Planning and nonparametric inderence for multistate time-to-event data such as diesease occurrences and disease durations

• Applicants: Professor Dr. Jan Beyersmann; Professor Dr. Martin Schumacher
• Subject Area: Epidemiology and Medical Biometry/Statistics
• Term: from 2011 to 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 189200139

The course of a chronic or acute disease often consists of a temporal sequence of events. Multistate models provide a general framework to analyse such disease trajectories. A common approach in clinical studies is to summarize such events in a composite endpoint; e.g. disease-free survival. A competing risks model allows to investigate disease occurrence and death in a more specific way. In case one is also interested in the duration of disease, a more complex multistate model that also takes death after disease occurrence into account is needed.Clinical studies that deal with complex questions need careful planning. In one of the subprojects we have investigated statistical methodology for planning such studies. As one example we considered the clinical question how to show that an innovative prophylactic treatment regimen reduces the occurrence of severe adverse immune reactions while not compromising relapse-free survival of patients after stem-cell transplantation. For rare events, we also considered the use of sampling designs where all patients with the event of interest, e.g. a hospital-acquired infection, but only a sample of patients without the event of interest, will be included in the study. The challenge here is that disease (infection) status is a time-dependent feature and has to be treated as such.In another subproject we investigated alternative approaches that consider these time-dynamic phenomena but allow for simultaneous statistical inference. Such approaches are often simulation and/or resampling based and thus computationally demanding. On the one hand, we have been able to theoretically justify an approach proposed about 20 years ago and, on the other hand, to apply that approach successfully in the analysis of a study on the safety of drugs during pregnancy.In the following we will investigate central issues of planning and statistical inference for studies that call for complex multistate models. In particular, we will concentrate on the so-called exposure-density sampling as an efficient alternative for studies in which exposure to a risk factor is rare. Resampling-/permutation-based statistical approaches are at the core with regards to statistical inference. The planned cooperation shall ensure that both equally important aspects are tied up as closely as possible.

• DFG Programme: Research Grants