Stress testing under different missingness mechanisms

Missingness patterns implemented in vrc_simulate

The vrc_simulate() function supports several additional missingness patterns on top of the explicit reporting model:

• Block dropout: missing region-time blocks (for example, system outages) with probability driven by conflict intensity and facility functioning.
• Age-selective dropout: preferential loss of older non-trauma deaths after conflict begins.
• MNAR overload: missing-not-at-random dropout driven by true death burden.

These mechanisms are designed to mimic failure modes commonly seen in VR and facility-based data streams.

Compare scenarios

library(vrcmort)

base <- vrc_simulate(R = 5, T = 48, t0 = 20, seed = 1, missing = list(type = "none"))
block <- vrc_simulate(R = 5, T = 48, t0 = 20, seed = 1, missing = list(type = "block", block_intercept = -1.5))
age <- vrc_simulate(R = 5, T = 48, t0 = 20, seed = 1, missing = list(type = "age_selective", age_dropout_strength = 1.0, age_dropout_old_from = 6))
combo <- vrc_simulate(R = 5, T = 48, t0 = 20, seed = 1, missing = list(type = "combined", block_intercept = -2.0, age_dropout_strength = 1.0, mnar_strength = 1.0))

c(
  base = mean(is.na(base$df_full$y)),
  block = mean(is.na(block$df_full$y)),
  age = mean(is.na(age$df_full$y)),
  combo = mean(is.na(combo$df_full$y))
)
#>       base      block        age      combo 
#> 0.00000000 0.28750000 0.01080729 0.21718750

Fitting and evaluation

A typical stress-test workflow is:

1. simulate a scenario,
2. fit the model,
3. compare inferred reporting collapse and latent mortality to truth.

fit <- vrc_fit(
  data = combo$df_obs,
  t0 = combo$meta$t0,
  mortality_covariates = ~ facility,
  reporting_covariates = ~ facility,
  chains = 4,
  iter = 1000,
  seed = 1
)

plot_reporting(fit)
plot_mortality(fit, value = "true_deaths")