以健保資料推估現居地演算法正確性驗證比較：以美兆健康資料為例

Abstract

前言： 健保資料庫是臺灣重要的全國性健康資源，然而其記錄的地址為投保地址，而非實際居住地。由於人口流動頻繁，戶籍地、投保地與實際居住地之間常存在落差，導致地理分析的正確性受到限制。目前臺灣已有三種居住地估算演算法被提出：林等人（2011）、古等人（2018）與余等人（2020）。臺灣疾病負擔中心採用全球疾病負擔研究架構，利用臺灣健康資料庫系統性量化可歸因於危險因子的疾病負擔，並建立臺灣健康指標資料庫；其居住地估算演算法即由古等人之方法延伸發展而來。由於余等人僅提供概念性架構而未完整說明操作步驟，因此本研究針對未明確定義之步驟加以補充，並對三種演算法進行比較，以找出正確率最高的方法。臺灣民眾就醫行為具有特定模式：輕症傾向至鄰近診所就診，重症則前往大型區域醫院。本研究整合美兆健康資料庫中的居住地資訊，評估就醫院所是否與實際居住地相符，藉此作為演算法選擇之依據，以提升以健保資料庫為基礎之空間流行病學研究的地理正確性。 方法： 本研究以美兆健康資料庫（2005–2017年）作為黃金標準，該資料庫的族群為美兆健康管理機構接受健康檢查之民眾。篩選後納入 283,286 名個案、共 309,625 筆記錄。透過與全民健保研究資料庫進行配對世代串連，取得門診與急診資料及承保記錄。串連後的健保資料分別套用三種居住地估算演算法：林等人（演算法一）、臺灣疾病負擔中心（演算法二）、余等人（演算法三）。三種演算法各自推估居住地後，與美兆自填居住地進行正確率比較，分別於縣市層級與鄉鎮市區層級進行整體及次族群分析。鄉鎮市區層級結果進一步分別以38分鐘旅運時間與共同行政邊界定義生活圈，評估生活圈正確率。

結果： 在縣市層級，臺灣疾病負擔中心演算法之整體正確率最高（83.21%，95% 群集穩健信賴區間：83.07–83.34%），林等人次之（82.36%，95% 群集穩健信賴區間：82.19–82.52%），余等人三個版本正確率相近，介於 78.98% 至 79.15% 之間。在鄉鎮市區層級，臺灣疾病負擔中心演算法同樣表現最佳（66.50%，95% 群集穩健信賴區間：66.33–66.67%），余等人 V1 次之（62.76%，95% 群集穩健信賴區間：62.61–62.92%），林等人最低（54.45%，95% 群集穩健信賴區間：54.23–54.67%）。由於余等人三個版本（V1、V2、V3）在兩層級之正確率差異很小且信賴區間高度重疊，本研究以 V1 作為代表版本進行後續比較。在生活圈正確率分析方面，以38分鐘旅運時間定義生活圈時，臺灣疾病負擔中心演算法正確率最高（87.32%，95% 群集穩健信賴區間：87.16–87.48%），林等人次之（86.76%，95% 群集穩健信賴區間：86.53–86.98%），余等人 V1 最低（80.80%，95% 群集穩健信賴區間：80.50–81.10%）。以行政邊界鄰接定義生活圈時，三種演算法排序相同：臺灣疾病負擔中心演算法為 79.90%（95% 群集穩健信賴區間：79.71–80.09%），林等人為 77.69%（95% 群集穩健信賴區間：77.44–77.94%），余等人 V1 為 61.86%（95% 群集穩健信賴區間：61.57–62.14%）。 結論： 綜合縣市層級、鄉鎮市區層級及生活圈正確度之分析結果，臺灣疾病負擔中心演算法在所有指標與次族群分析中均表現最佳。尤其在行政邊界鄰接生活圈之比較中，臺灣疾病負擔中心演算法（79.90%）與余等人 V1（61.86%）之差距達 18 個百分點，顯示臺灣疾病負擔中心演算法於掌握跨鄉鎮就醫行為方面具有明顯優勢。因此，本研究建議以臺灣疾病負擔中心演算法作為利用健保資料庫進行空間流行病學研究時之居住地估算首選。

Background: The National Health Insurance (NHI) database is a key national health resource in Taiwan; however, recorded addresses reflect enrollment locations rather than actual residences. Frequent population mobility creates discrepancies among household registration, enrollment, and actual residential addresses, limiting the accuracy of geographic analyses. Three residential location estimation algorithms have been proposed in Taiwan: Lin et al. (2011), Ku et al. (2018), and Yue et al. (2020). The Taiwan Burden of Disease Center (TBDC), which applies the Global Burden of Disease (GBD) framework to quantify disease burden in Taiwan, extended from Ku et al.'s algorithm for residential location estimation. Since Yue et al. provided only a conceptual framework without full operational details, this study supplements the undefined steps and compares all three algorithms to identify the most accurate approach. Given that Taiwan's healthcare-seeking patterns differ by disease severity—mild cases typically visiting nearby clinics and severe cases seeking regional hospitals—this study uses residential location data from the MJ Health Database to evaluate algorithm performance and improve the geographic accuracy of NHI-based spatial epidemiological research. Methods: This study used the MJ Health Database (2005–2017) as the reference standard. The database comprises individuals who underwent health examinations at MJ Health Management Institution. After screening, 283,286 individuals with 309,625 records were included. Through probabilistic linkage with the National Health Insurance Research Database (NHIRD) — Ambulatory Care Expenditures by visit and Registry for Beneficiaries were obtained. The linked NHI data were then applied to three residential location estimation algorithms: Lin et al. (Algorithm 1), TBDC (Algorithm 2), and Yue et al. (Algorithm 3). The estimated residences derived from each algorithm were compared against MJ self-reported residences, with accuracy assessed at both the county/city level and township level for overall and subgroup analyses. Township-level results were further evaluated using living circle accuracy, defined both by 38-minute travel time and shared administrative boundaries. Results: At the county/city level, the TBDC algorithm achieved the highest overall accuracy (83.21%, 95% cluster-robust confidence interval [CI]: 83.07–83.34%), followed by Lin et al. (82.36%, 95% CI: 82.19–82.52%), while the three versions of Yue et al. showed similar accuracy ranging from 78.98% to 79.15%. At the township level, TBDC again performed best (66.50%, 95% CI: 66.33–66.67%), followed by Yue et al. V1 (62.76%, 95% CI: 62.61–62.92%), and Lin et al. (54.45%, 95% CI: 54.23–54.67%). As the three versions of Yue et al. (V1, V2, V3) showed minimal differences in accuracy with highly overlapping confidence intervals at both levels, V1 was selected as the representative version for subsequent comparisons. Regarding living circle accuracy, when defined by 38-minute travel time, TBDC achieved the highest accuracy (87.32%, 95% CI: 87.16–87.48%), followed by Lin et al. (86.76%, 95% CI: 86.53–86.98%), and Yue et al. V1 (80.80%, 95% CI: 80.50–81.10%). When defined by shared administrative boundaries, the ranking remained the same: TBDC at 79.90% (95% CI: 79.71–80.09%), Lin et al. at 77.69% (95% CI: 77.44–77.94%), and Yue et al. V1 at 61.86% (95% CI: 61.57–62.14%). Conclusion: Across all metrics and subgroup analyses, the TBDC algorithm consistently demonstrated the best performance at the county/city level, township level, and living circle accuracy. Notably, in the shared administrative boundary living circle analysis, the gap between TBDC (79.90%) and Yue et al. V1 (61.86%) reached 18 percentage points, indicating that TBDC holds a clear advantage in capturing cross-township healthcare-seeking behavior. Therefore, this study recommends the TBDC algorithm as the preferred method for residential location estimation in NHI-based spatial epidemiological research.