Accounting for model uncertainty in estimating global burden of disease

Abstract

To illustrate the effects of failing to account for model uncertainty when modelling is used to estimate the global burden of disease, with specific application to childhood deaths from rotavirus infection. To estimate the global burden of rotavirus infection, different random-effects meta-analysis and meta-regression models were constructed by varying the stratification criteria and including different combinations of covariates. Bayesian model averaging was used to combine the results across models and to provide a measure of uncertainty that reflects the choice of model and the sampling variability. In the models examined, the estimated number of child deaths from rotavirus infection varied between 492,000 and 664,000. While averaging over the different models' estimates resulted in a modest increase in the estimated number of deaths (541,000 as compared with the World Health Organization's estimate of 527,000), the width of the 95% confidence interval increased from 105,000 to 198,000 deaths when model uncertainty was taken into account. Sampling variability explains only a portion of the overall uncertainty in a modelled estimate. The uncertainty owing to both the sampling variability and the choice of model(s) should be given when disease burden results are presented. Failure to properly account for uncertainty in disease burden estimates may lead to inappropriate uses of the estimates and inaccurate prioritization of global health needs.