Home Statistical Dictionary Calibration Estimator

Calibration Estimator

A general framework for adjusting survey design weights so that the weighted sample exactly reproduces known population totals for a set of auxiliary variables, while staying as numerically close as possible to the original design weights -- encompassing post-stratification, raking, and GREG as special cases.

Sampling Theory & Survey MethodsAdvancedFree

In Plain English

Calibration estimation is the umbrella framework behind a whole family of survey weighting techniques whose common goal is: adjust the sample's weights just enough that, when you total up known variables like age group or region using those weights, you get exactly the numbers you already know are true for the population, while changing the original design weights as little as possible in the process. Post-stratification, raking, and the generalized regression estimator are all specific recipes for doing this; calibration estimation is the general mathematical framework that shows they're all instances of the same underlying idea, minimizing a distance measure between old and new weights subject to matching known population totals exactly.

Definition

Calibration estimation adjusts a survey's initial design weights $d_{i} = 1/ π_{i}$ to new calibrated weights $w_{i}$ that exactly satisfy a set of calibration constraints, $\sum_{i \in s} w_{i} x_{i} = X$ , matching known population totals $X$ for a vector of auxiliary variables $x_{i}$ , while minimizing a chosen distance function $\sum_{i} G (w_{i}, d_{i})$ between the new and original weights, subject to that constraint. Different choices of the distance function $G$ recover different familiar estimators as special cases: a chi-squared distance yields the generalized regression (GREG) estimator, an exponential (raking) distance yields the iterative proportional fitting used in raking, and calibrating to a categorical cross-classification exactly recovers post-stratification; this unifying framework, developed by Deville and Sarndal (1992), is why calibration estimation is now the standard conceptual lens through which official statistics agencies design their weighting adjustments, rather than treating post-stratification, raking, and regression estimation as unrelated ad hoc techniques.

Formula

i \in s \sum w_{i} x_{i} = X, minimizing i \sum G (w_{i}, d_{i})

Notation

d_{i} = 1/ π_{i}

—Original design weight of unit i

w_{i}

—Calibrated (adjusted) weight of unit i

G (w_{i}, d_{i})

—Chosen distance function between calibrated and design weights

Properties

Calibration estimation's key insight is that post-stratification, raking, and the regression estimator, historically developed separately and taught as distinct techniques, are all solutions to the same general constrained-minimization problem, differing only in the choice of distance function G between the calibrated and design weights.
Calibrated weights simultaneously serve two goals often in tension elsewhere in statistics, they exactly reproduce known population totals (removing coverage and nonresponse bias correlated with the calibration variables) while staying as close as possible to the original design weights (preserving the efficiency properties of the original probability design), the distance function G is what governs that trade-off.
Because calibration weights are constrained to match known totals exactly, calibration estimation is only as effective as the calibration variables chosen, it corrects bias correlated with the variables actually included in the calibration, but provides no guarantee against bias on unmeasured characteristics not part of the calibration model.

At a Glance

Goaladjust weights to exactly match known population totals for auxiliary variables

Unifiespost-stratification, raking, and GREG as special cases of one framework

Key referenceDeville and Sarndal (1992)

Constraintminimizes distance from original design weights, subject to exact totals matching

Last updated September 8, 2026← Back to the Dictionary