Add analytic SE and CI for IPW estimator via influence function

[genrichez]

Adds analytic standard errors and confidence intervals for PropensityScoreWeightingEstimator using the influence function approach from Lunceford & Davidian (2004).

Problem
The IPW estimator has no _estimate_std_error or _estimate_confidence_intervals. When a user asks for CIs, the base class silently falls back to a bootstrap. This is slow and the user never asked for it.

Solution
The estimator computes a Hajek (ratio) weighted mean in each treatment arm. The influence function for this is:
IF_i = (T_i/e_i) * (Y_i - mu1) / (sum(T/e)/n) - ((1-T_i)/(1-e_i)) * (Y_i - mu0) / (sum((1-T)/(1-e))/n)

Then SE = sqrt(Var(IF) / n) and CI = estimate +/- z * SE.
I added:
_compute_influence_function() - computes the per-unit IF for ATE, ATT, and ATC
_estimate_std_error() - returns sqrt(Var(IF)/n)
_estimate_confidence_intervals() - returns estimate +/- z * SE
The stored arrays (_ipw_T, _ipw_Y, _ipw_ps) are saved during estimate_effect so the IF can be computed later without re-fitting.
Important detail: Hajek vs Horvitz-Thompson
All weighting schemes here are actually ratio estimators because the code divides by sum_dy_weights:

est = data["d_y"].sum() / sum_dy_weights - data["dbar_y"].sum() / sum_dbary_weights

This is sum(T*Y/e)/sum(T/e) - ..., which is the Hajek form. Using the simpler Horvitz-Thompson IF (which assumes (1/n)*sum(...) without normalization) gives SE about 3x too large. The implementation uses the correct ratio-linearized IF for all schemes.
Tests
Added 13 new tests covering:

1. SE is positive and finite
2. CI covers the true ATE
3. CI width shrinks with sample size
4. SE scales as 1/sqrt(n)
5. Analytic SE is close to bootstrap SE
6. All 3 target units (ATE, ATT, ATC)
7. All 3 weighting schemes
8. Influence function has mean near zero
9. Heterogeneous effects: ATT differs from ATE with positive selection
10. Existing 2 tests pass unchanged.

Known behavior
The IF-based SE is slightly conservative (about 30% wider than the true sampling variability) because it treats propensity scores as known. This is standard - same behavior as R's WeightIt and Stata's teffects ipw. It means the CI has over-coverage (safe) rather than under-coverage.
Reference
Lunceford, J.K. & Davidian, M. (2004). Stratification and weighting via the propensity score in estimation of causal treatment effects: a comparative study. Statistics in Medicine, 23(19), 2937-2960.
Closes #1640

[genrichez]

Thanks for checking it out, fixed in the latest commit along with the other suggestions

[github-actions]

---

Note

The pull request was not created — a fallback review issue was created instead due to protected file changes: #1646

🤖 This is an automated response from Repo Assist.

Thanks for the swift fixes, @genrichez! I've reviewed the updated commit and everything looks correct:

• ATT control-arm normalizer ✅ — now correctly uses sum_w0_att / n instead of p, matching the ratio-linearization derivation
• ATC treated-arm normalizer ✅ — symmetrically uses sum_w1_atc / n
• hasattr guard ✅ — clear error message when estimate_effect hasn't been called first
• DataFrame target_units comment ✅ — clearly documents the population-ATE fallback

The math is solid across all three target units and the tests are comprehensive. This PR is in great shape! 🎉

Generated by 🌈 Repo Assist, see workflow run. Learn more.

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@11c9a2c442e519ff2b427bf58679f5a525353f76