| Summary: | 碩士 === 國立臺北大學 === 統計學系 === 98 === Owing to aging, how to efficiently distribute the limited resources becomes very important. In turn, we need to know how to estimate medical cost accurately and efficiently. However, it is common to have dropouts when collecting medical cost. The naive estimators ignoring the unobservable data may be biased. Lin (1997) suggested partitioning the study duration and estimating the cost of each interval and then constructing the estimate by summing up the cost from each interval. Furthermore, to take into account of the unobservable data, Lin (1997) and Bang and Tsiatis (2000) proposed weighted estimators that used the survival probability and uncensored probability as the weight, respectively. This thesis compares the performance of these estimators under various scenarios.
In addition, the medical cost may be related to many covariates. Baser (2006) suggested using the general linear model for the longitudinal data to model the partitioned cost, where a random intercept is included. This thesis extends the model to a more general parametric model. Furthermore, similar to the weight concept in Lin (1997), we suggest using the uncensored probability as the weight which is estimated from the Cox proportional hazards model. The performance of these models is evaluated using simulations. Finally, the proposed model is implemented on the data extracted from Health Insurance database for patients with the colorectal cancer.
|
|---|
