LATEBREAKER
Methods/Statistics
Do you think you’re better off alone? Causal inference with imbalanced dependence between exposure groups Joshua Nugent* Joshua Nugent Laura Balzer
In many studies, participants are clustered into groups, such as neighborhoods or schools. In such settings, we typically assume that the clusters are the independent unit, but this is sometimes an oversimplification of the underlying dependence structure. In the SEARCH-IPT trial, for example, public health officials were (artificially) grouped together for randomization. In the intervention arm, the groups became mini-collaboratives in which an intervention for tuberculosis prevention was delivered; in the control arm, participants continued with standard practice. As a result, intervention participants were dependent, while control participants were effectively independent (Figure). For such trials, we show that targeted minimum loss-based estimation (TMLE) accounting for the imbalanced dependence structure and leveraging machine learning (Super Learner, SL) improves power while maintaining nominal confidence interval coverage. Its application in SEARCH-IPT resulted in a 20% efficiency gain. Additionally, unlike more traditional methods (e.g., GEE), our approach has the flexibility to estimate wide range of target parameters. We recently generalized our theoretical and practical results to observational studies, where imbalanced clustering is common but less frequently studied. For example, health officials choosing to participate in mini-collaboratives would be dependent, while health officials declining participation would be effectively independent. For such observational studies, we show that TMLE with SL yields meaningful improvements in precision as well as reductions in bias due confounding.
