疾病率時空分析方法：台灣砷相關癌症之應用

Abstract

烏腳病是1950至1990年台灣西南沿海地區的流行病。本文欲利用台灣癌症登記資料比較砷相關癌症在烏腳病流行地區與非流行地區的發生率。然而1995年起癌症登記資料的品質才趨於穩定，一般的發生率長期趨勢難以看出烏腳病防治成效。再者烏腳病是一種地方性流行病，發生率的地理分布也是重要的議題。本文提出一種時空地圖的繪製方法。我先使用隨機效應年齡-年代-世代模型得到各個行政區域的時空參數估計，再對時空參數進行內插，求得精細至經緯度的時空參數。歷經前述步驟可繪製出疾病率在年齡、年代、與世代的等高線地圖，還能製作影片播放疾病率的時空變化。我也進行電腦模擬比較經驗貝氏法、核密度估計法、克里金法、以及穩定克里金法等內插方法，結果顯示穩定克里金法的表現為最佳。最後，我使用本文的方法分析台灣癌症登記中的砷相關癌症資料，發現烏腳病流行地區與非流行地區的癌症發生率差距在1958-1962出生世代後逐漸消失，這個轉折點恰好是烏腳病防治的起始時間，時空地圖也顯示砷相關癌症發生率群聚在西南沿海地區，在1960出生世代後逐漸消失。本文的方法不僅在電腦模擬有優於傳統方法的表現，在實例應用中也展示具有公共衛生價值的發現。

Blackfoot disease was endemic to southwestern Taiwan during the 20th century. This study aims to use the Taiwan Cancer Registry dataset to compare incidence rates of arsenic-related cancers between Blackfoot disease-endemic areas and the remaining areas of Taiwan. However, not until 1995 that the Taiwan Cancer Registry could provide high-quality data. Therefore, the long-term trend in the incidence rates cannot be used to evaluate the effectiveness of Blackfoot disease prevention. In addition, the geographical distribution of arsenic-related incidence rates is also an important issue. This study proposed a spatiotemporal mapping method. First, the random-effects age-period-cohort model is fitted to obtain spatiotemporal parameter estimations in each administration area. Then, these estimations are further interpolated to make predictions on each longitude and latitude. With the procedures, we can draw contour maps of disease rates according to age, period, and cohort and create a movie to show spatiotemporal dynamics vividly. The Monte-Carlo simulation compared interpolation methods, including empirical Bayes, kernel density estimation, kriging, and stabilized kriging. The stabilized kriging performed better among all. Finally, we analyzed arsenic-related cancers from the Taiwan Cancer Registry dataset. The incidence gap between Blackfoot disease-endemic areas and the remaining areas of Taiwan shrunk after the 1958-1962 birth cohort, coinciding with the initial prevention of Blackfoot disease. In addition, spatiotemporal clusters of high incidence rates were identified in southwestern Taiwan, and the clusters also started to dissipate after the 1960 birth cohort. Our method demonstrated appropriate ability and contributed to public health.