Yamaguchi T, Ohashi Y, Matsuyama Y
Department of Biostatistics/Epidemiology and Preventive Health Sciences, School of Health Sciences and Nursing, University of Tokyo, Hongo 7-3-1, Bunkyo-ku, Tokyo 113-0033, Japan.
Stat Methods Med Res. 2002 Jun;11(3):221-36. doi: 10.1191/0962280202sm284ra.
In randomized clinical trials comparing treatment effects on diseases such as cancer, a multicentre trial is usually conducted to accrue the required number of patients within a reasonable period of time. The fundamental point of conducting a multicentre trial is that all participating investigators must agree to follow the common study protocol. However, even with every attempt having been made to standardize the methods for diagnosing severity of disease and evaluating response to treatment, for example, they might be applied differently at different centres, and these may vary from comprehensive cancer centres to university hospitals to community hospitals. Therefore, in multicentre trials there is likely to be some degree of variation (heterogeneity) among centres in both the baseline risks and the treatment effects. While we estimate the overall treatment effect using a summary measure such as hazard ratio and usually interpret it as an average treatment effect over the centre, it is necessary to examine the homogeneity of the observed treatment effects across centres, that is, treatment-by-centre interaction. If the data are reasonably consistent with homogeneity of the observed treatment effects across centres, a single summary measure is adequate to describe the trial results and those results will contribute to the scientific generalization, the process of synthesizing knowledge from observations. On the other hand, if heterogeneity of treatment effects is found, we should carefully interpret the trial results and investigate the reason why the variation is seen. In the analyses of multicentre trials, a random effects approach is often used to model the centre effects. In this article, we focus on the proportional hazards models with random effects to examine centre variation in the treatment effects as well as the baseline risks, and review the parameter estimation procedures, frequentist approach-penalized maximum likelihood method--and Bayesian approach--Gibbs sampling method. We also briefly review the models for bivariate responses. We present a few real data examples from the biometrical literature to highlight the issues.
在比较癌症等疾病治疗效果的随机临床试验中，通常会进行多中心试验，以便在合理时间内招募到所需数量的患者。进行多中心试验的基本要点是，所有参与的研究者必须同意遵循共同的研究方案。然而，即使已尽一切努力使疾病严重程度的诊断方法和治疗反应的评估方法标准化，例如，这些方法在不同中心的应用可能会有所不同，而且从综合癌症中心到大学医院再到社区医院，这些方法可能存在差异。因此，在多中心试验中，各中心在基线风险和治疗效果方面可能会存在一定程度的差异（异质性）。虽然我们使用风险比等汇总指标来估计总体治疗效果，并通常将其解释为各中心的平均治疗效果，但有必要检查各中心观察到的治疗效果的同质性，即治疗与中心的交互作用。如果数据与各中心观察到的治疗效果的同质性合理一致，那么单一的汇总指标就足以描述试验结果，并且这些结果将有助于科学推广，即从观察中综合知识的过程。另一方面，如果发现治疗效果存在异质性，我们应该仔细解释试验结果，并调查出现差异的原因。在多中心试验的分析中，通常使用随机效应方法来模拟中心效应。在本文中，我们重点关注具有随机效应的比例风险模型，以检查治疗效果以及基线风险方面的中心差异，并回顾参数估计程序，即频率学派方法——惩罚最大似然法——和贝叶斯方法——吉布斯抽样法。我们还简要回顾了双变量反应模型。我们给出一些生物统计学文献中的实际数据示例，以突出这些问题。
