Skip to contents

Changelog

Source: NEWS.md

cfcompare 0.2.1

  • Package website built with pkgdown: https://takuma1102.github.io/cfcompare/.
  • The comparison tables of the TROP paper, the official Python package, the official Stata command and cfcompare (as of July 2026 and as of October 2026), together with the notes on this package’s implementation choices, moved from the README to the article “TROP implementations compared: the paper, Python, Stata and R”; the README links to it.

cfcompare 0.2.0

Alignment with Stata trop v0.2.9 and Python trop v0.1.11 (October 2026)

Changes since the July 2026 development release on GitHub / R-universe, following the official Stata command’s (v0.2.9) and Python package’s (v0.1.11) summer releases; see the comparison tables (as of July 2026 and as of October 2026) in the article “TROP implementations compared: the paper, Python, Stata and R”.

  • trop(anchor = "pooled") now pools over treatment spell groups – treated cells with the same start and end period, pooled across units – solving eq. (2) once per group with weights anchored to that group, and returns the per-group effects in $tau_groups (start/end, unit and cell counts, estimate; the ATT is their cell-weighted average). This is the group(time) estimand of the official Stata command (v0.2.6+). Simultaneous permanent adoption is a single group, so block-design results are unchanged; staggered adoption now gets one anchor per cohort and a unit that switches in and out of treatment one anchor per spell, instead of a single anchor spanning the whole treated set. The group unit distance is to the average path of the group’s units over the periods in which none of them is treated, with each comparison unit’s own treated cells masked (the Stata masking); the placebo-CV criterion, the event study, counterfactual_matrix() and autoplot() use the same grouping.
  • Pooled time weights are now centred on the true midpoint of the spell, (first + last) / 2 (the Stata command’s convention since v0.2.4), so a single treated cell gets the same estimate under the pooled and per-cell anchors. The previous centre, half a period later (T - tp/2 in 0-based indexing, the official Python package’s convention), remains available as trop_control(time_center = "python") and is what trop_matrix() always uses, for like-for-like checks against that package.
  • New trop_control(pooled_treat_distance = list(unit =, time =)): overrides the distances of a pooled group to itself – the members’ unit distance (otherwise their distance to the group average, hence generally non-zero) and the in-spell time distance – e.g. list(unit = 0, time = 0) gives the group’s own units and periods the full weight. The equivalent of the Stata command’s pooled_treat_distance(unit(#) time(#)) option (v0.2.6+); it applies to the pooled anchor, the placebo CV and the pooled counterfactual, is stamped into $lambda like the nuclear-norm floor so downstream re-solves reproduce it, and is reported by print().
  • New trop_control(boot_resample = c("stratified", "controls")): "controls" keeps the treated units fixed and resamples only the control units (the default resample = "controls" of the Python package’s trop.bootstrap, v0.1.11+); "stratified" (default, unchanged) resamples treated and control units separately. Applies to the ATT and event-study bootstraps; the fit records the scheme in $boot_resample. The resampling now indexes through sample.int(), so a single treated unit is resampled from itself rather than from 1:unit.
  • se = "jackknife" now deletes one unit at a time over every unit (treated and control), as the Stata command’s vce(jackknife), instead of over the treated units only; replicates that leave no treated unit (or fail) are skipped and the variance uses the number of finite replicates. It therefore no longer requires two treated units. The event-study jackknife follows the same scheme. (se = "auto" still picks the jackknife when there are at least two treated units and the bootstrap otherwise.)

First development release on GitHub / R-universe (July 2026)

  • Native R implementations of the Triply RObust Panel (TROP) estimator of Athey, Imbens, Qu & Viviano (2026).
  • panel_compare() runs DID, SDID, SC, MC, TROP, gsynth and augsynth on one dataset and returns a common tidy ATT schema with shared plots.
  • trop_matrix() provides a matrix-in form whose results can be numerically compared with the official Python trop matrix-in signature for direct cross-checking; verified to match the reference to numerical tolerance on the weighted-TWFE special case.* autoplot() on a trop fit now draws the TROP time weights theta_s = exp(-lambda_time * |t - s|) as a filled band along the bottom (synthdid-style), replacing the earlier bar strip, and labels it so the band reads as time weights rather than observation counts. The band is shown for uniform weights too (lambda_time = 0, flat) and can be hidden with show_weights = FALSE.
  • compare_se_modes() compares the TROP bootstrap, jackknife and placebo standard errors on a single fit: the cross-validated penalties are chosen once and reused, so the point estimate is identical across rows and only the uncertainty differs (the same fixed-penalty pattern as trop_ablation()).