Merge pull request #426 from igerber/spillover-conley-phase2

Phase 2: Conley space-time HAC + panel-estimator wire-up (MPD/TWFE)

Author: igerber

CHANGELOG.md

@@ -8,7 +8,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Added
-- **Conley (1999) spatial-HAC standard errors via `vcov_type="conley"`** on cross-sectional `LinearRegression` / `compute_robust_vcov` (Phase 1 of the spillover-conley initiative). Keyword arguments: `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"` evaluated on pairwise distance `d_ij/h`, matching R `conleyreg`; both kernels emit a `UserWarning` if the resulting meat has a materially negative eigenvalue. Conley 1999's explicit PSD Bartlett formula is the 2-D separable product window on a lattice (Eq 3.14); the 1-D radial pairwise specialization that diff-diff and R `conleyreg` implement is a practitioner convention that is not formally PSD-guaranteed). 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 estimators (`DifferenceInDifferences`, `TwoWayFixedEffects`, `MultiPeriodDiD`) reject `vcov_type="conley"` at fit-time with `NotImplementedError`** — Phase 1's cross-sectional Conley does not handle the time dimension. Applying it over (unit, time) rows would treat same-unit cross-time pairs as `d_ij = 0 → K = 1`, mishandling the space-time HAC. Practitioners needing Conley with a panel design should pre-collapse to per-unit first-differences and call `compute_robust_vcov` directly on a single-period regression. Phase 2 will add the space-time product kernel (Driscoll-Kraay) for full panel support. `SyntheticDiD(vcov_type="conley")` raises `TypeError` (uses bootstrap variance, not analytical sandwich); `set_params` mirrors the constructor rejection. `vcov_type="conley"` + `cluster_ids=` / `weights=` / `survey_design=` raises `NotImplementedError` (combined product kernel + Bertanha-Imbens 2014 weighted-Conley deferred to follow-up phases). `n > 20_000` emits a `UserWarning` about the dense O(n²) distance-matrix memory; sparse k-d-tree fast path is queued for Phase 2. Helpers live in new module `diff_diff/conley.py` (`_haversine_km`, `_pairwise_distance_matrix`, `_bartlett_kernel`, `_uniform_kernel`, `_validate_conley_kwargs`, `_compute_conley_vcov`); `compute_robust_vcov` in `diff_diff/linalg.py` imports the dispatch helpers. R `conleyreg` parity (Düsterhöft 2021, CRAN v0.1.9) on three benchmark fixtures (`benchmarks/data/r_conleyreg_conley_golden.json`, regenerable via `benchmarks/R/generate_conley_golden.R`); 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`. Phase 1 of a multi-phase spillover-conley initiative; subsequent phases (space-time product kernel + sparse fast path + panel-estimator support; ring-indicator spillover-aware DiD per Butts 2021; mechanical extension to IF-aggregation and sandwich-derived estimators; survey design support) are tracked in `TODO.md` under "Tech Debt from Code Reviews" → spillover-conley rows.
+- **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")` continues to raise `NotImplementedError`** because `DiD.fit()` has no `unit` column declaration; the error message redirects users to `MultiPeriodDiD` / `TwoWayFixedEffects`. **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"` + `cluster_ids=` / `weights=` / `survey_design=` raises `NotImplementedError` (combined product kernel + Bertanha-Imbens 2014 weighted-Conley deferred to follow-up PRs). TWFE's default auto-cluster on the Conley path is silently dropped (the explicit `cluster=` raises with a clear redirect). `inference="wild_bootstrap"` + Conley raises (incompatible inference modes). `n > 20_000` emits a `UserWarning` about the dense O(n²) distance-matrix memory; sparse k-d-tree fast path is a follow-up. **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 (sparse k-d-tree fast path for `n > 20_000`; Conley + `cluster_ids` combined kernel; 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 remains queued as a separate notebook PR.
 - **`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).

README.md

@@ -124,7 +124,7 @@ Full guide: `diff_diff.get_llm_guide("practitioner")`.
 - [Honest DiD](https://diff-diff.readthedocs.io/en/stable/api/honest_did.html) - Rambachan & Roth (2023) sensitivity analysis: robust CI under PT violations, breakdown values
 - [Pre-Trends Power Analysis](https://diff-diff.readthedocs.io/en/stable/api/pretrends.html) - Roth (2022) minimum detectable violation and power curves
 - [Power Analysis](https://diff-diff.readthedocs.io/en/stable/api/power.html) - analytical and simulation-based MDE, sample size, power curves for study design
-- Conley spatial HAC SE (`vcov_type="conley"`) on cross-sectional `LinearRegression` / `compute_robust_vcov` - Conley (1999) spatial-correlation-aware SEs with parity vs R `conleyreg`
+- Conley spatial HAC SE (`vcov_type="conley"`) on cross-sectional `LinearRegression` / `compute_robust_vcov` plus panel `MultiPeriodDiD` / `TwoWayFixedEffects` (with `conley_lag_cutoff` for within-unit Bartlett temporal HAC) - Conley (1999) spatial-correlation-aware SEs with parity vs R `conleyreg` on cross-sectional + panel fixtures
 
 ## Survey Support