Resolve the antibiotic-class panel used for profile enumeration

Defines the set of classes entering the \(2^D\) profile enumeration for one hospital x pathogen pair from the fit-time eligibility report (fitted_model$eligibility_report), rather than "tested at least once". A class is included only if its marginal n_tested/n_resistant/n_susceptible thresholds were met at that hospital for that pathogen – this applies to every residual structure, including identity, so an identity-residual panel is never narrowed for reasons that only matter for estimating a correlated residual matrix. For correlated-residual fits specifically, classes that lack sufficient pairwise co-testing with every other candidate class at that hospital are additionally dropped iteratively (the class involved in the most insufficient pairs is dropped first, repeated until all remaining pairs clear the threshold), since \(\Omega\) for an under-co-tested pair would otherwise be prior-dominated.

Usage

.resolve_profile_class_panel(
  class_cols,
  hospital,
  pathogen,
  eligibility_report,
  upper_re_col,
  pathogen_col,
  residual_structure = "identity"
)

Value

List with:

classes

Character vector of class_cols entering the panel, in class_cols order.

excluded

Character vector of class_cols NOT entering the panel (empty if none).

method

Character; "marginal_only" (identity residual, or no pairwise report available), "marginal_plus_pairwise" (correlated residual), or "no_eligibility_report_available" (defensive fallback for fit objects built some other way).

reason

Character or NA; NA when nothing was excluded, otherwise a human-readable breakdown of which classes were excluded and why (insufficient marginal support vs. insufficient pairwise co-testing).

Details

The eligibility report passed in is computed by fit_bayesian_multivariate_probit() on event_data after all experiment-specific filtering (pathogen filter, eligible_pairs semi-join, all-NA-event drop) – i.e. on the exact row population that was fitted, not on an earlier unfiltered wide table. See the eligibility-report construction in fit_bayesian_multivariate_probit().