Statistical and Epidemiological Methods
The institute has a strong research focus on causal inference, mediation analysis techniques and prediction models.
Prediction modelling
Michael Edlinger places his research focus on the validation of prediction models. Edlinger et al. published a paper about risk prediction models for discrete ordinal outcomes, where they assessed the calibration and the impact of the proportional odds assumption. They concluded that non-proportional odds models require more parameters to be estimated from the data, and hence suffer more from overfitting. Despite larger sample size requirements, the authors generally recommend multinomial logistic regression for risk prediction modelling of discrete ordinal outcomes.
Mediation analysis
Another focus of our work is the field of causal inference. We are developing and applying mediation analysis techniques to epidemiological research problems. Taking into account potential outcomes, so called counterfactuals, these techniques aim to detect the underlying mechanisms and to clarify why and how exposure produces the outcome of interest. Josef Fritz and Hanno Ulmer are currently working on statistical methods that allow multiple mediators in complex situations and they are applying these methods to a study, with the aim of better understanding of the underlying mechanism of the effect of obesity on risk of end-stage chronic kidney disease.
Meta-analysis
Our institute also has expertise in individual-participant data meta-analysis, for instance, in the context of the Proof-ATHERO consortium. Meta-analysing these large-scale data can provide more reliable and more precise effect sizes than analysing data from a single study, as it is associated with greater statistical power. Combining data from various studies allows the identification of sources of heterogeneity. Furthermore, we are using methods for the analysis of longitudinal studies, such as mixed-effects models, clinical prediction models to help predict cardio-metabolic or infectious diseases, time-to-event analysis to quantify the risk to develop certain diseases and statistical methods to evaluate causality.