[vrjdev]

Description

DoubleMLVector.sensitivity_benchmark rebuilds the sub-models of the short-form model with _initialize_models(), so that they use the reduced covariate set. The new sub-models had no learners, because in the vector API the learners are only stored on the sub-models. Refitting the short-form model therefore always raised ValueError: Learner 'ml_l' is required but not set ..., and the method could not be used at all.

• doubleml/double_ml_vector.py: after rebuilding, each new sub-model gets the learners of the sub-model it replaces.
  • The learners are taken from the deep copy, so nothing is shared with the original model.
  • Copying _learners keeps the full learner state created by set_learners (e.g. ml_g cloned from ml_l for the IV-type score). It works for any DoubleMLVector subclass, whatever the arguments of its set_learners.
• doubleml/plm/tests/test_plr_vector_vs_plr.py: the new test_sensitivity_benchmark_equal checks that PLRVector.sensitivity_benchmark equals the legacy DoubleMLPLR.sensitivity_benchmark (rtol=1e-9). It covers both scores, both learners, and n_rep 1 and 3. All 8 cases fail before the fix.

Reference to Issues or PRs

Fixes #440

Comments

This only affects the unreleased vector API from #394.

PR Checklist

• The title of the pull request summarizes the changes made.
• The PR contains a detailed description of all changes and additions.
• References to related issues or PRs are added.
• The code passes all (unit) tests.
• Enhancements or new feature are equipped with unit tests.
• The changes adhere to the PEP8 standards.