Interactive demo · Nonparametric Statistical Inference · Chapter 2

Influence functions and the delta method

Move a small mass ε of the empirical distribution to a point x: the plug-in estimate changes by about ε L̂(x), where L̂ is the empirical influence function. The same function gives the standard error: sê = √(ΣL̂(Xi)2)/n, the nonparametric delta method. Pick a data set and a functional, then drag x and ε. For the median the difference quotient does not settle down (the empirical cdf has no density), which is why (2.36) needs a density estimate.

Data

Functional T

Quantile level p

Contamination point x

Mass ε (log10)

Try these

T((1 − ε)F̂n + εδx) is computed exactly as the functional of a weighted sample (weights (1 − ε)/n on the data and ε at x). Quantiles use the book's definition inf{x : F̂n(x) ≥ p}; their standard error uses a Gaussian kernel density estimate at the sample quantile with the rule-of-thumb bandwidth 0.9 min(s, IQR/1.34) n−1/5.

In the lecture notes