The effect of proximity to a registered sex offender's residence on single-family house selling price.

(15.) If points with quality ratings of 2 to 6 are included, the accuracy of the study suffers because the system places poorly geocoded points at the centroid of the zip codes. This would result in a cluster of sold houses (and/or offenders) located in one place, which is obviously not the case. No sexual predators or habitual offenders were eliminated from the study due to bad geocodes.However, 4 sexually oriented offenders and 11S sales transactions were eliminated for this reason.

(16.) ArcView[R] GIS (Redlands, Calif.: Environmental Systems Research Institute (ESRI), Inc.). In determining the proximity measures used in Equation (1), it was required that the offender had been in residence for at least one week before the purchase contract was signed. At some point in the year, an offender may have lived closer to a sold property than the one used to calculate the proximity measure. If the offender was not in residence prior to the sale, there is no way to attribute any price effect to them.

(17.) TIGER[R] Line Files (Washington, D. C.: United States Department of Commerce, Bureau of the Census, 1992).

(18.) Stepwise regression is the appropriate process to use in this case because the major objective is not to predict the value of the dependent variable; the major objective is the analysis of the independent variables, in particular PROX. In a predictive model, multicollinearity inhibits the analysis of independent variable effects due to the instability of the regression coefficient of each independent variable. In stepwise regression, an independent variable enters the model only if it explains variation in the dependent variable that is not already explained by variables in the model. Therefore, the predictability of the dependent variable is maximized while multicollinearity is minimized. Of course, predictability could be higher by including all possible independent variables in the model because the addition of any variable cannot lower the [r.sup.2]. But the possible presence of multicollinearity in this situation makes the interpretation of an independent variable problematic.

(19.) Analyzing explanatory variables in a regression model where all explanatory variables are qualitative (e.g., 0, 1) is basically equivalent to performing an ANOVA (analysis of variance). In this study, ANOVA is inappropriate because the study model contains both qualitative and quantitative variables (e.g., lot size). When a regression model has both qualitative and quantitative explanatory variables, this is basically equivalent to performing an ANCOVA (analysis of covariance). ANCOVA requires the analyst to test an important additional assumption beyond those in ANOVA. This test is described later in the article.

(20.) By eliminating observations from inner ring(s), the price effect from the inner ring(s) is also eliminated. Hence, any price effect discovered would be due solely to the difference between the houses in the subject ring and those located farther away.

(21.) The study attempted to do the same for offenders subject to limited disclosure by estimating the model after eliminating observations located within the significant range for passive notification offenders. However, because there are so many passive notification offenders, the model lost full rank for all rings. Because the effect of limited disclosure offenders should dominate the effect of passive notification offenders, we do not consider this a major problem, but there may be an interaction effect that the model is not capturing.

(22.) Variables that did not enter and remain in the model include SPRING (transactions that occurred in March, April, and May), FALL (transactions that occurred in September, October, and November), BATH2 (houses with more than one, but less than three, full bathrooms), and 22 location variables.

(23.) The holdout category to control for number of bathrooms was BATHI (where the house had one full bathroom), which was the most prevalent value in the sample. Three observations were recorded in the data as having less than one full bath. Some older houses in the sample area have less than full amenities. These three observations were included in the holdout category.

(24.) Values for this variable were obtained from the Montgomery County Treasurer's Office. If the mailing address for property tax purposes was the same as the property address, it was assumed that the buyer intended to live in the property and the variable was coded 1. If the mailing address differed from the property address, it was assumed that the buyer did not intend to live in the property and the variable was coded 0.

(25.) The holdout category to control for seasonal effects was SUMMER, which was the most prevalent value in the sample.

(26.) The tax district with the most observations in the sample was used as the holdout category for LOC. Montgomery is an extremely heterogeneous environment with neighborhoods that include extremely wealthy satellite cities, inner-city neighborhoods, new upscale suburban neighborhoods, and rural communities. The tax districts, although not perfect, divide the heterogeneous county into somewhat homogeneous subgroups. Inclusion of a variable to describe individual house condition would be preferable. This information was unavailable for the present study.

(27.) H. White, "A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity," Econometrica 48 (1980): 817-838.

(28.) SAS/STAT[R] User's Guide: Version 6, 4th ed., vol. 2 (Cary, NC: SAS Institute, Inc., 1989). Whenever dealing with microeconomic data there is the possibility that the error terms do not follow the classical assumption of homoskedasticity (i.e., that the variance of the error term is constant for all observations). If the error terms of the equation are heteroskedastic (i.e., the variance of the error term is related to the size of a particular independent variable), ordinary least squares estimators will be inefficient (i.e., they will not have the minimum variance). Use of an incorrect functional form can result in incorrect estimators. Because the tests conducted indicated that no specific functional form was significantly better than the others, we present the linear model results for expository expedience.

(29.) David A. Belsley, Edward Kuh, and Roy E. Welch, Regression Diagnostics: Identifying Influential Data and Sources of Collinearity (New York: John Wiley and Sons, 1980).

(30.) Barbara G. Tabachnick and Linda S. Fidell, Using Multivariate Statistics (New York: Harper & Row, 1983).

(31.) Blocking variables is a recommended method to correct for a violation of constant covariates. In this process, the range of the culprit quantitative variable (e.g., AGE, which, for example, ranges from 0 to 90) is divided into subgroups and these are treated as a series of qualitative variables (e.g., AGE1 if AGE [greater than or equal to] 0 or [less than or equal to] 10, AGE2 if AGE >10 or [less than or equal to] 20, and so forth).

(32.) See for example, James E. Larsen, "Money Illusion and Residential Real Estate Transfers," Journal of Real Estate Research 4, no. 1 (1989): 13-19, and James E. Larsen, "Leading Residential Real Estate Sales Agents and Market Performance," Journal of Real Estate Research 6, no. 2 (1991): 241-249, where the study areas examined included the county analyzed in the present study.

(33.) The number of observations mentioned here (802) does not match the total number of observations shown for the first two rings in the lower portion of Table 3 because (as explained previously) observations located within 0.3 mile of an offender subject to limited disclosure were eliminated from the sample before the test reported in the lower portion of Table 3 was conducted.

James E. Larsen, PhD, is a professor of finance in the Raj Soin College of Business at Wright State University in Dayton, Ohio. Larsen has published numerous articles in journals such as The Appraisal Journal, Real Estate Appraiser and Analyst, Journal of Real Estate Research, Journal of Real Estate Practice and Education, Real Estate Appraiser, and Journal of the American Real Estate and Urban Economics Association. He is a member of the Ohio Real Estate Commission's Education and Research Fund Advisory Committee, and is author of Real Estate Principles and Practices, published by John Wiley & Sons, Inc. Contact: james.larsen@wright.edu

Kenneth J. Lowrey is a lecturer in the Department of Urban Affairs and Geography at Wright State University (WSU) in Dayton, Ohio. He teaches geography and geographic information systems (GIS). His latest research is on the topic of gated communities and has been published in Urban Geography. Lowrey has produced numerous maps for use on the WSU campus and in southwestern Ohio. He is also the WSU representative to the Ohio GIS-Net, a statewide GIS organization representing the state universities in Ohio. Contact: kenneth.lowrey@wright.edu

Joseph W. Coleman, PhD, is an associate professor of management science and information systems in the Raj Soin College of Business at Wright State University in Dayton, Ohio. Coleman has published numerous articles in journals such as The Appraisal Journal, IEEE Transactions on Reliability, and Communications in Statistics: Simulation and Computation. Contact: joseph.coleman@wright.edu