Imagine an RCT that studied patients with cerebrovascular disease. The trial compares administration of aspirin with or without an experimental surgical procedure. Assume that, although the investigators conducting the trial do not know it, the underlying true effect of the surgical procedure is 0; patients in the surgical arm of the study do neither better nor worse than those in the aspirin-only arm.
Of 100 patients randomized to surgery, 10 experience the primary outcome of the trial, a stroke, during the 1-month preoperative period, and their surgery is cancelled. Of the 90 patients who undergo surgery, 10 have a stroke in the subsequent year (Figure 11.4-1). What will happen to the patients in the control group? Because randomization, when sample sizes are large, will create groups with the same fate or destiny and because we have already established that the surgical procedure has no effect on outcome, we predict that 10 control group patients will have a stroke in the month after randomization and another 10 will have a stroke in the subsequent year.
The principle that dictates that we count events in all randomized patients, regardless of whether they received the intended intervention, is the intention-to-treat principle. When we apply the intention-to-treat principle in our study of cerebrovascular surgery for stroke, we find 20 events in each group and, therefore, no evidence of a positive treatment effect. However, if we use the logic that we should not count events in patients in the surgical group who did not receive surgery, the event rate in the experimental groups would be 10 of 90 (or 11%) compared with the 20% event rate in the control group—a reduction in relative risk of 45% instead of the true relative risk reduction (RRR) of 0. These data reveal how analyses restricted to patients who adhered to assigned treatment (sometimes referred to as per-protocol analyses, efficacy analyses, or explanatory analyses) can provide a misleading estimate of surgical therapy's effect.