Primary Submission Category: Design of Experiments
Emulating Factorial Designs for Multiple Concurrent Binary Interventions
Authors: Nicholas Bakewell,
Presenting Author: Nicholas Bakewell*
The causal inference literature has focused on univariate binary interventions. However, many interventions are implemented concurrently, and ignoring between-intervention dependence or treating concurrent interventions as nuisance parameters may bias estimates. Recent multiple causes literature uses linear factor models, copulas, or other approaches (e.g. Bahadur expansion, saturated models) to capture dependence among interventions; however, these may ignore residual dependence, lack uniqueness, use inefficient parameterizations or cannot disentangle main and interaction effects. Focusing on settings where few binary interventions are selected based on subject-matter knowledge, Target Trial Emulation and potential outcomes frameworks are used to conceptualize this as a 2^k factorial design. Appealing to sparsity-of-effects, main effects and two-way interaction estimands are of interest. Integrating Empirical Likelihood (EL), calibrated covariate balancing weights, and joint marginal structural models (MSMs), estimation approaches are outlined for point and time-varying interventions. Simulations compare the proposed approach to existing joint MSMs estimation approaches via logistic regression-based inverse probability weighting and covariate balancing propensity score. Existing approaches scale poorly with the number of interventions/time periods, are more sensitive to misspecification of the joint distribution of interventions and do not account for simultaneous inference.