NEWS

bacipair 0.1.0

Documentation and CI

The package has a logo.
README carries status badges and lists related packages.
CI runs lintr (report-only) and test coverage.

Bug fixes

synth_control_baci() no longer requires the simulator-only oracle column ite; it runs on a real panel and attaches ite/oracle_unit only when the input carries them (F1).

Contract

The synthetic-control estimator's donor-pool-too-small and insufficient-pre-treatment-years declines are now a classed bacipair_abstention condition (also carrying orchestra_refusal), per the orchestra's refusal/abstention naming convention, rather than a bare stop() (F4).

Infrastructure

Package is now under git version control and carries a standard R-CMD-check GitHub Actions workflow (F10).

Documentation

Added a README with a three-line purpose statement, install instructions and a minimal runnable example (F8).
Added an "Estimand and estimators" section to the getting-started vignette giving the ATT definition, the paired-DiD two-way fixed-effects specification, the CR1 cluster-robust variance formula and the simplex-constrained synthetic-control objective, all previously stated in prose only with no checkable mathematics (F9).
Added three captioned, interpreted figures to the getting-started vignette: pre-treatment fit and post-treatment gap for the best- and worst-fitting treated units, per-unit estimated ATT against the simulator oracle, and the donor-weight matrix (F12).
Restated the vignette's kill-gate claim: it now says plainly that the shipped synth_control_baci() is the uncorrected estimator, and quotes the uncorrected audit-grade bias (0.0347 pp against the 0.15 pp floor) rather than implying reproduction of a corrected extension that is not in the package API (F2).
Inlined the kill-gate floor and achieved-bias numbers directly in the vignette text instead of pointing the reader at deliverables/RDT-20/... paths that do not ship with the package (F13).
Replaced the vignette's unsupported "small but real residual bias" characterisation of the paired-DiD estimator with the measured relative bias (37 per cent of the oracle effect, about twice the cluster-robust standard error) (F14).
Expanded the vignette's unexplained CER/ACCU/HIR/MRV acronym cluster on first use, and removed a + used as a prose conjunction.
Corrected @param documentation of additionality_mean_pp / additionality_sd_pp in simulate_staggered_panel(): the draw is an untruncated Normal, not truncated-Normal as previously documented, and can realise a negative additionality for a pair flagged as genuinely additional (F3, documentation only -- the simulator's numerical behaviour is unchanged in this release).
Corrected the @return documentation of baci_paired_did()'s coefficients element: the returned vector holds the ATT interaction and non-base-year dummies on the demeaned design, not an intercept, treated-arm dummy or post-treatment dummy, all three of which are absorbed by demeaning and were never returned (F16).
Added @examples to the two exported functions that previously had none, is_bacipair_abstention() and write_oracle_recovery_log().

Initial scaffold release.

Public API

simulate_staggered_panel() -- staggered-treatment HIR-style BACI panel simulator with explicit oracle panel-average and per-unit ATT.
baci_paired_did() -- paired difference-in-differences on a BACI panel with proponent-cluster-robust standard errors (CR1).
synth_control_baci() -- time-varying synthetic-control comparator with simplex-constrained donor weights and rainfall / elevation / baseline-cover predictors.
proponent_cluster_bootstrap() -- proponent-level resampling bootstrap for the synthetic-control or paired-DiD point estimate.
goodman_bacon_decomp() -- Goodman-Bacon (2021) two-by-two decomposition of staggered-DiD weights, exposed as a guarded wrapper around bacondecomp::bacon().
bacipair_targets_plan() -- targets-skeleton helper that returns a list of targets::tar_target calls covering simulate -> SC fit -> paired-DiD fit -> bootstrap -> Goodman-Bacon decomposition -> oracle-recovery JSON log.

Reproducibility

Every public function is deterministic given a seed.
No silent error handling: invalid inputs trigger stop() with diagnostic messages.
R CMD check --as-cran clean (zero warnings, zero notes other than first-release notes).
Vignette getting-started.Rmd covers the synthetic ground-truth-recovery audit and the published kill-gate JSON log.

Citation

If you use this package, please cite both the package and the source methodology paper:

Macintosh, A., et al. (2024). Australian human-induced native forest regeneration carbon offset projects have limited impact on changes in woody vegetation cover and carbon removals. Communications Earth & Environment, 5, Article 149. https://doi.org/10.1038/s43247-024-01313-x