Interactive demo · Nonparametric Statistical Inference · Chapter 1

Pointwise and uniform coverage

For n Bernoulli(p) trials the Wald interval p̂ ± zα/2√(p̂(1−p̂)/n) has coverage tending to 1 − α for each p, but for every n there are values of p near 0 or 1 where it almost never covers: the worst-case coverage does not tend to 1 − α. That is the difference between a pointwise and a uniform asymptotic confidence interval. Hoeffding's interval p̂ ± √(log(2/α)/(2n)) covers with probability at least 1 − α for every p and every n. All numbers are exact sums over the Binomial law.

Sample size n

True proportion p (bottom panels)

Level α

Tolerance δ

Try these

The worst coverage over p is taken over the grid 0.001, 0.002, ..., 0.009, 0.01, 0.02, ..., 0.99, 0.991, ..., 0.999. n(δ) is the smallest N such that the coverage is at least 1 − α − δ for every n from N to 3000; the dots in the bottom right panel are precomputed for α = 0.05 and δ = 0.01.

In the lecture notes