The Bonferroni Confidence Rectangle c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 1 / 23 Suppose y = Xβ + ε, ε ∼ N(0, σ 2 I). Let c01 β, . . . , c0m β be m estimable functions. c01 β . . We want to find an m-dimensional rectangle that contains . with c0m β probability at least 1 − α. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 2 / 23 We may prefer a rectangular confidence region to an ellipsoidal region for ease of interpretation. α2−α3 (−1.6,8.6) (10.4,5.4) (6,1) (0,0) α1−α2 (1.6,−3.4) c Copyright 2012 Dan Nettleton (Iowa State University) (13.6,−6.6) Statistics 611 3 / 23 Let Ej (α) denote the event q q c0j β ∈ c0j β̂ − tn−r,α/2 σ̂ 2 c0j (X0 X)− cj , c0j β̂ + tn−r,α/2 σ̂ 2 c0j (X0 X)− cj . From our previous result, we know P[Ej (α)] = 1 − α c Copyright 2012 Dan Nettleton (Iowa State University) ∀ j = 1, . . . , m. Statistics 611 4 / 23 Note that P(∩m j=1 Ej (α)) ≤ 1 − α with strict inequality in most cases. (An exception: c1 = · · · = cm ). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 5 / 23 We will show that P(∩m j=1 Ej (α/m)) ≥ 1 − α. ∴ (l1 , u1 ) × · · · × (lm , um ) is a rectangular confidence region with the desired coverage probability, where q α lj = c0j β̂ − tn−r, 2m σ̂ 2 c0j (X0 X)− cj q α uj = c0j β̂ + tn−r, 2m σ̂ 2 c0j (X0 X)− cj c Copyright 2012 Dan Nettleton (Iowa State University) and ∀ j = 1, . . . , m. Statistics 611 6 / 23 We say the Bonferroni intervals (lj , uj ) j = 1, . . . , m have simultaneous coverage probability at least 1 − α ∵ P[c0j β ∈ (lj , uj ) c Copyright 2012 Dan Nettleton (Iowa State University) ∀ j = 1, . . . , m] ≥ 1 − α. Statistics 611 7 / 23 A collection of subsets of a sample space S is called a sigma algebra, denoted B, if (i) ∅ ∈ B (ii) A ∈ B =⇒ Ac ∈ B (iii) A1 , A2 , . . . ∈ B =⇒ ∪∞ j=1 Aj ∈ B. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 8 / 23 Given a sample space S and an associated sigma algebra B, a probability function with domain B satisfies (i) P(A) ≥ 0 ∀A∈B (ii) P(S) = 1 (iii) A1 , A2 , . . . ∈ B pairwise disjoint =⇒ P(∪∞ j=1 Aj ) = c Copyright 2012 Dan Nettleton (Iowa State University) P∞ j=1 P(Aj ). Statistics 611 9 / 23 It follows that ∀ set A ∈ B, P(A) + P(Ac ) = P(A ∪ Ac ) = P(S) = 1. ∴ ∀ A ∈ B, P(A) = 1 − P(Ac ). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 10 / 23 Boole’s inequality: For any sets A1 , . . . , Am ∈ B, P(∪m j=1 Aj ) ≤ c Copyright 2012 Dan Nettleton (Iowa State University) m X P(Aj ). j=1 Statistics 611 11 / 23 Proof: Let B1 = A1 B2 = Ac1 ∩ A2 B3 = Ac1 ∩ Ac2 ∩ A3 .. . Bm = Ac1 ∩ · · · ∩ Acm−1 ∩ Am . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 12 / 23 Then Bj ⊆ Aj ∀ j = 1, . . . , m. B1 , . . . , Bm are pairwise disjoint, and m ∪m j=1 Bj = ∪j=1 Aj . Thus, m P(∪m j=1 Aj ) = P(∪j=1 Bj ) m X = P(Bj ) c Copyright 2012 Dan Nettleton (Iowa State University) j=1 ≤ m X P(Aj ). j=1 Statistics 611 13 / 23 Bonferroni’s Inequality: For any sets A1 , . . . , Am ∈ B P(∩m j=1 Aj ) ≥1− m X P(Acj ). j=1 Prove this inequality using basic results from probability theory. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 14 / 23 Proof: m c P(∩m j=1 Aj ) = 1 − P((∩j=1 Aj ) ) c Copyright 2012 Dan Nettleton (Iowa State University) c = 1 − P(∪m j=1 Aj ) m X ≥ 1− P(Acj ). j=1 Statistics 611 15 / 23 Now prove that P[∩m j=1 Ej (α/m)] ≥ 1 − α. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 16 / 23 Proof: P[∩m j=1 Ej (α/m)] ≥ 1 − c Copyright 2012 Dan Nettleton (Iowa State University) = 1− = 1− = 1− m X j=1 m X j=1 m X j=1 m X P[Ej (α/m)c ] (1 − P[Ej (α/m)]) (1 − (1 − α/m)) α/m j=1 = 1 − mα/m = 1 − α. Statistics 611 17 / 23 Under what circumstances does the Bonferroni confidence rectangle c01 β . . (l1 , u1 ) × · · · × (lm , um ) give exact 1 − α coverage of . ? c0m β c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 18 / 23 Looking back at the proof of Bonferroni’s inequality, we have P(∩m j=1 Aj ) =1− m X P(Acj ) j=1 iff c P(∪m j=1 Aj ) = c Copyright 2012 Dan Nettleton (Iowa State University) m X P(Acj ). j=1 Statistics 611 19 / 23 c P(∪m j=1 Aj ) ⇐⇒ P(Aci = ∩ m X P(Acj ) j=1 Acj ) = 0 ∀ i 6= j. Thus, we have P[∩m j=1 Ej (α/m)] = 1 − α ⇐⇒ P[Ei (α/m)c ∩ Ej (α/m)c ] = 0 ∀ i 6= j. This says that the Bonferroni confidence rectangle is exact iff the failure of any one interval to cover its c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 20 / 23 target parameter implies that all other intervals cover their target parameters with probability 1; i.e., c0j β ∈ / (lj , uj ) =⇒ c0i β ∈ (li , ui ) ∀ i 6= j. Thus, in practice P[∩m j=1 Ej (α/m)] > 1 − α. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 21 / 23 Note that if E1 (α/m), . . . , Em (α/m) are independent events, then P[∩m j=1 Ej (α/m)] = c Copyright 2012 Dan Nettleton (Iowa State University) = m Y P[Ej (α/m)] j=1 m Y (1 − α/m) = (1 − α/m)m j=1 > 1−α for m > 1. Statistics 611 22 / 23 Under independence, 1/m P[∩m )] = j=1 Ej (1 − (1 − α) c Copyright 2012 Dan Nettleton (Iowa State University) m Y P[Ej (1 − (1 − α)1/m )] j=1 = [(1 − α)1/m ]m = 1 − α. Statistics 611 23 / 23