(I1
In"!

. A.
.0:qu
unfilnoWWFfi
y .1020‘0 vb.—
oa-Nllxafltfll o F
I. u
‘5 .

1,2‘1 I .
3111‘ '1
HIE:

1
.1 Q .
{31%
My.

1 . :‘Ezié'
I 7}

. W;
. 3?}:
9g;

'5‘. n t...
it! It :“flfihtl‘llb
. - ‘Ifluflu‘u

~l|

o
:34 -
‘tubllf 1. u
'1' I.“
- ‘EJZIW-
t’l.“ ’10: . u .1

lit‘dXC‘ Ouloc'lfi'nl... t!

gIIOn I'll . \ .r

.ll‘lll'l‘ I

I l u
‘ x .13- V
.nl‘nil.vutv.|‘s|u.us\knll\1 I?! .. . - ‘ .
.1: . 3.! (11.50413... . 1 . .
i1‘{-ulanu.n.. . - ~
Yin n 1.0”!) , , y . -
.‘v.\r\ou.|ph.at\lr\ol. lg... . I. ..1.-W.|.Hb1xl\ .
r4431. . o ‘Puo.nu.lnu.u.-§ .. -551. .
.‘ll1v‘l I..ll4buollut.c!otlrld1~. o .. . 1'1 ... dsJ an,
.‘1 .0 i- -
{I‘ll-Kala. .. 1.10-1
azlupo: . .1! .‘b
.II I‘Q‘Ili. I
I‘ll“
. .

u n
a .1.

.1. .

no {‘A.Nl.uo
.

_ “Wkinfiunhflrd

1‘09:

gaudy”...

. VA.“I~I.0“ .-.vau1‘llnh‘.
114 , . in- 9.31%. :{ v.21.

t)

 

. $1“, .‘5;

|
,~
.0!
a!

‘ 2'.
vv.-
0‘
o
‘.
,n

I
1‘1. - 2

“4%
Mrs:

 

 

 

 

This is to certify that the

dissertation entitled

MULTIPLE OBJECTIVE DECISION SUPPORT
FOR FARM MANAGERS

presented by

James w. Pease

has been accepted towards fulfillment
of the requirements for

Ph . D. degree in Agricultural Economics

 

 

.H’? W

Major profes

Date JUIL'”; I98L

MS U is an Aflimau’vc Action/Equal Opportunity Institution 0-12771

 

MSU

 

 

 

RETURNING MATERIALS:
Place in book drop to
remove this checkout from

 

LIBRARIES _
~ your record. FINES w1ll
be charged if book is
returned after the date
stamped below.
ém£1/;f‘jl
p.193» .

 

 

 

MULTIPLE OBJECTIVE DECISION SUPPORT
FOR FARM mGERS

BY

James N . Pease

A DI SSERTATI EN

Submitted to
‘ Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Agricultural Economics

1986

 

”8‘. 7 .5)‘ 1,7

, MUM MUTE} (var. ~‘7r‘ '“W'nf’ ml}! Kant

_ . ,, HQWFKJ “’5
V) ‘ .. '3
m — .au. '4. ”NH
I
m“
o. T» "H‘CIDP U‘nx' "F'*f:*sc-:) ti meek-13 :5 w.» -.'.~.'.’(1:V.'.e'

"ME for dosnp 9‘ .~ t’fll'

t ‘r v_ I a in. (”I »‘K§"ul'._‘:l"$.
TM concept oi dec'h-v ~.'»;:L--’-'= «:1’.“ r (.4; , V a .‘ l
"(and bone» CBQ4CIQIKW- ‘ : u u- “-"s A. woman
oekteg while armor" ' in»? . 2' c

letlooetizt mo 3* uni-n :dw . - e' in «nu to
the role of human mam! .v. < . 'ar n \' 'ex‘ me

Copyright by

"06.10“, and We (W‘afim. m (‘5' :92. ;... .e-zeuch.

- 1986
”to"! rmrzhe’s ha¢¢~ Minnied «Tow-an momma: H

Patlflelist eeetmplxo-‘vs. The“: ~-;:~-‘a°.!m-u revs-t! the Imueo

'0’”th processing ',..p.(‘¢'{. c ‘t‘o :u .... raga. in.» ”new. $5.“;

- II (Vesuvius! bv want: .‘a ;¥P’~;€1':M. serial prawn-mg of «U.

‘, Md“ (”0“!th noun». ' ~~ ‘e<'- mart Lam. ~m'y ma

.WI 0. context variables. “no t my mum-I. serrate are” 96

'N“. wt dutch pernl' imezoo :n’cvalleo or we we

”the“ end the Input 04 ersrar'emwics m and later-echoes

  
 
 

_: -.~

_ gig“ wan stunner», decium meters and “HMO" '60".
. "'I'A'I' n

 
  
  
  
  

Went rowel-ch has alumnae m- imic-u of mum.
to “meat decisions. later-gun attire cruel-u
' “new poem «fer use mt. II! “we”...
I 7 g-fle‘mu cite «eunu- e0 “Rivera‘s” elm

teteeecttve mmm “it M“ {M M

 

ABSTRACT
MULTIPLE OBJECTIVE DECISION SUPPORT FOR FARM MANAGERS
BY

James N. Pease

The principal issue addressed in this study is the conceptual
framework for design of effective decision aids for farm managers.
The concept of decision support emphasizes the design of tools which
extend human capacities, aiding documented weaknesses in decision
making while supporting intuitive human abilities.

Rationalist and Behavioraiist paradigms differ with respect to
the role of human cognition, acceptable research context and
procedures, and the operational objectives of decision research.
Behavioral researchers have documented common violations of
rationalist assumptions. These violations reflect the limited
information processing capacity of the human cognitive system, which
is characterized by selective perception, serial processing of data,
limited computational capacity, limited short term memory and
dependence on context variables. This study outlines certain areas of
decision support which permit limited integration of the two
perspectives, and the impact of characteristics of and interactions
between decision situations, decision makers and decision tools.

Behavioral research has emphasized the importance of multiple
objectives in management decisions. Interactive multiple criteria
decision making (MCDM) procedures offer the prospect for integrating
optimization procedures with evaluation of decision prospects along

multiple dimensions. interactive Multiple Goal Programming (IMGP) is

James N. Pease

selected as a technique promising such a synthesis, and a decision aid
for support of land rental decisions of cash grain operators was
designed and implemented. Risk and return objectives were modeled as
competing objectives in the Goal-directed Search Model (GOALDIR).
Elicitation procedures were designed to obtain data for farm-specific
application of the aid, including probability elicitation.

Preliminary field testing of the aid indicated that:

i. Elicitation procedures to obtain farm-specific data may be
unacceptable because they consume considerable amounts of
the manager’s time.

2. The probability elicitation procedure used is feasible, but
its reliability is untested.

3. Subjects found graphic representations easier to interpret
than numeric data.

4. Subjects found explicit consideration of multiple objectives
reasonable, but were confused by optimization procedures.

A decision support research agenda is proposed to investigate
elements of synthesis between behavioral and rationalist principles.
Particular benefits are seen from integration with respect to
probability elicitation, preference relationships, decision rules and
graphical representations. Frameworks for evaluation must be

developed to accelerate progress in decision aid design.

ACKNOWLEDGEMENTS

For their guidance and assistance, I wish to thank the members of
my thesis committee: Roy Black, Al Shapley, Ralph Hepp, Harold Riley
and Carl Liedholm. I am especially grateful to my major professors:
to Roy Black, who pushed and pulled and prompted throughout the whole
research process, and to Al Shapley, who always gave good advice and
friendship.

Finally, the traditional acknowledgement to one’s spouse seems
inadequate when the spouse has also gone through the same long

process. Judy, we survived.

TABLE OF CONTENTS

LIST OF TABLES-IIIIOIIIIIIIIIIIIIIIIIIIIIIIIIIIlllllIIIllv

LIST OF FIGURES-ee-Inlelolllleleulel'lllelllllleeeee-eeevi

Chantal: PA:

I. INTRODUCTION
1.1 Motivation of the Study........................1
1.2 An Evolutionary Approach to Decision Support
Research and Design............................2

Ii. PARADIGMS OF DECISION MAKING

2.1 introduction...................................6
2 2 Management and Decision Making.................6
2 3 The Rationalist Perspective...................11
2.4 The Behavioralist Perspective.................15
2 5 Observed Errors in Decision Making............21
2 6 The Rationalist Reply.........................30
2 7 Some Elements of a Framework for Integration..38
2.7.1 Preferences............................4O
2.7.2 Probabilities..........................4i
2.7.3 The Limits of Logic....................42
2.7.4 Feedback and Learning..................43

III. DECISION AIDS

3.1 Introduction..................................45
Aids for the Decision Process.................46
Characteristics of Decisions..................51
Characteristics of Decision Makers............55
Characteristics of Decision Aids..............58
Consequences of Decision Aids.................69

@0000
.

IV.

E .
A ”Nu-t 6‘]wa

IPLE CRITERIA DECISICN TECiNIDUES
Introduction..................................78
Multiple Objectives in Agricultural Research..79
Multiple Criteria Decision Making: History
and Classification............................83
Mathematical Formulation of Multiple
Objective Techniques..........................89
4.4.1 The Method of Geoffrion................97
4.4.2 The Method of Zionts and Halienius.....98
4.4.3 The Method of Steur...................100

& A...
I.

iii

 

V.

4.5

4.4.4 The Method of Nijkamp and Spronk......102
Selection of a Multiobjective Technique......187

A MULTIPLE OBJECTIVE DECISION AID

UIUIUI

.1
.2
.3
.4
.5

6

UIUIUI

5.7

Introduction.................................111
Advantages of IMGP ..........................112
IMPG Consistency with Decision-theoretic
Principles...................................118
An Example IMGP Formulation..................125
Land Rental Decisions........................131
Modeling the Decision Problem................136
5.6.1 Assumptions...........................136
6 2 The Mathematical Model................137
6 3 Data Elicitation......................144
.6.4 Matrix Generator......................150
6 5

6 6

LP So‘verllll III-III. ll IIII'III IIIIIII16°
Goal-directed Solution Display
PrograMeae II I II... II I I II III I l I ll IIIIII162

UIUIUIUIUI

Preliminary Empirical Evaluation.............i67
5.7.1 Interview of First Agent..............168
5.7.2 Interview of Second Agent.............171
5.7.3 Interview of Third Agent..............173
5.7.4 Interview of First Operator...........175
5.7.5 Interview of Second Operator..........179
5.7.6 General Observations..................180

VI. CONCLUSIONS AND PROPOSED RESEARCH AGENDA

6.1 Introduction.................................183
6.2 Conclusions from the Literature..............183
6.3 Conclusions from Model Development...........186
6.4 Conclusions from Field Testing...............188
6.4.1 The Rental Decision...................188
6.4.2 Data Elicitation......................189
6.4.3 Multiple Objectives and Optimization..i92
6.4.4 Context of Management Assistance......196
6.5 Elements of a Decision Support Research
Agenda.......................................197
6.5.1 Probability Elicitation...............198
6.5.2 Preferences...........................200
6.5.3 Solution Procedures...................203
6.5.4 Graphics..............................208
6.5.5 Evaluation............................210
Appendix A
A. GOALDIR MODEL VARIABLES AND EOUATIONS.............213
8. SPREADSHEET TEMPLATES FOR GOALDIR DATA ELICITATION
NTIH PROTOCOLS FOR YIELD AND PRICE ELICITATION....218
C. PASCAL PROGRAM LISTING--MULTIVARIATE NORMAL

mom NmBER GmEmTORIIIIIIIIIIIIIIIIIIIIIIIIIII229

iv

 

LIST OF TABLES

Lam Ban
5.1 Risk and Return Solutions, Kennedy/Francisco

ProbIMIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII125
5.2 Estimated Yield Correlations, Major Crops.........139
5.3 Estimated Price Correlations, Major Crops.........140
5.4 Yield Penalty Matrix for Untimely Field

omr‘tion’ IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII141
5.5 Percentage Points Required to Dry Corn for

Different Planting/Harvest Periods................142
5.6 Estimated Field Hours Available for Planting

and Harvesting Operations, East Lansing...........143
5.? Yield Penalty Matrix for Untimely Field

M.ti°nslIllllIIIIIIII-IIIIIIIIIIIIIIIIIIIIIIIII157

LIST OF FIGURES

Email: 133
5.1 Functional Relationship Between a Preference

Function (f) and One Objective Function (g).......113
5.2 Utility Function with cl-t Risk...................124
5.3 Risk-Return Efficient Solutions, Kennedy and

Francisco PPODIMIIIIIIIIIIIIIIIIIIIIIIIllIIIIIIII127
5.4 Best (b) and Horse (w) Objective Function

Values for Three Steps of the IMGP Solution

Process, Kennedy and Francisco Problem...........138
5.5 Information Flow in Goaldir Decision Aid..........145
5.6 Monitor Displays of Hypothetical Conviction

Heights and Implied Probabilites..................151
5.7 Best/Horst Solution Values of Risk and Return,

Goal-Directed SEBPCI‘l Program.-...............-....I64
5.8 Activities Implied by Current Solution,

Goal-Directed Search Program......................165
5.9 Summary of Risk/Return Values for Previous

Steps, Goal-Directed Search Program...............166
8.1 Land Elicitation Template.........................222
8.2 Variable Cost Elicitation Template--Screen 1......223
3.3 Variable Cost Elicitation Template--Screen 2......224
8.4 Yield Probability Elicitation Template--Screen 1..225
B.5 Yield Probability Elicitation Template—-Screen 2..226
B.6 Yield Probability Elicitation Template--Screen 1..227
8.7 Yield Probability Elicitation Template--Screen 2..228

vi

CHAPTER I

INTRODUCTION

I-IWQLUQ!

This research has evolved from a concern that the apparently
boundless optimism with which many agriculturalists view new
developments in decision theory and computer technology ignores
fundamental gaps in knowledge of management decision processes. The
set of formal principles which economists have developed and utilized
with some success for prediction of aggregate variables and for
prescriptions for economic policy may be an insufficient basis for
advice to individual farm managers. Although there is little concrete
evidence to evaluate the impacts of training in decision-theoretic
principles, such impacts are not immediately obvious to an observer of
farm manager behavior.

One reaction to managers’ slow adoption of sophisticated decision
techniques is to impugn the vision and judgement of farm operators.
Managers are ’stubborn’, 'fixed in their ways’, or ’don’t know what is
in their best interest’. A contrasting perspective is adopted in this
research. There is intelligence and purpose in most decisions of farm
operators. The obstacles to better decisions are to be found
principally in poor understanding of current decision processes and of
managers’ perceived constraints and objectives. If more careful

attention is paid to these factors, and conceptual frameworks and

2
operational techniques are modified accordingly, progress in decision
support techniques for farm managers can be expected.

A related motivation for the research is concern for the role of
computer technology in support of farm management decision making.
Automation of many management functions is currently feasible with
extremely fast and powerful electronic technology, and some
agriculturalists contend that computerization brings unambiguous
benefits for farm businesses and families. The position taken here is
that computers and humans each have relative advantages for some
operations and types of decisions. Efforts should be devoted to
description of decision processes, determination of relative strengths
and weaknesses of humans and computers in different types of decisions
and assistance for managers to extend their capabilities with decision
tools which may or may not include computer technology. The
overriding motivation of such a perspective is to support manager

decision making, not displace it.

1.2 An v lut'onar A roach to D ' n Su r Re ea ch a d e i n
The current study does not follow the traditional research
pattern of problem definition, literature review, model development,

empirical testing. The ’problem’ examined here is the design of
effective decision tools for farm managers. The research approach is
that variously described in the decision support literature as
’evolutionary’, ’adaptive’ or ’iterative’ design. The objective of
this approach is not a polished, finished decision aid, but
development of an aid which embodies promising concepts of decision
support, and which can be expeditiously implemented, evaluated and

reformulated on the basis of research experience. The principal

3

objective of aid development is not analysis of model solutions, but
analysis of concepts and techniques which integrate conceptual
frameworks of decision making, which may be isolated and more
rigorously tested, and which open new avenues for applied decision
research.

The current orientation for decision research is not derived from
a single theoretical perspective. No single discipline provides an
adequate basis for management decision support research. Instead, a
synthesis of elements from contrasting perspectives are sought. Two
competing paradigms of decision making are defined and investigated in
Chapter 2, and their relative advantages for supporting individual
decision makers are discussed. The research literature of the two
camps (not surprisingly) emphasizes the importance of different
concepts and methodologies for describing decision processes. Issues
which are particularly emphasized by the school defined as
’rationalist’ are optimization, quantification and the conceptual
distinction between preferences and beliefs. Particularly emphasized
by the contrasting ’behavioral’ school are limited cognitive
information processing capacity, multiple preferences and aspirations
in decisions, the practical limits of rational decision making and the
importance of feedback and learning. Issues which are generally
accepted as equally important in both paradigms include the assumption
of fixed preferences (at least as an operational construct for
research) and the quantification of beliefs as probabilities.

Both decision paradigms are primarily concerned with unaided
individual decision making. An applied research perspective of

decision support has developed (concurrently with improvements in

4

computer technology) which contends that computer-based aids can
improve decisions. Some researchers have focused their investigations
on the computer tools themselves, with relatively little attention to
individual characteristics of users. Other researchers contend that
individual problem-solving style is an important determinant of
success in decision aid usage. Chapter 3 examine the relatively new
and unstructured literature on the interaction between characteristics
of decisions, individuals and aids.

To synthesize and operationalize the decision concepts of the
contrasting paradigms, a third body of literature is examined in
Chapter 4. Given the relative emphasis on optimization procedures by
one school and on multiple objectives by the other, multiple objective
mathematical models are examined that use conditional optimization for
solution search. Interactive methods which iterate towards solutions
preferred by the decision maker in response to articulated preference
information are shown to operationalize many concepts from both
paradigms of decision making. One multiple objective technique, the
interactive Multiple Goal Programming (IMGP) method, promises
considerable flexibility for such integration.

To evaluate the feasibility of this technique for implementation
as a real time microcomputer decision aid, a decision problem was
sought which is perceived by farm operators as important to their
business, for which many operators may desire assistance, and which is
characterized by multiple attributes or objectives (see Chapter 4).
The cropland rental decision for cash grain farms was indicated by
operators at extension meetings as important, and was chosen for aid

development. The consequences of financial stress in the cash grain

 

5
sector have included a more active land rental market (as some farmers
have liquidated their holdings) and a desire for more careful
consideration of existing and potential rental agreements. Many
operators must improve their revenues to relieve debt burden or
compensate for lower government support, but are hesitant to make
risky rental decisions which could jeopardize their financial
viability. The risk and return preferences of managers can be modeled
as distinct objectives in a decision aid implementation of the IMGP
technique. Although certain problems could be foreseen with this
implementation (such as the effect of rotation cycles and the
’lumpiness’ of rental decisions), the principal characteristics
desired for a case study of decision aid design were satisfied.
Chapter 5 documents development and field testing of the goal-directed
search (GOALDIR) decision aid.

. The final chapter completes a cycle of the evolutionary approach
to decision support research. Principal conclusions from the
dispersed literature of decision theory, decision support and multiple
criteria modeling, from the model development and from the field
testing are discussed. The decision support research agenda described
in the final section constitutes the last phase of the current

research cycle (and the first phase of the next).

 

CHAPTER II

PARADIB‘tS OF DECISICN WING

2.1 intrgduction

Kuhn defines a paradigm as a set of '. . . universally recognized
scientific achievements that for a time provide model problems and
solutions to a community of practioners.‘ (1970: viii) A paradigm is
thus a professionally acceptable network of concepts, theories and
methods used to investigate a certain body of phenomena. The
following chapter first examines the empirical and conceptual
relationship between management and decision making, and proceeds to
describe and analyze two contrasting research paradigms with respect
to the phenomenon of individual decision making. Each paradigm offers
different evidence concerning decision making ability and particular
approaches to aiding decision makers. The paradigms generally do not
take into account the insights offered by the other perspective. The
chapter concludes with a short discussion of some elements which may
provide a basis for integration of the two perspectives at the level

of individual decision support.

2.2 Managgment and Desision Making ‘

What is farm management? Initially, a clear distinction must be
made between the ggactigg of farm management (as an empirical
phenomenon) and the 352g! of farm management (as a conceptual
perspective of the phenomenon). The former connotation pertains to

the actual behavior of farm operators, while the latter pertains to

6

the analysis of empirical phenomena through the lens of a research
paradigm. Researchers sometimes do not clearly distinguish between
the practice of farm management (what is) and their particular
disciplinary and logical perspective of ’correct’ management (what
ghguld be). For example, textbooks on the subject often define farm
management as '. . . the subdivision of economics which considers the
allocation of limited resources within the individual farm..' (Heady
and Jensen 1954: 6).
A definition of farm management more closely corresponding to the
practice of farm management is given by Dillon. Farm management is:
The process by which resources and situations are manipulated by
the farm manager in trying, with less than full information, to
achieve his goals. (Dillon 1980: 258)
Several important elements should be noted in this definition.
First, farm management is clearly defined as a process, with no
specified beginning or end. The definition emphasizes that managers
act with incomplete information about the current state of the system
managed, about functional relationships and possibly about the
satisfaction which can be expected from actions. Finally, the
indication that farm managers may attempt to achieve goals other than
profit maximization illustrates that a multi-dimensional perspective
may be necessary to model farmers’ decisions. This definition makes
the generalization that a single person decides and executes
managerial actions, an assumption which has become progressively less
tenable as farm spouses and other family members provide more input to
both routine and strategic management decisions. The goals sought by
the management group may thus be a less than perfectly reconciled set

of personal and interpersonal preferences.

What distinguishes ’management’ from ’decision making’?
Conceptually and empirically, management performs functions extending
beyond the reflective and anticipatory role of making decisions. A
typical conceptual framework of management functions is given by
Johnson and Halter (1961):

1. Problem definition

2. Observation

3. Analysis

4. Decision

5. Execution

6. Responsibility bearing

Decision making as conceptualized includes at least the analysis
and decision functions, and may be defined to include as well the
problem definition and observation functions. Many researchers equate
‘ management with decision making. Johnson states, 'Hhile all of the
managerial functions depend on each other, the decision function is so
crucial that management is sometimes referred to as decision-making.‘
(1977: 18) A prominent text on decision analysis (Anderson et al.
1977) does not mention management in its index. Management functions,
however, include execution and control processes not generally
considered within conceptual frameworks of decision making or decision
research experiments.

Although clearly acknowledging the critical importance of all
management functions in the farm firm, this study is limited to
analysis of individual decision making. The general conceptual scheme

of decision functions accepted here is the widely recognized framework

of Simon (1965), who identifies the following phases of decision
making:

1. Problem identification. In this stage, the decision maker
may progress from a vague sense of dissatisfaction
concerning the present situation with respect to
preferences, to an identification of the source of
dissatisfaction and causal linkages to other elements of the
current situation.

2. Qgsigg. In this stage, the decision maker is involved in
generation of alternative actions which may be both feasible
and promise to change the current state to another thought
to better satisfy preferences.

3. fingigg. Alternative actions are compared either against
each other or against some criteria, and an alternative may
be chosen. '

There are several competing paradigms of decision research, each
with differing theoretical foundations, research methods and research
objectives. The two schools of thought examined here are alternative
treatments of the same phenomena, and may be labeled the Rationalist
and the figngvigralist schools of thought. Other terms used in the
literature for the rationalist school include decision-theoretic,
normative, prescriptive and axiomatic. The behavioralist school is
also sometimes called empiricist or descriptive. Most of these terms
carry more emotive than descriptive meaning, and the labels are used
here without endorsing any such connotations.

Both approaches examine the efficiency with which means are

utilized to reach goals, and accept the positivist or conditionally

1B

normative view that nothing scientific can be said about questions of
intrinsic goodness or badness of goals. No intrinsic preference is
attached to means, and both tend to regard goals or preferences as
fixed for purposes of the research, although behavioral research is at
least open to examinations of the dynamics of preference (March 1978).
The contrasting perspectives are both theories of ’innocent’ decision
making, disregarding any influence of other people’s preferences in
the decision maker’s choice process. They attempt to extract
individual decision makers from the social structures, organizations
or groups within which most decisions are actually made. in general,
the common research focus of these perspectives is goal-seeking
behavior or means-ends analysis (Checkland 1985).

The approaches differ principally with respect to the role of
human cognition, acceptable research context and procedures, and the
operational objectives of decision research. Behavioralist
investigators place the structure and mechanisms of cognition at the
center of decision research, while rationalist investigators do not
attempt to probe unseen cognitive processes. Research methods
encouraged and acceptable within one framework are unlikely to receive
acceptance within the other. By describing actual decision processes
in considerable contextual detail, behavioral research attempts to
induce general principles of how decisions are made. Rationalist
research, on the other hand, applies logical principles to a broad
range of decision types and contexts with the objective of predicting
or prescribing decisions. The following sections describe the two

perspectives in more detail and examine issues of dispute.

2.3 The Rationalist Perspectivg

Rationalist decision research seeks to describe, predict or
prescribe the choice of an optimal alternative from a set of
alternatives. Decision makers are assumed to have an internally
consistent ordering of preferences whereby outcomes can be compared in
terms of subjective value and a decision rule by which a preferred
alternative-consequence can be selected. On the basis of known
preferences and expectations, the decision maker examines alternatives
and their consequences in view of perceived constraints, orders the
consequences in terms of their subjective worth, and selects the best
alternative. Choices which conform to a minimal set of logical
assumptions are formally representable in a real-valued function of
preference relations. There is no explicit treatment in the theory of
a process whereby alternatives are developed or discovered and their
consequences determined, how probabilities are formulated, or how
preferences themselves are established or modified.

The appeal of rationalist models of decision making is
principally their consistency with rules of logic, their parsimony and
generality in depicting decision situations, and their prescriptive
value in generating solutions in accordance with the measured values
and preferences of decision makers. Research is driven by what Blaug
(1980) calls a ’hypothetico-deductive method’, in which theoretical
implications generate hypotheses which are then tested empirically.
Rationalist models are principally outcome-oriented, evaluated by
their logical consistency and comparison of empirical results with the
calculated ’best’ solution. The process of individual decision making

consists of implementation of the logical process, with no explicit

12
regard for what might be called ’intuitive’ human decision processes.
The primary purposes of rationalist research, therefore, are
prediction of actions by assumedly rational decision makers and
prescription of actions which gngglg be chosen by a decision maker
with given preferences. Prediction may be limited to decisions of
aggregate or of ’representative’ decision makers. There is relatively
little emphasis on the descriptive validity of the theory; that is,
whether the theory reasonably describes any particular decision making
process and how actual decisions may conflict with those suggested by
the theory. Indeed, one perspective within the rationalist school
contends that it is irrelevant whether decision makers consciously
follow the logical processes implied by rationalist assumptions, only
that the decision is made ’as if’ they understood logical principles
(Friedman and Savage 1948).

Modern rationalist decision theory, with origins in the Bernoulli
concept of utility, constitutes an attempt to formalize human decision
processes in non-deterministic environments. If the environment is
considered deterministic and preferences are fixed (as in the static
theory of economics), there is no ’decision’ in the normal sense of
the term. Instead, measurement of known alternative-outcome pairs and
comparison against the individual’s preference structure suffices to
predict actions. The first formal set of axioms linking probability
and utility was developed by von Neumann and Morgenstern (1944) as the
Expected Utility Hypothesis (EUH). The EUH is based upon the
assumptions that: 1) individuals have a stable and consistent set of
preferences (evaluative judgements about the world); 2) individuals

have a consistent set of expectations about events (predictive

13

judgements); and 3) preferences and expectations are independent (no

wishful

thinking). (Hogarth 1980)

The following are logical axioms underlying the Expected Utility

Hypothesis:

1'

Orgering: All pairs of alternatives can be compared by
either a preference relation or an indifference relation.
For any risky outcomes A and 8, individuals will consider A
preferred to B, 8 preferred to A, or will be indifferent
between A and B.

Transitivitz: Preference relations are transitive across
alternative-outcome pairs. If A is preferred to B, and B is
preferred to C, then A is preferred to C.

Continuity: If transitivity applies across three
alternatives, then for some unique probability (p), an
individual will be indifferent between choices of the
intermediate outcome with probability p on the one hand, and
an exhaustive probability mixture of the most and least
preferred outcomes (weighted by p and i-p) on the other
hand. If A is preferred to B, which is in turn preferred to
C, then there exists some probability p such that the
individual is indifferent between piB and (phA f (1-p)XC).
lgggpgnggggg: Pairwise preference for risky outcomes is not
affected by identical probability or outcome changes in the
original outcomes. If A is preferred to 8, then (piA f (1-
p)!C) is preferred to (pia f (1-p)!C), where C is some other

outcome. (Schoemaker 1982)

14

The representation of beliefs or expectations for both repetitive
and unique events in a manner consistent with mathematical probability
is also presumed in EUH and in all of rationalist theory. The work of
Ramsey (1931) and DeFinetti (1937) developed the concept of personal
or subjective probability. These beliefs are purely subjective, based
upon any experience, rules of thumb, historical data or other factors
which the decision maker wishes to use to develop expectations. Such
expectations expressed quantitatively as probabilities, however, are
required to conform to the criterion of coherence (Savage 1954). This
requires that the quantitative expression of beliefs (subjective
probabilities) are mutually exclusive and exhaustive, and that
assessments of disjunctive and conjunctive events conform to the
addition, product and equivalence rules. Subjective probabilities are
not required to agree with any external standards of likelihood or
expectation, but they are expected to express the individuals’s degree
of belief that particular events will occur. it bears repeating that
subjective probabilities must be independent of preferences for
outcomes.

Although rationalist theory does not indicate how expectations
are or should be formulated, it does state that dynamic revision of
probability estimates upon receipt of new information should be
carried out in a manner consistent with Bayes’ Theorem. This theorem
presents a logical framework to predict or prescribe probability
revision. Generally, it states that the new probability estimate
(expectation or belief) that a particular outcome will occur should
equal the product of the prior probability estimate times the

likelihood that the new information is correct, the result being

15
normalized between 8 and 1. Bayes’ Theorem does not indicate how
information of less than complete reliability should be treated, nor
how information should be utilized which is not expressed as a
probability.

Although not strictly required by the EUH, utility functions are
usually considered uni-dimensional, with the single argument of wealth
or income. In some cases, however, a multi-argument utility function
is assumed, but all arguments other than income or wealth are assumed
to be held constant. In response to observed anomalies in decision
making, the rationalist approach has been extended to Multiple
Attribute Utility (MAU), functions with multiple, incommensurable
preferences. This will be discussed below as one aspect of the

response to behavioralist criticisms.

2.4 Thg Behgvigraligt Perspective

Normative models gain their generality and power by ignoring

content in favor of structure and thus treat problems out of

context. However, content gives meaning to tasks and this should
not be ignored in trying to predict and evaluate behavior.

(Einhorn and Hogarth 1981: 61)

The focus of behavioral decision research is the decision process
itself, that is, the strategies used to select, combine or alter
information and reach decisions within specific types of decision
contexts. Since cognition is unobservable, research efforts center on
the observable selection of information from the environment, which
may reveal characteristics of the decision process. Principal issues
investigated include assessment of uncertainty and expectations
assessment of uncertainty and expectations (Hogarth 1988), decision

rules (Svenson 1983), information search (Bettman and Jacoby 1975) and

multiple attributes or objectives (Fishburn 1978). In addition, much

16

behavioral research has involved comparison of the performance of
human decision makers against the implications of rationalist
assumptions, and has raised important questions about observed
discrepancies. Behavioral decision models stress the multiplicity and
ambiguity of goals and the inconsistency of judgements resulting from
inherent limitations on human ability to perceive, combine and
evaluate information about the decision situation. Humans, it is
argued, make frequent and systematic errors in assessment of
expectations and logical operations, and apply rules which generate
sub-optimal solutions.

Behavioral decision theory is much less formalized and more
empirically based than its rationalist counterpart. It is constructed
rather more on the inferences of past research results than on the
implications of a formal theoretical framework. Development of
behavioral decision theory can reasonably be said to have occurred in
parallel and in reaction to developments in rationalist theory. Most
behavioral research attempts to construct a more realistic descriptive
model of how limited human memory and computational capability
interact with complexities of the decision problem and its context.

The decision makers in behavioral models lack some of the

relevant knowledge; they may, in addition, fail to make use of

some of the knowledge which they do have, or to which they at
least have access. Their problems are incompletely structured,
and the variables of interest are incompletely specified.

(Loasby, 1976)

The behavioralist approach stems principally from the seminal
work of Herbert Simon and his colleagues. Although the behavioralist
argument has deepened and extended to areas not initially

contemplated, the cornerstone of the framework is that humans are not

17
capable of the cognitive operations implied by the rationalist
perspective. All decision making is behavior within cognitive
constraints. Simon (1976) contends that humans can be viewed as
limited capacity information processing systems. The cognitive
processing equipment is basically serial in organization, that is, it
can handle only one operation at a time, and solution requires a large
number of operations. The analogy to a computer is made explicit:
Man and computer can both recognize symbols (patterns), store
symbols, copy symbols, compare symbols for identity, and output
symbols. These processes seem to be the fundamental concepts of
thinking as they are of computation. (Simon 1976: 71)
The minimal components of a general information processing system
are:
1. Memories containing discrete symbols
2. Receptors for sensing the environment
3. A set of primitive operators which interpret sensory input
and transform memory contents
4. A set of rules which combine operators and memories to
generate whole programs for information processing.
Therefore, a valid explanation of an observed behavior consists of an
external (to the individual) program of data, rules and operators
which reproduces the observed behavior. (Newell et al. 1958)
The principal functional characteristics of the human information
processing system are:
1. Selective perception or limited attention. For example,
Hogarth (1988) reports it has been estimated humans perceive

only 1/78th of the contents of the visual field at a time.

18

2. Sequential processing of data. The mechanisms of the
cognitive system operate serially, performing only one
function at a time.

3. Limited computational capacity. The system has only limited
capability to perform numerical functions.

4. Limited short term memory. Miller (1956) concludes that
humans can retain in active memory only five to nine
’chunks’ of information.

5. Dependence on context variables. Through experience, humans
construct mental patterns of co-occuring variables. They
use these patterns to select information in decision
situations characterized by little direct information about
the decision variables.

These information and computational restrictions limit both the
number of alternatives that can be simultaneously considered and the
amount and accuracy of information that is actually considered in the
decision situation. As a response to cognitive information processing
limitations and overwhelming information input from the environment,
humans employ functional processes such as sequential and selective
attention to stimuli and to one’s own goals, ’efficient forgetting’ of
information, dependence on information from the context of a problem,
and cognitively simple decision rules (’rules of thumb’ or
heuristics).' Application of these procedures in many situations (even
relatively simple ones) will result in decisions which are sub-optimal
when compared to rationalist models. Simon points out the severe
cognitive demands on the human decision maker or ’organism’ implied by

rationalist models:

19

The organism must be able to attach definite payoffs (or at least

a definite range of payoffs) to each possible outcome. This, of

course, involves also the ability to specify the exact nature of

the outcomes-~there is no room in the scheme for ’unanticipated
consequences’. The payoffs must be completely ordered--it must
always be possible to specify, in a consistent way, that one
outcome is better than, as good as, or worse than any other.

And, if the certainty or probabilistic rules are employed, either

the outcomes of particular alternatives must be known with

certainty or at least it must be possible to attach definite

probabilities to outcomes. (Simon 1979: 10)

Because of information processing constraints, decision makers
seldom search all the possible alternatives for the optimal solution.
Instead, alternatives are searched until one is found that is ’good
enough’ according to the decision maker’s preferences. This
’satisfycing’ principle is probably the best known of the concepts
associated with behavioral theory.

Behavioral researchers do not attempt to depict the human
decision maker as cognitively handicapped. Human decision making has
many strengths which may never be matched by mechanical devices. Very
complex problems can be solved even though they are not completely
understood. Intuition and creativity are certainly strong points of
cognition. Situations in which rational procedures are inappropriate
also come to mind. If decisions are required quickly or if the
decision is of limited importance, logical considerations may be too
costly. Humans often do the best they can, forming interpretations
based upon experience when faced with decisions in which preferences,
information, constraints and possible alternatives all may be ill-
defined. Further, simple rules of thumb may in many cases generate

solutions which are not substantially different from optimizing

solutions.

28

Decision researchers who investigate biases in human information
processing argue that biases reveal much about the psychological
processes that govern decision making. In addition, research on
biases indicates which principles of rationalist decision making are
counter-intuitive or ’unnatural’, and suggests procedures which might
improve the quality of decision making.

Recent developments in behavioral research have criticized the
information processing approach as inadequately representing decision
processes. Most of these developments have implications far beyond
the limited problem solving orientation of Simon’s framework. Certain
research suggests that choice may precede search and evaluation, that
is, action and feedback are sometimes substituted for classical
decision making (March and Olsen 1976). Personal decision habits have
also been studied in a wide range of cognitive style studies. This
research emphasizes inherent tendencies by individuals to approach
decisions in a certain manner rather than the information processing
emphasis on cognitive capacity (Keen 1978). Neisser (1963) contends
that the sequential processing concept of the information processing
framework is the result of an inappropriate analogy of human to
computer processing. The concepts of short term and long term memory
have also been challenged (Glass et al. 1979). Another broad area of
behavioral research includes cognitive schemas or scripts, which are
coherent sequences of events expected by the individual on the basis
of past experience (Abelson 1976). Many of these developments do not

imply willful choice or goal-seeking behavior in the same sense as

 

21

information processing or rational choice models, and will not be

pursued here I.

2.5 Obgerveg Errors in Decision Making

A primary issue of behavioral research has been evaluation of
humans’ ability to estimate uncertainty in terms of probabilities.
Subjective probabilities are quantitative expressions which reflect
the individual’s degree of belief about an event (Ramsey 1931).
Expression of beliefs as probabilities provides an interpersonal
language for expressing uncertainty, and through the probability
calculus allows analysis of the logical relationships between
uncertain events. It is not appropriate to assume that individuals
carry around a complete set of probability distributions in their
minds. Instead, a set of vaguely formulated beliefs are combined with
information from memory and information from the environment to give a
probability estimate. Obviously issues of probability elicitation
methods cannot be separated from issues of belief formation and use in
decision making. Of particular concern is whether probability
elicitation methods provide a valid representation of beliefs and
whether probabilities are stable, consistent and in accord with the
rules of probability.

The principal objectives of most behavioralist studies of
probabilistic judgements has been to compare behavior to that implied
by logical axioms and to determine how underlying cognitive processes

are affected by the interactions between the decision characteristics

 

For a discussion of these developments in cognitive science, see
Gardner 1985).

22

and cognitive limitations of the decision maker. Hundreds of
experiments have been carried out with non-expert and experts in
laboratory settingsl. The performance of non-expert subjects in
laboratory settings tends to show that '. . . man is not a good
statistician.I (Keen 1977: 45)

In simple tasks, such as estimating repetitive series like
drawing white or black balls from an urn, people seem to judge
probabilities fairly well (Peterson and Beach 1967). However, in
research involving unique events in which relative frequency has less
meaning, people tend to perform very poorly (Lichtenstein et al.
1977). in particular, people tend to be overconfident in their
probability assessments, ignoring such factors as sample size (Tversky
and Kahneman 1971), regression towards the mean (Kahneman and Tversky
1973) and the reliability of their data base (Peterson 1973). People
also seem to have a very poor intuitive sense of variation and
covariation (Kahneman and Tversky 1972). Individuals tend to
overestimate the probability of conjunctive events and underestimate
the probability of disjunctive events (Bar-Hillel 1973). Individuals
tend to recall their predictions as better than was the case (Fischoff
and Beyth 1975) and seem to retain memory for successes while
forgetting failures (Langer and Roth 1975).

A substantial body of research has been dedicated to comparison
between Bayesian probability estimates and estimates given by research

subjects in experimental tasks. Edwards (1968) termed humans

For reviews of this extensive literature, see Spetzler and Stael
von Holstein 1975, Lichtenstein et al. 1982 or Nailsten 1983.

23
‘conservative Bayesians’ from the results of his studies, which showed
that individuals underweight new information. On the other hand, a
considerable number of experiments have found that individuals often
ignore prior probabilities in favor of new information (Tversky and
Kahneman 1980). The errors in Bayesian tasks have led most
behavioralist researchers to conclude that '. . . in his evaluation of
evidence, man is apparently not a conservative Bayesian: he is not a
Bayesian at all.‘' (Kahneman et al. 1982).

The evidence is less pessimistic in studies with experts (persons
who have substantive knowledge of the topic or with training in
probability theory). These studies have included subjects such as
weather forecasters (Murphy and Ninkler 1977), physicians (Lusten
1977), psychologists (Beenen 1970), security analysts (Bartos 1969)
and military intelligence officers (Johnson 1977). Given outcome
feedback, experience with probabilities, substantive knowledge about
the topic and a reliable elicitation technique, experts can provide
relatively accurate probability estimates (Nailsten and Budescu 1983).
Performance in the absence of one or more of these elements may not be
better than that of non-experts.

Behavioral researchers have investigated the cognitive sources of
biases in probabilistic judgements. Because of information processing
limitations, decision makers use simple rules of thumb or heuristics
to estimate likelihood. The three heuristics described by Tversky and
Kahneman (1982) are the best known:

1. Rgpggsentgtivenesg. This concerns a judgement whether event

A belongs to set 8. Using this heuristic, the degree to

which the ‘essential’ characteristics of A are

 

24
representative of set B determines the degree of belief
(probability) that A belongs to B. This rule can explain
some errors in probability assessments such as Bayesian
errors, insensitivity to sample size, misconceptions of
randomness or of regressiveness, and overconfidence.
Anderson et al. (1977) gives an example of a farmer who
judges a current short dry spell as representative of the
beginning of a past drought, ignoring the prior or
historical probability of drought in the area.
Avgilapilitz. Using this rule for judging uncertainty, the
frequency of an event is estimated by how easily similar
events can be recalled. If two events occur simultaneously,
an illusory sense of correlation may develop due to this
rule. For example, a farmer’s judgement of the incidence of
mechanical failure among tractors of a particular brand
could be unduly influenced by a neighbor’s bad experience
with that brand.
Anghorigg gpg Adjugtment. This heuristic concerns the
tendency of individuals to make estimates based on some
‘natural’ anchor point. Subsequent adjustments from that
point based on information from the decision environment is
often insufficient from the perspective of rationalist
choice. This rule may cause errors in revision of
probabilities or of probabilities for disjunctive and
conjunctive events. An example might be estimation of next
year’s yield by taking this year’s figure and adjusting it

by some percentage.

25

A second important area of behavioral research has been the
choice rules used by subjects to edit the set of alternatives and
choose a preferred solution. Considerable research and casual
observation indicates that decision makers commonly make choices based
upon processing of information about multiple dimensions of
alternative actions. Choice rules can be categorized as compensatory
or non-compensatory. Compensatory rules permit tradeoffs between
dimensions or attributes of alternatives, while non-compensatory rules
do not. Among non-compensatory rules are included:

1. Qppjunctive rules. Alternatives which do not have

satisfactory levels of any attribute are eliminated.

2. Qigjgngtive rules. Alternatives are evaluated only on their
best attribute.

3. Lgxicographic rples. Attributes are hierarchically ordered
in importance, and alternatives are evaluated sequentially
down the hierarchy, with choice being made as soon as one
alternative has a better value than any other on a higher
ordered attribute.

4. Elimination by aspects. Using this rule, attributes are
selected randomly and alternatives that do not have the
characteristic are eliminated. The process continues until
only one alternative remains. (Hogarth 1980)

Compensatory rules include simple linear formulations reflecting
relative weights of importance and scale values of the attributes
(Hogarth 1980). Another rule is the ’ideal point’ rule, in which
alternatives are evaluated by their proximity to ideal values of the

various attributes or dimensions (Zeleny 1982). A further choice rule

 

26
uses the principle of dominance to eliminate all alternatives except
those which have better values of one attribute.

Non-compensatory rules are much easier to apply in most choice
situations characterized by incomplete data, incommensurable
dimensions, information overload, time pressures and large numbers of
alternatives (Slovic et al. 1977). However, choice may sub-optimal
when evaluated in a utility context. Subjects may use many different
rules in a multi-stage decision process, with non-compensatory rules
at initial stages and compensatory rules for a reduced set of
alternatives. Prediction of the type of decision problem and context
which evoke a particular rule has been problematic.

Schoemaker (1982) reviews some of the most notable experiments
demonstrating violations of rationalist principles. In one of the
first behavioralist experiments, Mosteller and Nogee (1951) found that
subjects would change preferences during deterministic and stochastic
repeated choice tests. Tversky (1969) also illustrated systematic
violations of transitivity in deterministic and stochastic choice.
Lichtenstein and Slovic (1971) conducted experiments in which subjects
were asked to select a preferred gamble and to name certainty
equivalents between gambles. In many cases, the preferred gamble was
seen to have a lower certainty equivalent than other gambles.

Kahneman and Tversky (1979) are among many behavioralist

researchers who have replicated Allais’ paradox of the following form:

27

(2.1)
Problem 1:
p= .10 $5 million
/
A [J p= .89 $1 million 8 [1 pH 1.0 $1 million
\
p- .01 80
Problem 2:
p! .10 $5 million p- .11 $1 million
/ /
C [I D [I
\ \
p= .90 30 p. .89 S0

A large majority of subjects choose 8 over A, but C over D.

Since the second problem is formulated by subtracting a $1 million
outcome of probability= .89 from A and 8, the experiment clearly shows
choice inconsistent with rationalist theory. Outcomes with certainty
seem to be weighted more heavily than outcomes which are merely
probable. The experiment has been replicated many times with
probabilities and payoffs much less extreme than those above.
Responding to the criticism that subjects would not commit such errors
if they understood rational decision making principles, Slovic and
Tversky (1974) explained the error to violators. A large proportion
of such violators would not agree to change their choice.

Subjects demonstrated persistent violations of continuity and
transitivity in experiments conducted by Coombs (1975). In these
tests of the EUH, nearly half the subjects incorrectly ordered three
lotteries in which one was a probability mixture of the other two
(implying that its preference should be intermediate). In addition,
many researchers have documented examples of choice situations in

which different representations of formally identical problems

28
affected choice. In these experiments (Kahneman and Tversky 1979),
individuals demonstrated risk averse or risk preferring behavior
depending whether identical options were presented as gains or losses.
Regarding choices as either gains or losses seems to serve as an
editing procedure for decision processes which operate in ways which
violate rationalist axioms.

Reported biases of EUH principles also result from unreliable
elicitation procedures. Hershey et al. (1982) present a comprehensive
review of utility assessment procedures and discussion of sources of
bias observed in experiments. They conclude that confounding effects
of the measurement process and the problem context cast doubt on
practical applications of the EUH.

Kahneman and Tversky (1979) have developed an alternative theory
of individual decision making under risk, called prospect theory.

This model describes two phases of decision processes, consisting of
an initial editing phases and a subsequent phase of evaluation. The
coding phase consists of coding (evaluating outcomes as gains or
losses), combination (combining probabilities of identical outcomes),
segregation (separating risky and riskless components) and
cancellation (discarding components shared by alternatives). Many
anomalies of preference can be satisfactorily represented with these
editing procedures. The evaluative phase of prospect theory resembles
utility theory, in that the components include a value function, an
expectation, and a compensatory rule. The value function differs from
a utility function in that it is defined on deviations from a
reference point, it is concave for gains and convex for losses, and it

is steeper for losses than gains. The expectation weights are based

29
on probabilities, but typically do not sum to unity and small
probabilities are overweighted. Although this theory presents a
significant attempt to reconcile rationalist theory and behavioralist
observations, there has been little additional research using prospect
theory as a conceptual framework.

A critical and as yet unanswered question involves the
‘reasonableness’ of using probability and choice rules which violate
rationalist principles. In most everyday situations, individuals
would seem to be foolish to carry out an extensive rational decision
process. Simple rules may also be necessary in decisions made under
time pressures or when the situation itself is not well defined or
understood. If these intuitive rules are functional in many
circumstances, how could ‘switching rules’ be developed which indicate
when the intuitive rules are likely to cause errors and when
rationalist rules should be used? Another important question is
whether utilization of particular decision rules is under the
individual’s cognitive control. If the mechanism which calls these
rules is sub-attentional or automatic, then the prospective for
helping individuals recognize decisions in which these rules are
likely to cause errors and to correct those errors is not bright.

As noted above, a persistent finding of behavioral decision
research is that individuals evaluate attributes (or characteristics)
of alternatives rather than determining preference for each
alternative in a holistic manner. Since values of attributes are
often subjectively determined, this implies innumerable operational
problems for uni-dimensional decision theory. Indeed, this single

discovery has sparked the entire field of rationalist Multiple

30
Attribute Utility theory, with the expressed aim of determining
utility preferences for incommensurable attributes. Behavioral
researchers have indicated that adding more arguments to the
individual utility function may resolve many of the preference

inconsistencies observed in their decision experiments.

2.6 The Rationalist Rgply

Rationalist researchers have strenuously defended their own
perspective. One response to behavioral criticism by rationalist
decision researchers has been to discount the quality of research
which indicates violations of theoretical principles. After three
decades of experiments, this argument no longer seems tenable. At
least as careful attention to experimental methods has been
demonstrated in behavioral research as in problem-oriented rationalist
research. Many results have been replicated in various settings with
different types of decision makers. For example, the Allais paradox
has been replicated many times by researchers such as Kahneman and
Tversky (1979) and Ellsberg (1961).

A related criticism concerns the experimental subjects and
controls of behavioral research, which is said to involve trivial,
artificial choices with naive, disinterested subjects. It is the
contention of rationalist researchers that savvy decision makers
presented with realistic incentives would not commit such errors.
However, this position is not substantiated by evidence from carefully
designed experiments. As noted above, subjects sometimes refuse to
change their ’irrational’ choices even after researchers explain their
errors (Slovic and Tversky 1974). Grether (1980) found evidence

supporting the representativeness heuristic and resulting biases in an

31

experiment testing Bayesian estimation, even though monetary
incentives were provided to subjects. Grether and Plott replicated an
experiment showing preference inconsistencies, controlling for a wide
Variety of economic, psychological and experimental method effects.
Despite their controls, they concluded '. . . the preference reversal
phenomenon which is inconsistent with the traditional statement of
preferences remains.‘ (1979: 634).

Preference reversals and probability assessment errors have also
been observed in laboratory environments with expert subjects. Slavic
et al. (1977) cite several such research experiments. Trained
psychologists were observed to disregard sample size in probability
assessments (Tversky and Kahneman 1971). A naive strategy of
predicting closing prices for selected stocks was better than
predictions of stock market experts (Stael von Holstein 1972).
Psychology graduate students made no better decisions in their areas
of expertise than in areas of general knowledge (Lichtenstein and
Fischoff 1976) .

The validity of laboratory experiment results for predicting
behavior in non-laboratory decisions is sometimes criticized. Ebbesen
and Konecni (1980) contend that results are very specific to simple
laboratory decision problems, pecauge of the observed use of sub-
optimal strategies. They base this contention on results of their own
study of decisions in real situations (eg. setting of bail, driving a
car), which indicates that real-world decisions involve many more
factors related in complex patterns than do laboratory problems.

Ebbesen and Konecni contend that:

32

There is considerable evidence to suggest that the external

validity of decision making research that relies on laboratory

simulations of real world decision problems is low . . .

Researchers should provide external validity evidence for claims

that causal models derived from laboratory data apply to

decisions in real-world settings. (1980: 42—43)

There is also some non-laboratory collaboration of behavioralist
results. Gamblers pursued sub-optimal strategies in on-site casino
experiments (Bond 1974). Kunreuther et al. (1978) interviewed
homeowners in flood plains and earthquake zones to determine whether
their disaster insurance decisions were in accord with utility theory.
One-half the samples had no knowledge of disaster insurance. They
found that approximately one-third of the remaining homeowners did not
act in conformity with expected utility optimization. Apparently
homeowners did not perceive the hazard in the same manner as expected
by policy makers. Researchers in psychology and education were seen
to regularly design experiments with inadequate statistical power,
reflecting the same errors as experts in laboratory experiments (Cohen
1962). A classic case of such bias was reported in Berkson et al.
(1940), which showed that instructors of laboratory technicians
demanded more accuracy in blood cell counts than was possible given
sampling variation.

Another response to the behavioralist critique has been to
broaden the scope of rationalist theory. One argument includes the
‘cognitive costs’ of decisions (Shugan 1980). Observed ‘biases’
appear to violate rationality principles because the experiments do
not consider the ‘cost of thinking’. This notion recognizes cognitive
constraints and yet preserves the concept of rationality. It would

appear, however, that the cognitive cost perspective runs counter to

the concept of bounded rationality, since it imposes an additional

 

33

stage of cost/benefit analysis on decision making (marginal analysis
of additional ‘decision investments’). In addition, it should be
noted that this makes a substantial part of rationalist theory
untestable and unfalsifiable, and does not suggest any feasible way to
improve decisions.

The cognitive cost position is also related to the principle of
‘meta-rationality’. Toda (1988) points out that people make ‘meta-
decisions’ in order to avoid the necessity of making a unique decision
in each future occasion. A cognitive control mechanism will monitor
and adapt rules which are persistently dysfunctional. In other words,
people are ‘global maximizers’ with local inconsistencies (Elster
1979). Also related are concepts of adaptive rationality (Day and
Groves 1975) and evolutionary rationality (Nelson and Hinter 1973).

If people are adaptively rational, plus preferences and the
environment are stable over some period, then behavior will tend to
approach the behavior postulated in individuai ‘calculated
rationality’ models. Evolutionary rationality, like adaptive
rationality, emphasizes the intelligence of decisions in relation to
the social or economic system.

Rules of behavior achieve intelligence not by virtue of conscious

calculation of their rationality by current role players but by

virtue of the survival and growth of social institutions in which

such rules are followed and such roles are performed (March 1978:

593).

However, such concepts as systemic and adaptive rationality or
cognitive cost, unlike individual ‘calculated’ rationality, do not

provide a basis for understanding behavior as a consequence of pre-

established preferences. Such a re-definition of rationality also

34
'. . . pays the cost of destroying the practical relevance of
normative prescriptions for choice.‘ (March 1978: 597).

A positivist perspective concerning the behavioralist critique is
to dispute the relevance of experiments to test the descriptive
validity of rationalist axioms. it is only necessary that individuals
act ‘as if’ they understood the actual principles (i.e. only
predictive power Justifies a model) or that individual biases will
average out in the aggregate (only macro-rationality matters). The
predictive superiority of EUH models, however, has not been fully
substantiated (Robison i982), and the notion that only predictive
power matters is ‘epistemologically unappealing’ (Samuelson 1963).
This perspective also allows no grounds for aiding management decision
making.

Some rationalist research has explored the possibility of
modifying those requirements of rational decision theories which seem
to be the cause of observed preference reversals. The independence
axiom has received particular attention. Machina (1982) develops a
version of the Expected Utility axioms without the independence axiom.
As noted by Robison (1982), this reformulation explains some
preference inconsistencies, but makes expected utility only a local
measure of preference.

Application of the EUH axioms is not theoretically restricted to
single argument utility functions, although utility defined on the
single argument of wealth or income has been the norm. Efforts have
been directed during the past two decades to extend the rationalist
approach to decision making with multiple preferences. Adapting the

theory to account for multi-argument utility is relatively

35
straightforward, but formulation of testable hypotheses and
operational elicitation techniques has proved extremely difficult.
Theoretical and applied research has concentrated on conditions
necessary for decomposition of a multiple argument utility function
into a series of single dimension functions which can be assessed with
uni-dimensional methods and then aggregated together to provide a
global measure of utility. The most common approach is to assume that
multiple argument utility functions are additively separable. Even if
the separability conditions are not strictly satisfied, Yntema and
Torgerson (1967) show that main effects are usually much more
important than interaction effects. This indicates that the ordering
of alternatives may not be affected by ignoring interactions.
Operational techniques have been developed to test separability
conditions and elicit utilities, but restriction to primarily riskless
choice situations and the expense and complexity of elicitation have
limited broader use of the techniques.

Hultiattribute utility formulations are often developed as
descriptions/explanations of behavioral inconsistencies with uni-
dimensional utility models defined over income or wealth. Bell
proposed a utility function formulation defined over two attributes,
income or wealth and ex post decision regret. The latter attribute
consists of '. . . the difference between the value of the outcome
that occurred and the value of the outcome that would have occurred
had the other alternative been selected.‘ (1980: 29-38) Bell shows
how this formulation can be used to explain a wide range of observed
preference reversal phenomena, such as the Allais and Ellsberg

paradoxes. This formulation, it should be noted, makes utility

36
measurements much more bound to the context of any particular
decision, as do formulations which regard preference as defined over
gains and losses rather than over final asset position.

The strong theoretical ties to utility theory and the committment
to dealing with the realistic complexity of decision structure make
multiattribute decision analysis very attractive, its operational
problems notwithstanding. This type of analysis '. . . tend(s) to
blur the distinction between descriptive and prescriptive theory.“
(Pitz, 1984: 157). The scope of its empirical applications includes
complex problems with multiple objectives, alternatives with multiple
attributes, intangible or incommensurable factors, extended time
horizons and risk in various formulations. The operational base of
multiple attribute formulations usually depends upon the
decomposability of assessed utility functions, which introduces a
level of complexity not found in single attribute utility models.

A further response accepts the results which show that people
commit systematic errors in decisions, and counters that behavioralist
researchers should be in the forefront of developing educational
techniques to correct observed errors and bring behavior into
agreement with rationalist principles. However, if the cognitive
limitations demonstrated by Simon and others are inherent in
cognition, education efforts may not be effective. It also should be
noted that decision makers recognize their decision procedures are not
’rational’ but persist in decision strategies which practically
guarantee choices violating rationalist principles. March (1978)
suggests that such persistent behavior suggests a form of

intelligence. Actions can often be seen as reasonable if contextual

37
information about a decision is examined, even if logical procedures
could not have generated the same decision. ’Reasonable’ decision
practices require refinement (not replacement) by the techniques of
choice theory.

No alternative behaviorally-based theory exists which covers the
broad range of phenomena treated in rationalist theory. As a
‘stalking horse’, the latter theory has been the focus of most
behavioral research and has illuminated aspects of human decision
making which have stimulated modifications in rationalist theory.
Keen and Scott Morton state:

Without the precision and formalism of rationalist theory, we

would almost certainly have made less progress in developing

descriptive insights; it has provided axioms to be challenged,
hypotheses to be opposed by counterexamples, and a vocabulary
that we need to use even to disagree with it. For example, the
concept of consistent, absolute utility functions has been
invaluable in all theories of decisionmaking, especially those

that argue such functions are nonexistent. (1978: 65)

Schoemaker (1982) notes that rationalist decision theory has been
a ‘research heuristic’ which summarizes current scientific knowledge,
encourages development of mathematical tools in ways which tend to
have secondary benefits, and leads to unexpected hypotheses and
applications. However, he also cautions that use of the closely
related ‘rationality heuristic’ and ‘optimality heuristic’ leaves the
scientist susceptible to biases of attribution (assuming that since
optimality models fit natural data well that nature therefore
optimizes), the illusion of explanation (allowing tautology to replace

explanation), and the bias of searching for confirming rather than

disconfirming evidence.

38
2.7 l t f P am k f nt r 'on

This chapter has examined the issues of debate between two
contrasting perspectives of individual decision making. Although many
of their concepts and underlying purposes of research are similar, the
perspectives have tended to develop as opposite poles of a dialectic,
with little room for communication or compromise on issues of theory
or operational methods. Yet a need exists for more coordinated
development, using the results of one school for development of the
other. As mentioned above, the behavioral school would certainly not
have advanced so far if it had not used the rationalist model as a
basis of its research. On the other hand, there seems little doubt
that rationalist theory is partly inconsistent with what is known
about real decision processes. Hultiattribute formulations form part
of the rationalist response to new insights brought to light by
behavioralist researchers, as do attempts to re-formulate utility
theory in ways which explain preference inconsistencies.

However, most concessions to the opposing perspective have been
grudgingly made. The possibility for more active collaboration is not
bright, due in some part to disciplinary intransigence. Each
discipline tends to frame real world phenomena through its own
conceptual lens, and with increasingly specialized tools and language.
Thus there seems less possibility of integration at the disciplinary-
theoretical level. However, as one moves to problems with empirical
content, out of the constructed laboratory problems and out of the
realm of hypothetical choices by naive subjects, there seems to be
much more fruitful ground for integration. Researchers working with

individual decision makers must help them to formulate problems from

39
what Ackoff (1974) calls ‘messes’, to focus and clarify their
preferences and expectations, to generate alternatives and to select
actions. Context, which has been shown to be critical in human
decision making, cannot be reflected in axiomatic constructs.
Therefore, in individual problem-oriented work, researchers seeking
development of their discipline may find problems and solutions which
do not correspond to the situation at hand. The complex and unique
nature of individual problems requires concepts and techniques from
many disciplinary perspectives, yet the very mechanism which drives
disciplinary development may inhibit needed interdisciplinary
research.

The distinction made between prescriptive and descriptive
research also inhibits development of processes to aid individual
decision makers. A prescriptive model indicates what action ought to
be taken only if the ‘small world’ of the model incorporates all the
lcomplexity of the ‘large world’ of the decision maker. Few models can
hope to capture more than a small part of the problem scope, the
causal relationships or the preference/belief structure of particular
individuals. we can hope that the optimal solution of the ‘small
world’ model is a good solution to the ‘large world’ problem, but if
the decision maker chooses not to follow such advice, the researcher
should be more willing to question the model formulation that to
impugn the intelligence of the decision maker. Such denigration of
intelligence in real decision making is fostered by the ‘prescriptive
research’ label.

At the level of individual problem solving, the behavioralist

perspective falters due to its tendency to see every situation as

40

unique, and its inability to categorize some decisions as better than
others. The rationalist perspective falters due to its tendency to
ignore context, to ignore human aspects of decision processes, and to
make sharp distinctions between a single ‘optimal’ decision and all
other ‘irrational’ decisions. A complete framework for integration of
the opposing perspectives at this level is beyond the scope of the
present research, but it seems possible to outline four areas where
working agreement on principles and coordination of research could
considerably increase our ability to improve decision making. They
are: l) the structure of preferences, 2) the nature of expectations,
3) the limits of logic in human thought processes, and 4) improvement

of decision making through feedback and learning.

2.7.1 Preferences

The two perspectives tend to agree that the concept of a fixed
preference structure is the only tractable framework for research on
individual decision making (for an opposing viewpoint, see March
1978). However, it should be realized that in many situations
preferences may be ill-defined and ambiguous, and that the decision
maker may identify or refine preferences during the decision process.
He should reflect the observed fact that decision makers desire multi-
faceted outcomes from their actions, and that not all these outcomes
are either easily measured or at all commensurable. Our models must
reflect the observation that aspiration levels are formulated to avoid
having to precisely define preferences, and that it is usually too
difficult to evaluate choices in terms of final outcomes instead of
gains and losses. And we must reflect that the most treasured

preference of an individual is the desire for control over his/her own

41

decisions. There is intrinsic value in the decision process beyond
the extrinsic value of outcomes. Therefore, the more control which
the decisionmaker has in the solution of the ‘small world’ problem,
the more likely that the corresponding action will be taken in the

‘large world’.

2.7.2 Probabilities

This issue provides the best opportunity for collaboration
between behavioralist and rationalist research. The logic of
mathematical probability provides both a common language and an
established set of operations with which we can express and evaluate
our expectations about future events. Other than the requirement of
coherence, however, the rationalist perspective considers probability
formulation as an unobservable primitive. That is, there is no a
priori basis for judging correctness of probability Judgements. Yet
probability assessments can be wildly inaccurate even though they are
precise expressions of belief. The most important factor which varies
between individual choices in some situations may be differences in
expectations rather than differences in preferences. Behavioralist
research into the cognitive formulation of beliefs and the probability
assessment errors to which humans are prone could aid the improvement
of expectations.

Nallsten and Budescu (1983) cite several examinations of
probability elicitation techniques. As noted above, training,
experience and feedback appear to be key elements in the ability of
certain individuals to give accurate probability estimates. It

appears that progress is being made to identify characteristics of

42
elicitation techniques which induce bias in probability encoding.
Hogarth (1980) begins the difficult task of developing operational
techniques to teach decision makers how to avoid bias in probability
Judgements. He suggests that progress depends upon an understanding
of the sources of biases and effective teaching of probability
principles. Probability training methods could be developed with
stress on those principles which are counter-intuitive. in many
situations, improvements may be possible by confronting individuals
with differences between their subjective probabilities and other

‘objective’ probability evidence.

2.7.3 The Limits of Logic

A third area for integration is the recognition that there are
bounds on the human ability to perform operations implied by
rationalist decision theory, and yet humans make very good decisions
in most circumstances. it should be recognized that '. . . human
judgement is more limited than flawed..' (Hammond 1975: 76). There
are significant limits to computational ability and short term memory
retrieval and storage. But beings as limited as portrayed in many
behavioralist experiments could not have built civilization and sent
people to the moon. Humans have the ability to make judgements which
are much too complex for computer models (imagine trying to program a
robot to cross a busy street!). Our decision strategies serve us
well, and we seem to recognize that in many cases too much rationality
is unreasonable (Harschak 1975). Nonetheless, there are many
situations in which our experiential or intuitive strategies fail us,
and logical processes offer the best opportunity to improve decisions.

Researchers and advisors should work with decision makers to identify

43

what are good but ‘irrational’ strategies, what are clearly identified
reasoning errors, how to identify situations which are likely to
stimulate these errors, and what reasoning mechanism can be
substituted at acceptable cost to the decision maker.

Rational choice researchers often attempt to replace the flawed
judgement of the human decision maker with the ‘objective’ judgement
of a dispassionate model. But only humans can attribute meaning and
purpose to actions and consequences, and only they will have to bear
responsibility for such actions. Therefore, we need to develop
mechanisms to aid humans in overcoming systematic errors while
complementing their strengths, such as integration of information of
varying types and quality, long term memory associations, abstract
reasoning, creativity and imagination. This decision support
perspective points out the need for a people-tool conceptual approach

to decision making.

2.7.4 Feedback and Learning

People turn to outsiders for help when they don’t understand the
‘mess’ that is bothering them. The ambiguity of choice situations can
mean that their preferences are not clearly defined, that the
structure and causal relationships of the problem are not understood,
or that they do not feel confident with their computational or
comparison ability. Sophisticated decision makers do not want ready-
made solutions (or at least don’t want jggt solutions), they want to
learn how to better deal with the ambiguity and complexity of choice.
Feedback can be distinguished by: 1) its potential to focus attention

on unconfounded cause and effect relationships: 2) its frequency and

44
or rapidity of results after actions are taken: and 3) the amount of
feedback from actions. When attention is focused on available
feedback, learning may occur which will benefit the quality of
decisions in future situations. Learning should be a central focus
for integration of rationalist and behavioralist research at the
individual level. Some decision models or aids provide a simulated
decision structure within which the decision maker can receive clearly
defined and rapid feedback from possible actions. This opportunity is
unique, because in the ‘large world’ we don’t necessarily know there

13 something to learn, or what is to be learned, or whether we have

 

learned anything (Einhorn and Hogarth 1981). Actions which would
probably not be attempted in the ‘large world’ can be tried with the
model and feedback evaluated. An awareness of the decision model
structure can lead to new insights of the real problem or re-
formulation of the model. The model can be seen as externalizing and
structuring the ‘thought trials’ mentioned by Hogarth (1981) as
crucial for feedback and learning.

Ackoff states that '. . . to learn is to increase one’s
efficiency or effectiveness over time under constant conditions.‘
(1974: xiii). Perhaps if we can improve our ability to design and
develop decision models and aids that enable individuals to learn
under the above definition, we can also improve our capacity to help

decision makers adapt under changing conditions.

CHAPTER III

DECISION AIDS

3.1 [ntrgguction

Decision aids are tools and techniques which extend the human
intellect and externalize some cognitive functions. They may be as
simple as a pencil and paper (to aid calculations or memory) or as
complex as multi-million dollar combinations of hardware and software
for managerial decision support. Decision aids may assist a single
function of decision making (such as a memory aid) or automate all
computational, memory and control functions of the decision process
(such as might be expected in a nuclear plant).

A set of decision tools may also be termed a decision support
model if its scope includes a significant portion of the decision
process. Aids to support programmable or fully structured decisions
relieve the individual of all decision making functions (except
possibly responsibility bearing). Other aids, such as probability
elicitation techniques, may play a part in a larger decision aid
system or function alone as a judgement aid for an otherwise unaided
decision process. The person-tool combination accomplishes the
functions of making a decision, and each has some degree of influence
on the outcome. The principal issue for aid designers is the
appropriate division of labor between man-made tool and human decision

maker (Pitz 1983).

45

46

Development of effective decision aid techniques has been an
important concern for centuries. The simple technique of listing pro
and con reasons for a decision was described by Benjamin Franklin in a
letter to Joseph Priestley in 1772 (Bigelow 1887). The potential of
decision tools has increased dramatically in recent years due to the
revolution in electronics and subsequent software development. Yet
the success of decision aids in improving decisions depends upon
complex and little understood interactions between the characteristics
of: 1) decision situations or tasks, 2) the individuals who use the
tools, and 3) the tools themselves.

This chapter describes the functions and purposes of decision
aids from the decision theory perspectives discussed in the previous
chapter. The critical characteristics and interactions of decision
task, decision makers and decision aids are examined, and the chapter

concludes with a discussion of evaluation issues for decision aids.

3.2 Aids for the Dggision Process

A decision aid is a set of tools and techniques designed to help
humans make decisions. This may entail support for elementary
information processing functions such as memory or computation, or
assistance for complex routines such as problem structuring,
generation of alternatives or choice of the best alternative. it can
be seen that an important (but not the only) purpose of decision aids
is to help the decision maker avoid information processing biases
described above. Some decision aids are developed to support personal
decisions, in which the individual will take primary or full
responsibility for the consequences. Other aids attempt to provide

objective support for decisions within an organizational context,

4?

where individuals must communicate and justify their decisions to
peers, superiors or outsiders.

Although decision aids are obviously designed to support the
decision maker to solve the problem at hand, they may have significant
secondary effects. By externalizing the cognitive decision process,
the decision maker may gain a greater understanding of personal
preferences, beliefs and causal relationships (Hammond 1975). This
may lead in turn to improved general problem solving ability in
related or dissimilar decisions. However, few aid designers have
attempted to tap this potential.

The development of a decision aid presupposes that individuals
have difficulty reaching ‘good’ decisions (according to some criteria)
and can be assisted to improve their decisions. Somewhat tritely, it
could be said that people use decision aids when they are sure that
they don’t know what or how to decide. People use decision aids when:

1. they feel unable to make a decision;

2. they want to better understand a decision they must make;

3. the problem is complex or uncertain; or

4. the stakes are high. (Jungermann 1980)

Simon (1965) conceptualizes phases of the decision making process
as identification, design and choice. Nith respect to the
identification phase, decision aids may perform certain intelligence
functions of problem finding. Davis (1985) describes these functions
as the ability to search internal and external data resources.
Decision aids for these functions are distinguished by the type of
search permitted, such as structured continuous search, structured ad

hoc search and unstructured search. Aids for problem finding include

48

the traditional report generators used to compare performance with
historical data, past projections and extra-organizational data. Aids
for the intelligence phase may also include problem structuring
functions, such as description and clarification of objectives.

Decision aids may also support the design phase by developing and
comparing alternatives. Davis (1985) proposes that support for this
phase should provide iterative procedures to assist the decision maker
to understand the problem and generate non-obvious alternatives.
Finally, aids for the choice phase of decision include the range of
computational and choice rules embodied in optimization procedures,
statistical inference, dominance analysis or even ad hoc rules.

Issues of rationality with respect to unaided decision making
have been examined above. The behavioralist evidence indicates that
in many contexts individual decisions fall short of that predicted by
rationalist principles, but many researchers argue that criticism of
unaided decisions begs a more important issue. Pondy contends that
'. . . what is needed is a theory of decision tools that improve the
performance of person-tool combinations, not merely a descriptive
theory of how humans unaided by tool extensions perform less than
optimally.’ (1982: 311) The focus of decision support is on improving
decisions with tools or extensions to human capabilities, rather than
the behavioral/rationalist issue of rationality. The primary issues
of decision research are the focus, techniques and application of
decision aids in real problem situations and real decision makers.
Rationalist and behavioralist researchers whose concern is to improve

decisions thus meet on the common ground of decision aid technology,

49

and must both deal with design, implementation and evaluation of
decision aids.

Decision aid models may be similar to any other research model
except that the principal purpose is to generate possible solutions to
problems faced by specific decision makers and thus must incorporate
to some extent the person’s participation. In development and
application of decision aids, both rationalist and behavioralist
researchers attempt to improve problem solving behavior. A somewhat
apocryphal examination of decision aids developed according to the
research orientation of these perspectives may help identify
significant differences in their purposes and techniques.

Rationalist decision aids presume, in general, that the decision
maker can unambiguously state personal preferences and expectations if
requested in the correct manner and that preferences and expectations
can be represented numerically. They do not generally assume that the
decision maker has all the relevant information about the problem
prior to use of the aid, but do assume that new information is
integrated with prior beliefs in a consistent fashion. Once
preferences and probabilities (expressed numerically as utilities and
probabilities) are elicited, the decision tool is often constructed so
as to calculate a unique optimal solution for a uni-dimensional
problem. In the uni-dimensional problem, the decision maker may
function something like a data input device to the analytical model.
If the individual accepts the rational decision axioms, accurately
verbalizes personal preferences and beliefs, and the ‘small world’
model sufficiently corresponds to the real world problem, then the

model solution gives the most preferred action for that individual.

50

If the decision maker does not accept the model solution, errors may
have occurred in model specification, measurement of preferences or of
beliefs; or the decision maker is making a choice in a manner
inconsistent with self-interest. Computational errors which might be
committed by the unaided individual are avoided by passing that
responsibility to the numerical model.

Behavioral decision aids are also developed to help individuals
avoid biases in important decisions. They do not necessarily assume
that a problem has been identified, but usually assume that an
underlying preference structure exists. The principal focus of a
behavioralist aid is usually assistance for the decision maker to
avoid the information processing errors identified in behavioralist
decision research. The aid thus places particular emphasis on
sequential display of limited amounts of data, retrieval of
information from long term memory, integration of information in short
term memory, explicit consideration of multiple consequences or
attributes and utilization of a series of decision rules. Behavioral
aids have benefitted from behavioral experiments demonstrating
weaknesses of individual probability assessments, and methods of
probability elicitation are chosen to minimize errors. Some
behavioral decision aids go beyond assistance for information
processing to reflect the effects of personality or cognitive style on
decision making. The contention is not only that all cognitive
capability is limited, but that individuals exhibit considerable
differences in their preferred problem solving style, and that better
decisions may result from the congruence of the representation

techniques of the aid and the decision style of the individual.

51

3.3 thragtgristigs 9f Qgcigiggs

There is no accepted, well-defined taxonomy of decision types.
Even within restricted subject domains, there have been few attempts
to categorize decisions. Generally, researchers have attempted to
delineate some of the characteristics which distinguish individual
problems. Castle et al. (1972) gives five major characteristics:

1. Importance of the decision

2. Frequency of recurrence of the decision

3. Time pressure when a decision is required

4. Revocability of the decision

5. Number of available alternatives

Decision aids are more likely to be used for important decisions,
in which the additional time and cost can be justified, although some
types of aids may be of_great benefit for frequently recurring
’smaller’ decisions. However, once the rules used by the aid are
learned by the decision maker, an aid is often no longer necessary.
Time pressure is an important characteristic of most decisions.
Jungermann notes, however, that '. . . time pressure, as perceived by
the decision maker, is a function not only of the actual or perceived
time available, but also of the complexity and significance of the
decision to be made.‘' (1980: 12) Decisions that must be made under
perceived time pressure are not as likely to stimulate use of decision
tools. Revocable decisions usually indicate less need for aids, since
decision makers can try a possible solution, observe its consequences
and decide whether to make a second choice. Similarly, the presence

of few alternatives may also reduce the usefulness of decision aids,

52
since the lower ‘cost’ of evaluating a few alternatives may induce
individuals to make unaided decisions. However, the presence of few
alternatives in a given decision situation may be more a result of
restrictions which the decision maker has placed on the scope of the
problem and the reduced effort invested in generating alternatives
than a characteristic of the decision type.

Other researchers have included further important characteristics
of decisions which affect the benefits of aids. Jungermann (1980)
distinguishes decisions in which the status quo is a viable
alternative from those in which a new action must be chosen. The
decision arises because new alternatives become available, or because
the status quo has not fulfilled current aspirations. Also critical
is the time horizon of decision consequences. Since more uncertainty
is associated with long time horizon decisions, aids may be more
fruitfully employed in decisions with long range effects.

Hogarth (1988) emphasizes characteristics of decisions which
combined with psychological factors may influence receptiveness to
decision aids:

1. Prgplem cgmplgxity may stimulate consideration of decision
aids. Large amounts of information or lack of knowledge of
solution techniques may affect use of decision aids.

2. Prgggggral gncggtainty implies that decision makers are not
sure of the procedures necessary to solve a problem. In the
presence of procedural uncertainty, individuals may be more

likely to utilize aids:

53

3. Psyghglggicgl regret: An individual is more likely to resort
to an aid when care in judgement is necessary to avoid
psychological ‘costs’ of regret from making mistakes. In
other words, a valid use of an aid is to corroborate
judgement.

4. gmgtignal stress: Problems which have important consequences
for which the individual must bear responsibility, or which
must be made under time pressure, or which are characterized
by substantial uncertainty are likely to be associated with
significant emotional stress. Decision tools may contribute
to lowering stress levels in such situations. Beach and
Mitchell (1978) group decision characteristics as inherent
in the decision or in the environment (context). Inherent
characteristics are unfamiliarity, ambiguity, complexity and
instability, while characteristics of the environment
include irreversability, significance, accountability and
time/resource constraints. If the decision maker perceives
these characteristics, the situation would be more likely to
initiate demand for decision aids.

In a managerial context, Simon and Hayes (1976: 121) distinguish
characteristics of ‘structured’ versus ‘unstructured’ problems. They
consider problems to be well structured or ‘programmable’ if: 1) the
conditions which define the existence and structure of the problem are
well known, 2) the procedures necessary to generate or identify
alternatives are feasible, and 3) unambiguous criteria exist for
choosing a best solution. Completely structured decisions can be

automated, and the decision maker’s judgement replaced by that of a

54
numerical algorithm. An example of such an aid might be a ration
formulation programming model. Less structured problems which are
nevertheless partially programmable are those in which individual
preferences are regarded as important, or characterized by features
such as novelty, time pressures, lack of knowledge, or non-
quantifiable factors. As more ill-structured problems are
encountered, the individual may find that the nature of the problem
itself, the information or procedures required, or the criteria for
selecting an action are not well understood. Several of the most
important managerial problems have these ill-structured
characteristics, such as significant expansion of the business, inter-
generational transfer or a decision to disband the business.

Sorry and Scott Norton (1971) combine the structured-unstructured
dimension of managerial decisions with the operational level at which
decision making takes place.

1. Operational performance decisions can be made while

performing the operation

2. Operational control decisions result from monitoring

effectiveness of operations

3. Hanagement control decisions relate to the acquisition and

efficient use of resources

4. Strategic planning decisions involve setting policies and

choosing objectives

Problems which are both well-structured and related to
operational performance are most likely to be automated, with humans
primarily performing an information input function. If problem

characteristics are more similar to the unstructured strategic

55
decisions such as product planning, decisions must rely more
extensively on judgement. A further extension of this framework is
presented by Gordon et al. (1975). Management decision types can be
classified by process (structured-unstructured), by level (strategic-
tactical-operational), by functional area (production, marketing,
finance, etc.), or by decision output (discrete choice, scale,
schedule, allocation, design or plan). Again, each factor can be seen

to influence the receptiveness to a decision aid.

3.4 Chgracteristicg of Decisign Makers

Extending well beyond human information processing limitations
discussed above, many psychology researchers consider how individual
personality, motivational and stress-related differences interact with
problem characteristics to affect decision making and the use of
decision tools. According to the cognitive style perspective, the
problem solving habits of the individual is consequence of
psychological characteristics. Cognitive style examines the approach
to decision making, not the person’s ability. Keen and Bronsema
(1981) contend that systematic differences between individuals
significantly influence their use of information and decision aids in
a manner quite distinct from the cognitive information processing
limitations.

Cognitive style refers to the process behavior that individuals

exhibit in the formulation or acquisition, analysis, and

interpretation of information or data of presumed value for

decision making. (Sage 1981: 64B)

Motivational characteristics such as ego-involvement may also

affect the utilization of decision aids (Jungermann 1980). Other

motivational features include whether the individual considers a

56
formal technique appropriate for the decision (who considers a
rational technique appropriate for choice of a spouse?) and why some
people are paralyzed when faced with decisions. Stressful decisions
affect problem solving behavior in ways which often result in
dysfunctional decisions (Janis and Mann 1977). Whether decision
makers under stress are more or less likely to utilize decision aids
is unclear.

Huber (1983) has criticized research emphasis on cognitive and
motivational types as a basis for decision tool or system design. He
contends that there is insufficient evidence to support cognitive
style research results as guidelines for designing decision tools, and
that the effort is both unwise and unlikely to bear fruit. Progress
in aiding individuals to make better decisions is, according to Huber,
more likely within the context of complementing human decision
strengths and in alleviating fairly well understood information
processing limitations.

Pitz (1983) examines three general strengths of humans over
automated procedures: 1) ability to rapidly encode, store and retrieve
complex patterns of information from memory: 2) extremely rapid
evaluation and classification of complex perceptual information; and
3) ability to make creative inferences in ill-structured decisions.
Even within narrowly defined subject areas, there is no mechanical
substitute for human ability in these functions. The automatic
integration of perceptual information with other information stored in
memory is extremely fast and usually quite accurate. Computerized
simulations of perception and integration are as yet limited to very

structured and simplified tasks (i.e. robot sensing on assembly

57
lines). The human ability to integrate perceptual data with
information from memory and to develop non-obvious insights is far
beyond the capacity of current computer models, and is likely to
remain so for the foreseeable future, despite progress in artificial
intelligence experiments in well-defined problem areas.

Decision tools should complement these human strengths, while at
the same time mechanically assist humans to transcend known
limitations of short term memory, inconsistency in judgement and
errors in learning. In addition, a good decision aid might help the
individual to retrieve information from long term memory, or to
recognize connections between previously unassociated items in ways
that are conducive to creative decision making. However, since
meaning, purpose and responsibility are essential ingredients in most
decisions, there will almost always be advantages to having human
perceptual skills and creative problem solving skills to evaluate
information, just as there are advantages in reliance on mechanical
aids for computational operations. The difficult task for decision
support designers is thus design of optimal person-tool combinations.

Integrating what he states are the most important characteristics
of decision makers with respect to the design of decision tools and
systems, Bennett (1983) makes four relevant observations:

1. Decision makers use conceptualizations such as diagrams or
graphs much more readily than written or numerical
information. Tools should utilize such representations.

2. Decision makers need short-term (and in some cases, long-
term) memory aids. Decision tools and systems should

incorporate more efficient and easily used memory aids.

58

3. Decision makers have different styles and skills. A
decision tool or system should not enforce a particular
style or skill level. The additional cost of flexibility,
Bennett points out, will be balanced by the benefits from
utilization by more decision makers.

4. Decision makers require personal control over decision tools
and systems. This does not necessarily imply personal
operations, but decision maker understanding of the support
process should be sufficient for the decision maker to

evaluate and direct operation.

3.5 r r' ' f De ' ' A'

Decision aids, like problems and decisions makers, can be
examined by appraisal of their characteristics. Tools are constructed
for different purposes, with different foci, and with different
techniques, each of which interacts with the problem and decision
maker characteristics to have considerable impact on the quality of
the decision. The implicit purpose of a decision tool is often very
difficult to assess. Often a decision ‘aid’ may be constructed to
accept input of probabilities and utilities and replace the cognitive
decision process of the individual with the logical process of a
computer model. Other aids attempt to explicitly complement the
aforementioned strengths of humans and provide support for known
limitations. These might be called the ’narrow’ and ’broad’
perspectives, respectively.

Mechanisms to complement decision making may consist of data
representations which stimulate non-obvious inferences or suggest new

alternatives, memory aids which allow consideration of many

59

information ‘chunks’ and which can be arranged in ways which may evoke
pattern or trend recognition, or output representations which
facilitate comparisons of multi-dimensional consequences.

Portions of the decision process which are externalized
(extracted from the unobservable cognitive process) may include
systematic assessment of preferences and beliefs, extensive data
computations, or application of complex decision rules. ’Narrow’
decision tools which attempt to externalize most of the decision
process run the risk of committing what Mittroff and Featheringham
(1974) call ‘Type III error’. That is, if the structure, decision
elements and operations imposed by the aid differ substantially from
the real world problem as perceived by the decision maker, the model
may generate a solution for the wrong real world problem. A subtle
but not very operational distinction might be made between decision
aids which facilitate better, consistent decisions and those which
achieve those results while fostering creative approaches to the
problem. The potential of a decision tool to stimulate new approaches
to the problem, to highlight problem structure and causal
relationships, and to thus create benefits which extend beyond the
current decision is an important (but neglected) area of research.

In terms of decision aid purpose, one could also distinguish
decision tools in terms of their consistency with rationalist decision
theory. Decision analysts can provide aids with different types of
decision rules, assumptions about problem structure, and data
requirements. The ‘best’ aid is useless if it requires rules, problem
structure, or data that do not exist or cannot be effectively

generated. A ‘good enough’ decision aid uses unaided judgements as a

60

benchmark in generating solutions which should be on the average
better than unaided decisions (Keen 1977). Other aids might stimulate
‘better’ decisions, but have characteristics which are not
complementary to characteristics of the problem or of the decision
maker. A ‘good enough’ aid might then be considered as an automated
rule of thumb or an efficient approximation to a more complex model.

Decision tools can also be examined according to their focus.
Jungermann (1980) classifies existing aids for personal decision into
three general categories: 1) aids for reaching a decision: 2) aids for
sticking to a decision; and 3) aids for improving general problem
solving behavior. General knowledge of decision skills could improve
individuals’ sense of control over their lives and their confidence in
decisions, but only a few aids have been developed which specifically
attempt to improve decision making ability. Instruction in
rationalist decision procedures and rules is the usual focus of this
type of aid. Unsystematic application and evaluation has restricted
ability to draw conclusions regarding effectiveness in improving
subsequent decisions. Jungermann notes that '. . . it might be more
effective to teach people, as early in their lives as possible,
tolerance of uncertainty and ambiguity, cognitive flexibility and
avoidance of biases that influence judgment or hinder learning from
experience.‘I (1988: 22)

Another focus of decision aids is the volitional problem of
following decision (a behavioral intention) with action (actual
behavior). Most literature on problem solving ignores the distinction
between the principally intellectual exercise of making a decision and

the sometimes emotionally charged atmosphere surrounding actions to

61

carry out the decision. Lack of confidence and anticipations of
regret are important emotional factors which undermine decisions
reached through logical processes, but logical principles are not
particularly useful in dealing with such feelings. Research on
behavior modification and cognitive therapy (Fischoff 1983) might be
of some use in designing aids which foster confidence and show how
anticipated regret can affect the quality of a decision.

The principal focus of existing decision aids in managerial
contexts is usually the static framework of problem structuring and
solving. Within this framework, decision aids may concentrate on:

1) problem structure, 2) assessment of probabilities, 3) assessment
of preferences, or 4) generation of one or more solutions. Any
particular aid may exhibit a mixture of these foci. The principal
objective of problem structuring aids is to '. . . provide some better
understanding of the interrelationships among elements of the

problem . . .' (Pitz 1983: 210), in particular the decision maker’s
perspective of which problem characteristics, alternatives, future
events, and consequences should be taken into consideration.
Examination of resources (including data resources) can also be viewed
as an aid to problem structuring. Few decision aids integrate problem
structuring with problem solving tools.

Jungermann (1980) states that focussing only on structure might
be appropriate when better understanding of structure provides a
sufficient basis for decision makers to make an otherwise unaided
decision, or when the individual is particularly receptive to an
analyst’s suggested solution. It may also be a good focus when

rational procedures are not seen as appropriate by the decision maker.

62

Pitz states that this '. . . may be the most critical stage in the
decision analysis [and is) . . . far more important than the

small amount of research devoted to the topic might imply.‘ (1983:
210) There are several reasons why more management research effort
has not been carried out in this area:

1. Such an aid is more a behavioral analysis procedure than a
computerized decision aid. It is unclear how the advantages
of computers (rapid computation, mass data storage, etc.)
can be fruitfully used in this area. Thus, such an aid
would probably suggest analyst-decision maker interaction,
with the computer as a facilitating memory or graphical aid.

2. The amount of analyst input is usually too great for all but
the most important decisions. Also, if the ‘problem’ is to
structure a decision situation, the range of skills required
of the analyst is consequently broadened.

3. It is unclear how the benefits of a problem structuring aid
could be evaluated.

The second concentration area of problem structuring and solving
aids is evaluation of consequences or preference assessment. This
must, of course, be kept quite separate from the predictive Judgements
of probability assessments. In single attribute decision making,
methods are well developed for eliciting measurement of von Neumann-
Morgenstern utility functions (Hershey 1982), although few attempts
have been made to automate the process. The individual is not
required to comprehend the measurement process. Decisions with
multiple consequences of importance to the decision maker may require

multi-dimensional utility assessment. The actual measurement

63

procedures in these cases are much the same as uni-dimensional
procedures (Keeney and Raiffa 1976), but also usually involve
assessment of relative weights of importance for the various
attributes, scaling of attributes and tests to validate separability
conditions of the multi-attribute utility functions. The time
required for verification of the latter conditions tends to make MAU
approaches unmanageable as decision aids in all but the most important
decisions.

Aids for assessment of probabilities are likely to be a major
area of development in the coming decade. A wide variety of encoding
techniques have been developed, and efforts have been made to devise
reliable testing procedures to ensure that bias is not introduced by
the encoding method itself (Nallsten and Budescu 1983). Unlike
utility assessment, much stricter conditions for internal consistency
can be placed on subjective probability statements. Also, behavioral
decision research (as discussed above) has identified common errors in
subjective probability estimation which aids should help individuals
to avoid. In some situations, the external validity of probability
estimates (correspondence with historical or known frequencies) can be
determined and discrepancies displayed to the decision maker.
Estimates may be improved by providing relevant external probability
data before assessing subjective probabilities, provided that care is
taken to avoid reported biases in Bayesian processes. 'Graphic
computer representations may significantly improve the individual’s
ability to integrate probabilistic information and state probability

estimates in accord with personal beliefs.

64

Evaluation of alternatives is another focus of problem solving
aids, particularly decision-theoretic aids. Preconditions include
that problem structure has been well defined, both preferences and
probabilities have been assessed, a suitable decision rule chosen, and
alternatives have been identified (either implicitly or explicitly).
Then a computerized aid can be used to compare alternatives and select
a set of preferred actions or a unique best solution. The power of an
automated procedure is most evident in aids with this focus, as the
search, memory and computational capability of computers are well
suited to these actions. It should be noted, however, that the above
prior problem conditions require that the individual’s problem
conception is veridically represented by the analytical model. Even
if these conditions exist, however, a danger exists that the decision
maker may reject the model solutions because cognitive control of
decision making has been abdicated to the analytical model. The
cognitive control element may be critical to the success of aids with
this focus.

Decision tools can also be characterized by the techniques
utilized. The principal characteristics (not mutually exclusive)
which distinguish aids are:

1. degree of quantification;

2. context-flexibility:

3. representation of problem structure;

4. type of analytical model (optimization/simulation or

deterministic/stochastic);

5. numerical/graphical input or output;

65

6. interactive or batch processing; or

7. operation by the decision maker or by an intermediary.

Complete quantification of monetary and non-monetary values is
usually necessary for analytical decision aids. The ‘fuzziness’ of
human language seems to serve decision makers well in normal
instances, and it is clear that added precision may be traded off
against significance in decision modeling. In some decision aids
dealing wholly with decision structure, quantification is kept to a
minimum as the focus is on providing clarity of relationships between
problem elements. Fuzzy set theory, in which transition from
membership to non-membership of objects in sets is gradual rather than
abrupt, can represent the ambiguity of natural language. Its
application to decision support has been proposed by Bellman and Zadeh
(1978) as an alternative to the precise definition of values and
probabilities, but no implementations have as yet been reported.

Techniques can also be characterized by the flexibility of
application to a range of problem environments. Although it is
normally infeasible to develop a unique decision aid for each
situation, an aid may promise sufficient payoff to warrant such
development. Certainly a decision tool for a major corporation would
have to be made highly specific to the corporate environment and
probably to the specific problem. As mentioned above, individuals
have different educational characteristics and problem solving style
which affect the effectiveness of aids (Keen and Scott Norton 1978).
Development of a single standardized model or aid for a wide range of

decision makers is certainly cost-efficient for the designer, but

66

standard models may be inappropriate to any specific problem and thus
may be considered unsatisfactory by decision makers.

If one objective of an aid is to assist the individual to better
understand problem structure, the techniques used to represent
analytical model structure are of considerable importance. Such a
representation may be as simple as a decision tree, or as complex as
CPH charts or influence diagrams (Bodily 1985). Even though problem
solution rather than structuring may be the principal objective of the
aid, suitable representation of structure may increase the decision
maker’s confidence in model solutions, or may initiate an interactive
process of model reformulation. whether such reformulation is
feasible depends on the flexibility of the aid.

Techniques for choice of decision rule are not currently
incorporated in decision aids, although concepts proposed by some
decision analysts include flexible, user-controlled application of a
range of decision rules to specific problems (Keen and Scott Norton
1978). Normally, rules for selection of a subset of actions (or a
single alternative) from the set of feasible alternatives are pre-
defined by the tool developer. Simulation models, as opposed to
optimization models, do not impose decision rules and allow
individuals to select alternatives according to personal decision
rules.

The deterministic or probabilistic nature of the analytical
model base of the aid is another distinguishing technique. Of course,
the problem itself usually is affected by probabilistic factors to
some extent, but the tool designer chooses to represent the problem

with or without probabilistic elements. A decision maker with minimal

67

training in probability theory might feel more confident with a
deterministic model if this more closely accords with his/her
perception of the problem, but it is seldom obvious whether such a
solution is better or worse than that produced by a more complex (but
possibly less understood) probabilistic model. Application of
different decision rules in various stages of the decision process is
proposed by Keen and Scott Norton (1978). Simple strategies such as
conjunctive rules may first be used to reduce the set of alternatives,
followed by a detailed optimization procedure applied to the reduced
set. Application of heuristic rules is probably most justified in
problems where structure, elements and causal relationships are less
clearly defined.

Techniques for input and output representations distinguish
decision aids, and the type of representation interacts with the
individual’s decision style to affect usefulness and learning from the
aid. Spatial representations such as graphical output tap the
perceptual resources of humans and greatly increase information
integration. As Davis notes, 'A graph is a ‘chunk’, yet it may
provide the same input of data as a large number of data items that
would each use one chunk of (cognitive) capacity.‘ (1985 : 246) The
interactive technology necessary to generate high quality, flexible
and fast graphical output in a field environment is now becoming
available. Techniques for model input through non-traditional media
have also been developed. Digitizers, touch screens and voice
synthesizers have the possibility to facilitate user input (Johnson

and Loucks i980).

68

Decision aids may be operated in either interactive or batch
mode. With a batch process, the complete input for model computation
is collected and later processed by the computer (although ‘later’ may
only be a few minutes). Presentation of model output is then made to
the decision maker at a later time. Interactive operation, on the
other hand, entails input and output between the computer model and
its human operators throughout at least some portion of the solution
process. With microcomputer processing of decision aids, the above
distinction becomes less obvious than with centralized computer
processing, both in terms of time necessary for solution and user
dialogue with the model. A further technique which has become
possible with faster microcomputer technology is iterative solution of
decision models. If operation of the model is sufficiently rapid, the
output from one complete cycle of the model can be used as information
input to the decision maker, who may revise his expectations, modify
the problem or experiment with different goals and solve the model
with these modifications. This input/output iterative process may
continue as long as the decision maker wishes to explore different
formulations and solutions.

Finally, decision aids differ according to the principal operator
of the computer model. If the decision maker is well associated with
the hardware and software operation of the aid, the aid may be used
without further assistance. It is generally felt that decision makers
will reap more benefits from aids which they can operate without
analyst mediation. However, formulation of the analytical problem and

operation of the model often must be mediated by a trained analyst.

69

This increases the costs of tool utilization, but may permit a broader

learning experience than with solo operation.

3.6 Qggggggenggs 9f Qgcisign Aigs

Although computerized decision tools have been developed and
disseminated for more than a decade, there is a dearth of evidence
examining impacts of aid utilization on management decision behavior.
One problem is the non-experimental environment of real business
decisions. It is difficult to derive experimental controls which
allow determination of decision tool impacts. Also, most managers are
not (for obvious reasons) willing to devote enough time to carefully
controlled experiments and much prefer their own intuitive evaluation
of the tool.

A further problem relates to the intellectual perspectives of the
previous chapter. From the behavioral perspective, the purpose of a
decision aid is somewhat ambiguous. If the principal purpose of
behavioral research is development of models to describe actual
decisions (what Hobbs I985 calls ‘imitative validity’), there is no
reasonable basis for development of decision aids. Standards for
‘better’ decisions or decision procedures do not necessarily exist.
Many behavioral researchers (including some of those who have
contributed most to our knowledge of judgemental biases) contend that
consistent choice in similar contexts is a minimal basis for improving
overall quality of decisions, and that a decision aid should provide
both a structure and procedures to encourage such consistency.
Further, if consistency is extended to include consistency with the

rules of logic and probability, the rationalist perspective is

70
reached. At the extreme of this perspective, researchers contend
implicitly that aids prescribe ‘best’ solutions which decision makers
should accept if they are ‘rational’. Evaluation of such aids may be
primarily a priori determination of logical coherence of the
underlying analytical model.

A more pragmatic approach accepts that each perspective has some
validity. Certainly decision tools should be compatible in some sense
with existing decision procedures. They should establish structure
and procedures which encourage consistent decisions for similar
settings, and they should be internally consistent with rules of logic
and probability. The principal dispute is an issue of emphasis rather
than substance. Yet all involved should recognize that the principal
purpose of a decision tool is to assist real decision makers and
suggest ‘better’ decisions, however defined.

Two orientations to evaluation of decision tools can be discerned
from these perspectives. An ogtsgmg-orignted approach focuses on the
consequences of the decision made using the aid. Three general
evaluation issues can be identified:

1. Are solutions logical? An a priori evaluation of the aid’s
consistency with logical operations based on its assumptions
can be carried out.

2. was the model solution a reasonable problem solution? Are
solutions within the range of observed real world actions?
Do decision makers think the solutions are reasonable? How
do the model solutions compare with those of alternative
models? Essentially, one asks whether the problem has been

modeled in a realistic (veridical) fashion. Both the

71

researcher and the user will have points of view on the
adequacy of the model.

3. Has the solution implemented? Here the consequences
sufficient to justify use of the aid? An ex post evaluation
of the decision outcome can also be carried out, although
the often lengthy period between decision, action and
consequences and the effects of uncontrollable or
unforeseeable intervening factors usually makes experimental
verification of the aid’s value extremely difficult.

A second evaluation orientation is process-grientgg. This type
of evaluation attempts to assess the consequences of the decision tool
with respect to changes in the decision process, regardless of the
actual solution or its realized consequences. Some justification for
the aid, it is argued, can be determined by evaluating whether the
decision is made correctly, apart from whether the correct decision is
made. However, evidence for improved decisions can only be indirect.
Three aproaches can be identified:

I. Before/after changes in procedures with respect to the
phases of decision processes can be evaluated. Information
search, generation of alternatives, computations, and
assessments of preferences or of expectations are all
factors which would indicate whether the tool has affected
the decision process.

2. Comparison of the procedures of the aid with those of
alternative aids. Both this approach and the previous
approach can include evaluation in terms of time and

resources. This may include a range of factors from the

72

difference in time spent on the decision to cost-
effectiveness evaluation.

3. The manager’s conception of the decision process is also a
critical factor influencing adoption and utilization of the
aid. This includes several factors. The decision maker’s
subjective evaluation of the aid’s value is the overall
determining variable. Several other elements can influence
that evaluation. The manager’s confidence in the model
solution, consideration (or understanding) of the
appropriateness and complexity of the analytical
formulation, and decision benefits from alternative
procedures will reflect on his/her determination of the
benefits. Evaluation of the aid’s ease of use will indicate
estimates of time, skill and resource costs necessary for
the procedure.

Finally, a process-oriented evaluation should investigate changes
in the decision process with respect to understanding and learning
about the decision problem specifically and problem solving methods in
general. Here both theory and methods fail in evaluation.
Considerable evidence exists that a major benefit of decision aids is
the improvement in understanding gained by examining the problem in
the perspective of the decision model (Humphreys and McFadden 1980).

A good decision aid can create considerable benefits to the decision
maker through learning effects by highlighting problem structure,
causal relationships, preferences, expectations, decision procedures,
resource constraints, and even simply by displaying formerly

disaggregated data in exploratory aggregate formulations. Yet little

73

theory guides the researcher to design decision tools to facilitate
such learning while solving complex decision problems (Keen and Scott
Norton 1978). The ‘silver lining’ approach to the issue is to regard
any such benefit as coincidental, but to utilize this in facilitating
adoption of the technique. The ‘bitter pill’ approach regards the
logical decision procedure as difficult but necessary for individuals
to master. Given time, decision makers will ‘learn’ that the value of
correctly made decisions outweighs the cost of learning complex
skills.

Evaluation of learning effects is particularly complicated by the
absence of a clear definition of learning. Even from the
behavioralist perspective, there are multiple interpretations of the
concept. The simplest definition is derived from the operant
conditioning research tradition of psychology. Learning is operant
conditioning, or in other terms the conditioning of a response to a
particular stimulus, rather like a Pavlovian dog. With the demise of
this research approach, it was recognized that human learning involved
motivational and memory factors in combination with complex patterns
of stimuli. Until the last decade, human learning research was still
limited primarily to memory and motor skill or simple problem solving
performance in laboratory experiments (Langley and Simon 1981), and
succeeded to some extent in distinguishing characteristics of ‘expert’
performance. The resulting research area of artificial intelligence
has focused on expert performance with suprisingly little attention to
the learning processes either of the experts or of ordinary humans
(Simon 1981). It is widely accepted, however, that learning involves

at least three somewhat distinct processes (Rumelhart 1981):

74

I) accretion, or the process of accumulating knowledge (facts,
beliefs) in memory: 2) restgugturing, the process whereby whole new
knowledge structures and procedures are created; and 3) tggigg, a
process involving modification of existing structures and procedures.
How these principles can be used to form a basis for experiential
learning in complex problems has received little attention, primarily
due to the lack of methodological framework. Langley and Simon (1981)
hypothesize that although researchers are unsure how cognitive
processes are affected in learning, certain principles define the
conditions within which learning can take place.

I. Kggwledgg gf rggult : Change in performance must be
detectable

2. Qggeratign gf altggngtivgg: The individual must be able to
attempt alternative behaviors

3. Qaggal attriggtion: Results must be attributable to specific
components of the decision environment

4. Hindsight: Past performance must be re-evaluated in terms of
subsequent results and causal attributions

5. ngrnigg frgm iggtructign: The quality and content of
examples, decision rules, and causal relationships suggested
by an instructor affects both what is learned and how
rapidly learning takes place

6. Agtgmatizatiog: Continued practice causes improvement in
speed and accuracy of performance

The rationalist perspective of learning is regretably quite

barren. This is attributable principally to the static nature of

rationalist decision models and the lack of interest in descriptive

75

modeling of decision processes. A prominent decision-theoretic text
in agriculture (Anderson et al. 1977) mentions learning only twice, in
bibliographic references to changing subjective probability
distributions by Bayesian probability revision.

‘Learning’ from a rationalist decision theory perspective thus
indicates incorporation of new information into expectations expressed
as probabilities. Rationalist decision theory has no clear
perspective on formulation or reformulation of knowledge about
procedures for arriving at decisions. The only recognized procedural
knowledge comes from the rules of mathematical logic, and the research
carried out to investigate ‘learning curves’ for adoption rates of
innovations generally does not extend beyond the correlational
analysis of static factors associated with adoption data. Procedural
‘learning’ is assumed to be the adoption of formal decision procedures
by the decision maker as he/she realizes that such procedures generate
results most likely to satisfy decision objectives.

What is learned and how is such learning facilitated by
utilization of a decision aid? Generalizable answers to these
questions are scarce. Certainly much depends upon interactions
between characteristics of the decision problem, the decision maker
and the tool, but how these characteristics interact to stimulate (or
inhibit) learning is unclear. Perhaps the best approach to evaluation
of learning within the constraints of current knowledge is to examine
the most common claims for possible learning effects of decision aids.

i. Pr l tru r : The decision maker may re-formulate

his/her perception of this and similar problems by exposure

to the parsimonious analytical model of the decision aid.

76

The relative importance of objectives, the critical nature
of certain constraints, and other causal relationships
between elements of the problem may stimulate a better
quality decision in the present instance or a broader
knowledge base for similar or repeated future decisions.
Finally, knowledge of the model structure serves as an
evaluative mechanism for the decision maker to determine
whether such a structure is considered appropriate for this
type of decision.

Prgglgg ‘gnfglding’: Through utilization of the decision
aid, the individual may learn general problemesolving
techniques such as decomposing complex problems into smaller
ones more amenable to analysis. The critical decomposition
between what is desired (preferences or utilities) and what
is believed (beliefs or probabilities) forms the basis for
more consistent, reasonable decisions. Similarly, the aid
may facilitate learning the implications of different
decision assumptions and of conflicting decision objectives.
Intggggtive grogessing and/or simulation: Interactive
processing is felt to greatly facilitate learning of problem
structure and decomposition through relatively immediate
communication between aid and user. It also facilitates
simulation-based aids to stimulate learning of action-
outcome relationships by allowing the user to define and
test numerous strategies in a real-time environment.
figaghigal ggtgut: Two related aspects are claimed to affect

learning through graphical representation of output. First,

77

graphs summarize data in a spatial perspective. Learning
may be facilitated by decreasing information processing
requirements through ‘chunking’ (Newell and Rosenbloom
I981), thus allowing limited attention resources to be
focused on learning model structure and causal
relationships. Second, graphs may be remembered better than
tables, allowing more accurate comparison of model
solutions. Because graphs can also be comprehended faster
than tabular data, the time necessary to make decisions may
decrease in repeated decisions.

Since learning necessarily refers to changes in knowledge
structures, procedures and cognitive strategies over time, evidence
supporting the above claims can only be obtained through long term
testing and evaluation of repeated utilization of decision aids. The
short term learning effects of changes in decision makers’ conception
of the problem, knowledge of causal relationships and confidence in
their ability to make better decisions are helpful, but not
conclusive, evidence of learning applicable to future decisions.
Nevertheless, even the short term learning effects of aided decision
making have not been examined in either laboratory or field
environments (Keen and Scott Norton 1978). Methodological problems
are the principal cause of this dearth of research. It is unclear
what is learned, how it is learned, and what are the relative costs
and benefits of different aspects of learning through decision aids.
Whatever evidence that can shed light on these issues is likely to

bring us closer to Pondy’s ideal of a theory of decision tools.

CHAPTER IU

MULTIPLE CRITERIA DECISION TECHNIQUES

4.1 Intrgguction

Behavioral decision research has emphasized the importance of
such factors as aspiration levels, sequential attention to multiple
goals and action/feedback as functional mechanisms to compensate for
information processing limitations in decision making. Rationalist
research emphasizes consistency with respect to decision postulates.
It has been suggested above that the design of decision aids may
benefit from a synthesis of the two perspectives on decision making.
This chapter will investigate a class of techniques which offers
promise for development of aids.

In many decisions, individuals weigh alternatives along multiple
dimensions or seek multiple objectives from a single decision or
choice. Often a decision aid and underlying model may be constructed
as a one dimensional problem to take advantage of powerful
optimization techniques without substantially misrepresenting the
decision from the perspective of the decision maker. However,
explicit consideration and modeling of multiple criteria perceived by
the decision maker as important to problem structure, analysis and
solutions may provide potential for greater benefits from an aid than
would be the case with a uni-dimensional model. This might be the
case if the decision maker considers the situation to be characterized

by multiple criteria and may respond negatively to less realistic

78

79
formulations, or if the problem requires clarification of the decision
maker’s objectives and tradeoffs between objectives. Multiple
criteria techniques, developed principally by operations researchers,
offer the power of optimization procedures with the flexibility to
handle such multiple criteria problems.

This chapter first briefly discusses research on multiple
objectives of farm operators within the agricultural economics
literature. Then the wide range of multi-criteria computational
procedures is discussed, with particular attention to multiple
objective programming techniques and interactive approaches which
might be operational within a farm management context. The
Interactive Multiple Goal Programming (IMGP) technique developed by
Nijkamp and Spronk (I981) receives particular scrutiny for its
relatively flexible data requirements, interaction with the decision
maker, adaptability to group decision processes, and potential for
decision maker learning with respect to the impacts of conflicting

objectives.

4.2 Mgltiglg ijectivgg in Agricultural Rgsearch

Although neo-classical economic theory is primarily based on the
assumption of uni-dimensional preference, management researchers in
agriculture have always recognized the importance of multiple
objectives to farm managers. Many objectives are poorly formulated by
the manager, are conflictive and/or incommensurable and are of varying
importance in any particular decision. However abstract and poorly
articulated, multiple objectives have been shown to be important to
farm managers. In fact, the suspicion has been voiced by some farm

management researchers that at least in part '. . . differences in the

88

financial performance of farm firms (growth, profitability, leverage,
liquidity) may be attributed to differences in the composition,
ordering and weighting of farmers’ goals, rather than to shortcomings
in management ability or to attitudes towards risk.“ (Robison et al.
1984: 28)

Smith and Martin (1972) utilized factor analysis techniques to
identify significant goals of Arizona ranchers. Although no attempt
is made to determine the comparative importance of these objectives in
business decisions, it is shown that social ties to the community
along with monetary goals emerge as significant objectives for
managers. Gasson (1971) showed that British farmers are strongly
motivated by objectives intrinsic to farm work rather than by economic
goals. Also, managers were seen to re-order goal priorities depending
upon the type of decision under consideration.

Studies by Cary and Holmes (1982), Hatch et al. (1974), Smith and
Capstick (1976), Harper and Eastman (1988), and others elicited farm
managers’ relative rankings of goals such as profit, consumption,
risk, credit borrowing, leisure and esteem. Although relative
rankings vary widely between individuals, most of these studies
indicated risk and security goals as most important, followed by firm
growth, living standard and income objectives. For example, Fernandez
(1982) used conjoint analysis to rank order combinations of
objectives, and found 56 percent of surveyed farm managers to consider
risk, defined as the probability of bankruptcy, as the most important
factor motivating decisions, while income and leisure objectives were

considered most important by 36 percent and 8 percent of managers.

81

Some researchers have attempted to model management decisions on
the basis of multiple objectives of representative firms. Problems
arise in determining the level of abstraction necessary to adequately
operationalize multiple objectives. Some objectives are hierarchical
(Georgescu-Roegen 1954), with no tradeoffs between levels of the
hierarchy, while others are directly competitive and can be traded off
against one another. The most appropriate computational technique for
such multi-criteria analysis is not well understood in agricultural
research. Wheeler and Russell(1977) applied goal programming<GP) to
the decision problem of planning a mixed crop and livestock farm.
They propose that farmers’ goals can be divided into three classes:

I. Non-quantifiable goals, such as.a desire to be considered a
top farmer.

2. Minimum achievement goals, such as remaining in farming.
The exact achievement level is not relevant.

3. Ouantifiable goals, which may be interrelated or
incompatible. Goals relevant to farm planning include
income goals, working capital goals, labor cost goals and
labor utilization goals.

They construct an illustrative goal programming model, assuming
three goals: 1) maximize gross margin, 2) minimize incidence of
negative cash flow, and 3) stabilize labor utilization through the
year. The effect of different possible goal priority weights on
solution activities generated by the model is analyzed and discussed.

Barnett et al. (1982) utilized multi-dimensional scaling to
estimate goal weights of Senegalese farmers for the assumed goals of:

I) producing sufficient food in bad years, 2) reducing cash

82

expenditures, 3) improving income, 4) achieving more leisure time, and
5) obtaining higher yields. A goal programming model was formulated
and used to estimate crop mix, net income, acreage cultivated and
credit usage. Although resulting solutions were similar to observed
activities in the area, the model did not outperform a similar profit
maximization model. No attempt was made to evaluate the model on any
basis other than its ability to imitate observed resource allocation.

Romero and Rehman (I984, 1985) describe various multiple criteria
techniques, such as lexicographic and weighted goal programming,
vector maximization and compromise programming. They illustrate
application of some techniques to an orchard planning exercise. Of
considerable interest is discussion of the similarity between risk
programming techniques such as quadratic or Mean Absolute Deviation
(MOTAD) programming and a certain class of multiple objective
procedures. These procedures use the constraint method, in which each
objective of a multi-criteria problem is sequentially optimized while
all others are parametrically varied over the relevant range. Thus,
all non-dominated solutions of the problem are generated. Romero and
Rehman state that both Expected Value-variance (E-U) and Expected
Value-Mean Absolute Deviation (E-M) '. . . can be legitimately taken
as a MOP [Multi-objective] model with two objectives.‘ (I985: 183)
They further describe the Target MOTAD programming technique developed
by Tauer (1983) as a hybrid between multi-objective and goal
programming, using the constraint method to generate solutions with
the risk objective measured as deviations from a goal (target) value.

None of the above studies have attempted to operationalize a

multi-criteria model as a decision aid for farm managers, although

83
each suggests how such an approach might be implemented. Selection of
a multi-criteria technique for aiding farm management decisions
requires careful consideration of technique characteristics and
requirements with respect to what is known about unaided decision
making and about characteristics of successful decision aids in other
contexts. Certain techniques will probably have advantages for
particular types of decisions, but what is sought here is a flexible
technique which can be used in many different contexts. The following
sections examine the various multi-criteria techniques available with
respect to their suitability as a computational model for a decision

aid.

4.3 Mgliiglg Criteria Decision Making: History and Classificaiion

It is the contention of multiple criteria decision researchers
that most economic decisions are made by decision makers on the basis
of multiple, conflicting objectives, and that the objective of
decision analysis should be to help decision maker(s) structure the
decision problem in such a way as to clarify preferences and to allow
the decision maker(s) the final judgement as to which action best
satisfies preferences. They criticize ’traditional’ single criterion
decision research as being in fact a methodology of measurement.
Elicitation and measurement, followed by computerized search
procedures '. . . become the substitute for decision making . . . The
decision is implicit in the measurement.‘I (Starr and Zeleny 197?: 25)
However, Rosenthal (1985) cautions researchers not to overstate their
case and dismiss single objective formulations so lightly. A

combination of carefully designed single objective model, multiple

84
reformulations and intelligent post-optimality analysis may be at
least as beneficial as a multiobjective formulation.

Decision theorists’ concern for what has sometimes been referred
to as the ’curse of multi-dimensionality’ was first expressed by von
Neumann and Morgenstern (1944). Writing with respect to the social
exchange economy, they expressed that such an optimization problem was
not a maximization problem, but a ’peculiar and disconcerting’ mixture
of maximization problems. That kind of problem, they state, is not
treated in classical mathematics. Koopmans (1951) first defined the
concept of an efficient vector, the vector of solution values for
multiple objectives which is undominated by any other feasible vector
of solutions. Kuhn and Tucker (1951) introduced the mathematical
programming model of vector maximization and derived necessary
conditions for optimal solutions.

Certain key concepts should be defined before classification and
analysis of multiple criteria decision research. These terms may not
be used in multiple criteria decision making (MCDM) research in quite
the same manner as used in economics research.

I) An attribute is a descriptor of objects or processes in the
real world as interpreted by each individual. The number of
possible attributes of any object or process is infinite,
but only a subset of all perceivable attributes are relevant
to a particular decision situation. The purpose of
attribute measurement is the unambiguous definition of
attributes and attribute magnitudes. Example of attributes

are height, heat and beauty.

2)

3)

4)

85
An objective is functionally related to the attributes of an
object. It indicates both preference and a preferred
direction of change. Examples are maximization of profits
or minimization of costs.
A gggi is an objective with a specific desired level. Thus,
reaching the moon by 1978 was a goal, target or aspiration
level. The terms ’goals’ and ’objectives’ tend to be used
interchangeably in much of the literature, although in many
instances a goal could as easily be considered a constraint.
Qritgrig can be either attributes or objectives salient to a
given decision. When problems involve mostly attributes,
research is said to involve the theory of choice, while the
cases dealing mostly with objectives are referred to as
involving the theory of decision making. Multiple Attribute
Decision Making (MADM) research usually involves selection
from a limited set of explicit alternatives, while Multiple
Objective Decision Making (MODM) concerns itself with the

design and selection of implicit alternatives.

A significant advance in multiple criteria decision making

occurred with the development of goal programming by Charnes and
Cooper (1961). Major contributors to this line of multiple criteria
methods are Ijiri (1965), Lee (1972) and Ignizio (1976). Following a
very different tact, researchers such as Yntema and Torgerson (1961)
laid the foundations of modern multiattribute decision theory with
research on necessary conditions for utility function decomposition.
Particularly noteworthy in the development of Multiattribute Utility

Theory (MAUT) are Keeney and Raiffa (1976) and Fishburn (1965).

86

Linear multiobjective programming techniques, meanwhile, have
developed through the works of Geoffrion (1968), Evans and Steuer
(1973) and Zeleny (1982). An annual international conference on
multiple criteria decision making began in 1972, encompassing all
areas of multiobjective research for decision making. Later, journals
such as Managemeni Scignce and Computgrs and Operations Research
devoted special issues to multiple criteria developments, and several
volumes exclusively dedicated to multiple criteria problems have been
published. MCDM is truly a multi-disciplinary research area, with
major contributions from operations research, management science,
psychology, finance, resource development, and systems science.

The three major branches of MCDM decision research are Multiple
Objective Programming, Goal Programming and Multiattribute Utility
approaches. Each has numerous variations which are reviewed in Hwang
and Masud (1979), Goicoechea et al. (1982), Cohon (1975) and Chankong
et al. (1985). Several authors have attempted to identify the
critical characteristics which distinguish multiple criteria
approaches. Chankong et al. (1985) develop a typology of five
dimensions which can be used to differentiate MCDM approaches:

1. Stochastic versus deterministic formulation

2. Implicit or explicit constraints (with implicit constraints,

alternatives are discrete, while with explicit constraints,
alternatives are implicit and may be infinite)

3. Methods and assumptions with respect to preferences:

a. Techniques which utilize the dominance principle and

assume that preferences are monotonic

87
b. Techniques which require elicitation of preferences by
priorities, weights, goals or ideals
c. Techniques which require elicitation of preferences by
tradeoffs
d. Techniques which utilize some form of global preference

4. The number of decision makers which can be accomodated

5. The number and type of objectives possible in the
formulation

Although each dimension is important, the authors contend that the
characteristics of stochastic/deterministic formulations and
explicit/implicit constraints are suitable to classify most MDCM
techniques.

Another classification of the approaches, according to Hwang and
Masud (1979), is formulated according to: I) the stage of decision
making at which preference information is elicited from decision
makers, and 2) the type of preference information required. According
to the first classification factor, research approaches fall into one
of the following categories.

1) No articulation of preference information. These approaches
derive the ’best’ solution according to a geometric
definition of closeness to a ’bliss point’. The essence of
the problem thus involves finding the measure of proximity
most acceptable to the decision maker.

2) ’A priori’ articulation of preferences. The decision maker
must make global judgements regarding tradeoffs between
objectives before the formulation of the problem. The MAU

approach falls in this category.

88

3) Posterior articulation of preferences. In this approach, a
subset of non-dominated solutions are determined, from which
the decision maker chooses the most satisfactory solution,
'. . . making implicit trade-offs between objectives based
upon some previously unindicated or nonquantifiable
criteria.‘ (Hwang and Masud 1979: 243) Only very general
assumptions need be made about preferences in order to
generate the efficient set. Linear and non-linear multiple
objective programming methods are found in this category.
Since the number of nondominated solutions might be very
large, these methods are often implemented in an interactive
environment, thus becoming virtually indistinguishable from
the following category.

4) Progressive articulation of preferences. This line of
research is principally context-specific, in that the
decision maker’s preferences are progressively revealed
during an exploration of the feasible set. Some methods
require explicit information about rates of tradeoffs
between objectives, while others require little more than
that the decision maker indicate the objective function
values with which he/she is dissatisfied. The decision
maker’s participation is much more intensive than in other
approaches.

With the continued development of powerful interactive computer

systems, the latter approach has received much attention in MCDM
research. The interests of operations researchers in obtaining

solutions acceptable to decision makers and of behavioral researchers

89
in avoiding some of the perceived inadequacies of the rationalist
approach have stimulated development of a wide range of research

efforts.

4.4 Mgthgmatical Fgrmglation 2f Mgltiple Objegtivg Technigggg

The general mathematical representation of the multiple objective
decision for any of the Hwang and Masud classifications is:

(4.1)

Max F = F[f1(x)....fy(x)l

subject to
x f X

A set of feasible solutions 5 is contained in n-dimensional
space, and a set of scalar-valued objective functions f are defined on
X. The model has n decision variables, m constraints and k
objectives. Any or all of the functions may be nonlinear. Rosenthal
(1985) notes that there is no precise mathematical definition of this
maximization because there is no natural ordering of the vector
function F.

With no such ordering, given two feasible alternatives y and 2,
there may be no definite answer as to whether F(y) is greater
than, less than, or equal to F(z) . . . A reasonable statement of
the problem is: find a feasible solution x so that the most
preferred vector of objective function values, F(;) is attained.
(Rosenthal 1985: 135)

The use of the subjective term ’preferred’ indicates that
additional rules or information must be imposed on the problem in
order to achieve a unique solution. This is accomplished in various
ways by multiobjective techniques, either by eliciting preferences

from affected decision makers or by utilizing mathematical rules as

substitutes for preference mechanisms. The following section

9%

discusses some of the more important methods within each of the Hwang
and Masud (1979) classifications.

The first classification, whose sole methodology is the ’method
of global criterion’ or ’compromise programming (Zeleny 1982),
requires no further information from the decision maker other than the
quantification of constraints and objectives. A vector of ’ideal’
values of the decision variables is calculated by solving the k
succesive single objective problems:
(4.2)

Max fJ(x) j= 1....k

subject to
x C X
The unique solution is found according to some global criterion which
minimizes deviations from an ideal solution consisting of the vector
of solutions from the above optimizations.
(4.3)

*

Min d" = [(ffix) - fig n / 3(3)) 9

subject to
x 6 X
The difference measure is considered to be a proxy for human
preferences. That is, decision making is assumed to function as if
the geometric concept of ’closeness’ were applicable to cognitive
processes. Choice of the parameter p is arbitrary, but can greatly
affect the computed model solution. Choice of p-I reflects a

disregard of for deviations of an individual objective from the ideal

- only the sum matters. If p-2, the function is the typical sum of

91

squares distance. As p becomes very large, only the objective with
maximum deviation matters. As p is decreased from 1 to 8, the
smallest deviation is given relatively more weight in the total sum.
The advantage of the global criterion approach is that no specific
preference information is required. There is, however, no empirical
or theoretical evidence whether preferences are realistically modeled
with distance measurements.

The second class of multiple objective decision approaches (a
priori articulation of preferences) includes both Multi-attribute
Utility (MAU) and Goal Programming (GP) approaches. MAU is the only
axiom-based multiple criteria technique. The MAU decision problem is
formulated as:

(4.4)

Max U = u<+J<5)) j= 1....k

subject to
x f X

U is the von Neumann-Morgenstern utility function with multiple
arguments fJ. The utility function must be defined over all
objectives and measured before solving the decision problem. Such
estimation becomes very difficult even for the most elementary
problems. Therefore, in empirical problems the concepts of
preferential and utility independence are applied to determine whether
the utility function can be considered either additively or
multiplicatively decomposable into single dimension functions (Keeney

and Raiffa I976). The major advantage of the approach is that if U

has been accurately measured, it will insure (by definition) the most

92
satisfactory solution to the decision maker. The major difficulty is
that the decision maker is required to articulate rather precise
judgements about preferences in a situation devoid of information
about the problem at hand. Optimal actions based upon such utility
assessment are also likely to be time and state dependent. Empirical
applications of MAU to public policy are documented in Drake et al.
(1972), while Green and Wind (1973) explore marketing applications of
MAU.

Goal programming (GP), originally proposed by Charnes and Cooper
(1961) and further developed by Ijiri, Lee and Ignizio is one of the
earliest operational techniques reflecting multiple objectives. Goal
programming approaches attempt to determine one or more solutions
which come ‘as close as possible’ to satisfying aspirations or
targets. It is sometimes considered as an operational basis for
modeling satisfycing behavior, but Spronk (1988) shows how GP can
incorporate optimizing behavior. Goal programming is based on well-
established mathematical programming methods (usually with linear
formulations) and documented behavioral characteristics of decision
processes (use of aspiration or target achievement levels and
hierarchical ordering of objectives). A common formulation of a
multiple goal problem is:

(4.5)

Min (Pf. d+., P-. d-.) j= 1....k
jjjj

subject to

f (x) - d+. + d-. = f.*
J - J J J

x C X

93

The objective function reflects preferences for positive and
negative deviations dfj and d--‘j from the aspired levels fJx of the
objective functions fJ, and £115 a vector of activity variables in the
feasible region. The deviations are weighted by the respective Pf and
Pb. The feasible region delimited by the activity variables x is
assumed convex, as is its projection in objective function space.

Two major approaches to multiple goal programming differ in their
interpretation of the weights P. The preemptive weights version is
consistent with a lexicographical interpretation of utility preference
for multiple objectives (Lee 1972). That is, objectives are
considered to be ordered in hierarchies, with satisfaction of higher
ranking objectives required before considering solutions to lower
objectives. Such non-compensatory preferences for objectives are
inconsistent with classical utility analysis.

In the second approach, the weights P are considered
compensatory, which therefore focuses the problem on estimation of
preference weights between conflicting objectives. This weighted
linear multiple goal programming formulation is also commonly
utilized. Note that this formulation is general enough to accomodate
a wide variety of formulations, with the typical maximization of each
objective function as a special case. Using as an example a model
with a single objective function:

(4.6)

min ( wf. df. f w-. d-. ) j= 1....k
J J J J

subject to

x G X

94

It can be seen that since d+ + d- = 8, the choice of w determines
the type of optimization. If wej =1 and '_j =0, minimal
overachievement of the objective is sought. The converse is true for
wej =0 and wfj =1. If both are set equal to one, exact attainment of
the aspiration level is sought. Setting wfj =1 and w-‘j =-I could be
used to minimize the objective function, while reversing the values
will maximize the function. As an approximation of a continuous
concave preference function, a piecewise linear objective function can
be formulated to indicate different marginal contributions of the
activity variables at varying levels of objective function
achievement.

The disadvantages of goal programming are similar to those of
most programming models. Uncertainty associated with parameter
estimates and other commonly encountered violations of model
assumptions compound empirical difficulties. Added to these
difficulties are the required a priori specification of two other
types of parameters: the relative weights between objectives and the
relevant aspiration levels of the decision maker with respect to each
objective. However, the latter drawbacks can be alleviated through
interactive procedures. Zeleny (1982) shows that in addition to
incompatibility with traditional utility concepts, if aspirations
levels are set too pessimistically, goal programming may fail to
identify unbounded solutions and may indicate solutions which are in
fact dominated. Goal programming has been used in a very broad range
of private and public applications, including manpower planning
(Charnes et al. 1968), production planning (Goodman 1974) and academic

resource allocation (Lee et al. I980). In agricultural research,

95
applications have included intergenerational transfer (Dobbins 1978)
and farm planning (Wheeler and Russell 1978, Flinn et al. 1988 and
Barnett et al. 1982).

In the ‘a posteriori’ class of decision models, the non-dominated
subset of all feasible solutions is identified and presented to the
decision maker for choice of the most preferred solution. This class
of techniques basically requires that the decision maker’s preference
function is monotonic in the objective functions. The Multiple
Objective Linear Programming (MOLP) or Vector Maximum approach is the
best known of this class. It is formulated as:

(4.7)

max f.(x)
J

subject to

kai) ; bk 1 a! j

x G X

In this formulation, each f is a linear function of the activity
variables x. The complete set of efficient extreme points of the
convex set formed by the linear constraints is generated. Any
interior non-dominated solution can be expressed as the linear
combination of adjacent extreme points. This process removes any
subjectivity from the mathematical formulation, since the operation of
generating the efficient set is well defined. There are numerous
techniques for determining first an initial non-dominated extreme
point and subsequently the entire set of non-dominated extreme points.
Applications of MOLP to problems such as transportation, portfolio

planning and production planning are documented in Zeleny (1982). An

96
application to capital budgeting in agriculture was described by
Candler and Boehlje (1971). The principal disadvantage of the
approach is the lack of assistance for the decision maker if a large
number of undominated solutions are generated. Therefore, ‘a
posteriori’ techniques are usually combined with interactive
approaches which elicit local preferences for intermediate undominated
solutions.

Interactive and iterative methods for progressive articulation of
preference information have received increasing attention in applied
decision analysis. As stated by Cohon:

The philosophy of iterative approaches to multiobjective problems

is an appealing one: involve the decision maker directly in the

solution process in a manner that will allow that person’s best-

compromise solution to be discovered. (I978: 211)

The search for the preferred solution (or a small set of equally
preferred solutions) involves an exchange of information at each
iteration between decision maker and decision analyst. In response to
local information about non-dominated solutions generated by the
analyst, the decision maker is asked to express local preferences
between objectives. This information is then used by the analyst to
generate more preferred solutions. The decision maker can thus learn
about relative tradeoffs between objectives implied by decision
constraints. Since the decision maker controls the direction and
termination of the decision process, the solution obtained should have
a better chance of being implemented.

Local preference information may be more accurate than global
information required by ‘a priori’ methods. Some techniques require

explicit information from the decision maker concerning rates of

tradeoff between objectives, while others require little more than

97
that the decision maker indicate which objective should be improved.
All these techniques, however, are critically dependent upon the
accuracy of local preference information. Some do not guarantee that
a single preferred solution can be obtained in a reasonable number of
iterations, and considerable input may be required from the decision
maker. Several interactive methods of this class have been
implemented for private or public planning. No interactive
multiobjective method has been applied in agricultural research. Four
methods are examined briefly for their suitability to farm management
planning: the methods of Geoffrion et al. (GDF), Zionts and

Wallenius, Steuer and Nijkamp and Spronk (IMGP).

4.4.1 The Method of Geoffrion

The method proposed by Geoffrion, Dyer and Feinberg (1972) is the
earliest interactive method, and is formulated as follows:
(4.8)

max U = UIf

(33...: (5)]

I V

subject to
x c X

The objective functions f and the constraints are assumed to be
explicitly known by the decision maker. The preference function U(i)
is assumed only implicitly known (but everywhere differentiable and
with positive first derivative). The procedure assumes: I) the
constraint set is convex and continuous, 2) 0(1) is differentiable and
concave with respect to g, 3) each fJ is concave with respect to x, 4)
dU / de ) 8 in the neighborhood of any local solution xi, where fJ is

some reference objective.

98.

The model is solved using the Frank-Wolfe nonlinear programming
algorithm, which belongs to the class of programming techniques known
as gradient methods. These techniques solve for global optima using a
series of linear approximations to a preference function. The
solution procedure consists of two repeated steps. In the first step,
the optimal direction of improvement in the utility function is
determined by eliciting information about relative tradeoffs between
objectives. The decision maker must be able to accurately express the
marginal rate of substitution between objectives (using some objective
as a reference). The second step involves calculation of the optimal
amount of change in the direction of greatest improvement. This
information is usually elicited from the decision maker in response to
graphical display of effects of various possible step sizes on the
objective functions.

Hwang and Masud state that '. . . the lack of a systematic
assessment procedure for tradeoffs is a drawback of this method . .

.' (1979: 121) Certainly the information demands of the decision
maker are not trivial. Nevertheless, it has been used in some
empirical planning efforts (Geoffrion et al. 1972). Further
developments of this technique have attempted to alleviate the rather

stringent information requirements.

4.4.2 The Method of Zionts and Wallenius

The method of Zionts and Wallenius (1976) is formulated in a
manner somewhat similar to that of GDP. All objective functions and
the constraint set must be linear in the decision variables. Again,
the preference function is assumed only implicitly known by the

decision maker. The preference function may be assumed linear in the

99

various objective functions, or more generally be assumed a concave
function of the objectives. In the latter case, the information
requirements of the decision maker are somewhat more stringent, but
the authors indicate the additional requirements are not excessive.
The method involves a cycle of four steps. First, an arbitrary

set of weights (XJ) is chosen to create a linear proxy for the
utility function:

(4.9)

max XJZI XJ fj(x) j= 1....k

subject to

x C X

This function is optimized to produce a non-dominated solution.
Then a set of ‘efficient’ variables is identified from the non-basic
variables, having the characteristic that if introduced into the
basis, such variables cannot increase one objective function without
decreasing at least one other function. A set of 'ij’ which are
decreases in the objective function {J stemming from introduction of
the non-basic variable xi into the solution are estimated around the
solution point found in the optimization. A number of tradeoffs
implied by the "ij are presented to the decision maker, who expresses
whether the tradeoffs are desirable or undesirable or whether he/she
is indifferent to the proposed tradeoffs. Essentially, the decision
maker is allowed to explore whether a move to one of the adjacent non-
dominated solutions would be desirable. Ouestions are posed as,
“Would you accept a decrease of 6 units in objective 1 for an increase

of 2 units in objective 2 and 1 unit in objective 3?’ A new set of

188

weights XJ is found which is consistent with all responses.

An optimum is reached when the decision maker rejects all proposed
tradeoffs. The optimum is reached in a finite number of iterations,
since each tradeoff increases the implicit utility function.
Information demands on the decision maker are considerably less with
this method compared to the GDF method.

This approach does require that all the constraint and objective
functions can be linearly approximated. If the implicit utility
function is assumed concave, the procedure is slightly modified and
the number of iterations required is significantly increased. There
is also a possibility for suboptimization if a proposed solution is
surrounded by other equally efficient solutions with approximately the
same utility level. Zionts and Deshpande (1978) describe how the

method was used in a U.S. Department of Energy planning model.

4.4.3 The Method of Steuer

The method of Steuer (1976) is an extension of the ‘a posteriori’
articulation of preferences techniques. As discussed above, these
techniques generate the subset of feasible solutions which are
undominated. However, the number of non-dominated solutions may
impose severe information processing requirements on the decision
maker. This general class of interactive methods thus seeks
techniques to reduce the number of possible solutions to manageable

size. The general model is to find all x, 5 so as to:

181

(4.18)
k
max 2. x. f.(x)
J=1 J J ‘
Subject to
x 6 X
x.“ x. =1
J=l j

The set of undominated feasible solutions can be unmanageably
large for realistic problems. This is because the linear weighting
parameters XJ can take on any values from zero to one in the
absence of information about decision maker preferences for the
various objectives. Steuer’s interactive method therefore relies on
an reduction of weighting parameter ranges based upon iterative
elicitation of the most preferred solution from a selection of
proposed efficient solutions. The tighter are the interval ranges for
the XJ parameters, the smaller will be the non-dominated, preferred
set of extreme points.

At each iteration, the current interval ranges of XJ are used
to generate efficient solutions consistent with the weights. To
reduce demands on the decision maker, only a subset of such efficient
solutions are presented to the decision maker (based upon the distance
function concept of ’closeness’ of solutions). The decision maker
only has to indicate the most preferred solution in this set, after
which the weighting parameter intervals are tightened in a manner
consistent with the preferred solution and another iteration

initiated. Provision is made for ’backtracking’ if the decision maker

182
is no longer satisfied with the solutions restricted by a previous
decision. The process continues until a small portion of the non-
dominated set is identified which contains the globally preferred
solution. The principal advantages of the procedure are:

1. The relaxation of information requirements from the decision
maker, both in terms of the weights associated with
objective functions and the number of candidate solutions
presented by the decision maker. At each iteration, the
decision maker must only select the most preferred solution
from a small set of proposed solutions.

2. The procedure converges in a finite number of iterations.

The method requires the assumption of an additively decomposable
utility function, and the computational procedures implied by the
gradient cone analysis utilized in this efficiency approach are not
trivial. Steuer and Schuler (1978) describe an application of this

method to forestry management.

4.4.4 The Method of Nijkamp and Spronk

Nijkamp and Spronk (1981) present an interactive goal programming
method which does not require a priori information about goals, or
explicit a priori information about the decision maker’s preference
function. It is not necessary to specify weighting factors for
objectives, since the decision maker implicitly states which objective
function is considered to be the most important at each proposed
solution. Nor must the decision maker specify by how much the
objective should be improved. Information known a priori, however,
can be incorporated within the procedure if certain objective function

values are considered important by the decision maker. Of course,

183

explicit definition of the relationships between activities,
constraints and objectives is required. The decision maker is
required only to express local preferences for prospective solutions
presented as vectors of objective function values. The method also
uses well known linear programming techniques, which greatly
simplifies its application and interpretation. The general
formulation of the Nijkamp and Spronk method (called the Interactive
Multiple Goal Programming method - IMGP) is the same as the normal
goal programming model:
(4.11)

min U = U(P+‘j dfj , P-J d-J) j= 1....k

subject to

X

f. (x) - d+. t d-. = f.
J '° J J J

x 4 X

The functions fj and gi are assumed linear in the decision
variables. The preference function U is assumed to be only implicitly
known to the decision maker and is assumed monotone nondecreasing in
both the objective functions and the decision variables. The decision
maker does not have to explicitly indicate the weighting factors (Pf
and P-), since the order of selection during the interactive process
indicates the importance of particular objectives to the decision
maker. The aspiration levels f* are minimally acceptable levels of
the objective functions and the decision maker is not required to
stipulate initial values. However, if certain values are of
particular interest to the decision maker, they may be incorporated in

the IMGP procedure.

184

This interactive version of linear multiple goal programming
attempts to avoid the problem of specifying a priori target values by
starting with a set of minimum goal levels which are satisfied by a
large number of alternatives. The set of minimally acceptable goal
values is called the ’pessimistic’ solution. These values may be
defined by the decision maker, or may be computed through the
mathematical formulation of the problem. The MOLP problem
corresponding to the general IMGP problem is (assuming all objectives
are to be maximized):

(4.12)

max f1(x)

max fk(x)
subject to

f.(x) - d+. + d-. = +.*
J- J J J

x 6 X

Each of the k single objective maximization problems is solved
sequentially , and the implied values of all other objectives are
calculated. The vector of all optimum values is called the
’optimistic’ solution. In general, the optimistic solution will be
infeasible. It will serve as an ideal point which should satisfy
preferences, but is not feasible considering conflicting objectives
and limited resources. As mentioned above, the set of minimally
acceptable objective values is the ’pessimistic’ solution. If the
decision maker does not indicate preferred values, the vector of

pessimistic values is formulated as the minimum of each objective

function in any of the k successive single objective problems used to

185

calculate the optimistic values. The final solution preferred by the
decision maker and consistent with the feasible set should be found
between the ideal and pessimistic solutions.

Trial solutions are proposed to the decision maker (starting with
the pessimistic solution), who is asked to indicate which objective
function should be improved. The indicated function is increased,
either to the next higher aspiration level indicated by the decision
maker or (arbitrarily) one-half the distance to the maximum level
consistent with the pessimistic levels of all other objectives. The
trial value of the indicated objective is augmented to the constraint
set as a ’hard’ constraint and new ideal values of the other
objectives are calculated. The decision maker is then presented
output showing the necessary decrease in potential improvement of some
of the (k-I) objectives implied by increasing the indicated objective.
If the decision maker considers the consequent effect on the other
ideal values too extreme, the amount of improvement in the objective
is scaled back until an increase is found that is acceptable in terms
of its impact on other objectives. Then the decision maker is asked
if the trial solution is acceptable, or whether another objective
should be increased. The process continues until a single preferred
compromise solution is reached or a non-dominated (preferably small)
set of solutions is obtained to which the decision maker is
indifferent (improvements in any objective are matched by unacceptable
deterioration in other objectives).

Sensitivity analysis can be carried out to provide the decision

maker with more information once a small non-dominated set of

186
solutions is identified. As in traditional linear programming,
sensitivity analysis can help explore the effects of relaxing right
hand side constraints (in this case either resource constraints or
current minimally acceptable objective levels). This information may
help the decision maker investigate further tradeoffs.

IMGP combines many advantages of goal programming and interactive
decision procedures within a traditional linear programming framework.
It can incorporate multiple objectives, both satisfycing and
optimizing behavior, and compensatory or hierarchical treatment of
objectives. Only local preferences with respect to well-specified
solutions are required, not global preferences. Information on
priorities among objectives or aspiration levels within objectives is
not required, but can be incorporated if available.

The decision process is structured, in that the decision maker is
directed towards final consideration of solutions that are non-
dominated, but maintains considerable flexibility through control of
the solution procedure by the decision maker. Increased participation
in the decision process should lead to better understanding of the
problem and of the tradeoffs between objectives. Considerable
emphasis is thus placed on the learning potential of the problem.
Spronk states that '. . . the interactive approach can be regarded as
an operational application of learning theory.‘ (1988: 184) Learning
may be stimulated by use of the procedure for a ’quasi-simulation’ of
changes in priorities of objectives, sensitivity analysis of
constraint relaxation, and possible reformulation of the entire model

describing the decision problem.

197 _

The procedure benefits from use of well-known linear programming
techniques. Spronk (1988) states that existing linear programming
packages can be adapted by addition of external processing routines to
compute the necessary steps efficiently. Convergence of the procedure
to a unique solution is assured if the structure of decision maker
preferences is such that a unique solution exists (the ’satisficer’ is
a caveat).

IMGP shares some disadvantages with traditional linear
programming. The assumptions of LP which restrict its application in
some situations are well known (see Hillier and Lieberman 1988).
Another disadvantage may be the large number of objectives which
created the need for a multiple objective analysis in the first place.
The information processing required of the decision maker in a problem
a with many objectives could be excessive. A further drawback is the
progression of IMGP from dominated solutions to non-dominated
solutions. An objection could be made that such a procedure should
permit the decision maker to explore the much smaller set of non-
dominated solutions. Although the procedure permits correction of
errors in expressed local preferences, the IMGP procedure assumes that

a consistent preference function exists.

4.5 Selection of a Multiobjective Technigue

The selection of a multiple objective decision support technique,
it appears, is itself a multiple objective problem. Such a technique
should satisfy at least the following requirements:

1. The procedure should structure and define the decision

problem in order that the decision maker may define

188

alternative actions and investigate and clarify how
preferences relate to actions and to possible outcomes.
The procedure should aid computation to investigate a larger
number of alternatives than otherwise possible, and should
define a subset of feasible efficient solutions from which
the decision maker can select a set of preferred solutions,
or (in some cases) can identify a single preferred
alternative.

The procedure should permit solution of the decision model
in a manner analogous to an unaided sequential decision
process.

The procedure should be adaptable to group decision making

processes.

With respect to the behavioralist/rationalist debate on decision

making, multiobjective techniques offer varying potential for an

operational synthesis of the two perspectives. The following are

important elements to consider in choosing a formulation for a

multiobjective decision:

1.

Mgltigle griteria grgblem solution. By definition, any of

the techniques discussed above reflect this as an important
factor in problem solving. However, techniques vary in the
number of objectives which can be modeled in formulations
designed to operate in real time. Further, the accuracy and
relevance of information reflecting tradeoffs between
objectives may distinguish the techniques. This behavioral

decision orientation of any multiple criteria technique

189

contrasts with the rationalist orientation of a revealed or
measured uni-dimensional preference structure.

Target or aspiratign valugs. Behavioral research has
suggested that individuals anchor their decisions on
context-specific values of desired objectives. Use of
target values in a technique may permit satisfycing
behavior. This contrasts with the rationalist emphasis on
’smooth’ preferences.

Inigrggtive operation in real iimg. Complexity in modeling
a decision problem must be traded off against the need for
fast turnaround with managers whose time and attention span
is limited. The bottom line in terms of decision support is
to keep waiting time for model computations to a minimum.
The rationalist perspective does not require this as a
criteria for judging the ’goodness’ of a decision model.
Data input. Most techniques will require the same type and
quantity of data for a particular problem, but each multi-
criteria procedure will vary in its requirements for
elicitation and incorporation of preference information.
Given a decision support focus, techniques which permit
interactive declaration of preferences between objectives
are of special interest.

Ouiggt information. Most multi-criteria procedures are
based on optimization algorithms, and thus are to some
extent consistent with a rationalist approach. As Willis

and Perlack (1988) indicate, the maximum information which

118
can be generated by a multi-criteria technique is the
complete set of nondominated solutions and all pairwise
local tradeoffs among objectives. Since the number of non-
dominated solutions is often very large, this characteristic
runs counter to the objective of rapid turnaround and may
violate information processing restrictions of the decision
maker. Therefore, most techniques will present a subset of
non-dominated solutions to the decision maker, either chosen
by some rule or by a priori interactively elicited
preference information. The relative tradeoff weights may
be considered and manipulated directly by the decision
maker, or indirectly through indications of preference for

one objective or another.

CHAPTER U

A MULTIPLE OBJECTIVE DECISION AID

5.1 Introduction

The current research began with an examination of the principal
issues of dispute between two schools of thought with respect to
individual decision making. The major argument of the behavioralist
school is that certain innate cognitive functions can be dysfunctional
and may cause individuals to take actions that are not in their best
interest. The perspective of decision support seeks to construct
tools which complement human strengths in perception, pattern
recognition and creative inference, while supplementing recognized
weaknesses in computational ability, probabilistic inference, and
decision rules. The framework of decision support research is thus
improving decision behavior within known information processing
constraints.

Decision makers are also observed to evaluate alternatives and
take strategic actions with the intention of seeking multiple and
sometimes conflicting objectives or goals. Mathematical programming
decision models have sometimes reflected hierarchical relationships
.among multiple objectives as ’hard’ constraints, but there have been
few attempts to implement a decision aid which explicitly models a
multiple objective decision problem. The previous chapter examined
the range of relatively new multiple criteria techniques to determine

if any offered some possibility for synthesis of rationalist decision

111

112
principles and behavioralist decision insights within an operational
decision aid. This chapter describes the decision model and aid
constructed during the current research for supporting an important

planning decision of cash grain operations.

5.2 Aggantages of IMGP

Some of the advantages of IMGP as a multiobjective solution
procedure have been mentioned in the previous chapter. Unlike
lexicographic or weighted goal programming, it is not required that
the decision maker specify a priori aspiration levels and weights
reflecting the relative preference for different objectives. The
decision maker is not asked to provide sufficient information to
generate a measure of global preference, but instead must express only
local preferences for the specific combination of objective function
levels generated at each iteration.

Spronk (1988) indicates additional advantages of the IMGP
approach. A considerable range of preference relationships can be
modeled by the procedure. Figure 5.1 demonstrates some functional
relationships between one objective and a preference function. Cases
(a) and (b) are relatively straightforward cases of monotone
increasing and decreasing preference. The general formulation of a
goal programming model for these cases is:

(5.1)

min Pf1 df1 f P-I d-l

subject to

x G X

9(x) - d+l + d-1 = g

Amy zofieozse m>Heomsmo mzo gz<
Acv zoaeozsa muzwmmuuea < zmmzemm aaxmzooh<ome o<zoaeuz=e F.m mmsufia

owmp xcocam "mucaom

 

 

11:3

 

 

     

 

 

 

 

 

 

 

 

mm mm .a u
_ - - -
u u e n n.
n . u L n e
. . .
m m u - m
I-.. L . :1-..
av Any Luv
#0 Amvm All Amvm T Amvm T
L a e

  

Adv Anv . Adv

 

 

 

114

Pf, P- are weights assigned to positive and negative deviations
(d- or d») respectively. The set of activity variables 5,is contained
in the feasible region X. The final equation reflects positive and
negative deviations from the aspiration level 9*. For case (a), 9%
can be set at some unattainable level and Pf set to zero, thus
minimizing deviations below the aspiration level or in other words
maximizing the objective function 9(5). Case (b) can be handled in an
analogous manner. Spronk (1988) calls this ’one-sided goal
programming’. Of course, with only one objective function, case (a)
can be modeled directly as:
(5.2)

max 9(5)

subject to

5 f X

Case (b) could be similarly formulated directly as a minimization
model. If the preference relationship of case (c) occurred, then the
objective function 9(5) could be split into two other functions (91(5)
and 92(5)), respectively monotone increasing and decreasing in
preference. Assuming gX is known:
(5.3)

min (df1 , d-z)
subject to

5 6 X

to
H

A
X
v
.g.
0.

I
IA
ID

to
N

I".
X
v
+
Q.
+

IV

to

115
The functional relationships in cases (c) and (d) can be
similarly solved by minimizing d- or df respectively. These
satisfycing formulations do not produce unique solutions because of
preference indifference in the ’flat’ areas of the function. Case (f)
is somewhat different, in that there are two aspiration levels (91*

and 92*). This might occur if the decision maker prefers values in

the interval between 91* and 92! to values outside that range. An
example might be a preferred range of inventory holdings. Note that
the function is monotone increasing below 91*, monotone decreasing
above 92! and flat (indicating no additional preference) between 91!
and 92*. This can be expressed as a two-sided formulation, minimizing
negative deviations below 91* and positive deviations above 92*.
Spronk also describes how an objective function 9 can be formulated as
a piece-wise linear function of an activity variable x in order to
approximate a smooth concave functional relationship between activity
variables and objectives.

IMGP as an interactive technique also lends itself to decision
support. The relatively small number of responses required of the
decision maker in an IMGP session are concrete evaluations of the
desirability of objective function magnitudes and indications of a
desire for improvement in one or more objectives. This contrasts with
the cognitively more complex rates of tradeoff required by the GDF
procedure or the greater number of evaluations required by the Zionts-
Wallenius procedure. Although IMGP can reflect aspiration levels,
which behavioral scientists contend are widely used by decision

makers, the decision maker is not forced to give a priori established

targets. Instead, vectors of objective function values can be

116

formulated mathematically which represent the worst goal values which
are necessary to accept (which Zeleny (1982) calls the ’anti-ideal’)
and the best goal values possible (the ’ideal’). Priority weights for
goal deviations need not be explicitly expressed by the decision
maker. Instead, priority is implied by the decision maker choosing
objectives to be improved. The flexibility of IMGP with respect to
functional relationships, aspiration levels and priority weights thus
considerably extends the applicability of goal programming techniques
and appears to make it the best currently available candidate for
decision support applications. If constraint functions are linear and
objective functions can be reasonably approximated as linear or piece-
wise linear in the activity variables, then use of efficient linear
programming algorithms in the decision aid may enable fast enough
turnaround to provide real time decision support.

IMGP can be shown to be a more general formulation of both MOTAD
and Target MOTAD (Romero and Rehman 1985). Given the two objective

(goal) function of return and risk:

(5.4)
fl = 222 c. x.
J-I-J J
R = 2:5 df /s
r=1 r
where df = max (0, T - 2 5 c . x.)
I" T‘J J

r=1
The first objective is the sum of expected net returns (55) per
unit of activity x.i times the activity level. The second objective is

risk, which is identified in this discrete formulation as a

117

non-decreasing function of deviations from a target level T. If T
equals expected returns, then the model corresponds to a MOTAD
formulation. If instead the target level is some fixed value, the
model is a Target MOTAD formulation. Risk is the average deviation,
or the net return for each state of nature (crj) weighted by the
activity level (x3) and divided by the probability of occurrence
(I/s). In goal programming terminology, we want risk to be ’as close
as possible’ to the minimal level of zero. We also want maximal
overachievement of the return objective. This implies a one-sided

goal programming formulation with two goal functions, as shown below:

(5.5)
min (d-l, dfz)
subject to
5 G X
n
Z._ c . x. + d+ 2 T r= 1....s
J-I rJ J 2r
5
xr=1d+2/S $.00
llfd- =u*

1
The deviation variable of returns (d-I) represents the amount by
which expected returns fall below the unattainable aspiration level

Bx. Similarly, the deviation variable d+ measures the amount by‘

2
which returns in state r fall below the target return level. The
objective of this model is to minimize some functional formulation of

(d-1’ dfz). With preemptive weights, one objective function would be

optimized (receive absolute priority), then the other. This would

118

make no sense in the current example, since the one-sided formulation
of both objectives forces the solution to be either the maximum
expected return (if the return objective has priority) or zero (if the
risk objective has priority). With weighted goal programming, a
linear function of (d- , df

1 2

reflecting the acceptable rate of tradeoff between objectives. It

) is formulated with a priori weights

appears unreasonable to expect decision makers to express a fixed
tradeoff rate between risk and return without knowledge of possible
magnitudes of the two objectives. IMGP does not require an explicit
formulation of weighted objectives. Thus the interactive goal
programming formulation seems to have particular potential to
represent this type of risk-return decision. While MOTAD and Target
MOTAD are designed to trace the entire set of solutions which are
efficient in risk-return space, IMGP can be effectively used for
decision support to reduce the number of feasible and efficient
solutions to a small set (possibly a singleton) of feasible, efficient
and preferred solutions. Target MOTAD also requires a target value.
Although this introduces an additional parameter (T) which must be
estimated or elicited during model formulation, it can be seen as an
implementation of the behavioral concept of target or aspiration
levels in decision making. By expressing some income level as
particularly important, the decision maker expresses an abrupt change

in marginal preferences for income.

5.3 iM§P Consigtency with Dgcigion-theoretic Principles

This section will examine logical properties of the general IMGP

model, in particular the conditions under which (given a fixed

119
preference structure) the procedure converges to a unique optimal
solution according to decision maker preferences. Given the formal
equivalence between a particular formulation of IMGP and the more
traditional single objective MOTAD and Target MOTAD risk-return
models, the consistency of general risk-return models with stochastic
dominance and expected utility principles is also discussed.

IMGP is not intended to model the complete preference structure
of an individual. Instead, IMGP is intended to determine the most
preferred element (or subset of preferred elements) within a
preference structure characterized as a total quasi-ordering. The
procedure assumes that such a preference relation exists. Spronk
(1988) contends that IMGP can be used to model various functional
relationships between objectives and weakly convex preference
relations (which include lexicographic ordering or traditional utility
relationships. In particular, Spronk shows that when the feasible
region in activity variable space is described by linear constraints
(i.e. is convex) and the objective functions are linear (thus implying
that the feasible region in objective function space is also convex),
IMGP converges to an (-neighborhood of the most preferred vector 9*
(which may not be unique) if the decision maker is able to indicate at
least one solution to be satisfactory, that responses are correct
according to personal preferences and that preferences do not change
during the solution process. Piece-wise relationships can be used to
approximate concave functions to any degree of accuracy.

The two-criteria IMGP model in this particular implementation is
formally identical to that of a Target MOTAD model as described by

Tauer (1983). The latter model can be considered a multiple criteria

128
model with return and risk as objectives. That is, a preference
function U defined on m evaluative (variously called objective or
goal) functions 91(5) .. gm(5), where x is a vector of activity
variables, each of which is a concave function of x (as is the
preference function) is a more general model than a real-valued
function increasing in mean and decreasing in risk. Expected return
is defined as the sum of expected net return per unit of activity
multiplied by the activity level. Risk is defined as the probability-
weighted sum of return deviations below some target return level.
This risk-return model is discussed by Fishburn (1977) as a special
case of mean-risk dominance models in which the mean of a probability
distribution function F is denoted as ”(F) and risk is defined by a
two-parameter function:
(5.6)
o
R(F) = (t - x) dF(x) o > 8

Deviations are counted from the target level t, while a
indicates the decision maker’s attitude towards below-target
deviations. If t = E(x) and o = 2, the result is a mean-variance
model, while if t = E(x) and a = 1, then a mean absolute deviation
(MAD or MOTAD) model ensues. If t is fixed across distributions, then
the result is a target semivariance model for o = 2 and target MOTAD
for o= 1.

Risk is thus a non-decreasing function of returns, and is equal
to zero at or above the target level. According to this type of risk-
return model, individuals are risk-neutral for returns above the
target level, and have risk preferences for returns below target

determined by the parameter a. Risk averse behavior is implied by

121

a ) 1, risk neutral behavior by o= 1, and risk preferring behavior

by a < 1. The class of o-t risk models has been generalized from

the specific mean-variance and mean-MAD models because of both
theoretical and empirical criticisms of those formulations. Mean-
variance models require that the utility function be of quadratic form
or that return distributions be normally distributed and utility
functions be of exponential form (Goldberger 1964). Mean-variance
efficient sets may not be consistent with second degree stochastic
dominant (SSD) sets (Levy and Hancock 1978), and have been shown to
perform poorly in rankings of alternative distributions (Pope and
Ziemer 1984). The same criticisms (and more) apply to the MOTAD
model. More generally, a behavioral assumption of any risk
formulation based on a parameter which varies between distributions is
that the individual considers variability of returns to be risk and
consequently to be avoided, regardless of whether such variability is
below or above the expected value and regardless of the resulting
wealth position or obligations of the individual.

Empirical concerns were expressed by Mao (1978) and others that
managers in investment decisions often associate risk with failure to
achieve some target return. Fishburn summarizes these observations,
'. . . most individuals in investment contexts do indeed exhibit a
target return - which can be above, at, or below the point of no gain
and no loss - at which there is a pronounced change in the shape of
their utility functions.‘I (1977: 122) He concludes, “The idea of a
mean-risk dominance model in which risk is measured by probability-
weighted dispersions below a target seems rather appealing since it

recognizes the desire to come out well in the long run while avoiding

122

potentially disastrous setbacks or embarassing failures to perform up
to standard in the short run.‘ (Fishburn 1977: 118)

Porter (1974) developed the concept of target semivariance and
showed solutions derived by this criterion to form a subset of the SSD
set. Fishburn (1977), as noted, extends Porter’s results to the more
general o~t class of models. He proves that a-t efficient sets
are implied by (subsets of) first degree stochastic dominant sets
(FSD) for all o'; 8, second degree stochastic dominant sets for all
a 2 1, and third degree dominant sets (TSD) for all a 2 2,
except when means and risk are identical. Tauer (1983) proves the
specific case that target MOTAD is implied by the 880 set.

With respect to preference functions, Fishburn (1977) states that
a decision maker’s preferences satisfy a mean-risk utility model if
and only if there exists a real-valued function U in mean and risk
such that for all distributions F and G:

(5.7)
F --> G iff

U(fl(F), R(F)) > u<u<s), R(G))

R() is a risk function as defined above, U is monotonically
increasing in U and decreasing in R, and '--)' indicates 'is
preferred to.‘ Such utility models are not necessarily consistent
with von Neumann-Morgenstern (vNM) utility axioms, in particular since
single-peakedness requires more stringent assumptions. In order for
the two models to be congruent, Fishburn (1977) shows that vNM utility

must be reformulated as:

123

(5.8)

II
x

u(x) for x 2 t

u(x) = x - k(t-x)a for x < t
where k is a positive constant defined as:

k = u(t) - u(t-I) - 1
u(ttl) - u(t)

Fishburn examines a variety of vNM utility functions published in
the literature and found certain evidence of support for congruence
between the a-t and vNM models. Several utility functions
demonstrated a significant change in shape at some fixed income level
at, above, or below the point of no loss or gain. However, risk
neutrality (linearity) above the target level was shown to hold in
only a minority of cases. Figure 5.2 shows representative utility
functions for individuals with three attitudes towards deviations of
income below a target value.

Fishburn (1977) emphasizes that although mean-risk models are
only approximations to more complex preference relations, they have a
considerable degree of compatibility with stochastic dominance and
expected utility criteria, and offer computational advantages over
some other techniques for modeling individual choices. However,
selection of a particular o-t model depends upon alternative
computational facilities for the particular decision situation, and
relevance of the solutions depends upon choice (or measurement) of the
relevant a level and identification of the situation and time-

specific target level.

121:

 

 

 

 

 

FIGURE 5.2 UTILITY FUNCTION WITH ok-t RISK

125

5.4 An Example IMGP Formulation

To illustrate a simple risk-return IMGP formulation, the farm
planning problem illustrated by Kennedy and Francisco (1974) is
utilized. The problem consists of allocating resources to cropping
and livestock activities in order to generate efficient combinations
of risk and return values. The activities possible include wheat,
barley and sorghum crops and pasture, which can be sold or used as
feed for lambs or merinos (adult sheep). Resources include 378 acres
of crop land, 838 acres of pasture land, 3888 annual hours of labor
and 312888 (Australian) of operating capital. Material balance
equations are also formulated for supply and demand of feed for winter
and for the rest of the year. Only pasture supplies winter feed, but
grains can supply feed in other seasons. Estimated gross margins are
calculated from historical enterprise gross margin data for each
enterprise estimated over each of the past five years.

Assuming a target income value (possibly a breakeven point or
fixed costs plus living expenses) of $18888, Table 5.1 presents the

efficient risk-return combinations.

TABLE 5.1 RISK AND RETURN SOLUTIONS, KENNEDY/FRANCISCO PROBLEM

Expected Returns Risk;

Plan 1 $10300 57
2 10749 161

3 10843 190

4 11136 320

5 11730 1433

 

* Risk is defined as the mean deviation of income below the target
level of $18888.

126
Figure 5.3 presents the solutions in risk-return space, with the
extreme points connected and labeled as A thru E. The solution
process consists of six steps:

1. Sequentially optimize the single objective problems:

(5.9)

IX
m
><

X c . x. + dfp 2 T r= 1....s

Calculate the implied values of the other objective in each
formulation. Call the optimal values Hband Rb

(for ’best’), and the corresponding values of the other
variables “w and Rw(for ’worst’). Formulate the

potential matrix P as follows:

8
. b b.
P8 3 1 “a Re 1
:11” R'”'

1400

1200

1000

800

600

400

200

 

 

 

r I I I I I
10000 10400 10800 11200 11600 12000

FIGURE 5.3 RISK-RETURN EFFICIENT SOLUTIONS, KENNEDY AND FRANCISCO
PROBLEM

128

Present the values to the decision maker as (flab,R b) and

8
we”,R0“). The latter is the worst solution that need be
accepted, while the former represents an unattainable ideal
solution of best return and lowest risk. Ask the decision
maker: Given the worst and best values implied by the
problem constraints, which objective should be improved?
Using the response, improve the worst value of the objective
indicated by the decision maker by one-half the amount
between the current worst and the current best value.

Append the adjusted value as a constraint to the
optimization model of the other objective, and re-solve.

The best value of the objective not indicated to be improved
has now been altered (worsened) because of the more
restrictive constraint imposed by the improved objective.
Formulate the second matrix as:

P 3

Two diagonal elements will have been changed, one by halving

the amount between the previous worst and best values, the

other because of tradeoffs imposed by the improved

objective. Present the decision maker with the current and

previous solutions and the same question as step 2.

a. If the decreased potential of the unimproved objective
is too great, the decision maker may indicate the
solution is unacceptable. If so, reduce the current

worst value of the improved objective to one-half the

129
amount between the current worst and the previous worst
value. Return to step 4.
b. If the solution is acceptable, return to step 4.

6. If the decision maker cannot indicate which objective should

be improved, display the activity values of the current
solutions. This may initiate further iterations.
Otherwise, the individual is indifferent between solutions.
As the difference between best and worst values narrows, at
some point the problem may be collapsed to the current best
value of one objective.

Figure 5.4 shows a hypothetical session with the Kennedy-
Francisco problem. The decision maker has improved minimally
acceptable expected returns in steps 1 and 3 and risk in step 2, and
has accepted as a final solution the plan with expected returns of
$11445 and mean deviation below target of $855. In other words, the
decision maker has discovered a solution which is only three percent
less than the profit-maximizing (best return objective) solution, but
which is characterized by forty percent less risk. Sensitivity
analysis of an accepted solution can be presented to the decision
maker to investigate the effect of changes in net return coefficients
and right hand side constraints. For example, sensitivity analysis of
the crop land constraint of the above solution shows that basis
changes occur if crop land increases by more than 188 acres or
decreases by 78 acres. If the potential crop land base is not
certain, sensitivity analysis may indicate the need for an additional

session with other projected values.

'130

zubmomm cummuz<¢m oz< >amzzmx .mmuuozm zo_h:46m max_ mo mauhm mmmzp mom mu24<> zo~Huzzm m>-umamo sz Raga: az< any hmum e.m manual

now

now

nno

“KN

mmv—

mum

mmvn

no

. 3

. a
xn_¢

x.c.ae .a.~co.om

nee.—
nvv——
owe——
nee——
ova.—
own—u
some.

can.—

a

a:

ccouom

pogo:

.nom

um aoam

.N doom

u. aoaw

no no.m

ooow—

com?—

com.—

comm.

oovmp

 

on

pg

 

 

cow

coo

com

com

comp

coe—

131

Given the described potential of IMGP as a multiobjective
decision support technique and the formal equivalence of a two
objective risk-return IMGP formulation to a linear MOTAD or Target
MOTAD model, an initial test of the empirical usefulness of IMGP can
be carried out by constructing a risk/target return decision support
aid. If severe problems are discovered in testing a relatively simple
implementation, it is unlikely that the technique would prove useful
in a broader context of more objectives, different functional forms or
larger problems. The following sections document the goal-directed
search aid (GOALDIR) developed in this research for supporting land

rental/crop mix decisions for cash grain operators.

5.5 Lgng Renial Qgcipigns

Within constraints of current labor, machinery and financial
resources and longer term committments for crop rotations and leasing,
cash grain operators must evaluate each year the set of land rental
opportunities and select cropping activities for both owned and rented
land. In the current economic environment, a careful decision may
mean the difference between failure or continuation of the business.
Contraction or failure of farm businesses has fostered a more active
than usual land rental market in many locations of Michigan. A wide
variety of rental arrangements are utilized, including cash and share
rent with various obligations for landlord and tenant, dates of
payment and adjustments for realized yields and prices. Often the
rental decision is a decision with multi-year consequences, the
operator knowing that if land available this year is not rented, it
may well not be available again for the foreseeable future. If a

neighbor rents the land, the implicit right to continue renting the

132

tract as long as desired may have been granted by the landlord.
Therefore, decisions made this year may affect resource utilization
and growth possibilities for years to come.

Crop rotations also instill a multi-year character to the rental
decision. If rotation requirements were to completely determine what
crops must be planted on a specific tract, the only flexibility in
determining overall crop mix would be the decision to rent or not rent
the land. At the other extreme, if any potential crop can be planted
on each land tract, then the decision involves both whether to rent a
tract and how much of each crop should be planted. With several land
tracts and only a few potential crops, the possible permutations
quickly multiply. Real decisions usually are not characterized by
either extreme. Rotation requirements are extremely important for
some tracts (particularly for owned land), while crops on other tracts
may depend almost entirely on economic prospects. Rental and crop mix
decisions are also strongly affected in the current economic
environment by government feed grain programs. Given massive stocks,
continued high production and expected low prices for most cash grains
over the next several years, there is considerable incentive to
participate in government price support programs, which limits acreage
for program crops.

A common decision procedure then involves consideration of
machinery and labor resources, crop rotations, government price
support programs, potential yields and prices (in the short or longer
run) in the selection of a preferred crop and rental mix with owned
cropland. The set of activities actually selected is considered to be

the best compromise between the business objectives, which may be (and

I33

usually are) competing to some degree. The abstract objective with
highest priority is usually ’to secure the highest possible current
period profits’. It is an abstract objective in many cases because
operators have only the vaguest notion of their variable and fixed
costs and poorly formed expectation of crop prices, and thus can make
only ’seat of the pants’ conjectures of crop mix and rental impacts on
profits. Operators also usually attempt to avoid risk, perhaps
conceived as ’to avoid disaster level losses’. Although avoidance of
risk can be incorporated as an attitude towards profit distributions
as in the expected utility framework, it is often more realistic and
useful for decision analysis purposes to consider risk as a second I
(and competing) objective in the decision process. Indeed, operators
seem to distinguish these as separate goals (Harper and Eastman 1988).
A more precise definition of these objectives will be given below as
the decision model is described.

Other objectives, possibly of lesser importance in this decision
context, might be ’to provide adequate opportunity for business
growth’, or ’to maintain high soil fertility’. Given the natural
fixed supply of land within reasonable distance from the home, access
to land resources has great influence on the growth potential of the
business. Rejection of rental opportunities this year may force the
operator to rent land in more distant locations in later years in
order to accomplish growth objectives. In addition, if land becomes
available for sale, it may well be the current rentor who is given
first chance to purchase the land. Soil fertility objectives are
accomplished through multi-year programs of crop rotations and

fertilization plans. Conceptually, these objectives can be

134

incorporated into a multi-period model. The current decision model,
however, assumes that the current period objectives of returns and
risk are predominant and does not attempt to model other objectives.
Selecting activities which are hoped to result in the most
preferred level of the objectives is particularly prone to error
because of the lack of 1) accurate cost, yield and price data, and 2)
a solution procedure or algorithm to search for the best solution.
Without farm-specific data, the best intentions and solution
algorithms may be of little benefit. Thus, an argument can be made
that establishment of farm level record keeping systems is a
precondition for any substantial improvement in decision making. On
the other hand, careful introspection or elicitation procedures might
provide sufficiently accurate data on which to base a solution
procedure promising a better result in terms of the objectives. Even
though data may be available for cost and price expectations,
operators may not use procedures which are likely to indicate ’good’
solutions. ’Back-of-the-envelope’ calculations are prone to error,
and a simple static partial budgeting procedure in itself can cause
dramatic mistakes in a dynamic whole-farm context for several reasons.
First, revenue-cost partial budgeting of land tracts ignores the
importance of other objectives for the operator, principally risk
avoidance but also other objectives such as long term growth. Setting
data parameters at expected levels does not consider variability in
those parameters or the covariation between yields of potential crops
for a particular location or between locations, nor is price
covariation between crops considered. Even when variability of crops

and yields is considered, evaluation of individual tracts can indicate

135_

rental and crop mix choices which are not preferred with respect to
objectives because the solution procedure fails to to consider the
whole business context. For example, rental of one high net revenue-
high variability land tract may imply unacceptable risk for the
enterprise taken as a whole.

A number of potential stategies may be calculated by farm
operators, limited by the time and effort necessary to estimate
returns. Even if a proposed whole set of alternative crop acreages
and rental opportunities are considered simultaneously, risk avoidance
and other objectives are usually not considered formally. This ’hunt
and peck’ solution process offers no assurance that other strategies
not contemplated might offer better results at no additional cost.
Nevertheless, it is probable that the perspective, experience and
knowledge of the operator compensates for lack of adequate solution
procedures.

Although research as discussed in section 4.2 has shown that
multiple objectives are clearly important in farm decisions, the
manner in which individual operators formulate their objectives is
unclear. Growth and security objectives, for example, can be
formulated in several a priori valid ways. One formulation of a
growth objective might be an increase in the resource base of the
business, while another might be an increase in revenues through more
intensive use of existing resources. The formulation of a risk or
security objective is particularly critical for the operation.
Extension agents and directors interviewed for this research generally
considered that farm operators view the likelihood of net cash flow

from cropping activities falling below some critical value (usually

136

related to family living and other cash obligations) as an intuitive

measure of risk.
5.6 Modeling theLQecision Problem

5.6.1 Assumptions

Modeling of such a planning decision involves consideration of
the preference structure of the individual or individuals making the
decisions, the resource base and other constraints on actions, the
activities possible with the resource base, the relevant technical and
economic parameters and the particular mathematical formulation used
to solve the problem in a decision aiding context. Boundaries must be
placed on the real problem so that a mathematical model may be
constructed. In the end, it is hoped that the solutions generated by
the mathematical model will assist the decision maker to deal with the
real problem. It is assumed that the only objectives sought by the
operator are I) to obtain the highest possible expected net return and
2) to obtain the lowest possible risk mixture of cropping activities
on owned and rented land. Risk is defined as probability-weighted
deviations below some situation-determined minimally adequate level of
net returns, as in Target MOTAD (Tauer 1983).

Farm operators are assumed to accurately describe their beliefs
and expectations for yields for possible crop on owned and rentable
land tracts and for prices at harvest for each crop in terms of
probability distributions. An assessment procedure was implemented to
elicit those distributions as part of this research. The rental
decision is made after determining the landlord’s asking price (either

in cash or share) for a particular property and after determining

137

whether the property would be included (if eligible) in the government
feed grain program. Variable costs for each crop considered possible
to plant on the property are assumed known with certainty before
rental. Consistent with the assumption of known probability for
yields, this assumption implies that fertilizer and related costs are
estimated in a manner consistent with yield potential of the land.
Machinery resources are fixed for the relevant decision horizon. Own
and hired labor are assumed sufficient so as always to be less

restrictive than the machinery capacity.

5.6.2 The Mathematical Model

The model utilizes the efficiency principle, in that no solution
indicated by the model is dominated in terms of the objectives. Given
the optimization framework, and given that a decision aid must operate
in real time, all functional relationships are assumed adequately
represented by a linear programming (LP) formulation. Principal among
the relevant limitations of LP in the current problem is the
divisibility assumption. An LP model does not distinguish between
renting 8.1 or 188 acres, even though rental agreements must be made
for an entire tract. This issue will be discussed below with respect
to preliminary testing of the model. Also relevant is the necessary
condition of proportionality, which assumes that a fixed amount of
output is produced by a particular amount of input, no matter what the
level of the activity.

The mathematical formulation consists of optimization of a two-
objective IMGP model, with the objectives: 1) maximize expected net

cash revenue from cropping activities, and 2) minimize expected

138
deviations below a subjectively determined net cash revenue target,
subject to constraints on: 1) total tillable acres available on each
potential tract, 2) chance-constrained field hours available for each
planting and harvest period, 4) possible maximum acres for each crop,
and 5) wheat acres already planted the previous fall. In addition,
acreage of certain crops is restricted for tracts entered in
government feed grain programs. Since optimization of more than one
objective function is an ambiguous mathematical formulation, the
decision aid is formulated to sequentially solve single objective
models which differ only in the objective row and iteratively
redefined constraints on minimal/maximal levels of the other
objective. The relevant variables and model equations are listed in
Appendix 1.

Certain variables and parameters are elicited from the particular
farm operator, while others are imposed on the problem as reasonable
estimates for data which is otherwise difficult or impossible to
elicit. Elicited data includes price and yield distributions, land
acreages, rents and government program status, variable costs of
production, planting and harvesting machinery capacity and the target
net revenue level. The decision aid components used to elicit this
data will be described below. External data which is imposed for any
farm operation includes the correlation matrices of price and yield,
yield penalties for untimely field operations, corn moisture and
related drying costs, and estimated field hours for planting and
harvest operations. Sources for the external data are included in

Tables 5.2 through 5.6.

139

TABLE 5.2 ESTIMATED YIELD CORRELATIONS, MAJOR CROPSx

Corn Soy Navy Wheat
Corn 1.8 .4 .3 .1
Soy .4 1.8 .6 .2
Navy .3 .6 1.8 .1
Wheat .1 .2 .1 1.8

 

SOURCEs: USDA, Wheat Situation, Various Issues.
USDA, Feed Outlook and Situation Report, Various Issues.

X Yield correlations are estimated from detrended national average

yields, except navy beans for which Michigan yields were used.

148

TABLE 5.3 ESTIMATED PRICE CORRELATIONS, MAJOR CROPSx

Corn Soy Navy Wheat
Corn 1.8 .7 .3 .8
Soy .7 1.8 .4 .4 .
Navy .3 .4 1.8 .3
Wheat .8 .4 .3 1.8

 

SOURCES: USDA, Whegt Situgtion, Various Issues.
USDA, Feeg Outlook gggrSitugtion Report, Various Issues.

Price correlations are estimated from national average annual prices,
deflated to 1967=188, except navy beans for which Michigan
prices were used.

141
TABLE 5.4 YIELD PENALTY MATRIX FOR UNTIMELY FIELD OPERATIONS

PENALTY FOR UNTIMELY FIELD OPERATIONS--CORNx

Harvest Period
Planting
Period 9/1-9/15 9/16-9/38 18/1-18/31 11/1-11/38
4/21-5/18 8 1.88 8.99 8.95
5/11-5/28 8 8.95 8.94 8.98
5/21-5/38 8 8.82 8.85 8.81
6/1-6/38 8 8 8 8

PENALTY FOR UNTIMELY FIELD OPERATIONS--SOY BEANS

Harvest Period
Planting
Period 9/1-9/15 9/16-9/38 18/1-18/31 11/1-11/38
4/21-5/18 8 8.98 8.94 8.86
5/11-5/28 8 8.98 8.93 8.86
5/21-5/38 8 8.96 8.91 8.83
6/1-6/38 8 8.76 8.71 8.69

PENALTY FOR UNTIMELY FIELD OPERATIONS--NAVY BEANS

Harvest Period
Planting
Period 9/1-9/15 9/16-9/38 18/1-18/31 11/1-11/38
4/21-5/18 8 8 8 8
5/11-5/28 8 8 8 8
5/21-5/38 8 8 8 8
6/1-6/38 8.99 8.95 8 8

 

SOURCE: Estimated from Black, J. (1974) and Rotz and Black (1985)

X Yields are penalized by the indicated factor for field operations

carried out in the indicated planting/harvesting period.

142

TABLE 5.5 PERCENTAGE POINTS REQUIRED TO DRY CORN FOR DIFFERENT
PLANTING/HARVEST PERIODS

Harvest Period
Planting
Period 9/1-9/15 9/16-9/38 18/1-18/31 _ 11/1-11/38
4/21-5/18 8 15 9 6
5/11-5/28 8 19 13 18
5/21-5/38 8 21 15 12
6/1-6/38 8 8 8 8

 

SOURCE: Estimated from 'Corn-Soybeans Planning Guide", TELPLAN User’s
Guide No. 18. Department of Agricultural Economics, Michigan
State University, 1976.

X For corn planted and harvested in the indicated periods, the table
displays the number of percentage points that corn must be dried.
Cost of drying is calculated at $.825 per bushel.

143

ESTIMATED FIELD HOURS AVA£LABLE FOR PLANTING AND HARVESTING

OPERATIONS, EAST LANSING

 

Estimated field hours at 88% probability level for East Lansing
Planting operatons field time
appropriate for all major crops. Harvest operations field time

appropriate for soy, wheat and navy.

Planting Operations

Clay loam

WD PD
124 63
74 64
85 81
249 225

Sandy loam

WD

148

73

89

268

Harvesting Operations

86

81

167

133

TABLE 5.6
Clay

Planting X!
Period WD PD
4/21-5/18 188 48
5/1145/20 73 52
5/21-5/38 88 89
6/1 -6/38 247 228
9/1 -9/15 83 74
9/16-9/38 77 68
18/1-18/31 168 148
11/1-11/38 124 181
SOURCE:
X

area for various major soil
XX

PD = Poorly Drained,

78

72

149

118

types.

WD = Well Drained

96

91

189

159

Estimated from Rosenberg, et al. (1982).

PD

125

76

95

238

89

84

173

148

Sandy

WD
162

84

98

237

185
182
211

188

PD

139

74

81

251

99

95

197

171

144

Before discussing the data elicitation, it is convenient to
describe in more detail the software components of the decision aid.
Commercial microcomputer applications software was utilized to the
extent possible in order to hold down development costs and time.
Other components were developed in the Pascal programming language on
a microcomputer. The components are: 1) data elicitation spreadsheet
templates, 2) a matrix generator program which creates the LP tableau
according to the deterministic and probabilistic data input, and 3) a
linear program solver, and 4) a solution display program for iterative
reformulation of objective function levels. The farm operator
interacts with the decision aid through the analyst for the data
elicitation and solution display software, and not at all with the
matrix generator and LP solver. The model decision process can be

depicted as in Figure 5.5.

5.6.3 Data Elicitation

It was considered that response of operators to data and
probability elicitation would be better if they could see questions
and responses displayed on a computer monitor. Spreadsheet templates
(See appendix 2 for templates) were prepared for elicitation of:

1. land information

2. crop budget data

3. yield probability distributions

4. harvest price probability distributions

For each owned or rented tract, information is requested
concerning total and tillable acres, soil type (each tract is

classified according to its prevailing soil type), government program

11:5

QH< onmHUMQ mHQQ<o¢ 2H 30gb on9<2mOEZH mom mchHm

 

 

 

 

M<QmmHQ
ZOHBDAom

 

ml?

 

mm>aom

+1

 

 

m59<mmzmc
XHxB<2

 

 

// \\

fillllllllllllV

 

 

amwu<2<

 

 

T

 

ZOHB<BHDHAW
<B<Q

 

 

 

Amvzmees
onmHomo

 

 

 

111-6
eligibility, base acres and yields. This information is stored to be
entered as input to the matrix generator program.

Variable cost budgets for medium yields of relevant crops as
estimated by farm management specialists are stored as defaults in
another template. Since different tracts have varying yield potential
and corresponding crop variable costs, a budget may be associated with
a single land tract. Each default budget is divided into dollar
amounts for variable cost categories such as fertilizers, weed and
pest chemicals, fuel and lubricants, seed, repairs and maintenance and
other expenses. The operator may change any of the dollar amounts or
accept the defaults as suitable for a particular land tract or soil
type. Elicited budgets are then stored for later input to the matrix
generator.

Probability elicitation techniques differ widely in their
procedures and complexity of application, their assumptions about the
subject’s expertise and understanding of probability concepts, and
their reliance on feedback mechanisms to facilitate comparison between
stated beliefs and estimated probabilities. An elicitation technique
should produce probabilities which: 1) are consistent with the rules
of probability, and 2) are consistent with the belief structure of the
individual. Questions of technique accuracy (for known
distributions), reliability (consistency of an individual’s estimates)
and acceptability (the individual’s attitude towards the elicitation
technique) should be addressed in choosing an elicitation technique
(Ludke et al. 1988). Unfortunately, there is little evidence
indicating clear superiority of one technique over another, and there

is no theoretically superior method to compare and evaluate

11:7

techniques. Indeed, techniques which work well in one situation may
fail in another (Wallsten and Budescu 1983). Until more extensive
comparisons of techniques can be carried out, selection of elicitation
technique for decision aiding purposes should probably be based on
factors such as the following:

1. Confidence that results both adequately reflect the
individual’s beliefs and are consistent with formal
probability rules

2. Time necessary for elicitation

3. Visual and verbal feedback for the subject and encouragement
for adjustment.

Under these criteria, the 'conviction weights' method developed
by Nelson (1978) is a good choice. It is an indirect elicitation
procedure, in that the subject does not state numeric probabilities.
Indirect methods serve to ’reveal’ probabilities in a way particularly
suitable for farm operators, who are not assumed to have explicit
training or understanding of probability concepts. In this method,
the subject assigns weights to discrete outcome intervals according to
strength of conviction as to their likelihood. The subject is asked
to indicate the most likely (modal) interval, to which a weight of 188
is assigned. For all other relevant intervals, weights from 1 to 188
are assigned according to the perceived likelihood of outcomes in that
interval as compared to the most likely interval. Thus, an interval
which is perceived as one-half as likely as the most likely interval
would receive a weight of fifty.

Endpoints of the distribution are elicited, with zero probability

beyond these points. However, the subject may legitimately place

114,8

considerable probability in the last intervals. The probability of
each interval is calculated by dividing its weight by the sum of all
weights. This assures that probabilities sum to unity by design,
which of course may not reflect true (although incoherent) beliefs
(Wallsten and Budescu 1983). Feedback given to the subject may serve
to correct some inconsistencies with the subject’s beliefs. Another
advantage of the process is its avoidance of cumulative error, which
in some techniques may cause serious problems. The conviction weights
technique allows subjects to anchor their convictions on the most
likely interval, which is closely related to unaided assessment
procedures as described by cognitive scientists (Hogarth 1988).

Equal interval widths are used for both yield and price
distributions, since subjects may be confused by unequal intervals.
The width of yield intervals can be established during the elicitation
process, although many crop yield distributions lend themselves to
division into 18-15 intervals. For example, an interval of ten
bushels per acre of corn gives fifteen intervals between twenty and
one hundred seventy bushels. No research appears to have been carried
out which investigates the effect of interval width on elicited
distributions. An extremely wide interval (and consequently a smaller
number of intervals) would tend to overestimate dispersion in tails
with abruptly declining probability and could also bias estimates of
.central tendency. On the other hand, extremely narrow intervals may
exceed the discrimination capacity of the subject and encourage
frivolous estimates or ad hoc smoothing techniques. The automatic

recalculation of intervals with these templates and display of all

149
intervals on the computer monitor greatly facilitates the elicitation
process.

Farm operators are less likely to have a thorough understanding
of the causal factors underlying price distributions than of yield
distributions, and hence may not have well-formed expectations about
prices. To aid in focussing their expectations, default conviction
weights are displayed and used an an anchor for price elicitation. It
is recognized that a farm operator may be willing to accept an
expert’s estimates or to base his assessment on that of an expert for
planning decisions, just as he/she might accept an economic forecast
of interest rates for such purposes. These default weights were
elicited from a Michigan State University price analyst. No claim of
additional validity or accuracy is made for these weights, only that
they represent the opinions of an marketing expert. Although useful
as a reference, unthinking acceptance of the defaults would result in
rental activities which do not reflect the operator’s beliefs. For
this reason it is emphasized to the operator that there is no reason
why attention must be paid to the default weights, and that he/she is
free to accept or modify the weights as desired. However, there is no
reason to assume that the operator must disagree with the expert,
particularly when present market conditions indicate that there is
likely to be very little price movement in major grains. Whether
operators will indicate distributions substantially different from the
default is an empirical issue.

Figure 5.6 shows displays of probability distributions,

histograms and cumulative distributions corresponding to hypothetical

158

conviction weights. The histogram provides the subject with a visual
comparison between stated conviction weights and underlying beliefs.
In particular, the skewness of the elicited distribution and computed
histogram is an initial check on expectations above and below the
modal range. Explanation of the implied probabilities for modal and
near-modal intervals and cumulative probabilities in the tails are
described in terms of ’years out of ten’ for yields and ’chances out
of ten’ for prices. The explanation of yield distributions in terms
similar to historical frequencies is based on the contention that
yield variations (given timely field operations) are primarily based
on random weather variations, and that expectations are usually
related to historical yield frequencies. Of course, new technology,
new varieties or weather variations never personally experienced can
prove yield expectations to have been very inaccurate estimates if
historical records alone determine expectations. The operator is
encouraged to formulate yield expectations in terms of constant
technology and ’normal’ weather patterns.

The terms used by the interviewer and the sequence of questions
and explanations all may contribute significantly to elicitation
accuracy, replicability and acceptability. A complete protocol of the

elicitation procedure is reproduced in Appendix 2.

5.6.4 Matrix Generator

The matrix generator program (NORMGEN) is a Pascal program
developed during this research to calculate random revenue vectors and
other rows of an LP tableau. The data obtained in the elicitation

phase of the decision aid process serves as input to the matrix

[SCREEN I]

Name:

John O . Farmer

Crop:
Corn

Yield Unit:
Bu.

Farm No.:
1

Soil Type:
Brookston Loam

Interval:
9

[SCREEN 21

Yield
Range

 

8- 9
18- 19
28- 29
38- 39
48- 49
58- 59
68- 69
78- 79
88- 89
98- 99

188- 189
118- 119
128- 129
138- 139
148- 149
158- 159
168- 169
178- 179
188- 189

151

Yield
Range

 

18- I9
28- 29
38- 39
48- 49
58- 59
68- 69
78- 79
88- 89
98- 99
188- 189
118- 119
128- 129
138- 139
148- 149
158- 159
168- 169
178- 179
188- 189

Probability That Yield:

"3888888;

Weights

 

In Range Equal or Less Than Range
.83 XXX .83
.88 xxxxxxxx .11
.14 XXXXXXXXXXXXXX .25 Std. Dev.:
.17 XX¥¥§XXXXXXXXXXXX .42 18.12
.28 §¥XXXXXXXXXXXX¥XKXXXX%XXX!!! .78 Coef. Var.
.14 XXXXXXXXXXXXXX .85 .14
.11 XXXXXXXXXXX .96 Exp.Yield:
.83 XXX .99 131
.81 i 1.88

FIGURE 5.6 MONITOR DISPLAYS OF HYPOTHETICAL CONVICTION WEIGHTS AND
IMPLIED PROBABILITIES

152
generator program. Data imposed from external sources is also
utilized at this point. Random vectors of net cash revenue for all
possible crops on each land tract are generated based upon moments of
elicited price and yield distributions, variable costs, rents and
government payments. Since this is a significant part of the model
matrix, an explanation is warranted.

The matrix generator performs two functions: 1) simulation of
sample vectors of returns for cropping activities consistent with the
elicited data, and 2) generation of a data matrix in the format
necessary for the LP solver. A sample vector of returns is defined by
the following matrix equation:

(5.18)
NCF = (P’ Y) - V - R f G - D
NCF = Net cash flow per acre for all crops
for sample state of nature s.
P = Random vector of crop prices at harvest
Y = Random vector of yields per acre

V = Deterministic vector of cash variables costs
per acre

R = Deterministic vector of rental costs per acre
G = Random vector of government payments per acre

D = Random vector of drying costs per acre

Stochastic price and yield distributions determine random
government payments and drying costs. Total net revenue per acre is
calculated as the product of random yield times random price plus
government payments (dependent on random prices) minus variable costs,

rent and drying costs (dependent on random yield and planting/harvest

153

period). Therefore, a sample vector of returns, which can be regarded
as a sample ’year’ or state of nature, can be calculated by Monte
Carlo sampling from a multivariate distribution of yields and prices.
If the distribution of net revenues is judged to be reasonably
approximated by the multivariate normal distribution, then the moments
implied by the elicited marginals and estimates of correlation
coefficients completely describe that distribution. Anderson et al.
(1977) state that appeal to the central limit theorem for
Justification of the normality assumption becomes more reasonable as
the number of random variables increases or the more independent are
the random variables. While farm operators can be expected to have
fairly well-formed marginal distributions for yields and prices of
individual crops, they seldom have similar joint expectations for
correlated yields or prices, and have even less conception of
statistically independent underlying variables associated with weather
variability, such as sunlight or rainfall (King 1979). Joint
elicitation of random variables is a very time-consuming procedure of
untested reliability for both research and decision support purposes.
Such procedures have thus seldom been carried further than textbook
explanation (Anderson et al. 1977). Research has also indicated poor
results for attempts to directly elicit correlations between random
variables, and there seems to be little evidence that farm operators
are better equipped to estimate correlations than other subjects
(Tversky and Kahneman 1973). Historical aggregate data can be used to
estimate correlation coefficients for both yield and price variables
and thus used in conjunction with directly elicited marginal

distributions. Given generally short farm-specific yield data series

1511

and confounding time trends, it may be expedient to estimate
correlation coefficients outside the decision session from aggregate
county, state or (in the case of some price correlations) even
national level data. Although marginal yield variances from aggregate
yield data tend to underestimate individual farm variances,
correlations are much less affected than marginal variances, and
generally are not expected to significantly affect Monte Carlo
sampling.

A procedure to generate variates from a multivariate normal
distribution is described in Naylor (1966) and implemented in a
FORTRAN program by King (1979). The estimated correlation (and
corresponding covariance) matrix can be used to determine a unique
lower triangular matrix. When this matrix is multiplied by a vector
of independent standard normal variates and the resulting vector is
summed to a vector of marginal means, the result will be multivariate
normally distributed with covariance matrix closely approximating the
original covariance matrix and means approximating the original
estimated values. The only restriction placed on the
correlation/covariance matrix is that it be positive definite and
symmetric. In matrix notation:

(5.11)
X = C2 + E

E is an Nx1 vector of elicited means of the marginal
distributions, 2 is an Nx1 vector of independent standard normal
variates, C is a lower triangular matrix of dimension NxN calculated
from the estimated covariance matrix (constructed from the elicited

variances and the imposed correlations), and X is the resulting Nx1

155

vector of values for one state of nature drawn from the correlated
multivariate normal distribution with expected values and covariance
matrix approximately equal to the original estimates. Appendix 3
lists the Pascal procedure used to generate dependent multivariate
normal variates. The procedure depends importantly on the premise
that correlations are preserved through the transformations made by
the described process. King (1979) provides evidence to support that
assumption, reporting original and generated correlations which differ
by no more than 8.85 in his tests. Examination of the procedure
implemented in this research also showed only minor discrepancies
between input and output correlations. In the current decision model,
a variate is generated according to the above process for each
relevant crop yield on each tract and for each crop price. For
example, if there are ten land tracts and three crops are considered
for each tract, thirty three variates would be generated.

Another major exogenous factor affecting yield has not yet been
considered - uncertain field days available for planting and
harvesting periods. Farm operators are constrained from profitably
farming more acres by their capacity to complete field operations in a
timely fashion. Researchers such as Rotz and Black (1985) indicate
considerable decline in yield potential for crops planted or harvested
outside certain optimal field work periods. King (1979) suggests that
for modeling purposes, random variates for estimated field time
available can be generated, upon which yield potential would depend.
This would considerably increase the number of variates to be

generated and the complexity of model formulation. For applied

156
decision purposes, farm operators may be willing to accept a more
simplistic but plausible rule that yields of crops planted before or
after recommended periods are adjusted by a fixed penalty matrix based
on the mean yield reduction expected per day during the relevant
planting/harvesting period as determined in agronomic experiments.
Table 5.7 shows the penalty matrix used for the current model.

As noted, the more or less continuous decline in yield potential
outside the optimum periods is approximated by discrete penalties when
field operations cannot be completed in timely fashion. It is not
clear how to best establish reasonable expectations for field time
likely to be available in various periods of spring and fall at the
individual farm level. Extension experts state that farm operators
often use the rule of thumb that they would like to be able to handle
field operations in a timely manner in four of every five years.
Machinery complement purchases seem to be based on this heuristic.
Many optimization models have therefore used a chance-constrained
programming approach to estimate available field time limits such that
the probability that actual field time is less than that estimated is
less than 8.2 (corresponding to the heuristic), or in other terms:

(5.12)
Field time limit = Pr(h g H) g 8.2

The probability operator is denoted as Pr, H is actual field
time, and h is estimated field time. The estimates for field hours
available used in the current research are calculated from Rosenberg
et al. (1982), which presents a simulation model of ’go-no 90’ field
days based on temperature, precipitation and pan evaporation.

Predictions of the model were found to vary from historical records by

157

TABLE 5.7 YIELD PENALTY MATRIX FOR UNTIMELY FIELD OPERATIONS

PENALTY FOR UNTIMELY FIELD OPERATIONS--CORNx

 

Harvest Period
Planting
Period 9/1-9/15 9/16-9/38 18/1-18/31 11/1-11/38
4/21-5/18 8 1.88 8.99 8.95
5/11-5/28 8 8.95 8.94 8.98
5/21-5/38 8 8.82 8.85 8.81
6/1-6/38 8 8 8 8

PENALTY FOR UNTIMELY FIELD OPERATIONS--SOY BEANS

Harvest Period
Planting
Period 9/1-9/15 9/16-9/38 18/1-18/31 11/1-11/38
4/21-5/18 8 8.98 . 8.94 8.86
5/11-5/28 8 8.98 8.93 8.86
5/21-5/38 8 8.96 8.91 8.83
6/1-6/38 8 8.76 8.71 8.69

PENALTY FOR UNTIMELY FIELD OPERATIONS--NAVY BEANS

Harvest Period
Planting
Period 9/1-9/15 9/16-9/38 18/1-18/31 11/1-11/38
4/21-5/18 8 8 8 8
5/11-5/28 8 8 8 8
5/21-5/38 8 8 8 8
6/1-6/38 8.99 8.95 8 8
SOURCE: Estimated from Black, J. (1974) and Rotz and Black (1985)
X

Yields are penalized by the indicated factor for field operations
carried out in the indicated planting/harvesting period.

158

approximately fifteen percent on average over soil types. Table 5.6
presented the estimated field hours for four soil types in the East
Lansing area.

Field hours are computed from the predicted field days in
Rosenberg et al. (1982) by assuming six day work weeks and twelve hour
work days for the planting/harvesting equipment. Sufficient labor is
assumed available to operate machinery during these periods. It
should be noted that the East Lansing estimates represent
extrapolations from Bad Axe data, since Rosenberg presents only
summary estimates for that locations. The model itself is
unfortunately not available to estimate more location-specific
estimates for a wider range of Michigan sites.

Subsequent development of the model described in this research
should include more extensive investigation of field time constraints,
which is a major factor affecting selection of optimal cropping and
rental activities. In this model, as described above, additional
acres are planted/harvested according to the capacity of existing
machinery complements until the field time limit is reached for a
particular period. Field operations in subsequent periods may be
subject to additional yield penalties, thus forming incentives for
timely operations.

Machinery resource data is elicited from decision makers, and
machinery technical coefficients are then estimated from White (1978).
The acres covered per hour by a given planting/harvesting machinery

complement is estimated by:

159

(5.13)
Acres/hour = W M E

in which W - operating width of equipment

operating speed of equipment

E

efficiency factor for the given type of
equipment, reflecting necessary stops for
lubrication, turning at the end or rows, etc.
A vector of net cash revenues is then calculated by multiplying
the vector of random yields for each tract (adjusted by the yield
penalty matrix) to obtain production valued at harvest prices.
Government payments are then calculated by multiplying base acreage
times base yields times the difference between target and cash prices
(if the latter is above the government program loan rate). Diversion
payments for set aside acres are added and costs of planting a cover
crop on set aside acres are included. No consideration is given for
interest income implied by timing of payments. This amount is then
distributed across planted acreagel.
To summarize, the data input by the analyst to the matrix
generator includes the following:
1. Yield expected values and coefficients of variation for each
crop on each tract, and yield correlations between tracts
2. Price expected values and coefficients of variation for each

crop, and correlations between all prices

3. Variable costs for each crop on each tract

 

There is no requirement in the model that government program
crops be planted on all eligible acres. Nevertheless, the
difference in revenue per acre for all tracts with government
base acreage using likely relative prices indicates that
generally all eligible acres would be planted.

168
4. Total and tillable acreage of each tract and rental payments
if applicable
5. Government program information for each separately listed
Agriculture Soil Conservation Service farm number, including
base acreage and base yield
6. Chance-constrained estimated field hour time limits for
planting and harvesting periods
7. Technical coefficients for planting/harvesting operations
8. Operator-specified cash revenue target below which
deviations under any state of nature will be penalized.
The second function of the matrix generator program is to create
a data file suitable for generation of the LP tableau. This process
creates and stores a file of row and column labels, coefficients,
constraint equality directions and resource quantities. The
coefficients and resource quantities are calculated either as
described above from farm-specific data entered by the analyst or from
default values coded in the program. This function is specific to the
format required by the LP solver utilized in this research, although a
new procedure could be easily written to generate other formats. A
final function of the matrix generator program is to store a small
file containing basic information for each land tract for later use by

the solution display program.

5.6.5 LP Solver

The function of the LP solver within the decision aid is to read
and solve the linear system of equations created by the matrix
generator program and subsequent modifications of certain values

indicated by the farm operator through the solution display software.

161

Creation of unique LP solver procedures for a particular decision aid
is not a cost-effective investment of programming resources, and a
commercial LP solver is also likely to provide more extensive
capabilities and faster solution time. The disadvantages of a generic
LP solver include the requirement to purchase the software for each
installation and interfacing problems for input from the matrix
generator or output to the solution display software.

The microcomputer LP solver used in this research is a commercial
software product called LP88 1. This stand-alone set of programs
will solve problems as large as 588 constraints and 2588 variables.

It operates under the MS-DOS operating system and supports the Intel
8887 numeric microprocessor. All calculations are carried out in
double precision (16 digits), and additional error control is provided
by optional matrix inversion of basis columns. Control of the program
may be directly (interactively) through selection from menu displays
or indirectly (batch) from a file of commands. Data input and editing
can also be carried out from a previously written file or directly
from the keyboard with a spreadsheet-like data editor.

Output reports of various types can be generated, and sensitivity
analysis can be carried out on constraint or objective function
coefficients modifications. Solution values can also be stored in a
format readable by other programs, the basis of coefficients in the
optimal solution can be stored and reused. This option is extremely
useful for solving similar large linear systems, since a previous

solution usually provides a better starting point than a basis

 

Eastern Software Products, P.O. Box 15328, Alexandria, Virginia
22289

162

consisting of singletons and artificial variables. As described
below, the solution process of the decision aid consists of sequential
solution of moderately large but very similar linear systems of
equations. The time necessary to solve large systems from a
‘slack/singleton can be substantial, but subsequent solution time

utilizing a stored coefficient basis is reduced significantly.

5.6.6 Goal-directed Solution Display Program

The principal functions of the goal-directed search program
(GOALDIR) are to display the effects of the current and previous LP
solutions on the revenue and risk objectives of the decision problem,
accept input from the operator with regard to modification of
acceptable objective levels, and create files to alter the relevant LP
matrix and update a summary of previous steps. GOALDIR is a stand-
alone program developed in Pascal for the current research. It takes
as input three files: 1) the land information file created by the
matrix generator; 2) the current solution file created by the LP
solver; and 3) (after the first step) a file created during the
previous step which contains the LP solution from the previous step
plus a summary of all previous steps. The current solution consists
of the optimal values of activity variables resulting from
optimization of one objective subject to feasibility constraints and a
minimal/maximal constraint on the other objective.

The program first displays information to the farm operator
concerning the most and least preferable (best and worst) levels of
the two objectives within the constraints of the feasible set. These

values are displayed for both objectives in tabular and graphic form,

163

as shown in Figure 5.7. The graphs and tables show changes in the
best and worst levels of both objectives based upon the changes
indicated by the decision maker in the previous step. In addition, a
partial cumulative probability table is displayed which indicates:

1) the probability that net cash inflow from cropping activities will
fall below the pre-determined target level and 2) the cash inflow
resulting from the five worst states of nature. Since each state of
nature is equally probable (drawn from a random distribution with
equal likelihood), arranging net revenues from all simulated states of
nature in ascending order gives a discrete approximation to a
cumulative distribution of net cash revenue from cropping activities.
The table provides the decision maker with additional information
concerning the probabilities of very low revenue from the current
solution.

To evaluate the consequences of decisions from the previous step,
the decision maker may review in identical format the results from the
previous step. To review all previous trial solutions, the decision
maker can also review a summary table of objective levels. Finally,
the decision maker may review the cropping activities on each land
tract implied by the current solution. Sample screens are reproduced
in Figures 5.7, 5.8 and 5.9.

Basing decisions on the above information, the decision maker may
choose to alter the ’worst acceptable’ value of one objective. The
program requests the decision maker to indicate the objective, then
suggests a new value equal to one half the difference between the

current worst and current best values. The decision maker may accept

C U R R E N T
Objective

WORST

GLOBAL
CURRENT

)0

1500!

300

: 10‘

GLOBAL
CURRENT

150000 I

 

:1 C) (:1 C)

1(

Revenues and Probabilities

Possible Net

100

181

19300

19500

14

1000

(:1. I

(.

'3
2.

Revenue
Probability

Net

C) . CH

C) . (:lb

164

RISK

Objective

BEST

WORST

GLOBAL
CURRENT

C:

Q

 

CURRENT

4=a1ter values

summary,

last step, 2= current activities, 3=

1:

0=save/quit,

FIGURE 5.7 BEST/WORST SOLUTION VALUES OF RISK AND RETURN. GOAL-DIRECTED SEARCH PROGRAM

A C T I V I T I E S

C U R R E N T

Acres
Planted

Crop

Land
Num.

Total
Acres

Tract
Description

0
)(Z)

‘2':
.. J

‘-\

CORN
SOYBEANS

65‘

John Wesley Hardin

u

"\
--

c
.

10

1 .31)

v (_I( I

WHEAT

WHEAT
SOYBEANS

N

Abraham Lincoln

150

WHEAT

M

I)

28¢

Tomas Borge

165

FIGURE 5.8 ACTIVITES IMPLIED BY CURRENT SOLUTION, GOAL-DIRECTED SEARCH PROGRAM

S

S T E P

R E V I O U S

U M M A R Y O F P

q
3

(
\

I
'-

RISF

REVENUE

NET

166

Worst

Best

Worst

Best

CH

3

2(

1000(

:00

150(

0

Step

FIGURE 5.9 SUMMARY OF RISK/RETURN VALUES FOR PREVIOUS STEPS, GOAL-DIRECTED SEARCH PROGRAM

167
this default or enter another more preferred value. The program then
creates and stores two files: 1) a record of the previous solution and
updated summary of previous solutions and 2) a file containing
commands to the LP solver to modify the current tableau according to
the preferences expressed by the decision maker and to re-solve the
problem. As described above, the resulting new solution is then read
by the GOALDIR program and displayed to the decision maker. This
process of modifying with GOALDIR and re-solving with the LP solver

continues until the decision maker is satisfied with the solution.

5.7 Preliminary Empirical Evaluation

Preliminary testing of the decision aid was carried out with
three extension agents and two farm operators to evaluate operational
features of the GOALDIR system. Individuals were visited at their
home or place of business, interviewed about the rental and crop mix
decisions, and asked to represent the farm operator/decision maker in
a session with the GOALDIR aid.

Extension agents were interviewed for several reasons. First,
their experience in observing and assisting farm operators to make
these decisions can provide valuable information concerning the nature
of the decision, the ability of farm operators to master the concepts
and procedures, and the likelihood that managers would incorporate
such an aid into their decisions. Agents would also be the
intermediaries in most manager use of GOALDIR, since operation of the
aid requires computer operations skills not common among farm
operators. Extension agents were also included because of the

unfortunate timing of field testing for GOALDIR. During spring

168
planting, few farm operators can afford the time necessary for the
aid, especially since the rental decision has already been made for
the current year.

The interviews with extension agents and farm operators are
presented as structured documentaries, with particular discussion of:
1) the operator’s opinion of the rental decision; 2) the
characteristics of the individual interviewed; 3) the data elicitation
procedure, including land, budget, yield expectation and price
expectation elicitation; 4) model operation, including reactions of
the individuals to the multiple objective framework and the
optimization/dominance procedures utilized in the decision aid. Other
factors of interest are the relative effectiveness of graphical versus
tabular data in elicitation and solution display, and the analyst~

mediated interface of the aid with the user.

5.7.1 Interview of First Agent

The first agent interviewed is from Barry County, in which a
significant proportion of farm operators rent cropland. In the
opinion of the agent, rental decisions are made infrequently (not each
year) when a relatively small number of properties become available
for rent. A landlord generally does not put properties on the rental
market is the current rentor wishes to continue the arrangement.
Stable access to resources (land) is often more important than current
year profits for the rentor.

The agent has skills and understands principles to an extent
uncommon to most county extension personnel. He has a good command of

budgeting principles, and understands major concepts of statistics and

‘I-fl

 

 

169
probabilities. For example, he understands the distinction between
modal and mean values, the symmetry/asymmetry characteristic of
probability distributions, and the concept of variance. His
thoughtfulness and inquisitiveness were demonstrated clearly by
unsolicited suggestions for improvement of aid procedures.

The reaction to interactive data elicitation procedures was very
favorable, particularly with respect to yield and price assessments.
The use of variable cost budget defaults from aggregate data was
considered a good mechanism to prompt operators for cost data and to
induce them to place attention on cost factors in cropping decisions.
He considered this procedure to be useful for purposes other than the
rental decision considered here.

Approximately twenty minutes were required to explain the
probability elicitation procedure to the agent. Subsequently,
elicitation of each probability distribution required approximately
ten minutes. Hypothetical price and yield distributions were elicited
for several crops. Although the agent had little difficulty with
probability elicitation procedures, his familiarity with probability
concepts stimulated a tendency to directly state probabilities for
yield/price intervals rather than the indirect conviction weights
requested by the aid. There were no obvious signs of bias in
probability assessment.

With a sophisticated subject such as this agent, direct
elicitation of probabilities is probably a more efficient elicitation
procedure. Graphical output of yield probability density functions
(PDF) was specifically mentioned as helpful for checking the symmetry

of elicited distributions against intuitive expectations. The

178

displayed PDF table was examined and easily interpreted with little
explanation from the analyst, but the cumulative distribution function
(CDF) tabular data was largely ignored.

This interview used an earlier version of the price distribution
elicitation procedure which displays a default CDF price distribution
and requests direct CDF subjective probabilities for price. Intervals
of the default price distribution varied, with equal probabilities in
each interval. CDF probabilities were regarded as both difficult to
express and to interpret. This procedure was not acceptable to the
agent, who suggested that the procedure used for yield elicitation
also be used for price elicitation. In subsequent interviews, price
elicitation was conducted using the conviction weights method.
Explanations and data elicitation required approximately two hours.

A previously constructed hypothetical farm situation was then
explained to the agent and solved with GOALDIR (the agent playing the
part of decision maker). The decision framework of tradeoffs between
multiple, conflicting objectives was intuitively reasonable to the
agent. Expected return as probability-weighted returns from random
draws as a current returns objective, and probability-weighted
deviations below a target value as a risk objective were seen as
acceptable operational forms of important objectives sought in farm
rental decisions. However, the intuitive definition of risk
considered most reasonable by the agent is the probability of returns
falling below variable costs (akin to focus-loss).

Considerable attention was paid to the graphical output as

contrasted to numeric output. The agent asked questions about model

171
structure and function, and understood some of the principles
underlying linear programming and the IMGP procedure. Whatever was
not understood about the optimization procedures was accepted on faith
as reasonable.

This agent can be viewed as the most promising potential
intermediary for GOALDIR use in extension assistance for land rental
decisions. He has substantive experience and knowledge, and a
propensity for inquiry which mark him as an innovator. His
inquisitiveness was demonstrated by the suggestions made during the
interview, some of which stimulated reformulation of the decision aid.
For example, he suggested yield probability elicitation should be
carried out before variable cost budgets were elicited, so that yield
goals and fertilization levels could be established and made
consistent. An unrelated fertilization decision application was
suggested for the yield elicitation procedures. It was recommended
that price and yield elicitation should be carried out with the same
conviction weights method. For the solution display procedure, risk
should be indexed from 1 to 180 instead of values of mean deviation
below target, since the latter are not intuitively recognizable to
farmers. Finally, he recommended a probability graph instead of a

table for the solution display procedure.

5.7.2 Interview of Second Agent

The second agent interviewed is from Ionia County. There is
considerable renting of parcels for cash grain farming in this county,
with some operations renting from as many as twenty landlords.

Considerable flux has been observed in rental arrangements in recent

172
years, including some instances of rent adjustment at harvest
conditional on actual yield and price. Modification of rental
arrangements with current landlords is much more common than switching
of farms rented from year to year.

This agent had more difficulty with probability concepts. For
example, he was unable to distinguish between modal and mean yields
without explanation. In addition, he was much less familiar with
computer hardware and software operations than the first agent.
However, the agent does have sound budgeting skills.

A favorable reaction was expressed concerning interactive
elicitation as a useful method to induce farm operators to identify
relative variable costs. The simple procedure of displaying default
variable cost budgets and eliciting farm-specific budgets was
considered to be very useful for planning assistance to farm operators
in combination with agent consulting. Yield and price probability
distributions were elicited for several crops with minimal difficulty.
The method was explained and illustrated in approximately fifteen
minutes, and each distribution was elicited in twenty minutes or less.
Little preference was expressed between display of PDF versus CDF
probability information. The histogram was used as a visual check on
correspondence between elicited probabilities and intuitive judgement
with respect to symmetry. The agent expressed the opinion that yield
probability distribution procedures would be useful in several other
contexts, such as fertilization decisions.

Solution of the hypothetical farm rental problem proceeded much
the same as with the previous agent. The concept of multiple

competing objectives was not difficult to grasp, but the agent’s

173

relative lack of understanding of probability concepts made the
operational definition of risk and return difficult to comprehend.
Confusion was obvious when the agent was presented the best/worst
solutions of GOALDIR. Solution of the problem by goal-directed search
seemed backwards to the agent. Preference was expressed for a similar
aid procedure which would allow direct selection of trial rental
plans. This agent also focussed his attention on graphical display of
goal tradeoffs, with less attention on numeric tables.

Explanation and application of data elicitation procedures
required approximately two hours. Another hour was required for

illustration of the solution display procedure.

5.7.3 Interview of Third Agent

In Gratiot county, there has been considerable flux in the rental
market as some farm businesses undergo severe financial stress. There
is considerable land rental in this county, and even in mid-May some
land had not yet been rented. The financial crisis has made land
available which would not normally have been on the rental market.
Some operators with financial difficulties have taken apparently risky
decisions by bidding high cash rents for large acreages, in hopes of
significantly improving their financial status.

It is unlikely that this agent understands probability concepts
well enough to instruct farm operators in formal risk analysis
procedures. Some difficulty was encountered in understanding concepts
such as expected value, modal value, variance and skewness while
explaining the probability elicitation procedures, but after a short

practice the agent apparently used the concepts in comparing the

174

elicited probability values with his beliefs. This agent also was
less familiar with the capacities and limitations of computers.

The agent made relatively few comments about land and budget
elicitation. Although pleased with the detail and flexibility of the
budget elicitation spreadsheets, he expressed the opinion that few
farm operators would be able to provide records upon which to base the
variable cost information requested. In his opinion, this indicated
two problems for successful implementation of the aid: 1) the
procedure would take considerable time to elicit budget information,
or 2) use of default budgets might seriously misrepresent the actual
farm situation. However, the data elicitation procedures were again
viewed as quite valuable to structure the decision and stimulate farm
managers to base their decisions on data.

Once the probability elicitation procedure was explained,
probability distributions of yields and prices for several crops were
elicited with little difficulty. However, most distributions were
relatively ’tight’, more symmetric and had higher minimum levels than
those of the other agents. For yield distributions, this might simply
be a real reflection of differences in soil and climate conditions.

It could also indicate potential bias in the elicitation technique.
However, questions to the agent concerning these points indicated that
subjective beliefs about variability of prices and yields were not
well formed, and that the elicitation technique seems to accurately
represent beliefs which are not well thought out. This indicates a
need to encourage the agent to focus carefully on his opinions about

the specific random process in question.

175

Framing a decision problem with multiple objectives was
reasonable to the agent, and he could name other decisions in which
farm operators pursue multiple objectives. Treating return and risk
as competing objectives was also reasonable to the agent. However,
the GOALDIR operational formulation of return and risk objectives was
not well understood, primarily because of the agent’s lack of
knowledge about probability. An intuitive definition of risk to this
agent reflects primary concern for avoiding low yields. He understood
the purpose of the aid as quantifying tradeoffs between the competing
objectives, and focussed principally on the graphics display to
determine desired relative changes in the objectives. No questions
were asked concerning the model structure or functional relationships.
In general, the agent was willing to accept on faith the validity of
the model procedures to determine best/worst objective function
values. However, a clear preference was expressed for alternative
procedures which would allow more detailed analysis of individual
cause and effect relationships. For example, he would like to examine
the effects on risk and return objectives caused by varying expected
prices of particular crops.

This session lasted over three hours. Two hours were spent in
demonstrating elicitation software and estimating probability

distributions for three crop yields and three crop prices.

5.7.4 Interview of First Operator
The first farm operator is a partner with several other family
members in a large (6000 acres) cash grain operation. The crops

raised are corn, wheat, soybeans and dry beans. Beside cropping

176
enterprises, the partnership also operates a certified seed operation
and has close business connections with other non-farm firms.

This operator is an unusually well educated (university degree)
and knowledgeable operator with over four decades of farming
experience. His management style is innovative and dynamic, as
demonstrated by the steady growth and diversification of business
operations. Although he was somewhat less familiar with and
knowledgeable of probability concepts than the agents, he was more
comfortable with computer operations than two of the three agents.

Although the session lasted three hours, considerable time was
spent discussing the rental decision process in general. This
partnership rents no land on a yearly basis. Instead, all land is
leased for at least three years on a cash basis. There is an
agreement with most landlords to adjust the rental price according to
harvest time cash price. The partnership rents twelve farms. The
operator considered that in his area very few farms are available for
rent each year. If a parcel is available, the operator usually makes
the rental decision on the basis of labor availability, government
base acreage of the land, and the price outlook for the next several
years. It should be noted that most important decisions in the
operation are made with the participation of all active partners.
Therefore, communication between partners and discussion based on
facts are essential when decisions are made.

This operator had no clearly defined concept of business risk,
although from his perspective risk for the operation has its source
primarily in unexpected interest rate increases and secondarily in

price fluctuations. The operation uses forward pricing to protect

17?

itself from price fluctuations. The concept of a target income level
was quite natural to the operator, and he was able to state a value
almost immediately.

Not enough time was available to elicit complete budget and
rental information, but the templates were demonstrated and some
budgets elicited. The business keeps quite detailed records, and the
operator was able to give very precise variable cost information from
memory without difficulty.

The operator expressed considerable interest in the probability
elicitation procedure, asking for a copy for his operation. Only a
few minutes were necessary to explain the process, and after the first
distribution was elicited the operator became quite comfortable with
the procedure. Somewhat more time was necessary to obtain each
distribution than with the extension agents. Because of limited time,
yield distributions were elicited for three crops on only a single
soil type. Price distributions were also elicited for three crops.

The operator decided by himself that the interval range would
affect his estimates, and asked that an interval range chosen by the
researcher be reduced. There was some evidence from the operator’s
comments that conviction weights stated for the distribution tails may
have been confused with historical frequencies. It appeared somewhat
more difficult to the operator to anchor weights on the modal range as
elicitation proceeded to the distribution tails. Reminding the
subject that the weights should be stated relative to the most likely
interval should be done at frequent intervals so that errors are not

committed. There was also a tendency to smoothly decrease conviction

178
weights as elicitation progressed from the modal range towards the
tails. This may reflect use of an ad hoc smoothing rule to reduce
cognitive effort, or may simply reflect imprecise beliefs. The
operator had difficulty with probability concepts, and was not able to
interpret the difference between the most likely and expected values
of distributions. More extensive explanations of simple probability
concepts using graphics procedures would likely be rewarded by more
understanding of terms and concepts and more carefully considered
subjective probabilities. The only conclusions that can be made with
confidence are that the operator found the procedures easy to master,
that he agreed with the resulting probabilities, and that those
distributions were similar to distributions which could be calculated
from historical data.

Evaluating decisions based upon tradeoffs between competing
objectives was not difficult for the operator. Although the
operational definition of returns as used in the model was understood
by the operator, he had difficulty understanding the definition of
risk. The land rental decision may be primarily considered a
production decision, and risk associated with land rental appears to
be closely associated with expectations of low yield. Price risk may
not be explicitly considered in rental decisions, probably because of
relative price stability provided by government feed grain programs.

The operator had considerable difficulty in understanding the
decision process of choosing an implied set of activities according to
explicit tradeoffs between risk and return. The best/worst solutions
presented in the aid appeared to confuse the operator, and did not

improve comprehension of model operation. The graphical output was a

I79

focus of tradeoff decisions, with relatively little attention to

numeric output.

5.7.5 Interview of Second Operator

The second operation visited is a single family, mixed crop-
livestock farm. Sugar beets, dry beans, soy beans and corn are raised
on the owned land and five rented farms. Although there have been
several opportunities to rent additional land, the family has not
changed rental properties in the past three years. Individual rental
parcels are evaluated by estimating expected returns given various
rental arrangements which might be negotiated with the landlord.

One son has assumed primary responsibility for the livestock
operation, and the record keeping is handled jointly by the husband
and wife. Several family members participate in important decisions,
and a multi-person management unit can be identified.

The operator’s time was very limited, and relatively less data
was elicited than with the previous operator. He did not find the
probability elicitation procedure difficult, but comments indicated
that the elicited distributions may not have been careful or accurate
expressions of belief. The concept of a target income was reasonable
to him, as were the risk and return concepts used in the model.
However, search by objectives was definitely not a comfortable
procedure. The operator stated a clear preference for a procedure
which would allow him to select farm plans and then calculate the

impacts on risk and return objectives.

188
5.7.6 General Observations

Certain observations common to several interviews should also be
mentioned. The agents were in general more comfortable with
production issues rather than the economic issues treated in the
decision aid. This is not surprising, since all are agricultural
production specialists. Although the agents interviewed were selected
for their reputations as innovators, it is unlikely that they are
qualified without further training to instruct farmers in risk
analysis and associated probability concepts. Neither are they
sufficiently comfortable with computer operations to instruct
operators in use of programs of more than moderate complexity.

Tabular probability values were not well understood by most
subjects, and verbal summaries of the information such as, 'According
to your stated estimates, there is one chance in five of corn price
between $1.70 and $1.80 at harvest time ', were generally necessary to
interpret the tabular data. A tendency to ’smooth’ conviction weights
for some intervals from the modal range towards the distribution tails
was observed in some cases. As noted above, it is unclear whether
this procedure accurately reflects beliefs, or is instead a technique
to reduce the difficulty of carefully examining beliefs for each
interval.

Some subjects determined historical frequencies for yields within
some intervals, which then were transformed to express the desired
conviction weights. Operators pay close attention to the random
weather processes which affect yield variability, and appear to have
more carefully considered expectations concerning yields. This may

indicate that direct elicitation of yield probability distributions

181

may be more efficient in some cases. It is likely that direct
elicitation would be more successful with operators who keep
production records and who have some familiarity with probability
concepts, and for crops for which historical frequencies provide a
reasonable estimate (to the operator) of expected yields in the
current year.

Display of a default set of conviction weights for prices was
expected to serve the purpose of a distributional reference for
subjects with less clearly defined expectations for prices than for
yields. On the other hand, blanket acceptance of the default weights
may severely misrepresent the subject’s price expectations. The
interviews indicated that default weights did influence the elicited
weights to some degree, but that subjects substantially modified the
weights in accordance with their own beliefs. Although all subjects
were encouraged to examine probabilities resulting from expressed
conviction weights and to alter distributions which did not ’seem
right’ in their opinion, conviction weights were not modified in any
interview. This could be attributed to lack of interest or ill-
defined underlying beliefs. In any case, it tends to undermine the
advantages of interactive elicitation procedures, since convenient
modification was expected to be a major benefit of the interactive
process.

Partial rental of a parcel was indicated in the hypothetical
problem, but all subjects rounded or truncated solutions with little

comment. Although it is entirely possible that such a practice could

182
lead to decisions which are not optimal, it appears to create few
operational problems for users.

Distinguishing risk and return as competing objectives was an
intuitively reasonable framework for all subjects. There was evidence
that crop rotation and soil fertility objectives are also important
factors in some operators’ rental decisions. However, quantification
of objectives confuses some subjects. They are often more comfortable
with ill-defined, intuitive formulations of objectives other than cash
returns. Most subjects did not express full confidence in the
suggested solution activities because they did not understand the
procedures used to calculate solutions. Only the first agent inquired
about the structure and functional relationships of the model, and he
subsequently displayed a fuller understanding and more confidence in
the solutions. Development of user confidence in complex procedures
appears to be more difficult if the procedures are not in some way
similar to existing decision practices. Graphic representations were
carefully considered by all subjects in probability elicitation and
solution evaluation, suggesting that many farm operators may find
tabular data more difficult to interpret and utilize in these types of

tasks.

CHAPTER U1

CONCLUSIONS AND PROPOSED RESEARCH AGENDA

6.1 Introduction

Keen and Scott Norton (1978) describe development of decision
support systems as an iterative process, with repeated phases of
design, implementation, evaluation followed by subsequent
reformulation. The present research has examined issues of decision
theory and decision support, designed and implemented a multiple
objective decision aid, and conducted preliminary field testing with
extension agents and operators. One decision support development
cycle has been completed, and an assessment must be made of what has
been learned from the literature, model development and field testing.
A decision support research agenda is derived to continue the
difficult task of expanding farm management researchers’ capability to

assist management decision making.

6.2 Conclusions from the Literature

Rationalist and behavioralist theories of individual decision
making are in many ways alternative explanations of the same empirical
phenomena. The former paradigm attempts to develop context-
independent and objective decision making principles, while the latter
describes actual decision processes from the perspective of
subjective, cognitive elements. While rationalist decision theory
claims to provide the framework for determining the best possible

183

184
solution within the domain of defined problems (and hence the action
that should be taken), behavioral theory attempts to reveal the
largely hidden cognitive operations (what the individual is thinking
during decision making). The rationalist perspective has relatively
little consideration for the cognitive processes in decision making,
while the behavioralist perspective provides no clear framework or
standard for evaluating decisions.

Characteristics of the human information processing system have
been described here, and limitations with respect to logical decision
processes have been examined. Behavioralist research has documented
persistent violations of rationalist assumptions in a wide range of
decision contexts with subjects of varying skill, education and
experience. Selective attention is paid to multiple objectives and
multiple dimensions of choice alternatives. To avoid the time and
complexity of finding the best alternative, individuals often choose
alternatives with values that are only satisfactory. Decision rules
are utilized which often guarantee sub-Optimal outcomes, and errors
are common in forming and manipulating judgemental beliefs. The
general conclusion is inescapable that unaided individual decision
making is prone to errors vis a vis rationalist principles because of
the innate structure and operation of the cognitive apparatus.

But the behavioralist camp begs one question and leaves another
hanging. Individuals generally do not make completely unaided
decisions. They utilize a wide range of memory, representational and
computational aids, and thus to an unknown extent avoid serious
errors. Behavioral theory is largely silent on the issues of ’person-

tool rationality’. Although a satisfactory basis for examining

185

defects in existing decision processes, behavioral theory forms a very
incomplete framework for designing procedures to aid decisions. The
objective of management support is not only to detect biases, but to
aid individuals to make better decisions. On the other hand,
rationalist theory is often criticized as not providing a practical,
workable basis for aiding decisions. Procedures recommended under the
rationalist framework frequently assume data, computational and time
resources which are not available to decision makers.

A working framework for decision support involves a synthesis of
concepts and techniques from both decision paradigms. Hybrid decision
aids would utilize rational decision procedures to the extent possible
within the documented characteristics and constraints of human
cognition. Similarly documented strengths of human information
processing such as pattern recognition, inferential judgements and
long term memory recall should be supported by such decision aids.

Multiple criteria decision models (MCDM) provide a mathematical
structure for decision aids reflecting both behavioral and rationalist
principles. All reflect the consideration of multiple dimensions of
decision alternatives, yet provide powerful solution techniques
utilizing either dominance or optimization procedures. Interactive
multiple criteria models allow further opportunities for developing
greater user confidence and learning about problem structure and
preference tradeoffs. Interactive MCDM methods require articulation
of only local preferences and iterate towards preferred solutions.
Since the model is directed by user preferences, proposed solutions

may have a better chance of being implemented.

186
Of the currently available interactive MCDM methods, the
Interactive Multiple Goal Programming (IMGP) method best embodies many
features of a desired hybrid analytical model. Articulation of local
preferences, use of aspiration levels and flexible consideration of
various functional formulations and decision rules are characteristics

which recommend IMGP as a mathematical model for a decision aid.

6.3 Conclusiggs from Model Development

A two objective risk-return aid was developed on a microcomputer
to assess feasibility and potential of IMGP-based aids. The principal
conclusion is that interactive multiple objective modeling is
feasible. However, limitations of current microcomputer software and
hardware force severe compromises between realistic modeling of
decision problems and timely operation of decision aids. The data and
functional form requirements cause the resulting optimization models
to press the limits of current microcomputer technology. It is
expected that these restrictions will be alleviated as faster and more
powerful microprocessors become available. However, it is unclear
whether both faster operation and more realistic modeling can be
accomplished with the next generation of technology.

A related design problem from both the analytical and programming
perspectives involves user-specific and default data. Use of
’representative’ data in the decision aid greatly simplifies model
construction and improves turnaround time, but may seriously
misrepresent the decision as perceived by the user. Compromise from
the standpoint of aid design was reached by eliciting data that a farm

operator might have in farm records or about which the operator is

187
likely to have relatively well-formed opinions or expectations. Other
data which the user could not be expected to provide were imposed from
external, representative sources. Although the effects of these
compromises on software development is tangible, effects on model
solutions are unknown.

The LP solution procedure upon which the IMGP model is based
depicts alternatives as infinitely divisible. Solutions may be
indicated with fractional acreages or farms, an obviously unrealistic
answer to the real problem. Although a mixed integer integer
algorithm could be used to alleviate this problem, mixed integer
programs are significantly slower (as much as fifteen times slower)
than LP programs. Thus, size and realism of the decision model would
again have to be traded off against speed of model operation. In
addition, the IMGP procedure chosen for this implementation requires
substantial modification for problems not characterized by a convex
feasible set.

Simple graphical representations were utilized in the decision
aid. Plans for more complex graphics were abandoned because of
hardware-dependence of most graphics procedures, relatively slow
operation and difficulty in integrating graphics routines with other
software. However, both graphical and optimization procedures will
become much more practical for real time decision support as faster
microprocessors (as much as twenty times faster) and more
sophisticated software (with complex but flexible ’toolkit’ modules)
become more widely available.

Integration of the various functional routines that together

constitute the decision aid is a serious obstacle to cost-effective

188

development and implementation. This research used two generic
commercial programs for elicitation and LP solver functions. Few such
commercial programs are designed to be integrated into a larger
software system, and these were no exception. Particular problems
arise in passing data from one routine to another and in developing a
reasonably consistent interface between the aid and the user. The
former problem required rather intensive analyst input to operate the
aid, while the latter problem was not resolved. An alternative design
procedure is development of custom software for each decision aid,
which unfortunately increases development costs and often does not

approach the power and speed of commercial software.

6.4 Conclgsions from Field Testing

Testing of the GOALDIR aid provided empirical evidence with
respect to several decision-specific and general decision aid issues.
Validity of the decision model, data elicitation procedures, multiple
objective solution procedures and intermediary-user issues are

discussed below.

6.4.1 The Rental Decision

Correspondence between the decision model and the real world
rental decision is limited by necessary approximation of functional
relationships and data measurements. In addition, characteristics of
the real problem may differ from the model structure with respect to
time horizon, relevant objectives, and organizational context. That
is, the rental decision is not generally a portfolio analysis of all

owned and potentially rented cropland, but is considered by farm

189
operators as an evaluation of isolated rental tracts. Rather than
evaluating rental arrangements on the basis of current year expected
returns, operators make formal or informal multiple year committments
for particular properties on the basis of expected returns over a
longer time horizon.

In addition, rotational objectives should be treated explicitly
in a rental decision model. Some operators will refuse to depart from
a strict rotational cycle regardless of return prospects. Others are
willing to plant a significant proportion of cropland according to
expected relative returns. This behavior could be considered in a
multiple objective decision aid.

Although this research has examined individual decision making,
many farm businesses have multi-person management units. Decision
aids could be designed as forums for discussion and decision making by
relevant individuals in the firm. Formalization of the problem and
interaction between differing individual perspectives may

significantly improve decisions.

6.4.2 Data Elicitation

The data elicitation for a farm-specific model consumes much more
time than initially estimated. Elicitation of rental arrangements and
government program information for ten to fifteen potentially rentable
parcels may require an hour. Another hour may be required for
elicitation of variable cost budgets, particularly if such budgets are
specified for all possible crops on each parcel. At some point,
differences between variable cost budgets no longer have a significant

effect on solutions, and the time requirements of additional specific

19B
variable cost budgets are no longer worth the effort. Gross figures,
on the other hand, may destroy the very subtlety desired in modeling
the rental decision.

Much the same conflict between specific data and time-effective
data elicitation applies to probability elicitation procedures. At
least one-quarter to one-half hour should be used to carefully explain
the elicitation process and the specific random variable to be
measured. Although not part of the procedures in this analysis,
historical yield data used as a benchmark will probably facilitate
more reliable encoding of probabilistic beliefs. If records are
available and the procedure is well understood, each probability
distribution can be elicited in less than fifteen minutes. If yields
are elicited for each possible crop on each possible parcel, the total
. number of elicited distributions quickly multiplies. For example,
three crops which can each be grown on fifteen parcels requires forty
five distributions. Again it is unclear at what point additional
investment in probability elicitation efforts no longer has a
significant effect on the model solution.

At a minimum, it seems that a yield distribution should be
elicited for each crop on each major soil type. However, one yield
distribution per soil type combined with a price distribution for each
crop still implies a significant amount of time consumed in price and
yield elicitation. With only three crops and two major soil types, at
least two hours could be required for this stage of elicitation,
suggesting that the total time necessary for data elicitation may
easily exceed four hours. It is questionable whether farm operators

would be content to ’feed’ data to the model for that amount of time

191

without receiving any usable results or feedback. Investigation could
be carried out to determine the effects of various aggregation or
smoothing procedures to provide quick estimates rather than such
intensive data elicitation. Perhaps if the elicited data could also
be used in other decision procedures, such as fertilizer application
decisions, the additional time might be more justifiable.

Critical issues related to probability elicitation include the
reliability (freedom from random error) and validity (accurate
representation of beliefs) of expressed probabilities. It is clear
from this research and other field tests that the conviction weights
procedure is operational. That is, probability distributions can be
elicited with relative ease from individuals who have little knowledge
of probability concepts. However, there is no documented evidence
that such distributions elicited with the conviction weights method is
free from random error or is an accurate representation of beliefs.
Before much confidence can be placed in an aid purporting to support
decisions, questions of reliability and validity of competing
elicitation techniques must be addressed.

Within the context of this research, it is not possible to
determine whether display of default budgets and price distributions
helped subjects to more accurately express farm-specific data than
would otherwise have been possible. The willingness of subjects to
disagree with default price distributions elicited from a university
price analyst was somewhat surpising. This behavior indicates that
subjects consider the default information but express their own

judgements.

192

Simple graphics representation of probability distributions was
used by the subjects to compare beliefs with skewness implied by the
elicited probabilities. A graphics-based conviction weights procedure
would allow the user to manipulate bar graphs on a computer monitor
rather than to express numeric weights. It is possible that a
completely graphics-based elicitation procedure may be well accepted
by operators, which could provide a more efficient and conceivably
more accurate mapping of beliefs.

Decision aids to some extent will require farm-specific data. If
records are not available or if operator expectations must be
measured, interactive data elicitation is feasible and may contribute
to more accurate measurement of desired variables. Uisual feedback in
probability elicitation procedures may stimulate more careful
consideration of beliefs and expectations, and may have benefits which
extend beyond the current decision. Although data elicitation was
implemented with a spreadsheet program for this research, better
software integration would require database functions usually

available only with custom software.

6.4.3 Multiple Objectives and Optimization

Explicit incorporation of multiple objectives and optimization
solution procedures are related decision support issues. As discussed
in Chapter 4, evidence exists that farm operators seek multiple,
sometimes incommensurable and often conflicting management objectives.
The interviewed agents and operators reacted favorably to a multiple
objective model framed as tradeoffs between risk and return

objectives. However, the specific objective formulations were not

193
well understood. The risk objective in GOALDIR is defined as
probability-weighted deviations below the elicited target income
level. Most subjects, however, expressed the Opinion that a more
intuitive of risk is the probability of returns below target, akin to
Roy’s (1952) ’safety-first’ formulation.

Although subjects also reacted favorably to explanations of the
target value concept, it is unclear whether preferences are accurately
represented by the type of mean-risk below target model described by
Fishburn (1977). Further questions involve the reliability of
elicitation procedures for target values and the effects of errors in
estimating target values. Can the operator state without assistance a
specific value which reflects an abrupt change of preference for
returns? If record keeping is poorly handled in the firm, the current
decision aid assume that the operator has a clear subjective estimate
of factors such as fixed costs and living expenses which may be used
to formulate a target return value. The effects of estimation errors
in target values on model solutions are also not well understood.
Clearly, solution values in risk-return are altered by changes in
target values (Tauer 1983), as are underlying activity variables.
Tauer derives a set of simple crop plans while incrementing the target
return by equal amounts from a benchmark value. Solutions for which
expected return is constrained vary significantly from plans derived
using a different target. The effect on returns of changes in target
values results from the interaction of constraint functions and the
’hard’ risk constraint, and does not appear amenable to analytic

examination. Sensitivity analysis of target values in the LP

194
formulation is not possible because the target value appears in
several constraint functions.

As noted above, crop rotation and multi-year return objectives
could also be formulated and explicitly incorporated in multiple
objective models. Some operators may find that such an expanded model
constitutes a significantly improved depiction of the real decision,
while others may prefer to consider only current year risk-return
objectives. Much simpler formulations are possible with only one or
two objectives, but only the researcher’s judgement can determine
whether such a formulation reasonably describes the preference
structure. On the other hand, simultaneous consideration of many
objectives in a decision may violate information processing
limitations of the decision maker. Structuring decision aids as
tradeoffs between multiple objectives is reasonable to decision
makers, but research must be conducted to examine the impacts of
decision aids formulated with different numbers and types of
objectives.

A closely related set of issues concerns the solution procedure
of the decision aid. IMGP restricts choice to a subset of the
feasible region bounded by efficient values of the relevant
objectives, with best and worst values used to construct reference
solutions. The procedure forces the user to select preferred
tradeoffs between objectives, while alternative actions are implicit
in model solutions (a ’top-down’ approach). These characteristics
place IMGP in the class of optimization (or dominance) models, and
demonstrate the power of rationalist procedures to direct users

towards unambiguously better solutions in terms of the model

I95

objectives. However, agents and operators found the process confusing
and counter-intuitive.

Clear preferences were expressed for solution procedures which
would permit selection of possible actions and subsequent evaluation
of feasibility and consequences (a ’bottom-up’ approach). GOALDIR
utilizes an optimization procedure to incorporate powerful rationalist
principles for decision support, but the least intelligible concept
for the subjects was the top-down approach of the aid. This may be
attributable to the subjects’ slight understanding of the model
structure, functions and implied causal relationships, or lack of
confidence that the model adequately represents the real problem. The
top-down approach as applied to some ill-structured decisions may also
violate the manager’s sense of control over the decision process. If
key aspects of model structure and function are not understood and
accepted, aids which allow design of individual alternatives may be
preferred by managers.

Depending upon decision maker and problem characteristics, aids
based on optimization procedures may not provide significant
assistance for making better decisions. Further, optimization-based
aids rarely provide a framework for user learning about the problem as
described by Langley and Simon (1981, see Section 3.6). Perhaps the
most effective contribution of decision aids is to be found in
techniques to assist in achieving better comprehension of problem
structure and effective search of the problem space, not in

application of decision rules.

196
Graphics representations were beneficial for display of model
solutions generated by GOALDIR. Subjects focussed intently on
tradeoffs displayed as bar graphs, and paid relatively little
attention to the objective function values. When pattern recognition
or display of relative changes in values are sought in decision aids,
graphics appear to convey information very efficiently and are

preferred in many cases by operators.

6.4.4 Context of Management Assistance

None of the managers interviewed would be able to operate the aid
without assistance, nor could agents serve as intermediaries for
operation of the aid in its present form. Training for users and
intermediaries in the technical and decision principles embodied in
the aid may be a precondition for successful use of GOALDIR or any
sophisticated decision aid. In the context of agricultural decision
aids, most intermediaries are likely to be extension agents, credit
officers or commercial technicians. Given current computer and
decision skills of agricultural managers, it appears that the most
beneficial applications of decision aids will be to complement other
management assistance activities of technical specialists.

Many researchers contend that unassisted use of decision aids by
the manager can increase comprehension and confidence in solutions,
but little evidence exists on this issue. What does seem obvious is
that unassisted operation is a more cost-effective approach when the
scarce resource is the qualified management advisor. As personal
computers become less expensive and more widely disseminated in the

agricultural community, opportunities for assisting managers with

197
decision aids will increase. Training and aids for learning computer
operations and decision principles will be required before it can be
expected that managers will utilize complex decision aids without

assistance.

6.5 Elements of a Decision Support Research Agenda

Decision support research has to this point been almost
exclusively problem- and action-oriented. Relatively little
generalizable research has been conducted to examine interactions
between characteristics of decisions problems, decision makers and
decision tools, and consequent effects on chosen actions. The final
sections of this chapter present issues which should be investigated
in experimental and operational settings in order to accelerate
progress in developing successful aids to assist managers.

Isolation of features and examination of their impacts is
necessary to obtain generalizable research results for decision aid
design. A number of experimental designs can be developed which focus
on analysis of particular decision support issues. In order to
provide valid generalizations to support farm manager use of decision
aids, experiments must be designed which at some stage employ managers
as subjects. This is not to say that other individuals such as
college students should never be used, but the relevance of such
results is open to question until confirmed in a farm context. In
order to assess the effects of individual personality characteristics,
managers with clearly distinct problem solving styles should be
selected as subjects. The interaction of decision type with the

feature investigated should generally be examined by testing effects

198
across a range of decision types (although our knowledge of decision

taxonomy is admittedly poor).

6.5.1 Probability Elicitation

Probability elicitation in some form will be a feature of many
decision aids, since uncertainty at the firm level is typically poorly
reflected in aggregate estimates. Lack of adequate historical records
at the farm level often means that managers must provide relatively
unsupported judgements of critical price and production parameters,
which highlights the critical importance of the elicitation procedure
to accurately encode the beliefs of the individual. A further
management issue concerns the consistency of probability assessments
with objective data.

Behavioral researchers have provided extensive evidence of errors

in unaided probability judgements, but little research has examined
probability elicitation aided by decision tools. Principal errors in
assessing probabilities have been related by Kahneman et al. (1982) to
heuristics used to reduce information processing requirements.
Hogarth (1988) documents biases in probabilistic thinking, but claims
that the most important human error observed in probability
experiments is lack of consistency. If rules used to estimate
uncertainty were applied consistently, decisions would be improved on
average. Another particularly relevant observation is that estimates
of probabilities differ depending on the elicitation method.
Techniques vary in their capability to accurately measure beliefs.

Several questions must be answered with respect to probability

elicitation with farm managers:

199

1. What are the relative operational advantages of competing
elicitation techniques in a particular probability
assessment?)

2. How do techniques compare in terms of reliability, that is,
are probability encodings repeatable, stable and consistent
(Wallsten and Budescu 1983)?

3. How do techniques compare in terms of validity, that is, are
they valid measurements of beliefs?

4. A broader and somewhat separate issue concerns whether
probability estimates are ’correct’. Although the
traditional concept of subjective probability allows no
unambiguous determination of ’right’ or ’wrong’ judgements,
from a management support perspective individuals can
certainly be assisted to learn probability concepts and
integrate information in such a manner that estimates are
more accurate when evaluated after outcomes are known.

The current research has documented an implementation of the
conviction weights technique, a relatively simple indirect procedure
for eliciting probabilities. However, the elicitation issues
discussed above can be addressed only by careful experimentation over
time and between methods. An evaluation of information from farm
records can also be used to assess the consistency of elicited
distributions with historical data. A sample of farm managers should
be identified who keep sufficient records to provide a basis for
analyzing external validity of elicited probabilities. Principal
competing elicitation techniques should be applied with each manager

to assess probabilities for particular variables. Repeated interviews

203

should be conducted over a reasonably long time period to determine
the replicability and stability of assessments with each method. The
benchmark used to evaluate the accuracy of such distributions would be
detrended estimates obtained from farm records. Individuals would
also be asked to rank and evaluate operational features of the
techniques. The most accurate assessment technique may not be
reasonable for management support if its application is tedious or

other features are objectionable to managers.

6.5.2 Preferences

The simple objectives of risk and return have been assumed in the
GOALDIR aid to be the only preferences of the decision maker in regard
to rental decisions. Certainly this does not seem to be a completely
unwarranted assumption. However, for decision support purposes, an
examination should be made whether more efficient use of managers’
time in decision making is possible by modeling a problem with one
objective or a range of potential objectives. Deterministic LP models
used as decision aids simplify problems by assuming profit
maximization as the single objective, and sometimes achieve
considerable success by highlighting solution sensitivity and
bottlenecks or opportunities. Yet no one would suggest that the
decision maker seeks only a single objective. It is often unclear
that more complex models add significantly to problem comprehension or
to better quality decisions.

Section 4.2 discussed the relevant research on multiple
objectives in farm management. Much of the research consists of

elaboration of relatively general objectives gathered from previous

281

research or researcher experience followed by relative ranking of
objectives by the manager-subject. Relatively few attempts have been
made to extend this research to examine the impact of multiple
objectives in the context of specific decisions (exceptions are more
common in developing country agricultural research, e.g. Barnett et
al. 1982). The social and psychological issues involved in analysis
of multiple objectives of farm operators have received relatively
little attention. A principal reason is that, IThe setting for
evaluating multiple goals is complex. It introduces a broader range
of concepts than is commonly used in economic analysis, many of which
are not designed to yield optimal or equilibrium conditions.”
(Robison et al. 1984: 28)

As noted by Robison et al. (1984), aspiration levels and time
significance of objectives are also important features of preferences
relevant to particular decisions. Before systematic examination can
be made of the effects of multiple objectives for decision aids,
descriptive research should be carried out to examine the ordering and
composition of objectives sought by managers in important decisions.
However, there is no tradition of descriptive research within the farm
management field which examines individual managers at the level of
detail necessary to describe specific decision processes and examine
the complex effects of multiple objectives.

There is some descriptive research in related fields. In
corporate management research, Mintzberg et al. (1976) examined
decision processes by enlisting managers to keep detailed diaries of

activities, and researchers spent long periods (as long as five years)

282
gathering information about a single firm’s decision processes. They
were able to identify common procedures in decision making which could
then be used as general references for intervention in management
assistance for other firms. Bennett et al. (1982) studied farm
management from a social perspective in longitudinal research lasting
seventeen years. Adaptation to economic conditions was related to
multiple values and objectives were affected by social relationships.

This type of longitudinal research is rare in farm management
research principally for reasons of theoretical perspective and cost
considerations. However, the methods of other social sciences could
be adopted to better describe the objectives which motivate farm
managers in particular decisions. The long term research effort of
investigating managerial processes on individual farms would be
rewarded by details of the objectives sought by managers in specific
decision types over time. Many questions concerning the
identification of good managers could be answered with more detailed
descriptive information relating how good managers make their
decisions. This can best be investigated in the format of
longitudinal observation and interviews.

Does multiple objective modeling make any difference? Dobbins
and Mapp (1983) selected six goals for a recursive goal programming
model examining intergenerational transfers. The order of multiple
objectives (goals) was found to significantly affect production plans.
Barnett et al. (1982), on the other hand, concluded that their
representative farm planning problem provided no evidence that
multiple objective modeling was a significantly more accurate

predictor of farm organization. Few research projects have attempted

283

to investigate whether the quality of decisions made by individuals is
affected by the assumed preference structure of the decision model.
Management game experiments in a workshop environment would appear to
be the best framework for investigation of these effects. Comparative
evaluation can be made of single objective models, risk (two
objective) models and more complex multiple objective models in a
range of constructed problems. Given that there is no way to know a
’best’ answer without information about individual preferences,
evaluation would principally involve the relative demands of models on
decision makers. Individuals would rank the procedures in terms of
ease of use, information requirements, time requirements, usefulness

of model output and confidence in the model.

6.5.3 Solution Procedures

Simon (1983) presents three ways to represent problem solving

tasks:

1. §ggpgg, in which the focus concerns examination of the
current state and selective, successive application of
operators to move to another state until a solution is
found.

2. Reasoning, in which a system of logic allows deduction of
new statements from axioms and previously deduced
statements. Thus, solving a problem consists of

accumulating logical statements until the outcome is found.

284
3. Constraint satisfaction, in which a set of alternatives is
defined by constraints. Solving the problem consists of
narrowing down the set of alternatives that satisfy all
constraints.

Simon indicates that these metaphors for problem solving are not
mutually exclusive, and any problem solving algorithm may be viewed
'. . . now as search, now as reasoning, now as constraint
satisfaction.” (1983: 7) The IMGP procedure can be examined with any
of the metaphors. The reasoning of the optimization procedure depicts
the accumulation of inductive and deductive knowledge about the farm
planning problem, the constraints applied (including both ’hard’
resource constraints and ’soft’ goal constraints) allow giant steps
towards solutions by eliminating whole classes of alternatives, and
the analysis of a current state determines the search direction for
the next step. Other decision models, such as regression models and
stochastic budgeting, can be seen to reflect somewhat different types
of problem solving metaphors. These models may not define problem
solving tasks in a manner similar to the unaided decision maker’s
representation. Fundamental differences in problem representation and
problem solving procedures may induce rejection of the decision model
by the user. The particular question at issue here is the
dispreference expressed by the individuals interviewed for the top-
down solution procedure utilized in the decision aid. The more
general question, of course, concerns the conditions under which
optimization procedures are appropriate for decision aids.

Three factors which seem to influence acceptance of and

confidence in the solution procedure of an aid are: 1) the model

285

representation versus the unaided representation; 2) decision maker
control of the procedure; and 3) opportunities for learning. These
are examined below with reference to the current aid.

First, the decision aid assumes that the optimization model
reasonably reflects the structure of the real world problem. There
appear to be several plausible reasons why this may not be the case.
The aid assumes that all objectives and relative preferences are well
defined (even if unknown) before the decision aid session. The
relevant variables, specific values of those variables, causal and
functional relationships, and the extent of detail in those
relationships are all factors which determine the model and real
problem structure. If the individual understands the model but does
not agree with the model representation, confidence is unlikely to
develop. In addition, the decision maker and the model designer are
often unlikely to be in complete agreement on the representation of
uncertainty in the problem, or possible sequential choices, or the
extent of decomposition necessary for decisions to be made. Any of
these problems can occur even though the decision maker understands
the model structure. The particular aid developed here has not been
explained in any detail to decision makers, and thus cannot claim to
be understood by them. Rather than rejection of model structure, this
application was characterized principally by ignorance of structure.
This also is unlikely to instill confidence. However, even if it were
understood, problems could be expected in decision makers’ acceptance

of such factors as the operational definition of objectives and the

286
restricted functional relationships dictated by the research objective
of developing a microcomputer-based aid operating in real time.

The second factor which affects confidence in a decision model is
the extent of user control. Since managers must bear responsibility
for decisions, they demand justification for relinquishing control of
the process to any decision model. An optimization model, with its
focus on finding solutions implied by objectives, may not at first
seem controllable to the decision maker. The normal decision process
of intelligence, design and choice is automated in considerable part
by the IMGP model, and the decision maker performs only the functions
of data input and potential solution evaluation. The process is so
different from the intuitive decision process that lack of control
instills a subsequent lack of confidence. For the decision maker to
feel in control of the model, 1) the model structure and procedures
must be understood to some extent and accepted by the user (as would
be the case with a professional user), 2) the model must have a track
record for the user, or 3) the model can be accepted on faith. Since
none of these conditions were satisfied in the current application,
the lack of user control was probably the principal cause for
rejection of the optimization procedure. It could be said that one
cannot control or have confidence in something that is not understood,
but this is not strictly true. Most individuals in the United States
have considerable (perhaps too much) confidence in their ability to
control an automobile, even though they understand very little of the
mechanical and physical principles involved. Experience and faith

seem to be very adequate substitutes for understanding in this case.

287

A final reason why optimization procedures are rejected by
decision makers is the inherent diminished possibility for learning.
It can be expected that inquisitive managers wish to learn facts,
principles and procedures in one decision which may serve them in
other, possibly unrelated decisions. This is particularly true if the
decision must be made repeatedly or is not a ’life or death’ decision.
Six principles defining the conditions under which learning can take
place were discussed in Chapter 3. They are: knowledge of results,
generation of alternatives, causal attribution, hindsight (feedback),
learning from instruction and automatization. It can be seen that
optimization procedures such as used in the GOALDIR aid provide
relatively weak conditions for learning which can be applied to new
decision situations, particularly with respect to generation of
alternatives, causal attribution and hindsight. Inquisitive decision
makers who wish to gain new skills rather than merely solve the
current problem may not place great confidence in the aid.

This feature of decision aids can be tested in a similar manner
to that suggested for the multiple objective issue. To ensure valid
comparison of contrasting procedures, risk programming and stochastic
budgeting techniques should be compared in prepared problems presented
to the manager-subjects. Managers would be assigned randomly to
groups which make individual decisions with either the budgeting or
programming aid. To provide a benchmark for evaluation, expected
utility functions should be elicited prior to problem solving.
Resulting solutions could be compared in terms of utility costs.
Comparisons should also be made of the time necessary for competing

techniques and the relative computational costs of the models.

288

Finally, a user evaluation of features such as described for the

multiple objective issue should be undertaken.

6.5.4 Graphics

A prevalent feature of decision support techniques for management
is graphical representations to communicate information and to
summarize and analyze data. Because of information processing
limitations in cognition, behavioral researchers (among others) have
been strong proponents of graphics use in decision support.
Recognition, memory, interpretation accuracy and comprehension are
sometimes improved by visual stimuli. Responding to this argument,
decision aid designers have increasingly incorporated graphical
capability in their products.

Recently some researchers have suggested that, “. . . the
extravagant claims favoring graphic presentation formats may be
considerable overstated.I (Ives 1982: 21) There is now some dispute
whether users interpret graphic formats more accurately than tabular
data and whether graphics improves the quality of decisions.
DeSanctis (1985) reviews research on the role of graphics in human
information processing. Some support is found for the premise that
the ability to use graphics effectively is related to individual
characteristics or cognitive style. Information search and memory
recall seems to function differently if graphics displays are used.
However, for a range of dependent variables such as interpretation
accuracy and speed, recall and user confidence, there is little
support for better performance with graphics than with tabular data.

There is virtually no research on the effects of different graphic

289

types or features on decision making. DeSanctis (1985) proposes that
future research investigate the interaction of graph characteristics,
individual user characteristics and decision context in their effect
on decision strategy, problem comprehension and decision making.

There appears to be no research examining the relative advantages
of graphic displays on farm management decision making, despite
considerable effort to include such capability in decision support
software. Some evidence would seem necessary to justify this effort,
since real costs are involved in designing such features. In
addition, if both graphic and other data representations are displayed
in a decision aid, the cumulative result of cluttered displays may be
to actually diminish the effectiveness of decision making.

At a minimum, two types of experiments can be designed to examine
the issue of graphics representations. First, the specific question
of ’graphics versus tables’ can best be investigated with a gaming
experiment based on a standard design described by Remus (1984). A
production scheduling problem is designed with implied costs for
changes in workforce level, idle time and excess inventory. Optimal
rules for scheduling are a function of the previous period’s workforce
size and inventory and forecasts for future periods. Each subject
makes repeated scheduling decisions based upon either tabular or
graphic data. The difference in costs incurred by tabular or graphics
users can then be compared. A similar design could be applied to farm
managers in a workshop environment. A marketing game which requires

storage and pricing decisions based on inventory levels and price

218
forecasts could be designed. Optimal marketing rules would provide a
benchmark for establishing decision costs.

Another analysis should examine wider issues of graphics support
for farm manager decision making. Particular issues which should be
addressed are the impacts of different graphics features such as color
and form on comprehension, and preferences of decision makers for
graphics features in varying problem contexts. Experiments can be
also designed which test comprehension and which measure relative

preference for alternative graphics features.

6.5.5 Evaluation

Relatively little is to be learned in terms of effective decision
aid design from evaluation research which only attempts to measure the
impacts of an aid on the decision process of a particular individual
in a unique decision situation. A decision aid cannot be regarded as
a homogeneous entity, but instead is a composite of design features,
each of which makes a contribution to the observed consequences. Some
of these features may be related to decision theory principles - the
combination of beliefs (probabilities) and preferences (utilities),
optimization techniques and the like. Other features may be more
closely related to the behavioral issues discussed in this research -
multiple objectives, graphical display, aspiration values or flexible
decision rules. Evaluation of the impacts of a decision aid thus
examines effects of the specific combination of features embodied in
the aid. This combination may not be relevant or feasible in another

situation for which a decision aid is designed.

211

Characteristics of the decision or of the individual also
interact with those of the support technique. For example, Davis and
Olson (1985) suggest that graphical display has beneficial impacts for
inventory decisions. The same techniques may not be effective in
decisions for which pattern recognition or data summarization are a
less important part of the decision process. In terms of personal
characteristics, the cognitive or problem solving style of the
individual also interacts with characteristics of the aid. An
analytical thinker may wish to use an optimization technique while a
more intuitive thinker may prefer to ’hunt and peck’ for acceptable
solutions. Similarly, the substantive knowledge and problem solving
skills of the individual will interact with and have considerable
effect on the decision process.

Results from analysis of the separate decision support issues
discussed above can provide a thorough basis for design and evaluation
of decision aid features. Examination of individual decision features
in a variety of decision types and with subjects of differing problem
solving style is a more justifiable and generalizable research
perspective for decision support. Nevertheless, evaluation
methodologies must be improved to determine benefits of particular
decision aids. The complexity of such evaluations has hindered
development of decision support principles. Most decision support
researchers simply recommend a ’smorgasbord’ of evaluation
methodologies ranging from user assessment of aid features to
traditional cost-benefit analysis (Keen and Scott Morton 1978). Most
decision support evaluations have consisted of case studies with a

very small group of subjects. Hobbs (1985) contends that decision

212

support research (and multicriteria research in particular) has
suffered from ill-controlled experiments and poorly constructed
frameworks for interpreting results. He states that ’ideal’ MCDM
experiments have the following characteristics:

1. Methods are tested which represent features of distinctly
different approaches to decision making or are claimed to
represent significant methodological improvements

2. Well-controlled experiments are designed and implemented
with a large sample so that treatment-effect relationships
are unambiguous

3. Methods are tested across a range of problem types

4. Experiments are designed to be as similar as possible to
real decisions with respect to problem characteristics and
motivational features

Of course, no single experiment is likely to satisfy all criteria
simultaneously. Nevertheless, they are appropriate guidelines for
designing decision aid evaluation.

Relative advantages and disadvantages of alternative decision
aids is emphasized by Hobbs (1985). The value of a decision aid can
only be evaluated with respect to alternative feasible procedures.
Alternatives may be unaided decision processes or may be other aids
developed with a different perspective for support. The contrast
offered by simulation and optimization procedures is an obvious
example of possible comparative evaluation of decision aids. Other
examples can also be developed which compare and evaluate distinctly

different decision aids for the same problem.

APPENDICES

APPENDIX A

GOALDIR MODEL VARIABLES AND EQUATIONS

A.1 Variables Derived Outside the Model

 

GMrh =

REV"h - VCh

(Revenue per acre minus variable costs per acre for crop h in
state of nature r.)

REV"h =
YLDrh X PRICEPh
(Expected yield per acre times expected price for crop h in state
of nature r.)

YLD"h =
Draw from multivariate normal yield distribution, described by
elicited expected values and standard deviations, and covariance
matrix constructed from external data. Yields are estimated for
optimal planting/harvesting period klx, and adjusted by a penalty
matrix to reflect the effect of timeliness on yield.

PRICEPh =
Draw from multivariate normal price distribution, described by

elicited expected values and standard deviations, and covariance

matrix constructed from external data.

VC
Estimated variable cost per acre for crop h. Variable cost is
sum of direct cash costs including fertilizer and chemicals,
fuel, lubrication and repairs, seed, hired labor, irrigation

fuel, drying and hauling.

213

214

PLTh =
Estimated hours/acre for planting operations of crop h with
available equipment, calculated as implement width times speed
times efficiency factor.

HARh =
Estimated hours/acre for harvesting operations of crop h with
available equipment, calculated as implement width times speed
times efficiency factor.

PLT-HOURSK =
Estimated hours available in period k for planting operations,
such that the probability that actual hours falls below this
level is less than .2.

HAR.HOURSk =
Estimated hours available in period k for harvesting operations,
such that the probability that actual hours falls below this
level is less than .2.

TARGET =

Operator-specified target return level.

215
A.2 Model Variables

The following section defines the variables of the GOALDIR model.

A.2.1 Activity Variables
ACRESiJkl:
Acres of crop h on tract i, feed grain program status j, planted

in period k and harvested in period 1.

Y :
r
Shortfall of total net revenue below target for state of nature r.
EXP-DPi:
Expected deficiency payment for farm i.
RENT.:

1

Cash rental payment for farm i.

A.2.2 Objective Function Coefficients
EXP'GMhijkl:
Expected gross margin of crop h on tract i, feed grain program

status j, planted in period k and harvested in period 1, in

dollars per acre.
Number of sample years or states of nature.

A.2.3 Technical Coefficients

PLT :
h Expected field time necessary for planting operations of crop h,

in hours per acre.

HAR

 

TAR:
com

} NHEM

 

 

216
HARh:
Expected field time necessary for harvesting operations of crop
h, in hours per acre.
GMrhijkl:
Gross margin of crop h on farm i, feed grain program status j

planted in period k and harvested in period l for state of nature

r, in dollars per acre.

A.2.4 Constraints

LANDiJ:
Total tillable acres of farm i, government status j.

PLT_HOURSK:
Estimated field time available in planting period k, such that
the probability that actual field time is less than this value is
.2 or less, in hours.

HAR.HOURS,:
Estimated field time available in harvest period 1, such that the
probability that actual field time is less than this value is .2
or less, in hours.

TARGET:
Target gross margin, in dollars.

CORN.BASEi:
Corn base acreage for those farms in government program.

WHEAT-BASE1:

Wheat base acreage for those farms in government program.

217

A.3 Multiobjective Rental/Crop Mix Model

MAX 2222..
ijk

EXP_GM..
.1 ijl

, X ACRES .. + EXP.DP. - RENT.
.l hij i i

kl
Ch= 1..NUM.CROPS}

MIN 2 Yr / s

subject to:

XXXZth' ACRESijkl L LANDi

{i= 1..NUH-FARMS}

XXXhiJ (PLT i ACRESiJkl) i PLT_HOURSK

{k=1..NUM-PLT}

XXXhiJ (HAR X ACRESijkl) i HAR_HOURSI

{I=I..NUM.HAR}

XZXZXhiJk‘ (GMrhijkl X ACRESiJk] +

X. EXP-DP. - X. RENT. + Y 2 TARGET
l i l l r

Cr= 1..s}
22k] ACREShijkl i CORN_BASEi
(h = <corn))
' = 1..NUM.FARMS,
j = 1,3)
22k! ACREShijkl £ WHEAT.BASEi
(wheat)

{h
' 1..NUM_FARMS,
1,3)

J

APPENDIX B
SPREADSHEET TEMPLATES FOR GOALDIR DATA ELICITATION WITH
PROTOCOLS FOR YIELD AND PRICE ELICITATION

8.1 Protocol for Yield Elicitation

The subject is first presented with screen 1 (Figure 8.4), which
displays a template for identification information, yield ranges and
conviction weights. The interviewer selects with the subject a
reasonable starting yield (not necessarily zero) and interval range
which are calculated and displayed by the program. The total range
should be wider than the subject is likely to consider possible.

Interviewer- 'For this particular farm/soil type, and assuming

that you were able to plant and harvest in a timely manner, in

which of these ranges would yield be most likely to fall?‘

When the subject indicates a range:

Interviewer- “I’ll give that range a score of 188. Now for each

of the other yields that are possible here, I’d like you to put a

number from 1 to 188 indicating how likely that yield range is

compared to the most likely range. A yield range that is just as

likely would have a score of 188, while one that is only 1/18 as

likely would be given a score of 18. You might want to think of

it this way - if you always planted the same varieties, for every

ten years that you might get yield in the most likely range, you

would expect six years of yields in the range which you gave a

score of 68. If you prefer, let’s set the scores of the highest

and lowest possible yields first. You may never have gotten

218

219

these yields in years past, but what we’re looking for is what

yields you think are possible to get in very bad or very good

conditions. How likely is this low (high) yield compared to the
most likely yield?‘

(Subject gives weights for lowest and highest ranges).

“Now let’s fill in the scores between the lowest (highest) yield

and the most likely yield. For each yield range, tell me a score

from 1 to 188 that represents how likely you think that
particular yield is compared to the most likely range."

(Subject fills in scores).

After all weights are filled in to the subject’s satisfaction:

Interviewer- 'Do you consider yields above the most likely range

just as likely as yields below?II (If the response is not

consistent with weights given, then point this out to the subject
and adjust weights.)

It should be repeatedly emphasized to the subject that weights
can be changed at any time the subject determines that they do not
reflect their judgements about likelihood of yields in the range.
Screen 2 (Figure 8.5) is then displayed to the subject:

Interviewer- 'Based on the scores you gave for each yield range,

the program has calculated probabilities for each yield. In

other words (indicating the probability density function),
according to your scores these numbers represent the chances out
of 188 that yields will fall in each of these yield ranges. The

graph (indicating the graph) shows one asterisk for each 1

percent of probability, and the numbers here (indicating the

 

228

cumulative table) show the probability that yields will fall
below the upper limit of that yield range.”

Interviewer - l'Let’s look first at the graph. The shape
indicates in relative terms how likely are yields above the most
likely yield as below. The fatter it is above (below), the more
likely you’ve indicated yields are above (below) compared to
below (above). Does that general shape seem right to you?“ (If
not, then return to screen 1 to adjust weights.)
Interviewer-'Now, according to the scores you gave, there is
approximately a 1 in (truncate 1.8 divided by modal probability)
chance of getting the most likely yield. Does that seem right?"
(If not, then return to screen 1 to adjust weights.)
Interviewer-“Chances are approximately (truncate 18 times sum of
probabilities of 2-3 most probable ranges) out of ten of getting
between (low boundary of lowest interval) and (high boundary of
highest interval). Does that seem right?‘ (If not, then return
to screen 1 to adjust weights.)

Interviewer-'There are approximately (truncate 18 times sum of
probabilities of 1-3 lowest intervals)) chances out of ten of
falling below (high boundary of highest of these intervals).
Does that seem right?‘ (If not, then return to screen 1 to
adjust weights.)

Interviewer-'On the other end, there are approximately (truncate
18 times sum of probabilities of 1-3 lowest intervals)) chances
out of ten of getting yield above (low boundary of lowest of
these intervals). Does that seem right?‘ (If not, then return

to screen 1 to adjust weights.)

221

Interviewer-'If we weight your yields by the probabilities, the
weighted average yield is (expected yield). Does this seem right
to you? Are any of the probabilities out of line with what you
would expect?‘ (If so, return to screen 1 and adjust weights. If
the subject accepts the probabilities as correct, the weights,
yield interval and starting point are saved, and the process
continues to the next crop or soil type/farm. Note that the
standard deviation and coefficient of variation are not displayed

or stated to the subject.)

8.2 Protocol for Price Elicitation

Screens 1 and 2 (Figures 8.6 and 8.7, respectively) are virtually
identical, except that the only identification information necessary
is the name of the crop. For screen 1, default price ranges and
weights are displayed to the subject which are elicited from a price
analyst. It is stated emphatically to the subject that everyone has
their own beliefs about expected prices, and that they need not agree
with the default values. If the subject thinks the total range should
be broader than the default, new ranges are calculated, but the
interval width should remain equal so that the default weights are not
misleading.

Subsequent questions and explanations are identical to yield
elicitation, except that probabilities are always expressed in

”chances out of ten' rather than in "years out of ten'.

222

Land Information

Name:

mum-'---====m-mm-=fi—m-====m==nnmn—w======m=====
Farm No. :

Farm Description 1
Principal Soil Type
Total Acres

Total Cultivable Acres :
Rented? :
Rent per acre 0 I
Feed Grain Program? 1
Corn Base Acreage 1
Corn Base Yield 1

Wheat Acres 1

-===:-==============--— '
- — n.- '

 

Completed for each rented or owned farm.

FIGURE 3.] LAND ELICITATION TEMPLATEx

223

CASH EXPENSE PER ACRE FOR CROP:

Name:

Soil:

Farm(s):
Typical Your
Expense Cost

Fertilizer

Chemicals

Fuel, lubrication, maintenance
Drying

Hauling

Hired Labor

Total =)

FIGURE 8.2 VARIABLE COST ELICITATION TEMPLATE--SCREEN 1

 

224

 

 

CORN SOY WHEAT NAVY OATS

19.6 8.4 18.8 28 6.4

27.1 7.9 37.3 11.5 9.7

15.2 18.15 2.35 14 .65

32.9 19 23 26.1 24.4

38 8.9 8 8 8

21 6.3 12.6 5 6.6

8 8 8 8 8

8 8 1.5 1.5 1.5

—__ 145.8 68.65 87.55 78:1 49.25
i Budgets are not displayed until costs for the crop are elicited.

BARLEY

13

22.6

.65

24

8

12.6

CORN
SILAGE

17.5

38.3

14

14

FIGURE 8.3 VARIABLE COST ELICITATION TEMPLATE--SCREEN 2 X

Crop: -

Yield Unit: -
Farm No.: -
Soil Type: -

Interval: -

FIGURE 8.4 YIELD PROBABILITY ELICITATION TEMPLATE-‘SCREEN 1

226

Yield Probability That Yield:
Range In Range Equal or Less Than Range

 

- Standard
- Deviation

- Variation

- Expected
_ Value

FIGURE 3.5 YIELD PROBABILITY ELICITATION TEMPLATE--SCREEN 2

227

 

Crop:
Name:
Price Default Your
Range Weights Weights
Interval: -

FIGURE 8.6 PRICE PROBABILITY ELICITATION TEMPLATE--SCREEN 1

228

Price Probability That Price:
Range In Range Equal or Less Than Range

 

- Standard
- Deviation

- Coeff. of
- Variation

‘ Expected
_ Value

FIGURE 8.7 PRICE PROBABILITY ELICITATION TEMPLATE-“SCREEN 2

APPENDIX C
PASCAL PROGRAM LISTING--MULTIVARIATE NORMAL RANDOM NUMBER GENERATOR

Procedure Mvnorm;
(Generates DRAWS vectors of VARIATES number of
correlated, jointly normal distributed variables
in the matrix MULTI_RND.
This program written by Dr. J. Roy Black,
Department of Agricultural Economics, Michigan
State University.)

Const
variates = 58; (number of random variates desired)
draws = 35; (number of vectors of random variates)
Type
all-arrays = arrayIl..variates] of real;
sqr.mat = arrayIl..variates] of all-arrays;
rnd.mat = arrayII..draws] of all_arrays;
Var
x,means : all-arrays;
c,covar : sqr.mat;
multi_rnd : rnd_mat;
i,j : integer;

Procedure Get-data(var covar: sqr.mat;
var means: all-arrays);
(XThis is a dummy procedure. Any data input procedure to enter
the COVAR covariance matrix and MEANS mean values array can be

used!)
begin
end;
Procedure Ltm( covar : sqr.mat;
var c : sqr.mat);

(COVAR is covariance matrix, C is the calculated
lower triangular matrix used to compute random variates)

Var

sum : real;
i,j,k : integer;

229

238

Begin
fillchar(c,sizeof(c),8); (Initialize matrix)

(first diagonal element equal to square
root of corresponding covariance element)
c[1,l] := sqrt(covar[1,11);
for j := 2 to variates do
(rest of lst col. equal to
corresponding COVAR element divided
by 1st diagonal element)
cIj,11 := covarIj,1] / c[1,11;
(this loop determines lower triangular)
( elements of columns > 1)

for i := 2 to variates-l do

(diagonal elements from 2 to 2nd to last)
(equal to the square of corresponding)
(COVAR element of that row minus)
(the sum of squares of that row)
= 1 to i-1 do
= sum + sqr(cIi,k]);
cIi,iJ := sqrt(covarIi,iJ - sum);
:= 8

(off-diagonal elements of C equal to
(corresponding COVAR element minus the
(product of elements J and i of column k,)
(all divided by diagonal element)
for j := i+1 to variates do
begin
for k := 1 to i-I do

sum = sum f cIj,kJ X cii,k];
cIj,iI := (covarlj,i] - sum) / c[i,i1;
end;
sum := 8;

(last diagonal element equal to square root}
(of the corresponding COVAR element minus)
(sum of squares of off-diag. elements)
for k := 1 to variates-1 do
sum := sum + sqr(c[variates,k]);
civariates,variates) := sqrt(covar[variates,variates] - sum);

end; (procedure Ltm)

231

Procedure Normal ( c : sqr_mat);
(computes one vector of variates)
means : all-arrays;
var x : all-arrays);
Const
twoxpi ' 6.283185;

Var
i,j : integer;
21,22 : real;
2 : all_arrays;
Begin

22 := random;
for i := 1 to variates do
(computes VARIATES number of standard normal elements)
begin
21 := random;.
in1 := sqrt(-2.8 X ln(zl) X cos(twoxpi X 22));
22 := 21;
end;
x := means;
for i := 1 to variates do
(computes variates number of normal,)
(correlated elements)
for j := 1 to i do

inI := inJ + cIi,jJ X szl;
End; (procedure normal)

Begin (main procedure)

get-data(covar,means);
(dummy data entry procedure not shown here)
Ltm(covar,c);
(create lower triangular matrix)
for i := 1 to draws do
begin
Normal(c,means,x);
(create one sample vector)
multi-rnd[i] := x;
(pass each vector to a DRAWS by VARIATES matrix)
end;

end; (Procedure Mvnorm)

REFERENCES

 

REFERENCES

Abelson, R. 1976.
“Script Processing in Attitude Formation and Decision Making“, in
J. Carroll and J. Payne (eds.). Qggpition gnd Socigl Behavior.
Hillsdale, New Jersey: Erlbaum.

Ackoff, R. 1974.
'Beyond Problem Solving". Decision Sciences 5: x-xv.

Anderson, J. 1988.
Cognitive Psycholooygand Its Implications. San Francisco:
Freeman.

Anderson, J., J. Dillon and B. Hardaker. 1977.
Agricultural Decision Applysis. Ames, Iowa: Iowa State
University Press.

Bar-Hillel, M. 1973.
'On the Subjective Probability of Compound Events'.

Organizational Behavior and Human Performance 9(3): 396-486.

Barnett, D., 8. Blake, and B. McCarl. 1982.
“Goal Programming via Multidimensional Scaling Applied to
Senegalese Subsistence Farms'. American Journal of quicultural
Economics 64: 728-727.

Bartos, J. 1969.
"The Assessment of Probability Distributions for Futur- Security
Prices“. Ph.D. Dissertation, Indiana University.

Beach, L. and T. Mitchell. 1978.
'A Contingency Model for the Selection of Decision Strategies".
Academy of Management Review 3(3): 439-449.

Beenen, F. 1978.
“Psychiatric Diagnosis and Subjective Probability". Acta
Psychologica 34: 328~337.

Bell, D. 1982.
"Regret in Decision Making”. Opergtions Research 38: 961-981.

Bellman, R. and L. Zadeh. 1978.

'Decision Making in a Fuzzy Environment”. Management Science
17(4): 141-164.

232

233

Bennett, J. 1983.

Building Decision Support Systems. Reading, Massachusetts:
Addison-Wessley.

Bennett, J., S. Kohl and G. Binion. 1982.

Of Time and the Enterprise. Minneapolis: University of Minnesota
Press.

Berkson, J., T. Magath and M. Horn. 1948.
'The Error of Estimate of the Blood Cell Count as Made with the
Hemocytometer'. American Journal of Physiology 128: 389-323.

Bettman, J. and J. Jacoby. 1975.
"Patterns of Processing in Consumer Information Acquisition".
Papers in Consumer Psychology No. 158. West Lafayette, Indiana:
Purdue University.

Bigelow, J. (ed.). 1887.
Thg Complete Works of Benjamin Franklin. New York: Putnam.

Blaug, M. 1988.
The Methodology of Economics. Cambridge: Cambridge University
Press.

Black, J. R. 1974.
'Farm Planning Guide: Corn vs. Beans.“ Department of
Agricultural Economics Staff Paper 74-12, Michigan State
University, East Lansing, Michigan.

Bodily, S. 1985.
Modern Decision Making. New York: McGraw-Hill.

Bond, N. 1974.
'Basic Strategy and Expectation in Casino Blackjack“.
Organizational Behavior and Human Performance 12: 413-428.

Boussard, J-M. and M. Petit. 1967.
'Representation of Farmers’ Behavior Under Uncertainty With a

Focus-Loss Constraint'. Journal of Farm Economics 49(4):
869-888.

Candler, W. and M. Boehlje. 1971.
“Use of Linear Programming in Capital Budgeting with Multiple

Goals". American Journal of Agricultural Economics 53(2):
325-338.

Cary, J. and W. Holmes. 1982.
'Relationships Among Farmers’ Goals and Farm Adjustment Strategies:
Some Empirics of a Multidimensional Approach". Australian
Journal of Agricultural Economics 26(2): 114-138.

Castle, E., M. Becker and F. Smith. 1972.
Farm Buainess Management. New York: Macmillan.

234

Chankong, V., Y. Haimes and J. Thadathil. 1985.
“Multiple Criteria Optimization: A State of the Art Review”, in
Y. Haimes and V. Chankong (eds.), Decision Makingfiwith Multiple
Objectives. New York: Springer-Verlag.

Charnes, A. and W. Cooper. 1961.
Management Models and Industrial Applications of Linear

Programming. New York: Wiley.

 

Charnes, A. and R. Nilhaus. 1968.
'A Goal Programming Model for Manpower Planning". Management
Science Report 115. Pittsburgh: Carnegie-Mellon University.

Checkland, P. 1985.
"From Optimizing to Learning: A Development of Systems Thinking

for the 1998s'. Journal of the Operations Research Society 36(9):
757-767.

Cohen, J. 1962.
'The Statistical Power of Abnormal-social Psychological Research".
Journal of Apnormal Social Psychology 65: 145-153.

Cohon, J. 1978.

Multiobjective Programming and Planning. New York: Academic
Press.

Coombs, C. 1975.
'Portfolio Theory and the Measurement of Risk.‘ in M. Kaplan and S.
Schwartz (eds.), Human Judoamentaand Decision Processes. New
York: Academic Press.

 

Davis, G. and M. Olson. 1985.

Management Information Systems: Conceptual Foundations, Structure
and Development. New York: McGraw-Hill.

Day, R. and T. Groves (eds.). 1975.
Adaptive Economic Models. New York: Academic Press.

DeFinetti, B. 1974.
Theory of Probability: A Critical Introduction. Translated by A.
Machi and A. Smith. New York: Wiley.

DeSanctis, P. 1985.
''Computer Graphics as Decision Aids: Directions for Research".
Decision Sciences 15: 463-487.

Dillon, J. 1988.
'The Definition of Farm Management. Journal of Agricultural
Economics 31(2): 257-258.

Dobbins, C. 1978.
"The Intergenerational Transfer of the Farm Firm: A Recursive Goal
Programming Analysis“. Ph.D. Thesis, Oklahoma State University.

235

Dobbins, C. and H. Mapp. 1983.
'A Recursive Goal Programming Model of Intergenerational Transfers
by Farm Firms', in K. Baum and L. Schertz, Modeling Farm
Decisions for Policy Analysis. Boulder, Colorado: Westview.

Drake, A., R. Keeney and P. Morse. 1972.
egalysis of Public Systems. Cambridge, Massachusetts: MIT Press.

Ebbesen, E. and V. Konecni. 1988.
“On the External Validity of Decision-Making Research: What Do We
Know About Decisions in the Real World?', in Wallsten, T. (ed.),

Cognitive Processes in Choice and Decision Behavior. Hillsdale,
New Jersey: Erlbaum.

Edwards, W. 1968.
“Conservatism in Human Information Processing' in B. Kleinmutz

(ed.), Formal Representation of Human Judgement. New York:
Wiley.

Einhorn, H. and R. Hogarth. 1981.
“Behavioral Decision Theory: Processes of Judgment and Choice”.
Annual Review of Psychology 32: 53-88.

Ellsberg, D. 1954.
'Classic and Current Notions of ’Measurable Utility”. Economic
Journal 64: 528-556.

Elster, J. 1979.
Ulysses and the Sirens. Cambridge, Massachusetts: Cambridge
University Press.

Evans, J. and R. Steuer. 1973.
'A Revised Simplex Method for Linear Multiple Objective

Programming'. Mathematical Programming 5(1): 54-72.

Fernandez, A. 1982.
I'Conjoint Analysis: An Application to the Importance of Farmers’
Goals“. M.S. thesis, Department of Agricultural Economics, Purdue
University.

Fischoff, B. 1983.
'Decision Analysis: Clinical Art or Clinical Science?', in L.
Sjoberg, T. Tyszka and J. Wise (eds.), Human Decision Makino.
Doxa, Sweden: Lund.

Fischoff, B. and R. Beyth. 1975.
I"I Knew It Would Happen’ - Remembered Probabilities of Once-
future Things“. Organizational Behaviopaand Human Performance
13: 1-16.

Fishburn, P. 1965.
“Independence in Utility Theory with Whole Product Sets".

Operations Research 13: 28-45.

 

236

Fishburn, P. 1977.
I'Mean-Risk Analysis with Risk Associated with Below-Target
Returns". American Economic Review 67(2): 116-126.

Fishburn, P. 1978.
'A Survey of Multiattribute/Multicriterion Evaluation Theories",
in Zionts, 3. (ed.), Multiple Criteria Problem Solving. New York:
Springer-Verlag, pp. 181-224.

Flinn, J., S. Jayasuriya and C. Knight. 1988.
I'Incorporating Multiple Objectives in Planning Models of Low-

resource Farms'. Australian Journal of Agricultural Economics
24: 35-45.

Friedman, M. and L. Savage. 1948.
'The Utility Analysis of Choices Involving Risk". Journal of

Political Economy 56: 279-384.

 

Gardner, H. 1985. I
The Mind’s New Scienge. New York: Basic Books.

Gasson, R. 1973.
I'Goals and Values of Farmers“. Journal of Agricultural Economics
24: 521-537.

Geoffrion, A. 1968.
'Proper Efficiency and the Theory of Vector Maximization'.
Journal of Mathematical Analysis and Applications 22(3): 618-638.

Geoffrion, A., J. Dyer and A. Feinberg. 1972.
'An Interactive Approach for Multi-criteria Optimization with an
Application to the Operation of an Academic Department'.

Managgmant Sgience 19(4): 357-368.

Georgescu-Roegen, N. 1954.
“Choice, Expectations and Measurability". anrterly Journal of
rEconomics 68: 583-534.

Glass, A., K. Holyoak and J. Santa. 1979.
Cognition. Reading, Massachusetts: Addison-Wesley.

Goodman, D. 1974.
''A Goal Programming Approach to Aggregate Planning of Production
and Workforce'. Management Science 28(12): 1569-1578.

Goicoechea, A., D. Hansen and L. Duckstein. 1982.

Multiobjgctivg Decision Analysis with Engineering and Business
Applications. New York: Wiley.

Goldberger, A. 1964.
Econometric Theory. New York: Wiley.

237

Goodman, D. 1974.
'A Goal Programming Approach to Aggregate Planning of Production
and Workforce". Management Science 28(12): 1569-1578.

Gordon, L., D. Miller and H. Mintzberg. 1975.
“Normative Models in Managerial Decision Making“. National
Association of Accountants Publication Number 7578. New York:
National Association of Acountants.

Gorry, G. and M. Scott Morton. 1971.
'A Framework for Management Information Systems“. Sloan Management
Review 13: 55-78.

Green, P. and Y. Wind. 1973.
Multiattribute Decisions in Marketing. Hindsdale, Illinois:
Dryden.

Grether, D. 1988.
“Bayes Rule as a Descriptive Model: The Representativeness
Heuristic“. anrterly Journal of Economics 95(3): 537-558.

Grether, D. and C. Plott. 1979.
“Economic Theory of Choice and the Preference Reversal
Phenomenon.“ American Economic Review 69: 623-638.

Hammond, K. 1975.
I'Externalizing the Paramaters of Ouasirational Thought“, in M.
Zeleny, (ed.), Multiple Criteria Decision Making. Berlin:
Springer-Verlag.

Harper, W. and C. Eastman. 1988.
'An Evaluation of Goal Hierarchies for Small Farm Operators”.
American Journal of Agricultural Economics 62: 742-747.

Harsh, 3., L. Connor and G. Schwab. 1981.

Managing the Farm Business. Englewood Cliffs, N.J.: Prentice-
Hall.

Hatch, R., W. Harman and V. Eidman. 1974.
'Incorporating Multiple Goals into the Decision-making Process: A
Simulation Approach to Firm Growth Analysis". Southern Journal of

Agricultural Economics 6(1): 183-118.

Hazell, P. 1971.
“A Linear Alternative to Ouadratic and Semivariance Programming
for Farm Planning Under Uncertainty”. American Journal of

Agricultural Economics 53: 664-665.

Heady, E. and H. Jensen. 1954.
Farm Management Economics. Englewood Cliffs, New Jersey:
Prentice-Hall.

238

Hershey, J., H. Kunreuther and P. Schoemaker. 1982.
“Sources of Bias in Assessment Procedures for Utility Functions".
Managment Science 28(8): 936-954.

Hillier, F. and G. Lieberman. 1988.
Introduction to Operations Research. San Francisco: Holden-Day.

Hobbs, B. 1985.
“Experiments in Multicriteria Decision Making and What We Can
Learn From Them: An Example“, in Y. Haimes and V. Chankong (eds.),
Decision Making with Multiple Objective . New York: Springer-
Verlag.

Hogarth, R. 1981.
“Beyond Discrete Biases: Functional and Dysfunctional Aspects of
Judgemental Heuristics“. Psychological Bulletin 98: 197-217.

Hogarth, R. 1988.

Judgement and Choice: The Psychology of Decision. Chichester,
New York: Wiley.

Huber, G. 1983.
“Cognitive Style as a Basis for MIS and 033 Designs: Much Ado
About Nothing“. Management Science 29: 567-579.

“Decision Support Systems: Their Present Nature and Future
Applications“, in G. Ungson and D. Braunstein, Decision Making;

An Interdisciplinary Inggiry. Boston: Kent Publishing.

Huber, G. 1988.
Managerial Decision Making. Glenview, Illinois: Scott-Foresman.

Humphreys, P. and W. McFadden. 1988.
“Experiences with MAUD: Aiding Decision Structuring Versus
Bootstrapping the Decision Maker“. Acta Psychologica 45: 51-69.

Hwang, C. and A. Masud. 1979.
Multi le b'ective Decision Making - Methods and Appligations.
New York: Springer-Verlag.

Ignizio, J. 1976.
Goal Programming_and Extensions. Lexington, Massachusetts:
Heath.

Ijiri, Y. 1965.
Management Goals and Accgunting for Control. Amsterdam: North-
Holland.

Ives, B. 1982.
“Graphical User Interfaces for Business Information Systems".
MIS Quarterly Special Issue, pp. 15-47.

239

Janis, I. and L. Mann. 1977.
Decision Making; A Psycholooigal Analysis of Conflict. Choice and
Commitment. New York: Free Press.

Johnson, E. 1977.
“The Perception of Tactical Intelligence Indications: A
Replication“. Technical Paper 282, U.S. Army Research Institute
for the Behavioral and Social Sciences, Alexandria, Virgina.

Johnson, G. 1977.
“Some Lessons from the IMS.“ Paper presented at the Agricultural
Development Council-CIMMYT conference on Risk and Uncertainty in
Agricultural Development, El Batan, Mexico.

Johnson, G. 1957.
“Agricultural Economics, Production Economics and the Field of

Farm Management“. ournal of Farm Economics 39(2): 441-458.

Johnson, G. and A. Halter. 1961.
“Introduction“, in A Study of Managerial Processes of Midwestern
Farmers. G. Johnson, A. Halter, H. Jensen and D. Thomas (eds.).
Ames, Iowa: Iowa State University Press.

Johnson, G., A. Halter, H. Jensen and D. Thomas (eds.). 1961
A Study of Managerial Processes of Midwestern Farmers. Ames,
Iowa: Iowa State University Press.

Johnson, L. and D. Loucks. 1988.
“Interactive Multiobjective Planning Using Computer Graphics“.
Computers and Operations Research 7: 89-97.

Jolly, R. 1983.
“Risk Management in Agricultural Production“. American Journal of

Agricultural Economics 65(5): 1187-1113.

Jungermann, H. 1988.
“Speculations about Decision-theoretic Aids for Personal Decision

Making“. Acta Psychologica 45:7-34.

Kahneman, D., P. Slovic, and A. Tversky. 1982.
Judgement Under Uncertainty: Heuristics and Biases. Cambridge,
Massachusetts: Cambridge University Press.

Kahneman, D. and A. Tversky. 1979.
“Prospect Theory: An Analysis of Decisions Under Risk.“
Econometrica 47: 263-291.

Kahneman, D. and A. Tversky. 1973.
“On the Psychology of Prediction“. Psychology Review 88:
237-251.

ZIiO
Kahneman, D. and A. Tversky. 1972.

“Subjective Probability: A Judgment of Representativeness“.
Cognitive Psychology 3: 438-454.

Keen, P. 1977.
“The Evolving Concept of Optimality“, in Starr, M. and M. Zeleny,

Multiple Criteria Decision Making. Amsterdam: North Holland.

Keen, P. and G. Bronsema. 1981.

' “Cognitive Style Research: A Perspective for Integration“. Center
for Information Systems Research Paper No. 82. Cambridge,
Massachusetts: Sloan School of Management, Massachusetts
Institute of Technology.

Keen, P. and M. Scott Morton. 1978.

Decision Support Systems: An Organizational Perspectiv . Reading,
Massachusetts: Addison-Wesley.

 

Keeney, R. and H. Raiffa. 1976.
Decisions with Multiple Objective . New York: Wiley.

Kennedy, J. and E. Francisco. 1974. *
“On the Formulation of Risk Constraints for Linear Programming“. v/
Journal of Agricultural Economics 25(2): 129-146.

King, R. 1979.
“Operational Techniques for Applied Decision Analysis Under
Uncertainty“. Ph.D. thesis, Michigan State University, 1979.

Koopmans, T. (ed.). 1951.
“Activity Analysis of Production and Allocation“. Cowles
Commission Monograph 13. New York: Wiley.

Kuhn, H. and A. Tucker. 1951.
“Nonlinear Programming“, in J. Neymann (ed.), Proceedings of the
Second Berkeley Symposium on Mathematical Statistics and
Probability. Berkeley: University of California Press.

Kuhn, T. 1978.
The Structure of Scientific Revolutions. Chicago: University of
Chicago Press.

Kunreuther, H., R. Ginsberg, L. Miller, P. Sagi, P. Slovic, B. Borkan
and N. Katz. 1978.

Disaster Insurance Protection: Public Policy Lessons. New York:
Wiley.

Langer, E. and J. Roth. 1975.
“Heads I Win, Tails It’s Chance: The Illusion of Control as a
Function of the Sequence of Outcomes in a Purely Chance Task“.

Journal of Perspectives on Social Psychology 32: 951-955.

2&4

Langley, P. and H. Simon. 1981.
“The Central Role of Learning in Cognition“, in J. Anderson
(ed.). Qgggitive Skills agg Their Acquisition. Hillsdale, NJ:
Lawrence Erlbaum.

Lawrence, K. and J. Burbridge, 1976.
“A Multiple Goal Linear Programming Model for Coordinated
Production and Logistics Planning.“ Internatiogal Journal of
Production Research 14(2): 215-222.

Lee, S. 1972.
Goal Programming for Decision Analysis. Philadelphia: Auerbach.

Lee, 3., J. Horn and H. Brisch. 1988.
“A Multiple Criteria Analysis Model for Academic Policies,
Priorities and Budgetary Constraints“, in G. Fandel and T. Gal

(eds.), Multiple Criteria Decison Making Theory and Application.
New York: Springer-Verlag.

Levy, H. and G. Hanock. 1978.
“Relative Effectiveness of Efficiency Criteria for Portfolio

Selection“. Journal of Financial and Quantitative Analysis 5(1):
63-76.

Lichtenstein, S. and B. Fischoff. 1976.
“00 Those Who Know More Also Know More About How Much They Know?".
Oregon Research Institute Bulletin 16(1). Eugene, Oregon: Oregon
Resarch Institute.

Lichtenstein, S. and B. Fischoff. 1982.
“Calibration of Probabilities: The State of the Art to 1988“, in
D. Kahneman, P. Slovic and A. Tversky (eds.), Judgment Under
Ungertainty: Heuristics and Biases. New York: Cambridge

University Press.

Lichtenstein, S. and P. Slovic. 1971.
“Reversals of Preference Between Bids and Choices in Gambling
Decisions.“ Journal of Experimental Psychology 89: 46-55.

Loasby, B. 1976.

Choice, Complexity and Ignorance: An Enquiry into Economic Theory
and the Theory of Decision Making. Cambidge: Cambridge
University Press.

Ludke, R., F. Strauss and D. Gustafson. 1977.
“Comparison of Five Methods for Estimating Subjective Probability

Distributions“. Organizational Behavior and Human Performance
19: 162-179.

Lusted, L. 1977.
A Stggy of the Efficacy of Diagpostic Radiolooic Procedures.
Efficacy Study Committee of the American College of Radiology,
Chicago, Illinois.

211.2

Machina, M. 1982.
“’Expected Utility’ Analysis Without the Independence Axiom".
Econometrica 58(2): 277-323.

Mao, J. 1978.
“Survey of Capital Budgeting: Theory and Practive". Journal of
Finance 25: 349-368.

March, J. 1982.
“Theories of Choice and Making Decisions“. Society Nov/Dec 1982,
pp I 29-39 a

March, J. 1978.
“Bounded Rationality, Ambiguity, and The Engineering of Choice“.
Bell Journal of Economics 9(2): 587-618.

March, J. and J. Olsen. 1976.
Ambiguity and Choice in Organizations. Bergen, Norway:
Universitetsforlaget.

Marschak, J. 1975.
“Guided Soul-searching for Multi-criterion Decisions“, in
M. Zeleny (ed.), Multiple Criteria Decision Making. Berlin:
Springer-Verlag.

Miller, G. 1956.
“The Magical Number Seven Plus or Minus Two: Some Limits on Our

Capacity for Processing Information“. Psychological Review 63:
81-97.

Mintzberg, H. 1973.

The Nature of Managerial Work. Englewood Cliffs, NJ: Prentice-
Hall.

Mintzberg, H., D. Raisinghani and A. Theoret. 1976.
“The Structure of ’Unstructured’ Decision Processes“.
Administrative Science Quarterly 21: 246-275.

Mittroff, I. and T. Featheringham. 1974.
“On Systematic Problem Solving and the Error of the Third Kind“.
Behavioral Science 19(6): 383-393.

Mosteller, F. and P. Nogee. 1951.
“An Experimental Measurement of Utility“. Journal of Political
Economy 59(5): 371-484.

Murphy, A. and R. Winkler. 1977.
“Reliability of Subjective Probability Forecasts of Precipitation
and Temperature“. Journal of the Royal Statistical Society 26:
41-47.

Naylor, T., J. Balintty, D. Burdick, and K. Chu. 1966.
Computer Simulation Technigues. New York: Wiley.

2143

Neisser, U. 1963.

“The Multiplicity of Thought”. British Journal of Psychology 54:
1-14.

Nelson, A. and T. Harris. 1978.
“Designing an Instructional Package: The Use of Probabilities in
Farm Decision Making“. American Journal of Agricultural
Economics 68(5): 993-997.

Nelson, R. and S. Winter. 1973.
“Towards an Evolutionary Theory of Economic Capabilities“.
American Economic Review 63: 448-449.

Newell, A. and P. Rosenbloom. 1981.
“Mechanisms of Skill Acquisition and the Law of Practice“, in J.

Anderson (ed.), Cognitive Skills and Their Acguisition.
Hillsdale, New Jersey: Erlbaum.

Newell, A., Shaw, J. and H. Simon. 1958.
“Elements of a Theory of Human Problem Solving“. Psychological
Review 65: 151-166.

Nijkamp, P. and J. Spronk (eds.) 1981.
Multiple Criteria Analysis. Aldershot, England: Gower.

Peterson, C. (ed.) 1973.
Special Issue: Cascaded Inference. Organizational Performagce
and Human Behavior. 18: 315-432.

Peterson, C. and L. Beach. 1967.

“Man as an Intuitive Statistician“. Paychological Bulletin 68:
29-46.

Petit, M. 1976.
“Farmers’ Adoption of Technical Innovations“. European Journal of
Agricultural Economics 1976(3): 293-322.

Pitz, G. 1983.
“Human Engineering of Decision Aids“, in P. Humphreys, O. Svenson

and A. Vari (eds.), Analyzing and Aiding Decision Processe .
Amsterdam: North Holland.

Pitz, G. and N. Sachs. 1984.
“Judgement and Decision: Theory and Applications.“ Annual Review
of Psychology 35: 139-163.

Pondy, Louis. 1982.
“On Real Decisions“, in G. Ungson and D. Braunstein, Decision
Makipg: An Interdisciplinary Inguiry. Boston: Kent Publishing.

Pope, R. and R. Ziemer. 1984.
“Stochastic Efficiency and Sampling Errors“. American Journal of

Agricultural Economics 66(1): 31-48.

21.11:.

Porter, R. B. 1974.
“Semivariance and Stochastic Dominance: A Comparison". American
Economic Review 64(2): 288-284.

Ramsey, F. 1931.
The Foundations of Mathematics. New York: Harcourt Brace.

Remus, W. 1984.
“An Empirical Investigation of the Impact of Graphical and Tabular

Data Presentations on Decision Making“. Management Science
38(5): 1984.

Robison, L. 1982.
“An Appraisal of Expected Utility Hypothesis Tests Constructed
From Responses to Hypothetical Questions and Experimental

Choices.“ American Journal of Agricultural Economics 64: 367-
375.

Robison, L., P. Barry, J. Kliebenstein and G. Patrick. 1984.
“Risk Attitudes: Concepts and Measurement Approaches“, in P.
Barry (ed.), Risk Management in Agrigulture. Ames, Iowa: Iowa
State University Press.

Romero, C. and T. Rehman. 1984.
“Goal Programming and Multiple Criteria Decision-making in Farm

Planning: An Expository Analysis“. Journal of Agricultural
Economics 35: 177-198.

Romero, C. and T. Rehman. 1985.
“Goal Programming and Multiple Criteria Decision-making in Farm

Planning: Some Extensions“. Journal of Agricultural Economics 36:
171-185. '

Rosenberg, S., C. Rotz, J. Black and H. Muhtar. 1982.
“Prediction of Suitable Days for Field Work“. Paper No. 82-1832.
American Society of Agricultural Engineers.

Rosenthal, R. 1985.
“Concepts, Theory, and Techniques: Principles of Multiobjective
Optimization“. Decision Sciences 16: 133-152.

Rotz, C. and J. R. Black.
“Machinery Requirements and Cost Comparisons Across Tillage
Systems“, in F. D’Itri, A Systems Approach to Conservation
Tillage. Chelsea, Michigan: Lewis Publishing.

Roy, A. 1952.
“Safety-first and the Holding of Assets“. Econometrica
28: 431-449.

21:5

1981.
“Analogical Processes in Learning“, in J. Anderson (ed.),

Cognitive Skills and Their Acguisition. Hillsdale, New Jersey:
Erlbaum.

Rumelhart, D. and 0. Norman.

Sage, A. 1931.

“Behavioral and Organizational Considerations in the Design of

Information Systems and Processes for Planning and Decision
Support“.

IEEE Transactions on Systems, Man and Cybernetics
11(9): 648-678.

Samuelson, P. 1963.

“Discussion: Problems of Methodology“. American Economic Review
53(supplement): 227-236.

Savage, L. 1954.

The Foundations of Statistics. New York: Wiley.

Schoemaker, P. 1982.

“The Expected Utility Model: Its Variants, PUrposes, Evidence and
Limitations.“ Journal of Economic Literature 28: 529-563.
Shugan, S. 1988.

“The Cost of Thinking“. Journal of Consumer Research 7(2): 99-
111.

Simon, H. 1983.

“Search and Reasoning in Problem Solving“. Artificial
Intglligence 21: 7-29.

Simon, H., 1981.

“Cognitive Science: The Newest Science of the Artificial“, in 0.
Norman (ed.), Perspective on Cognitive Scienc . Hillsdale, New
Jersey: Erlbaum.

Simon, H., 1979.

“Information Processing Mode‘s

.; of Cognition." Annual Review of
Psychology 38: 313-396.

Simon, H., 1979.
Models of Thought. New Haven, Conneticut: Yale University Press.
Simon, H., 1976.

“From Substantive to Procedural Rationality“, in S. Latsis (ed.),
Method and Appraisal in Economic . Cambridge, Massachusetts:

Cambridge University Press, pp. 129-148.

Simon, H. 1965.

The Shape of Automation. New York: Harper and Row.

Simon, H. 1959.

“Theories of Decision Making in Economics and Behavioral
Sciences“. American Economic Review 49: 253-283.

214.6

Simon, H. and J. Hayes. 1976.
“Understanding Complex Task Instructions“, in D. Klahr (ed.),
Cognition and Instruction. Hillsdale, NJ: Lawrence Erlbaum.

Slovic, P. and A. Tversky. 1974.
“Who Accepts Savage’s Axioms?“. Behavioral Science 19-368-373.

Slovic, P., B. Fischoff and S. Lichtenstein. 1977.
“Behavioral Decision Theory“. Annual Review of Psychology 28: 1-
39.

Smith, A. and W. Martin. 1972.
“The Socioeconomic Behavior of Cattle Ranchers, With Implications
for Rural Community Development in the West“. American Journal of
agricultural Economics 54(2): 217-225.

Smith, B. and D. Capstick. 1976.
"Establishing Priorities Among Multiple Management Goals".

Southern Journal of Agricultural Economics 8(2): 37-43.

 

Spetzler, C. and C. Stael von Holstein. 1975.
“Probability Encoding in Decision Analysis”. Management Science
22: 348-358.

Spronk, J. 1988.
Interactive Multiple Goal Programming. Boston: Martinus Nijhoff.

Stael von Holstein, C. 1972.
“Probabilistic Forecasting: An Experience Related to the Stock

Market“. Organizational Behavior and Human Performance 8: 139-
158.

Starr, M. and M. Zeleny. 1977.
“MCDM - State and Future of the Arts“, in M. Starr and M. Zeleny

(eds.), Multiple Criteria Decision Making. Amsterdam: North
Holland.

Steuer, R. 1976.
“Multiple Objective Linear Programming with Interval Criterion
Weights“. Management Science 23(3): 385-316.

Steuer, R. and A. Schuler. 1978.
“An Interactive Multiple Objective Linear Programming Approach to

a Problem in Forest Management“. Management Science 26(2): 254-
269.

Svenson, O. 1983.
“Decision Rules and Information Processing in Decision Making",
in L. Sjoberg, J. Wise and T. Tyezka (eds.), Human Decision
Making. Doxa, Sweden: Lund.

2147

Tauer, L. 1933. /
“Target MOTAD“. American Journal of Agricultural Economics
65(3): 686-618.

Thomson, K. and P. Hazell. 1972.
“Reliability of Using the Mean Absolute Deviation to Derive

Efficient E,V Farm Plans“. American Journal of Agricultural
Economics 54: 583-586.

Toda, M. 1983.
“What Happens at the Moment of Decision? Meta-decisions, Emotions
and Volitions“, in L. Sjoberg, J. Wise and T. Tyezka (eds.),
Human Decision Making. Doxa, Sweden: Lund.

Tversky, A. 1969.
“Intransitivity of Preferences.“ Psychology Review 76: 31-48.

Tversky, A. and D. Kahneman. 1981.
“The Framing of Decisions and the Psychology of Choice“. Science
211: 453-458.

Tversky, A. and D. Kahneman. 1988.
“Causal Schemas in Judgments Under Uncertainty“, in Fishbein, M.
(ed.), Progress in Social Psychology. Hillsdale, New Jersey:
Erlbaum.

Tversky, A. and D. Kahneman. 1973.
“Availability: A Heuristic for Judging Frequency and Probability".
Cognitive Psychology 5: 287-232.

Tversky, A. and D. Kahneman. 1971.
“The Belief in the ’Law of Small Numbers’“. Psychological
Bulletin 76: 185-118.

Ungson, G. and D. Braunstein (eds.). 1982.
Decision Making: An Interdisciplinary Inguiry. Boston: Kent
Publishing.

United States Department of Agriculture (USDA), Various issues.
“Wheat Situation.“ Economics, Statistics, and Cooperative
Service. Washington, D.C.: USDA.

United States Department of Agriculture (USDA), Various issues.
“Feed Outlook and Situation Report.“ Economic Research Service.
Washington, D.C.: USDA.

Von Neumann, J. and O. Morgenstern. 1944.
Theory of Games and Economic Behavior. Princeton, New Jersey:
Princeton University Press.

Wallsten, T. and D. Budescu. 1983.
“Encoding Subjective Probabilities: A Psychological and
Psychometric Review“. Management Science 29(2): 151-173.

21.18

Wheeler, 8. and J. Russell. 1977.
“Goal Programming and Agricultural Planning“. Operations
Research Quarterly 28: 21-32.

White, R., 1978.
“Determining the Capacity of Farm Machines.“ Extension Bulletin
E-1216. Cooperative Extension Service, Michigan State
University, East Lansing, Michigan.

Willis, C. and R. Perlack. 1988.
“A Comparison of Generating Techniques and Goal Programming for
Public Investment, Multiple Objective Decision Making“. American
aggrnal of Agricultural Economics 62(1): 66-74.

Wirth, M. 1968.
“Managerial Decision Processes“, in The Management Factor in

Farming: An Evaluation and Summary of Research. Minnesota
Experiment Station Technical Bulletin No. 258, 1968.

Woolsey, R. and H. Swanson. 1975.
Operations Research for Immediate Application. New York: Harper
and Row.

Yntema, D. and W. Torgerson. 1967.
“Man-computer Cooperation in Decisions Requiring Common Sense“, in
W. Edwards and A. Tversky, Decision Making. Harmondsworth,
England: Penguin.

Zeleny, M. 1982.
Multiple Criteria Decision Making. New York: McGraw-Hill.

Zelany, M., 1973.
“Compromise Programming.“ in J. Cochrane and M. Zelany (eds.).
Multiple Criteria_Qecision Making. Columbia, South Carolina:
University of South Carolina.

Zionts, S. and D. Deshpande. 1981.
“Energy Planning Using a Multiple Criteria Decision Method“, in
Nijkamp, P. and J. Spronk (eds.), Multiple Criteria Analysis:
Operational Methods. London: Gower Press.

Zionts, S. and J. Wallenius. 1976.
“An Interactive Programming Method for Solving the Multiple
Criteria Problem“. Management Science 22: 652-663.

 

Ill mum mm '

 

 

‘ -
_. . _
—'v-
i - _
— _
a -