MerQuR+Statistics Glossary

Statistics Glossary → Delta method

Delta methodconcept

A way of approximating the standard error of a function of estimated parameters by using a first-order Taylor expansion of that function.

A model returns coefficients and their standard errors, but the quantity of interest is often a transformation of those coefficients: an odds ratio (exp(β)), a ratio of two coefficients, a predicted probability or a marginal effect. The delta method represents the function by a linear approximation around the estimate: with one parameter, Var(g(θ̂)) ≈ [g'(θ̂)]² · Var(θ̂). With several parameters, the vector of derivatives (the gradient) is combined with the coefficients' covariance matrix.

The method relies on large-sample theory; it works well when the estimator is approximately normal and the function is not strongly curved near the estimate. In small samples, or with highly nonlinear transformations (e.g. ratios whose denominator is close to zero), the resulting symmetric confidence interval can be misleading; bootstrap, profile likelihood or Fieller's method are then more reliable. The delta method is not an estimation method in itself but a tool for carrying the uncertainty of an existing estimate onto another scale.

An everyday example

Suppose you know that a shop's average number of customers per day is 50 with a standard error of 2, and you assume weekly turnover is proportional to the square of that number. The delta method gives the uncertainty of 50 squared, 2500, as roughly 2 · 50 · 2 = 200: a small wobble in the mean produces a wobble about 2 · 50 times as large in its square.

In MerQur

It is used to compute standard errors of transformed quantities in Logistic Regression, GLMM, Structural Equation Modeling, Item Response Theory, Weighted Frequency, Nonlinear Regression, Parametric Survival (AFT) and the Heckman sample-selection models.

Analyses where this concept appears

Logistic RegressionGeneralized Linear Mixed Models (GLMM)Structural Equation Modeling (SEM)Item Response Theory (IRT / Rasch)Weighted Frequency (Survey Freq)Nonlinear RegressionParametric Survival (AFT)Sample Selection and Truncated Regression (Heckman)

Related concepts

Standard errorAverage marginal effect (AME)