Are We Really That Blind?

ABSTRACT

In double-blind randomized clinical trials, it is common practice to randomize patients using randomly permuted blocks. In this article, it is demonstrated that before unblinding statistical inference of the treatment effects can be conducted, yielding consistent and rather precise estimates even in the presence of an additive block effect. With an even greater precision, the within-group standard deviation on which power calculation are usually based can be inferred from blinded data. The use of blocks of random lengths as suggested by ICH-E9 in the (unlikely) case that previous treatment allocation can be guessed by strong pharmacological effects, merely complicates the analysis but blinded inference can still be conducted without much extra loss of information. On the one hand, one might argue that this possibility of blinded inference takes away the need of conducting interim analyses for administrative or business reasons or for sample size reestimation. On the other hand, however, it most probably will have a disputable, positive or negative effect on the conduct of the remainder of the trial. If regulators and the pharmaceutical world at large want to avoid this possibility, then other unrestricted, biased coin, or more general dynamic allocation randomization procedures may be less controversial alternatives. It at least provides another strong argument in favor of using large blocks as the precision of blinded inference decreases with increasing block lengths.

If blinded inferences are deemed a useful replacement of interim analyses in nonpivotal trials, then further guidelines will be needed on consequent decision-making aspects.

Key Words:

• Blinded inference
• Maximum likelihood
• Mixture distributions
• Normal distribution
• Randomly permuted blocks.

ACKNOWLEDGMENTS

The author wants to thank his colleagues Ton de Haan, Kit Roes, and Martin Struijs, as well as the reviewers, for useful suggestions that considerably improved the article.

Notes

Note: Per scenario, cumulative simulated data were used. In scenarios with small effects and limited data, the likelihood can be unimodal with a maximum at δ = 0, instead of bimodal with a local minimum at δ = 0. ML estimates were obtained using SAS-IML macro NLPNRA.

Note: Per scenario, cumulative simulated data were used. ML estimates were obtained using SAS-IML macro NLPNRA.

Note: Per scenario, cumulative simulated data were used. ML estimates were obtained using SAS-IML macro's NLPNRA and NLPFDD.