CONDITIONAL PROBABILITY OF SIGNIFICANCE FOR EARLY STOPPING IN FAVOR OF H₀

ABSTRACT

Early stopping is often desirable in a large clinical trial if data accumulated during the trial has little evidence to support the alternative hypothesis or the null result seems inevitable. This motivates, for example, the conditional power and predictive power approaches for early stopping in favor of the null. We investigate using conditional probability of significance (CPS), which measures the strength of evidence against the null in the data, for early stopping. Numerical results show that when the number of looks is small, stochastic curtailment based on conditional power, predictive power and CPS are similar after controlling for the significance level and power. However, when the number of looks is large, the conditional power approach projecting the current trend becomes very different from other methods. The CPS method, despite having a simple functional form of stopping times, is competitive to other existing methods over a range of fixed power values.

Keywords:

• Conditional power
• Fixed sample test
• Predictive power
• Repeated significance test
• Stochastic curtailment
• Stopping time

ACKNOWLEDGMENTS

We wish to thank the referee for comments leading more insights into this new approach. This research has been partially supported by NIH AI24643 for YGW.