Methods of survival analysis applied in oncology — assumptions, methods and common pitfalls

Abstract

Survival analysis is the analytical foundation of studies on cancer-related mortality or disease progression. Analytical techniques used for this purpose share one common trait — the uncertainty of the event’s occurrence in individuals, whose observation time has been censored due to study termination, withdrawal due to events other than prespecified endpoints or loss to follow-up. The main advantage of such analytical methods is the possibility to estimate individual hazard (risk of event occurrence) at any given timepoint of observation and consequently of expected survival time depending on a plethora of clinical variables and treatment modalities. Due to a huge and ever-expanding amount of oncologic data using survival analysis as a primary measure of outcome, the ability to interpret such results and to know the assumptions and workings of particular methods is slowly becoming ubiquitous. Analytical methods deployed on survival data feature univariate nonparametric ones (the log-rank test), multivariate modeling techniques (assuming proportional or additive risk), automated neural networks and classification-regression trees. The purpose of this review was to present and discuss in detail the available range of analytic and exploratory methods used in biostatistics and oncology. The spectrum of common analytical and interpretative problems, methods of designing and planning clinical trials and verifying the veracity of published data may provide a valuable addition in the process of clinical application of evidence based oncology.
Onkol. Prak. Klin. 2011; 7, 2: 89–101