Academic journal article The Journal of Real Estate Research

Predicting House Prices with Spatial Dependence: A Comparison of Alternative Methods

Academic journal article The Journal of Real Estate Research

Predicting House Prices with Spatial Dependence: A Comparison of Alternative Methods

Article excerpt


This paper compares alternative methods for taking spatial dependence into account in house price prediction. We select hedonic methods that have been reported in the literature to perform relatively well in terms of ex-sample prediction accuracy. Because differences in performance may be due to differences in data, we compare the methods using a single data set. The estimation methods include simple OLS, a two-stage process incorporating nearest neighbors' residuals in the second stage, geostatistical, and trend surface models. These models take into account submarkets by adding dummy variables or by estimating separate equations for each submarket. Based on data for approximately 13,000 transactions from Louisville, Kentucky, we conclude that a geostatistical model with disaggregated submarket variables performs best.

The hedonic method is increasingly being used for price index construction, mass appraisal, and other purposes. With respect to price index construction, the hedonic method yields indices that are used for multiple purposes, such as tracking housing markets, analysis of real estate bubbles, or investment benchmarking. For mass appraisal, the estimates yielded by hedonic models are used as a basis for the taxation of properties, but in some countries also to assess the value of properties for mortgage underwriting and for performance analyses of real estate portfolios. The method is also well suited to assess the impacts of externalities, such as increased noise levels resulting for instance from the extension of an airport, on house values.

Caution, however, should be exercised when devising hedonic models. Appropriate variables must be selected carefully and measured accurately. And, as with all regression models, errors should be independent from one another, else parameter estimates will be inefficient and confidence intervals will be incorrect. Both theory and empirical research suggests that the independence assumption is unlikely to be valid in a standard ordinary least squares (OLS) context. Basu and Thibodeau (1998), for instance, argue that spatial dependence exists because nearby properties will often have similar structural features (they were often developed at the same time) and also share locational amenities. Consistent with theory, much empirical analysis has concluded that house price residuals are spatially dependent.

Multiple authors have analyzed alternative methods for constructing and estimating hedonic models with spatial dependence in the context of mass appraisal.1 For example, Dubin (1988) compared geostatistical and OLS techniques, as did Basu and Thibodeau (1998). Other efforts include: Can and Megbolugbe (1997), who investigate a spatial lag model; Pace and Gilley (1997), who develop lattice models; Fik, Ling, and Mulligan (2003), who explore a trend surface model; Thibodeau (2003), who considers the importance of spatial disaggregation in a geostatistical model; Valente, Wu, Gelfand, and Sirmans (2005), who develop what they refer to as a spatial process model; and Case, Clapp, Dubin, and Rodriguez (2004), who compare various approaches.

One difficulty in comparing these studies is that they use different data, and their results may be data-dependent. A contribution of the present paper is to compare several methods using the same data set. We use a data set from Louisville, Kentucky, containing approximately 13,000 house sales for 1999. Our approach is similar to that of Case, Clapp, Dubin, and Rodriguez (2004), but we consider their best model (contributed by Case et al. and hereafter referred to as CCDR) in comparison to other methods that have performed well in other studies. We focus specifically on the best models from Thibodeau (2003, henceforth Thibodeau), Fik, Ling, and Mulligan (2003, henceforth FLM), and Bourassa, Cantoni, and Hoesli (2007, henceforth BCH). Another contribution of the present paper is to perform each type of analysis using 100 random samples of the data for estimation purposes to insure that the results are not specific to a particular sample. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.