[Repo Assist] fix(estimators): delta-method CI/SE for LinearRegressionEstimator with effect modifiers

[github-actions]

🤖 This is an automated PR from Repo Assist, an AI assistant for this repository.

Problem

LinearRegressionEstimator._estimate_confidence_intervals and _estimate_std_error raised NotImplementedError whenever effect modifiers were present (the code had a # TODO: Looking for contributions comment). This forced users to fall back to expensive bootstrapping, or receive an error, even though the exact analytic solution is straightforward for a linear model.

Closes #336

Root Cause

The ATE for a linear model with effect modifiers is a linear combination of OLS coefficients:

ATE = (b_t + b_{t·x1}·mean(x1) + b_{t·x2}·mean(x2) + ...) × (t1 – t0)
    = gradient @ beta
```

where `gradient = mean(X_t1 − X_t0, axis=0)` is the vector of mean feature differences under `treatment_value` vs `control_value`. By the delta method:

```
Var(ATE) = gradient @ cov_params @ gradient
SE(ATE)  = sqrt(Var(ATE))

This is exact for linear models (not an approximation) and requires no simulation.

Fix

Added _delta_method_std_error() helper to LinearRegressionEstimator that:

1. Builds feature matrices X_t1 / X_t0 with treatment fixed to treatment_value / control_value.
2. Computes gradient = mean(X_t1 - X_t0, axis=0).
3. Returns sqrt(gradient @ cov_params @ gradient).

_estimate_confidence_intervals and _estimate_std_error now call this for the effect-modifier case instead of raising NotImplementedError. The non–effect-modifier path is unchanged.

The approach naturally handles:

• Multiple effect modifiers (including categorical ones, via the existing OneHotEncoder reuse).
• Arbitrary treatment_value / control_value scaling.
• Multiple confounders.

References

• Gelman & Hill, ARM, Chapter 9 — delta method for linear combinations of regression coefficients.
• [Original issue paper cited in Add asymptotic confidence intervals for average treatment effect for linear regression with effect modifiers #336]((scholar.princeton.edu/redacted)

Test Status

21 new tests added, covering:

• No NotImplementedError for CI/SE with effect modifiers.
• CI shape is (1, 2) for the effect-modifier case.
• CI bounds are finite and properly ordered (lower < upper).
• CI contains the point estimate.
• SE is positive and finite.
• Multiple effect modifiers.

All 21 tests in test_linear_regression_estimator.py pass. Linting matches the pre-existing baseline (no new flake8 errors introduced).

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