Which patients to include in the analysis?

Intention-to-treat (ITT) analysis has become the accepted approach to analyzing data in clinical trials.1 Nevertheless, considerable uncertainty exists among researchers and users of the medical literature about what an ITT analysis actually means.2 This editorial will describe the ITT principle and reflect on situations in which adherence to the principle does not preclude postrandomization exclusions,3 a topic of considerable debate in studies of blood transfusion as evidenced by the recent letters from Vamvakas4 and Blumberg.5 Clinical trials that randomize patients to an experimental intervention or a control group do so to achieve prognostic balance—in the absence of intervention effects, patients in both groups will experience a similar rate of adverse events. Figure 1 shows a randomized trial in which 200 patients have been allocated to a surgical intervention or a medical control group. Forty percent of the medical patients suffer an adverse outcome compared to 24 percent in the surgical group (a 40% reduction in relative risk). Assuming no other methodologic problems, and ignoring the play of chance, this estimate reflects the true benefit of the intervention. The ideal study: 100% compliance, adherence, and follow-up. The true benefit of the surgical intervention is 40 percent RRR. R = randomization; RRR = relative risk reduction. Many things may happen, however, between randomization and outcome assessment. Patients might withdraw from the study, deteriorate and cross over to the other arm, fail to adhere to the allocated treatment, or suffer an event before receiving the intervention. Such circumstances demand considered decisions about how to treat events in the analysis. How should a researcher count patients who crossed over from one intervention group to the alternative group. Intuitively one might say those patients should be counted according to the treatment they actually received (if one did, one would be conducting an "as-treated" analysis). Consider a study about a surgical technique that is tested against a medical intervention. Medical patients with an unfavorable prognosis deteriorated and crossed over to the surgical intervention (Fig. 2). As a result, the prognosis across all patients in the control group has improved (event rate dropped to approx. 30%) while the underlying prognosis of the surgical patients worsened. This is a typical situation of unequal groups and the resulting imbalance of an as-treated analysis makes surgery look much less effective than it truly is (Fig. 1). Its true relative risk reduction is diminished from 40 percent to an apparent effect of 10 percent. As-treated analysis: Patients with a worse prognosis crossed from the medical to the surgical intervention and responded to surgery less than average (*). The apparent benefit of the intervention is 10 percent in RRR. R = randomization; RRR = relative risk reduction. What would happen in these situations if our top priority were to maintain the prognostic balance between the groups? To achieve this, the researcher would need to perform an ITT analysis. There, all patients are counted in the groups to which they had been randomized, irrespective of the treatment they actually received (Fig. 3). Now, the surgical group maintains the true effect, a 24 percent event rate, while the medical control group looks spuriously better as the beneficial effect of surgery on the patients who had crossed is counted in favor of the medical intervention. In balance, surgery looks beneficial, albeit the estimate of the treatment effect is spuriously smaller (some call it "more conservative") than it would be if everyone received the treatment to which they were allocated (Fig. 1). ITT analysis: 50% of the medical patients crossed over to surgical treatment, but are counted in the medical group. The apparent benefit of the surgical intervention is 25 percent in RRR. R = randomization; RRR = relative risk reduction. Imagine a new drug being tested against placebo. Let's assume the drug is actually ineffective but side effects cause a poor adherence by 40 percent of the patients in the experimental group. Intuitively one might want to exclude such patients, counting only those who followed the protocol (an as-treated analysis; Fig. 4). Patients with poor compliance may have a poorer prognosis than patients who comply with their treatments (indeed, in a wide variety of situations, they do).6 If that is the case, removing nonadherent patients from the treatment group will destroy the prognostic balance that randomization achieved and leave the remaining patients in the treatment group with a better prognosis. Under these circumstances, it is impossible to tell whether an apparent treatment benefit is due to a true treatment effect or to the bias created by removing the high risk patients from the intervention arm. As-treated analysis: Patients in the intervention group poorly adhere to the drug or drop out. The apparent effect of the intervention is a RRR of 20 percent when in truth there is none. R = randomization; RRR = relative risk reduction. What options has a researcher when patients suffer an event before they actually get the intervention? Imagine a trial on patients with an increased risk of stroke where patients are allocated to a medical intervention or a combination of medical and surgical intervention (Fig. 5). Assume that the surgical intervention has no impact on the course of the disease. Some patients (here 10/100) will suffer an event before the operation. Intuitively, one might regard it as unfair to apply an ITT analysis that counts events against an operation that the patients had not received. Performing an as-treated analysis and removing those patients with early stroke (i.e., with a worse prognosis) from the surgical group, however, but retaining comparable patients in the medical group, would make the surgical intervention look spuriously superior to the medical intervention. Again, by removing poor prognosis patients from one group but not the other the investigators have biased their results. Following the ITT principle and counting patient events against the groups to which they have been randomized maintains the prognostic balance and, in this scenario, correctly identifies the absence of benefit. As-treated analysis: Patients suffer an event before they actually receive surgery. The apparent effect of the surgical intervention is a RRR of 45 percent when in truth there is none. R = randomization; RRR = relative risk reduction. Following the ITT principle is not fully satisfactory. Clinicians and patients are interested in how a treatment works when patients receive it, not when they don't. To the extent that patients don't receive an effective intervention, the apparent effect from an analysis that adheres to the ITT principle will underestimate the effect. The unsatisfactory choice is between following the ITT principle and finding out about the true effect of the intervention as administered (however, unsatisfactorily) or an as-treated analysis that is liable to provide a biased estimate of the real effect. Because it leads to an unbiased estimate, researchers and methodologists have reached a consensus that observing the ITT principle is the superior approach. There are situations, however, in which one can exclude randomized patients from the analysis without introducing bias.3 Such situations go to the heart of the argument between Vamvakas and Blumberg concerning trials addressing the impact of leukoreduction of allogeneic blood transfusions in inducing postoperative infections. The peri-and intraoperative administration of the blood transfusion required the randomization of the patients even before the need for the intervention (the transfusion) could be established. As expected, a substantial proportion of patients with low to moderate intraoperative blood loss did not require transfusion, while those with more severe blood loss and a worse prognosis qualified for the transfusion (Fig. 6). Postrandomization exclusion: after randomization, 40 percent of patients in both groups do not require a transfusion and can be dropped from the analysis. R = randomization; RRR = relative risk reduction. Omission of patients from the analysis will not introduce bias if the reason for omission is independent of, and could not be influenced by, the arm to which the patient was randomized. If this is the case, one is in effect removing patients at random; that being the case, groups that were prognostically balanced by random allocation will remain balanced. Thus, as long as the decision to transfuse is uninfluenced by the patient's allocation to leukoreduced or nonleukoreduced allogeneic blood transfusions, one can omit untransfused patients without introducing bias (Fig. 6). Because it is hard to imagine that the transfusion decision could have been influenced by what transfusion would have been given if it were needed, the omission of untransfused patients seems well justified. Indeed, their omission is the superior analytic approach, because their inclusion adds random error and thus diminishes the power of the analysis. In summary, a study without dropout, crossover, poor adherence, or early events would provide the starting point for the ideal analysis that revealed a true estimate of the treatment effect. There, both ITT and the as-treated analysis would include the same number of patients and events. As the number of unforeseen incidents increases, the true effect of the intervention will become increasingly blurred, independent of the type of analysis, as treated or intention to treat. Because it leads to an unbiased estimate, researchers and methodologists have reached a consensus that observing the ITT principle is the superior approach. In some transfusion studies, however, it is justifiable to remove patients from analysis after they have been randomized.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.