A Highly Powered Large Outcomes Randomised Controlled Trial Will Usually, but Not Always, Have Balanced Groups at Baseline.

OBJECTIVE

The importance of randomised trials for discerning an intervention's effect rests on randomisation's ability to create groups with (on average) equal hazards of future outcomes. The frequency with which this does not occur in well-designed large trials with a high event rate was studied.

RESEARCH DESIGN AND

METHODS

Data from 8215 participants in the Outcomes Reduction with an Initial Glargine Intervention (ORIGIN) trial's biomarker study were used to simulate multiple trials. Participants were repeatedly randomised to two groups. After each randomisation, the hazard of death in people assigned to one versus the other was calculated. A set of 1000 randomisations was repeated 1000 times, and the number of pairs for which the hazard ratio was < 0.85 or > 1.18 was tabulated within each set of 1000. Whether such imbalances in hazard could be detected or mitigated by adjusting for baseline covariates was explored using adjusted Cox models.

RESULTS

In a median of 3 (IQR 2, 4) of 1000 randomisations (0.3%) the hazard ratio for death of assignment to one group versus the other was < 0.85 or > 1.18 by chance alone. This imbalance was not apparent from baseline data and could not be fully mitigated by adjustment for baseline independent predictors of death using data from four randomisations that yielded imbalanced hazards.

CONCLUSION

One out of every 300-400 large, two-arm outcomes trials have imbalances in the baseline hazard of the primary outcome that are large enough to misrepresent the true effect of the intervention.