Objective: To formalize identifiability conditions for electronic phenotyping when diagnosis codes are informative of latent clinical severity, and to characterize the bias of code-only and chart-review-calibrated prevalence estimators. Materials and Methods: We embed code-based phenotyping in the coarsening-at-random framework: the true clinical state is the latent cause, the observed diagnosis code set is the candidate set, the coding process is the masking mechanism, and a chart-reviewed patient is a singleton candidate set. We identify informative coding (sicker patients are coded more) as the failure of the non-informative-masking condition C2, and we analyze identifiability and bias under this failure. Results: A glass-ceiling theorem shows that (prevalence, code sensitivity, code specificity) is jointly non-identifiable from code data alone. An identifiability theorem shows that a chart-reviewed subsample covering both classes identifies the coding model and yields the Rogan-Gladen plug-in estimator. A code-frequency consistency theorem shows the fitted model reproduces the empirical code frequency exactly at any interior MLE, so marginal-fit diagnostics are blind to prevalence bias. A bias bound shows the code-only bias is controlled by the severity-coding correlation and the calibrated estimator's residual bias by the case-mix gap. Simulation confirms all four results; the bias sign can flip from negative to positive depending on whether sensitivity deficit or false-positive inflation dominates. Discussion: The framework names the failing assumption (C2) and quantifies the bias, unifying Hui-Walter latent class models, Rogan-Gladen correction, and verification-bias correction in one vocabulary. Conclusion: Chart review is a necessity, not a luxury; its identifying value is bounded by case-mix representativeness.
Alexander Towell (Thu,) studied this question.