Artikel
Marginal or process-based - which models of multivariate survival are relevant to therapeutic trials?
Suche in Medline nach
Autoren
Veröffentlicht: | 8. September 2005 |
---|
Gliederung
Text
Purpose
Analysis of multivariate time-to-event data is an emerging area of applied statistics since appropriate software has become available in several packages. As its statistical modeling requires both careful consideration of context of application and adequate understanding of the involved stochastics, extended discussion is obviously necessary to discourage uneducated software expeditions that will certainly trigger conflicting opinions between academia and industry without benefit to the profession.
Material and Methods
Focus will be on therapeutic or nosograhic studies in chronic diseases in which progression and sojourns in clinically discernible states or episodes of short-term attacks are an endpoint of evaluation. Psychic seizures, migraine attacks, post-traumatic complications, post-medication adverse events are examples. A classification of problems into 'concurrent' or 'recurrent' multiple events is helpful. The range of available methodology is then broadly demarcated by marginal models with and explicit stochastic process models without introduction of 'mathematical artefacts' like copula or frailty terms.
Results
Multiple-events times of the 'recurrent' type will most often benefit from multi-state stochastic process models in terms of more meaningful analyses and interpretations; for statistical inference on their intensity functions, the methods based on counting processes with multiplicative intensities are available and provide a rich tool box for sophisticated analyses. In some applications, (trend ) tests for ordered alternatives on intermittently censored time-to-event data will be of interest. Multiple-events times of the 'concurrent' type will benefit from the marginal models, in general. Even then, parametric and nonparametric statistical models of multivariate data based on explicit stochastic processes can be tractable and useful. This is seen for safety data from randomized clinical trials, or in applications to dental cohort studies.
Conclusion
There are no 'universal' approaches, though some are more versatile than others. Reservations about marginal approaches are motivated by mathematical artefacts that may conceal that extra insight into a complex stochastic system that is achievable with 'more-than-marginal data', a common feature in designed clinical follow-up studies.