1 Introduction

The importance of location in determining housing prices is widely recognized. Controlling for location and the spatial structure of markets is thus essential to explaining house price differentials and deriving accurate coefficient estimates in hedonic house price models. However, spatial issues have not always been given adequate attention in hedonic applications (Bowen et al. 2001; Páez et al. 2001). Key econometric issues include spatial dependence and spatial heterogeneity (Anselin 1988, 1990). While housing markets are likely to be characterized by both, we focus specifically on the issue of spatial heterogeneity as it has received less attention in the literature.

Within the housing market context, the issue of spatial heterogeneity centers on whether the marginal prices of housing attributes are constant throughout a metropolitan area or vary over space. Most empirical models have conceptualized a metropolitan area as a single unified market, and while neighborhood effects may be incorporated into regression models through varying intercepts, the coefficients of structural attributes are held constant throughout the market (Orford 1999). If spatial heterogeneity exists, stationary coefficient models will produce parameters that are in essence an “average” value of the parameter over all locations. A failure to incorporate spatial heterogeneity will result in biased coefficients and a loss of explanatory power and may obscure important dynamics relating to the operation of housing markets.

This paper seeks to add to our understanding of the role that spatial heterogeneity plays in housing markets by comparing two methods that allow spatially varying parameters in an analysis of the Tucson, Arizona housing market. The spatial expansion method, pioneered by Cassetti (1972), allows parameters to vary over space in a traditional OLS regression framework by interacting house characteristics with locational information. Geographically weighted regression (GWR) in essence specifies a separate regression model at every observation point, thus enabling unique coefficients to be estimated at each location (Brunsdon et al. 1996).

The remainder of the paper is organized as follows. We begin with a review of the pertinent literature followed by an overview of the Tucson housing market. Next, the data and methodology employed in the study are detailed. We then compare the results of the models and discuss the spatial patterns observed in the data. In the final section we draw conclusions and suggest avenues for future research.

2 Spatial variation in housing attribute prices

There is good reason to expect that the price of housing attributes will exhibit spatial heterogeneity within large housing markets due to localized supply and demand imbalances (Dubin et al. 1987; Goodman 1981, 1998; Michaels and Smith 1990; Schnare and Struyk 1976). The supply of specific housing characteristics often exhibits strong spatial patterns within a metropolitan area. For example, near the center of a metropolitan area homes tend to be older and frequently lack features such as large garages, in comparison to those located at the suburban fringe. Housing is a unique good due to its fixed location and durability, and the characteristics of the housing stock may be difficult to change in response to changing demand. Thus the supply of certain types of housing and neighborhood characteristics may be highly inelastic, particularly over short periods of time (Schare and Struyk 1976).

Demand by households for specific structural and locational attributes is known to vary based on socioeconomic status, household status, race and ethnicity (Quigley 1985), as well as the location of household activities such as the workplace. Demand for some attributes, such as a high quality school district or for a house with a minimum number of bedrooms, may also be highly inelastic (Schare and Struyk 1976). Thus all housing within a large metropolitan area will not be substitutable. In addition, access to information and the actions of market participants such as realtors, lenders and appraisers may constrain households from participating in all segments of a large market (Michaels and Smith 1990).

Changes in household preferences for housing characteristics and locational attributes, as well as the characteristics of neighborhoods themselves, may result in spatial mismatches between supply and demand as the housing stock available within a particular geographic area may not match current demand. Greater competition for those housing attributes that are in high demand, yet locally scarce, should result in higher marginal prices. Thus one would expect supply and demand imbalances to result in spatial heterogeneity within large housing markets.

One approach to dealing with spatial heterogeneity is to delineate the housing market into distinct geographic areas or submarkets and to estimate separate hedonic prices schedules for each (Schnare and Struyk 1976; Goodman 1981, 1998; Michaels and Smith 1990; Bourassa et al. 2003). However, housing submarkets are often problematic to define in practice, and this approach makes it difficult to generalize about the dynamics of the broader housing market or urban area. The focus on housing submarkets also posits that spatial heterogeneity is a discrete phenomenon and does not allow attribute prices to vary in a continuous manner over space.

A number of housing market studies have used variants of the expansion method pioneered by Cassetti (1972). This method recognizes that functional relationships may not be constant but vary over space, and explicitly allows parameter estimates to “drift” based on their spatial context (Jones and Cassetti 1992). This method is operationalized by “expanding” the parameters of stationary coefficient models.

Can (1992) utilized the expansion method to allow the parameter estimates of housing attributes to vary with neighborhood quality. The neighborhood interaction terms were significant for some variables, but a traditional specification with a spatial lag term performed almost as well the expansion model. Limitations of this study include the use of census tracts as a proxy for neighborhoods or submarkets, as well as its focus on demand as the sole driver of spatial heterogeneity, as similar attribute prices are estimated for neighborhoods with similar “quality” scores regardless of the supply of specific housing characteristics in each. Thériault et al. (2003) improve upon this approach with an expansion model that allows housing attributes to vary based on both accessibility and neighborhood attributes.

In a study of Tucson, Arizona, Fik et al. (2003) specified a “fully interactive” expansion model employing a second order polynomial expansion of housing attributes, the properties’ {x, y} coordinates, and dummy variables representing submarkets. The interactions between the absolute-location variables and structural attributes thus allowed the coefficients to vary over space. This model outperformed stationary model specifications and its explanatory power was far superior. A number of the spatial interactive terms were significant, indicating the presence of spatial heterogeneity in the prices of these attributes. A limitation of this study is that it relied on only three housing-attribute variables, thus making it difficult to assess whether observed spatial variation was the result of intrinsic parameter variation or due to the effects of omitted variables.

Incorporating absolute location into hedonic models in the form of a polynomial expansion of parcel coordinates Footnote 1 is appealing because it is difficult if not impossible to identify and accurately specify all locational influences that affect housing prices (Orford 1999). While a parcel’s coordinates are not a direct determinant of housing prices, they may serve a useful role in controlling for the influence of location on prices. A number of other authors have experimented with this approach in recent years (Clapp 2001; Pavlov 2000).

The geographically weighted regression is a local modeling approach that explicitly allows parameter estimates to vary over space (Brunsdon et al. 1996; Fotheringham and Brunsdon 1999; Fotheringham et al. 2002). Rather than specifying a single model to characterize the entire housing market, GWR estimates a separate model for each sale point and weights observations by their distance to this point, thus allowing unique marginal-price estimates at each location. This method is appealing because it mimics to some extent the “sales comparison” approach to valuation used by appraisers in that only sales within close proximity to the subject property are considered, and price adjustments are made based on differences in characteristics within this subset of properties.

The geographically weighted regression has not been widely utilized in the housing market context. The only application the authors are aware of is a recent study of the Toronto, Canada housing market by Farber and Yates (2006). In this study, GWR outperformed several alternative approaches and was able to explain more than 90% of the variation in housing prices and provided evidence for spatial heterogeneity in several housing attributes. Páez (2005) compares GWR and the expansion method in a simulation study, finding that both approaches are able to provide a reasonable representation of the spatial patterns inherent in the simulated data. However, no studies have provided a direct comparison between the two methods in the analysis of house prices.

3 Tucson housing market

Tucson is a mid-sized metropolitan area situated within the Sonoran Desert in Southern Arizona and is ringed by mountain ranges. Like many western cities, much of the land surrounding the urbanized area is publicly owned. The City of Tucson has been one of the nation’s fastest growing cities over the past several decades, with its population expanding from 330,000 in 1980 to 487,000 in 2000, while the greater metropolitan area (Pima County) saw its population swell from 530,000 to 850,000 during the same period. The population is diverse, with Hispanics accounting for nearly 30% of the total population (Mulligan et al. 2002).

Like most southwestern cities, Tucson’s development pattern is relatively dispersed and the urbanized area encompasses some 500 square miles. The metro’s transportation system is designed largely for private automobiles. While some central neighborhoods are densely populated, most residents live in suburban style settings. Tucson’s employment, too, is dispersed with its weak CBD functioning primarily as a government services center. Large area employers include The University of Arizona, Davis-Monthan Air Force Base, and Raytheon.

Tucson’s housing market is generally differentiated from north to south, with the highest priced housing found in the foothills of the Santa Catalina Mountains located to the northeast of Central Tucson (see Fig. 1). The northern half of the city typically contains newer, larger, more expensive housing units than those found in the southern and central areas, although new development is occurring at the urban fringe in all directions. The eastern and western portions of the market generally exhibit “average” house characteristics.

Fig. 1
figure 1

Study area

Full size image

Following a prolonged slump during the early 1990s, the Tucson housing market was generally stable during the mid-1990s. By the end of the decade prices began to appreciate again in real terms. During recent years house price appreciation has accelerated rapidly with median price increases of 13.4 and 30.6% in 2004 and 2005, respectively (National Association of Realtors 2006).

4 Data and methods

4.1 Data

Home sales data for the year 2000 were obtained through the Pima County Assessor’s Office. For tax valuation purposes, the assessor compiles an annual real estate sales database based on sales information recorded with the Pima County Recorders Office. All real estate sales within the county are required to be recorded for tax collection purposes; hence these data should contain all residential sales that occurred within the county during 2000. The master sales database contained 15,986 records.

The analytical database was constructed as follows: We first selected sales records representing detached single-family homes. These sales were then matched to a separate assessment “structural file” containing information on the characteristics of each dwelling. Next, these data were matched to a property “parcel” GIS coverage maintained by Pima County and each property was attributed with the coordinates of its parcel centroid. This resulted in a total of 14,154 matched records. A large number of records were missing lot sizes in the structural file, as the assessor no longer routinely tracks this information. In these cases we attributed the records with the area of the property’s parcel polygon from the GIS coverage. Footnote 2

We then selected a subset of records corresponding to the contiguous Tucson urbanized area, as much of the county is rural. Our data include all or portions of the municipalities of Tucson, South Tucson, Marana, and Oro Valley, as well as portions of unincorporated Pima County. Records that did not represent “arms length” transactions, were missing data elements, or were outliers in terms of price or structural attributes were also eliminated. Our final data set consists of 11,732 records, which represents the vast majority of valid single-family home sales recorded during the year 2000.

4.2 Models

Four models are estimated in the empirical analysis using 90% of the sales, or 10,569 observations. A random sample of 1,163 sales was generated and held out in order to test the predictive accuracy of the various approaches. All models use the natural log of sale price as the dependent variable. While there is no consensus in the literature regarding the appropriate functional form of hedonic house price models (Freeman 2003), experimentation with the Tucson housing data revealed that the semi-log form performed well in comparison to other common functional forms. Thus the coefficients can be interpreted as the percentage change in sale price attributable to a unit change in an independent variable.

Preliminary regressions established 13 property characteristics to be significant determinants of housing prices in Tucson. All variables, with the exception of LOTSIZE, as noted above, were taken directly from the assessor’s structural file. Variable descriptions are provided in Table 1.

Table 1 Independent variable descriptions
Full size table

We chose to reduce the number of variables entering our models through a principal components analysis because GWR is computationally intensive and the expansion method becomes intractable with a large number of explanatory variables. This is preferred to simply discarding variables as we wish to mitigate the potential for bias resulting from omitted variables. Five variables enter the models individually: SQFT, LOTSIZE, STORY, PRE1940, and CLASS. The remaining eight were reduced to two factors. Thus a total of seven variables are used to represent property characteristics (Table 2). Factor one indicates newer dwellings with modern features like enclosed garages and refrigerated air conditioning. Factor two represents homes with outdoor amenities such as pools or patios and a high average room size. Full details of the factor analysis are provided in the Appendix. Correlations between the seven independent variables are relatively low, as only corr(CLASS, SQFT) and corr(FACTOR2, SQFT) exceed 0.5.

Table 2 Descriptive statistics
Full size table

Our first model is a “global” specification that estimates a single set of parameters for the entire study area. Model 1 includes the seven housing-attribute variables detailed in Table 2, as well as nine variables representing absolute location in the form of a third degree polynomial expansion of the parcel coordinates. The raw coordinates were first transformed to deviations from the mean x and y values of all sales within the study area.

Model 2 is the spatial expansion. The seven housing-attribute variables are interacted with the nine absolute-location variables, thus allowing the marginal price of the housing attributes to vary over space. This results in 63 new independent variables, Footnote 3 in addition to the 16 included in Model 1.

While spatial dependence is not the focus of this paper, we include a spatially lagged dependent variable in Model 3 in order to mitigate potential bias resulting from the omission of this variable. The lag term is calculated as the distance weighted average price of each observation’s 15 nearest neighbors. Although maximum likelihood estimators have more desirable properties in the presence of spatial dependence, OLS has been employed under certain circumstances involving large data sets (Can and Mogbolugbe 1987; Farber and Yates 2006). We follow this approach here. Thus the only difference between Model 2 and Model 3 is the inclusion of the spatial lag term.

Model 4 is the GWR model, which was estimated using the GWR 3.0 software package. This model includes the same seven housing attributes as independent variables that were used in the prior models.

As outlined in Fotheringham et al. (2000), the standard hedonic model formulation specifies sale price as a function of a set of housing characteristics as follows:

$$y_i = \alpha_0 + {\sum\limits_k {\alpha_k} }x_ik + \varepsilon_i,$$
(1)

where x ik represents the ith observation of the kth independent variable. The GWR specification is similar, except that unique coefficients are estimated at each observation point:

$$y_i = \alpha_{0} + {\sum\limits_k {\alpha_{k} } }(u_{i},v_{i})x_{{ik}} + \varepsilon _{i},$$
(2)

where α k (u i , v i ) represents the regression coefficient for variable k at regression point i. In matrix notation, the parameters of a GWR model are estimated as follows:

$$\alpha(u_{i},v_{i}) = (X^{\rm T}W(u_{i},v_{i})X)^{{ - 1}}X^{\rm T}W(u_{i},v_{i})y,$$
(3)

where W(u i , v i ) is a spatial weighting matrix. We utilize a Gaussian function where d represents the Euclidian distance between the regression point and observation point, and h represents the bandwidth as follows:

$$W_{i} (u_{i},v_{i}) = \exp (- d/h)^{2}.$$
(4)

The results of GWR are sensitive to the choice of bandwidth, as weighting procedures that specify a wide bandwidth and allow for only minimal distance decay will produce results that are similar to a global model. Conversely, if the bandwidth is narrow only points in close proximity will be considered, which will lead to high variances in the estimators (Fotheringham et al. 2000).

We chose to use an adaptive spatial kernel that allows the bandwidth to vary based on the density of home sales around each regression point, thus encapsulating a smaller area where data are rich and a larger area where data are sparse. This ensures that an equal number of observations will receive a non-zero weighting at all regression points. The cross-validation method, which optimizes the choice through an iterative process based on a least squares criterion, was used to select the bandwidth. The resulting bandwidth is 544 observations.

5 Results

5.1 Model summaries

The results for Model 1, in which the marginal-price estimates are held constant throughout the study area, are depicted in Table 3. The global model, utilizing x, y coordinates to control for location, does a reasonable job of explaining variation in Tucson house prices as indicated by an adjusted R-squared of 0.88 and standard error of 0.164. All variables are significant at the 0.01 level, have the expected signs, and are of plausible magnitudes.

Table 3 Model 1 results: global
Full size table

The results for Model 2, the spatial expansion, are shown in Table 4. We report only coefficients that are significant at the 0.05 level, excluding interaction terms involving the two factors. The addition of location–attribute interaction terms results in a modest improvement in explanatory power as the adjusted R-squared increases to 0.89 while the standard error drops to 0.156. The coefficients for all seven “base” housing-attribute variables maintain the same signs as in Model 1 and only PRE1940 is no longer significant. A total of 33 location–attribute interaction terms are significant at the 0.05 level, indicating that the marginal prices of these attributes vary with locational context.

Table 4 Model 2 results: spatial expansion
Full size table

The addition of the spatial lag term in Model 3 results in an improvement in explanatory power as the adjusted R-squared increases from 0.894 to 0.91, while the standard error declines from 0.156 to 0.144 (Table 5). The spatial lag term is positive and significant, suggesting that housing prices are strongly influenced by the prices of nearby homes, or alternatively, features or externalities that homes in close proximity share have not been accounted for in our model. The estimates and significance levels of the housing attributes and location variables generally decline in comparison to Model 2. For example, several of the location variables become insignificant and the marginal-price estimate for LOTSIZE declines markedly.

Table 5 Model 3 results: spatial expansion with spatial lag term
Full size table

While the spatial lag term is clearly capturing important externality effects, the results still provide strong evidence for spatial heterogeneity, as 29 location–attribute interaction terms are significant at the 0.05 level or better. Location–attribute interaction terms are significant for all seven housing characteristics. For example, five CLASS interaction terms are significant, indicating a complex spatial pattern for this variable. Conversely, only two SQFT interaction terms are significant.

Table 6 contains the results of the GWR model. The adjusted R-squared Footnote 4 improves to over 0.92 and the standard error declines to 0.136. This improvement reflects the importance of localized spatial influences within the Tucson housing market. The GWR parameter estimates, which vary at each of the 10,569 observation points, are described by their median, minimum, and maximum value as well as interquartile range. In most cases, the parameter estimates from the global model are encapsulated within, or are very close to, the interquartile range of the GWR model. We note that pre-1940 vintage housing is not present throughout much of the Tucson area, thus there is no variability at the majority of observation points which results in an interquartile range of zero.

Table 6 Model 4 results: geographically weighted regression
Full size table

The parameter estimates for the seven independent variables vary widely over space. The P value from a Monte Carlo significance test indicates that the spatial variation in all seven is significant at the 0.01 level or higher. This provides strong evidence that the marginal prices of these housing characteristics are not constant, but vary over space within the greater Tucson area.

The interquartile ranges of the GWR estimates are of plausible magnitudes, however, the minimum and maximum values are extreme or counter intuitive in some cases. For example, the estimates for STORY range from  − 0.56 to 0.33, which implies that, all else equal, a multi-story house Footnote 5 sells for 56% less than a single-story home at one regression point and 33% more at another. The PRE1940 estimates range from  − 0.549 to 0.566. Negative values for the CLASS and LOTSIZE coefficients, too, are counterintuitive, as they indicate that homes of low structural quality sell for more than those with higher structural quality at some locations and that an additional square foot of lot reduces price in some areas. The negative estimates, however, are statistically significant within only a very small portion of the study area.

5.2 Spatial patterns

One advantage of GWR is that the spatial patterns inherent in the parameter estimates can be easily mapped and visualized. The dwelling size (SQFT) estimates are shown in Fig. 2. As expected, the estimates are positive and significant throughout the Tucson area and exhibit relatively smooth spatial trends. The highest marginal-price estimates are found within Central Tucson, where homes tend to be among the market’s smallest. The estimates also tend to be high in the exclusive Catalina Foothills area where demand for large homes is great. The marginal price of an additional square foot of living space is generally low near the eastern and southern peripheries where homes are typically larger. In the northwest, the SQFT estimates exhibit a more localized pattern.

Fig. 2
figure 2

GWR SQFT parameter estimates

Full size image

The spatial pattern of the LOTSIZE estimates is distinctly different from that of SQFT (Fig. 3). The marginal-price estimates are generally positive, though not significant at the 0.05 level, throughout much of the central and southern portions of the Tucson area. The estimates tend to be highest in the northwest and southeast areas. Surprisingly, a large pocket of negative estimates is evident within a portion of the Catalina Foothills area, which indicates a possible misspecification problem. One might expect lower marginal prices in this area because lots are generally among the market’s largest, but it is difficult to explain why the marginal price of an additional square foot of lot would be negative. This is most likely attributable to characteristics that larger lot homes in this area have in common but are not included in the model.

Fig. 3
figure 3

GWR LOTSIZE parameter estimates

Full size image

The estimates for several variables reflect complex, localized spatial patterns. For instance, while the coefficient for the STORY variable is negative and significant in the global model, the GWR parameter estimates indicate that ceteris paribus, multi-story homes sell for more than single-story homes within a large area east of Central Tucson (Fig. 4). High negative estimates for STORY are found just to the south of this area. Thus the GWR results suggest that the value of a multi-story home is dependent upon locational context.

Fig. 4
figure 4

GWR STORY parameter estimates

Full size image

5.3 Predictive accuracy

In order to assess the predictive accuracy of the GWR and the spatial expansion approaches, predicted prices were estimated for the 10% of sales withheld from the initial models. In GWR, this is accomplished by estimating new models at the 1,163 holdback sample locations using only the data from the original 10,569 observations. The holdback sample analysis indicates that allowing housing attribute prices to vary with absolute location strengthens house price prediction accuracy.

The results of this analysis, presented in Table 7, show a progressive improvement in prediction accuracy from Model 1 through Model 4. In the global model, about 57% of predictions are within 10%, and 83% within 20%, of actual sale price. The GWR model performs better on this metric, as nearly 65% of predicted prices fall within 10%, and 88% within 20%, of their actual value. As a measure of cross-validation, we report the R-squared of the regression of actual sale price and predicted sale price for each model. Models 3 and 4 perform comparably on this metric.

Table 7 Out of sample predictive accuracy: percent of predicted prices within specified range of actual price and R-squared between actual and predicted price
Full size table

6 Discussion

Our results show that both the spatial expansion and GWR methods of incorporating spatial heterogeneity result in an improvement in explanatory power and predictive accuracy over the stationary-coefficient model. The results also provide strong evidence for the presence of spatial heterogeneity within the Tucson market, indicating that the marginal prices of key housing attributes are not constant but vary with locational context.

The geographically weighted regression outperforms the spatial expansion method in terms of explanatory power and predictive accuracy. This difference is narrowed to some degree with the addition of the spatial lag term in the expansion specification. While the spatial expansion model is capable of picking up broad trends in the spatial structure of the housing parameters, GWR appears to be better able to represent the complex spatial patterns inherent in the Tucson data. This suggests that spatial heterogeneity may be due to discrete, localized influences, as well as those operating in a broad, continuous manner over space. Our results suggest that when explanatory power and predictive accuracy are the primary objectives, GWR is the superior approach.

A comparison of the absolute value of the GWR and spatial expansion residuals (Model 3) is depicted in Fig. 5. Dark points represent locations where the GWR prediction was closer to the actual sale price, while light points indicate locations where the spatial expansion model was more accurate. Locations with similar absolute prediction errors (within 0.05) are suppressed. In general, GWR is more accurate within Central Tucson, where housing tends to be dense and heterogeneous, as well as near the periphery of the study area. The spatial expansion predictions tend to be more accurate within the area immediately surrounding the central core.

Fig. 5
figure 5

Comparison of absolute prediction error for the GWR and spatial expansion models

Full size image

Although GWR outperforms the spatial expansion in terms of explanatory power and predictive accuracy, the expansion approach has greater flexibility and may be more suitable in other situations. For instance, when the primary objective is to explain the underlying determinants of residential housing prices, the expansion method may be superior due to its ability to accommodate a larger number of variables and interactions. The expansion framework is also more conducive to hypothesis testing than is GWR.

Do the complex spatial trends depicted by GWR reflect true spatial variation in the price of the seven housing-attribute variables? As argued earlier, there is good reason to expect variation in marginal prices within a large, dispersed market such as Tucson due to localized supply and demand dynamics, and the spatial pattern of the SQFT estimates appear to be plausible given our knowledge of the market. It is also plausible that the value associated with a multi-story house may vary with locational context. For example, in a neighborhood with desirable views and widely spaced homes, a multi-story home may be perceived as amenity. Conversely, a multi-story home may be perceived as a disamenity in an area with densely packed homes surrounded by less desirable land uses. If this is indeed the case, the stationary coefficient model has obscured this important relationship.

The negative estimates for LOTSIZE and CLASS are difficult to rationalize and are almost certainly due to some form of misspecification. Farber and Yates (2006) found a similar result for several variables in their study of Toronto. Omitted variables in particular likely influence our results. While both the GWR and spatial expansion models incorporate absolute measures of location, neither includes direct measures of the neighborhood, environmental, and accessibility attributes that underpin the value of “location”. If these influences are not adequately controlled for through absolute location and are spatially correlated with variables included in the models, they may contribute to the observed spatial heterogeneity. This holds true for structural attributes as well. Consequently, GWR may provide a means to identify such misspecification problems.

A potential limitation of both the spatial expansion and GWR models is that they in essence impose a continuous pattern on the spatial structure of the market. However, it is widely recognized that some locational attributes that might lead to spatial heterogeneity are discrete in nature. For example, school districts are known to play an important role in the determination of housing prices. Therefore, one would expect price shifts to be abrupt when moving across the boundary from a high-quality school district to a lower-quality district. If this is the case, it may be more appropriate to delineate housing submarkets.

The expansion method can incorporate discrete effects by including submarket or neighborhood interaction terms (Fik et al. 2003). However, this may not be practical in some situations as it would result in an even greater number of independent variables and require a priori knowledge of submarket boundaries. While it would be difficult to incorporate discrete influences in a GWR model, GWR may be a useful tool to determine whether segmentation is warranted as well as an aid in establishing meaningful submarkets boundaries.

7 Conclusions

Our comparison of the GWR and spatial expansion methods provides strong evidence that the marginal prices of key housing attributes are not constant throughout the Tucson market area, but vary with locational context. The parameter estimates for all seven housing attributes exhibited significant spatial variation in both models. We believe our results reflect both true variation in the marginal prices of these attributes due to localized supply and demand dynamics as well as potential misspecification problems such as omitted variables.

While both the spatial expansion and GWR approaches improve upon the results of the stationary coefficient model, GWR outperformed the spatial expansion specification in terms of explanatory power and predictive accuracy. The GWR results indicate that the spatial pattern of coefficient estimates is more complex than can be accounted for by a spatial expansion employing a third degree polynomial expansion of the homes’ x, y coordinates.

Regardless of whether the results are indicative of true parameter variation or misspecification, they highlight the complex spatial structure of housing markets and the need to explicitly address spatial heterogeneity in housing market models. A failure to do so may result in a loss of explanatory power, lead to erroneous conclusions, and obscure important housing market dynamics. GWR in particular provides a means to visualize the spatial structure of housing markets. Either method may be a viable alternative in situations where price prediction is the primary concern and locational information is difficult to obtain or when knowledge of local submarkets is unavailable.

Further research needs to be done to uncover the exact nature of spatial heterogeneity in housing markets, specifically whether it is a discrete or continuous phenomena or a combination of both. The counterintuitive GWR estimates found at some locations deserve further attention as well. A direct comparison of the spatial patterns of marginal-price estimates generated by each approach would be instructive. Additional cross-sectional studies would be useful in order to assess the stability of these results over time and space. Finally, while we have focused specifically on spatial heterogeneity in this study, the significant result for our spatial lag variable in Model 3 suggests that future research should strive to incorporate both spatial dependence and spatial heterogeneity in a formal spatial econometric setting.