研究方法 – CL 臺灣住宅價格指數

CL 指數之主要特點

📈 納入短期轉售交易Inclusion of Short-Term Sales

臺灣多數現有價格指數排除持有期間未滿六個月的交易。本指數考慮不動產短期投資行為，據此建構了納入此類短期交易的指數。

Most existing price indices in Taiwan exclude transactions where the holding period is less than six months. This index accounts for short-term real estate investment behavior and is therefore constructed to include such short-term transactions.

📋 資料全面Comprehensive Data Coverage

此指數反映臺灣幾乎所有不動產銷售交易。指數建構過程手動收集並整合2012Q3前109個地方地政事務所的交易記錄。

This index reflects nearly all real estate sales transactions in Taiwan. The index construction process involved manually collecting and integrating transaction records from 109 local land administration offices prior to 2012Q3.

🔁 降低選擇性偏誤Reduced Selection Bias

與標準重複交易指數不同，後者在價格成長方面可能存在正向選擇偏誤。本指數採用基於 McMillen (2012) 的配對估計法。此混合方式依據精確地理位置與房屋特徵進行交易配對，在最大化樣本涵蓋率與估計精度的同時識別獨特不動產。

Unlike standard repeat sales indices, which can suffer from positive selection bias regarding price growth, the CL index uses a matching estimator approach based on McMillen (2012). This hybrid approach matches transactions on precise geolocation and unit characteristics, allowing identifying unique properties while maximizing sample coverage and precision.

🔬 依建物品質及特徵調整High Granularity

本研究方法可建構品質調整後的價格水準，涵蓋特定不動產特徵，如建物面積、屋齡及距捷運站距離。

The methodology allows for the creation of quality-adjusted price levels that account for specific property amenities like floor space, building age, and distance to transit.

資料來源

此指數以臺灣內政部每季公布的實價登錄（APR）資料庫為基礎。自2012年8月起，所有不動產交易均須申報，使其成為目前最全面的臺灣住宅交易資料來源。

The index is based on the Actual Price Registration (APR) database (實價登錄) published quarterly by Taiwan's Ministry of Interior (MOI). Since August 2012, all real estate transactions must be registered, making this the most comprehensive source of Taiwan housing transaction data available.

原始資料檔案從內政部開放資料平台下載。

The raw data files are downloaded from the MOI Open Data Platform and organized by quarter and city code.

變數建構Variable Construction

建材標準化為6類（鋼骨、鋼筋混凝土、磚造、木造等）
Building materials standardized into 6 categories (steel, reinforced concrete, brick, wood, etc.)
屋齡計算方式：交易年份減去建築完工年份
Property age computed as transaction year minus construction completion year
新竹市+縣及嘉義市+縣透過將縣級觀測值重新編碼至市代碼予以整合
Hsinchu City+County and Chiayi City+County are consolidated by recoding county-level observations to city code

混合重複銷售特徵評估迴歸

此指數採用混合重複銷售/特徵評估法，延伸自 McMillen (2012) 及 Fang et al. (2015)，透過轉換以下迴歸中的時間固定效果來估計交易價格指數（引自 Chi、LaPoint & Lin, 2025）：

The index adopts a hybrid repeat-sales/hedonic approach in the spirit of McMillen (2012) and Fang et al. (2015) that transforms the time fixed effects in the following regression to estimate a transaction price index (Chi, LaPoint, and Lin, 2025):

\log P^c_{i,t} = \delta^c_t + \gamma^c_{\tilde{i}} + \beta^{c\prime} X_{i,t} + \varepsilon^c_{i,t}

其中：

where:

\(P^c_t = \exp(\delta^c_t)\) 為市場 \(c\) 在時間 \(t\) 的季度價格指數。
\(P^c_t = \exp(\delta^c_t)\) is the quarterly price index for market \(c\) at time \(t\).
\(i\) 代表不動產個體；\(t\) 代表季度；\(c\) 為地區與不動產使用類別之組合（例：臺北市住宅）。
\(i\) indexes a property; \(t\) denotes a quarter; \(c\) refers to a classification by regional market × property use category (e.g., Taipei residential).
\(\delta^c_t\)：時間固定效果——此即欲估計的指數值。
\(\delta^c_t\): time fixed effects — these are the estimated index values.
\(\gamma^c_{\tilde{i}}\)：不動產類型固定效果，控制所有時間不變之觀察或未觀察特徵。對於方法一，此為地址區塊級固定效果：若兩筆交易共享相同地址字串，則指派相同的面板識別碼（涵蓋約 85% 的交易）。
\(\gamma^c_{\tilde{i}}\): property type fixed effects, controlling for all time-invariant observed or unobserved characteristics of the property. For Method 1, these are block-level fixed effects: two transactions are assigned the same panel id if they share the same address string (~85% of transactions).
\(X_{i,t}\)：潛在隨時間變化的特徵控制向量，包含土地面積與建物面積的多項式、建物總樓層數，以及所在樓層（公寓及辦公用途適用）。
\(X_{i,t}\): a vector of potentially time-varying hedonic controls including a polynomial in land area and floor space, the number of floors in the building, and the unit floor (for apartments and office space).
\(\varepsilon^c_{i,t}\)：誤差項。
\(\varepsilon^c_{i,t}\): the error term.

基期標準化Base Period Normalization

所有指數以 2000Q1 = 1.00 為基準，對應本資料集的起始期，便於長期比較。

All indices are normalized so that 2000Q1 = 1.00. This corresponds to the start of the dataset and facilitates long-run comparisons.

最小觀測值門檻Minimum Cell Size

城市-季觀測值少於 27 筆重複交易的組合予以剔除，以避免不穩定的估計值影響指數。此門檻遵循重複交易指數建構的標準做法。

City-quarter cells with fewer than 27 repeat-sale observations are dropped to prevent unstable estimates from driving the index. This threshold follows standard practice in repeat-sales index construction.

四種房屋配對方法

本指數使用四種漸進嚴格辨識跨交易的相同實體不動產之方法，提供穩健性評估。

A key methodological contribution is the use of four progressively stringent criteria for identifying the same physical property across transactions. This allows a robustness assessment unavailable in any existing Taiwan index.

M1

地址區塊級配對 Block-Level Address Matching

以標準化地址區塊（移除單元/樓層識別符）分組不動產。最寬鬆的方法，涵蓋約 85% 的交易。

Properties are grouped by standardized block-level address (removing unit/floor identifiers). Most permissive method, covering ~85% of transactions.

M2

GPS 座標配對 GPS Coordinate Matching

以內政部登記的經緯度座標精確配對不動產。需有 GPS 資料，涵蓋約 18% 的交易。

Properties matched exactly on latitude/longitude coordinates registered with the MOI. Requires GPS data; covers ~18% of transactions.

M3

GPS + 四捨五入面積（±5 m²） GPS + Rounded Area (±5 m²)

GPS 座標加上四捨五入至最近 5 m² 的建物/土地面積，涵蓋約 7% 的交易。減少跨交易間細微測量差異造成的配對錯誤。

GPS coordinates plus building/land areas rounded to the nearest 5 m², covering ~7% of transactions. Reduces mismatches from minor measurement differences across transactions.

M4

GPS + 精確整數面積 GPS + Exact Integer Area

最嚴格：GPS 座標加上整數四捨五入的建物面積必須吻合，涵蓋約 5% 的交易。最小化錯誤的不動產配對。

Most stringent: GPS coordinates plus integer-rounded building area must match, covering ~5% of transactions. Minimizes false positive property matches.

發佈之CL指數使用方法一（最廣泛涵蓋）作為全國及城市級指數。方法二至四作為穩健性檢驗，用於房屋類型分層分析。

The main reported CL index uses Method 1 (broadest coverage) for the aggregate national and city-level indices. Methods 2–4 serve as robustness checks and are used in the property-type stratified analyses.

樣本篩選標準

僅納入建物+土地合併交易（排除僅土地或僅建物）
Include only building + land bundle transactions (excludes land-only or building-only)
限於涉及恰好一棟建物及一筆土地的交易（排除大宗/組合交易）
Restrict to transactions involving exactly one building unit and one land parcel (excludes bulk/portfolio deals)
剔除涉及 ≥ 15 筆地號或 ≥ 15 棟建物的交易
Transactions with ≥ 15 parcels or ≥ 15 units are dropped
剔除政府買方交易（以買方身分證字號代碼辨識）
Drop government-buyer transactions (identified from buyer identifier codes)
剔除家庭/關係人交易（透過備註欄位關鍵字搜尋辨識）
Drop family/insider transactions (identified from the remarks field via keyword search)
排除1946年以前建造的不動產（結構品質疑慮）
Exclude properties constructed before 1946 (structural quality concerns)
排除新建物（銷售時屋齡 = 1年）
Exclude new buildings (age = 1 year at time of sale)
要求完整的交易及建築日期；剔除格式不正確的日期
Require complete transaction and construction dates; drop malformed dates
剔除缺少 GPS 座標的觀測值（方法二至四所需）
Drop observations with missing GPS coordinates (required for Methods 2–4)

參考文獻

完整模型請見：

The full methodology is described in:

Chi, Chun-Che, Cameron LaPoint, and Ming-Jen Lin. (2025). "Flip or Flop? Tobin Taxes in the Real Estate Market." 第D節（附錄）描述品質調整定價動態之完整模型、四種不動產之配對方法、及與其他各指數之比較。 Appendix Section D describes the full methodology for quality-adjusted pricing dynamics, including the regression specification (Equations D.1-D.2), the four property-matching methods, and comparison with the Official and Sinyi indices.