strategies / calibration
strategy 02 · calibration
A candidate frame — its support built, before deployment — carries design weights that reflect the survey it came from, not the administrative totals a model needs to match. Calibration reweights it against thousands of hierarchical targets at once, by gradient descent on a capped relative-error loss, with hard bounds on how far any single weight can move.
The numbers below come from the same release registry the homepage reads — populace-us-2024-5da5a95, calibrated against 3,704 administrative targets spanning Census, IRS, BEA, and state agencies. Calibration adjusts weights only: no target value is ever written into a record.
populace-us-2024-5da5a95 · calibration.json, generated 2026-06-15. 82 of 3,704 targets still miss by more than 50% after calibration — mostly thin families with few observations (rent-and-royalty net losses is the extreme case: 28 targets, 0% within 10%). The per-family breakdown below is the same table the release registry exposes; nothing here is rounded or selected for effect.
Each target's contribution to the loss is capped, so a handful of targets with poor donor coverage cannot dominate the gradient and starve the rest of the surface of attention. Thousands of targets are fit jointly, not sequentially.
Design weights can move only up to a fixed multiple of their starting value (50× on the current release). Unbounded calibrators — including several classical methods — can send an individual weight arbitrarily high in pursuit of a thin target; the bound trades a small amount of in-fit accuracy for a hard cap on any one household's leverage.
The target surface spans Census population counts, IRS SOI income and deduction totals, BEA wages, and state-level adjusted gross income and program administrative totals — fit as one surface, not reconciled in separate passes.
classical-calibration comparison (raking, GREG, entropy balancing, bounded chi-square) is the calibration-paper's plan — not yet run; see the paper page
Calibration-target attainment (the numbers above) is one of the two validation currencies detailed on the evaluation strategy page — administrative targets hit, in-fit and held out. The other currency, support fidelity against survey holdouts under the popdgp harness, is calibration-blind by construction: it scores whether the weighted population still looks like the surveys that observe it, independent of whether calibration ever ran. Reading both pages together is the honest picture; neither one alone is.