摘要
目的 C统计量是评价Cox比例风险回归模型区分度的常见指标,然而,目前对C统计量的算法仍存在争议。本文将探讨C统计量的计算方法及其SAS实现,为编程输出Cox模型的C统计量提供参考。方法 运用PHREG过程估计研究观察期末的累积生存概率,判断实际生存时间与预期生存函数是否同趋势,并以此计算C统计量及其95%置信区间。以某注册登记研究为例,评价年龄、血压和心率对急性心衰患者出院后30d死亡率的预测区分度。结果 研究共纳入2836例急性心衰患者,年龄、基线收缩压和基线心率对出院后30d死亡的影响差异都具有统计学意义(均有P〈0.05),其中年龄(单位:岁;风险比(hazardratio,HR):1.029;95%置信区间(confi-denceinterval,CI):1.022~1.037)和心率(单位:次/分;HR:1.011;95%CI:1.007~1.014)为危险因素,收缩压(单位:mmHg;HR:0.992;95%CI:0.989~0.995)为保护因素。模型C统计量达到0.638(95%CI:0.570~0.704),可见模型具有一定的区分度,使用SAS程序能够得到所需结果。结论 C统计量是评价模型区分度的良好手段,并可以通过SAS程序求得。
Objective C statistics is one of the most widely-used indexes in accessing the discrimination of the Cox proportional hazards regression models. However, the calculation methods for C statistics have been controversial. Our study aims to investigate the calculation of C statistics and its implementation in SAS. Methods To calculate C statistics and its 95% confidence interval ( CI), we used PROC PHREG to predict the survival function, and decided whether the predicted survival probabilities was consistent with the actual survival times. Taking a registry study as an example, we e- valuated the discrimination of a Cox regression model which predicted the 30-day mortality after discharge in patients with a- cute heart failure. Results A total of 2 836 patients were included in the final analysis. Older age ( Unit: years; hazard ratio (HR) : 1. 029; 95% CI: 1. 022-1. 037), lower systolic blood pressure ( Unit: mmHg; HR: 0, 992; 95 % CI: 0. 989- 0. 995) and increased pulse rate ( Unit: beats/min; HR: 1.011 ; 95% CI: 1. 007-1. 014) were all statistically significant predictors for 30-day post-discharge death. The C statistics of the model was 0. 638 (95% CI: 0. 570-0. 704), indicating a certain degree of discrimination. Conclusions C statistics is a good index for accessing the discrimination of Cox regres- sion models, and it can be calculated by SAS programs.
出处
《中华疾病控制杂志》
CAS
CSCD
北大核心
2016年第9期953-956,961,共5页
Chinese Journal of Disease Control & Prevention
关键词
统计学
非参数
模型
统计学
流行病学方法
Statistics, the parameters
Models, statistical
Epidemiologic methods