Reproducibility of statistically significant phase III oncology trials: An In Silico meta-epidemiological analysis.

Purpose: The conventional assumption that P values ≤ 0.05 imply reproducible effects has come under recent criticism. This concern is particularly relevant in oncology, as phase III oncology trials, which directly inform practice, are usually not repeated. Using advanced modeling techniques, we investigated the relationship between P values and reproducibility in oncology.

Methods: We obtained the signal-to-noise ratio distribution in phase III oncology using outcomes from 632 two-arm superiority trials enrolling 496,219 patients. With this distribution, we estimated successful replication probability as the probability that a replicate trial, having the same design, effect size, and standard error, would have a two-sided P ≤ 0.05 and the same effect directionality as the original trial. We also estimated the following: the probability that the estimated effect had the same direction as the true effect (i.e., correct sign probability); the probability that the 95 % CI covered the true effect (i.e., coverage probability), and the ratio of the observed estimated effect to the true effect (i.e., exaggeration factor).

Results: The median exaggeration factor across all trials was 1.09 (IQR, 0.80-1.61). When P ≤ 0.05 in the original trial, mean correct sign probabilities were ≥ 97 % and mean coverage probabilities were between 93 % and 96 %. However, effects at P of 0.05, 0.01, and 0.001 had mean replication probabilities of 43 % (95 % CI: 35-45 %), 60 % (95 % CI: 53-61 %), and 77 % (95 % CI: 71-79 %), respectively. For trials with an overall survival primary endpoint that led directly to regulatory approval, the median replication probability was 66 %. A user-friendly web interface is provided to facilitate estimation of replication probabilities of individual trials.

Conclusions: While the direction of observed effects is likely correct when P ≤ 0.05, treatment effects at P of 0.05 in phase III oncology trials are unlikely to be replicated successfully. By itself, statistical significance should not be equated with high replication probability.