VALUING WATER QUALITY AS A

FUNCTION OF WATER QUALITY MEASURES

KEVIN J. EGAN, JOSEPH A. HERRIGES, CATHERINE L. KLING,
AND JOHN A. DOWNING

Employing a unique and rich data set of water quality attributes in conjunction with detailed household
characteristics and trip information, we develop a mixed logit model of recreational lake usage and
undertake thorough model specification and fitting procedures to identify the best set of explanatory
variables, and their functional form for the estimated model. Our empirical analysis shows that
individuals are responsive to the full set of water quality measures used by biologists to identify the
impaired status of lakes. Thus, changes in these quality measures translate into changes in the
recreational usage patterns and well-being of individual households. Willingness-to-pay (WTP)
estimates are reported based on improvements in these physical measures.

Key words: Nonmarket valuation, recreation demand, water quality.

More than three decades have elapsed since

formation are in short supply. The purpose
the passage of the 1972 Clean Water Act
of this article is to help fill this gap by
(CWA), yet progress toward meeting the
provid- ing information on the value of
stan- dards set forth in the CWA has been
water quality improvements as a function of
slow in the area of nonpoint source
detailed phys- ical attributes of the water
pollution. The most recent National Water
bodies involved. The water quality values
Quality Inventory (USEPA 2000)
are obtained from a recreation demand
categorizes 45% of assessed lake acres in the
model of lake usage in the state of Iowa,
United States as impaired, with the leading
combining trip and socio- demographic data
causes of these impairments being nutrients
from the Iowa Lakes Valu- ation Project and
and siltation. Moreover, few states have
an extensive list of physical water quality
developed the priority ranking of their
measures collected by Iowa State
impaired waters as required under Sec- tion
303(d) of the CWA. The task facing both the University’s Limnology Laboratory.1
EPA and state regulatory agencies remains Recreation demand models have long
daunting. The prioritization process requires been used to value water quality
information on the cost of remediation and improvements, but studies typically rely on
the potential benefits that will flow from wa- limited measures of water quality. The most
ter quality improvements. Both types of in- commonly used indica- tors are fish catch
rates (e.g., Morey, Rowe, and Watson 1993;
Chen, Lupi, and Hoehn 1999). However,
Kevin J. Egan is assistant professor in the Department of Eco- catch rates are themselves endoge- nous,
nomics at the University of Toledo, Joseph A. Herriges and
Cather- ine L. Kling are professors in the Department of
depending on both fishing pressure and the
Economics at Iowa State University, and John Downing is abilities of the anglers, and provide only in-
professor in the De- partment of Ecology, Evolution and direct measures of the underlying water
Organismal Biology at Iowa State University.
The authors would like to thank participants in seminars at qual- ity. Physical water quality measures,
Resources for the Future, the University of Minnesota, and the such as Secchi transparency and bacteria
Heartland Conference for helpful comments on earlier drafts of counts, are used only sparingly, in large part
this paper. All remaining errors are, of course, our own. This
jour- nal paper of the Iowa Agriculture and Home Economics because of limitations in available data.
Exper- iment Station, Ames, Iowa, Project No. 5141, was Phaneuf, Kling, and Herriges (2000) use fish
supported by Hatch Act and State of Iowa funds. Funding for
this project was also provided by the Iowa Department of toxin levels in their model of Great Lakes
Natural Resources and the U.S. Environmental Protection fishing, but the toxin levels were available
Agency’s Science to Achieve only for a lim- ited number of aggregate
sites in the region.
Results (STAR) program. Although the research described in the
article has been funded in part by the U.S. Environmental Protec-
tion Agency’s STAR program through grant R830818, it has 1
More specifically, physical (e.g., Secchi transparency),
not been subjected to any EPA review and therefore does not chemical (e.g., total phosphorus), and biological data (e.g.,
neces- sarily reflect the views of the Agency, and no official cyanobacteria) are collected, but for brevity we use the term
endorsement should be inferred. physical measures to collectively represent all three categories.
Amer. J. Agr. Econ. 91(1) (February 2009): 106–123
Copyright 2008 Agricultural and Applied Economics
Association
DOI: 10.1111/j.1467-8276.2008.01182.x
Egan et al. Valuing Water as a Function of Quality Measures 107

Parsons and Kealy (1992) use dummy vari- lack of direct infor- mation on how nutrient
ables based on dissolved oxygen levels and levels (phosphorus and nitrogen) impact
av- erage Secchi transparency readings to recreational usage is un-
capture the impact of water quality on
Wisconsin lake recreation. Similarly, Parsons,
Helm, and Bon- delid (2003) construct dummy
variables indi- cating high and medium water
quality levels for use in their analysis of
recreational de- mand in six northeastern
states. These dummy variables are based on
pollution loading data and water quality
models, rather than on di- rect measurements
of the local water quality. In these studies, the
physical water quality indi- cators are found to
significantly impact recre- ation demand, but
because of the limited na- ture of the measures
themselves, provide only a partial picture of
value associated with pos- sible water quality
improvements. Other pa- pers that have used
one or more measures of water quality
include von Haefen (2003), Atasoy, Palmquist
and Phaneuf (2006), Pha- neuf (2002), Kaoru,
Smith, and Liu (1995), Ribaudo and Piper
(1991), Russell and Vaughan (1982), and
Stevens (1966).
An alternative to physical measures of
wa- ter quality has been the use of perceived
or reported water quality measures
(Adamowicz et al. 1997; Jeon et al. 2005).
While perceived measures are likely to be
the direct drivers of behavior (McConnell
1993) and could be studied in a structural
model (Kaoru, Smith, and Liu 1995), their
major drawback is that information can
typically be gleaned only for the sites the
individual has visited. While im- portant
questions concerning the relationships
between perceived and observed measures
re- mains, it is likely that perceptions are
related to physical measures, the focus of
this work.
Bockstael, Hanemann, and Strand’s
(1986) analysis of beach usage in the
Boston-Cape Cod area has perhaps one of
the most ex- tensive lists of objective
physical water qual- ity attributes included
in a model of recre- ation: oil, fecal
coliform, temperature, chem- ical oxygen
demand (COD), and turbidity. However, the
study also points out one of the frequently
encountered problems in iso- lating the
impact of individual water quality attributes
—multicollinearity. Seven additional water
quality measures were available to the
analysts: color, pH, alkalinity, phosphorus,
ni- trogen, ammonia, and total coliform, but
were excluded from the analysis due to
correlation among the measures. While
these choices are certainly reasonable, the
fortunate in the context of setting logarithmic scale used for the these variables in our preferred
model below, the correlation reaches 0.9.
standards in many states, where nutrient
loadings are of particular concern.
The contribution of the current article
lies in our ability to incorporate a rich set of
physical water quality attributes, as well as
site and household characteristics, into a
model of recreational lake usage.
Importantly, the full set of water quality
variables used by biolo- gists to classify
lakes as impaired by the EPA, and therefore
potentially in need of policy ac- tion, are
included. Trip data for the study are drawn
from the 2002 Iowa Lakes Survey. The
survey was sent to a random sample of
8,000 Iowa households, eliciting
information on their recreational visits to
Iowa’s 129 principal lakes, along with
socio-demographic data and atti- tudes
toward water quality issues. The unique
feature of the project, however, is that a par-
allel inventory of the physical attributes of
these lakes was conducted by Iowa State
Uni- versity’s Limnology Laboratory. Three
times a year, over the course of a five-year
project, thirteen distinct water quality
measurements were taken at each of the
lakes, providing a clear physical
characterization of the condi- tions in each
lake. Moreover, because of the wide range
of lake conditions in the state, Iowa is
particularly well suited to identifying the
im- pact of these physical characteristics on
recre- ation demand. Iowa’s lakes vary from
a few clean lakes with up to fifteen feet of
visibil- ity to other lakes having some of
the highest concentrations of nutrients in the
world, and roughly half of the 129 lakes
included in the study are on the EPA’s list of
impaired lakes. An additional unique aspect
of Iowa lakes is that the diversity of land
uses in the watershed contributing to them −
leads to a relatively low degree of
collinearity among the physical and
chemical water quality measures, with
corre- lation coefficients ranging from 0.53
to 0.68, and typically lying below 0.4.2
Thus, Iowa lakes provide an almost ideal
“laboratory” for study- ing the effects of
biological water quality mea- sures on usage
and value.
A second unique contribution of this
study is the application of careful model
specifica- tion and fitting procedures to
identify the best set of explanatory
variables, and their func- tional form, for
the estimated model. Since economic
theory does not provide guidance

2
The biological measures, cyanobacteria, and total
Phytoplank- ton are more highly correlated. In the
108 February 2009 Amer. J. Agr. Econ.

to the analyst on these issues, ex ante selec-

2002 to 8,000 randomly selected households
tion of model variables and structure will in the state. Standard Dillman procedures
often fail to achieve the best model fit. On (1978) were used to ensure a high response
the other hand, specification searching on a
rate.4 Of the 8,000 surveys mailed, 4,423
given data set leads to the well-known
were returned. Allowing for undeliverable
problem of incor- rect standard error
surveys, this corre- sponds to an overall
estimates (Leamer 1983). Thus,We exploit response rate of 62%.
our large sample size by splitting the sample The survey sample was initially paired
into three equal parts. With the first sample,
down to 3,859 households as follows. Those
we extensively explore various house- holds who returnedthesurvey from
specifications of the model. The second
out of state were excluded (thirty-eight
sample was reserved for clean model observations), as it is not feasible to
estimation, alleviating bias in the estimated
ascertain whether these re- spondents have
standard errors of the parameter estimates permanently left the state or simply reside
stemming from the specification search
elsewhere for part of the year. Respondents
process. The third split of the sample was who did not complete the trip questions or
used to perform out-of-sample prediction to
did not specify their numbers of trips (i.e.,
provide an overall assessment of the model they simply checked that they had visited a
fit.3 given lake) were excluded (224 ob-
The remainder of the article is divided servations). Lastly, any household reporting
into five sections. The next section provides more than fifty-two total single-day trips to
an overview of the data, and a repeated the 129 lakes were excluded (133
mixed logit model is specified in the section observations). In the analysis that follows,
on the model. The mixed logit model allows only single-day trips are included to avoid
for a wide variety of substitution patterns the complexity of model- ing multiple-day
among the recreational sites and visits. Defining the number of choice
heterogeneity among households (see, e.g., occasions as fifty-two allows for one trip per
Train 1998; McFadden and Train 2000; week to one of the 129 Iowa lakes. While
Herriges and Phaneuf 2002). The the choice of fifty-two is arbitrary, we chose
specification search procedure and find- ings this cut-off as a conservative approach to in-
are reported in the section on specifica- tion sure our data sample is visitors to the lakes
and estimation, along with assessment of the and not residents who may claim many
out-of-sample predictions. In the section on recre- ational trips simply due to living on
welfare calculation, the policy implications
the shore of a lake.5
of the model are discussed in terms of recre-
Finally, because of the large number of re-
ational value of meeting the objectives of
spondents, the overall sample was randomly
the CWA via removal of all lakes from the
divided into three segments; specification,
im- paired water quality list. Conclusions
es- timation, and prediction portions, each
complete the article.
com- ponent using just under 1,290
observations. Once the estimation stage is
Data reached, the re- sults will be less prone to
pretest bias, and bias in the conventional
Two principal data sources are used in de- standard errors for the re- sulting parameter
veloping our model of recreational lake us- estimates induced by the ex- tensive
age in Iowa: the 2002 Iowa Lakes Survey specification search will be alleviated. For
and the physical water quality measures the bias to be eliminated the errors in the
collected by Iowa State University’s model must be uncorrelated across the indi-
Limnology Labora- tory. The focus of the viduals in the three samples, and the
survey was on gathering baseline “correct” model must be identified after the
information on the visitation patterns to specification search. To the extent that
Iowa’s 129 principal lakes, correlations do exist (say due to shared
unobservable site charac- teristics) or the
“correct” model is not iden- tified, then the
benefits from the split sample procedure
used here will be diminished. How- ever, the
bias in the standard errors should be
Survey was administered by mail in November
4
Complete details of the survey design and implementation
 can be found in Azevedo et al. (2003).
3 5
We are aware of only two other recreation demand studies Sensitivity analysis, raising the allowable number of trips
that have adopted this procedure: Creel and Loomis (1990, per year to ninety-five, indicated that the results do not change
1992) and Haener, Boxall and Adamowicz (2001). quali- tatively, including the welfare results.
Egan et al. Valuing Water as a Function of Quality Measures 109

Table 1. 2002 Iowa Lakes Survey Summary

Statistics tinct aspect of the lake ecosystem. Secchi
trans- parency is one of the most widely
Variable Mean Std. Dev. Min. Max.
applied lim- nological parameters and
approximately re-
flects the lake depth at which the bottom of
Total day trips 6.44 10.22 0 52 the lake can still be seen. Chlorophyll is an
Male 0.68 0.46 0 1 in- dicator of phytoplankton plant biomass
Age 53.34 16.09 15 82 which leads to greenness in the water. Three
Age2 3,104 1,779 225 6,724 nitrogen levels were gathered. Total nitrogen
is the sum +
School 0.66 0.46 0 1 of all dissolved and particulate forms. NH3
Household size 2.61 1.30 1 12 NH4 measures ammonium nitrogen that de-
+
rives from fertilizer or anaerobic conditions
and can be toxic. NO3 NO2 measures the
nitrate level in the water that derives from
reduced relative to having used a single aer- obic nutrient contributions. Total
sample for both specification and phosphorus is an important indicator of
estimation.6 nutrient conditions in freshwater systems
Table 1 provides summary statistics for and is usually the prin- cipal limiting
the full sample regarding trip and the socio- nutrient that determines phyto- plankton
demographic data obtained from the survey. growth. Silicon is important to di- atoms, a
The average number of total single-day trips key food source for aquatic organisms. The
for all 129 lakes is 6.44, varying from some hydrogen ion concentration of the water is
respondents taking zero trips and others tak- measured by pH with levels above eight in-
ing fifty-two trips. In total, 41.2% of the dicating extreme basicity. As table 2 notes,
sam- ple (1,588 respondents of the 3,859) all of the pH levels in this sample are
reports zero recreation trips. Of the clustered between 7.7 and 10. Plants need
remaining 2,271 respondents, 61% report carbon to grow and most carbon comes from
fewer than ten trips, with the number of alkalinity in lakes; therefore, alkalinity is an
respondents reporting a particular trip level indication of the availability of carbon to
gradually declining as the number of trips plant life. Inor- ganic suspended solids (ISS)
increases. Only 3.7% of the respondents consist of soil and silt suspended in the
who report positive trips take more than water through erosion, whereas volatile
forty. While on these visits, the survey suspended solids (VSS) con- sists of
respondents also reported engaging in a suspended organic matter. Increases in
wide variety of activities, with the top six either ISS or VSS levels decrease water
activities consisting of: recreational boating clarity.
(58.8%), fishing (47.4%), picnicking Table 2 demonstrates that there is
(46.6%), nature appreciation/wildlife consid- erable variation in water quality
viewing (44.9%), swimming/beach use conditions throughout the state.
(37.7%), and camping (28.5%). In general, In addition to trip and water quality data,.
the survey respondents are more likely to First, the travel costs, from each survey
be older, male, and to be more educated respondent’s residence to each of the 129
than the general population. Schooling is lakes, were needed. The out-of-pocket
characterized in terms of a dummy variable component of travel cost was computed as
equaling one if the individual has at- tended the round-trip travel distance multiplied by
or completed some level of post high school $0.25 per mile. The opportunity cost of time
education. was calculated as the estimated round-trip
Table 2 provides summary statistics on travel time multiplied by
the average water quality values for the
2002 sea- son. Samples were taken three
times through- out the year, in spring/early
summer, mid- summer, and late summer/fall
to cover the range of seasonal variation.
Each of the water quality measures helps to
characterize a dis-

6 out the concern with the correlated standard errors, and one of the
The authors would like to thank Spencer Banzhaf for pointing
reviewers for pointing out the concern with the misspecification
error. 8
The $0.25 per mile is used as a relatively conservative estimate
7
The unit of observation is the household. Although, we also of gasoline and deprecation costs per mile of driving. This esti-
in- mate is generally less than most official government reimburse-
clude the survey respondent’s individual socio-demographic ment rates. The round-trip travel distances and times were calcu-
data. lated using the software package PCMiler (Streets Version 17).
 110 February 2009 Amer. J. Agr. Econ.

Table 2. Water Quality Variables and 2002 Summary Statistics

Variable Mean Std. Dev. Min. Max.
Secchi transp. (m) 1.17 0.92 0.09 5.67
Chlorophyll (µg/L) 41 38 2 183
NH3 + NH4 (µg/L) 292 159 72 955
NO3 + NO2 (mg/L) 1.20 2.54 0.07 14.13
Total nitrogen (mg/L) 2.20 2.52 0.55 13.37
Total phosphorus (µg/L) 106 81 17 453
Silicon (mg/L) 4.56 3.24 0.95 16.31
pH 8.50 0.33 7.76 10.03
Alkalinity (mg/L) 142 41 74 286
Inorganic SS (mg/L) 9.4 17.9 0.6 177.6
Volatile SS (mg/L) 9.4 7.9 1.6 49.9
Cyanobacteria (mg/L) 295.8 833.1 0.01 7178.1
Total phytoplankton (mg/L) 304.8 835.2 3.99 7178.6

Table 3. Summary Statistics for Lake Site

Characteristics ond is a “wake” dummy variable that equals
one if motorized vessels are allowed to
Variable Mean Std. Dev. Min. Max.
travel at speeds great enough to create
wakes and
zero otherwise. About 66% of the lakes allow
Travel cost 135.09 29.23 93.90 238.61 wakes, whereas 34% of lakes are “no wake”
Acres 672 2,120 10 19,000 lakes. The “state park” dummy variable
Ramp 0.86 0.35 0 1 equals one if the lake is located adjacent to a
Wake 0.66 0.47 0 1 state park. The last dummy variable is the
State park 0.39 0.49 0 1 “hand- icap facilities” dummy variable,
Handicap 0.39 0.49 0 1
which equals one if handicap amenities are
facilities
Fish index 1.01 1.14 0 4 provided.10 Fi- nally, a “fish index” variable
is included that varies from zero to four
representing the num- ber of fish species for
which the lake is consid- ered one of the
“top 10” sites for the species in the state.11
one-third of the respondent’s average wage
rate. Table 3 provides summary statistics The Model
for the resulting travel cost variable. The
average price of a recreational trip to a lake
The repeated mixed logit model allows for
is $135.
corner solutions, integrates the site selection
Table 3 provides a summary of these site and participation decisions in a utility
characteristics. As table 3 indicates, the size of consistent framework, and controls for the
the lakes varies considerably, from 10 acres to count nature of recreation demand”
19,000 acres. Four dummy variables are The model assumes that the utility
included to capture different amenities at each individual i receives from choosing site j on
lake. The first is a “ramp” dummy variable choice occasion t is of the form
which equals one if the lake has a cement boat
ramp, as opposed to a gravel ramp or no boat (1) U ijt = V (X i j ; þi ) +  ijt , i = 1 , . . . , N ;
ramp at all. The sec-
j = 0,..., J ; t = 1,..., T

10
9 As one of the reviewers points out, the handicap facilities vari-
The “average wage rate” is calculated for all respondents as
able may proxy more generally for the ease of accessibility of the
their household’s income divided by 2,000 (assuming total annual
lake.
hours worked is forty hours per week for 50 weeks). 11
The candidate fish species are bluegill, crappie, largemouth
and
smallmouth bass, catfish, bullhead, walleye, sunfish, yellow
perch, and northern pike. See Iowa Department of Natural
Resources (2004b) for details.
Egan et al. Valuing Water as a Function of Quality Measures 111

where V represents the observable portion of

where ai is the random parameter on the
utility and, from the perspective of the re-
dummy variable, Dj, which does not appear
searcher, ijt represents the unobservable explicitly in equation (4) since it equals one
portion of utility. A mixed logit model is
1, . . . , J and zero for j 0. The observ-
for j =
defined as the integration of the logit
able information included in the stay-at-home
formula over the distribution of unobserved
option, represented by zi, is the set of socio-
random parame- ters (Revelt and Train
demographic characteristics such as gender,
1998). If the random parameters, þi, were
age, and education. Notice that with this spec-
known, the probability of observing
ification the socio-demographic factors do not
individual i choose alternative j on choice
influence thechoice among sites, but simply
occasion t would follow the logit form
the participation decision, as this information
does
(2) not vary across the sites. The xij matrix repre-
exp(Vijt (þi ))
Lijt(þ )i = J . sents the observable information that varies
.
exp[Vikt (þi )] across the lakes, including the travel costs
k=0 for each individual, lake characteristics, and
water
Since the þis are unknown, the quality measures.
correspond- ing unconditional probability, Therandom coefficient vectors for each
Pijt(0), is ob- tained by integrating over an indi- vidual, þi and ai, can be expressed as the
assumed proba- bility density function for sum of population means, b and a, and
the þis. The uncondi- tional probability is individual de- viation from the means, 6i and μ
now a function of 0, where 0 denotes the i , which rep- resents the individual’s tastes
vector of hyperparameters (e.g., the mean relative to the average tastes in the population
and standard deviation) characteriz- ing the (Train 1998). Therefore, we can note that
distribution of the random parameters
(i.e., the þis) in the population. This J J J
(5) þi xi j = b xi j + 6i xi j
repeated mixed logit model assumes the
random param- eters are i.i.d. distributed and
over the individuals so that
(6) ai = a + μi .
(3) Pijt (0) = ∫ ijt (þ) f þ|0)dþ.
(
L
The partitioned utility is then
No closed-form solution exists for this uncon- . J
ditional probability and therefore simulation (7) U þ˜ z i + μ i 0t , j = 0
is required for maximum likelihood ijt =
estimation of 0.12 b
J
+ a + ijt , j = 1,..., J
Following Herriges and Phaneuf (2002), a xi j μ
dummy variable, Dj, is included which equals where
one for all of the recreation alternatives (j = (8) 
1, ... , J) and equals zero for the stay-at-home i , j = 0; i = 1 , . . . , N ;
option (j = 0). Including the stay-at-home op- μ
0t
t = 1,..., T
= 6
  Ji x + μ +  , j = 1, . . . , J ; i = 1,
tion
in theallows a complete
population those set of choices,
individuals who ijt  ij
 i i jt
including
do not visit
always “stay any
at of the on
home” sites. It ischoice
every convenient ..., N ; t = 1,...,
to
occasion andindividual’s utility into the stay-
partition the T
at-home option or choosing one of the J sites, is the unobserved portion of utility. This por-
with tion is correlated over sites, j 1, . . . , J, and
choice occasions, t 1, . . . , T, =because of the
common influence of = the terms 6i and μ i. The
(4) U . J
þ˜ z i + i 0t , j=0 same preferences are used by the individual to
ijt = ai + 
þi x i j +
J
ijt
, j = 1,..., e lua si t each unobserved portion of utility is correlated
v te t time over sites and trips, the familiar
J
a ea e period. Independence of Irrelevant Alternatives
12
ch a Since the (IIA) assumption
Modified Latin Hypercube Sampling (MLHS) is used in the
simulation process (Hess, Train, and Polak 2006). Hess, Train, does not apply for mixed logit models. For ex-
and Polak demonstrate that MLHS provides better coverage of ample, consider ai, the random parameter on
the distribution space for the same number of draws versus the dummy variable Dj, which equals 0 for the
Halton se- quences or any of the Halton sequence variations
(such as shuffled Halton sequences). home option and 1 otherwise. An individual
112 February 2009 Amer. J. Agr. Econ.

who chooses the stay-at-home option for all

characteristics in determining recreation de-
choice occasions would have a negative
mand patterns, holding constant the manner
devia- tion from a (i.e., μ i < 0), the mean
in which both socio-demographics and other
of ai, while someone who takes many trips
site characteristics impact preferences. Specif-
would have a positive deviation (i.e., μ i >
ically, socio-demographic characteristics are
0), allowing the marginal effect of Dj to vary
assumed to enter through the stay-at-home
across individ- uals. However, the individual
carries their ai across choice occasions and option. They include age and household size,
sites. as well as dummy variables indicating gen-
We model the utility individual i receives der and college education (see table 1). A
from lake j on choice occasion t as quadratic age term is included in the model
to allow for nonlinearities. Site characteristics,
 identified in table 3, are included with random
˜
 þ j z i + i 0t , j=0 coefficients. This allows for heterogeneity in
(9) =  −þP + þq j individual preferences regarding site charac-
Uijt
Q
j j =Pij1 , . . . ,

J +i þ A + a
J
a water quality, site characteristics, and house-
hold characteristics identified in tables 1–3.
where the +vector
j zi consists
i i jt of socio- We Focus on modeling the role of water quality
demographic characteristics, Pij is the travel
cost from each Iowan’s residency to each of
the 129 lakes. The vector Qj denotes the
phys- ical water quality measures and Aj
represents the attributes of the lake. As
shown in equa- tion (9), notice that the
parameters on the lake attributes and the
dummy variable, Dj, are ran- dom. All of the
other variables are assumed to have fixed
parameters.

Specification and Estimation

While the repeated mixed logit model

provides the general framework for our
analysis, it does not determine the specific
variables to use in the model or the
functional form they should take. Moreover,
economic theory provides lit- tle guidance,
although valuable guidance can be obtained
from a number of sources, includ- ing
ecologists who understand the source and
degree of impairment in a watershed, previ-
ous studies of similar water quality concerns,
and focus groups in the affected population.
Nonetheless definitive functional forms and
choice of variables based on physical
science or other sources will be rare. In
order to inves- tigate the model specification
issue, we divided the full survey sample into
three portions, with one portion each
dedicated to model specification, estimation,
and out-of-sample prediction.

Specification
There are, of course, a large number of po-
tential model specifications given the range of
teristics, such as wake restrictions and site fa-
cilities, observed in previous studies (e.g.,
Train 1998). For example, some households
may pre- fer to visit less developed lakes
with wake restrictions in place, while others
might be at- tracted to sites allowing the use
of motorboats, jet skis, etc. To allow for a
range of possible reactions to the various
site characteristics, it is initially assumed
that the random parame- ters (þa) are each i
normally distributed with the mean and
dispersion of each parameter estimated.
Even restricting our attention to the wa-
ter quality characteristics in table 2, there
are a large number of potential model
specifica- tions. We focus on five groups of
water qual- ity characteristics for the Qj in
equation (9): Secchi transparency,
chlorophyll, nutrients (to- tal nitrogen and
total phosphorus), suspended solids
(inorganic and organic), and bacteria
(cyanobacteria and total phytoplankton).
The first four characteristic groups directly
affect visible features of water quality,
making it more likely that households
respond to them. Bacteria is included
because surveyed house- holds report it to
be the single most impor- tant water quality
concern (see Azevedo et al. 2003).
Our initial intent was to consider three pos-
sible specifications for each of these groups
of variables: inclusion linearly, inclusion log-
arithmically, or exclusion. However, prelimi-
nary analysis indicated that these variables in-
dividually and as groups were consistently
sig- nificant at a 5% level or better. Thus, we
chose to focus on determining whether each
group of factors should enter the model in a
linear or logarithmic fashion. This required
estimat- ing a total of 2 5 32 versions of the =
model. Table 4 provides a summary of the
results, with the various specifications listed
in terms of de- creasing log-likelihood. The
models in table 4 are non-nested
alternatives relative to each
Egan et al. Valuing Water as a Function of Quality Measures 113

Table 4. Recreation Demand Model Specifications Listed in Decreasing Log-Likelihood Values

Variable Group Specification
Secchi Depth Chlorophyll Nitr. and Phosp. Susp. Solids Bacteria Log-Likelihood
Linear Log Log Linear Log −39010.2
Linear Log Linear
Log Log Log
−39016.2 Linear Log Linear Linear
Log Log −39032.9 Linear Linear Linear
Log Log −39037.7 Linear Linear
−39029.7
Log Log Linear Log
Linear Linear
Log Log
−39045.6
−39034.7
Log Log Log Log
Linear Log
Linear Log
Log
Log
−39063.4 Log Linear Linear Linear Log
−39056.9
−39065.6Log
Linear Linear Linear Log Log
−39067.3 Linear Linear Log −39090.1
Linear
−39071.1Linear Log Log Log
Log Log −39087.6
Log Linear Log Linear Log
−39097.5Linear
Linear Log Log Log
Linear Log
Linear
−39104.0 Linear Log Linear Log
Linear Linear −39183.0 Linear Log Linear Linear
Linear −39196.9
−39172.1
Log Log Log Log
Log Log
Linear Linear
Linear −39198.5
Linear Linear Linear Log Linear
Log −39208.4 Linear Linear Log Linear Linear
Linear −39219.6 Log Linear Log Linear Linear
−39201.3Log Log Log Linear −39227.1
−39218.0Linear
Log Log Log
Linear Log
Linear Linear −39232.7
−39235.9
Log
Linear Linear Linear Linear Linear
Log Linear Linear Log Linear
−39231.9Log −39240.9 Log Linear Linear Log
Linear Linear
Linear −39242.5
Linear
−39236.3 Linear Log
Note: 200 MLHS draws are used for the simulations underlying this table.
Log Linear −39248.3

other, precluding a direct test of one model els that have bacteria entering logarithmically
over another.13 Since themodels all employ the being preferred to (i.e., having a higher log-
same number of parameters, Pollak and Wales’ likelihood relative to) those in which bacte-
(1991) likelihood dominance criteria for model ria enters linearly, regardless of the functional
selection, used in the current analysis, reduces form for the remaining variables. Given the
to a direct comparison of log-likelihood values. specification for bacteria, the models having
One interesting feature of the results is that chlorophyll entering logarithmically are gener-
the modeling ranking is lexicographic in terms ally preferred to those having it enter linearly.
of the bacteria specification, with all mod- The preferred model has Secchi transparency
and suspended solids entering the model lin-
13
early, with the remaining variables entering in
One can, of course estimate an “encompassing” model,
includ- ing both linear and logarithmic versions of the water a logarithmic fashion. This model is referred
quality vari- ables. Each model in table 4 could then be tested to as model A below.14
against this hybrid specification. As Greene (2000, p. 301)
notes, however, there are problems with this approach. First,
such an approach tests each
model against the hybrid model, not against each other. Second,
the hybrid model requires a substantial increase in the number of
parameters, increasing potential multicollinearity problems. 14
We also estimated exploratory models utilizing the
More- over, given that the loglikelihood value for the preferred quadratic function for the water quality measures. The results
model provides a lower bound on the loglikelihood value for indicate that a model with Secchi transparency and the
the hybrid model, the results in table 4 indicate that, at best, suspended solids en- tered quadratically (and the nutrients and
only the first two models would not be rejected using a bacteria levels remain- ing logged) fits better than model A.
likelihood ratio test. Thus, the hybrid strategy would provide However, we do not formally include the quadratic form in our
little guidance regarding the choice of models. specification search, as the ver-
sions of the model would increase from 32 to 243 (=35 ).
114 February 2009 Amer. J. Agr. Econ.

Table 5. Most Important Additional Water Quality Measures in

Recreational Choice
Secchi Transparency and Log-Likelihood y2
df =1
a. Single Additions
∗∗
Log(total phosphorus) −39275.4 67.2
∗∗
ISS −39279.9 58.2
∗∗
Log(chlorophyll) −39282.6 52.8
∗∗
Log(cyanobacteria) −39288.8 40.4
∗∗
Log(total nitrogen) −39303.9 10.2
Log(total phytoplankton) −39307.4 3.2
VSS −39308.1 1.8
Nothing −39309.0
b. Pairwise Additions
∗∗
Log(cyanob.)/log(total phytop.) −39173.1 271.8
∗∗
Log(chlorophyll)/log(total phos.) −39194.8 228.4
∗∗
Log(chlorophyll)/ISS −39243.2 131.6
∗∗
Log(cyanob.)/log(total phos.) −39250.9 116.2
∗∗
Log(cyanob.)/ISS −39262.4 93.2
∗∗
Log(total phos.)/ISS −39262.5 93.0
∗∗
Log(chlorophyll)/VSS −39270.4 77.2
∗∗
Log(chlorophyll)/log(total nit.) −39271.9 74.2
∗∗
Log(cyanob.)/log(chlorophyll) −39272.5 73.0
∗∗
Log(total phos.)/log(total nit.) −39274.2 69.6
∗∗
Log(total phosp.)/log(total phytop.) −39274.5 69.0
∗∗
Log(total phosp.)/VSS −39274.5 69.0
∗∗
Log(chlorophyll)/log(total phytop.) −39275.8 66.4
∗∗
VSS/ISS −39277.0 64.0
∗∗
Log(total phytop.)/ISS −39278.2 61.6
∗∗
Log(total nitrogen)/ISS −39278.5 61.0
∗∗
Log(cyanobacteria)/log(total nit.) −39282.0 54.0
∗∗
Log(cyanobacteria)/VSS −39288.2 41.6
∗∗
Log(total phytop.)/log(total nit.) −39303.0 12.0
∗∗
Log(total nitrogen)/VSS −39304.0 10.0
Log(total phytop.)/VSS −39307.3 3.4
Nothing −39309.0
Note: 200 MLHS draws are used for the simulations underlying this table. The y 2 test statistics reported in table 5 are
likelihood ratio tests for the exclusion of the additional water quality measures relative to a model using Secchi
transparency alone to capture water quality conditions. In each case, single and double asterisks are used respectively, to
indicate significance at the 5% and 1% critical levels.

In addition to investigating the appropri- Table 5 indicates the marginal benefit

ate functional form for the five water quality from gathering additional water quality
groups above, we also considered the ques- measures, suggesting the question of
tion as to which factors had the greatest im- whether the resulting gains in explanatory
pact on the overall fit of the model. Sec- power are worth it. Specifically, ranking the
chi transparency was included in the model cost of obtaining each of the measures from
due to its frequent availability. We then con- easiest to the most difficult and costly to
sidered which single factor had the greatest obtain, we have: (a) Sec- chi transparency,
marginal impact on the log-likelihood, as well (b) nutrients (total nitro- gen and total
as which pair of factors together had the great- phosphorus), (c) chlorophyll and the
est impact on log-likelihood. The results are suspended solids (inorganic and volatile),
re- ported in table 5. Total phosphorous is the
most important additional single factor, where
to- tal phytoplankton and cyanobacteria
emerged as the most important pair of
additional characteristics.
Egan et al. Valuing Water as a Function of Quality Measures 115

possible
and (4) bacteria (cyanobacteria and total
phy- toplankton). Focusing on the five 15
Hensher and Greene (2003) provide an excellent discussion of
groups of water quality measures we have the alternative distributional assumptions commonly used in mixed
considered, Secchi transparency is clearly logit models and their limitations.
the best single measure to include, as it is
easy to obtain and consistently has a
substantial and statistically significant
impact on recreational site selection and
participation decisions. Nutrient levels are
also relatively easy to collect and
phosphorous levels are the most important
additional mea- sure after Secchi
transparency. Moreover, nu- trient loadings
have consistently been a target of
agricultural and environmental policies and
understanding their impact on recreational
us- age patterns would seem worthwhile.
There- fore, if possible we recommend the
inclusion of the nutrient levels in the data
collection pro- cess. The bacteria levels are
also found to be consistently correlated with
recreational be- havior in our analysis.
However, they are also the most costly.
Choosing to include them in the data
collection process will be dictated in part by
the potential for cyanobacteria levels to
reach dangerous levels and threaten human
health. Questions in the Iowa Lakes survey
in- dicate that this was the single greatest
concern regarding water quality amongst
respondents. In our informal cost-benefit
analysis chloro- phyll and the suspended
solids fair the worst. They are relatively
more expensive to collect and also the least
important measures in ex- plaining
recreational usage patterns.
Finally, the specification search process
above assumes that the random parameters
associated with site characteristics are each
normally distributed. While this assumption
is common practice in the literature, there is
no a priori basis for this choice. Moreover, a
num- ber of authors have noted potential
problems with unbounded parameter
distributions (such as the normal) leading to
implausible results for some portion of the
population (Revelt and Train 1998). As a
final stage in the spec- ification search
process, we consider alterna- tive choices
for the random parameters in our model.15
Specifically, We grouped the site char-
acteristic parameters in (b) the five parame-
ters associated with the discrete variables for
paved ramps, handicap facilities, state parks,
wake restrictions, and the fish index, and (c)
participation parameter ai. For each set of
pa- rameters, we consider both normal and
trian- gular distributions, for a total of eight
 18
sets of distributional assumptions. The For the results in table 7, a total of 750 MLHS draws are used
in the simulation. Changing the distributional assumptions
advan- tage of the triangular distribution is across specifications required a large number of draws to obtain
that it al- lows for both positive and stability in the comparisons. The estimation results in table 8
negative parameter values and retains the are also with 750 MLHS draws.

basic shape of the nor- mal distribution

(i.e., with a definitive mode and a density
that diminishes away from that mode),
while avoiding the unbounded nature of
the normal distribution. As a basis of
com- parison, we also include a model in
which all but the participation parameter
are fixed (i.e., nonrandom). This structure
mimics the stan- dard nested logit
structure (the participation parameter is
normally distributed instead of the closed-
form extreme value distribution), nesting
together all the trip options as separate
from the stay-at-home option, and is
referred to as model A0 below.16 In
considering the var- ious distributional
assumptions, we also retain the
“preferred” functional form specification
for the water quality variables identified
in ta- ble 4. Clearly, the specification
search could be repeated for each
distribution scenario (re- quiring an
additional 8 32 256 models to be
estimated), but such a task is beyond the ∗ =
scope of the current paper.17 Our focus is
instead on illustrating the potential impact
that such dis- tributional assumptions can
have on the result- ing parameter
estimates and the overall perfor- mance of
the model in terms of prediction.
Table 6 provides the comparison of the
nine distribution scenarios in terms of
maximum log-likelihood values.18
Interestingly, the standard normal
specification’s performance is the second
worst of the nine scenarios, dominating
only model A0. Generally, triangular
distribu- tions appear to fit the data better
in terms of both the lake size and discrete
size character- istic parameters, with the
results being more mixed for the
participation parameter ai. The preferred
specification includes random trian- gular
distributions for all of the site charac-
teristics and a normal distribution for the
par- ticipation parameter. In the remainder
of this

16
There is, of course, a key distinction between model A0
and the standard nested logit specification in that the latter
assumes that the individual’s decisions over choice
occasions are independent, whereas model A0 allows for
correlation as induced by the shared element ai.
17
The distributional assumptions investigated here is by no
means exhaustive. In particular, discrete-factor and latent
class dis- tributions may also prove useful in capturing
variations in agent preferences, including the potential
multi-modal features of these underlying distributions. See,
e.g., Morey, Thacher and Breffle 2006; Hilger and
Hanemann 2006.
116 February 2009 Amer. J. Agr. Econ.

Table 6. Log-Likelihood Comparison Across Distribution Scenarios for

Recreational Choice
Site Characteristic Group Specification
Lake size Discrete Site Vars. Participation Parameter (ai) Log-Likelihood
Triangular Triangular Normal −38677.5
Triangular Normal Normal
Normal Triangular Triangular −38688.6
−38700.1
Triangular Triangular Triangular −38701.3
Triangular Normal Triangular
Normal
Normal Triangular Normal −38704.7
−38716.1
−38704.1
Normal Normal Normal Triangular
Normal
Fixed −40946.8
−38718.6Fixed Normal

paper, we refer to this specification as model found to be statistically significant factors in recreation site
choice.
A1. Since model A1 best fits the data, we
con- sider it to be the best model for
estimation, welfare analysis, and prediction.
In the next section, we turn to the estimation
of model A1 and other models for
comparison.

Estimation Results
Given the results from the specification
search, five models were used in the second
third of the sample: models A, A1, and A0
(varying the specification for the random
parameters of the model), a model including
only Secchi transparency as a measure of
water quality (referred to as model B
hereafter), and a model with all of the water
quality variables entering in a linear fashion
(model C).19 We include model B to
illustrate the consequences of re- lying on a
single measure of water quality, in this case
one that is often available to analysts. Model
C reflects what might be naturally con-
sidered as a default specification. The
resulting parameter estimates are presented
in tables 7 and 8.
In table 7a, focusing on model A1, all of
the coefficients associated with household
char- acteristics are significant at the 5%
level, ex- cept for age. Note that the socio-
demographic data are included in the
conditional indirect utility for the stay-at-
home option. Therefore, in model A1, males,
less-educated individuals, and larger
households are all more likely to

19
An additional model (model D), including all thirteen water
quality characteristics listed in table 2, was also estimated and
is available in the electronic appendix to this paper (Egan et al.,
forth- coming) available online through AgEconSearch. The
results from this expanded model are not substantially different
from the results for model A, with pH and Alkalinity also
take a trip to a lake. However, the size and
sign of the age, school, and household size
variables are sensitive to the model
specification used. The price coefficient is
negative and virtually identical across all
models.
The physical water quality coefficients
are reported in table 7b are relatively stable
across the various models. For all models,
the ef- fect of Secchi transparency is
positive and, in general, organic and
inorganic (volatile) suspended solids have a
negative impact (al- though ISS and VSS
are not statistically sig- nificant at the 5%),
indicating the respondents strongly value
water clarity. However, the co- efficient on
logged chlorophyll is positive, sug- gesting
that on average respondents do not mind (or
even prefer) some “greenish” water. The
negative coefficient on logged total phos-
phorus, the most likely principal limiting
nu- trient, indicates higher algae growth
leads to fewer recreational trips. High
logged total ni- trogen levels also have a
negative impact on recreational utility
associated with a site, al- though it is only
statistically significant at the 5% level for
model A. The log of the possibly toxic
cyanobacteria is negative, while the log of
total Phytoplankton is positive, indicating
that as cyanobacteria is a larger percentage
of the total Phytoplankton in the lake,
recreators are less likely to visit.
Finally, turning tothesite amenities, again
all of the parameters are of the expected
sign. As the size of a lake increases, has a
cement boat ramp, gains handicap facilities,
or is adjacent to a state park, the average
number of visits to the site increases.
Notice, however, the large dispersion
estimates in table 8b. For example, in model
A the dispersion on the size of the lake
indicates 11% of the population prefers a
smaller lake, possibly someone who enjoys
a
 Egan et al. Valuing Water as a Function of Quality Measures 117

Table 7. Repeated Mixed Logit Model Parameter Estimates—Fixed Parameters a (Robust Std.
Errors in Parentheses)
Model B Model C
Model A Secchi Linear WQ Model A1 Model A0
a. Household Characteristics
∗∗ ∗∗ ∗∗
Male −7.91 −6.65 −9.37 −6.63 −5.33
(6.57) (3.47) (2.16) (1.71) (1.60)
∗∗
Age 1.13 −0.23 −0.73 −0.35 −0.43
(0.32) ∗ (0.41) (0.47) (0.29) (0.28) ∗∗
∗ ∗
Age2 −0.0071 0.0064 0.0090 0.0064 0.0079
(0.0032) (0.0048) (0.0042) (0.0027) (0.0027)
∗∗
School −2.11 5.80 −2.59 5.70 −2.62
(2.29) (2.34) (2.37) (1.84) (1.74)
∗∗ ∗
Household 0.75 −0.68 0.36 −1.88 −1.41
(0.73) (0.69) (1.15) (0.56) (0.65)
∗∗ ∗∗ ∗∗ ∗∗ ∗∗
Price −0.49 −0.49 −0.49 −0.49 −0.50
(0.023) (0.024) (0.024) (0.023) (0.023)
b. Water Quality Attributes
∗∗ ∗∗ ∗∗ ∗∗ ∗∗
Secchi transp. (m) 2.40 1.77 1.73 2.49 1.38
(0.66) (0.35) (0.33) (0.58) (0.53)
∗ ∗∗ ∗
Log(chlor. (µg/l))b 2.37 0.022 2.42 1.52
(1.02) (0.16) (0.69) (0.67)
∗
Log(nitrogen (mg/l))b −1.16 −0.18 −0.84 −0.44
(0.53) (0.13) (0.51) (0.46)
∗∗ ∗∗ ∗∗ ∗∗
Log(phosph. (µg/l))b −2.83 −0.023 −2.87 −2.30
(0.96) (0.0078) (0.76) (0.75)
Inorganic SS (mg/l) −0.0057 −0.012 0.0013 −0.014
(0.029) (0.034) (0.029) (0.030)
Volatile SS (mg/l) −0.019 0.052 −0.014 −0.0074
(0.076) (0.78) (0.069) (0.069)
∗∗ ∗∗ ∗∗
Log(cyanobact. (mg/l))b − 2.44 − 0.0010 − 2.39 1.86
(0.40) −(0.030) (0.36) (0.38)
∗∗ ∗∗ ∗∗
Log(total phytop. (mg/l))b 3.44 0.0022 3.28 2.19
(0.56) (0.030) (0.49) (0.51)
LogLik −37,627 −37,759 −37,743 −37,568 −39,960
Note: All of the parameters are scaled by 10. single asterisk ( ∗) and double asterisk (∗∗ ) are used to denote significance at the 5% and 1% levels, respectively.
a For Model C, each of these variables enter linearly, rather than logarithmically.

more private experience. The large

models estimated in the previous subsection.
dispersion on the “wake” dummy variable
For each model, the predicted number of
seems partic- ularly appropriate given the
to- tal trips taken per year was computed
potentially con- flicting interests of anglers
for each household and used to calculate
and recreational boaters. Anglers would
the sample’s associated root mean square
possibly prefer “no wake” lakes, while
error (RMSE) in predicting the actual
recreational boaters would obviously prefer
number of trips taken. In addition, these
lakes that allow wakes. It seems the
calculations were made for the two
population favors a lake that allows wakes
subsamples of nonparticipants and partic-
by almost a two to one margin, with 63%
ipants, where the latter is defined as a
preferring a lake that allows wakes and 37%
house- hold having visited at least one of
preferring a “no wake” lake. Lastly, the
the lake sites in 2002. The results are
mean of ai, the trip dummy variable, is
summarized in table 9. None of the models
negative as expected.
perform particularly well in predicting the
actual participation pat- terns in the third
Out-of-Sample Prediction sample. The three models employing
unbounded distributions for their
The final third of the survey sample was random parameters (A, B, and C)
used to examine the out-of-sample overstate total trips by nearly 80%,
performance of the predicting total trips of roughly eleven,
when the average number of
118 February 2009 Amer. J. Agr. Econ.

Table 8. Repeated Mixed Logit Model Parameter Estimates—Random Parameters (Robust

Std. Errors in Parentheses)
Distributional Assumptions on Random Parameters
Model A Model B Model C Model A1 Model A0
Variable All Normal All Normal All Normal Triang/Normal Fixed/Normal
a. Parameter Means
∗∗ ∗∗ ∗∗ ∗∗ ∗∗
Log(Acres) 4.70 4.02 4.07 4.62 4.13
(0.65) (0.25) (0.29) (0.29) (0.27)
∗∗ ∗∗ ∗∗ ∗∗ ∗∗
Ramp 12.49 11.51 13.15 22.01 5.33
(3.09) (2.66) (2.43) (3.42) (1.65)
∗∗
Facilities 2.07 1.37 1.12 1.42 2.20
(1.13) (1.21) (1.13) (0.90) (0.70)
∗∗ ∗∗ ∗ ∗ ∗∗
State park 3.90 4.04 3.53 2.51 2.78
(1.05) (1.12) (1.40) (1.28) (0.78)
∗∗ ∗∗ ∗ ∗∗
Wake 3.79 3.88 1.59 2.08 3.25
(1.45) (1.26) (1.61) (0.97) (0.93)
∗
Fish index 0.043 −0.12 −0.94 −0.091 0.01
(0.78) (0.46) (0.45) (0.51) (0.33)
∗∗ ∗∗ ∗∗ ∗∗ ∗∗
a −11.90 −11.80 −11.93 −12.46 −11.78
(0.27) (0.30) (0.28) (0.28) (0.24)
b. Parameter Dispersionsa
∗∗ ∗∗ ∗∗ ∗∗
Log(acres) 3.80 3.00 3.40 7.43
(0.37) (0.54) (0.31) (0.46)
∗∗ ∗∗ ∗∗ ∗∗
Ramp 16.85 17.30 18.98 60.44
(2.35) (1.85) (1.57) (6.78)
∗∗ ∗∗ ∗∗ ∗∗
Facilities 13.46 13.95 14.21 29.17
(1.34) (1.33) (1.32) (2.73)
∗∗ ∗∗ ∗∗ ∗∗
State park 12.35 13.98 12.88 36.77
(1.25) (1.47) (1.16) (2.29)
∗∗ ∗∗ ∗∗ ∗∗
Wake 11.73 13.99 14.10 32.47
(1.33) (1.67) (2.94) (2.21)
∗∗ ∗∗ ∗∗ ∗∗
Fish index 6.36 5.41 4.62 10.50
(0.83) (0.69) (0.44) (1.21)
∗∗ ∗∗ ∗∗ ∗∗ ∗∗
a 2.76 2.76 3.01 2.93 2.29
(0.24) (0.14) (0.15) (0.10) (0.066)

Note: All of the parameters are scaled by 10, except a (which is unscaled). Single asterisk (∗) and double asterisk (∗∗ ) are used to denote significance at the
5% and 1% levels, respectively.
a The dispersion coefficients are the estimated standard deviations for the normally distributed parameters, and the estimated spread coefficients for the

triangularly distributed parameters.

Table 9. Trips from Alternative Model Specifications

Actual Model A Model B Model C Model A1 Model A0
Mean Trips 6.27 10.97 11.11 10.94 8.27 5.53
(10.73) (10.87) (10.65) (9.78) (9.60)
Nonparticipants 0 9.29 9.53 9.57 7.12 4.45
(10.33) (10.65) (10.40) (8.00) (5.38)
Participants 10.77 12.17 12.24 11.92 9.09 6.31
(11.00) (11.03) (10.83) (10.88) (11.73)

.
Note:trips
total Rootfor
mean
eachsquare
person,error
and reported
Yˆi is theinfitted
parentheses. We
total trips forcalculate
each ,
RMSE as RMSE = N −1 N ˆ
i (Y i − Y ) where
2
i Y
person. i =1
is the

trips is 6.27. Using bounded triangular

by almost 60%. The simplest of the three mod-
distribu- tions for the random parameters
els (model A0), which has all fixed
(model A1) not only provides a better fit to
parameters except one, while providing the
the model, but also reduces the
worst fit from a log-likelihood perspective,
overprediction out-of-sample
comes closest to
Egan et al. Valuing Water as a Function of Quality Measures 119

predicting the participation rate in the third J

sample, underpredicting total trips at 5.53 with a change in water quality from Q to Q
ver- sus the actual number of trips at 6.27. for individual i on choice occasion t
For all of (10)
the model specifications the RMSEs are quite ΣΣ
pants, underpredicting total trips by over
40%. exp (Vijt [Q; þi ])
In contrast, all of the models perform poorly
in predicting nonparticipation. While the ac- which is the compensating variation for the
tual nonparticipation rate in the third sample standard logit model. The unconditional
is 41.8%, all five models predict com- pensating variation does not have a
nonparticipation rates of less than 5.6%, closed form, but it can be simulated by
ranging from a low of
(11)
0.3% in model C to a high of 5.6% in model R . Σ J exp (Vi j t [Q J ; þr ])Σ
.
C Vit = R − 1
A0. Indeed, as table 9 reveals, models A, B, p ln . i
and C predict average total trips of over nine 1 r =1 þ j
per year Σ =0

for individuals who actually take no trips. .J i

Mod- els A1 and A0 perform better, but still − exp (Vijt [Q; þr ])ΣΣ
miss the mark in characterizing participation j
ln
for this subsample. =0
The results suggest that participants andwhere R is the number of draws and r repre-
nonparticipants are perhaps more distinct sents a particular draw from its distribution.
than the standard mixed logit structure The simulation process involves drawing
allows for. Employing bounded distributions val- ues of þi, calculating the resulting
for the ran- dom parameters of the model compen- sating variation for each vector of
appears to al- leviate, but not eliminate, the
draws, and finally averaging over the results
problems in the overfitting of the mixed for many draws.
logit model. The average annual compensating varia-
tions and predicted changes in total annual
trips per household are calculated for three
Welfare Calculations water quality improvement scenarios using
the parameter estimates from the five
Given the random parameters þi, the condi- models re- ported in the previous section.
tional compensating variation (CV) The first sce- nario improves all 129 lakes to
associated the physical water quality of West Okoboji
Lake, the clear- est, least impacted lake in
the state. Table 10 compares the physical
Table 10. Average Water Quality Comparisons water quality of West Okoboji Lake with the
average of the other 128 lakes. All of West
Okoboji Lake’s measures are considerably
improved over the other 128. For
West Other 9 Target Non-impaired Impaired
Okoboji 128 Lakes Lakes Lakes (64) Lakes (65)
Secchi transparency (m) 5.67 1.14 1.23 1.27 0.81
Chlorophyll (ug/L) 2.63 41.23 40.13 23.25 56.67
Total nitrogen (mg/L) 0.86 2.21 3.64 1.11 2.31
Total phosphorus (ug/L) 21.28 106.31 91.11 58.79 139.91
Inorganic SS (mg/L) 1.00 9.49 9.52 3.51 14.78
Volatile SS (mg/L) 1.79 9.40 8.42 6.02 12.93
Cyanobacteria (mg/L) 22.00 297.98 948.00 42.96 516.96
Total phytoplankton (mg/L) 24.06 307.04 953.37 47.48 528.65
120 February 2009 Amer. J. Agr. Econ.

Table 11. Counterfactual Welfare Analysis

a. CV Per Iowa Household ($’s/year)
All 128 Lakes 9 Zone Lakes 65 Impaired Lakes
Improved to W. Okb. Improved to W. Okb. Improved to Median
Model Uncond. Cond. Uncond. Cond. Uncond. Cond.
Model A 246.98 153.04 40.25 19.56 15.27 11.52
(25.23) (3.92) (2.61)
Model B 203.55 122.08 33.02 17.15 7.02 6.12
(32.16) (4.84) (1.40)
Model C 226.61 108.12 33.44 14.84 11.61 5.92
(17.60) (2.52) (1.78)
Model A1 188.88 151.42 28.92 19.45 10.08 10.55
(24.31) (3.68) (2.23)
Model A0 67.30 91.48 7.71 11.32 2.99 9.62
(17.98) (2.61) (2.46)
b. Increase in Total Trips
Model Uncond. Cond. Uncond. Cond. Uncond. Cond.
Model A 4.13 3.71 0.68 0.52 0.22 0.29
(0.59) (0.11) (0.07)
Model B 3.49 2.91 0.57 0.43 0.10 0.14
(0.74) (0.13) (0.03)
Model C 3.71 1.44 0.56 0.20 0.17 0.07
(0.41) (0.06) (0.03)
Model A1 3.51 3.29 0.55 0.46 0.15 0.24
(0.48) (0.09) (0.05)
Model A0 1.78 2.28 0.22 0.28 0.06 0.25
(0.42) (0.07) (0.06)

example, West Okoboji Lake has slightly reported in table 11. We consider the welfare
over five times the water clarity, measured
by Sec- chi transparency, of the other lakes.
The sec- ond scenario is a less ambitious,
though more realistic, plan of improving
nine lakes to the water quality of West
Okoboji Lake (see table 10 for comparison).
The state is divided into nine zones with one
lake in each zone, allow- ing every Iowan to
be within a couple of hours of a lake with
superior water quality. The nine lakes were
chosen based on recommendations by the
Iowa Department of Natural Resources for
possible candidates of a clean-up project.
The third and final scenario is also a policy-
oriented improvement. Currently of the 129
lakes, 65 are officially listed on the EPA’s
im- paired waters list and by 2009 the plans
must be in place to improve the water
quality at these lakes enough to remove
them from the list. Therefore, in this
scenario, the 65 impaired lakes would be
improved to the median physi- cal water
quality levels of the 64 non-impaired lakes.
The last two columns of table 10 com- pares
the median values for the non-impaired
lakes to the averages of the impaired lakes.
The resulting compensating variations and
changes in total trips under each scenario are
results from model A1 to be the best results
for policy analysis. For comparison
purposes, we present welfare results from
models A, A0, B, and C as well. Finally, the
relatively poor per- formance of the various
models in terms of out- of-sample
predictions, particularly in terms of
nonparticipants, raises concerns about the
re- sulting welfare measures. One approach
to the problem, as suggested by a reviewer,
is to con- dition on the observed choices in
the sample (see, e.g., von Haefen, 2003).
Table 11 pro- vides both unconditional and
conditional wel- fare measures, though we
would recommend the conditional estimates
as more credible.20
We start by contrasting the unconditional
and conditional welfare and trip measures.
As expected, conditioning on the observed
us- age patterns typically reduces the
compensat- ing variations (by as much as
40%), though the qualitative pattern of the
results remains generally the same. This is
consistent with the fact that,
unconditionally, the models gener- ally
over-predict both participation and the

20
We use 1,000 draws in the unconditional and conditional
wel- fare simulations. Moreover, for the conditional welfare
simula- tions, we also report Krinsky and Robb (1986) standard
errors using 200 simulated conditional welfare estimates.
Egan et al. Valuing Water as a Function of Quality Measures 121

number of trips taken by households, which in 21

West Okoboji Lake, while currently one of the most popular
lakes in the state is far from most population centers in Iowa.
turn is likely to inflate the welfare implications
of changing site attributes. In addition, condi-
tioning virtually eliminates the distinction be-
tween models A and A1, where the latter re-
lies on triangular (rather than normal) distri-
butions for the site characteristics. This seems
reasonable as well, in that conditioning will
reign in some of the extreme behavior implied
by tail values of the normal random parame-
ters. Also, as discussed by von Haefen (2003),
conditioning lowers the range of welfare esti-
mates across the five models. For the first sce-
nario the range is reduced by 65.7%, with a
74.7% reduction for the second scenario, and
a larger 84.5% reduction for the third scenario.
The same is true for the range of the expected
trip increases from the three scenarios (table
11b).
Turning to the specific scenarios, we see
for the first scenario both a substantial
change in household welfare and in the
predicted num- bers of trips taken to lakes.
For model A1, the conditional CV is $151
per Iowa household. This size of a welfare
change seems reasonable given the numbers
of trips taken in a typical year under
baseline conditions. The travel cost savings
alone from having closer (high quality) sites
at which to recreate would justify much of
these welfare gains.21 Model A yields very
sim- ilar conditional welfare gains, with
models B, C, and A0 yielding smaller
welfare gains (ranging from $91 to $122 per
household). The total trip changes exhibit a
similar pattern, with models A1 predicting a
3.3 increase in trips per year, whereas model
A yields a larger increase and models B, C,
and A0 yield smaller increases.
Shifting to the second scenario, we find a
similar pattern in terms of how the five
models predict both welfare and total trip
changes un- der these scenarios. The simpler
nested logit- style model of A0 suggests
smaller welfare and trip changes when
compared with the more complex mixed
logit models. Finally, under the third
scenario, a slightly different pattern
emerges, with the conditional welfare and
pre- dicted trip changes being similar among
mod- els A, A1, and A0, and roughly half as
large with models B and C.
A possible explanation emerges by first
fo- cusing on model B and A1 across the
three sce- narios. Model B, with only Secchi
transparency included for water quality,
reasonably prox- ies for the inclusion of all
eight water qual-
ity measures when the water quality suggest that Iowans place higher value on
improve- ment is large. For the first improving a subset of the lakes to superior
scenario, with all 128 lakes improved to water quality rather than providing adequate
the level of West Okoboji, Model B’s levels at all of the lakes (i.e., sufficient to
conditional welfare value is only dif- not be listed
ferent from model A1 by 19.4% ($122.08
ver- sus $151.42). For the third scenario,
which is a moderate water quality
improvement, model B’s conditional
welfare value is 42% less than model A1
($6.12 versus $10.55).
Comparing model C and A1 across the
three scenarios shows a similar pattern,
though the conditional welfare estimates
from model C are always lower than from
model B and there- fore more different
from model A1. Assuming model A1’s
conditional welfare estimates are the most
credible, comparing model A1’s con-
ditional welfare estimates to models B
and C indicates that inclusion of more
water quality measures (model C versus
model B) does not necessarily lead to
better welfare estimates, as it is important
the functional form of the wa- ter quality
measures, where we tested logged versus
linear.
Another message from table 11,
however, is consistent across all of the
five specifications. That is, the benefits
from improving a few lakes to a level of
high water quality exceeds the ben- efits
from modest improvements to all the im-
paired lakes. The CV benefits from
Scenario 2 are typically two to three times
those from Sce- nario 3. The annual
compensating variation for Scenario 2 is
$19 for each Iowa household (us- ing
model A1). As expected, this estimate is
10% of the value if all lakes were
improved to the level of West Okoboji,
even though the scenario involves
improving only 7% of the lakes. This
suggests that to maximize Iowan’s benefit
from improving a few lakes, policymak-
ers should disperse the changes
throughout the state.

Conclusions

Data on lakes usage at 129 of Iowa’s

princi- pal lakes was combined with
extensive physi- cal water quality
measures from the same set of lakes
gathered by the Iowa State Univer- sity
Limnology Lab. Analysis based on the
repeated mixed logit framework shows
that individuals are directly responsive to
physi- cal water quality measures. Three
improve- ment scenarios were studied and
122 February 2009 Amer. J. Agr. Econ.

as impaired by the Environmental Protection [Received July 2006;

Agency). accepted April 2008.]
By estimating the partial effects of a list
of physical measures, we have determined
which measures significantly affect
recreationists’ be- havior. Limnologists and
water resource man- agers can use this
information about what physical lake
attributes visitors’ trip behav- ior responds
to in designing projects for water quality
improvements. Our results indicate wa- ter
clarity is very important as evidenced by the
Secchi transparency. Also, high concentrations
of nutrients, in general, are found to
decrease recreational trips.
The findings of this study also have direct
relevance for environmental protection man-
agers and citizens concerned with water
quality in that they can be used to prioritize
clean-up activities to generate the greatest
recreational benefits for a given expenditure.
Not only can the findings be used to
determine which lakes to target and in what
order to clean them but also the most
efficient levels of improvement can be
identified.
Finally, our investigations into the econo-
metric specification for the mixed logit
model suggest that the results (both
unconditional welfare results and out-of-
sample predictive ability) can be sensitive to
the choice of the underlying distributions
employed for the ran- dom parameters of the
model. In general, both the model fit and
predictive ability improved in the current
application when the random pa- rameters
were bounded (in our case triangu- lar),
rather than unbounded. For the welfare
results, however, the difference between the
model with unbounded versus bounded ran-
dom parameters was nearly eliminated by
us- ing conditional welfare estimates.
Moreover, the conditional welfare estimates
substantially lowered the spread of welfare
estimates across the five model
specifications. The main reason for this is
that all of the models were found to do
poorly in predicting the behavior of non-
participants in our third sample. The condi-
tional welfare estimates control for this by
in- corporating the individual’s actual
observed choices at baseline conditions.
However, at the estimation stage, additional
research is needed to distinguish and
characterize non- participants, perhaps by
allowing for less sym- metry in the random
parameters of the model or some discrete
distributions in the mixture models.
References pothetical Choices Reflect Actual Behavior?”
Adamowicz, W., J. Swait, P. Boxall, J. Louviere,
and
M. Williams. 1997. “Perceptions vs. Objective
Measures of Environmental Quality in Com-
bined Revealed and Stated Preference Models
of Environmental Valuation.” Journal of
Envi- ronmental Economics and
Management 32:65– 84.
Arbuckle, K.E., and J.A. Downing. 2001. “The
In- fluence of Watershed Land Use on Lake
N:P in a Predominantly Agricultural
Landscape.” Limnology and
Oceanography 46:970–75.
Atasoy, M., R. Palmquist, and D. Phaneuf. 2006.
“Estimating the Effects of Urban Residential
Development on Water Quality Using Micro
Data.” Journal of Environmental
Management 79:399–408.
Azevedo, C.D., K.J. Egan, J.A. Herriges, and C.L.
Kling. 2003. Iowa Lakes Valuation
Project: Summary and Findings from Year
One. Final Report to the Iowa Department of
Natural Re- sources, August.
Bockstael, N.E., W.M. Hanemann, and I.E. Strand.
1986. Measuring the Benefits of Water
Quality Improvements Using Recreation
Demand Mod- els. Report Presented to the
Environmental Protection Agency under
Cooperative Agree- ment CR-811043-01-0,
Washington, D.C.
Chen, H.Z., F. Lupi, and J.P. Hoehn. 1999. “An
Em- pirical Assessment of Multinomial Probit
and Logit Models for Recreation Demand.” In
Her- riges, J.A., and C.L. Kling, eds. Valuing
Recre- ation and the Environment:
Revealed Prefer- ence Methods in Theory
and Practice, Chel- tenham, UK: Edward
Elgar, pp. 141–61.
Creel, M.D., and J.B. Loomis. 1990. “Theoretical
and Empirical Advantages of Truncated Count
Data Estimators for Analysis of Deer Hunt-
ing in California.” American Journal of
Agri- cultural Economics 72(2):434–41.
——. 1992. “Modeling Hunting Demand in the
Pres- ence of a Bag Limit, with Tests of
Alternative Specifications.” Journal of
Environmental Eco- nomics and
Management 22(2):99–113.
Dillman, D.A. 1978. Mail and Telephone Surveys
— The Total Design Method. New York:
Wiley.
Egan, K.J., J.A. Herriges, C.L. Kling, and J.A.
Downing. 2008. “AJAE Appendix: Valuing
Water Quality as a Function of Water Qual-
ity Measures.” Unpublished manuscript.
Avail- able at http://agecon.lib.umn.edu,
accessed February 2008.
Greene, W.H. 2000. Econometric Analysis.
Upper Saddle River, NJ: Prentice Hall.
Haener, M., P. Boxall, and W. Adamowicz. 2001.
“Modeling Recreation Site Choice: Do Hy-
Egan et al. Valuing Water as a Function of Quality Measures 123

American Journal of Agricultural Environmen- tal and Resource Economics

Economics 34(1):91–115.
83:629–42.
Hensher, D., and W. Greene. 2003. “The Mixed
Logit Model: The State of Practice.” Trans-
portation 30:133–76.
Herriges, J., and D. Phaneuf. 2002. “Inducing Pat-
terns of Correlation and Substitution in Re-
peated Logit Models of Recreation Demand.”
American Journal of Agricultural
Economics 84:1076–90.
Hess, S., K.E. Train, and J.W. Polak. 2006. “On the
Use of a Modified Latin Hypercube Sampling
(MLHS) in the Estimation of a Mixed Logit
Model for Vehicle Choice.” Transportation
Re- search: Part B: Methodological
40(2):147–63.
Hilger, J. and M. Hanemann. 2006.
“Heterogeneous Preference for Water
Quality: A Finite Mixture Model of Beach
Recreation in Southern Cali- fornia.”
California Sea Grant College Program,
Econ06-01.
Iowa Department of Natural Resources. 2004a.
Fishing Guide for Iowa Lakes.
——. 2004b, “The Value of Water Quality in Iowa
Lakes,” Iowa Conservationist 63(2):7–25.
Jeon, Y., J. Herriges, C. Kling, and J. Downing.
2005. “The Role of Water Quality
Perceptions in Modeling Lake Recreation
Demand.” Iowa State University,
Department of Economics, Staff research
papers.
Kaoru, Y., V.K. Smith, and J. Long Liu. 1995.
“Using Random Utility Models to Estimate the
Recre- ational Value of Estuarine Resources.”
Ameri- can Journal of Agricultural
Economics 77:141– 51.
Krinsky, I., and L. Robb. 1986. “On Approximating
the Statistical Properties of Elasticities.” Re-
view of Economics and Statistics 68:715–19.
Leamer, E.E. 1983. “Let’s Take the Con Out of
Econometrics.” American Economic
Review 73(1):31–43.
McConnell, K. 1993. “Indirect Methods for Assess-
ing Natural Resource Damages under CER-
CLA.” In J. Kopp, and V.K. Smith, eds. Valu-
ing Natural Assets, pp. 203–37. Resources
for the Future, Washington DC.
McFadden, D., and K. E. Train. 2000. “Mixed MNL
Models for Discrete Response.” Journal of
Ap- plied Econometrics 15:447–70.
Morey, E.R., R.D. Rowe, and M. Watson. 1993.
“A Repeated Nested-Logit Model of Atlantic
Salmon Fishing.” American Journal of
Agricul- tural Economics 75:578–92.
Morey, E.R., J. Thacher, and W. Breffle. 2006.
“Using Angler Characteristics and Attitudi-
nal Data to Identify Environmental Preference
Classes: A Latent-Class Model.”
Parsons, G.R., E.C. Helm, and T. Bondelid. American Journal of Agricultural Economics
2003. “Measuring the Economic Benefits 87(4):1061– 76.
of Water Quality Improvements to
Recreational Users in Six Northeastern
States: An Application of the Random
Utility Maximization Model.” Working
paper, University of Delaware, July.
Parsons, G.R., and M.J. Kealy. 1992.
“Randomly Drawn Opportunity Sets in a
Random Utility Model of Lake
Recreation.” Land Economics 68(1):93–
106.
Phaneuf, D.J. 2002. “A Random Utility Model
for TMDLs: Estimating the Benefits Water
Qual- ity Improvements.” Water Resources
Research 38:471–95.
Phaneuf, D.J., C.L. Kling, and J.A. Herriges.
2000. “Estimation and Welfare Calculations
in a Generalized Corner Solution Model
with an Application to Recreation Demand.”
The Re- view of Economics and Statistics
82(1):83–92.
Pollak, R. A., and T. J. Wales. 1991. “The Likeli-
hood Dominance Criterion: A New
Approach to Model Selection.” Journal of
Econometrics 47(2–3):227–42.
Revelt, D., and K. Train. 1998. “Mixed Logit
with Repeated Choices: Households’
Choices of Appliance Efficiency Level.”
The Review of Economics and Statistics
80:647–57.
Ribaudo, M., and S. Piper. 1991. “Estimating
Changes in Recreational Fishing
Participation from National Water Quality
Policies.” Water Resources Research
27(7):1757–63.
Russell, C., and W. Vaughan. 1982. “The
National Recreational Fishing Benefits of
Water Pollu- tion Control.” Journal of
Environmental Eco- nomics and
Management 9(4):328–54.
Stevens, J. 1966. “Recreation Benefits from
Water Pollution Control.” Water Resources
Research 2(2):167–82.
Train, K. 1998. “Recreation Demand Models
with Taste Differences Over People.” Land
Eco- nomics 74(2):230–39.
U.S. Environmental Protection Agency (USEPA).
2000. “Nutrient Criteria Technical Guid-
ance Manual: Lakes and Reservoirs.” Of-
fice of Water, Office of Science and
Technol- ogy, Report EPA-822-B00-001,
Washington,
D.C.
Von Haefen, R. 2003. “Incorporating Observed
Choice into the Construction of Welfare
Mea- sures from Random Utility Models.”
Journal of Environmental Economics and
Management 45:145–65.
von Haefen, R.H., D.M. Massey, and W.L.
Adamowicz. 2005. “Serial Nonparticipation
in Repeated Discrete Choice Models.”