Estimation of QT interval prolongation through model-averaging

The current method to analyze concentration-QT interval data, which is based on predictions conditional on a best model, fails to take into account the uncertainty of the model. Previous studies have suggested that failure to take into account model uncertainty using a best model approach can result in confidence intervals that are overly optimistic and may be too narrow. Theoretically, more realistic estimates are obtained using model-averaging where the overall point estimate and confidence interval are a weighted-average from a set of candidate models, the weights of which are equal to each model's Akaike weight. Monte Carlo simulation was used to determine the degree of narrowness in the confidence interval for the degree of QT prolongation under a single ascending dose and thorough QT trial design. Results showed that model averaging performed as well as the best model approach under most conditions with no numeric advantage to using a model averaging approach. No difference was observed in the coverage of the confidence intervals when the best model and model averaging was done by AIC, AICc, or BIC, although in certain circumstances the coverage of the confidence interval themselves tended to be too narrow when using BIC. Modelers can continue to use the best model approach for concentration-QT modeling with confidence, although model averaging may offer more face validity, may be of value in cases where there is uncertainty or misspecification in the best model, and be more palatable to a non-technical reviewer than the best model approach.