To the Editor:
We read with interest the randomized controlled trial in the journal by Shin et al. comparing postoperative delirium rates in patients undergoing major joint replacement surgery under spinal anesthesia using propofol versus dexmedetomidine sedation. Not only does this trial inform on which agent may result in decreased postoperative delirium, but it also illuminates the imperfections and vulnerabilities in Bayesian logic that underpin much of modern-day medical research. A power analysis was performed that assumed the incidence of delirium to be 2.3% in the dexmedetomidine group and 6.8% in the propofol group. A total of 772 subjects were required (336 per group) to find a difference (if present) 80% of the time (power = 0.8) with less than a 5% chance (P < 0.05) that the difference would be by chance. Eleven (3.0%) of the subjects in the dexmedetomidine group and 24 (6.6%) of those in the propofol group experienced delirium (P = 0.036) with a fragility index of 1.0.
The fragility index represents the number of subjects that would need to be moved from one outcome group to another to convert trial results to statistical nonsignificance (P > 0.05). A fragility index of 1.0 indicates that the addition of one subject to the dexmedetomidine group with delirium or removal of one subject from the propofol group with delirium would result in a P value exceeding 0.05, rendering the results statistically insignificant (table 1). Sources of this type of error would include the use of a range of propofol target effect site concentrations (1 to 2 μg/ml); the use of Confusion Assessment Method and Interviews of Caregivers for evidence of delirium that include the presence of confusion, agitation, or sedation; the characterization of level of intraoperative sedation by the Modified Observers Assessment of Alertness and Sedation scale and the loss of subjects (49) who failed to complete the protocol and thus were excluded from the results. Due to the necessarily inherent subjectivity of these metrics and the high number of exclusions, it can be appreciated how fragile statistical significance is if just one subject was misassigned.