Skip to contents

cfrnow

Real-time case fatality ratio (CFR) estimation from line-list data, using a Bayesian mixture-cure survival model. cfrnow is registered as an epidist model type, so the CFR and the onset-to-death delay both take brms formulas. You can put covariates (or a time-varying effect) on either.

Installation

Install the development version from GitHub with pak:

# install.packages("pak")
pak::pak("sbfnk/cfrnow")

pak pulls in the packages cfrnow builds on (epidist, brms, primarycensored and distspec), including the GitHub versions it needs.

Fitting runs through CmdStan, so you also need cmdstanr and a CmdStan install:

# install.packages("cmdstanr", repos = c("https://stan-dev.r-universe.dev", getOption("repos")))
cmdstanr::check_cmdstan_toolchain()
cmdstanr::install_cmdstan()

Why

The naive deaths / cases ratio runs low in real time, because recent cases have not had time to die yet. Switching to resolved cases, deaths / (deaths + recoveries), overshoots the other way: deaths resolve faster than recoveries, so the resolved set is top-heavy with deaths.

cfrnow conditions on neither. Each case is fatal with probability cfr; a fatal case dies after an interval-censored onset-to-death delay F; and a case still unresolved at the cut-off is right-censored, contributing 1 - cfr * F(t) (it is either non-fatal or fatal but not yet resolved). Hold F fixed and you recover the Ghani/Nishiura estimator (Ghani et al. 2005; Nishiura et al. 2009); here the model co-estimates F and carries its uncertainty through. Give it recovery dates and the two-outcome fit uses onset-to-recovery timing as well. primarycensored handles the onset interval-censoring and the real-time right-truncation of F.

Usage 

library(cfrnow)
#> Loading required package: distspec
#> 
#> Attaching package: 'distspec'
#> The following objects are masked from 'package:stats':
#> 
#>     Gamma, sd

# your own onset_date / death_date line list, or a simulated one
set.seed(1)
ll <- simulate_linelist(
  n = 400, cfr = 0.55, onset_days = 45,
  delay = Gamma(mean = 12.75, sd = 7)
)
d <- prepare_cfr_data(ll, obs_time = as.Date("2026-02-20"))
c(cases = d$n_cases, deaths = d$n_deaths, censored = d$n_cens)
#>    cases   deaths censored 
#>      400      180      220

Supply the onset-to-death delay and a cfr_prior as distspec distributions. A native delay parameter can be a Normal() prior (co-estimated) or a fixed number (held fixed; fixing the whole delay gives the Ghani/Nishiura estimator). The cfr_prior is a Beta() and matters because the CFR is weakly identified early on: Beta(1, 1) is uniform, Beta(1, 9) favours a low CFR, and Beta(6.6, 13.4) (mean 0.33) suits a high-fatality pathogen.

onset_to_death <- LogNormal(meanlog = Normal(2.41, 0.2), sdlog = Normal(0.51, 0.15))
fit <- fit_cfr(d, delay = onset_to_death, cfr_prior = Beta(1, 1))
summary(fit)
#>     quantity   mean   q2.5    q50  q97.5  rhat ess_bulk
#> 1        cfr  0.560  0.505  0.560  0.615 1.001     3100
#> 2 delay_mean 12.700 11.600 12.700 13.900 1.000     3300
#> 3   delay_sd  7.100  6.200  7.000  8.200 1.000     3000

Swap the family, fix the delay, or add recovery timing. And because the model is fitted through epidist, you can put covariates (or a smooth on time) on the CFR via a formula:

fit_cfr(d, delay = Gamma(shape = Normal(3, 1), rate = Normal(0.25, 0.1))) # gamma
fit_cfr(d, delay = LogNormal(meanlog = 2.41, sdlog = 0.51)) # fixed-F
fit_cfr(d,
  delay = onset_to_death,
  recovery_delay = LogNormal(Normal(2.9, 0.3), Normal(0.5, 0.2))
) # two-outcome

# covariate / time-varying CFR: carry the column through the prep, then
# reference it in the formula (the onset date rides along as `onset`).
dc <- prepare_cfr_data(ll, obs_time = as.Date("2026-02-20"), covariates = "age_group")
fit_cfr(dc, delay = onset_to_death, formula = brms::bf(mu ~ 1, cfr ~ age_group))

Pass obs_time = NULL to prepare_cfr_data() for a retrospective fit (it reduces to the naive ratio with the delay as a nuisance).

Your line list needs onset_date and death_date (NA for cases that have not died). Optional onset_lower/onset_upper widen the onset censoring, and a recovery_date column switches on the two-outcome fit that also times recoveries. summary() reports rhat/ess_bulk (and recovery_mean/_sd when recoveries are used) and flags cfr_low_information when the CFR posterior has barely moved from its prior, which is expected early on, when the estimate is prior-driven.

Assumptions and caveats

The correction fixes a timing bias; it cannot repair the data. Worth reading before you quote a number:

  • Deaths need to be ascertained completely. A death that never reaches the line list (a community death outside a treatment centre, say) looks like a survivor and drags the CFR down. This is the main worry wherever ascertainment is centred on treatment centres.
  • Put the death notification date in death_date, not the true date of death. A case counts as a death once its death_date is on or before the cut-off, so the notification date lets the censoring absorb any reporting lag and keeps cfr correct. The delay the model then works with is onset-to-notification, so read delay_mean/delay_sd (and any biological delay prior) in that light. With no reporting lag, plain death dates are fine.
  • The delay and CFR are taken as stationary and homogeneous unless you say otherwise: one delay and one CFR for the whole outbreak. Put a formula on cfr (a smooth on time, say) to let it vary.
  • The delay family is a modelling choice, not something the fit learns. Gamma and lognormal differ in tail weight (how fast a recent case counts as “probably cured”), and sparse data can’t tell them apart, so refit with the other one to check sensitivity.
  • Recovery timing leans on complete discharge data. In the two-outcome fit a case that really recovered but whose recovery went unrecorded stays censored and drifts toward the fatal branch over time, biasing cfr up. The death-only default (no recovery_date) sidesteps this, so prefer it where discharge recording is patchy.