本研究以國泰建設公司住宅新推個案市場調查資料，建立預售屋特徵價格模型。以「最小平方法(ordinary least squares, OLS)」為基準，比較DFFITS和「最小消去平方法(least trimmed squares, LTS)」異常點刪除技術的表現。LTS尋求配適多數樣本的迴歸參數，當樣本殘差值大於門檻值則賦予權重0，視為異常點，再以OLS校估參數，稱為「再加權最小平方法(re-weighted least squares, RLS)」。實證結果發現：1. RLS和DFFITS模型表現較OLS佳。2.住宅新推個案的異常點特色，來自特定區位和特定行政區的產品定位。3. RLS和DFFITS模型的房價指數，長期波動大致雷同，惟若觀察短期變化，不同異常點處理技術的房價波動不同，影響短期解讀。

The study employs the presale housing market survey data of the Cathay Real Estate Development Company to establish presale housing hedonic models. To observe the effects of outliers, ordinary least squares (OLS) is employed as a benchmark to compare the model performance of two outlier deletion techniques, DFFITS and least trimmed squares (LTS). LTS is aimed at fitting a regression model to most of the data while identifying the outliers as the points with large residuals. By giving a zero weight to the cases with residuals larger than a threshold value, the outliers are disregarded in the following OLS calibration process. The technique is referred to as reweighted least squares (RLS). The results demonstrate that: 1. the RLS regression and DFFITS models outperform the OLS models; 2. most outliers come from a specific area and product positioning from specific districts; and 3. in the long term, these different estimation techniques do not affect housing price indices-however, if we observe the movements season by season, the different estimation techniques yield different price movements, which affect the interpretation of short-term presale housing market data.