This Marie Skłodowska Curie Action (MSCA) titled ‘Supporting Causal Conclusions from Observational Survival Studies” (SCOUT) falls within the field of biostatistics. Broadly speaking, the challenges of a biostatistician either concern (1)prediction: quantifying how likely it is for an individual with certain characteristics (age, sex, blood lipids) to develop a certain disease (coronary heart disease) in, say, 10 years or (2) explanation: quantifying how much a certain intervention (e.g. lowering cholesterol levels with a defined statin prescription) would change the risk of developing the disease or (3) description: describing empirical associations in a dataset. As hinted in the title of this MSAC a major focus lay on explanation, also often denoted as ‘causal inference’. Particularly, we were not just interested in quantifying the effect of an intervention, but rather in decomposing the total effect into direct and indirect components (e.g. how much of the effect of the statin use is explained through a reduction in cholesterol and, are there other, perhaps more direct, effects of the intervention). In clinical studies such analyses can help to elucidate an underlying mechanism, to alter and further improve its components, or even to shift its target. Another key interest lay in the comparison of different models for time-to-event data. Such data occur in clinical or epidemiological research when one is interested in the time from an intervention to the occurrence of a particular event, but due to several constrains cannot observe the event in every individual. The most popular model to analyse such data is Cox’s proportional hazards model, mainly due its seemingly easy-to-report outcome measure. However, for causal inference an alternative model, Aalen’s additive hazards model, gained relevance, while the literature on its performance in prediction tasks is sparse. Therefore, we investigated gains and difficulties of these models in clinical case studies, archetypical for prediction and causal inference tasks. The work was carried out at the Section for Clinical Biometrics at the Medical University of Vienna (MUW) under the supervision of Assoc. Prof. Georg Heinze and led to a vibrant two-way transfer of knowledge between myself, my supervisor and the entire Prognosis Research Group.
