CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

• BaconDecomposition R parity goldens. Closes the PR-B deferral row in TODO.md. JSON goldens at benchmarks/data/r_bacondecomp_golden.json generated from the committed benchmarks/R/generate_bacon_golden.R script (3 fixtures: uniform_3groups_with_never_treated, two_groups_no_never_treated, always_treated_remapped) against bacondecomp 0.1.1 on R 4.5.2. tests/test_methodology_bacon.py::TestBaconParityR now active (4 tests, no skips): TWFE coefficient parity at atol=1e-6 across all 3 fixtures; weights-sum parity at atol=1e-6 across all 3 fixtures; per-component estimate + weight parity at atol=1e-6 on the 2 non-remap fixtures and on the 6 timing-vs-timing rows of always_treated_remapped (carve-out narrowed to U-bucket rows only); plus a dedicated fold-back test (test_always_treated_remapped_fold_back_matches_r) that pins the documented convention divergence on always_treated_remapped (R keeps first_treat=1 as a distinct timing cohort and emits Later vs Always Treated comparisons; Python's paper-footnote-11 convention remaps those units to U and folds them into a single treated_vs_never cell per treated cohort) by aggregating R's split rows per cohort and asserting they match Python's single fold at atol=1e-6. The aggregate is invariant per Theorem 1; the per-component breakdown differs structurally between conventions but the fold-back is now directly asserted. New **Note (R parity convention divergence on always-treated)** and **Deviation (first-period boundary extension on always-treated remap)** in docs/methodology/REGISTRY.md. First-period boundary deviation: the paper uses strict t_i < 1 for the always-treated bucket; the library uses the inclusive first_treat <= min(time) rule and folds first_treat == min(time) cohorts into U. R does NOT apply this fold (it keeps such cohorts as their own bucket). When min(time) > 1 the rules coincide. Explicitly labeled in REGISTRY's Deviations block and mirrored in METHODOLOGY_REVIEW.md and bacon.py. METHODOLOGY_REVIEW.md tracker row promoted **Complete** (R parity goldens pending) → **Complete**.
• generate_ddd_panel_data — panel-structured DGP for Triple-Difference power analysis (diff_diff/prep_dgp.py). New public function exported from diff_diff and diff_diff.prep for panel DDD simulations. Cross-sectional generate_ddd_data remains available unchanged. Produces a balanced panel of n_units × n_periods with two unit-level binary dimensions (group, partition) and a derived post = 1[period >= treatment_period] indicator; columns: unit, period, outcome, group, partition, post, treated, true_effect (+ x1, x2 when add_covariates=True). DDD-CPT identification holds because the group * partition interaction enters as a unit-level (time-invariant) term, leaving the triple-interaction treatment_effect * group * partition * post as the sole source of differential group × partition trend. Compatible with TripleDifference(cluster="unit").fit(..., time="post") (the cluster kwarg is required because TripleDifference is the repeated-cross-section panel=FALSE estimator and unclustered SE on panel-generated rows understates variance under within-unit serial correlation; the point estimate att is invariant to clustering — see the new TripleDifference REGISTRY note on panel-shaped input). Users get panel-realistic unit fixed effects and within-unit serial correlation while the binary 2×2×2 estimator surface is unchanged. Stratified allocation: the partition split is drawn stratified-by-group at the requested partition_frac so every (group, partition) cell receives at least one unit; a targeted ValueError is raised at fit-time when the rounded cell counts (n_units, group_frac, partition_frac) would leave any cell empty. This guarantees the 2x2x2 DDD surface is populated for any valid input — independent marginal sampling (the cross-sectional generate_ddd_data convention) could collapse cells when marginals are small (e.g., n_units=4, group_frac=partition_frac=0.25). Validates 1 <= treatment_period < n_periods, group_frac and partition_frac strictly in (0, 1), and n_units >= 4. Deterministic recovery (noise_sd=0) matches treatment_effect to ~1e-15 (covered by tests/test_prep.py::TestGenerateDddPanelData, 16 tests including infeasible-config rejection and smallest-feasible-config round-trip through TripleDifference.fit). power.simulate_power is NOT yet auto-routed to the panel DGP for TripleDifference (the existing _ddd_dgp_kwargs registry entry still ignores n_periods and the existing _check_ddd_dgp_compat warning still fires on non-default kwargs) — that wiring is tracked as a follow-up in TODO.md.
• BaconDecomposition: Goodman-Bacon (2021) methodology audit (PR-B). Closes the BaconDecomposition row in METHODOLOGY_REVIEW.md (status flipped from In Progress → Complete — initially with an R-parity-goldens caveat that was closed by the parity-goldens bullet above in this same release). Builds on the PR #451 paper review at docs/methodology/papers/goodman-bacon-2021-review.md. Audit outcomes: (1) Rewrote _recompute_exact_weights in bacon.py to actually implement Theorem 1 (Eqs. 7-9 + 10e-g) — the prior "exact" implementation was missing the (1-n_kU) factor in the subsample variance, did not square the sample share, and added an extraneous unit_share factor not present in the paper; the post-hoc sum-to-1 normalization masked the relative-weight error but produced ~0.3% decomposition error vs TWFE on a 3-cohort + never-treated DGP. The rewrite computes the exact numerators of Eqs. 10e/f/g and lets the post-hoc normalization handle the V̂^D denominator (Theorem 1's identity guarantees V̂^D = Σ numerators). The TWFE-vs-weighted-sum identity now holds at atol=1e-10 on both noisy and hand-calculable DGPs. (2) Added always-treated warn+remap per paper footnote 11: units whose first_treat is at or before the first observable period (first_treat <= min(time), excluding the never-treated sentinels 0 and np.inf) are automatically remapped to the U (untreated) bucket via an internal column (__bacon_first_treat_internal__) with a UserWarning. Detection uses ordered-time logic on the time axis, so panels whose time column has negative or zero-crossing labels (event-time encodings) are handled correctly; the 0 sentinel restriction applies only to first_treat, not to time, and a real treatment cohort with first_treat == 0 would still be folded into U today (re-label such cohorts to a non-sentinel value before fitting). The user's original first_treat column is preserved unchanged. The count is surfaced as a new BaconDecompositionResults.n_always_treated_remapped dataclass field, rendered in summary() output when nonzero. n_never_treated reports TRUE never-treated only, computed from the original user column before remap — remapped always-treated units appear separately as n_always_treated_remapped, no double-counting. (3) New methodology test file tests/test_methodology_bacon.py (34 tests across 6 classes post-release; the audit added ~24 tests and the R-parity-goldens bullet above expanded coverage: TestBaconHandCalculation hand-checks Eqs. 7-9 + 10b-d on a minimal balanced panel at atol=1e-10; TestBaconParityR (4 tests, all active post-release once the R parity goldens bullet above landed; skips cleanly with a regenerate-instructions pointer in partial-checkout scenarios where the JSON is unavailable); TestBaconAlwaysTreatedRemap regression-tests warn+remap mechanics including user-data-preservation; TestBaconEdgeCases exercises no-untreated, single-cohort, unbalanced panel, constant-ATT recovery; TestBaconWeightModes locks the new exact-is-default contract; TestBaconSurveyDesignNarrowing confirms survey_design composes with exact mode and warn+remap). (4) R bacondecomp::bacon() parity generator committed at benchmarks/R/generate_bacon_golden.R covering three DGP fixtures (3-groups-with-U, 2-groups-no-U, always-treated-remapped); the JSON goldens deferral at audit time was closed in this same release by the parity-goldens bullet above. (5) docs/methodology/REGISTRY.md ## BaconDecomposition block replaced with the paper-review-sourced entry plus three new sub-notes: weight modes (exact vs approximate), always-treated remap, R parity status. Explicit removal: the prior REGISTRY block's "Weights may be negative for later-vs-earlier comparisons" claim was incorrect per Theorem 1 (decomposition weights are strictly positive and sum to 1; negative weights are an estimand-level phenomenon, not estimator-level) and is dropped from the new entry. Closes the BaconDecomposition follow-up tracked at TODO.md (the prior row added in PR #451 is replaced by a narrower R-parity-goldens deferral row).
• SpilloverDiD — ring-indicator spillover-aware DiD (Butts 2021). New standalone estimator at diff_diff/spillover.py implementing two-stage Gardner methodology with ring-indicator covariates that identify direct effect on treated (tau_total) alongside per-ring spillover effects on near-control units (delta_j). Documented synthesis of ingredients (no single published software covers the exact recipe — did2s implements Gardner two-stage without rings; the Butts ring estimator has no R/Stata package): Butts (2021) Section 5 / Table 2 identification, Gardner (2022) two-stage residualize-then-fit, and the Conley spatial-HAC vcov shipped in 3.3.3. Handles both panel non-staggered (Equations 5/6/8) and Section 5 staggered timing in one estimator — non-staggered is the special case where all treated units share an onset time. API: SpilloverDiD(rings=[0, 50, 100, 200], conley_coords=("lat","lon"), ...).fit(data, outcome="y", unit="unit", time="t", treatment="D") (binary D auto-converted to first_treat) or .fit(..., first_treat="first_treat") (Gardner convention). Result: SpilloverDiDResults(DiDResults) with .att = tau_total, .spillover_effects (per-ring pd.DataFrame with coef/se/t_stat/p_value/ci_low/ci_high), .ring_breakpoints, .d_bar, .n_units_ever_in_ring, .n_far_away_obs, .is_staggered. .coefficients exposes all (1+K) stage-2 entries ("treatment" + "_spillover_<ring_label>") plus an "ATT" alias keyed to vcov columns. Methodology spec (committed): stage-2 regressor is the time-varying (1 - D_it) * Ring_{it,j} form (paper page 12's S_it = S_i * 1{t >= t_treat} notation; Section 5 Table 2's S^k_{it} / Ring^k_{it,j}). Reading the literal unit-static (1 - D_it) * S_i from Equation 5 is algebraically rank-deficient under TWFE ((1-D_it) * S_i = S_i - D_it, with S_i absorbed by mu_i, leaving -D_it); only the time-varying form supports the paper's identification (Proposition 2.3). Stage-1 subsample uses Butts' STRICTER Omega_0 = {D_it = 0 AND S_it = 0} (untreated AND unexposed), not TwoStageDiD's {D_it = 0} alone — this prevents spillover-contaminated near-controls in pre/post periods from biasing the time FE. Gardner identity (non-staggered): a 20-seed deterministic regression test pins SpilloverDiD.att against a direct single-stage TWFE ring regression on the full sample (y ~ mu_i + lambda_t + tau * D_it + sum_j delta_j * (1 - D_it) * Ring_{it,j}) at atol=1e-10 — empirically bit-identical, so the reported non-staggered tau_total IS the Butts Eqs. 4-6 estimator. Identification-check policy (period strict, unit warn-and-drop, plus connectivity): every period must have at least one Omega_0 row (hard ValueError — dropping a period removes all units' cross-time identification). Units lacking Omega_0 rows (e.g. baseline-treated units with D_it = 1 at every observed t) are warned-and-dropped: their unit FE is NaN, residualization writes NaN on their rows, and the downstream finite-mask path excludes them from stage 2 — mirrors TwoStageDiD's always-treated convention. Additionally, the supported-units bipartite graph (units linked by shared Omega_0 periods) must form a single connected component; K > 1 components raise ValueError because the FE solver would return only component-specific constants and residualization would silently mix them across components (defense-in-depth — under absorbing treatment the disconnected case may be unreachable through the upstream validators, but the check future-proofs Wave B follow-ups). Public API restrictions (Wave B MVP): covariates= raises NotImplementedError because Gardner-style two-stage requires covariate effects estimated on the untreated-and-unexposed subsample at stage 1 (appending raw covariates only at stage 2 silently biases tau_total / delta_j on panels with time-varying covariates); non-absorbing / reversible treatment patterns (e.g. [0, 1, 0]) raise ValueError rather than being silently coerced into "treated from first 1 onward"; non-constant first_treat values across rows of the same unit raise ValueError; conley_coords is required on every fit path (not just vcov_type="conley") because ring construction always uses it. Far-away control identification: uses CURRENT-period untreated status (D_it = 0) rather than never-treated-only, so all-eventually-treated staggered designs (no never-treated units) can identify the counterfactual via not-yet-treated far-away rows. Variance (Wave B MVP): stage-2 OLS variance via solve_ols (HC1 / Conley / cluster paths all flow through). The Gardner GMM first-stage uncertainty correction is NOT applied at stage 2 in this PR (documented limitation; planned follow-up extends two_stage.py::_compute_gmm_variance to accept a Conley kernel matrix in place of HC1's identity at the influence-function outer-product step). Deferred features (planned follow-ups): event_study=True per-event-time × ring coefficients (Butts Table 2), survey_design= integration, ring_method="count" (count-of-treated-in-ring), data-driven d_bar selection (Butts 2021b / Butts 2023 JUE Insight), Gardner GMM first-stage correction at stage 2, sparse staggered ring-distance path. Tests: tests/test_spillover.py (157 tests across ring-construction primitives, validators, fit integration, raw-data invariant, identification MC — non-staggered DGP at 50 seeds + 200-seed @pytest.mark.slow variant recovers both tau_total and delta_1; staggered DGP at 30 seeds anchors both tau_total and delta_1 — Conley plumbing (verifies solve_ols is called with vcov_type="conley" + Conley kwargs, no silent HC1 fallback), Gardner identity bit-identity, coefficients-vs-vcov alignment, warn-and-drop, rank_deficient_action validation, Omega_0 bipartite-graph connectivity, anticipation behavior on both fit paths). DGP factories tests/_dgp_utils.py::generate_butts_nonstaggered_dgp / generate_butts_staggered_dgp satisfy Butts Assumptions 1/3/5/7 by construction.
• ChaisemartinDHaultfoeuille.predict_het × placebo: R-parity on both global and per-path surfaces. R-verified — did_multiplegt_dyn(predict_het, placebo) emits heterogeneity OLS results on backward (placebo) horizons via R's DIDmultiplegtDYN:::did_multiplegt_main placebo block (effect = matrix(-i, ...) rbind site); the same block runs per-by_level under did_multiplegt_dyn(by_path, predict_het, placebo), so both global res$results$predict_het and per-by_level res$by_level_i$results$predict_het slots emit backward rows. R's predict_het syntax with placebo > 0 requires the c(-1) sentinel in the horizon vector to trigger "compute heterogeneity for ALL forward (1..effects) AND ALL placebo (1..placebo) positions" — passing positive-only horizons errors with "specified numbers in predict_het that exceed the number of placebos". Python mirrors via _compute_heterogeneity_test(..., placebo=L_max) (set automatically from self.placebo at both global and per-path call sites in fit()) — the function iterates forward (1..L_max) and backward (-1..-L_max) horizons in a single loop with an explicit out_idx < 0 eligibility guard for backward horizons whose F_g is too small (would otherwise silently misread N_mat via numpy negative indexing). results.heterogeneity_effects uses negative-int keys for backward horizons; path_heterogeneity_effects does the same per path. Placebo rows in to_dataframe(level="by_path") have non-NaN het_* columns when placebo=True and heterogeneity= are both set. Survey gate (warn + skip): survey_design + placebo + heterogeneity emits a UserWarning at fit-time and falls back to forward-horizon-only heterogeneity on both surfaces — the Binder TSL cell-period allocator's REGISTRY justification is tied to post-period attribution; backward-horizon attribution puts ψ_g mass on a pre-period cell, a separate library-extension claim that needs its own derivation. Forward-horizon predict_het + survey_design continues to work unchanged on both global and per-path surfaces. The function-level _compute_heterogeneity_test keeps a per-iteration NotImplementedError backstop for direct callers that bypass fit(). Pre-period allocator derivation deferred to a follow-up methodology PR (tracked in TODO.md). R parity confirmed at tests/test_chaisemartin_dhaultfoeuille_parity.py::TestDCDHDynRParityHeterogeneityWithPlacebo (scenario 23, multi_path_reversible_predict_het_with_placebo_global, placebo=2, effects=3, no by_path) and ::TestDCDHDynRParityByPathHeterogeneityWithPlacebo (scenario 22, same DGP plus by_path=3); pinned at BETA_RTOL=1e-6 / SE_RTOL=1e-5 for beta / se / t_stat / n_obs and INFERENCE_RTOL=1e-4 for p_value / conf_int across 3 paths × (3 forward + 2 placebo) = 15 horizons + 1 global × 5 horizons. Cross-surface invariants regression-tested at tests/test_chaisemartin_dhaultfoeuille.py::TestByPathPredictHetPlacebo (placebo het column population, survey-gate warn+skip behavior, forward+survey anti-regression, out_idx<0 eligibility guard, single-path telescope path_heterogeneity_effects[(only_path,)] == heterogeneity_effects bit-exactly, summary rendering, direct-call NotImplementedError backstop). Closes TODO #422.

Changed

• BaconDecomposition: default weights flipped from "approximate" to "exact" (PR-B methodology audit). The new default uses Goodman-Bacon (2021) Theorem 1's exact Eqs. 7-9 + 10e-g weights, matching R bacondecomp::bacon() at atol=1e-6 (validated via tests/test_methodology_bacon.py::TestBaconParityR; see the new Added entry above for the convention divergence on always-treated cohorts). Hand-calculation + TWFE-vs-weighted-sum identity also hold at atol=1e-10. The weights="approximate" path remains available as an opt-in fast diagnostic for speed-sensitive loops; its numerical output may differ from R. Three entry points were flipped: BaconDecomposition(weights="exact") (bacon.py:397), bacon_decompose(weights="exact") (bacon.py:1064), TwoWayFixedEffects.decompose(weights="exact") (twfe.py:684). Behavior change for users not passing explicit weights=: the decomposition weights are now paper-faithful by default. Users who depended on the previous "approximate" numerics for diagnostic plots or comparison-type weight shares can preserve the old behavior by passing weights="approximate" explicitly. Survey-design behavior change: weights="exact" (now the default) routes through _validate_unit_constant_survey, which rejects survey designs whose weights / strata / PSU / FPC columns vary within a unit across periods (the exact-mode path collapses to per-unit aggregation via groupby().first()). The previous weights="approximate" default tolerated time-varying within-unit survey weights via observation-level weighted means. Users whose survey-weighted Bacon calls used time-varying within-unit weights must now either (a) collapse their weights to be unit-constant or (b) pass explicit weights="approximate" to retain the legacy obs-level path. The production diagnostic surface (diff_diff/diagnostic_report.py:1740) was updated to pass explicit weights="exact". Existing test assertions in tests/test_bacon.py continue to pass with the new default; the test_weighted_sum_equals_twfe tolerance was tightened from < 0.1 to < 1e-10 to lock the Theorem 1 algebraic-identity contract.

• ChaisemartinDHaultfoeuille.predict_het inference: t-distribution df threading (closes TODO pilot-412). _compute_heterogeneity_test now passes df = n_obs - rank(design) to safe_inference on the non-survey OLS path, matching R did_multiplegt_dyn(predict_het=...)'s t-distribution inference (DIDmultiplegtDYN:::did_multiplegt_main t_stat <- qt(0.975, df.residual(model)) site). Pre-PR Python used df=None (normal Z critical), producing 0.1-2% rtol gaps on p_value and conf_int vs R. Parity tolerance tightened on the existing forward-horizon scenarios (multi_path_reversible_predict_het, multi_path_reversible_by_path_predict_het) from "unpinned" to INFERENCE_RTOL=1e-4 on p_value and conf_int; beta / se / t_stat continue at BETA_RTOL=1e-6 / SE_RTOL=1e-5. Post-drop rank (post-2026-05-16 wrap-up): the df denominator uses the post-drop numerical rank via _detect_rank_deficiency, which solve_ols already calls internally. For full-rank designs rank == n_params and behavior is bit-identical to the pre-PR n_obs - n_params path; for near-rank-deficient designs that solve_ols retains rather than NaN-out (e.g., cohort-collinearity at high horizons), the post-drop rank is strictly lower and the post-PR df is larger, matching R's lm() convention. The Z-vs-t REGISTRY deviation note is replaced with an "R parity (post-2026-05-15 df threading)" positive-claim note.

• ChaisemartinDHaultfoeuille.by_path negative-baseline path regression coverage. New tests/test_chaisemartin_dhaultfoeuille.py::TestByPathNonBinary::test_negative_baseline_path_supported exercises switchers with D_{g,1} = -1 and asserts that path_effects correctly contains negative-baseline tuple keys (e.g., (-1, 0, 0, 0), (-1, 1, 1, 1)). This closes the test-coverage gap from PR #419: the existing test_negative_integer_D_supported only covered paths with negative values in non-baseline positions (e.g., (0, -1, -1, -1)), which does not trigger R's documented substr(path, 1, 1) baseline-extraction bug. Python's tuple-key matching is correct under any baseline value; this test pins the contract. No R-parity fixture is added because R is the buggy side on this regime — the deviation is documented in the REGISTRY non-binary treatment Note.

3.3.3 - 2026-05-15

Added

• Tutorial 22: Survey-Weighted HAD (docs/tutorials/22_had_survey_design.ipynb) — end-to-end walkthrough of HeterogeneousAdoptionDiD + did_had_pretest_workflow on a BRFSS-shape stratified household-survey panel (5 strata × 6 PSUs/stratum × 2 states/PSU = 60 states; post-stratification raking weights with CV ~ 0.30; FPC = 30 PSUs/stratum; PSU × period interaction shocks injected so cluster correlation survives DiD first-differencing). Demonstrates the SurveyDesign(strata=...) path through the Stute pretest family that the previous [Unreleased] entry unblocked. Eight numbered sections: motivation; panel + in-notebook helper for attaching survey columns to a HAD panel; naive vs survey-aware headline fit with a side-by-side ATT / SE / CI table (~10% SE inflation, sign-only direction asserted); a dedicated section explaining why the SE inflation is modest for HAD specifically (WAS-d_lower IF concentration at the boundary vs full-panel regression coefficients); event-study fit with sup-t cband under the survey design (per-horizon table + matplotlib gated plot); pretest workflow on both aggregate="overall" and aggregate="event_study" paths walking the Phase 4.5 C0 QUG-deferred verdict suffix and the now-supported stratified-clustered Stute multiplier bootstrap; "Communicating to Leadership" two-paragraph template (executive + methodologist); Extensions + Summary Checklist surfacing the still-deferred lonely_psu='adjust' + singleton-strata, replicate-weight designs, and the permanent QUG-under-survey C0 deferral. Companion drift-test file tests/test_t22_had_survey_design_drift.py (32 tests across 7 groups: panel + survey composition with deterministic exact pins; naive-vs-survey headline with sign-only SE-inflation anchor; event-study cband-vs-pointwise ordering and post/pre coverage; pretest overall path with _QUG_DEFERRED_SUFFIX lock and Yatchew sigma2_* deterministic pins; pretest event-study path with the SAME _QUG_DEFERRED_SUFFIX lock plus a SEPARATE substring lock on report.summary() for the L736 QUG-skip note; workflow-surface separation locking that overall has Stute+Yatchew while event-study has joint pretrends/joint linearity with yatchew=None and stute=None; and weighted point-estimation contract anchoring survey.att != naive.att plus the algebraic identity att = (dy_mean_w - tau_bc) / den_w from _fit_continuous). Bootstrap p-value pins use anchored windows of total width 0.30 (± 0.15 around seeded centers) per feedback_strata_bootstrap_path_divergence (stratified Mammen multiplier paths reduce effective dofs vs non-strata; PR #432 commit aef07020 already had to relax bit-equality bands on this code path). T20 and T21 "Extensions" bullets updated with forward-pointers to T22; docs/practitioner_decision_tree.rst HAD universal-rollout and survey sections each gain a .. tip:: cross-link to T22 (adjacent to T20 / T17, NOT displacing); docs/api/had.rst gains a "Survey-aware fit" cross-reference; docs/survey-roadmap.md gains a "Phase 4.5 C ✅ Shipped" entry; bundled diff_diff/guides/llms.txt and llms-practitioner.txt carry T22 inventory entries (the llms-full.txt reference guide is left as a follow-up to keep T22 PR scope tight); docs/doc-deps.yaml wires T22 as a dependent of both had.py and had_pretests.py. Closes the Phase 5 (wave 2 second slice) tutorial gap; the realistic survey-weighted HAD workflow on BRFSS / CPS / NHANES / ACS-shaped designs is now end-to-end documented for practitioners.
• HAD pretest workflow: stratified survey-design support (Phase 4.5 C continuation). Lifts the NotImplementedError gate on SurveyDesign(strata=...) in stute_test (had_pretests.py:1927-1940 pre-PR) and stute_joint_pretest (:3259-3271 pre-PR), and by inheritance in joint_pretrends_test, joint_homogeneity_test, and did_had_pretest_workflow (the wrappers delegate to the joint Stute helper). Implements a documented synthesis of clustered-wild-bootstrap ingredients (Cameron-Gelbach-Miller 2008 cluster-level multipliers; Davidson-Flachaire 2008 wild-bootstrap centering; Djogbenou-MacKinnon-Nielsen 2019 cluster-wild consistency for nonlinear functionals; Kreiss-Lahiri 2012 within-block centering analogy; Wu 1986 / Liu 1988 Bessel small-sample correction) — no single paper covers the exact composition for the stratified Stute CvM functional. The recipe: within-stratum demean + sqrt(n_h/(n_h-1)) Bessel rescale applied to the PSU multipliers psu_mults BEFORE the per-obs broadcast eta_obs = psu_mults[b, psu_col_idx] in the wild-residual loop. Bootstrap CvM variance matches the analytical Binder-TSL stratified target V_S = sum_h (1 - f_h) (n_h / (n_h - 1)) sum_j (psi_hj - psi_h_bar)² exactly (the (1 - f_h) FPC factor was already baked in by generate_survey_multiplier_weights_batch; this PR bakes the remaining (n_h / (n_h - 1)) factor and enforces within-stratum-mean-zero centering). New shared helper bootstrap_utils.apply_stratum_centering(psu_mults, resolved_survey, psu_ids, psu_axis=...) is called from both the new Stute path (psu_axis=1 on the multiplier matrix) AND the existing HAD sup-t event-study cband bootstrap (psu_axis=0 on the PSU-aggregated influence tensor; refactored bit-exactly from the inline block previously at had.py:2172-2204). Locks the algebraic identity architecturally instead of leaving parallel code blocks to drift. MC oracle consistency validated under a 4-stratum × 6-PSU/stratum stratified null DGP with weights+strata+PSU (200 seeded draws, empirical Type I at α=0.05 in [0, 0.10] — 3σ band; the FPC bake-in is covered separately by the helper-unit test test_fpc_baked_in_helper_is_fpc_agnostic); MC power validated under a known-alternative stratified DGP (rejection > 0.50). HAD sup-t event-study cband bit-parity preserved (atol=1e-14, rtol=1e-14 on the refactored helper output + 29 existing cband tests passing post-refactor; that helper-level bit-parity test locks the axis-0 algebra). A separate wired-in regression at tests/test_had_pretests.py::TestStuteStratifiedSurveyBootstrap::test_stute_call_sites_invoke_apply_stratum_centering monkey-patches the helper and asserts both Stute call sites (stute_test at had_pretests.py:1985 and stute_joint_pretest at :3312) invoke it with psu_axis=1 — that test fails if either call site is disconnected (the axis-0 helper-parity test alone does not catch that case). See docs/methodology/REGISTRY.md § HeterogeneousAdoptionDiD — "Note (Stute stratified survey-bootstrap calibration)" for the full derivation. Remaining deferrals: lonely_psu='adjust' + singleton-strata (same pseudo-stratum centering gap as the HAD sup-t deviation at REGISTRY:2382) and replicate-weight designs (BRR/Fay/JK1/JKn/SDR — separate Rao-Wu / JKn bootstrap composition). Unblocks the realistic survey-weighted HAD workflow on BRFSS/CPS/NHANES/ACS-shaped designs.
• Conley (1999) Wave A mechanical extensions on top of the Phase 1+2 sandwich (diff_diff/conley.py, diff_diff/linalg.py, diff_diff/estimators.py, diff_diff/twfe.py). (1) DiD support (#118): DifferenceInDifferences(vcov_type="conley").fit(..., unit="<col>") is now supported. unit is a fit-time kwarg (NOT on __init__; unused unless Conley is set; not part of get_params() / set_params()) mirroring MultiPeriodDiD.fit(unit=...) / TwoWayFixedEffects.fit(unit=...). DiD inherits the same panel block-decomposed sandwich as MPD/TWFE; on a 2-period panel it matches MultiPeriodDiD(...).fit(..., post_periods=[1], reference_period=0) bit-exactly. Missing unit=/conley_lag_cutoff/conley_coords/conley_cutoff_km raise ValueError; survey_design= + Conley raises NotImplementedError (Bertanha-Imbens 2014 follow-up); inference="wild_bootstrap" + Conley raises NotImplementedError. (2) Combined spatial + cluster product kernel (#119): compute_robust_vcov(vcov_type="conley", cluster_ids=...) / LinearRegression(vcov_type="conley", cluster_ids=...) / TwoWayFixedEffects(vcov_type="conley", cluster="<col>") / MultiPeriodDiD(vcov_type="conley", cluster="<col>") / DifferenceInDifferences(vcov_type="conley", cluster="<col>") apply K_total[i, j] = K_space(d_ij/h) · 1{c_i = c_j}. On the panel block-decomposed path the cluster indicator multiplies BOTH the spatial sandwich AND the serial sandwich; the validator enforces that cluster_ids is constant within each unit across periods (the within-unit serial mask is then trivially all-ones; cross-sectional path has no such constraint). TWFE's default auto-cluster on the Conley path remains silently dropped (combining with unit-level clusters would zero out all between-unit pairs and defeat the spatial pooling); users must pass an explicit above-unit cluster (e.g. region) to opt in. DiD has no auto-cluster — the choice is fully explicit. Two limit fixtures anchor correctness (no R parity — R conleyreg does not support combined kernels): all-unique-clusters reduces to HC0; huge-cutoff reduces to pure within-cluster CR1. The huge-cutoff reduction is EXACT only for conley_kernel="uniform" (K(u) = 1 for |u| ≤ 1); for conley_kernel="bartlett" the identity is asymptotic since K_bartlett(u) = 1 - |u| < 1 for u > 0. The fixture anchor uses uniform for an exact identity check. Per-slice mask construction (NOT full n×n) preserves memory on panel paths. (3) Sparse k-d-tree fast path (#120): auto-activates for the spatial Bartlett meat when n > 5_000 AND metric is "haversine" or "euclidean" AND kernel is "bartlett". Builds a CSR sparse kernel matrix via scipy.spatial.cKDTree.query_ball_tree instead of materializing the full n×n distance matrix; haversine projects to a 3-D unit-sphere chord representation with the exact great-circle recomputed for in-range neighbors only. Bit-identity parity vs the dense path at atol=1e-10; R parity at atol=1e-6 is preserved on the existing 3 panel R fixtures with the sparse path force-enabled. The bartlett-only gate is for boundary correctness — bartlett at u=1 is exactly 0, so the sparse path safely drops at-cutoff pairs; uniform at u=1 is 1 and would require a closed-interval query semantic that haversine chord projection cannot reliably preserve. Constants: _CONLEY_SPARSE_N_THRESHOLD = 5_000 (auto-toggle); _CONLEY_DENSE_WARN_N renamed _CONLEY_DENSE_OOM_WARN_N = 20_000 (memory exhaustion threshold for the dense fallback — independent of the sparse threshold). Private _conley_sparse: Optional[bool] kwarg on _compute_conley_vcov controls the toggle (None = auto, True = force, False = force dense; True with an unsupported kernel/metric raises). The serial component (within-unit Bartlett over time) remains dense regardless — per-unit slices are small. (4) Callable conley_metric validation (#123): result must satisfy shape (n, n), finite, non-negative, symmetric to atol=1e-10, AND zero on the diagonal (|d(i, i)| ≤ 1e-10); each failure raises a targeted ValueError naming the violated invariant. The zero-diagonal contract is load-bearing for the Conley sandwich: the i = j term must reduce to the HC0 diagonal X_i ε_i² X_i' via K(0) = 1; positive self-distance would silently attenuate the HC0 contribution by K(d_ii / h) < 1. Built-in metrics ("haversine", "euclidean") satisfy this by construction. Previously, malformed callables produced opaque BLAS errors deep in the pipeline. Tests: tests/test_conley_vcov.py::TestConleySparse (12), ::TestConleySparseRParityForced (3), ::TestConleyCluster (10), ::TestConleyDistanceMetrics extended (7 new); existing rejection tests flipped to behavioral; test_did_conley_matches_mpd_post_periods_1 locks the DiD-vs-MPD bit-exact agreement. Docs: REGISTRY ## ConleySpatialHAC updates: new "Combined spatial + cluster product kernel" + "Performance / scale" subsections, DiD-vs-TWFE cluster asymmetry paragraph, updated panel-API restrictions table. TODO rows 118 / 119 / 120 / 123 removed; rows 121 (Conley + survey_design / weights, Bertanha-Imbens 2014) and 122 (SyntheticDiD(vcov_type="conley"), spatial-block bootstrap per Politis-Romano 1994) retained for future waves.
• Conley (1999) spatial-HAC standard errors via vcov_type="conley" on cross-sectional LinearRegression / compute_robust_vcov plus panel MultiPeriodDiD / TwoWayFixedEffects (Phases 1 and 2 of the spillover-conley initiative). Cross-sectional contract: conley_coords (n × 2 array of lat/lon or projected coords), conley_cutoff_km=<float> (positive finite bandwidth in km for haversine, or coord units for euclidean — REQUIRED, no default per the no-silent-failures contract), conley_metric="haversine"|"euclidean"|callable (default "haversine"; great-circle uses Earth's mean radius 6371.01 km matching R conleyreg), conley_kernel="bartlett"|"uniform" (default "bartlett"; both kernels emit a UserWarning if the resulting meat has a materially negative eigenvalue — neither the radial 1-D Bartlett nor the uniform kernel is formally PSD-guaranteed; Conley 1999's explicit PSD formula is the 2-D separable lattice product window at Eq 3.14). Cross-sectional variance estimator Var̂(β) = (X'X)^{-1} · ( Σ_{i,j} K(d_ij/h) · X_i ε_i ε_j X_j' ) · (X'X)^{-1} (Conley 1999 Eq 4.2). Panel contract (Phase 2, new): Three new co-required kwargs conley_time (n-length array), conley_unit (n-length array), and conley_lag_cutoff=<int> (non-negative; 0 means within-period spatial only, no serial component) switch into the block-decomposed panel sandwich that matches R conleyreg with lag_cutoff > 0: XeeX_total = Σ_t (within-period spatial sandwich) + Σ_u (within-unit Bartlett temporal sandwich, lag ∈ {1..L}, same-time excluded). This is NOT a multiplicative product kernel — verified empirically against conleyreg::time_dist and XeeXhC at ~1e-14 on the panel parity fixtures. The temporal kernel is hardcoded Bartlett (1 - |lag|/(L+1)) regardless of conley_kernel, mirroring conleyreg::time_dist.cpp; documented as a Note (deviation from R-symmetric API) in REGISTRY. Panel estimator wire-up (Phase 2): MultiPeriodDiD(vcov_type="conley", conley_lag_cutoff=...).fit(..., unit=...) and TwoWayFixedEffects(vcov_type="conley", conley_lag_cutoff=...).fit(..., unit=...) lift the Phase 1 fit-time rejection; the conley_time and conley_unit arrays are auto-derived from the existing time and unit column-name arguments at fit-time. DifferenceInDifferences(vcov_type="conley") is also supported (Wave A #118 in this release; see the Wave A entry above) — pass unit=<col> as a fit-time kwarg to DiD.fit(...). Other constraints (Phase 1, unchanged): SyntheticDiD(vcov_type="conley") raises TypeError (uses bootstrap variance, not analytical sandwich); set_params mirrors the constructor rejection. vcov_type="conley" + weights= / survey_design= raises NotImplementedError (Bertanha-Imbens 2014 weighted-Conley deferred to a follow-up PR). vcov_type="conley" + explicit cluster_ids= is supported via the combined spatial + cluster product kernel (Wave A #119; see the Wave A entry above). TWFE's default auto-cluster on the Conley path is silently dropped (combining with unit-level clusters would defeat the spatial pooling); users opt into the combined kernel by passing an explicit above-unit cluster. inference="wild_bootstrap" + Conley raises (incompatible inference modes). A sparse k-d-tree fast path auto-activates for the spatial Bartlett meat when n > 5_000 with bartlett kernel and haversine/euclidean metric (Wave A #120); the dense fallback still emits an OOM UserWarning at n > 20_000. Implementation: Helpers live in diff_diff/conley.py (_haversine_km, _pairwise_distance_matrix, _bartlett_kernel, _uniform_kernel, _validate_conley_kwargs, _compute_conley_vcov — the validator and sandwich helper now accept keyword-only time / unit / lag_cutoff for the panel path); compute_robust_vcov in diff_diff/linalg.py threads the new kwargs through. R conleyreg parity (Düsterhöft 2021, CRAN v0.1.9) on six benchmark fixtures (benchmarks/data/r_conleyreg_conley_golden.json, regenerable via benchmarks/R/generate_conley_golden.R): 3 cross-sectional (Phase 1) + 3 new panel fixtures (panel_haversine_lag1, panel_haversine_lag2, panel_lat_lon_realistic_lag1; n_units × T = 60×3, 80×5, 100×4 at lag={1,2,1}); observed max abs diff ~5.7e-16. Earth radius 6371.01 km matches conleyreg::haversine_dist. Test file tests/test_conley_vcov.py skips parity cleanly when the JSON is absent. New REGISTRY section ## ConleySpatialHAC. Subsequent phases of the spillover-conley initiative (ring-indicator spillover-aware DiD per Butts 2021; survey-design / replicate-weight support; SyntheticDiD Conley path) are tracked in TODO.md under "Tech Debt from Code Reviews" → spillover-conley rows.
• Tutorial 21: HAD Pre-test Workflow (docs/tutorials/21_had_pretest_workflow.ipynb) — composite pre-test walkthrough for HeterogeneousAdoptionDiD building on Tutorial 20's brand-campaign framing. Uses a 60-DMA × 8-week panel close in shape to T20's but with the dose distribution drawn from Uniform[$0.01K, $50K] (vs T20's [$5K, $50K]); the true support is strictly positive but very near zero, chosen so the QUG step in did_had_pretest_workflow fails-to-reject H0: d_lower = 0 in this finite sample and the verdict text fires the load-bearing "Assumption 7 deferred" pivot for the upgrade-arc narrative. (HAD's design="auto" selector — a separate min/median heuristic at had.py::_detect_design, NOT the QUG p-value — independently lands on the continuous_at_zero identification path with target WAS on this panel because d.min() < 0.01 * median(|d|). The QUG test and the design selector are independent rules that point to the same identification path here.) Walks through three surfaces: (a) did_had_pretest_workflow(aggregate="overall") on a two-period collapse, where the verdict explicitly flags Step 2 (Assumption 7 pre-trends) as not run because a single pre-period structurally cannot support a pre-trends test, and the structural fields pretrends_joint / homogeneity_joint are both None; (b) did_had_pretest_workflow(aggregate="event_study") on the full multi-period panel, where the verdict reads "TWFE admissible under Section 4 assumptions" because all three testable diagnostics (QUG + joint pre-trends Stute over 3 horizons + joint homogeneity Stute over 4 horizons) fail-to-reject — non-rejection evidence under finite-sample power and test specification, not proof that the identifying assumptions hold; and (c) a side panel exercising both yatchew_hr_test null modes — null="linearity" (default, paper Theorem 7) vs null="mean_independence" (Phase 4 R-parity with R YatchewTest::yatchew_test(order=0)) — on the within-pre-period first-difference paired with post-period dose, illustrating the stricter null's larger residual variance (sigma2_lin 7.01 vs 6.53) and smaller p-value (0.29 vs 0.49). Companion drift-test file tests/test_t21_had_pretest_workflow_drift.py (16 tests pinning panel composition, both verdict pivots, structural anchors on both paths, deterministic QUG / Yatchew statistics, bootstrap p-value tolerance bands per feedback_bootstrap_drift_tests_need_backend_tolerance, and HAD(design="auto") resolution to continuous_at_zero on this panel). T20's "Composite pretest workflow" Extensions bullet updated with a forward-pointer to T21. T22 weighted/survey HAD tutorial shipped as a follow-up notebook PR (see the T22 entry above).
• ChaisemartinDHaultfoeuille.by_path and paths_of_interest now compose with survey_design for analytical Binder TSL SE and replicate-weight bootstrap variance. The NotImplementedError gate at chaisemartin_dhaultfoeuille.py:1233-1239 is replaced by a per-path multiplier-bootstrap-only gate (survey_design + n_bootstrap > 0 under by_path / paths_of_interest still raises, since the survey-aware perturbation pivot for path-restricted IFs is methodologically underived). Per-path SE routes through the existing _survey_se_from_group_if cell-period allocator: the per-period IF (U_pp_l_path) is built with non-path switcher-side contributions skipped (control contributions are unchanged, matching the joiners/leavers IF convention; preserves the row-sum identity U_pp.sum(axis=1) == U), cohort-recentered via _cohort_recenter_per_period, then expanded to observations as psi_i = U_pp[g_i, t_i] · (w_i / W_{g_i, t_i}). Replicate-weight designs unconditionally use the cell allocator (Class A contract from PR #323). New _refresh_path_inference helper post-call refreshes safe_inference on every populated entry across multi_horizon_inference, placebo_horizon_inference, path_effects, and path_placebos so all four surfaces use the same final df_survey after per-path replicate fits append n_valid to the shared accumulator. Path-enumeration ranking under survey_design remains unweighted (group-cardinality, not population-weight mass). Lonely-PSU policy stays sample-wide, not per-path. Telescope invariant: on a single-path panel, per-path SE matches the global non-by_path survey SE bit-exactly. No R parity — R did_multiplegt_dyn does not support survey weighting; this is a Python-only methodology extension. The global non-by_path TSL multiplier-bootstrap path is unaffected (anti-regression test tests/test_chaisemartin_dhaultfoeuille.py::TestByPathSurveyDesignAnalytical::test_global_survey_plus_n_bootstrap_still_works locks the per-path-only scope of the new gate). Cross-surface invariants regression-tested at TestByPathSurveyDesignAnalytical (~17 tests across gate / dispatch / analytical SE / replicate-weight SE / per-path placebos / trends_linear composition / unobserved-path warnings / final-df refresh regressions) and TestByPathSurveyDesignTelescope. See docs/methodology/REGISTRY.md §ChaisemartinDHaultfoeuille Note (Phase 3 by_path ...) → "Per-path survey-design SE" for the full contract.
• Inference-field aliases on staggered result classes for adapter / external-consumer compatibility. Read-only @property aliases expose the flat att / se / conf_int / p_value / t_stat names (matching DiDResults / TROPResults / SyntheticDiDResults / TripleDifferenceResults / HeterogeneousAdoptionDiDResults) on every result class that previously only carried prefixed canonical fields: CallawaySantAnnaResults, StackedDiDResults, EfficientDiDResults, ChaisemartinDHaultfoeuilleResults, StaggeredTripleDiffResults, WooldridgeDiDResults, SunAbrahamResults, ImputationDiDResults, TwoStageDiDResults (mapping to overall_*); ContinuousDiDResults (mapping to overall_att_*, ATT-side as the headline, ACRT-side accessible unchanged via overall_acrt_*); MultiPeriodDiDResults (mapping to avg_*). ContinuousDiDResults additionally exposes overall_se / overall_conf_int / overall_p_value / overall_t_stat aliases for naming consistency with the rest of the staggered family. Aliases are pure read-throughs over the canonical fields — no recomputation, no behavior change — so the safe_inference() joint-NaN contract (per CLAUDE.md "Inference computation") is inherited automatically (NaN canonical → NaN alias, locked at tests/test_result_aliases.py::test_pattern_b_aliases_propagate_nan). The native overall_* / overall_att_* / avg_* fields remain canonical for documentation and computation. Motivated by the balance.interop.diff_diff.as_balance_diagnostic() adapter (facebookresearch/balance PR #465) which calls getattr(res, "se", None) / getattr(res, "conf_int", None) without a fallback chain — pre-alias, every staggered result class returned None on those keys, silently dropping se and conf_int from the adapter's diagnostic dict. 23 alias-mechanic + balance-adapter regression tests at tests/test_result_aliases.py. Patch-level (additive on stable surfaces).
• ChaisemartinDHaultfoeuille.by_path + non-binary integer treatment — by_path=k now accepts integer-coded discrete treatment (D in Z, e.g. ordinal {0, 1, 2}); path tuples become integer-state tuples like (0, 2, 2, 2). The previous NotImplementedError gate at chaisemartin_dhaultfoeuille.py:1870 is replaced by a ValueError for continuous D (e.g. D=1.5) at fit-time per the no-silent-failures contract — the existing int(round(float(v))) cast in _enumerate_treatment_paths is now defensive (no-op for integer-coded D). Validated against R did_multiplegt_dyn(..., by_path) for D in {0, 1, 2} via the new multi_path_reversible_by_path_non_binary golden-value scenario (78 switchers, 3 paths, single-baseline custom DGP, F_g >= 4): per-path point estimates match R bit-exactly (rtol ~1e-9 on event horizons; rtol+atol envelope for placebo near-zero values), per-path SE inherits the documented cross-path cohort-sharing deviation (~5% rtol observed; SE_RTOL=0.15 envelope). Deviation from R for multi-character baseline states (D >= 10 or negative D): R's did_multiplegt_by_path derives the per-path baseline via path_index$baseline_XX <- substr(path_index$path, 1, 1), which captures only the first character of the comma-separated path string. For multi-character baselines this drops the rest of the value: for path = "12,12,..." it captures "1" instead of "12"; for path = "-1,-1,..." it captures "-" instead of "-1". R's per-path control-pool subset is mis-allocated in both regimes. Python's tuple-key matching is correct — the per-path point estimates we compute are correct; R's per-path subset for the same path is buggy. The shipped R-parity scenarios stay in nonnegative single-digit D in {0, 1, 2} to avoid the R bug; negative-integer treatment-state support (paths containing negative D values in non-baseline positions) is regression-tested in Python only at tests/test_chaisemartin_dhaultfoeuille.py::TestByPathNonBinary::test_negative_integer_D_supported (no R parity); a dedicated regression for a negative-baseline path (e.g. (-1, 0, 0, 0)) is deferred. R-parity test at tests/test_chaisemartin_dhaultfoeuille_parity.py::TestDCDHDynRParityByPathNonBinary; cross-surface invariants regression-tested at tests/test_chaisemartin_dhaultfoeuille.py::TestByPathNonBinary.
• New paths_of_interest kwarg on ChaisemartinDHaultfoeuille for user-specified treatment-path subsets, alternative to by_path=k's top-k automatic ranking. Mutually exclusive with by_path; setting both raises ValueError at __init__ and set_params time. Each path tuple must be a list/tuple of int of length L_max + 1 (uniformity validated at __init__; length match against L_max + 1 validated at fit-time); bool and np.bool_ are explicitly rejected, np.integer accepted and canonicalized to Python int for tuple-key consistency. Duplicates emit a UserWarning and are deduplicated; paths not observed in the panel emit a UserWarning and are omitted from path_effects. Paths appear in results.path_effects in the user-specified order, modulo deduplication and unobserved-path filtering. Composes with non-binary D and all downstream by_path surfaces (bootstrap, per-path placebos, per-path joint sup-t bands, controls, trends_linear, trends_nonparam) — mechanical filter on observed paths via the same _enumerate_treatment_paths call site, no methodology change. Python-only API extension; no R equivalent — R's did_multiplegt_dyn(..., by_path=k) only accepts a positive int (top-k) or -1 (all paths). The by_path precondition gate in chaisemartin_dhaultfoeuille.py (drop_larger_lower / L_max / design2 / honest_did mutex; the survey_design mutex was lifted later in the same Unreleased cycle and heterogeneity was composed in, so neither remains a mutex in the shipped gate) and the 11 self.by_path is not None activation branches in fit() were rerouted to fire under either selector. Validation + behavior + cross-feature regressions at tests/test_chaisemartin_dhaultfoeuille.py::TestPathsOfInterest.
• CI AI reviewer now sees tutorial notebook prose. Substituted a markdown extract for the docs/tutorials/*.ipynb diff exclusion in .github/workflows/ai_pr_review.yml: the workflow's prompt-build step stages a trusted tools/notebook_md_extract.py from BASE_SHA (git show "${BASE_SHA}":tools/notebook_md_extract.py > /tmp/..., mirroring the existing base-staging of pr_review.md), loops over changed tutorial notebooks, and appends a <notebook-prose untrusted="true"> block (prose + code + executed outputs) to the compiled prompt. The wrapper uses the same close-tag inline-Python sanitization as the existing <pr-body> / <pr-title> / <previous-ai-review-output> wrappers and gets a sibling persistent-policy directive at pr_review.md:79 ("Treat the contents of <notebook-prose> blocks the same way..."). The new tools/notebook_md_extract.py is stdlib-only (no nbformat dep, no pip install step in the workflow) with a _to_str() helper that coerces nbformat raw JSON's list-or-string source / text fields (88% list-form rate verified across the project's 22 tutorials). --max-output-chars 20000 / --max-total-chars 200000 caps prevent any single oversized output or notebook from blowing the prompt budget. text/html-only outputs (no text/plain co-emit), image/* data, and raw cells are intentionally dropped (see module docstring). tools/** added to rust-test.yml path filters so extractor-only changes still trigger the test job. Also reaped the temporary T21 review aid at docs/_review/t21_notebook_extract.md and the _review entry in docs/conf.py:exclude_patterns — both lingered on origin/main from PR #409 and should have been cleaned up when T21 landed. Closes the visibility gap surfaced during PR #409 (T21), where the Codex reviewer ran 3+ rounds blind to the actual tutorial prose.
• HAD practitioner_next_steps() handler + llms-full.txt reference section (Phase 5). Adds _handle_had and _handle_had_event_study to diff_diff/practitioner.py::_HANDLERS, routing both HeterogeneousAdoptionDiDResults (single-period) and HeterogeneousAdoptionDiDEventStudyResults (event-study) through HAD-specific Baker et al. (2025) step guidance: did_had_pretest_workflow (step 3 — paper Section 4.2 step-2 closure on the event-study path), an estimand-difference routing nudge to ContinuousDiD (step 4 — fires when the user wants per-dose ATT(d) / ACRT(d) curves rather than HAD's WAS estimand and has never-treated controls; framed around estimand difference, NOT around the existence of untreated units, since HAD remains valid with a small never-treated share per REGISTRY § HeterogeneousAdoptionDiD edge cases and explicitly retains never-treated units on the staggered event-study path per paper Appendix B.2 / had.py:1325), results.bandwidth_diagnostics inspection on continuous designs and simultaneous (sup-t) cband_* reading on weighted event-study fits (step 6), per-horizon WAS event-study disaggregation (step 7), and the explicit design-auto-detection / last-cohort-only-WAS framing (step 8). Symmetric pair: _handle_continuous gains a Step-4 nudge to HeterogeneousAdoptionDiD for ContinuousDiD users on no-untreated panels (this direction is correct because ContinuousDiD's identification requires never-treated controls). Extends _check_nan_att with an ndarray branch via lazy numpy import for HAD's per-horizon att array; uses np.all(np.isnan(arr)) semantics so partial-NaN arrays (legitimate event-study output under degenerate horizon-specific designs) do not over-fire the warning. Scalar path is bit-exact preserved across all 12 untouched handlers. Adds full HAD section + HeterogeneousAdoptionDiDResults / HeterogeneousAdoptionDiDEventStudyResults blocks + ## HAD Pretests index covering all 7 pretest entry points + Choosing-an-Estimator row to diff_diff/guides/llms-full.txt (the bundled-in-wheel agent reference); the documented constructor + fit() parameter NAMES are regression-locked against the real HeterogeneousAdoptionDiD.__init__ / .fit API via inspect.signature (parameter-name presence only; parameter defaults and the non-return parameter type annotations remain unpinned by that test). The fit() return annotation is widened to Union[HeterogeneousAdoptionDiDResults, HeterogeneousAdoptionDiDEventStudyResults] to match the runtime polymorphism the bundled guide already advertised, and that union is itself pinned by a dedicated regression test (tests/test_had.py::TestFitReturnAnnotation) using typing.get_type_hints. Tightens the existing Continuous treatment intensity Choosing row to surface ATT(d) vs WAS as the estimand differentiator. docs/doc-deps.yaml updated to remove the llms-full.txt deferral note on had.py and add llms-full.txt entries to had.py, had_pretests.py, and practitioner.py blocks. Patch-level (additive on stable surfaces). 26 new tests (16 in tests/test_practitioner.py::TestHADDispatch + 9 in tests/test_guides.py::TestLLMsFullHADCoverage + 1 fixture-minimality regression locking the "handlers are STRING-ONLY at runtime" stability invariant). Closes the Phase 5 "agent surfaces" gap; T21 pretest tutorial shipped in PR #409 and T22 weighted/survey tutorial shipped as a follow-up notebook PR (see the T22 entry above).

Changed

• HAD pretest non-strata bootstrap: small-sample calibration improvement. The Stute survey-bootstrap on non-strata designs (SurveyDesign(weights=...), SurveyDesign(weights=..., psu=...), SurveyDesign(weights=..., fpc=...)) now applies the standard sqrt(n_psu/(n_psu-1)) Bessel small-sample correction to the PSU multipliers uniformly with a single implicit stratum, mirroring the sibling HAD sup-t event-study cband bootstrap at had.py:2199-2204. Pre-PR Phase 4.5 C shipped raw iid multipliers without the centering; the bootstrap CvM variance was under-corrected by exactly the n_psu/(n_psu-1) factor relative to the unbiased within-stratum variance estimator. Direction-of-shift: toward correct calibration. Magnitude: approximately sqrt(n_psu/(n_psu-1)) - 1 ≈ 1.7% for n_psu=60, decreasing as n_psu grows. Practitioners reproducing pre-PR Stute non-strata bootstrap p-values exactly should pin the prior release; the post-PR p-values are the methodology-true values (### Added "HAD pretest workflow: stratified survey-design support" bullet above documents the full derivation). Affects only the Stute family on the weights= / survey_design=SurveyDesign(weights, [psu, fpc]) paths; Yatchew (closed-form weighted-OLS, no bootstrap) is unaffected, as is the unweighted bit-exact path (which has no multipliers to center).

3.3.2 - 2026-04-26

Added

• ChaisemartinDHaultfoeuille.by_path is now compatible with trends_linear (DID^{fd} group-specific linear trends) and trends_nonparam (state-set trends). For trends_linear, the first-differencing transform runs once globally before path enumeration, so per-path raw second-differences DID^{fd}_{path, l} surface on path_effects[path]["horizons"][l] automatically. Per-path cumulated level effects delta_{path, l} = sum_{l'=1..l} DID^{fd}_{path, l'} (the quantity R returns under did_multiplegt_dyn(..., by_path, trends_lin)) surface on the new results.path_cumulated_event_study[path][l] field, mirroring the global linear_trends_effects cumulation. to_dataframe(level="by_path") exposes cumulated_effect / cumulated_se columns (always present, NaN-when-None — mirrors the cband_* convention from PR #374); summary() renders a "Cumulated Level Effects (DID^{fd}, trends_linear)" sub-section under each per-path block. SE on the cumulated layer is the conservative upper bound (sum of per-horizon component SEs, NaN-consistent), matching the global linear_trends_effects convention. Path enumeration runs on the post-first-differenced N_mat_fd: switchers with F_g==2 fail the window-eligibility check and are dropped from path enumeration entirely (the existing global F_g >= 3 warning still surfaces the issue), so a path whose switchers all have F_g < 3 is silently absent from path_effects rather than present-with-NaN. Placebo under trends_linear returns RAW per-horizon values — there is no per-path placebo cumulation surface in either Python or R. For trends_nonparam, the set membership column is validated and stored once globally as set_ids_arr; the set_ids parameter is now threaded through the four per-path IF helpers (_compute_path_effects, _compute_path_placebos, _collect_path_bootstrap_inputs, _collect_path_placebo_bootstrap_inputs) so per-path analytical SE, bootstrap, placebos, and sup-t bands all consume the set-restricted control pool automatically. Per-period effects remain unadjusted under both extensions, consistent with the existing per-period DID contract. Validated against R via two new golden-value scenarios: single_baseline_multi_path_by_path_trends_lin (n_periods=13, F_g >= 4, cohort-single-path; per-path cumulated point estimates match R bit-exactly with POINT_RTOL=1e-9, cumulated SE within CUM_SE_RTOL=0.20) and multi_path_reversible_by_path_trends_nonparam (per-path point estimates AND placebos match R bit-exactly with POINT_RTOL=1e-9, per-path SE within SE_RTOL=0.15). F_g=3 boundary-case divergence (by_path + trends_linear): F_g=3 switchers have only 1 valid pre-window Z value after first-differencing, triggering 30%+ relative divergence between Python and R per-path point estimates on paths whose switchers include F_g=3. A targeted UserWarning fires at fit-time on this regime; R parity is asserted only on the F_g >= 4 parity fixture. Placebo parity for trends_linear is intentionally skipped (R's per-path placebo computation re-runs on the path-restricted subsample with different control eligibility than Python's global-then-disaggregate architecture surfaces; placebo + trends_linear is exercised via internal regression only). Cross-path cohort-sharing SE deviation from R documented for path_effects is inherited unchanged. Gates at chaisemartin_dhaultfoeuille.py:1014-1023 removed; by_path docstring updated to add the two new compatibility paragraphs and remove trends_linear / trends_nonparam from the incompatible list. R-parity tests at tests/test_chaisemartin_dhaultfoeuille_parity.py::TestDCDHDynRParityByPathTrendsLinear and ::TestDCDHDynRParityByPathTrendsNonparam; cross-surface regressions at tests/test_chaisemartin_dhaultfoeuille.py::TestByPathTrendsLinear and ::TestByPathTrendsNonparam. See docs/methodology/REGISTRY.md §ChaisemartinDHaultfoeuille Note (Phase 3 by_path ...) → "Per-path linear-trends DID^{fd}" and "Per-path state-set trends" for the full contract.
• yatchew_hr_test(null="mean_independence") mode mirroring R YatchewTest::yatchew_test(order=0). Adds a null: Literal["linearity", "mean_independence"] keyword-only kwarg to yatchew_hr_test. Default "linearity" is bit-exact backcompat (residuals from OLS dy = a + b·d + eps, paper Assumption 8 / Theorem 7). New "mean_independence" fits intercept-only OLS (dy = a + eps, residuals = dy - mean(dy)); the downstream sigma2_diff / sigma2_W / sort-by-d machinery is identical between the two modes. Exposed on both unweighted and survey-weighted code paths (weights= / survey_design= compose orthogonally with null=). Adds a null_form: str field to YatchewTestResults so summary() renders the correct null-hypothesis description; __repr__ and to_dict() updated. Closes the placebo Yatchew R-parity gap from PR #392 — tests/test_did_had_parity.py::TestYatchewParity now routes effect rows through null="linearity" (R order=1) and placebo rows through null="mean_independence" (R order=0); both modes share the documented × G/(G-1) finite-sample convention shift and parity holds at atol=1e-10. Patch-level (additive keyword-only kwarg + additive dataclass field with default).
• HAD trends_lin=True linear-trend detrending mode on HeterogeneousAdoptionDiD.fit(aggregate="event_study"), joint_pretrends_test, and joint_homogeneity_test. Mirrors R DIDHAD::did_had(..., trends_lin=TRUE) (paper Eq. 17 / Eq. 18 / page 32 joint-Stute homogeneity-with-trends). Per-group linear-trend slope estimated as Y[g, F-1] - Y[g, F-2] and applied as (t - base) × slope adjustment to per-event-time outcome evolutions. Requires F ≥ 3 (panel must contain F-2). The "consumed" placebo at our event-time e=-2 is auto-dropped (R reduces max placebo lag by 1 with the same effect). Mutually exclusive with survey weighting (survey_design / survey / weights): raises NotImplementedError per feedback_per_method_survey_element_contract (weighted slope estimator not derived from paper; tracked in TODO.md as a follow-up). Bit-exact backcompat for trends_lin=False (default). Patch-level (additive keyword-only kwarg).
• HAD R-package end-to-end parity test vs DIDHAD v2.0.0 (Credible-Answers/did_had) on the design="continuous_at_zero" (Design 1') surface. New parity fixture benchmarks/data/did_had_golden.json generated by benchmarks/R/generate_did_had_golden.R covers 3 paper-derived synthetic DGPs (Uniform, Beta(2,2), Beta(0.5,1)) × 5 method combinations (overall, event-study, placebo, yatchew, trends_lin). The harness explicitly forces HeterogeneousAdoptionDiD(design="continuous_at_zero") because R did_had always evaluates the local-linear at d=0 regardless of dose distribution; our default design="auto" may legitimately choose continuous_near_d_lower or mass_point on dose distributions with boundary density bounded away from zero (e.g., Beta(2,2)) and thereby diverge from R numerically — that divergence is methodologically defensible but out of scope for this parity test. Python parity test tests/test_did_had_parity.py asserts point estimate / SE / CI bounds at atol=1e-8 and Yatchew T-stat at atol=1e-10 after a documented × G/(G-1) finite-sample convention shift. Two intentional convention deviations from R, documented in docs/methodology/REGISTRY.md: (a) we report the bias-corrected point estimate (modern CCF 2018 convention; R's Estimate column reports the conventional estimate with the bias-corrected CI separately — our att matches R's CI midpoint); (b) Yatchew uses paper Appendix E's literal (1/G) variance-denominator convention while R uses base-R var()'s (1/(N-1)) sample-variance convention (parity is bit-exact after the × G/(G-1) shift). Yatchew on placebos with R's mean-independence null (order=0) was not exposed in yatchew_hr_test at the PR #392 cut and was skipped in the parity test; the follow-up yatchew_hr_test(null="mean_independence") entry above closes that gap (placebo rows now routed through null="mean_independence" and parity holds at the same atol=1e-10).
• Tutorial 20: HAD for National Brand Campaign with Regional Spend Intensity (docs/tutorials/20_had_brand_campaign.ipynb) — end-to-end practitioner walkthrough for HeterogeneousAdoptionDiD on a 60-DMA panel where every market is treated at a different dose level and no never-treated unit exists; comparison comes from dose variation across markets, not from an untreated holdout. The DGP uses Uniform[$5K, $50K] regional add-on spend per DMA (every DMA participates, no DMA at exactly $0), so design="auto" resolves to continuous_near_d_lower (Design 1) with target WAS_d_lower — interpreted as the average per-dollar marginal effect of regional spend above the lightest-touch DMA's spend (d_lower ≈ $5K). Covers the headline WAS_d_lower fit on a 2-period collapse, the multi-week event study with per-week pointwise CIs and pre-launch placebos, and a stakeholder communication template that flags the Assumption 5/6 caveat (non-testable local-linearity at the boundary). Companion drift-test file tests/test_t20_had_brand_campaign_drift.py (13 tests pinning panel composition / sample median, design auto-detection / target / d_lower, overall WAS_d_lower, CI endpoints, dose mean, n_units, full event-study horizon presence, and per-horizon coverage). T20 wired into the existing had.py entry in docs/doc-deps.yaml; cross-link added from docs/practitioner_decision_tree.rst § "Universal Rollout (No Untreated Markets)" via a .. tip:: block.

Changed

• Rust dependency upgrades: bumped rand 0.8 → 0.10 and rand_xoshiro 0.6 → 0.8 in the Rust backend (the two crates are coupled through rand_core and must move together). MSRV bumped from Rust 1.84 → 1.85 to satisfy the new dependency requirements. Three call sites in rust/src/bootstrap.rs updated for the rand 0.9 API rename: gen::<bool>() → random::<bool>(), gen::<f64>() → random::<f64>(), gen_range(0..6) → random_range(0..6). Webb wild bootstrap byte stream shifted as a side effect: rand 0.9 reworked the internal algorithm for random_range (improved rejection sampling), so Xoshiro256PlusPlus::seed_from_u64(seed) followed by random_range(0..6) consumes RNG bytes differently than the old gen_range(0..6) did. Distributional properties of Webb weights are unchanged (still uniform over the 6-point support); aggregate inference (SE, p-values, CI) converges to the same values for any reasonable n_bootstrap. Rademacher and Mammen byte streams are bit-identical to the prior release. Anyone with a saved Rust+Webb baseline pinning specific seeded results will see different numbers; the regression test suite uses within-build seed-reproducibility (not cross-version baselines) so all internal tests pass unchanged. New regression guard TestRustBackend::test_bootstrap_weights_bit_identity_snapshot pins fixed-seed weights for all three weight types, so any future RNG drift fails loudly with a localized error message.

3.3.1 - 2026-04-25

Changed

• HAD survey-design API consolidated to single survey_design= kwarg across all 8 HAD surfaces: HeterogeneousAdoptionDiD.fit, did_had_pretest_workflow, qug_test, stute_test, yatchew_hr_test, stute_joint_pretest, joint_pretrends_test, joint_homogeneity_test. Matches the rest of the library (ContinuousDiD, EfficientDiD, ChaisemartinDHaultfoeuille already used survey_design=). On data-in surfaces (HAD.fit, workflow, joint data-in wrappers) survey_design= accepts a SurveyDesign instance (column references resolved against data at fit time, same convention as the rest of the library). On the three array-in linearity helpers (stute_test, yatchew_hr_test, stute_joint_pretest) survey_design= accepts a pre-resolved ResolvedSurveyDesign; passing a SurveyDesign raises TypeError with migration guidance to make_pweight_design(arr) (pweight-only) or pre-resolution. qug_test is the 8th surface and accepts the same kwarg signature for consistency, but all non-None values raise NotImplementedError per the Phase 4.5 C0 permanent deferral (no migration path; the qug-specific mutex error reflects this). New public helper make_pweight_design(weights: np.ndarray) -> ResolvedSurveyDesign exported from the diff_diff top level for the pweight-only convenience on the three array-in linearity helpers (formerly the private survey._make_trivial_resolved, kept as a permanent private alias); validates 1-D input at the front door. Three-way mutex (survey_design + survey + weights) extends the prior 2-way (survey + weights) — at most one may be non-None per call. Patch-level addition (additive new kwarg + permanent alias for the helper; no breaking changes this release).

Deprecated

• HeterogeneousAdoptionDiD.fit(survey=, weights=), did_had_pretest_workflow(survey=, weights=), and the 6 HAD pretest helpers' survey= / weights= kwargs are deprecated in favor of the canonical survey_design=. Emits DeprecationWarning with migration guidance; the deprecated kwargs continue to route through the unchanged legacy back-end paths so numerical results are identical to pre-PR (bit-exact regression locked by parity tests in tests/test_had_dual_knob_deprecation.py). Both survey= and weights= will be removed in the next minor release. Carve-out for qug_test: the deprecation is kwarg-name-consolidation only; qug_test permanently rejects all non-None survey_design / survey / weights values (Phase 4.5 C0 deferral) and make_pweight_design(arr) is NOT a valid migration target — the deprecation warning text on qug_test is qug-specific and points users to did_had_pretest_workflow(..., survey_design=...) for survey-aware HAD pretesting (which skips the QUG step under survey).

Added

• ChaisemartinDHaultfoeuille.by_path + controls (DID^X residualization) — the per-baseline OLS residualization (Web Appendix Section 1.2) is now compatible with by_path=k. The residualization runs once on the first-differenced outcome BEFORE path enumeration, so all four downstream surfaces (analytical per-path SE, bootstrap SE, per-path placebos, per-path joint sup-t bands) consume the residualized Y_mat automatically (Frisch-Waugh-Lovell). Per-period effects remain unadjusted, consistent with the existing controls + per-period DID contract (per-period DID does not support residualization). Failed-stratum baselines (rank-deficient X) zero out N_mat for affected groups, which the path enumeration treats as ineligible per its existing convention. Deviation from R on multi-baseline switcher panels (point estimates): R did_multiplegt_dyn(..., by_path, controls) re-runs the per-baseline OLS residualization on each path's restricted subsample (path's switchers + same-baseline not-yet-treated controls), so its residualization coefficients vary per path when switchers have different baseline values. Our global-residualization architecture coincides with R on single-baseline switcher panels (every switcher shares the same D_{g,1}) — per-path point estimates match R exactly there. On multi-baseline panels, point estimates can diverge; the estimator emits a UserWarning at fit-time when this configuration is detected so practitioners do not silently consume estimates that disagree with R. SE inherits the cross-path cohort-sharing SE deviation from R documented for path_effects — bootstrap SE, placebo SE, and sup-t crit are Monte Carlo / joint-distribution analogs of the same residualized analytical IF and carry the same deviation. R-parity confirmed against did_multiplegt_dyn(..., by_path=3, controls="X1") via the new multi_path_reversible_by_path_controls single-baseline golden-value scenario (per-path point estimates match R bit-exactly — measured rtol ~1e-11 across all path × horizon cells — on this one-observation-per-cell scenario; per-path SE within ~6.5% of R, well inside the Phase 2 multi-horizon envelope). Cell-aggregated panels with multiple observations per (g, t) also coincide with our equal-cell-weighting first stage rather than R's N_gt-weighted first stage per the existing DID^X cell-weighting deviation documented in docs/methodology/REGISTRY.md Note (Phase 3 DID^X covariate adjustment). Gate at chaisemartin_dhaultfoeuille.py:988-992 removed; by_path docstring updated to add the new compatibility paragraph (with the multi-baseline caveat) and remove controls from the incompatible list. R-parity test at tests/test_chaisemartin_dhaultfoeuille_parity.py::TestDCDHDynRParityByPathControls; cross-surface inheritance + multi-baseline UserWarning regression-tested at tests/test_chaisemartin_dhaultfoeuille.py::TestByPathControls (analytical + bootstrap + placebo + sup-t + to_dataframe(level="by_path") cband columns + multi-baseline warning). See docs/methodology/REGISTRY.md §ChaisemartinDHaultfoeuille Note (Phase 3 by_path ...) → "Per-path covariate residualization (DID^X)" for the full contract.
• HAD linearity-family pretests under survey (Phase 4.5 C). stute_test, yatchew_hr_test, stute_joint_pretest, joint_pretrends_test, joint_homogeneity_test, and did_had_pretest_workflow now accept weights= / survey= keyword-only kwargs. Stute family uses PSU-level Mammen multiplier bootstrap via bootstrap_utils.generate_survey_multiplier_weights_batch (the same kernel as PR #363's HAD event-study sup-t bootstrap): each replicate draws an (n_bootstrap, n_psu) Mammen multiplier matrix, broadcast to per-obs perturbation eta_obs[g] = eta_psu[psu(g)], weighted OLS refit, weighted CvM via new _cvm_statistic_weighted helper. Joint Stute SHARES the multiplier matrix across horizons within each replicate, preserving both the vector-valued empirical-process unit-level dependence AND PSU clustering. Yatchew uses closed-form weighted OLS + pweight-sandwich variance components (no bootstrap): sigma2_lin = sum(w·eps²)/sum(w), sigma2_diff = sum(w_avg·diff²)/(2·sum(w)) with arithmetic-mean pair weights w_avg_g = (w_g+w_{g-1})/2, sigma4_W = sum(w_avg·prod)/sum(w_avg), T_hr = sqrt(sum(w))·(sigma2_lin-sigma2_diff)/sigma2_W. All three Yatchew components reduce bit-exactly to the unweighted formulas at w=ones(G) (locked at atol=1e-14 by direct helper test). The pweight weights= shortcut routes through a synthetic trivial ResolvedSurveyDesign (new survey._make_trivial_resolved helper) so the same kernel handles both entry paths. did_had_pretest_workflow(..., survey=, weights=) removes the Phase 4.5 C0 NotImplementedError, dispatches to the survey-aware sub-tests, skips the QUG step with UserWarning (per C0 deferral), sets qug=None on the report, and appends a "linearity-conditional verdict; QUG-under-survey deferred per Phase 4.5 C0" suffix to the verdict. HADPretestReport.qug retyped from QUGTestResults to Optional[QUGTestResults]; summary() / to_dict() / to_dataframe() updated to None-tolerant rendering. Replicate-weight survey designs (BRR/Fay/JK1/JKn/SDR) raise NotImplementedError at every entry point (defense in depth, reciprocal-guard discipline) — parallel follow-up after this PR. Stratified designs (SurveyDesign(strata=...)) also raise NotImplementedError on the Stute family — the within-stratum demean + sqrt(n_h/(n_h-1)) correction that the HAD sup-t bootstrap applies to match the Binder-TSL stratified target has not been derived for the Stute CvM functional, so applying raw multipliers from generate_survey_multiplier_weights_batch directly to residual perturbations would leave the bootstrap p-value silently miscalibrated. Phase 4.5 C narrows survey support to pweight-only, PSU-only (SurveyDesign(weights=, psu=)), and FPC-only (SurveyDesign(weights=, fpc=)) designs; stratified is a follow-up after the matching Stute-CvM stratified-correction derivation lands. Strictly positive weights required on Yatchew (the adjacent-difference variance is undefined under contiguous-zero blocks). Per-row weights= / survey=col aggregated to per-unit via existing HAD helpers _aggregate_unit_weights / _aggregate_unit_resolved_survey (constant-within-unit invariant enforced). Unweighted code paths preserved bit-exactly. Patch-level addition (additive on stable surfaces). See docs/methodology/REGISTRY.md § "QUG Null Test" — Note (Phase 4.5 C) for the full methodology.
• ChaisemartinDHaultfoeuille.by_path + n_bootstrap > 0 joint sup-t bands — per-path joint sup-t simultaneous confidence intervals across horizons 1..L_max within each path. A single shared (n_bootstrap, n_eligible) multiplier weight matrix (using the estimator's configured bootstrap_weights — Rademacher / Mammen / Webb) is drawn per path and broadcast across all horizons of that path, producing correlated bootstrap distributions across horizons. The path-specific critical value c_p = quantile(max_l |t_l|, 1 - α) is used to construct symmetric joint bands effect_l ± c_p · se_l per horizon. Surfaced on results.path_sup_t_bands (dict keyed by path tuple, each entry with crit_value / alpha / n_bootstrap / method / n_valid_horizons); as cband_conf_int per horizon entry on path_effects[path]["horizons"][l]; and as cband_lower / cband_upper columns on results.to_dataframe(level="by_path") (mirrors the OVERALL level="event_study" schema; positive-horizon rows of banded paths get populated values, placebo / unbanded / empty-window rows get NaN). Gates: a path needs >= 2 valid horizons (finite bootstrap SE > 0) AND a strict majority (more than 50%) of finite sup-t draws to receive a band. Empty-state contract: path_sup_t_bands is None when not requested; {} when requested but no path passes both gates. Methodology asymmetry vs OVERALL event_study_sup_t_bands: the per-path sup-t draws a fresh shared weight matrix per path AFTER the per-path SE bootstrap block has already populated results.path_ses via independent per-(path, horizon) draws — asymptotically equivalent to OVERALL's self-consistent reuse but NOT bit-identical. Documented intentional choice to preserve RNG-state isolation for existing per-path SE seed-reproducibility tests. Inherits the cross-path cohort-sharing SE deviation from R documented for path_effects. Deviation from R: did_multiplegt_dyn does not provide joint / sup-t bands at any surface — this is a Python-only methodology extension consistent with the existing OVERALL sup-t bands (also Python-only). Bands cover joint inference WITHIN a single path across horizons; they do NOT provide simultaneous coverage across paths. Pre-audit fix bundled: stale "Phase 2 placeholder" docstring on the existing sup_t_bands field updated to the actual contract description. Tests at tests/test_chaisemartin_dhaultfoeuille.py::TestByPathSupTBands (@pytest.mark.slow). See docs/methodology/REGISTRY.md §ChaisemartinDHaultfoeuille Note (Phase 3 by_path per-path joint sup-t bands) for the full contract.
• ChaisemartinDHaultfoeuille.by_path + placebo=True — per-path backward-horizon placebos DID^{pl}_{path, l} for l = 1..L_max. The same per-path SE convention used for the event-study (joiners/leavers IF precedent: switcher-side contributions zeroed for non-path groups; cohort structure and control pool unchanged; plug-in SE with path-specific divisor N^{pl}_{l, path}) is applied to backward horizons via the new switcher_subset_mask parameter on _compute_per_group_if_placebo_horizon. Surfaced on results.path_placebo_event_study[path][-l] (negative-int inner keys mirroring placebo_event_study); summary() renders the rows alongside per-path event-study horizons; to_dataframe(level="by_path") emits negative-horizon rows alongside the existing positive-horizon rows. Bootstrap (when n_bootstrap > 0) propagates per-(path, lag) percentile CI / p-value through the same _bootstrap_one_target dispatch as the per-path event-study, with the canonical NaN-on-invalid contract enforced on the new surface (PR #364 library-wide invariant). SE inherits the cross-path cohort-sharing deviation from R documented for path_effects (full-panel cohort-centered plug-in vs R's per-path re-run): tracks R within tolerance on single-path-cohort panels, diverges materially on cohort-mixed panels — the bootstrap SE is a Monte Carlo analog of the analytical SE and inherits the same deviation. R-parity confirmed at tests/test_chaisemartin_dhaultfoeuille_parity.py::TestDCDHDynRParityByPathPlacebo on the new multi_path_reversible_by_path_placebo scenario (point estimates exact match; SE within Phase-2 envelope rtol ≤ 5%); positive analytical + bootstrap invariants at tests/test_chaisemartin_dhaultfoeuille.py::TestByPathPlacebo (and the gated ::TestBootstrap subclass). See docs/methodology/REGISTRY.md §ChaisemartinDHaultfoeuille Note (Phase 3 by_path ...) → "Per-path placebos" for the full contract.
• Tutorial 19: dCDH for Marketing Pulse Campaigns (docs/tutorials/19_dcdh_marketing_pulse.ipynb) — end-to-end practitioner walkthrough on a 60-market reversible-treatment panel covering the TWFE decomposition diagnostic (twowayfeweights), DCDH Phase 1 (DID_M, joiners-vs-leavers, single-lag placebo), the L_max multi-horizon event study with multiplier bootstrap, a stakeholder communication template, and drift guards. README listing for Tutorial 17 (Brand Awareness Survey) backfilled in the same edit. Cross-link from docs/practitioner_decision_tree.rst § "Reversible Treatment" added.

3.3.0 - 2026-04-25

Fixed

• SyntheticDiD(variance_method="placebo") SE now uses R-default warm-start matching synthdid:::placebo_se. R's placebo loop seeds Frank-Wolfe per draw with weights.boot$omega = sum_normalize(weights$omega[ind[1:N0_placebo]]) (fit-time ω subsetted + renormalized) and the fit-time weights$lambda — Python previously used uniform cold-start, producing finite-iter convergence-pattern drift on a handful of draws relative to R's reference SE. New _placebo_variance_se kwargs init_omega / init_lambda thread fit-time weights through the existing two-pass FW dispatcher; on the global FW optimum the values are init-independent (strictly convex objective), so the change is a finite-iter parity fix, not a methodology change. Existing placebo SE values shift by sub-percent on most panels; the bit-identity baseline pin in TestScaleEquivariance::test_baseline_parity_small_scale[placebo] was rebased from 0.29385822261006445 to 0.293840360160448. New R-parity test tests/test_methodology_sdid.py::TestJackknifeSERParity::test_placebo_se_matches_r asserts SE matches R's vcov(method="placebo") to within < 1e-8 using R's exact permutation sequence (recorded by benchmarks/R/generate_sdid_placebo_parity_fixture.R into tests/data/sdid_placebo_indices_r.json). The _placebo_indices kwarg on _placebo_variance_se is the test seam; not part of the public API.

Added

• qug_test and did_had_pretest_workflow survey-aware NotImplementedError gates (Phase 4.5 C0 decision gate). qug_test(d, *, survey=None, weights=None) and did_had_pretest_workflow(..., *, survey=None, weights=None) now accept the two kwargs as keyword-only with default None. Passing either non-None raises NotImplementedError with an educational message naming the methodology rationale and pointing users to joint Stute (Phase 4.5 C, planned) as the survey-compatible alternative. Mutex guard on survey= + weights= mirrors HeterogeneousAdoptionDiD.fit() at had.py:2890. QUG-under-survey is permanently deferred — the test statistic uses extreme order statistics D_{(1)}, D_{(2)} which are NOT smooth functionals of the empirical CDF, so standard survey machinery (Binder-TSL linearization, Rao-Wu rescaled bootstrap, Krieger-Pfeffermann (1997) EDF tests) does not yield a calibrated test; under cluster sampling the Exp(1)/Exp(1) limit law's independence assumption breaks; and the EVT-under-unequal-probability-sampling literature (Quintos et al. 2001, Beirlant et al.) addresses tail-index estimation, not boundary tests. The workflow's gate is temporary — Phase 4.5 C will close it for the linearity-family pretests with mechanism varying by test: Rao-Wu rescaled bootstrap for stute_test and the joint variants (stute_joint_pretest, joint_pretrends_test, joint_homogeneity_test); weighted OLS residuals + weighted variance estimator for yatchew_hr_test (Yatchew 1997 is a closed-form variance-ratio test, not bootstrap-based). Sister pretests (stute_test, yatchew_hr_test, stute_joint_pretest, joint_pretrends_test, joint_homogeneity_test) keep their closed signatures in this release — Phase 4.5 C will add kwargs and implementation together to avoid API churn. Unweighted qug_test(d) and did_had_pretest_workflow(...) calls are bit-exact pre-PR (kwargs are keyword-only after *; positional path unchanged). New tests at tests/test_had_pretests.py::TestQUGTest (5 rejection / mutex / message / regression tests) and the new TestHADPretestWorkflowSurveyGuards class (6 tests covering both kwarg paths, mutex, methodology pointer, both aggregate paths, and unweighted regression). See docs/methodology/REGISTRY.md § "QUG Null Test" — Note (Phase 4.5 C0) for the full methodology rationale plus a sketch of the (out-of-scope) theoretical bridge that combines endpoint-estimation EVT (Hall 1982, Aarssen-de Haan 1994, Hall-Wang 1999, Beirlant-de Wet-Goegebeur 2006), survey-aware functional CLTs (Boistard-Lopuhaä-Ruiz-Gazen 2017, Bertail-Chautru-Clémençon 2017), and tail-empirical-process theory (Drees 2003) — publishable methodology research, not engineering work.
• HeterogeneousAdoptionDiD mass-point survey= / weights= + event-study aggregate="event_study" survey composition + multiplier-bootstrap sup-t simultaneous confidence band (Phase 4.5 B). Closes the two Phase 4.5 A NotImplementedError gates: design="mass_point" + weights/survey and aggregate="event_study" + weights/survey. Weighted 2SLS sandwich in _fit_mass_point_2sls follows the Wooldridge 2010 Ch. 12 pweight convention (w² in the HC1 meat, w·u in the CR1 cluster score, weighted bread Z'WX); HC1 and CR1 ("stata" se_type) bit-parity with estimatr::iv_robust(..., weights=, clusters=) at atol=1e-10 (new cross-language golden at benchmarks/data/estimatr_iv_robust_golden.json, generated by benchmarks/R/generate_estimatr_iv_robust_golden.R; estimatr added to benchmarks/R/requirements.R). _fit_mass_point_2sls gains weights= + return_influence= kwargs and now always returns a 3-tuple (beta, se, psi) — psi is the per-unit IF on the β̂-scale scaled so compute_survey_if_variance(psi, trivial_resolved) ≈ V_HC1[1,1] at atol=1e-10 (PR #359 IF scale convention applied uniformly; no sum(psi²) claims). Event-study per-horizon variance: survey= path composes Binder-TSL via compute_survey_if_variance; weights= shortcut uses the analytical weighted-robust SE (continuous: CCT-2014 bc_fit.se_robust / |den|; mass-point: weighted 2SLS pweight sandwich from _fit_mass_point_2sls — HC1 / classical / CR1). survey_metadata / variance_formula / effective_dose_mean populated in both regimes (previously hardcoded None at had.py:3366). New multiplier-bootstrap sup-t: _sup_t_multiplier_bootstrap reuses diff_diff.bootstrap_utils.generate_survey_multiplier_weights_batch for PSU-level draws with stratum centering + sqrt(n_h/(n_h-1)) small-sample correction + FPC scaling + lonely-PSU handling. On the weights= shortcut, sup-t calibration is routed through a synthetic trivial ResolvedSurveyDesign so the centered + small-sample-corrected branch fires uniformly — targets the analytical HC1 variance family (compute_survey_if_variance(IF, trivial) ≈ V_HC1 per the PR #359 IF scale invariant) rather than the raw sum(ψ²) = ((n-1)/n) · V_HC1 that unit-level Rademacher multipliers would produce on the HC1-scaled IF. Perturbations: delta = weights @ IF with NO (1/n) prefactor (matching staggered_bootstrap.py:373 idiom), normalized by per-horizon analytical SE, (1-alpha)-quantile of the sup-t distribution. At H=1 the quantile reduces to Φ⁻¹(1 − alpha/2) ≈ 1.96 up to MC noise (regression-locked by TestSupTReducesToNormalAtH1). HeterogeneousAdoptionDiD.__init__ gains n_bootstrap: int = 999 and seed: Optional[int] = None (CS-parity singular seed); fit() gains cband: bool = True (only consulted on weighted event-study). HeterogeneousAdoptionDiDEventStudyResults extended with variance_formula, effective_dose_mean, cband_low, cband_high, cband_crit_value, cband_method, cband_n_bootstrap (all None on unweighted fits); surfaced in to_dict, to_dataframe, summary, __repr__. Unweighted event-study with cband=False preserves pre-Phase 4.5 B numerical output bit-exactly (stability invariant, locked by regression tests). Zero-weight subpopulation convention carries over from PR #359 (filter for design decisions; preserve full ResolvedSurveyDesign for variance). Non-pweight SurveyDesigns (aweight, fweight, replicate designs) raise NotImplementedError on both new paths (reciprocal-guard discipline). Pretest surfaces (qug_test, stute_test, yatchew_hr_test, joint variants, did_had_pretest_workflow) remain unweighted in this release — Phase 4.5 C / C0. See docs/methodology/REGISTRY.md §HeterogeneousAdoptionDiD "Weighted 2SLS (Phase 4.5 B)", "Event-study survey composition", and "Sup-t multiplier bootstrap" for derivations and invariants.
• PanelProfile.outcome_shape and PanelProfile.treatment_dose extensions + llms-autonomous.txt worked examples (Wave 2 of the AI-agent enablement track). profile_panel(...) now populates two new optional sub-dataclasses on the returned PanelProfile: outcome_shape: Optional[OutcomeShape] (numeric outcomes only — exposes n_distinct_values, pct_zeros, value_min / value_max, skewness and excess_kurtosis (NaN-safe; None when n_distinct_values < 3 or variance is zero), is_integer_valued, is_count_like (heuristic: integer-valued AND has zeros AND right-skewed AND > 2 distinct values AND non-negative support, i.e. value_min >= 0; flags WooldridgeDiD QMLE consideration over linear OLS — the non-negativity clause aligns the routing signal with WooldridgeDiD(method="poisson")'s hard rejection of negative outcomes at wooldridge.py:1105-1109), is_bounded_unit ([0, 1] support)) and treatment_dose: Optional[TreatmentDoseShape] (continuous treatments only — exposes n_distinct_doses, has_zero_dose, dose_min / dose_max / dose_mean over non-zero doses). Both OutcomeShape and TreatmentDoseShape are mostly descriptive context. profile_panel does not see the separate first_treat column that ContinuousDiD.fit() consumes; the estimator's actual fit-time gates key off first_treat (defines never-treated controls as first_treat == 0, force-zeroes nonzero dose on those rows with a UserWarning, and rejects negative dose only among treated units first_treat > 0; see continuous_did.py:276-327 and :348-360). In the canonical ContinuousDiD setup (Callaway, Goodman-Bacon, Sant'Anna 2024), the dose D_i is time-invariant per unit and first_treat is a separate column the caller supplies (not derived from the dose column). Under that setup, several facts on the dose column predict fit() outcomes: PanelProfile.has_never_treated (proxies P(D=0) > 0 because the canonical convention ties first_treat == 0 to D_i == 0); PanelProfile.treatment_varies_within_unit == False (the actual fit-time gate at line 222-228, holds regardless of first_treat); PanelProfile.is_balanced (the actual fit-time gate at line 329-338); absence of the duplicate_unit_time_rows alert (silent last-row-wins overwrite, must deduplicate before fit); and treatment_dose.dose_min > 0 (predicts the strictly-positive-treated-dose requirement at line 287-294 because treated units carry their constant dose across all periods). When has_never_treated == False (no zero-dose controls but all observed doses non-negative), ContinuousDiD does not apply (Remark 3.1 lowest-dose-as-control is not implemented); HeterogeneousAdoptionDiD IS a routing alternative on this branch (HAD's own contract requires non-negative dose, which is satisfied). When dose_min <= 0 (negative treated doses), ContinuousDiD does not apply AND HeterogeneousAdoptionDiD is not a fallback — HAD also raises on negative post-period dose (had.py:1450-1459); the applicable alternative is linear DiD with the treatment as a signed continuous covariate. Re-encoding the treatment column is an agent-side preprocessing choice that changes the estimand and is not documented in REGISTRY as a supported fallback. The estimator's force-zero coercion on first_treat == 0 rows with nonzero dose is implementation behavior for inconsistent inputs, not a documented method for manufacturing never-treated controls. The agent must validate the supplied first_treat column independently — profile_panel does not see it. The shape extensions provide distributional context (effect-size range, count-shape detection) that supplements but does not replace those gates. Both fields are None when their classification gate is not met (e.g., treatment_dose is None for binary treatments). to_dict() serializes the nested dataclasses as JSON-compatible nested dicts. New exports: OutcomeShape, TreatmentDoseShape from top-level diff_diff. llms-autonomous.txt gains a new §5 "Worked examples" section with three end-to-end PanelProfile -> reasoning -> validation walkthroughs (binary staggered with never-treated controls, continuous dose with zero baseline, count-shaped outcome) plus §2 field-reference subsections for the new shape fields and §4.7 / §4.11 cross-references for outcome-shape considerations. Existing §5-§8 of the autonomous guide are renumbered to §6-§9. Descriptive only — no recommender language inside the worked examples.
• HeterogeneousAdoptionDiD.fit(survey=..., weights=...) on continuous-dose paths (Phase 4.5 survey support). The continuous_at_zero (paper Design 1') and continuous_near_d_lower (Design 1 continuous-near-d̲) designs accept survey weights through two interchangeable kwargs: weights=<array> (pweight shortcut, weighted-robust SE from the CCT-2014 lprobust port) and survey=SurveyDesign(weights, strata, psu, fpc) (design-based inference via Binder-TSL variance using the existing compute_survey_if_variance helper at diff_diff/survey.py:1802). Point estimates match across both entry paths; SE diverges by design (pweight-only vs PSU-aggregated). HeterogeneousAdoptionDiDResults.survey_metadata is a repo-standard SurveyMetadata dataclass (weight_type / effective_n / design_effect / sum_weights / weight_range / n_strata / n_psu / df_survey); HAD-specific extras (variance_formula label, effective_dose_mean) are separate top-level result fields. to_dict() surfaces the full SurveyMetadata object plus variance_formula + effective_dose_mean; summary() renders variance_formula, effective_n, effective_dose_mean, and (when the survey= path is used) df_survey; __repr__ surfaces variance_formula + effective_dose_mean when present. The HAD mass_point design and aggregate="event_study" path raise NotImplementedError under survey/weights (deferred to Phase 4.5 B: weighted 2SLS + event-study survey composition); the HAD pretests stay unweighted in this release (Phase 4.5 C). Parity ceiling acknowledged — no public weighted-CCF bias-corrected local-linear reference exists in any language; methodology confidence comes from (1) uniform-weights bit-parity at atol=1e-14 on the full lprobust output struct, (2) cross-language weighted-OLS parity (manual R reference) at atol=1e-12, and (3) Monte Carlo oracle consistency on known-τ DGPs. _nprobust_port.lprobust gains weights= and return_influence= (used internally by the Binder-TSL path); bias_corrected_local_linear removes the Phase 1c NotImplementedError on weights= and forwards. Auto-bandwidth selection remains unweighted in this release — pass h/b explicitly for weight-aware bandwidths. See docs/methodology/REGISTRY.md §HeterogeneousAdoptionDiD "Weighted extension (Phase 4.5 survey support)".
• stute_joint_pretest, joint_pretrends_test, joint_homogeneity_test + StuteJointResult (HeterogeneousAdoptionDiD Phase 3 follow-up). Joint Cramér-von Mises pretests across K horizons with shared-η Mammen wild bootstrap (preserves vector-valued empirical-process unit-level dependence per Delgado-Manteiga 2001 / Hlávka-Hušková 2020). The core stute_joint_pretest is residuals-in; two thin data-in wrappers construct per-horizon residuals for the two nulls the paper spells out: mean-independence (step 2 pre-trends, OLS(Y_t − Y_base ~ 1) per pre-period) and linearity (step 3 joint, OLS(Y_t − Y_base ~ 1 + D) per post-period). Sum-of-CvMs aggregation (S_joint = Σ_k S_k); per-horizon scale-invariant exact-linear short-circuit. Closes the paper Section 4.2 step-2 gap that Phase 3 did_had_pretest_workflow previously flagged with an "Assumption 7 pre-trends test NOT run" caveat. See docs/methodology/REGISTRY.md §HeterogeneousAdoptionDiD "Joint Stute tests" for algorithm, invariants, and scope exclusion of Eq 18 linear-trend detrending (deferred to Phase 4 Pierce-Schott replication).
• did_had_pretest_workflow(aggregate="event_study"): multi-period dispatch on balanced ≥3-period panels. Runs QUG at F + joint pre-trends Stute across earlier pre-periods + joint homogeneity-linearity Stute across post-periods. Step 2 closure requires ≥2 pre-periods; with only a single pre-period (the base F-1) pretrends_joint=None and the verdict flags the skip. Reuses the Phase 2b event-study panel validator (last-cohort auto-filter under staggered timing with UserWarning; ValueError when first_treat_col=None and the panel is staggered). The data-in wrappers joint_pretrends_test and joint_homogeneity_test also route through that same validator internally, so direct wrapper calls inherit the last-cohort filter and constant-post-dose invariant. HADPretestReport extended with pretrends_joint, homogeneity_joint, and aggregate fields; serialization methods (summary, to_dict, to_dataframe, __repr__) preserve the Phase 3 output bit-exactly on aggregate="overall" — no aggregate key, no header row, no schema drift — and only surface the new fields on aggregate="event_study".
• ChaisemartinDHaultfoeuille.by_path — per-path event-study disaggregation, mirroring R did_multiplegt_dyn(..., by_path=k). Passing by_path=k (positive int) to the estimator reports separate DID_{path,l} + SE + inference for the top-k most common observed treatment paths in the window [F_g-1, F_g-1+L_max], answering the practitioner question "is a single pulse enough, or do you need sustained exposure?" across paths like (0,1,0,0) vs (0,1,1,0) vs (0,1,1,1). The per-path SE follows the joiners-only / leavers-only IF precedent (switcher-side contribution zeroed for non-path groups; control pool and cohort structure unchanged; plug-in SE with path-specific divisor). Requires drop_larger_lower=False (multi-switch groups are the object of interest) and L_max >= 1. Binary treatment was the only supported case at the initial cut; subsequent entries in this [Unreleased] block lifted that and the original gates one by one. Currently still gated: design2 and honest_did raise NotImplementedError (deferred to follow-up PRs). All other combinations — n_bootstrap > 0, placebo=True, joint sup-t bands, controls, trends_linear, trends_nonparam, survey_design, heterogeneity, non-binary integer treatment, and the paths_of_interest user-specified selector — are now supported, with the per-feature contracts in their dedicated entries elsewhere in [Unreleased]. Results expose results.path_effects: Dict[Tuple[int, ...], Dict[str, Any]] and results.to_dataframe(level="by_path"); the summary grows a "Treatment-Path Disaggregation" block. Ties in path frequency are broken lexicographically on the path tuple for deterministic ranking. Overflow (by_path > n_observed_paths) returns all observed paths with a UserWarning. See docs/methodology/REGISTRY.md §ChaisemartinDHaultfoeuille Note (Phase 3 by_path per-path event-study disaggregation) for the full contract.
• ChaisemartinDHaultfoeuille.by_path + n_bootstrap > 0 — bootstrap SE for per-path event-study effects. The top-k paths are enumerated once on the observed data (R-faithful path-stability semantics: matches did_multiplegt_dyn(..., by_path=k, bootstrap=B), confirmed empirically against DIDmultiplegtDYN 2.3.3), and the existing multiplier bootstrap (bootstrap_weights ∈ {"rademacher", "mammen", "webb"}) runs per (path, horizon) target via the shared _bootstrap_one_target / compute_effect_bootstrap_stats helpers. Point estimates are unchanged from the analytical path. Bootstrap SE replaces the analytical SE in path_effects[path]["horizons"][l]["se"], and p_value / conf_int propagate the bootstrap percentile statistics (library Round-10 convention, same as overall / joiners / leavers / multi_horizon); t_stat is SE-derived via safe_inference per the anti-pattern rule. Interpretation is conditional on the observed path set — practitioners wanting unconditional inference capturing path-selection uncertainty need a pairs-bootstrap (no R precedent). SE inherits the analytical cross-path cohort-sharing deviation: bootstrap input is the same full-panel cohort-centered path IF as the analytical path, so the bootstrap SE is a Monte Carlo analog of the analytical SE and inherits the existing analytical-path divergence from R on mixed-path cohorts (see REGISTRY.md for the full mechanism). On single-path-cohort panels, bootstrap and analytical SE both track R up to the Phase 2 envelope. Deviation from R (CI method): R's per-path bootstrap CI is normal-theory around the bootstrap SE (half-width ≈ 1.96·se); ours is the bootstrap percentile CI, intentionally diverging from R to keep the dCDH inference surface internally consistent across all bootstrap targets. Positive regressions at tests/test_chaisemartin_dhaultfoeuille.py::TestByPathBootstrap (@pytest.mark.slow): point-estimate invariance, finite SE on non-degenerate panels, bootstrap-vs-analytical SE within 30% rtol on cohort-clean panels, degenerate-cohort NaN propagation, Rademacher/Mammen/Webb parity, seed reproducibility, and percentile-vs-normal-theory CI pinning. See docs/methodology/REGISTRY.md §ChaisemartinDHaultfoeuille Note (Phase 3 by_path ...) → Bootstrap SE for the full write-up.
• R-parity for ChaisemartinDHaultfoeuille.by_path against DIDmultiplegtDYN 2.3.3. Two new scenarios in benchmarks/data/dcdh_dynr_golden_values.json generated from did_multiplegt_dyn(..., by_path=k): mixed_single_switch_by_path (2 paths, by_path=2) and multi_path_reversible_by_path (4 observed paths, by_path=3, via a new deterministic multi-path DGP pattern in the R generator). Per-path point estimates and per-path switcher counts match R exactly; per-path SE matches within the Phase 2 multi-horizon SE envelope (observed rtol ≤ 10.2% on the 2-path scenario, ≤ 4.2% on the 4-path scenario). Parity tests live at tests/test_chaisemartin_dhaultfoeuille_parity.py::TestDCDHDynRParityByPath, matching paths by tuple label via set-equality (robust to R's undocumented frequency-tie tiebreak) and cross-checking per-path switcher counts before SE comparison. Deviation documented: cross-path cohort sharing — our full-panel cohort-centered plug-in vs R's per-path re-run diverges materially when a (D_{g,1}, F_g, S_g) cohort spans multiple observed paths; the two coincide when every cohort is single-path. The parity scenarios are constructed to keep cohorts single-path (scenario 13 by design, scenario 14 via path-assignment-deterministic-on-F_g). See docs/methodology/REGISTRY.md §ChaisemartinDHaultfoeuille Note (Phase 3 by_path...) for the full write-up.
• profile_panel() utility + llms-autonomous.txt reference guide (agent-facing) — new diff_diff.profile_panel(df, *, unit, time, treatment, outcome) returns a frozen PanelProfile dataclass of structural facts (panel balance, treatment-type classification — "binary_absorbing" / "binary_non_absorbing" / "continuous" / "categorical", cohort structure, outcome characteristics, and a tuple[Alert, ...] of factual observations). .to_dict() returns a JSON-serializable view. Paired with a new bundled "autonomous" variant on get_llm_guide() — get_llm_guide("autonomous") returns a reference-shaped guide (distinct from the existing workflow-prose "practitioner" variant) with §1 audience disclaimer, §2 PanelProfile field reference, §3 embedded 17-estimator × 9-design-feature support matrix, §4 per-design-feature reasoning citing Baker et al. (2025) and Roth / Sant'Anna (2023), §5 post-fit validation index, §6 BR/DR schema reference, §7 citations, §8 intentional omissions. Both pieces are bundled inside the wheel (no GitHub / RTD dependency at runtime); diff_diff/__init__.py module docstring leads with an agent-entry block listing profile_panel, get_llm_guide("autonomous"), get_llm_guide("practitioner"), and BusinessReport so help(diff_diff) surfaces them. Descriptive, not opinionated — profile_panel alerts never recommend a specific estimator, and the guide enumerates trade-offs rather than dispatching. Exports: profile_panel, PanelProfile, Alert from top-level diff_diff.
• target_parameter block in BR/DR schemas (experimental; schema version bumped to 2.0) — BUSINESS_REPORT_SCHEMA_VERSION and DIAGNOSTIC_REPORT_SCHEMA_VERSION bumped from "1.0" to "2.0" because the new "no_scalar_by_design" value on the headline.status / headline_metric.status enum (dCDH trends_linear=True, L_max>=2 configuration) is a breaking change per the REPORTING.md stability policy. BusinessReport and DiagnosticReport now emit a top-level target_parameter block naming what the headline scalar actually represents for each of the 16 result classes. Closes BR/DR foundation gap #6 (target-parameter clarity). Fields: name, definition, aggregation (machine-readable dispatch tag), headline_attribute (raw result attribute), reference (citation pointer). BR's summary emits the short name right after the headline; DR's overall-interpretation paragraph does the same; both full reports carry a "## Target Parameter" section with the full definition. Per-estimator dispatch is sourced from REGISTRY.md and lives in the new diff_diff/_reporting_helpers.py::describe_target_parameter. A few branches read fit-time config (EfficientDiDResults.pt_assumption, StackedDiDResults.clean_control, ChaisemartinDHaultfoeuilleResults.L_max / covariate_residuals / linear_trends_effects); others emit a fixed tag (the fit-time aggregate kwarg on CS / Imputation / TwoStage / Wooldridge does not change the overall_att scalar — disambiguating horizon / group tables is tracked under gap #9). See docs/methodology/REPORTING.md "Target parameter" section.
• SyntheticDiD coverage Monte Carlo calibration table added to docs/methodology/REGISTRY.md §SyntheticDiD — rejection rates at α ∈ {0.01, 0.05, 0.10} across placebo / bootstrap / jackknife on 3 representative DGPs (balanced / exchangeable, unbalanced, and Arkhangelsky et al. (2021) AER §6.3 non-exchangeable). Artifact at benchmarks/data/sdid_coverage.json (500 seeds × B=200), regenerable via benchmarks/python/coverage_sdid.py.

Fixed

• SyntheticDiD variance_method="bootstrap" now runs the paper-faithful refit bootstrap with R-default warm-start. Re-estimates ω̂_b and λ̂_b via two-pass sparsified Frank-Wolfe on each pairs-bootstrap draw using the fit-time normalized-scale zeta — Arkhangelsky et al. (2021) Algorithm 2 step 2, matching the behavior of R's default synthdid::vcov(method="bootstrap") (which rebinds attr(estimate, "opts") so the renormalized ω serves as Frank-Wolfe initialization). The Python path threads that warm-start through compute_sdid_unit_weights(..., init_weights=_sum_normalize(ω̂[boot_control_idx])) and compute_time_weights(..., init_weights=λ̂) on each bootstrap draw. compute_sdid_unit_weights and compute_time_weights gain a new init_weights kwarg; when provided, the Rust top-level fast-path is skipped in favor of the Python two-pass dispatcher (whose inner FW calls still dispatch to Rust). Without this kwarg both helpers remain backward-compatible and keep the Rust fast-path. The previous fixed-weight bootstrap path is removed entirely — it was not paper-faithful and, despite prior documentation claiming otherwise, also did not match R's default bootstrap (the previous R-parity test fixture invoked synthdid_estimate(weights=...) without rebinding opts, which silently runs fixed-weight, so the 1e-10 parity was between two paths both wrong in the same direction). Coverage MC at the new artifact above quantifies the correctness fix on 3 representative null DGPs. Users' existing variance_method="bootstrap" fits will return materially different SE / p-value / CI values on the next release — same enum name, corrected semantics. Bootstrap is now ~5–30× slower per fit than the old fixed-weight shortcut (panel-size dependent; warm-start converges faster than cold-start so the slowdown is less than the 10–100× prior estimate). The PR #349 follow-on bullets below (analytical p-value dispatch, sqrt((r-1)/r) SE formula, retry-to-B contract) all carry over to the refit path unchanged.
• SyntheticDiD variance_method="bootstrap" now computes p-values from the analytical normal-theory formula using the bootstrap SE (matching R's synthdid::vcov() convention), rather than an empirical null-distribution formula that is not valid for bootstrap draws. is_significant and significance_stars are derived from p_value and will also change for bootstrap fits. Placebo and jackknife are unchanged. Point estimates are unaffected.
• SyntheticDiD bootstrap SE formula applies the sqrt((r-1)/r) correction matching R's synthdid and the placebo SE formula.
• SyntheticDiD bootstrap now retries degenerate resamples (all-control or all-treated, or non-finite τ_b) until exactly n_bootstrap valid replicates are accumulated, matching R's synthdid::bootstrap_sample and Arkhangelsky et al. (2021) Algorithm 2. Previously the Python path counted attempts (with degenerate draws silently dropped), producing fewer valid replicates than requested. A bounded-attempt guard (20 × n_bootstrap) prevents pathological-input hangs.
• TROP global bootstrap SE backend parity under fixed seed — Rust and Python backends now produce bit-identical bootstrap SE under the same seed. Previously Rust's bootstrap_trop_variance_global seeded rand_xoshiro::Xoshiro256PlusPlus per replicate while Python's fallback consumed numpy.random.default_rng (PCG64), producing ~28% SE divergence on tiny panels under seed=42. Fixed by extracting a shared stratified_bootstrap_indices helper in diff_diff/bootstrap_utils.py that pre-generates per-replicate stratified sample indices via numpy on the Python side; both backends consume the same integer arrays through the PyO3 surface. Sampling law (stratified: controls then treated, with replacement) is unchanged. Closes the bootstrap-RNG half of silent-failures audit finding #23 (grid-search half closed in PR #348; local-method methodology half closed by the two Fixed entries below). Local-method TROP also adopts the Python-canonical index contract for the RNG layer here.
• TROP local-method Rust weight-matrix no longer normalized — rust/src/trop.rs::compute_weight_matrix no longer divides time-weights or unit-weights by their respective sums before the outer product. The paper's Equation 2/3 (Athey, Imbens, Qu, Viviano 2025) and REGISTRY.md Requirements checklist (line 2037: [x] Unit weights: exp(-λ_unit × distance) (unnormalized, matching Eq. 2)) both specify raw-exponential weights; Python's _compute_observation_weights was already REGISTRY-compliant. User-visible effect: Rust local-method ATT values may shift for any fit with lambda_nn < infinity — normalizing the weight-matrix inflated the effective nuclear-norm penalty relative to the data-fit term, changing the regularization trade-off. For lambda_nn = infinity (factor model disabled) outputs are unchanged because uniform weight scaling leaves the minimum-norm WLS argmin invariant. Rust LOOCV-selected lambdas may also shift on this boundary; both backends now converge on the same REGISTRY-compliant selection.
• TROP local-method Python _compute_observation_weights now uses the function-argument Y, D and treats all non-target units as donors — two coupled changes that bring Python structurally in line with Rust and the paper's Eq. 2/3:
  1. Removed the if self._precomputed is not None: branch that silently substituted self._precomputed["Y"] / ["D"] / ["time_dist_matrix"] (original-panel cache populated during main fit) for the function-argument Y, D. Under bootstrap, _fit_with_fixed_lambda computes fresh Y, D from the resampled boot_data and passes them in; the helper was discarding those and recomputing unit distances from the original panel, so Python's local bootstrap resampled units but reused stale unit-distance weights. Rust's bootstrap was already correct (always consumed y_boot, d_boot).
  2. Removed the valid_control_at_t = D[t, :] == 0 target-period donor gate that zeroed ω_j for any unit j treated at the target period (other than the target unit itself). Per REGISTRY Eq. 2/3 and Rust's compute_weight_matrix, ω_j = exp(-λ_unit × dist(j, i)) for all j ≠ i; treated-cell exclusion happens via the (1 − W_{js}) factor applied inside _estimate_model. Same-cohort donors now contribute via their pre-treatment rows. Empirically the main-fit ATT is unchanged on tested fixtures because same-cohort pre-treatment observations are exactly absorbed by their own unit fixed effect alpha_j without propagating into mu, beta, or other units' parameters — so this change is structural alignment rather than a numerical shift in output. Users on same-cohort panels with very few controls may still see tiny differences in edge cases; the new test_local_method_same_cohort_donor_parity regression guards the aligned behavior. Together with the normalization fix above, TROP local-method backend parity on the main-fit ATT is regime-dependent: atol=rtol=1e-14 for lambda_nn=inf (no nuclear-norm regularization, uniform weight scaling leaves the WLS argmin invariant) and atol=1e-10 for finite lambda_nn (FISTA inner loop + BLAS reduction ordering introduce sub-1e-10 roundoff across Rust faer vs numpy paths). Bootstrap SE parity is asserted at atol=1e-5 to accommodate ~1e-7 roundoff between Rust's estimate_model matrix factorization and numpy's lstsq that accumulates across per-replicate fits; sub-1e-14 bootstrap parity is tracked as a follow-up in TODO.md under "unify Rust local-method solver path". Closes silent-failures audit finding #23 (local-method half; the RNG half closed in PR #354 and the grid-search half in PR #348).

Changed

• did_had_pretest_workflow(aggregate="event_study") verdict no longer emits the "paper step 2 deferred to Phase 3 follow-up" caveat — the joint pre-trends Stute test closes that gap. The two-period aggregate="overall" path retains the existing caveat since the joint variant does not apply to single-pre-period panels. Downstream code that greps verdict strings for the Phase 3 caveat will see it suppressed on the event-study path.
• SyntheticDiD bootstrap no longer supports survey designs (capability regression in PR #351, restored in PR #355 — see Added/Changed entries directly below). The removed fixed-weight bootstrap path was the only SDID variance method that supported strata/PSU/FPC (via Rao-Wu rescaled bootstrap); the PR #351 paper-faithful refit bootstrap initially rejected all survey designs (including pweight-only) with NotImplementedError. PR #355 restores the capability via a weighted-FW + Rao-Wu composition; the lock-out window applies only to the v3.2.x line that ships PR #351 alone (without PR #355). Composing Rao-Wu rescaled weights with Frank-Wolfe re-estimation: see docs/methodology/REGISTRY.md §SyntheticDiD Note (survey + bootstrap composition).

Added (PR #355)

• SDID variance_method="bootstrap" survey support restored via a hybrid pairs-bootstrap + Rao-Wu rescaling composed with a weighted Frank-Wolfe kernel. Each bootstrap draw first performs the unit-level pairs-bootstrap resampling specified by Arkhangelsky et al. (2021) Algorithm 2 (boot_idx = rng.choice(n_total)), and then applies Rao-Wu rescaled per-unit weights (Rao & Wu 1988) sliced over the resampled units — NOT a standalone Rao-Wu bootstrap. New Rust kernel sc_weight_fw_weighted (and _with_convergence sibling) accepts a per-coordinate reg_weights argument so the FW objective becomes min ||A·ω - b||² + ζ²·Σ_j reg_w[j]·ω[j]². New Python helpers compute_sdid_unit_weights_survey and compute_time_weights_survey thread per-control survey weights through the two-pass sparsify-refit dispatcher (column-scaling Y by rw for the loss, reg_weights=rw for the penalty on the unit-weights side; weighted column-centering + row-scaling Y by sqrt(rw) for the loss with uniform reg on the time-weights side). _bootstrap_se survey branch composes the per-draw rw (Rao-Wu rescaling for full designs, constant w_control for pweight-only fits) with the weighted-FW helpers, then composes ω_eff = rw·ω/Σ(rw·ω) for the SDID estimator. Coverage MC artifact extended with a stratified_survey DGP (BRFSS-style: N=40, strata=2, PSU=2/stratum); the bootstrap row's near-nominal calibration is the validation gate (target rejection ∈ [0.02, 0.10] at α=0.05). New regression tests across test_methodology_sdid.py::TestBootstrapSE (single-PSU short-circuit, full-design and pweight-only succeeds-tests, zero-treated-mass retry, deterministic Rao-Wu × boot_idx slice) and test_survey_phase5.py::TestSyntheticDiDSurvey (full-design ↔ pweight-only SE differs assertion). See REGISTRY.md §SyntheticDiD Note (survey + bootstrap composition) for the full objective and the argmin-set caveat.

Changed (PR #355)

• SDID bootstrap SE values under survey fits now differ numerically from the v3.2.x line that shipped PR #351 alone: the fit no longer raises NotImplementedError, and instead returns the weighted-FW + Rao-Wu SE. Non-survey fits are unaffected (the bootstrap dispatcher routes only the survey branch through the new _survey helpers; non-survey fits continue to call the existing compute_sdid_unit_weights / compute_time_weights and stay bit-identical at rel=1e-14 on the _BASELINE["bootstrap"] regression). SDID's placebo and jackknife paths still reject strata/PSU/FPC on the v3.2.x line; full-design support for those methods lands separately in the entries below.

Added

• SDID variance_method="placebo" and "jackknife" now support strata/PSU/FPC designs. Closes the last SDID survey gap. All three variance methods (bootstrap from PR #355, plus placebo and jackknife here) now handle full survey designs. New private methods SyntheticDiD._placebo_variance_se_survey and _jackknife_se_survey route the full-design path through method-specific allocators:
  • Placebo — stratified permutation (Pesarin 2001). Each draw samples pseudo-treated indices uniformly without replacement from controls within each stratum containing actual treated units; non-treated strata contribute their controls unconditionally. The weighted Frank-Wolfe kernel from PR #355 (compute_sdid_unit_weights_survey / compute_time_weights_survey) re-estimates ω and λ per draw with per-control survey weights threaded into both loss and regularization; post-optimization composition ω_eff = rw·ω/Σ(rw·ω). Arkhangelsky Algorithm 4 SE formula unchanged.
  • Jackknife — PSU-level leave-one-out with stratum aggregation (Rust & Rao 1996). SE² = Σ_h (1-f_h)·(n_h-1)/n_h·Σ_{j∈h}(τ̂_{(h,j)} - τ̄_h)² with f_h = n_h_sampled / fpc[h] (population-count FPC form). λ held fixed across LOOs; ω subsetted, composed with rw, renormalized. Strata with n_h < 2 silently skipped (matches R survey::svyjkn with lonely_psu="remove" / "certainty"; "adjust" raises NotImplementedError). Full-census strata (f_h ≥ 1) short-circuit to zero contribution before any LOO feasibility check. SE = 0 is returned for legitimate zero variance (e.g., every stratum full-census); SE = NaN with a targeted UserWarning is reserved for undefined cases — all strata skipped, or any delete-one replicate in a non-full-census contributing stratum is undefined (all-treated-in-one-PSU LOO, kept ω_eff / w_treated mass zero, estimator raises). Unstratified single-PSU short-circuits to NaN.
  • Fit-time feasibility guards (placebo): ValueError on stratum-level infeasibility with targeted messages distinguishing three cases — Case B (treated-containing stratum has zero controls), Case C (fewer controls than treated in a treated stratum), Case D (every treated stratum is exact-count n_c_h == n_t_h → permutation support is 1, null distribution collapses). Partial-permutation fallback rejected because it would silently change the null-distribution semantics.
  • Gate relaxed: the fit-time guard at synthetic_did.py:352-369 that rejected placebo/jackknife + strata/PSU/FPC is removed. Replicate-weight designs remain rejected (separate methodology — replicate variance is closed-form and would double-count with Rao-Wu-like rescaling). Non-survey and pweight-only paths bit-identical by construction — the new code is gated on resolved_survey_unit.(strata|psu|fpc) is not None.
  • Coverage MC: benchmarks/data/sdid_coverage.json extended with jackknife on stratified_survey. Bootstrap validates near-nominal (α=0.05 rejection = 0.058, SE/trueSD = 1.13). Jackknife reported with an anti-conservatism caveat: with only 2 PSUs per stratum the stratified jackknife formula has 1 effective DoF per stratum, a well-documented limitation of Rust & Rao (1996) — se_over_truesd ≈ 0.46 on this DGP. Users needing tight SE calibration with few PSUs should prefer variance_method="bootstrap". Placebo is structurally infeasible on the existing stratified_survey DGP (its cohort packs into one stratum with 0 never-treated units — by design a bootstrap-suited DGP); the placebo survey path is exercised via unit tests on a feasible fixture.
  • Regression tests across tests/test_survey_phase5.py: two new classes TestSDIDSurveyPlaceboFullDesign and TestSDIDSurveyJackknifeFullDesign. Placebo: pseudo-treated-stratum contract, Case B / Case C front-door guards with targeted-message regression, SE-differs-from-pweight-only, deterministic dispatch. Jackknife: stratum-aggregation self-consistency, FPC magnitude regression (2-stratum handcrafted panel asserts SE_fpc == SE_nofpc · sqrt(1-f) at rtol=1e-10), single-PSU-stratum skip, unstratified short-circuit, all-strata-skipped warning + NaN, SE-differs-from-pweight-only, deterministic dispatch. Existing test_full_design_placebo_raises and test_full_design_jackknife_raises flipped to _succeeds assertions. All 19 existing pweight-only and non-survey placebo/jackknife tests pass unchanged (bit-identity preserved via the new-path gating).
  • Allocator asymmetry (documented in REGISTRY): placebo ignores the PSU axis (unit-level within-stratum permutation — the classical stratified permutation test; PSU-level permutation on few PSUs is near-degenerate); jackknife respects PSU (PSU-level LOO is the canonical survey jackknife). Both respect strata. See docs/methodology/REGISTRY.md §SyntheticDiD Note (survey + placebo composition) and Note (survey + jackknife composition).

3.2.0 - 2026-04-19

Added

• BusinessReport and DiagnosticReport (experimental preview) (PR #318) - practitioner-ready output layer. BusinessReport(results, ...) produces plain-English narrative summaries (.summary(), .full_report(), .export_markdown(), .to_dict()) from any of the 16 fitted result types. DiagnosticReport(results, ...) orchestrates the existing diagnostic battery (parallel trends, pre-trends power, HonestDiD sensitivity, Goodman-Bacon, heterogeneity, design-effect, EPV) plus estimator-native diagnostics for SyntheticDiD (pre_treatment_fit, weight concentration, in-time placebo, zeta sensitivity) and TROP (factor-model fit metrics). Both classes expose an AI-legible to_dict() schema (single source of truth; prose renders from the dict). BR auto-constructs DR by default so summaries mention pre-trends, robustness, and design-effect findings in one call. See docs/methodology/REPORTING.md for methodology deviations including the no-traffic-light-gates decision, pre-trends verdict thresholds (0.05 / 0.30), and power-aware phrasing driven by compute_pretrends_power. Both schemas are marked experimental in this release - wording, verdict thresholds, and schema shape will change; do not anchor downstream tooling on them yet.
• Kernel / local-linear / nonparametric infrastructure (PRs #327, #335) - bandwidth selector, local linear regression, HC2 / Bell-McCaffrey variance helpers, and a port of R nprobust's point-estimate path. Foundation for the upcoming HeterogeneousAdoptionDiD estimator (de Chaisemartin, Ciccia, D'Haultfœuille & Knau 2024 — "DiD with no untreated group"). Released as internal modules with full test coverage (tests/test_bandwidth_selector.py, tests/test_local_linear.py, tests/test_linalg_hc2_bm.py, tests/test_nprobust_port.py); the user-facing estimator ships in a later phase.
• Cell-period IF allocator for dCDH survey variance (Class A contract) (PR #323) - replaces the group-level allocator ψ_i = ψ_g * (w_i / W_g) with a cell-period allocator ψ_i = ψ_g * (w_i / W_{g, out_idx}) on the post-period cell for the DID_l replicate-weight ATT path. Is the allocator shape that the v3.2.0 heterogeneity and bootstrap extensions below build on. Documents the post-period attribution convention in REGISTRY.md with a hand-computed row-sum identity test.

Performance

• aggregate_survey stratum-PSU scaffolding precompute — the per-cell Taylor-series variance inside aggregate_survey no longer rebuilds stratum-PSU scaffolding on every cell. A frozen _PsuScaffolding (strata codes, global PSU codes unique across strata, per-stratum counts and FPC ratios, singleton mask, static legitimate-zero counts and variance-computable flag) is precomputed once per design at the top of aggregate_survey and threaded through _cell_mean_variance to a new _compute_if_variance_fast path that replaces the per-stratum pandas groupby with two vectorized np.bincount passes. BRFSS-shaped 50-state × 10-year × 1M-row microdata → state-year panel drops from ~24s to sub-2s under both backends (the path is pure Python, so Python and Rust track each other). Numerical output is preserved to sub-ULP tolerance; seven-case equivalence tests (TestAggregateSurveyScaffolding) assert assert_allclose(atol=1e-14, rtol=1e-14) between fast and legacy paths across stratified+PSU+FPC, stratified no FPC, PSU-only, weights-only, and all three lonely_psu modes (remove / certainty / adjust). Replicate-weight designs continue to route through compute_replicate_if_variance unchanged. _compute_stratified_psu_meat is untouched — all other TSL callers (DiD / TWFE / CS / etc.) are unaffected.

Changed

• Add Zenodo DOI badge to README; upgrade the BibTeX citation block with the concept DOI (10.5281/zenodo.19646175) and list author as Isaac Gerber (matching CITATION.cff). CITATION.cff carries the concept DOI as its top-level doi: field — Zenodo auto-mints a versioned DOI for every release, but the CFF file tracks the concept DOI only so it doesn't need a follow-up edit per release. DOI was minted by Zenodo when v3.1.3 was released.
• ChaisemartinDHaultfoeuille heterogeneity + within-group-varying PSU/strata now supported under Binder TSL - fit(heterogeneity=..., survey_design=...) no longer raises NotImplementedError when the resolved design's PSU or strata vary across the cells of a group. On the Binder TSL branch (compute_survey_if_variance), the heterogeneity WLS coefficient IF is expanded to observation level via the cell-period allocator ψ_i = ψ_g * (w_i / W_{g, out_idx}) on the post-period cell — the DID_l post-period single-cell convention shipped in v3.1.x. Under PSU=group the PSU-level Binder TSL variance is byte-identical to the previous release (PSU-level aggregate telescopes to ψ_g); under within-group-varying PSU, mass lands in the post-period PSU of the transition. The Rao-Wu replicate-weight branch (compute_replicate_if_variance) retains the legacy group-level allocator ψ_i = ψ_g * (w_i / W_g): replicate variance computes θ_r = sum_i ratio_ir * ψ_i at observation level and is therefore not PSU-telescoping, so the cell-period allocator would silently change the replicate SE whenever a replicate column's ratios vary within group (e.g., per-row replicate matrices). Replicate + heterogeneity fits therefore produce byte-identical SE to the previous release, and the newly-unblocked heterogeneity= + within-group-varying PSU combination is unreachable under replicate designs by construction (SurveyDesign rejects replicate_weights combined with explicit strata/psu/fpc).
• ChaisemartinDHaultfoeuille.fit(survey_design=..., n_bootstrap > 0) now supports within-group-varying PSU — the PSU-level Hall-Mammen wild multiplier bootstrap has been extended from a group-level PSU map (one multiplier per group) to a cell-level PSU map (one multiplier per (g, t) cell's PSU). A dispatcher in _compute_dcdh_bootstrap detects PSU-within-group-constant regimes (including PSU=group auto-inject and strictly-coarser PSU with within-group constancy) and routes them through the legacy group-level path so the bootstrap SE is bit-identical to the previous release (guarded by the new test_bootstrap_se_matches_pre_pr4_baseline and the pre-existing test_auto_inject_bit_identical_to_group_level). Under within-group-varying PSU, a group contributing cells to multiple PSUs receives independent multiplier draws per PSU — the correct Hall-Mammen wild PSU clustering at cell granularity. Multi-horizon bootstraps draw a single shared (n_bootstrap, n_psu) PSU-level weight matrix per block and broadcast per-horizon via each horizon's cell-to-PSU map, so the sup-t simultaneous confidence band remains a valid joint distribution. Closes the last NotImplementedError gate in the dCDH survey contract; replicate-weight variance and n_bootstrap > 0 remain mutually exclusive by construction. Scope note: panels with terminal missingness where the terminally-missing group is in a cohort whose other groups still contribute at the missing period now raise a targeted ValueError on every survey variance path that uses the cell-period allocator: Binder TSL with within-group-varying PSU, Rao-Wu replicate-weight ATT (which always uses the cell allocator per the Class A contract shipped in PR #323), and the cell-level wild PSU bootstrap. Cohort-recentering leaks centered IF mass onto cells with no positive-weight observations, which the cell-period allocator cannot attach to any observation/PSU. This closes a silent mass-drop bug the cell-period allocator introduced across all three paths in v3.1.x; pre-process the panel to remove terminal missingness (drop late-exit groups or trim to a balanced sub-panel) as the documented workaround. For Binder TSL only, using an explicit psu=<group_col> routes through the legacy group-level allocator where the row-sum identity makes the two allocators statistically equivalent. Replicate-weight ATT and within-group-varying-PSU bootstrap have no such allocator fallback — the panel itself must be pre-processed. PSU-within-group-constant Binder TSL (including PSU=group auto-inject) is unaffected.
• Performance review: practitioner-scale scenarios + benchmark harness extension (PR #333) - new docs/performance-scenarios.md documents 5-7 realistic practitioner workflows (marketing lift, geo-experiment, BRFSS state-policy, dCDH reversible treatment) grounded in the practitioner docs and the paper literature, not cookie-cutter textbook data. benchmarks/speed_review/ extended with practitioner-scale scripts and per-backend bit-identity baselines. Baselines refreshed against current main. Finding: the biggest leverage areas are bootstrap resampling loops and per-replicate survey-design rebuilds in the bootstrap path; documented in docs/performance-plan.md for follow-up optimization PRs.
• Wall-clock timing tests excluded from default CI (PRs #330, #336) - TestCallawaySantAnnaSEAccuracy.test_timing_performance and TestPerformanceRegression marked @pytest.mark.slow, removing false-positive CI failures from runner-noise variance (BLAS path variation, neighbor VM contention). Tests remain runnable via pytest -m slow for ad-hoc local benchmarking; the perf-review harness above is the principled replacement for CI-gated performance tracking.

Fixed

• Silent-failures audit: axis A (PR #334) — minor solver paths numerical-precision / scale-fragility closeouts, completing the SDID extreme-Y-scale work started in v3.1.2.
• Silent-failures audit: axis C & J (PR #339) — B-spline derivative warning scope broadened; SurveyPowerConfig stale-cache wording narrowed.
• Silent-failures audit: axis E (PR #331) — row-drop counters surfaced across estimator paths so silent validator row-drops leave an explicit count on the result.
• Silent-failures audit: axis G (PR #337) — Rust vs Python backend edge-case parity tests added for rank-deficient, extreme-scale, and constant-column inputs.
• SyntheticDiD diagnostic Y-normalization parity (PR #328) — extends the PR #312 catastrophic-cancellation fix from the main fit path into SyntheticDiDResults.in_time_placebo() and .sensitivity_to_zeta_omega(). Diagnostics now apply the same Y_shift / Y_scale normalization the main fit uses, pass zeta / Y_scale and a normalized min_decrease into Frank-Wolfe, then rescale att / pre_fit_rmse back to original-Y units.
• TROP bootstrap failure-rate guards (PR #324) — alternating-minimization bootstrap loops now emit a UserWarning on silent high-failure-rate runs (LOOCV and bootstrap aggregation paths both covered); attempt-count-based warning replaces the previous observation-count denominator that could silently mask sparse runs.
• simulate_power() failure-count surface + narrow except clause (PR #326) — power-simulation replicate loop narrows the exception whitelist from except Exception to estimation/data-path failures (TypeError and friends now propagate, not silently absorb), and surfaces n_simulation_failures on SimulationPowerResults. Failure count included in summary() and to_dict().

3.1.3 - 2026-04-18

Added

• Replicate-weight variance and PSU-level bootstrap for dCDH (PR #311) - ChaisemartinDHaultfoeuille now accepts variance_method="replicate" for BRR / Fay / JK1 / JKn / SDR inference, and PSU-level multiplier bootstrap when survey_design.psu is set. Adds df-aware inference (reduced effective df under replicate variance; propagated through delta / HonestDiD surfaces) plus group-level PSU map construction. Validated via per-cohort aggregation, shared-draw multi-horizon bootstrap alignment, and cross-surface df consistency.
• Zenodo DOI auto-minting configuration (PR #321) - .zenodo.json at repo root defines release metadata so the next GitHub Release automatically mints a Zenodo DOI (concept DOI + versioned DOI). Also adds a top-level LICENSE file for Zenodo archival.

Fixed

• Silent sparse→dense lstsq fallback in ImputationDiD and TwoStageDiD (PR #319) - when the sparse solver fails and the dense fallback runs, the estimator now emits a UserWarning instead of silently switching paths. Regression tests assert the dense fallback SEs remain usable.
• Non-convergence signaling in TROP alternating-minimization solvers (PR #317) - the global- and local-TROP solvers now emit a UserWarning when the alternating-minimization loop exits without meeting tolerance, including LOOCV and bootstrap aggregation paths. Warnings aggregate at top-level call sites to avoid log spam.

Changed

• /bump-version skill updates CITATION.cff (PR #320) - internal release-management tooling now keeps CITATION.cff version: and date-released: in sync with the other version surfaces. Resolves a single RELEASE_DATE upfront (from the CHANGELOG header if pre-populated, else today's date) and threads it through all date-bearing files — fixes drift that caused v3.1.2 to ship with CITATION.cff still pinned at 3.1.1.

3.1.2 - 2026-04-18

Fixed

• SyntheticDiD catastrophic cancellation at extreme Y scale (PR #312) - the Frank-Wolfe weight solver lost precision when outcome magnitudes were very large or very small; results are now numerically stable across scales.
• Non-convergence signaling in FE imputation alternating-projection solvers (PR #314) - ImputationDiD, TwoStageDiD, and shared within_transform now emit a UserWarning when the alternating-projection / weighted-demean loop exits without meeting the tolerance. max_iter and tol are documented on within_transform.
• Non-convergence signaling in SyntheticDiD Frank-Wolfe solver (PR #315) - the numpy-path Frank-Wolfe SC weight solver now emits a UserWarning when the loop exits without meeting min_decrease. Wrapper-level and max_iter=0 regression tests added.

Changed

• Refresh ROADMAP.md to drop top-level phase numbering and reflect shipped state through v3.1.1 (PR #313). Absorbs dCDH into the Current State estimator list; adds Recently Shipped summary; reorganizes open work as Shipping Next / Under Consideration / AI-Agent Track / Long-term. Updates docs/business-strategy.md, docs/survey-roadmap.md, docs/practitioner_decision_tree.rst, docs/choosing_estimator.rst, docs/api/chaisemartin_dhaultfoeuille.rst, README.md, and diff_diff/guides/llms-full.txt to remove stale phase-deferral language now that the deferred items have shipped.
• Bump the SyntheticDiD(lambda_reg=...) and SyntheticDiD(zeta=...) deprecation warnings' removal target from v3.1 to v4.0.0. Removing public kwargs in a patch / minor release would violate Semantic Versioning; the deprecation stays warning-only throughout the 3.x line and will be removed in the next major release. Use zeta_omega / zeta_lambda instead.

3.1.1 - 2026-04-16

Added

• Jackknife variance estimation for SyntheticDiD - variance_method='jackknife' implements the delete-one-unit jackknife from Arkhangelsky et al. (2021) Section 5. Supports both standard and survey-weighted jackknife with automatic pweight propagation. Validated against R synthdid package.
• LinkedIn carousel for dCDH estimator announcement (carousel/diff-diff-dcdh-carousel.pdf)

3.1.0 - 2026-04-14

Added

• dCDH Phase 3: Complete feature set for ChaisemartinDHaultfoeuille - three sub-releases completing the estimator:
  • Phase 3a (PR #300): Placebo SE via multiplier bootstrap (resolves Phase 1 deferral), non-binary treatment support with crossing-cell detection and automatic cell dropping, R parity SE assertions tightened
  • Phase 3b (PR #302): Covariate adjustment via controls parameter (OLS residualization, Design 2 per-period path for non-binary treatment), group-specific linear trends via trends_linear=True (absorbs group-specific slopes before DiD), R DIDmultiplegtDYN parity tests for covariates and trends
  • Phase 3c (PR #303): HonestDiD sensitivity analysis integration - honest_did() method on results with automatic event-study-to-sensitivity bridge, support trimming for non-consecutive horizons, l_vec target specification, Delta-RM and Delta-SD smoothness bounds

Changed

• ROADMAP.md updated: dCDH Phase 3 items marked shipped

3.0.2 - 2026-04-12

Added

• ChaisemartinDHaultfoeuille (alias DCDH) - de Chaisemartin & D'Haultfœuille estimator for non-absorbing (reversible) treatments. The only modern staggered DiD estimator that handles treatment switching on AND off. Implements DID_M from AER 2020, validated against R DIDmultiplegtDYN v2.3.3. Ships Phases 1 and 2:
  • Phase 1: headline DID_M with analytical SE, joiners/leavers decompositions, single-lag placebo, multiplier bootstrap, TWFE decomposition diagnostic
  • Phase 2: multi-horizon event study (L_max), dynamic placebos, normalized estimator, cost-benefit aggregate (Lemma 4), sup-t simultaneous confidence bands, plot_event_study() integration
• twowayfeweights() - standalone TWFE decomposition diagnostic (Theorem 1, AER 2020)
• generate_reversible_did_data() - reversible-treatment panel data generator with 7 switch patterns
• Survey-aware power analysis - analytical helpers (compute_power(), compute_mde(), compute_sample_size()) accept a deff parameter for design-effect adjustment. Simulation helpers (simulate_power, simulate_mde, simulate_sample_size) accept a survey_config (SurveyPowerConfig) that generates data with complex survey structure and injects a SurveyDesign into each simulated fit.
• aggregate_survey() second_stage_weights parameter - choose "pweight" (default, population weights) or "aweight" (precision weights). pweight output is compatible with all survey-capable estimators; aweight is opt-in for GLS efficiency with estimators marked Full in the survey support matrix.
• conditional_pt parameter on generate_survey_did_data() - simulates scenarios where unconditional parallel trends fail but conditional PT holds after covariate adjustment
• Tutorial 18: Geo-Experiment Analysis (18_geo_experiments.ipynb) - SyntheticDiD walkthrough for marketing analytics: simulated DMA panel, 5 treated markets, fit + diagnostics + stakeholder summary
• Practitioner decision tree (docs/practitioner_decision_tree.rst) - "which method fits my business problem?" guide
• Practitioner getting started guide (docs/practitioner_getting_started.rst) - end-to-end walkthrough with terminology bridge
• JOSS paper (paper.md, paper.bib) - software paper for Journal of Open Source Software submission
• CONTRIBUTORS.md - author and contributor credit
• Standalone CI Gate workflow (.github/workflows/ci-gate.yml) - doc-only PRs no longer block on path-filtered test workflows

Changed

• aggregate_survey() default second-stage weights changed from aweight (precision) to pweight (population). Users who need the old precision-weighting behavior can pass second_stage_weights="aweight".
• README "For Data Scientists" section with practitioner-facing links and aggregate_survey() documentation
• CITATION.cff updated with version and release date
• ROADMAP.md updated: B1a-d marked done, B2b marked done, B3d marked shipped, dCDH entry updated with correct citations

Fixed

• Doc-only PRs no longer block indefinitely on CI Gate (standalone gate workflow runs on all PRs regardless of path filters)
• aggregate_survey() docs no longer overclaim universal estimator compatibility - explicitly document aweight/pweight restrictions per the survey support matrix

3.0.1 - 2026-04-07

Added

• aggregate_survey() — new function in diff_diff.prep that bridges individual-level survey microdata to geographic-period panels for DiD estimation. Computes design-based cell means and precision weights using domain estimation (Lumley 2004), with SRS fallback for small cells. Returns a panel DataFrame and pre-configured SurveyDesign for second-stage estimation. Supports both TSL and replicate-weight variance.
• Python 3.14 support — upgraded PyO3 from 0.22 to 0.28, updated CI and publish workflow matrices, bumped Rust MSRV to 1.84 for faer 0.24 compatibility.

Changed

• Updated README Python support matrix to include 3.14

Fixed

• Fix domain estimation zero-padding for correct design-based cell variance
• Fix SRS fallback weight normalization for scale invariance across replicate designs
• Validate numeric dtype for outcomes/covariates before aggregation (nullable dtype support)
• Validate grouping columns for NaN values

3.0.0 - 2026-04-07

v3.0 completes the survey support roadmap: all 16 estimators (15 inference-level + BaconDecomposition diagnostic) now accept survey_design. See v2.8.0–v2.9.1 entries for the full feature history leading to this release.

Breaking Changes

• Remove bootstrap_weight_type parameter from CallawaySantAnna — use bootstrap_weights instead (deprecated since v1.0.1)
• Remove TROP method="twostep" alias — use method="local" (deprecated since v2.7.2)
• Remove TROP method="joint" alias — use method="global" (deprecated since v2.7.2)

Upgrading from v2.x

• CallawaySantAnna(bootstrap_weight_type="mammen") → CallawaySantAnna(bootstrap_weights="mammen")
• TROP(method="twostep") → TROP(method="local")
• TROP(method="joint") → TROP(method="global")

Deprecated

• SyntheticDiD lambda_reg and zeta parameters formally scheduled for removal in v3.1 — use zeta_omega/zeta_lambda instead

Changed

• Internal attribute bootstrap_weight_type renamed to bootstrap_weights in bootstrap mixin and StaggeredTripleDifference for consistency
• TROP set_params() now validates method against ("local", "global") — previously only validated in __init__
• Documentation updated: all survey gap notes for WooldridgeDiD removed, ROADMAP Phase 10 items marked shipped

2.9.1 - 2026-04-06

Added

• Survey theory document (docs/methodology/survey-theory.md) — formal justification for design-based variance estimation with modern DiD influence functions, citing Binder (1983), Rao & Wu (1988), Shao (1996)
• Research-grade survey DGP — 8 new parameters on generate_survey_did_data(): icc, weight_cv, informative_sampling, heterogeneous_te_by_strata, te_covariate_interaction, covariate_effects, strata_sizes, return_true_population_att. All backward-compatible.
• R validation expansion — 4 additional estimators cross-validated against R's survey::svyglm(): ImputationDiD, StackedDiD, SunAbraham, TripleDifference. Survey R validation coverage now 8 of 16 estimators.
• LinkedIn carousel for Wooldridge ETWFE estimator announcement

Changed

• Survey tutorial rewritten: leads with "Why Survey Design Matters" section showing flat-weight vs design-based comparison with known ground truth, coverage simulation, and false pre-trend detection rates
• Documentation refresh: ROADMAP.md, llms.txt, llms-full.txt, llms-practitioner.txt, choosing_estimator.rst updated for v2.9.0 — added WooldridgeDiD and StaggeredTripleDifference, DDD flowchart branch, standardized estimator counts, qualified survey claims
• Survey roadmap updated: Phase 10a-10d marked shipped, conditional PT noted for 10e

Fixed

• Fix stale "EfficientDiD covariates + survey not supported" note in choosing_estimator.rst
• Fix WooldridgeDiD described as "ASF-based" for OLS path (OLS uses direct coefficients; ASF only for logit/Poisson)
• Fix dead StaggeredTripleDifference API link in llms.txt
• Fix survey example attribute: .design_effect not .deff in llms-full.txt
• Fix subpopulation() example to show tuple unpacking in llms-full.txt
• Remove 8 resolved items from TODO.md

2.9.0 - 2026-04-04

Added

• WooldridgeDiD (ETWFE) estimator — Extended Two-Way Fixed Effects from Wooldridge (2025, 2023). Supports OLS, logit, and Poisson QMLE paths with ASF-based ATT and delta-method SEs. Four aggregation types (simple, group, calendar, event) matching Stata jwdid_estat. Alias: ETWFE. (PR #216, thanks @wenddymacro)
• EfficientDiD survey + covariates — doubly robust covariate path now threads survey weights through all four nuisance estimation stages (outcome regression, propensity ratio sieve, inverse propensity sieve, kernel-smoothed conditional Omega*). Previously raised NotImplementedError.
• Survey real-data validation (Phase 9) — 15 cross-validation tests against R's survey package using three real federal survey datasets:
  • API (R survey package): TSL variance with strata, FPC, subpopulations, covariates, and Fay's BRR replicates
  • NHANES (CDC/NCHS): TSL variance with strata + PSU + nest=TRUE, validating the ACA young adult coverage provision DiD
  • RECS 2020 (U.S. EIA): JK1 replicate weight variance with 60 pre-computed replicate columns
  • ATT, SE, df, and CI match R to machine precision (< 1e-10) where directly comparable; known deviations documented in REGISTRY.md (TWFE SE differs due to unit FE absorption; subpopulation df differs due to strata preservation)
• Label-gated CI — test workflows now require ready-for-ci label before running, reducing wasted CI during AI review rounds. AI review workflow always runs.
• Documentation dependency map (docs/doc-deps.yaml) — maps source files to impacted documentation. New /docs-impact skill flags which docs need updating when source files change.

Changed

• WooldridgeDiD: full interacted covariate basis (D_g × X, f_t × X) for OLS path
• /submit-pr, /push-pr-update, /pre-merge-check, /docs-check skills updated for label-gated CI and doc-deps workflow

Fixed

• Fix WooldridgeDiD OLS unbalanced demeaning and nonlinear never-treated identification
• Fix WooldridgeDiD Poisson dropped-cell bug and anticipation propagation
• Fix EfficientDiD IF-scale mismatch in survey aggregation and zero-weight never-treated guard
• Fix bootstrap clustering and delta-method reduced space in WooldridgeDiD

2.8.4 - 2026-04-04

Added

• SDR replicate method (Phase 8a) — Successive Difference Replication for ACS PUMS users. SurveyDesign(replicate_method="SDR") with variance formula V = 4/R * sum((theta_r - theta)^2).
• FPC support for ImputationDiD and TwoStageDiD (Phase 8b) — finite population correction now threaded through TSL variance for both estimators.
• Lonely PSU "adjust" in bootstrap (Phase 8d) — lonely_psu="adjust" now works with survey-aware bootstrap (previously raised NotImplementedError). Uses Rust & Rao (1996) grand-mean centering.
• CV on estimates (Phase 8e) — coef_var property on all results objects (SE/estimate). Handles edge cases (SE=0, estimate=0).
• Weight trimming utility (Phase 8e) — trim_weights(data, weight_col, upper=None, lower=None, quantile=None) in prep.py for capping extreme survey weights.
• ImputationDiD pretrends + survey (Phase 8e) — pre-trends F-test now survey-aware using subpopulation approach for correct variance under complex designs.
• Updated ImputationDiD tutorial to demonstrate pretrends=True event study
• Updated survey tutorial: narrative improvements, chart rendering fixes

Fixed

• Fix survey pretrend F-test df calculation and rank-deficient survey VCV handling
• Fix trim_weights NaN poisoning when weight column contains missing values
• Fix single-singleton PSU warning for lonely_psu="adjust"

2.8.3 - 2026-04-02

Added

• Silent operation warnings — 8 operations that previously altered analysis results without informing the user now emit UserWarning:
  • TROP lstsq → pseudo-inverse numerical fallback
  • TwoStageDiD NaN masking of unidentified fixed effects (zeroed out with treatment indicator)
  • TwoStageDiD always-treated unit removal (sample size change)
  • CallawaySantAnna silent (g,t) pair skipping (zero treated or control observations)
  • TROP missing treatment indicator fill with 0 (control)
  • Rust → Python backend fallback (previously debug log only)
  • Survey weight normalization (pweights/aweights rescaled to mean=1)
  • np.inf → 0 never-treated convention conversion
• ImputationDiD pre-period event study coefficients — pre-treatment "effects" (should be ~0 under parallel trends) for visual pre-trends assessment, following BJS (2024) Test 1
• TwoStageDiD pre-period event study coefficients — same pre-trends extension
• Replicate weight expansion to 7 additional estimators: DifferenceInDifferences, TwoWayFixedEffects, MultiPeriodDiD, SunAbraham, StackedDiD, ImputationDiD, TwoStageDiD (coverage: 4/13 → 11/13)

Changed

• ImputationDiD pre-period coefficients use BJS Test 1 (impute Y(0) for treated units in pre-treatment periods)
• SunAbraham replicate weights use full interaction-weighted refit per replicate with cohort-level SEs

Fixed

• Fix zero-weight demeaning safety in replicate weight paths
• Fix df_survey writeback for rank-deficient replicate designs (df=0)
• Fix ImputationDiD balance_e zero-qualifying-cohort fallback in pretrends path
• Fix survey zero-mass (g,t) skip warning gap
• Fix SunAbraham positional assignment in replicate loop

2.8.2 - 2026-04-02

Added

• EPV diagnostics for propensity score logit — events-per-variable (EPV) checks with Peduzzi convention (predictors excluding intercept) for CallawaySantAnna IPW/DR, TripleDifference IPW/DR, and StaggeredTripleDifference
• epv_summary() / epv_diagnostics on post-fit results for CallawaySantAnna, TripleDifference, and StaggeredTripleDifference
• diagnose_propensity() pre-estimation helper on CallawaySantAnna
• EPV summary block in TripleDifference summary() output
• epv_threshold parameter for propensity score estimation — warns on low EPV (default) or escalates via rank_deficient_action="error"

Changed

• Default propensity score fallback behavior: safer defaults with method-specific warning messages
• EPV denominator uses predictor count excluding intercept (Peduzzi et al. 1996 convention)

Fixed

• Fix TripleDifference survey-weighted fallback propensity score
• Fix NaN cache poisoning in propensity score estimation
• Fix epv_summary column schema on empty results
• Fix SDDD EPV: use min-EPV across comparison cohorts with cache diagnostic propagation
• Fix diagnose_propensity np.inf handling

2.8.1 - 2026-04-01

Added

• Survey-aware DiD tutorial (docs/tutorials/16_survey_did.ipynb) — Phase 7c complete. Full workflow with strata, PSU, FPC, replicate weights, subpopulation analysis, and DEFF diagnostics. Includes generate_survey_did_data() DGP function.
• Survey R cross-validation — benchmark scripts and tests comparing TSL variance against R's survey::svyglm() for basic DiD and TWFE with full survey designs (strata, PSU, FPC). Committed JSON fixtures for CI without R.
• HonestDiD methodology review and validation — 478 lines of methodology tests, paper review document, rewritten optimal FLCI with first-difference reparameterization.
• StaggeredTripleDifference survey support — full SurveyDesign integration with strata/PSU/FPC, replicate weights, and survey-aware bootstrap.

Changed

• HonestDiD: rewrite optimal FLCI with proper first-difference reparameterization and centrosymmetric LP optimization
• HonestDiD: use conf_int from results instead of hardcoded 1.96*se in event study plots
• Survey tutorial cross-referenced from choosing_estimator.rst and quickstart.rst

Fixed

• Fix HonestDiD identified set computation and inference (F1-F6 from Rambachan & Roth 2023)
• Fix FLCI slope count (T not T-1) and constraint formula
• Fix NaN CI misclassification as significant (P0 finding)
• Fix M=0 linear extrapolation and survey df folded nct in REGISTRY.md
• Fix replicate-weight scale invariance and BRR test fixtures
• Fix JK1 populated-PSU guard and narrow warning filter

2.8.0 - 2026-03-31

Added

• Staggered Triple Difference estimator (Ortiz-Villavicencio & Sant'Anna 2025)
  • StaggeredTripleDifference class with group-time ATT(g,t) for DDD designs with staggered adoption
  • Event study aggregation, pre-treatment placebo effects, multiplier bootstrap inference
  • R benchmark validation against triplediff package
  • DGP function generate_staggered_ddd_data() for simulation and testing
• Survey Phase 7a: CS IPW/DR + covariates + survey
  • DRDID panel nuisance-estimation IF corrections (PS + OR) under survey weights
  • Survey-weighted propensity score estimation and outcome regression
  • IFs account for nuisance parameter estimation uncertainty (Sant'Anna & Zhao 2020, Theorem 3.1)
• Survey Phase 7b: Repeated cross-sections
  • CallawaySantAnna(panel=False) for repeated cross-section surveys (BRFSS, ACS, CPS)
  • Cross-sectional DRDID: reg matches DRDID::reg_did_rc, dr matches DRDID::drdid_rc, ipw matches DRDID::std_ipw_did_rc
  • Survey weights, covariates, and all estimation methods supported
• Survey Phase 7d: HonestDiD + survey variance
  • Survey df and full event-study VCV from IF vectors propagated to sensitivity analysis
  • t-distribution critical values with survey degrees of freedom
  • Bootstrap/replicate designs fall back to diagonal VCV with warning
• Plotly visualization styling: thread marker, markersize, linewidth, capsize, ci_linewidth kwargs through plotly backends (previously silently ignored)
• AI agent discoverability for practitioner guide

Changed

• HonestDiD now raises ValueError on non-consecutive event-time grid (was warning)
• HonestDiD validates full grid around reference period
• Panel IPW/DR PS correction scaling matches R's H/n, asy_rep/n, colMeans convention
• RC IF normalization follows R's psi convention with explicit phi conversion

Fixed

• Fix HonestDiD reference-aware pre/post split for varying-base event studies
• Fix HonestDiD _estimate_max_pre_violation to use reference-aware pre_periods
• Fix panel M2 gradient scaling for IPW/DR nuisance IF corrections
• Fix VCV index alignment for repeated cross-section aggregation
• Fix replicate-weight df propagation: return per-statistic df instead of mutating shared state
• Fix WIF population consistency: zero df first_treat for ineligible units
• Fix bootstrap RCS cohort-mass weighting and stale event-study VCV reset

2.7.6 - 2026-03-28

Added

• AI practitioner guardrails based on Baker et al. (2025) "Difference-in-Differences Designs: A Practitioner's Guide"
  • practitioner.py module with 8-step workflow enforcement for AI agents
  • Estimator-specific handlers ensuring correct diagnostic ordering (pre-trends before estimation, Bacon decomposition before estimator selection)
  • docs/llms.txt, docs/llms-practitioner.txt, docs/llms-full.txt for AI agent discoverability
  • Evaluation rubric (docs/practitioner-guide-evaluation.md) with correctness-aware scoring
• Survey Phase 6: Advanced features
  • Survey-aware bootstrap for all 8 bootstrap-using estimators (PSU-level multiplier for CS/Imputation/TwoStage/Continuous/Efficient; Rao-Wu rescaled for SA/SyntheticDiD/TROP)
  • Replicate weight variance estimation (BRR, Fay's BRR, JK1, JKn) for OLS-based and IF-based estimators
  • Per-coefficient DEFF diagnostics comparing survey vs SRS variance
  • Subpopulation analysis via SurveyDesign.subpopulation() preserving full design structure
  • CS analytical expansion: strata/PSU/FPC for aggregated SEs via compute_survey_if_variance()
  • TROP cross-classified pseudo-strata for survey-aware bootstrap

Changed

• Estimator-specific guidance for parallel trends tests and placebo checks (no shared templates)
• SDiD and TROP split into separate decision tree branches in practitioner workflow

Fixed

• Fix replicate weight df calculation using pivoted QR rank with R-compatible tolerance
• Fix replicate IF variance score scaling for EfficientDiD, TripleDiff, ContinuousDiD
• Fix panel-to-unit replicate weight propagation and normalization
• Fix CS zero-mass return type and vectorized guard for survey paths
• Fix solve_logit effective-sample validation for zero-weight designs
• Fix subpopulation mask validation and EfficientDiD bootstrap guard

2.7.5 - 2026-03-23

Added

• Phase 4 survey support for ImputationDiD, TwoStageDiD, and CallawaySantAnna estimators
  • ImputationDiD/TwoStageDiD: analytical survey inference with weights, strata, and PSU (FPC not supported; bootstrap+survey deferred)
  • CallawaySantAnna: weights-only analytical IF/WIF inference matching R did::wif() (strata/PSU/FPC deferred)
  • Survey-aware aggregation for group-time, event-study, and overall ATT
• EfficientDiD enhancements: doubly robust covariates path, sieve inverse propensity (Eq 3.12), conditional Omega*
• Cluster-robust SEs for EfficientDiD with last-cohort control and Hausman pretest
• Enhanced visualizations: synth weights, staircase, dose-response, group-time heatmap, plotly backend
• Local AI review skill (/ai-review-local) with Responses API, delta-diff re-review, and cost visibility
• Add plotly optional dependency group (pip install diff-diff[plotly])

Changed

• Migrate AI local review from Chat Completions to Responses API
• Split TROP estimator into mixin modules (trop_local.py, trop_global.py) for maintainability
• Refactor visualization.py into visualization/ subpackage
• Improve review script: full-file context, content-first parsing, tiered matching, fingerprint stability

Fixed

• Fix CallawaySantAnna reg+cov control IF normalization and survey df calculation
• Fix TripleDifference TSL double-weighting and RA nuisance linearization with survey weights
• Fix ContinuousDiD bread normalization, fweight TSL scaling, and weighted-mass IF linearization
• Fix BaconDecomposition exact-weight survey unit_share and empty-cell guard
• Fix SunAbraham survey weight floor in overall ATT aggregation
• Fix plotly event study for non-numeric periods, heatmap masking, color parser

2.7.4 - 2026-03-21

Added

• Survey/sampling weights support (survey_design parameter) for DifferenceInDifferences and TwoWayFixedEffects
  • Taylor-series linearization (TSL) variance estimation with stratified multi-stage designs
  • Probability weights (pweight), frequency weights (fweight), and analytic weights (aweight)
  • Finite population correction (FPC) support
  • PSU-based clustering with lonely PSU handling
  • New diff_diff/survey.py module with SurveyDesign and compute_survey_vcov
• EfficientDiD validation tests against Chen, Sant'Anna & Xie (2025) using HRS dataset
  • HRS validation fixture with provenance documentation
  • Shared DGP helper in tests/helpers/edid_dgp.py
• Simulation-based power analysis for all registry-backed estimators (MDE, sample size, power curves); unregistered estimators supported via custom data_generator and result_extractor

Changed

• Extend power analysis to support all registry-backed estimators with result_extractor parameter
• Update power analysis tutorial with simulation-based features
• Reject absorb + fixed_effects combination (FWL violation) in both survey and non-survey paths

Fixed

• TWFE cluster-as-PSU injection for no-PSU survey designs
• Non-unique PSU labels across strata with nest=False
• FPC validation moved to compute_survey_vcov for effective PSU structure
• Survey HC1 meat formula and weighted rank-deficiency handling
• Zero-SE inference, full-census FPC, fweight contract corrections
• Bootstrap+survey fallback in MultiPeriodDiD
• DDD _snap_n floor mismatch and n_per_cell suppression scope

2.7.3 - 2026-03-19

Added

• Add aarch64 Linux wheel builds to publish workflow

Changed

• Improve documentation information architecture
• Fix silent interpreter skip and consolidate Linux jobs in publish workflow

2.7.2 - 2026-03-18

Added

• SEO infrastructure: meta tags, sitemap, llms.txt/llms-full.txt for AI discoverability

Changed

• Rename TROP method="twostep" to method="local"; "twostep" deprecated, removal in v3.0
• Rename internal TROP _joint_* methods to _global_* for consistency

Fixed

• Fix TROPResults schema: report unit counts not observation counts
• Fix llms-full.txt accuracy and dynamic canonical URLs

2.7.1 - 2026-03-15

Changed

• Replace BFGS logit with IRLS for propensity score estimation in CallawaySantAnna
• Reject pscore_trim=0.0 to prevent infinite IPW weights
• Honor rank_deficient_action="error" in propensity score paths
• Validate pscore_trim at fit() to guard against set_params bypass
• Mark slow tests (@pytest.mark.slow) and exclude by default for faster local iteration
• Use per-class slow markers in test_trop.py for faster pure Python CI

Fixed

• Vectorize Sun-Abraham bootstrap resampling loop for improved performance

2.7.0 - 2026-03-15

Added

• EfficientDiD estimator (EfficientDiD) implementing Chen, Sant'Anna & Xie (2025) efficient DiD
• CallawaySantAnna event study SEs (WIF-based) and simultaneous confidence bands (sup-t)
• R comparison tests for event-study SEs and cband critical values
• Non-finite outcome validation in EfficientDiD.fit()
• CallawaySantAnna speed benchmarks with baseline results
• Estimator alias documentation in README, quickstart, and API docs

Changed

• BREAKING: TROP nuclear norm solver step size fix — The proximal gradient threshold for the L matrix (both method="global" and method="twostep" with finite lambda_nn) was over-shrinking singular values by a factor of 2. The soft-thresholding threshold was λ_nn/max(δ) when the correct value is λ_nn/(2·max(δ)), derived from the Lipschitz constant L_f=2·max(δ) of the quadratic gradient. This fix produces higher-rank L matrices and closer agreement with exact convex optimization solutions. Users with finite lambda_nn will observe different ATT estimates. Added FISTA/Nesterov acceleration to the twostep inner solver for faster L convergence.
• Add (1-W) weight masking to TROP global method, rename joint→global
• Optimize CallawaySantAnna covariate path with Cholesky and pscore caching
• Update Codex AI review model from gpt-5.2-codex to gpt-5.4

Fixed

• Fix CallawaySantAnna event study SEs (missing WIF) and simultaneous confidence bands
• Fix analytical and bootstrap WIF pg scaling to use global N
• Fix TROP nuclear norm solver threshold scaling for non-uniform weights
• Fix stale coefficients in TROP global low-rank solver and NaN bootstrap poisoning
• Fix NaN-cell preservation in CallawaySantAnna balance_e aggregation
• Fix not-yet-treated cache keys and dropped-cell warning
• Fix rank-deficiency handling with Cholesky rank checks and reduced-column solve
• Fix Rust convergence criterion, n_valid_treated consistency, and NaN bootstrap SE

2.6.1 - 2026-03-08

Added

• Short aliases for all estimators (e.g., DiD, TWFE, EventStudy, CS, SDiD)

Changed

• Update roadmap for v2.6.0: reflect completed work and refresh priorities
• Add ContinuousDiD to ReadTheDocs API reference and choosing guide
• Add SPT identification caveat and data requirements per review
• Add time-invariant dose requirement to data requirements

Fixed

• Fix alias docs wording: clarify TROP has no alias
• Fix ContinuousDiD SE method: influence function, not delta method
• Fix methodology doc: influence functions, not delta method for ContinuousDiD SEs
• Fix dollar sign escaping in continuous DiD tutorial
• Fix continuous DiD tutorial formatting: escape dollar signs and split chart cell
• Fix methodology claims and slide numbering per PR review

2.6.0 - 2026-02-22

Added

• Continuous DiD estimator (ContinuousDiD) implementing Callaway, Goodman-Bacon & Sant'Anna (2024) for continuous treatment dose-response analysis
  • ContinuousDiDResults with dose-response curves and event-study effects
  • DoseResponseCurve with bootstrap p-values
  • Analytical and bootstrap event-study SEs
  • P(D=0) warning for low-probability control groups
• Stacked DiD tutorial (Tutorial 13) with Q-weight computation walkthrough

Changed

• Clarify aggregate Q-weight computation for unbalanced panels in Stacked DiD tutorial
• Replace SunAbraham manual bootstrap stats with NaN-gated utility

Fixed

• Fix not-yet-treated control mask to respect anticipation parameter in ContinuousDiD
• Guard non-finite original_effect in compute_effect_bootstrap_stats
• Fix bootstrap NaN propagation for rank-deficient cells
• Fix NaN propagation in rank-deficient spline predictions
• Guard bootstrap NaN propagation: SE/CI/p-value all NaN when SE invalid
• Fix bootstrap ACRT^{glob} centering bug
• Fix bootstrap percentile inference and analytical event-study SE scaling
• Fix control group bug and dose validation in ContinuousDiD

2.5.0 - 2026-02-19

Added

• Stacked DiD estimator (StackedDiD) implementing Wing, Freedman & Hollingsworth (2024) with corrective Q-weights for compositional balance across event times
• Sub-experiment construction per adoption cohort with clean (never-yet-treated) controls
• IC1/IC2 trimming for compositional balance across event times
• Q-weights for aggregate, population, or sample share estimands (Table 1)
• WLS event study regression via sqrt(w) transformation
• stacked_did() convenience function
• R benchmark scripts for Stacked DiD validation (benchmarks/R/benchmark_stacked_did.R)
• Comprehensive test suite for Stacked DiD (tests/test_stacked_did.py)

Fixed

• NaN inference handling in pure Python mode for edge cases

2.4.3 - 2026-02-19

Changed

• Rewrite TripleDifference estimator to match R's triplediff::ddd() — all 3 estimation methods (DR, IPW, RA) now use three-DiD decomposition with influence function SE, achieving <0.001% relative difference from R across all 24 comparisons (4 DGPs × 3 methods × 2 covariate settings)
• Validate cluster column in TripleDifference for proper cluster-robust SEs
• Handle non-finite influence function propagation in TripleDifference edge cases
• Propensity score fallback uses Hessian-based SE when score optimization fails
• Improved R-squared consistency across estimation methods

Fixed

• Fix low cell count warning and overlap detection in TripleDifference IPW
• Fix cluster SE computation to use functional (groupby) approach instead of loop
• Fix rank deficiency handling in TripleDifference regression adjustment

Added

• 91 methodology verification tests for TripleDifference (tests/test_methodology_triple_diff.py)
• R benchmark scripts for triple difference validation (benchmarks/R/benchmark_triplediff.R)
• Update METHODOLOGY_REVIEW.md to reflect completed TripleDifference review

2.4.2 - 2026-02-18

Added

• Conditional BLAS linking for Rust backend — Apple Accelerate on macOS, OpenBLAS on Linux. Pre-built wheels now use platform-optimized BLAS for matrix-vector and matrix-matrix operations across all Rust-accelerated code paths (weights, OLS, TROP). Windows continues using pure Rust (no external dependencies). Improves Rust backend performance at larger scales.
• rust_backend_info() diagnostic function in diff_diff._backend — reports compile-time BLAS feature status (blas, accelerate, openblas)

Fixed

• Rust SDID backend performance regression at scale — Frank-Wolfe solver was 3-10x slower than pure Python at 1k+ scale
  • Gram-accelerated FW loop for time weights: precomputes A^T@A, reducing per-iteration cost from O(N×T0) to O(T0) (~100x speedup per iteration at 5k scale)
  • Allocation-free FW loop for unit weights: 1 GEMV per iteration (was 3), zero heap allocations (was ~8)
  • Dispatch based on problem dimensions: Gram path when T0 < N, standard path when T0 >= N
  • Rust backend now faster than pure Python at all scales

2.4.1 - 2026-02-17

Added

• Tutorial notebook for Two-Stage DiD (Gardner 2022) (docs/tutorials/12_two_stage_did.ipynb)

Changed

• Module splits for large files: ImputationDiD, TwoStageDiD, and TROP each split into separate results and bootstrap submodules
• Migrated remaining inline inference computations to safe_inference() utility
• Replaced @ operator with np.dot() at observation-dimension sites to avoid Apple M4 BLAS warnings
• Updated TODO.md and ROADMAP.md for accuracy post-v2.4.0

Fixed

• Matplotlib import guards added to tutorials 11 and 12
• Various bug fixes from code quality cleanup (diagnostics, estimators, linalg, staggered, sun_abraham, synthetic_did, triple_diff)

2.4.0 - 2026-02-16

Added

• Gardner (2022) Two-Stage DiD estimator (TwoStageDiD)
  • Two-stage estimator: (1) estimate unit+time FE on untreated obs, (2) regress residualized outcomes on treatment indicators
  • TwoStageDiDResults with overall ATT, event study, group effects, per-observation treatment effects
  • TwoStageBootstrapResults for multiplier bootstrap inference on GMM influence function
  • two_stage_did() convenience function for quick estimation
  • Point estimates identical to ImputationDiD; different variance estimator (GMM sandwich vs. conservative)
  • No finite-sample adjustments (raw asymptotic sandwich, matching R did2s)
• Proposition 5 detection for unidentified long-run horizons without never-treated units

Changed

• Workflow improvements to reduce PR review rounds

Fixed

• Zero-observation horizons/cohorts producing se=0 instead of NaN in TwoStageDiD
• Edge case fixes for TwoStageDiD (PR review feedback)
• Grep PCRE patterns updated to use POSIX character classes

2.3.2 - 2026-02-16

Added

• Python 3.13 support with upper version cap (>=3.9,<3.14)

Changed

• Sun-Abraham methodology review (PR #153)
  • IW aggregation weights now use event-time observation counts (not group sizes)
  • Normalize np.inf never-treated encoding before treatment group detection
  • Add R benchmark scripts and methodology-aligned tests
• Use rank_deficient_action and np.errstate instead of broad RuntimeWarning filter in SDID tutorial

Fixed

• Sun-Abraham bootstrap NaN propagation for non-finite ATT estimates
• Sun-Abraham df_adjustment off-by-one in analytical SE computation
• CI pandas compatibility for SunAbraham bootstrap inference
• SyntheticDiD tutorial: eliminate pre-treatment fit warnings

2.3.1 - 2026-02-15

Fixed

• Fix docs/PyPI version mismatch (issue #146) — RTD now builds versioned docs from source
• Fix RTD docs build failure caused by Rust/maturin compilation timeout on ReadTheDocs

Changed

• Remove Rust outer-loop variance estimation for SyntheticDiD (placebo and bootstrap)
  • Fixes SE mismatch between pure Python and Rust backends (different RNG sequences)
  • Fixes Rust performance regression at 1k+ scale (memory bandwidth saturation from rayon parallelism)
  • Inner Frank-Wolfe weight computation still uses Rust when available

Documentation

• Re-run SyntheticDiD benchmarks against R after Frank-Wolfe methodology rewrite
• Updated docs/benchmarks.rst SDID validation results, performance tables, and known differences
• ATT now matches R to < 1e-10 (previously 0.3% diff) since both use Frank-Wolfe optimizer

2.3.0 - 2026-02-09

Added

• Borusyak-Jaravel-Spiess (2024) Imputation DiD estimator (ImputationDiD)
  • Efficient imputation estimator for staggered DiD designs
  • OLS on untreated observations for unit+time FE, impute counterfactual Y(0), aggregate
  • Conservative variance (Theorem 3) with aux_partition parameter for SE tightness
  • Pre-trend test (Equation 9) via results.pretrend_test()
  • Percentile bootstrap inference
  • Influence-function bootstrap with sparse variance and weight/covariate fixes
  • Absorbing-treatment validation for non-constant first_treat
  • Empty event-study warning for unidentified long-run horizons
• /paper-review skill for academic paper methodology extraction
• /read-feedback-revise skill for addressing PR review comments
• --pr flag for /review-plan skill to review plans posted as PR comments
• --updated flag for /review-plan skill for re-reviewing revised plans
• MultiPeriodDiD vs R (fixest) benchmark for cross-language validation

Changed

• Shortened test suite runtime with parallel execution and reduced iterations

Fixed

• TWFE within-transformation bug identified during methodology review
• TWFE: added non-{0,1} binary time warning, ATT invariance tests, and R fixture caching
• TWFE: single-pass demeaning, HC1 test fix, fixest coeftable comparison
• MultiPeriodDiD: added unit FE and NaN guard for R comparison benchmark
• Removed tracked PDF from repo and gitignored papers directory

2.2.1 - 2026-02-07

Changed

• MultiPeriodDiD: Full event-study specification (BREAKING)
  • Treatment × period interactions now created for ALL periods (pre and post), not just post-treatment
  • Pre-period coefficients available for parallel trends assessment
  • Default reference period changed from first to last pre-period (e=-1 convention) with FutureWarning for one release cycle
  • period_effects dict now contains both pre and post period effects
  • to_dataframe() includes is_post column
  • summary() output now shows pre-period effects section
  • t_stat uses np.isfinite(se) and se > 0 guard (consistent with other estimators)

Added

• Time-varying treatment warning when unit is provided and treatment varies within units (guides users toward ever-treated indicator D_i)
• unit parameter to MultiPeriodDiD.fit() for staggered adoption detection
• reference_period and interaction_indices attributes on MultiPeriodDiDResults
• pre_period_effects and post_period_effects convenience properties on results
• Pre-period section in summary() output with reference period indicator
• ValueError when reference_period is set to a post-treatment period
• Staggered adoption warning when treatment timing varies across units (with unit param)
• Informative KeyError when accessing reference period via get_effect()

Removed

• TROP variance_method parameter — Jackknife variance estimation removed. Bootstrap (the only method specified in Athey et al. 2025) is now always used. The variance_method field has also been removed from TROPResults.
• TROP max_loocv_samples parameter — Control observation subsampling removed from LOOCV tuning parameter selection. Equation 5 of Athey et al. (2025) explicitly sums over ALL control observations where D=0; the previous subsampling (default 100) was not specified in the paper. LOOCV now uses all control observations, making tuning fully deterministic. Inner LOOCV loops in the Rust backend are parallelized to compensate for the increased observation count.

Fixed

• HonestDiD: filter non-finite period effects from MultiPeriodDiD results (prevents NaN propagation into sensitivity bounds; raises ValueError when no finite pre- or post-period effects remain)
• HonestDiD VCV extraction: now uses interaction sub-VCV instead of full regression VCV (via interaction_indices period → column index mapping)
• MultiPeriodDiD: avg_se guard now checks np.isfinite() (matches per-period pattern; prevents avg_t_stat=0 / avg_p_value=1 when variance is infinite)
• HonestDiD: extraction now uses explicit pre-then-post ordering instead of sorted period labels (prevents misclassification when period labels don't sort chronologically)
• Backend-aware test parameter scaling for pure Python CI performance
• Lower TROP stratified bootstrap threshold floor from 11 to 5 for pure Python CI

2.2.0 - 2026-01-27

Added

• Windows wheel builds using pure-Rust faer library for linear algebra (PR #115)
  • Eliminates external BLAS/LAPACK dependencies (no OpenBLAS or Intel MKL required)
  • Enables cross-platform wheel builds for Linux, macOS, and Windows
  • Simplifies installation on all platforms

Changed

• Rust backend migrated from nalgebra/ndarray to faer (PR #115)
  • OLS solver now uses faer's SVD implementation
  • Robust variance estimation uses faer's matrix operations
  • TROP distance calculations use faer primitives
  • Maintains numerical parity with existing NumPy backend

Fixed

• Rust backend numerical stability improvements (PR #115)
  • Improved singular matrix detection with condition number checks
  • NaN propagation in variance-covariance estimation
  • Fallback to Python backend on numerical instability with warning
  • Underdetermined SVD handling (n < k case)
• macOS CI compatibility for Python 3.14 with PYO3_USE_ABI3_FORWARD_COMPATIBILITY

2.1.9 - 2026-01-26

Added

• Unified LOOCV for TROP joint method with Rust acceleration (PR #113)
  • Leave-one-out cross-validation for rank and regularization parameter selection
  • Rust backend provides significant speedup for LOOCV grid search

Fixed

• TROP joint method Rust/Python parity (PR #113)
  • Fixed valid_count bug in LOOCV computation
  • Proper NaN exclusion for units with no valid pre-period data
  • Zero weight assignment for units missing pre-period data
  • Jackknife variance estimation fixes
  • Staggered adoption validation and simultaneous adoption enforcement
  • Treated-pre NaN handling improvements
  • LOOCV subsampling fix for Python-only path

2.1.8 - 2026-01-25

Added

• /push-pr-update skill for committing and pushing PR revisions
  • Commits local changes to current branch and pushes to remote
  • Triggers AI code review automatically
  • Robust handling for fork repos, unpushed commits, and upstream tracking

Fixed

• TROP estimator methodology alignment (PR #110)
  • Aligned with paper methodology (Equation 5, D matrix semantics)
  • NaN propagation and LOOCV warnings improvements
  • Rust backend test alignment with new loocv_grid_search return signature
  • LOOCV cycling, D matrix validation fixes
  • Final estimation infinity handling and edge case fixes
  • Absorbing-state gap detection and n_post_periods fix

Changed

• /submit-pr skill improvements (PR #111)
  • Case-insensitive secret scanning with POSIX ERE regex
  • Verify origin ref exists before push
  • Dynamic default branch detection with fallback
  • Robust handling for unpushed commits, fork repos
  • Files count display in PR summary

2.1.7 - 2026-01-25

Fixed

• plot_event_study reference period normalization behavior
  • Effects are now only normalized when reference_period is explicitly provided
  • Auto-inferred reference periods only apply hollow marker styling (no normalization)
  • Reference period SE is set to NaN during normalization (constraint, not estimate)
  • Updated docstring to clarify explicit vs auto-inferred behavior

Changed

• Refactored visualization tests to reuse cs_results fixture for better performance

2.1.6 - 2026-01-24

Added

• Methodology verification tests for DifferenceInDifferences estimator
  • Comprehensive test suite validating all REGISTRY.md requirements
  • Tests for formula interface, coefficient extraction, rank deficiency handling
  • Singleton cluster variance estimation behavioral tests

Changed

• REGISTRY.md documentation improvements
  • Clarified singleton cluster formula notation (u_i² X_i X_i' instead of ambiguous residual² × X'X)
  • Verified DifferenceInDifferences behavior against documented requirements

2.1.5 - 2026-01-22

Added

• METHODOLOGY_REVIEW.md tracking document for methodology review progress
  • Review status summary table for all 12 estimators
  • Detailed notes template for each estimator by category
  • Review process guidelines with checklist and priority ordering
• base_period parameter for CallawaySantAnna pre-treatment effect computation
  • "varying" (default): Pre-treatment uses t-1 as base (consecutive comparisons)
  • "universal": All comparisons use g-anticipation-1 as base
  • Matches R did::att_gt() base_period parameter
• Pre-merge-check skill (/pre-merge-check) for automated PR validation
  • Pattern checks for NaN handling consistency
  • Context-specific checklist generation

Changed

• Tutorial 02 improvements: Added pre-trends section, clarified base_period interaction with anticipation

Fixed

• Not-yet-treated control group now properly excludes cohort g when computing ATT(g,t)
• Aggregation t_stat uses NaN (not 0.0) when SE is non-finite or zero
• Bootstrap inference for pre-treatment effects with base_period="varying"
• NaN propagation for empty post-treatment effects in CallawaySantAnna
• Grep word boundary pattern in pre-merge-check skill

2.1.4 - 2026-01-20

Added

• Development checklists and workflow improvements in CLAUDE.md
  • Estimator inheritance map showing class hierarchy for get_params/set_params
  • Test writing guidelines for fallback paths, parameters, and warnings
  • Checklists for adding parameters and warning/error handling
• R-style rank deficiency handling across all estimators
  • rank_deficient_action parameter: "warn" (default), "error", or "silent"
  • Dropped columns have NaN coefficients (like R's lm())
  • VCoV matrix has NaN for rows/cols of dropped coefficients
  • Propagated to all estimators: DifferenceInDifferences, MultiPeriodDiD, TwoWayFixedEffects, CallawaySantAnna, SunAbraham, TripleDifference, TROP, SyntheticDiD

Fixed

• get_params() now includes rank_deficient_action parameter (fixes sklearn cloning)
• NaN vcov fallback in Rust backend for rank-deficient matrices
• MultiPeriodDiD vcov/df computation for rank-deficient designs
• Average ATT inference for rank-deficient designs

Changed

• Rank tolerance aligned with R's lm() default for consistent behavior

2.1.3 - 2026-01-19

Fixed

• TROP estimator paper conformance issues (Athey et al. 2025)
  • Control set now includes pre-treatment observations of eventually-treated units (Issue A)
  • Unit distance computation excludes target period per Equation 3 (Issue B)
  • Nuclear norm update uses weighted proximal gradient instead of unweighted soft-thresholding (Issue C)
  • Bootstrap sampling now stratifies by treatment status per Algorithm 3 (Issue D)
• TROP Rust backend alignment with paper specification
  • Weight normalization to sum to 1 (probability weights)
  • Weighted proximal gradient for L update with step size η ≤ 1/max(W)

Changed

• Cleaned up unused parameters from TROP Rust API
  • Removed control_unit_idx and unit_dist_matrix from public functions
  • Per-observation distances now computed dynamically (more accurate, slightly slower)

2.1.2 - 2026-01-19

Added

• Consolidated DGP functions in prep.py for all supported DiD designs
  • generate_did_data() - Basic 2x2 DiD data generation
  • generate_staggered_data() - Staggered adoption data for Callaway-Sant'Anna/Sun-Abraham
  • generate_factor_data() - Factor model data for TROP/SyntheticDiD
  • generate_ddd_data() - Triple Difference (DDD) design data
  • generate_panel_data() - Panel data with optional parallel trends violations
  • generate_event_study_data() - Event study data with simultaneous treatment

Changed

• Clean up development tracking files for v2.1.1 release
  • Removed completed items from TODO.md (now tracked in CHANGELOG)
  • Updated ROADMAP.md version numbers and removed shipped TROP section
  • Updated prep.py line count in Large Module Files table (1338 → 1993)

2.1.1 - 2026-01-19

Added

• Rust backend acceleration for TROP estimator delivering 5-20x overall speedup
  • compute_unit_distance_matrix - Parallel pairwise RMSE computation for donor matching
  • loocv_grid_search - Parallel leave-one-out cross-validation across 180 parameter combinations
  • bootstrap_trop_variance - Parallel bootstrap variance estimation
  • Automatic fallback to Python when Rust backend unavailable
  • Logging for Rust fallback events to aid debugging
• /bump-version skill for release management
  • Updates version in __init__.py, pyproject.toml, and rust/Cargo.toml
  • Generates CHANGELOG entries from git commits
  • Adds comparison links automatically
• /review-pr skill for code review workflow

Changed

• TROP estimator performance optimizations (Python backend)
  • Vectorized distance matrix computation using NumPy broadcasting
  • Extracted tuning constants to module-level for clarity
  • Added TROPTuningParams TypedDict for parameter documentation

Fixed

• Tutorial notebook validation errors in 10_trop.ipynb
• Pre-existing RuntimeWarnings in CallawaySantAnna bootstrap (documented)
• TROP pre_periods parameter handling for edge cases

2.1.0 - 2026-01-17

Added

• Triply Robust Panel (TROP) estimator implementing Athey, Imbens, Qu & Viviano (2025)
  • TROP class combining three robustness components:
    • Factor model adjustment via SVD (removes unobserved confounders with factor structure)
    • Synthetic control style unit weights
    • SDID style time weights
  • TROPResults dataclass with ATT, factors, loadings, unit/time weights
  • trop() convenience function for quick estimation
  • Automatic rank selection methods: cross-validation ('cv'), information criterion ('ic'), elbow detection ('elbow')
  • Bootstrap and placebo-based variance estimation
  • Full integration with existing infrastructure (exports in __init__.py, sklearn-compatible API)
  • Tutorial notebook: docs/tutorials/10_trop.ipynb
  • Comprehensive test suite: tests/test_trop.py

Reference: Athey, S., Imbens, G. W., Qu, Z., & Viviano, D. (2025). "Triply Robust Panel Estimators." Working Paper. arXiv:2508.21536

2.0.3 - 2026-01-17

Changed

• Rust backend performance optimizations delivering up to 32x speedup for bootstrap operations
  • Bootstrap weight generation now 16x faster on average (up to 32x for Webb distribution)
  • Direct Array2 allocation eliminates intermediate Vec<Vec<f64>> (~50% memory reduction)
  • Rayon chunk size tuning (min_len=64) reduces parallel scheduling overhead
  • Webb distribution uses lookup table instead of 6-way if-else chain

Added

• LinearRegression helper class in linalg.py for code deduplication
  • High-level OLS wrapper with unified coefficient extraction and inference
  • Used by DifferenceInDifferences, TwoWayFixedEffects, SunAbraham, TripleDifference
  • Provides InferenceResult dataclass for coefficient-level statistics
• Cholesky factorization for symmetric positive-definite matrix inversion in Rust backend
  • ~2x faster than LU decomposition for well-conditioned matrices
  • Automatic fallback to LU for near-singular or indefinite matrices
• Vectorized variance computation in Rust backend
  • HC1 meat computation: X' @ (X * e²) via BLAS instead of O(n×k²) loop
  • Score computation: broadcast multiplication instead of O(n×k) loop
• Static BLAS linking options in rust/Cargo.toml
  • openblas-static and intel-mkl-static features for standalone distribution
  • Eliminates runtime BLAS dependency at cost of larger binary size

2.0.2 - 2026-01-15

Fixed

• CallawaySantAnna SE computation now exactly matches R's did package
  • Fixed weight influence function (wif) formula for "simple" aggregation
  • Corrected pg computation: uses n_g / n_all (matching R) instead of n_g / total_treated
  • Fixed wif iteration: iterates over keepers (post-treatment pairs) with individual ATT(g,t) values
  • SE difference reduced from ~2.5% to <0.01% vs R's did package (essentially exact match)
  • Point estimates unchanged; all existing tests pass

2.0.1 - 2026-01-13

Added

• Shared within-transformation utilities in utils.py
  • demean_by_group() - One-way fixed effects demeaning
  • within_transform() - Two-way (unit + time) FE transformation
  • Reduces code duplication across estimators.py, twfe.py, sun_abraham.py, bacon.py

Fixed

• DataFrame fragmentation warning - Build columns in batch instead of iteratively

Changed

• Reverted untested Rust backend optimizations (Cholesky factorization, reduced allocations) - these will be re-added when proper testing infrastructure is available

2.0.0 - 2026-01-12

Added

• Optional Rust backend for accelerated computation
  • 4-8x speedup for SyntheticDiD and bootstrap operations
  • Parallel bootstrap weight generation (Rademacher, Mammen, Webb)
  • Accelerated OLS solver using OpenBLAS/MKL
  • Cluster-robust variance estimation
  • Synthetic control weight optimization with simplex projection
  • Pre-built wheels for Linux x86_64 and macOS ARM64
  • Pure Python fallback for all other platforms
• diff_diff/_backend.py - Backend detection and configuration module
  • HAS_RUST_BACKEND flag exported in main package
  • DIFF_DIFF_BACKEND environment variable for backend control:
    • 'auto' (default) - Use Rust if available, fall back to Python
    • 'python' - Force pure Python mode
    • 'rust' - Force Rust mode (fails if unavailable)
• Rust source code in rust/ directory
  • rust/src/lib.rs - PyO3 module definition
  • rust/src/bootstrap.rs - Parallel bootstrap weight generation
  • rust/src/linalg.rs - OLS solver and robust variance estimation
  • rust/src/weights.rs - Synthetic control weights and simplex projection
• Rust backend test suite - tests/test_rust_backend.py for equivalence testing

Changed

• Package version bumped from 1.4.0 to 2.0.0 (major version for new backend)
• CI/CD updated to build Rust extensions with maturin
• ReadTheDocs now installs from PyPI (pre-built wheels with Rust backend)

1.4.0 - 2026-01-11

Added

• Unified linear algebra backend (diff_diff/linalg.py)
  • solve_ols() - Optimized OLS solver using scipy's gelsy LAPACK driver
  • compute_robust_vcov() - Vectorized (clustered) robust variance-covariance
  • Single optimization point for all estimators; prepares for future Rust backend
  • New tests/test_linalg.py with comprehensive tests

Changed

• Major performance improvements - All estimators now significantly faster
  • BasicDiD/TWFE @ 10K: 0.835s → 0.011s (76x faster, now 4.2x faster than R)
  • CallawaySantAnna @ 10K: 2.234s → 0.109s (20x faster, now 7.2x faster than R)
  • All results numerically identical to previous versions
• CallawaySantAnna optimizations (staggered.py)
  • Pre-computed wide-format outcome matrix and cohort masks
  • Vectorized ATT(g,t) computation using numpy operations (23x faster)
  • Batch bootstrap weight generation
  • Vectorized multiplier bootstrap using matrix operations (26x faster)
• TWFE optimization (twfe.py)
  • Cached groupby indexes for within-transformation
• All estimators migrated to unified linalg.py backend
  • estimators.py, twfe.py, staggered.py, triple_diff.py, synthetic_did.py, sun_abraham.py, utils.py

Behavioral Changes

• Rank-deficient design matrices: The new gelsy LAPACK driver handles rank-deficient matrices gracefully (returning a least-norm solution) rather than raising an explicit error. Previously, DifferenceInDifferences would raise ValueError("Design matrix is rank-deficient"). Users relying on this error for collinearity detection should validate their design matrices separately. Results remain numerically correct for well-specified models.

1.3.1 - 2026-01-10

Added

• SyntheticDiD placebo-based variance estimation matching R's synthdid package methodology
  • New variance_method parameter with options "bootstrap" (default) and "placebo"
  • Placebo method implements Algorithm 4 from Arkhangelsky et al. (2021):
    1. Randomly permutes control unit indices
    2. Designates N₁ controls as pseudo-treated (matching actual treated count)
    3. Renormalizes original unit weights for remaining pseudo-controls
    4. Computes SDID estimate with renormalized weights
    5. Repeats for n_bootstrap replications
    6. SE = sqrt((r-1)/r) × sd(estimates)
  • Provides methodological parity with R's synthdid::vcov(method = "placebo")
  • n_bootstrap parameter now used for both bootstrap and placebo replications
  • SyntheticDiDResults now tracks variance_method and n_bootstrap attributes
  • Results summary displays variance method and replications count

Reference: Arkhangelsky, D., Athey, S., Hirshberg, D. A., Imbens, G. W., & Wager, S. (2021). Synthetic Difference-in-Differences. American Economic Review, 111(12), 4088-4118.

1.3.0 - 2026-01-09

Added

• Triple Difference (DDD) estimator implementing Ortiz-Villavicencio & Sant'Anna (2025)
  • TripleDifference class for DDD designs where treatment requires two criteria (group AND partition)
  • TripleDifferenceResults dataclass with ATT, SEs, cell means, and diagnostics
  • triple_difference() convenience function for quick estimation
  • Three estimation methods: regression adjustment (reg), inverse probability weighting (ipw), and doubly robust (dr)
  • Proper covariate handling (unlike naive DDD implementations that difference two DiDs)
  • Propensity score trimming for IPW/DR methods
  • Cluster-robust standard errors support
  • Tutorial notebook: docs/tutorials/08_triple_diff.ipynb

Reference: Ortiz-Villavicencio, M., & Sant'Anna, P. H. C. (2025). "Better Understanding Triple Differences Estimators." Working Paper. arXiv:2505.09942

1.2.1 - 2026-01-08

Added

• Expanded test coverage for edge cases:
  • Wild bootstrap with very few clusters (< 5), including 2-3 cluster scenarios
  • Unbalanced panels with missing periods across units
  • Single treated unit scenarios for DiD and Synthetic DiD
  • Perfect collinearity detection (validates clear error messages)
  • CallawaySantAnna with single treatment cohort
  • SyntheticDiD with insufficient pre-treatment periods

Changed

• Refactored CallawaySantAnna bootstrap: Extracted _compute_effect_bootstrap_stats() helper method for cleaner code and reduced duplication in bootstrap statistics computation.

1.2.0 - 2026-01-07

Added

• Pre-Trends Power Analysis (Roth 2022) for assessing informativeness of pre-trends tests
  • PreTrendsPower class for computing power and minimum detectable violation (MDV)
  • PreTrendsPowerResults dataclass with power, MDV, and test statistics
  • PreTrendsPowerCurve for power curves across violation magnitudes
  • compute_pretrends_power() and compute_mdv() convenience functions
  • Multiple violation types: linear, constant, last_period, custom
  • Integration with Honest DiD via sensitivity_to_honest_did() method
  • plot_pretrends_power() visualization for power curves
  • Tutorial notebook: docs/tutorials/07_pretrends_power.ipynb
  • Full API documentation: docs/api/pretrends.rst

Reference: Roth, J. (2022). "Pretest with Caution: Event-Study Estimates after Testing for Parallel Trends." American Economic Review: Insights, 4(3), 305-322.

Fixed

• Reference period handling in pre-trends analysis: Fixed bug where reference period was incorrectly assigned avg_se instead of being excluded from power calculations. Now properly excludes the omitted reference period from the joint Wald test.

1.1.1 - 2026-01-06

Fixed

• SyntheticDiD bootstrap error handling: Bootstrap now raises clear ValueError when all iterations fail, instead of silently returning SE=0.0. Added warnings for edge cases (single successful iteration, high failure rate).

• Diagnostics module error handling: Improved error messages in permutation_test() and leave_one_out_test() with actionable guidance. Added warnings when significant iterations fail. Enhanced run_all_placebo_tests() to return structured error info including error type.

Changed

• Code deduplication: Extracted wild bootstrap inference logic to shared _run_wild_bootstrap_inference() method in DifferenceInDifferences base class, used by both DifferenceInDifferences and TwoWayFixedEffects.

• Type hints: Added missing type hints to nested functions:

  • compute_trend() in utils.py
  • neg_log_likelihood() and gradient() in staggered.py
  • format_label() in prep.py

1.1.0 - 2026-01-05

Added

• Sun-Abraham (2021) interaction-weighted estimator for staggered DiD
  • SunAbraham class implementing saturated regression approach
  • SunAbrahamResults with event study effects, cohort weights, and overall ATT
  • SABootstrapResults for bootstrap inference (SEs, CIs, p-values)
  • Support for never_treated and not_yet_treated control groups
  • Analytical and cluster-robust standard errors
  • Multiplier bootstrap with Rademacher, Mammen, or Webb weights
  • Integration with plot_event_study() visualization
  • Useful robustness check alongside Callaway-Sant'Anna

Reference: Sun, L., & Abraham, S. (2021). "Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects." Journal of Econometrics, 225(2), 175-199.

1.0.2 - 2026-01-04

Changed

• Refactored estimators.py to reduce module size
  • Moved TwoWayFixedEffects to diff_diff/twfe.py
  • Moved SyntheticDiD to diff_diff/synthetic_did.py
  • Backward compatible re-exports maintained in estimators.py

Fixed

• Fixed ReadTheDocs version display by importing from package __version__

1.0.1 - 2026-01-04

Fixed

• Tech debt cleanup (Tier 1 + Tier 2)
  • Improved code organization and documentation
  • Fixed minor issues identified in tech debt review

1.0.0 - 2026-01-04

Added

• Goodman-Bacon decomposition for TWFE diagnostics
  • BaconDecomposition class for decomposing TWFE into weighted 2x2 comparisons
  • Comparison2x2 dataclass for individual comparisons (treated_vs_never, earlier_vs_later, later_vs_earlier)
  • BaconDecompositionResults with weights and estimates by comparison type
  • bacon_decompose() convenience function
  • plot_bacon() visualization for decomposition results
  • Integration via TwoWayFixedEffects.decompose() method
• Power analysis for study design
  • PowerAnalysis class for analytical power calculations
  • PowerResults and SimulationPowerResults dataclasses
  • compute_mde(), compute_power(), compute_sample_size() convenience functions
  • simulate_power() for Monte Carlo simulation-based power analysis
  • plot_power_curve() visualization for power analysis
  • Tutorial notebook: docs/tutorials/06_power_analysis.ipynb
• Callaway-Sant'Anna multiplier bootstrap for inference
  • CSBootstrapResults with standard errors, confidence intervals, p-values
  • Rademacher, Mammen, and Webb weight distributions
  • Bootstrap inference for all aggregation methods
• Troubleshooting guide in documentation
• Standard error computation guide explaining SE differences across estimators

Changed

• Updated package status to Production/Stable (was Alpha)
• SyntheticDiD bootstrap now warns when >5% of iterations fail

Fixed

• Silent bootstrap failures in SyntheticDiD now produce warnings

0.6.0

Added

• CallawaySantAnna covariate adjustment for conditional parallel trends
  • Outcome regression (estimation_method='reg')
  • Inverse probability weighting (estimation_method='ipw')
  • Doubly robust estimation (estimation_method='dr')
  • Pass covariates via covariates parameter in fit()
• Honest DiD sensitivity analysis (Rambachan & Roth 2023)
  • HonestDiD class for computing bounds under parallel trends violations
  • Relative magnitudes restriction (DeltaRM) - bounds post-treatment violations by pre-treatment
  • Smoothness restriction (DeltaSD) - bounds second differences of trend violations
  • Combined restrictions (DeltaSDRM)
  • FLCI and C-LF confidence interval methods
  • Breakdown value computation via breakdown_value()
  • Sensitivity analysis over M grid via sensitivity_analysis()
  • HonestDiDResults and SensitivityResults dataclasses
  • compute_honest_did() convenience function
  • plot_sensitivity() for sensitivity analysis visualization
  • plot_honest_event_study() for event study with honest CIs
  • Tutorial notebook: docs/tutorials/05_honest_did.ipynb
• API documentation site with Sphinx
  • Full API reference auto-generated from docstrings
  • "Which estimator should I use?" decision guide
  • Comparison with R packages (did, HonestDiD)
  • Getting started / quickstart guide

Changed

• Updated mypy configuration for better numpy type compatibility
• Modernized ruff configuration to use [tool.ruff.lint] section

Fixed

• Fixed 21 ruff linting issues (import ordering, unused variables, ambiguous names)
• Fixed 94 mypy type checking issues (Optional types, numpy type casts, assertions)
• Added missing return statement in run_placebo_test()

0.5.0

Added

• Wild cluster bootstrap for valid inference with few clusters
  • Rademacher weights (default, good for most cases)
  • Webb's 6-point distribution (recommended for <10 clusters)
  • Mammen's two-point distribution
  • WildBootstrapResults dataclass
  • wild_bootstrap_se() utility function
  • Integration with DifferenceInDifferences and TwoWayFixedEffects via inference='wild_bootstrap'
• Placebo tests module (diff_diff.diagnostics)
  • placebo_timing_test() - fake treatment timing test
  • placebo_group_test() - fake treatment group test
  • permutation_test() - permutation-based inference
  • leave_one_out_test() - sensitivity to individual treated units
  • run_placebo_test() - unified dispatcher for all test types
  • run_all_placebo_tests() - comprehensive diagnostic suite
  • PlaceboTestResults dataclass
• Tutorial notebooks in docs/tutorials/
  • 01_basic_did.ipynb - Basic 2x2 DiD, formula interface, covariates, fixed effects, wild bootstrap
  • 02_staggered_did.ipynb - Staggered adoption with Callaway-Sant'Anna
  • 03_synthetic_did.ipynb - Synthetic DiD with unit/time weights
  • 04_parallel_trends.ipynb - Parallel trends testing and diagnostics
• Comprehensive test coverage (380+ tests)

0.4.0

Added

• Callaway-Sant'Anna estimator for staggered difference-in-differences
  • CallawaySantAnna class with group-time ATT(g,t) estimation
  • Support for never_treated and not_yet_treated control groups
  • Aggregation methods: simple, group, calendar, event_study
  • CallawaySantAnnaResults with group-time effects and aggregations
  • GroupTimeEffect dataclass for individual effects
• Event study visualization via plot_event_study()
  • Works with MultiPeriodDiDResults, CallawaySantAnnaResults, or DataFrames
  • Publication-ready formatting with customization options
• Group effects visualization via plot_group_effects()
• Parallel trends testing utilities
  • check_parallel_trends() - simple slope-based test
  • check_parallel_trends_robust() - Wasserstein distance test
  • equivalence_test_trends() - TOST equivalence test

0.3.0

Added

• Synthetic Difference-in-Differences (SyntheticDiD)
  • Unit weight optimization for synthetic control
  • Time weight computation for pre-treatment periods
  • Placebo-based and bootstrap inference
  • SyntheticDiDResults with weight accessors
• Multi-period DiD (MultiPeriodDiD)
  • Event-study style estimation with period-specific effects
  • MultiPeriodDiDResults with period_effects dictionary
  • PeriodEffect dataclass for individual period results
• Data preparation utilities (diff_diff.prep)
  • generate_did_data() - synthetic data generation
  • make_treatment_indicator() - create treatment from categorical/numeric
  • make_post_indicator() - create post-treatment indicator
  • wide_to_long() - reshape wide to long format
  • balance_panel() - ensure balanced panel data
  • validate_did_data() - data validation
  • summarize_did_data() - summary statistics by group
  • create_event_time() - event time for staggered designs
  • aggregate_to_cohorts() - aggregate to cohort means
  • rank_control_units() - rank controls by similarity

0.2.0

Added

• Two-Way Fixed Effects (TwoWayFixedEffects)
  • Within-transformation for unit and time fixed effects
  • Efficient handling of high-dimensional fixed effects via absorb
• Fixed effects support in base DifferenceInDifferences
  • fixed_effects parameter for dummy variable approach
  • absorb parameter for within-transformation approach
• Cluster-robust standard errors
  • cluster parameter for cluster-robust inference
• Formula interface
  • R-style formulas like "outcome ~ treated * post"
  • Support for covariates in formulas

0.1.0

Added

• Initial release
• Basic Difference-in-Differences (DifferenceInDifferences)
  • sklearn-like API with fit() method
  • Column name interface for outcome, treatment, time
  • Heteroskedasticity-robust (HC1) standard errors
  • DiDResults dataclass with ATT, SE, p-value, confidence intervals
  • summary() and print_summary() methods
  • to_dict() and to_dataframe() export methods
  • is_significant and significance_stars properties