microcalibrate

Calibrate survey weights by providing your estimate matrix, target benchmarks, and initial weights to microcalibrate, which uses L0 regularization and PyTorch optimization to adjust weights until sample totals align with known population totals like income or employment.

L0 regularization penalizes non-zero weights during calibration, encouraging many weights to become exactly zero. This sparsity shrinks your active dataset—often by 60% or more—while maintaining accuracy for faster PolicyEngine simulations.

Yes, microcalibrate includes automatic hyperparameter tuning that optimizes the sparsity penalty (lambda) to balance target matching accuracy against dataset sparsity, eliminating manual lambda selection.

Use calibration with sparsity when you have known population benchmarks, need faster simulations without sacrificing representativeness, and can tolerate some weights becoming zero. It's ideal for PolicyEngine workflows requiring both accuracy and computational efficiency.

Provide three inputs: your survey weights (initial adjustment factors), an estimate matrix (survey characteristics aligned to population dimensions), and target values (known population totals for each dimension like total income or employment count).

Yes, microcalibrate accepts multiple population targets simultaneously—income, employment, demographics, or any measurable benchmarks—and optimizes weights to match all targets while enforcing sparsity across the entire dataset.