以往存活分析的成本估計都只看單一存活事件的壽命時間 (life time) 或是失敗時間 (failure time)。然而有些疾病會有復發的情形發生,像是腦中風或是乳癌等。 在研究中,常常失去病患的後續追蹤,導致不完整的資料觀察,此為右設限的資料類型。若是直接加總右設限的成本並且取平均數,則可能會造成錯誤的推論。故希望利用機率倒數加權 (inverse probability weighting) 的概念來正確估計真實復發成本。本篇論文將應用 Bang 與 Tsiatis (2000) 所提出的分割加權成本估計式 (partitioned weighted estimator) 推廣至復發事件時間,進而提出復發加權成本估計式 (recurrent weighted estimator)。 依照資料復發時間是否有順序與是否分層,總共有三種危險集合的設定方式,分別為 Andersen 與 Gill (1982), Wei、 Lin 與 Wiessfeld (1989),Prentice、 Williams 與 Peterson (1981) 所提出的邊際模型,以此建立三個估計式,並進行一系列的模擬來比較三種估計式估計結果的優劣。最後以某醫院的腦中風中心做為實例,估計成立腦中風中心前後的醫療成本是否有差異,並且分析與探討有可能造成差異的原因。
The existing estimator for the medical cost is usually focused on the cost for a single event. Since censoring is commonly seen when collecting the medical data, to obtain an unbiased cost estimate, this estimate has to be adjusted. The most common adjustment is to use the survival or censoring probability as a weight to compute a weighted cost estimator, which is an unbiased estimate. Nevertheless, some illnesses may recur, such ascerebral vascular accident or breast cancer and so on. The recurrent data may be also incompletely observed. To have a more accurate estimate, this thesis extends the partitioned weighted cost estimator proposed by Bang and Tsiatis (2000) for a single event to the multiple events. The probability of recurrence depends upon the risk set which can depend on the occurrence of the event time and the number of recurrence. Three well-known estimates are proposed by Andersen and Gill (1982), Wei, Lin and Wiessfeld (1989), Prentice, Williams and Peterson (1981), respectively. Using these three estimators, this thesis suggests three partitioned weighted estimators for the multiple events. Monte Carlo simulations are performed to compare the performance of these estimators under various scenarios. Finally, a real data is used to illustrate the usage of these estimators.