Feature/wooldridge did

[wenddymacro]

Methodology references (required if estimator / math changes)

• Method name(s):
• Paper / source link(s):
• Any intentional deviations from the source (and why):

Validation

• Tests added/updated:
• Backtest / simulation / notebook evidence (if applicable):

Security / privacy

• Confirm no secrets/PII in this PR:

[igerber]

Thanks for this contribution — the Wooldridge ETWFE estimator fills a real gap in the library. The nonlinear outcome support (logit/Poisson via the ASF approach) is something we've had on our roadmap, and this is the right paper to implement it from.

I've done an independent review of Wooldridge (2023) against your implementation. The REGISTRY.md section is solid overall, but there are items that need to be addressed before we can move forward.

Files to remove
These should not be part of the PR:

• uv.lock — we don't use uv
• docs/superpowers/CHECKPOINT.md — implementation scaffold
• docs/superpowers/plans/2026-03-18-wooldridge-did.md — same
• docs/superpowers/specs/2026-03-18-wooldridge-did-design.md — same

PR template

Please fill in the methodology references, validation, and security fields in the PR description.

Methodology notes from paper review

I did an independent review of Wooldridge (2023) and have a few items to flag:

1. Delta-method gradient verification. The paper does not write out the explicit gradient for the nonlinear ASF-based ATTs (it cites Wooldridge 2010, problem 12.17). Your implementation hand-codes this gradient. Please add a numerical gradient check (finite differences) as a test to verify correctness — this is the most error-prone part of the nonlinear path and the whole value proposition of this estimator depends on it.
2. Aggregation weights. The paper describes aggregation conceptually (Section 3.1, p. C50) but doesn't give explicit weight formulas with equation numbers. Please confirm your weight definitions match Stata jwdid_estat and document the exact formulas in the REGISTRY.md section.
3. Covariate centring. Wooldridge centres covariates at cohort means (X̂_{ig} = X_i - X̄_g, p. C48) in the treatment interaction terms. This affects interpretation of δ_{gs} on the index scale. Please verify this is implemented correctly.
4. Proposition 3.1 equivalence. The paper proves that imputation and pooled estimation produce identical ATTs when using the canonical link function. This is a testable property - please add a test that verifies this equivalence numerically (e.g., compare imputation-based ATTs to pooled ATTs for logit/Poisson on simulated data).

Validation
We need at least one numerical benchmark against Stata jwdid or R etwfe — verifying that ATT estimates match to a reasonable tolerance. See benchmarks/R/ for examples of how we validate other estimators.

Additionally, a test comparing OLS ETWFE ATT(g,t) to CallawaySantAnna ATT(g,t) on the same dataset would be valuable, since they should be approximately equivalent under linear PT with never-treated controls.

Code conventions

• The nonlinear paths (logit/Poisson) build their own sandwich vcov manually. Our convention is to route all inference through compute_robust_vcov in linalg.py - see CLAUDE.md under "Unified linalg.py backend". Please refactor to use that where possible.
• All inference fields (t_stat, p_value, conf_int) must be computed together using safe_inference() from diff_diff.utils. Please verify this is the case throughout.
• The Poisson FE path (Section 3.3 of the paper) does not suffer from incidental parameters - this is a unique property of the exponential mean. If your implementation uses unit dummies for Poisson, this is fine, but please document the distinction from the logit path where unit dummies would cause incidental parameters problems.

Happy to answer questions on any of this. This is a meaningful contribution and we want to make sure it lands well.