Skip to contents

Weight health and convergence status of a finished calibration, from its output directory

Source: R/calc_weight_health.R
calc_run_weight_health.Rd

Recomputes calc_weight_health (and so the 1.4.0 calc_convergence_status) for an existing run_MOSAIC() output directory from its stored likelihoods, so runs produced before the diagnostic existed (MOSAIC 1.1.x, 1.3.0) can be assessed without re-running anything. It reads 2_calibration/samples.parquet, 1_inputs/config.json and 1_inputs/control.json (and 3_results/summary.json when present, for the location label). It never modifies those files.

Usage

calc_run_weight_health(
  dir_output,
  write = TRUE,
  coarsening_alpha = NULL,
  weight_col = NULL,
  n_batches = 10L,
  best_subset_ratio = NULL,
  ess_best_target = NULL,
  min_members = NULL,
  max_members = NULL,
  thresholds = NULL
)

Arguments

dir_output

A run_MOSAIC() output directory.

write

Logical; write 3_results/weight_health.json. Default TRUE.

coarsening_alpha

Effective number of independent observations of the posterior. Default NULL means the run's control$targets$coarsening_alpha, or the package default for a run that predates it.

weight_col

Deprecated since MOSAIC 1.4.0 and ignored (with a warning): the status is assessed on the best-fit members.

n_batches

Number of seed batches. Default 10.

best_subset_ratio, ess_best_target, min_members, max_members

Optional overrides of the best-fit settings (see Details).

thresholds

Optional named list overriding the status bands.

Value

Invisibly, a list with location, dir_output, timestamp, fields (the flat summary.json schema 2 fields) and weight_health (the full calc_weight_health() result).

Details

n_obs is the number of scored location-day cells with a finite cases or deaths observation inside the scored window, after the burn-in and the deaths-era start of control$likelihood.

The best-fit settings come from the run's control when it is a 1.4.0 control (it carries targets$status_thresholds): targets$best_subset_ratio, targets$ESS_best, targets$min_ensemble_members, targets$max_ensemble_members and targets$status_thresholds. An older control is assessed on the 1.4.0 defaults, because its ESS_best and member caps meant something else. Arguments override both. A run is fixed-budget when its control has a positive calibration$n_simulations; an adaptive run has exhausted its budget when its draws reach calibration$max_simulations_total. The growth slope is replayed over the draws in simulation order (.mosaic_ess_replay(), as for a fixed-budget run), so a stalled growth is flagged. Simulations attempted are the rows of samples.parquet.

See also

calc_weight_health, calc_convergence_status

Other calibration-metrics: calc_convergence_diagnostics(), calc_convergence_status(), calc_model_agreement_index(), calc_model_cvw(), calc_model_ess(), calc_weight_health()

Examples

if (FALSE) { # \dontrun{
wh <- calc_run_weight_health("output/national/MOZ/v2026-10-08.01")
wh$fields$convergence_status
wh$fields$convergence_reasons
} # }