| dc.description.abstracteng | The aim of precision medicine is to identify the treatment that provides the best response for a patient. For this purpose, predictive biomarkers play a crucial role. Due to their ability to define subgroups of patients that respond differently to treatment, they are highly useful. Biomarker-related subgrouping of the patient population may have different reasons, e.g. an improved benefit-risk balance, and can be supported by different sources of evidence, e.g. clinical data or pharmacological evidence investigating the biochemical or physiological effect of a drug on cells, organs, and systems. To assess the usefulness of a biomarker stratifying the patient population considering the presented evidence, a classification scheme with five increasing levels of evidence with regard to the expected molecular mechanism and the clinical evidence was proposed as part of this dissertation. Additionally, for each of the categories, an example of a biomarker-drug pair were suggested.
As the mechanism of action of a drug is not always fully understood or maybe even unknown, data-driven identification of differential treatment effects in subgroups suggesting treatment-by-subgroup or more precisely treatment-by-biomarker inter-
actions is of interest to inform further research. Various data-driven subgroup identification methods have been proposed. However, neutral and systematic comparisons of their performance in simulation studies are rare. Therefore, I conducted a simulation study in order to compare five popular approaches regarding their capability to select a target population for subsequent trials. Although most of the methods performed well in settings with larger effects or more substantial sample sizes, all methods have difficulties in more realistic drug development settings with sample sizes that are not sufficiently large for identifying treatment heterogeneity across the population.
Pooling data from multiple trials can increase the sample size on which subgroup identification is performed. When pooling data from multiple studies, however, the between-trial heterogeneity must be taken into account, as otherwise spurious subgroups might be identified. Therefore, I proposed the metaMOB approach for subgroup identification in individual participant data (IPD) meta-analysis. The proposed approach combines commonly made assumptions in random-effects meta-analysis regarding between-trial heterogeneity and the generalized mixed-effects model tree algorithm based on model-based recursive partitioning (MOB). Using a Monte-Carlo simulation study, I showed that metaMOB is an appropriate subgroup identification method for IPD resulting from multiple heterogeneous trials.
Discrete time-to-event data needs specialised methods for data analysis including subgroup identification. Although MOB is applicable to a wide range of different outcome measures, e.g. normal, or binary, it is not suitable for discrete time-to-event
data. For discrete time-to-event models which are based on binary outcome models, I could show that the type I error rate of the M-fluctuation test used as splitting criterion in MOB is inflated. I illustrated the inflated type I error rate in a simulation
study and proposed a revised version of MOB for discrete time-to-event data, which is based on a resampling procedure and which controls the type I error more closely. | de |