integrating property value and local recreation models to...

Integrating Property Value and Local RecreationModels to Value Ecosystem Services in

Urban Watersheds

Daniel J. Phaneuf, V. Kerry Smith, Raymond B. Palmquist, andJaren C. Pope

ABSTRACT. This paper outlines a new revealedpreference method to estimate the effects of changesin land use associated with residential developmenton water quality and the implied ecosystem services atthe watershed level. The analysis integrates datadescribing several types of behavior and uses hedonicproperty value and random utility models for localrecreation to consider the multiple impacts ofecosystem services on household well-being. Severalpolicy examples drawn from changes in Wake County,North Carolina, are used to demonstrate how spatialdifferences in residential development are reflected inthe model’s estimates of the economic costs ofdeterioration in watershed quality. (JEL Q51, Q57)

I. INTRODUCTION

One of the more important challengesfacing environmental policy analysts today

is the need to measure the gains or lossesresulting from changes in ecosystem servic-es.1 Land use changes are significant factorsin influencing the amount and quality of theservices provided by watershed-based eco-systems. Recently Hascic and Wu (2006)have undertaken a comprehensive analysisof the effects of land use as well as otherproxy measures for human activities on thewatersheds in the lower 48 states of theUnited States. They consider 2,100 water-sheds defined based on the USGS’s eight-digit hydraulic units. Based on a count ofthe water samples in each watershed withconcentrations for one or more of fourwater quality measures (phosphorus, am-monia, dissolved oxygen, and PH) exceed-ing national reference points, they foundthat transforming forests to urban land usesconsistently increases this index of deterio-ration in water quality. Moreover, reduc-tions in water quality increase the numberof species at risk using the same watershedframework. Their research provides a keymotivation for integrating the ecologicalimpacts of alternative land uses into anyevaluation of the private benefits and netsocial costs of transforming the landscape.

Land Economics N August 2008 N 84 (3): 361–381ISSN 0023-7639; E-ISSN 1543-8325

E 2008 by the Board of Regents of theUniversity of Wisconsin System

1 There are a number of activities supporting this. TheU.S. Environmental Protection Agency has established acommittee of the Science Advisory Board to considermethods for valuing the protection of ecological systemsand services. Related to this activity the Agency has alsodrafted an Ecological Benefits Assessment Strategic Plans.A recent National Academy of Sciences volume onvaluing ecosystem services (see Heal and committee2005) observed that: ‘‘The fundamental challenge ofvaluing ecosystem services lies in providing an explicitdescription and adequate assessment of the links betweenthe structures and functions of natural systems, thebenefits (i.e., goods and services) derived by humanity,and their subsequent values’’ (p. 2).

The authors are, respectively, associate professor,Department of Agricultural and Resource Economics,North Carolina State University and Director of CEN-REP; W. P. Carey Professor, Department of Economics,Arizona State University and Resource for the FutureUniversity Fellow; professor, Department of Economics,North Carolina State University; and assistant professor,Department of Agricultural and Applied Economics,Virginia Tech. Funding was provided by the US EPASTAR grant program through grant number # R-82950801 to the authors. Thanks are due Vic Adamowicz,Anna Alberini, Nancy Bockstael, and Trudy Cameron, aswell as participants in NBER 2005 Environment andPublic Economics Workshop and two referees forcomments on an earlier draft, to Melissa Brandt, MichaelDarden, Vincent McKeever, Eric McMillen, NatalieRockwell, and Bryan Stynes for assistance in collectingthe survey data used in this analysis, and to Kenny Pickleand Vicenta Ditto for assistance in preparing severaldrafts of this manuscript.

Unfortunately, the literature has beenunable to establish convincing linkagesbetween private locational choices and theamenities provided by healthy watersheds.This information is important becausepolicy analyses often need to considerwhether and how these non-market losseswill count in relation to the tangible privatebenefits from increased development. Inev-itably, this type of assessment defines animplicit tradeoff between urban develop-ment and ecosystem services. While thefinal policy choices are based on manyconsiderations, it is important from theperspective of efficient land use that one ofthese factors be the public cost of thedevelopment-related changes in land cover.

This paper develops a new revealed-preference method to address these needs.Our model integrates data describing severaldifferent types of behavior to allow themultiple impacts of ecosystem services to beconsidered simultaneously. Most benefitestimation methods focus on a single-choicemargin. They might consider the factorsrelevant to the selection of a house andneighborhood or the reasons for choosing arecreation site for a one-day trip, but notboth. In these contexts, ecosystem servicesbecome the attributes of each potentialchoice alternative. By contrast, policies oftenaffect the quality of local neighborhoods andlocal recreation sites simultaneously.

We use the concept of short-run andlong-run decisions to adapt a static modeland allow residential and recreationalchoices to be combined. In making adecision about a home, we hypothesize thatan individual considers, along with conven-tional housing characteristics, the availabil-ity and quality of local recreation sites forshort outings. The water quality of thesesites offers a way to describe how one aspectof ecosystem services contributes to thequality of recreation sites available tohomeowners. A quality-adjusted measureof the available recreation choices in aneighborhood is treated as an attribute ofthat neighborhood. Implementing this logicrequires data that connect homeowners’recreation behavior to their housing choic-

es. Moreover, to link each of these decisionsto land use/ecosystems service outcomesrequires a consistent geo-coded mapping ofthe economic and environmental data.Comprehensive data of this type are notgenerally available. They were assembledfor our analysis from a special purposesurvey of home buyers in Wake County,North Carolina, together with detailedrecords of their home sales and a widearray of public data and special reports onthe watersheds in this county. This type ofspatial integration was one of the recom-mendations of the Hascic and Wu analysis.

To illustrate the payoff, we selected twotypes of current policy questions associatedwith local land use decisions. For our firstexamples, we construct estimates fromrecent reports of the differences in popula-tion growth for selected communities inWake County. Our estimates suggest thesecosts are not confined to the communitiesexperiencing rapid expansion and theirimmediate spatial neighbors. The costsspread and the pattern depends on the scaleof the growth impulse, the pre-existingdensity of residential development, and thenature of the integration of hydrologicalunits within the watershed. The secondexample compares two types of benefits ofan urban lake: the proximity amenities ofliving close to a lake and the local recreationbenefits provided to the general populationin the area.

By integrating a spatially delineateddescription of ecosystem services with aconsistently matched description of house-hold choices, it was possible to identify theseparate ways these services contribute towell being. A recreation site such as a lakecontributes amenity services to the housesnear it. With public access, the lake alsocontributes to recreation use and this role isreflected in our index of accessible recrea-tion. Changes in its water quality affect thevalue of recreation access and, ultimately,the amenity value of proximity as well.2 Our

2 Our model does not at this point include the water-quality effects on the amenity services provided by lakesto nearby homes. Our logic could with adequate data beexpanded to reflect this effect as well.

362 Land Economics August 2008

model exploiting the information conveyedby considering multiple choice marginsallows us to distinguish these separateinfluences.

The remainder of the paper is organizedas follows. Section 2 provides a descriptionof Wake County, N.C. Section 3 reviewsliterature related to the role of waterresources for property value models. Sec-tion 4 describes the conceptual basis for ourapproach, and Section 5 summarizes thedata. Section 6 outlines our estimationmethods and results, Section 7 details ourchecks for the sensitivity of our estimates toimplementation decisions, and Section 8reviews the policy scenarios. The lastsection discusses the more general implica-tions of our findings.

II. STUDY AREA

Over the past decade, North Carolina’slandscape has changed rapidly. It wasranked fifth in land area developed duringthe middle 1990s according to statisticsfrom the Center on Urban and Metropol-itan Growth, and that pattern seems to haveaccelerated in the area around ResearchTriangle Park. For example our study area,Wake County (shown in Figure 1a), cur-rently has the second largest population inthe state. Population grew by 48% between1990 and 2000 and added another 15%between 2000 and 2004. This growthgenerates new housing units (4,208 buildingpermits were issued in 1990 compared with13,779 in 2000) and an increased proportionof land in impervious surface. There are 12political jurisdictions in the county. Fig-ure 1b shows their land area and location inrelation to major water bodies in the region.

A key consideration in developing aneconomic description of the role of theservices provided by watershed-based eco-systems for household location choicesconcerns the relationship between the spa-tial unit of measurement for watershedcharacteristics and that for economic anal-ysis. In our case, the two decisions relevantto the analysis involve the selection of ahome and the local recreation outings made

after a home is chosen. For the former,neighborhood attributes must be integratedinto the description of household choices.Distances to employment centers, shop-ping, schools, and so on are obvious partsof this characterization. It is more difficult,however, to measure directly neighborhoodattributes that are not reflected by distancemeasures. Often submarkets arise based onthese differences. The submarkets mayreflect different ethnic or demographicgroups present in areas, heterogeneity inhousing prices or characteristics (e.g., ma-ture older neighborhoods versus new sub-divisions), or density of development. Tocapture the influence of these factors weused the Board of Realtors’ MultipleListing Service (MLS) definitions for sub-markets. These 18 sub-areas are shown inFigure 1c. The MLS definitions were usedin the design of our sampling plan and inthe definition of our quality adjusted indexof recreation accessibility described below.

An important issue in implementing ourmodel arises in establishing a concordancebetween the political and economic defini-tions for spatial units and the hydrologicallayout of the county. This task would not bedifficult if there were one accepted defini-tion for each perspective—the economicand hydrological models. There is not, anda number of alternatives are possible foreach type of analysis. For example WakeCounty commissioned a comprehensivewatershed management plan in 2000, re-sponding to the rapid growth. The environ-mental engineering firm preparing thereport (CH2MHILL) refined its watersheddefinition, disaggregating the USGS hydro-logical units and identifying 80 hydrologiczones for the county. One might also use thewater bodies (lakes and rivers) in the countyand their surrounding areas as a frameworkthat might be more recognizable to house-holds. Similarly for the economic definitionthere are alternative ways to define the sub-markets.3 The final decisions about how to

3 The county has a unified school district so thisdistinction is not as relevant as it might be expected to bein other locations.

84(3) Phaneuf et al.: Valuing Urban Water Quality 363

FIGURE 1ECONOMIC, HYDROLOGICAL, AND JURISDICTIONAL DIVISIONS OF WAKE COUNTY,

NORTH CAROLINA


define the relationships linking economicand hydrologic areas require compromisesthat may influence conclusions. As a result,we evaluated the sensitivity of our findingsto some of these choices and discuss ourevaluations below.

III. RELATED LITERATURE

Property value models have been appliedto a large set of environmental problems.However, there have been fewer hedonicstudies of water pollution and its effects onwatershed services. This is probably be-cause water and water quality affect prop-erties in a variety of ways and it is not clearwhether these effects are fully reflected inthe standard hedonic framework. Waterquality can have a direct amenity effect onproperties, but only for properties that areadjacent or very close to the water body.This can enhance the services that would beconsidered as delivered by watershed eco-systems (see Boyd and Banzhaf 2007). It isoften difficult to find a sufficient number ofwaterfront properties and enough variationin water quality to conduct such a study.4 Itis also reasonable to expect that nearbywater bodies will affect properties that arenot immediately adjacent to them. Theyprovide areas for uses such as swimming,fishing, and taking walks. Water quality canaffect these local recreation opportunitiesand this in turn can influence propertyvalues.

The study of local recreation trips forshort outings is also a neglected area ofresearch; so it is not surprising that thisconnection seems to have been overlooked.Nonetheless, there is some existing litera-ture relevant to our analysis. There are ahandful of studies on the effect of water

quality on waterfront property values.5 Theearliest of these was done by ElizabethDavid (1968) for waterfront properties inWisconsin. David used three categories ofwater quality based on the opinions ofgovernment officials. For their study ofPennsylvania streams, Epp and Al-Ani(1979) used an objective measure, a dummyvariable for pH below 5.5, and a subjectivemeasure, a dummy variable for the per-ceived water quality problems. More re-cently Leggett and Bockstael (2000) used ameasure of fecal coliform in their carefullydesigned study of the coastal properties onthe Chesapeake Bay. Finally, waterfrontproperties on Maine lakes were studied inMichael, Boyle, and Bouchard (2000) andPoor et al. (2001). The latter study used anobjective measure, clarity measured bySecchi disk, and a subjective measure,perceived clarity from a survey.

A second set of related research concernsthe potential for double counting measuresof site specific amenities in travel costrecreation site demand and hedonic prop-erty value models. McConnell (1990) firstraised these questions within a conceptualmodel for a lake that is used for recreationand also contributes to the amenitiesenjoyed by homeowners in the vicinity ofthe lake. Considering the theoretical imple-mentation of both a two-stage hedonicmodel and a travel cost model, he concludesthat if the lake only has value for recreation,the two models yield identical results.However if the lake also has amenity valuefor the properties separate from the on siterecreation use, then the hedonic modelcaptures all benefits and the travel costmodel captures only part. Adding pollutionto his model as a gauge of reductions in lakequality (or its amenities) yields similarresults.

Parsons (1991) also raised the issue thatrecreation and residential location decisions

4 Some studies, such as Michael, Boyle, and Bouchard(2000), have selected lakes and the homes around them asthe unit of analysis. This strategy inevitably confronts thequestions of whether homes around different lakes arepart of the same market (and hence have the samehedonic price function) and whether the homes near lakesare a separate market from other homes selling in thesame areas.

5 We restrict our review to studies that were publishedin refereed journals and used market sales prices. Thereare some other studies in government reports and filingsin litigation. Many of these are discussed in Palmquistand Smith (2002). Other unpublished studies are cited inthe papers discussed here.


may be intertwined. If residential decisionsare partially based on recreation opportu-nities, then the price variable in a travel costmodel will be endogenous. Parsons propos-es several types of potential instrumentsthat could be used in a travel cost model toaddress the endogeneity problem. Hisempirical experiment with a simple travelcost model suggests that the bias due to theendogeneity may not be trivial. Randall(1994) also discusses the problems with theprice variable used in travel cost models,including the difficulty sorting out therecreation and residential location deci-sions. Our framework overcomes the issuesraised by Parsons and Randall. In theprocess it illustrates the implementationchallenges that may well explain whyMcConnell’s conclusions did not lead to alarger number of applications of a revisedhedonic model.

IV. METHOD

The types of recreation that have receivedthe most attention in the recreation demandliterature probably have little influence onresidential location choices. Most studieshave considered full-day or multiple-daytrips with significant travel costs for gettingto the sites. The recreation sites of interest inour analysis involve local recreation. Thatis, these are short trips to recreationdestinations with short on-site stays. Thetime-cost of accessing these local sites canvary substantially within an urban area. Asa result, local recreation opportunities mayinfluence a household’s choice of a residen-tial location and these decisions will bereflected in the housing values in differentparts of the city.

Our framework uses a distinction instatic models that is often intended toreflect differences in costs of adjustment.It classifies choices based on a prioriassumptions about which choices wouldbe more difficult and costly to change.This formulation is what usually underliesthe long-run/short-run terminology adopt-ed when static models are used to illus-trate how adjustment and adaptation can

influence the conclusions of comparativestatic analysis. The long-run/short-runformulation permits a conceptual divisionto be made between different margins ofchoice. In this context, a choice marginrefers to a resource allocation decisionmade by an individual or household thatleads to the acquisition of both a privategood and the services of a non-marketgood. Long-run choices involve the selec-tion of a neighborhood and the purchaseof housing. Short-run choices involve tripsto local recreation sites for short outings.The short-run choices are made condi-tional on the long-run residential locationdecision. It is this conditionality that isisolated through the specification of onechoice as long-run and the other as short-run. Once the location choice is made, ahousehold allocates remaining time andmoney resources to market goods, leisure,and recreation site usage. Since the loca-tion choice influences the choice set forshort-run decisions, it seems reasonable toexpect that when deciding on a residentiallocation the household considers theportfolio of amenities conveyed by eachlocation, including its accessibility torecreation and the quality of the accessiblerecreation opportunities.

To represent these decisions more for-mally, we assume a two-stage householddecision process. Preferences are a functionof recreation trips to local sites x(q), anumeraire good z, and housing services h(a,q) where a is a vector of housing attributesand q is a measure of environmental quality.The vector a is broadly defined to includeall structural, neighborhood, and spatialcharacteristics of the house, including dis-tances to the available recreation sites. Thepreference function is given by

U ~ U x qð Þ, h a, qð Þ, z, eð Þ, ½1�

where e is an error term denoting individual(or household) heterogeneity that is unob-served by the analyst.6 For short-run

6 Household and individual will be treated as synon-ymous.


decisions involving x(q) and z, h(a,q) istreated as quasi-fixed. If m* is the house-hold’s annual income and ph(a,q) is theannual cost of a unit of housing, the short-run decision problem is described by

maxx,z

U x qð Þ, z h a, qð Þjð Þ s:t: m ~ pxx qð Þz z, ½2�

where m 5 m* 2 ph(a,q) is income net ofhousing costs and px is the distance-dependent price of a recreation trip. Thefirst-order conditions for this problemimply solutions for the recreation demandand market goods conditional on thehousing choice. The conditional indirectutility function for the short run problem isdenoted by V 5 V(px, m, q, e).

Recreation demand models can be usedto recover an estimate of V which is thenused to compute the realized ex postbenefits from visits to the recreation sites.The Hicksian surplus from a pattern ofannual recreation that is selected with givenprices and quality conditions is implicitlydefined by

V px, m, q, eð Þ~ V pcx, m z CS, q, e

� �, ½3�

where pcx denotes the Hicksian choke price

for visits to the recreation site and CS iscompensating surplus. The change in com-pensating surplus due to a change in q isdefined by

V px, m, q0, e� �

~ V px, m z DCS, q1, e� �

, ½4�

where q0 and q1 denote the original andnew values for q, respectively. Equations[3] and [4] imply household-specific sur-plus measures CS(q,e) and DCS(q0,q1,e) thatare functionally dependent on both ob-served and unobserved household hetero-geneity.

To see how additional choice margins canbe incorporated into the analysis considerhow the value of trips at a given qualitylevel might influence other decisions. Equa-tion [3] provides a summary of the gainsdue to the household’s ability to take tripsto the recreation site. It is the ex post benefitfrom the access and quality conditions ofrecreation trips from the individual’s chosen

housing location. In making a residentialchoice, the framework assumes that at thetime of the decision households considerwhat these recreation benefits would be foreach possible neighborhood being consid-ered. If recreation opportunities and/orenvironmental quality vary spatially, eacharea provides different potential for recre-ation benefits. As a result, the expectedrecreation benefits can be interpreted as anattribute of the location. These expectedbenefits from recreation at a given residen-tial location are given by

ECS qð Þ~ E CS q, eð Þ½ �: ½5�

In this relationship, the expectation opera-tor is with respect to the factors associatedwith the heterogeneity across households inthe location.7 Equation [5] is not a house-hold-specific measure. It measures theaverage recreation benefits that would beexpected by choosing to live in one locationcompared to no access. Of course, inpractice the choice is based on eachneighborhood’s relative value, so the de-fault of no access becomes irrelevant. Theexpectation is across diverse householdsconditional on the level of q at each specificlocation. Using a long-run perspective, wehypothesize that this value would becapitalized into housing prices in equilibri-um.

The long-run component of our modelconsiders the housing choice. Householdsevaluate many attributes of a house and itslocation including how different spatiallocations convey different potential forlocal recreation outings. Since recreationdecisions are made conditional on thelocation choice, we replace x(q) in thepreference function with ECS(q) whenevaluating the housing choice stage. Thisfunction implies that for each value of qassociated with a location, the householdconsiders the future potential recreationalbenefits and conditions its choice on theseand the other features conveyed by that

7 In this expression, the expectation refers to e. Inpractice, we will average over both observed andunobserved sources of heterogeneity.


location. More formally, the objectivefunction in the long-run is

maxU ~ U ECS qð Þ, h a, qð Þ, z, eð Þ,

subject to: m� ~ ph a, qð Þz px~xx z z, ½6�

where ~xx is the optimized value of x given q.With the actual values of x and z assumed

selected conditional on location, the valueof h(a, q) is determined through the spatialchoice of a and q recognizing the connectionthrough ECS(q). The first-order conditionsin this case show that in choosing alocation, the household balances the bene-fits of a higher level of q with the marginalincrease in housing cost:

Lu

LECS: LECS

Lqz

Lu

Lh: Lh

Lq

Lu

Lz

~Lph

Lq: ½7�

At the margin, households choose a resi-dential location such that the marginal valueof increased ecosystem services from theexpected recreation activities plus the incre-mental benefits from the services that aredirectly available at the location are bal-anced against the marginal purchase price.

To implement this logic, assume an urbanwatershed is divided into J areas corre-sponding to well-defined real estate mar-kets. The total number of recreation sites inthe watershed is K, and the quality associ-ated with a given site k is qk. The spatiallayout of the landscape and existing ame-nity levels convey a similar portfolio ofaccess to recreation opportunities for eachresident in a market area j. Given observa-tions on visits to local sites by the residents ofthe area, it is possible to estimate a randomutility model of site choice. For simplicityassume the indirect utility for a visit byperson i to site k is linear and given by

Vik ~ ak z btik z dqk z eik,

i ~ 1, . . . , N, k ~ 1, . . . K , ½8�

where tik is a measure of the time cost forvisiting site k, (a, b, d) are parameters to be

estimated, and eik is a random error termdistributed type I extreme value. With thisdistribution for the error term estimation ofthe site choice model in [8] is straightfor-ward, and the expected utility per trip for aperson in the sample is

EVi ~ lnXK

k~1exp V̂Vik

� �h iz C, ½9�

where V̂ik is the predicted deterministiccomponent of utility for person i and C isa constant. Under the linear specificationfor utility the average compensating varia-tion per trip (expressed in units of time) foran individual is obtained by dividing theexpression in equation [9] by b̂.

The expected utility available to a personfrom the K recreation sites varies across thewatershed due to the variation in the timeneeded to access the sites. Closer proximityto sites of higher quality will, on average,convey a higher per trip utility level.However, equation [9] is an individual-specific reflection of the expected utilityavailable from the portfolio of sites andtheir attributes. It is not an appropriateindex for measuring the overall marketreflection of the potential to derive benefitsfrom recreational outings to local sites

Markets could be expected to take ac-count of information that buyers could learnas part of evaluating alternative potentialneighborhoods and that sellers would con-sider when marketing their homes againstpotential competitors. This assessmentwould not reflect individual heterogeneityin the recreation benefits. Instead, it wouldbe a composite index, such as an overallaverage, of the potential to generate recre-ation benefits for the residents of a neigh-borhood. In some respects the logic thatunderlies the definition of this index mightbe seen as parallel to what is used in definingprice indexes. This parallel offers somehelpful intuition but there are also importantdistinctions. The first follows from the factthat there are no mechanisms that assurethat these indexes describing the accessibilityof a particular home’s location to constantquality opportunities for local recreation willbe equalized, as we might expect for private


goods and services traded in the samemarket. The second arises because we arenot imposing conditions for consistency onthe index. For example, one consistencycondition holds when a price index times thecorresponding quantity index for an aggre-gate of a set of heterogeneous goods equalstotal expenditures on the aggregate good.8

The definition for the average that wepropose is given by

ECS j ~ N{1j

XNj

i~1lnXK

k~1exp V̂Vik

� �h i.b, ½10�

where Nj is the number of person-tripsoriginating from market area j. By averag-ing out all observed and unobserved house-hold heterogeneity [10] provides a ‘‘qualityadjusted recreation access index’’ for repre-senting the recreation opportunities con-veyed by a given neighborhood.

A focus on the aggregate is possiblebecause the sum of an independent set ofdecisions (across individuals or choiceoccasions for the same individual) yieldsthe same measure for the Hicksian value ofaccess as would be derived by assumingchoices arose from a representative individ-ual with a specific nonlinear preferencefunction (see Anderson, De Palma, andThisse 1988). Thus relying on the RUMlogic for our empirical model is consistentwith describing an individual’s recreationchoices as if they took place at the intensivemargin.9 This format is what underlies CS

in equation [3] and ECS in equation [5].Equally important, we can use this indirectutility function to describe the housingchoices for a representative consumer andthereby consistently motivate the hedonicanalysis as well (see Anderson, De Palma,and Thisse 1992).

It is possible to use the access index inconjunction with a hedonic model tomeasure how recreation opportunities arereflected ex ante in housing prices. Forexample, define the hedonic price equationfor home sales occurring in the watershedby the semi-log form,10

ln p ~ a0 zXs

l ~ 1

bla z c1ECSj z c2q dð Þz u,

½11�

where q(d ) is a function describing theneighborhood amenity effect of a resourceq. We hypothesize it will be related to someindex of the distance d to the site. Estima-tion of equation [11] allows separate iden-tification of the direct and indirect effectson housing prices of ecosystem services. Inspecial cases when the effects are small andlocalized, one can assume that the equilib-rium price schedule which is being describedby the hedonic price equation would notchange with a change in external factorsinfluencing these non-market services.Then, when the error term is assumed tobe due to unobserved characteristics, it ispossible to directly calculate the exactwelfare measure as a simple differencebetween equilibrium prices under baselineand changed environmental conditions(Palmquist 1992). For our two policyapplications discussed in Section 8 belowwe maintain these assumptions. The annualcompensating surplus measure for propertyi in sub-market j arising from a change in

8 For an early but comprehensive review of therelationship between economic and non-economic ap-proaches to the definition of index numbers see Allen(1975). See pp. 44–47 for a discussion of the various‘‘tests’’ for properties of index numbers.

10 Cropper, Deck, and McConnell’s (1988) simulationexperiments suggest that when the independent variablesin hedonic models are replaced with proxy variables orthe specifications are likely to be incomplete, simplerspecifications for the price function such as the semi-loghave superior properties based on estimates of themarginal willingness-to-pay.

9 This connection at the aggregate level is importantto our analysis because it implies we can describe thechoice of sites for local recreation within a discrete choiceframework, recognizing they are not likely to becoordinated over the course of a year. Rather, thedecision to visit a park to walk the dog depends onspecific circumstances at the time it is done. Nonetheless,when considered from the seasonal perspective theaggregate of these choices over individuals can beconsistently represented as if it were a marginal choice.Moreover, the linear form for Vik yields an indirect utilityfunction for the representative agent that contains thesame log sum relationship and has the same analyticalform for aggregate welfare measure (see Anderson, DePalma, and Thisse 1988, equation 8).


the recreation index due to a change in q isgiven by

CSi ~ h | p0i { p1

i

� �

~ h | p0i | 1 { exp {c1 ECS j q0ð Þ

��

{ ECS j q1ð Þ�

; ½12�

where h is an annualization factor, p0i is the

observed sale price for property i underbaseline conditions, and p1

i represents theprice (including the original regression errorterm) for property i under the new qualitylevel. In more general cases where thehedonic price schedule changes or some ofthese assumptions are not met, it is possibleto approximate and possibly bound will-ingness to pay measures for changes inwatershed quality.11

V. DATA

Our approach requires spatially explicitdata on home sales, recreation decisions byhomeowners, and water quality for thearea of our application, Wake County,North Carolina. As part of a largerproject examining water quality in thecounty and state, we constructed adatabase that integrates property salesdata obtained from the Wake CountyRevenue Department, survey data ob-tained from a sample of homeowners inthe county, and data from individualwater quality monitoring stations assem-bled from a variety of state, municipal,and private sources.

Records for approximately 100,000 pri-vate home sales occurring between 1992 and2000 are contained in our database. Ta-ble A-1 in the Appendix provides a descrip-

tion of a subset of the structural andneighborhood variables coded for eachproperty in the data. The level of resolutionin structural and other attributes is morecomplete than in most hedonic studies.Fulcher (2003) and Palmquist and Fulcher(2006) provide a detailed description andanalysis of these data, suggesting specifica-tions for the structural and neighborhoodvariables that we employ here as well. Weuse sales from 1998 and 1999, totalingmore than 26,000 observations.12 The meanprices for sales during these years are$179,221 and $184,574, respectively. Homeshave on average 1,933 square feet of heatedliving space, 2.5 bathrooms, are 11 yearsold, and sit on 0.46 acre lots.

The component of our database describ-ing household behavior was obtainedthrough a mail survey sent to Wake Countyhomeowners between May 2003 and Sep-tember 2003. The objective of the question-naire was to gather household specificinformation for a proportion of the homesrepresented in our sales data, allowing us tolink location decisions to other householdactivities, and the economic characteristicsof the residents. To select the sample weused the Wake County home sales recordsto identify a subset of owner-occupiedproperties. Properties were randomly se-lected subject to four filters. First, weexcluded properties that sold for less than$50,000. Second, the county was dividedinto four quadrants based on an aggrega-tion of the MLS zones described above.Nine thousand properties were drawn suchthat they were evenly distributed across the

11 Our formulation treats the hedonic price function’sparameter estimates as approximately equal to the truevalues. There are a number of other interpretations andstrategies for estimating the unobserved heterogeneityassociated with the hedonic function under baselineconditions. We evaluated several alternative predictionstrategies including methods that adjust for bias due tothe nonlinear form of the price function. Use of thesealternative prediction methods and associated welfaremeasures made no significant difference to the welfareeffects reported below.

12 We limited our analysis to these two years becauseduring 1999 Wake County undertook a reassessment ofproperty values for tax purposes. This process takes placein the county at seven-year intervals. It can result in largechanges in property taxes due to a realignment of theassessed values of older homes with their market values.As a result, following a reassessment sale prices of housesexperiencing these changes can be affected. Preliminaryestimates of hedonic models immediately after thereassessment confirmed this discrete change in the marketprice schedule. Our survey was designed to contacthouseholds who purchased homes in Wake County wherewe have the corresponding record of the housing sale. Asa result we restricted the sample to housing sales thatpreceded the reassessment period.


four aggregate real estate zones. Third, weincluded checks to assure that the hydro-logical division of the county was alsoreasonably well represented by the salesdata. The sampled properties were evaluat-ed to assure a sufficient number of obser-vations fell in each of the CH2MHill sub-hydrologic units. For each sub-hydrologicarea we determined if the initial draw of9,000 properties resulted in at least 20observations from the area. For the areasthat did not meet this criterion, we ran-domly selected additional observations toraise the number in each hydrological areato 20. For areas with an insufficient numberof sales, we simply selected all that met ourcriteria. Finally, property owners’ namesand addresses were verified using thecurrent Wake County property tax records.Only properties for which the sales recordfrom our hedonic database could be cross-linked to the currently listed owner wereincluded in the final sample. This resulted in7,554 matched names and addresses, eachof whom was sent a mail survey followingthe Dillman (1978) protocol with twomailings and a reminder postcard. We hada 32% response rate that provided slightlymore than 2,000 completed surveys.13

We collected two types of recreation databased on feedback from two focus groupsconducted as part of the design of thesurvey questionnaire: (1) information onstate-wide visits to lakes, streams, andcoastal areas; and (2) information on ‘‘localoutings.’’ We define local outings as shortexcursions to sites close to home involvingat most a few hours of combined on-site

and travel time. Our hypothesis is that thequality of environmental sites used foroutings close to home is more likely toinfluence property values than that for thesites at more distant locations. For ourrecreation analysis we therefore considerthe influence of watershed quality on visitsto sites within Wake County. Forty-eightwater recreation sites were identified in thecounty.14 Our survey provides informationon 1,187 respondents who reported makingtrips of this type between May and Novem-ber 2002. Records on over 14,000 local tripsare available for analysis, with each partic-ipating household taking an average 12trips.

The water quality component of thedatabase combines technical indicators ofambient water quality from 12 separatesources.15 For this application, we usereadings from monitoring stations locatedin Wake County and focus on chemicalmeasures. The specific variables used for theRUM analysis are total suspended solids

13 Our relatively low response rate led to concernsabout non-response bias. To investigate the degree towhich this may be present we compared the proportion ofcompleted surveys to socio-economic characteristics atthe census block group level using a grouped logit model.The results suggest people in blocks with a higherproportion of white residents, older homes, and recentarrivals to the area were slightly more likely to return thequestionnaire. There was no significant effect due tomedian income or median house value. Thus, while whitesmay be slightly over represented in our sample of homeowners, there is little evidence of systematic bias in thesample that would compromise our ability to gaugehomeowners’ use of water recreation resources.

14 Our survey included a map with a legend listingsites identified a priori by us, as well as space forrespondents to identify sites not included in the list. Waterrecreation was broadly defined to include both contactand near-shore uses of area resources.

15 Chemical monitoring data were obtained from theNorth Carolina Department of Environment and NaturalResources for both public and private monitoringnetworks. These include monthly readings from ambientmonitoring stations throughout the county between 1994and 2000 for 61 variables. Pollutant loadings from majorNPDES point sources were collected from electronic andpaper sources. Nine variables were collected from themonthly reports of these sources for the Neuse River.Four types of biological summaries are available. Singlesamples collected on benthic and aquatic habitat charac-teristics in August 2001 by CH2MHill for Wake Countyare summarized in our database. Periodic readings for thestate benthic communities were collected by NorthCarolina Division of Water Quality from 1982 to 2003,for Neuse River Basin sites and from 1983 to 2001, for theCape Fear River Basin. Additional data sources are asfollows. Chemical data for four variables describing waterquality for major lakes in the Neuse and Cape Fearwatersheds are available periodically from 1981 to 2002.The U.S. Geological Survey (USGS) also reportschemical and flow data for sites within the upper Neuseand Cape Fear basins monthly from 1989 to 2001. Allthese databases can be linked either through the latitudeand longitude of the sampling location or otheridentifying information to our various geographic areadefinitions.


(measured in lab nephelometric turbidityunits), total phosphorus (measured in mil-ligrams per liter), and ambient dissolvedoxygen (measured in milligrams per liter).

VI. ESTIMATION AND RESULTS

RUM Estimation

Estimation of a RUM model requiresspecification of the choice set, calculation ofthe implicit price of a visit to each site, andattachment of environmental quality char-acteristics to each site. There are twopossibilities for defining an object of choicein the choice set: (1) a specific point in spacebased on a named destination such as apark with a lake; or (2) a spatial unit definedby physical hydrology. The former is bettersuited for understanding visitor use andbenefits for a specific site. The latter ispreferred when the focus is more on waterquality conditions in a general area becausethe technical definition is designed torepresent the location where the waterquality readings are measured. Definingchoice alternatives based on watershedsaligns the spatial resolution in the choicemodel with the physical conditions that giverise to variation in water quality levels. Bydividing the study area into high-resolutionwatersheds, the full variability in waterquality across the landscape can be exploit-ed in the model, and choices are more likelyto reveal distance/quality tradeoffs. For thisreason, we focus our primary attention on awatershed-based choice set. Our evaluationof the sensitivity of the results considers theeffects of a site-specific choice set and isdiscussed separately.

We define the choice set to be the set ofhydrological (CH2MHill) zones describedin Section 2. Trips observed in the samplewere assigned to one of the 80 zones. TheRUM model’s objective is to explain thechoice of one of these zones for a local tripas a function of the implicit price andwatershed quality in the zone. We evaluatethe effects of water quality in two ways. Thefirst directly includes measures of waterquality in the specification of the condi-

tional indirect utility for each site. Thesecond includes alternative specific fixed-effects in the indirect utility function thatcapture all features of the sites (observableand unobservable) and estimates the effectsof water quality variables in a second-stageanalysis. Our primary results are based onthe first approach. Section 7 discussesalternative site definitions and measures ofthe effects of site characteristics as part ofrobustness checks for the model. Theresource cost of a visit is the round triptime needed to travel to the zone, calculatedbetween the person’s address and the centerof each hydrological zone. GIS softwarewas used to make these calculations for allsampled households and all zones. Theaverage round-trip travel time for tripsobserved in our sample is 35 minutes.

The CH2MHill study of the watersheds inWake County provided a qualitative expertassessment for each of the hydrological unitsincluded in our choice set. Nevertheless, dueto the limited variability in the outcome oftheir assessments, we focus on technicalmeasures of ambient water quality toparameterize our quality variables. Ourdatabase includes readings for total sus-pended solids (TSS), total phosphorous(TP), and dissolved oxygen (DO) taken inthe county after January, 1998. Monitoringstation readings were linked to a hydrolog-ical zone based on the location of the stationgenerating the reading. The empirical distri-bution of all available readings attached to ahydrological zone was then used to generatesummary measures. For phosphorous andsuspended solids the 90th percentile in eachzone is the summary measure, while fordissolved oxygen the 25th percentile is used.Table 1 contains a summary of thesemeasures across the 80 hydrological zones,as well as a summary of watershed size.16

16 To our knowledge, our database provides therichest characterization of urban water quality used inan economic model. Nonetheless the monitoring stationnetwork is not sufficiently dense to individually coverevery hydrological zone. Of the 80 zones in our model, 33are covered by a monitoring station, necessitatingaggregation to cover the remaining. Our aggregationstrategy followed the USGS definitions for increasing


Our empirical specification for the con-ditional indirect utility function for a visit toa site in hydrological zone k is

Vk ~ btimek z dsizek zc1TSSk z c2TPk

z c3DOk z ek, k ~ 1, . . . , 80, ½13�

where the error terms are distributed type Iextreme value. Estimates of the utilityfunction parameters are shown in Table 1.The estimates are statistically significantand consistent with prior expectations. Ingeneral, people visit areas of the county thatare closer to their home, have lower levels ofTP and TSS, and higher levels of DO.17

Accounting for the different units in whichthe pollutants are measured, under baselineconditions TP has the largest disutilityeffect relative to TSS and DO.

Hedonic Estimates

The RUM model provides the character-ization of preferences for recreation and

water quality that is used to constructour quality adjusted recreation index. TheMLS zones in the county serve to definethe neighborhoods for measuring the recre-ation access index. This index, definedin equation [10], provides a quality-adjusted measure of the average benefitsfrom a trip originating from each zone.Inclusion of the index in our hedonicspecification links a quality-adjusted mea-sure of the expected accessibility to localrecreation to the current choice of homelocation.

We use a non-linear function of distanceto the nearest water body to proxy theresidential amenity affect of water resourcesin the county. In conventional propertyvalue models, location on, or near a lake isassumed to convey amenity services to thehomeowners. To compute our distancemeasure we obtained a GIS shape filecontaining all lakes in the county. Thedistances between each house and all lakeswere calculated and the distance to thenearest lake was determined. Since theamenity effect is likely to diminish withgreater distance we use the function,

lakedist ~ max 1 { d=dmaxð Þ1=2, 0

h i, ½14�

to create the distance variable, where d isdistance in miles and dmax is the cutoffpoint, set to 1/2 mile for this application.The index is between zero and one and isconvex. Twenty-two percent of properties

TABLE 1

RUM MODEL SPECIFICATION AND ESTIMATED PARAMETERS

Variable Description Estimate (t-Value)

Travel time Time in minutes for round trip from respondent’s home to center of eachhydrological unit

20.0750 (2145.27)

Size Size of hydrological unit in square miles. Mean (std. dev.): 12.55 (9.36) 0.1027 (106.60)TSS Summary of total suspended solids. Calculated for each hydrological unit

as 90th-percentile of empirical distribution of readings. Mean (std.dev.): 65.36 (102.92)

20.0016 (221.03)

TP Summary of total phosphorous. Calculated for each hydrological unit as90th-percentile of empirical distribution of readings. Mean (std. dev.):0.846 (0.132)

21.914 (230.17)

DO Summary of dissolved oxygen. Calculated for each hydrological unit as25th-percentile of empirical distribution of readings. Mean (std. dev.):6.52 (0.857)

0.465 (43.79)

17 The chemical measures tend to proxy conditions atsites perceptible to individuals. For example, suspendedsolids tend to affect visible water clarity. Phosphorouslevels are predictors of algae blooms and water smell,while dissolved oxygen is a predictor for the health ofaquatic life.

watershed size (i.e., 14-, 11-, and 8-digit aggregate units).For each zone missing a monitoring station we assigned acovered zone located in the same 14-, 11-, or 8-digithydrological unit with preference given to the lowest levelof aggregation available.


in the sample are within a half mile of thenearest lake.

Estimates from the hedonic model forlake distance and the recreation index areshown in Table 2, with the remainingparameter estimates given in Table A-1 inthe Appendix. The results are consistentwith prior expectations. Both the qualityadjusted recreation access index and prox-imity to lakes capitalize into housing pricesas expected. Thus, changes in qualityadjusted access conditions (either throughcloser proximity or better water quality) canbe measured with the hedonic model. Thisfinding is consistent with our hypothesisthat the ecosystem services we associatedwith watersheds seem to influence prefer-ences in different ways. The results are alsosuggestive of the importance of consideringthese different channels for householdresponse and adjustment for benefits as-sessment.

VII. ROBUSTNESS CHECKS

Our evaluation of the sensitivity of ourresults to implementation decisions consid-ered a number of variations in the models.Our overall findings support the modelingstructure. It appears to be quite robust to awide range of alternative specificationchoices. Several aspects of this analysisdeserve further discussion. The first isrelated to Parsons’ (1991) discussion ofpossible price endogeneity in travel costmodels. Parsons raised the issue with atraditional travel cost model, where bias inthe estimate of the price coefficient can havea large impact on the welfare measures. Inour model, the welfare impacts are capturedin the hedonic estimation rather than therecreation demand. The RUM stage of our

model is only used to generate a quality-adjusted index of access to local recreation.Nonetheless, if the travel time parameter inthe RUM model were affected by endo-geneity, the index might be affected as well.To gauge the extent to which this may be aproblem in our model we estimated themodel using a sub-sample of people whowere new arrivals to the county at the timethey purchased their houses. New arrivals(who are less familiar with the county’sgeography than long-time residents) may beless likely to choose a location basedprimarily on access to recreation sites.Thus, the simultaneity issues suggested byParsons’ argument should be reduced inthis sub-sample. In spite of a much-reducedsample size we find estimates of the RUMparameters that are nearly identical to theirfull-sample counterparts. This finding couldbe because the new arrivals were equallyconcerned with recreation, or that theendogeneity issue is not significant. Eitherway, our index represents the choices buyersface in making their location decisions, andour welfare measures are based on thosechoices.

The second issue relates to the choice setused for the RUM model.18 There areseveral issues in structuring a definitionfor the choice set for the model of localrecreation decisions. One of the mostimportant concerns the implied spatiallinkage of water quality measures to eachchoice alternative. Our final model uses ahydrologic definition for choice alternativesbecause this formulation offered the most

TABLE 2

HEDONIC ENVIRONMENTAL VARIABLES AND ESTIMATED PARAMETERS

Variable Description Estimate (t-Value)

Lake distance index Proxy for amenity impact of lake proximity5max[12(d/dmax)K, 0], d 5 distance to thenearest lake, dmax 5 K mile

0.016 (3.23)

Index of recreation access Average value of ECS from RUM model calculatedfrom trips occurring in each MLS zone

0.0174 (6.41)

18 For a detailed discussion of a similar issue inmarket research, the framing of the consideration set forchoice models, see Horowitz and Louviere (1995).


authentic link between the spatial definitionof the site and quality. As we noted thereare 80 such choice alternatives in thecounty. Our sample only visited 23 of theselocations. The time costs of an outing andexisting quality attributes of the other 57alternatives may explain why they were notvisited. Alternatively, they may simply notbe part of homeowners’ consideration set.The former rationale suggests we shouldinclude them in the choice set for the modelwhile the latter suggests not.

A completely different logic would sug-gest that we use named sites and nothydrologic units as choice alternatives. Thisperspective seems more likely to correspondto the framework people use. It is the waythey reported their recreation outings in oursurvey. It has the disadvantage of providinga less reliable basis for linking the waterquality measures since the monitoringnetwork is defined based on the hydrologicsystem. It was not intended to characterizewater quality at specific lakes or riverreaches. As with the hydrologic definitionfor choice alternatives, only a subset of thesites was visited by our homeowners: 39 ofthe 48 named sites for local outings.

Finally there are two strategies weidentified as possible ways of reflecting thequality attributes associated with ecosystemservices for these alternatives—includingthe measures of water quality in the RUMspecification directly (as in equation [8]) orusing alternative specific constants in theRUM model and then, in a second stage,estimating how the fixed effects are influ-enced by the water quality measures. There

is not an unambiguous basis for decidingwhich was best. Our final results adoptthe most complete specification of choicealternatives with what we considered to bethe most reliable measure for the waterquality characterizations of each choicealternative.

Our robustness analysis considered sixalternative specifications for the RUMmodel: (1) three versions of the hydrologicalsite definition (all 80 zones, the 23 visitedzones, and a 23-zone model with fixedeffects and the second-stage model formeasuring the effects of site attributes);and (2) the analogous three versions for thenamed site framework. In each case themodels were used to construct our index ofthe quality-adjusted potential for localrecreation from each neighborhood. Ta-ble 3 summarizes the results for the alter-native specifications of our hedonic model.The estimates of the effects for otherstructural and neighborhood variables werenot influenced by which model was used toconstruct the recreation index. Three con-clusions follow from this comparison. First,the characterization of what is a choicealternative does not appear to influence therelevance of both the quality of ecosystemservices and local recreation to propertyvalues. Second, and equally important, theimplementation decisions linking quality tochoice alternatives whether hydrologic ornamed sites are also not important to thisconclusion. Finally, the role of quality andaccess remains influential to housing priceswhen we reduce the choice set to only thesites visited.

TABLE 3

COMPARISON OF HEDONIC ESTIMATES USING DIFFERENT RUM MODELS

RUM Model Use for Index

Hedonic Regression–Recreation Index Coefficient

Estimate t-Statistic

80 watershed site RUM (in paper) 0.0159 5.8523 watershed site RUM 0.01379 9.2523 watershed site RUM–fixed effects 0.01620 21.47

48 named site RUM 0.02514 16.0739 named site RUM 0.02831 17.5239 named site RUM–fixed effects 0.02249 26.03


Some caveats should be noted. Therelative size of the estimated parametersfrom the hedonic models across the variousformulations used to develop the index forquality-adjusted local recreation does nothave a direct interpretation. As we alter thechoice set and the choice alternative defini-tion, the marginal effect of a change inwater quality also changes. As a result, therealized marginal effect on the indexes willbe different. So a direct comparison of howthese modeling decisions influence marginalvalues of the hedonic model is not possible.Equally important, a comparison cannot bemade for a given policy scenario because thespatial structure of the recreation-choicemodel influences how the policy is translat-ed into numerical changes in each model.Thus, the same policies might appear tohave different implications due to thistranslation and not as a result of the effectsof the measurement and model implemen-tation issues.

A further qualification stems from thejudgments about which water quality mea-sures are important to site choices in thevarious RUM versions. These conclusionswere influenced by the definitions used forthe choice alternatives, choice set, andmodeling strategy (i.e., the use of fixedeffects). The variations in results could beinterpreted based on the co-linearity in thepollution measures and the limited degreesof freedom available for some modelingapproaches. Nonetheless, it would be amistake to conclude that all were equallydecisive in terms of households’ localrecreation decisions.19

Before turning to the policy analyses usedto illustrate the model in the next section,two interrelated qualifications should benoted. First, we do not attempt to compareour estimates with what would be derivedfrom applying a repeated RUM analysis ofthe same local recreation choices. To do soin an informative way would require someresolution of questions about the timehorizon connecting ex ante expectations

reflected in the hedonic capitalization withthe ex post, single season, orientation of theconventional recreation model. Our qualityadjusted index of accessibility is an instan-taneous measure of the availability of localrecreation. Second, our analysis assumesthe changes in quality or in site availabilityare not large enough to modify the hedonicprice equilibrium. We selected examplesthat seem likely to meet this condition.Larger changes affect whether existinghouseholds would want to alter theirchoices and affect the population inducedto enter or leave a region. This latter featurechanges the population that is considered asa potential visitor to recreation sites. Thus,the same equilibrium process influencingthe hedonic matching has implications forrecreation models. It is another reflection ofhow non-market factors can be involved inlimiting who is a market participant andtherefore the character of the marketequilibrium (see Phaneuf, Carbone, andHerriges 2007 for further discussion).

VIII. POLICY ANALYSES

As we noted at the outset, an importantmotivation for our model is to expand therange of environmental policies that can besubjected to a benefit-cost analysis. Themethodological limits imposed by conven-tional techniques are especially importantfor evaluating changes in ecological servicesrelated to watersheds. To illustrate how ourcomposite model can be used to addressthese issues we considered two types ofexamples. The first stems from the RaleighNews and Observer’s lead article on June 30,2005. It described the rapid growth ofseveral communities in Wake County.Betwee’n 2000 and 2004, the fastest growingtown in North Carolina was Morrisville (inwestern Wake County) with 122% growth.Five towns in Wake County grew by morethan one-third during those four years.Such growth can bring water qualityproblems. To illustrate this point, we adaptthe information in this news account toconsider how the relationship betweenwater quality and property values would

19 An appendix available from the authors reportsthese model estimates.


interact in our model.20 In related research(Atasoy, Palmquist, and Phaneuf 2006), wehave studied the effects of residentialconstruction and residential land use onwater quality using spatial econometricsand a micro-panel data set for WakeCounty.21 These results can be used topredict the effect on water quality due to therapid growth in some of these communities.The predicted changes in water quality canthen be used in our choice margins model toinfer changes in property values. It is, ofcourse, important to acknowledge that thisprocess considers only one effect of growth.

Morrisville on Crabtree Creek grew by6,376 people between 2000 and 2004 leadingto the 122% growth. Morrisville lies inhydrologic unit W15 (a larger spatialaggregate than the CH2MHill zones usedhere) in Atasoy, Palmquist, and PhaneufSince the data in that study ended in 1999,the new growth immediately follows thestudy period. The population growth wasconverted to new houses dividing by 2.3persons per residence. This prediction forthe increment to housing stock was thenconverted to new houses per square kilo-meter in the hydrologic unit to match thevariable definition used in the Atasoy,Palmquist, and Phaneuf water-quality mod-el. The spatial econometric results from thewater-quality model were then used topredict the change in total phosphorousfrom the change in housing stock. Theadditional houses by the end of 2004 arepredicted to increase total phosphorous inthe hydrologic unit by 30%. It would also betransported to downstream hydrologic

units, but the decay is fairly rapid. Therewould be a 2.1% increase in the nextdownstream hydrologic unit, and it be-comes negligible further from the originalchange.22 These predictions are based onthe housing stock once construction wascomplete. Because the rate of constructionin Morrisville between 2000 and 2004 wasfar outside the range of the data used inAtasoy, Palmquist, and Phaneuf we havenot predicted the effects of the constructionon water quality.

The predicted pollution changes werelinked to the CH2MHill zones lying in theaffected hydrological zones to analyze theestimated welfare effects of the changes.Column 2 of Table 4 shows the averageeffect on property values for each of the 19MLS zones included in the study. Thesewere calculated using the equation [12] andan interest rate of 5% (to annualize thevalues) for the predicted difference. As-suming the change we consider is suffi-ciently small that the hedonic price func-tion does not shift, the figures provide anestimate of the annual household-levelwelfare measure for the water qualityreduction. The welfare impacts vary spa-tially between the MLS zones. The largesteffects were about $16 and $19 per year inthe areas most directly affected and fell toan average of less than a dollar in the zonesthat were farther removed from the growth.The variability between these two extremesreflects the underlying hydrology andpopulation distribution in the county. Thelargest effects are found in the areas of theprojected growth, population centers nearthe growth areas, and the downstreamneighborhoods.

A second community that is rapidlygrowing is Wake Forest on Smith Creek,where 4,539 residents were added for 35%growth over the four years. Because of thespatial layout, the houses per squarekilometer are substantially lower for Wake

22 Atasoy, Palmquist, and Phaneuf (2006) also con-sidered TSS and found that the existing housing stock didnot affect TSS, although housing construction did have asignificant effect, as expected.

21 In Atasoy, Palmquist, and Phaneuf (2006) we usewater-quality monitoring data for total nitrogen, totalphosphorous, and total suspended solids over a five-yearperiod. We control for agricultural sources and pointsources of pollution. In-stream transport of pollutants iscaptured by a spatial lag. The variables of primaryinterest are new residential construction and the existinghousing stock.

20 Not all of the areas of rapid growth could be used inthese scenarios. The land in some of the communities wasin two or more drainage basins, and the reported growthcould not be allocated between the basins. Therefore, anyscenario would have been based on arbitrary allocationsof the growth. This was not true of the growth inMorrisville and Wake Forest, which is discussed below.


Forest. Most of the growth took place inhydrologic unit W4 in the Atasoy, Palm-quist, and Phaneuf study. The increase inhouses is predicted to cause a 3.5% increasein total phosphorous in the hydrologic unit.For this scenario the level of construction iscomparable to the range of our earlier data,so we can predict the effect of the actualconstruction of houses on the pollutantloadings as well. If the construction wasspread evenly over the four-year period,total phosphorous would increase by 4.5%.New construction also increases total sus-pended solids and would result in a 3%increase. Finally, in-stream transport wouldresult in a 0.56% increase in TP and a 0.1%increase in TSS in the next hydrologic unitdownstream.

Column 3 of Table 4 shows the predictedmean welfare effects of this Wake Forestscenario for the MLS zones in the county.As with the Morrisville scenario, there issubstantial variation in the welfare mea-sures across MLS zones. The size, however,is more than an order of magnitude smaller.MLS zones 14 and 21 contain the projectedgrowth and have the largest annual welfare

decreases of $1.11 and $.60 per year,respectively.

Our second example uses a different typeof scenario that focuses on the amenityeffects of a lake on nearby homes. Almostall the lakes in Wake County are manmade.We wanted to consider the impacts that thepresence of a lake has for county residents.These impacts take two forms: people withhouses close to the lake gain amenities fromits proximity, while all residents in thecounty potentially receive local recreationservices. Lake Lynn, in the northern part ofthe county, was selected as an example forour valuation exercise. The 54-acre lake hasboth public access and neighboring houses.The lake is large enough that it receivesvisits, but it is not one of the majorreservoirs in the county. It ranks 19th insize of among the 52 officially named lakesin Wake County. The houses close to thelake are fairly typical of those in oursample, with a mean sale price of $213,315and a range between $117,000 and $368,000in 1999 dollars.

We consider both the recreation lossesand the amenity consequences if Lake Lynn

TABLE 4

WELFARE BOUNDS FOR GROWTH SCENARIOSa

MLS ZoneMorrisville Growth

Scenario ($)Wake Forest Growth

Scenario ($)Lake Lynn Recreation

Scenario ($)Lake Lynn Amenity

Scenario

1 4.27 0.09 12.38 —2 4.49 0.08 15.81 2.133 1.02 0.05 2.81 —4 3.35 0.01 4.80 —5 15.73 0.02 3.69 —6 1.03 0.02 1.88 —7 3.25 0.20 8.32 —8 1.33 0.16 4.55 —9 1.37 0.00 1.47 —

10 18.58 0.00 1.86 —11 0.54 0.03 1.81 —12 0.16 0.07 0.53 —13 0.12 0.06 0.40 —14 0.30 1.11 1.05 —15 2.99 0.00 1.52 —16 0.36 0.01 0.45 —17 0.88 0.00 0.66 —18 0.47 0.00 0.88 —21 0.65 0.60 2.23 —

a Average dollars per year per homeowner.


were not available. The mean recreationlosses are reported in Column 4 of Table 4.These losses again vary throughout thecounty and the magnitude depends on theproximity of the MLS zone to the lake,ranging from a mean of about $16 per yearto less than a dollar. On the other hand, theloss in amenity value from having a housethat was within one-half mile of Lake Lynnis more localized. Column 5 of Table 4shows the mean effects of the loss. Since theamenity effects only occur within a one-halfmile of the lake, the average effect in theMLS zone is small, about $2, and there is noeffect in the other MLS zones. However, themean value obscures the magnitude of theimpacts on some individuals. The mean lossfor houses within one-half mile of the lake isabout $35 per year and the maximumamenity loss at the closest house wasestimated at $266 per year or about $5,000in value. By comparison, the individualmaximum annual loss of recreation valuewas $41.13 for Lake Lynn.

Thus, this example illustrates howurban lakes and streams generate benefitsthat accrue to a much larger population.Amenities arising from proximity to theresource itself can be much larger for theclosest houses, but these are limited innumber. This distinction is importantbecause the amenities that are usuallymeasured are related to the nearbyhomes. However, when recreation is alsoconsidered, it is apparent that the aggre-gate effects of local recreation on proper-ty values can dominate these amenityeffects.

IX. IMPLICATIONS

There is widespread policy interest inmeasuring the economic value of enhance-ments in ecosystem services. Addressingthe challenges posed by these policy needsrequires methods capable of incorporatingthe multiple, spatially delineated pathwaysthat these services take in influencingpeople. Past applications might lead areader to believe that revealed-preferencemethods were not up to the task. In this

paper, we employ three data sources anda new strategy for linking models toaddress this problem. This structure dem-onstrates how recreation site choices andproperty value data can be used togetherto gauge how the services provided by anurban watershed are capitalized intohousing values. We find evidence thatproximity to water resources, access torecreation sites, and the water quality atthese sites are positively related to prop-erty values. Our analysis exploits adistinction between two choice marginsto isolate these effects.

Developing the model required new datacollection designed to map housing marketsand neighborhoods into the different spatialdefinitions for the hydrological areas thatcomprise watersheds. Publicly availabledata do not have the resolution requiredto link housing choices, recreation siteselections, household characteristics, andwatershed attributes. As a result, new datacollection activities were required as well asefforts to establish consistency in thegeographic roles of economic and hydro-logic data. Using the resulting model, weillustrate how to evaluate the environmentalcosts of rapid suburban housing growth inWake County. Our policy examples suggestthere is substantial variation in the welfareimpacts of decreases in water quality, bothin magnitude and across space from thedifferent growth scenarios. They also sug-gest that both the local recreation value andthe amenity value of urban watersheds aresignificant.

At this stage, our analysis provides aproof of concept that consistent use ofspatial linkages, together with an assumedhierarchy in individual choices, can helpto recover separate roles for several of theenvironmental services we associate withurban watersheds. Our estimates for theincremental environmental costs of resi-dential growth to existing homeowners areplausible. Nonetheless, these resultsshould be interpreted as a first step inthe process of enhancing the spatialdimensionality of non-market valuationmethods.


APPENDIX

TABLE A-1

REMAINING VARIABLES FROM HEDONIC ESTIMATION MODEL

Variable Description Estimate (t-Statistic)

age Age of structure, calculated as sale year-year built 20.001 (218.12)baths Number of bathrooms 0.0135 (9.51)acreage Lot size in acres 0.0469 (41.37)regheatarea Main heated living area in square feet 0.0002 (136.52)detgarage Dummy variable indicating presence of detached garage 0.0543 (13.39)fireplaces Number of fireplaces 0.0614 (29.55)deck Deck area in square feet 0.0001 (19.34)floordum1 Dummy variable indicating presence of hardwood floors 20.0099 (23.26)scrporch Screened porch area in square feet 0.0002 (18.12)atticheat Attic heated area in square feet 0.0001 (32.69)bsmtheat Basement heated area in square feet 0.00007 (16.3)garage Garage area in square feet 0.0002 (57.03)poolres Dummy variable indicating presence of residential swimming

pool0.0309 (4.4)

bsmtdum1 Dummy variable indicating presence of full basement 0.0973 (26.2)bsmtdum2 Dummy variable indicating presence of partial basement 0.0928 (26.92)walldum1 Dummy variable indicating presence of brick walls 0.012 (5.49)yrdum99 Dummy variable for 1999 sale 0.026 (19.62)encporch Enclosed porch area in square feet 0.0002 (10.29)opnporch Open porch area in square feet 0.0001 (16.83)condadum Dummy variable indicating house is of condition A (highest) 0.0948 (27.08)condcdum Dummy variable indicating house is of condition C 20.1054 (224.49)condddum Dummy variable indicating house is of condition D 20.2388 (218.26)commute 2000 census block mean commute time 0.0018 (15.48)grade Numeric grade assessed by revenue department 0.0063 (122.28)value Median house value for 2000 census block group 8.00E-07 (49.58)percent Percentage owner occupied housing in 2000 census block

group20.0001 (23.94)

taxrate Property tax rate per $100 in value 0.0226 (6.33)Constant 10.332 (913.56)

R-squared 0.926Dep. variable ln(price)Number of

observations26,305


References

Allen, R. G. D. 1975. Index Numbers in Theory andPractice. Chicago: Aldine Publishing Company.

Anderson, Simon P., Andre de Palma, andJacques-Francois Thisse. 1988. ‘‘A Representa-tive Consumer Theory of the Logit Model.’’International Economic Review 29 (3): 461–66.

———. 1992. Discrete Choice Theory of ProductDifferentiation. Cambridge, Mass.: MIT Press.

Atasoy, Mary, Raymond B. Palmquist, and DanielJ. Phaneuf. 2006. ‘‘Estimating the Effects ofUrban Residential Development on WaterQuality Using Micro Data.’’ Journal of Envi-ronmental Management 79 (4): 399–408.

Boyd, J., and H. S. Banzhaf. 2007. ‘‘What AreEcosystem Services? The Need for StandardizedAccounting Units.’’ Ecological Economics 63(2–3): 616–26.

CH2MHill Inc. 2003. ‘‘Wake County Final Plan,’’Wake County Comprehensive Watershed Man-agement Plan. Charlotte, N.C. February.(This document and other related documentscan be found at http://projects.ch2m.com/WakeCounty).

Cropper, Maureen L., Leland B. Deck, andKenneth E. McConnell. 1988. ‘‘On The Choiceof Functional Form for Hedonic Price Func-tions.’’ Review of Economics and Statistics 70(Nov.): 668–75.

David, E. L. 1968. ‘‘Lakeshore Property Values: AGuide to Public Investment in Recreation.’’Water Resources Research 4:697–707.

Dillman, Don A. 1978. Mail and TelephoneSurveys: The Total Design Method. New York:John Wiley & Sons.

Epp, D., and K. S. Al-Ani. 1979. ‘‘The Effect ofWater Quality on Rural Non-Farm ResidentialProperty Values.’’ American Journal of Agricul-tural Economics 61 (3): 529–34.

Fulcher, Charles M. 2003. Spatial Aggregation andPrediction in the Hedonic Model. PhD. diss.,North Carolina State University.

Hascic, Ivan, and Jun Jie Wu. 2006. ‘‘Land Useand Watershed Health in the United States.’’Land Economics 82 (May): 214–239.

Heal, G. M., and committee. 2005. ValuingEcosystem Services: Towards Better Environ-mental Decision Making. Washington, D.C.:The National Academies Press.

Horowitz, J., and J. Louviere. 1995. ‘‘What is theRole of Consideration Sets in Choice Model-ing?’’ International Journal of Research inMarketing 12 (1): 39–54.

Leggett, Christopher G., and Nancy E. Bockstael.2000. ‘‘Evidence of the Effects of Water Qualityon Residential Land Prices.’’ Journal of Envi-ronmental Economics and Management 39 (2):121–44.

McConnell, K. E. 1990. ‘‘Double Counting inHedonic and Travel Cost Models.’’ LandEconomics 66 (2): 121–27.

Michael, Holly J., Kevin J. Boyle, and RoyBouchard. 2000. ‘‘Does the Measurement ofEnvironmental Quality Affect Implicit PricesEstimated from Hedonic Models?’’ Land Eco-nomics 76 (May): 283–98.

Palmquist, Raymond B. 1992. ‘‘Valuing LocalizedExternalities.’’ Journal of Urban Economics 31(1): 59–68.

Palmquist, Raymond B., and Charles M. Fulcher.2006. ‘‘The Economic Valuation of ShorelineTwenty-Five Years Later.’’ In Explorations inEnvironmental and Natural Economics: Essaysin Honor of Gardner M. Brown, Jr., ed. RobertHalvorsen and David F. Layton, 208–23.Northampton, Mass.: Edward Elgar.

Palmquist, Raymond B., and V. Kerry Smith.2002. ‘‘The Use of Hedonic Property ValueTechniques for Policy and Litigation.’’ InInternational Yearbook of Environmental andResource Economics, ed. Tom Tietenberg andHenk Folmer, 115–64. Northampton, Mass.:Edward Elgar.

Parsons, George R. 1991. ‘‘A Note on Choice ofResidential Location in Travel Cost Models.’’Land Economics 67 (Aug.): 360–64.

Phaneuf, Daniel J., Jared C. Carbone, and Joseph A.Herriges. 2007. ‘‘Non Price Equilibrium for Non-Marketed Goods.’’ Resources for the FutureDiscussion Paper No. 07-25-REV. Wash-ington, D.C.: Resources for the Future.

Poor, P. Joan, Kevin J. Boyle, Laura O. Taylor,and Roy Bouchard. 2001. ‘‘Objective versusSubjective Measures of Water Clarity in He-donic Property Value Models.’’ Land Econom-ics 77 (Nov.): 482–93.

Randall, Alan. 1994. ‘‘A Difficulty with the TravelCost Method.’’ Land Economics 70 (Feb.):88–96.


integrating property value and local recreation models to...

Documents