[DresdenGman]

Closes #607

Problem

In IV2SLS (and the other IV estimators sharing the same base class), the weighted arrays _wy, _wx, and _wz were always constructed by multiplying the input data by the square root of the weights. That multiplication promotes the arrays to float64 and allocates full-size copies, even when every weight is 1 and the weighted data is identical to the input — including the default unweighted case, where the weights are populated with 1.0 internally. For large low-dtype inputs, this causes the memory blow-up reported in #607: full-size float64 copies are materialized unconditionally before estimation begins.

Change

linearmodels/iv/model.py: when all weights are 1, _wy, _wx, and _wz now alias the original arrays instead of allocating multiplied float64 copies, preserving input dtypes and avoiding the extra full-size allocations. (Weights are normalized to have mean 1.0 before these arrays are built, so any constant weight vector reaches this check as all ones.) Any non-constant weight vector takes the existing materializing path unchanged.

Multiplying by a weight of exactly 1 is exact, so for float64 and integer inputs the estimation results are unchanged. One deliberate consequence: float32 inputs now stay in float32 through the weighting step instead of being silently promoted to float64, so a float32 fit agrees with the float64 fit of the same data within float32 precision rather than exactly (the tests below use rtol 1e-6 for parameters and 1e-5 for the covariance).

Scope note: #607 also mentioned PanelOLS. PanelOLS does not unconditionally materialize weighted copies — it already branches on np.any(weights != 1.0) to take a fast path for unweighted data (the same idiom used here) — so this PR is scoped to the IV models, where the unconditional copy occurred.

Tests

9 new test functions in linearmodels/tests/iv/test_model.py:

• Unit weights alias the original arrays (no copy, dtypes preserved); non-constant weights still materialize.
• Results with no weights and with constant weights are identical (parameters and covariance).
• A float32 fit with all-ones float32 weights matches the float64 fit of the same data within the tolerances above; an integer-input fit with constant integer weights matches the float64 results exactly.
• Equivalence against an independently constructed reference: 2SLS computed by hand from materialized float64 arrays using numpy.linalg.pinv agrees with the new path to 1e-8 (measured deviation ~1e-13).
• Boundary behavior is unchanged from baseline: a single non-unit weight takes the materializing path, with the stored arrays scaled by the square root of the weights; NaN weights follow the existing missing-value contract (the observation is dropped, verified against a manual row deletion).

Verification

IV test suite (same command before and after): baseline 9521 passed / 2400 skipped / 0 failed → with this change 9530 passed / 2400 skipped / 0 failed. The delta is exactly the 9 new test functions; the affected subset (test_model, test_weight_config, test_data, test_gmm) is 140 passed.

Panel suite re-run: 3680 passed (PanelOLS code untouched).

An A/B comparison of 20 result items between unweighted and all-ones-weighted runs of the new code is bit-identical; equivalence to the previous materialized path is covered by the independent-reference test above.

Peak memory on the reproduction case: 21.2 MiB → 10.7 MiB.

Known limitation (performance only): weights that are constant only up to floating-point rounding (within 1 ulp of a constant after normalization) still take the materializing path. Results are correct; the shortcut is simply not taken.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.68%. Comparing base (5b85e0d) to head (8f81bfe).

Additional details and impacted files

@@           Coverage Diff            @@
##             main     #862    +/-   ##
========================================
  Coverage   99.68%   99.68%            
========================================
  Files         199      199            
  Lines       37875    37998   +123     
  Branches     2436     2437     +1     
========================================
+ Hits        37756    37879   +123     
  Misses         51       51            
  Partials       68       68            

Flag	Coverage Δ
adder	99.67% <100.00%> (+<0.01%)	⬆️
subtractor	99.67% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.