Compute Observed-Plus-Imputed Resistance Profile Probabilities

For each event, retains every observed (tested) AST cell exactly and computes a probability distribution only over the genuinely untested (NA) antibiotic-class cells in that event's profile panel. Any enumerated profile inconsistent with the event's observed cells receives probability 0 – this function never overwrites an observed R/S result.

Usage

compute_event_profile_probabilities(
  fitted_model,
  n_posterior_draws_for_profiles = 2000L,
  outcome_col = NULL,
  nonfatal_values = c("Discharged", "Survived", "Alive", "discharged", "survived",
    "alive"),
  seed = 123L,
  n_gibbs_burnin = 10L,
  n_gibbs_kept = 20L,
  posterior_draw_indices = NULL,
  event_indices = NULL
)

Arguments

fitted_model

List returned by fit_bayesian_multivariate_probit().

n_posterior_draws_for_profiles

Integer. Number of posterior draws averaged over for both the event-level profile probabilities and the draw-level aggregates. Subsampled without replacement when total draws exceed this value. Default 2000L.

outcome_col

Character or NULL. Patient outcome column in fitted_model$event_metadata. When NULL, all events are treated as having a known outcome and R_NF is not computed separately.

nonfatal_values

Character vector. Outcome values for the non-fatal cohort (R_NF). Default covers common discharge/survival labels.

seed

Integer. Random seed. Default 123L.

n_gibbs_burnin, n_gibbs_kept

Integer. Only used when fitted_model$residual_structure == "correlated" – see .gibbs_conditional_profile_probs(). Defaults 10L/20L.

posterior_draw_indices

Optional unique indices into the fitted posterior draws. When supplied, these replace random draw subsampling.

event_indices

Optional canonical event indices to include. When supplied, only observed events with these indices are processed.

Value

Named list: event_profiles (event-level posterior mean observed-plus-imputed profile probabilities, with n_classes_observed/n_classes_missing per event) and aggregate_draws (per-draw R_ALL [truly all events in the hp pair], R_KNOWN_OUTCOME [known-outcome cohort], and R_NF [nonfatal cohort], used by aggregate_profiles_for_daly() for credible intervals). Both tibbles contain all \(2^D\) profiles per hospital-pathogen pair and carry profile_generation_method and panel_reason columns.

Details

Identity residual structure (fitted_model$residual_structure == "identity"): classes are conditionally independent given the linear predictor, so the missing-cell probabilities are computed analytically as \(P(Y_{ed}=1 \mid \theta) = \Phi(\mu_{ed})\) for each missing class \(d\), and the profile probability over the missing dimensions is the exact product of per-class Bernoulli probabilities – no latent-variable simulation is used, so there is no added latent-profile-simulation noise. Rows carry profile_generation_method = "conditional_analytic_identity".

Correlated residual structure: the missing latent dimensions are sampled conditional on the observed resistance SIGNS of the tested classes via a Gibbs sampler on the truncated multivariate normal – see .gibbs_conditional_profile_probs(). This replaces the earlier unconditional latent-\(Z\) simulation, which resampled tested cells along with untested ones and was therefore not observed-plus-imputed. Rows carry profile_generation_method = "conditional_gibbs_correlated" and ARE eligible for downstream DALY use (subject to the same panel-support and sampler-acceptability gates as identity-residual profiles – see aggregate_profiles_for_daly()). Because Gibbs is only run for n_gibbs_burnin + n_gibbs_kept iterations per draw rather than to full convergence, correlated-residual profiles carry additional Monte Carlo error beyond identity-residual profiles at the same n_posterior_draws_for_profiles; increase n_gibbs_kept (and/or n_gibbs_burnin) for lower-noise estimates at higher compute cost.

Both residual structures produce, per event per posterior draw, a profile probability vector over the panel's \(2^{D_p}\) enumerated profiles that (a) sums to exactly 1 and (b) assigns exactly 0 to any profile inconsistent with that event's observed cells – enforced by construction in both cases (an exact product for identity; Gibbs never visiting an inconsistent pattern for correlated), and checked with a hard stop() for the identity path since it has no other source of Monte Carlo noise to blur a genuine bug.

Class panels: the antibiotic-class panel enumerated for each hospital x pathogen pair is drawn from the approved eligibility rules computed at fit time (fitted_model$eligibility_report: marginal n_tested/n_resistant/n_susceptible thresholds, plus pairwise co-testing sufficiency for correlated-residual fits) – not from "tested at least once". See .resolve_profile_class_panel().

Estimand: The posterior distribution over the observed event case-mix, conditional on each event's own observed AST results. This is labelled "observed_stewardship_event_mix".