Pharmacometric Investigations of Prediction Precision and Advances of Models for Composite Scale Data

Abstract: Clinical trials are needed to evaluate new treatments. In late-stage clinical trials, failures are mostly due to lack of efficacy. Fit-for-purpose analysis methods will likely increase the success rates and advance drug development by providing higher precision to support decisions such as go/no-go, dose selection, or sample size. This thesis presents new methods for analysis of composite scale data, and comparisons of prediction precision of new and standard analysis methods. Composite scale data arise from questions/items rated with integers. A total score can be derived, which is discrete and bounded. Item response theory (IRT) models are the natural choice for such data, since they use the item-level information. However, when only the total score is available they cannot be used. The bounded integer (BI) model is a new method for discrete, bounded outcomes. With composite scale total score data, it had superior fit compared to standard methods, because it respects the nature of the data. Further, a new method, formally linking IRT models to models for total score, was developed. The expected mean and variance, given an IRT model, was implemented in BI and continuous variable models. This improved fit, allowed estimation of IRT parameters, and allowed comparison of different model types.The prediction precision of both outcome and parameters were investigated with different methods, ranging from t-test to mechanistic pharmacometric models, for composite scale and continuous data. The most suitable method depended on the purpose, for example mechanistic models are superior at establishing a drug’s site of action.In conclusion, the choice of method should be based on the primary question, and also the data collected. The method should not be more complex than necessary, and the nature of the data respected. This thesis will help modellers select the most appropriate analysis method for a problem at hand.