Capture-recapture methods—useful or misleading?
Disease registers are used for two main purposes: to measure the incidence or prevalence of a disease, or to study its natural history. For example, the WHO MONICA collaboration was established in the early 1980s to register myocardial infarction and stroke in different populations worldwide, and thus allow comparisons of incidence to be made. 1 Similarly, cancer registries are routinely used to provide data for comparisons of incidence of different cancers between areas of the UK. These purposes clearly require a different breadth of data than a register intended to study the natural history of a disease. For instance, a stroke register in South London, UK, was established not only to measure the incidence of stroke in this area, but also to follow stroke patients over time in order to examine factors affecting outcome and risk factor management. 2 Quality criteria for stroke incidence registers have been defined, emphasizing the importance of complete, community-based case ascertainment. 3 In either case, a register needs to provide an accurate, unbiased estimate of the number of cases of disease in the population. This should be done by estimating the number of cases missed by the register. In practice however, it is often assumed that if the register has been conducted ‘carefully’ then it will be 100% complete, or at least have missed so few cases that the implications of the study will be unaffected by incompleteness. For example, the MONICA study set out criteria (more than 10% of fatal cases not hospitalized, more than 5% of non-fatal cases not hospitalized, 28-day case fatality less than 40% and a ratio of fatal cases to stroke deaths from routine mortality statistics greater than 1) for assuming equal completeness of case ascertainment across different centres. 4 However, an estimate of the completeness of the register was not made, and there was no evidence that the criteria were necessary or sufficient for completeness. Using the number of cases from a register to estimate incidence assumes that no cases are missing. 5 In addition, if the cases missed differ from those observed (e.g. less severe cases are less likely to be registered) then using only the observed cases will lead to biased inferences.