Measuring cancer survival in populations: relative survival vs cancer-specific survival

Two main methods of quantifying cancer patient survival are generally used: cancer-specific survival and relative survival. Both techniques are used to estimate survival in a single population, or to estimate differences in survival between populations. Arguments have been made that the relative survival approach is the only valid choice for population-based cancer survival studies because cancer-specific survival estimates may be invalid if there is misclassification of the cause of death. However, there has been little discussion, or evidence, as to how strong such biases may be, or of the potential biases that may result using relative survival techniques, particularly bias arising from the requirement for an external comparison group.

In this article we investigate the assumptions underlying both methods of survival analysis. We provide simulations relating to the impact of misclassification of death and non-comparability of expected survival for cause-specific and relative survival approaches, respectively.

For cause-specific analyses, bias through misclassification of cause of death resulted in error in descriptive analyses particularly of cancers with moderate or poor survival, but had smaller impact in analyses involving group comparisons. Relative survival ratio (RSR) estimations were robust in relation to non-comparability of comparison populations for single RSR but were less so in group comparisons where there was large variation in survival.

Both cause-specific survival and relative survival are potentially valid epidemiological methods in population-based cancer survival studies, and the choice of method is dependent on the likely magnitude and direction of the biases in the specific analyses to be conducted.