UC Berkeley

Researchers in the field of public and global health continue seeking ways to reduce the disproportionate burden of disease on marginalized communities and in resource-constrained settings, for example, in low-middle income countries (LMIC). While progress towards this goal has been made by increasing the uptake of evidence-based practices (EBP) in LMIC, many barriers to sustainable implementation of EBP in LMIC remain. This dissertation is comprised of three studies which harness data-adaptive methods as a tool for supporting uptake of EBP in LMIC.

The first study sought to inform allocation of viral load tests by proposing a differentiated care approach for persons living with HIV. Our results indicate that, in comparison to current non-targeted approaches, a hypothetical machine learning approach may reduce testing frequency relative to resource-rich settings while obtaining similar sensitivity, and while maintaining delay time to viremia detection low. Relative to WHO viral load testing standards, this approach greatly improved the sensitivity and delay time to detection of viremia.

The second study sought to determine whether easily attainable maternal-infant characteristics can predict risk of preterm birth, while maintaining accuracy similar to that of more expensive gestational age (GA) dating methods. Our results indicate that, among women who entered antenatal care (ANC) in the first trimester, an algorithm based on simple maternal-infant characteristics can predict GA and preterm risk within a clinically valuable margin of error. Therefore, this has the potential to inform clinical care surrounding the time of delivery and can be used for preterm birth rate reporting. The last study sought to compare methods for estimating intervention effects in small-scale cluster randomized trials (CRT), which are commonly used, particularly in resource-constrained settings. We assessed the robustness of various methods in analyzing hierarchical data, both in simulation and using a real-data example. We concluded that analyses of small-sample CRT require careful consideration surrounding weighting scheme, covariate adjustment, and target parameter specification.