A research team at Debre Markos University and University of Gondar employ a causal forest double machine learning framework to estimate the effect of tuberculosis preventive therapy on antiretroviral adherence in a large Ethiopian cohort. Leveraging orthogonalized random forests, they quantify a modest average adherence reduction and reveal clinical subgroups that experience differential treatment responses.
Key points
- Applied causal forest DML using random forest propensity and outcome models with orthogonalization to estimate TPT effect on ART adherence.
- Found a 3.14 percentage point decrease in adherence (ATE=−0.0314; 95% CI [−0.0373, −0.0254]; p<0.001) in a cohort of 4,152 Ethiopian HIV patients.
- Identified treatment effect heterogeneity: improved adherence in patients with advanced WHO stage, longer ART duration, higher BMI, older age; reduced adherence in those with higher CD4, functional impairment, CPT use.
Why it matters: Causal machine learning reveals treatment effect heterogeneity in TPT’s impact on ART adherence, enabling personalized TB prevention strategies in HIV care.
Q&A
- What is causal forest DML?
- How is ATE different from CATE?
- Why is orthogonalization important?
- What assumptions underlie this causal analysis?
