Address JOSS review feedback (#147, #148, #149)

[TomeHirata]

Addresses the issues raised in the JOSS review (openjournals/joss-reviews#11047).

Changes

• Outcomes shape handling in predict_dte #147 – outcome shape handling: all fit methods now flatten single-column treatment_arms, outcomes, strata and treatment_indicator (e.g. shape (n, 1)) to 1-D and raise a clear ValueError for other shapes. predict_dte now returns identical results for Y and Y[:, None].
• Statistical validation with known baselines #148 – QTE bug: predict_qte returned wrong values (e.g. 1 instead of 5 for a constant shift of 5) because the quantile search returned the largest location with F(y) <= q. It now returns the generalized inverse, the smallest y with F(y) >= q. Added tests/test_statistical_baselines.py with hand-computed baselines (DTE/PTE/QTE), a stratified-vs-simple equivalence check, and a simulated normal location-shift DGP with known DTE/QTE for the simple and adjusted estimators (including a CI-width check for the adjusted one). The existing mock-based test_predict_qte relied on the buggy behaviour and now uses a CDF mock with a known quantile shift.
• Confounding adjustment language in docstrings #149 – confounding language: docstrings of the adjusted estimators, the docstring examples (now randomized treatment) and the Oregon tutorial now describe ML adjustment as variance reduction for randomized experiments, not confounding correction.
• Minor points
  • Empty folds: cross-fitting now raises an informative ValueError suggesting fewer folds when a fold leaves no training data for the target arm.
  • Missing values: fit rejects NaN in treatment_arms, outcomes, strata and treatment_indicator; covariates are not checked (depends on the base model). Documented in the get-started guide.

Not included

Bootstrap inference options for the partial-compliance (predict_ldte/predict_lpte) methods. This is a new statistical feature rather than a fix and needs a design that respects the stratified sampling scheme, so I left it for a follow-up.

Testing

pytest (72 tests) and ruff check . pass.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.