Refs #571.

…e B2 (#583)

Pin the absorbing LPDiD estimator against the method authors' own R recipes
(danielegirardi/lpdid) with an alexCardazzi/lpdid cross-check gate:
- benchmarks/R/generate_lpdid_golden.R: in-R panel (+ interior-gap unit) and 6
  variants (variance-weighted, reweight, pmd, direct-covariate, pooled,
  RA-point); writes committed lpdid_test_panel.csv + lpdid_golden.json.
- tests/test_methodology_lpdid.py: skip-guarded parity (att/se to ~1e-12,
  cross-platform asserted at 1e-6/1e-7).
- benchmarks/python/coverage_lpdid_ra.py + lpdid_ra_coverage.json: ungated
  Monte-Carlo study validating the RA influence-function SE calibration (~0.95).

Resolves the two provisional REGISTRY deviation notes in the library's favour
with no estimator change: the RA SE matches the Stata teffects convention
(point-anchored, SE pinned + coverage-validated; no R-package analogue), and the
pooled estimand matches the authors' fixed-composition recipe (correcting the
prior "horizon-stacked" wording). no_composition documented as more paper-faithful
than the R packages (B1-tested). Ticks the REGISTRY B2 checklist box.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…awaySantAnna (#582)

* feat(staggered): materialize non-estimable (g,t) cells as NaN in CallawaySantAnna

Uniformly materialize a NaN entry (with a machine-readable skip_reason) for every
non-estimable (g,t) group-time cell across all CS estimation paths (no-covariate
regression, covariate regression, IPW/DR, repeated cross-section, survey-weighted)
instead of omitting it. Previously only the covariate-singular case materialized
NaN; the other paths dropped the cell silently from the grid.

Cells carry no influence-function entry, so they are excluded from every
aggregation (simple/group/calendar/event-study), balance_e, and bootstrap -- all
aggregate point estimates and SEs, plus event-study n_groups / by-group n_periods,
are numerically unchanged and continue to match R did's aggte(). A fit where no
cell is estimable still raises ValueError. to_dataframe("group_time") now includes
the NaN rows and a skip_reason column.

Documented per-cell surface deviation from R's att_gt (which omits the rows).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(staggered): uniform no-IF NaN cells + cover covariate paths (review #582)

- The covariate-regression non-finite cell now materializes via _nan_gt_entry
  with NO influence-function entry, matching the other paths and the documented
  REGISTRY/helper contract (previously it wrote a zero-IF entry and ran batch
  inference). Aggregates and SEs are unchanged -- the cell was finite-masked /
  IF-membership-filtered out either way; now the "NaN cells carry no IF entry"
  invariant holds uniformly across all paths.
- Extend the materialization test to cover covariate IPW/DR (panel + RCS) paths.
- Remove the now-implemented CallawaySantAnna NaN-cell row from TODO.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(staggered): treat non-finite ATT(g,t) as non-estimable in general/RCS paths (review #582 P1)

The general (IPW/DR) and RCS estimable builders treated `att_gt is not None` as
estimable even when att_gt was non-finite (NaN/inf): they stored effect=att_gt
(which could surface inf), kept the influence-function entry, and did not count
the cell in the consolidated skip total. Both now branch on
`att_gt is None or not np.isfinite(att_gt)` first and materialize via
_nan_gt_entry(skip_reason="non_finite_regression") with NO IF entry, so the
documented contract (non-estimable cells are NaN entries, carry no IF, excluded
from the bootstrap) holds uniformly across every path. Aggregate estimates and
SEs are unchanged (these cells were finite-masked / IF-filtered out either way).
Adds a regression test mimicking the inf-ATT-with-IF leak.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(staggered): guard non-finite ATT in the no-covariate path too (review #582 P0)

Completes the uniform invariant: every CallawaySantAnna estimable-cell builder
(no-covariate vectorized, covariate regression, general IPW/DR, RCS) now routes a
non-finite ATT(g,t) to _nan_gt_entry(skip_reason="non_finite_regression") with NO
influence-function entry, skipping batch inference and bootstrap membership. The
no-covariate diff-in-means ATT is finite given n_t,n_c>0, but a non-finite outcome
(inf survives the NaN-only valid mask) could otherwise store inf as the effect and
produce t_stat=inf / p=0 / infinite CI via safe_inference_batch. Aggregate
estimates and SEs are unchanged. Adds a regression test injecting an inf outcome
through the no-covariate path.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(staggered): omit all-non-estimable relative-time buckets from event study (review #582 P1)

_aggregate_event_study() appended an all-NaN row (effect=NaN, se=NaN, n_groups=0)
for a relative-time bucket whose cells are all non-estimable, instead of omitting
the bucket. With NaN-cell materialization this surfaced new all-NaN event-study
rows where the bucket previously had no cells (and thus no row) -- an aggregate-
surface change vs the prior omit behavior and R did::aggte(). The bucket is now
dropped when finite-filtering leaves no cell (tracked via a kept-periods list so
the result lists stay aligned), matching _aggregate_by_group, which already omits
all-NaN groups. Adds a test asserting an all-non-estimable relative time is absent
from event_study_effects.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs(staggered): drop nonexistent "calendar" aggregation from CS NaN-cell notes (review #582 P3)

CallawaySantAnna's aggregation options are simple / event_study / group / all
(there is no calendar aggregation). Remove "calendar" from the REGISTRY edge-case
Note and the CHANGELOG entry listing which aggregations exclude materialized NaN
cells.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(staggered): report real treated/control counts on non-estimable cells (review #582 P2)

The per-cell helper sentinel returns (and the covariate-reg empty-control batch
site) hardcoded n_treated=n_control=0 for materialized NaN cells even after the
observation masks had been computed, so group_time_effects / to_dataframe could
show zero counts for a cell that actually had treated (or control) observations.
The zero-control / zero-weight exits in _compute_att_gt_fast and _compute_att_gt_rc,
and the covariate-reg empty-control batch site, now return the observed counts;
missing-period exits (masks not yet built) keep 0. Display-only metadata --
estimates, SEs, and aggregation are unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(staggered): handle materialized NaN cells in ported CS tests

Three test_csdid_ported.py tests relied on non-estimable (g,t) cells being
OMITTED from group_time_effects; with the new NaN-cell materialization those
cells are present as NaN, so the tests' membership / golden-iteration guards no
longer skip them. Update them to preserve their intent on the FINITE cells:
- test_some_units_treated_first_period: a first-period cohort (no base period) is
  now all-NaN (missing_period) rather than absent -> assert it is all-NaN.
- test_zero_pretreatment_outcomes: skip NaN pre-cells (the last cohort under
  not_yet_treated has no controls); finite pre-cells are still ~0.
- test_golden_fewer_periods: skip NaN cells (gapped panel where base g-1 is
  unobserved -> missing_period; R falls back to an available base) -> R-parity on
  the finite cells.

No source change; the cells are correctly non-estimable, only now visible.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…584)

* feat(lpdid): non-absorbing treatment (entry-effect estimands), Phase C1

Implement Dube, Girardi, Jorda & Taylor (2025) Section 4.2 non-absorbing
treatment for LPDiD via a new `non_absorbing` parameter:

- "first_entry" (Eq. 12): effect of entering treatment for the first time and
  staying treated; reuses the absorbing clean control, restricts only the
  treated set. Bit-identical to the absorbing path on absorbing panels.
- "effect_stabilization" (Eq. 13, `stabilization_window=L`): units whose
  treatment has been stable for >= L periods serve as clean controls, so
  estimation is feasible with few/no never-treated units.

Default `non_absorbing=None` is unchanged (absorbing path, still rejects
non-absorbing input). Mode-aware clean-sample masks evaluate window conditions
via cumulative treatment-change/level lookups with a documented "untreated
before the first observed period" boundary convention; placebo horizons use the
full pre-span window so pre-trends are uncontaminated; a per-horizon clean-
treated indicator threads through the estimator / RA / reweight / pooled paths
so re-entry events are classified correctly. Non-absorbing modes require a
gap-free panel within each unit's observed span.

Pure-Python validation (tests/test_lpdid.py::TestLPDiDNonAbsorbing): absorbing
reduction, single-cohort reduction, re-entry mechanism, boundary retention,
negative-horizon placebos, non-negative weighting, stabilized-control
admission, equal-weight recovery, and DGP recovery; absorbing tests + R-parity
goldens unchanged. Exit-event dynamics, R-package parity (PR-C2), and
survey-design support are tracked follow-ups.

* fix(lpdid): non-absorbing pooled-pre uses deepest reach-back horizon

Codex P1: `_build_pooled_sample(kind="pre")` passed horizon=0 to the
non-absorbing masks, so the effect_stabilization clean window only covered
[t-L, t] instead of the pooled-pre reach-back to the most-negative horizon
([t - max(L, -h), t-1]). A unit with a prior treated spell at t-3 (clean at
t-1) leaked into a [-3, -2] pooled-pre placebo and biased it. Pre windows now
use min(horizons); the absorbing branch keeps horizon=0 (not-yet-treated at t
already implies a clean pre-span, so its R-parity goldens are unchanged).

Adds a deterministic regression test (spell entrants excluded from the
pooled-pre sample; verified to fail at 0.286 before the fix).

…y absorbed FE (#586)

N>1 absorbed fixed effects used single-pass sequential demeaning, which is the
exact (weighted) Frisch-Waugh-Lovell residualization only on balanced
orthogonal-FE panels; on unbalanced panels it was a biased approximation
(coefficients off by ~1e-2 in tested cases).

Add an N-way method-of-alternating-projections engine demean_by_groups() in
utils.py; route the DiD/MultiPeriodDiD absorb= paths and the shared two-way
within_transform() through it, fixing TWFE / SunAbraham / BaconDecomposition on
unbalanced unweighted panels too. Lift the weighted-multi-absorb rejection
(now supported via weighted MAP).

Single-absorb and balanced-panel results are byte-stable; the weighted
within_transform output is bit-identical; R-parity goldens unchanged.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…C2 (#587)

* Add LPDiD complex-survey-design support (Phase D1)

Adds a `survey_design=` argument to `LPDiD.fit()` (a `SurveyDesign` with
probability weights + optional strata/PSU/FPC), matching the library-wide
fit()-time convention. On the variance-weighted default path each horizon's
long-difference regression is fit by WLS on the survey weights, and the SE is
the stratified-PSU Taylor-linearization (Binder TSL) sandwich with
`df = n_PSU - n_strata`, reusing `diff_diff/survey.py` (`compute_survey_vcov`).

The design is re-resolved on each realized (post-clean-control) sample so
weights/strata/PSU align with the regression rows; with no explicit PSU the
unit is injected as the PSU. Fails closed to NaN on under-identified samples.
Rejects `survey_design` with `reweight=True` (the equally-weighted /
regression-adjustment IF path), replicate-weight designs, and non-pweight
types (deferred follow-ups). `LPDiDResults` gains `survey_metadata` /
`n_strata` / `n_psu`, a `"survey_tsl"` vcov_type, and a Survey Design block in
`summary()`. The non-survey path is byte-for-byte unchanged.

Validated against `survey::svyglm` on the stacked long difference (numeric
golden parity is the D2 follow-up); 15 new pure-Python invariant tests
(reduction/unit-clustering, FPC-shrinks-SE, stratification, lonely-PSU,
NaN-consistency, weighting-moves-point, metadata, rejection paths).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(lpdid): report survey PSU count as headline G for only_event survey fits

CI-codex P2: under a survey design the effective variance cluster is the PSU
(cluster_name reports the PSU column), but for only_event=True fits (pooled is
None) headline_n_clusters fell back to the panel unit count -- so an explicit
PSU design with n_psu != n_units could display the unit count mislabeled as
G=<psu>. Per-row event-study n_clusters and inference were already computed on
the realized survey design, so this was a metadata/labeling issue only, not a
wrong SE/p-value. Fix: when a survey design is active, seed headline_n_clusters
from the panel-level effective PSU count (the pooled-post override still prefers
the realized survey-sample count when available). Regression test added
(only_event=True, explicit PSU, n_psu != n_units).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(lpdid): build survey sandwich on the kept-column design (rank-deficient contract)

CI-codex P1: `_estimate_survey_sample` computed the Binder TSL sandwich on the
UNREDUCED design and recomputed `response - design @ coef`. When the rank
handler drops a redundant direct-inclusion covariate / absorbed dummy / lag
(setting that coef to NaN while `treatment_entry` stays identified), the NaN
coef propagated through the residuals and the full-design X'WX bread
singularized, collapsing an otherwise-identified treatment SE/t/p/CI to NaN --
violating the library's rank-deficient contract that the non-survey solve_ols
path honors.

Fix: keep solve_ols's returned residuals (original-scale, computed on the
identified reduced design) and build `compute_survey_vcov` on `design[:, kept]`
where `kept = isfinite(coef)`, mapping treatment back to its kept-column index.
If treatment itself is dropped, the effect is NaN and the SE stays NaN;
`rank_deficient_action="error"` still raises from solve_ols. Regression test
(duplicate + constant covariate, `silent`) asserts the treatment SE stays
finite and equals the non-redundant reference fit, and that `error` raises.

CI-codex P2: type-check `survey_design` before `_survey_columns` accesses its
attributes, so a non-SurveyDesign argument raises the intended TypeError rather
than an incidental AttributeError (test added).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ull-design convention (#589)

* docs(survey): waive zero-weight-PSU SE-invariance item; lock Lumley full-design convention

Re-examined the TODO row proposing the survey TSL finite-sample correction
count only positive-weight PSUs so the SE is invariant to zero-weight
(subpopulation / padded) rows. Investigation shows the premise conflicts with
the library's documented, R-validated convention:

- `_compute_stratified_psu_meat`'s per-stratum correction
  (1 - f_h)*n_PSU_h/(n_PSU_h - 1) and PSU-mean centering intentionally keep
  genuine-subpopulation zero-weight PSUs. This is the full-design domain
  estimator of Lumley (2004 Section 3.4) / R survey::svyrecvar(subset()),
  already documented in REGISTRY section "Subpopulation Analysis".
- The ATT is exactly invariant; the survey SE is deliberately NOT invariant to
  genuine-subpopulation zeroing (it should differ from a naive physical subset
  -- that is the whole point of subpopulation()). R produces the matching SE
  (only df differs).
- Zero-weight rows that reuse an existing PSU label are already bit-invariant.
  The only invariance-violating shape -- appending synthetic new all-zero PSUs
  -- arises in no estimator path (domain padding goes through prep.py's
  zero-padded full-design cell variance, which retains the real PSU layout).

Forcing the meat to positive-weight-only counting would break the documented
Lumley/R parity, so the item is waived (no estimator behavior change):

- TODO.md: move the row from Actionable Backlog to "Won't-fix / waived
  (decisions on the record)" with the Lumley/R justification.
- REGISTRY.md: add a Note in section "Subpopulation Analysis" making explicit
  that the TSL meat finite-sample correction counts zero-weight PSUs by design.
- tests/test_survey.py: add TestZeroWeightPsuConventionWaiver regression-lock
  (inert existing-PSU padding is bit-invariant; subpopulation zeroing keeps the
  full PSU structure so its SE differs from a naive subset). A future
  positive-weight-only change would collapse the two and trip the test.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(survey): add direct _compute_stratified_psu_meat full-design unit test

Addresses the CI review's actionable P3 (Documentation/Tests): the SE-level
test only asserts that subpopulation zeroing differs from a physical subset,
which catches a full positive-weight-only rewrite but could miss a partial
edit (e.g. changing only the finite-sample denominator while still centering
over the zero PSU).

Adds a direct unit test on _compute_stratified_psu_meat with crafted PSU
scores including one all-zero-score PSU (a fully zeroed subpopulation PSU),
asserting the exact full-design meat formula (n_PSU_h including the zero
PSU) and that it is NOT the positive-weight-only meat. Any change to the
centering OR the denominator now trips the lock.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(survey): model a true zero-weight PSU in the direct meat fixture

Addresses the re-review's actionable P3: the direct meat test represented its
all-zero PSU via zero score rows only, with weights left all ones. A future
denominator-only edit that reads resolved.weights to drop positive-weight PSUs
would not have been caught. Set PSU 2's weights to 0 so the fixture models a
true fully zero-weight subpopulation PSU. The current meat ignores weights
(it operates on scores), so the expected value is unchanged; the change only
hardens the regression lock.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…od) (#592)

…(Phase D2) (#591)

Release 3.6.1. Changes since 3.6.0:
- LPDiD non-absorbing (reversible) treatment with entry-effect estimands
  (Dube, Girardi, Jorda & Taylor 2025) + complex-survey-design support
  (survey_design=), each R-parity validated.
- TROP non-absorbing (on/off) treatment support, opt-in local method
  (Athey, Imbens, Qu & Viviano 2025).
- Weighted multiple absorbed fixed effects (absorb=[a, b, ...]) via
  iterative alternating-projection demeaning.
- CallawaySantAnna materializes non-estimable (g,t) cells as NaN.
- Fix: BusinessReport appendix render failures now surfaced.
- R-parity validation backfill for the LPDiD absorbing/non-absorbing/survey
  paths; survey zero-weight-PSU SE-invariance item waived (Lumley
  full-design convention); SciPy lower-bound doc alignment.

Promotes the CHANGELOG [Unreleased] section to [3.6.1] - 2026-07-01 and
syncs the version across __init__.py, pyproject.toml, rust/Cargo.toml,
llms-full.txt, and CITATION.cff.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>