Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Coarsened but Confused: Why Composite Exposures Often Fail in Causal Inference
Authors: Nicholas Bakewell,
Presenting Author: Nicholas Bakewell*
In health research, high-dimensional binary treatments are often reduced to composite exposures (CEs) via coarsening functions, typically weighted linear combinations (e.g., medication indicators summarized as a unit-weighted linear combination to form a polypharmacy CE). In causal inference, CEs may be treated as deterministic nodes, often assuming no direct effects from underlying indicators to distal outcomes and informational equivalence. This implies CEs are causally efficacious and sufficient summaries containing necessary information from underlying indicators. However, the latter may not hold, as coarsening functions are often non-invertible and outcome-agnostic. While discussed under multiple versions of treatment theory, this literature implicitly assumes CEs are sufficient, resulting in tautological identification arguments. Further, estimands of CEs represent weighted-averaged effects based on underlying version distributions, but do not allow meaningful per-indicator interpretation as done in practice, and inference is invalid as it ignores uncertainty in the CE. Assumptions under which CE effects may be interpreted as such are formalized: equal conditional effects, no interactions, monotonicity, exchangeability under permutation, sufficiency, and homogenous effect modification across indicators. Simulations demonstrate that even when assumptions hold, further categorization induces bias. These results highlight fundamental problems with CEs for causal inference.