Interactive demo · Probability and Statistical Inference · Chapter 7

Plug-in estimators and their standard errors

A statistical functional T(F) is estimated by plugging in the empirical cdf: T(F̂n). Often T(F̂n) ≈ N(T(F), τ²/n), where τ² = E LF(X)² is the variance of the influence function LF, the effect of one observation at x on T; plugging F̂n into τ² gives the estimated standard error sê and the interval T(F̂n) ± 1.96 sê. Draw many samples from a known F and see where this works (means, medians, correlations) and where it does not yet work at realistic n (variances and, above all, skewness of skewed data).

Functional T(F)

Population F

Sample size n

Number of samples

Try these

Sample r uses its own seeded generator. "nerve data as F" and "Old Faithful as F" draw with replacement from the 799 nerve times and the 272 eruption–waiting pairs, so T(F) is known exactly. Quantiles use the book's definition F̂n−1(p) = inf{x : F̂n(x) ≥ p}; their sê needs the density at the quantile, estimated by a Gaussian kernel with the normal-reference bandwidth. The nerve data have atoms, so the quantile formula for τ² does not apply to that population.

In the lecture notes