Causal Estimation of Artificial Intelligence Readmission Risk Under Variable Community Health Worker Visit Records: A Negative-Control Outcome Evaluation of Healthcare Resource Misallocation

Predictive models that flag patients for readmission risk are increasingly used to allocate post-discharge resources, but their training labels are drawn from community health worker (CHW) visit records that vary in completeness across clinics and payers. the present analysis treats that variation as a causal identification problem rather than a data-cleaning nuisance. It formalises the readmission-risk estimator as a function of both true clinical need and visit-record observability, and shows that the resulting misallocation is not a bias in the conventional sense but a stable equilibrium in which under-recorded patients are systematically under-served. A negative-control outcome design is proposed to detect and bound this misallocation: a placebo outcome that cannot be affected by the intervention but shares the same recording mechanism is used to estimate the recording-driven component of the risk score. The design is characterised analytically and evaluated through a replication-based power analysis under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64). Power was estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis indicates that 52 clusters are required to detect a 5-percentage-point effect at 80% power. The contribution is a design, not an estimate: a negative-control outcome framework that makes the recording mechanism visible in readmission-risk evaluation, together with the operating characteristics that determine whether such an evaluation is feasible in a given health system.