Abstract
BACKGROUND: Nationally representative tuberculosis (TB) prevalence surveys provide invaluable empirical measurements of TB burden but are a massive and complex undertaking. Therefore, methods that capitalize on data from these surveys are both attractive and imperative. The aim of this study was to use existing TB prevalence estimates to develop and validate an ecological predictive statistical model to indirectly estimate TB prevalence in low- and middle-income countries without survey data. METHODS: We included national and subnational estimates from 30 nationally representative surveys and 2 district-level surveys in India, resulting in 50 data points for model development (training set). Ecological predictors included TB notification and programmatic data, co-morbidities and socio-environmental factors extracted from online data repositories. A random-effects multivariable binomial regression model was developed using the training set and was used to predict bacteriologically confirmed TB prevalence in 63 low- and middle-income countries across Africa and Asia in 2015. RESULTS: Out of the 111 ecological predictors considered, 14 were retained for model building (due to incompleteness or collinearity). The final model retained for predictions included five predictors: continent, percentage retreated cases out of all notified, all forms TB notification rates per 100 000 population, population density and proportion of the population under the age of 15. Cross-fold validations in the training set showed very good average fit (R-sq = 0.92). CONCLUSION: Predictive ecological modelling is a useful complementary approach to indirectly estimating TB burden and can be considered alongside other methods in countries with limited robust empirical measurements of TB among the general population.