Causal Inference
When measurement mediates the causal effect of interest Joy Z. Nakato* Joy Z. Nakato Brian Beesiga Janice Litunya Jaqui Mwango Kara Marson Judith A. Hahn Carol S. Camlin Diane V. Havlir Maya L. Petersen Moses R. Kamya Jane Kabami James Ayieko Gabriel Chamie Laura B. Balzer
In many studies, participants with measured outcomes differ meaningfully from participants with missing outcomes. A common approach to address missingness considers a hypothetical intervention to ensure complete measurement outcome. This approach fails when measurement mediates the causal effect of interest. Consider, for example, the OPAL study, a cluster randomized trial to compare bar-based recruitment strategies on uptake of HIV pre-exposure prophylaxis (PrEP) in rural Kenya and Uganda. Since starting PrEP is contingent on testing HIV-negative at a health clinic, clinic-based HIV testing is the key measurement variable. However, “intervening” to test all would block the indirect effect of the bar-based recruitment strategy (exposure) on PrEP uptake (outcome; Figure). To evaluate the total causal effect in such settings, we develop a general framework to define and identify a novel causal estimand, while still accounting for differential missingness. For the corresponding statistical estimand, we develop and apply a novel Two-Stage TMLE that also accounts for clustering and small sample sizes. Simulations demonstrate the practical performance of our approach as well as the limitations of more traditional approaches.
