Nonparametric Dispersion and Equality Tests Nathaniel E. Helwig AssistantProfessorofPsychologyandStatistics UniversityofMinnesota(TwinCities) Updated04-Jan-2017 NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide1 Copyright Copyright (cid:13)c 2017 by Nathaniel E. Helwig NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide2 Outline of Notes 1) Dispersion (Ansari-Bradley): 3) Equality (Kolmogorov-Smirnov): Overview Overview Procedure Procedure Example 1 Example 3 2) Dispersion/Location (Lepage): Overview Procedure Example 2 NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide3 DispersionTest(Ansari-Bradley) Dispersion Test (Ansari-Bradley) NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide4 DispersionTest(Ansari-Bradley) Overview Problem(s) of Interest Like two-sample location problem, we have N = m+n observations X ,...,X are iid random sample from population 1 1 m Y ,...,Y are iid random sample from population 2 1 n We want to make inferences about difference in distributions Let F and F denote distributions of populations 1 and 2 1 2 Null hypothesis is same distribution (F (z) = F (z) for all z) 1 2 Using the location-scale parameter model, we have F (z) = G([z −θ ]/η ) and F (z) = G([z −θ ]/η ) 1 1 1 2 2 2 θ and η are median and scale parameters for population j j j NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide5 DispersionTest(Ansari-Bradley) Overview Assumptions Within sample independence assumption X ,...,X are iid random sample from population 1 1 m Y ,...,Y are iid random sample from population 2 1 n Between sample independence assumption Samples {X}m and {Y}n are mutually independent i i=1 i i=1 Continuity assumption: both F and F are continuous distributions 1 2 Location assumption: θ = θ or θ and θ are known 1 2 1 2 NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide6 DispersionTest(Ansari-Bradley) Procedure Parameter of Interest and Hypothesis Parameter of interest is the ratio of the variances: V(X) γ2 = V(Y) so that γ2 = 1 whenever V(X) = V(Y). The null hypothesis about γ2 is H : γ2 = 1 0 and we could have one of three alternative hypotheses: One-Sided Upper-Tail: H : γ2 > 1 1 One-Sided Lower-Tail: H : γ2 < 1 1 Two-Sided: H : γ2 (cid:54)= 1 1 NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide7 DispersionTest(Ansari-Bradley) Procedure Test Statistic Let {Z }N denote the order statistics for the combined sample, and (k) k=1 assign rank scores (cid:26) 1,2,3,..., N, N,...,3,2,1 if N is even R∗ = 2 2 k 1,2,3,..., N−1, N+1, N−1,...,3,2,1 if N is odd 2 2 2 to the combined sample {Z }N . (k) k=1 The Ansari-Bradley test statistic C is defined as n (cid:88) C = R j j=1 where R is the assigned rank score of Y for j = 1,...,n j j NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide8 DispersionTest(Ansari-Bradley) Procedure Distribution of Test Statistic under H 0 Under H all (cid:0)N(cid:1) arrangements of Y-ranks occur with equal probability 0 n Given (N,n), calculate C for all (cid:0)N(cid:1) possible outcomes n Each outcome has probability 1/(cid:0)N(cid:1) under H n 0 Example null distribution with m = 3 and n = 2: Y-ranks C ProbabilityunderH 0 1,2 3 1/10 1,3 4 1/10 1,4 3 1/10 1,5 2 1/10 2,3 5 1/10 Note: thereare(cid:0)5(cid:1)=10possibilities 2 2,4 4 1/10 2,5 3 1/10 3,4 5 1/10 3,5 4 1/10 4,5 3 1/10 NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide9 DispersionTest(Ansari-Bradley) Procedure Hypothesis Testing One-Sided Upper Tail Test: H : γ2 = 1 versus H : γ2 > 1 0 1 Reject H if C ≥ c where P(C > c ) = α 0 α α One-Sided Lower Tail Test: H : γ2 = 1 versus H : γ2 < 1 0 1 Reject H if C ≤ [c −1] 0 1−α Two-Sided Test: H : γ2 = 1 versus H : γ2 (cid:54)= 1 0 1 Reject H if C ≥ c or C ≤ [c −1] 0 α/2 1−α/2 NathanielE.Helwig (UofMinnesota) NonparametricDispersionandEqualityTests Updated04-Jan-2017 : Slide10
Description: