What should be done and what should be avoided when comparing two treatments?

The gold-standard approach to compare two treatments, or more generally to assess the efficacy and safety of an intervention versus a comparator (which could be no intervention at all, a placebo or sham intervention, or another intervention), is a well-designed and well-conducted randomized clinical trial (RCT). There are issues on how to design, conduct and analyze RCTs, from a scientific, ethical or regulatory point-of-view. Those have been summarized in many textbooks [1,2] or documents from regulators, such as guidelines issued by the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH) (https://www.ich.org/page/search-index-ich-guidelines). Key elements of the validity of an RCT can be classified into internal validity and external validity components [3]. Internal validity relates to selection bias (whether allocation to each group is biased), performance bias (whether other interventions or care differ between groups, apart from the intervention being evaluated), detection bias (whether the outcome is measured the same way in both groups) and attrition bias (whether protocol deviations and loss to follow-up occurred or were handled differently in both groups). External validity refers to the ability of the trial results to be generalized to the target population, and relates to patients’ characteristics, treatment regimens, settings in which the trial was conducted (e.g. experience or specialization of centers and care providers), and the definition of outcomes. Assigning the treatment group at random as in RCTs eliminates selection bias, and ensures that the two groups have—on average—the same characteristics, measured or unmeasured. Another key feature is the experimental design. The prospective nature of clinical trials and the requirement to set up a precise protocol allow to limit the other threats to internal validity (sources of bias), for instance using blinding to limit performance and detection bias, specifying precisely how outcomes should be measured, implementing actions to limit attrition, etc. Despite their clear methodological assets, RCTs also have limitations. They are generally very expensive, and require a long time to obtain results, especially when clinical outcomes are mid- or long-term outcomes or rare events. There are also situations where randomization is felt unethical, or will not be accepted by participants or physicians. This can be the case when one wants to compare hematopoietic cell transplant (HCT) to non-HCT; most often there is no randomization, and patients receive HCT or not according to the availability of a donor. It is also not possible, from a financial and logistical point-of-view, to conduct RCTs for pairwise comparisons of all treatments available for a condition, or in all relevant subgroups of patients. Last, there has been evidence that participants to clinical trials generally differ from the target population. This is caused partly by stringent eligibility criteria (e.g. excluding patients vulnerable to adverse effects such as elderly patients or those with comorbidities) [4,5], and also by a selection of participant even within eligibility criteria [6]. Patients in RCTs are also often treated in more experienced centers, and with a better follow-up. This leads to impair the external validity of trials [7,8].

Observational data may provide valuable information to compare treatments [9,10]. In observational studies, the treatment assignment is not controlled by investigators, but simply follows the individual decisions of physicians and patients. Different sources of data can be used. Since those data reflect the usual care practice, observational data are also often referred as real-world evidence. The lack of randomization in observational studies may result in differences in patients’ characteristics (observed or unobserved) between the treatment groups. Such differences may bias the estimate of treatment effect. This is called confounding or indication bias. There are however several statistical approaches allowing to estimate treatment effects from observational data. In this article, we review the most popular, and provide some insights to understand how, and under which conditions, they should be used.

留言 (0)

沒有登入
gif