Lareau and R ae, 1989, preferences for siting municipal landfills Opaluch et al., 1993, prefer-
ences for various energy programs Johnson and D esvousges, 1997, preferences for
waterfowl hunting M acK enzie, 1993, wild salmon manage-
ment options R oe et al., 1996, preferences for recreational activities G an, 1992; Adamowicz et
al., 1994, and values for protecting threatened woodland caribou populations Adamowicz et al.,
1998.
The purpose of CJ is to analyze choice in a multi-attribute context. Individuals are presented
choice alternatives with varying values of at- tributes, and are asked to choose the best or
rankrate the
alternatives. This
hypothetical choice setting appears to mimic real choice set-
tings in requiring the individual to simultaneously consider many dimensions of alternatives. The
potential of CJ procedures for contingent valua- tions of ecosystems is direct. F irst, allowing one
of the attributes to be a ‘price’ can be used to reveal implied valuations of individual attributes.
Second, the multidimensional nature of the alter- natives allows a more realistic representation of
complex ecosystems where a variety of dimensions may be important to individuals. Third, the
choice is referendum-like and thereby emulates public decision situations.
CJ is defined as ‘‘any decompositional method that estimates the structure of consumer’s prefer-
ences…given his or her overall evaluations of a set of alternatives that are prespecified in terms of
levels of different attributes. Price typically is included as an attribute’’ G reen and Srinivasan,
1990. As a general illustration of the CJ proce- dures, suppose we are interested in valuing a
wetlands system and its component services. Sup- pose wetlands services include storm protection
S , recreational fishing F, and water treatment W . Let the utility level of the individual depend
upon the level of these services as well as some price placed on the bundle of service, P. Then
U = f S , F, W , P . In particular, suppose that U = aS + bF + cW + dP, a general linear utility
function. CJ provides estimates of the ‘part- worths,’ a, b, c and d. In general, knowing these
part-worths permits the valuation of any wetlands system since cd, for example, measures the mar-
ginal value placed upon water quality treatment. The CJ procedure establishes several levels of
the attributes. F or example, S may be measured by the probability of a storm impacting the indi-
vidual, and may have several values, say, 0.10, 0.20 and 0.30. R ecreational fishing service, F, may
be measured by average fish catch per day, say, 0, 3, 6 and 10. Water treatment services may be
simply measured by whether toxics are in the water, Yes or N o. Several prices may be used, say,
10, 50, and 100 per household per year. CJ then presents the respondent with combinations
of the four attributes, and asks the respondent to select the best, to rank them or to provide a rating
of them. In this example, there would be 3 × 4 × 2 × 3 = 72 possible combinations of attribute lev-
els. In order to simplify the choice, only a subset of these combinations would be used. Various
software programs e.g. SPSS Conjoint Ortho- plan facilitate the selection of a reasonable num-
ber of choice combinations. They typically select combinations that permit the estimation of only
main effects. This is termed orthogonal design, and would not allow determining the interaction
effects of, say, levels of W on P.
This paper reports on a study of watershed quality improvement for two streams in Western
Pennsylvania, both degraded from acid mine drainage. CJ allows joint valuation of these poten-
tially substitute goods. The study design includes distance from the two sites as a value-determining
variable and allows determination of the spatial extent of the market for users and non-users.
Several choice models were used to determine the sensitivity of valuation results to model specifica-
tion. Section 2 outlines the experimental design, including the CJ design as well as survey ques-
tions and procedures. Section 3 derives the R U M - based estimation models, and Section 4 derives
specific estimating equations. Section 5 presents the results of model estimation, including distinc-
tions between user and non-user valuations. Sec- tion 6 is the conclusion.
2. Experimental design
The study areas are two sub-basins of the Lower Allegheny Watershed in Western Pennsyl-
vania. This 2400 sq. mile watershed has a total of 3000 miles of stocked and natural freshwater
streams providing opportunities for warm and cold water fishing. Also, some reaches permit
boating and hiking. R oughly 15 of these stream miles are degraded, primarily from acid mine
drainage PA D EP, 1993. The sub-basins in which the two streams of interest are located,
Loyalhanna Creek subbasin 18C and Cone- maugh R iver subbasin D , differ significantly in
their extent of degradation. Loyalhanna Creek’s sub-basin has 60 of its stream miles degraded,
and the Conemaugh R iver’s sub-basin has only 3 of its miles degraded.
H abitat evaluation by U S EPA and PA state fisheries biologists distinguishes between stream
reaches that are Severely Polluted, M oderately Polluted and U npolluted. A Severely Polluted
reach has been determined by biologists to be incapable of supporting fish and other organisms.
F ishing conditions would be poor to non-existent. Wildlife could not rely on these streams for food
and habitat. H owever, human health is not typi- cally affected. M oderately Polluted reaches will
support only some fish and other organisms. R e- productive conditions for fish are poor, and
fishing is supported but catch would be limited. U npolluted reaches are those where fish and other
organisms can thrive. By these designations, Loy- alhanna Creek is currently M oderately Polluted
and the Conemaugh R iver is Severely Polluted.
A mail survey was administered during the summer of 1996 to a sample of area residents.
Individuals were presented with several choice scenarios and were asked to choose between the
status quo and various combinations of stream quality improvements for the two streams. Each
alternative had a price attached. A sample choice is shown below.
The survey used in this study is available from the authors. A brief letter presenting commonly
asked questions and answers about water quality in the study region provided an introduction to
the questionnaire. A map of the streams in the watershed, shaded according to the degree of
degradation followed the introductory letter. A major difficulty was defining the ‘good’ for the
choice alternatives. After discussion with fish and wildlife managers, we decided to use a mixed
habitat-fisheries characterization. The traditional water quality ladder Smith and D esvousges,
1986 was not appropriate for the types of water quality improvements evaluated in this study. The
resulting characterization of Severely Polluted, M oderately Polluted and U npolluted, based on
survivability of fish and other organisms, was described to survey recipients. The areas of stream
valuation were illustrated on the map provided. Instructions to recipients noted that stream clean-
up would be costly, resulting in higher prices or taxes. Persons were asked to consider a price they
would pay each year for the next 5 years.
The Loyalhanna Creek segment had two possi- ble quality conditions, M oderately Polluted and
U npolluted; the Conemaugh R iver segment had three possible quality conditions, Severely Pol-
luted, M oderately Polluted and U npolluted. The pilot and prior studies Smith and D esvousges,
1986 suggested the use of five payment levels 15, 45, 90, 180, 360. The policy issue in the
watershed was a determination of valuation for stream quality improvements, so degradation of
Loyalhanna Creek to Severely Polluted was not relevant. Consequently, there were a total of 25
possible alternatives to the status quo in the full choice set.
CJ allows several methods for presenting the choice alternatives to respondents. F irst, respon-
dents could compare the status quo to each alter- native in a binary choice, requiring 25 binary
comparisons. Second, respondents could be pre- sented with all 25 alternatives plus the status quo,
and be asked to rank or rate say, 1 to 10 the 26 options. Both of these methods would require
considerable effort, and the ranking or rating may be a difficult mental task. CJ has established
procedures for reducing the number of choice alternatives presented to an individual, and still
obtain the part-worths, or marginal valuations of choice attributes. A reasonable, orthogonal subset
of the 26 alternatives could be presented to the individual and the individual would either rank or
rate these. The orthogonal subset would be se- lected in order to determine the main effects of
each attribute SPSS, 1999. Such a design would not allow determining the dependence between
valuation of one stream improvement and the improvement in the other stream. H owever, we
were interested in both the main and interaction effects; we wished to determine how the value of
improving Stream A depended upon the improve- ment in Stream B. So we could not use an orthog-
onal design. Instead, we opted to use a full binary choice design. Each respondent was given five
choice sets, with prices varying across sets. There were five blocks of choice sets, with each block
administered to one-fifth of the sample. An exam- ple of a single choice set for an individual is
shown below:
The sampling design was to establish a 8 × 10 grid of cells, each 9 miles by 9 miles for a total of
6480 sq. miles, completely covering the Lower Allegheny Watershed and parts of adjacent water-
sheds. Zipcodes were selected for each cell, with one, two or three zipcodes used for cells with
more zipcodes included for cells closer to the streams of interest. Out of a total 460 possible
zipcodes within the grid sampling area, 133 were selected for sampling and a commercial mailing
database was used for selecting households. A pilot mail study of 200 persons resulted in a low
response rate of 10 to the initial questionnaire. F ollow up phone calls suggested that a problem
might have been complex instructions and choice, so the instruction
and choice sections were modified to make them simpler.
A final sample of 3958 households was drawn and first class postage surveys mailed to them.
Only one mailing was sent. A total of 510 house- holds responded, and 372 surveys were returned
as undeliverable. This was only a 14 response rate, excluding undeliverables, which is low. H ow-
ever, this rate is not far below that of other general population CJ samples Adamowicz et al.,
1994. Only 367 of the 510 returned surveys were usable as some respondents did not answer all
valuation related questions, particularly income.
We undertook a phone survey of non-respon- dents after an initial mailing of 600 in order to
examine reasons for non-response M itchell and Carson, 1989. A total of 380 non-respondents
were called. N early 20 had inoperable phone numbers. F orty percent stated that they had not
received the survey or had misplaced it so a second survey was sent to them. Twenty-five per-
cent refused to participate in the phone interview. F ifteen percent, 56 households, had received the
survey, had not returned it but did participate in the phone interview. Of these 56 households, 15
stated they were not interested in water quality issues or felt water quality was just fine. This does
suggest a survey response bias, implying that val- uations derived from the survey responses may be
overstated. Twelve of the 56 said they did not have time to fill out the survey. This response may
also suggest a bias. Other reasons for non-re- sponse included ‘It was too confusing,’ ‘N ot ap-
plicable to them,’ ‘D o not do surveys,’ or ‘Too old or sick to participate.’
The survey respondents’ characteristics differed slightly from the general population in the Lower
Allegheny Watershed. The respondents were 68 male compared to 48 in the watershed and
99 white compared to 96 in the watershed. The mean household size was 2.9 persons com-
pared to 2.6 persons in the watershed; mean age was 51 years compared to 52 years in the water-
shed; and median household income was 37 500 compared to 30 100 for the sampling area. This
suggests the sample is more male and higher income than the population. G eneralizing sample
results to the population must consider these so- cio-economic differences, as well as the potential
sampling non-response bias.
3. Econometric models
The basis for an econometric model is an indi- rect utility function implied by the choice of levels
of consumption of n goods, including ecological services offered by streams. Individual i maxi-
mizes the utility function:
U
i
= U x
i
, q
i
, c
i
; individual i = 1,…,T 1
subject to the budget constraint: I
i
− p
i
x
i
+ w
i
= 0 2
where x
i
= x
i
1
…x
il
…x
in
, vector of quantities of goods, l = 1,…,n; q
i
= q
i
1
…q
il
…q
in
, vector of qualities i.e. characteristics of goods, l = 1,…,n;
c
i
= c
i
1
…c
ik
…c
is
, vector of personal characteris-
tics, k = 1,…,s; I
i
, income; p
i
, vector of goods prices, including stream visit costs; w
i
, exogenous addition to income.
The indirect utility function is then: V
i
= V p
i
, q
i
, c
i
, I
i
+ w
i
3 R andom utility maximization R U M M cF ad-
den, 1981 provides an empirical formulation for estimation of Eq. 3. Let an individual face a set
of M discrete choices, j = 0,…M , where the status quo is j = 0 and j = 1,…M are alternatives. The
alternatives include various combination of the exogenous variables: goods quality, prices and
incomes. U nder R U M the indirect utility function is partitioned into systematic and random, unob-
servable components:
V
ij
= V
ij
+ o
ij
; i = 1,…,T ; j = 1,…,M 4
where V
ij
is the systematic or predictable portion of utility and o
ij
is the random portion of utility. The presence of o
ij
in the indirect utility function is due to the inability to observe V . While we can
observe elements entering into the systematic component, V
ij
, and predict, ex ante, which of the alternatives,
j = 0,…,M , the
individual will
choose, prediction will not be perfect due to the presence of o
ij
. The difference in utilities between alternatives j
and k is determined from Eq. 4: V
ij
− V
ik
= V
ij
− V
ik
+ o
ij
− o
ik
5 We cannot observe V
ij
− V
ik
directly, only its sign. If the individual chooses j over k, V
ij
− V
ik
\ 0. The probability this choice will be made is
given by: P
i, j \ k
= Pr[V
ij
− V
ik
\ 0]= Pr[o
ik
− o
ij
B V
ij
− V
ik
] 6
Assume a linear approximation to the system- atic component of V ,
V
ij
= z
ij
g
j
+ h
ij
7 where h
ij
is a fixed effect and z
ij
the vector of elements in Eq. 3. Letting y
i
V
ij
− V
ik
, V
ij
− V
ik
z
ij
g
j
− z
ik
g
k
+ h
ij
− h
ik
, o
i
o
ij
− o
ik
, Eq.
5 can be written: y
i
= x
i
b + o
i
8 where x
i
= 1, z
ij
, z
ik
and b = h
ij
− h
ik
, g
j
, − g
k
. Estimation of b in Eq. 8 requires observable values of y
i
and an assumed error distribution for o
i
.
3
.
1
. Binary choice BC
U nder binary choice BC, estimation of Eq. 8 requires defining M cF adden, 1974:
y
i
= 1 if y
i
\ y
i
= 0 if y
i
5 9
Therefore, Pry
i
= 1 = Pr − o
i
B x
i
b = Fx
i
b, where F · is the cumulative distribution func-
tion for o
i
. M cF adden 1974 has shown that when o
ij
is iid Weibull, F · is a logistic cumulative distribution function. M aximum likelihood meth-
ods can be used to estimate b in Eq. 8. One deficiency of using logit models to estimate coeffi-
cients in Eq. 8 is that the relative probabilities of choosing between two choice alternatives are con-
strained to be constant regardless of whether other alternatives are available. F or example, the
relative probability of choosing a configuration of Stream A and B quality improvements relative to
the status quo is independent of whether a third alternative, improving Stream C, is available. This
condition is known as the Independence of Irrele- vant Alternatives IIA, and can result in unrealis-
tic probability estimates Judge et al., 1985, p. 70.
1
1
The implications of IIA for the valuation estimates in this study can be considered. The procedure for valuation is to
present the respondent with a binary choice, where Alternative 0 can be represented by w
, A , B
and Alternative 1 by w
1
, A
1
, B
1
with A and B representing quality conditions of the two streams and w the accompanying price. The estimated
probabilities of choice between the two alternatives are used to infer the value of prices that would make the probabilities of
selecting the two alternatives equivalent. The logistic model holds the relative probabilities of these two alternatives con-
stant, even if the experiment is run by introducing a third stream. The result is that marginal valuations of quality im-
provements in Streams A and B are independent of introduc- ing a third stream to the choice set. This is likely to result in
overestimates of marginal stream improvement values when there are other substitute streams that could be introduced
into the choice set.
When any element of z
ij
is constant across alternatives and the corresponding element of g
j
is also constant, the effect of that element on choice
is not identifiable. H owever, a constant element of z
ij
may interact with other non-constant elements requiring its inclusion in an interactive, or varying
coefficients model G riffiths et al., 1993. F or example, while income remains constant across
choices, the effect of environmental quality im- provements on choice may depend on income.
The inclusion of h
ij
in Eq. 7 allows for fixed effects over individuals and alternatives. There
may be a threshold of utility increase before an individual selects an alternative to the status quo.
Letting j = 0 represent the status quo, h
i1
− h
i0
= h
i
\ 0. A more complex possibility of a fixed effect
arises from potential experimental design and elic- itation procedures. F or example, if respondents
act as if the ‘right’ response is to choose the alternative to the status quo, h
i
\ 0. A fixed effect
could also be present if respondents have a desire for change h
i
\ 0 or no change h
i
B 0. The
procedure for introducing fixed effects into the binary choice model has been developed by
Chamberlain 1980, 1984 and Baltagi 1995. The procedure requires a panel data set, as in the
present study, and conditions the likelihood func- tion on the summed 0, 1 responses across each
individual’s revealed choices. F or example, the present study asked each respondent to make a
choice in five different scenarios; the maximum conditioning value is then five.
4. Specific forms for estimation of shadow values
