Background: Reported tuberculosis (TB) incidence globally continues to be heavily influenced by expert opinion of case detection rates and ecological estimates of disease duration. Both approaches are recognised as having substantial variability and inaccuracy, leading to uncertainty in true TB incidence and other such derived statistics.
Methods: We developed Bayesian binomial mixture geospatial models to estimate TB incidence and case detection rate (CDR) in Ethiopia. In these models the underlying true incidence was formulated as a partially observed Markovian process following a mixed Poisson distribution and the detected (observed) TB cases as a binomial distribution, conditional on CDR and true incidence. The models use notification data from multiple areas over several years and account for the existence of undetected TB cases and variability in true underlying incidence and CDR. Deviance information criteria (DIC) were used to select the best performing model.
Results: A geospatial model was the best fitting approach. This model estimated that TB incidence in Sheka Zone increased from 198 (95% Credible Interval (CrI) 187, 233) per 100,000 population in 2010 to 232 (95% CrI 212, 253) per 100,000 population in 2014. The model revealed a wide discrepancy between the estimated incidence rate and notification rate, with the estimated incidence ranging from 1.4 (in 2014) to 1.7 (in 2010) times the notification rate (CDR of 71% and 60% respectively). Population density and TB incidence in neighbouring locations (spatial lag) predicted the underlying TB incidence, while health facility availability predicted higher CDR.
Conclusion: Our model estimated trends in underlying TB incidence while accounting for undetected cases and revealed significant discrepancies between incidence and notification rates in rural Ethiopia. This approach provides an alternative approach to estimating incidence, entirely independent of the methods involved in current estimates and is feasible to perform from routinely collected surveillance data.
Keywords: Binomial mixture models; Incidence; Spatial analysis; Tuberculosis.