Model Selection for the Bimodal Effect in Individual Cost-Effectiveness Analysis

• Main Authors: WU,JING-YI, 吳軍毅
• Other Authors: LIN,JIAN-FU
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/13383369883876338189

碩士 === 國立臺北大學 === 統計學系 === 104 === Individualized cost-effectiveness analysis (CEA) is often used in the assessment of a medical decision. Individualized CEA uses regression models to predict cost and effect simultaneously; then, the predicted values for cost and effect are used to calculate the individualized incremental cost-effectiveness ratio (ICER). In surgical researches, the marginal distribution of the effect variable is often bimodal. It is unclear how to select the best prediction model for the bimodal distribution of the effect variable to calculate the individualized ICER. In this study, we use simulation to assess four different prediction models for the bimodal distribution, including linear regression, beta regression, boosted beta regression and ordinal regression model with cubic spline. We simulated linear and non-linear bimodal distributions based on eight scenarios. We used 10-fold cross validation to compare different prediction models and different scenarios. We also compared different prediction models by analyzing random samples from a real population database. Finally, we conducted a sensitivity analysis by calculating individualized ICER based on random samples from a real population database. We found that linear beta regression had best performance in terms of prediction error from corss-validation for bimodal distributions. However, in real data analysis, we found that ordinal regression with cubic spline had best performance for bimodal distributions. In the sensitivity analysis, we found that linear beta regression and ordinal regression cubic spline were all suitable for calculating individualized ICER. we also found that the performances were often interfered by those outliers in a random sample. The prediction models' performances are usually influenced by different contexts. Bimodal distributions, the number of covariates, collinearity among covariates and outliers all affect the prediction models' performances. We suggest conducting a sensitivity analysis before we select a prediction model to calculate individualized ICER.