Interactive demo · Nonparametric Statistical Inference · Chapter 7

Confidence balls for many Normal means

From Zi = θi + σnεi, i = 1, …, 100, a 95% confidence ball for θ can have a fixed radius (the χ² ball, radius σn√q, about σn√n) or a random one. Baraud's ball tests which subspaces Sd = {θ : θj = 0 for j > d} are compatible with the data and uses the smallest radius ρS among them; the pivotal ball centres at the SURE-chosen projection with radius from a Normal approximation. Panel (a) shows the radii ρS (Figure 7.2); panel (b) shows the radii of the three balls over many simulated data sets, with their coverage.

Mean vector θ and noise σn

γ for Baraud's ball

Simulated data sets

Try these

α = 0.05 and αS = α/(n + 1) for each of the 101 nested subspaces. The constants ρS²/σn² of (7.57) and the test levels c(S) were computed once with exact noncentral χ² probabilities and are loaded from a data file; everything else runs here. Data set r uses its own seeded generator (seed + 7919 r), Zi = θi + σnεi in index order. The pivotal radius uses the positive-part fix of the book.

In the lecture notes