Skip to contents

Changelog

Source: NEWS.md

aftsplex 0.3.3

  • two_stage_bootstrap() now returns a classed "two_stage_bootstrap" object with a summary() method: a quantile table of the corrected curve and its 95% percentile CI, plus the replicate yield, the exposure anchor, and the Phase-1 sigma_w_sq spread across the replicates whose calibration succeeded. Previously summary() fell through to the generic list output. The returned object also gains an x_ref element; existing element access is unchanged.

aftsplex 0.3.2

aftsplex 0.3.1

Robustness fixes surfaced by an adversarial stress-test of 0.3.0. All changes are input guards and a bug fix; valid-input behaviour and the byte-identical generate_aft_data() baseline are unchanged.

  • two_stage_bootstrap() no longer crashes on the serial path (workers = 1, the default) when bootstrap replicates fail. A failed replicate returned NULL and list[[r]] <- NULL deleted the element, shrinking the result list and triggering “missing value where TRUE/FALSE needed”; the documented drop-and-warn behaviour now works on both the serial and parallel paths.
  • fit_me_calibration(), gls_combine(), compute_ise(), sigma_u_for_reliability(), build_sigma_u(), and generate_aft_data() now validate their inputs and raise informative errors instead of silently propagating NaN/Inf or returning a meaningless result: constant/too-small validation samples, non-finite or uninformative (zero-slope) surrogates, singular surrogate covariances, unsorted or length-mismatched compute_ise() arguments, non-positive var_x/ratios/sigma_u_sq, out-of-range rho, and negative x_sd.
  • Clarified that the serial (workers = 1) bootstrap is the canonical bit-for-bit reproducible path; parallel runs use independent L’Ecuyer streams that are reproducible across worker counts but differ from the serial result for the same seed by design.
  • Documentation: new applied-lpa article presenting the dose-response on the applied scale – the time-acceleration ratio against light-physical-activity minutes, anchored at a 200 min/day reference, with the lower-support caveat and a two-stage vs single-stage interval-width comparison. The quickstart and the simulation-study daily-scale section now use the manuscript’s daily-LPA log scale (x_mean = log(300), x_sd = 0.45), correcting an earlier raw-minute parameterisation. The faq article gains a parallel-execution / reproducibility note.

aftsplex 0.3.0

Realistic-scale data-generating process, two new simulation diagnostics, and a robustness pass on the bootstrap/SIMEX failure modes.

  • generate_aft_data() gains x_mean, x_sd, and f_args so the latent exposure and dose-response can be expressed on a meaningful scale (e.g. daily light-physical-activity minutes, mean ~300 with a 150-minute threshold). The defaults reproduce the previous arbitrary-unit output bit-for-bit.
  • New exported sigma_u_for_reliability() returns a measurement-error covariance targeting a given mean per-surrogate reliability, for sweeping measurement quality (e.g. 0.5 / 0.65 / 0.8).
  • The simulation-study vignette gains a “daily-minutes scale” section with a reliability sweep (Oracle / Naive / SIMEX ISE across reliability) and a per-surrogate ladder (each surrogate alone vs the GLS combiner via the existing surrogate_pattern argument; no engine change).
  • Robustness pass (a user hit an all-NA bootstrap curve with no message, then an opaque error at plotting). two_stage_bootstrap() now isolates each replicate so a degenerate resample drops to NA instead of aborting the run, and emits a single summary warning (replicates dropped, grid points with no finite value, all-NA point estimate, or all replicates failed). The full-sample calibration gives a clear message on collinear/constant surrogates instead of a raw Lapack error, and pointwise summaries are NA-safe.
  • simex_aft_spline() returns NA at grid points with too few converged perturbed fits (one summary warning) instead of erroring with “0 (non-NA) cases”; a failed naive fit now gives an informative error.
  • plot_curves() drops entirely-NA curves with a warning and stops with a clear message when nothing is left to plot, instead of an empty panel.
  • New “Troubleshooting” article documenting every warning/error message, and a “coverage diagnostics” article.

aftsplex 0.2.0

Revision round addressing reviewer feedback. All additions are backward-compatible via defaults; existing output and the SIMEX regression are unchanged.

  • two_stage_bootstrap() gains nested to toggle between the full two-stage (double) bootstrap and a single-stage interval that holds the Phase-1 calibration fixed, so the Phase-1 contribution to CI width can be quantified.
  • two_stage_bootstrap() and simex_aft_spline() gain x_ref to anchor the centred curves (and every bootstrap replicate) at a clinically meaningful reference exposure before quantiles are taken.
  • two_stage_bootstrap() and simex_aft_spline() gain support_probs, which flags grid points outside the exposure support (returned as in_support) and warns that they are spline extrapolations rather than fits.
  • plot_curves() and plot.simex_aft_spline() gain time_ratio to plot the exponentiated (time-ratio) curve with a reference line at 1, and render out-of-support segments dashed with a lighter ribbon.
  • two_stage_bootstrap() gains workers for parallel execution of the outer bootstrap loop via future.apply (per-task L’Ecuyer streams; serial default unchanged). Adds future and future.apply to Suggests.

aftsplex 0.1.0

  • Initial release.