空間微觀模擬方法於微資料的人口特徵估計與空間結構的誤差分析

Abstract

目標：在微觀研究中，總體資料難以反應多統計項目間的交互關係，經常會以微資料作為分析材料。微資料的調查難度隨規模增加，空間微觀模擬方法的出現，讓研究者能以少量的抽樣微資料，以各地區統計資料為加權依據，估計出人口微資料模型。但目前的驗證過程中，無法檢驗微資料交叉項目誤差量。方法：本研究使用2000年台澎地區各鄉鎮市區普查原始資料，以最佳組合法、重複比例擬合法和廣義迴歸法，三種空間微觀模擬方法估計居民的微資料。將估計之微資料與真實的微資料，從資料結構與空間結構兩方面進行個體尺度的誤差驗證與討論。結果：最常使用之總絕對誤差法，低估重複比例擬合法與廣義迴歸法的誤差量。雖然如此，這兩方法估計之微資料，於多數地區依然可以反映真實微資料的統計特性，替代真實資料進行後續分析，但最佳組合法於多項誤差檢驗中與真實微資料有明顯差異。將誤差依發展程度分組討論後，發現各地區誤差的空間分布受抽樣過程與區域差異影響。而造成各項目估計時誤差的因素，除了區域差異外，當項目的欄位數量較多時，誤差量也有偏高的狀況。結論：空間微觀模擬方法的估計結果能取代真實微資料作為分析材料，但研究者在操作時應選擇適合研究區的估計方法，並在空間差異明顯時，將地區分組再進行估計，及減少估計的項目欄位數量，將能降低空間微觀模擬方法的估計誤差。

Source：Micro-data was often used in the microsimulation research, but a large number of population micro-data surveys are difficult. Spatial microsimulation models are being used to create simulation micro-data for geographical areas. The models combine sample records with benchmark data for areas by re-weighting sample records to fit statistical data for each area. However, the ways of validation are debatable. Because those ways are compares the simulated micro-data to the constraint data used in the model, the interaction of variables can’t be verified. Method：We re-conducted the spatial microsimulation for all townships in Taiwan. By using three methods, including Combinatorial Optimization (CO), Iterative Proportional Fitting (IPF) and Generalized Regression (GREGWT), combined sample records with benchmark data for areas with the Taiwan Census raw data in 2000. By the individual scale method, we compared the data structure error and the spatial structure error between the realistic micro-data and the simulation micro-data. We analyze the reason of error distribution by regional differences and variables selection and make the recommendations of spatial microsimulation model. Result：Although Total Absolute Error underestimate the error by IPF and GREGWT, the simulation micro-data can still replace the real micro-data for analysis. The simulation micro-data of CO are significant differences with realistic. The error spatial distribution was affected by the sampling process and regional differences. Variable fields number, data distribution and regional differences are the influence factors of the variable estimated error. Conclusion：Spatial microsimulation model can replace realistic data with selecting the appropriate method, grouping the zones and decreasing the number of variable fields.