样本量及删失率对生存分析模型有效性和偏倚性的影响

The Effect of Sample Size and Censoring Proportion on the Power and Bias of Survival Analysis Models

目的探讨样本量及删失率对生存分析模型有效性和偏倚性的影响,为根据资料特点选用适合的生存分析方法提供参考依据。方法以实际数据为总体,通过抽样模拟和删失模拟,比较不同样本量和删失率下,Cox模型和加速失效时间模型的有效性和偏倚性及其变化趋势。结果对于大样本(≥500)或删失率较小(≤30%)的数据,两种模型的有效性和偏倚性相近,分析结果均较为可靠。当样本较小(300～400)或删失率较高(40%～60%)时,加速失效时间模型的分析结果更为可靠。结论 Cox模型对于大样本或删失率低的数据更为实用。加速失效时间模型的分析结果在样本量较少或删失率较高时更为可靠。

Objective To explore the effect of sample size and censoring proportion on the power and bias of survival analysis models. To provide the basis of survival analysis models choice based on characteristics of data. Methods Sampling simulation and censoring simulation are used to evaluate the power and bias of Cox model and accelerated failure-time model and their trends at different sample size and censoring proportion. Results For the survival data with more than 500 samples or censoring proportion is less than 30% ,two kinds of models have the similar power and bias and their results are reliable. When the sample size is between 300 and 400 or censoring proportion is between 40% and 60%, accelerated failure-time model is the better choice. Conclusion For the survival data with large sample size and small censoring proportion, Cox model is more practical than accelerated failure-time model, otherwise the accelerated failure-time model is a better choice.