Table 4 Recommendations for improved meta-analysis of Phase II trials of binary outcomes

Framework

Use a logistic regression model to model the binomial distribution of the data within studies, and to avoid continuity corrections given a zero event in one arm.

Choice of model

Do not make decisions to use a fixed-effect or random-effects model based on I² or tests for heterogeneity.

Heterogeneity

State a priori that a random-effects model will be used to account for heterogeneity in treatment effects.

Uncertainty

Use a Bayesian framework to account for all parameter uncertainty and external evidence (such as the between-study variance) and to enable direct probabilistic inferences. However, a sensitivity analysis to the choice of prior distributions is required.

Prediction intervals

Report 95% prediction intervals as they reveal the potential treatment effect in a new population, and inform subsequent Phase III decisions.

Bias

Use skeptical prior distributions for the treatment effect if there is evidence to suggest the Phase II trials may be biased in favor of the treatment.