Developing HIV risk assessment tools with pooled complex survey data from 15 African countries

Southern, Eastern, and West African countries host 15% of the global population, yet account for more than half of new global HIV acquisition events. Tools for identifying adults at greatest risk of acquiring HIV can guide focused HIV prevention. Historically, risk assessment tools have had challenges with both internal and external validity due to limitations with the data sources from which they were developed and the models used to construct them. Using data from complex sample surveys to construct risk assessment tools may improve external validity, but these data sources pose analytic challenges due to design features and missing data. We develop risk assessment tools for women and men by fitting Lasso regression models to pooled, complex survey data from 15 countries in Africa. Models were trained on the full population and internally cross-validated. Performance was evaluated using area under the receiver-operating-characteristic curve (AUC), sensitivity, and specificity. Both the male and female models had favorable predictive performance and were parsimonious, with strong external validity.

Cross-validation

External validity

Lasso regression

Missing data

Model validation