Home Statistical Dictionary Taylor Series Linearization

Taylor Series Linearization

A variance-estimation technique for complex survey designs that approximates a nonlinear statistic by a first-order Taylor series expansion around the true population value, then computes the variance of that linear approximation using standard design-based variance formulas.

Sampling Theory & Survey MethodsAdvancedFree

In Plain English

Taylor series linearization solves a specific problem: standard design-based variance formulas (like the ones for a simple sample mean or the Horvitz-Thompson total) are built for linear statistics, sums and weighted averages, but many statistics survey analysts actually want, ratios, regression coefficients, correlation coefficients, are nonlinear functions of the data. Linearization's trick is to use calculus, specifically a first-order Taylor series expansion, to approximate the nonlinear statistic by a linear stand-in near the true population value, and then apply the well-understood linear variance formulas to that stand-in instead. It's the classical, formula-based alternative to the more computational replicate-weight methods (jackknife, BRR, bootstrap), and it's what most standard survey-analysis software packages use by default under the hood.

Definition

Taylor series linearization estimates the sampling variance of a nonlinear statistic $\hat{θ} = g (\hat{Y}_{1}, \dots, \hat{Y}_{k})$ , a smooth function of one or more estimated totals or means, by taking a first-order Taylor expansion of $g$ around the true population values, $\hat{θ} \approx θ + \sum_{j = 1}^{k} \frac{\partial g}{\partial Y _{j}} (\hat{Y}_{j} - Y_{j})$ , and then applying the standard design-based variance formula for a linear combination of totals (already well understood under the survey's design) to this linear approximation, $Var (\hat{θ}) \approx d^{⊤} Var (\hat{Y}) d$ , where $d$ is the vector of partial derivatives evaluated at the sample estimates. Taylor series linearization is the classical, closed-form approach to complex survey variance estimation, requiring the analyst (or, in practice, statistical software) to derive the relevant partial derivatives for each new type of statistic, in contrast to replicate-weight methods (jackknife, BRR, bootstrap), which sidestep explicit derivative calculations entirely by repeatedly recomputing the statistic on perturbed versions of the sample; linearization is generally faster to compute for simple, well-studied statistics with known derivative formulas already built into standard survey software, while replicate weights are often preferred for complex or novel statistics where deriving the necessary derivatives would be impractical.

Formula

Var (\hat{θ}) \approx d^{⊤} Var (\hat{Y}) d, d_{j} = \frac{\partial g}{\partial Y _{j}}

Notation

g (\cdot)

—The nonlinear function defining the statistic of interest

d

—Vector of partial derivatives of g, evaluated at the sample estimates

Var (\hat{Y})

—Design-based variance-covariance matrix of the underlying estimated totals

Properties

Taylor series linearization is the default variance-estimation method built into most general-purpose survey analysis software (including the standard 'svy' commands and packages used across statistical software), because the necessary derivative formulas have already been worked out and implemented for the most common statistics, means, totals, ratios, regression coefficients.
For genuinely novel or complex statistics, quantiles, certain robust estimators, some machine learning outputs, deriving the required partial derivatives for Taylor linearization can be difficult or effectively impossible, which is exactly the situation where replicate-weight methods are preferred, since they estimate variance purely by recomputation, without needing any closed-form derivative.
Taylor series linearization and replicate-weight methods (jackknife, BRR, bootstrap) generally agree closely for well-behaved statistics under standard survey designs, they are considered largely interchangeable approaches to the same underlying complex-survey variance estimation problem, and the choice between them in practice is often driven by software availability and the specific statistic being computed rather than by a fundamental preference for one over the other.

At a Glance

Mechanismfirst-order Taylor expansion turns a nonlinear statistic into a linear approximation

Then appliesstandard design-based variance formulas to that linear approximation

Requiresderiving partial derivatives of the statistic with respect to underlying totals

Alternative toreplicate-weight methods (jackknife, BRR, bootstrap), which avoid derivatives entirely

Last updated September 8, 2026← Back to the Dictionary