Drives ppc_draws$generate_state(s) once per posterior state (see simulate_probit_posterior_predictive), streaming each state's complete and observation-masked replicate through every requested discrepancy statistic and accumulating only the compact per-state scalar contributions – never an S x N_events x D array. Statistic families (Stan User's Guide-style posterior predictive checks, adapted to AMR resistance data):

marginal: Per-class resistance rate, pooled and per eligible hospital x pathogen x class cell (reusing fitted_model$eligibility_report$marginal).
resistant_count: Number of resistant classes per event, among whatever was tested for that event, summarised (mean/variance/median/ p90/max); plus, for fully-observed eligible panels, the full count distribution.
pairwise: RR/RS/SR/SS co-resistance proportions for every sufficiently co-tested class pair (reusing fitted_model$eligibility_report$pairwise).
complete_profile: All-susceptible/all-resistant/most-common-profile frequency, distinct-profile count, Shannon entropy, Simpson concentration, and the top-3 most frequent OBSERVED profiles (profile_top_observed_frequency – a purely data-dominant descriptive summary, deliberately NOT a claim of clinical importance; see the file header of probit_posterior_predictive.R for the distinction and the deferred clinical_priority_profiles extension point), restricted to adequately-supported fully-observed hospital x pathogen panels (enumerate_binary_profiles- style class panel via .resolve_profile_class_panel()).
hospital_heterogeneity: SD/IQR/MAD/range of per-hospital resistance proportions among hospitals meeting min_hospital_support, per class.
admission_clustering, patient_clustering: Within-/between-group same-class agreement and Hamming distance, only when the relevant role is declared eligible (see random_effect_roles and random_effect_eligibility_audit below); skipped cleanly (support_status != "supported"), never as a failure, when unsupported.

Usage

compute_probit_ppc_statistics(
  fitted_model_or_ppc_draws,
  n_states = 500L,
  seed = 123L,
  preserve_observation_mask = TRUE,
  statistics = c("marginal", "resistant_count", "pairwise", "complete_profile",
    "hospital_heterogeneity", "admission_clustering", "patient_clustering"),
  random_effect_roles = NULL,
  random_effect_eligibility_audit = NULL,
  min_hospital_support = 30L,
  min_complete_profile_events = 30L,
  ci_level = 0.95
)

Arguments

fitted_model_or_ppc_draws: Either a fitted model object (as for simulate_probit_posterior_predictive) or an "amr_ppc_draws" object already returned by that function – the latter avoids re-deriving posterior draw setup when statistics are recomputed for the same simulated replicates.
n_states, seed, preserve_observation_mask: See simulate_probit_posterior_predictive. Ignored when fitted_model_or_ppc_draws is already an "amr_ppc_draws" object (its own settings are used).
statistics: Character vector of statistic families to compute. Default: all of "marginal", "resistant_count", "pairwise", "complete_profile", "hospital_heterogeneity", "admission_clustering", "patient_clustering".
random_effect_roles: Optional named character vector/list mapping the name of a GROUPING COLUMN present in fitted_model$event_metadata (the NAME of each entry) to one of "hospital", "admission", "patient" (the VALUE – a closed set of ROLE labels, never an assumption about what any particular experiment's columns are actually named), e.g. c(admission_id = "admission"). Deliberately decoupled from whether that column is declared as a random-effect block in this particular fit – e.g. tagging a column that is NOT modelled as a random intercept with role "admission" is exactly how a fit lacking an admission random effect gets checked for whether it fails to reproduce observed within-admission clustering (see Stan User's Guide- style synthetic recovery tests in test-probit-predictive-synthetic-recovery.R). NULL (default) means no admission/patient-role clustering statistics can run (reported with support_status = "role_not_mapped").
random_effect_eligibility_audit: Optional data frame with columns block (role: "hospital"/"admission"/"patient") and status ("eligible_primary"/"eligible_sensitivity_only"/ "not_eligible") – e.g. the pathogen-specific random-effect eligibility audit produced by the analysis layer. When NULL, a generic structural fallback check is used instead (see simulate_probit_posterior_predictive's file header for the general never-hardcode-eligibility-in-the-package philosophy this mirrors).
min_hospital_support: Integer; minimum tested events for a hospital to enter the hospital-heterogeneity comparison. Default 30.
min_complete_profile_events: Integer; minimum fully-observed events for a hospital x pathogen panel to enter complete-profile / full resistant-count-distribution statistics. Default 30.
ci_level: Numeric; determines the replicated_q025/ replicated_q975 quantile levels ((1-ci_level)/2 and its complement). Default 0.95.

Value

A tibble with columns statistic_name, stratum, observed_value, replicated_mean, replicated_sd, replicated_q025, replicated_q50, replicated_q975, ppc_tail_probability, ppc_two_sided, n_replications, support_status.

References

Stan Development Team. "Posterior and Prior Predictive Checks." Stan User's Guide. https://mc-stan.org/docs/stan-users-guide/posterior-predictive-checks.html

Compute AMR-specific posterior predictive discrepancy statistics

Usage

Arguments

Value

References