Medizinische Universität Graz Austria/Österreich - Forschungsportal - Medical University of Graz
Gewählte Publikation:
Svendova, V.
Statistical modelling of effect sizes and rankings from multiple biomedical studies
Doktoratsstudium der Medizinischen Wissenschaft; Humanmedizin; [ Dissertation ] Graz Medical University; 2017. pp.
[OPEN ACCESS]
FullText
- Autor*innen der Med Uni Graz:
- Svendova Vendula
- Betreuer*innen:
- Berghold Andrea
- Schimek Michael
- Altmetrics:
- Abstract:
- Replicating research experiments or studies can often lead to different conclusions, sometimes even contradicting each other. To resolve potential conflicts, researchers often pool related studies to obtain a combined, more reliable result. This pooling of information is called meta-analysis. Assuming reasonable consistency, one can combine experiments in any research area. A large number of statistical methods for meta-analysis have been developed over the years. Most of these methods are tailored for specific tasks, such as combining clinical trials or genomic experiments, and cannot be immediately applied to other problems. A more general group of meta-analytical methods are based on rank order data. These methods work with ranks instead of the measured values themselves, the latter of which are not always available, and therefore are not limited by the data type and not disturbed by different data transformations, presence of outliers, or requirements regarding their statistical distribution. Nevertheless, the generality of rank-based methods comes at a price: the relative differences between the measured values are lost and as a consequence they could not, until now, estimate the common study signals that have produced the observed ranks. This thesis proposes a new approach that combines the advantages of rank-based methods, while achieving the ultimate goal of meta-analysis: estimating those signals that are causal for the ranks. Moreover, the standard errors of the signal estimates are estimated by a non-parametric bootstrap, and the stability of the observed rank positions is assessed. The proposed approach is tested on simulated data under various scenarios, as well as applied to real data combining studies and experiments from clinical and genomic research. The simulations showed that the proposed approach can estimate the underlying signals accurately, as well as estimate the derived rank positions. As expected, better estimates were achieved when the agreement between the studies or experiments was high. The size of the standard errors reflected the uncertainty of the signals, and the amount of overlap of the standard error ranges was indicative of the rank instability. When applying the method to the real-world applications mentioned above, promising results were obtained. Finally it was demonstrated that the proposed method is a useful meta-analytic tool in biomedical research. The main drawback of the method is its computational demands, which, however, could be relaxed by further optimisation of the algorithm. The thesis is accompanied by supplementary figures and tables from the simulations, as well as the R source code for the algorithm. In conclusion, the submitted thesis presents a new, general tool for meta-analysis of ranked lists, which can estimate the underlying signals that inform the observed rankings, the standard errors of the signals, and the involved rank stabilities.