Observer bias in randomized clinical trials with binary outcomes: Systematic review of trials with both blinded and non-blinded outcome assessors

Nordic Cochrane Centre, Rigshospitalet Department 3343, Blegdamsvej 9, 2100 Copenhagen Ø, Denmark.
BMJ (online) (Impact Factor: 16.38). 02/2012; 344(3):e1119. DOI: 10.1093/ije/dyt270
Source: PubMed

ABSTRACT Background: We wanted to evaluate the impact of nonblinded outcome assessors on estimated treatment effects in time-to-event trials. Methods: Systematic review of randomized clinical trials with both blinded and nonblinded assessors of the same time-to-event outcome. Two authors agreed on inclusion of trials and outcomes. We compared hazard ratios based on nonblinded and blinded assessments. A ratio of hazard ratios (RHR) < 1 indicated that nonblinded assessors generated more optimistic effect estimates. We pooled RHRs with inverse variance random-effects meta-analysis. Results: We included 18 trials. Eleven trials (1969 patients) with subjective outcomes provided hazard ratios, RHR 0.88 (0.69 to 1.12), (I-2 = 44%, P = 0.06), but unconditional pooling was problematic because of qualitative heterogeneity. Four atypical cytomegalovirus retinitis trials compared experimental oral administration with control intravenous administration of the same drug, resulting in bias favouring the control intervention, RHR 1.33 (0.98 to 1.82). Seven trials of cytomegalovirus retinitis, tibial fracture and multiple sclerosis compared experimental interventions with standard control interventions, e.g. placebo, no-treatment or active control, resulting in bias favouring the experimental intervention, RHR 0.73 (0.57 to 0.93), indicating an average exaggeration of nonblinded hazard ratios by 27% (7% to 43%). Conclusions: Lack of blinded outcome assessors in randomized trials with subjective time-to-event outcomes causes high risk of observer bias. Nonblinded outcome assessors typically favour the experimental intervention, exaggerating the hazard ratio by an average of approximately 27%; but in special situations, nonblinded outcome assessors favour control interventions, inducing a comparable degree of observer bias in the reversed direction.

Download full-text


Available from: Ann Sofia Skou Thomsen, May 19, 2015
  • Source
    • "When intraand inter-individual variability are low, and sample size relatively small (under 100 in the Gensburger example cited below), blinding may have little effect on study outcome (Gensburger et al., 2012). However, it has been shown that blinding can affect outcome, even with binary outcomes as determined from a meta-analysis of a large population sample (over 4000 in the example cited; Hróbjartsson et al., 2012). There is, additionally, the famous case of Benveniste who published a paper in Nature (which was never retracted) that identified an apparent activity of antibodies at a 'concentration' below that necessary to provide a single molecule in the solutions used in the experiment (Davenas et al., 1988). "
    [Show abstract] [Hide abstract]
    ABSTRACT: This Editorial is part of a series. To view the other Editorials in this series, visit:;; and © 2015 The British Pharmacological Society.
    British Journal of Pharmacology 07/2015; 172(14):3461-3471. DOI:10.1111/bph.12856 · 4.99 Impact Factor
  • Source
    • "We summarized the impact of nonblinded outcome assessment as the ratio of odds ratios (ROR 5 OR nonblind /OR blind ). The standard error of ROR was calculated as the square root of the sum of the variances of the two ORs [16] [23]. An ROR !1 indicates that nonblinded assessors exaggerate the benefit of the intervention. "
    [Show abstract] [Hide abstract]
    ABSTRACT: Objectives: To examine the impact of not blinding outcome assessors on estimates of intervention effects in animal experiments modeling human clinical conditions. Study Design and Setting: We searched PubMed, Biosis, Google Scholar, and HighWire Press and included animal model experiments with both blinded and nonblinded outcome assessors. For each experiment, we calculated the ratio of odds ratios (ROR), that is, the odds ratio (OR) from nonblinded assessments relative to the corresponding OR from blinded assessments. We standardized the ORs according to the experimental hypothesis, such that an ROR <1 indicates that nonblinded assessor exaggerated intervention effect, that is, exaggerated benefit in experiments investigating possible benefit or exaggerated harm in experiments investigating possible harm. We pooled RORs with inverse variance random-effects meta-analysis. Results: We included 10 (2,450 animals) experiments in the main meta-analysis. Outcomes were subjective in most experiments. The pooled ROR was 0.41 (95% confidence interval [CI], 0.20, 0.82; I-2 = 75%; P < 0.001), indicating an average exaggeration of the nonblinded ORs by 59%. The heterogeneity was quantitative and caused by three pesticides experiments with very large observer bias, pooled ROR was 0.20 (95% CI, 0.07, 0.59) in contrast to the pooled ROR in the other seven experiments, 0.82 (95% CI, 0.57, 1.17). Conclusion: Lack of blinding of outcome assessors in animal model experiments with subjective outcomes implies a considerable risk of observer bias.
    Journal of Clinical Epidemiology 06/2014; 67(9). DOI:10.1016/j.jclinepi.2014.04.008 · 5.48 Impact Factor
  • Source
    • "Le recours à un évaluateur du critère de jugement aveugle est toujours utile (évaluateur qui se distingue du thérapeute). Si cet évaluateur n'est pas aveugle, le biais induit dans la mesure du critère de jugement est désormais bien documenté [8]. Il faudra donc éviter les contacts entre le thérapeute non aveugle et cet évaluateur. "
    Gynécologie Obstétrique & Fertilité 02/2013; 41(2):144–146. DOI:10.1016/j.gyobfe.2012.12.005 · 0.58 Impact Factor
Show more