Valuation of Unconventional Oil and Gas Development

Valuation of Unconventional Oil and Gas Development By Andrew J. Boslett Doctor of Philosophy of Environmental and Natural Resource Economics University of Rhode Island Public discourse regarding the local economic, environmental, and socio-cultural impacts of unconventional oil and gas development has been intense, especially in those areas of the country that are relatively unfamiliar with extractive industry. At the center of this debate is the contrast between the economic and financial benefits that accrue to local governments and landowners versus the environmental and socio-cultural costs borne by the public-at-large. Understanding how local citizens value unconventional oil and gas development is an important policy consideration for federal, state, and local governments. The overarching theme of this work, consisting of three manuscripts, seeks to contribute to the debate regarding the impacts of energy consumption and development. In two manuscripts, I use the hedonic valuation approach to value the benefits and costs of shale gas and oil development. I do this through the context of a statewide moratorium on development in New York and a long-standing severance between the surface and mineral estates in the western Colorado property market. My results indicate that homebuyers significantly value both the financial benefits and environmental costs of unconventional oil and gas development. In Manuscript 1, I find that New York properties that were most likely to experience the financial and environmental impacts of Marcellus Shale development decreased in value by 23% as a result of the moratorium, which under certain assumptions indicates a large and positive net valuation of development. In Manuscript 2, I find that the homebuyers have large and significant valuations of the environmental costs of development on the order of 35% of sale prices for those properties that have an unconventional well within a mile of the property’s extent. In my third manuscript, I again apply hedonic valuation and value the benefits and costs of silica sand mining in western Wisconsin. The great increase in the application of hydraulic fracturing in oil and gas production has led to an increase in demand for silica sand. This type of sand is an important ingredient in hydraulic fracturing. Silica sand mining has a number of local benefits and external costs. I use the hedonic valuation methodology to value sand mining’s impacts, focusing on property views and local air quality. I find strong evidence that both changes in view and air quality are negatively capitalized into housing prices. I also find evidence of appreciation for those properties that are not as subject to those environmental quality changes.

: Robustness checks for the 5 mile border restriction and shale thickness model .49 Table 4: Heterogeneous impacts of the moratorium by acreage and for resolution towns 50

INTRODUCTION
Shale gas development (SGD) has dramatically changed the US energy landscape in the last decade. The Energy Information Administration (2013) predicts that the US will shift from being a net importer to a net exporter of natural gas by 2020 and domestic production will increase 44% by 2040. Much of the attention on SGD has been on the Marcellus Shale, which extends over 95,000 square miles across New York, Ohio, Pennsylvania, and West Virginia (Kargbo et al., 2010). Given the current scale of SGD and expected growth in the future, it is critical to understand to the local valuation of SGD. This paper seeks to answer this question using a hedonic framework, as housing prices should reflect the future stream of benefits and costs tied to the property. Empirically, this is hindered in two ways. First, the location of wells may be endogenous. Second, expectations about SGD form in advance of actual drilling, and if expectations are capitalized into housing prices, then a 4 simple before-after comparison may lead to incorrect inference about the valuation. We mitigate these confounding factors by specifically focusing on expectations and using an exogenous shift in expectations to reveal valuation.
Just as hydraulic fracturing was beginning its exponential increase in Pennsylvania, New York State implemented a de facto moratorium on hydraulic fracturing on July 23, 2008, citing uncertainty about health and environmental impacts (State of New York's Executive Chamber, 2008). 1 The state extended the moratorium multiple times between 2010 and 2014 (e.g., Wiessner, 2011) and, on December 17, 2014, the New York Department of Environmental Conservation implemented a permanent ban (Kaplan, 2014). These decisions were highly contentious, as evidenced by several dozen towns in New York passing resolutions in support of SGD in the spring and summer of 2012 and 15 towns are currently considering secession (Mathias, 2015). 2 To date, there has been no hydraulic fracturing in New York.
This paper exploits changes in expectations that resulted from New York's moratorium on drilling and measures this event's impact on housing prices. Importantly, the moratorium did not mark a change in the amount of hydraulic fracturing in New York expectations about future SGD are the only thing that changed.
We estimate the effect of the statewide moratorium using a difference-indifferences methodology. We use Pennsylvania as a counterfactual because 1 There is considerable heterogeneity in state regulation on shale gas development as a result of different political, hydrological, and geological dynamics (Kulander, 2013; Richardson et al., 2013. Some states have used a more lenient approach to regulation. For example, Pennsylvania had no specific regulations concerning hydraulic fracturing until early 2010 . Since then, Governor Tom Corbett's signed Act 13, prohibiting any local regulation or restrictions on shale gas well production (Begos, 2012). Like New York, New Jersey and Maryland have enacted regulations to restrict or ban hydraulic fracturing. 2 These resolutions could not supersede state law, but were meant to send a signal to state politicians in Albany and were in contrast to the more common local bans and moratoria implemented elsewhere in the state.
expectations about future SGD were likely similar to those in pre-moratorium New York, but in contrast with New York, those expectations were realized. Our aim is to identify the change in prices for properties in New York that are most likely to be impacted by SGD (both positively and negatively), relative to price changes for similar properties in Pennsylvania. We use private well water use as a proxy for properties likely to experience SGD. 3 These are essentially rural properties outside of municipal water supply boundaries, meaning they have the space requirements for drilling.
Further, contaminated well water is one of the most common and serious environmental costs.
The design of our preferred sample is motivated by a border discontinuity and underlying shale geology. We begin with property transactions data for two Pennsylvania and three New York counties along the border. In the vein of recent border discontinuity designs (e.g., Grout et al., 2011;Turner et al., 2014) and specifically those that use state borders (Holmes, 1998;Rohlin et al., 2014), we restrict observations to be within five miles of the border in order to minimize unobserved differences in price determinants and best model the counterfactual for New York residents. Even after these restrictions, there are still substantial shale geology differences across the border. Thus, we further restrict observations to be in a specific band of shale thickness, a geological characteristic that strongly affects the amount of gas or oil in a reservoir (Advanced Resources International, 2013). These restrictions are meant to improve the similarity of expectations about future SGD. Post-moratorium spillovers across the border are a threat to identification. However, we contend that 6 these effects are minimal due to pre-moratorium expectations about spillovers, the rapid pace of drilling stemming from high initial prices, the area comprising a single labor market, and southerly flow of surface water.
Using the 5-mile border and shale geology restrictions, our results suggest that the statewide moratorium decreased New York property values 23.1% for those properties most likely to experience SGD. Relaxing the sample restrictions leads to smaller estimates in the range of 10-21%, which suggests that effects are heterogeneous across our New York counties and that accounting for shale geology is critical for understanding expectations. We estimate a series of robustness checks that test additional shale geology restrictions, test for spillover effects across the state border, and use municipal water properties as an additional control, and results are consistent with point estimates in the range of an 18-26% drop in housing values.
We interpret these results as a positive net valuation of SGD by buyers and sellers in New York and Pennsylvania. However, this interpretation relies on two assumptions: the expected probability of SGD in pre-moratorium New York is 1 and the expected probability of post-moratorium SGD is 0 and New York and Pennsylvania property owners and buyers accurately valued the negative and positive aspects of SGD prior to the moratorium. We estimate several models that bolster our confidence in these assumptions. However, if either of these assumptions are false, we are still recovering the effect of the moratorium on property values, which is driven by expectations over financial benefits and environmental externalities of SGD, and this is an important estimate for areas considering bans on hydraulic fracturing. Further, the estimates serve as a validation that expectations are capitalized into property values.

7
One of the models we use to test the assumptions needed for an interpretation of net valuation is a more traditional model of the effect of proximity to drilling using only our Pennsylvania observations. The results suggest no price impacts of proximity. While one interpretation is that the impacts of drilling are small, we interpret this to mean that ex ante expectations established in the initial expansion of SGD in Pennsylvania were capitalized into property values and were accurate ex post leading property values not to change. These results corroborate our claim that New York households near the border have accurate expectations about SGD, which in turn supports a rational expectations assumption in hedonic valuation.
There are two major contributions of this paper. First, we provide new evidence of local impacts of SGD. Existing hedonic studies (Gopalakrishnan and Klaiber, 2014;Muehlenbachs et al., 2014) find negative impacts of nearby drilling for well-water dependent properties as large as -22%. However, Gopalakrishnan and Klaiber (2014) also find that negative effects dissipate to a statistical zero 6-12 months after a permit is issued. Our results lead to very different conclusions. One reason may be that both of these studies either use data exclusively from western Pennsylvania or derive most of their identifying variation from western Pennsylvania. A concern is that split estates, where mineral rights are sold separately from the property, are common in western Pennsylvania due to the area's more extensive history of resource extraction . In contrast, split estates are relatively uncommon in our focus area of eastern Pennsylvania and south-central New York. Thus, our data are more likely to recover net effects of SGD because property owners hold mineral rights and will benefit from royalties and lease payments. Our interpretation of Gopalakrishnan and Klaiber (2014) 8 and Muehlenbachs et al. (2014) is that their estimates capture the negative externality of SGD near private well water, which is critical to understand, but mostly exclude the financial benefits because of the area of study. Consistent with this interpretation are recent survey findings that indicate a majority of property owners that do not hold the mineral rights to their property are dissatisfied with local drilling, whereas a majority of property owners holding mineral rights are satisfied (Collins and Nkansah, 2013). 4 While the split estate issue is perhaps the most critical, there are other differences between our study and others that could lead to different estimates of the local impact of SGD. We incorporate physical attributes of shale geology into the analysis, which existing valuation studies have not utilized. This appears to be important to creating valid counterfactuals in a difference-in-differences framework.
Further, our treatment group has no direct experience with SGD, though they seemingly would learn about it as SGD expanded right across the border. Additionally, we are estimating area-level impacts that capture impacts occurring to whole areas, as opposed to a proximity analysis that captures differential impacts for properties nearby drilling.
This focus may average away some of the negative effects of SGD if property owners in NY expect that they would be minimally impacted by negative externalities since the placement of future shale gas wells is unknown.
The second contribution is to add to our understanding of how expectations are capitalized into property values. While many hedonic papers implicitly assume expectations exist and recent structural models have incorporated expectations (e.g., Bishop and Murphy, 2011;Ma, 2013), we offer a particularly clean, reduced-form 9 illustration of how expectations factor into prices. The effect of the New York moratorium is to change expectations, whereas the results of the proximity analysis using only Pennsylvania properties support the idea of rational expectations because no price changes occur once drilling commences. This work also complements hedonic studies that show new information can cause capitalization of dis-amenities, even when levels of dis-amenities do not change (e.g., Pope, 2008;Guignet, 2013).

BACKGROUND
The first objective of this section is to catalog various estimates of benefits and costs of SGD, which is critical for putting our estimates of the net valuation of SGD in context. Given the private and dispersed nature of financial benefits, it is a contribution of this paper to compile these estimates. The second objective is to give a timeline of SGD in Pennsylvania and SGD regulation in New York.

Financial benefits
During shale gas extraction, owners of sub-surface mineral rights may sign a mineral lease contract with energy production companies, granting them the right to develop mineral deposits underneath their property (Pennsylvania Department of Environmental Protection, 2012). The two primary monetary benefits associated with shale gas production are lease signing bonuses and royalty payments. A lease signing bonus is an initial payment, based on acreage, for signing a gas lease contract (Weidner, 2013). Due to the uncertainty of natural gas production, this is perhaps the most important element of the lease (Hefley et al., 2011). The payment level is based on a number of variables, including geological factors, landowner-stipulated restrictions, nearby drilling results, and the current state of the natural gas market (Weidner, 2013).
The average per acre signing bonus is $2,700 (Hefley et al., 2011), though this can vary from $50 to almost $6,000 (Humphries, 2008; Green, 2010; Eichler, 2013; Rieley, 2014. The other major monetary benefit is royalty payments, which are recurring payments on a proportion of natural gas production. The minimum royalty rate, set by law, is 12.5% of the value of extracted natural gas (Pennsylvania Department of Environmental Protection, 2012). However, the negotiated rate can be much higher, depending on the same factors that determine lease payments (Weidner, 2013).
According to a Penn State University Extension associate, Marcellus Shale gas production has generated a cumulative total of $160 million in royalties for landowners in Bradford County, Pennsylvania as of late 2012 (Loewenstein, 2012).
SGD infrastructure-related benefits can also serve as economic windfalls for landowners. Surface rights owners can receive monetary payments for allowing pipeline, compressor station, and water impoundment construction on their properties. These are often one-time payments. A payment for pipeline easement construction is based on the length of the constructed pipeline and can range from $5-25 per linear foot (Messersmith, 2010). Due to the nuisance factor associated with compressor stations (e.g., Litovitz et al., 2013), payments for their construction can range from hundreds of thousands to millions of dollars (Clark, 2014). Payments for water impoundment construction can range from $40,000-70,000, but could potentially be lower or higher given their intended size and permanency (Clark, 2014).
Lastly, governments have the ability to raise revenue through taxing shale gas 11 development, which in turn would have public finance implications. These public finance measures could then be capitalized in housing prices through improvements to public goods and services in local municipalities, such as schools. Pennsylvania did enact an "impact fee" in 2012 through Act 13, retroactive to 2011 activity, which charged a fee on a per-well, per-year basis. 5 As SGD expands these impact fees could be a considerable source of income for local government; in 2012 impact fees in Pennsylvania brought in $202 million (Rabe and Hampton, 2015). The distribution of the fees can go to a variety of sources, such as county and municipal governments, various environmental and non-environmental state government agencies, and the state's legacy fund (Powelson, 2013).These various estimates of monetary benefits highlight the variation and uncertainty of the how much revenue could be expected from future SGD.
For additional details, see online appendix Table A1.

Costs
There are also a number of potential landowner costs of nearby shale gas development, which are primarily driven by environmental impacts. The hydraulic fracturing process is highly water intensive, so much of the focus on the environmental costs of SGD revolve around water quantity and quality impacts. Shale gas development has led to large increases in wastewater management needs (Rahm et al., 2013). In Pennsylvania, regional wastewater generation has increased by 540% since 2004 (Lutz et al., 2013). In terms of water quality, Jackson et al. (2013) find increased levels of 12 methane contamination in groundwater in heavy-SGD areas, while  find evidence of surface water pollution as a result of SGD waste disposal and management processes. In 2014, the Pennsylvania Department of Environmental Protection released a list of more than 250 instances where SGD operations impacted water quality in the state.
In addition, recent research has shown increased air pollution in areas close to shale gas extraction and processing infrastructure (e.g., Litovitz et al., 2013;Rich et al., 2014). Increased air pollution associated with shale gas development may have significant public health implications (e.g., McKenzie et al., 2012). Although the mechanisms are unclear, Hill (2012) finds significant impacts of shale gas extraction on the birth weight of children born in nearby homes. Additional environmental costs have been identified as concerns such as seismicity (Frohlich, 2012), forest loss and fragmentation (Drohan et al., 2012), and ecosystem services and local biodiversity (Evans and Kiesecker, 2014;Kiviat, 2013). For additional details regarding environmental and social impacts of SGD, please refer to Table A2 in the online appendix.  Although excitement regarding the economic benefits of SGD grew as reports of lease activity became public, there were still significant concerns regarding the environmental, social, and public health aspects of drilling (Wilber, 2014). Citing the fact that the state was relying on a previous environmental impact statement of oil and gas 14 drilling from 1992 that did not address the many unique environmental issues associated with SGD (Lustgarten, 2008) (Hakim, 2012). However, this policy was never enacted.

Timing of drilling and regulation
After six years of legislative and executive order action, the situation culminated in a permanent statewide ban on SGD in December 2014 (Kaplan, 2014), driven largely by lingering public health concerns (NY Department of Health, 2014). As a result of this series of policies, no unconventional natural gas development has occurred in New York, which is reflected in Figure 2.
Despite the statewide nature of the moratorium, New York is a home rule state that grants legislative authority to local governments to enact local legislation that may 15 limit state-level intrusion into local matters (Stinson, 1997). Given this history and the discontent with the moratorium, 45 New York towns passed resolutions in support of SGD in the spring and summer of 2012 (FracTracker, 2014). Fourteen of these towns are in our sample counties and are shown in Appendix Figure A3. These resolutions were passed by town councils and were not voted on by residents, but likely reflect residents' sentiments. The resolutions had no impact on the ability for gas companies to operate in New York, but were intended to apply political pressure to state policy makers and signal to industry that these towns are supportive of SGD. One of the major landowner groups driving the passage of the resolutions, the Joint Landowners Coalition of New York, sued Governor Cuomo in order to expedite the state's environmental and public health review of SGD (De Avila, 2014).
On the other side of the debate, 176 New York towns implemented local bans or moratoriums on hydraulic fracturing in the event that the statewide moratorium was lifted (FracTracker, 2014). Most of these townships were located in areas of the state that were unlikely to experience significant SGD from the Marcellus Shale due to geological limitations (e.g., low thickness). In our three NY counties, there were only two towns -Owego (Tioga County) and Wayne (Steuben County)that passed moratoria on shale gas development, both in 2012.

CONCEPTUAL FRAMEWORK
In this section we present a hedonic property model that incorporates the phenomena of interest, the valuation of expected shale gas development through the enactment of a moratorium. The hedonic valuation methodology, originally presented by , posits that the price of a heterogeneous good can be decomposed into implicit prices associated with its individual characteristics. By separating the price of the good into its implicit prices, the technique can help illuminate the value of each characteristic. The standard hedonic model assumes that all negative and positive discounted cash flows will be capitalized into the transaction price if there is full information about those attributes that derive benefits and costs.
We apply the hedonic valuation concept to shale gas development through housing prices. The price function is given as Ph=Ph (L,S,N,Q(D),G(D)) where L is a vector of lot characteristics, S is a vector of structural characteristics, N is a vector of neighborhood characteristics, Q is a vector of environmental characteristics, and G are geological characteristics that allow for possible shale gas development (e.g., land overlying shale with retrievable gas). Shale gas development is represented by D.
Geological attributes, G(D), derive value from financial amenities such as lease and royalty payments that gas companies pay to homeowners to gain access to the shale. The dis-amenities are represented by the effect on environmental characteristics, Q(D). We assume that development of the shale would also reduce environmental quality. 6 A homebuyer derives utility from these attributes and a composite good Y, and is expressed as U (Y, L, S, N, Q(D), G(D)). The homebuyer maximizes utility with respect to a budget constraint and the expected utility gained from these attributes in relation to the composite good Y.
The change in expectations reveals the value of shale gas development to the area with a moratorium and provides an event that can uniquely identify this change in expectations.
This valuation of expectations implicitly includes the negative and positive local amenities of shale development in Equation 1.
Two assumptions are required to interpret the change in prices as net valuation.
First, we assume that the probability of SGD is one without the moratorium and zero with it. As discussed in Section 2.3, given that leasing contracts were being signed and drilling permits were being applied for in New York and given the eventual well density in Pennsylvania near the border with New York (Figure 2), we think that residents believed that drilling would come to New York with certainty, especially those areas near the border with Pennsylvania. It is less clear whether expectations dropped to zero following the moratorium, given that it was initially temporary and then made permanent. If the perceived probabilities of the SGD are strictly within the bounds of one and zero, before and after a moratorium, then our model would underestimate the net value of shale gas development. We detail in Section 6.2 two robustness checks that allow for expectations to change slowly or change in the years after the moratorium and results are similar.

METHODOLOGY
We develop a model that identifies the impact of the New York statewide moratorium on housing prices, and thus reveals the net valuation of expectations about SGD. We employ a difference-in-differences model, which compares properties in New York before and after the moratorium to similar properties in Pennsylvania. As discussed in more detail in the next section, our preferred sample is comprised of properties within five miles of the state border with similar shale geology and only includes private well properties, which are most likely to experience SGD. This choice is motivated by intuition and prior findings; Gopalakrishnan and Klaiber (2014) and Muehlenbachs et al. (2014) find that private water properties are most price responsive to proximate drilling. We define the treatment group to be New York properties and the treatment is the moratorium.
Equation 3 is our main specification: where is the sales price of property i, is a binary variable equal to one if the property is located in New York, is a binary variable equal to one if the transaction occurs after the New York State moratorium on SGD, and is a set of housing, location, and temporal controls. also includes a constant to capture the omitted group of properties located in Pennsylvania that transact before the moratorium.
Finally, is the error.
The interpretation of the model coefficients is as follows. 1 is the premoratorium price difference between properties in New York relative to Pennsylvania. 2 is the price change from pre-moratorium to post-moratorium for Pennsylvania properties.
The key coefficient in Equation 3 is 3 , which is the double difference estimate. This term identifies the effect of the moratorium on New York properties, relative to Pennsylvania properties. As discussed in Section 3, our expectation about the sign and magnitude of this coefficient is ambiguous. It could be positive if New York households are concerned about the environmental dis-amenities of SGD and value the delay or ban of SGD. Alternatively, 3 could be negative if households anticipated economic gains from SGD and house prices had already capitalized that expectation. Lastly, 3 could be zero if the moratorium did not change expectations or perceived benefits and costs of SGD are small. 7 While the prior section laid out assumptions required to interpret coefficients as 7 One might think this type of specification and data could also be used to estimate the area level net value of SGD for Pennsylvania. However, we feel this is untrue precisely because expectations in both Pennsylvania and New York would muddle the comparison. A better comparison would be to compare Pennsylvania to some area with no possibility of SGD, with data prior to 2006 marking the pre-treatment time.
net valuation, there are also assumptions required for the difference-in-differences design to be valid. First, we assume that Pennsylvania serves as a good counterfactual for New York, in terms of house price dynamics. One potential concern is that areas in our study had different reactions to the US housing market collapse, which is correlated with the timing of the moratorium. In the Section 5, we show that our sample of Pennsylvania and New York homes follow a similar price trend pre-moratorium. Also, by focusing on observations close to the border, we hope to mitigate unobservable determinants of price trends. 8 Our sample choice of five border counties was meant to improve the treatmentcontrol comparison. Our refinement to focus in particular on properties within 5 miles of the border with similar shale thickness furthers the strength of the good counterfactual assumption. However, using bordering counties implicitly assumes that spillover effects are minimal. Spillover effects could be either environmental or economic. Environmental spillovers would occur if water or air pollution from SGD were to travel into New York from Pennsylvania. Evidence from Gopalakrishnan and Klaiber (2014) and Muehlenbachs et al. (2014) suggests that effects of water pollution are localized at about 2km. SGD in our study area is limited to the Susquehanna River Basin, which flows south. Thus, any surface water contamination is also likely to flow south further into Pennsylvania rather than north into New York. 9 Economic spillovers are increases in employment and spending across the border that indirectly or directly affect the housing market. Our estimates are unlikely to be affected by any economic spillover because our sample is restricted to a small area of just five miles on either side of the border, and thus can be thought of as a single labor market. 10 An additional argument that applies to these two types of spillovers is that New York residents and potential buyers would have formed expectations about drilling in Pennsylvania and those expectations would be capitalized into prices prior to the moratorium. Thus, while spillovers may occur, they should be expected and already accounted for in house prices.
Second, we assume that the treatment (moratorium) had no effect on the control (Pennsylvania). The main concern here is whether the moratorium on drilling in New York increased drilling in Pennsylvania. We argue that the pace of development in Pennsylvania (and elsewhere) was so rapid in the 2008-2011 timeframe that the lack of drilling in New York had no effect on prices or scarcity in Pennsylvania. Another way for drilling to be impacted would be if horizontal drills could cross state boundaries and extract New York gas from Pennsylvania, but this is in fact illegal. 11 Third, we assume that the implementation of the New York statewide moratorium was exogenous to the counties in this study. We believe this is a safe assumption for two true, then the New York moratorium may have increased prices in Pennsylvania. We argue that this effect would have been minimal given the evidence of highly localized environmental impacts of drilling from Gopalakrishnan and Klaiber (2014) and Muehlenbachs et al. (2014). 10 We additionally examined cross border migration to see if individuals relocated from New York to Pennsylvania after the moratorium. The results, presented in Figure A1  reasons. One, it is a statewide moratorium, not just a moratorium for the three sample New York counties, and much of the support for the moratorium came from regions in New York outside of this sample. Two, many of the sample towns were and still are against the moratorium as evidenced by the fact that 14 of 37 towns in our New York sample passed resolutions in support of SGD during the spring and summer of 2012, while only two towns passed a moratorium.

DATA
This study was conducted with property transaction data from five counties along the New York -Pennsylvania border: Chemung, Steuben, and Tioga counties in New York; Bradford and Tioga counties in Pennsylvania. We specifically chose these five counties because 1) the two Pennsylvania counties constitute one of the major clusters of drilling in that state, 2) all five counties are primarily agricultural and rural in character and thus make for good comparison, and 3) they border each other so that unobservable determinants of house prices likely follow similar dynamics.
We obtained transactions and property characteristics data from January identification straightforward. However, New York's data on water supply access is in parcel centroid format, which represents every parcel in the state by its center point.
Using parcel boundaries provided by county and regional planning departments, we connected sales data to water supply data using Geographic Information Systems (GIS).
However, a portion of our sold parcels do not overlay a centroid. In order to identify the water supply for each parcel in our transaction set, we follow Muehlenbachs et al. (2014) and create buffers of 100 meters around all public water supply parcel centroids. Then, we assume that all parcels falling outside of these buffers are dependent on well water. Figure A2 in the Appendix presents all transactions in our five counties by water type.
The figure makes clear several points. First, private water supply properties are almost exclusively outside of town boundaries. Second, Pennsylvania has a larger share of private water properties than New York. Further, there are very few public water properties within five miles of the border, especially in Pennsylvania.
In total, our original dataset includes 26,138 property transactions across all five counties from 2006 to 2012. We include only single-family residential and mobile homes with private water, which leaves us with 8,466 observations. 13 We drop all observations that sold for less than $10,000 or more than $1,000,000 in 2011 CPI adjusted dollars.
Further, we hypothesize that lot size is a key property characteristic for forming expectations about benefits and costs to SGD. Pennsylvania has larger lot sizes on average, so we drop observations that fall outside of the 5% and 95% of the lot size distribution to ensure common support between our Pennsylvania and New York samples. Lastly, we drop eight Pennsylvania transactions that occur prior to the moratorium that are located within two miles of a permitted well. We do this such that all transactions pre-moratorium have expectations about SGD, but no realized impacts.
Our analysis of the moratorium uses sales in the time span 2006-2011. 2006 marks the beginning of exploration and lease signings in Pennsylvania and New York.
At this point, both properties in Pennsylvania and New York will begin to capitalize expectations about the benefits and costs to hydraulic fracturing, but have yet to experience it. We use 2011 as a cutoff because local resolutions begin to be passed in early 2012. With these cuts, we are left with a sample of 4,976 transactions.
While choosing counties along the border goes a long way towards removing unobservable differences between New York and Pennsylvania observations, we develop four samples that further restrict observations. First, in the vein of a border discontinuity design, two samples are created that limit observations to be within 15 miles of the border and then within five miles of the border. These samples are intended to further minimize possible bias stemming from unobservable, time-varying processes that differentially affect housing prices across the state boundary. Second, we further restrict the 15-and 5-mile samples to only include properties that have similar shale geology. Figure 3 shows the thickness of shale deposits, which is a key driver of extraction potential. 14 On average, our Pennsylvania counties have thicker shale deposits than in New York, with thickness increasing towards the southeast. In order to ensure that expectations about SGD are similar on either side of the border, we restrict observation to be in the 100-200 feet range of thickness. Our preferred sample, shown by the dashed region of Figure 3, satisfies both the 5-mile border restriction and shale thickness restriction and includes 1,018 observations. We note that the only significant difference is in shale thickness between the samples which is dramatically reduced by using the restricted sample. There is also convergence of socioeconomic characteristics as we restrict our sample to tracts just along the border in our study counties, as shown in Appendix Table A4.
As discussed in Section 4, the critical assumption for our difference-in-differences design to be valid is Pennsylvania must a good counterfactual for New York. The most common way to support this assumption this is to compare pre-treatment price trends, and now having introduced the data, we can do just that. Figure 4 displays price trends for 2006 through July 2008 for the preferred sample. Price trends are similar between New York and Pennsylvania for private water properties, which further bolsters our confidence that the counterfactual created by the control counties is appropriate. In contrast, the pre-moratorium price trends for public water properties do not coincide, which motivated us to not use these properties in our difference-in-differences design.
One reason for the non-parallel trends could be the small number of Pennsylvania public water properties near the border. We could expand the sample in order to include more public water properties (and this does indeed improve the alignment of pretreatment trends), but that would defeat the purpose of the border discontinuity. In addition, we tested whether characteristics of transacted properties were different across states after the moratorium. The results presented in Table A3 of the online appendix show that most characteristics, most importantly lot size, are not statistically different across states. but discount the possibility of adverse health or environmental consequences, then our estimate may reflect lost benefits more than the net value. However, in Sections 6.2 and 6.3, we present results that bolster our confidence in these two assumptions. Table 3 provides a series of robustness checks that probe several key assumptions 15 Another way in which expectations can affect the calculation of net values is if households have a perceived duration of the moratorium. For example, many people (authors included) had the impression that the initial moratorium would last five years, and then the New York State government would make a decision. In this case, the house price reduction is not the lost value of SGD, but the cost of waiting five years for SGD. At a 5% interest rate, the annualized value of SGD would be $7,618. Given results in Section 6.2 that households do not seem to be updating their beliefs, we think the assumption that people believed the moratorium to be temporary is inconsistent with the data. However, future work could examine more recent data to determine if the permanent ban in 2014 led to any price change.

Robustness checks and extensions
of our difference-in-differences model and our interpretation of the results as revealing net value of SGD. 16 Each column builds on our main result of Column 5 Column 4 additionally requires transactions to overlay shale with similar depth. 17 Low depth areas of the shale have relatively lower reservoir pressure and higher water content, which may reduce potential oil and gas recovery. However, high depth areas of the shale 16 In the appendix, we present additional robustness checks. Table A5 shows the set of control variables. Table A7 estimates models similar to Table 2, but using the level of prices and results are similar. Table A8 tests for robustness of only including single family homes and using propensity score matching to trim the sample for covariate balance, and results are similar. Our research design also assumes that environmental spillovers across the border are minimal. In Column 5, we remove observations within 1 mile of the border in hope that any expected or actual spillover is contained to that distance. The estimated coefficient is -0.231, identical to the main result. 18 Our final robustness check, Column 6 in Table 3, further probes how well Pennsylvania serves as a counterfactual in terms of house price trends. Here, we use a triple difference specification and include properties that are served by municipal water supply. These properties are in towns and are unlikely to receive either the financial benefits or the main environmental damages associated with water contamination. Their purpose is to serve as an additional control for differential market trends in New York and Pennsylvania. However, our border discontinuity with a five mile distance requirement is designed to mitigate these types of unobservable variables. In addition, the pre-treatment trends did not run parallel (Figure 4), and there are very few public water properties on the Pennsylvania side (Appendix Figure A4). Thus, we do not believe this is the best research design, but investigate it as a robustness check. The triple difference term NY*PostMoratorium*Private, which is now the coefficient of interest, is interpreted as the change in price due to the moratorium for New York private water supply properties, relative to Pennsylvania private water properties, and relative to the differential change between public water properties in New York and Pennsylvania. The coefficient estimate is -0.179, which is smaller than the main result and less statistically significant. While this result is broadly consistent with the main result, the reduced magnitude is consistent with a pre-moratorium downward trajectory of municipal water properties in New York relative to Pennsylvania. In sum, the range of estimates from Table 3 support our assumptions and suggest that the moratorium reduced property values between 18% and 26% for New York properties overlaying high-quality shale deposits.
In Table 4, we extend our analysis of the effect of the statewide moratorium by examining whether there were heterogeneous price impacts based on lot size and future resolution status of a property's township. 19 Instead of using our preferred 5 mile distance band in this analysis, we apply a border distance band of 15 miles in order to include more towns that ultimately passed resolutions in 2012. Larger lots may be able to capture more financial benefits through leases and royalties. Owners and potential buyers may expect these financial benefits and so the price decline post-moratorium will be amplified for larger properties. The first column in Table 4 presents results from a specification that builds on our double difference framework by interacting lot size (log acres) with each of the three double difference variables. Resulting parameter estimates indicate that the double difference coefficient of interest is similar in magnitude to the results from Table   2, Column 3. The interaction with lot size is negative (as expected), but is statistically indistinguishable from zero. 20 We interpret these results as indicating that there is no heterogeneity across lot sizes, which we may be due to the high degree of uncertainty in financial benefits.
The second column of Table 4 explores whether the future resolution status of a town affects price impacts. As noted above, the statewide moratorium was unpopular in certain parts of New York. Towns that eventually pass resolutions likely supported SGD in the 2006-2008 range. Residents and homebuyers may have had larger expectations about the financial benefits, and thus may have a larger price decrease post-moratorium.
We estimate a specification that is similar to the preferred specification, but additionally interacts ResolutionTown (= 1 if a property is located in a town that passes a resolution in 2012) with the double difference variable NY*PostMoratorium. The results indicate that residential property prices declined -12.8% post-moratorium, similar to the estimate found in Table 2, Column 3. However, we find that this effect does not change in resolution towns. Although the parameter estimate on the triple difference term is negative (-9.8%), it is not statistically significant. This result suggests that the underlying likelihood of SGD as viewed by homeowners and prospective buyers is no different in resolution towns versus other towns. However, this modeling approach is focused on selection and underlying characteristics, not the effect of the resolution itself. In the appendix, we present an analysis of the property value impacts of passing a resolution.
Our results suggest no effect.

Proximity analysis
In this section we develop a more traditional model that estimates the effect of proximity to drilling on house prices. Similar to Gopalakrishnan and Klaiber (2014), we estimate variations on the following model: where is the number of wells within a given distance of a property, is a binary variable equal to one if the property depends on well water, and is a set of housing, location, and temporal controls. Gopalakrishnan and Klaiber (2014) estimate several specifications varying distance cutoffs and time windows, and we follow suit.
We implement both a one mile and two mile cutoff for the measure of proximity, and estimate the impact of new wells six and 12 months prior to transaction, as well as any time prior to transaction. We interact the binary variable with to allow for heterogeneous effects of proximity over private versus public water. We use all observations in our threes counties of Pennsylvania from 2006 to 2012. Identification comes from comparing houses within a given distance band of a drilled shale well to houses further away and houses that predated the drilling.
While this model has a traditional interpretation of valuation of proximity, this model also estimates the accuracy of households' pre-drilling expectations. Because all transactions in this analysis occurred after households began signing leases with gas companies, we claim that expectations of SGD net benefits had already been capitalized into sales prices. Under this interpretation, a post-drilling change in house prices would suggest that expectations were incorrect. Specifically, if prices decline post-drilling, as is the case in prior studies, this would indicate that households' expectations had either overvalued the financial benefits or undervalued the environmental costs.

35
The results presented in Table 5 suggest that proximity to wells has no effect on house prices. Across all six specifications, neither the public nor private well water coefficients are statistically different from zero. These results are qualitatively robust to changes in functional form. We interpret this to mean that ex ante expectations established in the initial expansion of SGD in Pennsylvania were on average accurate ex post. In turn, these results support our claim that New York households near the border have accurate expectations about SGD. However, an alternative interpretation is there is no economically significant benefit or cost of living in proximity to drilling.

CONCLUSION
This paper proposes to view hedonic analysis through a different lens by focusing on the formation, realization, and change of expectations to inform environmental policy.
We provide area-level impacts of the statewide moratorium on SGD for households For our preferred border discontinuity sample, our point estimate translates into a $25,531 loss in equity per house and an annualized loss of $1, 649 per house. In order to estimate the total property market impact to these three counties, we performed the following back-of-the-envelope calculation.   Observations are single family residences and mobile homes located outside of public water supply areas. Housing prices are CPI-adjusted to 2011 levels. The 5 mile border restrictions includes only those observations that are located within 5 miles of the NY-PA state border. The shale trim restriction includes only those observations that overlay shale that is 100 -200 feet thick. * if the differences in means divided by their combined standard deviation is greater than 0.25.  to 2011. All observations are located outside of public water supply areas and, thus, rely on private well water. The dependent variable is the natural log of sale price, CPI-adjusted to 2011 values. Property variables include number of beds and bathrooms, a quadratic in square feet of living area, a quadratic in property acreage, a cubic in property age, an indicator variable for mobile home, and an indicator variable for whether some property characteristics were imputed. Standard errors are shown in parentheses and are estimated using township-level cluster-robust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.  Unless otherwise noted, observations represent single family residence and mobile home property transactions from 2006 to 2011. All observations are located outside of public water supply areas and, thus, rely on private well water. All observations are located within 5 miles of the border and overlay shale from 100 to 200 feet thick. The dependent variable is the natural log of sale price, CPI-adjusted to 2011 values. Property variables include number of beds and bathrooms, square footage of finished living area and its squared term, a quadratic in property acreage, and a cubic in property age. Model 1 excludes observations from 2008 for a cleaner break in expectations. Model 2 tests for the robustness of the results with an additional year's worth of sales data from 2012. Model 3 drops all observations in the shale thickness band from 100 -150 feet, thus only leaving those in the 150 -200 feet thickness band. In Model 4, we add shale depth trim to our main model (4,000 -5,000 feet of shale depth) to better control for potential geological differences between the two states. Model 5 drops all properties that are located within a mile of the state border. Model 6 applies a triple difference approach, using properties in both private and public areas. Private is a variable that indicates whether a property is located in a private water supply area. Standard errors are shown in parentheses and are estimated using township-level cluster-robust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively. R-squared 0.400 0.395 Note: Observations represent single family residence and mobile home property transactions from 2006 to 2011. All observations are located outside of public water supply areas and, thus, rely on private well water. All observations are located within 15 miles of the border and overlay shale from 100 to 200 feet thick. The dependent variable is the natural log of sale price, CPI-adjusted to 2011 values. Property variables include number of beds and bathrooms, square footage of finished living area and its squared term, a quadratic in property acreage, and a cubic in property age. The first model interacts the DD with the natural log of acreage and replaces the quadratic in property acreage with the natural log of acreage. The second model interactions the DD with a binary variable that indicates whether an observation was located in a town that passed a local resolution in support of shale gas development in 2012. Standard errors are shown in parentheses and are estimated using township-level cluster-robust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.  to 2012 for $10,000 to $1,000,000 (in 2011 CPI-adjusted dollars). Property characteristics include # of beds, # of baths, quadratics for acreage and finished square footage, a cubic in age, and binary variables that indicate whether or not the property is in a private water supply area, whether it is a mobile home, and whether it is missing values for beds, baths, finished square footage, or age. Standard errors are shown in parentheses and are estimated using township-level cluster-robust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.

Appendix
This appendix provides information, data, and results that supplement, but are not critical to, the analysis in our main paper.
Tables A1 and A2 provide information regarding the monetary benefits and environmental costs, respectively, which may accrue to landowners in areas with active shale gas development. These supplement the discussion in Sections 2.1-2.2 of the main text. Although we focus on environmental costs, other costs from SGD could be social in nature, such as increased inequality and vulnerability of poor community members  Table A3 gives summary statistics for housing characteristics post-moratorium.
This table is identical in structure to Table 1 from the main text. The purpose of this table is to test whether characteristics of transacting properties differed after the moratorium. If the types of properties did change, then this would discredit our main difference-indifferences estimates. The table shows that transacting New York properties are older than Pennsylvania properties. The table also shows that shale thickness is greater in the New York sample, as it is pre-moratorium, and the distance to spud wells is greater in New York than in Pennsylvania, which is expected. Most importantly, lot size is not significantly different between the states. Thus, we see no reason to believe that changes in the types of properties that transact will affect our estimates. whether New York and Pennsylvania are significantly different from each other in terms of demographic, economic, and social characteristics. We break up the states into three subsets: (i) full state, (ii) study counties, and (iii) all census tracts in our study counties that are adjacent to the state boundary. By testing for statistical differences across different subsets of the states, we identify whether these characteristics converge as we move closer to our primary study area adjacent to the border. Results indicate that there are significant differences in demographic, economic, and social characteristics between New York and Pennsylvania at the state level. Although the absolute differences start to decrease when we restrict the census tract subset to only include those within our five study counties, we still observe significant differences in racial characteristics, rental occupancy percentage, and the unemployment rate. Once we restrict our sample to include only those census tract observations along the state border, we find that the only statistically-significant difference between the two states is median age. This is an important result, as it strengthens our argument that by focusing our analysis along the state border, we improve the comparability between our treatment and control groups.  estimates on all covariates included in our models. All estimates have the expected sign.
In Table A7, we estimate Table 2 from the main paper using Linear-Linear models with the understanding that the value of shale development is likely to be similar across the spectrum of sale prices. Our model estimates indicate that New York 54 properties that are most likely to experience shale development value it in the range of $10,000 to $28,000. Given mean sale price values for our subsample, the results are qualitatively similar whether you use a Log-Linear or Linear-Linear, making the average effect of the moratorium the same regardless of using logs or levels. Table A8 provides a series of robustness checks on our results from Table 2, Column 5 (additional to those found in Table 3), while Table A9 tests the robustness of our results from Table 2, Column 3 (by estimating a similar series of models from Table   3). Results from both tables provide additional support to those found in the main text. Figure A1 examines household movement across the border. One threat to our identification strategy is spillovers across the border caused by the moratorium. Using IRS migration data from our five sample counties, we compare the number of people moving from New York to Pennsylvania and vice versa. Figure A1 shows parallel trends both before and after the moratorium, giving credence to our identification assumptions. Figure A2 shows the locations of residential property sales in our study area from 2006 to 2011 by water type. Figure A3 shows the locations of towns in our study area that have passed SGDsupporting townships, while Figure A4 shows shale depth across our study area with an outlined 5-mile state border restriction buffer.

Analysis of Property Price Impacts of Local Resolutions in Favor of SGD
In addition to examining the effect of the statewide moratoria, we explore the effect of town-level resolutions in support of SGD on housing prices. While local resolutions had no direct effect on actual drilling activity, it is possible that passing a 55 resolution did affect expectations about the future likelihood of local SGD, especially in light of the state's position on home rule authority and the proposal floated in 2012 that would allow drilling only in communities that support it. In addition, bans were more popular in New York as a whole and towns in several counties north of our sample counties were very active with bans and moratoria (e.g., Livingston, Ontario, Tompkins and Yates counties).
Similar to the moratorium, these resolutions mark a potential change in expectations, especially when seen in the context of the greater number of local bans occurring elsewhere in the state and the floated policy to allow SGD in southern NY counties that expressed support, and this change in expectations could be capitalized into housing prices. Figure A3 shows the geographic distribution of local resolutions in our sample counties.
Like our moratorium analysis we work in a difference-in-differences framework, but now we use only transactions occurring in New York after the statewide moratorium, still restricting the sample to private well water properties. The dimensions of comparison now are location within a town that passes a local resolution versus location within a town that did not pass a local resolution and pre-versus post-resolution passage.
Treatment is passage of the local resolution. Equation 4 is the model we estimate: where is sales price, is a binary variable equal to one if the property is located in a town that passes a resolution in support of SGD, and is a binary variable equal to one if the transaction occurs after July 31, 2012. Because resolutions were being passed at various times March -July 2012 and it often took multiple meetings of discussion before passing, we exclude observations January -July 2012 so that the pre-post comparison is clean.
The coefficient of interest in Equation 1 is 3 , which is the double difference estimate. This term identifies the average change in prices for properties in resolution towns from before to after the resolution, relative to price changes for other New York properties located in towns that do not pass a local resolution. As discussed above, we expect this coefficient to be either zero or positive. It could be zero if passing a resolution does not change expectations about the likelihood of SGD in these areas. 3 will be positive if expectations do change and residents of the resolution towns feel the net benefits of SGD are positive, which seems likely since the residents passed the resolution in the first place.
The assumption of exogeneity may not hold with local resolutions as it does with the statewide moratoria. There could be characteristics of towns that pass resolutions that are correlated with house price changes. 21 While estimates from Equation 4 cannot be interpreted as causal, we still feel there is value in investigating the correlation of prices and resolution decisions.

Lease Signing Bonus
An initial payment, based on acreage, for signing a gas lease contract (Weidner, 2013); Most important element of lease given uncertainty of natural gas production (Hefley et al., 2011) $2700 per acre is average signing bonus amount, though this can vary considerably (Hefley et al., 2011); Average per acre payments can range from $50 to almost $6,000 (Green, 2010;Eichler, 2013;Riely, 2014) Royalty Payments Recurring payments on a proportion of natural gas extracted from the property (Weidner, 2013) State law guarantees minimum 12.5% royalty rate, though the negotiated rate may be higher (Weidner, 2013)  Air Quality Impacts Conventional air pollution increases associated with shale gas extraction and processing (e.g., Litovitz et al., 2011) Local, Regional -Changes in overall greenhouse gas emissions due to natural gas displacement of coal use in energy production (e.g., Howarth et al., 2011;Cathles et al., 2012) Global N/A

Water Impacts
Groundwater pollution due to shale gas development activities (e.g., Osborn et al., 2011;Jackson et al., 2013) Local, Regional -Surface water contamination due to shale gas development waste disposal and management processes  Local, Regional -Significant pressure on wastewater management capacity (e.g., Rahm et al., 2013) Local, Regional -

Miscellaneous Effects
Shale gas infrastructure-driven forest and farmland fragmentation (e.g., Racicot et al., 2014) Local, Regional -Decreases in wildlife habitat availability and quality in and around shale gas infrastructure (e.g., Kiviat, 2013) Local, Regional, Global -Increased seismicity in and around unconventional wells and wastewater injection sites (Frohlich, 2012) Local, Regional -66 Observations are single family residences and mobile homes located outside of public water supply ares. Housing prices are CPI-adjusted to 2011 levels. The 5 mile border restrictions includes only those observations that are located within 5 miles of the NY-PA state border. The shale trim restriction includes only those observations that overlay shale that is 100 -200 feet thick. * if the differences in means divided by their combined standard deviation is greater than 0.25.   Property variables include number of beds and bathrooms, square footage of finished living area and its squared term, a quadratic in property acreage, and a cubic in property age. Standard errors are shown in parentheses and are estimated using township-level cluster-robust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.   values. Property variables include number of beds and bathrooms, square footage of finished living area and its squared term, a quadratic in property acreage, and a cubic in property age. The Shale Thickness restriction includes only those observations that overlay Marcellus Shale that is 100 -200 feet thick. The Shale Depth restriction includes only those observations that overlay Marcellus Shale that is 4000 -5000 feet below the surface. Standard errors are shown in parentheses and are estimated using township-level cluster-robust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.  All observations are located outside of public water supply areas and, thus, rely on private well water. All observations are located within 5 miles of the border and overlay shale from 100 to 200 feet thick. The dependent variable is the natural log of sale price, CPI-adjusted to 2011 values. Property variables include number of beds and bathrooms, square footage of finished living area and its squared term, a quadratic in property acreage, and a cubic in property age. Model 1 includes only observations classified as single family residences (no mobile homes included). Model 2 applies propensity score matching, where we use a probit model of all property characteristics listed above and 100-150 v. 150-200 shale thickness bins to predict NY. We then cut all observations outside of the 5% and 95% of the distribution of the propensity score. We then run our main specification on the reduced dataset. In Models 3 and 4, we cut all observations within a mile of the border from NY and PA, respectively. This is done to test for the effect of environmental spillovers (see Model 6 in Table 3 in the main text for an analogous regression where both sides are cut from the model). Instead of using a shale thickness restriction, Model 5 keeps only those observations that are 4,000 -5,000 feet below the surface, which is in the common support of both states (see Marcellus Center for Outreach and Research at Pennsylvania State University for maps). Standard errors are shown in parentheses and are estimated using township-level cluster-robust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.  In Model 4, we use a shale depth trim instead of a shale thickness trim (4,000 -5,000 feet of shale depth). Model 5 tests for the impact of environmental spillovers by cutting all NY and PA observations that are located within a mile of the state border. Model 6 applies a triple difference approach, using properties in both private and public areas. Private is a variable that indicates whether a property is located in a private water supply area. Standard errors are shown in parentheses and are estimated using township-level cluster-robust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.  Observations represent single family residence and mobile home property transactions sold in NY from the NY moratorium in 2008 to late 2013. All observations are located outside of public water supply areas and, thus, rely on private well water. Due to the uncertainty of actual resolution dates, we cut all observations from January to July of 2012, which is in the range of all resolution passages that we have information on. The dependent variable is the natural log of sale price, CPI-adjusted to 2011 values. Property variables include number of beds and bathrooms, square footage of finished living area and its squared term, a quadratic in property acreage, and a cubic in property age. ResolutionTown indicates whether a sold property was located in a township that eventually passed in a resolution, while PostResolution indicates whether the property was sold after July 2012. Standard errors are shown in parentheses and are estimated using township-level clusterrobust inference. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.

Introduction
Shale and tight oil and gas basins have emerged as important sources of energy in the United States through innovations in hydraulic fracturing and horizontal drilling. This development has led to significant impacts on residents and landowners that are close to drilling activities. There are environmental and health risks associated with unconventional oil and gas development related to groundwater contamination (Osborn et al., 2011), surface water pollution , wastewater management (Rahm and Riha, 2012), and infant health (Hill, 2013;McKenzie et al., 2014). However, there are positive impacts as well, especially for those residents that lease their mineral rights to drilling companies in return for royalty payments and lease signing bonuses (Fitzgerald, 2014;Hardy and Kelsey, 2015). 1 Royalty payments can be paid as high as 20% of the value of production and lease signing bonuses can reach into the thousands of dollars per acre leased (e.g., Brasier et al., 2011;Kelly-Detwiler, 2013).
There is a growing body of literature in economics that seeks to estimate the local impacts of unconventional oil and gas development through hedonic valuation, with remarkable variation in estimates. Gopalakrishnan and Klaiber (2014) examine price responses in southwestern Pennsylvania and find that prices can decline by as much as 22% for private well-water dependent properties, but generally are smaller and short lived.  use data from across Pennsylvania and find that shale development can lead to depreciation as high as 17% for well water properties within 1.5 1 The United States is unique in that private citizens can own the minerals underneath their property. In most countries, it is the government or the crown that owns all subsurface rights . Since U.S. citizens can own their properties' minerals, they can financially benefit from drilling through leasing. The possibility of financial benefits has led to more public support for unconventional oil and gas development in the United States than in other countries (e.g., Stevens, 2010;Gény, 2010). Private mineral rights ownership has been crucial in driving development in the United States (Wang and Krupnick, 2013). kilometers of an unconventional drill site. They also find that drilling can have a small positive impact on property sale prices in public water supply areas. Delgado et al. (2016) use data from Northeastern Pennsylvania and find no robust impact of nearby drilling. Boslett et al. (2016)  Importantly, their estimates are attenuated in areas with significant mineral severance. In Colorado, Bennett and Loomis (2015) find mixed results that are often not statistically significant. Also in Colorado, James and James (2015) find that a one kilometer decrease in distance away from an unconventional well is associated with a 7 to 20% decrease in sale price. However, this effect can be mitigated if the property is above a horizontal well lateral, which suggests royalty payments from leasing are capitalized into prices. Using zip code level aggregate data from Texas, Weber et al. (2016) find evidence of both appreciation and depreciation due to shale development. There remain large variations in estimates of shale gas development on property prices without sufficient explanation.
One potential reason for the contrasting results is that mineral right ownership is not accounted for explicitly in these studies. Mineral ownership is important because it firmly demarcates property owners who financially benefits from unconventional gas development versus those who do not. Without mineral ownership information, it is not known whether the recovered valuation is a net impact of financial positives and environment/health negatives, purely the negatives, or a weighted average of the two. In Pennsylvania, which is the focus of several of the above studies, mineral right ownership varies substantially and systematically across the state. In Western Pennsylvania, where there is a legacy of energy extraction, mineral rights are more often severed from surface rights than in northeastern Pennsylvania, where energy extraction only started recently . The valuation studies focused on Eastern Pennsylvania estimate less negative (or even positive) impacts of shale gas. 2 While the important distinction between private well water and public water properties has been identified by the existing hedonic valuation literature, the critical issue of mineral rights ownership has not been resolved.
This has not been a lack of foresight or understanding by researchers of the importance of mineral ownership in how people are impacted by drilling. Rather, this information is extremely difficult for researchers to obtain. 3 In this paper, we resolve the issue of unknown mineral rights by exploiting a historical severance in mineral rights ownership, the Stock-Raising Homestead Act of 1916 (SRHA1916). Over the course of the 19 th century, the United States expanded its boundaries through territorial acquisition, often exchanging land rights for homesteading starting with the Homestead Act of 1862. In the late 1800s and early 1900s, the federal 2 Consistent with this interpretation of the valuation results, homeowners that do not own mineral rights have an increased perception of environmental risk  and frustration (Collins and Nkansah, 2015). 3 Mineral right ownership information is held in county deeds offices and is not commonly included in property deeds. The chain of title can be unclear, especially when the mineral estate was historically separated from the surface estate. One may need to go back to the original land grant to confirm the title of the mineral estate. Charting mineral rights ownership over time is the full-time job of a title abstractor (Wilson, 2014). Suffice it to say, it would be difficult for researchers to successfully obtain mineral rights ownership information for a large property transaction database. One of the authors of this study spent a day at Pennsylvania's Bradford County's Register and Recorder office researching mineral rights transfers and can attest to this. See Table 1A in the Appendix on a review of the previous literature and how each paper contextualized the issue of mineral rights ownership. government recognized both the increasing value of energy for economic growth and the government's inability to properly identify "mineral" lands, which they kept in federal ownership, and "nonmineral" lands, which they disbursed for homesteading. In response, the federal government passed the SRHA1916, which continued the tradition of land disbursement, but the federal government retained ownership of minerals in all land disbursed after 1916 (Gates, 1977;Harrison, 1989).
To build our dataset of transactions, we identify residential properties in Colorado located on land originally distributed under the SRHA1916. Thus, the federal government owns the mineral rights for each of these properties and current residents do not benefit financially from lease and royalty payments. Our study area is on the western slope of Colorado, centered in Garfield, Mesa, and Rio Blanco counties. Western Colorado was one of the major areas of post-1916 homesteading and this region is one of the primary locations of unconventional oil and gas development in the state.
Our hedonic analysis suggests that houses within one mile of an unconventional drill site sell for 34.8% less than comparable properties without proximate drilling. This result is robust across various subsets of the data and alternative regression specifications, including a repeat sales model and a matching model. When multiplied by the average house price of $183,300, this discount translates to a price reduction of $63,788, which equals $3,952 when annualized by a 30-year mortgage and a 5% interest rate. We interpret this price difference as the household valuation of the external environmental and health costs associated with proximity to unconventional oil and gas development.
Our findings corroborate the negative valuations found in other papers, but are much larger -60% larger in magnitude than the largest existing negative estimate. This 82 disparity demonstrates the importance of understanding mineral rights, as financial benefits of drilling are capitalized into housing prices and can adulterate estimates of external costs. Supporting this conclusion, we also estimate hedonic models using Western Colorado properties with unknown mineral rights ownership, mirroring the setup of prior studies, and find much smaller and statistically insignificant impacts of proximity.
Our paper is structured as follows. We start in Section 2 with a conceptual framework that discusses the issue of mineral rights ownership in valuation of unconventional oil and gas development. In Section 3, we outline the history of SRHA1916. In Section 4, we discuss our data set and how we obtained it. We then follow with a discussion our methodological approach and the assumptions we use in our interpretation of our model results. In Section 5, we present our results, and Section 6 concludes.

Conceptual Framework
In this section, we outline the potential biases in hedonic valuation of unconventional oil and gas development. The net benefits and costs of oil and gas development vary across mineral estate classifications. Only residential properties that are unified with their property's mineral estate can receive direct financial benefits, such as a lease signing bonus and production-based royalties. However, all properties receive the environmental costs of nearby oil and gas development. Previous hedonic valuation work has not incorporated the ownership of mineral rights into valuation frameworks due 83 to the lack of data. Thus, these studies have typically contextualized their estimates as net valuations of local oil and gas development. 4 We define the price of property i as ) In this framework, mineral ownership is defined as a binary variable, i M , equal to 1 if the property's surface estate is connected with its mineral estate and equal to 0 otherwise. The net valuation of unconventional oil and gas development for property i can be decomposed as follows: (1) Thus, the valuation of development, as capitalized by housing prices, is a net valuation of the environmental impacts of development (first term) and the financial benefits of development, contingent on ownership of the property's mineral estate (second term). For the sake of simplicity, we re-write the equation as: where C and B refer to the true values of the capitalization of the financial benefits and environmental costs (Table 1). If a property has mineral rights, then the estimated valuation will be the net impact, B C  . However, if a property is a split estate, then the estimated valuation will be the external costs of being proximate to drilling.
In order for researchers to estimate valuation with real data, a residential property market level analysis is required. We define  as the proportion of properties unified with their minerals. Using the terminology from Table 1, we have: The estimated valuation is conditional on the proportion of treated properties with mineral rights, which is unknown to the researcher. Knowing either B C  or C is useful for understanding local impacts and for guiding policy. But a weighted average of the two with an unknown weight yields imprecise guidance.
In this paper, we take advantage of the historical split in mineral rights from surface rights caused by SRHA1916, which is detailed in the next section. Thus, in our sample,  = 0 and we can isolate the negative environmental costs, C, from the financial benefits of development. This framework assumes there are no area level impacts of oil and gas development on properties which differ by mineral rights ownership, which would impact the estimate of proximity to wells. This potential issue is easily addressed with information about mineral rights ownership over the entire population of sale, or restricting the sample to only those properties without mineral rights ownership, i.e.  = 0, as we do in this study.

The Stock-Raising Homestead Act of 1916
In the late 1800s and early 1900s, the federal government passed a series of homestead acts that encouraged immigration and development in the western United States. The exact mandate varied across the acts; typically, settlers received 160-640 acres in return for a promise of ranching, cropping, or timber management on the property (Bureau of Land Management, 2006).
Early homestead acts apportioned both surface and subsurface rights to homesteaders. This was the case for the original Homestead Act of 1862, signed by President Abraham Lincoln after the separation of the Confederacy from the United States. These lands were selected for disbursement due to their perceived "nonmineral" character, as delineated by General Land Office's entrymen and surveyors. All lands with mineral potential were retained by the federal government; however, demarcation was imperfect due to technology limitations (Leshy, 1987).
Recognizing its limitations in mineral assessment, the federal government passed the SRHA1916. This policy effectively discontinued the mineral-nonmineral classification system (Harrison, 1989). Homesteading individuals were granted no more than a section of land for ranching and forage crop production, conditional on making permanent improvements on the land within three years of the entry date. However, the federal government would retain mineral ownership of all lands disbursed through the Stock-Raising Homestead Act: "That all entries made and patents issued under the provisions of this Act shall be subject to and contain a reservation to the United States of all coal and other 86 minerals in the lands so entered and patented, together with the right to prospect for, mine, and remove the same." 5 In addition, homesteaders and other later surface right owners had to allow access to the land for subsurface exploration and production: "Any person who has acquired from the United States the coal or other mineral deposits in any such land, or the right to mine and remove the same, may reenter and occupy so much of the surface thereof as may be required for all purposes reasonably incident to the mining or removal of the coal or other minerals, first, upon securing the written consent or waiver the homestead entryman or patentee; second, upon payment of the damages to crops or other tangible improvements to the owner thereof…or, third, in lieu of either of the forgoing provisions, upon the execution of a good and sufficient bond…to secure the payment of such damages to the crops or tangible improvements of the entryman or owner." 6 This condition precludes the surface owner from preventing mineral exploration.
The intent of the act was to continue the practice of homesteading and agricultural development of the west without compromising the federal government's interest in mineral exploration (Tanke and Putz, 1982). If the surface owner had the right to prevent production from the property, the mineral estate would have no value. More recent 5 The Statues at Large of the United States of America from December, 1915to March, 1917. Session 2. Chapter 9. Section 9. Page 864. 6 The Statues at Large of the United States of America from December, 1915to March, 1917. Session 2. Chapter 9. Section 9. Page 864. legislation has maintained surface use access for mineral rights interests while providing some limited protection to surface owners through accommodation doctrine. This principle allows only "reasonable" use of the surface by the mineral owner. The mineral owner may only access the surface if there are no other alternatives that could avoid interference with the present surface uses (Johnson, 1998). However, mineral rights dominance has largely been kept in place, as it is difficult for surface owners to prove that the use of the surface by the mineral owner is not reasonable .
On privately-owned lands with federal mineral ownership, the mineral lessee must make a "good faith effort" to secure surface owner consent to access the property (Bureau of Land Management, 2007). However, the surface estate owner is only entitled to compensation associated with damages to crops and agricultural-related improvements.
Thus, in the context of shale gas extraction, a homeowner may be exposed to water contamination, air pollution, noise, and visual disamenities, but not be entitled to compensation. 7 The SRHA1916 led to a significant amount of private land with federal mineral ownership in the western United States. Out of approximately 300 million acres conveyed to private individuals through the various homesteading acts (Loomis, 2002), nearly 60 million acres have been split from their underlying subsurface estates as a result of the SRHA1916. In Colorado alone, these lands total 5.2 million acres. Figure 1 shows lands with federal mineral ownership across Colorado.
In a split estate, the surface property owner cannot financially benefit from drilling. This situation is common in areas of the country that have experienced historical energy development and mining (e.g., Pender et al., 2014;Railroad Commission of Texas, 2015). Surface ownership can be split from its mineral rights through the issuance of a severance deed, in which a private landowner sells the land but retains the minerals. This situation is analogous to mineral ownership law in other countries: private land owners neither control the course of oil and gas development, nor financially benefit from it. The mineral estate is essentially dominant, which means that the surface owner must allow access to the mineral owner for exploration.

Empirical Setting and Data
Colorado has a long history of oil and gas development. According to data from the U.S. Geological Survey (Biewick, 2008), most Colorado counties have experienced oil and gas exploration since the early 1900s. Some counties, especially those outside of the intermountain areas, have seen a large increase in oil and gas development since the 1940s.
More recently, the new technologies of horizontal drilling and hydraulic fracturing have been used for extraction in Colorado. According to Drillinginfo, over 11,000 horizontal wells were drilled in Colorado between 2000 and 2014. Figure 1 displays the spatial distribution of this drilling. There are generally three drilling hotspots: northwestern Colorado, Weld County on Colorado's Front Range, and La Plata and Montezuma counties in southwestern Colorado. Our study focuses on northwestern Colorado because there is extensive federal mineral ownership in this area. Bennett and 89 Loomis and (2015) and James and James (2016) examine impacts in Weld County, which is more densely populated, but has much less federal mineral ownership that northwestern Colorado. Montezuma County contains many tribal lands, which complicates analysis due to jurisdictional complexity of mineral development and policy (West, 1992).
In this study, we use residential property transaction data from Garfield, Mesa, and Rio Blanco counties in western Colorado. We received this data from each county assessor's office. All transactions in our analyses occurred from 2000 to the end of 2014.
All transaction data contains property characteristic information including the number of bedrooms and bathrooms, living area, age of the property, and its classified property or land use. We include only those transactions that are defined as residential or agricultural with a residential building (N = 55,114). All mobile homes are dropped from the analyses (N = 2,819). 8 The data allow us to observe multiple sales per property, not just the most recent. All transactions that had more than seven bedrooms were dropped out of concern that the properties were apartment buildings (N = 16).
In order to focus on properties without mineral rights, we use the federal mineral ownership data from the Bureau of Land Management's Colorado GIS office. 9 We overlay federal oil and gas ownership data layers with parcel boundaries. We include properties in our final sample that are completely contained within the federal mineral ownership boundaries (N = 871). 10 While this cuts 98% of the transactions in our 8 Results from our main models are robust to including mobile homes and are available in the online appendix. 9 "Statewide Federal Mineral Ownership" data product. 10 There were data scale and alignment issues between our parcel data and our mineral ownership data. As a result, our restriction is conservative. There are likely other properties that are severed from their oil and gas rights by the SRHA but are not fully within the federal mineral ownership extent. We test our results database, this restriction is necessary to identify the subset of properties that do not benefit financially from nearby drilling. It is almost certain that some properties outside of the federal mineral ownership boundary do not own mineral rights, but it is unknown on a property basis. Thus, this is the sample that provides the best estimate of the external costs of local oil and gas development. As shown in the results section, our coefficient estimates change considerably when we include properties with unclear mineral rights ownership. Lastly, we cut all observations in the below the 5 th and above the 95 th percentile of the sale price distribution to remove the influence of outliers. Our final sample is 783 transactions.
We received directional and horizontal well development data from Drillinginfo.
In our three counties, there were 4,374 horizontal wells drilled from 2000 to 2014. Figure   2 maps the distribution of these wells relative to federal mineral ownership. In ArcGIS, we calculate the distance to the closest well at the time of sale. In addition, we calculated additional spatial statistics associated with distance to the closest municipality (U.S. Census definition) and the percentage of the property in an agricultural use from the National Land Cover Dataset 2001 to use as control variables in our regression model. Table 2 provides summary statistics for our preferred sample of split estate properties, as well as the complete set of properties. Properties with federal mineral ownership are less expensive, relative to the population of properties. Split estate properties typically are less expensive, are larger in lot size, and have relatively less agricultural land. Interestingly, split estate properties are also more likely to have a using less restrictive definitions of a split estate (i.e., 75 and 90% of property within federal mineral ownership boundary) and find robust results (available in the online appendix). nearby horizontal well. 11 As discussed in the introduction, prior hedonic valuation papers have focused on differences in impacts between municipal water and private well water properties (Gopalakrishnan and Klaiber 2014, Boslett et al. 2016). In Colorado, data does not exist on which properties have private vs. public water supply.
However, our primary sample of split estate properties are typically outside of municipal boundaries, which is the best proxy for public water supply. 12 Therefore, our estimates are applicable to households that face the risk of groundwater contamination and are comparable to studies that focus on private water properties.

Methodology
We use the hedonic price method to estimate the effect of drilling proximity on housing prices. Our basic specification is: p is the sale price of property i in spatial unit s in year t. Prices are adjusted to 2015 levels using the Consumer Price Index.
ist Well is a binary variable that indicates whether property i has an unconventional well within a given distance buffer at the time of sale.
' ist X is a vector of structural, locational, and environmental explanatory variables. s  are spatial fixed effects; we present specification using both county and census tract as the 92 spatial unit. t  are year fixed effects. Collectively, these fixed effects control for unobserved price determinants across space and time.
ist  is the error.
 is our coefficient of interest and is interpreted as the impact of having an unconventional well within r miles of the property on residential sale prices. Since we are only analyzing properties with federal mineral ownership, it reflects the marginal value of the environmental costs of having an unconventional well within the given spatial buffer.
Our goal is to define the radius r so that it captures the full spatial extent of negative externalities, but this is a priori unknown. Following Linden and Rockoff  Klaiber (2014) and , who use a one mile and a two kilometer buffer, respectively, in their analyses. The similarity in coefficient estimates between the 0-0.5 mile and 0.5-1 mile bins is at first surprising because we would expect greater externalities closer to the drill sight. However, Figure 3 also displays the frequency of observations by distance and we see clustering of observations around 0.5 miles, which is why it is difficult to discern differential impacts less than and greater than 0.5 miles.

93
Going forward, we define r equal to one, and hence ist Well is a binary variable equal to 1 if there was a well drilled within one mile of the property before its sale.

Assumptions
There are a number of assumptions needed to interpret our estimates as the valuation of the environmental costs of unconventional oil and gas development. First, we assume that the assignment to treatmentin this case, close proximity to a horizontal wellis exogenous. It is unlikely that split estate owners can strongly dictate whether drilling happens, as the surface owner cannot prevent the mineral estate owner from accessing subsurface resources. Impacted parties can protest the inclusion of a parcel in an oil and gas lease sale (Bureau of Land Management Regulation 43 CFR 3120.1-3), but we were unable to find any protests directly from homeowners in our three study counties. 13 A potential concern is that early oil and gas development could influence the likelihood of severance. In our case, the likelihood of severance is not a concern because of our study area's historical severance of mineral rights.
Second, we assume that property buyers and sellers have full information about local drilling activity and its potential for environmental impact. This is reasonable given the scale of planning and land disturbance associated with well permitting and drilling (e.g., Moran et al., 2015) and significant local discussion regarding the impacts of drilling (e.g., Williams, 2008;Lustgarten, 2009;Harmon, 2014). This region of Colorado has 13 There are examples of homeowners protesting the inclusion of a parcel in a 2012 lease sale in Delta County, outside of our study area. Most protests submitted to the Bureau of Land Management concerned environmental and social issues associated with drilling on publically-owned land, not split estate concerns on privately-owned land.
also been the primary site for multiple assessments of the health risks associated with unconventional development (e.g., Kassotis et al., 2013). As a result, home buyers are likely conscious of the potential benefits and costs of unconventional oil and gas development.
Third, property buyers and sellers are informed of the property's mineral right status and they understand its ramifications. There has been significant mineral development in western Colorado over the last century (Biewick, 2008). As a result, citizens are likely familiar with both oil and gas development and mineral ownership laws. The state legislature also passed a law in 2001 that required pre-transaction notification of the potential for a split estate (Garfield County Energy Advisory Board, 2007). There has been much public discussion since 2000 regarding a law that would mandate disclosure of mineral severance prior to all real estate sales. Although this legislation has not been passed due to issues secondary to the disclosure requirement (Moreno, 2011), it has been a major point of policy discussion in the state legislature over the study's time period. Additionally, federally-owned minerals were never disbursed with the land, so a title search is relatively quick and the information is publically accessible on the Bureau of Land Management's website.
Fourth, we assume that our estimates are not impacted by positive spillover effects of unconventional oil and gas development, including labor opportunities and improved public finances. Although these can be important benefits of local oil and gas development (e.g., Weber, 2012; Newell and Rami, 2015), these benefits are likely to be received at the regional level and are unlikely to be related to drilling adjacency or to mineral rights ownership.

95
Fifth, we assume that the financial benefits that a split estate owner can receive from local development are negligible. Landowners who do not own mineral rights are unable to receive lease or royalty payments from on-site production. However, landowners can receive compensation through a surface use agreement, which formally outlines where drilling and surface disturbances can happen on the property. Drilling companies must make a "good faith effort" to come to a surface use agreement with a landowner in a split estate situation. If the two parties are unable to come to an agreement, then the drilling company can rely on a performance bond (Bureau of Land Management, 2007). This alternative option does not cover all damages that a surface owner may face from drilling, such as contaminated drinking water, drilling-related noise, air pollution, and visual changes to the landscape. Additionally, performance bonds are typically meant to compensate for damages to cropland, but not pastureland (Fitzgerald, 2010). Since lands subject to SRHA1916 were originally intended for ranching and were deemed largely unsuitable for cropping, it is unlikely that property owners are reliant on a performance bond to indemnify them against damages related to drilling. It is more often than not that the surface owner will come to a surface use agreement with the drilling lessee (Hill and Rippley, 2004;BLM, 2006;Fitzgerald, 2010

Main Results
In

Robustness checks
In Table 4, we test for the robustness of this general result across alternative specifications and subsets of the data. In Column 1, we include an additional control variable, which is the number of vertical (conventional) oil and gas wells drilled within a mile of a property between 1980 and 1999. The concern is that past drilling is likely correlated with unconventional drilling, and thus if there is a negative impact on prices of past drilling, our estimates in Table 3 could be misattributing the variation from past drilling to current drilling. Since recent vertical well development may be an exploratory precursor to later horizontal well development, we use pre-2000 data to avoid potential endogeneity issues. In Column 2, we restrict our sample to only properties that are within 20 miles of an eventual unconventional well site. One concern with our full sample used in Table 3 is that some observed sales are far from drilling and may be a poor control group. By restricting the spatial distance, we hope to mitigate any bias that results from distant control observations. Column 3 further restricts the sample to be within 10 miles of an eventual unconventional well site. In Column 4, we restrict observations to be within 2005 to 2014. No proximate drilling took place within one mile of any sample properties prior to 2005. In our full sample, these pre-2005 properties are purely control observations. If there are structural changes in the housing market not captured by year fixed effects, then these observations may not be a good control and our estimates may be biased. In Column 5, we exclude all observations that sold more than once over our time period. Although fifteen years is a long time, there may be unobservable differences in the price dynamics of properties that sold multiple times over the study's period. In 98 Column 6, we estimate a repeat sales model and include only properties that sold more than once. Including property-level fixed effects better controls for unobservable property characteristics that could be correlated with proximity to drilling.
The coefficient estimates across these six columns are largely consistent with the main results. Magnitudes range from -0.307 to -0.381 and all estimates are statistically significant. Table 4 indicates that our estimates of the effect of drilling with one mile of a residential property are largely robust to alternative specifications and subsets of the data. 15

Matching analysis
In this section, we shift to a matching approach in order to better control for observable differences between our control and treatment groups (e.g., Abbot and Klaiber, 2013; Ghanem and Zhang, 2014;Ferraro et al., 2015). The main goal of matching is to avoid the issue of selection bias and to create valid treatment-control comparisons through pairing on observable covariates (e.g., Caliendo and Kopeinig, 2008;Angrist and Pischke, 2009). This occurs when the estimated relationship between treatment status and outcome is driven by inherent differences in covariate distributions between treatment and control groups.
We use matching to further test the robustness of our regression model results. 15 In the Appendix, we provide additional robustness checks. In Tables 2A and 3A, we find qualitatively similar results when we define our treatment variable as the number of wells drilled within one mile of the property's extent, or when we use a distance bin approach. At the suggestion of a reviewer, we estimate our models in levels, as opposed to logs, and find similar results (Table 3A). In line with the structure of Table  4, we provide additional robustness check in Table 5A. All results from the Appendix support our main findings.
We first estimate a propensity score model of the probability of treatment as a function of property-specific variables: Well is a dummy variable equal to 1 if property i has an unconventional well within r miles of the property. In this case, ' it X is a vector of structural and locational explanatory variables used in our regression models, along with the number of vertical oil and gas wells drilled within a mile of the property from 1980 to 1999, as in Column 1 of Table 4. Including year and tract fixed effects would be ideal to control for temporal price trends and spatial unobservables, but given our limited sample size these match criteria are infeasible.
The propensity score is calculated using estimated coefficients from Equation (5).
We then match treated observation to control observations using nearest neighbor matching with replacement. We match each treatment observation to its closest three control observations (3 -1 nearest neighbor matching). We apply a 0.05 caliper on the propensity score. Figure 4 provides the propensity score distributions for our control and treatment groups, pre versus post-matching, and shows that matching significantly reduces the difference between the distributions. 16 Table 5 presents estimates of the treatment effect for our matching models. In Column 1, we estimate the difference in means between our treated observations and our matched control observations. The estimated difference in log prices is -0.263 and is 16 Following Rosenbaum and Rubin (1985), Sianesi (2004), and Kassie et al. (2011), we test the balancing between our matched treatment and control groups through mean standardized differences and the pseudo R² and likelihood ratio test of joint significance. We find that the mean standardized differences in our variables are reduced, our Pseudo-R² is reduced, and that the joint significance of the matching covariates is rejected, post-matching. These results are available in Table 8A in the Appendix. statistically significant. In Columns 2-4, we use the matched sample to ensure covariate balance, but we return to a least squares framework to account for price dynamics and spatial unobservable variables. Control observations can be used more than once, so we weight each transaction proportional to the number of times it is used in the matching process using weighted least squares. 17 In Column 2, we control for the estimated propensity score in our regression. The coefficient on proximity is -0.241, quite similar to the matching estimate. In Column 3, we add year fixed effects. The coefficient here is -0.439, which is a substantial increase in magnitude over Columns 1-2. Adding year fixed effects had a similar impact on coefficient magnitude in Table 3. In Column 4, we lastly add tract fixed effects and the resulting coefficient is -0.353, nearly identical to the main results in Table 3. In conclusion, our matching model improves the similarity of our treated and control observations, but results are similar to the regression models.

The effect of proximity when mineral rights ownership is unknown
We now seek to understand how valuation estimates change when mineral rights ownership is unknown, as is the case in all prior papers in this literature. We now include properties that are not completely contained within the federal mineral ownership boundaries. It is uncertain whether the mineral rights are unified with the property or split and owned by another party.
In Table 6, we estimate Equation (4) with our expanded sample (N = 47,073). 17 All transactions not matched to another observation are given zero weight.
Column 1 presents a specification identical to our preferred specification of Column 4 in Table 3, which includes property characteristics and year and tract fixed effects. The coefficient estimate is -0.057 and is statistically insignificant. This is substantially smaller in magnitude than estimates using only split estate properties. We interpret the disparity in coefficients as resulting from inclusion of properties that are tied to mineral rights and thus are able to financially benefit from nearby drilling, which offsets the negative impacts. However, we are unable to recover a comparison of benefits and costs using these estimates because the distribution of mineral estate ownership is unobserved.
The second column in Table 6 estimates separate proximity effects for properties with federal mineral ownership (our main sample) and properties without federal mineral ownership. Federal mineral ownership is a binary variable equal to 1 if the property is completely contained within the federal mineral ownership boundaries. The estimated impact for properties with unknown mineral ownership is -0.031, similar to Column 1.
The coefficient on the interaction between the indicator for within one mile and federal mineral ownership is -0.318, similar to the main result in Table 3. Further, the interaction coefficient and is statistically significant meaning that the impact of proximity is statistically different for properties without mineral rights than properties with unknown mineral rights, and these are all properties in the same three counties. 18 18 We again estimate similar models using count and distance bin-based treatment variables. These results are available in the Appendix. In Table A6, our results suggest that the impact of nearby unconventional oil and gas development on housing prices is statistically significant and negative. We estimate that each well drilled within a mile decreases sale price by -0.7 to -0.2%. This is a significant attenuation relative to properties split from their mineral rights by the federal government, where we find a much higher estimate of -2.0%. In Table A7, we find that the effects of development within different spatial rings around the property varies across models. We generally find negative effects within two miles of development, though they are only strong and statistically significant when using county fixed effects. When using our preferred model with census tract fixed effects, our results indicate that having a well within one mile of the property reduces sale price by -4.3%.

Conclusion
We seek to quantify the negative externalities associated with unconventional extraction of oil and gas using hedonic valuation of residential properties. We exploit the Stock-Raising Homestead Act of 1916 to identify properties that do not have mineral rights, and thus cannot benefit financially from lease payments or royalties to isolate the external costs of unconventional oil and gas development. This approach resolves a significant issue in the valuation of unconventional oil and gas development, which has caused uncertainty in the interpretation of estimates from previous studies.
The results of our hedonic analysis suggest that houses within one mile of an unconventional drill site sell for 34.8% less than comparable properties without proximate drilling. This discount translates to a price reduction of $63,788, which equals $3,952 when annualized by a 30-year mortgage and a 5% interest rate. We interpret this price difference as the household valuation of the external environmental and health costs associated with proximity to unconventional oil and gas development. Further, our findings are 60% larger in magnitude than the largest existing negative estimate. While there are differences across study areas and identification strategies across studies, we interpret the disparity resulting from our studies ability to identify split estate properties.
The results of this paper can inform how the United States and other countries proceed with energy development. The suite of energy options available to consumers have benefits and costs that are received at global, regional, and local levels. Our estimates of the local external costs of unconventional oil and gas development should be considered with respect to those incurred from other forms of energy production, including coal-fired power plants (Davis, 2011), wind turbines , and nuclear power facilities (Gawande et al., 2013).
For those states that allow local regulation of oil and gas development, optimal local policy responses to unconventional oil and gas development should consider these results alongside others which find positive valuations of development (Boslett et al., 2016). Mineral rights ownership is clearly important to valuation of local energy development. Policy-makers should account for mineral rights in policy development.
They can then take measures to further support or regulate development (e.g., Zirogiannis et al., 2015) as a function of the level of local ownership of the minerals.
Our findings are relevant and applicable to a broader geographic area than Colorado. Prior hedonic valuation research provides great insight on local valuation of unconventional oil and gas development, especially on its perceived water quality risks.
However, external validity outside of the United States is limited in these studies because private citizens in European and many other countries do not own subsurface minerals.
Our study may provide a better metric for external costs of unconventional oil and gas exploration in these cases because it accounts for the critically important issue of mineral rights ownership.   We received residential property transaction, structural, and parcel data from Garfield, Mesa, and Rio Blanco county assessment and geographic information systems (GIS) offices. We received location data for all horizontal wells drilled from 2000 to 2015 and all vertical wells drilled from 1980 to 1999 in Colorado from Drillinginfo. We calculated the percentage of each property in an agricultural use using National Land Cover Dataset 2001 data. We calculated the distance to the closest municipality using U.S. Census Bureau TIGER data from 2010.
N N N Y Notes: Observations represent single family residential properties sold from 2000 to early 2015 in Garfield, Mesa, and Rio Blanco counties (N = 783). We truncate the data set to exclude the 5 and 95 percentiles of sale price. The dependent variable is the natural log of sale price (CPI-adjusted to 2014 values). Property variables include quadratics of # of bedrooms and bathrooms, parcel acreage, property finished living area, and property age. Location variables include quadratics of distance to the closest municipality and the percentage of the property in an agricultural use. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
In Model 1, we include a control for the # of vertical oil and gas wells drilled from 1980 to 1999. In Models 2 and 3, we restrict our analyses to only those observations that are within 20 and 10 miles, respectively, of a well drilled pre or post-sale. In Model 4, we restrict our analysis to only those observations sold from 2005 to 2014. In Model 5, we only include those observations that were sold once between 2000 and 2015. In Model 6, we use a panel approach and only use those properties that were sold more than once from 2000 to 2015. We use property-level and year fixed effects. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. In Columns 1, we use nearest neighbor matching (3 -1) and match on the # of bedrooms and bathrooms, living area, property age, and acreage, as well as the distance to the closest municipality, the percentage of the property in an agricultural use, and the number of vertical wells drilled within a mile from 1980 to 1999. We match with replacement and with a caliper of 0.05. In Column 2 through 4, we use our treatment observations and matched control observations from our Nearest Neighbor matching in Column 1 and estimate weighted regression models (in line with Equation 1 and Table 3). We weight based on the number of times each control observation was matched to a treatment observation. In Column 2, we only control for the estimated propensity score. In Column 3, we add year fixed effects. In Column 4, we add the structural and locational variables defined above. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. Y Y Notes: Observations represent single family residential properties sold from 2000 to early 2015 in Garfield, Mesa, and Rio Blanco counties. We truncate the data set to exclude the 5 and 95 percentiles of sale price. The dependent variable is the natural log of sale price (CPI-adjusted to 2014 values). Property variables include quadratics of the # of bedrooms and bathrooms, parcel acreage, property finished living area, and property age. Location variables include quadratics of distance to the closest municipality and the percentage of the property in an agricultural use. Federal Mineral Ownership is a binary variable equal to 1 if the property is completely within the boundaries of federal mineral ownership. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.  Notes: Unconventional well location data was provided by Drillinginfo. Data on federal ownership of oil and gas is from the Bureau of Land Management Colorado's GIS department. We created our private lands estate classifications using federal mineral ownership and statewide surface ownership data. Unknown Mineral Ownership is the intersection between "Private" surface ownership and mineral ownership categories that do not include the federal ownership of the oil and gas estate, while Federal Mineral Ownership is the intersection between "Private" surface ownership and federal mineral ownership categories that do include the oil and gas estate. Properties indicated as Unknown Mineral Ownership may be split estate properties if the mineral rights were at one point severed due to alternative mechanism (e.g., a severance deed).
118 Figure 3: Distance bin parameter estimates and histogram of well distance Notes: In the LHS graph, we highlight parameter estimates for distance bins in line with Equation 4. We define our treatment variable as a bin variable (see Equation 5). Each estimate reflects the impact of the closest well from the property drilled within the distance bin range, relative to the omitted category of having the closest well drilled beyond two miles. In the RHS graph, we highlight the distribution of distance to closest unconventional well. We only include those wells that were drilled before the sale of the home. Notes: These graphs highlight the difference in propensity score distribution, pre versus post-matching. Variables included in the propensity score estimation include the # of bedrooms and bathrooms, parcel acreage, property finished living area, property age, distance to the closest municipality, the percentage of the property in an agricultural use, and the number of vertical oil and gas wells drilled within a mile from 1980 to 1999.

Post-Matching
This Appendix provides information that is supplementary but not critical, to the analyses that included in the main paper.   Table 3 in the main text. The signs on the variables generally match our expectations. Property price increases in acreage and living area but with diminishing returns. Older houses sell for a discount. Property price decreases in the number of bedrooms and bathrooms. This is unexpected, but this is not a concern given correlation between both variables and living area. We generally find that property price decreases as one moves further away from municipalities, though at a diminishing rate. This is somewhat expected given the amenities associated with living in or near towns.
In Table 3A, we define treatment as the number of wells drilled within 1 mile.
The structure of the table is similar to that of Table 3 in the main paper. We find qualitatively similar results when using this alternative treatment definition. Each well drilled within a mile is associated with a decrease in sale price of 2%.
At the advice of an informal reviewer, we estimate Equation 4 in levels instead of logs in Table 4A. We again find qualitatively similar results. In our preferred specification in Column 4, we estimate that having an unconventional well within a mile of a property decreases sale price by $52,000. This is similar to an analogous estimate using the mean sale price in our study area and our estimate from Table 3, Column 4.
In Table 5A, we test for the robustness of our results in Table 3 of the main paper across a number of additional models. In Column 1, we include only those properties that are within three miles of a well drilled from 1980 to 1999. This model tests whether or not our treatment effect estimates are associated with differing levels of historical oil and gas development. In Columns 2 and 3, we relax our definition of a split estatethose properties with federal mineral ownershipto allow properties with 90% and 75% coverage, respectively, within federal mineral ownership boundaries. In Columns 4 and 5, we use alternative well distance restrictions to our 20 and 10 mile distances used in Columns 2 and 3 from our robustness table in the main text. In Column 6, we cut observations below the 5% and above the 95% distribution in acreage. We do this to avoid the influence of outlying observations based on property size. Lastly, in Column 7, we include mobile homes while also controlling for them using a binary indicator.
Our results from our additional models suggest that our results are robust. Our results are similar when we relax our definition of the split estate. Although these definitions are still restrictive and are likely including observations with federal mineral ownership, they suggest that we could increase our sample size and still obtain similar results. Our results are also similar to our main results when we use alternative restrictions based on distance to the closest well and when we cut outlying observations by acreage.

122
In Table 6A, we provide a brief summary of pre and post-matching statistics from Table 5 in the main text. Mean standardized differences in the matching covariates are reduced. Also, the Pseudo-R² is reduced and the p-value of the likelihood ratio test of joint significance of the matching covariates is significantly higher, post-matching. These results suggest that there are no differences in observed characteristics between control and treatment observations, post-matching.
In Table 7A, we use our full sample in line with Table 6 in the main text. Instead of only including properties with federal mineral ownership, we now include properties with unknown mineral ownership. However, we now define treatment as the number of unconventional wells drilled within a mile of the property. In Column 1, we find that each well within a mile is associated with a 0.3% decrease in property sale price. This effect is not significant and is much smaller than the estimates found in Table 2A (1.7 to 2.0%). In Column 2, we differentiate the effect of well development by whether or not the property is within federal mineral ownership boundaries. We find that the effect of well development is much larger for those properties with federal mineral ownership (-1.6%). For those properties with unknown mineral ownership, the effect is close to zero and insignificant. These findings support those from Table 6 in the main text. Since those properties with unknown mineral ownership are likely to include properties with unified mineral rights, our treatment effect for these properties is likely to be a net effect of financial benefits and external costs of development. Argue that their estimated valuation of the environmental costs is a lower bound due to the inclusion of properties with mineral rights ownership. Highlight the need for future work to differentiate the impacts of development on homebuyers by mineral rights ownership.
New York & Pennsylvania Note that their estimated valuation does not capture positive impacts of development on nearby property prices if properties that are close to development are split from their mineral rights Delgado et al. (2015) Northeastern Pennsylvania Argue that their estimated valuations capture both the positive effects and negative effects of development due to their choice in study area, which may have less severance.

James and James (2015) Weld County, Colorado
Argue that their identification strategy avoids the mineral rights issue by controlling for both distance to the closest well and whether or not the property overlays the well's horizontal lateral Bennett and Loomis (2015) Weld County, Colorado Argue that some of their estimated positive impacts are due to the potential that properties near wells are more likely to own their minerals Note that there is relatively low local ownership of mineral rights in the study area, which suggests that the estimated appreciation in housing prices after development is likely due to the expansion of the tax base (rather than expected financial benefits of drilling) # of Wells < 1 Miles -0.017*** -0.020*** -0.020*** -0.020*** (0.0009) (0.003) (0.003) (0.003) N N N Y Notes: Observations represent single family residential properties sold from 2000 to early 2015 in Garfield, Mesa, and Rio Blanco counties (N = 783). We truncate the data set to exclude the 5 and 95 percentiles of sale price. The dependent variable is the natural log of sale price (CPI-adjusted to 2014 values). Property variables include quadratics of # of bedrooms and bathrooms, parcel acreage, property finished living area, and property age. Location variables include quadratics of distance to the closest municipality and the percentage of the property in an agricultural use. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
N N N Y Notes: Observations represent single family residential properties sold from 2000 to early 2015 in Garfield, Mesa, and Rio Blanco counties. We truncate the data set to exclude the 5 and 95 percentiles of sale price. The dependent variable is sale price (CPI-adjusted to 2014 values). Property variables include quadratics of # of bedrooms and bathrooms, parcel acreage, property finished living area, and property age. Location variables include quadratics of distance to the closest municipality and the percentage of the property in an agricultural use. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
In this table, we provide additional robustness checks to those shown in Table 3. In Column 1, we restrict our sample to only include those observations that are in historic oil and gas drilling areas, measured by being within a distance of 3 miles to the closest vertical well drilled from 1980 to 1999. In Columns 2 and 3, we ease the split definition to include those properties that have 90% and 75%, respectively, of their acreage within the federal mineral ownership boundaries. In Columns 4 and 5, we restrict our observations to those within 15 and 5 miles, respectively, of a pre or post-sale horizontal well. In Column 6, we cut off the 5% tails of the acreage distribution. In Column 7, we include mobile homes and control for this type of property with a binary indicator. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.  (3 -1) with replacement a 0.05 caliper. We matched on the # of bedrooms and bathrooms, living area, property age, distance to municipality, percentage of the property in an agricultural use, and the number of vertical wells drilled within a mile from 1980 to 1999. Y Y Notes: Observations represent single family residential properties sold from 2000 to early 2015 in Garfield, Mesa, and Rio Blanco counties. We truncate the data set to exclude the 5 and 95 percentiles of sale price. The dependent variable is the natural log of sale price (CPI-adjusted to 2014 values). Property variables include quadratics of the # of bedrooms and bathrooms, parcel acreage, property finished living area, and property age. Location variables include quadratics of distance to the closest municipality and the percentage of the property in an agricultural use. Federal Mineral Ownership is a binary variable equal to 1 if the property is within the boundaries of federal mineral ownership. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.

Introduction
The demand for silica sand has increased measurably since the beginning of the boom in unconventional oil and gas development. Silica sand is one of the key ingredients in hydraulic fracturing, a process that has spurred significant oil and gas production from previously inaccessible shale resources. This type of sand is the most often-used proppant in hydraulic fracturing. A proppant is injected to maintain the fractures created during hydraulic fracturing (Benson and Wilson 2015). Hydraulic fracturing can require over 450 tons of proppant per well (King 2016). Although sand mined in the Midwest United States has historically been used in a variety of applications, most often in glass-making, the proportion of sand purchased by the oil and gas industry has increased over time. It comes as no surprise that silica sand mining is increasingly known as 'frac' sand mining.
Though the academic literature has explored many of the impacts of the unconventional oil and gas boom, there has been considerably fewer studies on local impacts associated with silica sand mining. These impacts can be both a blessing and a curse. In terms of positive impacts, the increase in demand for silica sand from the oil and gas industry has led to employment and income creation, a welcome boost for rural communities that have not recovered from the great recession (Prengaman 2012). Sand mining may also lead to growth in other sectors of the economy, including the rural rail economy (Davies 2012a). Local landowners and farmers may also receive significant signing bonuses and royalty payments from leasing their land to mining companies.
Others have sold their land outright to mining companies for prices above and beyond assessed value (e.g., Davies 2012b; Lenker and Auch 2013;Dirr 2014;Mertens 2014).
Yet, these positive benefits are juxtaposed against environmental quality concerns. Despite its name, 'frac' sand mining involves no hydraulic fracturing and is instead extracted using surface mining techniques. This type of mining can take a significant toll on the landscape by leveling hills and causing habitat fragmentation.
These mines can be very large; the average (maximum) size of permitted or operational frac sand mines in Wisconsin is 341 (4000)  total production of silica sand (Content 2015).
In this paper, I use hedonic pricing theory to value the local impacts of silica sand mining in Wisconsin. This technique has a wide history of use in the economic literature and has recently been applied to understand how people value unconventional oil and gas development (e.g., . To my knowledge, this is the first economic study to use hedonic valuation to understand how citizens value silica sand mining's impacts.
My approach relies on regression models that relate the level of local silica sand mining to residential property sale prices. I first estimate a series of preliminary models exploring how general proximity to silica sand mining has influenced residential property sale prices from 2011 to 2016. I then explore the potential for heterogeneous impacts by impact of mine. I focus on the capitalization of viewshed impacts and air quality impacts associated with mining. These are two of the major environmental costs of sand mining. I differentiate the effect of mining by whether or not it is in view of the residential property. I hypothesize that having to view mining activity may negatively influence sale price beyond the effect of mere proximity. My approaches are similar to those used in the hedonic valuation of the impact of wind turbine views on housing prices (e.g., Heintzelman and Tuttle 2012;. 41 41 This work builds on an extended literature incorporating property views into hedonic valuation models. Much of this literature has focused on views of environmental amenities or natural land uses (see Bourassa et al. 2004 for an extensive review). More recent work has focused on viewshed impacts associated with energy infrastructure, including wind turbines and nuclear energy stations (e.g., Heintzelman and Tuttle 2012;Dickes et al. 2013;.
Following this analysis, I explore how air pollution and dust from mines influence the sale price of residential properties. Given the rural character of my study area, I do not have spatially-explicit air pollution data. Instead, I use an indicator of cardinal direction away from each mine as a proxy for likelihood of air pollution associated with mining. I argue that those properties on the eastward side of sand mines are more likely to be impacted by dust and particulates from mining activity. An estimated differential effect between properties in opposite directions away from the mine is likely to be connected with greater air pollution effects on the eastward side of mines.
In my preliminary models, I find only weak evidence that silica sand mining has a significant impact on residential property prices. When I differentiate by whether a property is in-view or out-of-view of a mine, my results suggest that sand mining has a negative impact on sale prices for those properties in-view of a mine. Each in-view frac sand mine within two to four miles of the properties is associated with a 6 to 11% decrease in property sale price. However, I do not find a significant impact of sand mining that occurs within one mile of the property. Out-of-view sand mining generally has positive effects on sale price, but these are not significant.
My hypothesis regarding air quality impacts is supported by my findings. Sand mining from the west can have a negative impact on sale prices of downwind residential properties. At the same time, I find significant evidence of appreciation for properties that are upwind of mining activity. Each eastward sand mine within a mile of a property is associated with an increase in property sale price of 12%. This effect decreases to around 10% for mines within one to two miles and to 3-4% for mines within two to four miles.
These effects are much stronger than the effects for corresponding westward mines.
Given the above results, I examine how silica sand mining influences residential properties across indicators of both views of and the direction of mining activity.
Estimates suggest that silica sand mining from the west significantly reduces residential property sale price, though the effect is considerably stronger for those properties that are in view of a sand mine. I also find sale price appreciation for residential properties that are out-of-view of eastward sand mining.
My results are suggestive of highly-localized positive capitalization of mining, as well as significant negative capitalization of viewshed and air quality impacts for those properties that are close but not adjacent to sand mining. Appreciation associated with silica sand mining makes sense, as Wisconsin sand mines have expanded outward over time 42 and mining companies have been active in leasing or buying land outright for mining (e.g., Davies 2012b; Dirr 2014). Appreciation may be linked to (1) property sales to mining companies at prices above assessed values, or (2) expectations for this type of sale price premium in the future (especially for those properties that are in close proximity to mining activity but have not been sold to mining companies yet). It may also be linked to property price compensation packages used by frac sand mine companies to alleviate concerns from local residents regarding property price impacts. These packages are used to compensate property owners for any losses in value, post-mining. According to an unnamed county councilman in western Wisconsin, a sand mining company will typically pay a premium over the listing price if a nearby residential property is unsold six months after listing. It is currently unclear how often this type of transaction occurred 42 Based on aerial footprints of mines across time from Ted Auch, FracTracker Alliance. in our data set. If it happened often, it may contribute to my findings of no change in price or appreciation for certain properties.
At the same time, depreciation associated with silica sand mining appears to be driven by both view and air quality changes. This is especially true for those properties that are close but not adjacent to mining activity. Since I do not find any significant evidence of environmental costs being capitalized within one mile, it seems as though the benefits of mining are highly localized around mines but the costs are diffused outwards from the mine to at least four miles away.
To my knowledge, this is the first economic study to explore the economic impacts related to silica sand mining (concurrent with Kalinin et al. 2016). Although Parker and Phaneuf (2013)  This paper is structured in the following manner. In Section 2, I provide background on silica sand mining and its economic, environmental, and socio-cultural implications. In Section 3, I discuss my study area of western Wisconsin and outline my data. In Section 4, I highlight my methodological approach. In Section 5, I present my empirical results and finish with some brief conclusions in Section 6.

Silica sand mining in Wisconsin
Although sandstone basins are located in many areas of the United States, the sandstone deposits in Wisconsin have characteristics that make it ideal in hydraulic fracturing applications. This sand has a high silica content and mostly consists of quartz, which is very hard (7.0 on the Mohs scale). As a result, it is able to resist pressure and hold open fractures from hydraulic fracturing. The sand deposits in the region are also generally close to the surface. This is in part a result of the fact that this general area of Wisconsin avoided significant glaciation during the last ice age. 43 This anomaly has made it easier for companies to access and retrieve the sand through surface mining (Benson and Wilson 2015).

Silica sand mining and its impacts
Academic research on the economic and socio-cultural impacts of silica sand mining has been fairly limited. Deller and Schreiber (2012) explore the relationship between general mining activity and various economic and socio-cultural variables, though their methodology was unable to differentiate silica sand mining from other types of mining. They also use county-level data at the national scale. Other economic studies have been either non-academic or associated with stakeholder group funding. Power and Power (2013) raise caution over high expectations of economic impact from silica sand mining, but their suggestions are extrapolated from general mining activity across the United States and rest of the world. Using IMPLAN and data from Illinois' LaSalle County, Harger (2015) estimates large direct and indirect employment and income effects from local silica sand mining. 44 In a review of literature related to the economic and environmental impacts of industrial activity and mining, Parker and Phaneuf (2015) suggest that silica sand mining may negatively impact property values and tourism in Pepin County, Wisconsin. However, this conclusion was based on anecdote and insights from the hedonic valuation literature of general industrial activity.
In the sociology literature, Pearson (2013) highlights community expectations regarding the financial and employment benefits of silica sand mining versus the environmental costs received by the general public in the vicinity of the mines. Although some rural landowners may receive hundreds of thousands of dollars for selling their land to mining companies, he generally describes the local community as opposed to silica sand mining on environmental, public health, and socio-cultural grounds. In later work, Pearson (2016) argues that the rapid pace and intensity of silica sand mining can reduce residents' sense of place and community.
A number of studies have explored the environmental impacts of silica sand mining. Air pollution and increased silica dust in the local atmosphere is a prominent concern in public media, though studies have found mixed results regarding the magnitude of impact. Richards and Brozell (2015) were unable to detect differences in upwind/downwind respiratory crystalline silica levels. However, Walters et al. (2015) found PM2.5 levels above the Environmental Protection Agency's annual standard.
Although each silica sand mine can use as much as 2.5 million gallons of water a day (Wisconsin Department of Natural Resources 2012) and silica sand mining may change groundwater aquifer recharge (Parsen and Gotkowitz 2013), there have been no studies that evaluated pre and post-mining water quality in areas around silica sand mines.
However, in 2012, 20% of active mines and processing plants in Wisconsin were cited for environmental violations (Prengaman 2013). Locke (2015) argues that this is an increasing concern given that silica sand mines are often located in unzoned areas without formal planning procedures.

Related hedonic valuation literature
There is a relatively limited literature on the hedonic valuation of mining activity or its associated impacts. In early work, Hitzhusen et al. (1997) use lakeside property prices to estimate the economic cost of pollution associated with abandoned strip mines in Ohio. Their results indicate that close proximity to abandoned mines reduces the sale price of properties. In later work, Damigos (2006)

describes a National Resource Damage
Assessments that used hedonic valuation to value environmental costs of the Eagle Mine in Colorado. The mine reduced sale price by $25,000 for those properties within six miles. Williamson et al. (2007) explore the property price impacts associated with acid mine drainage-impaired streams in West Virginia. They observe impacts of -$4,800 for those properties located within a quarter-mile of an impaired stream.
To my knowledge, there are no studies using hedonic valuation to infer how residents value the impacts of silica sand mining. Although Parker and Phaneuf (2013) discuss some of potential implications of mining on housing prices, their insights are based on a survey of the related literature on other types of mining and industrial activity.

Study Area
My study area is focused in thirteen counties in western Wisconsin (Figure 2).
These counties are primarily rural with some scattered larger cities and towns, including Chippewa Falls, Eau Claire, La Crosse, Marshfield, and Sparta. All counties overlay sandstone formations that are sources of 'frac' sand and have experienced recent mining.
There are 95 'frac' sand mines within the counties (out of 101 in the entire state). Many of these mines were permitted between 2011 and 2016. Only seven of the mines were permitted for reclamation prior to the mid-2000s, which is generally agreed as the starting point to the boom in unconventional oil and gas development (Bauer 2014). There is some limited evidence that another sixteen mines were in place in 2004, based on aerial imagery from the National Agriculture Imagery Program (NAIP). However, this evidence is highly subjective, as it is difficult to differentiate mining and agriculture in aerial imagery. All other facilities are dropped from the analysis.

Data
The data set provides coordinate location and production status for all silica sand mines in the state. However, the data does not include any information on when the mines were first permitted or excavated. Most permitting and regulation of sand mining in Wisconsin is done at the local level (e.g., Schuessler 2015). Although sand mines must abide by state air and water regulations, all reclamation permitting is done either by the county or the township where the mine is located. In order to commence non-metallic mining, each company must receive a reclamation permit (Haines 2012). I received reclamation permitting dates for 93 frac sand mines in Wisconsin through correspondence with local county and township officials in zoning and land information departments. 47 I use this as my proxy for before-after status of mining for each property transaction. If a property is close to mining activity that was permitted for reclamation before the time-of-sale, then the property is considered "treated" by silica sand mining. If the mine was permitted for reclamation after the time-of-sale, then the property is considered a "control" observation since silica sand mining was not occurring at the time- 46 Wisconsin's Department of Natural Resources' "Locations of industrial sand mines and processing plants in Wisconsin",Microsoft Excel spreadsheet. 47 There were eight frac sand mines from which I could not find associated reclamation permitting information. These eight mines were prospective mines as of January of 2016 that had only just applied for reclamation permitting. of-sale. Thus, for all analyses in this study, properties that are "treated" with nearby sand mining are those that were sold after a mine was first permitted for reclamation.

The effect of silica sand mining on local property prices
In my analyses, I estimate hedonic valuation models to value the impacts of silica sand mining as perceived by home buyers in western Wisconsin. This approach has a long history of use in applied economics research (see  for an early outline of the model). The main idea underlying the approach is that the price of a heterogeneous good is associated with its attributes and characteristics. The price of the good can then be decomposed as a function of its characteristics. This approach is particularly useful in eliciting how homebuyers value the impacts associated with neighborhood amenities and disamenities.

147
My first models relate the sale price of a home to the number of silica sand mines that are within a series of distance bins away from the property: I infer a valuation of the impacts of silica sand mining using my β parameters. In these models, I do not identify separate impacts so any valuation from these parameters is net of the benefits and costs of nearby mining. Given the possibility of both positive and negative impacts associated with silica sand mining, I have no prior expectation of the signs of these parameters.
To test for the robustness of these results, I estimate the model with alternative subsets of the data. I describe each robustness check in Section 5.

The effect of silica sand mining's viewshed changes on local property prices
Next, I explore the possibility of heterogeneous impacts based on certain characteristics of the spatial relationship of each property to silica sand mining. I first estimate how changes in views associated with silica sand mining impact nearby residential property prices. The landscape and land use fragmentation impacts of silica sand mining may lead to greater property price impacts for those properties that are in sight of a nearby sand mine, ceteris paribus.
I estimate the following regression model: This distance bin-based approach is similar to that used in Tables 2 and 3 except I  As an additional source of robustness, I follow  and use a slightly different approach to identification. My viewshed approach described above defines treatment using a bin framework, where I differentiate the effect of in-view and out-ofview silica sand mining across different spatial areas. I now use a strict r mile buffer 150 approach and limit my observations to only those within r miles of a mine at any time. I again include counts of both visible and non-visible frac sand mines, permitted prior to the sale of the property, within r miles. (3) ) (  is the number of westward frac sand mines located within X and Y miles and ist Y X Eastward ) (  is the number of eastward frac sand mines located within the same distance for residential property i at time t in census tract s. For each given X -Y spatial area, 1  and 2  indicate the impact of each westward and eastward sand mine, respectively, on residential property sale price.
I again have no prior expectations of the sign of 1  or 2  . However, I hypothesize that 1  will be attenuated or more negative relative to 2  as a result of presumed higher levels of ambient dust and pollution for properties with westward-situated mining activity.
I use the difference between 1  and 2  as a first-order approximation of the capitalization of local air quality impacts in housing prices. The primary assumptions in this interpretation is that (1) air quality effects are felt predominately by those on the eastward side of mines, and (2) the only difference between properties on either side of the mine is the relative levels of air pollution and dust. However, a concern is that my control group -those mines with eastward mines -are likely to incur air quality impacts, as well. To the extent that my control group suffers from spillover effects associated with air quality impacts, my estimate is an underestimation of the valuation of air quality impacts of silica sand mining in housing prices.

Heterogeneity across both viewshed and cardinal direction
In  ) ln( Each variable is a separate count of frac sand mines within r miles of each potential combination of visibility and direction. Parameter interpretations are similar to those above. They represent the differential impacts of each type of mine, specified by visibility and direction, on residential property sale price. Comparisons in parameters from 1  through 4  allow a more thorough exploration of the implications of air pollution and views of frac sand mines on prices for 153 residential properties near frac sand mines. Each combination of visibility and cardinal direction is associated with a different combination of impact. I hypothesize that 1  will be the most negative or, if all parameters are positive, strongly attenuated parameter, due to the fact that these properties are subject to both the viewshed and air pollution impacts of silica sand mining. I also hypothesize that 4  will be the most positive or, if all parameters are negative, strongly attenuated parameter, due to the fact that these properties do not face either the viewshed or the air pollution impacts of silica sand mining. I have no prior expectation of the relative differences between 2  and 3  since I do not know whether the air pollution impact (i.e., 2  ) or the viewshed impact (i.e., 3  ) will be more strongly valued in housing prices.
As a final note, this analysis allows us to estimate which properties are most impacted by the viewshed and air pollution impacts of silica sand mining. Although I can approximate the valuation of viewshed impacts by differencing my 1  from 2  or 3  from 4  , there may be concerns that those properties that are in view and to the east of silica sand mining are also more subject to air pollution costs than those out-of-view but also to the east of silica sand mining. This may be the case if topography serves as a source of deposition of dust for those out-of-view properties to the east of mining. In this case, the difference of 1  and 2  may capture viewshed impacts and the increased impact of air pollution based on viewshed.
With that concern acknowledged, however, I still believe that this is an important analysis. By controlling for both viewshed and cardinal direction, I obtain a more nuanced understanding of the differential impact of silica sand mining across spatial and topographical characteristics.  properties that are within four miles of a frac sand mine, pre-sale. Properties share generally similar characteristics, though properties near silica sand mining are also closer to the interstate highway system and railroad system. This makes sense given the transportation needs of silica sand mining production and transportation.
My results in Table 2 suggest that nearby silica sand mining can have a significant and negative impact on residential property prices, though this effect seems to only exist for properties with mines located between two to four miles away. However, this effect is not statistically-significant when including census tract fixed effects. Across all models, I uncover no evidence that mining has an influence on price within two miles.
These results are somewhat surprising. The typical assumption in the literature is that environmental impacts diminish with increasing distance away from the source. In the mining context, one would expect attenuating effects of proximity to sand mining as one moves further away from mining activity. However, I find some limited evidence of a significant effect of mining beyond two miles, but no effect of sand mining within two miles.
Since census tract fixed effects help control for unobserved differences in housing price across the study area, Column 4 in Table 2 Table 3 provide further support for results in Column 4 of Table 2. I am unable to find any significant evidence that silica sand mining has a significant net effect on sale price. Although I find slightly larger estimates of some of our treatment effects, none of these differences are significant.

The effect of silica sand mining's viewshed changes on local property prices 156
My results from Tables 2 and 3 suggest that sand mining has at most a limited effect on the sale price of residential property prices. I now explore whether this finding of only a limited relationship is masking potential heterogeneity of valuation outside of a pure distance measure.
I first explore whether the impact of local silica sand mining is different for those properties that are in-view versus out-of-view of mining activity. In Table 4 My results in Table 4 suggest that silica sand mining can be negatively capitalized into residential properties prices if it is in-view of the property. This is especially true if the mines are located from 1 to 4 miles away from the property. I estimate that each inview frac sand mine from 1 to 2 miles and 2 to 4 miles is associated with 5-7 and 10-11% decreases, respectively, in sale price. Most estimated effects in these two bins are statistically significant.
However, I find no statistically-significant impacts within 1 mile of the property.
In fact, the magnitude of the effect of in-view silica sand mining seems to increase as one 157 moves further away from the property. This result is again surprising. One would expect that the impact of views would generally be smaller for properties that are further away from mining.
I find some limited evidence that out-of-view silica sand mining has a statistically-significant effect on nearby residential property prices. However, my estimates are only statistically significant for one subset of the dataonly those properties in census tracts with silica sand miningand only within the 1 to 2 mile ring. I find that effects of out-of-view mining attenuate with increasing distance away, which is the opposite of the relationship between estimated effect and distance away for in-view mining.
These results provide some insight as to why I find significant effects only beyond a certain distance away of mining activity. I find positive but statistically insignificant effects for properties within one mile of out-of-view silica sand mining. This may be driven by highly-localized expectations of financial benefits associated with silica sand mining. As one moves further away from mining, these expectations diminish, as suggested by out-of-view effects attenuating with increasing distance.
Yet, at the same time, I find that the effect of in-view mining increases with increasing distance. It is possible that those properties beyond a mile of a mine are not likely to receive financial benefits from leasing land for mining or selling land to mining companies. If this is the case, then these properties face viewshed impacts without a strong likelihood of compensation. For those properties that are in closer proximity to silica sand mining (i.e., within a mile), there may be a stronger likelihood of financial benefits, which may compensate for some of the viewshed impacts. This explanation is 158 speculative, but does make sense given the estimated relationships between in-view and out-of-view mining and distance away from mining activity.
As an additional form of robustness, I follow  in Table 5 and focus exclusively on those properties that are close to sand mining activity, pre or post-sale. In doing so, I differentiate those properties sold after mining occurs by visibility of the frac sand mines. Thus, I compare properties in-view and out-of-view of silica sand mining after it occurs versus properties sold in the same area, pre-mining. By focusing in such a small area, I remove the potential for omitted variable bias associated with differences between control areas without mining versus treatment areas with mining. All of these models suggest that visible sand mines have a negative impact on the sale price of nearby residential properties. My estimates are statistically-significant for the three and four mile buffers. I estimate that each visible sand mine within three miles is associated with a 6% decrease in sale price. This effect increases in magnitude to 7-8% at four miles.
However, I do not obtain statistically-significant effects for the one and two mile buffers. Estimates for these buffers are also smaller in magnitude than those found in the larger buffers. Although this is surprising, it matches up with what I found in Tables 2   through 4. There are two main reasons why this may be happening. First, I have a very 159 limited sample in these regions very close to mining activity. When I relax my econometric specification and only control for year and census tract fixed effects, I obtain stronger estimates in the -12 to -7% range that are close to statistical significance at the 10% level. Second, the potential financial benefits of silica sand mining that accrue to landowners may complicate any comparison between models across buffer sizes. In-view properties within one mile of a mine may have greater expectations of potential financial benefits from mining activity than those that are in-view of mines, but are further away.
Those that are relatively far away from mining may have viewshed impacts without the potential for compensation.

Heterogeneity across cardinal direction
In Table 5, I estimate Equation 3 using the same spatial areas defined in Table 2 and 3. I follow the structure of Table 3 and use various subsets of the data in order to test for robustness of the result across different spatial restrictions. Instead of differentiating by view of mining activity, I now differentiate properties sold after mining occurs by the cardinal direction away from the frac sand mines. I hypothesize that properties with westward mines are more likely to be negatively impacts by air pollution impacts from silica sand mining than those properties with eastward mines.
Within the 2 to 4 mile ring away from the property, I generally find that each silica sand mine from the west is associated with a decrease in sale price on the order of 2 to 3%. However, I do not find significant effects for properties within two miles.
Importantly, close proximity to eastward frac sand mines is associated with an increase in property sale price, relative to pre-mining properties. This effect is statistically significant across all treatment areas. The effects attenuate with increasing distance away, 160 as expected. The impact of each eastward frac sand mine within a mile is on the order of 12-13%. For eastward frac sand mines located within 1 to 2 miles, the impact is on the order of 10%. From 2 to 4 miles, the effect is around 3.5%.
Negative estimates associated with westward sand mining provide supportive evidence of capitalization of air pollution and ambient dust impacts of silica sand mining.
At the same time, positive estimates for eastward sand mining is suggestive of the capitalization of either (1) actual or expected property prices premiums from direct sales of land to mining companies, or (2) expected payments from leasing land for silica sand mining.

Heterogeneity across both viewshed and cardinal direction
In viewshed and air quality impacts decrease in value across all four buffer sizes, though the effect is only statistically significant from two to four miles (8-10% for each frac sand mine). There is no effect of westward, non-visible mines. Additionally, I find evidence 161 of appreciation for those properties that are not impacted by either environmental cost (i.e., those properties with non-visible, eastward mines). This effect is statistically significant across all models and attenuates from 20% at 1 mile to 6% at 4 miles. My estimates are not statistically significant for eastward, visible sand mines. However, the difference in estimates for in-view and out-of-view eastward mines is suggestive of there being a strong viewshed impact on property prices.
In  Table 7. Parameter estimates across the models are similar to those found in Column 4. Some estimates are larger than from the main sample, but not significantly so.
From these results, it appears as those the properties that are most impacted by silica sand mining are those properties that are both in-view and east of the mines. This makes sense, as they are most subject to the viewshed and air pollution impacts of silica sand mining.

Conclusions
Although these results are exploratory, they are meaningful as a first step in a longer analysis of the local economic impacts of silica sand mining. Overall, I find evidence of both appreciation and depreciation associated with nearby silica sand mining.
My results suggest that silica sand mining's impacts on local views and air quality can be a source of depreciation in home values. Properties with westward mines decrease in value, post-mining, whether or not they have a view of the mine. In-view, westward mines have a larger influence on sale price than those out-of-view, though I am unable to find significant viewshed effects for eastward mines. This may be driven by the small number of observations with viewable, eastward mines. This may be a cause of concern 162 in Tables 4 and 5 With that said, these findings reveal an intuitive and consistent story. Silica sand mining seems to influence homebuyers, especially through viewshed and air quality impacts. These results support anecdotal evidence from local residents in Wisconsin, as well as the insights from Parker and Phaneuf (2013), that silica sand mining has environmental impacts that may manifest in housing price decreases. At the same time, 163 some of my results suggest the possibility of limited price appreciation. Although the mechanism is not clear, this may be due to expectations regarding future leasing activity or direct sales to mining companies for a large premium. Future work in this literature should focus on addressing the above issues to provide stronger policy information for decision-makers and stakeholders in Wisconsin and the surrounding region. Notes: All distance variables are in miles. Distance to the closest interstate highway and railway is based on U.S. Census Bureau's Tiger Geodatabase data from 2010. Agricultural % values are based on National Land Cover Database (NLCD) 2011 data. As a note, the parcel data that I used were not amenable to calculating statistics within parcel boundaries due to some overlapping parcel boundaries. Thus, I used focal statistics in ArcGIS and calculated local statistics within 500 meters around each parcel's centroid. Our variable of interest is a count variable indicating the number of sand mines within the given buffer at the time-of-sale. Observations represent single family residential properties sold from 2011 to early 2016 in our study area in Wisconsin. We truncate the data set to exclude the 5 and 95 percentiles of sale price. We only use those residential properties that are within twenty miles of a sand mine permitted before or after the sale of the home. The dependent variable is the natural log of sale price (CPI-adjusted to 2015 values). Property variables include quadratics of acreage, distance to the closest interstate highway, distance to the closest railway, and the percentage of land in an agricultural use within 500 meters of the parcel's centroid. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.  Table 2. In Column 1, I include no price restrictions. In Column 2, I only include observations from those census tracts that contain or are within two miles of a frac sand mine. In Column 3, instead of only cutting observations that are beyond twenty miles away from a sand mine, I now additionally cut those beyond ten miles away from a mine. In Column 4, I only include observations from counties from which I received property structural data from the state's Multiple Listing Services (i.e., Chippewa, Eau Claire, Jackson, Monroe, and Trempealeau). In Column 5, I only include those observations from Column 5 for which I have structural data. I control for these variables, including the number of bedrooms, bathrooms, and the finished living area of the property. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. of interest are a series of count variables indicating the numbers of in-view and out-of-view sand mines within given buffers at the time-of-sale. Observations represent single family residential properties sold from 2011 to early 2016 in our study area in Wisconsin. I include only those observations that overlay a sandstone basin with sand suitable for hydraulic fracturing. I truncate the data set to exclude the 5 and 95 percentiles of sale price. The dependent variable is the natural log of sale price (CPI-adjusted to 2015 values). Property variables include quadratics of acreage, distance to the closest interstate highway, distance to the closest railway, and the percentage of land in an agricultural use within 500 meters of the parcel's centroid. I include year fixed effects and census tract fixed effects. Columns 1 and 2 restrict our control observation subset to be located within 20 and 10 miles, respectively, of a mine at any time. Column 3 only includes observations from those census tracts that contain or are within two miles of a frac sand mine. Column 4 only includes those observations with structural data from the Multiple Listing Services in Wisconsin. I control for these variables, including the number of bedrooms, bathrooms, and the finished living area of the property. Census tracts are based on U.S. Census 2010 boundaries. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
In this table, I continue to explore the impact of views of silica sand mining on residential property sales. In odd columns, I only include those properties within r miles of a frac sand mine, pre and post-mining. In even columns, I do not make the above restrictions for control observations. In these columns, I use all observations with twenty miles of a sand mine, conditional on sale price and property type restrictions. My controls are the same as those used in Table 2. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. explore the effect of silica sand mining based on whether it is westward or eastward of a property. I do this in an effort to understand how air pollution and ambient dust levels influence sale price. My indicator of cardinal direction is based on the direction field for each frac sand mine in our data set. I hypothesize that residential properties with frac sand mines that are situated to the west are more likely to be negatively influenced by air pollution and ambient dust. Columns 1 and 2 restrict our control observation subset to be located within 20 and 10 miles, respectively, of a mine at any time. Column 3 only includes observations from those census tracts that contain or are within two miles of a frac sand mine. Column 4 only includes those observations with structural data from the Multiple Listing Services in Wisconsin. I control for these variables, including the number of bedrooms, bathrooms, and the finished living area of the property. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. the effect of mining on residential property sale price by both views and air quality level. Each variable represents a count of one combination of viewshed level (i.e., in-view or out-of-view of a frac sand mine) and cardinal direction (i.e., eastward or westward of a frac sand mine). I again only include those properties within r miles of a frac sand mine, pre and post-mining. Thus, I am comparing those observations with westward, visible mines; westward, non-visible mines; eastward, visible mines; and eastward, non-visible mine, post-mining, to observations close to the future, but not current, mines, and observations that are relatively far way (> 2 -4 miles, depending on model) from mining activity. My controls are the same as those used in Table 2. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.  Table 7, using a four mile buffer. In Column 1, I do not use a sale price restriction. In Column 2, I only use those observations that are within ten miles of a mine. In Column 3, I only use those observations from census tracts that contain or are within two miles of a frac sand mine. In Column 4, I include observations from only those counties from which we received limited structural information. In Column 5, I only include those observations from which we received limited structural information. I control for quadratics of the number of beds, the number of bathrooms, living area, and acreage. I also include year fixed effects and census tract fixed effects. I relax our control variables and do not include any locational variables (e.g., distance to interstate highway) due to our small sample size. Standard errors are shown in parentheses and are estimated using tract-level cluster-robust inference: *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.

Figure 1: Silica sand regions in Wisconsin
Notes: Silica sand mine locations are from the Wisconsin Department of Natural Resources. I followed Locke (2015) in her work on silica sand mining and zoning to determine the boundaries of silica sand basins. These boundaries are based on Wisconsin Department of Natural Resources and United States Geological Survey data. Wisconsin DNR data was used to understand the location of different types of bedrock, including sandstone, while the USGS data was used to classify the sandstone's age. Only those sandstone basins from the Cambrian and Ordovician ages are shown. the underlying assumption of this approachthat consumers in different markets have similar distributions of tastes and preferencesis questionable. More recent approaches include Ekeland, Heckman, and Nesheim (2004), Bishop and Timmins (2015), and Bishop and Timmins (2016). Ekeland et al. (2004) support the general intuitions behind the second stage approach used in  and suggest two approaches to identification: (1) a non-parametric transformation approach, and (2) a non-linear instrumental variables approach. Bishop and Timmins (2015) employ a maximum likelihood approach to estimate the marginal bid function that is most likely given consumer choices and their demographic information. Bishop and Timmins (2016) apply a panel data approach where they observe multiple sales per household. Since each sale is subject to a different hedonic price function, they are able to estimate two amenity choices under different supply conditions. They use this information to infer individualspecific demand functions. Due to the computational difficulties and theoretical issues associated with estimation of the supply and demand functions in the second stage, most papers restrict their focus to estimating the hedonic price function. However, there have been limited attempts to estimate both the first stage hedonic price function, as well as the underlying demand function (e.g., Bishop and Timmons 2015;Bishop and Timmons 2016).