Build the daily reported-CFR matrix mu_jt from annual estimates

Expands per-location, per-year reported case fatality ratios (from est_CFR_hierarchical) into the [location x day] matrix the engine reads as config$mu_jt. Values are interpolated linearly on the logit scale between mid-year points, and held flat before the first and after the last estimated year.

Usage

make_mu_jt(
  cfr_estimates,
  location_name,
  date_start,
  date_stop,
  interpolation = c("linear_logit", "step")
)

Arguments

cfr_estimates: Data frame with columns iso_code, year and either logit_mean or cfr_estimate (for example est_CFR_hierarchical()$predictions, or the cfr_hierarchical_estimates.csv it writes).
location_name: Character vector of ISO codes, in config row order.
date_start, date_stop: First and last simulation dates (Date or character).
interpolation: "linear_logit" (default) or "step" (the calendar-year value applies to every day of that year).

Value

Numeric matrix with length(location_name) rows and one column per day from date_start to date_stop; every value in (0, 1).

Details

Days past the last estimated year carry that year's value forward, so a simulation window that runs beyond the WHO annual record is filled with the most recent estimate rather than an extrapolated trend.

The values are the GAM's logit-scale centres, i.e. the median of each year's reported-CFR distribution, not its mean. That is the exact prior centre for the integrated deaths likelihood (whose year deviations have mean zero on the logit scale). A simulation that draws deaths directly at mu_jt – a scenario run or a prior-predictive check – runs 5-10% below WHO-annual deaths (2018-25); the logit-normal mean would over-predict, because annual CFR is lower in high-case years.

The default "linear_logit" centre and the integrated likelihood's year deviations use different time bases on purpose. The deviations are calendar-year levels blended over 30 days either side of 1 January (.d7_basis() in calc_log_likelihood_deaths_integrated.R), while the centre interpolates between 1 July anchors, so from January to June the centre sits partway toward the previous year's value. The resulting offset is at most the year-to-year change in the GAM centre (under 0.07 logit, about 3 percent of the CFR, for 2022-26 estimates) and is absorbed by the year deviation; years past the last estimate are held flat, so forecast windows are unaffected.

Examples