구간 중도절단 자료에서 Cox 비례위험모형의 추정방법 비교

Abstract

Objectives: Our purpose was to compare the methods for the regression coefficients estimation in a Cox proportional hazards model with interval-censored data. Methods: The methods included the mid-point and right-endpoint imputation methods and 'intcox' method implemented following the algorithm proposed by Pan (1999, Journal of Computational and Graphical Statistics). Their performance was evaluated based on the estimated bias and mean squared error and the empirical coverage rate for the regression coefficients. Three methods were also discussed with real data analysis. Results: The 'intcox' method had the better performance than the other imputation methods. In particular a 95% coverage rate of the 'intcox' method was very close to 0.95, while the imputation methods were pretty less than 0.95. As a right censoring rate decreases, all the methods underestimated the true value, but the 'intcox' method seemed to be more stable than the others owing to smaller bias and mean squared error. With analyzing the breast cosmetics data, the effect size of treatment based on the 'intcox' method was the largest among three methods, but there was no difference in statistical significance among them. Conclusions: In many clinical studies the imputation methods were often used for dealing with the interval-censored data because they can be easily implemented using commercial softwares. However they may entail biased results as shown in simulation results such as low coverage rate. We recommend the ‘inxcox’ method or multiple imputation methods rather than the single imputation methods.