Skip to contents

Estimate Resistance Profile Probabilities via Convex Optimisation

Source: R/daly_resistance_profiles.R
estimate_profiles_convex.Rd

The public-facing solver for Pathway 1. Accepts pre-computed marginal and pairwise resistance rates (from compute_marginals_from_data() / compute_pairwise_from_data(), or from validate_aggregate_inputs() for GBD-style pre-computed inputs), enumerates all \(2^n\) binary resistance profiles, builds the constraint matrix, and solves the simplex-constrained weighted least-squares QP:

Usage

estimate_profiles_convex(
  marginals,
  pairwise = NULL,
  panel_map = NULL,
  exclude_near_zero = TRUE,
  zero_threshold = 0,
  top_n_classes = NULL,
  lambda = 1e-08,
  sigma_sq = 1,
  solver = c("osqp", "quadprog"),
  n_cores = 1L,
  pathogen_col = "pathogen",
  class_col = "antibiotic_class",
  rate_col = "marginal_resistance",
  n_tested_col = "n_tested"
)

Arguments

marginals

Tibble. Output of compute_marginals_from_data() or a validated aggregate table. Must have at minimum pathogen_col, class_col, and rate_col.

pairwise

Tibble or NULL. Output of compute_pairwise_from_data(). When NULL, independence \(P(A \cap B) = P(A)P(B)\) is assumed for all class pairs. Default NULL.

panel_map

Named list or NULL. If supplied, classes for each pathogen are restricted to the panel. When NULL, all classes present in marginals are used. Default NULL.

exclude_near_zero

Logical. Drop classes with marginal_resistance <= zero_threshold. Default TRUE.

zero_threshold

Numeric. Classes at or below this value are considered near-zero. Default 0.

top_n_classes

Integer or NULL. Cap classes per pathogen to top N by n_tested. Limits 2^n explosion. Default NULL.

lambda

Numeric. Ridge regularisation added to QP Hessian diagonal. Default 1e-8.

sigma_sq

Numeric. Uniform constraint variance. Default 1.

solver

Character. "osqp" (default, preferred) or "quadprog" (fallback).

n_cores

Integer. Parallel cores. Default 1L.

pathogen_col

Character. Default "pathogen".

class_col

Character. Default "antibiotic_class".

rate_col

Character. Default "marginal_resistance".

n_tested_col

Character. Default "n_tested".

Value

Tibble with one row per (stratum x) pathogen x profile:

Stratum columns

Any geography / year / outcome columns from marginals.

pathogen

Pathogen name.

profile_set_type

"aggregate_convex".

profile_class_set

Ordered class names joined by "|".

profile_delta

Binary profile label, e.g. "RSS".

profile_probability

Estimated probability \(\hat{p}_\delta\).

estimator

"convex".

convergence_flag

Logical; FALSE if QP failed and uniform distribution was returned.

identifiability_flag

Logical; TRUE when the constraint matrix is rank-deficient (system is underdetermined).

max_abs_residual

Maximum absolute constraint residual.

notes

Semicolon-separated notes: capped pairs, fallback pairs.

Details

$$\hat{p} = \arg\min_{p \in \Delta} \|Mp - v\|^2 + \lambda \|p\|^2$$

Stratification columns (geography, year, outcome) present in marginals are detected automatically and carried through to the output.

Examples

if (FALSE) { # \dontrun{
marg   <- compute_marginals_from_data(result$data_wide, result$col_map_resolved, panel_map)
pw     <- compute_pairwise_from_data(result$data_wide, marg, result$col_map_resolved, panel_map)
out    <- estimate_profiles_convex(marg, pw, panel_map)
} # }