A subnational analysis of burden of disease in South Africa: mortality levels, causes of death and their forecasts.

Abstract

Empirical and model based approaches have provided estimates of South Africa’s national and provincial mortality. However, little is known about district-level mortality patterns and differentials. The purpose of this research is to provide reliable estimates of these over three aims. Firstly, district-specific all-cause deaths for each age, sex and year are estimated by adjusting observed vital registration (VR) death numbers using a Bayesian regression model that concurrently addresses under-reporting of deaths and the random noise which typifies small-area samples. For the second aim, district-specific causeof- death proportions for selected causes are determined using a Dirichlet-Multinomial regression model that leverages the South Africa province cause-of-death estimates from the 2017 Global Burden of Disease study to correct the district cause-of-death numbers from VR. In the final aim, the results from the previous aims are forecasted within a compositional data framework (CoDa) by first applying Singular Value Decompositions (SVDs) to the estimated all-cause and multiple decrement life table death matrices for all district-years. Next, estimated time-varying parameters from the resultant low-rank matrix approximations are forecasted using additive models that assume first order autoregressive residuals. Finally, full life table death matrices are reconstructed by combining these forecasts with the non-varying principal components estimated using the SVD-CoDa.