THE ESCALATION ERROR PROBLEM AND RELATED
STUDIES..................................... 6

Introduction................................ 6
The Accounting Issue.......................... 7
The Principal-Agent Setting................. 9
The Issue of Information Asymmetry.......... 12
The Formation of the Manager's Reputation... 14
Related Escalation Error Literature.......... 15
The Present Research........................ 20

3 DEVELOPMENT OF MODELS AND HYPOTHESES.......... 22

Introduction................................ 22
The Principal-Agent Model................... 22
The Information System...................... 27
The Effect of the Manager's Talent Level.... 31
Information Asymmetry Effects of the
Monitoring Variable ....................... 33
The Effects of Timing of Information
Disclosure........ ........................ 36
Effects of the Environment.................. 37
The Decision Model............... ............ 38
Supplementary Analyses...................... 39


iii









4 RESEARCH DESIGN AND DATA ANALYSIS METHODS..... 40

Introduction................................ 40
Experimental Design......................... 40
Overview.................................. 40
Experimental Task.......... ..... ........... 41
Subjects................................. 48
Incentive Structure....................... 50
Operationalization of Independent
Variables............................... 51
Measurement of Dependent Variable......... 54
Data Analysis Methods....................... 55
The Decision Model............ ............. 55
Predicted Relationships................... 57
Supplementary Analyses.................... 58
Summary..................................... 62

5 RESEARCH RESULTS .............................. 67

Introduction................ ................ 67
Results of Hypothesis Testing............... 67
Additional Testing.......................... 74
Supplementary Analyses.................... 74
Manipulation Checks....................... 81
Results for Revised Experiment............ 86
Summary..................................... 92

6 SUMMARY AND CONCLUSIONS....................... 112

Summary of the Study........................ 112
Summary and Discussion of Results............ 112
Contributions and Implications.............. 116
Contributions to Managerial Accounting
Research.................................. 116
Implications for Understanding
Escalation Errors........................ 118
Limitations................................. 120
Directions for Future Research.............. 120

APPENDICES

A EXPERIMENTAL PROTOCOL......................... 122

B EXPERIMENTAL INSTRUMENTS...................... 126

C CALCULATION OF LOGICAL PAYOFFS AND BIDS........ 146

REFERENCES............................................ 150

BIOGRAPHICAL SKETCH................................... 152















Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

AN EXPERIMENTAL INVESTIGATION OF
MANAGERS' ESCALATION ERRORS

By

Robin Rae Radtke

August 1992

Chairman: Bipin B. Ajinkya
Major Department: Accounting

An escalation error results when the manager continues

with a course of action that has proven to be nonoptimal.

These errors can represent costly losses to the firm. The

purpose of this study is to investigate the source of

managers' escalation errors.

It was hypothesized that when the manager commits an

escalation error, he may be attempting to conceal his

previous incorrect choice and thereby protect his

reputation. The protection of the manager's reputation may

provide greater benefits to him than the benefits resulting

from making a correct choice. A principal-agent model was

developed to examine how the reputation value of the manager

influences the manager's decisions in an escalation error

scenario. Four factors were posited to directly influence

escalation behavior: the talent level of the manager, the








presence of a monitor, the time at which the manager

discovers he has made an escalation error and the difficulty

level surrounding the manager's decision.

The model was modified to fit an experimental setting

using student subjects as surrogates for actual managers and

supervisors. A manager was expected to escalate when he was

untalented, had no monitor present, discovered that he had

made an escalation error at a late stage and faced difficult

environmental factors.

The hypotheses were tested using a logit model. The

results for the three independent variables--the manager's

talent level, the time at which the manager discovers he has

made an escalation error and the difficulty level

surrounding the manager's decision--support the predictions.

The expected effect of the presence of a monitor was not

supported. A revised experiment was run with some

modifications to re-test the effect of the monitor. The

results of these sessions again showed an insignificant

effect for the monitor construct. These experimental

results may help researchers gain new insight into the

escalation error problem in designing further studies.















CHAPTER 1
INTRODUCTION AND BACKGROUND


Introduction


This dissertation investigates the escalation error

phenomenon. An escalation error results when an individual

continues with a course of action that has proven to be

nonoptimal for the individual. Behavioral studies

constitute the bulk of the previous research in this area.1

These studies have isolated several factors associated with

escalation errors. The purpose of this study is to examine

the causes of escalation errors from an accounting

standpoint.

The causes of escalation errors cannot be identified

without a thorough understanding of the escalation error

problem. The definition of an escalation error must be

applied carefully to each case, as what constitutes

nonoptimality, and evidence of nonoptimality, will vary

between cases. In general, negative feedback that leads an

individual to believe an alternative course of action would

produce a more favorable outcome constitutes evidence of

nonoptimality. If after receiving this information the


1 See Brockner (1992) for a good review of this
literature.











individual continues with the originally chosen course of

action, an escalation error is said to exist.

Escalation errors are not unique to any one setting.

The Vietnam War is a prominent example that demonstrates how

potentially disastrous an escalation error can be. The

following memo emerged in the early stages of the conflict:

The decision you face now is crucial. Once large
numbers of U.S. troops are committed to direct combat,
they will begin to take heavy casualties in a war they
are ill-equipped to fight in a non-cooperative if not
downright hostile country-side. Once we suffer large
casualties, we will have started a well-nigh
irreversible process. Our involvement will be so great
that we cannot--without national humiliation--stop
short of achieving our complete objectives. Of the two
possibilities, I think humiliation would be more likely
than the achievement of our objectives--even after we
have paid terrible costs. (Memo from George Ball to
President Lyndon Johnson, July, 1965; source: The
Pentagon Papers, 1971.)2

This memo shows the sentiments of Ball and undoubtedly many

others. The war continued on for many years, however, even

with these strong assertions of the impending horrors ahead.

This example shows the depths to which individuals can

become entrenched in escalation errors and points out the

importance of recognizing negative feedback to avoid an

imminent escalation error.

Escalation errors in business may be very costly as

well. Many business decisions could lead to escalation

errors, but one in particular is the focus of this study.



2 This example is referenced in Staw (1976, 1981) and
Staw and Fox (1977) as an important case of an escalation
error.










The choice between alternative investment projects, or more

specifically the capital budgeting replacement decision, is

examined. This scenario provides the opportunity to

investigate the escalation error phenomenon in a business

setting. By doing so, the determinants which cause a

manager to engage in escalation behavior may be isolated and

the relationships between these determinants can be

investigated to provide additional insight into the problem

of escalation errors.


Motivation for the Research


When capital budgeting decisions are made with

inappropriate information, an inefficient allocation of the

firm's available capital results which is inconsistent with

the business goal of maximizing firm value. Consequently,

the issue of determining what causes the manager to make

these decision errors is of interest. Empirical evidence

suggests that escalation errors also occur in business

settings.3 Explanations from the body of psychological


3 DeBondt and Makhija (1988) and Statman and Sepe
(1989) are examples of empirical studies. Both of these
studies look at share price movements in reaction to the
specified events. DeBondt and Makhija test the validity of
the sunk cost hypothesis as a basis for committing
escalation errors in the context of the U.S. nuclear power
program. They examine the market reaction to all plant
completions and cancellations (over $50 million) prior to
March 1984. Their results are mixed and do not support a
'powerful' sunk cost effect. They point to prudency reviews
ordered by U.S. Public Service Commissions, however, as
anecdotal evidence of costs incurred and supported by the
sunk cost hypothesis. Statman and Sepe examine the reaction










research present many theories for the occurrence of

escalation errors. Kanodia, Bushman and Dickhaut (1989)

offer an alternative explanation based on economic

rationality.

Behavioral studies have tested the hypothesized

psychological factors. Many of these studies have actually

used business or investment decision contexts.4 The

current study incorporates factors of interest from a

managerial accounting standpoint (agency aspects) and

represents an empirical test of the Kanodia, Bushman and

Dickhaut (1989) theory. By designing this study with

accounting and economic factors specifically in mind,

additional insight into understanding escalation errors in

terms of accounting can be gained.

The variables of interest in this study that have not

been previously investigated (in an experimental study) are

the talent level of the manager, the necessity and the

impact of information asymmetry, the effect of the timing of

information disclosure and the difficulty level surrounding

the manager's decision. In an experimental setting each of



of the market to project termination announcements. They
find the mean market-adjusted return to be positive. They
conclude that on average shareholders consider project
termination announcements good news, because managers will
no longer throw good money after bad (and are finally
abandoning the sunk costs of failed projects).

4 Staw (1976) and Staw and Fox (1977) both use the
allocation of research and development funds between two
competing divisions as their financial decision case.









5

these variables can be manipulated so a fully crossed design

can be established. This allows the impact of each factor

to be analyzed. As a result, issues more specifically

related to accounting can be investigated in the escalation

error scenario.


Overview of Research Method and Organization


The impact of the variables of interest on the

escalation error decision was examined in an experimental

study. The true purpose of the study was disguised such

that subjects knew only that they were involved in an

experiment dealing with investment project choices.

Subjects earned lottery tickets based on decisions made in

the study. The tickets were entered into a lottery with

seven cash prizes.

The chapters are organized as follows. Chapter 2

discusses the escalation error problem and related

literature. The research models and hypotheses used in the

current study are developed in Chapter 3. Chapter 4

describes the research design used to test the hypotheses.

The results of the experiment are described in Chapter 5.

Chapter 6 includes a summary of the study and conclusions.















CHAPTER 2
THE ESCALATION ERROR PROBLEM AND RELATED STUDIES


Introduction


The discussion in this chapter focuses on the problem

of managers committing escalation errors when making

investment project or capital budgeting replacement

decisions. The accounting issue is discussed first. An

analysis of relevant and irrelevant accounting information

governs how the manager should choose between alternative

capital assets. If the manager allows irrelevant accounting

information such as sunk costs to influence his choice, he

may commit an escalation error and stray from the optimal

choice. The framework for the current study is presented

next. This includes a discussion of the principal-agent

setting and the related issues of information asymmetry and

the formation of the manager's reputation. This framework

provides the setting within which a manager can justify an

escalation error in terms of a benefit to his reputation

value. The chapter then considers the related escalation

literature in the accounting and psychology fields.

Finally, the link between the adopted framework and the

current study is presented.










The Accounting Issue


As discussed in Chapter 1, an escalation error results

when an individual continues with a course of action that

has proven to be nonoptimal for the individual. The

individual of interest from an accounting standpoint in this

study is the manager. The manager continually makes

important decisions that affect both the well being of the

firm and his career. The accounting information the manager

has at hand governs his choices. Appropriate use of this

information should lead to optimal choices. Misuse of this

information may lead to escalation errors. Determining what

information the manager should use and how he should use it

is the crux of this discussion.

Accounting information for managerial purposes consists

of projected and historical data. An old management

accounting adage states that only relevant information

should be used in the decision making process from the

viewpoint of the firm. Information is considered irrelevant

when it is equal for all decision alternatives. Thus,

historical data from past decisions is considered irrelevant

information. Past costs and revenues may be useful in

predicting future costs and revenues, but have no direct

relevance for future decisions.1




I See Horngren and Foster (1987) chapter nine for a
good discussion of this issue.










One problem facing the manager on a recurring basis is

the capital budgeting replacement decision. This decision

entails the replacement of an old capital asset (such as a

machine) with a new asset. Once the decision to replace an

old asset has been made, the primary issue becomes what type

of new asset to purchase. Many cost figures are available

when considering this problem.2 They include the cost of

the new asset, the disposal value of the old asset, the

purchase price of the old asset and the book value of the

old asset. Only the relevant costs in the above list should

be utilized.

If several new assets are being considered for

purchase, then the cost of the replacement asset will vary

across decision choices and as such is relevant information.

The disposal value, the purchase price and the book value of

the old asset do not vary across current decision

alternatives and are irrelevant information. The disposal

value is equal to the expected proceeds of the sale of the

old asset. The purchase price and the book value of the old

asset are based on historical information. The purchase

price of the old asset is considered a sunk cost,3 since


2 Tax considerations inherently play a role in the
capital budgeting replacement decision. In the current
study tax issues are ignored to make the analysis more
tractable.

3 Sunk costs are defined as costs which have been
incurred in the past and cannot be changed by future
actions. As such, sunk costs are always irrelevant.
Individuals, and managers in particular, often fixate on











the cost of the old asset has already been incurred and

cannot be recovered (except through productive use of the

asset). The book value of the old asset (the purchase price

of the old asset less accumulated depreciation) is also

irrelevant information since it is a meaningless combination

of a historical cost and an arbitrary accumulated expense

account. Therefore, once the manager has decided to replace

an old asset, the only relevant information he should

consider is the cost of the replacement asset.4'5


The Principal-Agent Setting


From a normative standpoint, the analysis of the

relevance of various pieces of accounting information is

always undertaken from the point of view of the firm. When

the manager uses the appropriate relevant information, the


sunk costs since they feel responsible for the capital
outlay. This fixation can lead to escalation errors when
the manager refuses to replace an old asset that has not
earned an acceptable return on the initial investment.

4 The other cost figures mentioned may be important in
making the decision of when to replace the old asset. They
would also be salient in a setting where tax considerations
are involved. In the current study however, neither
condition is applicable. Once the manager realizes the old
asset should be replaced, the decision between replacement
assets is the relevant issue in the escalation error
scenario.

5 In a broader context, many other relevant items are
considered by the manager when making the asset replacement
decision. Such factors as differences in technologies and
production capacities between potential replacement assets
are obviously evaluated. Additionally, purchase terms and
agreements with different companies may enter into the
decision.










firm value is maximized. This analysis leaves no room for

the commitment of escalation errors. Empirical evidence

suggests, however, that escalation errors do occur in

business settings.6 The proposed framework based upon a

principal-agent setting recognizes the opportunity for the

occurrence of escalation errors.

In the principal-agent setting, the economic relevance

of information may change when agency costs between

principals (owners) and agents (managers) are considered.

In a strict principal-agent problem, agents attempt to

maximize their personal utilities. In the pure context of

the business decision, on the other hand, the manager should

attempt to maximize the utility of the firm, since the

manager is hired as an agent of the firm and takes on the

responsibility of acting in the best interests of the firm.

In some cases the personal utility of the manager and the

utility of the firm may not be maximized by the same course

of action. When the manager puts the maximization of his

personal utility before that of the firm, a divergence

between these two utilities may lead to suboptimal decisions

for the firm. Within this context an escalation error is

defined as continuing with a course of action that is

nonoptimal for the firm, but optimal for the manager.





6 See DeBondt and Makhija (1988) and Statman and Sepe
(1989) for examples of empirical studies.








11

The maximization of personal utility on the part of the

manager is in accordance with the manager's desire to

establish a reputation. This is supported by the

proposition of Holmstrom and Ricart i Costa (1986) that the

manager's compensation is a function of both the firm value

and the manager's reputation value. They assert that some

sort of contracting is necessary in order to prevent the

manager from fixating solely on his personal utility. They

develop a model of the manager's reputation based on

learning. Once the manager realizes that his performance is

used as a signal of his competence, he tries to influence

the evaluation process through his choice of actions. A

dilemma arises when the financial value and reputation value

of his actions are different. The difference in these

values causes the preferences of managers and shareholders

(owners) to diverge.

From the point of view of the firm, investment projects

should be evaluated on the basis of net present value.7

Narayanan (1985) posits that managers, on the other hand,

often prefer investment projects that pay off quickly, and

look to additional measures such as the payback criterion



7 The net present value analysis is the generally
accepted method in capital budgeting. This coincides with
the firm's objective to maximize future inflows given a
limited amount of resources. Many alternative analysis
techniques using cash flows, net incomes, interest rates and
time periods also exist. The true value of an investment
project to the firm is not disclosed, however, when using
these alternative methods of analysis.










for justification. By choosing investment projects with

earlier payoffs, managers are attempting to accelerate the

enhancement of their human capital value. This behavior

could be detrimental to the firm if those projects that have

the shortest payback period are not the projects with the

highest net present value (given equivalent risk). The

manager may then be sacrificing long-term financial benefits

of the firm for short-term gains to his reputation value.

The escalation error scenario represents a situation where

the financial value and reputation value of investment

project decisions may be inversely related. If the manager

chooses and continues with the nonoptimal project in

attempts to receive quicker payoffs, his reputation value

may be enhanced, while the firm value decreases in the long

run. This study investigates under what circumstances the

manager's reputation value is more important to him than the

firm value in the escalation error scenario.


The Issue of Information Asymmetry


In a principal-agent setting the issue of information

asymmetry is particularly important. Information asymmetry

exists when one individual (usually the agent) has more

information than another (usually the principal). In this

case the agent may be motivated to use his superior

information to his benefit, at the expense of the principal.

In the current scenario, if the manager has private










information on which to base his asset replacement or

investment project decision, he may be able to act

nonoptimally from the point of view of the firm without

direct detection.8 Without information asymmetry the

incentive for the manager to commit escalation errors

disappears. In a world of only public information the

interests of the manager and the firm coincide and the

manager cannot form a reputation spuriously. Consequently,

escalation errors may be viewed as a cost of the agency

relationship.

Narayanan (1985) shows that only under asymmetric

information does the manager have incentives to make

suboptimal decisions which are consistent with short-term

goals instead of long-term goals. Suboptimal decisions may

yield short-term profits, but are not in the best interests

of the firm in the long-term. In garnering short-term

profits the manager is attempting to bolster his reputation,

and therefore his wages. This is an example of a case in

which the personal utility of the manager becomes more

important to him than the utility of the firm under

asymmetric information. Without information asymmetry the

agency contract causes the personal utility of the manager

and the utility of the firm to coincide. Consequently, for



8 A device such as a monitor, which reports
information to the principal independent of the agent, may
be used to reduce the level of information asymmetry between
the two parties.










this study, information asymmetry is considered a

prerequisite for committing escalation errors.


The Formation of the Manager's Reputation


Since multiple studies suggest the manager may be

concerned about his reputation in an information asymmetric

setting,9 the formation of the manager's reputation is an

important issue. Kreps and Wilson (1982) show how imperfect

information promotes reputation formation. They posit that

uncertainty of players about the payoffs of other players is

the simplest type of imperfect information that yields

reputation formation in finitely repeated games. This is

not an extreme condition, as players' payoffs typically are

not common knowledge.

The theory of reputation formation under imperfect

information is tested experimentally by Camerer and Weigelt

(1988). They posit that the behavior of players in an

experiment can be predicted by a sequential equilibrium

model of reputation formation in an incomplete information

repeated game. Evidence of reputation-building is found in

all of their experimental sessions of an abstracted lending

game. They conclude that incomplete information about a

player's type (or privately known characteristics) leads





9 See Holmstrom and Ricart i Costa (1986) and
Narayanan (1985).








15

other players in the game to form beliefs about the player's

type. These beliefs become the player's reputation.10

Imperfect or incomplete information is usually present

in a typical principal-agent setting. In the current

context, managers invariably have more information available

to them than do the owners of the firm. This condition

lends itself to reputation formation on the part of the

manager and to possible escalation errors if there are

apparent gains to the manager's personal utility.


Related Escalation Error Literature


Kanodia, Bushman and Dickhaut (1989) model the

escalation error problem. They posit that the manager may

engage in escalation behavior in an attempt to maintain

uncertainty about his true management talent. When others

can only infer the talent level of the manager from his

actions and the resulting consequences, these actions

acquire a reputation value.11 Consequently, the manager

will avoid switching to an alternative (superior) course of



10 In the Camerer and Weigelt (1988) experiment some
players were borrowers and some were lenders. It was
posited that borrowers should be concerned about
establishing a good reputation for repayment of funds
borrowed in order to secure future amounts. The results
showed that the majority of the borrowers did not renege at
the first opportunity they had, or as often as predicted.
This constitutes evidence of reputation-building behavior.

11 This condition fits the imperfect or incomplete
information criteria of Kreps and Wilson (1982) and Camerer
and Weigelt (1988).










action to avoid admitting a past mistake and damaging his

reputation as a manager. By maintaining the appearance of

good management, the manager's opportunities in a labor

market should improve. This theory of the reputation value

of the manager is an economic rationale for escalation

errors.

Other studies have investigated escalation errors from

a behavioral standpoint. These studies do not explain

escalation errors in terms of an individual's reputation

value, but rather in terms of the psychological effects of

various factors on the individual's attempt to rationalize

his errant behavior. Empirical evidence suggests that

individuals are generally prone to take sunk costs into

account and commit escalation errors when they make

decisions. Staw (1976) states that individuals seek to

maintain or restore the appearance of rationality to their

previous behavior. This is known as self-justification

theory. The theory posits that individuals cognitively

distort negative consequences into a perceived positive

outcome. Therefore, an individual who has made a choice

that led to negative consequences may continue in that

course of action in an attempt to justify his prior behavior

or to demonstrate the ultimate rationality of his original

course of action. The important factors behind the theory

of self-justification in the investment decision are posited

to be the level of personal responsibility an individual had











for the decision and the outcomes resulting from the

decision. In an experimental setting Staw shows that

business school students invest substantially greater

amounts of resources when they are personally responsible

for previous negative outcomes. He concludes that they are

justifying their actions to themselves by attempting to

appear competent in previous actions as opposed to future

actions.

Staw and Fox (1977) investigate the persistence of the

escalation process over time in an experimental setting.

They design an experiment in which the commitment of

resources to a course of action is the dependent variable

and the independent variables are personal responsibility,

efficacy of resources and time. They replicate the personal

responsibility effect of Staw (1976) using undergraduate

business students. Additionally, high-responsibility

subjects show the greatest commitment to a course of action

immediately following the receipt of negative consequences.

Efficacy of resources is also significant. Under high

efficacy of resources, subjects invest substantially more

resources than under low efficacy of resources. The

investment decisions for all subjects, however, are highly

unstable over time.

Staw (1981) points out that a cost which is sunk

economically for the firm may not be sunk psychologically

for the decision maker. He develops a model of the








18

commitment process and identifies four major determinants of

commitment. These are 1) motivation to justify previous

decisions, 2) norms for consistency, 3) perceived

probability of future outcomes and 4) perceived value of

future outcomes. He states that rational and objective

decision makers should not be influenced by the first two

determinants since they are retrospective instead of

prospective determinants. Analogously, in accordance with

subjective expected utility models, the maximization of

future utility should be the goal of the decision maker.

An experimental study by Staw and Ross (1980) uses both

undergraduate students and practicing managers enrolled in

an MBA program to assess a case description of an

administrator's behavior. The three variables manipulated

in the case descriptions are consistency versus

experimentation in the administrator's course of action,

minimum versus maximum commitment of resources to the course

of action and the ultimate success versus failure of the

administrator's efforts. Main effects for each of these

independent variables are found such that an administrator

is rated highest when he follows a consistent course of

action, allocates minimum resources and is ultimately

successful. Additionally, practicing managers participating

in the study view consistency in actions as a much more

important characteristic of a successful manager than

undergraduate students, even when negative results accrue as










a consequence of previous actions. Continuing with a

previously chosen investment apparently benefits the manager

more than switching to a superior investment. This suggests

that managers consider sunk costs when making capital

budgeting decisions and are maximizing their personal

utility instead of the utility of the firm.

In summary, the previous behavioral studies have

identified personal responsibility as a crucial factor

consistent with the commitment of an escalation error.

Other factors such as the efficacy of resources, consistency

in actions and the ultimate success of the chosen course of

action have also been shown to significantly affect the

escalation decision. A model of the commitment process has

also been presented which coincides with accepted subjective

expected utility models.

The accounting study by Kanodia, Bushman and Dickhaut

(1989) provides the primary source for the hypotheses of the

current study. Hypotheses concerning the manager's talent

level as well as others pertaining to the principal-agent

setting are developed. The current study adopts much of the

theory of Kanodia, Bushman and Dickhaut (1989) and a limited

amount from the body of behavioral literature. This

research provides an empirical test of the Kanodia, Bushman

and Dickhaut (1989) theory, as well as some extensions.










The Present Research


The purpose of the current study is to integrate the

issues discussed in this chapter into a single experiment.

The analysis of the accounting issue provides the link

between the escalation error phenomenon and accounting. The

principal-agent setting and the related issues provide a

ready-made framework within which to conduct an analysis.

The previous behavioral studies have laid the groundwork for

an experimental investigation of the escalation error

phenomenon. The current study draws on these bases and the

paper by Kanodia, Bushman and Dickhaut (1989).

This study explicitly examines the effect on the

manager's decision to commit an escalation error related to

1) the talent level of the manager, 2) the presence of a

monitor of the manager, 3) the time at which the manager

discovers he has made an escalation error and 4) the

difficulty level surrounding the manager's decision. These

four factors have been chosen for the current study because

none have been included in the same form in previous studies

and all are very important from an accounting standpoint.

The importance of personal responsibility to the escalation

decision has already been shown in multiple studies (Staw,

1976; Staw and Fox, 1977). Consequently, all subjects in

the current study are always personally responsible for all

of their decisions. The relevance of consistency in actions

is recognized in the experimental instruments and is









21

discussed in detail in the experimental design section of

the paper. Therefore, these concepts from the behavioral

studies of the escalation phenomenon are not manipulated,

but are incorporated in the experimental task as constant

factors.















CHAPTER 3
DEVELOPMENT OF MODELS AND HYPOTHESES


Introduction


The following discussion first introduces a principal-

agent model which serves as a structure within which to test

the hypotheses. Second, the related information system used

in the implementation of the experiment is presented. The

development of the research hypotheses is described next.

The decision model developed to describe the anticipated

effect of each of the independent variables on the manager's

decision to commit an escalation error is introduced. The

expected effects of each of the independent variables are

presented. The chapter concludes with a discussion of some

supplementary analyses which are usually associated with

this type of experimental study.


The Principal-Agent Model


A principal-agent model is developed as the framework

within which the hypotheses of this study are tested. The

model is based on Kanodia, Bushman and Dickhaut (1989).

Their model includes constructs representing managerial

talent, information asymmetry and the time at which the

manager discovers he has made an escalation error. The

22










current model is extended and altered to include the

constructs of this study.

Kanodia, Bushman and Dickhaut (1989) model an economy

consisting of two periods in which escalation behavior

versus switching behavior is studied. In this economy,

self-employed managers must choose between two projects, A

and B, in period one. There are also two possible

underlying states of nature, 0A and 80. The project inflows

accruing to managers as a result of implementing a project

are highly dependent upon a match between the project chosen

and the actual underlying state of nature. In stage one of

period one, managers receive a private signal which reveals

a certain level of information about the underlying state of

nature. Each manager then chooses to implement either

project A or B. Talented managers have a higher probability

of discerning the actual underlying state of nature from the

information signal than do untalented managers.

At stage two of period one, each manager receives new

information that reveals the true underlying state of

nature. Each manager then has the option of continuing to

implement the chosen project or switching to the alternative

project. The differential project inflows from switching to

the correct project always exceed the cost of switching, so

that it is always beneficial to switch from the firm's

viewpoint. Consequently, the manager will escalate only

when he perceives that escalating will provide him with a










higher wage in the impending labor market. At stage three

of the first period, project inflows accrue in accordance

with the project chosen/actual state of nature combination.

In period two, each manager seeks employment. The better a

manager's results from his self-employed capacity, the

higher the wage he can demand in the job market. Thus, the

project inflows that accrued as well as the reputation value

of the manager are both involved when the manager seeks

employment. These two factors are the basis for analyzing

escalation behavior in the current model.

The current model has its basis in the Kanodia, Bushman

and Dickhaut (1989) model and is modified to fit an

empirical study. Before the beginning of the decision

rounds, the state of nature is randomly determined for each

manager, but is unknown to the managers. The managers also

receive limited information about the two alternative

investment projects. The sequence of events in each

decision round is

1) managers receive additional discriminating

information to aid them in their investment

project choices,

2) managers make their investment project choices,

3) the random factor drawing takes place for each

manager and

4) investment project wages accrue to the managers.










Managers receive wages, r, which depend upon the

investment project payoff at the end of each decision round.

There are two possible values for the wages, rL and rH,

representing low and high wages. The wages are determined

by three factors: the manager's choice of investment

projects (a), a stochastic realization of the state of

nature (s) and an additional stochastic realization of a

random factor (e), such that ri(a,s,e). The manager's

choice is between two investment projects, al and a2, and is

a function of his talent level, t, and the information

signal received, y, such that ai(t,y). The manager's talent

level impacts his choice, as the talented manager is more

likely to make a correct project choice (a choice which

leads to a high payoff). An information signal, y-E(y,,y2),

is received in each round by the manager. The signal

provides the manager with the probability of each state of

nature given the signal. The state of nature is randomly

determined between two possible states, sI and s 2 The

random factor, E, is defined as the probability of receiving

an inconsistent project payoff. An inconsistent project

payoff occurs when the manager chooses project al, the state

of nature is sI and the resulting wage is rL (or a2, S2 and

rL). This factor is introduced to allow additional

randomness in the model so the manager is not able to

discern the true state of nature early on with certainty.

The random factor starts relatively high and decreases in








26

magnitude with time. In the later rounds the probability of

receiving an inconsistent project payoff becomes minimal.

The last variable of interest to the manager is the

cost of switching projects, b. This variable is positive

and increases throughout the decision rounds. The cost of

switching is a function of the manager's wages and time such

that b(r,time). The variable b represents the penalty (the

sunk cost) of discovering the true state of nature in the

later rounds when the chosen project has begun

implementation. Since there are no project costs in the

model, the use of b seems to be a plausible way to manifest

a sunk cost effect. Therefore, in this context, b is a

relevant variable to be considered in the investment project

choice problem.

In the principal-agent framework the manager is

considered to be an expected utility maximizer. If the

manager is assumed to be risk neutral, maximizing expected

utility is equivalent to maximizing wages. Therefore, based

upon the previous discussion and definitions, the manager's

problem is

maximize {r(a,s,e) b. (1)
a, b










The Information System


In every principal-agent problem an information system

represents the various probabilities associated with the

possible combinations of the variables defined in the model.

This information system dictates which potential contracts

will be optimal. The basic information system adopted for

the current study and used in the implementation of the

experiment is presented below. The payoff structure can be

characterized as


with an e probability of not attaining. The information

signal is from an information system of the form


multiplying P(yijsj) by P(sj) gives 6j, which is


Yi\sj s 1 s 2


yj P (y, y s1) P y(Y1 s2)


Y2 PP(y21sO) P(Y2sS2)



















dividing 0i. by 0i gives 8(sj yi), which is


where P(yilsj) is the probability of signal yi given that the

state is sj, Oi. is the joint probability of signal yi and

state sj, i0 is the marginal probability of signal y,, and

6(s.jyi) is the probability of state sj given signal yi.

After receipt of the information signal, the manager's

strategy is as follows.

If yl is observed and al is chosen, then the expected wage

is 6(s1 y) (r (l-e) + rL(e)) + (s2 y1) (rL(1-e ) + rH(e)).

If yl is observed and a2 is chosen, then the expected wage

is 6(s yl) (r(l- + H()) + (s21 y) (r H(1- ) + rL(e)).

If y2 is observed and al is chosen, then the expected wage
is O( y) (rH(l-) + rL()) + (s21y2) (rL (1-C) + rH(E )).

If y2 is observed and a2 is chosen, then the expected wage
is e(s y2) (rL(I-e) + rH(E )) + 8(S21Y2) (rH(1-e ) + rL(e)).


Yi\sj s S2

Y11 012

Y2 021 822
yZ 8,, 8 I


y,\s SI S2

Y ( (S Y ) (S21Yl)

YZ 2(SI Y2) (s 2 y 2)









29

Let a represent the two possible choices when y, is observed

and / represent the two possible choices when y2 is

observed. Given Oi as previously defined, the manager's

problem as stated in Equation 1 becomes

maximize {81a + 602}. (2)
ai(Y)


The variable b is the cost of switching projects.

Presumably, the manager switches whenever he determines that

the increase in his expected wages from the alternative

project is greater than the cost of switching.

Alternatively, if

n
E E(U(rH(a,s,e) rL(a,s,e))) > b (3)
i=1

where i = 1 is the round of the switch and n = the last

round, then the manager switches.

It is evident that the manager's strategy is to

maximize his utility, or equivalently his wages if the

manager is assumed to be risk neutral. To do this he should

use all the information signals he receives over the rounds

to determine the underlying state of nature. The manager's

decision for the first round is

1) estimate the probability of each state of nature

given the signal that has been received and

2) choose the project which has a greater probability

of matching the underlying true state of nature.










After the first round, the manager's decision problem

becomes much more complex.

1) If the prior round's payoff was not high, assess

probabilities of possible causes: wrong project

was chosen or an inconsistent project payoff

occurred with probability e. If the prior round's

payoff was high, this is evidence (strong but not

conclusive) of a correct project choice.

Alternatively,

P(rH) = (0( + 022) (l-e) + (012 + 21) (e) (4)


P(rL) = (011 + 822) () + (012 + 821) (-e) (5)

Computing the probability of receiving rH or rL

given the occurrence or non-occurrence of the

random factor e gives

P ( r )
P( rH) = (6)
P(rH)



(012 + 021) (e)
((01, + 022) (l-e) + (012 + 021) ())



P ) (611 + 22) ()(7)
P(eIr ) = (7)
S (011 + 022) () + (012 + 021) (l- ))



P (l-eC rrH) + (011 + 022) (1-) (8)
( n 0 + 022) (1-C) + (012 + 021) (e))












P(l- (12 + 021) (1- )
( ((11 ) ( + 22 12 + 021) (1-))


In the current setting,

P(C:rL) > P(1-c:rL) (10)

for the first several rounds. This shows that a

low project payoff is more likely to be due to the

random factor than to an incorrect project choice

in the early rounds only, as this relationship is

reversed in the later rounds. Throughout the

rounds, however, a high project payoff is most

likely not due to the random factor, as P(erH) is

always small. This holds true for any one round

looked at independently.

2) Analyze information from all previous rounds as

well as the current round to determine whether to

remain with the previously chosen project or to

switch to the alternative project.

This model is modified somewhat to fit an experimental

setting, but its essence remains unchanged. The hypotheses

of the current study are developed within this structure.


The Effect of the Manager's Talent Level


The talent level of the manager is a primary variable

of interest. Managerial talent is typically described as

the ability to organize and supervise the various inputs









32

needed for investment and production. Kanodia, Bushman and

Dickhaut (1989) refine this definition into the idea of

"foresight." A manager with superior foresight can

anticipate future developments much sooner than a manager

who is not farsighted. This equates to the superior ability

of the talented manager to discern the true state of nature.

Talented managers are more likely to make correct choices in

the initial stages of an investment decision and are

therefore less likely to face the problem of a superior

investment being available later. Occasionally talented

managers may make initial incorrect choices and find

themselves faced with the question of whether to escalate or

switch projects. In this case they should feel less

pressure to escalate. Since they are talented and probably

already have favorable reputations, they can therefore

sustain the adverse effects of admitting a past mistake.

Consequently, talented managers make correct project choices

a higher percentage of the time than untalented managers.

Untalented managers are unable to discern the true

state of nature with certainty until much later than

talented managers. Consequently, they learn that the

alternative investment should have been chosen initially,

and are then faced with the question of whether to switch to

the superior investment (assuming that the initially chosen

investment is not at the stage of completion and can still

be abandoned). Therefore, the untalented manager is more











likely to be faced with the dilemma of whether to escalate

with the inferior project or switch to the superior project

than the talented manager. Additionally, the untalented

manager should feel more pressure to escalate in an attempt

to bolster his dubious reputation. Given this, the

following hypothesis describes the expected talent effect:

H1: Escalation errors are made more frequently by
untalented managers than by talented managers.


Information Asymmetry Effects of the Monitoring Variable


Since information asymmetry is considered a

prerequisite for committing escalation errors,' the

presence or absence of information asymmetry is not

manipulated in this study. Information asymmetry exists

between owners and managers throughout the study. When the

manager has private information, the possibility for agency

problems exists. The agency problem of particular interest

in the current study is adverse selection. An adverse

selection problem exists when the manager has incentives to

hide or misrepresent his private information. When the

manager keeps his private information to himself, he is

judged only by his observable actions and their results.

This allows the manager to choose or continue a course of

action which may provide acceptable results, while an

alternative superior course of action is available. Since



1 This was noted previously, see Narayanan (1985).










the manager has private information about the superior

course of action, he will not switch and therefore admit to

a previous incorrect choice without an incentive to do so.

Therefore, the manager escalates and is compensated based on

his acceptable results while concurrently protecting his

reputation. Since the owner knows the manager has

incentives to act nonoptimally, the question arises as to

what measures the owner may take to deter such action.2

Penno (1983) investigates several issues of information

asymmetry in managerial accounting. He specifically

addresses the issues of monitoring and reporting/incentive

systems. In a reporting/incentive system, the manager

reports some of his private information to the owner. The

owner then revises the standards by which the manager is

evaluated. A major problem with a reporting/incentive

system is that the manager may have incentives to distort

the reported information. This is possible since the

manager privately observes the information. If by

distorting the report of his private information the manager

can further his own self-interests, he may do so.



2 The possibility of signaling, while considered a
solution to the adverse selection issue, does not fit in
well with the escalation error problem. Signaling assumes
that the individual producing the signal is attempting to
honestly convey useful information to the market (see
Akerlof, 1970; Spence, 1973). In the escalation error
scenario, the manager is trying to conceal his private
information. This is contradictory to the premise of
signaling and therefore invalidates the use of signaling to
reduce information asymmetry in this study.










Participative budgeting is an example of this system.3 In

contrast, under a monitoring system the manager has no

opportunity to distort the information reported to the

owner. A monitor is a device which reports information to

the owner, independent of the manager. By Penno's

definition a monitor is inherently assumed to be noisy or an

imperfect signal of the actual information. An essential

tradeoff between a reporting/incentive system and a monitor

is manipulability versus accuracy.

In the current setting, the availability of a monitor

may act to reduce the level of information asymmetry between

the manager and the owner. If the manager knows there is a

possibility that additional discriminating information about

him will be disclosed via the monitor, he will be less

likely to commit an escalation error. This is posited in

the following hypothesis:

H2: The presence of a monitor reduces the incidence of
escalation errors.









3 Participative budgeting requires the employee to
report some of his private information to the employer. The
employer then uses the reported information to set a budget
by which the employee is evaluated. The value of
participative budgeting is in question if the employee
misrepresents his private information. Penno (1984) shows
that participative budgeting can be strictly valuable if the
employer has some control over the private information
available to the employee.










The Effects of Timing of Information Disclosure


Time is clearly an important factor in the escalation

error problem. Staw and Fox (1977) implement their time

construct by having subjects make three consecutive

investment decisions. Their study showed an unstable effect

of time on the amounts invested. In the current study the

manipulation of time is treated quite differently. This

manipulation takes the form of the point in time when the

manager discovers with certainty that he has made an

escalation error. In the Kanodia, Bushman and Dickhaut

(1989) model, the manager either discovers the true state of

nature immediately or after his initial project choice. The

current multi-period model affords the opportunity to vary

the point in time when the manager learns the true state of

nature. It is posited that the later the manager learns the

true state of nature and realizes that he has made an

escalation error, the more likely he is to continue with the

error.

In the context of the current study, it is posited that

this time effect may be due to two factors. First, the

longer the firm has undertaken the project, the larger the

project's sunk costs.4 Second, the manager's reputation is


4 This assumes that the project's sunk costs
(implementation or construction costs) continue to increase
monotonically until the completion of the project. This is
usually the case, as even if a large initial outlay is
required, some additional costs are typically incurred until
the project's completion.










more severely damaged by admitting a mistake made many

periods in the past. By doing so, he reveals that he is

untalented, and has either been unable to discern that the

alternative project is superior or has been hiding his

choice of the inferior project. Consequently,

H3: The later the manager discovers that he has made
an escalation error, the more likely he is to
continue with the error.


Effects of the Environment


The conditions under which the manager makes decisions

vary due to uncertainty. Although information signals are

often noisy or imperfect, the manager must use them.

Additionally, unforeseen economic factors may distort the

results from any chosen course of action, giving the manager

yet another imperfect feedback signal (project payoffs).

The possible combinations or series of signals and payoffs

will vary continuously in any given environment. In this

study the possible series are divided into difficult series

with many occurrences of noisy signals and unforeseen

economic factors, easy series with few occurrences and

intermediate or moderate series for those in between. The

manager is more likely to make an incorrect choice and

potentially an escalation error in a more difficult series.

Additionally, the more talented manager should always be

less likely to commit escalation errors regardless of the

series type due to his superior ability to discern the true









38

state of nature (see H1). This anticipated effect is stated

in the following hypothesis:

H4: Escalation errors are made by both talented and
untalented managers in a difficult series (of
signals and payoffs), by only untalented managers
in a moderate series and by no managers in an
easy series.


The Decision Model


Within the principal-agent framework, the hypotheses of

the current study are tested. The proposed model of the

manager's escalation error decision is

EE = f(TAL, MON, TIME, DIFF) (11)

where

EE = whether an escalation error is committed;

TAL = the manager's talent level;

MON = the presence (or absence) of a monitor of the
manager;

TIME = the timing of information disclosure to the
manager; and

DIFF = the difficulty level of the series (of
signals and payoffs) facing the manager.

Within this context, an escalation error is expected when

the manager is untalented, has no monitor present, discovers

that he has made an escalation error at a late stage and

faces a difficult series. Therefore, the expected

coefficients for the independent variables are negative for

TAL, negative for MON, positive for TIME and positive for

DIFF.










Supplementary Analyses


Aside from the four posited hypotheses, additional

exploratory premises are also investigated. These premises

pertain to the behavior of the experimental participants.

Trends in behavior over time are typically tested in this

type of experimental setting (experimental markets). These

premises will be developed in Chapter 4, following a

detailed explanation of the experiment.

In this chapter, the hypotheses to be tested and the

principal-agent model were integrated into the resulting

decision model. The experiment designed to test the

hypothesized effects is described in Chapter 4.















CHAPTER 4
RESEARCH DESIGN AND DATA ANALYSIS METHODS


Introduction


The research method used to test the hypotheses

presented in Chapter 3 is described in this chapter. First,

the design of the experiment that was conducted to test the

effects of the four independent variables is discussed in

detail. Next, the chapter describes the data analysis

methods used which take into account the discrete nature of

the dependent (and some of the independent) variables. The

data analysis section also includes a summary of the

predicted results and a discussion of the supplementary

analyses.


Experimental Design


Overview


The decision model in Chapter 3 was tested in a multi-

period market setting in an attempt to identify those

factors responsible for managers' escalation errors. The

setting included two types of participants: managers and

supervisors. A total of 156 graduate and undergraduate

students in the College of Business Administration at the











University of Florida participated in the study. Each

subject attended one experimental session. As an incentive

to participate, subjects earned lottery tickets based on

their decisions made during the experiment. The lottery

tickets were entered into a drawing in which seven cash

prizes were awarded.

The experiment was run over a two week period and

consisted of a total of 16 sessions. Each session lasted

about two hours and could accommodate up to 12 subjects (six

manager-supervisor pairs). Some sessions were run with less

than 12 subjects. This was done in an attempt to use all

willing subjects (some were only available during certain

time slots). Precisely, the number of sessions containing

each allowable number of subjects was three with 12, three

with 11, two with 10, four with 9, three with 8 and one with

7. In all sessions with an uneven number of subjects, the

experimental assistant acted as the last supervisor.


Experimental Task


A detailed explanation of the experimental protocol can

be found in Appendix A. A more succinct description of each

experimental session is presented here. In each session

there were four to six manager-supervisor pairs. Subjects

acting as managers made investment decisions to maximize

their experimental payoffs. Subjects acting as supervisors

gave investment decision feedback to managers and bid










competitively for managers' services at the end of each

sequence of rounds. Supervisors' wages were based upon the

investment project payoffs which resulted from managers'

decisions. In those sessions with an uneven number of

subjects, the experimental assistant provided the remaining

manager with the necessary feedback information, but did not

actively bid for managers' services. All bids from this

"dummy" supervisor were set equal to $1.50 (the minimum

allowable bid).

The steps in each experimental session can be found in

Figure 4-1. Managers and supervisors were randomly paired

in the first sequence of rounds. For every sequence after

the first, managers and supervisors were paired based on the

results of the bidding for managers' services. During each

sequence, managers and supervisors went through a group of

decision rounds in which each manager made an investment

project choice based on the information he received.1,2 The

information initially given to managers consisted of the two

possible investment projects, the two possible states of

nature, the probability of each possible state of nature,


1 Managers chose between two potential investment
projects. They received limited information about the two
projects, but were told they should attempt to discern which
investment project was favored by the economic conditions in
the market. Managers knew that determining which project
was superior based on these conditions would maximize
project payoffs.

2 If the manager is not attempting to build a
reputation, he should be making his project choices
consistent with Equation 1 from Chapter 3.











the two possible payoffs and some additional information

(see Specific Instructions for Managers in Appendix B).

At the end of the first sequence, project choices for

each round, as well as cumulative payoffs for the firm for

the first sequence, were made public for all managers.

Supervisors were given time to analyze each manager's

performance and then bid competitively for managers'

services. Each supervisor entered three written bids, one

bid apiece for each of three managers. The supervisor

entering the highest bid for a particular manager was

matched with that manager for the second sequence. If the

highest bid for a manager was from a supervisor who had

already been matched with another manager, then the

supervisor entering the next highest bid was matched with

that manager. This process continued until all managers and

supervisors were matched. If some managers were not bid

upon by supervisors who had not already been matched, then

the remaining managers and supervisors were matched by the

experimenter at the minimum allowable bid. The more

talented managers should have logically commanded the higher

bids from the supervisors, since supervisors' wages were

based upon the cumulative payoffs resulting from managers'

project choices in each sequence.3


3 The payoff to the firm for each round was either
$.50 or $1.00. The cumulative payoff to the firm for the
sequence was the total of the payoffs for each round less
the cost to switch whenever a switch between projects was
made. The cumulative payoffs to the firm each sequence were








44

A time line for the ordering of events in each decision

round can be found in Figure 4-2. Managers received an

information signal from their supervisors at the beginning

of each round. This signal provided them with additional

information about the true state of nature.4 Managers were

given time to analyze their information and then made their

investment project choices. Supervisors then informed the

managers of the project payoffs resulting from their

investment project choices.5 The project payoffs which

accrued at the end of each round served as an imperfect

signal of whether a correct investment project choice had









divided in the proportions of \ to the manager and 4 to the
supervisor.

4 The signals provided the managers with the
conditional probability of the state of nature given the
signal that was received. The probabilities associated with
these signals were calculated using the information system
presented in Chapter 3. These probabilities were reported
to the managers in the Specific Instructions for Managers.
The signals were designed such that signal one would be more
descriptive than signal two.

5 During every sequence each supervisor acted in
multiple capacities. First, each supervisor acted as an
information provider. This can be likened to the supervisor
steering the manager in the right direction to make a
correct project choice. Next, the supervisor acted in his
superior capacity in providing feedback to the manager about
his choice. Finally, the supervisor acted in his hiring and
firing capacity when deciding which managers he should bid
upon. In actuality, most supervisors act in all of these
capacities on a regular basis.











been made. This chain of events was repeated until the end

of each sequence.6

Cost of switching. In each round after the first,

managers chose whether to stay with the current investment

project or switch to the alternative investment project.

Managers had to pay a cost to switch investment projects,

which increased throughout the decision rounds. The cost of

switching was manipulated in a manner which induced managers

to switch investment projects, i.e., the expected benefits

from switching always exceeded the cost of switching.

Managers were explicitly informed that this was the case.

If managers switched projects more than once, the cost of





6 There were six sequences in each experimental
session. Managers and supervisors were led to believe that
there would be seven sequences based on their payoff
records. The subjects were not informed of the true ending
point of the experiment due to what can be termed the
finite-period paradox. Luce and Raiffa (1957) address this
issue in the context of temporal repetition of the
prisoner's dilemma. In the prisoner's dilemma, for any one
play of the game, each player does better when he does not
cooperate, regardless of what the other player does. The
Pareto superior outcome is only achieved, however, if both
players cooperate. When the prisoner's dilemma is repeated
an infinite number of times, both players gain from a
commitment to always cooperate. The problem arises with a
large, but not infinite, number of repetitions. In this
case, the final stage of a finitely repeated game becomes
equivalent to a single-stage game in which cooperation
cannot emerge. Inductive reasoning dictates that this
scenario repeats for the penultimate stage of the game and
so on for all stages back to the first. In this case,
ignorance of the ending point of the game is one method to
prevent uncooperative behavior. The current case is
considered similar enough to the prisoner's dilemma to
warrant a similar treatment.








46

switching started at the same initial value and increased as

previously noted.

Incentive to escalate. Since many of the experimental

variables were designed to discourage escalation behavior

(cost of switching and certain monitoring), subtle

incentives to promote escalation errors were included in the

experimental instructions.7 These incentives included a

suggestion about the importance of consistency in actions

and maintaining a smooth stream of earnings for enhancing

reputation. Both of these variables are consistent with the

commitment of an escalation error. Consistency in actions

is maintained when committing an escalation error, since the

same project is chosen as has been previously chosen.

Additionally, a smooth stream of earnings is maintained,

since the project payoff resulting from an escalation error

is probably low, as it has been in the past from choosing

the nonoptimal project.

Stressing the importance of these two variables is not

inappropriate for the manager, as both of these

characteristics are considered to enhance the reputation of

the manager. Staw and Ross (1980) show that consistency in


7 These incentives were based upon the results of the
previous behavioral research on escalation errors. As was
noted previously, all subjects were always personally
responsible for all investment decisions. Additionally, the
importance of consistency in actions was stressed in both
the typed experimental instructions as well as the verbal
explanations of the experimenter. These two factors were
found to be consistent with the commitment of escalation
errors in previous studies.










actions is viewed as an integral characteristic of a

successful manager. Several studies including Hepworth

(1953), Beidleman (1968) and Copeland (1968) note that a

great deal of emphasis is placed on the periodic earnings

figure. Managers accordingly are apt to feel pressure to

maintain the current level of earnings. It is hypothesized

that managers may actually try to avoid large variances in

the earnings figure from year to year, since variance

connotes risk and typically is looked upon unfavorably.

Consequently, both consistency in actions and maintaining a

smooth stream of earnings are posited to have a favorable

effect on the manager's reputation value and are consistent

with the manager committing escalation errors.

The experimenter explicitly explained two learning

sequences to the managers before the beginning of the actual

experiment. The conditions which would produce comparable

payoffs from switching and escalating were examined (e.g.,

switching payoffs exceeded escalating payoffs by $.50).

Managers were informed that supervisors were aware that a

switch between projects in a late round signalled an initial

incorrect project choice. They were also told that a late

round switch may lower supervisors' assessments of the

manager's skill level. Consequently, the potential benefit

of escalating in terms of increased bids from supervisors in

the future was mentioned. It was then left up to each

manager to determine whether the potential benefits from









48

escalating exceeded the certain increase in project payoffs

from switching.


Subjects


In any experimental study, trade-offs exist between

replicating real world circumstances and the tractability of

the experiment. The decisions made by subjects in this

experiment are simplified from their real world

counterparts. Many factors that would be present in actual

business decisions are missing in the experiment.

Nevertheless, by isolating the variables of interest, the

results of the study can be attributed to the experimental

variables.

Business students were targeted as the subject group

for the experiment instead of practicing managers for

several reasons. These students had the requisite (somewhat

uniform) business training to deal with the experimental

task, while using practicing managers would have introduced

a wide dispersion of experience and knowledge bases. Also,

they were readily available to participate in an

experimental session on campus. The use of practicing

managers would have most likely entailed moving the

experiment to the employment sites. Additionally, given the

requisite number of subjects, the procurement of this number

of practicing managers would have been extremely difficult.











Given these observations, business students were used as

experimental subjects.

Subjects were recruited from undergraduate and graduate

accounting courses, as well as undergraduate finance

courses. The lowest level subjects recruited were second

semester juniors and the highest level were final semester

master's students. When the students were initially

contacted, they were informed that to participate in the

experiment they would need to attend one two hour

experimental session. They were told that they would earn a

chance to win a cash prize in a lottery type drawing as a

result of their participation in the experiment (see

Appendix B for a copy of the introduction letter). Students

who were interested in participating in the experiment chose

which experimental session to attend over the two week

period of the experiment.

The total number of subjects participating in the

experiment was equal to 156. Of the total, 82 subjects

acted as managers and 74 subjects acted as supervisors.

Descriptive statistics about the subjects can be found in

Table 4-1. As shown in the table, the number of male

subjects was somewhat greater than the number of female

subjects. The average age was 23 and the average GPA was

3.30. The greatest number of subjects were master's level

accounting students. Senior level undergraduate accounting

students made up the next largest group. Junior level










accounting students and senior level finance students

comprised the only other large groups. The average payoffs

were $28.05 for the managers and $29.60 for the supervisors.

The average pre-test score for managers was 9.05, with a

maximum of 12 and minimum of 3. The solicited risk aversion

measure for the managers revealed 70 risk neutral subjects,

seven risk averse subjects and three risk seeking

subjects.8


Incentive Structure


To entice subjects to commit the necessary two hours of

their time to participating in the experiment, some

potential reward had to be offered.9 Traditionally, each

subject is paid a flat fee for participating which is

presumably equal to their reservation wage. As the number


8 The risk aversion measure was solicited as a check
on an assumption from the principal-agent model of Chapter
3. In the model it is stated that if managers are assumed
to be risk neutral, maximizing expected utility is
equivalent to maximizing wages. The measure used in this
study consisted of one question at the end of the Pre-
Experimental Questionnaire for Managers. This question
pertained to an equivalency calculation for a lottery ticket
and is admittedly a rough measure. A more detailed measure
was not used due to the very tight time constraints of each
experimental session. Nevertheless, the solicited measure
does show the majority of the subjects to be risk neutral,
thus validating the assumption of the principal-agent model.

9 In experimental market studies it is generally
accepted that cash incentives should be offered to subjects
in order to induce them to act as expected utility
maximizers. Without cash incentives subjects are not
motivated to care about their choices (act optimally), or
even show up if they are not required to do so for some
other reason (such as receiving class credit).










of subjects participating in the experiment becomes large,

the total amount required to pay the subjects becomes

prohibitively high. Recently Bolle (1990) has shown that

under certain circumstances (small decision costs and

anonymous choices) a randomized reward structure is

equivalent to rewarding each subject. These conditions were

met in the current experiment, as there were no decision

costs and no identification of subjects (by anything other

than their subject numbers) during the experiment. The

randomized reward structure for the experiment took the form

of a lottery drawing for seven prizes at the completion of

all experimental sessions. The seven prizes included one of

$200, two of $100 and four of $50. The subjects earned

lottery tickets in direct proportion to their cumulative

experimental payoffs, which should have induced them to give

a good effort during the experimental sessions. The winners

of the lottery were identified only by their lottery ticket

numbers and had to have the other half of the ticket stub to

claim their prizes.


Operationalization of Independent Variables


Manager's talent level. The manager's talent level was

measured in the experiment by administering a test to all

subjects acting as managers. The test was administered

after the learning sequences for managers and before the

actual experimental sequences began. The test consisted of










twelve statistical and business questions designed to

measure the subject's (manager's) aptitude for the

experimental task (see Appendix B for a copy of the Pre-

Experimental Questionnaire for Managers). The designation

of each manager's talent level (in terms of his score on the

questionnaire) was made by the experimenter after the

experimental sessions were completed.

Information asymmetry. The manipulation of the

information asymmetry construct was between-subjects. The

two possible states were certain monitoring and no threat of

monitoring. Each subject faced only one possible state to

enable him to discern the effects of the monitoring

condition over the multiple sequences.

When monitoring was present, additional discriminating

information about the manager was disclosed to the market

(the supervisors) in the form of optimal total payoffs for

the sequence. The disclosure of the optimal total payoffs

reduced information asymmetry between that particular

manager and the market. The difference between the optimal

total payoffs and the actual payoffs provided supervisors

with important information upon which to revise their

estimate of the talent level of the manager. For

experimental purposes the monitor was noiseless or a perfect

signal of the actual information.

Time. The manipulation of time was within-subjects.

Maintaining a constant time condition for each subject was










not considered necessary, since the effect of the time

construct was considerably easier for the subjects to

understand than the effect of the monitor construct. The

time variable was manipulated by informing managers of the

superior project for the sequence at different times during

the sequence. Two points were chosen within each twelve

round sequence. Managers were privately informed of the

superior project after either round six or round eight.10

Based on this, an escalation error is defined for

experimental purposes as continuing to choose the nonoptimal

project in any round after the manager is informed of the

superior project.

Environmental factors. The manipulation of the

difficulty level surrounding the manager's decision was also

within-subjects.11 This ensured that every manager faced


10 Many factors were involved in choosing to disclose
the superior project to the managers after rounds six or
eight. During the first half of each sequence managers
should attempt to discern the superior project themselves.
After this point in time, however, in order to ensure that
managers are actually committing escalation errors and are
not simply ignorant of which project is superior, the
superior project must be disclosed to them. Obviously, many
other points in time could be chosen to disclose the
superior project, but these two were considered to be
sufficient to describe the anticipated effect.
11 The specification of the time and environmental
factors variables as within-subjects is not consistent with
the use of the logit model for data analysis. Nevertheless,
the true manipulation of these constructs was within-
subjects as each subject faced all possible levels of each
construct. The reconciliation of this problem is presented
with the discussion of the logit model later in this
chapter.











each types of series the same number of times and had the

same opportunity to escalate. Each series of information

signals, realizations of the random factor e and resulting

payoffs was classified as to its level of difficulty. The

validity of escalation errors made by managers was ensured

by this classification. If this was not done, some

escalation errors may have been caused by certain series

being more difficult to interpret (to discern the true state

of nature for the sequence before it is revealed to the

manager) than others. By classifying the randomly generated

series into three groups, any spurious results were avoided.

The three groups were classified as easy, moderate and

difficult. Each group contained one prototype series which

fit the predetermined specifications for that group.


Measurement of Dependent Variable


The hypotheses discussed in Chapter 3 are based on the

dependent variable of whether an escalation error is

committed. An escalation error is defined for experimental

purposes as continuing to choose the nonoptimal project in

any round after the manager is informed of the superior

project. This definition is used as the basis for

classification of the subjects' choices of investment

projects. Their choices after either the sixth or eighth

round constitute either an escalation error or a correct

project choice. Based on the design of the experiment,











managers should always continue choosing the same project,

once they had made their choice between the two projects

after round six or eight.12 Consequently, the dependent

variable is dichotomous in nature for each sequence: either

an escalation error is committed or no error is committed.


Data Analysis Methods


The Decision Model


The manager's decision model as previously specified is

EE = f(TAL, MON, TIME, DIFF) (11)

where

EE = whether an escalation error is committed;

TAL = the manager's talent level;

MON = the presence (or absence) of a monitor of the
manager;

TIME = the timing of information disclosure to the
manager; and

DIFF = the difficulty level of the series (of
signals and payoffs) facing the manager.

The manager's decision model containing the variables

specified in hypotheses one through four is tested using the

logit model. The logit model is the appropriate model to




12 Subjects knew that there were no gains whatsoever
to be made from switching in any round after round seven or
nine (depending on when the superior project was revealed to
them). The experimental data showed that subjects
understood this fact, as once the subjects had chosen a
project after round six or eight they never switched
projects again.








56

use when the response variable is binary. The general logit

model is defined as


gjklm = 0 + +P P + PE (12)

where


gjklm is the logit (logit transformation) corresponding
to the possible levels of the explanatory variables;


0P is the intercept term;

B, C, D and E represent the four explanatory variables;

j, k, 1 and m represent the possible levels of each
explanatory variable; and


P represents the effect on the response variable of a
given set of observed values of the explanatory
variable.13

In the current case, variable B is the talent level of the

manager. This variable has thirteen possible levels in

accordance with the possible scores of zero to twelve on the

managers' pre-experimental questionnaire. Variable C is the

monitoring of the managers. The two possible levels are

presence or absence of a monitor. Variable D is the time

when the manager learns of the superior project for the

sequence. The two possible times are rounds six and eight.

Variable E is the difficulty level surrounding the manager's

decision. The three possible levels are difficult, moderate



13 See Andersen (1990) for a good discussion of the
logit model and its applications.









57

and easy. The response variable to be tested is whether an

escalation error was made by the manager (0=no, l=yes).

Each sequence for every manager is treated as an independent

observation since each potential combination of the time and

difficulty level variables represents a unique problem for

the manager.14


Predicted Relationships


When the manager's talent level is high it is expected

that escalation errors will not occur, therefore hypothesis

one will be supported by a significant negative coefficient

on the talent variable. When a monitor is present it is

expected that the manager will not escalate, therefore

hypothesis two will be supported by a significant negative

coefficient on the monitor variable. When the manager

learns of the superior project after round eight, he will be

more likely to escalate than if he learned of the superior

project after round six. Therefore, hypothesis three will

be supported by a significant positive coefficient on the

time variable. When the difficulty level of the series is

easy the manager will not be expected to escalate, while if


14 This treatment provides a resolution to the problem
of these two variables being manipulated within-subjects.
If each manager's multiple sequences are treated
independently, then these variables actually vary between
(instead of within) managers (or sequences). Arguably each
manager's multiple sequences are not truly independent, as
the managers should have been learning (and making better
decisions) as the experiment went on.











the level is difficult the manager will be expected to

escalate. Consequently, hypothesis four will be supported

by a significant positive coefficient on the difficulty

level variable.


Supplementary Analyses


The additional premises to be tested in this study

pertain to the total payoff to the firm for each sequence

and the bids of supervisors for managers' services. A trend

which can be analyzed in such experiments is whether the

market converges to equilibrium. At equilibrium no parties

earn excess profits and all parties act rationally to

maximize their payoffs. This implies that the parties

always make optimal choices to maximize the current period's

payoff. This equates to maximizing firm value and not the

manager's reputation value. Consequently, in the current

case there is no pure equilibrium value as has been defined

in previous experimental market studies.15 This is because

the experiment is specifically designed to promote

escalation errors (and induce the manager to be concerned


15 In pure experimental market studies based on
information economics, convergence to an analytical
equilibrium price is typically tested. These markets
usually contain as many as 40 periods of bargaining for a
hypothetical good. Over this extended time frame,
convergence is often observed. The current experiment does
not entail such an expansive time frame nor does it qualify
as a pure experimental market, since there is no analytical
equilibrium price. Nevertheless, trends resembling
convergence to equilibrium can be analyzed in the current
setting.










about his reputation) by increasing the uncertainty

surrounding the manager's decision of which project is

optimal, especially in the difficult sequences. Therefore,

it is highly illogical to expect the manager to make optimal

choices (to maximize the current period's payoff) throughout

the experiment. Consequently, an alternative measure of

pseudo-equilibrium must be employed.

A value called the logical payoff is used to proxy for

the equilibrium value of total payoffs for a sequence. This

value is calculated for each type of series by considering

all reasonably rational patterns of choices. The logical

payoff is simply the expected value (average) of these

patterns.

The three premises involving the logical payoffs are

P1) the market should be closer to the logical payoff for

the easy sequences than for the moderate and difficult

sequences, P2) the market should move toward the logical

payoffs over time (irrespective of series type) and

P3) there should be more dispersion from the logical payoff

when an escalation error is made than when no error is made.

These three premises are tested using generalized distance

measures, i.e., the variance and the average absolute value

of the actual payoffs from the logical payoffs. The

absolute values are reported in addition to the variances

since they are less affected by severely outlying











observations. The variances and absolute values are

computed as

n
E (APi LP)2 (13)
VAR = =1
n-1



EIAP LPJ (14)
ABS =-
n

where
APi = actual payoff for observation i;
LPi = logical payoff for observation i; and
n = total number of observations.

Premise P1 pertains to the magnitude of the variances

and absolute values in the easy, moderate and difficult

sequences. The lowest total measures are anticipated in the

easy sequences since there is less uncertainty involved and

correct choices should be made the majority of the time.

The highest measures are expected in the moderate sequences,

due to an interspersion of inconsistent signals and random

factors which should cause managers to receive many low

payoffs and switch quite often. The dispersion in the

difficult sequences should be quite low. The difficult

sequences lead managers to believe they are choosing the

correct project, when in fact they are choosing the

incorrect project. Consequently, managers receive many high

payoffs and usually switch only once.

The second premise, P2, concerning the movement of the

market toward the logical payoffs over time is tested for









61

the moderate and difficult sequences only. This is because

all subjects faced the same moderate and difficult sequence

twice, which was not the case for the easy sequence. The

premised relationship is expected as the subjects learned

more about the market over time. The third premise, P3,

posits that the variances and absolute values should be

greater when an escalation error is made than when no error

is made. The rationale for this is that more incorrect

project choices accompany the commitment of an escalation

error, thus increasing the difference between the actual

payoff and the logical payoff.

Other relationships that can be explored in this

experiment pertain to the bids from supervisors for the

managers' services.16 The movement of these bids over time

and the reaction of supervisors' bids to an escalation error

are both relevant issues. Two reference points are used to

evaluate these premises about supervisors' bids: the actual

average bid and the logical average bid. The logical

average bid is calculated based on the logical payoffs

discussed previously.

A decreasing trend is expected in the bids over time as

supervisors learn and understand the mechanics of the market

16 Each supervisor entered three bids for three
different managers at the end of each sequence (one through
six). Only the accepted bids, which were subtracted from
the supervisor's payoffs and added to the manager's payoffs
with whom the supervisor was paired for the next sequence,
were analyzed. The total number of accepted bids was equal
to 480; the total number of experimental sequences.











(this was shown to be the case in pilot studies).

Additionally, the variance of the bids should also decrease

over time, as a bidding strategy is adopted. Two measures

of variance based on the actual average and the logical

average are used to assess this trend. These variances are

calculated as

n
S(BID AAB)2 (15)
VARA- = n-



n
S(BIDi LAB)2 (16)
VARL n


where
VARA = variance from actual average bid;
VARL = variance from logical average bid;
BIDi = actual bid for observation i;
AAB = actual average bid;
LAB = logical average bid; and
n = total number of bids.


Summary


The experiment described in this chapter was

specifically designed to test the hypothesized effects of

the independent variables. Consequently, it could be argued

that the experiment is devised to promote escalation errors.

This is certainly not true, however, since economically

(from the viewpoint of the firm) it is never beneficial for

subjects to escalate. Furthermore, given the experimental

design, escalation behavior is only expected in a very small











percentage of sequences. Specifically, the untalented

manager is expected to escalate when no monitor is present,

when he is informed of the superior project after the eighth

round and when he faces a difficult sequence. Since the

experimental design is fully crossed, these circumstances

are only met approximately 8% of the time.17 Therefore,

during the majority of the experimental sequences, no

escalation errors are expected. This proves that the

experiment is certainly not contrived.

The data analysis technique discussed recognizes the

dichotomous nature of the dependent variable in the

manager's decision model. The supplementary analyses are

designed to test trends which are usually examined in this

type of experimental setting. The analyses presented in

Chapter 5 center around the manager's decision model and

incorporate ancillary analyses appropriate for such

experiments.












17 The 8% approximation is derived from the following
conditions. Approximately half of the managers are
untalented and half have no monitor present. If the first
sequence (which is easy for all managers) is discounted,
then four of the last five sequences are in the eighth round
treatment, and two of the last five sequences are difficult.
Therefore, 1/2 X 1/2 X 4/5 X 2/5 = 8/100 = 8%.















TABLE 4-1

Descriptive Statistics of Subjects

Total
Managers Supervisors Sample

Total Number 82 74 156

Sex
Female 34 34 68
Male 47 38 85

Average Age 22 23 23

Average GPA 3.27 3.32 3.30

Classification
3AC 8 12 20
4AC 26 11 37
6AC 1 3 4
7AC 24 31 55
3BA 3 3 6
4BA 13 8 21
7BA 3 3 6
8BA 1 0 1
Other 2 1 3

Average Payoff $28.05 $29.60

Pre-test for Managers
Average Score 9.05
Variance 4.14
Maximum Score 12
Minimum Score 3

Measure of Risk Aversion
Risk Neutral 70
Risk Averse 7
Risk Seeking 3

Classification key: Number equals class year (3=junior,
4=senior, 6=post-baccalaureate, 7 and 8=master's) and AC
equals accounting, BA equals business administration.

Note: Any numbers not adding up to the reported total are
due to some subjects not filling in all questions asked of
them.















Initial
random
pairings
bidding


12
rounds


New
pairings
based on
bidding


New
pairings
based on
bidding


IILLLLLIIIIII III I I iii ii Ih
Sequence #1 Sequence #2
A


New
pairings
based on
bidding


New
pairings
based on
bidding


Sequence #6
A


Bidding
for
managers
services


I


Bidding
for
managers'
services


Bidding
for
managers'
services


Bidding
for
managers'
services


FIGURE 4-1
Time Line for an Experimental Session















Managers
receive
information
signal
from
supervisors


Managers
make
investment
project
choices


Supervisors
inform
managers of
project
payoffs


FIGURE 4-2
Time Line for a Decision Round















CHAPTER 5
RESEARCH RESULTS


Introduction


The results of the tests of the hypotheses developed in

Chapter 3 are described in this chapter. First, the results

of the logit analysis for the independent variables is

presented. Second, the supplementary analyses related to

this type of experimental setting (experimental markets) are

described. Manipulation checks of the validity of the

experiment are then presented. A summary of the additional

sessions run to test the revised operationalization of the

information asymmetry construct are reported next. A

discussion of the results concludes the chapter.


Results of Hypothesis Testing


The main data analysis centers on the managers'

commitment of escalation errors. As such, the primary data

gathered in the experimental setting consist of the

sequences in which the managers committed escalation errors.

Descriptive data of the frequency of escalation errors for

each of the independent variables can be found in Table 5-1.

Each panel of the table shows the breakdown of the 33

escalation errors committed during the 480 experimental

67








68

sequences by levels of the independent variable.1 Panel A

shows that a greater proportion of escalation errors were

committed by managers with talent levels of eight and below

(10%=19/184), than were committed by managers with talent

levels of nine and above (5%=14/296). A smaller difference

is shown in Panel B, as managers escalated 8% (18/222) of

the time when a monitor was present and the rate dropped to

6% (15/258) when no monitor was present. Panel C shows a

marked difference between the two revelation points with a

2% (3/190) escalation rate after six rounds and a 10%

(30/290) rate after eight rounds. The largest differences

exist in Panel D where the easy series had a 0% (0/164)

escalation rate, the moderate series a 3% (5/158) rate and

the difficult series an 18% (28/158) rate. Additional

preliminary analysis shows Spearman correlation coefficients

between the dependent variable, EE, and the four independent

variables of -0.090 (p=0.0246) for TAL, 0.045 (p=0.1615) for

MON, 0.169 (p=0.0001) for TIME and 0.285 (p=0.0001) for

DIFF. These escalation rates and correlation coefficients

are initial indicators of which variables are likely to be

significant in the logit analysis.

Hypotheses one through four are tested using the logit

model (Equation 12 from Chapter 4), since the dependent



1 The 33 escalation errors committed during the
experiment equates to a 7% escalation rate (33 errors + 480
total experimental sequences). The escalation rate
predicted in Chapter 4 was 8%.









69

variable (commitment of an escalation error) is dichotomous

in nature (0=no and l=yes).2 Hypothesis one states that

escalation errors are made more frequently by untalented

managers than by talented managers. Each subject acting as

a manager was tested before the beginning of the actual

experiment on their knowledge of general business concepts

and probability theory (see Pre-Experimental Questionnaire

for Managers in Appendix B). The scores on this test ranged

from a low score of three to the maximum possible score of

twelve. Therefore, the first independent variable

representing the talent level of the managers has nine

levels. A negative relationship is expected between this

variable and the dependent variable, since a low score

represents an untalented manager who is expected to commit

escalation errors.



2 As mentioned previously, logit analysis assumes each
observation to be independent of all others. This is not
truly the case in the current experiment, as managers
invariably experienced a learning curve throughout the
experimental sessions. To mitigate this issue, additional
analyses were done with each manager representing one
observation. This treatment eliminates the problem of
independence of each manager's six sequences. Kendall's r7
was computed between the number of escalation errors
committed and time, as well as the number of escalation
errors and the difficulty level. T-tests of the average rT
versus zero showed significant results in both cases. These
results are consistent with those of the correlation
analysis of the dependent variable from the logit model, EE,
and the two independent variables, TIME and DIFF.
Additionally, in Chapter 4, the correlation analysis of the
manager's descriptive measures showed the number of
escalation errors to be negatively correlated with the pre-
test score representing talent. These tests support the
results of the logit analysis.











Hypothesis two states that the presence of a monitor

reduces the incidence of escalation errors. The independent

variable for this hypothesis is a simple dichotomous

variable where a zero represents the absence of a monitor

for the manager throughout the experiment, and a one

represents the presence of a monitor. It has been

established that a manager should only commit an escalation

error when information asymmetry exists. The presence of a

monitor reduces the information asymmetry between the

manager and the market, while the absence of a monitor

widens the asymmetry. Consequently, a negative coefficient

is anticipated for this variable, as managers should only

commit escalation errors when no monitor is present.

Hypothesis three states that the later the manager

discovers that he has made an escalation error, the more

likely he is to continue with the error. Again the

independent variable representing time is a dichotomous

variable. For this hypothesis, the time when the managers

were informed of the optimal project was varied across

sequences for each manager. A zero for this variable

represents the receipt of the information after six decision

rounds, while a one indicates after eight decision rounds.

Since the manager is expected to escalate when he has chosen

the same project consistently over a large time frame, a

positive coefficient is anticipated for this variable.








71

Hypothesis four states that escalation errors are made

by both talented and untalented managers in a difficult

series of signals and payoffs, by only untalented managers

in a moderate series and by no managers in an easy series.

This hypothesis can be broken down into two parts. The

first part fits in with the main analysis and posits that

more escalation errors are made in the difficult series than

in the moderate series and no errors are made in the easy

series. This variable has three levels representing the

three types of series. A zero represents the easy series, a

one the moderate series and a two the difficult series.

Since the manager is more likely to escalate in a difficult

series, the anticipated relationship is positive.

The results of the logit analysis show that the

hypothesized model fits the actual data (likelihood ratio

statistic=43.54 and goodness of fit=0.9998).3 The fitted

logit model is shown in Table 5-2. The p-values associated

with the parameter estimates of this model show that

hypotheses one, three and four are supported, while

hypothesis two is not.

The significant negative coefficient for the talent

variable is -0.235 (p=0.008). This supports the contention

that untalented managers make more escalation errors than


3 The likelihood ratio statistic for the logit model
follows a X2 distribution and compares the specified model
with the unrestricted model. The X2 statistic converts to a
goodness of fit measure which is analogous to a pseudo-r2 in
regression.








72

talented managers. The significant positive coefficient of

1.526 (p=0.011) for the time variable confirms that

escalation errors are made more frequently after the same

project has been chosen for many periods. The significant

positive coefficient for the difficulty level variable is

2.161 (p=0.000). This supports the hypothesis that more

escalation errors are made when the uncertainty surrounding

the manager's project choices is high. These three results

are consistent with the hypothesized effects for the three

independent variables.

The lack of significance for hypothesis two

(coefficient=0.064 and p=0.439) is somewhat troubling,

however, since information asymmetry is such an important

concept in principal-agent problems. The experimental

manipulation of the monitor was obviously not salient to the

subjects in their choices. Despite this result, managers

and supervisors both reported that the monitor did affect

their choices in post-experimental questionnaires. The

subjects apparently did not fully understand the meaning and

purpose of the monitor in the experiment. Changes in the

manipulation of the monitor were undertaken for the

additional sessions of the experiment (and are discussed

later in this chapter).

The second part of hypothesis four asserts that

escalation errors in the difficult series are committed by

both talented and untalented managers, while those in the











moderate series are committed only by untalented managers.

Of the 28 errors made in the difficult series, 19 were made

by untalented managers and nine by talented managers. This

supports the contention that both types of managers will

escalate in the difficult series. Of the five errors made

in the moderate series, two were made by untalented managers

and three were made by talented managers. This does not

seem to support the contention that only untalented managers

will escalate in the moderate series, but is obviously not

definitive due to the small sample size. As reported

earlier, no errors were made in the easy series. Overall,

the validity of hypothesis four is supported.

Much descriptive data were gathered from the subjects

in the post-experimental questionnaires. Spearman

correlation coefficients between these measures are shown in

Table 5-3. There are seven significant correlations between

the different variables. The manager's score on the

"talent" pre-test, administered before the experiment, is

correlated with three different variables.4 The pre-test

score is positively correlated with the manager's self-

reported GPA. This suggests that the pre-test may be a

somewhat accurate measure of the subject's actual talent or

intelligence level. There is an apparently spurious



4 Cronbach's a is a measure which reports the
reliability of a test. The a for the manager's talent pre-
test is 0.62, which is considered an adequate measure of
reliability.








74

positive correlation between the pre-test score and the male

managers, as no such correlation is expected. A significant

negative correlation exists between the pre-test score and

the number of escalation errors committed during the

experiment. This is additional evidence of the hypothesized

negative relationship between the talent variable and the

commitment of escalation errors. An anticipated positive

correlation exists between the manager's GPA and the

experimental payoff. An unexpected positive correlation

exists between the number of escalation errors and the

manager's GPA. An obvious positive correlation is found

between the manager's age and classification. A positive

correlation exists between the manager's experimental payoff

and the number of escalation errors. This correlation

supports the hypothesized positive effect of escalation

errors on the manager's reputation. Since the manager

actually earned less in a sequence when he escalated, he

apparently was able to make up for this discrepancy by

garnering higher bids from supervisors in future sequences.

The results of the correlation analysis provide support for

the main results of the logit analysis.


Additional Testing


Supplementary Analyses


Convergence to equilibrium. The three premises

previously mentioned in Chapter 4 pertaining to the logical









75

payoffs are tested in this section. The calculation of the

logical payoffs is shown in Appendix C. Table 5-4 shows the

logical payoffs for easy, moderate and difficult sequences

when the optimal project is revealed after both six and

eight rounds.

The logical payoffs for the two types of easy sequences

(see note following table) are inherently the highest of the

calculated payoffs. The difficult sequence has a higher

logical payoff than the moderate sequence due to several

factors. These factors are the result of the difficult

series being specifically designed to promote more

escalation errors. This was accomplished through a greater

incidence of both an inconsistent signal and the random

factor in the early rounds of a sequence. This combination

induced managers to choose the incorrect project and

rewarded them with spuriously high payoffs. Managers were

consequently led to believe that they were choosing the

correct project, when in fact they were choosing the

incorrect project. Therefore, they often continued to

choose the incorrect project until the superior project was

revealed to them. At this point, the managers normally

switched, thus incurring the cost to switch just once.

During the moderate sequence, managers were more likely to

switch more often since the incidence of inconsistent

signals and the random factor was interspersed. This

resulted in more low payoffs and higher total switching










costs. Therefore, the combination of inconsistent signals,

the random factor and the total switching costs incurred

cause the logical payoff for the difficult sequence to be

larger than that for the moderate sequence.

Premise P1 pertains to the magnitude of the variances

and absolute values in the easy, moderate and difficult

sequences. The lowest total measures are anticipated in the

easy sequences, while the highest are expected in the

moderate sequences. Table 5-5 shows the resulting variance

and absolute value calculations.

The variances reported suggest no clear-cut trend for

either revelation point. The absolute values, on the other

hand, display the anticipated relationships. For both

revelation points the easy sequences have the lowest

absolute values while the moderate sequences have the

highest absolute values. As mentioned previously, the

absolute values are considered to be better indicators of

the true variation in the samples since they are less

affected by severe outliers. Z-tests comparing the

different absolute values show significant results at a=0.05

or better for all comparisons except the easy versus the

difficult after six rounds.5 Premise P1 is therefore

supported by the absolute value calculations.




5 Z-tests on a standardized measure of the absolute
value (the absolute value as previously calculated divided
by the logical payoff) show comparable results.











The second premise, P2, posits that the market should

move toward the logical payoffs over time. These variances

and absolute values are reported in Table 5-6. From this

table it is apparent there is no movement toward the logical

average in the moderate sequences. The variances and

absolute values are very similar when the superior project

was revealed after six rounds and actually increase slightly

with revelation after eight rounds.

For the difficult sequences, however, there does appear

to be movement toward the logical payoffs. In the non-

escalation error sample, both variances decrease

significantly by over 50% at the second sequences, while the

absolute values decrease as well. The entire sample shows

similar decreases for revelation after eight rounds. The

figures for the non-escalating sample are shown in addition

to those for the entire sample since the three escalation

errors that occurred during the second sequence under

revelation after six rounds accounted for the majority of

the entire sample's variance and absolute value at that

revelation point. For the non-escalation error sample, Z-

tests are computed between the absolute values of the first

and second sequences at both six and eight rounds. The

resulting Z-score after six rounds is insignificant

(p=0.1736), but the Z-score after eight rounds is moderately

significant (p=0.0793). Since there were more sequences

with round eight revelation, this result is encouraging.











The movement toward the logical payoffs for the difficult

sequences but not for the moderate sequences is not

surprising, given the previous discussion concerning the

volatility of payoffs in the moderate sequences. Premise P2

is thus supported for the difficult sequences (especially

for round eight revelation), but not for the moderate

sequences.

Premise P3 posits that the variances and absolute

values should be greater when an escalation error is made

than when no error is made. The majority of the escalation

errors (28 of 33) were committed in the difficult sequences.

Additionally, the majority of these escalation errors (25 of

28) were committed after round eight revelation. Therefore,

the expected effect should be the most prominent under these

conditions. Table 5-7 presents these variances and absolute

values. In the difficult sequences the variances (and

absolute values for round six revelation) when an escalation

error occurred are greater than those for the non-escalating

sequences, as expected. Z-tests on the absolute values show

significant results after six rounds, but not after eight.

For the few escalation errors made in the moderate sequence

(five) the relationship is reversed. The small sample size

in this case, however, precludes this result from being

significant. Premise P3 is consequently not supported in

the difficult sequences after round eight revelation where

the majority of the escalation errors occurred.








79

Bidding strategies. The movement of supervisors' bids

over time and the reaction of the bids to an escalation

error are two relationships that are investigated. The

actual average bid and the logical average bid are used as

the two reference points in evaluating these trends. The

actual average bid for the 480 experimental sequences is

$2.10. The logical average bid, calculated based on the

logical payoffs discussed previously, is $2.43 (see Appendix

C). The actual average bid is somewhat below the logical

average bid partially because the expected payoffs for

managers and supervisors were not equal, as is assumed when

dividing the logical payoffs between managers and

supervisors. Pilot studies showed that supervisors often

bid quite low and therefore garnered higher total payoffs.

The minimum bid was raised for the actual experiment, but it

was still anticipated that supervisors may bid low. This

low bidding caused supervisors to earn more during the

experiment.6

6 When the experiment was initially devised, the
division of the firm's total accumulated payoffs between
managers and supervisors as well as the boundaries for
supervisors' bids were carefully considered. In information
economics models the expected payoffs of all participants
are equal. In the current experiment the division of total
payoffs and the boundaries for supervisors' bids were
established in an attempt to equate managers' and
supervisors' expected payoffs. Initial pilot tests showed
that supervisors often bid low and earned higher
experimental payoffs. Since the adopted division of the
total payoffs between managers and supervisors seemed
logical, the bidding boundaries were manipulated. This
manipulation obviously did not fully counteract the observed
behavior, since supervisors continued to bid low and earn








80

The optimal strategy for a supervisor in the experiment

was to bid low, due to the unequal expected payoffs. Some

supervisors identified this strategy quickly, while others

did not. Consequently, a decreasing trend is expected in

the bids over time. Additionally, the variance of the bids

should also decrease over time, as a bidding strategy is

adopted. Table 5-8 shows the average bids across the

sequences, and the two measures of variance based on the

actual average and the logical average. No decreasing trend

in either the average bids or the variances is evidenced.

After additional analysis of each supervisors' six bids, it

is apparent that very few supervisors adopted a discernable

bidding strategy. Even though the majority of the subjects

acting as supervisors appeared to be giving an earnest

effort during the experiment, they were apparently unable to

devise a distinct bidding strategy. The reason for this

lack of explicit bidding strategies is unknown, since the

supervisors' post-experimental questionnaire responses

suggest an understanding of the experimental constructs.

The second relationship to be investigated concerning

the supervisors' bids relates to the reaction of the bids to

an escalation error. A direct causal relationship cannot be

inferred since the supervisors were unable to discern with

certainty when an escalation error had been made. Any


more in the actual experiment. In future experimentation,
the division of the total payoffs will be manipulated in an
attempt to equate experimental payoffs.










observed relationship may be because the commitment of an

escalation error lowered the resulting payoff for that

sequence. The lower payoff may in turn have caused lower

bids from supervisors. This is substantiated by the

supervisors' post-experimental questionnaires, where payoffs

were found to be an important factor that influenced

supervisors' bids. The average bid following an escalation

error was $1.97, which is significantly different from the

overall average bid of $2.10 at a=0.010. It is difficult to

draw conclusions concerning the supervisors' bids, however,

since the significant decrease following an escalation error

cannot be directly traced to the commitment of the error.


Manipulation Checks


To ensure the validity of experimental manipulations,

all subjects filled out a post-experimental questionnaire

which can be found with the other experimental instruments

in Appendix B. Subjects were asked to respond to each

question on a five point scale with the options of strongly

agree, agree, no opinion, disagree and strongly disagree.

The questions were devised such that a response of agreement

or disagreement would display the subject's interpretation

of the importance of the construct. The constructs were

tested using the nonparametric sign test.7 Consequently,


7 For the sample of escalating managers (sample
size=25), the test statistic B is used in the sign test.
The entire sample of managers, the sample of non-escalating








82

the salience of a construct is validated by a statistically

significant result from the sign test. The analysis of the

responses to the questionnaires is in Tables 5-9 through 5-

12. The inferences that can be drawn from this analysis are

powerful, particularly those comparing the samples of

escalating and non-escalating managers.

Question one dealt with the importance of consistent

project choices. Question two examined the importance of

project payoffs to the exclusion of all other factors.

Question three asked about the importance of a late round

switch. Question four dealt with the disclosure of the

optimal project payoffs. Question five asked whether the

subject felt he was successful in earning the highest

payoffs possible during the experiment. Question six asked

whether the subject would make the same decisions in the

"real world."

For the entire sample of managers, questions one

through four and six are statistically significant with an a

of 0.05 or better. Specifically, the majority of the

managers (79%) strongly agreed or agreed that consistency in

project choices was important. Over half (56%) of the

managers strongly disagreed or disagreed that project

payoffs were the only factor they should be concerned about.



managers and the sample of supervisors constitute large
sample sizes (greater than 30). For these samples, the test
statistic B* is used as the large sample approximation of
the sign test.








83

The majority (54%) of the managers strongly agreed or agreed

that switching projects at a late round in a sequence would

have a negative effect on the amounts of bids from

supervisors in the future. Almost half (48%) of the

managers strongly agreed or agreed that the disclosure or

non-disclosure of their optimal total payoffs affected their

project choices. Over half (53%) strongly agreed or agreed

that they would make the same decisions if they were in the

"real world." Question five does not show a significant

effect. When asked whether they thought they were

successful in earning the highest payoffs possible during

the experiment, nearly equal percentages replied positively

as replied negatively.

A comparison of Tables 5-10 and 5-11 for the samples of

escalating and nonescalating managers shows that the means

are quite different for many of the questions. For the

sample of escalating managers, questions one through three

are statistically significant with an a of 0.05 or better.

The sample of nonescalating managers shows significance for

questions one and six at a of 0.05 or better. More

specifically, t-tests comparing the two sample means show

significant differences between the two samples for

questions two and three (p-value=0.038 for question two and

0.002 for question three). This constitutes strong support

for a basic difference between the two samples of managers.










A large difference exists between the two samples for

question three. The escalating managers clearly believed

that a switch in the later rounds of a sequence would lower

the amounts of bids from supervisors in subsequent

sequences, while nonescalating managers were not convinced

of this point. The difference in results for question two

shows that the escalating managers were more convinced that

project payoffs were not the only factor that they should be

concerned about. These two beliefs led escalating managers

to commit escalation errors. These results shown in the

post-experimental questionnaires present strong evidence of

a critical difference between the samples of escalating and

nonescalating managers. Escalating managers clearly

believed that late round switches would be detrimental to

them in the future sequences and that project payoffs were

not the only factor that they should be concerned about.

For the supervisors questions one, two, four and six

are statistically significant at a of 0.05 or better. All

of these questions are significant for the entire sample of

managers as well. Specifically, a large percentage (69%) of

the supervisors strongly agreed or agreed that consistency

in project choices was an important quality in a manager.

For question two, the majority of the supervisors (60%)

strongly disagreed or disagreed with the contention that

project payoffs were the only factor that they should be

concerned about when bidding for managers. The great









85

majority (71%) strongly agreed or agreed that the disclosure

of the optimal total payoffs for some managers affected

their bidding strategy. Almost half (49%) strongly agreed

or agreed that they would make the same decisions if they

were in the "real world." Questions three and five show

insignificant results. For question three, supervisors were

almost evenly split on the contention that a late round

switch was a signal of a poor manager. Supervisors were

also split on question five on whether they were successful

in earning the highest payoffs possible during the

experiment.

In general the majority of the constructs were proven

to be salient in the post-experimental questionnaires. The

importance of consistency in project choices, the contention

that project payoffs were not the only factor they should be

concerned about, the importance of the disclosure of optimal

project payoffs and the feeling that subjects would make the

same decisions in the "real world" were all proven to be

salient for both managers and supervisors. Only managers

felt that a late round switch would lower the amounts of

bids from supervisors in the future. Neither group felt

that they were successful in earning the highest payoffs

possible during the experiment.










Results for Revised Experiment


A revised experiment consisting of six sessions was run

in an attempt to discern whether a stronger manipulation of

the information asymmetry (the monitor) construct would

prove more salient to the subjects. A total of 55 student

subjects from one section of Financial Accounting 2 at the

University of Florida participated in the experiment. The

students earned class credit for participating (equal to 2%

of their final grade) as well as lottery tickets. The cash

prizes for the lottery drawing for the revised experiment

consisted of two prizes: one of $50 and one of $25.

Subjects. Descriptive statistics about the subjects

for this experiment can be found in Table 5-13. As shown in

the table, the numbers of male and female subjects were

almost equal. The average age was 23 and the average GPA

was 3.21. The vast majority of the subjects were junior and

senior level accounting students. The average payoffs were

$28.32 for managers and $27.54 for supervisors. The average

pre-test score for managers was 8.43, with a maximum of 11

and minimum of 5. The solicited risk aversion measure for

the managers revealed 26 risk neutral subjects, two risk

averse subjects and no risk seeking subjects. The average

age and GPA were similar to those for the subjects for the

main experiment (23 and 3.30). The average pre-test score

was significantly lower than in the main experiment (9.05).

This is possibly due to the lower classification level of










the majority of these subjects compared with the subjects

from the main experiment. A larger average payoff was

earned by the managers, in contrast to the relationship in

the main experiment. This was apparently due to the

modifications undertaken for the additional sessions.

As in the main experiment, additional descriptive data

were gathered in the post-experimental questionnaires.

Spearman correlation coefficients between these measures are

shown in Table 5-14. These correlations are quite different

from those of the main experiment. The significant positive

correlation between the manager's age and classification is

the only significant correlation in the same direction as in

the main experiment. The three most relevant correlations

of the first experiment--a positive correlation between the

manager's pre-test score and GPA, a negative correlation

between the manager's pre-test score and the number of

escalation errors committed and a positive correlation

between the manager's experimental payoff and the number of

escalation errors--are not present in the revised

experiment. The opposite correlations involving the pre-

test score suggest that this measure was not effective in

measuring the subject's aptitude for the experimental task

in the revised experiment. The negative correlation between

the manager's experimental payoff and the number of

escalation errors suggests that the hypothesized reputation

effect observed during the main experiment was not present










during the revised experiment. The lack of correlation

between these measures suggests that the revised experiment

is not entirely comparable to the first.

Modifications. The revised experiment was modified

slightly to promote a greater incidence of escalation errors

and to strengthen the information asymmetry manipulation.

There was no manipulation of the time variable (the

revelation point after either six or eight rounds) in the

additional sessions, as all managers were always informed of

the superior project after eight rounds. Additionally,

managers faced an extra difficult sequence in place of the

second easy sequence.

The change in the manipulation of information asymmetry

for these sessions was two-fold. First, instead of

disclosing the optimal total payoffs (as was done in the

main experiment), the superior project itself was disclosed.

A second change was also made, since the certain disclosure

or non-disclosure of the superior project was a perfect

signal of an escalation error (and seemed too blatantly

obvious). Instead of certain disclosure or non-disclosure,

there was a percentage chance of disclosure for each

manager. Half the managers had a 20% chance and half an 80%

chance of disclosure. At the end of each sequence managers

drew a card from a deck with these probabilities to

determine whether the superior project was disclosed. This

is considered a stronger manipulation than in the main











experiment. In the main experiment, the commitment of an

escalation error by a manager had to be interpreted by a

supervisor through a comparison of the optimal total payoffs

and the actual payoffs. In the revised experiment, the

superior project was actually disclosed. Consequently, if a

manager had an 80% chance of disclosure of the superior

project he should never have escalated, since there was a

very good chance the supervisors would find out that he made

an escalation error. With the 20% chance the manager should

have felt a very small chance of being "caught." It was

hoped that this manipulation would cause managers to only

escalate under the 20% chance of disclosure.

Results. There were a total of twelve escalation

errors committed during the six sessions.8 Exactly half of

the errors occurred in each treatment of the information

asymmetry construct. The results of the logit analysis are

shown in Table 5-15. Again the hypothesized model fits the

actual data (likelihood ratio statistic=21.71 and goodness

of fit measure=0.9149). An insignificant effect for the


8 The twelve escalation errors committed during the
revised experiment equates to a 7% escalation rate (12
errors + 168 total experimental sequences). This rate is
equivalent to that of the main experiment. The escalation
rate predicted for the revised experiment is 15%--half of
the managers are untalented, half have no monitor present
and three of the last five sequences are difficult--1/2 X
1/2 X 3/5 = 3/20 = 15%. In the main experiment the actual
and predicted escalation rates are almost equal. In this
case the actual escalation rate is half of the predicted
rate. This observation represents further evidence that the
revised experiment is not entirely comparable to the main
experiment.








90

monitor was again shown in these experimental sessions. The

post-experimental questionnaires for both managers and

supervisors showed no significance for question four which

deals with the effect of disclosure on the subjects'

decisions (the majority of the remaining post-experimental

questionnaire responses were similar to those of the main

experiment). This is in stark contrast to the significant

effect for this question in the post-experimental

questionnaires for both managers and supervisors in the main

experiment. This lack of results for the monitor variable

under the modified manipulation is quite troubling.

Apparently the subjects are given too much information about

the entire experiment to be able to discern the relevance of

each piece. Alternatively, the manipulation of the monitor

construct may have been unable to capture the theoretical

essence of the variable. Nevertheless, the theory that a

monitor should reduce escalation errors cannot be discarded

without numerous attempts to prove or disprove the saliency

of the monitor construct.

Another troubling result of the revised experiment is

the significant positive coefficient of the talent variable.

This relationship was also shown in the correlation

analysis. Specification testing for outliers shows only one

manager in the revised experiment as a severe outlier.9


9 An outlier is usually considered extreme if its
residual is larger than two standard deviations. From a
conservative viewpoint, only outliers which had values of










When compared to the main experiment, this manager is the

most severe outlier of both samples. When this manager's

observations are removed from the logit analysis, the talent

variable loses its significance. These results are shown in

Table 5-16. This outlier has a great influence on the

results of the revised experiment due in part to the small

sample size of this experiment (28 managers as compared to

82 in the main experiment). When the data from the main

experiment are combined with the data from the revised

experiment excluding the outlier, the coefficient on the

talent variable reverts to its expected sign and regains its

significance. Table 5-17 shows these results, which are

similar to those of the main experiment.

Another observation supports the assertion that the

revised experiment was not completely comparable to the main

experiment. The subjects in the revised experiment did not

have the same incentive system as those of the main

experiment. In the main experiment, the subjects were

recruited from their classes with the sole incentive of

earning lottery tickets and a chance to win a cash prize.



studentized t residuals greater than 3.00 were removed. The
identification of the severe outlier from the revised
experiment was straightforward and was based on a regression
with NEE as the dependent variable and TAL, MON and PAYOFF
as the independent variables. Only one studentized residual
greater than 3.00 was found. The severe outlier's
studentized residual was 3.27. When the data from the main
experiment was combined with that from the revised
experiment, the observed outlier's studentized residual
jumped to 4.11.










In the revised experiment, the subjects still had a chance

to earn lottery tickets and win a cash prize (although the

cash prizes were significantly smaller), but it is believed

that the main incentive for these subjects was the class

credit they received for participating in the experiment.

The credit was given unconditional of the subjects'

performance during the experiment, i.e., all students

received equal credit. By earning the class credit the

subjects were relieved of an out of class assignment. The

instructor informed the students that participation in the

experiment would probably be "easier" than the completion of

the assignment. Only three students out of the class of 58

did not participate in the revised experiment. During these

experimental sessions, an obvious lack of effort and

interest was shown by many subjects (and was noted by the

experimental assistant as well as the experimenter). Due to

these facts the revised experiment is not considered to be

analogous to the main experiment.


Summary


The results of the data analysis suggest that the main

experiment was successful in inducing escalation behavior.

Three of the four factors posited to influence this behavior

were significant: the talent level of the manager, the time

when the manager is informed of the superior project and the

difficulty level surrounding the manager's decision. The








93

talent factor should be closely related to the commitment of

an escalation error. The significance of the talent factor

in the experimental study suggests that the pre-experimental

questionnaire designed to measure talent was an accurate

gauge of the subject's actual aptitude for the experimental

task. The time factor is somewhat obvious and its

experimental significance supports the hypothesized effect.

The manipulation of the difficulty level surrounding the

manager's decisions is also logical. It's experimental

significance not only supports the predictions, but also

provides an internal validity check of the experiment. As

mentioned previously, the lack of significance for the

monitor manipulation in both the main experiment and the

revised experiment is troubling. In general this experiment

was successful in identifying which of the posited factors

influence the commitment of an escalation error.















TABLE 5-1
Frequency of Escalation Errors for Independent Variables

Panel A: Talent Level

Escalation Error

Talent Level Yes No

3 0 4

4 2 4

5 2 10

6 2 22

7 7 65

8 6 60

9 2 72

10 6 80

11 2 80

12 4 50

Panel B: Monitor Variable

Escalation Error

Monitor Yes No

Present 18 204

Absent 15 243

Panel C: Time Variable

Revelation Escalation Error

Point Yes No

6 Rounds 3 187

8 Rounds 30 260




University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - Version 2.9.7 - mvs