臺灣高齡人口長照供需落差分析與需求預測

Abstract

中文摘要 人口快速高齡化使長期照顧（以下簡稱長照）資源的區域配置成為重要的公共政策課題。本研究運用政府公開資料，探討臺灣各縣市長照供需落差的空間分布，並預測其未來走向，核心目的在於檢驗「衡量指標的選擇如何左右供需結論」。方法上，先以Python 自政府開放平臺蒐集人口、老化指數、身心障礙人口、老人福利機構與長照服務涵蓋率等資料，經清理後建立縣市主表與多維度供需指標；再整合面量圖視覺化、關聯式資料庫（SQLite/SQL）查詢、機器學習分類與多方法時間序列回測等技術。需求預測對老化指數同時採線性迴歸、年均複合成長率（CAGR）與移動平均外推至2030 年，並以遮蔽最後三年之回測（平均絕對誤差，MAE）作為選模依據。研究發現，將供給指標由「每千名老人核定床數」改為以失能需求為分母的「長照服務涵蓋率」後，真實供需缺口由都會集中轉為聚焦於離島與北部都會圈，顯示指標選擇會直接翻轉政策判斷；回測亦證實老化正在加速，線性外推將系統性低估未來需求。最終成果部署為公開互動儀表板，可供各界檢視與延伸應用。

Abstract Rapid population aging makes the regional allocation of long-term care (LTC) resources a pressing public-policy issue. This study uses open government data to examine the spatial distribution of the LTC supply–demand gap across Taiwan’s counties and to forecast its trajectory, with the central aim of testing how the choice of measurement indicator shapes supply–demand conclusions. Methodologically, Python is used to collect population, aging-index, disability, elderly-care-institution, and LTC service-coverage data from government open platforms; after cleaning, a county master table and multidimensional supply–demand indicators are constructed, and choropleth visualization, relational-database (SQLite/SQL) querying, machine-learning classification, and multi-method time-series backtesting are integrated. Demand forecasting extrapolates the aging index to 2030 using linear regression, compound annual growth rate (CAGR), and moving average, with a hold-out backtest (mean absolute error, MAE) guiding model selection. The results show that replacing “licensed beds per thousand elderly” with a demand-based “LTC service coverage rate” shifts the true gaps from metropolitan concentration toward the outlying islands and the northern metropolitan area, demonstrating that indicator choice can reverse policy judgments; backtesting further confirms that aging is accelerating and that linear extrapolation systematically underestimates future demand. The final results are deployed as a public interactive dashboard for review and further application.