Skip to contents

Observed-versus-Model Pairwise Co-resistance Validation

Source: R/daly_resistance_validation.R
validate_pairwise_calibration.Rd

For every adequately co-tested hospital x pathogen x class-pair combination, compares the observed pairwise resistance rate (\(n_{RR}/n_{\text{cotested}}\)) against the model's implied joint probability, computed differently by residual structure:

Identity

\(P(Y_{d_1}=1, Y_{d_2}=1 \mid \theta) = \Phi(\mu_{ed_1})\,\Phi(\mu_{ed_2})\) – the two classes are combined only through the shared fixed and random effects in \(\mu\), not through any residual correlation (the identity model has none). This checks whether the identity-residual model reproduces observed pairwise resistance through shared covariates and random effects; it is not evidence that a residual-independence assumption is correct. Persistent pairwise miscalibration despite good marginal calibration suggests unmodelled class-to-class dependence that a correlated-residual fit might capture.

Correlated

\(P(Y_{d_1}=1, Y_{d_2}=1 \mid \theta) = \Phi_2(\mu_{ed_1}, \mu_{ed_2}; \rho_{d_1 d_2})\) – the standard bivariate-probit identity \(P(Z_1>0,Z_2>0) = \Phi_2(\mu_1,\mu_2;\rho)\) for \((Z_1,Z_2)\sim N((\mu_1,\mu_2), [[1,\rho],[\rho,1]])\), with \(\rho_{d_1 d_2} = \Omega_{d_1 d_2}\) from the fitted residual correlation matrix (computed via pbivnorm). This actually uses the fitted correlation structure, unlike the identity-model product formula – checking whether \(\Omega\) reproduces observed co-resistance.

Usage

validate_pairwise_calibration(
  fitted_model,
  n_posterior_draws_for_validation = 2000L,
  seed = 123L,
  ci_level = 0.95,
  min_cotested = NULL
)

Arguments

fitted_model

List returned by fit_bayesian_multivariate_probit().

n_posterior_draws_for_validation

Integer. Posterior draws used. Default 2000L.

seed

Integer. Random seed (draw subsampling only). Default 123L.

ci_level

Numeric. Credible interval coverage. Default 0.95.

min_cotested

Integer or NULL. Eligibility threshold. When NULL (default), eligibility is taken from fitted_model$eligibility_report$pairwise.

Value

Tibble, one row per eligible hospital x pathogen x class-pair: n_cotested, observed_RR/RS/SR/SS, observed_pairwise_resistance, model_pairwise_mean/lower/upper, absolute_error, interval_contains_observed.