Estimating the Value of Water from Property Sales in an Arid High Environmental Amenity Region: A Difference-in- Difference in Approach

Estimating the Value of Water from Property Sales in an Arid High Environmental Amenity Region: A Difference-in- Difference in Approach Michael Brady Assistant Professor bradym@wsu.edu 509-335-0979 Pitchayaporn Tantihkarnchana Doctoral Student School of Economic Sciences Washington State University Selected Paper prepared for presentation for the 2015 Agricultural & Applied Economics Association and Western Agricultural Economics Association Annual Meeting, San Francisco, CA, July 26-28. Copyright 2015 by Michael Brady. All rights reserved. Readers may make verbatim copies of this document for non-commercial purposes by any means, provided this copyright notice appears on all such copies. 1

Abstract This paper reports results from an analysis of residential property sales in a high environmental amenity region of Washington State characterized increasingly by water scarcity problems. Kittitas County sits on the dry Eastern side of the Cascade Mountains and is in close proximity to the greater Seattle metropolitan area. While water law restricts any additional uses of water without purchasing an existing water right there was a domestic well exemption until 2009 that allowed wells to be dug without acquiring a right. This exemption became increasingly controversial as the number of new homes in rural areas increased in the 1990 s and 2000 s. A moratorium was put in place in 2009 that removed this exemption but was only applied to the upper half of the county. We exploit this fact to use a difference-indifference estimation strategy that helps remove other unobserved factors that would make a more straightforward hedonic analysis difficult. Our goal in estimating the hedonic regression is to estimate the impact of the moratorium under the idea that the value of water is capitalized into home prices following 2009. We compare sales of properties with a house and those without a house to provide a robustness check. A so-called difference-in-difference-in-difference model is estimated which provides additional controls thanks to the fact that the moratorium did not directly affect houses in municipalities which always had to acquire a water right. Our results are important in the development of water markets in the state which are an efficient way to reallocate water from low to high value uses, but are only just beginning to be used in the region. 2

1. Introduction The primary purpose of this paper is to estimate the value of a water right in Kittitas County, WA through a hedonic difference-in-difference analysis of single-family house transactions. We take advantage of the closing of a loophole in water law that allowed new homes to be built without acquiring a water right in one part of the county. Our immediate interest in the result is to inform the development of water markets in the region which are currently nascent but could become the central means of reallocating water towards higher uses at a time when natural supply is diminishing and demands are increasing. Significant drought response planning is currently being considered that includes multi-billion dollar storage enhancement projects. Having estimates of potential prices for water is critical for considering less costly alternative strategies like water markets for adapting to increased water scarcity. This research also seeks to contribute to the literature on the expansion of low density residential development in previously rural areas, often referred to as exurban development. Research aimed at understanding the drivers of low density residential development in the United States has received a great deal of interest in the economics and regional science literature the last twenty years as this type of development became the dominant form of residential land use change. Over 75% of land converted to residential use from another use was on a parcel of more than one acre from 1994 to 1997 (Heimlich and Anderson, 2001). A point of contention that has continued is over whether this pattern of growth is the result of household preferences or unintended consequences of policy related to land use zoning (Newburn and Berck, 2009), transportation infrastructure (Baum-Snow, 2007), or other related areas. A number of papers have focused on environmental amenity driven growth where the demand for open space amenities attracts home buyers in a way that leads to its demise (Chen et al., 2009). Nechyba and Walsh (2004) argue that many of the externalities cited in the literature as posing a welfare loss are in fact aesthetic judgments, such as reliance on the automobile and should not factor into a welfare accounting. Glaeser and Kahn (2003) push back against the argument that expanding residential uses onto farmland 3

poses a threat to society by noting that forests have expanded onto farmland to a much greater extent in recent decades. In this paper we do not directly seek to estimate the degree to which land use change in rural areas was the result of the water law loophole, but our estimate of the impact of the moratorium on price partially informs this question. Kittitas County is on the eastside of the Cascade Mountains in Washington State which receives much less rainfall and many more days of sun per year than the Seattle area. Critically, Interstate-90, which extends to Seattle, runs directly through the middle of the county and provides easy access from the Greater Seattle area. The Seattle-Tacoma-Bellevue Metropolitan Statistical Area has a population of 3.5 million people. Depending on departure and destination, the total driving distance from Seattle to Kittitas County is about 80 miles and much of that is on highway with a 70 mile per house speed limit. Therefore, it is not the conventional exurban growth or suburban sprawl in that it is too far for a daily commute. While hard data is not available to test this, anecdotal evidence is that the demand for lowdensity housing comes from second homes and retirees. Kittitas County contains two markedly different land cover types. The western half is on the downslope of the mountains and is largely forested with a number of large lakes. The eastern half extends to the Columbia River and is more of a desert type climate with few trees. Both areas contain outdoor recreation, scenic vistas, and a dry climate. As is the case in most of the Western U.S. property rights over water are governed by Prior Appropriations, which relies on the concept of first in use, first in right. During times of drought when surface water levels drop below in-stream floor requirements water is curtailed for out of stream uses where those with the most junior rights are curtailed first. In most cases, new uses of water for farming and municipalities are not permitted without acquiring a water right either through permitting by the Washington Department of Ecology or buying an existing water right. Central Washington is, for the most part, closed off to additional water rights. An historical exception to this requirement was provided by so-called permit exempt wells, which were part of the water code that had the aim of reducing the 4

regulatory burden of homes on large farms from having to acquire a water right. The rationale was that the domestic use was negligible and would not affect overall water availability. From an economic perspective, this constitutes an incompletely defined property right over water resources. It became increasingly popular over the last twenty years to use the exception for rural housing developments. The issue became controversial in the 2000 s as the number of permit exempt wells grew rapidly and senior water rights holders believed they were being curtailed more often as a result. While many of the senior rights are serviced by surface water the geology of the region has been found to be characteristic of hydraulic continuity between surface and groundwater where groundwater withdrawals lower surface water levels in relatively short (within season) time horizons. This is in contrast to groundwater withdrawals from deep aquifers that contain fossil water which was deposited many hundreds or thousands of years in the past and has no interaction with surface water. From the late 1990 s through the early 2000 s more than 100 new wells were drilled each year. If true, the impairment of senior rights holders as a result of permit exempt wells constituted a violation of state water law. This led to political action to put a moratorium on the practice of permit exempt wells. The first informal request for a moratorium was made in 2007. However, it was not until the middle of 2009 that a decision was made by the Washington State Department of Ecology to put in a temporary moratorium. While it was initially temporary it was made permanent without any gap in enforcement. An interesting aspect to this decision, which provides the motivation for the empirical approach used in this paper, is that the moratorium on exempt wells only applied to what is called the Upper Kittitas which covers the higher elevation forested portion of the county. Importantly, the line delineating Lower and Upper Kittitas is based on geological formations that influence the degree of hydraulic continuity, which is ideal in terms of statistical identification compared to a line drawn based on some aspect of the housing market. Another important aspect to the moratorium that helps in our empirical strategy is that it did not directly affect municipalities who are required to have a water right. Therefore, the change in housing 5

prices at the time of the moratorium in municipalities in Upper Kittitas provides a control group in addition to the Lower Kittitas. The increase in home values that may have occurred as a result of the moratorium provides a convenient way to estimate the value of a water right. This argument simply follows from the idea that the change in the price of an existing house with a permit-exempt well has the implied value of the water right capitalized into its value. This is an important time to have estimates of these values as water trading has started to increase although there is still only one significant water seller (Suncadia) in the area. Interestingly, media reports initially projected a price of $4,000 to acquire a water right for a new home. Estimates quickly tripled to over $12,000. The average residential rural lot size in the county is in the 2-4 acre range. For comparison, the discrepancy between irrigated and non-irrigated farmland in the region is roughly in the $4,000 to $7,000/acre. Greater confidence in the estimate of the value of a water right from house sales would inform the potential transfer of water from agricultural to residential uses. The effect of water rights on housing prices has been considered in other studies. Petrie and Taylor (2007) looked at the effect of requiring people to acquire agricultural irrigation permits in order to pump groundwater from Flint River, Georgia using a hedonic pricing framework. They found that irrigation permits have a positive impact on land value. Other studies include Crouter (1987), Butsic and Netusil (2007), Faux and Perry (1999), and Hartman and Anderson (1962). Three aspects of the exempt well moratorium allow us to take advantage of the difference-in-difference approach (Ashenfelter and Card, 1985; Card and Kruger, 2000). Our data is all residential real estate transactions in Kittitas County from 1988 to 2013. This provides the first difference as we can analyze sales before and after the moratorium. The second difference is derived from the fact that the moratorium only applied to the Upper Kittitas. They are still permitted in the Lower Kittitas. The third difference, allowing us to estimate what is typically referred to as a difference-in-difference-in-difference model (DDD), arises from the fact that the moratorium did not affect the cost of building a new home in 6

municipalities. Therefore, cities and towns in Upper Kittitas provide a control group within the treatment region. The primary set of data we analyze is for parcels with a house where we hypothesize that the moratorium will have had a positive impact on prices. For additional robustness we also analyze the sale of parcels without a house that are zoned for residential use. The opposite prediction would be made in this case where we hypothesize that the moratorium will have a negative impact on price because existing homes that drilled and used water previous to the 2009 moratorium are grandfathered in. Any bare land sales after 2009 would have to acquire a water right. This second model importantly provides a check against other explanations for a positive relationship between the moratorium and house prices. For instance, one could argue that economic growth in Seattle accelerated in 2009 due to the presence of firms such as Amazon, Inc. If this were the case, and the effect of the moratorium were negligible, then we would also expect a positive impact of the moratorium timing on residential parcels without a house. The paper proceeds with a review of the theoretical considerations of amenity driven growth relevant to this analysis. This is followed by a review of the Kittitas County data. The empirical model is then explained together with results. The paper concludes with a discussion of how the results from this paper inform the literature on drivers of residential land use change in a high environmental amenity region. 2. Data Description Data was acquired from the Kittitas, Chelan, and Okanogan County Assessors Offices. Sales data was available from 2000 through the end of 2013 for all three counties in the form of an ArcGIS geodatabase containing detailed records of all parcels in the county that includes fields on house characteristics, parcel size, and owner. A GIS parcel map allows for detailed analysis of locational and landscape characteristics of each parcel. Sales records extend earlier than 1988 although this year marks the beginning of a detailed record of all parcel sales in the county. The dataset is cleaned to include only arm-length sales which are identified by a positive sale price. The dataset identifies four types of parcels with number of 7

sales in parentheses: commercial, exempt, farm, and residential. The non-residential observations are dropped. Additional details on the sales records included in each regression are discussed below. There were a small number of observations deemed to be outliers based on price per square foot of house area. These values were likely errors. Even if they were not they were deemed to potentially influence results to too great of a degree. Parcel and House Characteristics Before proceeding to analyze sales data a description of the housing stock and residential parcels is provided perspective on the house attributes portion of the hedonic regression results. The total area of residential land in Kittitas County contained in the dataset is 16,752 acres. For context, Kittitas County covers a total of 1.47 million acres, much of which is publically owned by Federal or State government agencies (1.144 million acres). There are a total of 10,274 residentially zoned parcels with a house. We assume that the parcels that do not contain house characteristic information do not have a house. We think this is an acceptable assumption as the motivation for the creation of these datasets is for collecting property taxes, so there is a strong incentive for the Assessor s offices to have house information if it exists. Table 1. Number of residential parcels. Region Lower Kittitas Upper Kittitas City 3,168 1,422 Non-city 2,168 3,516 To focus on the portion of the distribution containing most residential parcels, the histogram in Figure 1 shows the distribution of residential properties by lot size cutting off the graph at 5 acres. The spikes at 1, 2, and 3 acres likely reflect the preference of developers for building on lots in whole numbers for the benefit of potential consumers. While there has been a proposal for allowing exempt wells again with a 8

minimum lot size there was no such requirement in the past. The only limit was a maximum of 5,000 gallons per day for a well, and that the well be placed on a parcel at least 1 acre. However, this well could service multiple homes as long as the pumping rate was not exceeded. Figure 1. Histogram of residential parcel lot size in acres with a house. The parcel database contains a number of other fields including floor area, basement area, garage area, number of bedrooms, number of bathrooms, and number of plumbing fixtures. All area measures are in square feet. Summary statistics for these variables are shown in Table 2. Table 2. Mean and median for house characteristics. floor area basement area age bedrooms bathrooms garage area mean 1,524.9 191.1 40.7 2.7 1.8 151.1 median 1,432 0 29 3 2 0 The age of house provides a means for looking at home building on an annual basis provides an initial clue into the impact of the recession and the moratorium. The drop in 2007 is dramatic. While one must 9

1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 Number of new homes built per year. be careful not to read too much into this sort of graph it is interesting to see that a jump home building in Lower Kittitas in 2010 while non-municipal Upper Kittitas saw a decline in homes built. 180 160 140 120 100 80 60 Lower noncity Upper noncity 40 20 0 Figure 2. Number of new homes built by year in non-municipal Upper and Lower Kittitas by year. 10

1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 Wells drilled 90 80 70 60 50 40 30 20 10 0 Figure 3. Number of domestic wells drilled per year in Kittitas County. Two additional variables included in the database used by the Assessor s office are (1) condition of the house, and (2) quality of the house. These are used to determine assessed value. We include them in our regression as a valuable indicator of house characteristics that are often otherwise unobserved. Specifically, the quality of the home and how well the home has been maintained. Quality reflects things like building materials. Both variables are scored on a value between 0 and 60 where most houses are graded on 10 point intervals from 0, 10, 20, and so on with some exceptions. A bar chart of each variable is shown in Figure 4 and Figure 5. 11

Figure 4. Bar chart of reported house condition. Figure 5. Bar chart of reported house quality. In addition to house characteristics a few other important locational features are included for each house. In addition to the whether the parcel is in Upper or Lower Kittitas it is known which parcels are in one of 12

the five municipalities in the county. This information is used for determining clustering for standard errors in the regression analyses. Table 3. Number of houses in municipalities. City name Number of houses Cle Elum 744 Ellensburg 2,889 Kittitas 282 Roslyn 555 South Cle Elum 120 Another location variable is distance to the nearest major water body. There are very scenic lakes and rivers in the region so it makes sense to hypothesize that homes closer to these garner a higher price. There was a consideration to include land cover to account for whether a parcel is forested. However, it was determined that the dummy variable for Upper and Lower Kittitas captures much of this. Also, parcel by parcel investigation showed that a number of parcels in Lower Kittitas have a good number of trees but are surrounded by land with almost no trees. Therefore, the parcel specific value would likely be misleading. Sales Data We now move to summarizing the sales data which extend from 1988 to 2013. This long time series is valuable for establishing a long-run trend in the DDD regression. Nominal prices were converted to real prices using the U.S. Census Fisher Price Index of Single Family Houses. All prices reported in the remainder of the paper are deflated. The S&P/Case-Shiller Home Price Index was also used but did not alter results. A time-series of the number of house sales per year separately for Upper and Lower Kittitas and for cities and non-cities is shown in Figure 6. It is not surprising that Upper Kittitas, noncity and Lower Kittitas, city have the most sales as they also have more houses relative to the other two categories. A total number of sales per year is provided in Figure 7 to clarify sample size by year. 13

1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 Number of house sales 500 450 400 350 300 250 200 150 Lower, city Upper, city Lower, noncity Upper, noncity 100 50 0 Figure 6. Number of house sales per year. 1200 1000 800 600 400 200 0 Figure 7. Total number of residential land sales in Kittitas County by year. 14

1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 Number of sales per year Focusing specifically on sales of parcels without a house, total by year for non-city Upper and Lower Kittitas is shown in Figure 8. Interestingly, the drop in sales activity actually occurred two years before the beginning of the real estate crash in 2007. 500 450 400 350 300 250 200 Lower Upper 150 100 50 0 Figure 8. Number of bareland zoned residential parcel sales in noncity Upper and Lower Kittitas by year. A time-series plot of mean and median price per square foot is shown in Figure 9. The climb in prices over the full time period and the decline is still present even after deflating by the national home price level. For comparison, Census data shows the national average in the vicinity of $70 to $80 per square foot nationally, but in the range of $80 to $110 for the Western U.S. Therefore, housing price per square is about 30% higher in Kittitas County compared to the region. 15

1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 $/square foot 180 160 140 120 100 80 60 40 20 0 Mean Median Figure 9. Price per square foot for Kittitas County residential land sales between 1988 and 2013. Data is available on which houses have a well and the year the well was drilled. The effect of the moratorium is as expected and obvious in Figure 3 which shows a steep drop in 2009. There was a drop from the peak of 2006 to 2007 and 2008 which may be related to the real estate crash. Many houses are sold more than once over this time frame, which can create problems for violating independence assumptions made in statistical inference. However, in our data set most houses only sold once during the 1988 to 2013 time period (. About half as many sold twice. A smaller number sold three or more times. 16

Table 4. Tabulation of houses by number of times sold between 1988 and 2013. Number of times sold Frequency 1 5,777 2 2,256 3 1,305 4 567 5 251 6 81 7 24 8 10 9 3 3. Empirical Analysis The DDD approach relies critically on being able to separate one s sample into treatment and control. In regards to Kittitas County, the DD approach is possible because the moratorium on exempt wells only went into action in the Upper Kittitas after July, 2009. Therefore, the sample can be divided into the treatment group of Upper Kittitas after 2009. An additional control group is provided within the treatment geographic region as the moratorium did not affect municipalities who already have a defined water right. There are a few points where one could question the reliability of these divisions. The greatest confidence is with the geographical boundary of the treatment and control areas defined by the Upper Kittitas (treatment) and the Lower Kittitas (control), which is shown in Figure 10. It also helps that the boundaries of the Upper Kittitas are defined by hydrological continuity between surface and groundwater which is not especially relevant to other characteristics that influence home buying decisions. That said, most of the Upper Kittitas is different in natural amenities than the Lower Kittitas. Kittitas County extends into the Cascade Mountains on its western half. The eastern half of the county is in the rain shadow and is arid. However, these are time invariant features that can be controlled for in a regression. 17

There is more uncertainty over defining the time-break. While the implementation date of the moratorium is very clear there is a chance that expectations of the moratorium going into effect could have had an effect previous to this date. An initial petition was received in 2007 which generated a great deal of political action among various parties including senior water rights holders, developers, and local politicians. If home buyers and developers were anticipating the moratorium there could have been an increase in prices to start using water for beneficial use in advance of the pending moratorium. While it was not done in this study a future analysis could attempt to analyze prices at a finer temporal scale in accordance to major news announcements about the potential of a moratorium. Figure 10. Kittitas County designating Upper Kittitas (gray) and Lower Kittitas (white). A second potential pitfall comes in considering the municipalities a control group. If one believed the entire Upper Kittitas is a single real estate market then the moratorium could also put upward pressure on prices in the municipalities. We believe there is enough evidence to be less concerned with this. The municipalities in Upper Kittitas, in particular, cannot be credited with the growth in the rural population in the region given their size. We also have some degree of confidence that the two real estate markets 18

are separate. The rural residential buyers are second home buyers or retirees who have a preference for cabins in the country. Another obstacle to identification is provided by the real estate market crash that occurred two years previous to the moratorium instigating a great deal of volatility, and also occurred at the same time as the initial push for the moratorium, around 2007. However, the real estate market crash is what makes the DDD approach so valuable. It allows us to focus on a divergence in trend between the treatment and control groups where the control groups were also affected by the moratorium. There is some chance that the real estate market crash affected Upper and Lower Kittitas differently in a way that is similar to the moratorium which presents the risk that the crash effect is attributed mistakenly to the moratorium. Having an additional difference provided by the fact that the well-exemption only directly affected parcels outside of municipalities helps account for this. As discussed previously, the original intention of the well-exemption was for homes on large farms so that there was a maximum allowed density to quality. Therefore, the moratorium on exempt wells did not directly affect costs associated with acquiring a water right for higher density housing. Any differential effect in the recession on low versus high density housing is controlled for by the fact that we have the comparison of Upper and Lower Kittitas that both have low density houses. With these descriptions of the treatment and control groups in mind, our empirical model can be written concisely as ln(v) = Xβ + Zγ + αd + ε (1) where V is the deflated sale price of the house, X is a vector of parcel characteristics (house characteristics, parcel size, and neighborhood characteristics), Z is vector of dummy variables, and D is the interaction dummy variable equal to one if all the conditions of the control group are met. Specifically, D = 1 in the DDD model if the house is in Upper Kittitas, is not in a municipality, and if the 19

sale occurred after July, 2009. The key parameter of interest is α as it isolates the effect of the moratorium, which can be written as a difference in conditional means (see Angrist and Pischke). To be more explicit about the interpretation of the key variable in the DDD model, it is the result of differencing values for the conditional mean of the sale price to isolate the treatment effect. The first quantity is the difference in the conditional mean of Upper (U) Kittitas non-city (N) house sales before (1) and after (2) the moratorium (subscript explanations were given in parentheses). The second quantity is the difference in the noncity sales prices before and after the moratorium. The last quantity is the difference in sale price for houses in cities before and after the moratorium. α DDD = (Y U,N,2 Y U,N,1 ) (Y L,N,2 Y L,N,1 ) (Y U,C,2 Y U,C,1 ) (Y L,C,2 Y L,C,1 ) (2) Part of the appeal of the difference-in-difference approach is the ease of estimation and interpretation of coefficient estimates. What has gained much more recognition, and was largely ignored in the early DD literature, is the care that one needs to take with inference. This was prominently pointed out by Bertrand et al. (2003) for the case of serial correlation in models with long time series. Additional issues arise due to the inclusion of variables that only vary at the group level (Moulton, 1990). These tend to be easier to address when the number of groups is large. Common examples include state, county, or school level data. Downward biased standard errors are a greater challenge when the number of groups is small. The analysis reported in this paper potentially suffers from both the serial correlation and small groups problem. Cluster-robust standard errors are used to account for unobserved group level variation over city/region groups and also clustering by year separately. Interacting groups and time to have NxT groups violates the independence between groups assumption so it is not advised. For example, a group for the city of Ellensburg in 2009 and Ellensburg in 2010 is clearly not independent. Therefore, the clustered standard errors reported here do not satisfactorily account for all the potential issues of inference. A solution for simultaneously accounting for serial correlation, group level shocks, and 20

uncertainty over the correct distribution of the t-statistics has been proposed by Donald and Lang (2007) and will be included in updated drafts of this paper. The results from the DDD model estimation are shown in Table 5. The sample extending from 1988 through 2013 has a total sample size of 11,731 observations. The model has an R 2 of 0.26. Following Bogart and Cromwell (2000), a time trend is included separately for the treatment and control region to allow the two regions to have different time trends. The next set of variables describe the house, parcel, or neighborhood. The signs are as would be expected. Some of the coefficients are not statistically significant although this is likely in part due to a good degree of collinearity. For example, larger houses tend to have more bathrooms. It is interesting that lot size is not significant. Given that this is a high environmental amenity region it makes sense that parcels near marquee locations like the large lakes in the region are smaller and more expensive. A non-scientific visual inspection of the parcel layer shows some evidence of this. Houses with more square footage are more expensive. From our results, a 10% increase in square footage increases the house price by 3.2%. For a 2000 square foot house that costs $200,000 this means that increasing the size to 2,200 sq. feet increases the price of the house by $6,400. Additional bedrooms and bathrooms increase the price at a 95% and 90% significance level. There is some evidence that older homes as measured by the age of the house are more expensive. An explanation is that the best parcels were developed first. Condition of the house shows no statistically significant relationship with price while quality does. It is reassuring to see that higher quality is associated with higher price in terms of house characteristics that would otherwise be unobserved. Not surprisingly, houses closer to the water were more expensive. An increase in distance of 10% lowers the house value by 11%. The remainder of the variables relate to the DDD estimation. We focus on the coefficient estimate for the triple interaction term d2009*dupper*dnoncity. The coefficient estimate of 0.3297 on this dummy variable is statistically significant at greater than 99% confidence level (p-value is 0.002). In terms of 21

measuring the impact of the moratorium on house price, the coefficient on this dummy variable calculated to imagine the counterfactual scenario of removing the moratorium is equal to 100*(exp(-α)-1, or -27%. This is a fairly dramatic number. The median home sale price in non-city Upper Kittitas in 2013 was $240,000. A drop of 27% corresponds to -$64,800. Table 5. DDD model estimation results for parcels with a house between the years 1988 to 2013. Coef. s.e. t-stat p-value Intercept 8.9604 0.7202 12.44 <0.001 Upper time trend 0.0602 0.0047 12.78 <0.001 Lower time trend 0.0339 0.0028 12.02 <0.001 Log lot size 0.0297 0.0228 1.3 0.241 Bedrooms 0.0217 0.0068 3.2 0.019 Bathrooms 0.0698 0.0325 2.15 0.076 Log floor area 0.3221 0.0656 4.91 0.003 Garage area 0.0003 0.0001 4.22 0.006 Condition 0.0065 0.0055 1.18 0.283 Quality 0.0127 0.0059 2.14 0.077 Log age 0.1169 0.0499 2.34 0.058 Log water distance -0.1122 0.0359-3.12 0.021 d2009-0.1497 0.0392-3.82 0.009 dupper -0.3457 0.1289-2.68 0.036 dnoncity 0.0321 0.0546 0.59 0.578 d2009*dupper -0.4427 0.0708-6.25 0.001 d2009*dnoncity 0.0542 0.0105 5.17 0.002 dupper*dnoncity 0.0089 0.1034 0.09 0.934 d2009*dupper*dnoncity 0.3297 0.0642 5.13 0.002 R2 0.26 # of observations 11,731 Notes: 1. All house "area" variables are in square feet. 2. Lot size is measured in acres. 3. Cities listed are dummy variables. 4. The prefix "d" denotes a dummy variable. 5. Cluster-robust standard errors are used where groups are each municipality, non-city Upper Kittitas, and non-city Lower Kittitas. 22

While the long time series helps to reveal the long-run trends between the treatment and control regions there may be other factors that increase that are difficult to observe that have changed over time. To see whether results are robust to shortening the time horizon results from the same regression limiting the sample to sales from 2000 on is shown in Table 6. For space we will not go back over the house, parcel, and neighborhood characteristics as most change little in terms of interpretation. The sample size is 7,274 and the R 2 is slightly lower at 0.24. The coefficient estimate is again significant at the 99% confidence level and is positive which corresponds to the moratorium increasing house prices in non-city Upper Kittitas. The magnitude is lower at 0.2095. As far as the impact of removing the moratorium this corresponds to a decrease in home prices of 18.1%. For the same median home this is $43,200. 23

Table 6. DDD model estimation results for parcels with a house between the years 2000 to 2013. Coef. s.e. t-stat p-value Intercept 8.5918 0.7292 11.78 <0.001 Upper time trend 0.1007 0.0139 7.22 <0.001 Lower time trend 0.0509 0.0104 4.89 0.003 Log lot size 0.0213 0.0129 1.65 0.15 Bedrooms 0.0052 0.0084 0.62 0.56 Bathrooms 0.1155 0.0396 2.91 0.027 Log floor area 0.2984 0.0654 4.56 0.004 Garage area 0.0003 0.0001 4.22 0.006 Condition 0.0055 0.0054 1.03 0.341 Quality 0.0162 0.0046 3.51 0.013 Log age 0.1118 0.0541 2.07 0.084 Log water distance -0.0954 0.0213-4.48 0.004 d2009-0.2394 0.0737-3.25 0.018 dupper -0.7536 0.2697-2.79 0.031 dnoncity -0.0073 0.0338-0.22 0.836 d2009*dupper -0.5559 0.1308-4.25 0.005 d2009*dnoncity 0.0966 0.0221 4.38 0.005 dupper*dnoncity 0.1036 0.0716 1.45 0.198 d2009*dupper*dnoncity 0.2095 0.0559 3.75 0.01 R2 0.24 # of observations 7,274 Notes: 1. All house "area" variables are in square feet. 2. Lot size is measured in acres. 3. Cities listed are dummy variables. 4. The prefix "d" denotes a dummy variable. 5. Cluster-robust standard errors are used where groups are each municipality, non-city Upper Kittitas, and non-city Lower Kittitas. Before proceeding to interpret these values in terms of the value of water we also report estimates from regression analyses of sales of parcels without a house, or bareland parcels. As explained previously, this is an important robustness check against an event like a surge in the demand for housing in non-city Upper Kittitas corresponding to the start of the moratorium. If we found a positive and statistically significant relationship on bareland sales then we would have concern that this sort of thing is what we were capturing rather than the moratorium. It would be much less likely that this was happening if we 24

observed a negative relationship as the demand for all zoned residential land, without or without a house, would increase. Of course, we drop all house characteristic variables. We also limit the sample to noncity houses since most of the land within the municipalities already have houses. This means that this constitutes simply a difference-in-difference model. We estimate this regression on two separate samples. Results in Table 7 include fewer observations than Table 8 because only sales are included on parcels that have never had a house. The second set of results also include sales for parcels that did not have the current house constructed on them at the time of the sale. There may be some additional uncertainty with the second sample as there could have been a house previously that was torn down. However, if that were the case it seems unlikely that the old house would have affected the value of the parcel very much. The reason for the second larger sample is to remove a potential selection bias of parcels that are still not developed. These are likely of lesser quality in terms of scenic view or any other measure if other parcels were developed earlier. By including parcels that now have a house on them but didn t previously we can try and limit that selection effect. All coefficient estimates are significant at a level greater than 99%. This likely signifies that additional covariates could be included such as whether the neighboring parcels are develop and proximity to recreation activities other than water. These sort of open space characteristics do introduce uncertainty in terms of measurement error so we do not try and include them here. We also do not see why they would alter results to a significant degree. Heteroskedasticity robust standard errors are used in both regressions. The standard errors are likely downward biased since we do not account for spatial autocorrelation although given their level of significance it seems likely that inference would not change substantially. The coefficient of interest is the interaction dummy term d2009*dupper, which is equal to -0.4945 in the smaller sample and -0.5287 in the larger sample. Since these are relatively close we will just consider their interpretation with a value of -0.5. The median sale price of this type of parcel in 2013 in Upper Kittitas was around $100,000. The coefficient estimate 25

corresponds in a counterfactual sense of removing the moratorium to a 32% increase in the value of the land, or $32,000. It makes sense that the percentage term is higher for bareland sales given the lower average cost of land without a house compared to one with a house. Table 7. DD model results for bareland parcels currently without a house. Coef. s.e.* t-stat p-value Intercept 12.7473 0.1474 86.49 <0.001 Upper time trend 0.0951 0.0025 37.97 <0.001 Lower time trend 0.0445 0.0040 11.06 <0.001 Log lot size 0.1503 0.0111 13.52 <0.001 Log water distance -0.2224 0.0146-15.19 <0.001 d2009-0.3216 0.0895-3.59 <0.001 dupper -0.7466 0.0624-11.96 <0.001 d2009*dupper -0.4945 0.1072-4.61 <0.001 R2 0.18 # of observations 8,662 Notes: *HK robust Table 8. DD model results for parcels currently without a house or that did not have the current house at time of sale. Coef. s.e.* t-stat p-value Intercept 12.4567 0.1233 101.03 <0.001 Upper time trend 0.0980 0.0020 49.49 <0.001 Lower time trend 0.0424 0.0033 12.67 <0.001 Log lot size 0.1714 0.0095 17.98 <0.001 Log water distance -0.1965 0.0123-15.97 <0.001 d2009-0.2745 0.0828-3.31 0.001 dupper -0.6911 0.0513-13.47 <0.001 d2009*dupper -0.5287 0.0982-5.38 <0.001 R2 0.21 # of observations 11,634 Notes: *HK robust 26

4. Discussion In total the estimates from these regressions are consistent with a story that the moratorium has increased the value of land with houses that will have their permit-exempt wells grandfathered in. Also, the decrease in the value of land without a house provides a critical robustness check as it limits the possibility of any explanation related to an overall increase in the demand for residential land in the Upper Kittitas coinciding with the moratorium. The attractive feature of these results is that we can directly interpret them as representing the capitalization of water values into home prices. The thinking is that the alternative for buyers is that to build a new home they would also have to go out and buy a water right. If doing so costs less than the price of an existing home then that would make sense. Our interpretation of the results is that they are in the realm of what could be expected although definitely on the very high end. However, we do not have any immediate explanation for why they are so high other than the value of water is high. There are a couple reasons to have some confidence in these values despite their magnitude. First, there is only one large water seller in Kittitas County so far and many parts of the county require additional geological research to determine whether they are suitable for mitigation from this right. This is a costly process in terms of time. Second, initial estimates of prices quickly jumped from $4,000 to $13,000 for a typical home as of the summer of 2014. It could be that these estimates are still low which would make sense given that there seems to still be only a smaller number of sellers engaging in transactions. Many sellers may be watching the increase in the price of existing homes with exempt wells and concluding that prices will continue to climb. Current prices would only have to double to reach the low end of our estimates after having tripled in the initial two or so years of trading. 27

5. Conclusion This papers reports results from a hedonic regression analysis of land sales in Kittitas County, WA using a difference-in-difference estimation strategy. The primary objective was to estimate the value of a water right in the region. We exploited a moratorium on permit exempt wells that previous to 2009 allowed building a new house and drilling a well without acquiring a water right. The implication of these results are important for providing estimates of water values in the nascent water markets that are emerging in the area, in large part due to the moratorium. We also see the results from this study as relevant to the larger literature on the unintended effects of resource policy (water, land, etc.) on the pattern of residential development. While we do not directly estimate the additional amount of residential land that results from allowing permit-exempt wells our results show large price increases following the moratorium. While it depends on the elasticity of demand for residential housing in the region it seems likely that there was more residential land development than there would have been otherwise. Future research will explore this question more directly. In the meantime, it appears as if buyers are willing to pay a significant premium to buy an existing house rather than attempting to buy an existing water right and build a new home as was common in the 1990 s and 2000 s in the region. 28

6. References Ashenfelter, Orley C., and David Card. "Using the longitudinal structure of earnings to estimate the effect of training programs." (1984). Baum-Snow, Nathaniel. "Did highways cause suburbanization?." The Quarterly Journal of Economics (2007): 775-805. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. How much should we trust differences-indifferences estimates?. No. w8841. National Bureau of Economic Research, 2002. Bogart, William T., and Brian A. Cromwell. "How much is a neighborhood school worth?." Journal of Urban Economics 47.2 (2000): 280-305. Butsic, Van, and Noelwah R. Netusil. "Valuing Water Rights in Douglas County, Oregon, Using the Hedonic Price Method1." (2007): 622-629. Card, David, and Alan B. Krueger. "Minimum wages and employment: a case study of the fast-food industry in New Jersey and Pennsylvania: reply."american Economic Review (2000): 1397-1420. Chen, Yong, Elena G. Irwin, and Ciriyam Jayaprakash. "Dynamic modeling of environmental amenitydriven migration with ecological feedbacks." Ecological Economics 68.10 (2009): 2498-2510. Crouter, Jan P. "Hedonic estimation applied to a water rights market." Land Economics (1987): 259-271. Donald, Stephen G., and Kevin Lang. "Inference with difference-in-differences and other panel data." The review of Economics and Statistics 89.2 (2007): 221-233. Faux, John, and Gregory M. Perry. "Estimating irrigation water value using hedonic price analysis: A case study in Malheur County, Oregon." Land economics (1999): 440-452. Glaeser, Edward, and Matthew Kahn. Sprawl and Urban Growth. NBER Working Paper Series (2003), w9733. Hartman, Loyal M., and Raymond L. Anderson. "Estimating the value of irrigation water from farm sales data in Northeastern Colorado." Journal of Farm Economics 44.1 (1962): 207-213. Heimlich, Ralph E., and William D. Anderson. Development at the urban fringe and beyond: impacts on agriculture and rural land. No. 33943. United States Department of Agriculture, Economic Research Service, 2001. Moulton, Brent R. "An illustration of a pitfall in estimating the effects of aggregate variables on micro units." The review of Economics and Statistics(1990): 334-338. Nechyba, Thomas J., and Randall P. Walsh. "Urban sprawl." Journal of economic perspectives (2004): 177-200. Newburn, David, et al. "Economics and Land Use Change in Prioritizing Private Land Conservation." Conservation Biology 19.5 (2005): 1411-1420. 29