[kbattocchi]

Summary

Stacked follow-up to #1069. Contains two independently confirmed inference fixes after auditing the codebase for other issues with multidimensional arrays; point-estimation logic is unchanged but we had bugs in our confidence intervals for DynamicDML and for OrthoForest's marginal effects when using an explicit treatment featurizer.

• DynamicDML: correct the cross-period Jacobian blocks so rows index current-period score coordinates and columns index future-period coefficient coordinates. The previous outer-product order transposed nonsymmetric off-diagonal blocks and therefore corrupted the sandwich covariance. Added comments to clarify the period/treatment/feature covariance ordering in the implementation.
• OrthoForest BLB inference: take the square root of the projected covariance before returning transformed marginal-effect standard errors. This bug was just variance being reported as standard error, not an axis transpose.

Regression coverage

• DynamicDML: independent numerical derivatives of stacked panel scores, independently solved point coefficients, and public effect-inference standard errors. Covers one/three periods, scalar/two-coordinate treatments, no/two heterogeneity features, and vector/single-column/two-output outcomes. Nonsymmetric cross-period coupling exposes the original bug. This estimating-equation/finite-difference reference reuses fitted nuisance residuals to isolate final-stage correctness. Each independent panel unit contributes one influence vector with periods kept together, giving coefficient sampling covariance rather than outcome variance.
• OrthoForest: fixed coefficient covariance with correlated errors; analytical derivatives and coefficient-error factors give independent expected standard deviations for identity, linear, and quadratic treatment transformations. Checks point estimates, standard errors, variances, confidence intervals, identity invariance, and scalar/vector shapes.
• With an explicit identity treatment featurizer, compare marginal-effect inference for each treatment coordinate against effect inference for a one-unit change in that coordinate, using the same fitted inference object, inputs, and confidence level (alpha=0.1). Check point estimates, standard errors, and both confidence-interval endpoints through effect_inference and effect_interval. Preserve scalar/column outcome shapes. Both inference calculations run unmodified against the same fixed estimator fixture.

Validation

• Before the implementation fixes: 24 failed, 18 passed across the original new regressions (six DynamicDML covariance cases and all 18 OrthoForest cases failed).
• After the fixes: 42 passed across those regressions.

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Copilot review overview

2 open issues

• projected_var can become slightly negative due to floating-point roundoff even when cov is PSD,… · New
• This test imports BLBInference from a private module (_ortho_forest), which increases coupling… · New

What changed in this PR

Fixes two inference bugs affecting reported uncertainty (not point estimates) in DynamicDML and OrthoForest, and adds regression tests to prevent recurrence.

Changes:

• Correct DynamicDML cross-period Jacobian block orientation used in sandwich covariance calculation.
• Fix OrthoForest BLB marginal-effect inference to return standard errors (sqrt of projected variance).
• Add focused regression tests covering multidimensional shape cases and transformation/Jacobian paths.

File	Description
econml/​orf/​_ortho_forest.py	Fixes BLB marginal-effect scaling to report standard errors by taking sqrt of projected variance.
econml/​panel/​dml/​_dml.py	Corrects Jacobian block outer-product order for DynamicDML covariance and clarifies ordering via comments/docstring.
econml/​tests/​test_orf.py	Adds regression tests validating BLB transformed marginal-effect stderr/CI behavior across shapes and transforms.
econml/​tests/​test_dynamic_dml.py	Adds regression test comparing DynamicDML covariance/effect inference vs. independent finite-difference score Jacobian reference.

_{Review effort: Lite}

[kbattocchi]

Agreed that roundoff can make a theoretically nonnegative projected variance slightly negative. I'd defer clipping here: unconditional maximum(..., 0) would also hide materially invalid covariance estimates, and the existing effect and constant-marginal inference paths take square roots without clipping. Numerical hardening should use a consistent, scale-aware policy across those paths, with near-singular regression cases. This fix stays focused on converting variance to standard error.

[kbattocchi]

Direct use of BLBInference is deliberate here. The comment beside factor explains that the stub supplies known coefficient covariance to isolate inference propagation from forest fitting. The inference methods themselves are real, and the test checks an independent factor-based standard-deviation reference plus agreement between identity-transform marginal and finite-contrast inference. A fitted-forest-only test would lose that controlled reference, so I'd retain this focused component test alongside the existing estimator tests.