MerQuR+Statistics Glossary

Statistics Glossary → Taylor linearization and replication variance methods

Taylor linearization and replication variance methodsconcept

The two main ways of obtaining standard errors of weighted estimates in complex sample designs while accounting for stratification and clustering.

Statistics such as ratios and regression coefficients are nonlinear functions of weighted totals. Taylor linearization turns such a function into a linear approximation using a first-order Taylor expansion; a 'linearized value' is computed for each observation, and its variance is found with the formula appropriate to the design (from the variability between primary sampling units within strata, with a finite population correction where needed). It is fast and is the default in most software.

Replication methods instead recompute the estimate on many modified copies of the data and estimate the variance from the spread among these copies. The jackknife (JK1, JKn) drops primary sampling units one at a time; BRR requires exactly two primary units per stratum and uses balanced half-samples; Fay's method softens BRR by shrinking weights by a factor ρ instead of setting them to zero; the Rao-Wu bootstrap resamples primary units within strata. Replication is handy for complex statistics and when the data provider publishes only replicate weights instead of design information. For non-smooth statistics such as the median the jackknife is unreliable, and BRR, Fay or the bootstrap is preferred.

An everyday example

A school survey divides the country into 20 regions, selects a few schools in each region and then pupils within each school. Because pupils in the same school resemble each other, 4,000 pupils carry less information than 4,000 independent people. Taylor linearization, or a jackknife that drops schools one at a time, accounts for this similarity and gives a realistic standard error.

In MerQur

'Variance method' (Taylor (default), Jackknife (JK1 / JKn), BRR, Fay BRR, Bootstrap (Rao-Wu), Supplied replicate weights) and the related fields (Fay coefficient, number of replicates, scale) in Weighted Mean, Frequency, Total, Linear and Logistic Regression; Survey-PHREG offers the same options.

Analyses where this concept appears

Weighted Mean (Survey Means)Weighted Frequency (Survey Freq)Weighted Total (Survey Total)Weighted Linear Regression (Survey Regression)Weighted Logistic Regression (Survey Logistic Regression)Survey-PHREG (Weighted Cox)

Related concepts

JackknifeBootstrap (resampling)Design effect (DEFF)