Add health insurance premiums to local area calibration, a matrix builder function, Modal model fitting #473
Conversation
- Add build_calibration_matrix.py runner for building X_sparse matrix with SNAP + health_insurance_premiums_without_medicare_part_b targets - Add test_sparse_matrix_builder.py with tests for person-level aggregation and combined target support - Update changelog_entry.yaml Closes #472 Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
- Add fit_calibration_weights.py for L0-regularized weight optimization - Add l0-python as optional dependency under [project.optional-dependencies] - Uses SparseCalibrationWeights from l0.calibration package - Builds calibration matrix for all 436 congressional districts - Filters to achievable targets before fitting Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace sparsity_log with calibration_log that tracks per-target performance (estimate, target, error metrics) across epochs. Update Modal runner to return both weights and log files. Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
| @@ -0,0 +1,64 @@ | |||
| import pytest | |||
There was a problem hiding this comment.
@juaristi22 You've taken the testing way further. Let me know if I should delete this file and let your PR bring in the tests (or is there anything to take over to your PR?)
There was a problem hiding this comment.
I think we can delete this file, my PR should cover it.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Avoid confusion with fit_calibration_weights.py which contains the actual fitting logic. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
baogorek
changed the title
| @@ -0,0 +1,64 @@ | |||
| import pytest | |||
There was a problem hiding this comment.
I think we can delete this file, my PR should cover it.
| @@ -0,0 +1,195 @@ | |||
| """ | |||
There was a problem hiding this comment.
I'm wondering if we need this file. I feel like the local_area_calibration_setup notebook has the tools to inspect the matrix already, and adding a new variable to inspect shouldn't take much effort. Meanwhile, the test in my PR and the fit_calibration_weights.py files already contain the full workflow to build the matrix and check the values of variables in them. What do you think?
| sim, | ||
| target_filter={ | ||
| "stratum_group_ids": [4], | ||
| "variables": ["health_insurance_premiums_without_medicare_part_b"], |
There was a problem hiding this comment.
| "variables": ["health_insurance_premiums_without_medicare_part_b"], | |
| "variables": ["health_insurance_premiums_without_medicare_part_b", "snap"], |
Running this and seeing the magic happen was cool! One small note, I think we also need to add snap to the list of variables because otherwise National SNAP (the total amount) target won't be included as it belongs to a stratum_group_id different from 4.
Summary
This has really morphed into a hodge-podge of updates. Technically, the health insurance premium variable is included in the matrix builder, but this PR doesn't include any specific testing of the variable. Thus, @juaristi22 's PR #475 will go in first to cover the testing of this variable.
I'm also going to hold off on linting until #475 goes in, so that the "Files changed" doesn't increase unnecessarily.
build_calibration_matrix.pyrunner script for building sparse calibration matrices with combined SNAP and health insurance premium targetstest_sparse_matrix_builder.pywith tests for person-level aggregation and combined target supportDetails
The
health_insurance_premiums_without_medicare_part_bvariable is a person-level target that needs to be aggregated to household level. The existingsparse_matrix_builder.pyalready handles this through PolicyEngine'smap_to="household"parameter, which automatically sums person-level values to household level while preserving totals.Closes #472
Test plan
pytest policyengine_us_data/tests/test_sparse_matrix_builder.py- 3 tests passpython build_calibration_matrix.pyfrom the local_area_calibration directory🤖 Generated with Claude Code