Add survey data support for SyntheticDiD and TROP (Phase 5)

[igerber]

Summary

• Add pweight-only survey support to SyntheticDiD and TROP — the last two estimators without survey integration
• SyntheticDiD: Both sides weighted (WLS regression interpretation) — treated means survey-weighted, control synthetic weights composed with survey weights post-optimization (ω_eff = ω * w_co, renormalized). Covariate residualization uses WLS. Placebo and bootstrap SE preserve survey weights on both sides.
• TROP: Survey weights enter ATT aggregation only — population-weighted average of per-observation treatment effects. Model fitting (kernel weights, LOOCV, nuclear norm) unchanged. Rust and Python bootstrap paths both support weighted ATT.
• Extract shared survey helpers (_resolve_pweight_only, _extract_unit_survey_weights) to reduce duplication
• Align to_dict() survey metadata schema across all result classes
• Add NaN finite guard in TROP local bootstrap _fit_with_fixed_lambda() and Rust bootstrap

Methodology references (required if estimator / math changes)

• Method name(s): Synthetic Difference-in-Differences (Arkhangelsky et al. 2021), TROP (Athey, Imbens, Qu & Viviano 2025)
• Paper / source link(s): AER 111(12) 4088-4118; AIQV (2025) working paper
• Any intentional deviations from the source (and why): Survey weighting is a novel extension not in original papers. SDID uses WLS regression interpretation (both sides weighted). TROP weights ATT aggregation only (kernel weights encode similarity, not representativeness). Both documented in REGISTRY.md.

Validation

• Tests added/updated: tests/test_survey_phase5.py (32 new tests), tests/test_trop.py (1 fix for tracking_fit wrapper)
• 403 tests pass (0 regressions), verified with both Rust and Python backends
• AI code review (gpt-5.4-pro): ✅ Looks good — no unmitigated P0/P1 findings

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

🤖 Generated with Claude Code

[github-actions]

Overall Assessment

✅ Looks good

Executive Summary

• No unmitigated P0/P1 issues found in the changed SyntheticDiD and TROP survey-weighting paths.
• The two methodology deviations introduced here are explicitly documented in the Methodology Registry, so they are P3 informational rather than defects: SDID survey weighting in docs/methodology/REGISTRY.md#L1146 and TROP survey weighting in docs/methodology/REGISTRY.md#L1400.
• I did not find new NaN/inference anti-patterns in the touched estimator code; the changed scalar inference paths still go through safe_inference() in diff_diff/synthetic_did.py#L486, diff_diff/trop.py#L842, and diff_diff/trop_global.py#L800.
• Parameter propagation appears complete in the visible call graph: TROP.fit(), trop(), global/local bootstrap helpers, Rust bootstrap entrypoints, result objects, and registry/TODO docs were all updated in sync, e.g. diff_diff/trop.py#L470, diff_diff/trop.py#L930, diff_diff/trop_local.py#L800, diff_diff/trop_global.py#L787, rust/src/trop.rs#L925, and rust/src/trop.rs#L1680.
• The remaining survey-design limitation is properly tracked in TODO.md#L57, so it is mitigated under the review rubric.
• Static review only: I could not execute the test suite in this environment because pytest is unavailable and the default python runtime is missing installed deps such as numpy.

Methodology

• Severity: P3. Impact: SyntheticDiD is affected, but the survey-weighted behavior is explicitly documented and the implementation matches that note: treated-side means are survey-weighted, control-side synthetic weights are composed post-optimization, covariate residualization uses WLS, and both placebo/bootstrap paths preserve survey weights in docs/methodology/REGISTRY.md#L1146, diff_diff/synthetic_did.py#L387, diff_diff/synthetic_did.py#L409, and diff_diff/synthetic_did.py#L455. The underlying SDID source material describes the estimator through a weighted least-squares / doubly weighted DiD lens, so this extension is directionally aligned rather than an undocumented methodology mismatch. Concrete fix: None. citeturn0search19turn2view1
• Severity: P3. Impact: TROP is affected in both method="local" and method="global", but again the deviation is documented: survey weights enter ATT aggregation only, while model fitting / LOOCV / nuclear-norm regularization remain unchanged in docs/methodology/REGISTRY.md#L1400, docs/methodology/REGISTRY.md#L1402, diff_diff/trop.py#L806, diff_diff/trop_global.py#L175, diff_diff/trop_local.py#L971, rust/src/trop.rs#L1030, and rust/src/trop.rs#L1796. The paper’s multi-treated formulation averages per-treated-cell effects, and its bootstrap algorithm resamples treated/control units separately, so the PR’s “aggregation-only” survey threading is consistent with the stated extension. Concrete fix: None. citeturn1search0turn3view1turn3view2

Code Quality

• No material findings in the changed code.

Performance

• No material findings in the changed code.

Maintainability

• Severity: P3. Impact: This PR extends the copy-pasted survey metadata summary / to_dict() pattern into two more result classes in diff_diff/results.py#L740, diff_diff/results.py#L819, diff_diff/trop_results.py#L208, and diff_diff/trop_results.py#L269. That increases sync risk, but the debt is already tracked in TODO.md#L63, so it is informational only. Concrete fix: None required in this PR; eventually extract the shared survey-metadata formatting / dict helper already noted in TODO.

Tech Debt

• Severity: P3. Impact: The PR correctly records the remaining survey limitation for these estimators in TODO.md#L57: strata/PSU/FPC are still rejected by diff_diff/survey.py#L433, and full design-based Rao-Wu style resampling is deferred. Under the review rubric, that tracking mitigates the limitation. Concrete fix: None in this PR.

Security

• No findings.

Documentation/Tests

• Severity: P3. Impact: The new TROP survey tests cover the global path only with a smoke check and a uniform-weight no-op check in tests/test_survey_phase5.py#L409 and tests/test_survey_phase5.py#L446. Only the local path gets an explicit non-uniform-weight regression in tests/test_survey_phase5.py#L511, even though the global survey logic lives in separate code paths in diff_diff/trop_global.py#L175 and diff_diff/trop_global.py#L988. A regression that silently ignored non-uniform survey weights in method="global" could still pass today. Concrete fix: Add a method="global" analogue of test_weighted_att_differs, ideally using heterogeneous unit-level treatment effects so the ATT must move under non-uniform weights.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: e4b3f1074eba026f21e7483b4c3eedd0c7fa3b8c

---

Overall Assessment

✅ Looks good

Executive Summary

• No unmitigated P0/P1 findings in the changed SyntheticDiD and TROP survey-support paths.
• The two estimator-level methodology changes are explicitly documented in the Methodology Registry, so they are P3 informational rather than defects.
• The prior re-review test gap on non-uniform survey weighting for TROP(method="global") is addressed by tests/test_survey_phase5.py:L531-L549.
• Parameter propagation looks complete across fit(), convenience wrappers, local/global bootstrap helpers, Rust bootstrap entrypoints, and result objects in the touched files.
• I did not find new inline inference anti-patterns or partial NaN guards in the modified estimator code.
• Static review only: I could not run the test suite here because pytest is not installed, and a direct Python smoke run was also blocked because numpy is unavailable in this environment.

Methodology

• Severity: P3. Impact: SyntheticDiD is affected, but the survey-weighted behavior is explicitly documented in docs/methodology/REGISTRY.md:L1146 and implemented consistently in diff_diff/synthetic_did.py:L387-L430, diff_diff/synthetic_did.py:L455-L486, and diff_diff/synthetic_did.py:L745-L928. The source method is framed as a weighted-least-squares generalization with unit and time weights, so this documented survey extension is not an undocumented methodology break. Concrete fix: None. citeturn0search0
• Severity: P3. Impact: TROP is affected in both local and global paths, but the registry explicitly documents that survey weights enter ATT aggregation only while model fitting, LOOCV, and nuclear-norm regularization remain unchanged in docs/methodology/REGISTRY.md:L1400. The implementation matches that note in diff_diff/trop.py:L487-L839, diff_diff/trop_global.py:L176-L220, diff_diff/trop_global.py:L489-L1042, diff_diff/trop_local.py:L800-L1063, and rust/src/trop.rs:L924-L1081, rust/src/trop.rs:L1680-L1815. The paper defines ATT over treated cells, keeps the weighting structure in the model-fitting objective, and bootstraps by sampling treated and control units separately, so this documented extension is methodologically coherent rather than an undocumented mismatch. Concrete fix: None. citeturn2view0turn3view0turn3view3

Code Quality

• No findings. The modified scalar inference paths still use safe_inference() rather than inline t_stat/p_value/CI code in diff_diff/synthetic_did.py:L486-L493, diff_diff/trop.py:L841-L843, and diff_diff/trop_global.py:L799-L801.

Performance

• No findings.

Maintainability

• No findings in the changed logic.

Tech Debt

• Severity: P3. Impact: The remaining limitation is properly tracked: full design-based survey resampling and strata/PSU/FPC support are still deferred for these two estimators in TODO.md:L57, with the same scope reflected in docs/survey-roadmap.md:L73-L78. Under the review rubric, that tracking mitigates the limitation. Concrete fix: None in this PR.

Security

• No findings.

Documentation/Tests

• No findings. The previous re-review concern about missing non-uniform-weight coverage for TROP(method="global") appears addressed by tests/test_survey_phase5.py:L531-L549. The new survey test file also covers local/global smoke paths, uniform-weight no-op checks, pweight-only guards, metadata serialization, SDID bootstrap/placebo smoke coverage, and the local NaN-bootstrap regression in tests/test_survey_phase5.py:L121-L384 and tests/test_survey_phase5.py:L395-L787. Static limitation only: I could not execute those tests in this environment because pytest is missing, and a direct Python smoke attempt was blocked by missing numpy.