Is it possible to increase the reliability of estimated incidence and prevalence rates by disease modelling?

Background In order to monitor the impact of health policy, morbidity estimates must be timely and reliable. Registries in general practice (GP) are key sources for morbidity estimates, as the general practitioner is the gatekeeper of health care. However, morbidity estimates between different GP registration networks vary considerably, and these differences could not be explained by characteristics of the patient population. The aim of this research is to see whether these differences can be explained by variation between networks in the accuracy of distinguishing new (incident) cases from existing (prevalent) cases, and if so, whether modelling this enables more reliable estimates. Methods We used 2010 data from five Dutch GP registration networks and data on four chronic diseases (chronic obstructive pulmonary disease, diabetes, heart failure, and osteoarthritis). We fitted a joint model (DisMod) using all information on morbidity (incidence and prevalence) in each network as well as mortality rates in those with and without the disease. The model assumed stable incidence and survival probabilities over time and included a misclassification factor that allows for misclassification of a percentage of prevalent cases as incident cases. We fitted this model by maximum likelihood and compared the prevalence and incidence estimates for each network with the observed prevalence and incidence estimates. Findings Osteoarthritis of the knee showed large misclassifications (between 2 and 24%), especially in episode-based registrations. Before taking this misclassification into account, prevalence rate estimates from different networks were 1·0–2·8%, while they were 1·3–3·9% afterwards. Incidence rates changed from 2·4–4·5 per 1000 person-years before taking misclassification into account, to 0·8–3·0 per 1000 person-years afterwards. Also, for the other diseases, including a misclassification term mostly affected incidence estimates (lowering them) but did not systematically decrease the variation between prevalence and incidence estimates from different networks. Interpretation Episode-based registers cannot reliably deliver first incidence rates for chronic diseases requiring low levels of professional health care. For the other diseases, modelling misclassification rates does not systematically decrease the variation between registration networks, and only modestly influences estimates. However, such a modelling exercise gives qualitative insight into the reliability of estimates