> - u Y'a-
.- A... ‘ “b‘vfigvmgkr
t.\\.uu~. .21.. a
"‘3‘2‘ gut .L

. ~ ~.
wpo: . ‘V' . "112;; s
u

m, '
33%: '

fists; .3 h 75—",

" Eff“ “ m’w.»%.§l

" ' “ . far-3x333”

- ‘ . ~ , 4:; w .. «ram

‘4‘} “3‘" vi“? a- ‘ fiafl
. 'm-z - *2 .v.

" ~x‘ti‘flfiyt"

A, .Fwfifltf‘gg.‘

mgfifilzaa: ;~

W a

If:

'2
“:5
”'5‘ -‘
$333 Ta:

5%“:

. .. ‘1:

fiat ..

,thri'lz. ~ ~ ~‘ .x’fi—n-i:

“mm 5‘» fl. . . > _ y: ‘
J. J IN I ‘ 4-1”.

we.“ »

~ 3102.. .
‘1"”*~

4W3
.. 45....
_ a _
aim» '{Ti}:i‘.‘§{i:¢‘?“£"$%3”
,.—.¥n%;.,.
“we .

v. .u.
.m

’ ~" 4.3“

‘ fié'r 21;»

L hi2 “Wit“?

”k: w
«1

\\.

41;"... :L

:‘ 'Uz‘ “{3 ma.

3” mg”.

. n-w \d—u. ..
w- “.13... ..-¢

”:9; fl,“ .
' ‘HT’§_;§;\:§' \ .'
.g w'v

u

.- ..-.‘i‘z§.;‘¢é;l an

'.~\

‘. :.~ 1.50?
’ “' . t;

..

v.

«a my .

ll \,.!,vu
1‘

413’ A
‘4 .4
" .q’vim

’53.; “‘5. 3:“
Efi§~§5
“‘~" .2“ 3..

wvéé,:~zv:"

,;_ "1,,
1%??3“

'1
1373;.
I"!
1 ,

,

" a

a.
T1:

7.9

‘ :21 33532,»:
v ”49.? :m-
". Wit

2‘ ,
,.
-\

v5o‘l‘4‘5?’ 1‘
'a' V

 

 

THF- 7 ram an

Illu'inmuumlllmllllm

3 1293 01051 54

T

jlfllll

This is to certify that the

dissertation entitled

THE ECONOMIC VALUE OF ENVIRONMENTAL RISK
INFORMATION: THEORY AND APPLICATION
TO THE MICHIGAN SPORT FISHERY

presented by

Douglas J. Krieger

has been accepted towards fulfillment
of the requirements for

Ph . D . degree in Agricultural
Economics

 

 

/
/ V “jot professorx

Date Agril 7’ 1994

MS U is an Affirmative Action/Equal Opportunity Institution 0-12771

 

 

LIBRARY
Michigan State
Universlty

 

 

 

PLACE N RETURN BOX to removethte checkout from your record.
TO AVOID FINES return on or before dete due.

DATE DUE DATE DUE DATE DUE

' S I ' .9 u: I

. i

I' n V ‘ 1
LE ' -
[III I

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

”TD:
l I I

MSU to An Afflmetlve ActlorVEquel Opportunlty lnetltulon
W

 

 

 

ulna-9.1

THE ECONONIIC VALUE OF ENVIRONMENTAL RISK
INFORMATION: THEORY AND APPLICATION
TO THE MICHIGAN SPORT FISHERY

By

Douglas J. Krieger

A DISSERTATION

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

DOCTOR OF PHILOSOPHY

Department of Agricultural Economics

1994

ABSTRACT

THE ECONOMIC VALUE OF ENVIRONNIENTAL RISK
INFORMATION: THEORY AND APPLICATION
TO THE MICHIGAN SPORT FISHERY

By

Douglas J . Krieger

Tests of sport fish in the Great Lakes region have found traces of a variety of chemical
residues. In sufficient doses some of these chemicals are suspected of causing adverse human
health effects. In response to this problem, all of the governmental jurisdictions bordering the
Great Lakes provide anglers with "Public Health Advisories" which list sites and species
known to be contaminated and offer advice on how to reduce the risk of exposure.

This study asks whether Michigan's current advisory has value to anglers and how that
value might be enhanced. Specifically, the study examines the value of two advisory alterna-
tives -- testing a greater number of sites and providing a list of relatively safe sites. The report
first reviews a Bayesian definition of information and its value. It proves that two common
definitions of information value are algebraically identical. It also clarifies a confusion in the
literature between information value and the welfare effects of changes in environmental
quality.

The study then applies the framework to an evaluation of Michigan's advisories. It
demonstrates that information value is nondecreasing in informativeness and fineness. It also
shows that increased testing increases the informativeness of advisory information while listing
safe sites increases fineness. The conceptual model also concludes that advisory value depends

on (I) the possibility of behavioral change in response to information. (2) the perceived

 

severity of health effects associated with exposure to contaminated fish, and (3) the perceived
accuracy of the advisory.

Focus group and interview pretesting produced a contingent valuation survey that was sent
to licensed anglers in Michigan. The analysis concludes that anglers are probably willing to
pay little for additional testing that does not also list safe sites. However, willingness to pay
for an advisory that lists safe sites is positive and increasing in the number of sites tested.
Nonlinear functional forms suggest that marginal WTP may be decreasing as the number of

sites tested increases.

Acknowledgements

In completing this research and my doctoral program, I owe the greateSt debt to my major
professor, John Hoehn. John's emphasis on quality research contributed considerably to the
development of my research and writing skills and to the overall quality of my graduate educa-
tion. John's dedication to teaching and his breadth of knowledge made him an excellent major
professor and mentor. He was supportive and accessible and deserves a large measure of the
credit for both the quality of my graduate education and of the dissertation research. Of
course, I accept responsibility for the remaining deficiencies in either area.

The other members of my committee also contributed significantly to my education and
dissertation research. Eileen van Ravenswaay's insights and experience were particularly help-
ful in the process of designing the survey. She also taught me a great deal about how to be an
effective teacher. Allan Schmid consistently challenged me to view economics and my
research from a different perspective. His influence will hopefully prevent my thinking from
becoming too narrowly focused by my formal training in economics. Ted Graham-Tomasi,
though a late addition to my guidance committee, helped with the conceptual and empirical
components of the research.

Many others also contributed to my research. I would like to thank John Hesse of the
Michigan Department of Public Health and Douglas Jester of the Fisheries Division of the
Michigan Department of Natural Resources for their help in defining the practical aspects of

the research problem. I would also like to thank Christine Waggoner of the MDNR'S Surface

Water Quality Division for her help in clarifying aspects of the current advisory program and
the practical implications of alternatives. Finally, I wish to thank the Michigan Sea Grant Col-
lege Program for the generous support of my doctoral research.

In my many years at MSU too many people to mention individually have helped me
educationally, personally, administratively, and financially. My family and the many friends
and colleagues I've found here helped me maintain a health perspective on life. The friends
I've made are more important than any degree that I will take from here. Also, without
Gayle's love and support, the hours away from the office would have been much less fulfilling

and the presciently announced celebratory pig roast would not have been.

Table of Contents

Chapter One - Michigan's Public Health Advisories .......................... 1
Michigan's Public Health Advisories ................................ 2
Information Content of Advisories .............................. 3
Sampling and Testing Procedures ............................... 4
Information Alternatives ........................................ 5
Thesis Objectives ............................................ 8
Outline of Research ........................................... 8
Chapter Two - Economic Comparison of Information Systems .................. 10
Information and Probabilistic Beliefs ............................... ll
Representing Information: The Likelihood Function ................... 12

The Effects of Prior Beliefs and Information on Perceived Risks . . . . ..... . . 15
Information Value ........................................... 18
Information Value: Two Definitions ............................ 19
Information Value and Related Welfare Effects .. ..................... 23
Determinants of Information Value ............................. 26

The Value of Fineness and Informativeness ........................... 28
Fineness .............................................. 34
Informativeness ......................................... 37
Summary and Implications ..................................... 40
Chapter Three - The Value of Alternative Public Health Advisories ............... 42
Anglers' Willingness to Pay for Information .......................... 42

 

Welfare Effects of Contamination and Information .................... 44

Advisory Characteristics and Value ................................ 46
Advisory Value and Behavioral Change .......................... 47
Value of Full Disclosure. . . .. ................................ 48
Value of Testing More Sites . . ............................... 50

Accuracy, Perceived Severity of Health Effects, and Advisory Value ........... 59
Perceived Accuracy of Information ............................. 60
Perceived Severity of Mistake ................................ 62

Summary ................................................ 63

Chapter Four - Data Collection: A Contingent Valuation Mail Survey of Licensed Anglers 65

Survey Design and Implementation ................................ 66
Survey Design .......................................... 67
Eliciting WTP .......................................... 68

Description of Advisory Alternatives ......................... 68
Payment Vehicle ..................................... 73
Bid Elicitation ....................................... 75
Explanatory Variables ................................... . . 76
Program Variables ............................ . ....... 76
Behavioral Change .................................... 79
Risk ............................................. 80
Information Accuracy .................................. 82
Sampling Frame and Survey Implementation ....................... 82

Survey Results ............................................. 86
Socio-Economic Characteristics of Respondents ..................... 88
Angler Behavior ......................................... 91
Ex-Post Value of the Current Advisory .......................... 92
Perceived Risk .......................................... 95

vi

Information Quantity and WTP ............................... 96

Summary ................................................ 97
Chapter Five - Estimation of a Discrete Choice Model ...................... 100
Estimating WTP from Discrete Data .............................. 100

An Empirical WTP Function for Advisory Information ................... 102
Variable Definitions ..................................... 102

Program Variables ................................... 103

Perceived Severity of Outcome ............................ 105

Anticipated Behavioral Change ............................ 108

Perceived Accuracy of Advisory ........................... 110

Education ......................................... 111

Estimation and Tests of Hypotheses ............................... 111
Model Specification ...................................... 112

WTP for Full Versus Partial Disclosure ......................... 113

WTP for Full Disclosure .................................. 116
Predicted Probability of Accepting the Advisory Program ........... 123
Testsonypotheses ........... . ................ . 123
Perceived Severity of Outcome ............................ 126

Behavioral Change ................................... 126

Perceived Accuracy of Advisory ........................... 126

Education ......................................... 127

Summary ............................................ 127

Chapter Six - Summary and Conclusions ............................... 128
Conceptual Conclusions ...................................... 128
Designing an Effective CV Survey ............................... 129
Policy Implications ......................................... 133
Appendix A - Questionnaire and Related Materials ........................ 135

viii

Questionnaire ............................................ 135

Survey Cover Letter ........................................ 143
Reminder Postcard ......................................... 144
Cover Letter for First Follow-up ................................ 145
Cover Letter for Second Follow-up ............................... 146
Appendix B - Specification Tests and Alternative Models .................... 148
Specification Tests ......................................... 148
Marginal Effect of Non-Program Variables ....................... 149

Tests for Exclusion of Non-Program Level Forms ................... 149
Significance of Behavioral Change Variables ...................... 149
Bibliography............................. ................... 153

ix

 

List of Tables

Table 2.1 - Informativeness ........................................ 29
Table 2.1 - (Continued) .......................................... 30
Table 2.2 - Fineness ............................................ 32
Table 2.2 - (Continued) .......................................... 33
Table 3.1 - Likelihoods and Posteriors for Partial Disclosure ................... 52
Table 3.2 - Likelihoods and Posteriors for Full Disclosure .................... 55
Table 4.1 - Item Nonresponse ...................................... 87
Table 4.2 — Tests for Nonresponse Bias ................................ 89
Table 4.3 - Socio-Economic Characteristics of Respondents .................... 90
Table 4.4 - Site Types Used ....................................... 92
Table 4.5 - Diversity in Use of Site Types .............................. 93
Table 4.6 - Behavioral Response to Current Advisory ....................... 94
Table 4.7 - Perceived Risk of Adverse Health Effects . . . ..................... 96
Table 4.8 - Probabilistic Perceptions of Risk and Severity of Outcome ............. 97
Table 5.1 - Variable Definitions .................................... 104
Table 5.2 - Estimated WTP for Full and Partial Disclosure ................... 114
Table 5.3 - Nonlinear WTP Functions ................................ 117
Table 5.4 - Final WTP Estimates ................................... 119
Table 5.5 - Total and Marginal WTP for Full Disclosure Information ............. 120
Table 5.6 - Program Costs Associated with 50 Percent Probability of Acceptance ..... 125
Table 6.1 - Pretest WTP Means .................................... 131
Table A.l - Factorial Design for Survey Variables ........................ 147
Table 8.1 - Model Specification .................................... 150

 

Table 8.2 - Exclusion of Behavioral Change Variables ...................... 151

Table 8.3 - Probit Model Estimates .................................. 152

xi

 

List of Figures

Figure 2.1 - Likelihood Probabilities .................................. 13
Figure 3.1 - Determination of Likelihood Probabilities ....................... 51
Figure 3.2 - Likelihood Probabilities with Imperfect Tests ..................... 61
Figure 4.1 - Initial Wording of WTP Questions ........................... 69
Figure 4.2 - Final Description of Current Advisory ......................... 71
Figure 4.3 - Intermediate Form of WTP Questions ......................... 72
Figure 4.4 - Final Form of WTP Question .............................. 74
Figure 4.5 - Behavioral Change Questions ............................... 81
Figure 4.6 - Pattern of Survey Response ................................ 85
Figure 4.7 - WTP for Program Alternatives .............................. 98
Figure 5.1 - Concepts, Variables, and Hypotheses ......................... 106

Figure 5.2 - Total and Marginal WTP for Full Disclosure Information . . . . . . . . . . . . 122

Figure 5.3 - Predicted Probability of Acceptance .......................... 124

xii

 

Chapter One

Michigan's Public Health Advisories

Tests of sport fish in the Great Lakes region have found traces of a variety of chemical
residues. In sufficient doses, some of these chemicals are suspected of causing adverse human
health effects. Of most concern are mercury and suspected carcinogens such as PCBs, DDT,
dieldren, toxaphene, and chlordanes (Bro et al. 1987, Humphrey 1974). Studies of accidental
exposure to mercury and PCBs suggest that health effects may include skin problems, develop-
mental retardation in children, nervous system damage, cancer, and death (Gaffey 1983, Kurat-
sune 1989, Jacobson et al. 1984, Marsh 1987, Eccles and Annau 1987). While acute health
effects most likely result from exposure levels much greater than those faced by Great Lakes
fish eaters, there remains the concern that fish consumers may suffer long term chronic health
problems (Bro et al. 1987).

Because fish can accumulate some toxins to levels thousands of times greater than that
found in the water, fish eaters suffer a potentially greater exposure than those who do not eat
fish (Connor 1984). Humphrey (1983, 1974) found blood concentrations of PCBs and mercury
to be significantly higher in consumers of Great Lakes fish than among those who did not eat
fish. Concentrations were positively related to the quantity of fish consumed. Also, women
who consume Great Lakes fish often have elevated levels of PCBs in their breast milk
(Schwartz et al. 1983).

In response to this problem, all of the governmental jurisdictions bordering the Great
Lakes provide anglers with information about chemical residues in fish. These warnings most
often take the form of public health advisories which list sites and species known to be con-

taminated and offer advice about how to reduce the risk of exposure (Hesse 1990).

 

 

In“. .Il't. II. t

The budget for chemical analysis and printing costs constrain the type and quantity of
information included in Michigan's current advisory. Within these constraints, the content of
the advisory is determined primarily by state fisheries and health risk experts with little public
input. While budget constraints and the opinions of experts are relevant considerations in
program design, cost benefit analysis suggests that the benefits to affected individuals are also
pertinent (Mishan 1976). However. the fact that the advisories are not priced makes [assess-
ment of their value (benefit) difficult.

This study develops and applies methods to estimate the value of advisory information
about sport fish contamination to anglers. Conceptually, it defines information value and
reviews the economic comparison of information systems. Empirically, it adapts the con-
tingent valuation method to obtain measures of willingness to pay for health risk information.
The case study focuses on estimating the value of two specific additions to Michigan's existing
public health advisory program. This knowledge should aid state agencies in making tradeoffs
between the costs of providing different types of information and the benefits to the affected
p0pulation.

The remainder of this chapter reviews the information contained in Michigan's advisory.
It also examines the information activities of other states and provinces in the Great Lakes
region. The review of information strategies used by other jurisdictions helps identify specific
alternatives to Michigan's current approach. Finally, the chapter outlines specific thesis objec-

tives.

Michigan's Public Health Advisories

The Michigan Department of Public Health (MDPH) has issued annual public health
advisories in some form since 1970. Since 1977, the Michigan Department of Natural
Resources (MDNR) has published the advisory in the back of the annual "fishing guide" dis-

tributed to anglers when they purchase a license. The MDPH also issues a news release

accompanying the annual advisory as well as periodic releases when new information becomes
available, about once or twice a year. They also make a special effort to target particularly
vulnerable groups. For example, through direct contact and articles in professional journals,
the MDPH has informed childbirth educators and general physicians about the Special health
concerns for women of childbearing age associated with consuming contaminated fish. Also,
the MDNR has recently begun to post advisories at public access sites where contaminants

have been found.

Information Content of Advisories

Michigan’s current advisory presents information in three forms: text, graphics, and
tables. The text of the advisory contains a brief description of some chemicals found in
Michigan fish, the sources of the chemicals, and possible health effects. It also warns that con-
centrations of contaminants are likely to be higher in larger or older fish, predator species,
fatty fish, and carp or catfish. It suggests that anglers who intend to eat their catch trim the fat
and skin from fillets and cook them so that fat can drain away. A graphic presentation
illustrates the suggested method of trimming fish.

Since 1989 the advisory has also included a textual warning about mercury in inland
lakes. This warning states that mercury contamination is widespread in the North Central
United States and Canada. It suggests that anglers eat no more than one meal per week of rock
bass, perch, or crappie over nine inches in length; largemouth bass, smallmouth bass, walleye,
northern pike, or muskie of any size from any inland lake. The advisory warns nursing
mothers, pregnant women, women who intend to have children, and children under the age of
15 against eating more than one meal per month of these fish.

Finally, the advisory also contains a table listing sites that have been found to be con-
taminated. For each site. the table lists the species and sizes of fish that are contaminated, the

specific chemical responsible for the advisory, and suggested consumption restrictions. The

 

advisory contains three specific categories of consumption advice based on the proportion of
sampled fish for which contaminant levels exceed state standards. When fewer than 10 percent
of tested fish contain excessive contaminant levels, the advisory does not suggest any consump-
tion restrictions. If 11 to 49 percent exceed these levels, consumption is recommended not to
exceed one meal per week. The advisory suggests no consumption if over 50 percent of the
sampled fish exceed trigger levels for any contaminant. In addition, nursing mothers, pregnant
women, women who intend to have children, and children under the age of 15 are warned

against eating any fish from either of the restricted categories (Hesse 1990).

Sampling and Testing Procedures

The state of Michigan has tested contaminant levels in sport fish regularly for more than
twenty years. The sampling and testing procedures currently in use have been in place since
1987. In the four year period from 1987 to 1990 the state tested 4,072 fish from 260 sites.
This amounts to an average of 1,018 fish from 65 sites annually.

Under the current program, the Surface Water Quality Division of the Michigan Depart-
ment of Natural Resources annually compiles a list of proposed test sites. Several criteria
influence site selection. First, the state tests a subset of a preselected group of ”trend monitor-
ing" sites every year on a four to five year rotation in order to identify trends in contamination
levels. Second, the state may choose to test a site if a contamination problem is suspected.
These "problem evaluation" sites might include collection points near known waste contamina-
tion sites or industrial discharges. In addition, a number of previously tested problem evalua-
tion sites are retested to determine if contaminant levels have changed. Finally, the MDNR
may select particular sites if it receives enough public requests for testing. Priority is also
given to sites that receive heavy fishing pressure. When a list of proposed sites is compiled, it

is distributed to other agencies within state government for comment. Sites may be added to or

removed from the list based on this feedback. The total number of sites tested each year is
constrained by the budget for chemical analysis which in 1992 totalled $320,000. ‘

Once a list of sites is agreed upon, employees of the Fisheries Division or the Surface
Water Quality Division of the MDNR visit the site and capture fish for testing. They take fish
by netting or shocking, depending on the characteristics of the site. The species sampled
depend on the reason for testing. If the site is being tested for mercury, two top predator
species are sampled. If the site is a problem evaluation site, carp and a top predator species
are sought. Sometimes a particular species may be targeted if specific data is needed.

Once fish are collected the MDPH analyzes them for the presence of 28 separate chemi-
cals. The preparation of fish for analysis depends on the species and how they are most often
prepared for consumption. For most Great Lakes species, tests are conducted on skinon fillets

of individual fish. For carp and catfish, skin-off fillets are used.

Information Alternatives

The advisories issued by various jurisdictions in the Great Lakes region differ in a number
of dimensions. These differences suggest a feasible set of alternatives to Michigan's current
advisories. Differences include (1) the way advisories are published, (2) the trigger levels used
to generate specific advice, (3) the specific categories of advice issued, (4) the sampling and
analysis of fish, and (5) the type of information included (Foran and Vanderploeg 1989, Hesse
1990).

Most jurisdictions publish advisories as part of the pamphlet of fishing regulations and
also distribute advisories in a separate booklet. Exceptions are Pennsylvania and Ohio which
currently publish advisories only through press releases. Michigan distributes a brochure

targeted at women of childbearing age and places posters at public access sites in addition to

 

‘ Personal communication with Christine Waggoner. Surface Water Quality Division. Michigan Department of

Natural Resources.

printing the formal advisory in the fishing regulations. Most states update advisories annually.
Exceptions are Wisconsin which updates twice a year and Minnesota which updates every two
years. Ontario publishes consumption advisories in a large booklet separate from fishing
regulations and updates annually.

Jurisdictions also differ in the specific warnings they issue and in the contaminant levels
that trigger the warnings. For example, the states bordering Lake Michigan issue three
categories of advice based on the percentage of fish in a sample that contain levels of any con-
taminant exceeding FDA action levels. New York also uses FDA action levels to initiate three
specific warnings but uses a method of combining measured concentration levels that accounts
for possible interactions of contaminants. Minnesota is the only state that does not link
advisories to FDA action levels. Minnesota uses the detection limit for PCBs and dioxins to
trigger consumption warnings. Detection limits for these chemicals are two orders of mag-
nitude smaller than FDA action levels. For mercury, Minnesota uses a trigger level an order
of magnitude smaller than FDA action levels. Ontario issues six categories of advice based on
trigger levels that are very close to FDA levels.

All jurisdictions except Ohio include special warnings for women and children. The
wording of these warnings differs, but they are essentially the same except for the age at which
the warning takes effect for children. All jurisdictions also offer preparation and cooking sug-
gestions that can reduce risk. Ontario and Wisconsin include much greater detail about types
of contaminants, sampling and analysis techniques, and potential health risks than other
advisories. Judging by the number of warnings issued, some jurisdictions appear to test a
much greater number of sites. For example, Illinois listed only seven sites in 1986 while
Ontario listed 1,700 in 1990.

Probably the greatest difference in the type of information provided is the inclusion of
information about fish that are found not to contain dangerous levels of contaminants.
Michigan and all Other states except Minnesota list a site in the advisory only if tests of fish

from that site indicate the need for consumption restrictions. Many tested fish do not contain

levels of contaminants sufficient to warrant a restriction, but this information is not published.
Ontario and Minnesota are the only jurisdictions that list test results from all water bodies and
species tested.

Alternatives to Michigan's current advisory range from eliminating the program altogether
to an expanded program similar to Ontario's. Not all of these alternatives are politically or
economically feasible. Discussions with state officials revealed two informational changes they
view as desirable and would like to incorporate in future monitoring plans and advisories. 3
The first is to publish results of all tests regardless of whether chemical concentrations warrant
a consumption restriction. During the period from 1987 to 1990 the state of Michigan tested
260 separate sites. This testing program generated 13 specific consumption warnings. Thus, a
program that listed all tested sites could have provided information about 247 relatively safe
sites. The primary constraint in implementing this policy change is printing costs.

A second alternative that interests state officials is to test a greater number of sites.
Michigan contains over 5,800 publicly accessible fishing sites. These include about 3,600
inland lakes, 2,200 sites on rivers and streams, and the Great Lakes. The current budget
permits testing about 65 sites per year. Some of the testing effort is directed at trend monitor-
ing sites so only about 30 new sites are tested each year. At current testing rates it will take
many years to gather information on all of Michigan's water bodies. Information about a
greater number of sites would provide anglers a broader base of knowledge about the risks
associated with their fishing choices. With this information they could make choices more
consistent with their preferences for risk bearing. Survey results suggest that anglers are also
interested in information about a greater number of sites. A survey of Ontario anglers in 1989
asked for suggestions for improving advisory information. Respondents most frequently sug-

gested increasing the number of sites and species tested 3 (Ontario 1990).

 

3 Personal communication with John Hesse. Chief. Environmental Health Assessment Division. Michigan Depart-
ment of Public Health.
3 Ontario's advisory already lists sites that were tested and found to be safe. Therefore. respondents did not sug-

gest listing safe sites among the desired improvements.

Thesis Objectives

This research has three specific objectives related to the problem of providing information
about fish contamination in Michigan. The first is to develop a conceptual foundation for
assessing the value of different types of contamination information. The framework links a
compensating variation measure of willingness to pay (WTP) to changes in expected utility
measured in a Bayesian model. The conceptual model generates hypotheses about the value of
different types of information. It also yields hypotheses about the distribution of information
benefits across different types of individuals.

The second objective is to develop a contingent valuation (CV) survey instrument to
measure WTP for specific information related to fish contamination. The survey assesses
anglers' perceptions of risk and their current angling practices. Focus groups provide insights
into anglers' cognitive processes concerning risk and the aspects of the contamination problem
they believe to be important. Few studies have used focus groups to explore attitudes about
environmental risks and the research yields some insights into their application and interpreta-
tion.

The third objective is to obtain estimates of anglers' WTP for two types of contamination
information: (1) information about which locations have been found to be safe, and (2)
information obtained from testing a greater number of sites. The state of Michigan should find
the estimates useful as a measure of benefits to compare with the costs of obtaining and distrib-

uting information.

Outline of Research

The remainder of this report develops the conceptual and analytical tools needed to

measure the value of information about contamination, describes the estimation of information

value for the case of sport fish contamination in Michigan, and develops the policy implica-

tions of the empirical findings. Chapter two reviews the theory of information value and
explores the conditions under which different information systems can be compared. Chapter
three analyzes the welfare impacts of alternative information systems and develops specific
behavioral and distributional hypotheses. Chapter four discusses the methods used for data
collection and analysis and reports descriptive statistics. Chapter five describes the estimation
of WTP measures and discusses results. Chapter six summarizes research findings, reviews

their policy implications, and suggests areas for future research.

Chapter Two

Economic Comparison of Information Systems

This research investigates the value of information about food and environmental con-
tamination. In particular, it asks whether current advisories about chemical residues in
Michigan's sport fish have value to anglers and how different types of information might
enhance that value. This chapter outlines a conceptual framework for the analysis of informa-
tion value. It also explores the economic comparison of different types of information.

The analysis uses the Bayesian model to describe how individuals incorporate information
into their decision processes. The chapter begins by defining information and explaining the
characterization of information in terms of a likelihood function. It also demonstrates how the
Bayesian framework combines prior beliefs and new information to form posterior beliefs.
The definitions and analysis of the section aid in understanding the remainder of the chapter.

The chapter also reviews the literature on information value and the economic comparison
of information alternatives. The review has two objectives. The first is to develop a clear
conceptual understanding of information and its value. The literature contains two definitions
of ex ante information value. The algebraic similarity of the two definitions has contributed to
contradictory conclusions about the sign of ex post information value. Both Graham-Tomasi
(1988) and Pope (1985) conclude that ex post information value may be negative. The analysis
of this chapter shows that information value is nonnegative ex ante and ex post. It
demonstrates that the confusion about ex post value is caused by a failure to distinguish
between nonnegative information value and the possibly negative welfare effects associated

with changes in the environment. The chapter also proves that the two definitions of ex ante

IO

11

information value are algebraically equivalent if probabilistic beliefs are updated using Bayes'
rule.

The second objective of this chapter is to systematically link characteristics of information
systems to economic value. The analysis defines an ordering of information systems based on
the concept of informativeness - a measure of how well an information system predicts states.
It then shows that informativeness is synonymous with economic value. It also defines the
concept of fineness as a special case of informativeness. Fineness measures the relationship
between signals and states that are relevant to decision making. Two propositions summarize
the relationships between informativeness, fineness, and economic value.

Characteristics of information users may also influence information value. However, the
links between individuals' characteristics and information value are weaker and more difficult
to assess empirically than those between information value and characteristics of information
systems. The analysis also develops the proposition that information value is nonincreasing in

the perceived utility of outcomes.

Information and Probabilistic Beliefs

The literature defines an information system as a set of signals (Hirshleifer and Riley
1979). A signal is a message about how likely each of several unknown states is to occur. For
example, a weather forecast service produces signals about the likelihood of future states of
"rain” or ”dry”. Michigan's public health advisory issues signals that distinguish between
states of ”safe" or ”unsafe" levels of chemical residues in fish. Specifically, the current
advisory issues an explicit signal to "restrict consumption" when sites are "unsafe". I An

information system that issues only one signal also produces an implicit signal associated with

 

' Unless specified otherwise. the examples in this chapter and the next do not distinguish between the two levels

of consumption restrictions suggested by the advisory. In the examples. the "restrict consumption" signal represents
both the signal to restrict consumption to no more than one meal per week and the signal to consume none of the
fish.

the absence of the explicit signal. For instance, the current advisory provides the explicit sig-
nal to "restrict consumption" and an implicit signal of "no report".

The Bayesian model assumes that individuals have prior beliefs about the probabilities of
different states. Information is defined as a signal that changes probabilistic beliefs (Chavas
and Pope 1984). For instance, prior to reading the public health advisory, an angler may per-
ceive the probability of contamination in fish from a favorite site to be low. This perception
represents the angler's prior belief about the probability of the state of "safe" relative to a state
of "unsafe". Suppose the angler then discovers that the advisory issues a signal to "restrict
consumption" of fish from that site. This message may alter prior beliefs about whether fish
from the site are "safe" or "unsafe" to eat. The Bayesian model defines the informed percep-
tions of state probabilities as posterior beliefs.

This section introduces the terminology and analytics of the Bayesian approach. It first
defines the likelihood function that the Bayesian framework uses to represent the information
content of signals. It then-introduces Bayes' rule which formally describes the formation of
posterior beliefs from prior beliefs and new information. Several examples demonstrate that
the prior beliefs and new information influence posterior beliefs in proportion to weights
attached to each. These weights represent the relative confidence people have in prior beliefs

and in information.

Representing Information: The Likelihood Function

The Bayesian likelihood function describes the relationship between states and signals. It
summarizes the informational content of signals. For each possible state, the likelihood func-
tion defines the conditional probability that the information system will generate a particular
signal given that the state is true. Thus, two states and two signals yield four likelihood
probabilities. The likelihood function generates twelve likelihood probabilities if there are four

states and three signals .

l3

 

 

 

 

 

 

 

 

States Signals
No
Consumption Restrict
Restriction Consumption
(NCR) (RC)
Safe P(NCR | safe) P(RC I safe)
Unsafe P(NCR | unsafe) P(RC I unsafe)

 

Figure 2.1 - Likelihood Probabilities

Figure 2.1 illustrates the likelihood probabilities associated with an information system
that provides signals of "no consumption restriction" (NCR) and "restrict consumption" (RC)
associated with "safe" and "unsafe" states of chemical contamination. P(NCRlsafe) represents
the probability (likelihood) that the information system generates a "no consumption restric-
tion" signal when a site is "safe". The likelihood P(NClsafe) is the probability of a "restrict
consumption" signal when a site is "safe”. Similarly, when a site is "unsafe", P(RClunsafe) is
the probability of a "restrict consumption" signal and P(NCRlunsafe) is the probability that the
information system identifies an "unsafe" site as "safe".

An information system must generate a signal regardless of the state that ultimately
occurs. Thus, the probabilities in the rows of Figure 2.1 sum to one -- P(RCIsafe) plus
P(NCRlsafe) equals one and P(RCIunsafe) plus P(NCRIunsafe) equals one. Likelihood
probabilities over states (columns in the figure) do not necessarily sum to one. Suppose a test
for contaminants never generates a false positive. This implies that P(RClsafe) equals zero
and P(NCRlsafe) equals one. It does not imply that the test never produces a false negative.
Therefore, P(RCIunsafe) does not necessarily equal zero and P(NCRlunsafe) does not neces-
sarily equal one.

Likelihoods reflect subjective perceptions of the link between states and signals. In simple

physical models. individuals may choose likelihoods that reflect the physical or statistical char-

14

acteristics of a testing environment. However, when the link between states and signals is less
certain, likelihoods may appear much more subjective. As an example of the former case. con-
sider the problem of estimating the proportion of defective units in a production run from a
representative sample. The binomial distribution describes the statistical relationship between
the true proportion of defectives and the number found in the sample (W inkler 1972, Freund
1962).

The binomial distribution

P(rln,p)= [ :]pf(1-p)n-r

describes the likelihood of observing r defective units in a sample of size n (the signal) given
that p is the true proportion of defectives (the state). For example. suppose testing finds three
defectives in a sample of 20 units. The test produces a signal that the proportion of defectives
is .15 (3 -:- 20). The likelihood of receiving this signal depends on the true proportion of
defectives. If the true state is ten percent defectives, the likelihood of receiving the above sig-
nal is.P(3|20,.1) = .19. Substituting other possible states for .1 in the binomial formula
yields the likelihoods of receiving the signal of 15 percent defectives given that other states are
true. The likelihood of finding 15 percent defectives in the sample is greatest when there are
really 15 percent defectives in the production run.

When available, individuals may base the choice of likelihoods on past observations of the
association between states and signals. Several studies of the value of weather information to
farmers used historical data on the relationship between forecasts and actual weather patterns to
calculate likelihoods (Baquet et al. 1976, Tice and Clouser 1982, Byerlee and Anderson 1969).
Specifically, these studies derive the likelihood of receiving a forecast of rain when it actually
will rain from the historical record of the association of actual rain with a rainy forecast.

These examples illustrate cases where likelihoods have a strong foundation in statistical or
physical processes. However, likelihoods essentially reflect subjective perceptions of the

information content of a signal. Even likelihoods from the examples above are subjective. For

15

example, farmers may believe that the accuracy of weather forecasts changes over time.
Similarly, plant managers may believe that sampling for defective items is not random.

Some situations seem to offer little substantive basis for the choice of likelihoods. This
seems particularly true in the case of chemical contamination of food and the environment
where even experts disagree about the associated health risks. The lack of consensus among
experts confronts consumers with a number of contradictory messages. How individuals evalu-
ate these messages depends on a number of factors. For example, an individual may question
the ability of tests to identify different concentrations of chemical residues or question whether
an information provider accurately reports test results. Factors such as these may affect the

perceived association between signals and states.

The Effects of Prior Beliefs and Information on Perceived Risks

Bayes' rule formalizes the way the Bayesian model combines prior beliefs and new
information to form posterior beliefs. Consider an information system that generates a set of
signals Y= {y,,y2,...,y,} related to a set of states 8: {s,,sz,...,sx}. Formally, Bayes' rule states

P(Sk) P(Yi I St)

(2.1) P(Stlya) = 1’09

 

where P(sk) is the prior probability of state st, P(sklyi) is the posterior probability of state sk
given signal yi, P(yilsk) is the likelihood probability of receiving signal yi given state st, and
P(yi) = 2k P(sk) P(yilsk) is the probability of receiving signal yi.

The relative effect of information on posterior beliefs depends on the relative confidence
an individual has in information relative to prior beliefs (Hirshleifer and Riley 1979). An indi-
vidual who is very confident of prior beliefs will not be easily swayed by new information.
Conversely, information of any quality may strongly influence an individual with little con-
fidence in prior beliefs. Confidence measures the strength of beliefs and is reflected in prior

and likelihood probabilities. An individual who is very confident that a particular state is true

16

will assign a relatively high prior probability to that state. Similarly, an individual with a
great deal of confidence in a signal will believe receipt of the signal to be very likely for one
state and relatively unlikely for other states. Confidence in a signal implies that one likelihood
probability associated with the signal is large relative to the others.

Greater confidence in prior beliefs or likelihoods implies greater posterior confidence.

Consider how posterior probabilities change with changes in priors and likelihoods.

aP(Sk i Yi) : P(Yi i sit) P(Yi) ' P(Yi I 502 P(Sk)

 

 

 

 

(2.2a) 3W5.) P(y,)2 .>_ 0 for k= j
(22b) 6% = - Pods-2:23.218.) P(sk) < 0 for k ¢j
(22C) (311%:3 = P(s.) ”"5552“? Pods.) _>_ 0 for 1:;
(2.2d) (rig—fl = ' P(S“) 11:83:)” "51) < o for k¢ j

As the prior probability of a state increases, so does the posterior probability of that state for
each signal (equation 2.2a). Increases in the prior probability of a state reduce the posterior
probability of other states (equation 2.2b). Similarly, for a given signal, greater likelihood
associated with a given state increases the posterior probability of that state while reducing the
posterior probabilities of other states (equations 2.2c and 2.2d). Thus, greater confidence in
the priors and/or the likelihoods increases the confidence of the posterior.

This discussion leads to a natural interpretation of confidence in terms of the spread of a
distribution. Greater confidence in prior beliefs implies that more probability is concentrated
on a single state. Greater confidence in a signal implies a greater likelihood that the signal is
associated with one state as opposed to Others. The relationship between formal measures of

the spread of the prior and likelihood distributions and posterior beliefs depends on distrib-

17

utional assumptions. For example, when priors and likelihoods are normally distributed,
Kihlstrom (1974) shows that the posterior estimate of the mean, E(,u), is the sum of prior and

likelihood means weighted by inverses of the prior and sample variances respectively.

4)

(2.3) Eta) = m “p

6
mes +

where E01.) equals the posterior estimate of the mean, 0 equals the inverse of the sample
variance, p, equals the sample mean, ¢ equals the inverse of the prior variance, and up equals
the prior mean. Kihlstrom defines 6 and (b as measures of confidence, Zellner (1971) refers to
them as precision parameters, and Winkler (1972) calls them measures of uncertainty.

As the variance of the prior declines (confidence in the prior increases), (1) increases, and
the weight attached to the prior mean increases. Thus, as confidence in the prior increases, the
effect of prior beliefs on posterior beliefs increases relative to the influence of information.
Similarly, increased confidence in information implies a smaller sample variance, an increase
in 0, and greater relative weight attached to information.

If priors and likelihoods are distributed according to the beta distribution, confidence is
defined in terms of the quantity or statistical quality of beliefs. Viscusi and O'Connor (1984)
investigate changes in risk perceptions for workers in the 'chemical industry in response to new
risk information. They describe priors and likelihoods with a beta distribution and show that
the posterior estimate of risk is the sum of prior risk and sample (information) risk weighted

by the relative number of observations reflected in priors and information.

 

(2.4) p'=£:ys+F%—Tp

where p‘ equals the posterior risk of accident, 5 equals the risk of accident conveyed by
information, p equals the perceived prior risk of accident, E equals the number of observations
reflected in information, and 7 equals the number of observations reflected in prior beliefs.

An experienced worker may develop perceptions of chemical risks from years of observa-

tion. Such a worker may believe 7 to be large relative to E. Viscussi and O'Connor's model

18

implies that the worker's posterior perceptions of risk would depend largely on prior beliefs --
confidence in prior beliefs is high relative to confidence in information. On the other hand, a
worker may have a great deal of confidence in a large scale study of chemical risks. This
belief implies a larger E and increases the influence of information on posterior perceptions of
risk.

Viscussi and O'Connor refer to the weights associated with s and p as measures of preci-
sion. These weights are conceptually equivalent to the measures of confidence developed in

the context of the normal distribution.

Information Value

The following analysis shows that information has value because it helps people avoid
mistakes. For example, an individual may wish to avoid chemical residues in wild game.
Prior to news reports of mercury in pheasants, a hunter may have consumed pheasant with
little perceived risk. Information about mercury contamination changed many hunters' percep-
tions of the health risks associated with eating pheasant (Shulstad and Stoevener 1978).
Because they were informed, these individuals had the option of changing consumption behav-
ior to avoid exposure. Without information they would have continued their pre-information
consumption -- a mistake if informed -- and may ultimately have suffered an unexpected
adverse health outcome.

This section reviews two definitions of information value - one that expresses information
value in terms of mistakes, and one that does not. The analysis proves that the two definitions
are algebraically equivalent. The definition that does not identify mistakes has been misinter-
preted to imply that information can make people worse off. The section shows this result to
be caused by a confusion between ex post signal value and the welfare effects associated with
contamination. It concludes by providing a clear distinction between the different welfare

effects associated with contamination and information. The analysis shows that information

19

value is always nonnegative and that welfare losses due to contamination are larger if con-

sumers are uninformed.
Information Value: Two Definitions

Suppose anglers' utility depends on attributes of a fishing experience. Anglers select
attributes by choosing levels of specific fishing behaviors, q={q,,q2,...,qL}. These behaviors
may include the choice of a fishing site and the species, size, and quantity of fish to consume.
Suppose anglers also care about the state of health. The advisory suggests that health may
depend on the level of chemical residues in fish consumed, s, and on behavioral choices that
avert risk. Thus, to the extent that some behavioral choices reduce the risk associated with
chemical residues, health, h, depends on s and q. However, current public health advisories
are not sufficient to identify with certainty which fish contain harmful chemical residues.
Therefore, anglers are uncertain of the health consequences associated with specific fishing
behaviors.

Under conditions of uncertainty, economic decision theory suggests that pe0ple maximize
expected utility over possible states of nature (von Neumann and Morgenstern 1944). Anglers
have only probabilistic knowledge about the occurrence of states of contamination. Under
these conditions, the angler's maximization problem is

max I u(q,h(q,s)) P(s) ds

q

where P(s) is the prior probability distribution over the alternative states or levels of con-
tamination. Individuals continually update probabilistic beliefs as new information becomes
available. Therefore, prior beliefs at any time are the posterior probabilities associated with
the last information received. Priors implicitly depend on all prior information. However. for
simplicity, the notation used in this paper suppresses the conditional aspect of prior beliefs.
Define q' as the vector of behavioral choices that maximizes expected utility given prior

beliefs. The angler's prior maximal expected utility is therefore

20

(2.5) up = J u(q',h(q',s)) P(s) ds.

Now suppose the angler receives a signal about contamination of fish. After receiving the
signal, y, the angler uses Bayes' rule to form posterior probabilities P(s I y). The angler's max-
imization problem is now

max I U(q,h(q,8)) P(Sly) ds.

q

Define qy as the Optimal action given signal y. The angler's posterior maximal expected utility

associated with receipt of signal y is
(2.6) u, = j U(q’.h(qy.s)) P(sm ds.

Each signal may cause a different change in beliefs and a different optimal action. The
Bayes' strategy defines the optimal behavior for each possible signal. Ex ante, prior to receipt
of a particular signal, an individual will be uncertain about which signal is forthcoming. The
expected utility associated with the information system will then be the expected utility of fol-
lowing the Bayes' strategy (Hirshleifer and Riley 1979). For example, suppose the informa-
tionsystem contains I signals, Y={y,,y2,...,y,}. The expected utility associated with the

Bayes' strategy is

(2.7) uy = I IU(q’,h(q’.s))P(sly)dsP(y)dy

where:
x
1’0) = Elfin.) 1’0 | St)

defines the probability of receiving signal y. Therefore, the ex ante expected utility associated
with information system Y is the expectation of ex post expected utilities of signals (equation

2.6) with respect to the probability of receiving each signal.

21

Anderson et al. (1977) define the ex ante value of information as the difference between
the expected utility of the Bayes' strategy and the expected utility of the prior optimal act

(equation 2.5)

(28) VY 2” UY ' up

I j u(qymtqas» P(s I y) ds P(y) dy j u(q'.h(q‘,s» P(s) ds.

Equation 2.8 states that ex ante information value is the difference between expected util-
ity with information and expected utility prior to being informed. This definition has two con-
ceptual limitations. First. it is not clearly consistent with the conceptual source of information
value as helping to prevent mistakes. A prior act is a mistake only when evaluated in light of
posterior perceptions. The Anderson et al. definition does not identify mistakes. Second, the
definition has no ex post analogue.

Hirshleifer and Riley (1979, I992) propose a definition of information value that is intui-
tively consistent with both ex post and ex ante values. They define both ex post and ex ante
values in terms of mistakes. To illustrate, consider an individual who chooses a course of
action without knowledge of a signal, y, that would change probabilistic beliefs. Once the sig-
nal is received, the individual may view prior behavior as a mistake. Given receipt of a signal

y, the expected utility associated with the (mistaken) prior act is
(2.9) u... = j u(q‘.h(q'.s» P(s I y) ds.

By definition, q‘ does not maximize expected utility given posterior beliefs P(sly).
Therefore, q‘ is a mistake when evaluated relative to posterior beliefs. Signal y makes behav-
ioral change possible and prevents the mistake. The utility value of a signal is the difference
between the expected utility of the optimal act when aware of the signal (equation 2.6) and the

expected utility of the mistake (equation 2.9).

(2.10) v =u —u

.v .v m

[ u(q>'.h(qY.s)) P(sly) ds - J u(q‘.h(q',s)) P(siy) ds

22

Equation 2.10 defines (in utility terms) the ex post value of signal y.
Hirshleifer and Riley define the ex ante value of the information system Y= {y1,y3....,y,}

as the expectation of ex post signal values with respect to signal probabilities

(2.11) vY = I { [u(qY.h(qY,s))P(sly)ds

- [U(q',h(q',5))P(51y)ds }P(y)dy-

The definitions of ex ante information value in equation 2.8 and 2.11 are algebraically
identical if posteriors are formed from priors using Bayes' rule. The following proof shows
the difference between the two definitions to be zero and thus proves the equivalence.

The Anderson et al. definition of equation 2.8 can be written as

(2.12) vn= I { I u(qy,h(q’.s))P(SIy)ds

- j u(q'.h(q'.s>) P(s) ds } P(y) dy

since i P(y) dy = 1.

The expression of equation 2.12 appears similar to Hirshleifer and Riley's definition in
equation 2.11. However, equation 2.12 uses prior beliefs to evaluate the prior act. Equation
2.11, on the other hand, uses posterior beliefs and defines the prior act as a mistake. The dif-

ference between equations 2.12 and 2.11 is

(2.13) vA-vy = j ju<q'.h(q‘.s» P(smds P(y)dy

- j u(q',h(q°.s» P(s) ds.

Expanding the posterior probabilities with Bayes' rule yields

23

j J u(q'.h<q‘.s)> ”5312;, 8) ds P(y) dy - j u(q'.h(q‘,s» P(s) ds

= j [ u(q‘,h(q‘,S))P(yIS)P(S)dsdy- [u(q',h(q‘.s»P(s>ds.

Changing the order of integration in the first term and rearranging yields

(2.14) [ u(q',h(q’,s)) P(s) I P(yls) dy ds - I u(q‘,h(q‘,s)) P(s) ds.

The Bayesian model implies j P(yls) dy = 1. Thus, expression 2.14 reduces to

j u(q',h(q‘.s» P(s) ds - ju<q'.h(q',s» P(s) ds = 0.

Therefore, the two definitions are the same.

Information value is nonnegative. Consider the definition of ex post signal value in equa-
tion 2.10. By definition, qy is the optimal act associated with posterior beliefs P(sly). Thus,
the second term on the right hand side of equation 2.10 can not be greater than the first.
Therefore, the ex post value of a signal is nonnegative. As the sum of ex post signal values,

ex ante information value must also be nonnegative.
Information Value and Related Welfare Effects

Graham-Tomasi (1988) and Pope (1985) conclude that ex post signal value may be nega-
tive. This section shows that this contradiction arises from a confusion between information
value and other welfare effects associated with contamination.

Consider an angler with beliefs P(s) about the probability of chemical residues in fish.

The angler's expected utility is

(2.15) up = I u(q’,h(q‘,s)) P(s) ds

where q' represents optimal behavior given P(s). Define this as the prior utility of the prior

act.

24

Suppose the angler is ignorant of a signal, y, about a contamination incident. If informed,
the signal would produce posterior beliefs P(sly) and new optimal behavior qY. To avoid
issues of risk perception, assume that informed perceptions of risk are correct. Ignorance does
not shield uninformed anglers from actual conditions. Both informed and uninformed anglers
face possible future health effects determined by P(s | y). In light of informed perceptions, the

prior optimal act is a mistake. The expected utility of the mistake, um, is

(2.16) u... = j u(q',h(q'.s» P(sly) ds.

Define this as the posterior utility of the prior act.
Contamination and ignorance of signal y cause a welfare loss equal to the difference

between the posterior utility of the prior act, um, and the prior utility of the prior act, up.

(2.17) u... - u. = j u(q',h(q',s»1>(sly) ds - [ notations» P(s) ds

By definition, q‘ is the optimal act given beliefs P(s). Therefore, the first term on the right
hand side of equation 2.17 can not be larger than the second and the expression is
unambiguously nonpositive.

Ignorance forces the uninformed angler to continue prior behavior, q‘, even though this
would be a mistake if informed. If aware of signal y, the angler would choose behavior qy and
avoid the mistake of continuing the prior optimal act. The posterior utility of the posterior act

is
(2.18) 11). = J u(qY,h(qY,s)) P(sly) ds.

Information prevents an individual from making the mistake of continuing prior behavior when
informed risk perceptions dictate a different choice. Thus, the informed angler experiences a
welfare change equal to the difference between the posterior utility of the posterior act, u,, and

the prior utility of the prior act up

(2.19) uy - up = I u(qY,h(qY,s)) P(sly) ds - I u(q',h(q',s)) P(s) ds.

25

Equations 2.17 and 2.19 differ only in that the angler becomes informed. The difference
v, = (u, -u,) - (um-u.)

= j u(qv.h(qr.s» P(sly) ds - [ u(q‘.h(q'.s» P(sly) ds

is thus the ex post value of the signal y. This equals the ex post value of signal y given by
equation 2.10.
The analysis above illustrates the three distinct welfare effects associated with a con-

tamination incident and information.

1. Contamination causes a gross welfare loss (equation 2.17) as individuals
continue prior behavior in ignorance of potential exposure to contaminants.
The gross welfare loss is the difference between the posterior utility of the
prior act and the prior utility of the prior act. This loss has two sources,
(1) environmental contamination, and (2) being unable to change behavior

in response to contamination.

2. Information allows people to choose behavior consistent with informed per-
ceptions of risk. Information has value because it prevents the continuation
of mistaken prior behavior. The ex post value of a signal is the difference
between the posterior utility of the posterior act and the posterior utility of
the prior act. Information value equals the change in gross welfare loss

associated with the opportunity to change behavior.

3. The opportunity to engage in risk averting behavior reduces the magnitude
of gross welfare loss to a welfare change net of information value. Net wel-

fare change (equation 2.19) is the difference between gross welfare loss and

26

information value. It is measured by the difference between the posterior

utility of the posterior act and the prior utility of the prior act.

The gross welfare loss associated with changes in environmental contamination and
ignorance (equation 2.17) is unambiguously nonpositive. Information value is nonnegative.
The net welfare change (equation 2.19) may be positive or negative. A'signal about environ-
mental contamination may cause an individual to forego valued activities in order to reduce
risk. A number of studies document sales or consumers' surplus losses as a result of news
about food or environmental contamination (Shulstad and Stoevener 1978, Swartz and Strand
1981, Foster and Just 1989, Smith et al. 1988, van Ravenswaay and Hoehn 1990, Mowen
1980, Brown 1969, Hamilton 1972, Schuker 1983). Environmental improvements may cause a
positive net welfare change. Improved environmental quality may encourage an individual to
give up some risk averting behaviors and reacquire foregone valued activities. These individu-
als will enjoy greater expected utility after being informed of the improvement than before -
they will experience a net welfare gain.

Pope (1985) and Graham-Tomasi (1988) mistakenly claim that ex post signal value may be
negative. They both confuse ex post signal value with the net welfare change of equation 2.19.
The similarity of the two definitions of ex ante information value contributes to this confusion.
The authors mistakenly try to extract a measure of ex post signal value from the definition of

information value given by equation 2.12.

Determinants of Information Value

Information value depends on, (1) prior risk, (2) behavioral change, (3) the utility func-
tion, and (4) the information system itself (Hilton 1981). Gould (1974) and Hess (1982) show
that the effect of prior risk on information value is indeterminant without restricting the form

of the utility function. Specifically, Gould shows that increased risk in the Rothschild-Stiglitz

27

(1970) sense 2 increases information value only when utility is linear in states. Furthermore,
even though increased Rothschild-Stiglitz risk implies a mean preserving increase in variance,
increased variance of the distribution of priors does not imply increased information value.

The relationship between information value and behavioral change is straightforward -—
inforrnation has value only if it changes behavior. Consider the definition of ex post signal
value given by equation 2.10. If behavior does not change, qy=q‘, and the expression reduces
to zero. Since ex ante information value is the expectation of ex post signal values, informa-

tion has no value ex ante if behavior is unchanged for every signal.

Proposition (Behavioral Change) - Ex post, a signal has no value if it does not
change behavior. Ex ante, information has no value if there is no possibility of

behavioral change in response to any signal.

The relationship between information value and the utility function is also relatively
straightforward. Consider the ex post value of a signal given by equation 2.10. The derivative
of ex post signal value with respect to the expected value of the mistake (the second term) is
negative. Thus, the lower the expected utility associated with the mistaken prior optimal act,
the greater the value of a signal.

The ex ante value of information is the expectation of ex post signal values. Since the
value of each signal is nonincreasing in the expected utility of the prior act, the ex ante value

of information is also nonincreasing in the expected utility of the prior act.

Proposition (Utilithof Mistake) - The value of information is nonincreasing in the

utility of the possibly mistaken prior optimal action.

 

3 A distribution Fl(x) is Rothschild-Stiglitz riskier than a distribution F2(x) if and only if the distributions have the
same mean but F:(x) has more weight near the mean.

28

Finally, the value of information depends on characteristics of the information system
itself. Control over specific aspects of an information system provide the greatest opportunity
for an information provider to influence the value of information to users. The links between
information system characteristics and value are the key results of this chapter. The following
section reviews the relationship between the fineness and informativeness of information and

its economic value.

The Value of Fineness and Informativeness

Information systems can be ordered by their "informativeness". This section shows that
informativeness is synonymous with economic value. Informativeness measures the
probabilistic strength of the association between a signal and a given state. A perfectly
informative signal identifies with certainty which state will occur. A perfectly informative
information system contains only perfectly informative signals. Perfect informativeness is
necessary but not sufficient for perfect information. Perfect information must also identify
states that are relevant to decisions (Nermuth 1982). The concept of fineness measures the
relationship between signals and decision relevant states (Marschak and Miyasawa 1968).
Fineness is a special case of informativeness.

The following example illustrates the concepts of informativeness, fineness, and decision
relevance in the context of the current public health advisory. With respect to mercury,
Michigan issues a warning to restrict consumption to no more than one meal per week if, (I)
the site is tested, and (2) mercury concentrations fall between .5 and 1.5 parts per million
(ppm). The advisory suggests no consumption if, (1) the site is tested, and (2) concentrations
of mercury exceed 1.5 ppm. Finally, the advisory does not issue any specific warning if, (1)
the site has not been tested, or (2) the site was tested but mercury concentrations were below .5
ppm. Thus, the advisory issues three signals relevant to mercury contamination. (I) no con-

sumption, (2) restrict consumption. and (3) no report.

29

Suppose anglers accept that the concentrations used to trigger different signals are relevant
to their utility and behavioral choices. Therefore, the signals issued by the advisory identify
the states that are relevant to anglers' decisions. Table 2.1 illustrates likelihoods that an angler
might assign to the advisory signals. It also shows the posteriors implied by the likelihoods

and prior beliefs.

Table 2.1 - Informativeness

Likelihoods P(Signal lState)

 

 

 

 

 

 

 

 

 

States Signals

No Restrict No

Report Consumption Consumption
(NR) (RC) (NC)

5 .5 ppm P(NRI S .5) P(RCI S .5) P(NCI 3.5)
t = .8 = .2 = 0

.5-1.5 ppm P(NRI .5-1.5) P(RC|.5-1.5) P(NC | .5-1.5)
= .8 = .2 = 0

21.5 ppm P(NRI21.5) P(RCI215) P(NCI215)
= _3 = O = .2

 

 

The likelihoods of Table 2.1 reflect varying degrees of informativeness. The angler in the
example believes the "no report" signal to be equally likely for each state. The likelihoods
imply that the "no report" signal provides no information about which state is more likely -
the signal is uninformative (Nermuth 1982). Since the "no report" signal provides no informa-
tion about the probability of occurrence of states, the posteriors associated with the signal
equal prior beliefs.

The "restrict consumption" signal is more informative. The angler believes the signal to

be issued with probability .2 when measured concentrations are between .5 and 1.5 ppm, with

30

Table 2.1 - (Continued)

Posteriors P(State | Signal)

 

 

 

 

 

 

 

 

 

 

 

 

 

States Signals
No Restrict No
Report Consumption Consumption
(NR) (RC) (NC)
5.5 ppm P(S.5|NR) P(S.5|RC) P(S.5|NC)
= .8-P(s.5) .2-P(s.5) = O-P(s.5)
P(NR) P(RC) P(NC)
=P(S.5) < P(s.5) =0
.5-l.5 ppm P(.5-1.5|NR) P(.5-1.5|RC) P(.5-1.5|NC)
= .8 - P(.5-1.5) = .2 - P(.5-1.5) ___ 0 - P(.5-l.5)
P(NREP) P(RC) P(NC)
= P(.5-1.5) < P(.5-1.5) = 0
21.5 ppm P(21.5|NR) P(21.5|RC) P(21.5|NC)
_ .8-P(21.5) = 0-P(21.5) = .2-P(21.5)
P(NREP) P(RC) P(NC)
= P(21.5) = 0 = l

 

 

 

P(NR) = .8 - P(NRI 5.5) + .8 - P(NR|.5-l.5) + .8 ° P(NRI 21.5)

P(RC) = .2 - P(RCI 5.5) + .2 - P(RC|.5-l.5) + o - P(RCI 21.5)

P(NC) = o - P(NCI $5) + o - P(NCI .5-1.5) + .2 - P(NCI 21.5)

 

31

probability .2 when concentrations are below .5 ppm, and with probability zero when con-
centrations exceed 1.5 ppm. Receipt of the "restrict consumption" signal provides some
information - it implies that actual concentrations are equal to or below 1.5 ppm. However, it
provides no information to distinguish between the states of 5.5 ppm and .5-1.5 ppm.
Because it contains some information, the "restrict consumption" signal alters prior beliefs. It
increases the perceived probability that concentrations are either less than .5 ppm or between
.5 and 1.5 ppm and eliminates the chance that concentrations are above 1.5 ppm.

Finally, the angler believes that the "no consumption" signal is perfectly informative —- it
reveals with certainty that concentrations exceed 1.5 ppm. The posteriors associated with the
signal reflect this certainty.

This example demonstrates that the informativeness of a signal does not depend on
absolute likelihoods. Instead it depends on relative likelihoods associated with a signal. The
distribution of likelihoods over states for a given signal determines the signal's influence on
prior beliefs. The example also links the earlier notion of confidence (loosely measured by the
spread of a distribution) to the concept of informativeness. The distribution of likelihoods for
the uninformative "no report" signal is flat (McGuire 1972) or diffuse (W inkler 1972). This
terminology implies a uniform distribution with no concentration of probability on any state.
An uninformative signal is synonymous with no confidence. Similarly, anglers have complete
confidence in the perfectly informative "no consumption" signal. The completely concentrated
distribution of likelihoods reflects this confidence. This research uses the term informative-
ness.

The example above assumed that the states identified by signals are relevant to anglers'
behavioral choices. Suppose instead that anglers believe they can safely eat one meal per
month of fish with mercury concentrations between 1.5 and 2.0 ppm. The existing "no con-
sumption" signal does not distinguish finely enough between contamination levels to be entirely
decision relevant to these anglers. Table 2.2 illustrates the likelihoods and posteriors associ-

ated with decision relevant states when signals remain as they appear in Table 2.1. In

32

Table 2.2 - Fineness

Likelihoods P(Signal lState)

 

 

 

 

 

 

 

States Signals

No Restrict No

Report Consumption Consumption
(NR) (RC) (NC)

s .5 ppm P(NRl s .5) P(RC] 5.5) P(NC| 3.5)
= .8 = .2 = 0

.5-1.5 ppm P(NRI .5-1.5) P(RC|.5-l.5) P(NC| .5-l.5)
= .8 = .2 = 0

1.5-2.0 ppm P(NRI1.5-2.0) P(RCI1.5-2.0) P(NC] 1.5-2.0)
= ,3 = O = .2

22.0 ppm P(NRI 22.0) P(RCI 22.0) P(NC| 22.0)
= ,3 = 0 = .2

 

 

 

 

 

this example, the "no consumption" signal is no longer perfectly informative because it does
not distinguish between decision relevant states.

The above examples illustrate that informativeness depends on both the ability of a signal
to identify a state 3 and on the description of identified states relative to states that are decision
relevant to information users. The remainder of this section relates the notions of fineness,

decision relevance, and informativeness to information value.

 

3 The state that a signal identifies is irrelevant. For instance, a "no consumption" signal that anglers perceive

identifies a site with mercury concentrations below .5 ppm 90 percent of the time is still quite informative.

33

Table 2.2 - (Continued)

Posteriors P(State | Signal)

 

 

 

 

 

 

 

 

 

 

 

States Signals
No Restrict No
Report Consumption Consumption
(NR) (RC) (NC)
P(3.5|NR) P(3.5|RC) P(3.5|NC)
3.5 _ .8-P(3.5) = .2-P(3.5) = 0-P(3.5)
ppm P(NR) P(RC) P(NC)
=P(3.5) > P(3.5) =0
P(.5-1.5 | NR) P(.5-1.5 | RC) P(.5-l.5 | NC)
.5-1.5 = .8 - P(5.-1.5) = .2 - P(.5-1.5) = 0 - P(.5—1.5)
ppm P(NR) P(RC) P(NC)
= P(.5-1.5) > P(.5-1.5) = 0
P(l.5-2.0|NR) P(l.5-2.0|RC) P(l.5-2.0|NC)
1.5-2.0 = .8 - P(l.5-2.0) _ 0 - P(l.5-2.0) = .2 - P(l.5-2.0)
ppm P(NR) P(RC) P(NC)
= P(l.5-2.0) = 0 > P(l.5-2.0)
P(22.0|NR) P(22.0|RC) P(22.0|NC)
22.0 = .8 - P(22.0) = o - P(22.0) = o-1>(22.0)
ppm P(NR) P(RC) P(NC)
= P(22.0) = 0 > P(22.0)

 

 

 

P(NR) = .8 - P(NRI 5.5) + .8 - P(NRI .515)
+ .8 - P(NR|1.5-2.0) + .8 - P(NRI 22.0)

P(RC) = .2 - P(RCI 5.5) + .2 - pacts—1.5) + o - P(RC] 1.5-2.0)
+ o - P(RCI 22.0)

P(NC) = 0 - P(NC| 3.5) + o - P(NC|.5-1.5) + .2 - P(NC|1.5-2.0)
+ .2 - P(NC| 22.0)

 

34

Fineness

Marschak and Radner (1972) develop a comparison of information systems on the basis of
fineness. The concept of fineness defines the description of signals. Marschak and Radner
define an information system Y to be finer than a system X if at least two signals in Y are sub-
sets of at least one signal in X. They prove that Y is at least as valuable as X if Y is finer than
X. The following analysis reviews the comparison of information values on the basis of fine-
ness. It then relates fineness to decision relevance.

Suppose that signals from X are related to signals from Y in the following manner.

x1 = {Yi’y4}
x2 = {Y2}
x3 = {Y3}

An individual who observes X can distinguish between three possible states and has three
possible behavioral responses. An individual who observes Y can distinguish between four
possible states and has four possible behavioral responses which include those associated with
observing X. Thus, a person who observes Y can be no worse off than a person observing X.
The concept of fineness imposes a partial ordering on information systems. The ordering is
only partial because information systems that use noncomparable partitions of the state space

can not be ordered (Marschak and Miyasawa 1968).

Proposition (Fineness) - Information value is nondecreasing in the fineness of signals

that identify decision relevant states.

This result has implications for the relative value of information systems that combine or

separate signals. Combining two or more signals in an information system into a single deci-

35

sion relevant signal can not increase the value of the system. Similarly, separating a signal
into two or more decision relevant signals can not decrease the value of an information system
(Marschak and Miyasawa 1968).

The above result applies only if the signals yl and y4 are decision relevant. Suppose the
signal x1 is decision relevant but yl and y, distinguish between states too finely to affect an
individual's utility. If the individual can costlessly conclude that yl and y4 provide the same
information as x1, then systems X and Y are equally valuable. However, information users
may incur monetary and nonmonetary costs associated with combining signals. For example.
it may take time, money, and significant cognitive effort to determine the decision relevant
states implied by irrelevant signals. These costs affect the expected utility associated with
information (Marschak and Miyasawa 1968).

To illustrate, include in an angler's utility function an argument that reflects the effort
required to interpret information, e(~). The ex ante expected utility associated with information

system Y is then

(2.20) I I u(q’,h(q’,s),e(y)) P(SIY) ds P(y) dy

where e(y) is the effort required to interpret or make use of signal y. Suppose e(y) is mini-
mized when Y contains all possible decision relevant signals. If Y aggregates some decision
relevant signals, more effort is required to identify decision relevant states. Similarly, if Y
issues signals that are finer than decision relevant states, information users may need to expend
effort to aggregate the signals to make them decision relevant. Furthermore, assume that util-
ity is nonincreasing in the effort required to use information - du(-)/de(-) 30.

Assume that systems Y and X define comparable partitions of the state space and that Y is
at least as fine as X. Furthermore, suppose signals from an information system Y require

more effort to use than those from a system X. Thus, e(y) 2e(x) V y; x. Then

36
(2.21) I J u(q".h(q",s),e(x)) P(ij) ds P(x) dx

2 j j u(qi.h<qy.s),e(y» P(sly) as P(y) dy

and X is at least as valuable as Y. 4 Marschak and Miyasawa (1968) emphasize that it is not
possible to order information system values without knowing e(-) and its effect on u(-).

Providing additional information may also cause people to forget other relevant informa-
tion. Magat et al. (1988) conclude that people may systematically forget some aspects of haz-
ard warnings if presented with too much information. This result suggests that effective
information programs must be carefully designed to provide information that is relevant to con-
sumer decisions without overloading them with irrelevant or confusing information.

The preceding discussion implies that adding signals to an information system does not
necessarily increase information value. Additional signals may actually reduce value if, (1)
they do not identify decision relevant states, (2) they increase the effort required to use
information (whether they are decision relevant or not), or (3) they overwhelm other important

signals.

 

‘ An information system X is at least as valuable as a system Y if the expected utility of the Bayes' strategy asso-

ciated with X is greater than or equal to that associated with system Y. Using the definition of ex ante information
value given in equation 2.8, information system Y is at least as valuable as system X if

I I U(q’.h(q’.s)) P(sly) ds P(y) dy - I U(q'.h(<f'.8)) P(s) ds

2 I I u(q".h(q',s)) P(slx) ds P(x) dx - I u(q',h(q',s)) P(s) ds

or

[ I u(q’,h(q’,s)) P(sly) ds P(y) dy

2 J J u(q‘.h(q",s)) P(slx) ds P(x) dx.

Thus. system X is at least as valuable as system Y if the expected utility of the Bayes' strategy associated with X
is not less than that associated with Y.

37

Informativeness

Fineness is a special case of a more general characteristic of informativeness (Marschak
and Miyasawa 1968, Marschak and Radner 1972). Intuitively, information should be more
informative as signals become more accurate predictors of states (Hilton 1981). Hilton's use of
the term accuracy with regard to informativeness can be misleading. To illustrate. consider the
example of a weather forecast service. The common interpretation of the term accuracy
implies that a forecast of rain most often predicts a state of rain. However, a forecast of rain
that is most often followed by dry weather, while being inaccurate, is still informative with
regard to decision making -- it informs an individual that dry weather is most likely.

With respect to a matrix of likelihoods such as Figure 2.1, increasing accuracy implies
increased diagonal elements and decreased off-diagonal elements. 5 Perfect accuracy implies
diagonal elements equal to one and off-diagonal elements equal to zero. Perfectly informative
information also implies likelihoods of zero and one. However, the ones need not lie on the
diagonal - signals do not necessarily predict their corresponding states. Therefore, perfectly
accurate information is a special case of perfectly informative information. This section
reviews the formal conditions under which one information system may be said to be at least as
informative as another. Accuracy as defined above is irrelevant to informativeness and value.
Therefore, the remainder of this report uses the term informativeness.

Blackwell (1953) defines relative informativeness in terms of stochastic transformations.
Suppose signals from an information system, Y={y,,y2,...,y,}, are not directly observed.
However, an information system, X={x,,x2,...,x,}, yields signals that are related to signals
from Y. An example might be the case where signals are relayed (perhaps imperfectly)

through a third party. Define h(y, x) as the function that relates signals from Y to signals from

 

5 This illustration assumes a likelihood matrix organized such that the signal of column n predicts the state of row

n.

38

X. This function has the same interpretation as the likelihood function -- it generates condi-
tional probabilities, P(xly), of receiving final signal x when the initial signal is y.
Combining probabilities P(xly) with the likelihoods associated with Y yields the

probabilities P(xls) that relate signals from X to states.

Y

(2.22) P(x I s) = 21% l 8) P(x | y)

To ensure that these likelihoods satisfy the restriction that likelihoods sum to one across signals
for a given state, h(y, x) must be nonnegative and for each y E Y, h(y, -) must be a

probability measure over X (Nermuth 1982). This implies

Y
(2.23) 23 P(xly) = 1.
y=l

The function h(y, x) is a stochastic transformation if it is nonnegative and satisfies equa-
tion 2.23. The stochastic transformation, in effect, introduces an additional layer of
uncertainty between signals from Y and states. Thus, in addition to imperfections in the
information system Y, there may be imperfect translation between signals from X and signals
from Y. Given this interpretation, system Y should be no less informative than X.

Blackwell demonstrates that an information system Y is at least as informative as a system
X if, (1) X is derived from Y as described by equation 2.22, and (2) the function generating
probabilities P(xly) is a stochastic transformation. Thus, information systems related through
stochastic transformations can be ranked on the basis of informativeness.

Information value is nondecreasing in informativeness (Marschak and Miyasawa 1968,
Kihlstrom 1984, Blackwell and Girschik 1954). The following proof (presented in summation
notation for convenience) is based on Marschak and Miyasawa.

Consider information systems Y={y,,y2,...,y1} and X={x,,x2,...,xK}. Further, suppose
that system Y is at least as informative as X. This implies a stochastic transformation h(y, x)

that generates probabilities P(x 1 y) such that

39

I
(2.24) P(xklsj) = 221130118.) P(thy.)

and

I
gently.) = 1.

In summation notation the expected utility of the Bayes' strategy associated with system Y

is
l 1
(2.25) u,,= ZIP(y.) ZIP(sjly) net-”tsp
i: j:
l 1
= )3 E P(sj) P(yilsj) u(qY‘,sj)

1 j 1

where qyi is the optimal act given signal yi.

Similarly, the expected utility of the Bayes' strategy associated with system X is

K I
(2.26) u,,= E E P(sj) P(xklsj) u(qus.)
k=1 j=I
K J I
= E B P(sj) 21>(yilsj)1>(x,|y,)u(qu,sj) (using equation 2.24)
it=1 i=1 i=1 .

where q“ is the optimal act given signal xk.

Rearranging terms yields

1 J
2: Italy.) Elpts,)1>(y.ls,)u(qe.s,).

l i

‘ K
(2.27) 11,, = 23
k=

I

By definition, qyi is the optimal (expected utility maximizing) action given likelihoods

associated with signal yi. Thus

40

(2.28) uh, 3 uby

01'

l 1
E: P(xtly.) gas) Peas.) u(qes).

l i

E.

l

J
P(x. | ya); P(S,) P(y. I 8,) U(q".5,)
,2

l/\
iil‘das
M...

Iil

Using the restriction Bi P(xk l yi) = 1 from equation 2.24, the second term reduces to

l

(2.29) Y P(s,) Po. 1 5,) u(q>",s,)

‘Ttl‘dc.
M...

'lil

which is the expected utility of the Bayes' strategy associated with system Y, mutliplied by a
constant that is less than one. Thus, from an earlier result (see footnote on page 36), system Y

is at least as valuable as system X.

Proposition (Informativeness) - Information value is nondecreasing in informative-

IIBSS .

This proposition holds for all prior distributions and all payoff functions.

Summary and Implications

The analysis of this chapter demonstrates that information can not make people worse off.
Information has positive ex post value if it causes a change in behavior. Ex ante, information
has positive value if there is a possibility that it will change behavior. Information value is
zero only if no behavioral change is possible. This result implies that information systems
should distinguish only between states that matter to information users -- the states should be

decision relevant. Information users will likely differ in the states considered to be decision

41

relevant. Information system design will benefit from knowledge of the user's perceptions and
information needs.

The chapter also shows that increasing the number of signals associated with decision rele-
vant states can not decrease the value of an information system. This result assumes that all
signals are equally costly to use. However, information systems may produce signals that dif-
fer in the cognitive, monetary, or time costs required to interpret or use them. Other things
equal, an information system that produces signals that are costlier to use will have lower
value. In terms of information system design, this result implies that the mental effort,
monetary costs, and time associated with interpreting and using information should be con—
sidered.

The analysis also concludes that information value is nondecreasing in informativeness.
Informativeness is a measure of the identifiability of states from signals. This result implies
that value can be enhanced by strengthening the perceived probabilistic link between a signal
and a particular state. To the extent that an information provider can increase users' percep-
tions of the informativeness (confidence) of signals, information value can be increased.

Finally, information value is nonincreasing in the perceived utility associated with making
a mistake. In the context of chemical residues in food or in the environment, this implies that
particularly vulnerable groups may value information more highly. An information provider

may wish to make a special effort to inform such groups.

Chapter Three

The Value of Alternative Public Health Advisories

This chapter applies the framework of the previous chapter to an evaluation of anglers'
willingness to pay for changes in Michigan's public health advisory. It first links the utility
based definition of information value to a monetary willingness to pay (WTP) measure based
on the expenditure function. It then examines the value of two proposed changes in the format
of Michigan's current public health advisory (1) testing a. greater number of sites, and (2)
including a list of tested sites where chemical residues pose little or no health risk. The analy-
sis links advisory alternatives to changes in informativeness and fineness and thus to changes in
value.

Beliefs about information accuracy also influence the perceived link between states and
signals. This chapter defines accuracy in the context of the public health advisory. It then
shows that increasing accuracy is synonymous with increasing informativeness (value.) The
chapter also develops the relationship between advisory value and the perceived severity of
health effects. The primary objective of the chapter is to develop specific, testable hypotheses

regarding the value of alternative advisory formats to individual anglers.
Anglers' Willingness to Pay for Information
The utility anglers enjoy from fishing depends on various aspects of the fishing experi—
ence. These may include the species, number, and size of fish caught (Vaughan and Russell

1982), characteristics of a fishing location (Graham-Tomasi 1986), or the presence or absence

42

43

of chemical residues in fish (Kikuchi 1986). In addition, anglers would be expected to posi-
tively value a present and future state of good health.

The presence of chemical residues in some of Michigan's sport fish implies that choices
about specific characteristics of fishing may affect anglers' health. The public health advisory
suggests that the risk of health effects depends in part on the size or species of fish consumed,
the frequency of consumption, the choice of fishing sites, the age and gender of the consumer,
and methods of preparing fish for consumption. Therefore, anglers must balance the utility of
specific aspects of the fishing experience against the value of perceived changes in health that
may result from different actions.

Define utility as a function of specific fishing behaviors, q= {q,,q2,...,qL}, and the state of
health, h. The expected state of health is a function of the state of contamination that is per-

ceived to exist, 5, and the behaviors chosen. Thus, the utility function is
(3.1) u = u(<1. h((i. 8))

The level of contamination, s, is not known. Therefore, the state of health, h(-), is a ran-
, dom variable. Furthermore, the functional relationships between behavioral choices and
exposure and between exposure and health effects are subject to a great deal of uncertainty
(Bro et al. 1987, Paradis et al. 1988). Anglers are assumed to choose fishing activities to max-
imize expected utility with respect to the perceived probabilities of different states and subject
to a budget constraint.

The previous chapter derived measures of information value and welfare changes by maxi-
mizing the expectation of the utility function of equation 3.1. The resulting measures defined
value in terms of utility. The remainder of this section defines the utility measures of the
previous chapter in terms of the expenditure function. The expenditure function yields
monetary measures of information value and welfare change.

The dual problem is to choose fishing behavior to minimize expenditures subject to a con-

stant level of expected utility. The expenditure function

(3.2) e = e(p", u°, P(s))

44

where p°={p,,p2,...,pL} is the initial vector of prices associated with attributes q, u0 is the
level of expected utility, and P(s) represents the prior distribution of s, is the solution to the
dual minimization problem. The expenditure function represents the minimum income neces-
sary to attain utility u° given prices p° and risk perceptions P(s) (Deaton and Muellbauer
1980).

Since fishing in public waters is not explicitly sold in a market, the price associated with
different fishing activities, p°, is not observable. However, anglers do incur costs in pursuing
their sport. These may include travel costs, boat rental, charter fees, bait and tackle costs, and
the value of time. These costs may vary across activities. For example, fishing for salmon or
trout on the Great Lakes requires access to a boat and specialized equipment. Also, an angler
may incur greater travel and time costs to reach sites with certain species or aesthetic charac-
teristics. License fees also differ by species. In Michigan a special stamp is required to catch

salmonoids (trout and salmon).

Welfare Effects of Contamination and Information

Contamination of sport fish can cause a welfare loss for anglers. The gross loss due to
contamination is the welfare change associated with an increase in risk when behavioral adjust-
ment is not permitted. An expenditure function that restricts behavior q to q° can not be
derived from empirical demand functions. However, the restricted expenditure function can be
defined in terms of the unrestricted expenditure function (Foster and Just 1989). Consider a
price vector p‘ that maintains u° and q° when a signal changes perceived risk from P(s) to

P(s | y). The restricted expenditure function can then be expressed as
(3.3) e(p°,u0,P(8|y),q°) = e(p‘.u0,P(SIy)) + (p0 - p‘)q°

where (p0 - p‘)q0 is an adjustment for the change in actual expenditures required to purchase

q°. Equation 3.3 is a monetary measure of the utility associated with a mistake. The monetary

45

measure of the ex post gross loss of welfare, V8, associated with contamination is thus

(3.4) v, = e(p°, u", P(s|y).q°) - e(p°, uO, P(S),q°)

e(p‘. U°, P(s I y), (1‘) + (11° - 13‘) (1° - e(p°, U". P(s). (1°)

601‘, 11°, P(Sly), q‘) - m0 + 0)" - 9‘) (1°-

This welfare effect is the change in income required to maintain utility u° when perceptions of
risk change from P(s) to P(s|y) but consumption is restricted to q“. Equation 3.4 is the
monetary equivalent of the gross welfare loss defined in utility terms in the previous chapter.
Ignorance restricts anglers to uninformed behavior. Information has value because it
permits anglers to adjust optimally to changed levels of risk. The ex post value of a signal that
causes a change in risk perceptions from P(s) to P(s I y) is the difference in the utility associated
with optimal adjustment and that associated with being forced to continue prior behavior in

light of changed perceptions of risk. The monetary measure of ex post signal value

(3.5) V

t e(p°, 11°. P(sly),q°) - e(p°, 11", P(s|y). q‘)

= e(p‘, u°. P(s|y)) - e(p°, u°, P(s|y). q‘) + (11° - p‘) (1"

is the difference in expenditure required to maintain initial utility under posterior risk percep-
tions when ignorance makes behavioral adjustment impossible.

Information value is nonnegative because the expenditure required to maintain u° when
consumption is restricted must be at least as great as when consumption is allowed to adjust.
For risk averters, the value of information is also nondecreasing in the relative riskiness of the
prior and posterior distributions over states, 5. Define a distribution, P(s), to be riskier than
P(s|y) when all risk averters prefer P(s|y) to P(s) (Rothschild and Stiglitz 1970). The greater
the difference in expected utilities associated with P(s) and P(s|y) (the greater the riskiness of
P(s) relative to P(s|y)), the greater the difference between prior and posterior consumption, q1
and q°, and the greater the difference between e(p°, u°, P(s|y),q°) and e(p°, u°, P(s l y), q‘).

The nonnegative value of information can not increase the gross welfare loss associated

with contamination. A monetary measure of the net welfare loss is the change in income

46

necessary to maintain it0 at prices p". This measure is net of information value because it
permits behavioral response to information. In terms of the expenditure function the net wel~

fare loss, V“, is

(3.6) v, = e(p°, U“, 9‘, q‘) - e(p°, 11", 9°, (1°)

e(p°, 11". 9‘. q‘) - m°

where m0 is initial income. The net welfare loss can also be derived by subtracting the gross
welfare loss (equation 3.4) from information value (equation 3.5). The net welfare change
measure implies that anglers are able to make behavioral adjustments to changes in risk percep-

tions. This research measures the value of information given in equation 3.5.

Advisory Characteristics and Value

Of the determinants of information value, information providers have the most control
over characteristics of the information system -- the fineness and informativeness of signals.
This research asks how the state can enhance the value of the public health advisory. There-
fore, the most important hypotheses are those that relate characteristics of information systems
to value.

Michigan's current public health advisory contains limited information. Of more than
5,800 fishing sites in the state, only 350 have ever been tested for the presence of chemical
residues. The MDNR tests about 30 new sites each year. The advisory informs anglers only
of the sites where chemical residues were found to exceed state .standards. It does not tell them
about relatively safe sites where testing found no chemical residues or residues below state
standards.

An advisory program that tells anglers only about unsafe sites discloses only part of the
collected information. Define such a program as one of "partial disclosure." A program that
also tells anglers about relatively safe sites fully discloses available information. Define such a

program as one of "full disclosure." The remainder of this report uses the terms partial and

47

full disclosure to describe respectively an advisory program that does not list safe sites and one
that does.

This research explores the value of testing a greater number of sites annually under both
partial and full disclosure advisories. Ex ante, nonzero information value depends on the pos-
sibility of behavioral change. If an angler would not change behavior for any possible signal
provided by the advisory, the advisory has no value. This section first explores situations
where the advisory may have no value. It then shows that full disclosure does not decrease the
fineness (value) of the advisory. It also shows that testing more sites does not decrease the
informativeness (value) of the advisory. It concludes with a statement of specific hypotheses

regarding the relative value of partial and full disclosure advisories.
Advisory Value and Behavioral Change

Information value is nonnegative. It is strictly positive ex ante only if (1) at least one of
the possible signals would cause an ex post behavioral change, and (2) at least one of the sig-
nals that would cause behavioral change has a nonzero probability of occurrence. This section
identifies groups of anglers for whom the advisory would not cause a change in behavior and
thus would have no value.

First, an angler will not find the advisory valuable if unaware of it. Furthermore, even if
aware of the advisory, an angler may not consider changing behavior if current behavior is
perceived to entail little or no risk. This group might include catch and release anglers who do
not eat fish, individuals who eat only species known to be safe from a tested site, or those who
do not believe in or do not care about the advisory. Also, anglers who believe with certainty
that current and anticipated future behavior poses no health risk will not anticipate receiving a
signal that would cause a behavioral change -- the probability of receiving a useful signal is
perceived to be zero.

Another group of anglers who might not change behavior in response to advisory informa-

tion are those for whom the utility cost of behavioral change is high. For these people. other

48

considerations may be more important than health concerns. For example, an angler may
choose a particular site for convenience -- it may be close to home or a favorite vacation spot.
An individual may also choose a site, species, or consumption pattern because of tradition.
The convenience or tradition associated with a site, species, or consumption pattern may out-
weigh health considerations. These anglers may not consider changing behavior regardless of
specific advisory warnings. ‘

Other anglers may have few affordable options for inexpensive food sources other than
sport caught fish. For example, some anglers may depend on sport fishing for food. These
individuals may also be unable to afford the time or travel costs to fish at a different site if the
advisory listed a currently used site as unsafe. For these anglers the consideration of nutrition
may outweigh health concerns.

This subsection identified groups of anglers who may not change behavior in response to
any type of advisory information. However, many anglers would anticipate a behavioral
change in response to the advisory. The remainder of this section examines how advisory
alternatives influence value for those who would consider a behavioral change. The analysis of
each alternative specifically considers the decision context and type of behavioral change the

alternative might bring about.
Value of Full Disclosure

The current advisory contains two signals about chemical residues at specific sites. First,
it issues a warning to "restrict consumption" if the site has been tested and chemical residues

have been found to exceed state standards. 2 Second, the advisory issues an implicit "no

 

‘ Kihlstrom (1974) finds that people who's preferences are strongly skewed in favor of one product or another are

less likely to demand information about product quality. This result implies that those with a strong preference for a
particular behavior (perhaps because of convenience or tradition) are less likely to change behavior and are thus less
likely to find information valuable.

3 In this discussion. "restrict consumption" encompasses both restricted signals from the current advisory - the
signal to eat no more than one meal per week and the signal to eat none of a certain species.

49

report" signal if (1) a site has not been tested, or (2) the site was tested but chemical residues
did not exceed state public health standards. This section shows that full disclosure of test
results does not decrease the fineness of information and thus does not decrease the value of
the advisory. It also provides an intuitive description of how full disclosure increases an
angler's welfare.

Sites that receive the "no report" signal in the current advisory fall into three categories.
They may be (1) tested and found to pose little or no health risk, (2) not tested and pose little
or no health risk, or (3) not tested and contain residues above state standards. A full dis-
closure advisory program creates a separate signal for sites in the first category. Such an
advisory would contain two explicit signals, "restrict consumption" and "no consumption
restriction", and an implicit signal of "no report". The "no consumption restriction" signal
would apply to tested sites that pose little or no health risk.

The "no consumption restriction" signal applies to a subset of the sites that receive a "no
report" signal under the current advisory. Therefore, from the definition in the previous chap-
ter, a full disclosure advisory provides finer information than the current advisory. Thus, the

addition of a signal of safety can not decrease the value of the advisory.

Hypgthesis (Eull Disclgsure) - A full disclosure advisory program will be at least as

valuable as a partial disclosure advisory.

Consider how listing safe sites actually improves an angler's welfare. Assume the angler
does not belong to the previously identified groups who would not consider a behavioral
change in response to the advisory. Further, suppose the angler wishes to avoid eating too
much contaminated fish. The quantity that constitutes "too much" may correspond to sug-
gested consumption restrictions from the advisory. It may also depend on personal percep-
tions.

Under the current advisory, an angler who wishes to avoid eating contaminated fish would

choose to fish at a "no report" site unless other considerations were more important. If fishing

50

at an unsafe site, the angler may choose not to eat the fish or at least choose to limit consump-
tion. The willingness to make behavioral choices in response to signals implies that the angler
attaches a strictly positive ex ante value to the advisory information.

Under the current partial disclosure program there is a chance that a "no report" site is
really unsafe. Therefore, there is a chance that the choice to eat fish only from "no report"
sites is a mistake. The addition of a "no restriction" signal to the advisory greatly increases
the chance of correctly identifying a safe site. Thus, it reduces the chance of making the mis-
take of eating too many contaminated fish.

This analysis also applies to an angler who decides not to eat fish from any site because of
the chance of eating contaminated fish. For this angler, continuation of the current advisory is
not valuable -- behavior is unchanged regardless of the signal. The addition of a "no consump-
tion restriction" signal increases the chance of correctly identifying a safe site and reduces the
chance of mistakenly eating contaminated fish. Greater certainty about the chance of con-
taminants at a site may cause the angler to begin eating fish from "no restriction" sites. For
this angler, a full disclosure advisory may have value while a partial disclosure program would

not.

Value of Testing More Sites

This section demonstrates that testing more sites can not reduce the value of either a full
or partial disclosure advisory program. The analysis first derives likelihoods for both advisory
formats. It shows that likelihoods depend on the probability of testing. It then models a
reduction in the probability of testing as a stochastic transformation of the likelihoods associ-
ated with reduced testing. The stochastic transformation implies that a program with more
testing is at least as valuable as a program with reduced testing. The section also provides an
intuitive illustration of how increased testing might increase advisory value.

The probability tree of Figure 3.1 illustrates how anglers might form likelihood

probabilities from knowledge of the current testing program. The first column lists the

51

 

 

 

 

 

 

 

 

State Test Partial Disclosure Full Disclosure
State Signals Signals
Restrict Restrict
Tested Consumption Consumption
Unsafe
\ Not Tested No Report No Report
_ 1 - P11" 1 - Pm
No
Consumption
Tested No Report Restriction
Safe
\ Nolt Tiested No Report No Report
1 ' PTS 1 ‘ PTS

 

Figure 3.1 - Determination of Likelihood Probabilities

posible states of contamination at a site -—_ "safe" or "unsafe."

Likelihoods assume that these

states are given. With the current limited testing program, a site may or may not be tested.

The second column lists possible "test states"

believes that the testing procedure is perfect —

Define the perceived probability of testing a safe site as P13.

probability of testing an unsafe site.

- tested and not tested. Assume the angler
tests always correctly identify a site's state.
Similarly, define P1U as the

Since the MDNR targets testing to sites suspected of

being unsafe, P13 does not necessarily equal Pru- Under these assumptions, the probability

that an unsafe site will be tested and found to be unsafe is the probability that it is tested, PTU

The current advisory issues a signal of "restrict consumption"

in this instance. The third _

column lists possible signals associated with the current partial disclosure advisory and their

associated likelihood probabilities.

52

Similarly, the advisory issues a "no report" signal for an unsafe site only when it has not
been tested. This occurs with probability l-Pw. A safe site receives a "no report" signal
whether it has been tested or not. Therefore, the likelihood probability of a "no report" signal
given that a site is safe is the probability of testing the site plus the probability that the site is
not tested. This probability is one. Table 3.1 summarizes the likelihoods and associated
posteriors for the current advisory. The probabilities PU and PS are the prior probabilities of
states of unsafe and safe respectively.

Testing more sites does not change the posterior probability that a "restrict consumption"
signal implies an unsafe site. The signal is perfectly informative regardless of the level of test-
ing. However, increased testing does increase the posterior chance that a "no report" signal
implies a safe site. Increased testing increases the chance that an unsafe site will be identified

and listed as such. Thus, the "no report" sites are more likely those that have been tested and

Table 3.1 - Likelihoods and Posteriors for Partial Disclosure

 

 

 

 

 

 

Likelihoods Posteriors
States Signals Signals
Restrict No Restrict No
Consumption Report Consumption Report
(RC) (NR) (RC) (NR)
Unsafe P(RClunsafe) P(NRlunsafe) P(unsafelRC) P(unsafelNR)
= P111 = l - Pm = 1 = P11(1'Pm)
PU(l-Pm)+Ps
Safe P(RC | safe) P(NR | safe) P(safe | RC) P(safe I NR)
= O = 1 = 0 P

 

S
Pu(1‘Pru)+ Ps

 

 

 

 

 

 

 

53

found to be safe. 3 Therefore, testing more sites increases the chance that a site chosen from
among the "no report" sites is safe. The following analysis uses stochastic transformations to
formally demonstrate that increased testing implies increased informativeness.

In matrix notation, “ the likelihoods associated with the current advisory from Table 3.1

are given by

(3.7)

where row one corresponds to a state of unsafe, row two corresponds to a state of safe, column
one represents the "restrict consumption" signal, and column two represents the signal of "no
report."

Now suppose the state tests fewer sites each year. Less testing reduces the probability
that an unsafe site will be tested. Represent this reduction by multiplying PTU by a positive

constant a belonging to the interval [0,1]. The resulting likelihood probabilities are

(3.8)

0 1

 

3 In the extreme, when the probability of testing is one, the "no report" signal is equivalent to providing a list of

safe sites.

‘ To simplify notation this chapter uses matrices to describe likelihoods and stochastic transformations. Consider
a setting with states 8 = {s,,s2,...,s,) and two alternative information systems, X= {xl,x2,...,xK} and
Y={y,,y2,...,yl}. The likelihoods associated with systems X and Y can be represented by the matrices XL (S x K)
and YL (S x I) respectively. The stochastic transformation from YL to XL takes the form of a matrix H (I x K)

hll h12 th

ha a
ny.X) = -I h22 h..K

hll hIZ hIK

where hik is the probability of receiving signal x“ from X when Y produces yi. The matrix H(y.x) must be a non-
negative, row stochastic (elements of rows sum to one) Markov matrix. This implies that but lies on the interval
[0,1] for each i and k (Kihlstrom 1984). In matrix notation, information system Y is at least as informative (valu—
able) as X if there exists a stochastic transformation H such that X=Y H.

54

The equation

(3.9)

aPTU l-ozPm PTU l-Pm oz l-oz

relates the likelihoods associated with testing fewer sites to the likelihoods associated with the
original level of testing.

The second matrix on the right hand side of equation 3.9 is a stochastic transformation.
Therefore, the likelihoods in the first matrix on the right hand side (the likelihoods associated
with testing more sites) are more informative (more valuable) than the likelihoods on the left
hand side of the equation (the likelihoods associated with less testing). The more informative
likelihood matrix also conforms to the intuitive notion of informativeness. PTU is larger than
ozPTU and l is larger relative to l-Pm than to 1-aPTU. Therefore, with a greater probability of
testing, the "restrict consumption" signal is more likely when a site is unsafe. The probability
of the "restrict consumption" signal for a safe site remains zero. 5 The "no report" signal is
relatively more likely when a site is safe.

Similarly, increased testing can not reduce the informativeness (value) of a full disclosure
advisory. The fourth column of Figure 3.1 lists the signals that a full disclosure advisory
would assign to different state and testing combinations. Likelihoods are the same as those
associated with a partial disclosure advisory except that a tested safe site receives a "no con-
sumption restriction" signal.

Table 3.2 summarizes the likelihoods and posteriors associated with a full disclosure
advisory. Testing more sites does not increase the probability that a "restrict consumption" or
a " no consumption restriction" signal identifies an unsafe or safe site respectively - the signals

are both perfectly informative and posteriors for both signals reflect certainty. However,

 

5 Note that in this example the " restrict consumption" signal is perfectly informative regardless ofthe level of test-

ing.

55

Table 3.2 - Likelihoods and Posteriors for Full Disclosure

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Likelihoods
States Signals
No
Restrict Consumption No
Consumption Restriction Report
(RC) (NCR) (NR)
Unsafe P(RC I unsafe) P(NCR l unsafe) P(NR I unsafe)
= Pm = O = 1 - PTU
Safe P(RC l safe) P(NCR | safe) P(N R l safe)
= 0 = PTS = l - PTs
Posteriors
States Signals
No
Restrict Consumption No
Consumption Restriction Report
(RC) (NCR) (NR)
Unsafe P(unsafe | RC) P(unsafe | NCR) P(unsafel NR)
= 1 = 0 ._.. Pti(1’P'm)
Putl-Pm)+Ps(1-Pm)
Safe P(safe | RC) P(safel NCR) P(safe | NR)
= O = 1 P50 - Pm)

 

 

 

 

= PU(l-PTU)+PS(1- 1T,)

 

 

56

testing more sites increases the chance that a "no report" signal identifies a safe site. The fol-
lowing analysis using stochastic transformations formally proves that more teSting does not
decrease the informativeness (value) of a full disclosure advisory.

Represent a reduction in the probability of testing a safe site, P“, by multiplying by a
constant B belonging to the interval [0,1]. Define the first column of the likelihood matrices to
correspond to a signal of "restrict consumption", the second column to a signal of "no con-
sumption restriction", and the third column to the signal of "no report." The likelihoods asso-
ciated with reduced testing are derived from the original likelihoods through the stochastic

transformation

(3.10)

orPTU O l - ozPTU
O BPm l - {S’Pr5

 

" 'l
= PTU 0 l-Pm a O l-a
0 Prs l-PrS 0 6 1-8
0 0 1

 

The transformation reduces the probability of receiving the perfectly informative signals of
"restrict consumption" and "no consumption restriction". It also increases the chance of an
uninformative "no report" signal for both states. Therefore, increased testing can not reduce
the informativeness (value) of the advisory.

An ordering based on informativeness does not require that a signal is correct in the sense
that a "restrict consumption" or "no report" signal most likely corresponds to a state of unsafe
and safe respectively. A signal can be no worse (no less valuable) than uninformative.
Likelihoods for an uninformative signal are equal across states. As any likelihood probability
associated with a signal increases relative to the others (whether correct or incorrect) the signal
becomes more informative (valuable). Suppose anglers perceive the "restrict consumption"

signal to be most likely when a site is really safe. As the perceived likelihood of receiving the

57

signal when a site is safe increases, so does its informativeness and value even though it is

incorrect. 6

The analysis above yields the following hypotheses.

Hypothesis (Testing More Sites) - An advisory program that tests a greater number
of sites increases the informativeness of both a full and partial disclosure advisory
program. Thus, increased testing does not reduce the value of either advisory

program.

Consider how testing more sites improves anglers' welfare. Suppose an angler's behavior
is guided by a desire to avoid eating too much contaminated fish. One possible response to a
partial disclosure advisory is a choice not to eat fish because the chance of eating contaminated
fish is too high. The analysis above demonstrates that testing more sites increases the chance
that a "no report" site is safe. Therefore, more testing increases the chance of identifying a
safe site and reduces the chance of mistakenly eating contaminated fish. For anglers who begin
eating fish because of a greater chance of identifying a safe site, increased testing has value.

Another possible response to a partial disclosure advisory is a choice to eat fish only from
"no report" sites. Because this choice is in response to an advisory signal, it implies a strictly
positive ex ante value to the current advisory program. Testing more sites may not change an
angler's strategy in response to signals. However, it does increase the probability of receiving

a "restrict consumption" (RC) signal relative to the probability of a "no report" (NR) signal.

Signal probabilities are given by

P(RC)
P(N R)

P(unsafe) P(tested)

P(unsafe)(l - P(tested)) + P(safe)

 

6 To avoid confusion. but with no loss of generality. the examples and analysis of this report assume that signals

are COUCI.

58

As P(tested) increases, P(RC) increases and P(NR) declines.

An increase in the probability of a "restrict consumption" signal increases ex ante
information value only if the expected utility associated with the "restrict consumption" signal
is greater than that associated with the "no report" signal. Consider an angler who would con-
tinue to fish at "no report" sites. The "no report" signal causes no behavioral change and has
zero ex post value. The ex post value of a "restrict consumption" signal is the difference
between the expected utility of eating fish from a "no report" site and the expected utility of
eating fish from a "restrict consumption" site (a mistake). The ex post utility of eating fish

from a "no report" site is

u = P(unsafelNR) ' u(eat, unsafe)

nr

+ P(safelNR) - u (eat, safe).

The "no report" signal does not identify states with certainty. Therefore, a "no report"
site may be "safe" or "unsafe". The expected utility of eating fish from a "no report" site is
the probability that the site is unsafe multiplied by the utility of eating unsafe fish plus the
probability that the site is safe multiplied by the utility of eating safe fish.

The "restrict consumption" signal always identifies an "unsafe" site. Therefore, the utility

associated with eating fish from a "restrict consumption" site is

u = u(eat, unsafe).

['0

The ex post expected utility of the "restrict consumption" signal is

um - um = (P(unsafelNR) - l) - u(eat, unsafe)

+ P(safelNR) - u(eat, safe) 2 0.

Therefore, testing more sites increases the probability of receiving the more valuable "restrict
consumption" signal relative to the less valuable "no report" signal. The ex ante value of the
advisory is the sum of ex post signal values weighted by signal probabilities. Therefore, an

advisory program that tests more sites is more valuable than one that tests fewer sites.

59

Testing more sites makes receipt of a more informative signal more likely, increases the
chance of correctly identifying the state, and reduces the chance of mistakenly eating con-
taminated fish. This analysis applies also to a full disclosure advisory format. Increased test-
ing increases the chance of receiving the perfectly informative "restrict consumption" and "no
consumption restriction" signals and reduces the chance of receiving the completely uninforma-
tive "no report" signal.

Increased testing with full disclosure also increases the number of sites available to an
angler who wishes to avoid chemical residues. Greater testing increases the number of known
safe sites. If an angler anticipates fishing at sites that are currently untested, more testing can
only increase the number of safe alternative sites.

The analysis above assumes anglers attempt to avoid eating too much contaminated fish.
Testing more sites will not change the decision setting for an angler who knowingly eats fish
from contaminated sites —- the alternative is still to fish at a "no report" site which is possibly

contaminated. Testing more sites will not increase advisory value for such anglers.
Accuracy, Perceived Severity of Health Effects, and Advisory Value

The analysis of the previous section assumes that the testing procedure perfectly identifies
the presence or absence of chemical residues in fish. This section examines the case where
anglers perceive tests to be imperfect. Anglers may perceive signals to be imperfect predictors
Of states because of (1) questions about the technological ability of tests to identify contaminant
levels, (2) doubts about the adequacy of scientific knowledge linking exposure to health effects,
or (3) a belief that the MDNR does not accurately report test results (Krieger and Hoehn
1991). Define these factors as affecting the perceived accuracy of the advisory. The first part
of this section uses stochastic transformations to demonstrate that increased accuracy is
synonymous with increased informativeness. This section also examines how advisory value
might differ among individuals with different perceptions of health outcomes resulting from

exposure to chemical residues in fish.

60

The following analysis maintains the assumption that signals are essentially correct - sig-
nals of "restrict consumption" and "no consumption restriction" most likely correspond to
states of unsafe and safe respectively. The term accuracy as used here refers to the strength of
the correct association. The conclusions of the analysis apply as well to a situation where sig-

nals are incorrect.

Perceived Accuracy of Information

The probability tree diagram of Figure 3.2 illustrates the formation of likelihoods when
the link between states and signals is not perfect. The figure differs from Figure 3.1 by the
addition of a column labeled "Reported Test Outcome" that accounts for possible inaccuracies
in reported signals. Define PUU as the probability that a test result of unsafe is issued when a
site is unsafe. Similarly, PSS represents the probability that a test result of safe identifies a safe
site. These probabilities may be less than one for any of the three reasons listed above. The
signals of "no report" and "restrict consumption" become more accurate with increases in PSS
and PCC respectively.

A perception that tests are inaccurate weakens the probabilistic link between states and
signals and reduces the informativeness of either a partial or full disclosure advisory. The
likelihoods associated with less accurate information can be derived from the more accurate

likelihoods of Table 3.1 using the stochastic transformation

(3.11)

P1, (1 - P55) 1 - P130 - P55) 0 1

pw+§m (Pm-1)(1-P,,) (l-PW)--:-TS (Pm-1)(1-P,,)
T3 T3

PTs (1 ' p35) 1 ‘ Prs(1 " Pss)

61

 

 

 

 

 

 

 

 

 

 

State Test Reported Partial Disclosure Full Disclosure
State Test Outcome Signals Signals
Restrict Restrict
Unsafe —— Consumption Consumption
/ Puu Pm PUU Pm Puu
Tested
Pro
No
Safe No Report Consumption
Unsafe l - PUU P111 (1 - Pun) Restriction
\\ Pm (1 ’ pm)
Not Tested No Report No Report
1 - 1 - Pm —— 1 - Pm
Restrict Restrict
Unsafe Consumption Consumption
/1-Ps P13 (1 - P55) P13 (1 - P55)
Tested
“\855 Safe No Report No
P13 P88 Consumption
Safe Restriction
\ PTS PSS
Not Tested No Report No Report
1 ' PTS 1 " PTS 1 ‘ PTS

 

Figure 3.2 - Likelihood Probabilities with Imperfect Tests

62

The second matrix on the right hand side of equation 3.11 is a stochastic transformation.
The first matrix on the right hand side is the matrix of likelihoods from Table 3.1. The left
hand side matrix contains the less accurate likelihoods derived from Figure 3.2. Therefore, a
less accurate partial disclosure advisory is no more informative (valuable) than a more accurate
advisory.

Similarly, the stochastic transformation

(3.12)
PTUPUU PTUU'PL’U) (I'Pm)
Prs(1' pss) Prs Pss (1‘ Pm)

" i
PTU 0 1- P1,, PW (l-PUU) 0
= 0 PTU 1-PTU (l-PSS) P35 0
0 o 1

 

 

relates the likelihoods associated with a less accurate full disclosure advisory and the more
accurate likelihoods associated with the same advisory. Therefore, increases in perceived
accuracy can not reduce the informativeness (value) of a full disclosure advisory.

The following hypothesis summarizes the results of this section.

Hypothesis (Perceived Accuracy of Tesg) - The value of either a partial or full dis-

closure advisory is nondecreasing in the perceived accuracy of the advisory.

Perceived Severity of Mistake

Information value is nondecreasing in the perceived utility loss associated with mistaken
prior actiOn. In the context of sport fish contamination, a mistake is an action that leads to
unacceptable exposure to chemical residues. What is unacceptable depends on the utility asso-

ciated with the perceived health consequences of exposure. Thus, the more severe the per-

63

ceived health consequences of exposure to chemical residues, the greater the ex post value of
the current advisory and the greater the ex ante value of alternative advisory formats.

Consider the definition of information value given by equation 3.5. The equation defines
the value of information as the change in income necessary to maintain initial (pre-information)
utility when behavioral adaptation to informed perceptions of risk is not possible. The greater
the perceived utility penalty associated with continuing initial behavior in light of increased
risk, the greater the change in income necessary to compensate an individual for being unable
to change behavior. Decreases in the perceived utility of q° given P(s I y) increase the expendi-

ture necessary to maintain u° and thus increase information value.

Hypothesis (Perceived Disutility of Exposure) - The greater the perceived utility loss

as a consequence of exposure, the greater the value of information.
Summary

A full disclosure advisory is at least as valuable as a partial disclosure advisory. Further-
more, the value of either advisory is nondecreasing in the number of sites tested. Whether the
value of the additional information is zero or strictly positive depends on whether anglers
change their behavior. Behavioral responses to increased testing might include (1) ceasing
fishing, (2) eating fewer fish, (3) eating smaller fish, (4) eating only certain species, (5) not
fishing at some sites, or (6) changing the way fish are prepared for consumption. Full dis-
closure may lead to site choices that favor safe sites or to changes in catch and consumption
patterns from safe sites.

The value of advisory information also depends on anglers' perceptions of advisory
accuracy. Perceived accuracy may depend on beliefs about the testing procedure or trust in the
integrity of the information source. This finding suggests two approaches to increasing

advisory value. First. educating anglers about the advisory program may increase the per-

64

ceived accuracy of the advisory and thus its value. Second, many anglers seem to believe the
MDNR to be too politicized to provide accurate information. Identifying the advisory with a

more trusted source may increase its perceived accuracy and thus its value.

Chapter Four

Data Collection:
A Contingent Valuation Mail Survey of Licensed Anglers

Most empirical estimates of information value depend on observed changes in market
behavior. For example, Foster and Just (1989) base their estimate of the ex post value of
information about milk contamination on changes in surplus measures for milk consumers.
Studies of the ex ante value of weather information to farmers estimate values with respect to
changes in production of a marketed output (T ice and Clouser 1982, Baquet et al. 1976, Byer-
lee and Anderson 1969).

Direct application of these empirical techniques to the problem of sport fish contamination
is difficult because no formal market exists for the contaminated good (sport fishing). In this
situation, non-market valuation methods can often produce estimates of value. This research
uses a contingent valuation (CV) mail survey of licensed anglers in Michigan to estimate WTP
for additional information about chemical residues in sport fish.

The CV approach asks individuals their WTP for a good in a hypothetical market setting.
Whether the resulting contingent choices correspond to actual market choices depends in part
on the description of the hypothetical market. Consumer behavior in actual markets depends
on a variety of factors. These include characteristics of the good, characteristics of the pay-
ment, and the market itself (Fischhoff and Furby 1988). These factors exist explicitly in a
hypothetical market only if they are included in the description of the market provided by the
researcher. If a CV survey does not adequately describe the factors relevant to a decision,

stated behavior will depend on what the respondent assumes. If respondents' assumptions do

66

not correspond to the market context envisioned by the researcher, the values obtained may not
be those desired.

The challenge of successful CV research is to clearly describe all relevant aspects of the
good, the payment, and a credible market, and to word questions clearly so as to avoid
ambiguity (Mitchell and Carson 1989). Thus, the design of the CV survey is crucial to obtain-
ing good data. The implementation of a mail survey may also affect the quality and quantity of
usable responses. Important considerations include the physical appearance of the survey
instrument, the wording, order, and format of questions, the timing of mailings. and the
process of following up on those who are slow to respond (Dillman 1978).

This chapter reports on a CV mail survey of 1,578 licensed Michigan anglers. The sur-
vey focused on anglers' WTP for changes in the current public health advisory. The chapter
first reviews the design and implementation of the survey. It illustrates the importance of
focus group and one-on~one interview pretesting. Pretesting ensured that questions were both
meaningful to respondents and consistent with research needs. The second section of the chap-
ter provides a statistical summary of survey responses. It examines respondents' socio—
economic characteristics, fishing behavior, and beliefs about the advisory. It also compares
respondents' socio-economic characteristics to those of other surveys of Michigan anglers and

reports the findings of a phone follow-up designed to assess possible nonresponse bias.

Survey Design and Implementation

A contingent valuation survey should clearly communicate ideas to potential respondents
(Mitchell and Carson 1989). If respondents interpret questions differently than researchers,
questions may not measure intended concepts. Questions that use language and concepts that
are familiar and meaningful to anglers facilitate clarity and reduce misinterpretation (Fischhoff
and Furby 1988, Desvousves and Smith 1988). This section reports on the use of a mail

pretest. focus groups, and one-on-one interviews in the design of the CV mail survey used in

67

this research. It focuses on the role of each in formulating clear and meaningful questions and
response categories. The section also describes the physical implementation of the mail sur-

vey.

Survey Design

The process of survey design began with three focus groups. Each group contained four
to six licensed anglers from the Lansing, Michigan area. The focus group discussions con-
centrated on exploring how anglers think about chemical residues in fish, the language they use
to talk about it, and how they think about and respond to the current advisory. Insights from
focus groups aided in designing a draft questionnaire.

Iterative one-on-one interviews with twelve licensed anglers from the Lansing, Michigan
area further refined the draft questionnaire. Interview participants were asked to think aloud as
they responded to the questionnaire. Each round of interviews stimulated changes in the
questionnaire draft that were tested in the next round. The iterative interviews identified ques-
tions that were unclear or ambiguous and helped assess the adequacy of response choices.
They were particularly useful in identifying assumptions respondents made about factors not
explicitly mentioned in the questionnaire. They also helped identify irrelevant and redundant
questions that could be removed or combined with other questions, thus reducing the length of
the questionnaire from 46 to 15 questions. Also, to improve readability, the printed format of
the questionnaire was increased from a booklet measuring 5% by 8% inches to one measuring
8% by 11 inches and the type size increased. Appendix A contains a complete copy of the
final questionnaire.

This section reports on the design of survey questions to measure anglers' WTP for
advisory alternatives and the factors hypothesized to affect WTP. The section first describes
the design of the central WTP questions. It then describes the measurement of explanatory
variables. The discussion emphasizes how mail, focus group. and interview pretests influenced

the final form of specific questions.

68

Eliciting WTP

A clearly described and credible hypothetical market is a crucial component of a CV sur-
vey (Mitchell and Carson 1989, Hoehn and Randall 1987). Important aspects of the market
include a description of the good to be valued, how it differs from what is currently available
to respondents, and a proposed method of payment that respondents accept and believe to be
credible. The survey also describes a decision setting that elicits WTP. This section describes
the evolution of the valuation questions used to elicit anglers' WTP for additional information
about chemical residues in Michigan's sport fish. It focuses first on the development of des-
criptions of the current and alternative advisory programs. It then explains the choice of a pay-

ment vehicle and bid elicitation format.

Description of Advisory Alternatives

A challenge in this particular survey was to clearly explain subtle changes in a complex
information program. Valid measures of WTP for additional information depend on respond-
ent's clear understanding of the content of currently available information. Pretests suggested
that anglers had very different perceptions of the information content of the current advisory.
Focus groups revealed that many anglers have greater confidence in the scope of advisory
information than is warranted by actual testing. Many participants stated a belief that most
sites had been tested. The state has actually tested less than three percent of all sites. A belief
that most sites have been tested implies a belief that a site not listed in the advisory as unsafe
has probably been tested and found to be safe.

To provide anglers with an accurate and common point of reference, the questionnaire
described the information content of the current advisory program. It told respondents how
many sites had been tested and that the advisory listed only those sites that were tested and
found to be unsafe. The initial survey draft (reproduced in Figure 4.1) contained three WTP

questions. The first question described the testing environment and information content

69

 

Fish from all of Michigan's Great Lakes have been tested for chemical residues.
Michigan also contains 11,000 or so inland lakes and over 36,000 miles of rivers and
streams. Only 331 tests have been conducted on these inland waters.

The current advisory contains a list of water bodies where fish have been found to be
unsafe. Water bodies that have been tested and found to be safe are not listed in the
advisory.

Suppose the state could no longer afford to test fish and print advisories unless fishing
license fees were increased. Would you rather, (I) eliminate the current advisory
program and keep fishing license costs the same, or (2) keep the advisory if it meant
increasing license costs by $3.00? (Circle one number)

1. Eliminate the advisory
2. Keep the advisory

Suppose that an up-to-date list of 600 sites that had been tested and found to be safe
could be printed in the advisory if money for printing was available.

Would you rather, (1) keep advisories as they are and keep fishing license costs the
same, or (2) include a list of 600 safe sites if it meant increasing license costs by
$2.00? (Circle one number)

1. Keep advisories the same
2. Include list of safe sites

Only‘331 tests for chemical residues have been conducted in Michigan's inland waters.
About 30 new sites are tested each year.

Suppose more sites could be tested if license fees were increased.
Would you rather, (1) continue to test about 30 sites per year and keep fishing license

costs the same, or (2) increase testing to 100 new sites per year if it meant increasing
license costs by $4.00? (Circle one number)

1. Continue current testing
2. Increase testing

 

Figure 4.1 - Initial Wording of WTP Questions

70

of the current advisory program. It then asked anglers' WTP to continue the current program.
The primary purpose of this question was to provide a clear description of the current
program. The second question proposed a full disclosure advisory and again asked for
respondents' WTP. The third question asked for anglers' WTP for testing a greater number of
sites each year.

One-on—one interviews revealed two problems with the initial format of the WTP ques-
tions. First, the questions did not adequately emphasize the important aspects of the current
advisory program or clearly identify the changes offered by the proposed alternatives. When
questioned during interviews, participants were often unable to clearly recall the information
contained in the current advisory or clearly state the differences between the advisory alterna-
tives. Second, the second and third questions described changes in the advisory in terms of
numbers of sites tested or listed as safe. However, the program description contained in the
first question did not identify the total number of sites. In the absence of an explicit
denominator, respondents made different assumptions about the proportion of sites tested under
the proposed programs.

To remedy these problems, the next draft of the questionnaire separated the description of
the current advisory program from questions of WTP for program alternatives. The descrip-
tion of the current advisory (Figure 4.2) included a more explicit definition of the total number
of fishing sites in the state. Discussions with MDNR officials refined a definition of sites that
reflects those used for testing and advisory purposes. The description also presented the
important aspects of the current program and proposed changes more clearly. Interviews were
particularly useful in improving the brevity and clarity of the description. Participants in sub-
sequent interviews seemed to quickly and accurately identify important features of the current
advisory program when presented in the form of Figure 4.2.

Figure 4.3 reproduces an intermediate form of the WTP questions presented with the
program description of Figure 42. The revised format improved on the initial questions of

Figure 4.1 by presenting important aspects of program alternatives in tabular form and by

71

 

Michigan's Public Health Advisory

There are more than 5,800 public fishing sites in Michigan - 2,200 sites on
rivers and streams, 3,600 inland lakes, and the Great Lakes. The state has
tested 350 of these sites for chemical residues in fish. About 30 new sites

are tested each year.
The current public health advisory tells you:

O that you should not eat too much fish from any inland
lake because of widespread mercury contamination, and

s it lists 50 sites where fish contain chemical residues
above state limits.

The advisory does not tell you about:

0 the 300 tested sites where chemical residues do not exist
or are below state limits.

The advisory program could be changed.

0 The advisory could list tested sites where chemical
residues do not exist or are below state limits.

0 More than 30 new sites could be tested each year.

These changes would increase the amount of information in the advisory

but they would also cost more money.

 

Figure 4.2 - Final Description of Current Advisory

72

 

Table 1 - Advisory Programs and their Cost to You

 

Current Program Program
Program Options Advisory A B

 

 

Lists tested sites where non-mercury

 

 

 

residues pose little or no health risk. No No Yes
Number of new sites tested each year. 0 30 0
Cost to you in highler license fees. $0 $5 $3

 

 

 

 

 

In the next two questions suppose anglers could vote on Programs A and
B. Vote for the program if it is worth the additional cost to you. Vote
against the program if it is not worth the additional cost.

1.

Would you vote for Program A if it permanently increased your

yearly license cost by $5.00, or vote against it and keep the Current
Advisory at no additional cost?

1. Vote for Program A

2. Vote against Program A and keep Current
Advisory

Would you vote for Program B if it permanently increased your
license cost by $3.00, or vote against it and keep the Current
Advisory at no additional cost?

1. Vote for Program B
2. Vote against Program B and keep Current
Advisory

 

Figure 4.3 - Intermediate Form of WTP Questions

73

placing greater emphasis on voting as a choice mechanism. Focus group results suggested that
information presented in tables was often more likely to be read and more easily understood
than textual information. The resulting tabular comparison of advisory programs in the revised
WTP format increased the ease with which interviewees understood and were able to
accurately recall program differences.

Interviews with the revised questionnaire draft revealed that respondents did not always
view the two alternative programs as independent. When asked WTP for testing more sites a
respondent might say, "well, if I'm already paying X dollars to list safe sites...". If respond-
ents do not perceive the programs to be independent, reported WTP will depend on the order
in which the questionnaire presents program alternatives and it will be difficult to determine
the separate value for full and partial disclosure advisories. The final form of the WTP ques-
tions (Figure 4.4) addressed the problem of independent programs by asking respondents to
value only one program alternative. The proposed alternatives consisted of a partial disclosure

or a full disclosure advisory, each with increased testing.

Pigment Vehicle
The questionnaire proposed a permanent license fee increase as the payment vehicle for

proposed program alternatives. The license fee payment vehicle is appealing because fees are
collected by the state and people link them directly to spending on state programs. An alterna-
tive payment vehicle of selling a separate advisory booklet was tested in focus groups and
interviews. The booklet did not adequately capture WTP for information because it is difficult
to exclude people from access to advisory information. Many people stated they would read
the booklet without buying it or share one with friends. The following comments typify reac-
tions to the booklet.

"...we couldn't [buy the booklet] or we'd have to go in with some friends or

go borrow somebody's book."

"And I think that's what would happen. I think a lot of people would stand at
the counter and look up their section and then go on."

74

 

Your Vote on Advisory Programs

The table below shows two advisory programs. The "Current Advisory"
is Michigan's current advisory program.

program that could be put in place.

"Program A" is a different

 

 

 

 

 

Current Program
Program Options Advisory A

Lists tested sites where chemical residues

are above state limits? Yes Yes
Lists tested sites where chemical residues

do not exist or are below state limits? No Yes
Number of new sites tested each year. 30 110
Cost to you in higher fishing license fees. $0.00 $.40

 

 

 

 

 

Suppose the Michigan Department of Natural Resources (DNR) sent you a
ballot to vote "for" or "against" Program A. If a majority of anglers vote
"for" Program A, it will replace the Current Advisory. If a majority vote
"against" Program A, the Current Advisory will be continued.

1. Would you vote for Program A if it permanently increased your
yearly license cost by $.40, or vote against it and keep the Current
Advisory at no additional license cost?

1. Vote for Program A

2. Vote against Program A and keep Current Advisory
3. Don't know or no opinion

 

Figure 4.4 - Final Form of WTP Question

75

"l'd probably stand there and look through it, look up my site and set it back
down."
By contrast, the license fee payment is more difficult to avoid if respondents wish to continue
fishing.

Some pretest participants objected to the license fee payment vehicle. Several focus group
participants thought it unlikely the money collected would actually be used to improve
advisories. As a basis for these beliefs, participants cited several past examples of state real-
location of collected funds. One participant phrased his concerns as follows:

"I wouldn't mind paying a little bit more in a fishing license if they earmarked
that's what it's going for. ...if they're not taking money away from the other
part of the license to do something else with."

The final questionnaire used the license fee payment vehicle instead of the booklet.
Respondents who reject the referendum because they object to the use of license fees are react-
ing to a relevant dimension of the described market context. Empirical WTP estimates will
ultimately reflect reactions to many contextual factors. This is also true of WTP estimates
derived from observed market behavior - actual market choices also depend on the many con-

textual dimensions that define a market.

Bid Elicitation

The study used a referendum format for bid elicitation. The referendum format is proba-
bly easier for respondents than open-ended questions (McConnell 1990, Hanemann 1985,
Cameron 1988, Freeman 1993). The decision to accept or reject a good at a given price is the
most common type of market transaction people make. Also, people are familiar with the idea
of voting on public programs. Ballots often contain measures offering public goods such as
schools, water and sewer systems, or roads, for a given increase in taxes or fees (Mitchell and
Carson 1989). In addition. the referendum format provides incentives for respondents to

honestly reveal WTP. Hoehn and Randall (1987) examine the problem of strategic overstate-

76

ment or understatement of values in CV surveys. They conclude that an individual's optimal
strategy in response to a referendum format is to state true WTP.

Finally, the survey also described the institutional context under which advisories were
funded and issued. A number of respondents throughout the pretest process voiced strong feel-
ings about the integrity of the MDNR. Typical of these responses were:

"People come from all over the United States to fish here and if the state

MDNR tells people just how contaminated some of these fish are, that would

scare a lot of them away."

"I think the MDNR is more politicized than the Department of Health. They

listen to too many special interests..."
In general, respondents seemed to view the MDNR as the source of the advisory even though
the advisories are actually issued by the MDPH. When questioned, respondents seemed to
believe the MDPH more likely than the MDNR to provide accurate information. To facilitate
comparison of different information sources, the final questionnaire design was a split sample

with both the MDNR and MDPH as the stated source of the advisory.

Explanatory Variables

The conceptual framework suggests that four classes of variables influence anglers' WTP
for advisory alternatives. These are (1) information content, (2) the possibility of behavioral
change to avert risk, (3) perceived accuracy of the advisory, and (4) the perceived severity of

health consequences resulting from consumption of contaminated fish.

Program Variables

The questionnaire described alternative programs in terms of (1) the number of new sites
tested each year, (2) whether the program fully or partially discloses test results, (3) the annual
cost in higher license fees, and (4) the state agencythat issued the advisory. The experimental

design consisted of combinations of three different testing levels, ten levels of program cost,

77

two levels of information disclosure, and two state agencies responsible for the advisory. A
complete factorial design over these factors defined 120 separate versions of the questionnaire.
Table A.1 in appendix A defines the questionnaire versions.

The questionnaire offered testing levels of 110, 620, and 1,240 new sites per year. Two
considerations influenced the choice of testing levels. First, the chosen levels define programs
that differ significantly in both the absolute number of sites tested and in the proportion of total
sites tested. A mail pretest of the survey revealed no significant difference in WTP for
programs that tested 100, 200, 300, or 600 sites annually. This result may reflect a poorly
designed pretest that did not adequately describe program differences. It may also imply that
respondents did not view the program differences to be meaningful. The larger range used in
the final survey makes it more likely that respondents perceive the programs to be significantly
different. A second consideration was to choose testing levels that were identified as physi-
cally feasible during interviews with state officials.

In addition to testing levels, the survey also presented the respondent with a program cost.
The literature addresses two issues concerning the structure of program costs (bids) offered in
dichotomous choice CV surveys - the choice of bid levels and the number of surveys to use
with each level. Two recent studies address these issues in the context of CV survey design
(Kanninen 1993, C00per 1993).

Kanninen and Cooper propose complex methods for the choice of bids that minimize the
variance of the WTP estimator. The general thrust of both approaches is to use prior knowl-
edge about the distribution of WTP to attempt to cluster bids around the true mean WTP. This
minimizes the number of surveys "wasted" on bids far from the mean. Kanninen pr0poses a
sequential interview process with each step using past responses for prior information about
true WTP. Her approach is not readily adaptable to mail surveys because of the slow response
time. Cooper suggests using a pretest to determine an empirical distribution of bids. He
divides the distribution into equal probability increments and sets bids at the boundaries. He
then uses a complex algorithm to jointly determine the variance minimizing number of bids and

the allocation of different bids across surveys.

78

Duffield and Patterson (1991) and Boyle et al. (1988) also address the distribution of bids
across surveys. Although their approaches are not Optimizing methods, they also employ an
empirical distribution of WTP and attempt to cluster the bulk of offered bids around mean
WTP.

This study employs a mail pretest to gain some prior knowledge about the distribution of
WTP. However, it does not use the methods reviewed above to determine optimal bids or
allocations. The pretest asked valuation questions that were somewhat different than those
used in the final survey. The questionable quality of the pretest data relative to the final WTP
questions seemed to dictate a conservative approach that reduced the risk of clustering many
offered bids around a point far from the true mean.

A mail pretest of a sample of 200 licensed anglers in the Lansing, Michigan area asked
anglers' WTP for advisory alternatives in an open ended format. The pretest produced 44.
completed surveys. Responses to the pretest defined a cumulative distribution of WTP. The
cumulative distribution was initially divided into 32 equal probability increments. The WTP
levels associated with the boundaries of these increments were considered as possible bid
levels. The final design set bid levels at every fourth boundary over most of the empirical dis-
tribution. However, to obtain more information about the relatively flat upper tail of the dis-
tribution, the higher bids were spaced more closely. Also, selected increments began at the
third boundary from a probability of zero.

The final design used bid levels of $.40, $.95, $1.45, $1.90, $2.85, $4.10, $5.55, $8.75,
and $14.50. The empirical probability that WTP was greater than or equal to these bids was
.94, .81, .69, .56, .44, .31, .19, .09, and .03 respectively. In addition, the design included a
relatively large tenth bid level chosen in hopes of eliminating any positive response and avoid-
ing the problem of arbitrarily truncating the empirical distribution. The upper bid amount was
$41.00. The final experimental design used equal numbers of surveys containing each bid.

In addition to the issues discussed above. focus group results raised one other important

consideration in the choice of bid amounts. When asked their WTP for a particular program,

79

participants often conditioned responses on perceptions of a reasonable cost of providing the
program. Respondents seemed concerned about not paying more than a fair share of the costs.
The following responses illustrate the nature of this concern.

"Well, I have no idea how much this [publishing a list of safe sites] costs so

it's really kind of hard to sit here and hem and haw over how many dimes I

would...actually give towards it."

"...for a 30 page pamphlet or something, five bucks, well, they're just making
money off it."

"What's it cost to test a site? ...So, how many fishing licenses do they sell in
a year? I mean, if they tack on 50 cents a fishing license, how much
money..."
Fortunately, the bids derived from the empirical distribution corresponded reasonably well
to estimates of actual program costs. State officials estimated that about 100 additional sites

could be tested for each $1.00 increase in license fees. ' This implies actual costs ranging

from about $1.10 to $12.40 for the testing levels used in the questionnaire.

Behavioral Change

The questionnaire asked how respondents had changed behavior as a result of the current
advisory. It also asked for anticipated behavioral change in response to a list of safe sites or a
report of unsafe for a respondent's favorite site and species. The advisory suggests a number
of changes in behavior that can reduce risk. The research used focus groups to explore which
of these actions anglers were aware of and which they were likely to use. These discussions
influenced the form of pretest questions dealing with behavioral change.

Pretest questions asked if respondents would make specific behavioral changes in response
to a given change in the advisory. Interviews revealed that response choices for this format

were not nearly rich enough. Participants often mentioned behaviors that were not included

 

' Personal communication with Christine Waggoner. Surface Water Quality Division, Michigan Department of

Natural Resources.

80

among the response categories offered in early drafts of the questionnaire. As an example of
the issues that arose when designing these questions, consider the question to assess behavioral
response to the listing of safe sites. Figure 4.5 lists both the pretest and final versions of this
question.

The final version of the question improves on the pretest version in two ways. First, the
answer to the question would likely depend on whether a respondent believed that currently
used sites would be listed as safe or not. The pretest version provides no means to determine
what assumptions respondents make regarding currently used sites. Second, the pretest version
seems to lead the respondent by asking for response to a specific behavioral change (fish at a
safe site) that the researcher believed to be important. The final version conditioned responses
on currently used sites and provides a list of alternatives derived from interviews - response
categories respondents believed to be relevant.

Risk

The format of questions for assessing risk perceptions changed very little between the
pretest and final versions of the questionnaire. The questionnaire asked respondents for their
guess about the probability of having health problems someday because of chemical residues in

fish. Response categories covered a roughly logarithmic scale. These were:

1. No chance 6. 1 in 100

2. 1 in a million 7. 1 in 10

3. 1 in 100,000 8. l in 5

4. 1 in 10,000 9. 1 in 2

5. l in 1,000 10. Certain to happen

Focus group and interview participants generally perceived very small risks associated with
chemical residues in fish. The logarithmic scale concentrates responses around small risks and
follows the approach of other studies designed to measure small perceived risks (van Ravens-
waay and Hoehn 1991).

People generally found the question about risk difficult. Subsequent interviews focused

on the source of difficulty. In general, respondents seemed to have little problem interpreting

81

 

Initial Question

1. If sites that were tested and found to be safe were listed in
the "Public Health Advisory" would you try to fish at those
sites rather than sites that were not listed? (Circle one num-
ber)

Final Question

1. In addition to the list of unsafe sites, suppose the advisory
listed all tested sites where chemical residues in fish did not
exist or were below state limits. If your favorite sites had not
been tested would you...? (Circle all that apply)

Continue to fish at your favorite sites

Stop eating fish from your favorite sites

Fish only at sites where chemical residues are
below state limits

When choosing a new site, be more likely to
go to a site where chemical residues were
below state limits

.“ 9‘59.“

 

Figure 4.5 - Behavioral Change Questions

82

risks as stated. One participant interpreted the logarithmic scale as linear but soon realized his
mistake. The hesitation in responding to the question arose primarily from uncertainty about
the accuracy of the guess. A typical reaction when asked to guess about the probability of a
health problem was:

"I don't know. really, really slim I think. I don't think it's no chance. I'll

guess one in a million. I'm really not sure."
The final version of the questionnaire also measured the degree of certainty associated with
perceived risk by asking how sure people were of their guess. Other researchers have also

successfully used this approach to assess respondents certainty about probability estimates (van

Ravenswaay et al. 1992).

Informgtion Accurggy

The conceptual model identifies three factors that may affect the perceived accuracy of the
advisories. These are (l) the perceived accuracy of tests to identify chemical residues in fish,
(2) the perceived adequacy of scientific knowledge linking exposure to health effects, and (3)
beliefs about whether the state will accurately report test results. Interviews and the mail
pretest revealed that few pe0ple questioned the accuracy of the tests themselves. Con-
sequently, the final form of the questionnaire eliminated the question to assess perceived test
accuracy. Pretesting resulted in relatively minor changes in the wording and format of ques-

tions addressing the other two sources of perceived accuracy.

Sampling Frame and Survey Implementation

Chemical residues in Michigan's fish potentially affect three groups of individuals. These
include currently licensed anglers, those who do not fish but would if residues were not pre-
sent, and those who do not fish but eat fish caught by others. For practical and conceptual

reasons. this research focused on currently licensed anglers. From a practical perspective,

83

licensed anglers are an easy group to identify -- the MDNR obtains names and addresses when
a license is purchased. Those who do not fish because of chemical residues and those who eat
fish caught by others are relatively difficult and expensive to identify.

The group who would fish if contaminants were not present is likely small. A casual
examination of fishing license sales records revealed no obvious decrease associated with the
appearance of advisories in 1970. The sampling frame does not include this group because of
the difficulty and expense of identifying and reaching them.

It is difficult to determine the number of people who eat fish caught by others. This
group includes people who purchased fish in restaurants and stores and those who receive fish
from acquaintances. Twenty percent of New York anglers reported giving away some of their
catch (Knuth and Velicer 1990). Furthermore, commercial fishing operations in Michigan
landed 15.7 million pounds of fish in the Great Lakes in 1988 (Michigan Department of Natu-
ral Resources 1990). This research examines the value of advisories designed to influence
sport angling behavior. Thus, while the group of people who consume fish they do not catch
is potentially large, they are not as likely to be directly influenced by the advisory.

The sampling frame for this research consisted of individuals who purchased a Michigan
fishing license for the 1991 fishing season. 2 The Fisheries Division of the MDNR provided a
random sample of 1,578 anglers licensed to fish in 1991. Information provided for each angler
included, name, address, birth date, and type of license purchased. Rodabaugh (1987) reports
that between 12.5 percent and 14.5 percent of surveyed anglers fished without purchasing a
license. This figure reflects the actions of anglers who reside within one mile of the Shiawas-
see River and may represent the more casual anglers. Whatever the composition of this group,
this survey did not reach them.

Design and implementation of the survey followed Dillman's (1978) total design method

(T DM). The TDM stresses the many small details which, taken together, have a potentially

 

3 Michigan fishing licenses are valid through the end of March of the year following their issue.

84

large impact on response rates and the quality of data from mail and phone surveys. Following
the TDM, the first mailing of 1,578 surveys was sent on Tuesday, February 9, 1993. The sur-
veys were followed in one week by a reminder postcard to prompt response and to thank those
who may already have responded for their participation.

Three weeks after the initial mailing, a second copy of the survey was sent to the 1,012
members of the sample population who had not yet returned a completed questionnaire.
Finally, seven weeks after the first mailing, a third copy of the questionnaire was sent by
certified mail to the 576 people who had not responded to the previous mailings. Of the 1,578
surveys originally sent, 230 were returned as undeliverable yielding a final sample of 1,348
anglers. The survey achieved an overall response rate of 73.4 percent of deliverable surveys
(990 returned surveys.) Figure 4.6 graphs the pattern of returns and illustrates the effect of
each contact with respondents. Appendix A reproduces the cover letters sent with each mailing
and the text of the reminder postcard.

Each contact prompted an increase in overall response. However, the magnitude of the
response decreased with each contact as the remaining individuals were less inclined to com-
plete the questionnaire. The first mailing obtained a response rate of 30 percent; the second,
25 percent; and the third, 24 percent. It is surprising that the final mailing elicited almost the
same rate of response as the second. The following comment from a returned questionnaire
suggests that the use of certified mail may have increased response rates relative to regular
mail.

"In-as-much that you have gone to the expense of sending this via certified
mail, I felt that it would behoove me to fill out and return. I do admire your
efforts in safeguarding the health and safety of Michigan sportsmen." (respond-
ent #2385)

However, the third contact also ran the risk of angering people. The following comment
expresses a common theme a bit more creatively than usual. This was written on the cover let-

ter and returned without a completed questionnaire.

85

Surveys Returned

 

 

 

 

 

 

 

 

 

Figure 4.6 - Pattern of Survey Response

86

"You can take this as my answer, if I were interested in answering this
questionnaire I would have sent the first one back. Now that you've wasted
enough of the taxpayer's money to mail me three envelopes @ $1.52 each, you
can save us all some money and use this paper constructively the next time you
visit your favorite john. Thank you very much for your time." (respondent
#1105)
The use of certified mail also angered some people who made a special, time consuming
trip to the post office to pick up what they assumed was important mail. The negative reaction
has prompted some researchers to abandon the certified mailing. Casual observation seems to

indicate little negative effect on response rates. 3 The certified mailing resulted in 150 addi-

tional completed questionnaires -- about 15 percent of the overall response.

Survey Results

The survey produced 789 complete questionnaires -- 79.7 percent of the 990 returned.
Neither item nor survey non-response seems to be significant in this study. Table 4.1 reports
non-response rates for each survey question. Not surprisingly, respondents were most reluc-
tant to reveal income. Item non-response presents a problem in this study only if the
referendum choice of those who chose not to respond to a question -- and are thus omitted
from the analysis - differs significantly from choices of respondents.

Respondents and non-respondents to specific questions differed significantly in this respect
only for the question that asked about objections to using a license fee increase to pay for addi-
tional information. The direction of causality in responses is difficult to determine. Respond-
ents may have refused to answer the valuation question because they objected to the payment
vehicle. Conversely, respondents may have registered a strong objection to the license fee
increase to justify not answering the valuation question.

The license information from MDNR also contained age and gender for the entire sample.

The proportions of male anglers in the overall sample (81.1%) and the respondent sample

 

3 Personal communication with John Stoll, University of Minnesota.

87

Table 4.1 - Item Nonresponse

 

 

Frequency
of Percent

Question Nonresponse Nonresponse
Question 15: Income 128 12.9
Question 2: Fishing frequency 59 6.0
Question 4: Risk 45 4.6
Question 9: Referendum 39 3.9
Question 13: Best information source 32 3.2
Question 7: Behavioral response

to current advisory 36 3.6
Question 3: Consumption frequency 36 3.6
Question 1: Other activities 36 3.6
Question 6: Perceived risk reduction 36 3.6
Question 8: Concern about residues 31 3.1
Question 10D: Advisory accuracy 30 3.0
Question 14: Education 28 2.8
Question 5: Certainty about risk 28 2.8
Question 108: Scientific knowledge 25 2.5
Question 11: Behavioral response

to safe list 24 2.4
Question 10A: Protest . 23 2.3
Question 10C: Likelihood of fatal outcome 20 2.0
Question 12: Behavioral response to

more testing 17 1.7

Complete Surveys 789 76.1

 

88

(82.0%) are statistically identical for oz=.5. Also, the average age of respondents, 44.9 years,
does not differ significantly from the average age of the overall sample (44.1 years) for
oz=.001.

Phone interviews with 23 survey non-respondents provided information about the extent of
survey non-response bias. The interview asked for (l) beliefs about whether the advisory
understates health risk from chemical residues, (2) the perceived likelihood that health effects
are fatal, (3) objections to increasing license fees to pay for better advisory information, and
(4) education level. The questions corresponded exactly to the wording of survey questions 10
(parts 4, 3, and 1) and question 14 respectively. These questions were chosen because they
were conceptually important and closely correlated empirically with WTP. Table 4.2 reports
results of statistical tests of the null hypotheses that mean responses to these questions were the
same for the populations of respondents and non-respondents. The tests reveal no statistically
significant difference between the two groups for a=.01. The remainder of this section

provides a statistical summary of collected data.
Soda-Economic Characteristics of Respondents

Information about the sample provided by the MDNR included data on gender and age.
In addition, the survey asked about the respondents' education and income. These variables
provide a means of comparing some socioeconomic characteristics of respondents with
samples obtained in other studies. Chief among comparable studies of Michigan anglers is the
1984 MDNR mail survey of 10,948 licensed anglers. Mahoney et al. (1986), Kikuchi (1986),
and Jones and Sung (1991) report results of the MDNR study.

Table 4.3 compares gender, age, education, and average income of respondents to this
survey with the MDNR survey. The proportion of male anglers responding to the MDNR sur-
vey is significantly greater than the proportion responding to this survey. The difference is

probably due to a regulation enacted for the 1985 fishing season that required spouses of

89

Table 4.2 - Tests for Nonresponse Bias

 

 

Respondent Nonrespondent Test Accept/
Variable Mean Mean Statistic Reject?
Understate risk? 2.998 2.565 1.71 Accept
Chance of fatality 3.262 3.143 .49 Accept
Objection to fee 2.685 2.391 1.24 Accept
Education 4.292 4.391 .36 Accept

 

Respondent means reported for "understate risk , chance of fatality", "objection to
fee", and "education" reflect 960, 969, 967, and 962 observations respectively. Non-
response means are based on 23 observations.

licensed anglers to purchase a license if they fish. Prior to the 1985 season, spouses of
licensed anglers were permitted to fish without a license.

The average age of respondents to this survey (44.9 years) differs very little from the
average age of respondents to the MDNR survey (44.2 years). However, information about
the MDNR sample is insufficient for statistical comparisons of average respondent age for the
two studies.

Respondents to this survey are slightly more educated than those from the MDNR survey.
Over ninety percent (90.7%) of respondents reported having at least a high school diploma
compared to 86.9 percent from the MDNR survey. This difference is significant for a = .01.
Similarly, 50.6 percent of respondents to this survey reported at least some college compared
to 46.2 percent from the MDNR survey. This difference is significant for or = .05. The 19.2
percent of respondents to this survey reporting a college degree is also significantly different
from the 17.3 percent reported by the MDNR survey.

The higher levels of education found in this study mirror education trends in the state. In

the period from 1984 to 1990 the proportion of the Michigan population with a high school

90

Table 4.3 - Socio-Economic Characteristics of Respondents

 

 

1984 MDNR
Sample Respondents Study ‘
Gender
Male 81.1 82.0 94.3
Female 18.9 18.0 5.7
Age 43.3 45.5 44.2
Education
Grade school only N.A.2 3.5 5.8
Did not finish high school NA 7.1 7.3
High school or GED NA 30.2 38.1
Vocational or technical school N .A 8.6 N.R.3
Some college NA 31.4 28.9
College graduate (BS or BA) NA 9.7 11.5
Some graduate or professional school NA 4.4 .8
Graduate degree (PhD, MD, MA, MBA) N .A 5.1 5.0
Average Income N .A. $39,230 $36,2024

 

‘ Reported in Mahoney, Jester and Stynes (1986) and Kikuchi (1986).
2 N.A. means "Not Available".

3 N.R. means "Not Reported".

‘ Converted to 1993 dollars from $26,285 in 1984 dollars.

Data reported in the "sample" column reflect a sample size of 1,578. Data on gender and
age of respondents is based on the 990 respondents who returned surveys. Data on
respondents' education and income is based on 962 and 862 observations respectively. Data
from the 1984 MDNR survey reflects 10,948 completed surveys.

91

degree increased from 68.0 percent to 77.0 percent. Over the same period the proportion with
college degrees rose from 14.3 percent to 17.3 percent (Statistical Abstract of the United
States, 1984, 1990). Therefore, the significant difference in education between respondents to
this survey and the MDNR survey does not necessarily imply that they represent different pop-
ulations.

Finally, the average annual income of respondents to this survey is $39,230. This is not
significantly different from the adjusted annual income of respondents to the MDNR survey
($36,202). The above results suggest that this survey and the 1984 MDNR survey reached the

same population.

Angler Behavior

Most respondents (78.0%) reported fishing at inland lakes or ponds. Fewer reported fish-
ing in rivers and streams (56.1%), the Great Lakes (54.6%), or other locations (5.2%). Other
types of fishing sites mentioned included private ponds and out-of-state locations. Anglers also
took more trips per year to inland lakes or ponds than to other types of fishing sites. Over all
types of sites, anglers reported fishing an average of 29.1 days per year. Table 4.4 lists the
percentage of respondents who reported fishing at each type of site and the average number of
days per year they reported fishing at that site type.

A significant number of anglers (30.3%) reported fishing at only one type of site.
Anglers may fish at only one type of site for several reasons. First, they may have a strong
preference for certain species of fish or for a particular type of fishing. Second, they may pre-
fer a specific site. Site preference may result from convenience - close to home or vacation
site - or merely habit. The data reported in Table 4.5 seem to support the notion that anglers
who fish at only one type of site are more casual than those who fish several types of sites.
Anglers who fish only one type of site fish an average of 17.9 times per year. Those who fish

two or three different types of sites fish an average of 26.7 and 43.6 days per year

92

Table 4.4 - Site Types Used

 

 

Percentage Average

of Anglers Number of

Who Use Days Fished
Site type Site Type Per Year
Inland Lakes and Ponds 78.0 17.6
Rivers of Streams 56.1 13.3
Great Lakes 54.6 1 1.0
Other 5 .2 1.4

 

Data in this table was taken from 931 surveys with complete
responses to questions about fishing behavior.

respectively. In addition, anglers who fish fewer types of sites are less likely to eat what they
catch.

Kihlstrom (1974) finds that people with strong preferences are less likely to change pur-
chasing behavior in response to information about the quality of a product. Similarly, anglers
with strong preferences for a particular site or type of fishing, for whatever reason, should be
less willing to change sites or species than those who enjoy a wider variety of fishing experi-
ences. The data in Table 4.5 seem to support this view. Anglers who fish only one type of
site are less likely to change behavior than those who fish two or three types. If the advisory
reported their favorite site to be unsafe they would be less likely to change sites and more

likely to stop fishing than anglers who fish several types of sites.

Ex-Post Value of the Current Advisory

Ex post, the current advisory is valuable only if anglers change their behavior because of

it. Table 4.6 lists the behavioral changes reported by survey respondents. Most anglers

(63.9%) reported making some change in their behavior in response to the current advisory.

93

Table 4.5 - Diversity in Use of Site Types

 

Number of Different Site Types Used

 

 

 

1 2 3
Percent of Anglers 30.3 34.8 33.0
Average Number of Days/Year 17.9 26.7 43.6
Consumption Frequency
Do not eat fish 20.5 16.7 9.9
Response to current advisory
Read advisory? 86.5 87.4 92.7
Eat less fish 24.0 22.7 27.9
Use different sites 10.2 16.9 19.5
Eat smaller fish 14.2 16.2 21.6
Eat different species 7.8 15.1 15.9
Prepare differently ‘ 21.2 30.8 41.9
Do nothing differently 39.9 35.0 32.0
Stop eating fish 6.4 4.6 2.3
Response to report that favorite site is pnsafe
Still fish at site 33.7 44.1 49.5
Eat fewer fish from site 16.3 20.2 23.6
Eat no fish from site 52.5 56.8 49.8
Fish at different site 34.8 41.0 39.7
St0p fishing 13.8 7.5 4.9

 

This table is based on 907 surveys for which questions 2, 3, 7, and 12 were answered.

94

Table 4.6 - Behavioral Response to Current Advisory

 

 

Single
Behavior
Behavioral Response Percent Percent
Did not change behavior 36.1
Have not read the advisory 13.6
Read advisory but do nothing different
Changed behavior 63.9
Prepare fish to eat differently 31.7 37.4
Eat fish less often 24.3 34.2
Eat smaller fish 18.3 10.9
Fish at different places 16.4 21.8
Eat different species of fish 14.7 15.0
Stopped eating fish 4.7 66.7

 

This table reports responses from 951 completed surveys.

The most common change was to prepare fish to eat differently (reported by 31.7 percent of
respondents). This finding is consistent with focus group results. Other changes, in decreas-
ing order of use were (1) eating fish less often (24.3%), (2) eating smaller fish (18.3%), (3)
fishing at different places (16.4%), (4) eating different species of fish (14.7%), and (5) st0p-
ping eating fish (4.7%).

More than one third (36.1%) of respondents reported doing nothing differently because of
the advisory. For this group, the current advisory has no value. However, 37.7 percent of
those who reported no behavioral change also said they had not read the advisory. These
anglers may value the current advisory if they had read it.

Table 4.6 also shows that anglers generally changed more than one aspect of their fishing
behavior in response to the current advisory. The column labeled "single behavior percent"

reports the proportion of respondents who made a particular behavioral change for whom that

95

response was the only behavioral change made. Overall, 29.4 percent of respondents reported
making only one type of change in behavior in response to the advisory.

The proportion of respondents who made a particular behavioral change their only
response may reflect perceptions of the relative effectiveness of a particular response in reduc-
ing risk. Following this reasoning, respondents believed stopping fishing, preparing fish dif-
ferently, eating fewer fish, fishing at different sites, eating different species, and eating smaller
fish to be progressively less effective methods for reducing risk. However, the ordering of
behavioral changes may also reflect the relative ease, both in physical and utility terms, of

implementing the change.
Perceived Risk

Most respondents think their chances of suffering an adverse health effect from chemical
residues in sport fish are small. Table 4.7 reports the distribution of perceived risk. The
logarithmic scale of possible responses renders the mean perceived risk of 5.2 in 100 a mis-
leading indicator of central tendency. A more revealing measure is the median of 1 in 100,000
chance of an adverse health effect.

The survey also asked respondents how certain they were about their reported chance of a
health problem associated with chemical residues in fish. Table 4.7 also reports the percent of
respondents for each level of perceived risk who were "very certain” of their guess. Not sur-
prisingly, those who reported extreme risks ("Certain to happen" or "No chance") were more
likely to be "very certain" their answer was correct.

This research also confirms that risk is a multidimensional concept. Rescher (1983)
defines risk as determined by probability and outcome. Fischhoff (1978) lists perceived
severity of the outcome as one of twelve factors that affect individual risk perceptions. This
research finds support for the interdependence of probabilistic concepts of risk and perceived

severity of outcome. The statistics of Table 4.8 demonstrate a strong correlation between

96

Table 4.7 - Perceived Risk of Adverse Health Effects

 

 

Proportion

Perceived Very Sure of
Risk Frequency Percent Chance
No Chance 244 25.8 47.7

1 in 1,000,000 338 35.8 21.0

1 in 100,000 122 12.9 5.7

l in 10,000 61 6.5 4.9

1 in 1,000 82 8.7 3.6

1 in 100 32 3.4 3.1

1 in 10 13 1.4 21.4

1 in 5 4 .4 0.0

1 in 2 3 .3 33.3
Certain to Happen 45 4.8 35.6

 

This table reports data from 944 surveys for which questions 4 and 5 were
answered.

probabilistic perceptions of suffering a health effect and the perceived likelihood the effect will

be fatal.

Information Quantity and WTP

Casual observation of the data supports the expected relationship between WTP and the
amount of information provided by alternative advisory programs. The graphs of Figure 4.7
illustrate the pr0portion of respondents who would vote for the program as a function of
program cost, number of sites tested, and whether safe sites are listed. The probability of
accepting the program declines as cost increases. Also, acceptance probability increases with
the number of sites tested. Listing safe sites produces a roughly parallel upward shift in
acceptance proportions for both program costs and number of tested sites.

These results do not account for other factors that likely affect anglers' WTP for the

advisories. Therefore, they do not illustrate anglers' WTP for testing more sites or providing

97

Table 4.8 - Probabilistic Perceptions of Risk and Severity of Outcome

 

Mean Perceived
Outcome Likely Fatal? Risk (x 103)

 

Strongly agree 219.1
Somewhat agree 90.6
Somewhat disagree 29.7
Strongly disagree 1.0

 

Table reflects data from 932 surveys for
which questions 4 and 10 were complete.

full disclosure. However, they do suggest that respondents understood and responded in a
coherent manner to the differences among alternative programs described in the questionnaire.
They also suggest that respondents took the questionnaire seriously enough to provide a con-

sidered response.

Summary

More than anything else, this stage of the research emphasized the importance of interac-
tion with potential respondents prior to and during the process of writing a questionnaire.
Without such interaction, researchers must trust that respondents interpret questions and
responses as they themselves do. They must also take for granted that questions are clear and
unambiguous and do not elicit strong emotions or beliefs that may influence responses. Such
assumptions in this study would have been a mistake. Pretest participants routinely interpreted
questions, made assumptions, and reached conclusions that were unanticipated and inconsistent
with the researchers' interests.

Through focus groups and repeated one—on-one interviews, the draft questionnaires were

revised. The revisions generated a set of questions that were interpreted in a like manner by

98

Proportion Accepting (96)

 

100

so ----------------------------------

   
 

60 -------- ' ' - ‘ - -----------------
Full Disclosure

Partial Disclosure
40 --------------------------------

20 -----------------------------

 

 

 

0 I I I T I I I I

0.40 0.95 1.45 1.90 2.85 4.10 5.55 8.75 14.5 41.00

Program Cost

Preportion Accepting (96)
60

 

55 --- --- --- --- --- --A—e ----------------

Full Disclousre

    
 
    

50 ..................................

451 ---------------------------------
Partial Disclosure

 

 

 

4o ----------------------------------
35 -----------------------------------
30 .

110 620 1240

Sites Tested

Figure 4.7 - WTP for Program Alternatives

99

most respondents. They also ensured that questions evoked interpretations consistent with the
researcher's intentions. For example, revisions to the WTP question corrected a common but
erroneous belief that the current advisory program had tested most sites. Revisions also
clarified the differences between advisory alternatives and eliminated the problem of inter-
dependent valuation when two several programs were offered. Also, revisions of questions
about anticipated behavioral change included a richer set of response choices that corresponded
more closely to those actually perceived as options by anglers.

Another specific lesson about the importance of pretesting involved the clarity of the
questionnaire. Respondents who know little about the advisories may be confused by explana-
tions that are clear to a researcher who is very familiar with the program. Without extensive
pretesting, this questionnaire would have been confusing to many respondents and results

would not have corresponded to research needs.

Chapter Five

Estimation of a Discrete Choice Model

This chapter defines the statistical model used to estimate WTP from collected data, des-
cribes the estimation process, and presents results. The first section reviews the conceptual
link between utility theory and discrete choice estimation methods. The second section defines
the statistical model used in the analysis. The remainder of the chapter reports and interprets

regression results for three alternative functional forms of the empirical model.
Estimating WTP from Discrete Data

Referendum style questions yield binary WTP data - respondents either accept or reject
the program at the proposed price. Studies with binary data generally use discrete choice
models to obtain parameter estimates for explanatory variables. Because of the binary nature
of the dependent variable, discrete choice models typically estimate probabilities of accepting a
program rather than estimating WTP directly. Suppose an individual angler's utility, u,
depends on information, 1, income, y, other observable personal characteristics (denoted by the
vector a), and unobservable factors. The error term, 5, represents the unobservable component

of utility. Thus, utility is defined as
(5.1) u = u(y,a,i) + e.

Let 1 equal one if the angler is informed and zero if uninformed. An angler will vote for an

informative advisory program offered at price b if

(5.2) u(y-b.a,1) + e, - u(y,a,O) - so 2 0

100

101

Define x as a vector such that x=(y,a). Also define
(5.3) xa + b6 = U(y-b,a,1) - U(y,a,0)

where a and 6 are parameters to be estimated. Then, the probability that a respondent will

accept the program offered at price b is

(5.4) Pr [u(y-b,a,l) + e, - u(y,a,O) - 60 2 0]
= Pr [el - 60 2 ~(xa + b6)]

= F[-(xoz + b6)/o]

where F(') is the cumulative distribution and a the standard deviation of the error difference,
60-61.

Similarly, the probability that a respondent will reject the program at price b is

(5.5) Pr [u(y-b,a,1) + e, < u(y,a,O) + 60]

= 1 - F[-(xa + b6)/o].

Most empirical applications of discrete choice models treat program cost, bi, as an
explanatory variable. They then obtain parameter estimates ale and file by maximizing the log

likelihood function

(5.6) log L = i:{t,log[1-F(-(x,a + bfi)/O)] + (14,) log[F(-(xia + biB)/o)]}.

where n is the number of observations, ti equals one if respondent 1 accepted the program at
price bi, and ti equals zero if the respondent rejected the program. If F(-) defines the cumula-
tive normal distribution, equation 5.6 defines a probit model.

The parameter estimates from equation 5 .6 represent marginal effects of explanatory vari-
ables on the probability of acceptance if a variance of one is assumed for the error difference,
eo-q. The marginal effect of explanatory variables directly on WTP (oz and B) are not
generally recovered because a. a, and 6 are not separately estimated.

Cameron and James (1987) show that the coefficient of cost in the model of equation 5 .6,

(,8/0), is a point estimate of -1/0. Therefore. a can be calculated. The estimate of 0 permits

102

calculation of the vector 0: from the parameter estimates oz/a of equation 5.6. Thus, Cameron
and James obtain parameter estimates that represent marginal effects of explanatory variables
on WTP.

In application, Cameron and James suggest a more direct approach that eases computation
of standard errors for estimates. They reparameterize the model of equation 5 .6 and maximize

the log of the likelihood function

(5.7) log L = "2 (t, log[l-F((bi - xia)/o)] + (14,) log[F((bi - x,a)/a)]}

with respect to a and a. This study uses Cameron and James' model to estimate WTP for

additional advisory information.

An Empirical WTP Function for Advisory Information

The preceding section showed how to obtain parameter estimates that represent the
marginal effects of explanatory variables on WTP. This section defines the variables used to
empirically explain variation in WTP for advisory information. Explanatory variables include
(1) characteristics of the proposed program, (2) anticipated behavioral change in response to
additional advisory information, (3) the perceived severity of health effects associated with
eating contaminated fish, and (4) the perceived accuracy of information. This section defines

empirical measures of these concepts from survey responses.

Variable Definitions

The questionnaire elicited discrete (accept/reject/no Opinion) WTP responses for additional
advisory information. The discrete variable CHOICE measures a respondent's vote on the
referendum offered by the questionnaire. CHOICE is the dependent variable in the estimated

equation. The remainder of this section defines the explanatory variables used in the analysis.

103

It also restates the research hypotheses in terms of the empirical variables. Table 5.1 sum-

marizes variable specifications.

Program Variables

The conceptual framework hypothesizes that response to the referendum depends in part
on characteristics of the offered program. Programs pr0posed in the questionnaire differ in (1)
whether they are full or partial disclosure, (2) the number of sites tested, and (3) the cost in
higher license fees. The current partial disclosure advisory program tests about 30 sites per
year. A partial disclosure program described in the questionnaire provides more information
than the current advisory only if it tests more than 30 sites. Therefore, respondents should
evaluate a proposed partial disclosure program on the basis of the number of sites it tests above
existing levels. PSITES represents a partial disclosure program in the empirical model. The
variable PSITES equals the number of sites tested under a proposed partial disclosure program
minus 30.

A full disclosure program provides additional information about all sites tested under a
proposed program, not just unsafe sites. Therefore, a full disclosure program that tests 30
sites per year provides more information than the current advisory program. Respondents
should base WTP for full disclosure on the total number of sites tested. The variable FSITES
is the number of sites tested under a proposed program and represents a full disclosure
program in the empirical model.

In addition to different marginal effects on WTP, full and partial disclosure programs may
have different intercepts. FULLD defines a dummy variable that takes the value one for a full
disclosure program and zero otherwise. The coefficient of FULLD estimates any shift in the
WTP function related to full disclosure.

In terms of the two types of programs offered in the questionnaire, the program specific

hypotheses of chapter three can be stated as:

104

Table 5.1 - Variable Definitions

 

Variable

Definition

 

 

CHOICE
FSITES
PSITES
F U LLD
FATAL

BCHAN GE

NCHANGE

USELIST
NOREAD

SCIENCE

COLLEGE

Referendum choice
0 = voted against the program
= voted for the program

Sites tested each year to which full disclosure applies
= number of sites when full disclosure
= 0 when partial disclosure

Sites tested each year to which partial disclosure applies
= number of sites minus 30 when partial disclosure
= 0 when full disclosure

Dummy variable for full disclosure
1 = full disclosure
0 = partial disclosure

Health effect likely fatal
1 = yes
0 = no

Anticipated behavioral response to unsafe favorite site and
species

1 = change site or stop fishing

0 = otherwise

Propensity for behavioral response to current advisory
1 = eats fish from unsafe sites or has not changed behav-
ior in response to current advisory
0 = otherwise

Anticipate a behavioral response to a list of safe sites?
1 = yes
0 = no

Has respondent read the advisory?
1 = no
0 = yes

Agrees that scientists understand the risks associated with
chemical residues in fish

1 = yes

0 = no

College education?
1 = yes
0 = no

 

 

direc
telea'

but

111081

link

Ulidl

ref:

V'cll

XS

Cit

Iii

CE

inc

1153

105

1. WTP for either a full or partial disclosure program is nonnegative.

2. WTP for a full disclosure program is not less than WTP for a program of

partial disclosure.

This research focuses on how to enhance the value of the advisory. The state has the most
direct control over the number of sites tested and whether full or partial information is
released. Therefore, the hypotheses relating program variables to WTP are the fundamental
hypotheses of the study.

With respect to measurement, the hypotheses relating program variables to WTP are the
most secure of the research. There is little measurement error in the program variables - the
link to underlying concepts is relatively strong. Pretest participants had little difficulty
understanding the differences between program alternatives. Furthermore, participants seemed
to carefully consider the program variables when making a choice about the proposed
referendum.

Figure 5.1 summarizes the link between concepts related to research hypotheses, measured
variables, and hypothesized effects on WTP. The first column lists concepts and correspond-
ing variables. Column two lists the hypothesized effect of variables on estimated WTP. The
X3 in columns three and four indicate whether a particular variable affects WTP for full dis-
closure, partial disclosure, or both. The table lists hypotheses in decreasing order of con-
fidence in measurement - the program hypotheses? are the strongest and the hypotheses con—
cerning perceived accuracy the weakest. Expectations with regard to the variable COLLEGE
are not expressed as a formal hypothesis. Its placement at the bottom of the table does not

indicate relative confidence in measures of the concept.

Perceived Severity of Outcome
The conceptual framework also suggests that WTP depends on the perceived severity of

health effects associated with eating contaminated fish. Several studies suggest a link between

1

106

 

 

Concepts and Hypothesized Full Partial
Variables Effect on WTP Disclosure Disclosure

1. Partial Disclosure

PSITES Nonnegative X
2. Full Disclosure

FSITES Nonnegative X
3. Perceived Severity

FATAL Nonnegative X X
4. Behavioral Change

NCHANGE Nonpositive X

BCHANGE Nonnegative X X

USELIST Nonnegative X

NOREAD Nonpositive X X
5. Perceived Accuracy

SCIENCE Nonnegative X X
6. Education

COLLEGE Nonnegative X X

 

Figure 5.1 - Concepts, Variables, and Hypotheses

107

perceptions of risk or severity and the likelihood that a health effect is fatal (Fischhoff 1978,
Caudill 1992, Smith and Johnson 1988). Fischhoff identified the perceived likelihood of fatal-
ity as one of twelve factors that affect peoples' perceptions of risk. Caudill used factor analy-
sis to explore perceptions of risk from groundwater contamination. He found the perceived
likelihood of a fatal health effect to be a significant component of the severity factor. The vari-
able FATAL in the empirical model measures the perceived likelihood of a fatal health effect
and represents anglers' perceptions of severity. The following hypothesis relates WTP to

beliefs about the likelihood of a fatal health effect.

3. WTP for either full or partial disclosure programs is nondecreasing in the

perceived likelihood that outcomes are fatal.

The variable FATAL does not correspond to the underlying concept of perceived severity
as well as measures of the program variables relate to intended concepts. Interview
participants seemed to have well defined beliefs about the likelihood that a health effect would
be fatal. However, FATAL does not capture beliefs about severity itself - it is only a measure
of the perceived likelihood of the most severe outcome, a fatality.

Pretest questions experimented with the terms "severity" or "seriousness" to measure
degrees of perceived severity of health outcomes. However, participants often interpreted the
questions as identical to the question that asked for a probabilistic guess about perceived risk.
Thus, the terms severity and serious often elicited a response about the probability of a health
effect rather than the intended concept of severity. The placement of the severity question
immediately following the probability question may have influenced this interpretation. Addi-
tional pretesting may have developed a question that adequately measured perceived severity.
However, lacking a good measure of perceived severity, this research relied on a measure of

the perceived likelihood of a fatal health outcome.

108

Anticipated Bebgvioral Change

The conceptual model also suggests that anglers will be willing to pay for advisory
information only if they anticipate changing their behavior in response. The questionnaire
obtained relatively poor measures of anticipated behavioral change. Even with extensive
pretesting, a mail survey is not well suited to capturing the richness of anglers' behavioral
response to advisory information. The ex ante nature of the questions also increased the effort
required of respondents to provide considered and meaningful answers. To decide about
behavioral responses, respondents had to consider how the information might affect future fish-
ing behavior without knowing the specific signals an advisory would provide.

This analysis includes two measures of anticipated behavioral change that apply to both
full and partial disclosure programs. The variable BCHANGE identifies anglers who would
change sites or stop fishing if the advisory listed a favorite species and site as contaminated.
The variable NOREAD identifies anglers who have not read the advisory. A failure to read
the current advisory may only mean that a respondent was unaware of the advisory. Once
made aware (by the questionnaire or through other channels), this group may value advisory
information. However, NOREAD may also identify respondents who know about chemical
residues but are not concerned enough to read the advisory. To the extent that NOREAD
identifies the latter group, it corresponds to anglers who should place little value on additional
information.

The analysis also includes two measures of anticipated behavioral change that apply only
to a single program. The variable USELIST indicates whether respondents anticipate changing
behavior in response to a list of safe sites. Since a partial disclosure program does not list safe
sites, USELIST applies only to the value of a full disclosure program. The variable
NCHANGE identifies anglers who are not disposed to behavioral change. This group includes
respondents who reported doing nothing differently in response to the current advisory. It also
includes those who currently practice consumption behavior at odds with advisory recom-

mendations -- those who reported eating one or more meals per week from inland lakes or the

109

Great Lakes. NCHANGE represents anglers' response to the current partial disclosure
advisory. It does not imply a lack of behavioral response to a full disclosure advisory. There-
fore, NCHANGE applies only to the value of a partial disclosure advisory.

Conceptually, anglers who do not anticipate using advisory information will be willing to
pay nothing for information. However, the questionnaire does not perfectly capture anticipated
behavioral change. Therefore, the empirical hypotheses stated here are weaker than the con-

ceptual hypotheses.

4. Respondents who anticipate changing behavior in response to advisory
information will be willing to pay at least as much for additional informa-

tion as those who anticipate no behavioral change.

a. Respondents who report that they would change behavior if the
advisory listed their favorite site and species as unsafe will be willing
to pay no less for advisory information than those who would not

change behavior.

b. Respondents who have not read the advisory will be willing to pay no

more for additional information than those who have read the advisory.

c. Respondents who would change behavior in response to a list of safe
sites will be willing to pay no less for a full disclosure advisory

program than respondents who would not change behavior.

(1. Respondents who have not changed behavior in response to the current
advisory or those who currently practice consumption behavior at odds

with advisory recommendations will be willing to pay no more for a

110

partial disclosure advisory than those who have changed behavior or do

not engage in risky consumption behavior.

Because of problems measuring anticipated behavioral change, the hypotheses listed above
are somewhat exploratory. Rejection of a hypothesis does not necessarily imply that a particu-
lar behavioral change measure is unimportant -- it may mean only that the concept was not

measured adequately.

Perceived Accurgcv of Advisory

The perceived accuracy of advisory information should also affect WTP. However, the
questionnaire obtains poor measures of perceived accuracy. Perhaps the most severe problem
is that it measures separate dimensions of accuracy -- beliefs about the accuracy of scientific
knowledge and the perceived accuracy of reporting. Measures of only some of the individual
dimensions of perceived accuracy identify some of the anglers who believe the advisory to be
inaccurate but not those who believe it to be accurate. For instance, a belief that any single
dimension is inaccurate implies a belief that the advisory as a whole is inaccurate. Therefore,
a respondent who believes any measured dimension to be inaccurate believes the advisory to be
inaccurate. However, a respondent who believes all measured dimensions to be accurate may
still believe the advisory to be inaccurate if they question the accuracy of an unmeasured
dimension. Because the questionnaire does not measure all dimensions of perceived accuracy,
it does not identify respondents who believe the advisory to be accurate.

The variable SCIENCE identifies anglers who believe the advisory to be inaccurate
because scientific knowledge is inaccurate. The questionnaire asked whether respondents
believed that the current advisory understated the risk from chemical residues in fish. The
question did not capture perceived accuracy because it did not distinguish between respondents
who believed the advisory overstated risk and those who believed it accurately portrayed risk.

A belief that reporting is inaccurate also implies a belief that the overall advisory is inaccurate.

111

However, the questionnaire did not obtain an adequate measure of perceived accuracy of
reporting. The empirical model includes only the variable SCIENCE as a measure of per-

ceived advisory accuracy.

5. WTP for either a full or partial disclosure advisory program is nonincreas-

ing in the perceived inaccuracy of scientific knowledge.

Because of measurement difficulties, expectations with regard to tests of the perceived accuracy

hypothesis are low.

Education
The variable COLLEGE identifies respondents who reported at least some college educa-
tion. Education should positively affect WTP for advisory information. More educated

anglers should better understand the complex advisory and its implications.
Estimation and Tests of Hypotheses

The analysis uses the estimation approach suggested by Cameron and James (1987). The
resulting estimated coefficients represent marginal effects of explanatory variables on WTP.
Equations were estimated with Limdep's ‘ MINIMIZE command. The command minimized
the negative of the log likelihood function
(5.8) log L = E {CHOICEi log[1-<I>((COST,-18x,)/a)]

+ (1-CHOICE,)log[<1>((COST,-Bx,)/o)]}
where:

(3xi = 13,18xi

<1) = the standard normal cumulative density.

 

' Limdep is a computer program that focuses on statistical analysis of limited dependent variables models (Greene

1991).

$

112

This section addresses the functional specification of the empirical model and tests of the

hypotheses stated in the previous section.

Model Specification

The functional relationship between WTP and the explanatory variables listed in Table 5.1
can take many forms. A desirable specification should (1) be consistent with the conceptual
framework, (2) facilitate tests of research hypotheses, and (3) accurately predict the
referendum choice. The conceptual analysis suggests that the marginal value of testing with a
full disclosure program is not less than with partial disclosure. Therefore, the specified
empirical model separates the effects of the two programs on WTP. The variables PSITES and
FSITES in the model of Table 5 .2 represent the marginal effects of testing an additional site
under partial and full disclosure respectively. The variable FULLD defines a shift variable for
a full disclosure advisory.

An additional specification issue concerns the role of non-program variables in the empiri-
cal model. The non-program variables are FATAL, BCHANGE, NOREAD, NCHANGE,
USELIST, SCIENCE, and COLLEGE. Two questions arise: (1) whether non-program vari-
ables affect WTP at the margin or whether they shift the WTP function, and (2) whether non-
program variables are significant determinants of WTP. Appendix B presents results of
likelihood ratio tests of hypotheses related to these questions. The tests fail to reject the
hypothesis that interaction forms of the non-program variables can be jointly restricted to zero.
Therefore, interaction forms of the non-program variables have no significant joint effect on
marginal WTP and they do not appear in the regression models. The tests also reject the
hypothesis that level (non-interacted) forms of non-program variables can be jointly restricted

'to zero. Therefore, the regression models include non-program variables in level form only.

113

WTP for Full Versus Partial Disclosure

Table 5.2 reports regression results for a linear WTP function. All non-program variables
enter the model as deviations from mean values. The model supports the hypotheses that WTP
for both full and partial disclosure advisories is nonnegative. For the average angler -— an
angler with average values for non-program variables -- the estimated coefficient of FSITES
represents marginal WTP for testing an additional site with full disclosure. An average angler
is willing to pay $0046 per tested site with full disclosure. Marginal WTP for full disclosure
information is different from zero at the ten percent significance level. Marginal WTP for par-
tial disclosure information -- the estimated coefficient of PSITES, $.0045 -- is essentially equal
to marginal WTP for full disclosure. However, estimated marginal WTP for partial disclosure
does not differ significantly from zero. This suggests that anglers are willing to pay less to test
a site when the advisory releases only partial information than when it fully discloses test
results.

The finding that marginal WTP for partial disclosure information is not statistically distin-
guishable from zero seems consistent with rational behavior. Additional testing with partial
disclosure produces, at best, a small improvement in an angler's decision environment. For
example, suppose an angler wishes to avoid eating contaminated fish. Short of not eating fish,
the angler's rational choice with partial information is to eat fish only from no report sites.
Assume the angler believes that 20 percent of the presently 5,500 no report sites are unsafe.
The chance of eating contaminated fish while fishing only at no report sites is thus .20.

Now suppose tests of 200 additional no report sites identify 40 (20 percent) unsafe sites.
The testing reduces the number of no report sites to 5,460 - 5,500 original no report sites
minus 40 sites newly listed as unsafe. It also identifies 40 of the 1,100 (20 percent of 5,500)
unsafe sites. Therefore. 1.060 unsafe sites still exist among the 5,460 remaining no report
sites. The revised chance that a no report site is unsafe is thus 19.4 percent -- a very small

reduction in risk relative to 20.0 percent.

114

Table 5.2 - Estimated WTP for Full and Partial Disclosure

 

 

. Coefficient
Variable Estimates
CONSTANT .3457

(2.202)
FSITES 0046*
(.0027)
PSITES .0045
(.0028)
FULLD 5643*
(3.151)
FATAL 7.245***
(2.120)
BCHANGE 1.702
(1.847)
NOREAD -2.658
(2.777)
NCHANGE -.3652
(2.645)
USELIST 15.612***
(2.979)
SCIENCE 2.688
(1.901)
COLLEGE 6.836***
(1.866)
Log-Likelihood 428.67
% Correct Predictions 71.9%

 

*

significant for a = .1
significant for a = .05
Significant for a = .01

**
***

Regression results for the linear model
based on 758 observations.

All. non-program variables enter as
devratrons from mean values.

Numbers in parentheses are standard
errors.

115

A partial disclosure advisory tells anglers about unsafe sites without telling them about
safe alternatives. In the absence of private information, the chance of unintentionally fishing at
an unsafe site is the chance that a no report site is unsafe. A full disclosure advisory, on the
other hand, tells anglers about relatively safe alternatives to either unsafe or no report sites.
To the extent the advisory accurately identifies safe sites, full disclosure eliminates the risk of
eating contaminated fish. Full disclosure would seem to bring about a non—trivial reduction in
risk for an angler who wishes to avoid contaminated fish and believes that some no report sites
are unsafe.

Estimated WTP for testing levels that imply no additional information provides a check of
the correspondence between the estimated model and the behavioral implications of economic
theory. The current advisory program tests 30 sites per year and provides partial disclosure of
test results. A model consistent with theory should yield zero WTP for a partial disclosure
advisory that tests 30 sites - an angler should not be willing to accept an increase in license
fees to continue the current advisory. The constant term of the model in Table 5.2 represents
WTP for testing 30 sites with partial disclosure. 2 The fact that the constant is not statistically
different from zero supports the implications of economically rational behavior.

The above argument also implies that estimated WTP for a full disclosure program that
tests no sites should be zero. An angler should not be willing to pay a positive amount for an
advisory program that provides no information. However, the positive and significant coeffi-
cient of the full disclosure dummy variable, FULLD, implies that anglers are willing to pay
$5.64 for a full disclosure advisory that tests zero sites. The following section asks whether

this counterintuitive result is an artifact of the linear model.

 

‘

" PSITES equals the number of sites tested minus 30. Therefore, the estimated constant term represents WTP for
testing 30 sites with partial disclosure.

116

WTP for Full Disclosure

Suppose the data actually describe a marginal WTP function that is positive and decreas-
ing and has an intercept near zero. A linear functional form fitted to the data may produce a
relatively large positive intercept and overestimate WTP for small levels of testing. Table 5.3
reports results of two nonlinear models - a quadratic and a square root model -- that examine
the possibility that WTP is a nonlinear function of testing. The linear model implies that WTP
for partial disclosure is insignificant. Therefore, the nonlinear models estimate WTP for full
disclosure only. The models do not include the variables PSITES and NCHANGE because
these variables apply only to partial disclosure. They also do not contain the variable FULLD
that shifts the linear model for full disclosure.

The first two columns of Table 5.3 report the quadratic and square root models respec-
tively. The conceptual results suggest that the constant term in both models should be zero --
anglers should not be willing to pay for no information. Statistically, the constant terms are
not significantly different from zero. Likelihood ratio tests imply that restricting the constant
terms to zero does not significantly affect the explanatory power of the estimated equations. 3
Therefore, the nonlinear models suggest that WTP for a full disclosure advisory that tests no
sites is effectively zero — a result consistent with conceptual expectations.

The estimates of Table 5.3 support the hypothesis that the WTP function for full dis-
closure may be nonlinear. The nonlinear terms are significant and suggest that marginal WTP
is decreasing in the number of sites tested. Furthermore, a comparison of the quadratic and

square root models with the restricted model in the final column suggest that the linear and

 

3 The likelihood ratio statistic associated with the null hypothesis that the constant term of the quadratic model can

be restricted to zero is 1.06. The likelihood ratio statistic for restriction of the constant term of the square root
model is .18. The X: critical value for a=.l and 1 degree of freedom is 2.71. Therefore. the tests fail to reject the
hypotheses that the constant terms in either model can be restricted to zero -- the restrictions are not significant.

 

117

Table 5.3 - Nonlinear WTP Functions

 

 

Quadratic Square Root Restricted
Variable Model Model Model
CONSTANT 3 .332 -3.518
g (3.242) (8.453)
FSITES .0190 -.0176
(.0127) (.0190)
FSITES2 -.1055E“
(.8962E'5)
FSITES.5 1.026
(.8719)
FATAL 8.175*** 8.175*** 15.205***
(3 .068) (3.068) (5.975)
BCHANGE .9659 .9659 .1225
(2.710) (2.710) (4.945)
NOREAD —3.356 -3.356 6.963
(3.989) (3.989) (7.166)
USELIST 15.812*** 15.812*** 29.393***
(3.279) (3.279) (7.594)
SCIENCE 2.391 2.391 7.266
(2.774) (2.774) (5.211)
COLLEGE 7.875*** 7.875*** 12.992***
(2.691) (2.691) (5.303)
Log-Likelihood -213.87 -213.87 -232.94
% Correct Predictions 74.8% 74.8% 73.3%

 

*
**
***

significant for a = .1
significant for a = .05
significant for a = .01

Regression results for the quadratic and square root models based on 397
observations for full disclosure of information.

All non-program variables enter as deviations from mean values.

Numbers in parentheses are standard errors.

 

118

nonlinear forms of FSITES are jointly significant in both nonlinear models. 4 Therefore, the
nonlinear models are statistically superior to the restricted model.

Table 5.4 reports regression results for the final linear and nonlinear models of WTP for
full disclosure. The linear model differs from the model of Table 5.2 because it estimates
WTP for full disclosure only. Therefore, it does not include the variables PSITES, FULLD,
or NCHANGE. The nonlinear models do not contain the insignificant constant terms found in
the models of Table 5.3. Little statistical basis exists to choose among the three models of
Table 5 .4. There is no statistical difference in explanatory power and predictive accuracies are
very similar -- 73.8 percent for the linear model, 74.0 percent for the quadratic, and 74.3 per-
cent for the square root. Economic theory suggests a declining marginal WTP -- a rationale
that favors the nonlinear forms. However, the analysis suggests no reason to prefer one non-
linear form over the other.

All three models of Table 5.4 support the hypothesis that WTP for full disclosure is non-
negative. As a consequence of the chosen functional forms, the nonlinear models imply nega-
tive marginal WTP for extremes in testing - more than 870 sites for the quadratic model and
more than 1,068 sites for the square root model. However, over a large relevant range of test-
ing around current levels, both models imply that marginal WTP is positive and decreasing - a
result consistent with economic theory. Table 5.5 illustrates total and marginal WTP values
for full disclosure implied by each of the three alternative models. Since marginal values from
a linear function are not particularly meaningful, they are not included in the table.

Table 5.5 shows that total WTP for full disclosure is positive for testing levels between
zero and 300 sites per year which represents a tenfold increase over current levels. Because of

the large positive constant, the linear model of column one yields greater total WTP for

 

4

The likelihood ratio statistic associated with restricting FSITES and FSITES2 in the quadratic model or FSITES
and FSITES" in the square root model to zero is 38.14. The x3 critical value for a=.01 and 1 degree of freedom is
6.63. Therefore. the test rejects both restrictions - the linear and nonlinear forms of FSITES contribute sig-
nificantly to explaining variation in estimated WTP.

Table 5.4 - Final WTP Estimates

119

 

 

Linear Quadratic Square Root
Variable Model Model Model
CONSTANT 6.041***
(2.264)
FSITES .0046* .0300*** -.0103
(.0027) (.0074) (.0073)
FSITES2 -.00002***
(.000006)
FSITES-5 .6756***
(.2322)
FATAL 7863*“ 8639*“ 8.027***
(3.044) (3.139) (3.016)
BCHANGE .7247 1.088 .8874
(2.672) (2.789) (2.669)
NOREAD 3.543 -3.208 -3.426
(3 .979) (4.1 14) (3.965)
USELIST 15.897*** 16.352*** 15.749***
(3.283) (3.348) (3.251)
SCIENCE 2.440 2.566 2.374
(2.777) (2.851) . (2.758)
COLLEGE 8.015*** . 5*** 7.889***
(2.700) (2.775) (2.680)
Log-Likelihood -214.60 ~214.40 ~213.96
% Correct Predictions 73.8% 74.0% 74.3%

 

*
**
***

significant for oz
significant for a
Significant for or

= .1
= .05
= .01

Regression results for the quadratic and square root models based on 397
observations for full disclosure of information.

All non-program variables enter as deviations from mean values.

Numbers in parentheses are standard errors.

120

Table 5.5 - Total and Marginal WTP for Full Disclosure Information

 

 

Linear Quadratic Square Root
Model Model Model
Total WTP for Full Disclosure
Sites Tested
0 $6.04 $0.00 $0.00
30 $6.18 $.88 $3.39
50 $6.27 $1.46 $4.26
100 $6.50 $2.82 $5.72
200 $6.96 $5.31 $7.49
300 $7.42 $7.44 $8.60
Marginal WTP for Full Disclosure
Sites Tested
0 $0300 N .D.
30 $.0289 $.0513
50 $0282 $.O374
100 $.0265 3.0234
200 $.0231 $0136
300 $.0196 $.0092

 

N .D. means not defined.

121

relatively low levels of testing -- below 100 sites per year. However, estimated total WTP
from the

square root model of column three surpasses that of the linear model at a level of testing of
about 300 sites per year. Similarly, the quadratic model yields greater total WTP than the
linear model for testing levels greater than or equal to 300 sites per year. Relative to the non-
linear models, the linear model inflates total WTP for levels of testing near the current level.

The two nonlinear models differ primarily in the rate of change of marginal WTP. The
quadratic model of column two implies a constant decline in marginal WTP as testing
increases. The square root model implies a relatively large marginal WTP for small levels of
testing (below about 50 sites). However, marginal WTP in the square root model declines
rapidly at first and less rapidly as testing levels increase. Figure 5.2 graphically illustrates the
relationship between marginal WTP for the quadratic and square root models.

Estimates of net benefits -- the difference between total WTP and total cost -- are probably
most relevant to the practical questions that motivate this research. Discussion with MDNR
officials produced a rough cost estimate of $.01 per angler to test an additional site. The
marginal cost of providing a full disclosure advisory also includes the cost of printing informa-
tion about an additional test. Pages must be added to the advisory in multiples of eight. Print-
ing cost per angler to add eight pages to the advisory is about $.01. 5 Therefore, additional
printing costs are near zero if additional information can be included in the existing space and
may reach $.01 per angler if additional pages must be added.

Figure 5.2 graphs total values for the three estimated functional forms - linear, quadratic,
and square root. The figure also graphs two total cost curves. The curve TCl corresponds to
a marginal cost of testing of $.01 per angler per site. TC2 represents the total costs associated
with a marginal cost of $.02 per angler per site.

The practical implications of the research are not particularly sensitive to either the choice

of functional form or assumptions about testing and printing costs. The total WTP curves lie

 

5 Personal communication with Ned Fogle. Michigan Department of Natural Resources. Fisheries Division.

122

Total WTP/Cost ($)

 

14

1 2 Quadratic

 

 

 

 

IIIIleIIW

600 800 1000 1200
Sites Tested

 

Marginal WTP/Cost ($)

 

0.04

0.03

0.02

0.01

 

-0.01

 

 

 

 

-O.02rirrrTrrrrrrrrTrrrrTIrr

0 200 400 600 800 1 000 1 200
Sites Tested

Figure 5.2 - Total and Marginal WTP for Full Disclosure Information

123

above the total cost curve TCl for all testing levels below 1,100 sites -- the level where TCl
intersects total WTP for the linear model. Therefore. net benefits are positive for all three
functional forms for all testing levels below 1,100 sites. Furthermore, even if actual testing
costs are twice the estimated $.01 per site per angler, all three models yield positive net bene-
fits for testing levels below 579 sites -- the level at which TC2 intersects total WTP from the

quadratic model.

Predicted Probability of Accepting the Advisorv Program

The probit model yields a normally distributed index of the probability that a respondent
will accept the referendum. Table 8.3 of appendix B reports the probity estimates from which
index values are derived. The cumulative normal distribution evaluated at the index defines
the probability of acceptance as a function of the explanatory variables that produce the index
value. Figure 5.3 graphs the predicted probability that an average angler will accept a full dis-
closure program as a function of cost. The graph illustrates predictions of the linear model for
each level of testing presented in the questionnaire. It also shows that the probability of
acceptance is decreasing in program cost. Although they are not graphed, the quadratic and
square root models also exhibit decreasing probability of acceptance as cost increases.

Table 5.4 shows the program costs that correspond to an estimated 50 percent chance of
acceptance for three rates of testing near current levels -— 50, 100, and 150 sites. Each of the
models predict that the probability of acceptance is increasing in the number of sites tested.
Furthermore, if testing costs are about $.01 per angler per site, the models predict that a

majority of anglers will accept a license fee increase that covers the cost of testing.
Tests of Hypotheses

A likelihood ratio test reported in appendix B fails to reject the hypothesis that the non-

program variables can be jointly excluded from the empirical model. The test implies that,

124

Program Cost ($)

 

70

60 .................. . ...................................................

.....................................................................

  
  

..................................................................

1,240 Sites

....................................................................

620 Sites

20 ........... f ......................................................
110 Sites

10 ..................................................................

 

 

 

0 IIIIIIIIIIIllIlIIIIITIIllIIIIIIIIIIITIIIIIIIIIHIIIIIIIIIIIIIIIIIIIIIIIIIIIIIWIIIIIIIIIIIIIIIIIIII

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of Acceptance - Linear Model

Figure 5.3 - Predicted Probability of Acceptance

 

125

Table 5.4 - Program Costs Associated with 50 Percent Probability of Acceptance

 

 

Sites Linear Quadratic Square Root
Tested Model Model Model

50 $6.22 $1.46 $4.26

100 $6.44 $2.82 $5.72

150 $6.67 $4.11 $6.72

 

jointly, the non-program variables contribute significantly to explaining variation in WTP.
This section reviews conclusions about the hypotheses related to non-program variables.
Whether survey questions related to these hypotheses accurately measured intended concepts is
an empirical question. Failure to reject a particular hypothesis suggests that the corresponding
variable may have captured the concept. Rejection of a hypotheses does not invalidate the
model -- it may merely mean that-the corresponding variable did not adequately measure the
intended concept.

Non-program variables enter the estimated equation only in level form. Therefore, they
shift the estimated WTP function and do not effect estimates of marginal WTP. All coeffi-
cients of non-program variables reported in Table 5.2 have signs consistent with the
hypotheses of Figure 5.1. Strict interpretation of the level forms and estimated coefficients
puts extreme demands on the estimation procedure and yields some counterintuitive conclu—
sions. It is unlikely, for instance, that anglers who believe possible health effects to be fatal
are willing to pay more at every level of testing than anglers who believe health effects are
unlikely to be fatal. Failure to reject a hypothesis suggests only that a particular variable

affects WTP, it implies little about the specific form or magnitude of the effect.

126

Perceived Severity of Outcome

The conceptual framework suggests that anglers who believe that the health effects of
eating contaminated fish are likely fatal will be willing to pay at least as much for advisory
information as those who believe fatality to be unlikely. The estimated coefficient of FATAL
in the linear model is positive and significant. This result supports the hypothesized relation-
ship between perceived likelihood of fatality and information value -- estimated WTP for

advisory information is nondecreasing in the perceived likelihood of a fatal health outcome.

BehgviorMhmg

Anticipated behavioral change in response to advisory information also appears to sig-
nificantly affect WTP. A likelihood ratio test reported in appendix B fails to reject the
hypothesis that the four behavioral change variables can be jointly excluded from the estimated
model. All four behavioral change measures have the expected sign -- BCHANGE and
USELIST have a nonnegative effect on estimated WTP, NOREAD and NCHANGE have non-
positive effects. The signs are consistent with the general hypothesis that anglers who
anticipate a behavioral change in response to additional information are willing to pay at least
as much as those who do not foresee making a behavioral change.

The variable USELIST is the only behavioral change measure with a significant coeffi-
cient. However, the fact that the coefficients of the remaining behavioral change variables

have the expected signs suggests that they captured some aspects of the intended concepts.

Perceived Accurgcy of Advisory

The conceptual framework also suggests that information value is nondecreasing in per-
ceived accuracy. The nonnegative coefficient of SCIENCE suggests that respondents who
believe advisories to be inaccurate because of insufficient scientific knowledge are willing to

pay no more for advisory information than other respondents.

127

We

While not stated as a formal hypothesis, the estimated coefficient of COLLEGE is also
consistent with conceptual expectations. The significance of COLLEGE as an explanatory
variable suggests that education is an important determinant of an angler's response to the

advisory.

Summary

The preceding analysis produces three primary conclusions. First, additional testing
under the current partial disclosure advisory likely has little value to anglers. Second, WTP
for a full disclosure advisory program is positive at current testing levels and increasing in the
number of sites tested. Finally, full disclosure produces positive net benefits for a large range
of testing levels around the current 30 sites per year. The final result is robust with respect to
alternative functional forms and assumptions about the costs of providing a full disclosure

advisory.

Chapter Six

Summary and Conclusions

This research had three primary objectives. First, it developed a Bayesian conceptual
framework that linked characteristics of information systems to economic value. Second, it
adapted CV techniques to obtain the empirical measures needed to test the conceptual
hypotheses in the context of sport anglers' WTP for information about chemical residues in
fish. Third, it asked the practical questions of whether Michigan's current public health
advisory has value to anglers and how the state might enhance advisory value. The process of
addressing these objectives produced useful insights in each of the three primary components
of the research. Conceptually, it clarified the definition of information value relative to the
welfare effects associated with contamination. It also emphasized the importance of pretesting
in the design of an effective CV survey. Finally, it produced WTP estimates that may guide
practical decisions about the future of Michigan's public health advisory program.

This chapter first reviews the findings from each of the components of the research listed
above. It also emphasizes the limitations of the findings and suggests areas where future

research may prove fruitful.

Conceptual Conclusions

The research identified two measures of ex ante information value used in the literature.
It showed them to be algebraically equivalent if posterior beliefs were formed using Bayes'
rule. The conceptual development also clarified the difference between information value and

the welfare loss associated with environmental contamination. Contamination of food or the

129

environment may cause a loss of welfare as people engage in risk averting behavior. However.
the loss would be greater if ignorance prevented the behavioral adjustment.

Confusion between nonnegative information value and negative welfare effects of con-
tamination may influence public information policies. Johnson (1988) does not distinguish
between WTP for information and welfare losses associated with EDB contamination of grain
products. Furthermore, he defines ex post WTP for information relative to a state of "correct"
information rather than the most recent posterior perceptions of risk. He concludes that
information value may be negative if it is incorrect. However, the health risks associated with
food and environmental contamination are never known with certainty. Furthermore, the best
scientific estimates of risks change over time with the accumulation of evidence. Johnson's
analysis suggests that it may not be in the public interest to release risk information when it
may ultimately turn out to be incorrect. The conceptual results of this research suggest that
health risk information cannot make people worse off and, therefore, should not be withheld.

Nonnegative information value does not imply that additional information will always
improve welfare. Information has strictly positive ex ante value only if it creates the pos-
sibility of behavioral change --"the information must be decision relevant. Furthermore, the
net value of information depends on its cost and the monetary and non-monetary costs of inter-
preting and using it. This suggests several practical considerations for the design of informa-
tion programs. First, information will be more useful (valuable) if it identifies states that are
relevant to information users' behavioral decisions. Second, informing people about risk
reducing behaviors may enlarge the set of perceived behavioral responses to information and
enhance the value of existing signals. Finally, information issued in a form that is easily

understood and applied by users has a greater net value than complex information.

Designing an Effective CV Survey

The contributions of this research to the practice of CV relate to both the measurement of

information value and the application of the CV approach. The research builds on the ex post

130

analysis of Foster and Just (1989) and adapts the CV method to the analysis of ex ante informa-
tion value. Results of focus group, interview, and mail pretests emphasize the importance of
contact with the respondent population prior to implementation of a CV survey.

The process of survey design was one of the more educational aspects of this research.
The CV survey instrument directly impacts the quality of the data and, ultimately, the analysis.
Extensive pretesting influenced the survey used in this research in several ways. First, focus
group research identified the language anglers used to think about chemical residues and the
advisory. It also identified anglers' interpretation of the current advisory and the behavioral
responses they considered effective. Second, iterative one-on-one interviews focused on
respondents' interpretation of specific questions. The interviews were crucial to ensuring that
questions measured intended concepts.

A comparison of WTP estimates from the pretest and final surveys emphasizes the impor-
tance of clear descriptions of the contingent market. The pretest survey asked for respondent's
WTP for (1) continuation of the current advisory program, (2) a partial disclosure program that
tested either 100 or 300 sites per year, and (3) a full disclosure program that listed either 200
or 600 safe sites. Figure 4.1 of chapter four illustrates the form of the three pretest WTP
questions. An open ended question that asked for respondents' maximum WTP followed each
’ of the questions. Table 6.1 lists mean WTP values for each program offered in the pretest.
Statistical analysis reveals no significant difference in WTP between mean bids for any of the
programs described for a level of significance of a= .01.

The similarity of the mean bids suggests that respondents may not have understood or
responded to the differences between programs. The uniformity of bids within surveys further
supports the notion that respondents did not interpret the programs as intended. A majority of
respondents (72%) stated the same WTP for full as for partial disclosure advisories and 56 per-
cent stated identical WTP for the current advisory, full disclosure, and partial disclosure with
increased testing. This result seems remarkable given the significant quantitative differences

between proposed programs. However. the result could arise if (1) respondents

131

Table 6.1 - Pretest WTP Means

 

 

- Number of Mean
Program Sites WTP
Partial Disclosure 30 $3.49
Partial Disclosure 100 $3.57
Partial Disclosure 300 $3.06
Full Disclosure 200 $2.82
Full Disclosure 600 $2.95

 

understand the differences but do not perceive them to be large, (2) they understand the dif-
ferences but do not view them as essential to the good or program to be valued, or (3) they do
not clearly perceive the differences due to poor question design.

Additional interviews improved the clarity of the questionnaire. In particular, the final
survey included an expanded description of the current advisory (Figure 4.2) and a tabular
comparison of program alternatives (Figure 4.4). Participants in interviews with the final
questionnaire format seemed to quickly and accurately identify important features of the cur-
rent advisory program. They also exhibited a firm understanding of the differences between
alternative programs.

The comparison of pretest and final WTP estimates suggests that poor survey design
begets questionable data. In this case study, meaningful WTP responses depended crucially on
a clear description of the contingent market. Pretesting, in turn, was a necessary step in the
evolution of a clear survey. In retrospect, had the research proceeded from the initial draft of
the survey without pretesting, many questions would not have measured intended concepts and

the empirical results would likely have been meaningless.

132

Despite extensive pretesting, the survey had limitations. A great deal of prior effort
focused on describing the existing program and proposed alternatives. As a result, measures of
program variables provided strong tests of program hypotheses. However, the survey could
probably have obtained better measures of many non-program explanatory variables. For
example, both the conceptual framework and qualitative impressions from pretesting suggest
that perceived accuracy is an important determinant of advisory value. The survey did not
obtain adequate measures of perceived accuracy. Therefore, the empirical importance of per-
ceived accuracy remains an unanswered question.

Also, the survey did not obtain good measures of anticipated behavioral change. The
questions were poorly thought out in terms of anglers' actual behavioral options. For instance,
the analysis ultimately relied on relatively complex combinations of responses to several ques-
tions to measure some aspects of behavioral change. More forethought and care in design
would have identified and obtained these measures more directly. In general, more prior effort
in designing questions related to non-program variables may have obtained better measures and
stronger, more informative empirical conclusions.

Quantitative measures of respondents' understanding of the pretest and final questionnaire
may strengthen the conclusions regarding the importance of pretesting. A simple test of com-
prehension could easily provide such measures. Also, the technique of verbal protocol analysis
from psychology may provide more concrete measures of how people perceive and respond to
different forms of pretest CV questions (Schkrade and Payne 1994). Such analysis could con-
tribute to a researcher's understanding of how respondents think about CV questions. Thus, it
could help ensure that respondents perceive and answer the questions the researcher intends to
ask.

Another issue concerns the process of estimating WTP from discrete (referendum) data.
The model used in this study assumes that WTP is normally distributed (Cameron and James
1987). However, discrete choice responses are often not normal (Duffield and Patterson 1991,

Cooper 1993). If the normality assumption is incorrect. the parameter estimates from the

133

probit model are inconsistent. Nonparametric estimation methods do not require a distrib-
utional assumption and may provide better parameter estimates if WTP is not normally distrib-
uted.

Several nonparametric techniques exist. Kristrom (1990) suggests a simple nonparametric
estimation approach. However, his approach is not easily applied to models that contain many
explanatory variables. Greene (1990) describes an alternative nonparametric estimator - the
maximum score estimator. Future research could compare the predictive abilities of the para-

metric probit model against a nonparametric approach.

Policy Implications

The research revealed that a majority of surveyed anglers (63.9%) have changed their
fishing behavior in response to the current advisory. Information has positive ex post value if
it changes behavior. Therefore, most anglers appear to value Michigan's current public health
advisory program. While the current advisory is valuable, the research also finds that many
anglers are willing to pay a positive amount for some types of additional information.

Additional testing with only partial disclosure of test results produces little improvement
in an angler's decision environment. It slightly increases the chance that fish from a site not
explicitly listed in the advisory are safe to eat. Consequently, respondents to the CV survey
were willing to pay nothing for additional testing under the current partial disclosure program.
However, full disclosure of test results provides anglers with known safe alternatives to con-
taminated sites and significantly reduces the chance of eating contaminated fish. Respondents
were willing to pay a significant amount for an advisory that fully disclosed test results.

The value of a full disclosure advisory program is positive at the current level of testing
and increasing with the number of sites tested annually. The results of this study do not reveal
the optimal level of testing. However, the benefits of additional testing with full disclosure

exceed the probable costs of providing the information for a wide range of testing around cur-

134

rent levels. This result is robust for different assumptions about both the form of the WTP
function and the cost of the advisory program.

The CV survey used in this research did not describe where additional testing would take
place. Therefore, WTP estimates depend on respondent's assumptions about the geographical
distribution of pr0posed testing. However, the perceived geographical pattern of testing may
strongly influence angler's WTP for information. For example, an angler who fishes only in
one area is not likely to value testing in other areas. Similarly, even anglers who are willing to
substitute sites between geographical areas will likely be sensitive to the relative proximity of
different locations. Travel costs and time constraints would imply a greater WTP for informa-
tion about nearby areas than for'those further away. Further research might help identify dif-
ferences in information value across different regions of the state. The information might pro—

vide guidance for targeting future testing.

Appendices

Appendix A
Questionnaire and Related Materials

 

MICHIGAN ’S
SPORT FISH
CONSUMPTION
ADVISORIES

 

 

Copyright 1992 Michigan State University

135

136

Do you regularly do any of the following? (Circle all that apply)

1. Firearm or bow hunting
2. Bird or wildlife viewing
3. Camping

4. None of the above

About how many times per year do you fish at the following types of
sites? (Fill in numbers)

Great Lakes

Inland lakes or ponds

Rivers or streams
Other

 

On average, throughout the year, about how often do you eat fish
that you catch in Michigan? (Circle one number)

I do not eat fish that I catch
Less than one meal per week
About one meal per week
Two or more meals per week

93"?!"

What do you think is the chance that you will someday have health
problems because of chemical residues in Michigan's sport fish?
(Circle one number)

No chance 6. 1 in 100
1 in a million 7. 1 in 10
1 in 100,000 8. 1 in 5
1 in 10,000 9. 1 in 2

1 in 1,000 10. Certain to happen

PM“???

137

5. How certain are you that your guess about the chance of a health
effect is correct? (Circle one number)

Very uncertain
Somewhat uncertain
Somewhat certain
Very certain

I have no idea

9'93“?!"

The public health advisory from the 1992 Fishing Guide is included with
this questionnaire. Questions that mention the ”advisory” refer to this
insert.

6. Has the advisory helped you to avoid health problems from chemical
residues in fish? (Circle one number)

1. Yes
2. No
3. Idon't know

7. As a result of the advice in the advisory, do you... (Circle all that
apply)

I have not read the advisory
Eat fish less often

Fish at different places

Eat smaller fish

Eat different kinds of fish
Prepare fish to eat differently
Do nothing differently

HP‘P'PP’P'."

8. Are you concerned about chemical residues or other contaminants in
other foods that you eat? (Circle one number)

Not at all concerned
Somewhat unconcerned
Somewhat concerned
Very concerned

PPS"?

138

Michigan's Public Health Advisory

There are more than 5,800 public fishing sites in Michigan - 2,200 sites on
rivers and streams, 3,600 inland lakes, and the Great Lakes. The state has
tested 350 of these sites for chemical residues in fish. About 30 new sites

are tested each year.
The current public health advisory tells you:

0 that you should not eat too much fish from any inland
lake because of widespread mercury contamination, and

0 it lists 50 sites where fish contain chemical residues
above state limits.

The advisory does not tell you about:

0 the 300 tested sites where chemical residues do not exist
or are below state limits.

The advisory program could be changed.

0 The advisory could list tested sites where chemical
residues do not exist or are below state limits.

0 More than 30 new sites could be tested each year.

These changes would increase the amount of information in the advisory

but they would also cost more money.

139

Your Vote on Advisory Programs

The table below shows two advisory programs. The "Current Advisory"
"Program A" is a different

is Michigan's current advisory program.
program that could be put in place.

 

 

 

 

 

Current Program
Program Options Advisory A

Lists tested sites where chemical residues

are above state limits? Yes Yes
Lists tested sites where chemical residues

do not exist or are below state limits? No Yes
Number of new sites tested each year. 30 110
Cost to you in higher fishing license fees. $0.00 $.40

 

 

 

 

 

Suppose the Michigan Department of Natural Resources (DNR) sent you a
ballot to vote "for" or "against" Program A. If a majority of anglers vote
"for" Program A, it will replace the Current Advisory. If a majority vote
"against" Program A, the Current Advisory will be continued.

9. Would you vote for Program A if it permanently increased your
yearly license cost by $.40, or vote against it and keep the Current
Advisory at no additional license cost?

1. Vote for Program A
2. Vote against Program A and keep Current Advisory
3. Don't know or no opinion

140

10. Do you agree or disagree with
the following statements?

 

Circle best response

It is OK to increase fishing

. Strongly Somewhat Somewhat Strongly No
license fees 10 pay for bCIIBI' Agree Agree Disagree Disagree Opinion
public health advisories.

The health risks from chemical

. . . Strongly Somewhat Somewhat Strongly No
I'CSldUCS In fISh are W611 Agree Agree Disagree Disagree Opinion
understood by scientists.

If chemical residues in fish

. . Strongly Somewhat Somewhat Strongly No
made SOIITCOIIB Sle, the liiIICSS Agree Agree Disagree Disagree Opinion
would probably be fatal.

The advisory understates the

. . Strongly Somewhat Somewhat Strongly No
health I'lSkS from chemical Agree Agree Disagree Disagree Opinion
residues in Michigan's fish.

11. In addition to the list of unsafe sites, suppose the advisory listed all
tested sites where chemical residues in fish did. not exist or were
below state limits. If your favorite sites had not been tested would
you... (Circle all that apply)

Continue to fish at your favorite sites

Stop eating fish from your favorite sites
Fish only at sites where chemical residues
are below state limits

. When choosing a new site, be more likely to
go to a site where chemical residues were
below state limits

‘3 SANH

12. What would you do if next year's advisory said you should not eat
any of your favorite species of fish from your favorite site? (Circle
all that apply)

1. I would still fish at the site

2. I would eat fewer fish

3 I would not eat fish from the site
4. I would fish at a different site

5 I would stop fishing

141

13. Which of the following groups do you trust to provide the best
information about contaminants in Michigan's sport fish? (Circle one

number)

P‘P'PP‘PE‘

Federal government

State government

A well known consumer's group
An environmental group

A university laboratory
Other

 

14. What is the highest grade of school you have finished? (Circle one

number)

23°99‘95“???

Grade school only

Did not finish high school

High school or GED

Vocational or technical school

Some college

College graduate (BS or BA)

Some college or professional school
Graduate degree (Phd, MD, MA, MBA)

15. Which choice below best describes your household's expected before
tax income from all sources for 1993? (Circle one number)

HP‘P'PPPI‘

$0 to $9,999 8. $70,000 to $79,999
$10,000 to $19,999 9. $80,000 to $89,999
$20,000 to $29,999 10. $90,000 to $99,999
$30,000 to $39,999 11. $100,000 to $109,999
$40,000 to $49,999 12. $110,000 to $149,999
$50,000 to $59,999 13. $150,000 to $199,999
$60,000 to $69,999 14. $200,000 and above

142

If you have any comments about this questionnaire please write them
on this page.

When you are finished with the questionnaire, please fold it in half,
place it in the enclosed business reply envelope, and return to:

Douglas Krieger

Project Director

Department of Agricultural Economics
Michigan State University

East Lansing, MI 48824-1039

Thank you very much for your help.

Chlorine bleach used in the paper industry contributes to dioxins in
Michigan waters. This questionnaire is printed on recycled paper made
from 100% post consumer stock and processed without chlorine bleach.

143

Survey Cover Letter

February 9, 1993

John Angler
123 Whitefish Dr.
Grayling, MI 99999

Dear Mr. Angler,

I am writing to you because Cyou are a licensed angler in Michi an. You may
have heard that chemical resi ues have been found in some of ichigan's sport
fish. The state is concerned that anglers are aware of the problem and know
how to rotect themselves. To inform anglers, the Michigan Department of
Public ealth issues an annual "Public Health Advisory" that is printed in the
back of the "Fishing Guide."

Testin fish and printing advisories is expensive. Because the state budget is
limite , choices must be made about how many sites to test and the amount of
information to publish. Knowing what kinds of information anglers find useful
will assist the state in makin these choices. This survey asks about your fishing
choices and how you use dif erent kinds of information about chemical residues.

I am interested in the opinions of different kinds of an lers. Your response to
this survey is important even if you do not fish very 0 en. It is important that
the questionnaire is filled out by the person it is addressed to even if another per-
son in your household fishes more often. The questionnaire should take a out
10—15 minutes to complete.

Your participation in the survey is completely. voluntary. You indicate your
voiunta agreement to particrpate by com leting1 andreturning the question-
naire. our responses Will be kept complete y con dential.

Thank you for your help with this research.

Sincerely,

Douglas J. Krieger
Project Director

144

Reminder Postcard

 

 

February 16, 1993

Last week you should have received a questionnaire asking
for your opinions about Michigan's public health advisories
concerning sport fish contamination.

If you have already filled out the questionnaire and returned
it I thank you very much for your help. If not, please do so
today. It is important that we receive you response if the
results are to accurately represent the opinions of Michigan
anglers. Your response will be greatly appreciated.

Sincerely,

Douglas J. Krieger
Project Director

 

 

145

Cover Letter for First F ollow-up

March 2, 1993

John Angler
123 Whitefish Dr.
Grayling, MI 99999

Dear Mr. Angler,

About three weeks ago I sent you a questionnaire asking for your Opinions about the
Public Health Advisories regarding chemical residues in Michigan's sport fish. To date
I have not received your completed questionnaire.

The results of this research will help the state design advisories that are more useful to
anglers. This survey is a chance for you to let the state know whether or not you find
the advisory information useful.

I am writing to you again to remind you to complete and return the questionnaire. I
sent the questionnaire to a small sample of licensed anglers. To be accurate, it is essen-
tial that each person in the sample return their questionnaire. The questionnaire should
be filled out be the person it is addressed to even if another person in your household
fishes more often.

Your participation in the survey is completely voluntary. You indicate your voluntary
agreement to participate by completing and returning the questionnaire. Your
responses will be kept completely confidential.

In case your questionnaire has been lost or misplaced, a replacement is included with
this letter.

Thank you for your cooperation.

Sincerely,

Douglas J. Krieger
Project Director

146

Cover Letter for Second Follow-up

March 30, 1993

John Angler .
123 Whitefish Dr.
Grayling, MI 99999

Dear Mr. Angler,

I am writing to you about my study of Michigan's public health advisories regarding
chemical residues in sport fish. I have not yet received a reply to the questionnaire I
sent you.

Your response is very important. Past experience suggests that the opinions of those
who do not return questionnaires are different from the opinions of people who do.
For the study to be accurate, it is essential that I hear from you.

The results of the survey will be used to make choices about the future of the advisory
program. State agencies are interested in insuring that the advisories meet the needs of
anglers like yourself. The more accurately the survey represents the Opinions of
Michigan's anglers, the more useful it will be in making decisions consistent with
angler's preferences.

I am sending this copy of the questionnaire by certified mail to insure delivery. In case
my previous surveys did not reach you, I have enclosed a replacement c0py of the
questionnaire. I would ask that you complete and return it as soon as possible.

Your participation in the survey is completely voluntary. You indicate your voluntary
agreement to participate by completing and returning the questionnaire. Your
responses will be kept completely confidential.

Your contribution to the success of this study is greatly appreciated.

Sincerely,

Douglas J. Krieger
Project Director

147

Table A.1 - Factorial Design for Survey Variables

 

 

Number of Full Number of Full
Survey Sites Tested Disclosure? Bid Survey Sites Tested Disclosure? Bid
Version (sites) (safesites) (bid) Version (sites) (safesites) (bid)
1 110 Yes .40 31 110 No .40
2 110 Yes .95 32 110 No .95
3 110 Yes 1.45 33 110 NO 1.45
4 110 Yes 1.90 34 110 No 1.90
5 110 Yes 2.85 35 110 No 2.85
6 110 Yes 4.10 36 110 No 4.10
7 110 Yes 5.55 37 110 No 5.55
8 110 Yes 8.75 38 110 No 8.75
9 110 Yes 14.50 39 110 NO 14.50
10 110 Yes 41.00 40 110 No 41.00
1 1 620 Yes .40 41 620 No .40
12 620 Yes .95 42 620 No .95
13 620 Yes 1.45 43 620 No 1.45
14 620 Yes 1.90 44 620 No 1.90
15 620 Yes 2.85 45 620 No 2.85
16 620 Yes 4.10 46 620 No 4.10
17 620 Yes 5.55 47 620 NO 5.55
18 620 Yes 8.75 48 620 No 8 .75
19 620 Yes 14.50 49 620 NO 14.50
20 620 Yes 41.00 50 620 No 41 .00
21 1240 Yes .40 51 1240 NO .40
22 1240 Yes .95 52 1240 No .95
23 1240 Yes 1.45 53 1240 No 1.45
24 1240 Yes 1 .90 54 1240 NO 1 .90
25 1240 Yes 2.85 55 1240 No 2.85
26 1240 Yes 4.10 56 1240 ' No 4.10
27 1240 Yes 5.55 57 1240 No 5.55
28 1240 Yes 8.75 58 1240 No 8.75
29 1240 Yes 14.50 59 1240 No 14.50
30 1240 Yes 41.00 60 1240 No 41.00

 

 

Appendix B

Specification Tests and Alternative Models

This appendix outlines specification tests for the empirical WTP model.
Specification Tests

Table B.l lists regression results for three models - an unrestricted model, a model con-
taining non-program variables only in level form, and a model that excludes both interaction
and level forms of the non-program variables. These models facilitate likelihood ratio tests of

two hypotheses.

1. Ho -- Non-program variables do not affect marginal WTP for advisory

information.

2. Ho -- Level forms of the non-program variables have no significant effect

on estimated WTP.

The tests fail to reject the first hypothesis but reject the second. Therefore, the final
model includes non-program variables in level form but not in interaction with program vari-

ables.

148

149

Marginal Effect of Non-Program Variables

The values of the log likelihood function for the unrestricted model and the model without
interaction terms provide a test of hypothesis one. The likelihood ratio statistic for the null
hypothesis that non-program interaction terms can be jointly restricted to zero is 2.16. The X2
critical value for 6 degrees of freedom and a=.01 is 16.81. Therefore, the likelihood ratio
test fails to reject the restrictions. Thus, the final model does not include interaction terms of

the non-program variables.
Test for Exclusion of Non-Program Level Forms

In addition to the restrictions of the final model, the model without non-program variables
also omits the level forms of non-program variables. The likelihood ratio statistic for this
hypothesis (73.94) is well above the )8 critical value of 16.81 for 6 degrees of freedom and
a=.01. Therefore, the likelihood ratio test rejects the hypothesis and the model that retains

level forms of the non-program variables is preferred.
Signifiwnce of Behavioral Change Variables

Table 8.2 lists regression results for a model that, compared with the final model of Table
B.1, jointly restricts the coefficients of the behavioral change variables to equal zero. The
likelihood ratio statistic for the null hypothesis that the behavioral change variables are not
jointly significant is 37.82. The )8 critical value for 4 degrees of freedom and a=.01 is
13.28. Therefore, the test rejects the hypothesis —- the behavioral change variables have a sig-

nificant joint effect on estimated WTP.

 

150

Table 8.1 - Model Specification

 

 

. Without
. Unrestricted Final Non-Pro ram
Var1able Model Model Variab es
CONSTANT .1668 .3457 -.1987
(2.249) (2.202) (2.414)
FSITES 0046* 0045* .0042
(.0028) (.0027) (.0028)
PSITES .0047* .0045 .0045
(.0029) (.0028) (.0031)
FULLD 5844* 5643* 6. 822**
(3.221) (3.151) (2.032)
FATAL 7.521“ 7.245***
(3.598) (2.120)
BCHANGE 2. 568 1.702
(3.183) (1.847)
NOREAD -4.216 -2. 658
(4.625) (2. 777)
NCHANGE -3.731 -.3652
(4.351) (2.645)
USELIST 16.430*** 15.612***
(5.261) (2.979)
SCIENCE .3668 2.688
(3.249) (1.901) _
COLLEGE 5553* 6.836***
(3 .267) (1 .866)
FATAL x (fisites + psites) -.0006
(.0044)
BCHANGE x ofsites + psites) -.0012
(.0040)
NOREAD x (fsites + psites) .0028
(.0060)
USELIST): ites + sites —.0009
0% p ) (.0063)
SCIENCE x mites? psites) .0038
(.0043)
COLLEGE x (fsites + psites) .0020
(0040)
Log-Likelihood -428.67 -429.75 466.72
% Correct Predictions 72.2% 71.9% 66.9%

 

*
**
***

= significant for a = 1
= significant for a =

205
significant for or = .01

Regression results based on 758 observations for which complete data exists for all variables.

Numbers in parentheses are standard errors.

151

Table B.2 - Exclusion of Behavioral Change Variables

 

 

Coefficient
Variable Estimates
CONSTANT .1473
(2.260)
FSITES .0042
(.0027)
PSITES .0047
(.0029)
FULLD 6010*
(3.181)
FATAL 9.158***
(2.183)
SCIENCE 3.070
(1.931)
COLLEGE 7.981***
(1.882)
Log-Likelihood -448.66

% Correct Predictions 68.6%

 

*
**
***

significant for a = .1
significant for a = .05
significant for a = .01

Regression results based on 758 observa-
tions.

All non-program variables enter as
deviations from mean values.

Numbers in parentheses are standard
errors.

152

Table 8.3 - Probit Model Estimates

 

 

Linear Quadratic Square Root
Variable Model Model Model
CONSTANT .1916E'l
(.1240)
FSITES .2523E'3* .1601E'3*** -.5725E‘3
(.1507E'3) (.4190E'3) (.4108E'3)
FSITES2 -.9202E‘°***
(.3512E'6)
FSITES-5 .3742E‘1***
(.1344E")
PSITES .2477E'3
(.1580E'3)
COST -.5543E"*** -.5341E'1*** -.5539E“***
(.6168E'2) (.7784E‘2) (.8041E'2)
FULLD .3128*
.1723
FATAL .4016*** .4614*** . ***
(.1194) (.1620) (.1622)
BCHANGE .9435E'l .5811E’l .4915E'l
(.1025) .1487 (.1485)
NOREAD -.1473 -.1714 -.1897
(.1507) (.2108) (.2106)
NCHANGE -.2024E‘l
(.1455)
USELIST .8654*** .8733*** .8723***
(.1518) (.1564) (.1559)
SCIENCE .1490 .1371 .1315
(.1047) (. 1495) (.1497)
COLLEGE .3789*** .427S*** .4370*** ‘
(.1017) (.1424) (.1426)
Log-Likelihood -429.75 -214.40 -213.96
% Correct Predictions 71.9% 74.0% 74.3%

 

*
**
***

significant for a = .1
significant for a = .05
Significant for a = .01

Regression results based on 758 observations for which complete data exists for all variables.

Numbers in parentheses are standard errors.

Bibliography

Bibliography

Anderson, Jock R., John L. Dillon, and J. Brian Hardaker. Agricultural Decision Analysis.
Ames, Iowa: Iowa State University Press, 1977.

Baquet, A. E., A. N. Halter, and Frank S. Conklin. "The Value of Frost Forecasting: A
Bayesian Appraisal." American Journal of Agricultural Economics 58 (August 1976):511-
20.

Blackwell, David. "Equivalent Comparison of Experiments." Annals of Mathematical Statistics
24 (June 1953):265-72.

Blackwell, David, and Meyer A. Girshick. Theory of Games and Statistical Decisions. New
York, NY: John Wiley and Sons, 1954.

Boyle, Kevin J., Michael P. Welsh, and Richard C. Bishop. "Validation of Empirical
Measures of Welfare Change: Comment." [and Economics 64 (February l988):94-98.

Bro, Kenneth M., William C. Sonzogni, and Mark E. Hanson. "Relative Cancer Risks of
Chemical Contaminants in the Great Lakes.” Environmental Management 11 (August
1987):495-505.

Brown, Joseph D. ”Effect of a Health Hazard 'scare' on Consumer Demand." American
Journal of Agricultural Economics 51 (August 1969):676-78.

Byerlee, D. R., and Jock R. Anderson. "Value of Predictors of Uncontrolled Factors in
Response Functions.” Australian Journal of Agricultural Economics 13 (December
1969):118-27.

Cameron, Trudy Ann. "A New Paradigm for Valuing Non-market Goods Using Referendum
Data: Maximum Likelihood Estimation by Censored Logistic Regression." Journal of
Environmental Economics and Management 15 (1988):355-79.

Cameron, Trudy Ann, and Michelle D. James. "Efficient Estimation Methods for 'closed-
ended' Contingent Valuation Surveys." Review of Economics and Statistics 69 (May
l987):269-76.

Caudill, James D. Yhe Valuation of Groundwater Pollution Policies: The Difi’erential Impacts
of Prevention and Remediation. Doctoral Dissertation, Department of Agricultural Eco-
nomics, Michigan State University, East Lansing, MI. 1992.

Chavas, Jean-Paul, and Rulon D. Pope. "Information: It's Measurement and Valuation."
American Journal of Agricultural Economics 66 (December l984):705-10.

Connor, Michael Stewart. "Comparison of the Carcinogenic Risks from Fish Vs. Groundwater
Contamination by Organic Compounds." Environmental Science and Technology 18
(August 1984):628-31.

154

Cooper, Joseph C. "Optimal Bid Selection for Dichotomous Choice Contingent Valuation Sur-
veys." Journal of Environmental Economics and Management 24 (January 1993):25-40.

Deaton, Angus, and John Muellbauer. Economics and Consumer Behavior. Cambridge: Cam-
bridge University Press, 1980.

Desvousges, William H., and V. Kerry Smith. "Focus Groups and Risk Communication: The
'science' of Listening to Data." Risk Analysis 8 (December l988):479-84.

Dillman, Don A. Mail and Telephone Surveys: The Tatal Design Method. New York: John
Wiley & Sons, Inc., 1978.

Duffield, John W., and David A. Patterson. "Inference and Optimal Design for a Welfare
Measure in Dichotomous Choice Contingent Valuation." Land Economics 67 (May
l99l):225-39.

Eccles, Christine U., and Zoltan Annau. "Prenatal Exposure to Methyl Mercury." In The
Toxicity of Methyl Mercury, Eds Christine U. Eccles and Zoltan Annau, 114-30.
Baltimore, Md.: Johns Hopkins University Press, 1987.

Fischhoff, Baruch, and Lita Furby. "Measuring Values: A Conceptual Framework for Inter-
preting Transactions with Special Reference to Contingent Valuation of Visibility."
Journal of Risk and Uncertainty 1 (June 1988):147-84.

Fischhoff, Baruch, Paul Slovic, Sarah Lichtenstein, Stephen Read, and Barbara Combs. "How
Safe is Safe Enough? a Psychometric Study of Attitudes Towards Technological Risks and
Benefits." Policy Sciences 9 (April 1978): 127-52.

Foran, Jeffery A., and David Vanderploeg. "Consumption Advisories for Sport Fish in the
Great Lakes Basin: Jurisdictional Inconsistencies." Journal of Great Lakes Research 15,
no. 3 (1989):476-85.

Foster, William, and Richard E. Just. "Measuring Welfare Effects of Product Contamination

with Consumer Uncertainty." Journal of Environmental Economics and Management 17
(November 1989):266-83.

Freeman, A. Myrick ID. The Measurement of Environmental and Resource Values: Theory and
Methods. Washington, DC: Resources for the Future, 1993.

Freund, John F. Mathematical Statistics. Englewood Cliffs, N.J.: Prentice-Hall, Inc., 1962.

Gaffey, William R. "The Epidemiology of PCBs." In PCB '3: Human and Environmental Haz-
ards, Eds Frank M. D'Itri and Michael A. Kamrin, 279-98. Boston, MA: Butterworth
Publishers, 1983.

Gould, John P. "Risk, Stochastic Preference, and the Value of Information." Journal of Eco-
nomic Theory 8 (May 1974):64—84.

Graham-Tomasi, Theodore. "The Economic Value of Damages to the Pigeon River. Michigan,
Due to an Abnormal Discharge of Sediment from the Lansing Club Dam." Unpublished
Paper. Michigan Department of Natural Resources, 1986.

------ . "A Theoretical and Empirical Approach to the Value of Information in Risky Markets: A
Comment." Review of Economics and Statistics 70 (August 1988):543-45.

155

Greene, William H. Econometric Analysis. 2d ed. New York. NY: Macmillan Publishing
Company, 1990.

------ . LIMDEP, Version 6.0: User's Manual and Reference Guide. Bellport, N.Y.:
Econometric Software, 1991.

Hamilton, James L. "The Demand for Cigarettes: Advertising, the Health Scare, and the

Cigarette Advertising Ban." Review of Economics and Statistics 54 (November 1972):401-
11.

Hanemann, W. Michael. "Some Issues in Continuous- and Discrete-response Contingent
Valuation Studies." Northeastern Journal of Agriclutural and Resource Economics 14
(April 1985):5-13.

Hess", James. "Risk and the Gain from Information." Journal of Economic Theory 27 (June
1982):23l-38.

Hesse, John L. "Summary and Analysis of Existing Sportfish Consumption Advisory Programs
in the Great Lakes Basin." Unpublished Paper. Michigan Department of Public Health,
1990.

Hilton, Ronald W. "The Determinants of Information Value: Synthesizing Some General
Results." Management Science 27 (January 1981):57-64.

Hirshleifer, J., and John G. Riley. "The Analytics of Uncertainty and Information - an
Expository Survey." Journal of Economic Literature 17 (December 1979): 1375-421.

----. The Analytics of Uncertainty and Information. Cambridge Surveys of Economic Litera-
ture. Cambridge, England: Cambridge University Press, 1992.

Hoehn, John P., and Alan Randall. "A Satisfactory Benefit Cost Indicator from Contingent
Valuation." Journal of Environmental Economics and Management 14 (June l987):226-
47.

Humphrey, Harold E. B. "Mercury Concentrations in Humans and Consumption of Fish Con-
taining Methylmercury." Mercury Project - Update Report, Michigan Department of Pub-
lic Health, Lansing, MI. 1974.

—----. "Population Studies of PCBs in Michigan Residents." In PCB 's: Human and Environ—
mental Hazards, Eds Frank M. D'Itri and Michael A. Kamrin, 299-310. Boston, MA:
Butterworth Publishers, 1983.

Jacobson, Joseph L., Sandra W. Jacobson, Greta G. Fein, Pamela M. Schwartz, and J. K.
Dowler. "Prenatal Exposure to an Environmental Toxin: A Test of the Multiple Effects
Model." Developmental Phsychology 20 (1984):523-32.

Johnson, F. Reed. "Economic Costs of Misinforming About Risk: The EDB Scare and the
Media." Risk Analysis 8 (December 1988):261-69.

Jones, Carol Adaire, and Yusen D. Sung. "Valuation of Environmental Quality at Michigan
Recreational Fishing Sites: Methodological Issues and Policy Applications." Unpublished
Paper. School of Natural Resources and Economics Department, University of Michigan,
Ann Arbor, Mich. 1991.

156

Kanninen, Barbara J. "Design of Sequential Experiments for Contingent Valuation Studies."
Journal of Environmental Economics and Management 25 (July 1993):S 1-1 1.

Kihlstrom, Richard E. "A 'Bayesian' Exposition of Blackwell's Theorem on the Comparison of
Experiments." In Bayesian Models in Economic Theory, Eds Marcel Boyer and Richard
E. Kihlstrom, 13—32. Studies in Bayesian Econometrics, vol. 5. Amsterdam, The Nether-
lands: Elsevier Science Publishing Company, 1984.

mm. "A Bayesian Model of Demand for Information About Product Quality." International
Economic Review 15 (February 1974):99-118.

Kikuchi, Hideo. Segmenting Michigan '3 Sport Fishing Market: Evaluation of Two Approaches.
Doctoral Dissertation, Department of Parks and Recreation Resources, Michigan State
University, East Lansing, Michigan. 1986.

Knuth, Barbara A., and Christine M. Velicer. "Receiver-centered Risk Communication for
Sportfisheries: Lessons from New York Licensed Anglers." Paper Presented at the
American Fisheries Society Annual Meeting, Pittsburgh. PA. 1990.

Krieger, Douglas J., and John P. Hoehn. "Angler Perceptions of Fish Contamination Risks and
Consumption Advisories in Michigan: An Experience with Focus Group Research."
Department of Agricultural Economics Staff Paper #91-60. Michigan State University,
East Lansing, Mich. 1991.

Kristr'dm, Bengt. "A Non-parametric Approach to the Estimation of Welfare Measures in Dis-
crete Response Valuation Studies." [and Economics 66 (May 1990): 135-39.

Kuratsune, Masanori. "Yusho, with Reference to Yu-cheng." In Halogenated Biphenyls, Ter-
phenyls, Naphthalenes, Dibenzodioxins and Related Products, Eds R. D. Kimbrough and
A. A. Jensen, 381-400. Amsterdam, The Netherlands: Elsevier Science Publishers B.V.,
1989.

McConnell, Kenneth E. "Models for Referendum Data: The Structure of Discrete Choice
Models for Contingent Valuation." Journal of Environmental Economics and Management
18 (March 1990):19-34.

McGuire, C. B. "Comparisons of Information Structures." In Decision and Organization, Eds
C. B. McGuire and Roy Radner, 101-30. Minneapolis: University of Minnesota Press,
1972.

Magat, Wesley A., W. Kip Viscusi, and Joel Huber. "Consumer Processing of Hazard Warn-
ing Information." Journal of Risk and Uncertainty 1 (June 1988):201-32.

Mahoney, Edward M., Douglas B. Jester, and Daniel Stynes. "A Study of Michigan's Sport
Fishing Market." Report to the Fisheries Division, Michigan Department of Natural
Resources. Department of Parks and Recreation Resources, Michigan State University,
East Lansing, MI, 1986.

Marschak, Jacob, and Koichi Miyasawa. "Economic Comparability of Information Systems."
International Economic Review 9 (June 1968): 137-74.

Marschak, Jacob. and Roy Radner. Economic Theory of Teams. New Haven: Yale Univeristy
Press, 1972.

157

Marsh, David O. "Dose-response Relationships in Humans: Methyl Mercury Epidemics in
Japan and Iraq." In The Toxicity of Methyl Mercury, Eds Christine U. Eccles and Zoltan
Annau, 45-53. Baltimore, Md.: Johns Hopkins University Press, 1987.

Michigan Department of Natural Resources. [980—1990 Michigan Fisheries: A Foundation for
the Future. Fisheries Special Report No. 13. Edited by W.C. Latta. Lansing, Mich. 1990.

Mishan, Edward J. Cost-benefit Analysis. New York: Praeger, 1976.

Mitchell, Robert Cameron, and Richard T. Carson. Using Surveys to Value Public Goods: The
Contingent Valuation Method. Washington, DC: Resources for the Future, 1989.

Mowen, John C. "Further Information on Consumer Perceptions of Product Recalls." In
Advances in Consumer Research, edited by Jerry C. Olson, 519-23. Association for Con-
sumer Research. Ann Arbor, MI, 1980.

Nermuth, Manfred. Information Structure in Economics. Lecture Notes in Economics and
Mathematical Systems. Heidelberg, Germany: Springer-Verlag, 1982.

Ontario Ministries of the Environment and Natural Resources. "Guide to Eating Ontario Sport
Fish." Toronto, Ontario, 1990.

Paradis, Gilles, Chris Archibald, and Tom Kosatsky. "Risk of Disease Associated with Con-
sumption of Sport Fish Caught in the Montreal Region." 1988.

Pope, Rulon D. "The Impact of Information on Consumer Preferences." In Proceedings of
Research on Effectiveness of Agricultural Commodity Promotion Seminar. 69-79.
Washington, DC: USDA, Economic Research Service, 1985.

Rescher, Nicholas. Risk: A Philosophical Introduction to the Theory of Risk Evaluation and
Management. Washington, DC: University Press of America, 1983.

Rodabaugh, Gary L. An Investigation Into Processes Contributing to Voluntary Exposure of
Michigan Anglers to Contaminated Waterways and Contaminated Fish. Doctoral Disserta-
tion, Department of Fisheries, Michigan State University, East Lansing, Michigan. 1987.

Rothschild, M., and J. Stiglitz. "Increasing Risk: 1. A Deinition." Journal of Economic Theory
2 (December 1970):225-43.

Schkrade, David A., and John W. Payne. "How People Respond to Contingent Valuation
Questions: A Verbal Protocol Analysis of Willingness to Pay for an Environmental

Regulation." Journal of Environmental Economics and Management 26 (January
1994):88-109.

Schuker, Raymond E. "The Impact of Saccharin Warning Labels on Sales of Diet Soft Drinks
in Supermarkets." Journal of Public Policy and Marketing 2 (1983):46-56.

Schwartz,-Pamela M., Sandra W. Jacobson, Greta Fein, Joseph L. Jacobson, and Harold A.
Price. "Lake Michigan Fish Consumption as a Source of Polychlorinated Biphenyls in
Human Cord Serum, Maternal Serum. and Milk." American Journal of Public Health 73
(March 1983):293-96.

Shulstad. Robert N.. and Herbert H. Stoevener. "The Effects of Mercury Contamination in

Pheasants on the Value of Pheasant Hunting in Oregon." Land Economics 54 (February
1978):39-49.

158

Smith, Mark E., Eileen 0. van Ravenswaay, and Stanley R. Thompson. "Sales Loss
Determination in Food Contamination Incidents: An Application to Milk Bans in Hawaii."
American Journal of Agricultural Economics 70 (August 1988):513-20.

Smith, V. Kerry, and F. Reed Johnson. "How Do Risk Perceptions Respond to Information?
the Case of Radon." The Review of Economics and Statistics 70 (February 1988):1-8.

Swartz, David G., and Ivar E. Strand, Jr. "Avoidance Costs Associated with Imperfect
Information: The Case of Kepone." Land Economics 57 (May 1981): 140-50.

Tice, Thomas F., and Rodney L. Clouser. "Determination of the Value of Weather Informa-

tion to Individual Corn Producers." Journal of Applied Meteorology 21 (April 1982):447-
52.

van Ravenswaay, Eileen 0., and John P. Hoehn. Consumer Willingness to Pay for Reducing
Pesticide Residues in Food: Results of a Nationwide Survey. Department of Agricultural
Economics Staff Paper #91-18. Michigan State University, East Lansing, MI, 1991.

------ . "The Impact of Health Risk on Food Demand: A Case Study of Alar and Apples."
Department of Agricultural Economics Staff Paper, # 90-31. Michigan State University,
East Lansing, Mich. 1990.

van Ravenswaay, Eileen, Jennifer B. Wohl, and John P. Hoehn. Michigan Consumers' Percep-
tions of Pesticide Residues in Food. Department of Agricultural Economics Staff Paper,
No. 9256. Michigan State University, East Lansing, MI, 1992.

Vaughan, William J., and Clifford S. Russell. "Freshwater Recreational Fishing: The National
Benefits of Water Pollution Control." Research Paper. Resources for the Future, Wash.,
DC. 1982.

Viscusi, W. Kip, and Charles J. O'Connor. "Adaptive Responses to Chemical Labeling: Are
Workers Bayesian Decision Makers?" American Economic Review 74 (December
1984):942-56.

von Neumann, J., and O. Morgenstern. Theory of Games and Economic Behavior. Princeton,
NJ: Princeton University Press, 1944.

Winkler, Robert L. Introduction to Bayesian Inference and Decision. 563. Series in Quantita-
tive Methods for Decision Making. New York: Holt, Rinehart and Winston, Inc., 1972.

Zellner, Arnold. An Introduction to Bayesian Inference in Econometrics. New York: John
Wiley and Sons, Inc., 1971.

"Illlllllllllllllllllllls