Specifying prospective compartmental models of chronic disease

BackgroundDetermining the impact of preventive interventions on future health and economic outcomes in a given population often requires parameterising disease models with forecast disease incidence, remission and case fatality rates. This paper outlines a method to specify these forecasts, using period estimates over several years from source data that are not necessarily coherent from a cohort perspective. MethodsFor a given chronic disease, the model is specified by first obtaining and smoothing source data inputs. We used Global Burden of Disease (GBD) period estimates of disease prevalence (p) incidence rates (i) and mortality rates (d) from 1990 to 2019 as a starting point, from which case fatality rates (f) were calculated. Values were smoothed by birth cohort within each calendar year using a quadratic Bezier curve, from which remission rates (r) were inferred. Next, each measure was projected for 20 years using log-linear regression by year, with 10-year age knots, produced independently by sex. Finally, a stochastic optimization algorithm was used to adjust i,f and r to ensure longitudinal (cohort)-coherence such that they matched simulated forecasts of p and d, which were treated as the reference targets. Model outputThe method is illustrated by presenting model output for the 20 highest disability-adjusted life year-causing chronic diseases in Australia. Model fit was assessed visually, as well as through the relative and absolute differences between calibration output and targets (p and d). A subset model using GBD input data from 1990-2009 only was also produced and compared to 2019 GBD data to investigate the models predictive validity. ConclusionsOur method is a useful optimization procedure to derive cohort-coherent estimates of chronic disease in simulation models. For this analysis, readily available GBD period rates were used as input data; however, other raw data sources could also be applied for a given disease.