Compatible simultaneous lower confidence bounds for the Holm procedure and other Bonferroni based closed tests

Transcription

1 Compatible simultaneous lower confidence bounds for the Holm procedure and other Bonferroni based closed tests K. Strassburger 1, F. Bretz 2 1 Institute of Biometrics & Epidemiology German Diabetes Center, Düsseldorf 2 Novartis Pharma AG Tuesday, July 10/ MCP 2007, Vienna

2 Notation Parameter of interest: Data: Thresholds : ϑ = (ϑ 1,..., ϑ k ) Θ IR k X X (δ 1,..., δ k ) Θ Hypotheses: H i : ϑ i δ i, i I = {1,..., k} Corresponding p-values : p i = p i (X), i I

3 The Closure Principle Construct level-α tests for all intersection hypotheses H J = j J H j with = J I Reject H i, iff all H J with i J can be rejected. Such closed test procedures control the familywise error rate (FWER) at level α.

4 A Class (B) of Bonferroni Based Closed Tests Basis: Local α-levels α i (J) 0, i I, J I with α i (J) α, J I. i I An intersection hypothesis H J is rejected, if min j J (p j α j (J)) 0. Bonferroni s inequality ensures the control of the FWER at level α. Closure test: Reject all those hypotheses H i with max J:i J I min j J (p j α j (J)) 0

5 Simultaneous Compatible Confidence Bounds Let R = R(x) be the index set of rejected hypotheses H i R = {i I : max min(p j α j (J)) 0} J I:i J j J Problem: Find lower confidence bounds L i = L i (X), i I, such that P ϑ (L i (X) < ϑ i, for all i I) 1 α and x X : i R(x) L i (x) δ i.

6 Examples (Single-step) Bonferroni test R = {i : p i α/k} Local α levels: α i (J) = α/k, i I, J I (Step-down) Bonferroni-Holm test Ordered p-values: p (1) p (2) p (k) R = {(i) : max j=1,...,i (k j + 1)p (j) α} { α/ J if i J, Local α levels: α i (J) = 0 else. Simultaneous confidence bounds?

7 Construction Principles Bofinger (1987). Step down procedures for comparison with a control. Aust. J. Stat. 29, Finner (1994). Testing Multiple Hypotheses: General Theory, Specific Problems, and Relationships to Other Multiple Decision Procedures. Habilitationsschrift. Universität Trier, FB IV Mathematik/Statistik. Finner & Straßburger (2006). On equivalence with the best in k-sample models. J. Am. Stat. Assoc. 101, Finner & Straßburger (2007). Step-up related simultaneous confidence intervals for MCC and MCB. Biometrical J 49, 40-5 Hayter & Hsu (1994). On the relationship between stepwise decision procedures and confidence sets. Journal of the American Statistical Association 89, Stefansson, Kim & Hsu (1988). On confidence sets in multiple comparisons. Statistical decision theory and related topics IV, Pap. 4th Purdue Symp.; West Lafayette/Indiana 1986, Vol. 2,

8 Basis: p-value Compatible Local Confidence Bounds For the construction of simultaneous confidence bounds one needs local, lower confidence bounds L i (α ) = L i (X, α ) fulfilling: For all i I and α [0, α] it holds, P ϑ (L i (X, α ) < ϑ i ) 1 α, x X : L i (x, α ) δ i p i (x) α.

9 Compatible Confidence Bounds for Bonferroni Based Closure Tests Compatible confidence bounds for multiple tests in B can be constructed as follows: min J J :i J max{δ i, L i (α i(j))} if i R L i = min J J :i J L i (α i(j)) if i I \ R, where J J J = {J I : min j J (p j α j (J)) > 0}. intersection hypothesis H J can not be rejected

10 The Subclass (SCB) of so Called Short-Cut Procedures Hommel, Bretz und Maurer (Stat. Med. 2007) showed that there exists an easy implementable short-cut for all procedures in B with α i (J) α i (J ) for all i I and all J, J with i J J I. Examples: (weighted) Bonferroni-Holm test, Test of apriori ordered hypotheses, Fallback and Gatekeeping procedures.

11 Simplified Confidence Bounds for Short-Cut Procedures Simultaneous compatible confidence bounds for multiple tests in SCB reduce to: min J I\R max{δ i, L i (α i(j))} if i R, L i = L i (α i(i \ R)) if i I \ R.

12 Further Reduction for α-exhaustive Multiple Tests Necessary condition for maximizing the number of rejected hypotheses: α i (J) = 0, for all i, J with i J and = J I. Reduced confidence bounds: max{δ i, L i (α i( )) if R = I, L i = δ i if i R I, L i (α i(i \ R)) if i I \ R. In case of i R I the lower bound L i only reflects the test decision ϑ i > δ i.

13 A Numerical Example Observations: X = (X 1, X 2,..., X 5 ), X i N(ϑ i, 1) Hypotheses: H i : ϑ i 0, i = 1,..., 5 FWER: α = 5% p-values: p i = 1 Φ(X i ) Loc. conf. bounds: L i (α ) = X i u 1 α i x i p i

14 (Single-step) Bonferroni Test Local levels: α i (J) = 0.01, i I, J I i x i p i α/k L i Rejected hypotheses: R = {1, 2} Confidence bounds: L i = x i u 1 α/5 = x i 2.326

15 (Step-down) Bonferroni-Holm Test Local levels: α i (J) = 0.05/ J, i I, J I i x i p i α/i L i Rejected hypotheses: R = {1, 2, 3} Confidence bounds: α i (I \ R) = α/ I \ R = α/2 { 0 if i R L i = x i u 1 α/2 = x i if i I \ R

16 Comparison of Single-step & Step-down Confidence Bounds i Single-step L i Step-down L i Step-down versus single-step bounds / Assets and drawbacks: (+) H 3 can be rejected by the step-down procedure. (+) (moderate) larger confidence bounds for ϑ 3, ϑ 4 and ϑ 5 ( ) much smaller confidence bounds for ϑ 1 and ϑ 2