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

OBJECTIVE: To evaluate the impact of non-blinded outcome assessment on estimated treatment effects in randomised clinical trials with binary outcomes. DESIGN: Systematic review of trials with both blinded and non-blinded assessment of the same binary outcome. For each trial we calculated the ratio of the odds ratios--the odds ratio from non-blinded assessments relative to the corresponding odds ratio from blinded assessments. A ratio of odds ratios <1 indicated that non-blinded assessors generated more optimistic effect estimates than blinded assessors. We pooled the individual ratios of odds ratios with inverse variance random effects meta-analysis and explored reasons for variation in ratios of odds ratios with meta-regression. We also analysed rates of agreement between blinded and non-blinded assessors and calculated the number of patients needed to be reclassified to neutralise any bias. DATA SOURCES: PubMed, Embase, PsycINFO, CINAHL, Cochrane Central Register of Controlled Trials, HighWire Press, and Google Scholar. ELIGIBILITY CRITERIA FOR SELECTING STUDIES: Randomised clinical trials with blinded and non-blinded assessment of the same binary outcome. RESULTS: We included 21 trials in the main analysis (with 4391 patients); eight trials provided individual patient data. Outcomes in most trials were subjective--for example, qualitative assessment of the patient's function. The ratio of the odds ratios ranged from 0.02 to 14.4. The pooled ratio of odds ratios was 0.64 (95% confidence interval 0.43 to 0.96), indicating an average exaggeration of the non-blinded odds ratio by 36%. We found no significant association between low ratios of odds ratios and scores for outcome subjectivity (P=0.27); non-blinded assessor's overall involvement in the trial (P=0.60); or outcome vulnerability to non-blinded patients (P=0.52). Blinded and non-blinded assessors agreed in a median of 78% of assessments (interquartile range 64-90%) in the 12 trials with available data. The exaggeration of treatment effects associated with non-blinded assessors was induced by the misclassification of a median of 3% of the assessed patients per trial (1-7%). CONCLUSIONS: On average, non-blinded assessors of subjective binary outcomes generated substantially biased effect estimates in randomised clinical trials, exaggerating odds ratios by 36%. This bias was compatible with a high rate of agreement between blinded and non-blinded outcome assessors and driven by the misclassification of few patients.