Correcting population-based survival for DCOs - why a simple method works and when to avoid it.

A high proportion of cancer registrations solely based on a death certificate (DCOs) indicates poor data quality and biases cancer survival estimates. Intensive trace-back of registrations initiated after death (DCIs) can reduce the proportion of DCOs to an acceptable level and also improve data quality in other areas (such as increasing the information on disease extent, morphology and treatment) but is expensive in staff time. Our approach - based on a proportional hazards model for DCOs relative to all other cases - can be used to predict what the likely effect of the trace-back will be on survival and to justify the extra work involved. It can also be used to correct results from other sources (including historical data) especially when these sources contain high percentages of DCOs. Of course, the ability to make this correction is no excuse for omitting trace-back of DCI cases when resources permit. With our model the true survival tends ultimately to (1-p) *S where p is the proportion of DCOs and S is the observed survival, which is a simple correction noted by others. The worse the assumed survival of DCOs is relative to all other cases, the earlier is the time for the maximum difference between observed and true survival. Correction to the later part of survival curves is easy and an example is shown using EUROCARE data. This paper shows why the simple method works and suggests that researchers should always think about adjusting their survival estimates with regard to the percentage of DCOs. This paper also shows when the simple correction can (on 5-year survival estimates) and cannot (on 1-year survival estimates, generally) be used to adjust survival figures when comparisons are made across regions or countries with differing percentages of DCOs. We also present examples of some hazard ratios found in practice.