|
编者按：除了传统的解剖分期，美国癌症联合委员会第八版分期手册为乳腺癌推出了结合生物因素的预后分期。那么，预后分期与传统解剖分期相比，能否对乳腺癌相关生存提供更精细的分层？此前，德克萨斯大学MD安德森癌症中心和加利福尼亚癌症预防研究所发表的验证研究报告通过乳腺癌患者单中心队列和大样本人群数据库，对预后分期与解剖分期进行了比较，结果发现预后分期提供了更精准的预后信息，从而支持将其用于乳腺癌分期的临床实践。 2018年6月7日，《美国医学会杂志》肿瘤学分册在线发表默克研究中心、亨利福特医疗集团的通讯，对上述比较验证研究所用统计学检验方法进行了质疑。 该研究【1】分别根据预后分期和解剖分期，通过哈勒尔一致性指数，确定模型的预测性能。作者利用R软件包比较C进一步确定预后分期与解剖分期的哈勒尔一致性指数：
对于删失数据（观察值只有部分可知的数据，根据观察值不可观察的区域，分为右删失、左删失、区间删失数据），众所周知哈雷尔一致性指数可能高估一致性。该研究未报告两个队列的删失数据比例。根据该研究报告的生存曲线，两个队列大约75%的患者未观察到任何事件，并且研究结束时被删失，尤其IA～IIB期病变患者。此外，根据纽约大学和默克研究中心报告【2】，为了提供有效推论，R软件包比较C方法需要临床可能无法实现的高条件。如果条件无法实现，比较C方法可以引起严重偏倚和假阳性增加。一种替代方法是根据哈佛大学达纳法伯癌症研究所【3】提出的删失加权推算逆概率（R软件包生存C1），但是如果删失比例高，偏倚可能无法忽略。另一种方法是通过多因素比例风险回归模型或比例比值比回归模型，随后通过纽约纪念医院斯隆凯特林癌症中心的R软件包CPE【4】或纽约大学和默克研究中心的R软件包分离【5】，推算并比较一致性指数。 该研究还报告了通过赤池信息量准则比较模型拟合指数。对于单因素分析，可以直接推算哈勒尔一致性指数和删失加权C统计量的逆概率，而不用模型。尚不清楚为何需要模型推算一致性指数并进一步通过赤池信息量准则比较模型。纽约纪念医院斯隆凯特林癌症中心【4】R软件包CPE推算需要通过多因素比例风险回归模型，不过该模型的拟合优度检验比赤池信息量准则更重要，因为违反比例风险假设可以引起推算偏倚。 参考文献
相关阅读 JAMA Oncol. 2018 Jun 7. [Epub ahead of print] Statistical Tests Used to Validate the American Joint Committee on Cancer Eighth Edition Prognostic Stage Compared With the Anatomic Stage in Breast Cancer. Yilong Zhang; Xiaoxia Han. Merck Research Laboratories, Rahway, New Jersey; Henry Ford Health System, Detroit, Michigan. Weiss et al validated the American Joint Committee on Cancer Eighth Edition prognostic stage and compared it with the anatomic stage in breast cancer in 2 large cohorts. The authors used the Harrell C index to qualify the models' predictive performance based on prognostic stage and anatomic stage, respectively. The authors further determined the significance between the Harrell C index of the prognostic stage and anatomic stage using the R package compareC. In the MD Anderson cohort, the Harrell C indices for the prognostic stage and the anatomic stage are 0.8357 and 0.7370 (P<.001). In the California Cancer Registry, the Harrell C indices for the prognostic stage and the anatomic stage are 0.8426 and 0.8097 (P<.001). With censored data, it is well known that the Harrell C index can overestimate the C index. Weiss et al did not report the proportion of censored data for the 2 cohorts. Based on the Kaplan-Meier curves in the article, the 2 cohorts have approximately 75% subjects for whom no event was observed and who were censored at the end of the study, especially those with stage IA to IIB disease. Furthermore, to provide a valid inference, the method implemented in the R package compareC requires a strong condition that might not hold in practice. If the condition does not hold, the compareC method can induce a serious bias and inflated type I error. An alternative way is to use the inverse probability of censoring weighting estimator proposed by Uno et al (R package SurvC1), but the bias may be nonnegligible if the censored proportion is high. Another way is to assume a Cox proportional hazards (PH) model or proportional odds model and then apply the method proposed by Gonen and Heller (R package CPE) or Zhang and Shao (R package evacure) to estimate and compare the concordance indices. The authors also report the Akaike information criterion (AIC) to compare model fits. For univariate analysis, the Harrell C index and the inverse probability of censoring weighted C statistics can be estimated directly without assuming a model. It is not clear why a model is required to estimate the C index and further compare the model by using the AIC. The Gonen and Heller estimator requires a Cox PH model, yet the goodness-of-fit test of the Cox PH model is more important than the AIC because the violation of the PH assumption can lead to a biased estimator. DOI: 10.1001/jamaoncol.2018.0884 |
|
|
