Abstract
This paper examines a Massachusetts policy that encourages communities to raise money through referenda for preservation and affordable housing. I use difference-indifferences, fixed effects, and quantile regression to compare home prices before and after such referenda in two towns. I include covariates representing existing land uses, zoning, and historic resources to estimate the value of these amenities. Standard regression techniques indicate weak effects of the referenda, while the estimated coefficients on land use and historic preservation confirm that preservation has a positive effect on property values. The quantile regression sheds light on some heterogeneity that goes unnoticed in standard regression results. (JEL Q24, R14)
I. Introduction
Preservation of local land and historic resources provides a number of amenities that may be valued by consumers. Among these are recreational opportunities, wildlife habitat, cultural amenities, or simply the views provided for passersby. Also, such resources may be valued simply for providing an absence of development, independent of other amenity values (Irwin 2002). The loss of these resources to development erodes these benefits.1
Relatedly, consumers have historically been, and continue to be, concerned about local land uses and zoning laws when purchasing a house. Most consumers prefer to be convenient to shopping and employment centers, but also insulated from the associated dis-amenities (noise, traffic, etc.). Uncertainty about future land uses also plays a role; an open field could remain open space or become a residential development or a shopping mall. These possible uses are likely to have very different consequences for property values.
In recent years, states, counties, and local communities have increasingly used voter referenda to raise funds for the preservation of natural and historic resources. According to the Trust for Public Land, voters cast ballots in more than 1,500 referenda on preservation between 1996 and 2005. More than 1,200 of these passed, raising almost $30 billion for preservation.2 These referenda span 43 different states and seem to indicate an intense and growing interest in preservation amid pressure from ever-expanding metropolitan areas. A common assumption about preservation policies is that they will, in addition to preserving local resources, have a positive effect on property values by preserving amenities and restricting the supply of housing. If this is true, then an obvious question is whether residents who are indifferent to preservation will vote for the policies in order to increase their property values.
In this study, I focus on one particular preservation policy, the Massachusetts Community Preservation Act (CPA).3 I examine the effect of the policy on property values as well as the impact of local land uses and zoning laws. To test the impacts of the CPA, I implement a treatment/control strategy comparing two towns that passed the CPA to two very similar towns that did not. By collecting a time-series of data on property transactions spanning years before and after implementation of the CPA, I attempt to measure the short-run property-value effects of the policy at the local level using a difference-indifferences estimator, as well as fixed-effects analysis on a subsample of my data containing “repeat sales”—properties that sold more than once during the study period. To examine the effects of local land uses and zoning laws I include covariates in the full-sample analysis to represent open space, different developed land uses, and zoning regulations. I also use quantile regression to reveal how estimated effects of all of the explanatory variables vary across the distribution of home sales by price, thus relaxing the assumption that the effects of any of the covariates are homogenous.
Estimates of the effect of the referendum on property values are mixed and insignificant, while the estimated coefficients on land-use and historic district covariates confirm that, in this region, preservation appears to have a positive effect on property values. Since the tax costs of the CPA are incurred immediately, but benefits are uncertain and delayed, the insignificance of the referendum effect is not surprising. In turn, the hypothetical average voter, if indifferent to preservation amenities, does not have an incentive in the short run to support preservation referenda. This leads to a bit of a puzzle, however, about why the referenda have passed. It turns out that heterogeneity in property-value impacts, as evidenced by the quantile regression results, in combination with other individual incentives, can account for passage. In one of the two case studies, the referendum is seen to significantly increase property values for homes in the top 35% of homes by value, while the other, smaller, program negatively impacts property values in some quantiles.
II. Preservation in Massachusetts: The CPA
In September 2000, the CPA was signed into law in Massachusetts to enable towns to leverage their local property taxes with state funds to help in the preservation effort. It allows towns in Massachusetts to charge property owners a tax surcharge of up to 3% on their annual property tax bill, and provides state matching funds to supplement those raised locally. Both locally raised and matching funds must be spent in three areas: open-space preservation, historic preservation, and community housing, with a minimum of 10% spent on each. The proceeds are held in special accounts and are not available to address other local spending priorities. It is important to note that everything about this policy is implemented and decided at the local level, subject to only basic restraints.4
In this paper, I focus on two towns, Acton and Chelmsford, that have implemented the CPA.5 In Acton, a proposal including a 1.5% surcharge and exemptions for low-income households and the first $100,000 of property value passed a town meeting in April of 2002. On November 5, 2002, that proposal passed in a referendum vote with 55% of voters approving the initiative. Chelmsford passed the CPA with 61% support, including a smaller 0.5% surcharge and an exemption for the first $100,000 of property value, on April 3, 2001.6 The proposal was voted onto the ballot at a town meeting in February of that year.7 It is worth noting that the CPA charge is an itemized charge on each town’s tax bills so that residents are reminded every year that the CPA has been enacted and are aware of exactly how much it is costing them.
The tax surcharge in Acton went into effect for the tax year 2003. Since then, three years’ worth of funding has been allocated. The CPA has raised a combined total over those three years of approximately $3 million, half of which came as matching funds from the commonwealth of Massachusetts. Of this, more than half, approximately $1,850,800, has been set aside for the preservation, acquisition, or development (for the purposes of recreation) of open space. Another $650,000 has been spent or set aside for community housing, and $435,505 for historic preservation. In Acton, as a result of the CPA, the average single-family tax bill in 2005 was approximately $100 higher (1.5% of the reported average single-family tax bill of $6,900).8
The CPA surcharge in Chelmsford raises approximately $400,000 in taxes and state matching funds each year, less than half of what is raised in Acton. Similarly, the average tax bill in Chelmsford in 2005 was increased by only about $20 as a result of the CPA. As of January 2007, only $575,000 had been spent or allocated out of a total fund of approximately $1.6 million. Of the money allocated, $135,000 had been spent on open space, $103,000 on historic preservation, and $337,700 on community housing. Two of the three funded community housing projects were for senior housing, and the town has identified 20 parcels totaling 359 acres for possible future open-space preservation.9
The towns that implement the CPA do not have to be specific about projected uses of the money raised. In practice, Acton and Chelmsford certainly were not before the referendum, and plans continue to be quite vague. Much of the money raised has simply been set aside for future acquisition of open space or future use in other as-yetunknown projects. Town long-term plans are vague at best and set forth amorphous goals such as the “preservation of rural character.” This seems to indicate that, even after a few years of implementation, the physical outcomes of the CPA are largely uncertain and likely not to be realized for several more years.
III. Previous Literature and Theoretical Predictions
The existing literature measuring the value of open-space preservation is deep and has established a largely positive relationship between publicly owned or protected open space and property values.10 This relationship between open space and property values, however, is heterogeneous, and some studies have found no or even a negative relationship, including those by Shultz and King (2001), Smith, Poulos, and Kim (2002), and Geoghegan, Lynch, and Bucholtz (2003). There have been no studies, to my knowledge, that employ quantile regression techniques to the issues of land use, zoning, and historic preservation, while there is every reason to believe that there will be some heterogeneity across the distribution of homes. Indeed, I know of only one hedonic study using quantile regression, that of Zietz, Zietz, and Sirmans (2008), and it does not include any land-use or preservation variables.
In addition to open space, a small literature suggests that there are positive amenity values capitalized into home sales prices from historic preservation. Mason (2005) provides a nice survey of many strands of literature relating to the value of historic preservation, most of which point to positive net impacts of historic preservation. Specifically, historic designation seems to have a positive effect on the property values of designated properties.11
These hedonic studies, however, largely fail to look at the ongoing effects of specific policies.12 Kotchen and Powers (2006) and Nelson, Uwasu, and Polasky (2007) look at which factors influence whether a town holds a preservation referendum and also whether a town passes a referendum given that one is held, using community-level census data. In effect, these studies examine the causes of preservation referenda, while this paper concentrates on the effects.
The CPA can be thought of as having many different possible impacts on property values. First, while it increases the quantity of presumed amenities, it also raises taxes. In the classic model of Brueckner (1982), average property values in a community will be maximized at the optimal level of public spending (and associated taxation), meaning there are two intuitive effects: a demand effect13 and a tax effect,14 which work in opposite directions. An important further effect of preservation is a supply effect, since the CPA effectively provides funds for towns to remove property from the real estate market. Finally, one complicating factor is the effect of community (or “affordable”) housing, which is less clearly a public amenity and may, in fact, reduce property values.15 All in all, the effect of the CPA on property values is an empirical question—one on which I hope to shed some light.
IV. Data and Emperical Strategy
The dataset that I have gathered includes all single-family residential property sales for the years 1996–2005 for four towns in the northwestern suburbs of Boston: Acton, Andover, Chelmsford, and Tewksbury. As discussed above, both Acton and Chelmsford have adopted the CPA, and the passage of the CPA serves as the treatment. Andover and Tewksbury are included in the dataset as controls.16 Each treatment town is paired with a control town in a case study format: Acton with Andover, and Chelmsford with Tewksbury.17 The data gathered include sales price, sales date, and a number of characteristics of the home. These data were purchased from the Warren Group, a real estate and financial information firm. In addition to the information above, the sales dataset contains information on the street address and the latitude and longitude of the properties. I am able to merge this locational information with a number of external “layers” available from the Massachusetts Office of Geographic and Environmental Information (MassGIS). Data obtained from MassGIS includes information about open-space parcels, land uses, location of major highways, highway exits, rail corridors, commuter-rail stations, historic landmarks, town centers or villages, and location of solid-waste facilities and facilities deemed to be of “greatest potential environmental significance” by the Massachusetts Department of Environmental Protection’s Bureau of Waste Prevention. In addition, I collected some time-series data at the town level, including monthly unemployment rate data from the U.S. Bureau of Labor Statistics, and town revenue and expenditures from the Massachusetts Department of Revenue, Division of Local Services.
Full–Sample Analysis
The basic strategy is to pair each CPA town with a similar town that did not adopt the CPA. I then use a dummy variable to denote those transactions that took place in an affected town after a referendum was passed. This technique is commonly referred to as difference-in-differences estimation.18 Existing local land uses and historic resources are included as covariates in order to estimate their respective marginal effects on sales prices. This, then, allows me to approximate how adding an additional acre of open space (or preventing the conversion of an acre of existing open space to a developed use) near a home will affect the sales price of that home.
Analysis of town selection. To minimize the likelihood of an endogeneity problem in my analysis, the towns in my case studies were chosen using town-level census and geographical data to be as similar as possible so that population sorting is minimal. To begin this selection process, I compare all towns in Massachusetts using town-level data on population (2000), population change (1990–2000), the share of town land undeveloped in 1999, the percentage change in that share between 1985 and 1999, age characteristics, average income, average home value, the share of homes that is owner occupied, and the average residential tax burden. Using this data, I use a probit model to compute a predicted probability of a town holding a CPA referendum, as of 2004. This predicted probability is the propensity score. Calculated predicted probabilities for the four chosen towns are shown in Table 1. Note that for my chosen case studies, the predicted probabilities of holding a referendum are not statistically significantly different.
Summary of Predictions, with Standard Errors in Parentheses
I then compare each treatment town to each nontreatment town in Massachusetts by comparing the calculated propensity scores along with the vector of town-level characteristics.19 Pairs are eliminated from consideration if calculated propensity scores are statistically significantly different from one another. I then calculate a statistic, the sum of squared differences between each pair of towns over this vector. Comparisons for which this statistic is small are examined in detail, with close attention paid to any economically significant differences in any variable, and narrowed to a set of 10 possible pairs.20 Using this measure of closeness, the pairs analyzed here, Acton/Andover and Chelmsford/Tewksbury, are the two closest (have the two smallest values for the statistic, the sum of squared differences) of the final 10 pairs under consideration.21
Following the probit model discussed above, I also perform weighted least-squares estimation of the log (odds ratio) on town and referendum characteristics using data on those Massachusetts towns that held a referendum on the CPA.22 This model enables me to calculate a predicted “yes-vote” share for each town, assuming, for each comparison, that the characteristics of the hypothetical referendum in the control town would be the same as those of the actual referendum in the affected town. The results of these predictions are also in Table 1.23 For neither case study are the differences in this predicted probability statistically significant. This is indicative of the towns being quite similar as regards their observable attitudes toward the CPA. That is to say, based on observable characteristics, the towns in each case study are predicted to have the same propensity for holding a referendum and then passing the CPA.
Local fixed effects. In addition to the careful choice of towns for this study, I employ local-area fixed effects. Including dummy terms for each census block group controls for any static, local characteristics that may influence both property values and the town’s propensity to pass the CPA. I further allow error terms to be clustered at the census-block level. This makes the standard error estimates robust to spatial autocorrelation at the census-block level.24 This use of clustering and small-scale spatial dummy variables is one version of spatial econometric methods and follows Davis (2004) among others. Essentially, the spatial weighting matrices are implicitly specified to allow for spatial dependence within census-block groups and spatial autocorrelation within census blocks. That is, the spatial weighting matrices (one for the spatial dependence and the other for spatial autocorrelation) are implicitly discrete, where entries are 1 if two observations are in the same census-block group (spatial dependence) or census block (spatial autocorrelation) and zero otherwise. As shown below, my results are robust to changes in the spatial scale of both fixed effects and error clustering. Intuitively, these controls are using the time-series nature of my data to control for the fact that populations sort between neighborhoods based on characteristics that, to me, are unobservable. Left uncontrolled this sorting could lead to both omitted variables bias and spatial autocorrelation. What remains after inclusion of these neighborhood effects is property-level heterogeneity, which is controlled for with property-level covariates and dynamic local factors, perhaps coincident with passage of the CPA, which can be controlled for using time-series data on monthly unemployment rates, annual property tax rates, and public expenditures—those dynamic factors most likely to be acting at the local level on prices.
An equally important part of the analysis is the inclusion of covariates representing static land-use, geographic, and structural characteristics of the properties. I used ArcGIS to gather information on a number of different land uses.25 In the preferred model, described below, I include four categories of land use: passive open space, low-density residential, other residential (all other categories of residential land use), and commercial and industrial land. The passive open-space category allows me to measure the effect on home prices of converting undeveloped forest or agricultural land to developed land uses.26
For each of the land uses in my dataset, I measure the total acreage of parcels of that type which intersect a buffer zone around the property, measured from the latitude/longitude point provided in my data. For a graphical example of this calculation, see Figure 1. This measure of land uses is appealing since it takes account of the size of different land-use parcels in addition to their proximity to the home—an effect that would not be captured by measuring the share of the buffer zone taken up by each type of land use. The estimated effects may vary depending on the size of the buffer zone considered. There is a certain amount of multicollinearity between different land uses (for instance, those properties which are near commercial land are more likely to also be near higher density residential land, highways, etc.), which makes the measurement of their capitalized value difficult. This problem is made worse by the consideration of larger buffers. In addition, as buffer size increases there is less variation in the land-use measures across observations. This is because the buffers of different homes will increasingly intersect each other, becoming more and more similar in their land-use characteristics. The “cleanest” measure of land-use effects is, then, the smallest, a 0.1-mi buffer.27 Unfortunately, this is also the narrowest measure of land-use effects, and clearly cannot be considered a comprehensive measure of local land-use impacts on property values. I primarily report estimates for this smallest of buffers, but also discuss how estimates vary for different buffer sizes.
Example of Land-Use Measure
In this paper, I do not attempt to measure a direct effect of preserving existing historic resources. Instead I aim only to get a rough idea about the direction of the effect of historic resources on property values. To do this, I use two measures. The first is a dummy variable for whether or not a home is in a recognized historic district.28 One could imagine this effect going either way since it is likely to have positive aesthetic effects, but often also implies limitations on a homeowner’s ability to alter existing structures. The second is a dummy variable for whether or not there is a recognized historic landmark within 1 mi of a transaction. This captures less formal neighborhood effects.29
I normalize all sales prices according to the Office of Federal Housing Enterprise Oversight’s (OFHEO) House Price Index (HPI), in addition to employing year and month fixed effects. The HPI is compiled at the metropolitan division level.30 Normalization places all observations of sales price in constant dollar terms. The year and month dummies serve to control for extra-municipal price trends and seasonal effects, respectively.
As with any hedonic analysis, omitted variables bias is a potential problem. The variables chosen to represent the effects of open space and historic preservation are not comprehensive—they are surely leaving some effects out and possibly standing in for others. Nonetheless, the analysis performed here is designed to be representative of the effects of preservation on property values. Another problem of the hedonic analysis might be an endogeneity problem similar to that above. For instance, if there is something about an area that impacts property values and makes it more attractive for, say, commercial/industrial development, then I might be erroneously attributing low property values to the commercial/industrial development. However, the local-area fixed effects should ameliorate this problem, since fixed effects will control for any such common factor. Similarly with historic neighborhoods, by using fixed effects, if there is something about a neighborhood, say that it is filled with older, historic, mansions, then that might be correlated both with property values and with its likelihood of being designated a historic area. However, the fixed effects will net this out and capture heterogeneity within that small area only, which is likely to be closer to homogeneous as regards any common factors.
Repeat Sales Analysis
In addition to the full-sample analysis described above, I perform a property-level fixed-effects analysis by looking at homes that have sold more than once in the 10 years included in my study. There are 2,663 such properties in my dataset, of which only 694 have at least one sale both before and after the CPA passed in the respective towns. The fixed effects effectively eliminate static, otherwise-unobserved, property-level characteristics from the error term of the regression, leaving only those factors that are changing over time coincident with the CPA.31 However, the small sample size will limit the power of the analysis to measure the referendum effect.32 As in the full-sample model I include some town-level time-series data to try to capture other dynamic effects.
Quantile Regression Analysis
In the prior analyses I implicitly assume that property-value effects, in percentage terms, are constant across the distribution of homes by price. In fact, one would expect instead that the effects of many of the covariates would differ depending on the value of the home. In particular, very low value homes (those assessed at less than $100,000), are exempt from the surcharge associated with the referendum, implying that they would receive any benefits from the referendum without having to pay the costs.33 As another example, more expensive homes will tend to have larger lots, which likely makes nearby open space less valuable. I test these hypotheses of heterogeneity using quantile regression, a procedure originated by Koenker and Bassett (1978).
Whereas standard least-squares regression models estimate the mean relationship between a dependent variable and a number of independent covariates, quantile regression estimates the effects of covariates on a dependent variable at different points of the distribution. So, if the referendum affects homes in the lower 10% quantile of homes (according to price) differently than homes in the 50% quantile and the 90% quantile, quantile regression allows for the consistent estimation of these different effects.34 Using the same covariates as in the full sample model above, I estimate quantile regression coefficients for 19 quantiles (0.05, 0.1, 0.15, etc.). Fixed effects are included at the census tract level.
V. Analysis
Results of the econometric analyses are presented in Tables 3 through 7. The fullsample specification presented in Table 3 uses the entire dataset. The second specification, presented in Table 6, employs property- level fixed effects on a dataset comprised of properties that sold more than once between 1996 and 2005.
Full Sample Analysis
I employ a simple ordinary least squares model using a log-log/semilog specification for the continuous variables. Estimates of the referendum effect vary depending on the case study, although they are never found to be statistically significant. In the case of Acton, the referendum effect in the preferred model is estimated to be positive and insignificant, with a coefficient value of 0.011, implying a 1.1% increase in property values. In Chelmsford, the estimated effect is negative and generally small, and also not statistically significant, with a value of 20.004, corresponding to a 0.4% decline in property values. These point estimates are robust within small ranges to changes in the specification of the model. Based on this analysis, then, one would have to conclude that there is no consistent or measurable effect of the referendum on property values, at least in the short term, in these two communities.
The estimates of land-use effects provide evidence of a preference for open space over some developed land uses, and a general preference against commercial/industrial development. In the preferred model (shown in Table 3), passive open space has a positive and significant effect in Chelmsford/Tewksbury but no significant effect in Acton/Andover. In each case study, commercial/industrial land uses are estimated to have significant negative impacts. Together these results imply that passive open space is preferred to commercial/industrial land. This argument is strengthened by the estimated effects of zoning—homes in areas zoned residential are of higher value than those in other zones. The other measured land uses do not have significant effects in Acton/Andover, though in Chelmsford/Tewksbury low-density residential development has a significant positive effect. The varying estimated effects for the two case studies are not out of line with the existing literature and may be partially explained by the relative prevalence of different land uses. As Table 2 shows, the average home in Acton or Andover has much more open space intersecting its 0.1-mi buffer than one in Chelmsford or Tewksbury.35
Summary Statistics
Full Sample Regression Results
Using the coefficients provided in Table 3, I calculate the predicted change in sales price from having one additional acre of each land use intersect the 0.1-mi buffer. These marginal effects are reported in Table 4 for each land use, for both the mean and median home, by normalized price, in each case study. As an easy heuristic, suppose that a 10-acre plot of farmland intersecting a property’s buffer is developed. Commercial/industrial development of this parcel would cause a statistically significant decline in price in the range of $500 and $1,000. The results for residential development, that it has either no statistically significant effect, or in the case of Chelmsford/Tewksbury, a positive effect of almost $400, suggest that some types of residential development may be more valuable than open space.36 Alternatively, open space may be undervalued if its future use is uncertain. That is, low-density residential development may be seen as an improvement since it resolves this uncertainty.
Estimated Marginal Effects of Land Use and Preservation ($)
The choice of buffer size affects the estimates obtained for the value of various land uses, as seen in Table 5. In Acton/Andover, moving to even a 0.25-mi buffer results in a negative and significant effect for passive open space; while, increasing the buffer to 1 mi results in positive significant estimates for both classifications of residential housing, and insignificant negative estimates for open space and commercial/industrial land. This non-robustness is likely to be related to the measurement problems discussed above. Estimates for Chelmsford/Tewksbury are considerably more stable, with there being no qualitative change in the estimates from increasing the buffer to 0.25 mi and than 0.5 mi. There is a reversal, however, when considering a 1-mi buffer. At that range, open space is negative and significant while commercial/industrial land is positive and significant. In addition to the measurement problems associated with larger buffers, this may also reflect a preference for having commercial strips close, but not too close.
Land-Use Coefficients at Various Buffer Sizes
The measures of historic resources are mostly positive in both case studies, but not always significant. The historic site dummy is significant in Chelmsford/Tewksbury, but neither measure is significant in Acton/Andover. There are substantial differences between the towns in the number of historic properties, which helps to explain these results. In particular, Andover has many more historic sites than the other towns in the study, which limits the successful estimation of the effect of the historic sites dummy variable (nearly 85% of observations in Andover are within 1 mi of a historic site) and reduces the marginal effect of being near one of those sites or in a historic district. In contrast, there are relatively few such sites in Chelmsford or Tewksbury, and almost no homes are located in recognized historic districts (less than one-third of 1% of observations).37
Evidence about the relationship between sales prices and proximity to other elements of development is, again, mixed, but mostly of the expected sign and sometimes significant. For instance, in both case studies, proximity to active rail lines is negatively associated with property values (the estimated coefficient is positive, implying that property values increase as distance to active rail lines increases). Interestingly, in Chelmsford/Tewksbury, prices are simultaneously negatively related with proximity to rail lines but positively related to proximity to commuter rail stations (this effect is significant in some specifications).38 In Acton/Andover, home prices significantly increase with distance from highways as well as solid waste facilities. There is no significant effect from proximity to town centers. Finally, another interesting result of the analysis is the negative impact that a parcel’s being zoned for conservation has on its sales price. Clearly, the restrictions on land use that come from this zoning designation dampen the value of the property.39
To begin to understand how local land uses and the effect of the referendum may interact, I estimate the preferred model, now including interaction terms of the referendum with the land-use variables. The interaction terms capture heterogeneity in the effects of the referendum depending on the land-use characteristics around each home. In Acton/Andover, the interaction terms are not significant and do not significantly change the estimates from the base model. In Chelmsford/Tewksbury the original land-use estimates are left qualitatively unchanged, although they do become more significant. The referendum variable changes sign, becoming positive, but remains insignificant. In addition, the interaction terms are significant: positive for commercial/industrial and negative for the other three land-use variables. This implies that the referendum effect is more likely to be positive in areas near commercial and industrial land uses and less likely to be positive in areas near open space or residential development. This matches the intuition that homes in more developed areas are more likely to benefit from preservation.40
I also tried a specification that included interaction terms between the referendum and all other explanatory variables.41 On the whole, the interaction terms were not significant, with a few exceptions. In Acton/Andover, being zoned commercial increases the effect of the referendum, as does proximity to a solid waste site (being closer makes a positive referendum effect more likely), while the number of bathrooms, strangely, makes the referendum effect more likely to be negative. In Chelmsford/Tewksbury, the two significant interaction terms are the historic district dummy and the home’s age. Both indicate a positive effect: being in a historic district or being an older home increases the likelihood of the referendum effect being positive. This is likely due to additional funds being available for historic preservation.
Repeat-Sales Analysis
Results for the base fixed-effects model are presented in Table 6. Obviously, the relevant sample sizes in this analysis are smaller than above. Taking the two case studies together, there are 2,663 properties that sold more than once during the 10 year study period, yielding 5,741 sales. This is about 40% of the observations included in the full-sample analysis. Additionally, there are only 694 observations on which to identify the referendum effect (those properties that sold at least once both before and after the referendum in CPA towns), 243 in Acton and 451 in Chelmsford. The results are similar to those of the full-sample analysis, with a positive effect in Acton/Andover and a negative effect in Chelmsford/Tewksbury, but neither of these estimates is significant.
Repeat-Sales Analysis
The estimates of the referendum are quite robust to changes in the specification of the model, as seen in Table 7. Due to the loglinear specification, these estimates can be interpreted as the percentage change in price attributable to the passage of the referendum.
Fixed-Effects “Referendum” Estimates
In Acton, the point estimates, with two exceptions, indicate very small price changes of no more than 0.6%. The estimates for Chelmsford are considerably larger in absolute magnitude, implying price declines of between 3.8% and 8.0%, but are still not statistically significant.
Among the most important of the robustness checks I consider is the effect of redefining the referendum variable. I consider two alternative definitions. The first redefinition takes into account that the effects of the CPA may not be fully felt until the first tax year in which the surcharge will be in place, but that people will begin to anticipate and prepare for the change ahead of this date. The referendum variable, then, takes on a value of 0.5 at the time of passage and gradually increases, month by month, until the tax year begins, at which point it takes a value of 1. The second alternative definition uses the nature of CPA passage to its advantage by taking a value of 0.5 when it is announced that the CPA will be on the ballot and a value of 1 when it is passed. Neither of these alternative measures affects the estimates of the referendum effect.42
In order to test the validity of my parsimonious use of town-level time-series variables, I experiment with specifications including the residential tax rate and the log of per capita town expenditures (including school/education spending). These values are assigned to each sale observation by the year of the sale. The effects of this re-specification are small. In Chelmsford, the inclusion of these variables increases the magnitude of the estimated referendum effect but reduces its statistical significance. Unfortunately, these public spending variables exhibit very little variation, since they are only measured annually and, obviously, at the town level.
The degree to which my choice of HPI normalization affects the analysis is tested by dropping normalization altogether as well as by instead normalizing according to the Consumer Price Index (CPI). In both alternatives, the Chelmsford estimate remains basically unchanged, while the Acton estimate becomes substantially more negative, although not significant. This suggests that when the normalization is omitted, the model attributes some price trends that are actually happening at a larger geographic level to the referendum effect, which is obviously incorrect. The normalization then is important as it captures extra community price trends. Note that using CPI normalization produces similar effects to not normalizing at all, since the CPI is calculated at the MSA level and not at the smaller Metropolitan Division level, so that there is no geographical variation in this calculation in my dataset.
A Quantile Regression Approach
Figure 2 presents a series of summary charts showing the quantile regression estimates for different quantiles in each of the two case studies for a set of the included explanatory variables. The dark lines represent the coefficient estimates at each quantile, and the dotted lines represent the upper and lower bounds of the 95% confidence intervals, achieved through bootstrapping. These analyses were conducted with the same 0.1-mi buffer used in the full-sample analyses above.
Quantile Regression Results
In Acton/Andover, there is a distinct upward trend in the effect of the referendum. That is, more-valuable houses experience a positive effect while less-valuable houses are more likely to experience a negative effect, and the positive effect is statistically significant for homes above the 65th percentile. This important result indicates that the CPA had a significant positive impact on property values for roughly 35% of homeowners. This result may stem from a distribution of preferences related to the distribution of home prices; if consumers/residents of higher-priced homes have stronger preferences for preservation, then one would expect the referendum to be more likely to have a positive effect, despite the additional implied tax price. In Chelmsford/Tewksbury, there is no definitive trend across the quantiles, although the referendum coefficient is negative throughout and significant for the 55th, 85th, and 90th percentiles. The referendum in Chelmsford was a much smaller program, and it is reasonable to think that there are threshold effects from public spending on open space that would cause Acton’s program to have positive effects while Chelmsford’s does not.
The effects of local-level land uses are also interesting to examine in the quantile regression format. In Chelmsford/Tewksbury, additional open space has a positive impact on property values, although the magnitude of this effect declines over the distribution. This suggests that private land, which is correlated with higher sales prices, may be a substitute for open space. Commercial/industrial land generally has a negative and significant impact, in line with the standard regression, although the absolute magnitude again seems to decline over the distribution, particularly in Acton/Andover. Medium- to high-density residential land is generally positive in Chelmsford/Tewksbury, and occasionally significant, while it is mostly negative in Acton/Andover, and significantly so in the higher end of the distribution. Finally, low-density residential land is positive and significant in both case studies at the lower end of the distribution but insignificant at the high end of the distribution. Much like in the case of open space, this may suggest that more expensive homes, with more land, are less affected by surrounding residential development, which is intuitive. This argument is strengthened by considering that the land area of a property is not only positively related to price, but also increasing over the distribution; it is more important to more expensive properties.
The effects of historic districts and historic sites do not exhibit a strong trend up or down across the distribution of homes and are rarely statistically significant. The historic site dummy is significantly positively related to home prices in most of the range for Chelmsford/Tewksbury.
Zoning variables do exhibit some trends. In both studies, being zoned commercial or industrial has a negative impact on sales price across most of the distribution, an effect that was not significant in the standard regression. Interestingly, this does not hold in the upper end of the distribution, where the effect becomes insignificant. As noted above, conservation zoning also has a negative impact, and this is marked in the upper half of the distribution, but insignificant at the lower end. This suggests that higher-end properties are more substantially impacted by the restrictions this implies.
The effect of the interior square footage of a property, like that of land area, is positive and increasing over the distribution. The effect of the number of bedrooms, on the contrary, is positive but decreasing over the distribution. Together these suggest that lower-end buyers care more about the number of bedrooms, while higher-end buyers care about the size of those rooms. This seems consistent with an argument that bedrooms act like a necessary good, while the size of those rooms is more of a luxury good. Finally, the age of a home is negatively related to price, but the magnitude of this effect declines over the distribution. This may be because more expensive older homes are likely to be well taken care of and have “character,” whereas cheaper older homes are more likely to be “fixer-uppers.” All of these results are consistent with those of Zietz, Zietz, and Sirmans (2008).
Altogether, the quantile regression analysis makes an important case for being wary of standard hedonic estimates that do not account for heterogeneity across the distribution. There are important, significant, impacts that may be hidden in the “average” coefficient estimated by standard ordinary least squares, like the positive impact of the referendum in the higher-end homes of Acton.
VI. Discussion
The evidence presented in this paper indicates that the CPA, in the short term, has no overall statistically significant effect on property values in the towns I studied. However, there is evidence from the quantile regression approach that the CPA did have statistically significant effects on homes in some price ranges. Nonetheless, these results suggest that the CPA and similar preservation policies do not produce a widespread increase in property values, even given the state matching funds, which amount to, if not free, then certainly discounted money for local projects. This further suggests that the tax effect together with the affordable housing requirement are counterbalancing the demand and supply effects. In the model of Brueckner, the towns in this study may be at the top of the curve where the negative and positive marginal effects balance out, indicating that Acton and Chelmsford, already, are at or near the optimal level of preservation. If indeed the costs and benefits are balancing out, then one conclusion could be that the CPA has produced a median, annual, gross benefit of around $100 per household in Acton and of around $20 per household in Chelmsford (the respective, approximate, median tax costs of the policies). However, this conclusion should be drawn with caution, since while one cannot reject the null hypothesis of no net effect, neither can one prove this hypothesis.
Another explanation for the lack of a strong property-value effect is the uncertainty associated with the CPA in these communities. Most of the money, as I have noted, has not yet been spent or, in the case of Chelmsford, even allocated among the three goals of the policy. In this situation, if a resident owns property adjacent to a plot of privately owned woodland, she likely does not know whether that land will be preserved through the CPA, developed, or used through the CPA to provide community housing. If the first, her property values would likely increase, but if the last, likely be depressed. At the aggregate level, this uncertainty about where and how CPA funds will be used might prevent significant appreciation in sales prices in the wake of the CPA.
An interesting puzzle is, if there is no general, significant, positive capitalization of the CPA into property values, what motivated voters to pass the CPA and face the increased tax implications? There are a couple of possible explanations. A first explanation revolves around local heterogeneity in expectations and incentives. In Acton, 55% of voters supported the CPA while 61% did the same in Chelmsford. This implies that 55% and 61% of voters, respectively, expected to be made better off by its passage. Given the turnout numbers provided earlier (59% in Acton, 9% in Chelmsford), however, only a much smaller share of eligible voters would have to expect to be made better off, and we show that approximately 35% of voters in Acton could have anticipated such a benefit from property values. Another group of voters with strong incentives to support the CPA was renters, who occupied 16% of housing units in Chelmsford and 24% in Acton in 2000.43 Finally, those residents who are exempt from the surcharge (those whose assessed home value is less than $100,000 or with “low” incomes) also have an incentive to support the policy. It is still quite possible, then, for property values to decline or hold steady on average, yet rise for enough people to result in CPA passage by the observed margins.44
A second explanation involves a lack of perfect foresight either about the direct effects of the CPA on the community or about its effect on property values. In particular, given the vagaries of the policy’s implementation, voters might have expected one set of outcomes but observed thus far the unfolding of a different set in practice. For instance, some CPA project proposals in both towns have been rejected, including, in one case, a proposal to improve the local Pop Warner football field. If supporters of this proposal, or others that may have been turned down, expected it to go through given the passage of the CPA, they may have supported the CPA.
Finally, there are some values, widely recognized in the environmental economics literature, to open space, such as existence values, that may not be capitalized into home prices. Such values could provide incentives for people, particularly those with environmentalist tendencies, to vote for the CPA even in the face of negative financial implications. Together with imperfect expectations and the possibility that some voters correctly anticipated significant appreciation of their properties, existence values may easily have been enough to pass the referendum despite the lack of general property-value appreciation.45
Despite the findings on the effect of the referendum, I find that undeveloped open space is preferred to some types of land development, as evidenced by the hedonic analysis. In particular, preservation of a 10- acre plot of open space being considered for commercial/industrial development near the median home is estimated to have a positive price impact of between $500 and $1,000. The analysis further points to the fact that larger parcels of land may be a substitute for undeveloped open space. In particular, it is shown in the standard regression analysis that low-density residential parcels have comparable value to open space. The quantile regression results further support this since higher-value homes, which are likely to have more land, experience less of a positive impact from nearby open space, and for these homes, enlarging the parcel has a higher positive marginal effect on value. This implies that those able to afford large lots may, instead of relying on the public provision of open space, provide it privately for themselves.
As for policy implications, this analysis points to the fact that cost-benefit analyses of preservation programs should not rely solely on static hedonic estimates when predicting property-value effects, nor assume homogeneity of the impacts across properties. Instead, it is important to acknowledge that property values may be stable or possibly hurt in the short run, and that the actual effect will be different across the distribution of homes. Also, the role that uncertainty about the benefits seems likely to be playing in the results suggests that proper planning and the specification of the local intentions for use of CPA funds may be critical to the realization of property- value appreciation. That is, properly allocated, CPA funds may well generate desired public goods, which would ensure that the CPA provides net positive effects on property values. Perhaps the resolution of the uncertainty alone could have a substantial positive impact.
Appendix
Referenda Texts
Acton
Shall the town of Acton accept section 3 to 7, inclusive of chapter 44B of the General Laws, as approved by its legislative body, a summary of which appears below?
Summary
Sections 3 to 7 of Chapter 44B of the General Laws of Massachusetts, also known as the Community Preservation Act, establish a dedicated funding source for the acquisition, creation, and preservation of open-space, for the acquisition and preservation of historic resources, for the acquisition, creation, and preservation of land for recreational use, preservation and support of community housing and rehabilitation or restoration of such open space, historic resources, land for recreation use and community housing. In Acton, the Community Preservation Act will be funded by an additional surcharge of 1.5 percent of the annual tax levy against real property and, as available, by matching funds provided by the state. If adopted, the following will be exempt from the surcharge: property owned and occupied as a domicile by any person who would qualify for low income housing or moderate income senior housing in the town, as defined by Section 2 of said Act; and $100,000 of the value of each taxable parcel of residential real property. Any other taxpayer receiving and exemption of real property authorized by Chapter 59 of the General Laws or any other law shall be exempt from the surcharge under this Act. A Community Preservation Committee composed of local officials designated by the Act and of local citizens will study the needs, possibilities and resources of the town regarding community preservation and will make recommendations on the uses of the funds. Town Meeting must approve any such appropriations from the Community Preservation Fund before funds can be expended to acquire any particular parcel of land or take similar action. All expenditures pursuant to the Act will be subject to annual audit.
Chelmsford
Shall the Town accept sections 3 to 7 inclusive of Chapter 44B of the General Laws as approved by its legislative body?
Summary
Sections 3 to 7 of Chapter 44B of the General Laws of Massachusetts, also known as the Community Preservation Act (CPA) established a dedicated funding source to acquire, create, and preserve open space, historic resources, land for recreational use, and community housing, and to rehabilitate and restore such open space, historic resources, land for recreational use and community housing acquired or created as provided under the CPA. In Chelmsford, the CPA will be funded by an additional 0.5 percent surcharge on the annual tax levy on real property and by matching funds provided by the state. The following exemption from surcharge, permitted under Section 3(e) of said Act, will apply: $100,000 of the value of each taxable parcel of residential real property. A Community Preservation Committee must be created pursuant to by-law and will make recommendations on the use of the funds. Town Meeting must approve any such recommendation before funds can be expended to acquire any particular parcel of land. All expenditures pursuant to the Act will be subject to an annual audit.
Footnotes
The author is assistant professor, Economics and Financial Studies, School of Business, Clarkson University. I would like to thank my dissertation committee members, Dallas Burtraw, Lucas Davis, Michael R. Moore, Emre Ozdenoren, and Stephen W. Salant, for encouragement and advice, as well as Soren T. Anderson, Jamie Emerson, Stephen Holland, Edward Knotek, Matthew Kotchen, Peter Morrow, Jody Schimmel, and seminar participants at Camp Resources XIII, the University of Michigan, the 15th Annual Canadian Resource and Environmental Economics Working Group, Yale University, the University of North Carolina–Greensboro, the U.S. EPA National Center for Environmental Economics, the University of Cincinnati, the University of Hawaii-Manoa, Clarkson University, Union College, the University of Rhode Island, and Boise State University for helpful comments and suggestions on earlier drafts. Funding for this project was provided by Clarkson University, the University of Michigan through the Michael J. Moore Dissertation Prize, and a grant from the Rackham Graduate School Discretionary Fund, as well as from Resources for the Future through the Joseph L. Fisher Dissertation Fellowship. All mistakes, of course, are my own.
↵1 In addition to these amenity costs of land development, new residential development imposes very real costs to communities from expanded demand for local services. Fausold and Lilieholm (1996) cite a number of cases of “fiscal impact analysis,” which indicate net fiscal deficits incurred by municipalities from residential development (that is, additional costs of public services exceed new tax revenues). For one recent example in the media, see “Driven by Development,” New York Times, August 17, 2005, page B1.
↵2 Trust for Public Land, Landvote Database (available at www.tpl.org).
↵3 The CPA is one example of similar preservation policies being considered and implemented throughout the country. New Jersey, as one other example of the use of referenda for preservation, also provides state funds to match those raised locally for preservation, in addition to low-interest loans for acquisition, through the Green Acres Planning Incentive Program.
↵4 The tax surcharge may include any or all of three tax exemptions: for low-income households, for the first $100,000 of property value, and for commercial and industrial properties. There is a two-part process for the local adoption of the CPA. First, the language of a ballot initiative (and obviously, the terms of the town’s implementation: the chosen surcharge rate and exemptions) must be decided and then placed on the ballot through either a town meeting or petition drive. After this, the approved referendum must pass in a townwide election. After 5 years have passed, the town may vote to revoke the CPA through the same two-part procedure described above; failing this, the CPA will remain in place indefinitely. As of October 2008, 133 towns in Massachusetts had adopted the CPA in various forms, and none had yet revoked it after passage. Once in place, decisions about the use of CPA funds are made at the direction of an appointed, local Community Preservation Committee and must be separately approved as a part of the annual town meeting.
↵5 For details as to how the study towns were selected, see Section IV.
↵6 There were a total of 8,469 votes in the Acton referendum, which amounts to a 59% voter turnout. In Chelmsford there were a total of 2,366 votes, amounting to a turnout of about 9%. This low turnout in Chelmsford is likely because the referendum occurred during an April town election in which there were few other issues on the ballot. In April 2007, after the study period, Chelmsford voted to increase its surcharge to 1.5%.
↵7 The text of these ballot initiatives is reproduced in the Appendix.
↵8 Budgetary information comes from the 2007 Acton Community Preservation Plan, available at http://doc.acton-ma.gov/dsweb/Get/Document-13574/2007+CPPlan+08.2006.pdf.
↵9 This information comes from the Chelmsford Community Preservation Committee’s Community Preservation Informational Forum, held on January 17, 2007.
↵10 See the survey by McConnell and Walls (2005) for a summary of both revealed- and stated-preference studies.
↵11 One example is presented by Leichenko, Coulson, and Listokin (2001), who use a hedonic study to examine the effects of historic designation in various Texas cities.
↵12 One exception is presented by Riddel (2001), who uses a dynamic model over 20 years to examine how property values in Boulder, Colorado, changed with a coordinated series of municipal open-space acquisitions made during that period. She finds a lag of about 1 year before changes in open space are capitalized into housing prices.
↵13 Irwin (2002) showed that people may be willing to pay a premium for open space that is preserved and no longer developable. This likely reflects that, as assets, the price of homes and land parcels should reflect not only current conditions but expectations of future conditions, and implies that, even if the CPA does not increase open space, but it preserves existing resources for the future, this should be capitalized positively into property values.
↵14 In a seminal paper, Oates (1969) demonstrated the result that higher tax rates unaccompanied by corresponding increases in public services result in lower property values. To the extent that tangible positive outcomes from the CPA come several years after the tax increase and are uncertain, one might expect the tax effect to dominate. To be clear, future benefits of the CPA will also be capitalized immediately but will be capitalized at a discounted rate due to uncertainty (what resources will be preserved) and the delay in their realization.
↵15 Massachusetts has set as a goal for all towns to have 10% of housing units deemed “affordable,” and this appears to be a major reason why community housing was included as part of the CPA. In Massachusetts, affordable housing is housing that could be afforded by a family with an income of 80% of the local median household income. Affordable housing shares in the study towns are as follows as of February 2007: Acton, 6.6%; Andover, 8.7%; Chelmsford, 6.8%; Tewksbury, 5.1%. Since I do not have complete information on the location of affordable housing units, it is not possible for me to explicitly control for this effect as part of my analysis. However, presuming that affordable housing shares do not change significantly during the period of study, this is one of many factors controlled for by fixed effects. At a local level, provision of such housing (either through construction of units or by “buying down” existing units) may reduce home values if, for instance, such units are less attractive to passersby or are perceived to be in areas of relatively higher crime. In addition, new affordable housing units tend to be built as part of higher-density development (higher, in fact, than would otherwise be allowed by local zoning in many cases). There is anecdotal evidence of a general unease with such developments—certainly they are often controversial. Also, population sorting may lead to decreases in property values around community housing, since higher-income people are likely to sort away from these areas. Just how large the effect of community housing will be remains unclear. Nguyen (2005) provides a survey of the literature on the effects of affordable housing on property values. Her conclusions are that the effect depends on many factors but is often negative and generally “small.” To the extent that community housing funds are put primarily toward senior housing, property-value effects may be minimal.
↵16 A proposal in Andover to implement the CPA failed to pass the town meeting in March 2002. In March 2008, however, the CPA did go on the ballot in Andover, and the measure failed. Tewksbury recently passed the CPA in April 2006, after the end of the study period. Accounting for the fact that residents of Tewksbury may have anticipated the CPA’s passage through a modified referendum dummy variable has no impact on the results. However, dropping 2005 sales in Chelmsford/Tewksbury, while leaving the referendum coefficient estimates unchanged, does make the estimates significant at the 10% level. This suggests that, in 2005, Tewksbury’s future passage of the CPA may have influenced relative home sale prices and weakened the estimated treatment effect. In that sense, estimates and conclusions drawn below are conservative. Realistically, we would not expect much spillover, since, while the Tewksbury referendum was announced in May 2005, it was not passed until April 2006, and not by a wide margin: 53% to 47%.
↵17 This is done to account for heterogeneity in effects of the explanatory variables across towns. As will be explored below, these groupings are extensively tested, and results support pooling of the towns into these two pairs (that is, effects are relatively homogenous for towns within each case study), but not across case studies.
↵18 One example of pairwise difference-in-differences estimation applied to property values is presented by Davis (2004), who examined the effects of a cancer “cluster” on property values by comparing property values in two otherwise similar, neighboring counties before and after several cases of cancer were diagnosed in one of the two counties.
↵19 I cannot use the propensity score matching approach to estimate the effect of the CPA on property values (the average treatment effect on the treated), since I consider only two treatment towns. The propensity score is used only to decide which sets of towns to compare.
↵20 Any pair that exhibits an important difference in any variable is eliminated from consideration. For instance, if a pair of towns has statistically similar propensity scores but a large difference in population density, this pair would be eliminated from consideration. Geography also plays a role; I eliminate pairs for which there are large geographical differences.
↵21 Ideally, more pairs would have been included in the analysis, but that was not feasible under important time and budgetary constraints. This is the focus of future work.
↵22 The probit and WLS models discussed in this section are essentially replications of the work of Kotchen and Powers (2006). I then use the results of these models to calculate predicted probabilities.
↵23 Results of the underlying regressions for these predicted probabilities are available from the author.
↵24 As a reference, the average census block group in my sample is 1.29 mi2. The average census block consists of 0.05 mi2.
↵25 These land uses are commercial, industrial, four types of residential (multifamily, high density, medium density, and low density), passive open space (forest, pasture, and cropland), recreational open space (including parks and golf courses), power lines, water, wetlands, and other. Note that I separately measure the effects of active and passive open space, and wetlands. Parks and spaces that are widely used for recreation will be capitalized very differently into home values than, say, open woodlands, and to mix these effects together would produce inconsistent results. Similarly, wetlands tend to be valued very differently than other forms of open space (see Bin and Polasky 2003) and are often off limits to development (or at least subject to thorough regulation), warranting their exclusion from a broad-based openspace measure.
↵26 School quality is omitted from my analysis. This is because of a lack of variation (like other New England states, Massachusetts school district boundaries coincide with town boundaries), and also because, presuming that relative school quality between the comparison towns does not change in any significant way (this is backed up by an examination of test scores), the effect of general school quality is picked up in the fixed effects (local level in the full-sample analysis, property level in the repeat-sales analysis).
↵27 A 0.1-mi buffer creates a circle around each parcel, with an area of about 20 acres.
↵28 To get the full measure of how historic districting impacts property values, one would want to do a treatment-control analysis much like that performed here for the CPA. This is outside the scope of this study.
↵29 Other specifications including the absolute number of registered historic places within the buffer and the distance to the nearest historic landmark are also tested and achieve qualitatively similar results. These specifications, however, seem overly restrictive in measuring the effects of historic designation.
↵30 Acton, Chelmsford, and Tewksbury are contained in the Cambridge-Newton-Framingham Metropolitan Division, and Andover is contained in the Essex County Metropolitan Division. All are part of the Boston- Cambridge-Quincy Metropolitan Statistical Area (MSA). The fact that Acton and Andover are in separate divisions implies some heterogeneity in price trends that would not be captured by the year dummies but could otherwise confound our analysis by attributing to the CPA an effect that is occurring at a larger geographical level. The HPI is calculated using sales prices from “conforming” mortgage transactions—transactions involving mortgages for less than a federally set limit. In 2004, that limit was $333,700. For a detailed description of the HPI’s calculation see work by Calhoun (1996).
↵31 In estimating this model I could instead assume a random-effects formulation. This requires one additional assumption: that the property-level random effect is independent of the regressors. A Hausman test suggests that this assumption is violated in the case of Chelmsford/Tewksbury, which further implies that fixed-effects estimation is more appropriate. In Acton/Andover, in contrast, the Hausman test suggests that random effects would be more appropriate. However, in this case, the fixed-effects estimator is still consistent, but not efficient, whereas in Chelmsford/Tewksbury the random-effects estimator would be inconsistent. For this reason, and since there is no significant difference in the coefficients in Acton/Andover between the fixed- and random-effects estimates, I report only the fixed-effects estimates.
↵32 In addition to the small sample size, another complication is the possible introduction of a bias in the sample, since this analysis will oversample “high-turnover” properties. A comparison of summary statistics between the restricted and unrestricted samples yields no significant differences in property characteristics. This indicates that such a turnover bias is likely to be very small or nonexistent.
↵33 This $100,000 threshold approximately corresponds to the 2nd percentile in Acton and the 3 percentile in Chelmsford. Since there are so few observations below this threshold, it is not possible for me to take better advantage of this discontinuity.
↵34 For an intuitive summary and review of quantile regression see work by Koenker and Hallock (2001).
↵35 It is possible that the true value function is concave in different land uses. I provide one test of this by estimating the base model with quadratic terms included for the land-use variables. Neither the linear nor the quadratic open-space terms are significant in Acton/Andover, while in Chelmsford/Tewksbury they are both significant and in the expected direction (increasing and concave). Full results are available from the author.
↵36 Since the particular measure of land use used in this analysis is novel, it is difficult to compare these results to others in the literature. Irwin (2002), however, reports the value of converting a 1-acre parcel of undeveloped pastureland to low-density residential and finds that such a conversion reduces sales price by $1,530, or 0.89%, in contrast to my findings. More generally, the hedonic results on home characteristics presented here are broadly consistent with some 125 hedonic studies summarized by Sirmans, Macpherson, and Zietz (2005) and a meta-analysis performed by Sirmans et al. (2006). For instance, the 2006 study finds an average coefficient on square footage of 0.0037, and a standard deviation of 0.0024. My estimated coefficients of 0.002 and 0.001 are thus within a 95% confidence interval of Sirmans et al.’s (2006) results. Similar analyses lead to the same conclusion of consistency for lot size, bedrooms, bathrooms, and age.
↵37 These results are qualitatively consistent with prior literature (see Mason 2005; Leichenko, Coulson, and Listokin 2001), although the point estimates are relatively low.
↵38 The station variable was left out of the Acton/Andover analysis because of its insignificance and strong correlation with the rail-line variable.
↵39 Again, these results are at least qualitatively consistent with the literature. Looking only at statistically significant results, point estimates for Andover for distance to a solid waste facility indicate an increase in sales price of approximately $600 for each mile. This is an order of magnitude smaller than the estimate of Reichert, Small, and Mohanty (1992) as cited by Boyle and Kiel (2001), although a 1-mi increase in distance amounts to a 250% increase in distance for the average home in Acton, which makes the comparison difficult. Strand and Vågnes (2001) estimate the effect of distance to railroads on property values and find an elasticity of 0.059, which compares nicely to my estimates of 0.037 and 0.034 for Acton/Andover and Chelmsford/Tewksbury, respectively.
↵40 In the above analysis, I do not allow there to be differences in the regression coefficients across towns in the same case study. The estimated regression coefficients are a weighted average of the actual coefficients for each individual town. It is worth asking what, if any, differences exist between the towns in each study. To do this, I rerun the preferred regressions, this time including a set of interaction variables that interact with a town dummy (Acton in Acton/Andover and Chelmsford in Chelmsford/Tewksbury) with each of the regressors (not including month and year dummies). If there are significant differences between the regression coefficients for different towns, then the coefficients for the interaction terms will be significantly different from zero. In Acton/Andover, there are statistically significant differences in the estimated coefficients for 5 of the 19 regressors, only 2 of which imply differences in the sign of the effect (low-density residential land is estimated to have a negative effect in Andover, and a positive effect in Acton, while a property’s land area has essentially a zero coefficient in Andover and a positive coefficient in Acton). In Chelmsford/Tewksbury, only 2 of 20 estimated coefficients are significantly different, and neither implies a qualitative difference. These results support the pooling of the towns chosen for the case studies. Along these same lines, one could argue for pooling all of the data into one analysis, relying on the fixed effects to eliminate the endogeneity problem. Such an analysis provides no additional insight and pinpoints the differences between the two case studies, which warrant them being studied separately.
↵41 Again, full results are withheld for space reasons but are available from the author.
↵42 A third robustness check attempts to account for the fact that the referendum may have different impacts in the short run than in the long run. In this case, following Donovan, Champ, and Butry (2007), I define two separate referendum variables: one to represent the first year after passage and a second to represent the later years after passage. Neither variable has a significant effect, nor is there a statistically significant difference between the coefficient estimates in the short and long terms.
↵43 This is a conclusion based on the literature on property tax incidence. This literature concludes that the burden of property taxes falls largely or entirely on homeowners and not on renters (DiPasquale and Wheaton 1996, 342–44). Rental demand curves in a particular town in a metro area are likely to be highly elastic, while rental supply curves will be highly inelastic. Further, suppose that, as a result of the public goods provision implied by the CPA, each renter’s willingness to pay for housing increased by x%. If the supply were perfectly inelastic, then the price of rental housing would increase by as much as x%. Likewise, so would the total consumer surplus, which implies that each renter who votes should vote “yes.” If the supply were more elastic, the increase in net consumer surplus would be even larger. Keep in mind that an increase in demand and rental price occurs only after willingness to pay has risen as a result of increased public goods provision—there would be no (or very little) increase in rents based solely on the increasing tax in this standard model.
↵44 For concreteness, consider the case of Chelmsford. According to the 2000 census, there are 12,812 housing units in Chelmsford, of which 2,061 are rentals. On the CPA referendum in 2001, there were 1,437 “yes” votes and 929 “no” votes. The 1,437 observed “yes” votes could be accounted for solely by renters. Therefore, every property owner could have anticipated a decline in property values, 929 voting “no” and the remainder not participating. Alternatively, each of the renters could have stayed home, and the 1,437 “yes” votes could have come from just a small fraction of the 10,751 property owners. Of the 9,314 other property owners, 929 would account for the “no” votes, leaving 8,385 not participating. In either case, passage of the proposal is entirely consistent with perfect foresight about changes in property values and a null or even negative price change on average.
↵45 The welfare implications of the CPA are, unfortunately, unclear. In a supply-demand framework, the stable general price results found here could result from a number of different scenarios. In the first, suppose that there is no net change in either supply or demand for property from the CPA (demand would increase due to the additional amenities but could decrease by the same amount due to the increased taxation; supply, in the short run, may be unaffected). In this case, there is no net change in welfare in the CPA community (although there may be, statewide, a social welfare loss since matching funds are drawn from statewide taxation). Alternatively, the price outcome could be derived from a decline in supply (due to preservation) and an associated decline in demand due to dominance (in at least the short run) of the tax effect and/or the community housing component. In this case, social welfare in the community must decline. A final possibility is that demand and supply both increase (demand from the dominance of the amenity effect and supply from the “takings” effect discussed above), in which case social welfare will have increased. Unfortunately, it is impossible, with the existing data, to disentangle these possibilities.
