LEARNING IN A HEDONIC FRAMEWORK: VALUING BROWNFIELD REMEDIATION

• Author: Lala Ma
• Date: 01 August 2019
• DOI: http://doi.org/10.1111/iere.12389
• Published date: 01 August 2019

INTERNATIONAL ECONOMIC REVIEW

Vol. 60, No. 3, August 2019 DOI: 10.1111/iere.12389

LEARNING IN A HEDONIC FRAMEWORK: VALUING BROWNFIELD

REMEDIATION∗

BYLALA MA1

University of Kentucky, U.S.A.

Incomplete information in property value hedonic models can bias estimates of marginal willingness to pay

(MWTP). Using brownfield remediation as an application, this article recovers hedonic values from a dynamic

neighborhood choice framework that allows households to learn about brownfield contamination in a Bayesian

fashion before choosing where to live. I find that ignoring learning yields nontrivial biases to the MWTP estimate.

This has important implications for hedonic valuation if agents are imperfectly informed. Estimates are used to

calculate information’s value had it been withheld from the public and to assess heterogeneity in information’s

value along site and homebuyer demographics.

1. MOTIVATION

In the absence of prices for environmental amenities, hedonic property value models have

been widely used as a revealed-preference approach to measuring the value individuals place

on nonmarketed goods. Rosen (1974) provides the theoretical foundations for interpreting

the hedonic gradient of a nonmarketed good as an individual’s marginal willingness to pay

(MWTP) for the good. Based on this model, many researchers have focused on recovering

consistent estimates of the MWTP for amenities using changes in property values under quasi-

experimental settings.2These studies exploit an exogenous change in some amenity of interest

(usually due to a policy shift) to establish a causal relationship between the amenity change

and changes in local housing prices. The capitalization of the policy change into housing prices

is then interpreted as the average MWTP for the amenity. What is crucial for these studies to

recover unbiased estimates of the MWTP is that prices in the prepolicy period represent a valid

baseline to which postpolicy prices can be compared. However, this is likely to be complicated

by the substantial provision of information to homeowners that accompanies policies before

the policy is implemented. For example, if consumers infer from provided information that

cleanup of a nearby contaminated site, otherwise known as a brownfield, is imminent (forward-

looking behavior) or gain information about the severity of the contamination (learning from

new information), then the baseline period prices need to be adjusted to consider these two

effects, as they may drive potential distortions to the MWTP estimate if left unaccounted for.

This article considers distortions in policy valuation of nonmarketed goods from not account-

ing for two types of household behaviors—an expectations bias from forward-looking behavior

∗Manuscript received April 2016; revised September 2018.

1I would like to thank Peter Arcidiacono, Patrick Bayer, Glenn Blomquist, V. Joseph Hotz, Nick Kuminoff, Carlos

Lamarche, Hugh Macartney, and especially Chris Timmins for their invaluable advice. I am grateful to the editor,

Holger Sieg, and four anonymous referees for comments that significantly improved the article. I would also like to

thank seminar participants at Duke University, University of Kentucky, Camp Resources, and the AERE Summer

Conference for their helpful comments, Richard DiSalvo for excellent research assistance, and Kevin Haninger and

contacts at the EPA and the Massachusetts Department of Environmental Protection for access to and assistance with

the brownfields data. I gratefully acknowledge support under the Joseph L. Fisher Doctoral Dissertation fellowship

from Resources for the Future. Please address correspondence to: Lala Ma, University of Kentucky, 225C Gatton

Business and Economics Building, 550 S. Limestone Street, Lexington, KY 40506. E-mail: lala.ma@uky.edu.

2Some recent examples include Davis (2004), Chay and Greenstone (2005), Linden and Rockoff (2008), Linn (2013),

Currie et al. (2015), Muehlenbachs et al. (2016), and Haninger et al. (2017).

1355

C

(2019) by the Economics Department of the University of Pennsylvania and the Osaka University Institute of Social

and Economic Research Association

1356 MA

and an information bias due to learning from information provision. To address these two

sources of bias, I incorporate learning into a dynamic hedonic framework. Using brownfield

remediation as an application, I model a household’s residential location decision as a dynamic,

discrete neighborhood-choice problem, where households update their knowledge of brown-

field hazard information in a Bayesian fashion before making decisions on where to live. I

collect a new data set on brownfield hazard information, which allows me to test my model and

estimate the bias.

The primary contribution of this article is to build upon previous models of locational sorting

(Epple and Sieg, 1999; Epple et al., 2001; Bayer et al., 2004) and extend models of dynamic

residential choice (Bishop, 2012; Bayer et al., 2016) to allow for learning from exogenous infor-

mation provision. As evidence of learning has been found in various other settings from health

plan to college major choice, it seems sensible that it would impact residential choice as well.

That I find significant bias from assuming fully informed or myopic agents has important impli-

cations for willingness to pay estimates recovered from hedonic models. Hedonic estimates are

crucial inputs into cost–benefit analysis of policies affecting nonmarketed amenities; researchers

and practitioners alike should be cognizant of the potential bias from ignoring these dynamics.

As such, this work is also related to the growing literature on hedonic valuation, including sev-

eral important, recent innovations on the identification and estimation of demand for housing

and nonmarketed goods (Ekeland et al., 2004; Bajari and Benkard, 2005; Heckman et al., 2010;

Bajari et al., 2012; Kuminoff and Pope, 2014; Landvoigt et al., 2015; Epple et al., 2018). A related

contribution is methodological. Models of residential location usually involve large choice sets,

which become computationally overwhelming once dynamics are introduced. Various meth-

ods have been used to deal with this state–space issue.3In my setting, the problem is further

complicated by uncertainty when households make choices based on their beliefs, which would

require reestimating the discrete choice problem for each guess of the belief parameters. To

circumvent this, I set up preferences such that I can recover beliefs outside of neighborhood

choice. Although this requires learning over time to take place at the neighborhood level, doing

so yields significant computational savings, especially given the size of the choice set.

The main results find an MWTP that is estimated to be $499.01 to remove contamination

(e.g., volatile organic compounds [VOCs]) from a specific pollution medium (e.g., soil) when

accounting for learning and forward-looking behavior. Viewed as a perpetuity and using the

average amount of contamination found at a brownfield site, this translates into an MWTP

of 12.4% of mean housing values. I find evidence that households are forward-looking with

respect to contamination and are learning about contamination over time: The MWTP to clean

an average brownfield site of 12.4% is approximately two times larger than one that ignores

dynamics altogether. My model attributes about half of the total bias (42–49%) to ignoring

forward-looking behavior and the remaining bias (53–60%) to ignoring learning. In particular,

the bias from ignoring learning implies that households slowly discover that contamination is

worse than what they had originally thought. As such, when we use data on housing values

to infer MWTP and assume that households are informed about site hazards, we are falsely

attributing high housing values to households not caring (enough) about brownfields when, in

fact, the households are just uninformed about the extent of the contamination.

Using the estimates from the model, I then calculate the value of the information regarding

contamination had it been withheld from the public. I find that most of the value of an assess-

ment comes from the toxicity information that it conveys instead of its implications for cleanup.

The results also suggest that information is most valuable in areas where contamination is the

worst. I additionally assess heterogeneity in information’s value across different sociodemo-

graphic groups based on the race and income of homebuyers. Higher income households value

contamination removal more than lower income households, but I find that homebuyers with

low-income backgrounds value assessments twice as much as their higher income counterparts.

3This article’s empirical methodology builds upon previous work including Hotz and Miller (1993) and Arcidiacono

and Miller (2011).

LEARNING IN A HEDONIC FRAMEWORK 1357

These results highlight the importance of accounting for learning in obtaining estimates for

environmental preferences and the potential for information provision to not only mitigate

pollution exposure but also affect pollution’s distribution across sociodemographic groups.

Section 2 describes the two information-related sources of bias in more detail, as well as other

research that has addressed these biases. As part of the model is tailored to learning about a

specific amenity, I follow in Section 3 with a background on the amenity of interest, a brownfield,

and a description of the collected data. Section 4 then outlines a model of neighborhood choice

that incorporates learning and forward-looking behavior. Section 5 gives summary statistics and

discusses other data sources, and Section 6 describes the estimation procedure. I present the

results in Section 7. Finally, Section 8 concludes.

2. TWO SOURCES OF INFORMATION BIAS

2.1. Expectation Bias. Rosen (1974) models households as static utility maximizers. How-

ever, if individuals are forward-looking, their choices of amenities in the current period will

reflect expectations for how these amenities may evolve, which will subsequently be reflected in

transaction prices. This is most applicable in decisions relating to durable goods, such as that to

purchase a house, given the size of the investment as well as the large costs associated with mov-

ing. Using a static framework to interpret MWTP from changes in housing prices will therefore

confound the estimate with individual expectations about the future. This is exacerbated in the

case of valuing the impacts of an amenity-related policy if there is sufficient information provi-

sion before the policy is implemented. Information provided may cause the policy of interest to

be anticipated and drive an even greater expectations bias in the estimated MWTP.

Beginning with Epple and Sieg (1999) and Epple et al. (2001), researchers have modeled

and estimated the sorting process that underlies the hedonic equilibrium to recover preferences

for nonmarketed attributes. In the first application to the study of the environment, Sieg et al.

(2004) use an equilibrium sorting model to estimate the value for nonmarginal changes in

air quality.4By placing additional structure on household/regulator decision processes, these

models are able to incorporate important features beyond traditional hedonic models such

as allowing for general equilibrium feedback effects in the presence of large policy changes.

More recently, the urban and environmental economics literatures have seen dynamic structural

models such as those of Bishop and Murphy (2011), Mastromonaco (2011), Bishop (2012), and

Bayer et al. (2016), who specify individual preferences and maximizing behavior to account for

the forward-looking nature of household decisions. They employ recent advances in models

of dynamic discrete choice (Hotz and Miller, 1993; Arcidiacono and Miller, 2011) to estimate

the parameters of a household’s MWTP curve for time-varying neighborhood amenities (e.g.,

crime and air pollution). The same logic applies to fixed amenities as well, since households

may change their expectations of how likely a fixed (dis)amenity will exist in the future based

on, among other things, information provided in the time leading up to the policy change.

2.2. Learning Bias. A second source of bias relates to learning about the current quality of

some attribute of interest from information provision instead of its value in the future. Public

disclosures may change what is known about an amenity. Many policies regarding the cleanup

of a hazard will provide information to affected households before any remedial actions are

taken. Furthermore, information regarding the amenity may be released in multiple waves.

Suppose a household lives near an aesthetically displeasing site and then learns that the site

is not only unattractive but is also contaminated with low levels of carcinogens. It is likely

that estimated preferences for the site will be different depending on whether baseline prices

were measured before the information was released or after. Practically, this means that the

estimated MWTP will depend on when one takes prices as baseline even if agents are assumed

4For additional applications to value environmental goods, see Walsh (2007), Banzhaf and Walsh (2008), Klaiber

and Phaneuf (2010), and Tra (2010).