Minimizing bias in randomized trials: the importance of blinding.

Abstract

IN 1992, HANSSON ET AL 1 PROPOSED A NOVEL DESIGN, THE prospective randomized open trial with blinded endpoint assessments. The lack of blinding of investigators and patients simplified the conduct of the trial, which would become more similar to routine medical practice than the blinded design. The use of blinding for the adjudication of outcomes would preserve the benefits of a fully blinded trial. A number of trials have used this design to evaluate antihypertensive agents and more recently antidiabetic agents. These trials were thought to produce valid and, perhaps, more generalizable results. Studies of trial results for specific classes of treatments support the importance of investigator blinding. In a crosssectional analysis of 192 trials that compared a statin drug with another statin or a nonstatin drug, the studies that described adequate blinding were much less likely to report findings that favored the test drug than studies that did not describe adequate blinding (odds ratio, 0.41; 95% confidence interval [CI], 0.23-0.73). Hence, it is important to evaluate whether the absence of investigator blinding may have influenced the results of specific important trials. The Rosiglitazone Evaluated for Cardiac Outcomes and Regulation of Glycemia in Diabetes (RECORD) trial, for instance, was an open-label trial that enrolled patients with type 2 diabetes who were taking either metformin or a sulfonylurea and randomized them to receive either rosiglitazone or the combination of metformin plus a sulfonylurea. The trial met the noninferiority criterion for the primary outcome of cardiovascular hospitalization or death. Moreover, rosiglitazone was not significantly associated with the risk of myocardial infarction (MI) (hazard ratio, 1.14; 95% CI, 0.80-1.63), a concern that had been raised by an earlier meta-analysis. However, event rates in the control group were unexpectedly low in this open trial with blinded end-point assessment. The independent review by the US Food and Drug Administration (FDA) of case-report forms from the RECORD trial by Marciniak has provided empirical estimates of the potential bias associated with an open-label design when, in an industry-sponsored trial setting, investigators were aware of the treatment assignment. Of the 2220 patients randomized to rosiglitazone, case-report forms were reviewed for 278 (12.5%). Of the 2227 patients randomized to control, case-report forms were reviewed for 271 (12.2%). The review included a random sample of 100 case-report forms plus others that were targeted because they might have represented problems. In general, the blinding of the eventrelated material sent to the end-points committee was good. Among the 549 case-report forms, however, Marciniak identified 70 problem cases (12.8%). These 70 problem cases included events that were missed (n=14), ascertained events that were not referred for adjudication (n=8), cases that had insufficient information (n=18), and other adjudication issues (n=22). Based on the random sample, he estimated that if all the case-report forms had been re-reviewed, a total of 283 problem cases would have been identified. The appendix of Marciniak’s report (see summary, page 92) provides details about the problem case-report forms in the RECORD trial. For instance, one patient had an MI that was not referred for adjudication (case A). Hospitalization of another patient for pulmonary edema was not adjudicated (case B). Another patient with intracerebral hematoma presented with a seizure, but information about the intracerebral hematoma was deleted from the case-report form and replaced with epilepsy; the dossier was not sent for adjudication (case C). Information was not obtained about a long hospitalization for a severe stroke (case D), and the event was labeled as noncardiovascular. Errors occur in all studies, and insofar as they are nondifferential between treatment groups, they may not systematically bias hazard ratios or other summary measures although precision will be reduced. The FDA review, however, provides evidence of differential outcome ascertainment by treatment group in the RECORD trial (TABLE). Among the 549 case-report forms, the prevalence of problem cases was higher in the rosiglitazone group (16.2%) than in the control group (9.2%). Marciniak also classified the problem cases according to the group that had been favored by the error

DOI: 10.1001/jama.2010.1161
02040602011201220132014201520162017
Citations per Year

195 Citations

Semantic Scholar estimates that this publication has 195 citations based on the available data.

See our FAQ for additional information.

Cite this paper

@article{Psaty2010MinimizingBI, title={Minimizing bias in randomized trials: the importance of blinding.}, author={Bruce M . Psaty and Ross L. Prentice}, journal={JAMA}, year={2010}, volume={304 7}, pages={793-4} }