Valuing preferences over stormwater management outcomes including improved hydrologic function


  • Catalina Londoño Cadavid,

    1. Department of Agricultural and Consumer Economics, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA
    2. Escuela de Ingeniería de Antioquia, Antioquia, Colombia
    Search for more papers by this author
  • Amy W. Ando

    Corresponding author
    1. Department of Agricultural and Consumer Economics, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA
    • Corresponding author: A. W. Ando, Department of Agricultural and Consumer Economics, University of Illinois at Urbana-Champaign, 326 Mumford Hall, MC-710, 1301 W. Gregory Dr., Urbana, IL 61801-3605, USA. (

    Search for more papers by this author


[1] Stormwater runoff causes environmental problems such as flooding, soil erosion, and water pollution. Conventional stormwater management has focused primarily on flood reduction, while a new generation of decentralized stormwater solutions yields ancillary benefits such as healthier aquatic habitat, improved surface water quality, and increased water table recharge. Previous research has estimated values for flood reduction from stormwater management, but no estimates exist for the willingness to pay (WTP) for some of the other environmental benefits of alternative approaches to stormwater control. This paper uses a choice experiment survey of households in Champaign-Urbana, Illinois, to estimate the values of several attributes of stormwater management outcomes. We analyzed data from 131 surveyed households in randomly selected neighborhoods. We find that people value reduced basement flooding more than reductions in yard or street flooding, but WTP for basement flood reduction in the area only exists if individuals are currently experiencing significant flooding themselves. Citizens value both improved water quality and improved hydrologic function and aquatic habitat from runoff reduction. Thus, widespread investment in low impact development stormwater solutions could have very large total benefits, and stormwater managers should be wary of policies and infrastructure plans that reduce flooding at the expense of water quality and aquatic habitat.

1. Introduction

[2] Urban stormwater runoff causes many environmental problems. Conventional stormwater management has been designed primarily to reduce floods. However, a new generation of decentralized stormwater solutions can produce important ancillary environmental benefits. Previous research has estimated values for surface water quality [Carson and Mitchell, 1993; Van Houtven et al., 2007; Johnston et al., 2005] and for flood reduction from stormwater management [Bin and Polasky, 2004; Zhai et al., 2006, 2007], but no estimates exist for the values of some of the other environmental benefits of alternative approaches to stormwater control. This paper fills that gap by using a choice experiment survey of households in Champaign-Urbana, Illinois, to estimate the values of multiple attributes of stormwater management outcomes. This work adds to the valuation literature by exploring the combined effects of heterogeneous status quo situations and state-dependent preferences on total willingness to pay (WTP) for a public good that has variable benefit levels across space and trade-offs between attributes.

[3] Urbanization causes environmental problems by interfering with hydrological cycles. Roads and buildings create impervious surfaces which limit water infiltration and increase stormwater runoff during storms. Runoff contributes to flooding and water pollution, and hydrologists have pointed out further that diminished infiltration starves streams of groundwater that supports base flows during dry periods [National Research Council, 2009; Zhang and Schilling, 2006]. Historically, urban stormwater has been controlled primarily with large-scale engineering solutions that convey the water directly to streams, rivers, and detention ponds. These technologies, however, make stream flows excessively fast and heavy during storms, scouring stream beds, and further degrading aquatic habitat in urban water bodies [Brabec, 2009; National Research Council, 2009].

[4] New strategies now exist for mitigating stormwater runoff. Such low impact development (LID) tools include elements such as bioswales, pervious pavement, cisterns, and green roofs [U.S. Environmental Protection Agency, 2000] which capture, temporarily store, and infiltrate or evapotranspirate stormwater. The results can include better water quality, increased water table recharge, and healthier aquatic habitat [National Research Council, 2009]. The U.S. Environmental Protection Agency (EPA) is considering new regulations [U.S. Environmental Protection Agency, 2007] that might require developers to ensure that new development and significant redevelopment manages a significant amount of rainfall on-site; this would effectively require widespread implementation of LID development approaches, but the total benefits of that change are hard to estimate given gaps in the literature on the benefits of stormwater management.

[5] This paper is the first to present joint estimates of the monetary values of flood frequency reductions and environmental improvements from stormwater management. The results help to understand which type of flooding people care about most and consequently should be prioritized in terms of management in urban areas. This paper also measures the relative importance of ecological benefits in consumer WTP for stormwater management projects, which can help federal and local policy makers to evaluate the benefits of new stormwater regulations that implement LID techniques.

[6] Our work also contributes to research regarding the importance of current conditions to the value of environmental improvements. Flood frequency belongs to a category of environmental disamenities for which households in a single community might experience highly variable status quo conditions. Thus, a single intervention in the environment can have variable welfare effects on households depending on the conditions they currently experience. We test for state-dependent preferences in our analysis and use those results explore the implications of those preferences for the total welfare effects of policies with effects on multiple attributes of the environment.

2. Related Economic Literature

[7] One of the main negative consequences of stormwater runoff is flooding. Some research has estimated monetary values for the costs of flooding. For example, the effects of flooding on housing prices have been evaluated using hedonic property price functions [Bin and Polasky, 2004; Harrison et al., 2001]. Those studies find that houses located in flood-prone areas have a 4%–12% lower market value than equivalent houses located in a zone without flood risk. However, hedonic price methods can only measure some elements of the benefits of stormwater management because they cannot capture the value to individuals of changes in environmental services with indirect benefits far from their homes such as improvement of water quality, habitat for aquatic species, and water-table recharge [Birol et al., 2006; Novotny et al., 2001].

[8] Another line of work has estimated the value of surface water quality in rivers and streams [e.g., Carson and Mitchell, 1993; Van Houtven et al., 2007; Johnston et al., 2005; Whitehead, 2006] yet only a little of that research has been in urban or urbanizing areas [Bateman et al., 2006]. Some research has studied the values people place on dimensions of environmental quality in freshwater systems that are more complex than pollution levels [Loomis et al., 2000; Wilson and Carpenter, 1999]. However, research directly related to stormwater management outcomes is extremely limited. Several studies on the subject of stormwater have examined attitudes and behavior toward stormwater pollution [Dietz et al., 2004; Jorgensen and Syme, 2000] but do not quantify monetary values of the outcomes of stormwater management. Clark et al. [2002] try to investigate the relative importance of flood-control and ecological restoration objectives in watershed management practices, but that research was unable to identify separate values for the two types of stormwater management outcomes due to limitations on sample size and survey design.

[9] Research in behavioral economics implies that individuals in a community might have preferences over a given level of an environmental quality attribute (such as flooding or water pollution) that vary if they have heterogeneous experiences of that attribute. Several valuation studies have found evidence of state-dependent preferences in the environmental arena. Tait et al. [2012] and Moore et al. [2011] document state-dependent preferences for water quality when the status quo varies across space, but the importance of this feature of consumer preferences has not been explored in settings that involve trade-offs between elements of environmental quality.

3. Methods

3.1. Choice Experiment Technique

[10] Economists use many methods to estimate the values of environmental goods or disamenities. One set of methods is stated preference approaches, so called because they estimate values by describing a hypothetical environmental good or scenario to people and then asking those people to state in some way what they would be willing to pay to have (or to avoid) it. Two advantages of stated preference methods are: (1) they can be used to estimate the benefits of changes in environmental quality that are entirely hypothetical and cannot be observed in a real data set, such as improvements in urban hydrology that might arise in the future from widespread adoption of LID, and (2) they can measure what economists call nonuse values (values someone has for helping the environment even they will not benefit directly from it) as well as use values (things like health or recreational benefits a person enjoys directly).

[11] This paper applies a particular kind of stated preference tool, choice experiment (CE), or conjoint analysis, valuation methodology, to evaluate people's WTP for several elements of stormwater management outcomes. CE methodology has become increasingly widespread, with applications to areas ranging from valuation of environmental goods to consumer product marketing [Alriksson and Öberg, 2008; Hoyos, 2010]. CE methods can estimate total economic values for an environmental good that is comprised of a set of attributes that can be varied independently of one another [Holmes and Adamowicz, 2003; Hoyos, 2010]. In the process, CEs yield estimates of the value of each of a set of attributes of a good individually. Thus, the results can estimate the values of multiple alternative scenarios in which one or more attributes are varied simultaneously [Adamowicz et al., 1998; Alriksson and Öberg, 2008].

[12] We use CE methods for several reasons. First, the outcomes of a stormwater management strategy, basement flooding, street flooding, backyard flooding, water quality, cost, and infiltration/aquatic habitat quality, can actually vary in different directions from one another depending on the nature of a city's strategy. For example, basement flooding can be mitigated by temporarily flooding streets as miniature detention basins; all flooding can be reduced with large sewer infrastructure, but that can worsen water quality. Thus, it makes sense to describe a scenario of stormwater management outcomes in which the attributes vary separately from each other. Second, given this feature of stormwater runoff management, CE methods provide valuable information to policy makers by determining which type of flooding and features of environmental quality people care about the most and consequently should be prioritized. Third, CE analysis can readily be adapted to test for state-dependent preferences because it presents respondents with multiple scenarios and attribute values at multiple levels.

[13] Here we summarize the intuition, theory, and practice of CE methodology following Louviere et al. [2000] and Holmes and Adamowicz [2003]. In a typical CE survey, respondents are asked to answer multiple questions in which they compare and choose between two or more alternative designs of an environmental scenario to be valued. Each alternative consists of a set of attributes than can be quantitative or qualitative. Often the respondent is given the option of choosing a “status quo” alternative, a situation without any change in the scenario away from current conditions [Hoyos, 2010]. The researcher chooses several levels of each attribute that can appear in a given choice scenario. Observing choices between scenarios that have varied levels of attributes permits the researcher to quantify an individual's willingness to substitute between attributes. A person's WTP is calculated by modeling the influence of attributes on the probability that the person chooses one scenario over the others. As long as monetary cost is included as an attribute, one can estimate the marginal value of each of the noncost attributes, and one can calculate total WTP to move from the status quo to another alternative with its own set of attribute values [Louviere et al., 2000; Meyerhoff et al., 2009].

[14] The conceptual economic framework for CE analysis lies in Lancaster's [1966] theory of demand, which assumes that an individual benefits from the features of a good rather than the good itself. CEs are based on random utility maximization (RUM) theory [Louviere et al., 2000] where the different attributes contribute to a person's well-being (utility) together with a random component to capture the unobserved differences.

[15] In this framework, the utility associated with choice j, Uj, is comprised of a certain (vj) and a stochastic (εj) element:

display math(1)

where xj is a vector of noncost attributes, pj is the monetary cost of choice j, and β is a vector of preference parameters. Uj is indirect utility; it is unobservable to the econometrician, should decline with undesirable characteristics such as a higher frequency of flooding, and should increase with desirable characteristics such as higher environmental quality. In accordance with neoclassical economic theory of consumer behavior, individuals are assumed to pick the alternative that gives them the highest utility, i.e., the individual chooses an alternative j over l if and only if Uj > Ul. Choice is deterministic from the standpoint of the individual but stochastic from the point of view of the researcher; the random error term εj reflects the researcher's uncertainty about the utility the individual obtains from a given option.

[16] Usually a linear functional form is assumed for the utility function [Pendleton and Mendelsohn, 2000]. Then, for K attributes,

display math(2)

[17] By differentiating (2) with respect to each of the attributes, we see that the β parameters represent marginal utilities of noncost attributes inline image and −βp =  ∂U/∂p captures the marginal utility of money because an increase in the cost of the hypothetical project directly decreases the amount of income the respondent has available to spend on other things. The ratio between any two parameter estimates is the marginal rate at which a respondent can substitute between attributes k and m while holding utility constant (MRSkm = βk/βm). Marginal WTP for attribute k is given by −βk/βp [Louviere et al., 2000; Holmes and Adamowicz, 2003]. Total WTP for a change between two scenarios (xj0 to xj1) is given by

display math(3)

3.2. Econometric Methods

[18] We employ three econometric methods to estimate parameters from our choice experiment data. In this section we explain the standard conditional logit (CL) approach in detail and describe the two other approaches we use: the mixed multinomial logit (MMNL) which controls for unobserved heterogeneity and a weighted conditional logit (WCL) which controls for possible nonresponse bias.

[19] Following equation (1), the probability of observing the outcome in which an individual chooses alternative l in choice set C can be written as

display math(4)

[20] A typical assumption in econometric implementations of RUM models is that errors are independently and identically distributed (IID) with a type I extreme value distribution [Holmes and Adamowicz, 2003; McFadden and Train, 2000; Pendleton and Mendelsohn, 2000]. This leads to the CL; this is the most common model for analyzing choice data because has a simple and closed form for probabilities.

[21] In this paper, we used the MMNL in addition to the standard CL because the MMNL models preference heterogeneity and has the capacity to deal with the fact that every respondent answers several choice questions, individuals are likely to have unobserved preference heterogeneity, and one person's choice question responses are likely to be correlated with one another. Details on this methodology can be found in Louviere et al. [2000]. For the MMNL, the utility of individual i choosing alternative j with K attributes becomes

display math(5)

[22] Error term εij has a type I extreme value distribution. The β parameters are assumed to be random and distributed independently of εij. In particular, the coefficients have a fixed and a random component:

display math(6)

[23] In practice, the MMNL can be estimated with many different assumptions for each of the ηki terms. It is common to assume ηki is normally distributed, though econometricians often employ a triangular or lognormal distribution for the distribution of a parameter if that parameter (such as the coefficient on the cost attribute) is expected to have a bounded range [Hensher and Greene, 2003]. The results of a MMNL yield estimates of the median and standard deviation of the distributions of each of the random coefficients' distributions. If the standard deviations are statistically significant, then it is important to have controlled for unobserved heterogeneity.

[24] The third econometric model we employ is not commonly used in the CE literature, but it allows us to control for possible nonresponse bias. The WCL is a weighted version of the standard CL model in which the weights are estimates of the respondents' “propensity” to have returned the survey [Hindsley et al., 2011]. Those estimates are derived from a probit regression of whether or not a household did return the survey as a function of demographic features of household's location [Cameron et al., 1996]. Because the correction for nonresponse bias is not appropriate for a MMNL model [Hindsley et al., 2011], we cannot carry out a regression that compensates for both problems.

[25] To run the probit regression, we merge information we have about how many surveys were delivered and returned in each block with 2010 census data on households: block-level data on average household size, age and gender of household head, proportion of households with children and adults over 65 years old, and block group data on median income and level of education. This is similar to the approach of Cameron et al. [1996]. The estimated probability that household j does return the survey, πj(Xj), is the same for each surveyed household in a block because they have the same characteristics Xj, respondents, and nonrespondents alike. In contrast to sample selection correction in regressions with continuous dependent variables, one does not calculate and include an inverse mills ratio as a covariate in a CL setting to control for nonresponse bias. Instead, one uses the predicted probabilities from the probit regression to calculate propensity-score based weights [Hindsley et al., 2011; Manski and Lerman, 1977]. The weight, Wj, for an observation associated with respondent j is calculated as

display math(7)

which represents the odds that a respondent is a member of the random sample of nonrespondents given its characteristics. The weights are used as sampling or probability weights.

3.3. Survey Design

[26] We developed a survey to measure the values of flood reduction and environmental quality changes that are connected to stormwater management. We identified the attributes for our choice scenarios through informal interviews with professors of engineering and landscape architecture, community members, and personnel from local city government offices in two public meetings held by the University of Illinois. We chose attributes that community members were concerned about and that stormwater-management design could actually influence; the wording of the survey was also informed by these public meetings as we learned what kind of language was familiar in a useful way to city residents and what language might provoke unhelpful responses from survey respondents. We pretested our original survey design with two focus groups of Champaign-Urbana residents and modified the final survey in response to feedback from that process.

[27] The survey provided respondents with background information about stormwater management problems and controls and then presented respondents with six choice questions, each of which asked them to choose between a pair of hypothetical stormwater management projects that had varied values of the following six attributes (see Table 1): the frequencies of street, backyard, and basement flooding; surface water quality; rainfall infiltration; and cost (in the form of an annual stormwater utility bill). Each attribute had four levels including the current situation. Respondents could also choose to have no new stormwater-management projects in their town; this opt-out option leaves flooding and environmental quality the same and entails no cost. An example of a choice question is shown in Figure 1.

Table 1. Attributes and Levels in the Choice Experiment Questionsa
  1. a

    This table lists all the possible levels that each of the scenario attributes can take in a given choice scenario. Status quo levels are shown in bold font.

Type of floodingNumber of street floods within one block of your house50% less frequent than current
25% less frequent than current
Current frequency
25% more frequent than current
Number of floods in your backyard50% less frequent than current
25% less frequent than current
Current frequency
25% more frequent than current
Number of floods in your basement50% less frequent than current
25% less frequent than current
Current frequency
25% more frequent than current
Environmental attributesQuality of water in nearby streamsBetter quality: swimmable
Better quality: fishable
Current situation: boatable
Worse quality: polluted
Water infiltrationMore infiltration: high (90%–100%)
Current situation: medium (75%–89%)
Less infiltration: low (51%–74%)
Less infiltration: very low (0%–50%)
CostAnnual stormwater utility bill$0
Figure 1.

Example of choice question used in the survey.

[28] Attribute levels were specified as changes relative to the current situation because respondents in different areas are likely to experience heterogeneous status quo levels of flooding frequency. The survey included levels that are both lower and higher than the status quo to reflect the fact that stormwater control techniques can sometimes increase the likely occurrence of one type of flooding in order to decrease the frequency of another one, and stormwater management can either improve or degrade environmental conditions.

[29] For the water quality attribute, we used a modified “water quality ladder” applied in valuation research by Carson and Mitchell [1993], which translates technical water quality measures into simple categories which nonexperts can easily understand. The ladder had four categories (from best to worst quality: drinkable, swimmable, fishable, and boatable) that depend on levels of conventional pollutants. In our survey, boatable was the status quo level; we explained that and provided a simple description of each category.

[30] Recent research on stormwater management and LID has emphasized that LID strategies can improve measures of environmental quality other than water pollution [National Research Council, 2009]. LID infrastructure such as rain gardens and permeable concrete increase infiltration while reducing runoff; this decreases the amount of impervious surface in urban areas, increases water table recharge, and reduces extreme fluctuations in the volume and speed of water flowing in streams [Federal Interagency Stream Restoration Working Group, 1990]. Thus, we included an attribute of local environmental conditions which is summarized in the survey as “infiltration.” The survey instrument had a section with a simple explanation of the benefits of infiltration including water table recharge, pollution control, and improved general aquatic ecosystem health; the explanation made clear how increased impervious surface (and thus, decreased infiltration) is associated with low fish populations and biotic integrity in local streams [FitzHugh, 2001; Fitzpatrick et al., 2005]. A positive coefficient on the infiltration rates associated with hypothetical scenarios indicates that respondents value the health of local aquatic ecosystems. The categories of infiltration we used (very low, low, medium, and high) translate into exact percentages of rainwater that infiltrates (instead of becoming runoff) as shown in Table 1. The choice questions were followed by two sets of simpler questions. Respondents answered a demographic questionnaire and a set of questions about their experiences with flood frequency and their willingness to allow installation of decentralized stormwater controls on their property.

[31] In order to decide exactly which combinations of attributes respondents should be asked to choose between, CE survey design uses statistical methods to develop an experimental design (the combination of attributes and levels that result in different alternatives or profiles included in the choice questions). We used an orthogonal fractional factorial main effects design to assign attributes' levels in the scenarios presented in choice questions; this is standard in the choice experiment literature [Holmes and Adamowicz, 2003; Louviere et al., 2000; Street et al., 2005]. Such a design avoids correlation between the levels of multiple attributes in the choice alternatives with which people are presented. We created a design with 36 choice sets using a Macro for SAS 9.2 [Kuhfeld, 2009] and then blocked it into six sets so each survey had six choice questions for the respondents to answer. A limitation of our experimental design is that it does not permit interactions between attributes to be estimated. However, main effects often capture most of the variance in a CE model [Louviere et al., 2000].

[32] We administered this survey to households in the twin cities of Champaign and Urbana, Illinois. According to the U.S. Census Bureau, out of the 366 metropolitan statistical areas, Champaign-Urbana is the 191st largest in population with over 230,000 people in 2011. This area is typical of small growing urban communities, with two downtown cores and expanding residential and commercial development at the fringes that is increasing impervious surfaces and burdens on storm sewer infrastructure. According to the Clark Dietz, Inc. [2009], significant surface flooding (one to three feet deep) occurs in several locations of the city when rainfall events exceed a 1 year return frequency (that is, the average recurrence interval of events of that intensity is 1 year). Water quality in streams around this area is only “boatable.”

[33] We distributed the survey to 1000 randomly selected residents in the early summer of 2010 and spring 2011. The houses were chosen randomly from U.S. TIGER/Line® block-level shapefiles [Geography Division, 2009]. We used a variation of a drop-off/pick-up method of survey administration [Steele et al., 2001], delivering surveys directly to respondents' front doors and picking them up 2 days later.

[34] The definitions of the variables used in the econometric models are presented in Table 2. The dependent variable for the regressions is a discrete indicator of which of the options in a given choice question was chosen by the respondent. The coefficients on infiltration and on water quality that is swimmable or fishable are all expected to be positive because infiltration is an amenity and both swimmable and fishable water are cleaner than the status quo of boatable. We expect the coefficients on street flooding, basement flooding, polluted water, and cost to be negative because flooding is a disamenity, polluted water is worse than the status quo category, and the coefficient on cost is minus the marginal utility of money. We expect the coefficient on basement flooding interacted with owning a basement to be negative, because basement flooding will matter more to people with real basements. We also expect the interaction term with flooding experience to have a negative coefficient because people with large status quo flood problems may be more concerned about a given percentage flood increase.

Table 2. Definition of Variables Used in the Model and Expected Signsa
VariableDescriptionExpected Sign
  1. a

    The dependent variable for the regressions is a discrete indicator of which of the options in a given choice question was chosen by the respondent in question.

CostAnnual stormwater utility bill
Street floodingFrequency of street flooding (% change)
Backyard floodingFrequency of backyard flooding (% change)
Basement floodingFrequency of basement flooding (% change)
Basement flooding × (basement owner?)Interaction between frequency of basement flooding and dummy for basement owners
Basement flooding × (basement owner?) × (flooding experience?)Interaction between frequency of basement flooding, dummy for basement owners, and dummy for people who report basement flooding in the last 2 years
InfiltrationRate of infiltration+
Water quality =  swimmableDummy for water quality = swimmable+
Water quality =  fishableDummy for water quality = fishable+
Water quality = pollutedDummy for water quality = polluted

[35] However, people are not randomly assigned houses in which to live. Recent research on locational sorting [Bayer and Timmins, 2005] points out that individuals can and sometimes do choose houses according to their personal preferences regarding the state of the environment at different locations. Hence, it could be that people who live in houses that are flood-prone have relatively low WTP to avoid flooding in comparison with other people in this housing market. In that case, the flood-experience variable would be endogenous [Englin and Cameron, 1996; Whitehead, 2006]. We do not have sufficient data to use an instrumental variable approach to control for this [Whitehead, 2006]. Thus, the coefficient on the experience interaction term may pick up two competing effects: a given person might be WTP more for flood prevention if they are suffering more from floods, but people in flood-prone areas might be of a type that worries less about a given level of flooding.

4. Results

[36] We obtained a total of 140 responses and a final useable data set of 131. In CEs, a unit of observation is a choice question rather than a respondent. Because each survey has multiple choice questions, this sample size is sufficiently large to identify individual coefficients and permit robust hypothesis testing. The response rate from this round of surveys is not high. This is an increasingly common feature of paper surveys [Groves, 2006; De Leeuw and de Heer, 2002], which may have been heightened in this case by the short amount of time between dropping off and picking up the surveys. Methodological research finds that response rate alone is a poor predictor of the presence or extent of nonresponse bias in survey findings, and that response rates may be less correlated with the extent of bias in cases like ours where response was affected by survey administration features that are not highly correlated with the subject of the survey [Groves, 2006]. We can evaluate the severity of nonresponse bias in our study by comparing the results of the CL and WCL regressions.

[37] Some summary statistics for several of the nonchoice questions in the survey are shown in Table 3. We can compare some of these statistics to census data for the towns in our study area. Our sample is close to the census report in having an average household size around two people. Slightly more women than men answered our survey (we have 41% men, as opposed to 50% in the census data). Our respondents are slightly older than the full population (54 as opposed to 38 and 37 years old in the two cities), though the census average age lies just within the standard deviation of our sample mean. Our respondents are slightly more well educated than the full population, with 64% of our sample reporting a bachelor's degree or higher (compared to 50% and 55% in the census data for the two cities); Champaign-Urbana is home to a major university, a community college, and two hospitals. One major difference is driven by the fact that we did not survey people in large apartment buildings (hence, 85% of our sample owns their residence, as opposed to 46% and 35% in the census data for the two towns). Our results may not be representative of people who live in such dwellings.

Table 3. Summary Statistics for Selected Nonchoice Questions
VariableMeanSDMinMaxCensus Dataa
  1. a

    Source: U.S. Census Bureau, 2010 census data, American FactFinder. The total population of Urbana in 2010 had a total population of 41,250; the total population of Champaign was 81,055.

  2. b

    Approximate average age of adults ages 18 or over, calculated from census population table.

  3. c

    Percentage of adults age 25 and older with a bachelor's degree or higher.

  4. d

    Self-reported flooding events in the last 2 years.

Dummy for gender: 1 if male0.410.490150.9%50.1%
Dummy: 1 if college degree or more0.640.480149.8%c55.4%c
Number of people in household2.120.99152.252.12
Dummy: 1 if house owner0.850.350145.7%35.0%
Dummy: 1 if belongs to environmental group0.130.3401  
Number of years at current home16.1112.550.2555  
Dummy: 1 if house has basement or crawl space0.890.3101  
Self-reported street flooding frequencyd7.2119.850105  
Self-reported backyard flooding frequencyd1.622.98020  
Self-reported basement flooding frequencyd3.817.49040  
Dummy: 1 if ever seen LID infrastructure0.600.4901  

[38] Other statistics cannot be compared to census data. Our sample is not uniformly allied with environmental groups. Over 89% of the homes have a basement or crawl space. Table 3 also makes clear that self-reported flood frequency experience is highly variable in our sample; some respondents never experience flooding of any kind near their homes, while others report flood problems every time it rains.

[39] The choice data were analyzed using a CL, a WCL, and a MMNL regression as described in section 'Econometric Methods'. The MMNL model was estimated with maximum simulated likelihood using a program developed for STATA by Hole [2007]. We assumed most of the random attribute parameters were normally distributed but used a lognormal distribution for the cost parameter to constrain the sign of the parameter over its range. Some studies specify at least one of the coefficients (usually the coefficient on the cost variable) to be fixed rather than random to avoid problems with convergence [Revelt and Train, 1998] but since our model converged within a reasonable number of iterations we allow all the coefficients to be random. The results of the first-stage probit regression for the WCL are shown in Table 4. None of the individual coefficients are significant even at a 10% level, but the test for joint significance of all the variables in the regression is significant at a 5% level. These estimated regression equation was used to generate weights for the WCL.

Table 4. Probit Regression of Probability of Survey Responsea
  1. a

    N = 999. The unit of observation is a household to which a survey was delivered. The dependent variable is a dummy variable equal to one if the household returned the survey. Independent variables have the same values for all households in the same census block, except income and education which are calculated at block group level.

  2. b

    Standard errors in parentheses. *p < 0.05, **p < 0.01, ***p < 0.001.

  3. c

    This variable measures the percentage of households in the block with that characteristic.

ln(household median income)−0.117 (0.152)
Median age0.00748 (0.0110)
Average household size−0.291 (0.254)
Household has children under 17c1.407 (0.890)
One-person householdc0.0210 (0.646)
Educational attainment of head of householdc
High school−0.0823 (1.872)
Attended college−1.932 (1.794)
College degree0.383 (1.363)
Grad school−0.00383 (1.351)
Head of household whitec0.0112 (0.728)
Head of household blackc−0.842 (0.930)
Head of household over 65 years oldc0.598 (0.746)
Head of household malec−0.364 (0.360)
Head of household owns residencec0.180 (0.346)
Constant term0.438 (2.099)
Prob > χ20.0157

[40] The results of the main regressions of interest are presented in Table 5. The likelihood ratio test for all three models has a very small associated p value. The first two columns of Table 5 have broadly similar results. Nearly, all the coefficient signs and levels of significance are unchanged by the survey-response propensity weighting; even the absolute values of the estimated marginal values given in Table 6 are similar across the two regressions, though the point estimates of those marginal values are slightly smaller in the WCL results. The main difference between the weighted and unweighted CL results is that weighting causes the backyard flooding variable to become insignificant; however, that variable was only significant at the 10% level in the unweighted CL results. Overall, the regression results in Table 5 indicate that sample selection bias is not a severe problem in our analysis. In contrast, it appears to be important to control for individual heterogeneity. In the MMNL specification, many of the parameters have significant standard deviations around their mean; unobserved heterogeneity is significant for these attributes. In addition, the MMNL yields much more conservative estimates of marginal WTP values than the CL. It seems that the CL model is dominated by the MMNL model for analysis of our data. Thus, we focus the rest of our discussion on the MMNL results.

Table 5. Estimated Regression Coefficientsa
Variable(1) CL(2) WCL(3) MMNL
  1. a

    Standard errors in parentheses. *p < 0.05, **p < 0.01, ***p < 0.001.

  2. b

    SD is the estimated standard deviation of the distribution of a parameter across the sample of respondents in the MMNL model. The sign of the estimated SD is irrelevant and must be interpreted as being positive in every case [Hole, 2007].

Cost−0.0148*** (0.00312)−0.0156*** (0.00344)−0.0463*** (0.01356)10.8256 (18.84614)
Street flooding−0.00502* (0.00240)−0.00617* (0.00261)−0.00623 (0.00444)−0.0168 (0.00929)
Backyard flooding−0.00694* (0.00277)−0.00424 (0.00307)−0.0184** (0.00568)0.0158 (0.00866)
Basement flooding0.000766 (0.00570)0.000924 (0.00497)−0.00455 (0.0117)−0.00630 (0.0144)
Basement flooding × (basement owner?)0.00170 (0.00642)0.000164 (0.00591)−0.00415 (0.0138)2.138** (0.735)
Basement flooding × (basement owner?) × (flooding experience?)−0.0256*** (0.00392)−0.0209*** (0.00428)−0.0236* (0.00972)0.0176** (0.00594)
Infiltration0.00949*** (0.00232)0.00996*** (0.00255)0.0228*** (0.00562)0.0132 (0.00981)
Water quality = swimmable0.0938 (0.180)0.0531 (0.205)2.080*** (0.389)0.0200** (0.00713)
Water quality = fishable−0.0637 (0.186)−0.0470 (0.211)1.804*** (0.350)1.094* (0.451)
Water quality = polluted−1.827*** (0.224)−1.731*** (0.241)−1.859* (0.767)−0.419 (0.615)
Log likelihood−633.88577−447.4639−483.22939
LR test p value0.0000.0000.000
Table 6. Estimated Marginal WTPa
Variable(1) CL(2) WCL(3) MMNL
  1. a

    95% confidence intervals in parentheses. *p < 0.05, **p < 0.01, ***p < 0.001.

  2. b

    Number gives marginal WTP for flood reduction in dollars per percentage increase in flooding frequency.

  3. c

    No calculation given because the affiliated parameter was not statistically significant.

  4. d

    Number gives marginal WTP for improved infiltration (and hence, aquatic habitat) in dollars per percentage change in infiltration rate.

  5. e

    Number gives marginal WTP for a change in water quality in dollars per categorical change from “boatable” level.

Street floodingbn.a.c−0.394* (−0.763, −0.025)n.a.c
Backyard floodingb−0.468* (−0.855, −0.082)n.a.c−0.397** (−0.687, −0.106)
Basement floodingb (if respondent is basement owner with flooding experience)−1.559*** (−2.287, −0.831)−1.289*** (−1.077, −0.600)−0.698** (−1.129, −0.268)
Infiltrationd0.641** (0.247, 1.034)0.637** (0.252, 1.022)0.493** (0.179, 0.808)
Water quality =*** (23.108, 66.804)
Water quality =*** (19.598, 58.352)
Water quality = pollutede−123.446** (−193.800, −52.817)−110.698** (−178.791, −42.605)−40.170* (−78.394, −1.946)

[41] In the MMNL model results, all coefficients have the expected signs and all of the environmental variables are significant at a 5% level or better. The coefficient on infiltration is positive and significant; people place positive value on hydrological improvements other than pollution reduction that are associated with LID-style stormwater management. In addition, the coefficients on “swimmable” and “fishable” are positive and the coefficient on “polluted” is negative. The status quo of “boatable” lies between those “polluted” and “fishable,” so people clearly gain utility from improved water quality. However, statistical tests fail to reject the hypothesis that the coefficients on “swimmable” and “fishable” are equal to each other in each of the three regressions (the p values are 0.2395, 0.1023, and 0.3510 for the CL, WCL, and MMNL, respectively). Either our survey respondents do not place more value on having water in which they could swim instead of just fish, or they did not carefully distinguish between different levels of improvement when answering the survey questions.

[42] Backyard flooding is significant in the MMNL regression but street flooding is not. In regressions not reported in this paper, we verify that the insignificance of street flooding is not changed if we add interaction terms of street flooding with past street flooding experience, a dummy for whether or not the respond was younger than 65 years old, or a dummy for whether the responded had experienced basement flooding (in case one type of flooding sensitizes them to all); none of those coefficients are significant at even the 10% level. Basement flooding is not significant without interactions; these results suggest that people are willing to pay to reduce basement flood rates, but only if they have a basement and have experienced such flooding in the last 2 years. The fact that the standard deviation on the coefficient for street flooding is significant indicates that some people in the sample might have positive WTP to reduce the frequency of street flooding, but the mean WTP is not significantly different from zero.

[43] Table 6 summarizes marginal WTP estimates for attributes; these were calculated using the median values of all parameters for both models. The numbers in Table 6 represent the annual amount of money a representative person or household is willing to pay for a unit change in each attribute. In the case of flooding, the figures represent the value of reducing the frequency of flooding by one percentage point. Thus, to make basement flooding 50% less frequent, people who have basements and have experienced basement flooding would be willing to pay around $35/yr. Note that if the locational sorting discussed earlier is a major factor in this housing market that value is biased down relative to the amount the average resident of Champaign-Urbana would be willing to pay to reduce flooding.

[44] It seems from the results that people are willing to pay for improving environmental quality of streams, which is consistent with previous literature on valuation of water quality. Respondents are willing to pay over $38/yr for a discrete improvement of quality from boatable to fishable and would be willing to pay $40/yr to avoid further deterioration of water quality in streams. These findings certainly fit within the range found by previous work in the literature; other studies of the value of improved water quality in U.S. surface waters have found values as low as less than 10 dollars and as high as hundreds of dollars [Johnston et al., 2005; Van Houtven et al., 2007] depending on factors such as the size of the water quality change, the methodology used, geographic variation, and household characteristics.

[45] In addition to traditional water quality, the hydrological properties of the watershed seem to matter. People are willing to pay almost half a dollar a year to improve a percentage point in infiltration, which translates to around $34/yr to go from the worst to the best possible category of infiltration rates in their watershed (from 25% to 95%).

[46] We use these results to explore the total benefits (or losses) to citizens in this community of projects that change multiple stormwater outcomes at the same time; in the discussion later we describe estimates based on median marginal WTP values and on marginal values from the lower bounds of the 95% confidence intervals (Table 7). Given that 58% of houses in our sample have a basement and have experienced flooding and the population of the survey area is around 50,000 households [U.S. Census Bureau, 2010], citizens of this urbanizing area would be willing to pay approximately $580,000 ($220,000)/yr for stormwater control that reduces basement flood frequency by 25%. Furthermore, if a stormwater management project also improved environmental quality (25% improvement in infiltration rates and an improvement in water quality to a fishable level), this community would obtain additional benefits valued at about $2,700,000 ($1,300,000)/yr.

Table 7. Approximatea Total WTP for Several Hypothetical Changes ($1,000)
ScenarioType of Estimate(1) CL(2) WCL(3) MMNL
  1. a

    These benefit and loss calculations make the simplifying assumption that the median parameter values estimated by our regressions apply across the full population of the communities surveyed. We based our calculations on around 52,000 total households in the area from the 2010 census data.

  2. b

    Using the median marginal WTP value as shown in Table 5.

  3. c

    Using the lower bound of the confidence interval shown in Table 5.

  4. d

    Using the upper bound of the confidence interval shown in Table 5.

  5. e

    No calculation given because the affiliated parameter was not statistically significant.

Reduce basement flooding 25% (only basement owners with flooding experience benefit)Medianb$1305$1079$585
Upper boundc$1914$1656$945
Lower boundd$696$503$224
Improve infiltration 25%Medianb$857$852$660
Upper boundc$1383$1367$1081
Lower boundd$331$337$239
Improve water quality to fishableMedianbn.a.en.a.e$2086
Upper boundcn.a.en.a.e$3123
Lower bounddn.a.en.a.e$1049
Decrease water quality to pollutedMedianb−$6560−$5925−$2150
Upper boundc−$10,373−$9570−$4196
Lower boundd−$2827−$2280−$104
Reduce flooding 25% but decrease water quality to pollutedMedianb−$5295−$4846−$1565
Upper boundc−$8459−$7914−$3251
Lower boundd−$2131−$1778$120

[47] However, the presence of heterogeneous status quo conditions means that in general, caution must be taken when deciding whether to undertake a stormwater management project that entails trade-offs between attributes of the outcome. For example, if the project that reduced basement flood frequency by 25% also reduced water quality of streams in the area from boatable to polluted, median parameter estimates indicate that the net benefit of the project would be negative, around −$1,500,000, though if we use conservative estimates of the harm done by harming water quality and the benefits of reducing basement floods, then the net impact is negligibly positive. Some households benefit from the flood reduction, but all are harmed by the decrease in water quality; the balance can easily be negative for the community as a whole.

[48] The calculations in Table 7 should be interpreted with caution. If, for example, the 87% of households that did not return the survey place zero value on all of the attributes, then median WTP for flood reduction and environmental improvement would be zero and the hypothetical projects in Table 7 would have little value. However, this result would be inconsistent with the extant literature that finds positive values for flood reduction and water quality, and when we control explicitly for possible nonresponse bias we do not find evidence that this is a very serious problem in our study.

5. Conclusion

[49] In this paper, we have used a CE to evaluate people's preferences over stormwater management outcomes in an urbanizing area. We find significant unobserved heterogeneity in the coefficients on most of the attributes in our study, lending further support to the growing body of CE evidence that simple CL estimation is often not appropriate because of individual heterogeneity. In contrast, we do not find strong evidence of serious nonresponse bias in our regression results even though our survey-response rate was fairly low.

[50] Our results find that people are willing to pay to reduce flood frequency (especially in the case of basement flooding), but the value of flood reduction depends on how much flooding people currently experience. We also find that citizens place large value on changes that would improve hydrologic function in a watershed, in addition to being willing to pay for improvements in conventional pollution-related stream water quality. This is the first research to estimate the benefits of modern stormwater associated with improvements in the environment other than reduced water pollution. The findings imply that these benefits are significant; policy makers and managers would benefit from more research on this subject that covers more urbanized areas and that samples citizens who live in apartment buildings as well as single-family homes.

[51] Implementing LID technologies has proved to be less expensive than conventional development in a number of case studies throughout the Unites States, with monetary savings ranging from 15% to 80% [Braden and Ando, 2011; U.S. Environmental Protection Agency, 2007] and environmental benefits relative to conventional stormwater management strategies that include pollution reduction, groundwater recharge, reduced water treatment costs, and habitat improvements. Cities across the country are developing a wide range of policies to improve stormwater management, and EPA is evaluating regulations that might mandate significant on-site management of stormwater nationwide. Our results imply that such regulations can have large monetized benefits relative to the costs of new LID-style development, especially if care is taken to design policies to improve hydrological function in urbanized areas instead of focusing entirely on minimizing flood risks. However, our findings also contain a cautionary lesson: policies and municipal storm sewer projects that worsen aquatic habitat in a quest to reduce flooding that affects only a subset of households in an area may have questionable net benefits for the community as a whole.


[52] This paper is based in part on work supported by Fulbright Colombia Colciencias-DNP and by USDA-NIFA Hatch projects ILLU-470-316 and ILLU-470-323. The authors are grateful for comments and advice from John Braden, Sahan Dissanayake, Rob Johnston, Noelwah Netusil, Rich Ready, members of the program in Environmental and Resource Economics at the University of Illinois, members of the W2133 Multistate Research group, attendees of annual workshops held by AERE and AAEA, three anonymous reviewers, and an Associate Editor.