Package index • hvtiRbootstrap

Resample and fit — %bootreg

Start here. Fits a model on each of many bootstrap replicates and records which terms each one kept. PROC= becomes the fitter argument, and a term the model did not select is left missing — which is what makes the counts downstream read as selection frequencies.

boot_select(): Build a model by bootstrap resampling

Summarise the replicates — %SUMBOOT and %cluster

Reach for these once you have replicates. boot_summary() gives the per-variable selection frequency and coefficient distribution; boot_clusters() answers the question it cannot — how often at least one member of a correlated group survived. Both are held to exact parity with the macros.

boot_summary(): Summarise bootstrap replicates into selection frequencies
boot_clusters(): Aggregate selection frequencies over clusters of correlated variables

Report over a screen

The model-agnostic reporting layer, extracted so that the logistic, linear, Cox and hazard bootstrap reports are thin templates rather than near-copies of one another. boot_validate() checks that the screen a runner wrote has the shape a report reads; the rest turn it into tables. The ones that group their rows take an optional phase, so a multiphase hazard screen and a single-phase logistic screen run down the same code path; boot_dropped() and boot_health() have nothing to group and take none. boot_bag() is the way in from this package’s own screen: boot_select() returns a wide object and everything here reads a long one.

boot_bag(): A screen the reporting layer can read
boot_validate(): Check that a bootstrap bag has the shape a report reads
boot_provenance(): Where a bootstrap screen came from
boot_seeds(): Every seed a bootstrap screen used
boot_frequencies(): Selection frequencies, with the error they carry
boot_dropped(): Candidates the screen never saw
boot_concepts(): Selection frequencies grouped by concept
boot_health(): Did the bootstrap screen actually run?

Pool a chunked run

Needed only when a screen was run in chunks, which is how a run of days is made restartable. boot_pool_chunks() folds them into one object of the same shape, refusing to pool chunks that did not draw from the same data or run the same screen; boot_shortfall() says whether what you pooled is the run you launched, which nothing inside a chunk can know.

boot_pool_chunks(): Pool chunked bootstrap runs into one screen
boot_chunk_files(): Find the chunk files of a chunked bootstrap run
boot_shortfall(): Is a pooled screen the run that was launched?

Fitters — one per PROC=

Passed to boot_select(), one per model family. Use these when choosing the model type; write a new one when adding a family, rather than a new pipeline. Each returns the coefficients its model kept, or NULL when the fit failed and the replicate should be redrawn.

fit_linear(): Fit a linear model for one bootstrap replicate
fit_logistic(): Fit a logistic model for one bootstrap replicate
fit_cox(): Fit a Cox proportional-hazards model for one bootstrap replicate

Band an estimate — %BNMNR and %BNPREV

The other branch. Everything above selects variables: refit on each replicate, count which terms survived, and a term the model did not choose is left missing. This does not select anything — it resamples to put a band around an estimate, so the replicates are a distribution rather than a vote and there is no missingness to read. statistic is the fitter contract with the selection semantics removed. No argument sets a coverage level, because no macro in the family takes one: both the 95% and the 68% band come back, in columns named for their coverage.

boot_predict_ci(): Band an estimate by bootstrap resampling

Package overview

The shape of the port: three macros and three functions at the core, with the pooling and reporting layers that grew around them.

hvtiRbootstrap hvtiRbootstrap-package: hvtiRbootstrap: Bootstrap Model Building for the HVTI CORR Group