Estimating incidence of stroke using capture-recapture models
Background:Incidence of stroke (and other cardiovascular diseases) is often estimated using population-based registers. No register is 100% complete, and cases missed may differ from those registered, for example by having less severe disease. Failure to estimate the number of cases missed will lead to under-estimation of the true incidence of disease. Comparisons between areas, or examination of trends over time, will also be biased if the probability of being registered varies between them. Methods: Recently developed capture-recapture models including covariates were used to estimate incidence of stroke using data from a population-based register in South London, UK. The model parameters were used to estimate ascertainment-adjusted age-standardised incidence rates. Confidence intervals which allow for the uncertainty in the estimation of population size were calculated. Results:The crude capture-recapture model (not including covariates) under-estimated the number of non-fatal strokes. Probability of capture by the register was associated with both demographic variables and stroke-severity indicators. A capture-recapture model including these covariates led to more plausible estimates of the numbers of both fatal and non-fatal strokes, and suggested that the stroke register was approximately 88% complete. Adjusting for the number of cases missed increased the estimated incidence from 1.31 (95% confidence interval 1.21 to 1.42)to 1.49 (95% confidence interval 0.38 to 2.60) per 1,000 people. Conclusions: Using these methods, data from an incomplete register can be used to calculate both incidence and age-standardised incidence. However, sparse data in subgroups can lead to wide confidence intervals for adjusted rates. Routine registers might be made more cost-effective by using the combination of sources and covariates which most accurately estimates the population size, instead of aiming for 100% completeness. Comparisons of incidence between registers should be based on rates which are adjusted for the number and type of cases missed.