Executes the optimized bootstrapping loop for smoking transition probabilities. This function pre-calculates mortality risks once to save memory, then iterates through `B` bootstrap samples, saving intermediate results to a temporary directory before combining them into final output tables.
Usage
run_bootstrap_pipeline(
config,
survey_data,
pops,
tob_mort_data,
tob_mort_data_cause,
B = 100,
seed = NULL
)Arguments
- config
A list containing model configuration parameters (e.g., country, years, ages).
- survey_data
A data.table or data.frame containing the base survey data.
- pops
A data.table containing population denominators.
- tob_mort_data
A data.table containing general tobacco mortality data.
- tob_mort_data_cause
A data.table containing cause-specific tobacco mortality data.
- B
Integer. The number of bootstrap iterations to run. Defaults to 100.
- seed
Integer. Master seed for the resampling. Taken from `config$seed` by the caller. Supplying it makes the whole run reproducible; passing NULL restores the old unseeded behaviour and warns, because an unseeded run cannot be reproduced or audited.
Value
A list containing seven data.tables with all bootstrap iterations combined:
init, quit, quit_no_init, relapse, net,
trend and survey_prev.
The master seed and the per-iteration seeds are attached as attributes.
Details
The fitted smoking trend surface is collected too. estimate_quitting has always fitted it on every iteration in order to solve for quitting; it just discarded it afterwards. Each replicate is thinned to the ages, years and smoking states the prevalence targets need before it is written to disk, because the full grid at B = 1000 is roughly 38 million rows.
Alongside the fitted surface, each iteration also saves the design-weighted survey aggregates for the same ages and years (aggregate_survey_prev), so the prevalence targets can be built either from the model fit or from the survey data directly, from the same draws under the same seed. Years inside the target range that the survey does not cover are checked against the original data once and must stay identical across iterations.
The central estimates written out by `process_country()` are bootstrap medians, not point estimates, so they are a function of the random draws. Without a seed, two runs of identical code on identical data return different numbers, and a diff against a previous delivery cannot separate a genuine change from resampling noise. This bit us in July 2026 when the project team queried initiation probabilities that had moved between deliveries.
Rather than seeding once and relying on the loop running in order, we draw `B` iteration seeds up front from the master seed and set the seed at the top of each iteration. Iteration `i` then depends only on the master seed and `i`, so results are identical whether the loop is run start to finish, resumed part way, or parallelised later.