Allocated but not treated: the silent 16% – Authors' reply

Randomised controlled trials generally seek to estimate the effect of assignment to an intervention (ie, the intention-to-treat effect) rather than the effect of adhering to an intervention (ie, the per-protocol effect). Intention-to-treat analysis is the only type of analysis to guarantee that the patient groups being compared remain balanced with respect to their baseline characteristics (differing only due to chance), thus allowing valid randomised comparisons of these groups.¹ Of course, in any trial, some patients assigned to an intervention might not receive it (ie, drop out), whereas others who are not assigned to the intervention might subsequently receive the intervention (ie, drop in). Because both of these effects reduce the difference between randomised groups in exposure to the randomised intervention, intention-to-treat analyses tend to underestimate the effect of full adherence.² Although it can be tempting in such circumstances to try to estimate the effect of adhering to the intervention (perhaps through analyses restricted to individuals who adhered to their assigned intervention), such analyses are problematic, involve strong assumptions, and can be seriously biased.² It is with this background that we consider Todd C Lee and colleagues’ request.

In the RECOVERY trial 4116 patients with COVID-19 who were admitted to hospital and had hypoxia and systemic inflammation were randomly assigned to either receive tocilizumab plus usual care or usual care alone.³ Allocation to tocilizumab resulted in a 15% proportional reduction in 28-day mortality (31% among those allocated tocilizumab vs 35% among those allocated usual care, rate ratio 0·85, 95% CI 0·76–0·94; p=0·0028), with consistent proportional mortality reductions in all prespecified subgroups of patients, including those receiving systemic corticosteroids.³ This significant reduction in mortality among all patients who were randomly assigned was detected despite about one in six patients allocated tocilizumab not receiving it (based on the 1964 (97%) of 2022 patients where information on treatments received was known). Mortality in this 16% of patients was somewhat higher than in other patients allocated to receive tocilizumab, but direct comparison of people who did not receive tocilizumab after allocation with any other subgroup (as suggested by Lee and colleagues) would not be informative about the effect of tocilizumab. This analysis could be misleading because there is no way of reliably knowing which participants allocated to usual care would also not have received tocilizumab had the randomisation process happened to allocate them to the tocilizumab group.

From a purely observational perspective, however, it is perhaps worth noting that the baseline characteristics of the 317 patients allocated to tocilizumab who did not receive it were broadly similar to the 1647 patients allocated to tocilizumab who did (appendix).