Incorporation of hierarchical structure into estimation and projection package fitting with examples of estimating subnational HIV/AIDS dynamics.

OBJECTIVES

The article aims to give Spectrum/estimation and projection package (EPP) users and the scientific community a basic understanding of the underlying statistical model used to incorporate hierarchical structure in HIV subnational estimation, and to show how it has been implemented in the Spectrum/EPP interface for improving subepidemic estimation. The article also provides recommended default settings for this new model.

METHODS

We apply a generalized linear mixed-effects model on antenatal clinics prevalence data to get area-specific prevalence and uncertainty estimates, and transform those estimates to auxiliary data. We then fit the EPP model to both the observed data and auxiliary data.

RESULTS

We apply the proposed methods to four countries with different levels of data availability. We compare the out-of-sample prediction accuracy of the proposed method with varying auxiliary sample sizes and EPP without auxiliary data.

CONCLUSION

We find that borrowing information from data-rich areas to data-sparse areas using our proposed method improves EPP fit in data-sparse areas. We recommend using the sample size estimated from generalized linear mixed-effects model as the default auxiliary sample size.