| Summary: | This is a data-driven study using clinical data provided by UK international company. Their objective was to evaluate biomarkers as surrogate endpoints of the clinical outcomes and specifically to investigate differences between two treatments based on biomarkers. The study raised three issues which are common in clinical trials: surrogate validation (especially with multiple surrogates), missing value imputation and treatment comparison. Biomarkers have been widely used as surrogate endpoints in clinical studies. In the current study, multiple biomarkers were suggested as substitutes for the clinical scores. However, the validation of these is needed. Prentice's validation approach has been popular since the late 1900s. Many researchers have developed their methods based within Prentice's framework, such as Prentice's criteria, Freedman's Proportion Explained (PE), two-level validation (Buyse and Molenberghs, 1998) and the unifying validation using Likelihood Reduction Factor (LRF) by Alonso et al (2006). This thesis concentrates on the latter two approaches, extending it to the case of multiple (>2) surrogates with mix variable types of endpoints in the context of repeated measures. Incomplete data or missing values is an issue common to many fields and is a considerable complication here, with both categorical and continuous variables affected. Multiple biomarkers were involved simultaneously in the imputations. The possible methods studied and compared are K Nearest Neighbours imputation, Expectation Maximization algorithm and Additive Regression with Mean Matching Prediction. EM imputation was employed and adapted to handle multiple continuous variables and complete longitudinal data, using the basic idea of the chained equation, and in data with a 'File-matching' structure, where the biomarkers were never jointly observed. Further areas of development are identified, particularly to enhance the flexibility of coping with a mixture of types of variables. As a result, more sophisticated analyses can be applied on the completed data to investigate the treatment effects. The treatment differences thus are discriminated by biomarkers using a mixed effects model with multiple modelling levels.
|
|---|
