JOURNAL OF REAL ESTATE RESEARCH Appraisal Using R. Kelley .

JOURNAL OF REAL ESTATE RESEARCHAppraisal UsingGeneralized AdditiveModelsR. Kelley Pace*Abstract. Many of the results from real estate empirical studies depend upon using acorrect functional form for their validity. Unfortunately, common parametric statisticaltools cannot easily control for the possibility of misspecification. Recently, semiparametricestimators such as generalized additive models (GAMs) have arisen which canautomatically control for additive (in price) or multiplicative (in ln(price)) nonlinearrelations among the independent and dependent variables. As the paper shows, GAMscan empirically outperform naive parametric and polynomial models in exsamplepredictive behavior. Moreover, GAMs have well-developed statistical properties and cansuggest useful transformations in parametric settings.IntroductionMany functional forms of the variables and parameters lead to pricing functions thatagree with the information amassed by the substantial theoretical and empirical workin hedonic pricing and mass assessment.1 Consequently, the exact specification toadopt remains one of the central uncertainties of empirical work, especially since the‘‘wrong’’ functional form leads to all sorts of disastrous consequences for traditionalestimators. In response to this problem, many nonparametric estimators have beenproposed which adapt to the data and do not require an a priori functionalspecification. However, nonparametric estimator performance typically declines as thedimensionality of the problem increases.2 As a compromise, various semiparametricestimators have arisen that possess the adaptive traits of nonparametric regressionwhile retaining the estimation efficiency of parametric estimators. Projection pursuit(Friedman and Stuetzle, 1981), neural nets, alternating conditional expectations(Brieman and Friedman, 1985), additivity and variance stabilization (Tibshirani,1988), regression trees (Brieman, Friedman, Olshen and Stone, 1984), P (Brieman,1991), multivariate adaptive regression splines (Friedman, 1991) and sliced inverseregression (Duan and Li, 1991) estimators represent some of the efforts in thisdirection.3In real estate, various applications of nonparametric and semiparametric regressionhave appeared from time to time. For example, analysis of pairs of houses matchedon all but two or fewer characteristics via graphical methods would qualify asnonparametric estimation. Isakson (1986) used a form of nearest neighbor*Department of Finance, Louisiana State University, Baton Rouge, LA 70803 or kelleypace@compuserve.com.77

78JOURNAL OF REAL ESTATE RESEARCHnonparametric estimation.4 Sunderman, Birch, Cannaday and Hamilton (1990)explored the relation between assessed value and market price using a bivariate splineregression estimator. Meese and Wallace (1991) and Pace (1993b) appliednonparametric multivariate regression estimators to real estate data. Meese andWallace employed locally weighted regression (see Cleveland and Devlin, 1988) toform hedonic price indices. They conducted diagnostics on the sample fit to documentits generally good performance. Using two hedonic pricing examples, Pace (1993)demonstrated the kernel nonparametric regression estimator could out-performordinary least squares (OLS) in ex-sample prediction. Pace (1995) applies thesemiparametric index estimator (y5g(Xb)1«) of Hardle and Stoker (1989) and Powell,Stock and Stoker (1989) to real estate data and showed it could compete with OLSand the kernel regression estimator.Coulson (1992) used a model incorporating some parametric components and abivariate spline estimator for the nonparametric component. Anglin and Gencay(1993) used a model incorporating parametric components and a multivariate kernelestimator for the nonparametric component to investigate the hedonic pricingfunctional form. Their semiparametric estimates clearly outperformed their bestparametric ones.Both the Coulson and the Anglin and Gencay estimators fall in the category ofGeneralized Additive Models (GAMs). GAMs constitute perhaps the simplest classof semiparametric estimators in terms of computation and visualization. Essentially,GAMs in Equation (1) estimate the dependent variable as a sum of functions of theindependent lly, GAMs include linear models.y5X1b11X2b21.1Xkbk1«.GAMs can extend their range in the same way linear models extend theirs throughtransformations and functions of the individual �5(X2X3)1.1ƒk(Xk)1«.Effectively, Coulson used a model involving the first two terms while Anglin andGencay used a model involving the first and the fifth terms. Specifically, Anglin andGencay used a kernel estimator involving six dimensions or characteristics. Naturally,the dependent variable, y, could represent a transformation of some other variable(e.g., y5ln(z) or y5z1 / 2), which means GAM could also include multiplicativemodeling of z.Graphs of the estimated transformation ƒ̃i(Xi) versus Xi constitute one of the mainproducts of the GAM estimator. These may have interest in their own right or canserve as guides to transforming variables in the ordinary linear model. Alternatively,VOLUME 15, NUMBERS 1 / 2, 1998

APPRAISAL USING GENERALIZED ADDITIVE MODELS79these estimated transformations allow one to check on the linearity of and y in aposited linear model.This article examines the computation of the GAM estimator in section two andapplies the estimator in section three. Specifically, beginning with a typicalsemilogarithmic specification using 442 observations from the Memphis MultipleListing Service (MLS), the GAM estimator suggests transformations that lead to alinear double logarithmic model. For comparison, section three includes thesemilogarithmic and double logarithmic GAM and polynomial regression models.Section three also includes a cross-validation prediction experiment that shows thesuperiority of the GAM and the retransformed linear model to the originalsemiparametric and polynomial regression models. Section four summarizes the keyresults.Computation of Generalized Additive ModelsAs mentioned previously, the GAM estimator is one of the simplest semiparametricestimators to compute and visualize. One minimizes some loss function, typicallysquared error, through the choice of functions as opposed to individual parameters.min ẽ9ẽ where ẽ5(y2ƒ 1(X1)1ƒ 2(X2)1.1ƒ k(Xk)).(2)Hastie and Tibshirani (1990) extensively discussed the use of the backfitting algorithmthat iteratively minimizes Equation (2) an estimated function at a time. Let i(i51,2,.k) represent the individual estimated functions (ƒ̃i (z)) and j ( j51,2,.m)represent the iteration. For each iteration j one minimizes Equation (2) with respectto each of the estimated functions ƒ̃i(z). One continues the iterations until convergence.Hastie and Tibshirani prove this algorithm will converge to an unique solutionindependent of the starting values for symmetric smoothing functions such assmoothing or regression splines. Interestingly, if for all i ƒ̃i(z)5Xibi, the backfittingalgorithm yields, albeit slowly, the least-squares solution for a squared-error lossfunction.One can employ a variety of methods to estimate the functions ƒ̃i(z). For example, onecan employ the kernel method, locally weighted smoothing, smoothing splines,regression splines, nearest neighbor and polynomials.5The advantage to GAM, as opposed to purely nonparametric methods, lies in thereduction of the problem of estimating nonparametric surfaces to a sequence ofbivariate smoothing problems. These allow (1) visual inspection of the smooth; and(2) the estimates converge as rapidly as parametric estimators.In the following estimates, I used smoothing splines as the bivariate or scatterplotsmoother. Smoothing splines minimize Equation (3):min [(y2ƒ i (Xi))9(y2ƒ i (Xi))1lE (ƒ 0(t)) dt]2i(3)

80JOURNAL OF REAL ESTATE RESEARCHwhere l represents a roughness penalty. If ƒ̃i (Xi) had a linear form, the secondderivative of a linear function, ƒ 0i (Xi), would be 0. Alternatively, if ƒ i (Xi) rapidlychanged with Xi, the second derivative, ƒ̃0i (Xi), would have a large magnitude. If lequals 0, the smoothing spline would cause ƒ̃i (Xi) to match every point in y, resultingin no error. If l equals infinity, the heavy penalty on roughness would cause ƒ̃i (Xi) toreturn a linear fit, resulting in the least squares regression line.Naturally, the parameter l greatly affects the smoothing splines behavior. A smallvalue of l means ƒ̃ (Xi) is very flexible in the same way a high order polynomial isflexible. As most individuals do not have much prior information concerning l, Buja,Hastie and Tibshirani (1989) provided a way of measuring the equivalent degrees-offreedom sacrificed by making ƒ̃i (Xi) very flexible. This greatly reduces the difficultyof smoothing parameter selection. The equivalence between l and degrees-of-freedommakes it possible to perform approximate inference for GAM.As a final note concerning l, by appropriate selection of li5g(Xi) one could maximizethe linearity of ƒ i (li). This could greatly reduce the value of ƒ i0 (li) which reduces thesensitivity of the overall solution to an inappropriate choice of l. We shall use thistechnique latter in the actual estimation.Polynomials constitute the traditional way of modeling functions of Xi ( ƒ̃i (Xi)5a1bXi1cX 2)i1.1pXip21) in linear models. A series of polynomials leads to a modellinear-in-the-parameters which least squares can fit directly. The difference betweennonparametric smoothers such as smoothing splines or the kernel method andpolynomial regression lies in the local nature of the nonparametric estimator fits versusthe global nature of the polynomial regression estimator fit. If the (y, Xi) plot linearlyover part of Xi but have a curved portion over another part of Xi, nonparametricestimators can follow this (even with two degrees-of-freedom). A second degreepolynomial will have some curvature over all of Xi. Hence, polynomials of limiteddegree do react to nonlinearities. Their global fit means any nonlinearity polynomialsdetect in ƒ̃i (Xi) for some values of Xi will be spread over all Xi.Finally, GAM are a generalization of generalized linear models (GLM) of McCullaghand Nelder (1989). GLM parametrically fits models y5g(Xb )1« for differentdistributions (with different variance specifications). Consequently, one can easilyapply GAM using other distributions such as Poisson, gamma and multinomial.Hence, one can estimate count or duration data, survival data and probabilities withthe same flexibility in functional form.Estimation ResultsThis section provides an empirical illustration of the advantages of GAM using realestate data. Specifically, the first subsection provides the models and variables usedin the latter subsections, the second subsection discusses the data, the third subsectionfocuses upon the graphs of the estimated functions versus their arguments, the fourthVOLUME 15, NUMBERS 1 / 2, 1998