Skip to contents

Write ensemble trajectory channels to per-location CSV

Source: R/write_trajectory_csv.R
write_trajectory_csv.Rd

Exports the ensemble central daily trajectory of every captured channel (the artifact's $summary[[channel]]$median field, which is not always a median; see "What is lost") to a plain-text CSV, one file per location, as trajectories_<LOC>.csv.

Usage

write_trajectory_csv(
  trajectories,
  dir_out,
  channels = NULL,
  digits = 6L,
  verbose = TRUE
)

Arguments

trajectories

A mosaic_trajectories object, or a path to a trajectories_ensemble.rds file.

dir_out

Directory to write into; created if absent.

channels

Optional character vector selecting channels. Default NULL writes every channel present.

digits

Significant figures to round values to. Default 6.

verbose

Logical; print one message per file written.

Value

Invisibly, the character vector of written file paths.

Why this exists

run_MOSAIC() reduces the per-member channel arrays into a mosaic_trajectories object and persists it to 2_calibration/trajectories_ensemble.rds. That file is an R binary, so the MOSAIC-results promoter classifies it as heavy and records it store:"pending" — declared in the manifest but never committed. The consequence is that only reported_cases and reported_deaths (via the prediction CSVs) reach the archive, while incidence, new_symptomatic, disease_deaths, the compartments and the derived channels do not, and are unreadable outside R even when present.

A text CSV is copied into git by the existing promoter with no schema or promoter change.

Format and size

WIDE, not long: one row per (location, date) with one column per channel. A long table repeats location and date once per channel, which for 24 channels measured 2.72 MB against 582 KB wide on a national model (4.7x), and 410 KB against 202 KB after gzip.

Values are rounded to digits significant figures. At the default 6 this measured 582 KB per national model (202 KB packed) against 970 KB (373 KB packed) at full precision — model output does not carry 15 significant figures of information.

The per-member $lines component is deliberately NOT exported: at 6 significant figures it measured 62.7 MB raw / 7.8 MB packed for a single national model, which belongs in blob storage rather than git.

Deaths and the population balance

When the run integrated the reported CFR out (MOSAIC >= v0.96.0), reported_deaths and disease_deaths are the members' deaths redrawn from the calibrated CFR. The population compartments and N carry the engine's own fatal draws at the prior mu_jt (the fatal share p_fatal of symptomatic onsets, a few percent, never enters Isym), so N balances against those, not against the redrawn disease_deaths. disease_deaths are true deaths, i.e. reported deaths / rho_deaths, and rest on the pinned rho_deaths.

What is lost

The summary carries one central line per channel and no credible intervals. The central line is not the same statistic for every column:

  • reported_cases and reported_deaths follow the run's per-channel central_method (default weighted median for cases, weighted mean for deaths) over the engine-level member trajectories, matching predicted_central in the prediction CSVs; disease_deaths follows the deaths central_method.

  • mass_balance is a ratio of the compartments' weighted means, CFR a ratio of 28-day rolling sums of the weighted-mean reported deaths and cases, and epidemic_frac the weighted mean of the reconstructed epidemic flag.

  • I_total is the sum of the Isym and Iasym medians.

  • Every other channel is the weighted median.

Intervals for these channels require the per-member data. reported_cases and reported_deaths keep their intervals in the prediction CSVs.

Examples

if (FALSE) { # \dontrun{
# From a run directory, or to backfill an already-promoted model:
write_trajectory_csv(
  file.path(dir_output, "2_calibration", "trajectories_ensemble.rds"),
  file.path(dir_output, "3_results", "predictions")
)
} # }