research paper

L₀ regularization for subnational microsimulation calibration

The paper evaluates target-informed sparse sampling for pruning Populace's full U.S. candidate microsimulation dataset while preserving calibration accuracy.

anchor result

Post-L₀ refit keeps 57,240 of 337,704 candidate records with a 4.74% Populace loss.

method

Hard Concrete gates select records while calibration weights are fit to the target surface.

source

Public populace candidate data, administrative targets, and reproduction code.