Comparison between specifications of linear regression and spatial-temporal autoregressive models in mass appraisal valuation for single storey residential property
Property valuation is an area of interest for property owners, real estate agents, government bodies and researchers. There are various approaches to estimate a property value. Among them, the statistical and spatio-temporal methods incorporate the location and time in the valuation modelling. These...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2013
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/60054/7/FK%202013%2055IR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Property valuation is an area of interest for property owners, real estate agents, government bodies and researchers. There are various approaches to estimate a property value. Among them, the statistical and spatio-temporal methods incorporate the location and time in the valuation modelling. These models however, are not widespread as the simple linear models due to scarcity of proper data and incomprehensive research findings on their implementation issues. Effects such as normality treatment, definition of neighbourhoods and weights and choice of autocorrelation parameter and parameter estimation are some of the complexities that are inherent to these models. This study therefore, was designed to investigate different aspects of spatial and spatio-temporal autoregressive modelling. Further, the performance of these models compared to the standard linear model that is widely used in mass appraisal of real properties, was studied. Datasets of transacted terrace houses over the period 1999-2009 from Selangor, Malaysia were obtained and geocoded for analyses using cadastral and topographic maps and online mapping services. A complete data analysis was carried out on the datasets. Furthermore, various spatial, temporal and spatio-temporal neighbourhood and weighting schemes, optimization algorithms and lag and error modelling scenarios were created and tested with the data. A hold-out validation was performed for different sets of experiments. The best set of parameters that could produce more accurate results in the validation process, were selected and their associated neighbourhood and weights were used to compare with the linear models. The experiments were replicated on three different treatments based on removal of outliers and transformation of variables with high value of skewness. The results showed that although there was a strong presence of spatial autocorrelation in the dataset, especially when the outliers are removed, the results of linear and spatio-temporal models are mixed. The best result using criteria of coefficient of determination and the uniformity level of prediction belonged to the spatio-temporal lag and spatial lag models respectively. The error variant of the abovementioned models could only reduce the problem of heteroscedasticity in regression error residuals. Linear regression model could provide better uniformity level at the expense of very low R2 and higher heteroscedasticity in residuals. It was also found that the graph based neighbours would increase the chance of the spatial model to predict better. Furthermore, the row-standardized or stochastic weight matrices showed to be more effective compared to other weighting schemes. Finally, it was demonstrated that incorporating the space and time interaction (S×T or T×S) autocorrelation in the spatio-temporal model along with higher time interval between dates of transactions in temporal neighbourhood selection would produce more reliable results in prediction for spatio-temporal autoregressive models. |
|---|