Abstract
Views of natural areas and green space may have value quite apart from access to those lands. Using 25 years of home sales data from St. Louis County, Missouri, and modern geographic information system tools to measure views, we estimate a hedonic property fixed-effects model that captures the effects of changing land cover on house sale prices. Unlike previous studies, our approach minimizes bias from omitted variables and uniquely captures changes over time. We find that forest views negatively affect home prices, whereas farmland views have positive effects. Changes in relative scarcity of these land types over time may explain the findings. (JEL Q51, R14)
I. INTRODUCTION
Economists have relied on hedonic property value techniques to study the capitalization of local amenities into home prices since Rosen’s (1974) seminal paper. A review by McConnell and Walls (2005) identified more than 40 peer-reviewed studies that used hedonics to analyze various kinds of natural lands and open space, including neighborhood parks, greenbelts, forest preserves, farmland, and more. A few of those studies, and some papers published since, have emphasized the value of views of natural areas. Including a measure of views in a hedonic regression, along with measures of access or proximity, enables one to distinguish the passive aesthetic values from other active use-oriented values. In suburban and exurban areas where farmland and forests are being converted to development, the aesthetic value of open spaces—and even the simple absence of nearby development—may be particularly important (Irwin 2002; Smith, Poulos, and Kim 2002).
Early studies measured views in unsophisticated ways, often by physically observing properties and recording the absence or presence of water, forest, or other natural areas in the view (Kulshreshtha and Gillies 1993; Benson et al. 1998; Tyrväinen and Miettinen 2000; Bourassa, Hoesli, and Sun 2004; Loomis and Feldman 2003). With the proliferation of very large datasets on property sales and property characteristics—many hedonic models are now estimated on tens of thousands of observations—physically looking at properties to determine views is impractical. In recent years, new geographic information system (GIS) techniques and highly detailed spatial data have offered the opportunity to more carefully and systematically measure views, and some studies have begun to apply these methods (Paterson and Boyle 2002; Bin et al. 2008; Sander and Polasky 2009; Cavailhes et al. 2009; Hamilton and Morgan 2010). However, many of the studies rely on relatively small samples and do not employ the most up-to-date econometric methods. In particular, in recent years, economists who work with hedonic models have become increasingly concerned about misspecification of the hedonic price function and bias in estimates of the capitalization of local amenities in home prices due to omitted variables (Kuminoff, Parmeter, and Pope 2010; Parmeter and Pope 2013). Unobserved neighborhood characteristics that are important determinants of household locations, and thus equilibrium house prices, are often correlated with the amenity variable of interest and possibly other explanatory variables. This can introduce bias in estimates and create a challenge for identification.
In our study, we take advantage of a large sample of properties that have sold more than once over our time period of 1988 through 2012 to estimate a property fixed-effects model, purging the effects of time-constant omitted variables (Palmquist 1982; Livy and Klaiber 2013). We include sale year fixed effects to control for temporal trends. This approach is an improvement over much of the prior literature examining the value of views, which relied on cross-sectional variation. By contrast, our methodology allows us to evaluate the hedonic price impacts of changing land uses over time. By holding the property constant and measuring the effect of a changing natural landscape, we are able to assess a critical issue in many suburban and exurban areas: the impacts on home prices of the gradual conversion of open space to development.1 We do compare the property fixedeffects model with two simpler hedonic specifications: one that pools the data and includes spatial and sales period fixed effects, and one that includes interactions of spatial and sales period fixed effects, an increasingly common technique in hedonic analyses. We also perform a series of robustness checks on the model.
For our sample of properties in southern St. Louis County, Missouri, we estimate both the capitalization of views of and the proximity to natural areas. We use modern GIS techniques based on a 10 m digital elevation model (DEM) to measure views, and we measure physical proximity to natural areas by identifying the percentage of such lands in a buffer around properties. This approach allows us to separate the passive aesthetic values, as captured in the view, from the more active use-oriented values. We estimate the effect of views of and proximity to farmland, forests, and grassy recreational lands, along with a measure of the diversity of land uses in one’s view, using three time periods of land cover data.
The results suggest that proximity to all three kinds of open space positively affects home prices, but the effects of views are more mixed. The greater the percentage of forest in a property’s viewshed, the lower the property price, all else being equal. Although this result seems counterintuitive, other studies have found a negative or zero effect from forest views (Paterson and Boyle 2002; Sander and Polasky 2009). Views of grassy land covers and farmland have positive effects, though only the farmland coefficients are statistically significant. We find that the diversity of view has a negative effect on home prices, indicating that the greater the mix of land covers in a property’s viewshed, the lower the home value. This finding is consistent with some other studies in suburban and exurban areas, where the mix includes developed uses (Sander and Polasky 2009). Incorporating views in the model is important; a statistical test of the results compared to those of a model with just proximity rejects the proximity-only model.
It is difficult to say exactly what underlies our results, but we hypothesize that two factors are at work. The first is the topography of the study region. The average elevation is relatively low and there is limited landscape relief. This could explain why the more open farmland views appear to be preferred over views of the relatively dense hardwood forests in the area. Second, in our study region over the 25-year period for which we have home sales data, developed land cover has increased mostly at the expense of agricultural lands. Forest cover, on the other hand, has increased slightly, due at least in part to creation of an extensive greenway along a major river. Thus, the relative scarcity of alternative land cover types may partially explain the results—as farmland and grassy open spaces have declined, these pastoral views have increased in value. These types of findings about changes in land uses over time can be important in communities considering various open-space protection policies.
II. REVIEW OF THE LITERATURE
Early studies of the value of a view typically included dummy variables for whether a property had a view of a particular amenity such as an ocean, lake, or forest. Information was obtained from tax assessment records, real estate listings, or in some cases, observations made in person by the researcher. Kulshreshtha and Gillies (1993), using data from Saskatoon, Saskatchewan, find that the presence of a river view increases prices and that the value varies significantly by neighborhood. Benson et al. (1998) visually inspected properties in Bellingham, Washington, and assessed the extent of ocean, lake, and mountain views. Their results suggest that unobstructed ocean views increase the price of an otherwise comparable home by almost 60%; they also find that the value of a view varies inversely with distance. This very high value of an ocean view is consistent with findings by Bourassa, Hoesli, and Sun (2004), who estimated a hedonic price model on 5,000 properties in Auckland, New Zealand, and who also visually inspected properties to measure views. These authors find that surrounding natural land views are valuable as well. Tyrväinen and Miettinen (2000), using data on apartment sales in two towns in Finland, estimated the value of a forest view, along with access to forested recreation areas and forest preserves. They find that having a forest view tends to increase prices, all else being equal, by about 5% for an average house. In a study focused on the effects of declining lake levels on home prices, Loomis and Feldman (2003) used data on home sales in north-central California and find that lake proximity has value, as do lake views.
Studies began to move toward more sophisticated measures of views based on GIS data and tools in the early 2000s. Several studies have employed land cover GIS datasets in combination with a DEM that measures the topography of the landscape. Paterson and Boyle (2002) used data on 500 home sales in Connecticut over an 18-month period and find that the greater the share of forest surrounding a property, the higher the price, but the greater the share of the view that is forest, the lower the price; total size of the view has no statistically significant effect on home prices. Sander and Polasky (2009) took a similar approach in a study of the value of views in Ramsey County, Minnesota, a suburban area outside St. Paul. In their analysis, they include the size of the view; the percentage of the view that is forest, grasslands, or water; the diversity of the view (i.e., the number of land types present in the viewshed as a percentage of all possible types); the standard deviation of the elevations in the viewshed (as a measure of landscape relief); and a dummy for whether the property has a view of downtown St. Paul. Results show that the greater the size of the view, the greater the house price, all else being equal, and the larger the percentage of the view that is in grassy surfaces or water, the greater the price. Having a forest view has no statistically significant effect on prices. Cavailhes et al.’s (2009) hedonic study in the suburbs of Dijon, France, is similar to the studies by Paterson and Boyle (2002) and Sander and Polasky (2009), but the authors distinguish the differential impacts of land cover variables in six rings around a property, assuming that the impact of views on property prices varies by distance, which the results confirm.
Several recent studies have focused on more methodological advances with GIS; some of these have constructed a three-dimensional landscape from land cover data and a DEM. The viewscapes in these models typically account for the presence of buildings and trees and other vegetation that might obstruct views. Lake et al. (2000) and Baranzini and Schaerer (2011) used this approach; the former used data from Scotland and the latter looked at rents on properties in Geneva, Switzerland. None of the view variables are statistically significant in Lake et al.’s (2000) study, but Baranzini and Schaerer (2011) find that rents are higher the larger the view of natural areas and the smaller the view of development.
A three-dimensional view, which can account for buildings and other obstructions, is more easily constructed when LiDAR data are available for an area. LiDAR is a remote sensing technology that illuminates a target with a laser and then analyzes the reflected light.2 Hamilton and Morgan (2010) used LiDAR data in a study of the effect of ocean views on house sale prices in Pensacola, Florida, between 1998 and 2007, but the sample size in this study is very small at only about 100 properties. Bin et al. (2008) also used LiDAR data in a study of the value of ocean views. Their study area is a coastal county in North Carolina; they controlled for whether the home is in a Special Flood Hazard Area and the distance to the nearest beach, both of which are important determinants of home prices, and find that views are important.
In summary, the literature on views generally finds that ocean views have high value, and in many settings, that value declines with distance. Results for views of land-based natural amenities are more mixed, with some studies finding that forest views have positive value, some finding a negative impact, and others finding no statistically significant effect. This could be due to variations in the study region’s topography and land use, or due to variations in study methodology and design. In the studies that distinguish proximity to natural lands from views of those lands, the findings suggest that both aspects of open space matter.
Table 1 lists the studies, along with information about their data and the use of fixedeffect techniques in the econometric analysis. As the table makes clear, many of the studies have relatively small sample sizes and do not incorporate spatial, time, or property fixed effects. For studies with sales over many years, this is problematic. While the more recent studies reflect advances in accurately measuring views through more sophisticated GIS techniques and in incorporating different view elements, they have not generally made ad-vances in the hedonic methodology, particularly to control for omitted variable bias— practices that are becoming commonplace in recent hedonic studies. This brings into question the reliability of some of the findings. We are able to address this concern in our study area, as we describe in Section IV, after discussion of our study area and data.
III. STUDY AREA AND DATA
Our study area is the southern portion of St. Louis County, Missouri, specifically, all properties within a 5 mile buffer of the Meramec Greenway. The greenway is a 108 mile long stretch of protected lands that follows the Meramec River back from its confluence with the Mississippi River into its headwaters in the Ozarks, but we restrict attention to the portion in St. Louis County. The greenway has trails and river access points and is mostly forested lands. Bluffs rise above part of the river, which is scenic and frequently used for recreational activities. Most of the rest of the region is a fairly level plateau, but the Ozark Mountains begin in the southwestern part of the county, which is the least developed. Figure 1 shows a map of St. Louis County and our study region, which is 295 square miles in size and includes suburban, exurban, and rural communities (but excludes the more urbanized areas in and around the city of St. Louis). The area includes 59 municipalities, in whole or in part, and 15 high school districts. We chose this area as our study region because it includes (1) a relatively cohesive set of suburban and exurban areas and (2) a substantial amount of “green” land cover types, as well as some changes in that land cover over time.3
Map of Study Area in St. Louis County, Missouri
We obtained property sales data from the St. Louis County Revenue Department for the years 1988 through 2012. These data include a variety of structural characteristics of buildings, such as number of stories, square footage, number of bedrooms, and lot size, among other attributes. Georeferenced parcel boundaries, along with land uses, were obtained from the St. Louis County GIS Service Center.4 Using unique locator numbers, we were able to match the GIS data to the sales data. Of the properties in our study area, 47% are residential, including duplex/townhome, multifamily, and single-family homes; all of the residential properties are used in our hedonic analysis. From the GIS Service Center, we also obtained GIS shape files of the 100-year floodplain, as defined by the Federal Emergency Management Agency, and used these files to identify any parcel lying wholly or partially within the 100-year floodplain. We obtained 10 m elevation data from the National Elevation Dataset, maintained by the U.S. Geological Survey,5 which we use to calculate the centroid elevation for each residential property.
To estimate views and proximity of different land types, we use the National Land Cover Database (NLCD) from the U.S. Geological Survey. This dataset has land cover data available for 1992, 2001, and 2006, in 21 different categories in 1992 and 19 categories in 2001 and 2006. We collapse the number of categories into six groupings for purposes of our analysis—farmland, grassy recreational areas, forest, development, barren land, and open water—and we use only farmland, grassy recreational areas, and forest as our measures of natural area views and proximity.6 We do this because many of the individual classes are not present in our study area (or present only in very small amounts) or because we felt that a distinction is not important for many of the individual categories in terms of views. Table 2 shows the 2001 and 2006 NLCD categories and our own reclassifications.7
Table 3 shows the percentage of land in the study region in various land cover categories for the three periods, 1992, 2001, and 2006. Forested lands have increased slightly, while farmland and grassy recreational lands have declined. The decline in these kinds of open spaces has come largely as a result of conversion to developed uses. Population increases and an increase in average lot sizes for residential parcels have both contributed to land conversion. The increase in forest cover is attributable, in part, to additional properties added to the Meramec Greenway over time.8
As stated above, we use the NLCD land cover data to calculate our proximity and viewshed variables. We assign land cover data to home sales based on the NLCD year that is closest to the home sale year, as shown in Table 4. Although land cover does not change by a large amount from year to year, this assignment will be more accurate for some homes than others. For example, the 1992 NLCD land cover information is likely to be more precise for homes sold in 1992 than for homes sold in, say, 1988 or 1989. Moreover, land cover will be measured with more error for homes soldinyears furtherfromanNLCD year. Below, we describe two robustness checks of the model in which we alter the assumptions shown in Table 4 to address these two issues.
For proximity, we measure the percentage of each land cover type within a 200 m buffer of each property. A distance of 200 m (or 1/8 of a mile) is easily walkable and thus provides a measure of accessible open space.9 We calculate a GIS-based viewshed for each property using the 10 m DEM from the National Elevation Dataset.10 The Viewshed tool available in ArcGIS11 computes the visible areas from one or more observation points based on the elevation information provided by the DEM. In our study, the viewshed is measured by simulating the view, from each residential property, of an observer whose eyes are five feet above ground level. Figure 2 shows an example. The simulation of view is made for 360 degrees around the observation point in our data. As Figure 2 illustrates, the amount of the view depends on elevation: property 2 in the figure has a larger viewshed than property 1, in part because its elevation is higher and in part because of surrounding parcel elevations. Figure 3 shows the same two properties and the surrounding land cover in 2006, along with buffers of 0.1, 0.3, 0.5, 0.7, and 1.0 km for reference. Property 1 has more of a mix of land covers nearby and more developed areas beyond 0.5 km, whereas most land cover near property 2 is forest with some grassy areas.
Viewsheds as Calculated for Two Properties
Land Cover Type in Buffers (0.1–1 km) for Two Properties
Table 5 shows summary statistics for the viewshed and proximity variables, as calculated for the properties in our dataset, along with some selected structural characteristics for the properties and sale prices. Diversity of land cover types is computed as the percentage of possible land cover types present in a viewshed. For instance, a property that provides views of five different land cover types out of six total types would have a diversity index of 83%. We focus our attention on the property fixed-effects model, which includes only houses that have sold more than once during the study period; but because we also show econometric results for the full sample of home sales, Table 5 shows the summary statistics for both samples. It is clear that the samples are quite similar; we return to this point in the next section.
Comparing Tables 3 and 5, we can see that the percentage of land cover in forest or farmland for the entire region is significantly larger in each year than the percentage of an average home’s view of forest or farmland. For example, in 2006, forests make up about 14% of an average home’s view but nearly 27% of the land cover in the region; farmland is about 3% of an average home’s view but 10% of the region’s land cover. These results reflect the fact that these types of lands are typically relatively far from residential areas.12 Recreational grasses, on the other hand, account for roughly the same percentage of land cover as the average home’s view (approximately 14% in 2006). This is likely the result of subdivision open spaces, golf courses, and parks with playing fields, which tend to be located closer to residential properties.
IV. METHODOLOGY
As we noted in the introduction, omitted variable bias is now a well-recognized problem in hedonic property models. Some recent studies have investigated the problem and present various methods for minimizing its influence. It is now the norm, for example, to include spatial fixed effects to control for unobserved neighborhood factors that vary across geographic space. In models with cross-section/time-series data, sales period fixed effects are also often added to control for general trends over time and temporal shocks. Some studies incorporate interactions between spatial and time fixed effects; these control, to some degree, for unobserved spatial variables that vary differentially over time. In some studies, researchers have interacted time fixed effects with other predetermined variables. Linn (2013), for example, interacts time fixed effects with median lagged neighborhood (defined by grid squares) prices.
One of the most successful approaches for addressing omitted variable bias is a quasiexperiment in which the researcher exploits a structural break in the data that occurs at a point in time or space, or a discontinuity due to the application of a particular policy or regulation. Researchers then employ a difference-in-difference or regression discontinuity design to more carefully identify the effects of the variables of interest. Studies using these methods have assessed the impacts of air pollution, crime, education, airport noise, flood risk, and the presence of Walmart stores, among other issues (Black 1999; Chay and Greenstone 2005; Linden and Rockoff 2008; Pope 2008a, 2008b; Kousky 2010; Pope and Pope 2012).13
As pointed out by Bajari et al. (2012), the quasi-experimental approach is not feasible in all settings. A source of quasi-randomness that generates exogenous variation in the variables of interest may simply not be available, or may be available only under unrealistic assumptions. In addition, a structural break or discontinuity that does exist may identify impacts only over a narrow range. In the case of open space, for example, a hedonic model estimated on home sales before and after creation of a new park, using data on locations near and far from that park, might provide only limited information about the value of parks more generally in the region.
Kuminoff, Parmeter, and Pope (2010) suggest a “generalized difference-in-difference” approach, in which time dummy variables are interacted with all the covariates. The authors argue that this technique controls for changes in the shape of the equilibrium price function over time and reduces the bias from timevarying omitted variables. They recommend also including spatial fixed effects to control for omitted variables that vary spatially in each time period.
Our preferred specification here is a property fixed-effects model, with sale year dummy variables. This approach, with only repeat-sale properties used in the estimation, will both remove the effects of time-constant omitted variables—at both the property and neighborhood level—and control for temporal shocks. We are thus identifying changes in view or proximity within a given parcel. Our approach has advantages over a pooled model including fixed effects for zip codes, census tracts, or other geographic areas, which will typically only partially correct for omitted spatial variables and not control at all for missing house characteristics. Economists have long recognized the potential of the property fixed-effects approach (Palmquist 1982), but the technique is not often employed because of the data demands; because only properties that sell more than once are used in the analysis, the number of observations often drops significantly from a full sample of home sales. In our setting, however, we have home sales data over a 25-year period, thus we have a substantial number of repeat sales in our dataset. As the note to Table 3 shows, the repeat sale sample includes 130,702 home sales, about 53% of the full sample. A drawback of the approach is that the property fixed effects cannot capture the effects of property renovations. We do find, however, that the distributions of key variables used in the model are quite similar across the two samples, as shown in Table 3. This reduces concerns that repeat sale properties are different on key observables from those that do not sell more than once over our time period.
The property fixed-effects model does not fully identify shifts in the equilibrium price function over time, however, an advantage that Kuminoff, Parmeter, and Pope (2010) claim for the generalized difference-in-difference approach. In theory, one could employ the generalized difference-in-difference technique within the property fixed-effects framework, perhaps getting the best of both worlds. However, in our setting, this approach is not fully implementable. Our natural lands view and proximity variables do not vary continuously over time because the NLCD land cover data is available only for three discrete years. This means that only two time dummies, covering multiple years, could be interacted with the other covariates. In a robustness check in the next section, we show results from such a model, but the inferences from it are limited given the data limitations we face.14
equation [1] gives the form of the property fixed-effects model we estimate:
[1]
where Pit is the inflation-adjusted sale price of property i in year t, Vijt is the share of property i’s view in land cover j in year t, Bijt is the share of property i’s surrounding land (in a 200 m buffer) in land cover j in year t, VDit is the diversity (or mix) of land covers in the view for property i in year t (the number of land cover types as a percentage of all types), Tt is the sale year fixed effect, αi is the individual property fixed effect, and εit is an idiosyncratic error term. The only time-varying property characteristic we have for our data is the age of the house at the time of sale, ageit, and we include this as well.
For the pooled models that we estimate for comparison, we implement a standard hedonic regression on a range of property characteristics, including the view and proximity variables:
[2]
All variables in equation [2] are as previously defined, and Xit is a vector of house characteristics for property i in year t, and θz is the census tract–level fixed effects; thus, we interact spatial and time fixed effects to purge the effects of neighborhood omitted variables that vary over time. We also estimate a version of equation [2] that includes no fixed effects, in order to compare our results more directly to the literature, much of which does not include either spatial or time fixed effects. The house characteristics we include in the model are house size (in square feet), lot size (in acres), number of bedrooms, number of plumbing fixtures, house-style dummy variables, and whether the house is in the 100-
year floodplain. We are also able to include two additional view variables that we cannot include in the property fixed-effects model because they do not change over time: the total size of the view and a dummy for whether the house has a river view. The pooled models thus not only serve as a check on the property fixed-effects model, but also let us examine the effect of total view size and river views on sale prices.
V. RESULTS
Table 6 shows the econometric results for the three versions of the hedonic model. The first and second specifications show the pooled models in which all 246,029 observations of home sales are used. The third specification shows the results of the property fixed-effects model.
The view and proximity variables are in the first several rows of the table. Looking first at the property fixed-effects model, all of the measures of proximity to natural areas appear to increase home prices: the greater the percentage of a 200 m buffer around a property in either farmland, grassy recreational lands, or forest, the higher the sale price; however, all of the impacts are fairly modest. A 10% increase in farmland, recreational grasslands, and forest cover in the buffer increases sale prices by 2%, 1.4%, and 0.6%, respectively.
The effects of views are more mixed. Farmland views are valuable: a 10% increase in the amount of farmland in a home’s viewshed leads to an increase of almost 3% in its price—a larger effect than the proximity variable. But forest views have a negative effect on home prices, and grassy recreational lands have no statistically significant effect. Having a mix of views appears to affect home prices negatively: a 10% increase in the mix of views reduces the price by 1%. This is consistent with findings in other studies that estimate hedonic models in suburban and exurban areas, where the mix includes developed uses along with natural areas (Sander and Polasky 2009; Acharya and Bennett 2001). Other researchers have underscored this difference between suburban and rural areas; diverse views appear more valuable in the latter (Bastian et al. 2002).
We feel that our results about views could be explained by two aspects of our study region. First, the general topography in the study region is relatively flat, as Table 4 shows, and the landscape does not have sharp relief. This may tend to increase the value of open views, such as farmland and grassy lands, and decrease the value of forests, especially the dense hardwood forests of the study region. With less elevation change, the forests may be associated with less expansive views. As we point out in the literature review, some other studies have also found that a forest view can have a negative effect or zero effect on house prices (Paterson and Boyle 2002; Sander and Polasky 2009). Second, and perhaps more important, a substantial amount of agricultural land has been converted to development over the years of our study, whereas forest cover has actually increased slightly. The relative scarcity of farmland visà-vis forestland may explain why increases in farmland views tend to raise prices while increases in forest views tend to decrease them. This effect carries over to the proximity variables as well; while coefficients on all three open space variables are positive, the farmland coefficient is larger than the coefficients on the forest and recreational grassy lands variables.
The results from the pooled models, especially the specification without any fixed effects, differ substantially from the property fixed-effects model. The differences are especially pronounced for the view variables. In the pooled model without fixed effects, all of the view variables (except the diversity measure) have the opposite sign from the property fixed-effects specification. Including the sale year/census tract interaction fixed effects improves the model, but a few key differences from the property fixed-effects model remain. Interestingly, the two pooled models also differ from each other in whether the effects of the river view and total view size are statistically significant. The size of the view positively affects home prices, and a river view negatively affects home prices in both specifications, but neither effect is statistically significant in the model with spatial and sale year interaction effects. These findings call into question some of the extant literature on the value of views, as the studies do not, for the most part, include either spatial or time fixed effects.
Some of the other coefficients also differ across the specifications. House age shows up as a larger determinant of house prices in the property fixed-effects model than in the pooled models: the average house price drops by approximately 6.5% as the house gets one year older. The effect appears much smaller in the pooled models and is statistically insignificant in the model without fixed effects of any kind. Some of the house characteristics are similar across the two pooled models, but not all. And the coefficient on the floodplain dummy variable, negative in both of the pooled models, is statistically significant only in the model without fixed effects.
Accounting for the views of open space as distinct from proximity appears to be important, particularly in the case of surrounding forestland. In a version of the property fixedeffects model in which we omit the view variables and include only the 200 m buffer variables, the estimated coefficients differ from those in the model with both views and proximity. An F-test comparing the simpler model without the view variables to the results with both the view and buffer variables rejects the hypothesis that the buffer variables alone are sufficient (the F-statistic is 110.03). Table 7 shows the coefficients on the buffer variables for the two models; column 2 excludes the view variables and column 3 is the reprinted results from Table 6. Proximity to forests appears to have no statistically significant effect on home prices when the view is excluded, a result that is probably attributable to the variable’s conflation of the positive effect of nearby access with the negative effect of forest views. Once these effects are separated, as in the Table 6 results, proximity to forest becomes valuable (though the effect on house price is relatively small). The farmland and grassy area buffer variables also appear to conflate access and views, though the impact on the estimated coefficients is smaller. Both coefficients are larger in magnitude when the view variables are excluded, but that result is partially attributable to the view and partially to the access. It is worth noting that while there is some correlation between views and the buffer variable, they are not highly correlated, and as Figures 2 and 3 show, they can differ significantly across properties.
VI. ROBUSTNESS CHECKS
As discussed, our land cover variables are available only for three specific years, and as such, our assignment of land cover to house sales in the intervening years will include some measurement error. Thus, to test the robustness of our results, we estimate the property fixed-effects model using alternative assignments of the land cover data, as shown in Table 8 (the assignment used in the baseline model is shown again for comparison). In the first specification, we drop observations for years that are not close to the NLCD years. Although this yields fewer observations for our analysis, the remaining sales will more accurately reflect land cover that exists at the time of sale and thus may do a better job of measuring the effects of changes in land cover on home prices. In the second robustness check, we use only the 1992 and 2006 land cover and assign property sales as shown in the third column of Table 8. We do this because the latter two years of the NLCD, 2001 and 2006, are relatively close together, especially given the nearly decadal gap between the 1992 and 2001 datasets. And, as the summary statistics in Tables 3 and 5 show, land cover shares did not change by as much between 2001 and 2006 as they did between those years and 1992. As Table 8 shows, we omit observations that lie in the middle of these two land cover “windows,” as it is difficult to know which of the two land covers to assign in those cases. This again reduces the number of observations in our econometric analysis over the baseline model. Finally, in a third robustness check, we employ both of these approaches: use of only the 1992 and 2006 NLCD data and a tighter sales window around the two NLCD years. This last approach greatly reduces the sample size.
Table 9 shows the results of these estimations, which indicate that the findings are robust to these alternative uses of the land cover data. The signs on the coefficients on the view and proximity variables are the same as in the baseline model in Table 6 and nearly identical in magnitude and statistical significance. These results give us some confidence that the data capture the true effects of changing land cover on home prices in southern St. Louis County.
Table 10 shows results of another robustness check: the generalized difference-in-difference (GDID) estimation procedure proposed by Kuminoff, Parmeter, and Pope (2010) and implemented by Kuminoff and Pope (2013), discussed above. The technique is designed to allow the marginal effects of the variables of interest, the views and buffers in our case, to vary over time. We define a dummy variable equal to one if the home sale occurred after 2001 and zero otherwise and interact that dummy with each of our covariates. With only three changes over time in the view and buffer variables, this is the finest temporal break that we can analyze—in other words, we cannot divide the data into smaller time intervals and still identify our variables of interest. We choose 2001 as the break point because it ensures that each property will have two land covers and thus a change in views and proximity to natural areas and also divides the data into two roughly equal numbers of years.
The results show that there are differences between the estimated coefficients for the two time periods for most of our variables. Thus there has likely been a shift in the hedonic price function over time. However, with the ability to capture only one break point, due to our data limitations, we feel that interpretations of the results in Table 10 are somewhat limited. The signs and general magnitudes of most of the coefficients do not differ greatly from the baseline model. The most important differences show up for the forest and grassy recreational land buffers. In our baseline model, increases in the amount of forest or grassy lands in a 200 m buffer around a property increase its price, all else being equal. In the GDID model, increases in the forest buffer have no effect in the early period and a small negative effect in the later period, while increases in the grassy lands buffer have a positive effect and then no effect. The signs and magnitudes of the three percentage view variables, which are the focus of our analysis, are quite similar to those of the baseline model.
In an additional robustness check, not included in the paper, we estimated a version of the property fixed-effects model that includes the viewshed variables as measured in separate buffers around the property, increasing in distance. Some studies have found the effect of a view (especially a view of water) on property values to vary by distance (Benson et al. 1998; Bourassa, Hoesli, and Sun 2004; Bin et al. 2008). Similar to the technique used by Cavailhes et al. (2009), we constructed buffers around a parcel and measured the percentage of the view in each buffer that is in each of the land cover classes; we estimated one version of the model using three buffers, and one with six buffers. Our results show some small differences in the impact of the view by distance, but the basic results about the relative values of farmland, forest, and grassy recreational areas remain unchanged from the baseline model. For this reason, and because construction of the buffers has a degree of arbitrariness, we do not present the results here.
We also speculated that the effect of a view may be different depending on the elevation of the property—in other words, that the same percentage view in, say, farmland may be more valuable for a house at a higher elevation than for one at a lower elevation. Thus, in alternative specifications of the model, we (1) interacted the view variables with property elevation and (2) interacted the view variables with a dummy for properties with elevations above the median elevation of all properties.
However, our results were not markedly different from those reported here, and given the limited variability in elevations across properties, we conclude that the model without the elevation interaction variables is preferred.
VII. CONCLUSION
Hedonic studies of open space and natural areas have shown that the capitalization of such amenities in home prices varies greatly depending on the type of natural lands and various attributes of those lands. One aspect that has received limited attention in the literature is the value of natural area views. When using the hedonic approach, which assumes that values are capitalized in property prices, separating this aspect of value from more use-oriented aspects may be important, especially in areas on the urban fringe that are rapidly developing and thus seeing substantial land use changes over time.
With improving GIS techniques, it is possible to measure views more accurately than in the past, and some recent studies have made use of these approaches. They have not, however, coupled these improved methods for identifying a property’s viewshed with newer econometric techniques that address the serious omitted-variable bias problem that is often present in hedonic models. In this study, we combined modern GIS methods with a property fixed-effects model to control for time-constant omitted variables and estimated the value of views of, and proximity to, farmland, forestland, and grassy recreational open spaces in southern St. Louis County, Missouri. With 25 years of home sales, we were able to capture changing land cover over time— namely, the conversion of farmland to development and the increase in forestland that resulted largely from the designation of a major greenway—and assess the value of these alternative land covers as captured in home prices.
Econometrically, we found that simply estimating a pooled model over the 25-year period, even when including spatial-time interaction fixed effects, yielded noticeably different results from the property fixed-effects model, especially with regard to our main variables of interest, the views of natural areas. The property fixed-effects model should purge most of the effects of unobserved time-constant variables, as it identifies changes for the same houses over time. These results appear to confirm the warnings in some recent studies about hedonic models estimated on pooled cross-section/time-series data.
We found that proximity to farmland, forest, and grassy lands, as measured by a 200 m (approximately 1/8 mile) buffer around a property, has a positive and statistically significant effect on home prices. Views have more mixed effects. Farmland views have value: as the percentage of a home’s viewshed comprising farmland views increases, its sale price increases. But as the percentage comprising forest views increases, the sale price declines, and grassy views have no statistically significant effect on home prices. Incorporating both views and proximity is important; a model with only proximity and no views yielded different results, and statistical tests showed that both variables are important to include in the model.
We hypothesize that our view results are due to two important facts about the region we study. The first is topography: relatively low elevations and a lack of landscape relief may make the dense hardwood forests of the region less valuable than the more open, pastoral views associated with farmland. The second reflects the important changes in the landscape over time. A substantial amount of farmland has been converted to development, whereas forest cover has actually increased slightly over time. The relative scarcity of farmland may make it more valuable relative to other open space types.
Our results suggest that changing land use patterns over time can have impacts on property values. This has implications for land use policy in suburban and exurban communities. Controlling the pace and location of development through zoning, special use (e.g., agriculture) districts, purchase and transfer of development rights programs, impact fees, and other mechanisms are common in many areas. In our area of study, St. Louis County has an active public lands acquisition program, as do surrounding jurisdictions, much of it supported by a one-tenth of 1 cent sales tax surcharge. Our findings suggest that consideration of the multiple ways in which the landscape affects property values is important for optimal design of such policies.
Acknowledgments
This research was partially funded by a grant from the National Oceanic and Atmospheric Administration’s Climate Program Office, Climate and Societal Interactions, Sectoral Applications Research Program. We appreciate helpful comments from Nicolai Kuminoff, Joshua Linn, Lucija Muehlenbachs, and two anonymous referees.
Footnotes
The authors are, respectively, research director and senior fellow; fellow; and senior research assistant, Resources for the Future, Washington, D.C.
↵1 Economists are increasingly recognizing that the conditions under which hedonic prices measure actual willingness to pay are very restrictive (Kuminoff and Jarrah 2010; Kuminoff and Pope 2013). We abstract from those concerns here as our focus is purely on the capitalization of views and proximity in home prices.
↵2 LiDAR can be costly, and thus the data are not available in all locations.
↵3 Proper housing market segmentation is a longstanding issue in hedonic studies (Straszheim 1974). While we cannot be sure that our region of study is a homogenous submarket, we feel that it strikes a balance between homogeneity and a sufficient sample size to conduct the analysis.
↵4 These data are available on the St. Louis County GIS Service Center’s website: www.stlouisco.com/OnlineServices/MappingandData.
↵5 These data are available for download from the National Elevation Dataset website: http://ned.usgs.gov/.
↵6 In the pooled model, we include a river view variable, but because neither this variable nor the open water land cover variable change over time, we cannot include any water view variables in the property fixed-effects model.
↵7 The NLCD data are available for download from the NLCD website: www.mrlc.gov/nlcd06_data.php. The data (at a spatial resolution of 30 m) are generated using remotely sensed imagery for the extent of the United States.
↵8 The greenway was officially created in 1975 but has seen publicly acquired lands added to it steadily over time. Buyouts of properties that repeatedly flooded, using Federal Emergency Management Agency funds, form a portion of the greenway protected lands; many of these parcels were added in the mid-1990s. The Great Rivers Greenway park district uses funds from a dedicated sales tax to purchase lands (Kousky and Walls, 2014; also see the Meramec Greenway website: http://meramecgreenway.org/).
↵9 We also estimated a version of the model using a 400 m (1/4 mile) buffer and found results very similar to the results we show below, with a 200 m buffer.
↵10 LiDAR is not comprehensively available for the region. Moreover, while a three-dimensional approach is better in areas where buildings and other structures tend to block views, our study region comprises low-rise residential and commercial buildings and a great deal of open natural areas; thus the combination of GIS land cover and elevation data should be sufficient in our setting.
↵11 Information on the Viewshed tool can be found in the Help Resources for ArcGIS: http://resources.arcgis.com/en/help/main/10.1/index.html#//009z000000v8000000.
↵12 Some hedonic studies have found that potentially developable open space is less valuable than public lands or private open space under a conservation easement (Geoghegan 2002; Irwin 2002), but we are unable to distinguish these lands in our dataset. Most farmland in St. Louis County is not under easement and is thus developable, but forested lands may be protected or unprotected.
↵13 See Parmeter and Pope (2013) for a review.
↵14 While we could estimate a model in which time dummies are interacted with our view and proximity variables, without changes in those variables over time, the estimated coefficients are not capturing what the generalized difference-in-difference technique is designed to capture, changes in the capitalization of those amenities in home prices. This issue with our land cover data also prevents use of Bajari et al.’s (2012) technique, in which the property fixed-effects model is estimated using past sale prices of the same property as covariates.
