【作者】 金欢；

【导师】 赵耐青； 涂冬生；

【作者基本信息】 复旦大学 ， 流行病与卫生统计学， 2011， 博士

【摘要】 据WHO官方网站公布的信息,恶性肿瘤目前在世界范围内是导致死亡的最主要的原因之一。2008年约有760万人死于恶性肿瘤,约占总死亡人数的13%；预计到2030年,世界范围内死于恶性肿瘤的人数将超过1100万人。调查显示,我国恶性肿瘤发病数正以年均3%～5%的速度递增。在城市,肺癌、胃癌和乳腺癌的发病率最高；在农村,胃癌、食管癌和肝癌是前三位的高发癌症。因此,针对抗肿瘤治疗方法的研究越来越引起了社会各界的广泛重视。寻找抗肿瘤治疗效果证据的最安全最有效的途径是临床试验。临床试验的主要目的是研究治疗的有效性与安全性。肿瘤临床试验作为临床试验的一个特殊分支,具有自己的特点：首先,结局指标往往是某结局事件发生的时间；其次,随访时间通常较长,病人在随访中更容易发生删失,因此最常用的分析方法是生存分析。由于肿瘤临床试验的特殊性,在肿瘤临床试验的设计和数据分析中,可能遇到各种各样的问题。到目前为止,仍然存在一些问题尚未解决。本论文的目的就是探索与研究在肿瘤临床试验中遇到的三个问题的解决方法。第一个问题,在一类探索新治疗有效性的研究中,研究者为在常规治疗下发生了进展(一般定义为肿瘤增大25%)的带瘤病人尝试新的治疗,为了在有限的成本下能够对新治疗的效果进行评价,研究者怎样利用这些病人在之前接受的常规治疗下的肿瘤进展时间或其相关函数作为对照,与新治疗下的肿瘤进展时间或其相关函数进行比较,来对新治疗的效果进行评价。在比例风险假定成立时,我们给出了相关生存数据在非劣性假设下,用非参数的加权秩检验进行分析的方法,对新治疗的效果进行评价。并且通过随机模拟的方法对检验的效果进行了评估,发现效果很好。该方法不仅限于以上情形,更可以应用于任何配对或相关生存数据的非劣性假设检验中。另外,我们在无删失和存在删失的条件下,分别给出了非劣性假设检验相关生存数据所需样本量的计算公式,并在计算出的样本量下通过随机模拟并计算实际的统计效能(power)。从结果中不难看出计算出的样本量完全能够满足power的要求。因此,我们给出的样本量公式可以为临床试验的设计提供参考。我们利用一个乳腺癌临床试验的例子说明了检验的方法及样本量计算公式的可行性。第二个问题,在临床试验得到的配对或相关生存数据中,如果不满足比例风险假定,我们对随访过程中任意时间点上的风险比可能更感兴趣。针对在比例风险假定不满足时,如何计算任意时间点上的风险比的问题,我们给出了不需要比例风险模型假定的非参数的核方法来估计任意时间点上的风险比的及其置信区间。目前为止,这种核方法仅被用于当两个治疗组的生存时间独立的情形之下。我们将非参数核方法用于估计边际风险率,并将近年来提出的MOVER方法用在构造两个相关的生存数据在随访过程中任意时间点的边际风险率之比的置信区间中。在构造单个边际风险率的置信区间时,我们分别采用了"delta"法与"transform"法。通过随机模拟,我们可以看出当样本量很大并且时间点不是很靠近边界时,两种方法的效果都很好。我们也将这两种方法运用于实际的临床试验数据中,用以说明其可行性。第三个问题,在平行对照的临床试验中,将病人随机分为两组,新治疗组与对照治疗组,在中期分析中,发现所试验的治疗比对照治疗好,使得原来设计的试验不得不终止,并允许对照治疗的病人选择开始接受新治疗。此时,由于随访时间较短,得到的仅是治疗的短期疗效指标的结果,为了观察新疗法的长期效果,因此继续对病人进行观察。我们希望能得到如果没有发生治疗转换,一直继续随访,最终如何评价两种治疗效果的差异。我们比较了利用反事实风险率构造的SCC法、利用逆概率删失加权的IPCW方法和简单ITT法、在ITT集上调整了与治疗转换和结局都有关的协变量的COX回归相比的效果。通过随机模拟,我们发现SCC法与IPCW法都能校正治疗转换效应,并且与ITT和COX相比更加能够反映两种治疗的真实差异。于是,虽然SFDA要求使用ITT法进行分析,但在遇到病人发生治疗转换时,SCC与IPCW法可依作为ITT法的补充。我们也将SCC法与IPCW法用于解决加拿大国家癌症研究所的临床试验组(NCIC CTG)负责的一个随机双盲安慰剂对照的肿瘤临床试验中遇到的问题。更多还原

【Abstract】 According to the World Health Organization’s official website, cancer is a leading cause of death worldwide. In 2008, cancer accounted for 7.6 million deaths, which is around 13% of all deaths. Deaths from cancer worldwide are projected to continue to rise to over 11 million in 2030. Recently, surveys revealed that the number of newly developed cancer patients is increasing by 3% to 5% per year in China. In urban dis-tricts, the top three cancers with the highest incidence rate are lung cancer, stomach cancer and breast cancer, whereas in the rural areas, stomach cancer, esophageal can-cer and liver cancer are the three with the highest incidence rate.The safest and most effective way to find the evidence of anti-cancer treatment is through clinical trials. The main purpose of clinical trials is to evaluate the efficacy and safety of anti-cancer treatment. Cancer clinical trial as a special branch of clinical trials has its own features:first, the follow-up time is usually relatively long; second, the most common endpoints are time-to-event; in addition, the major analyzing me-thod of cancer clinical trials is survival analysis.Because of the features above, during the design and the data analysis of cancer clinical trials, one might come across all sorts of problems. Until now, there still are some of these problems being unsolved. The main purpose of this thesis is to study on three problems we met in cancer clinical trials.First, in exploratory research, to make a preliminary evaluation of the efficacy of the new treatment, researchers will try the new treatment on those cancer patients with cancer progression (usually defined as 25% increase of the tumor size) under the standard treatment. To control the research cost, how to evaluate the efficacy of the new treatment by using the performance of treatment before the new treatment as a control to be compared with the performance of the new treatment.When proportional hazard assumption holds, we derived a class of non-parametric weighted rank tests for paired survival times to evaluate the efficacy of the new treatment. Monte-Carlo simulations were performed to evaluate the per-formance of the proposed tests and the results were satisfactory. The proposed me- thods are not only limited to the scenario above, they could be used to verify any non-inferiority hypothesis for paired or correlated survival data.In addition, we also derived the sample size formula under no-censoring or cen-soring situation for these tests. Through simulations, the statistical powers were gen-erated under the calculated sample size. The results showed the sample size could gain enough statistical power. So, the sample size formula could be used as a refer-ence to design clinical trials. As an example, a phaseⅡclinical trial of metastatic breast cancer was taken to demonstrate the proposed tests and to illustrate the feasibil-ity of the sample size formula.Second, if the proportional hazard assumption does not be satisfied by the paired or correlated survival data from the clinical trials, the inferences on hazard ratio at each specific time point of follow-up may be of more interest. How to detect the dif-ference between the two treatment groups? We used the recently proposed non-parametric kernel method, which doesn’t need the proportional harzard assump-tion, to estimate the time-dependent hazard ratio and its confidence interval. Until now, this kernel method had only been used in the situation when the survival times of two treatment groups are independent.We used the non-parametric kernel method to estimate each marginal hazard rate and the variance estimates recovery (MOVER) method to construct the confidence intervals of time-dependent hazard ratios based on confidence limits of each hazard rate. Two methods were proposed when constructing confidence limits for each mar-ginal hazard rate:"delta" method and "transform" method. Simulation results showed that both methods can perform well, especially when sample size is large and the time points are not too close to the boundary. The proposed methods were also applied to datasets from real clinical trials to show their feasibility.Third, in parallel controlled clinical trials, we randomly assign patients into two groups, the new treatment group and the control group. After the interim analysis, if a short-term endpoint of the new treatment, such as disease free survival, is significant-ly better than the controlled treatment, the trial has to be stopped, and those patients in the control arm are offered a chance to switch to the new treatment. The conclusion we drawn after interim analysis is only the short-term effect of the treatment. To ob-serve the long-term effect of the treatment, we continue to follow up those patients. Our concern is if those patients who crossed over to the new treatment group who did not switch and were continued to be followed up, how to detect the real difference between the two treatment groups.We compared the SCC method which based on counterfactual hazard rates and the IPCW method which uses the inverse probability of censoring rate to weigh the patients to simple ITT method and the COX regression adjusting the covariate both related to switch probability and outcome based on ITT set. Through simulations, we found the SCC and the IPCW method can both adjust the treatment crossover effect and outperform the ITT and the COX method in revealing the real difference between the two treatments. So, although the State Food and Drug Administration recommends the ITT method in the clinical trial analysis, we still could use the results from the SCC and the IPCW method as supplemental analysis when there is treatment crossov-er. Real data from a randomized double-blind placebo-controlled clinical trial from the National Cancer Institute of Canada Clinical Trial Group is used to support the results.更多还原