individual prediction regions for multivariate ... · better detect abnormal alues.v recentl,y wang...

19
HAL Id: hal-00856736 https://hal.archives-ouvertes.fr/hal-00856736v3 Submitted on 12 May 2014 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Individual prediction regions for multivariate longitudinal data with small samples Didier Concordet, Rémi Servien To cite this version: Didier Concordet, Rémi Servien. Individual prediction regions for multivariate longitudinal data with small samples. Biometrics, Wiley, 2014, 70 (3), pp.629-638. hal-00856736v3

Upload: others

Post on 03-Oct-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

HAL Id: hal-00856736https://hal.archives-ouvertes.fr/hal-00856736v3

Submitted on 12 May 2014

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.

Individual prediction regions for multivariatelongitudinal data with small samples

Didier Concordet, Rémi Servien

To cite this version:Didier Concordet, Rémi Servien. Individual prediction regions for multivariate longitudinal data withsmall samples. Biometrics, Wiley, 2014, 70 (3), pp.629-638. hal-00856736v3

Page 2: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♥ ♣rt♦♥ r♦♥s ♦r ♠trt ♦♥t♥ tt s♠ s♠♣s

❱ ❯♥rsté ♦♦s ❯ ♦①♠sr ♥tr ♥ ♦♦ ♦①♦♦② ♦♦s

r♠sr♥t♦♦s♥rr

strt

♦♦♣ s ♠♦r ♥ ♠♦r s ♥ ♠♥♦♣♥ ♦♥tr♦ t♦ ♥t② ♥♦r♠rsts ♥ ♥ ♥ rr♥t② ♦♦♣s r ♠♦st② rr ♦t r ② r s♥ rr♥ ♥trs tt ♦♥t♥ t s ♦sr ♥ 100(1 − α) ♦t②♥♦♣ ♥s srt♦♥s ♦ t ♦t♦♥ ♦ t rs ♦r t♠♥ s♠♣ ♦ N t②♥♦♣ ♥s ♦s ts rr♥ ♥trs t♦ ♥③ ② t♥ ♥t♦ ♦♥t t ♣♦ss t ♦ ♦rs ♥ s♦♠ ♣r♦s♦srt♦♥s ♦ ts rs ♦t♥ ♥ t ♥ s t②♥♦♣ ♦r r ts ♥③ ♥trs s♦ ♦♥t♥ 100(1 − α) ♦ ♦srs ♦♠♣t t ♣r♦s ♦sr s ♥ ts ♥ ♥r ♠t♦s t♦ ts ♥trs r t t② ♦ ♦♥② r ② r ♦♦♣tr t ♣♦ss ♦rrt♦♥s ♦r t♠ t♥ t♠ ♥ ts rt ♣r♦♣♦s ♥r ♠t♦ t♦ ♦♥t② ♦♦♣ sr ♦rrt rs ♦r t♠ s ♠t♦♦♦② rs ♦♥ ♠trt ♥r ♠① ts ♠♦ ❲ rst ♣r♦ ♠t♦ t♦st♠t t ♠♦s ♣r♠trs ♥ ♥ s②♠♣t♦t r♠♦r N r ♥♦ t♥ r (1−α) ♥③ ♣rt♦♥ r♦♥ ♦♠t♠s t s♠♣ s③ N s♥♦t r ♥♦ ♦r t s②♠♣t♦t r♠♦r t♦ rs♦♥ ♣rt♦♥ r♦♥t s ♦r ts rs♦♥ ♣r♦♣♦s ♥ ♦♠♣r tr r♥t ♣rt♦♥ r♦♥s tts♦ ttr ♦r s♠ N ♥② t ♦ ♠t♦♦♦② s strt ② t♦♦♣ ♦ ♥② ♥s♥② ♥ ts

②♦rs ♦rrt ♦r rt ♦♥t♥ ♦♦♣ trt ♠① ss♥♠♦ P♥ st♠t♦r Prt♦♥ r♦♥ r♥ ♥trs

Page 3: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♥tr♦t♦♥

♦♥t♥ ♦♦♣ ♦ ♦♦ rs s ♠♦r ♥ ♠♦r s ♥ ♣r♥t ♠♥t ♦♥ssts ♦ ♠♦♥t♦r♥ t ♠rrs ♦ ♠♣♦rt♥t ♥t♦♥s ♦r t r② tt♦♥ ♦ s♦②♣r♦rss sss t s♥ ♣s ♦r ①♠♣ t ♣r♦stt s♣ ♥t♥P s s t♦ tt ♣r♦stt ♥r ♥ ♠♥ s♠ ♥ ♦ ♦♦♣ s s②st♠t②♦♥ t t♥rs s♥ tr t ♥ t t♦ tt t ♥♥♥ ♦ ♦st② ♥s♣♦rt ②♥ ♦r tts ♥t♦♣♥ ♦♥tr♦ t♦rts tr② t♦ ♥r③ t s ♦ ♦♦ ♣ss♣♦rt ♦♥ssts ♦ ♦♥t♥ ♦♦♣ ♦ s♦♠ ♥♦♥♦s sst♥s♦ ♥trst ♥ ♦rr t♦ tt ♥♦r♠ rt♦♥s ♥ ♥ ♥ ♦tts t ❩♦r③♦♥ ♦ss st♥r ♠t♦ ♦ ♦♥ ts ♦♦♣s s t♦ s t s♦ rr♥ ♥trs s ♥trs ♦♥t♥ ① ♣r♥t s② ♦ ♠sr♠♥ts tt ♥ ♦sr ♥ t② ♥s ♦r ts ♠t♦ srs r♦♠ sr s rst t♦s ♥♦t s ♥ ♥♦r♠t♦♥ t② ♥ ♥ ①tr♠ s ♦tst rr♥ ♥tr ♦r s♦♠ ♦tr ♥s s ♥s t rr♥ ♥trr ♣t♦♦ ♦♥ ts ♥trs r t ♥ ♥ ♥rt r♠♦r r② r t♦t t♥ ♥t♦ ♦♥t t ♣♦ss ♦rrt♦♥s t♥ t♠ ♥② t♦s ♥♦t ♦♥t ♦r tr ♦t♦♥ ♦r t♠ t♥ ♥ ♥ ♥ rr♥ ♥trs ♦r ♣rt♦♥ ♥trs ♠tt ts ② ♦♥ t♦♥strt♦♥ ♦ rr♥ ♥ s ♦♥ t ♦sr s ♥ t② ♥♥ t♥ ♥t♦ ♦♥t s♦♠ ♦rs s s s① trtr ♦♥ ts sts ♣♥t ♥ t s ♠t♦♦♦② s t♦ s ♥r♥♦♥♥r ♠① ts ♠♦s ♥ ♥s ❱r ♥ ♦♥rs ♥ ♥ t♥♥ ♥ ts♠♦s t ♦srt♦♥s r s② ss♠ t♦ ♥♣♥♥t ♦♥t♦♥ t♦ t ♥s♣ ♣r♠trs ♦♠♣♦♥ s②♠♠tr② ss♠♣t♦♥ ♦ ♦r ♥♦ t ♦♣♠♥t♦ r ♠t♦s t♦ tt ♥♦r♠ rt♦♥s ♦ ♦♥t♥ rs s r♠♥♠t ♦tts t ♣r♦♣♦s ②s♥ ♣♣r♦ t♦ ♦♠♥ ♣♦♣t♦♥r♠ts ♥ ♥s trs♦s rtss ts ♠t♦ s t ♥ ♥ ♥rtr♠♦r rs ♦♦♣ s s② ♣r♦r♠ ♦♥ sr ♠rrs ② ♦♥sq♥t ♥ rr♥ ♥trs r t ♦♥ ♠rr ♥♣♥♥t② ♥tt♦♥ sststt ♥ r♦♥s s♥ s♠t♥♦s ♥♦r♠t♦♥ ♦♥ ♦rrt rs ♦ ♣ t♦ttr tt ♥♦r♠ s ♥t② ❲♥ ♥ ♥ ♣r♦♣♦s ♠t♦ t♦ ♣rt♦♥ r♦♥s ② s p ♦rr t♦rrss ♣r♦ss t♦ ♠♦ t t♦♦rrt♦♥ ♦ r t t♠ t ♦rrt♦♥s t♥ r♥t rs s ss♠ t♦ ①♦r t♠ s r♦♥s r t s♥ t s②♠♣t♦t strt♦♥ ♦ sttst ♥ ♦tr♦rs t strt♦♥ ♦ t sttsts s ♦♠♣t ss♠♥ tt t ♠♦ ♣r♠trs r♥♦♥ st♠ts r s t♦ ♦♠♣t t ❲ ts ♣♥ ♠t♦ s s② t♦ s tsr② ♥tr ♦s ♥♦t r♥t ♥ ①t ♦r rt ♦r t ♣rt♦♥ r♦♥ s t♦s ♥♦t ♦♥t ♦r t ♠♣rs♦♥ ♦ t ♣r♠tr st♠ts s ♥ r ♣r♦♠♥ t s♠♣ s③ s s♠ r♥♦rs♥ ♥ ♦① ♥ ts s ss♠♥ ttt s②♠♣t♦t r♠♦r ♦s ♦ t♦ ♦♥sr ♣♣r♦①♠t♦♥ rr♦rs r♦rs♣ tt♥t♦♥ s t♦ ♣ t♦ ts ♣r♦♠ t♦ ♦♥tr♦ t r ♦r rt ♦ t t

Page 4: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♣rt♦♥ r♦♥♥ ts rt ♣r♦♣♦s t♦ ♥ ♥ ♣rt♦♥ r♦♥ r♦♠ ♣r♦s ♦srt♦♥s♦ ts rs rr ♦t ♥ t s♠ ♥ ♥ ♠♦ ♣r♠tr st♠ts ♦srt♦♥s ♦t♥ ♥ ♥ ♥ r ss♠ t♦ ♦rrt ♦r t♠ ♦rrt♦♥t♥ r X1 t t♠ t1 ♥ r X2 t t♠ t2 s ♥♦t ss♠ t♦ qt♦ t ♦rrt♦♥ t♥ X1 t t♠ t2 ♥ X2 t t♠ t1. s s t♦ ② strtrt♦♦rrt♦♥s tt ♥♥♦t rt② st♠t ② ♦♥♥t♦♥ ♠t♦s t ❳♣r♦r ♥ ♦r t ♥♠ ♣ r♦r ♣r♦♣♦s s♣ st♠t♦♥♠t♦ ♠♦ s t♦ ts ♣rt♦♥ r♦♥s s t ♦♥ t♦♥ ♥ t st♠t♦♥♦ ts ♣r♠trs s ♣r♦r♠ ♥ t♦♥ ♣rt♦♥ r♦♥s r t♥ t ♥ t♦♥ s♥ ♣♥ ♣♣r♦ r r♥t ♦rrt♦♥s ♦ t s②♠♣t♦t ♦♥♥ r♦♥r ♣r♦♣♦s ♥ ♦♠♣r ♥ t♦♥ s ♦rrt♦♥s ♠ t ♦rrt♥ t ♣♥st♠t♦♥ ♦ t ♣rt♦♥ r♦♥ rst t♦ ♦♠ r♦♠ t ❯ ♥ ❱♦♥ ♦♥s t t tr s ♥s♣r ② t ♦r ♦r♥ ♥ ♦♥s t r tst ♦♥ ♥② ♥s♥② ♦♦♣ ♥ts s t♥ trt ♥ t♦♥ ♥② sss♦♥ s ♣r♦ ♥ t♦♥

♠♦

t s ♥♦t Xi = [Xi1 : · · · : Xir] t ♠sr♠♥ts ♣r♦r♠ ♥ t ith ♥ ♦ s♠♣ ♦ s③ N t♦rXij ♦♥t♥s t ni ♦srt♦♥s rr ♦t ♦r t♠ ♦r t jth

r ♦r ♣rs② Xijk s t ♦sr ♦r t ith ♥ ♦r t jth rt t♠ tik ❲t♦t ♦ss ♦ ♥rt② ♥ ss♠ tt ti1 ≤ ti2 ≤ . . . ≤ tini

♦t tt t rs r s♣♣♦s t♦ ♠sr t t s♠ t♠ ♦r ♥ ♥ t t♠♠srs ♠② r r♦♠ ♦♥ ♥ t♦ ♥♦tr ❲ ss♠ tt ♣ t♦ ♠♦♥♦t♦♥tr♥s♦r♠t♦♥

Xi = Biβ +TiΦi + ζi

r Bi ♥ Ti r ♥♦♥ r♥ ♦rt ♠trs ♦ ♠♥s♦♥s ni × p ♥ ni × qrs♣t② β = [β1 : · · · : βr] s p×r ♠tr① ♦ ♣r♠trs s t♦ sr t ♣♦♣t♦♥♠♥ Φi = [Φi1 : · · · : Φir] ♥ ζi = [ζi1 : · · · : ζir] r rs♣t② q× r ♥ ni× r ♠trs♦ ♥♦sr ss♥ r♥♦♠ ts r♥ ♦ t ♦♠♣♦♥♥ts ♦ t r♥♦♠♠tr① ζi s ss♠ t♦ ② strtr

cov (ζijk, ζij′k′) = Σjj′ρtik−tik′jj′ k > k′ ♥ cov (ζijk, ζij′k′) = Σjj′ρ

tik′−tikj′j k < k′

r ρjj′ ∈ [0, 1] ♥ Σjj′ = αjj′σjσj′ ♥♠rs αjj = 1, ∀j ∈ 1, . . . , r αjk = αkj ∈[−1, 1] ∀j 6= k ∈ 1, . . . , r ♥ ρjk ∈ [0, 1] ∀j, k ∈ 1, . . . , r σj r♣rs♥ts t st♥rt♦♥ ♦ t jth r t ♠sr♠♥t t♠ ♦rrt♦♥ t♥ ζijk ♥ ζij′k′s ss♠ t♦ ρtik−tik′

jj′ ♦r k > k′ ♥ ρtik′−tikj′j ♦r k < k′ s ♠♥s tt ♦ ♥♦t ss♠

tt t ♦rrt♦♥ t♥ t jth r ♥ ζi ♠sr t t♠ k ♥ t j′th r ♥

Page 5: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

ζi ♠sr t t♠ k′ s t s♠ s t ♦rrt♦♥ t♥ t jth r ♥ ζi ♠srt t♠ k′ ♥ t j′th r ♥ ζi ♠sr t t♠ k ♠♦r r♥ t t♣♣r ♦ ❲♥ ♥ ♥ s tt t② ss♠ tt t ♦srt♦♥ t♠s tik r q②s♣ ♥tr ♥♠rs ♥ tt ♦r j ♥ j′ cov (ζijk, ζij′k′) = Σjj′ρ|t−t′| r ρ|t−t′| st ♦rrt♦♥ ♦ ♥ t♦rrss ♣r♦ss ♦ ♦rr p ♠tr① ♦ t ♦r♥ ♦ t ζi s r♥♦r♥ ♠tr① s t s s②♠♠tr ♥ ♣♦st♥t ♠tr① s t r♦♥r ♥ r ♣r♦ts ♦ t♦ ♣♦sts ♠trst ts ♠♦ rt♥ s s② t♦ ♥rst♥ ts ♠t♠♥s♦♥ ♥tr ♦s♥♦t tt t st♠t♦♥ ♦ ♣r♠trs ♥ t strt♦♥ ♥t♦♥ ♦ Φi ♥ ζi s rrt ts ♠♦ ♥ t♦r r♠♦r t♦ tt rtr st♠t♦♥s t s ♥ψi = vec(Φi) t t♦r ♦t♥ ② st♥ t ♦♠♥s ♦ Φi ♦♠♥s ❲ ss♠ ttψi ∼

iidN(0,Ω) r♥ ♠tr① Ω = [ωjm]jm s ♦♣rtt♦♥ t q × q r♥

♠trs ωjm = cov(Φij,Φim)♠r② t t♥ st rr♦r ζi ♥ st♦r ♦♠♥s ♥t♦ t♦r εi =vec(ζi) ∼ N(0,Λi(ρ,Σ)) ♠tr① Λi(ρ,Σ) ♥ rtt♥ s D

−1i R

−1i D

−1i r

D−1i = diag(σ1, . . . , σ1, σ2, . . . , σ2, σr, . . . , σr) t σj r♣t ni t♠s s t s

ss♠ t♦ ♦♥st♥t ♦r t♠ ♠tr① R−1i (ρ) s ♦♣rtt♦♥ t ni × ni

♠trs ωijk t

ωijk(ρ) = (♦rr (ζijl, ζikf ))l,f∈1,...,ni= αjkρ

|tif−til|

jk

♥ αjj = 1. ♠tr① ωijk(ρ) ♦♥t♥s t ♦rrt♦♥ t♥ t jth ♥ kth r tt r♥t s♠♣♥ t♠s εis r ss♠ ♠t② ♥♣♥♥t ♥ ♥♣♥♥t ♦ t ψis t s ♥Yi = vec(Xi) Ai = 1r ⊗Bi ♥ Zi = 1r ⊗Ti ❯s♥ ts ♥♦tt♦♥s t ♠♦ ♥ rrtt♥ s

Yi = Aiθ + Ziψi + εi

r θ = vec(β) s ♠♦ ♠② ♣♣r t♦ st♥r ♥r ♠① t ♠♦ ♦s♣r♠tr ξ = (θ,Ω,Σ,ρ) ∈ Ξ ♥ s② st♠t s♥ st♥r sttst s♦tr♦r t ♦r♥ ♠trs ♦ ts ♠♦ r ② strtr ♥ tr st♠t♦♥♥s r ♦♣♠♥t s ♦♥ ♥ t♦♥ ss♠ tt nw ♦srt♦♥s ♦ t r rs r t t♠s (t1, . . . , tnw) ♥ ♥♥ t s ♥♦t U ∈ R

r×1 t tr s tt ♦sr t t♠ tu > tnw

♦r t r rs ♥ ts ♥ ♥ ❲ ss♠ tt

(W′U

′)′= Aθ + Zψ + ε

r Z = (Z′w Z

′u)

′ A = (A′w A

′u)

′ r ♥♦♥ ♠trs ♥ ε = (ε′w ε′u)

′ r♥♦♠ ♠tr①(W′

U′)′ s ss♠ t♦ ♥♣♥♥t ♦ t Yis ❲ r ♦♦♥ ♦r r♦♥ Rα

ξ (W) s♦tt P

(U ∈ Rα

ξ (W)∣∣W)= 1− α.

♦ s r♦♥ ♥ t♦ t♥s r♥♦♠ s♠♣ ♦ ♥s (Yi)i∈1,...,N

tt ♥s t ♣♦♣t♦♥ ♣r♠trs ξ t♦ st♠t ♥ s♦♠ ♦srt♦♥s ♣r♦r♠♥ t ♥ ♦ ♥trst W ❲ ♣r♦ ♥ tr st♣s rst r♦♥ Rα

ξ (W)

Page 6: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

② ss♠♥ tt ξ s ♥♦♥ s♦♥② ♣♥ t st♠t ξ ♦ ξ ♦t♥ s♥ ts♠♣ (Yi)i∈1,...,N ♥t♦ Rα

ξ (W) t♦ t Rαξ(W) ♦rs s t st♠t ξ s

r♥♦♠ r ts ♣♥ st♠t♦r ♦s ♥♦t r♥t ♦r ♦ 1 − α. s s trs♦♥ ② ♣r♦♣♦s ♥ ♦♠♣r tr r♥t ♦rrt♦♥s ♦r t st♠t ♣rt♦♥r♦♥ s♥ ts ♣♥ st♠t♦r ♥ t tr st♣

st♠t♦♥ ♦ ♣r♠trs

t s ♥♦t Λi = Λi(ρ,Σ)) = var (εi) ss♠ ♦r tt t ♣r♠tr ξ s ♥♦♥♥ q t♦ ξ0 = (θ0,Ω0,Σ0,ρ0) ♥

(Y i −Aiθ0

Ψi

)∼ N

(0,

(ZiΩ0Z

′i +Λi ZiΩ0

Ω0Z′i Ω0

))

s♦ ② r ♠♠ ❩♥ ♦t♥

mi,0= Eξ0

(Ψi|Y i −Aiθ0) = Eξ0(Ψi|Y i) = Ω0Z′i [ZiΩ0Z

′i +Λi]

−1[Y i −Aiθ0]

V i,0= Vξ0(Ψi|Y i) = Ω0 −Ω0Z

′i [ZiΩ0Z

′i +Λi]

−1ZiΩ0.

♦ st♠t t ♣r♠tr ξ ♣r♦♣♦s t♦ s t ♦rt♠ ♠♣str t t trt♦♥ k ξk−1 s ♦rt♠ tr♥ts t♥ t♦st♣s t st♣ ♦♠♣ts Q(ξ, ξk−1) = E

[logL (Y 1, . . . ,Y n; ξ) |Y 1, . . . ,Y n; ξk−1

]♥

t st♣ s♦s ξk = arg supξQ(ξ, ξk−1). s t Y i r ♥♣♥♥t Q(ξ, ξk−1) =∑i Eξk−1

[logL(Y i,Ψi)|Y i] .♠♥♥ tt

−2 logL(Y i,Ψi) = (Y i −Aiθ −ZiΨi)′Λ

−1i (Y i −Aiθ −ZiΨi)+log |Λi|+Ψ

′iΩ

−1Ψi+log |Ω|,

ts ♦t♥

−2Q(ξ, ξk−1) =N∑

i=1

tr(Λ−1i ZiV i,k−1Z

′i) + (Y i −Aiθ −Zimi,k−1)

′Λ

−1i (Y i −Aiθ −Zimi,k−1)

+ tr(Ω

−1V i,k1

)+m′

i,k−1Ω−1mi,k−1 + log |Λi|+ log |Ω|.

♠♥♠③t♦♥ ♦ −2Q(ξ, ξk−1) t rs♣t t♦ Ω ♥ θ s rs♣t② t♦

Ωk =1

N

N∑

i=1

[V i,k−1 +mi,k−1m

′i,k−1

],

θk =

(N∑

i=1

A′iΛ

−1i Ai

)−1 N∑

i=1

(A′

iΛ−1i (Y i −Zimi,k−1)

).

Page 7: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

t s ♥♦t U i = Y i −Aiθ −Zimi,0 ♥ P i = ZiV i,0Z′i +U iU

′i ❲ ♥t t♦ ♦♠♣t

argminΣ,ρ

−2Q(Λ,Λ0) = argminΣ,ρ

tr

(N∑

i=1

Λ−1i P i

)+

N∑

i=1

log |Λi|

tt sargmin

Σ,ρ−2Q(Λ,Λ0) = argmin

Σ,ρtr

(N∑

i=1

DiRiDiP i

)− 2

N∑

i=1

log |Di|.

t s ♥♦t J j,i =δD−1

i

δσ−1

j

−2∂Q

∂σ−1j

= 2

[N∑

i=1

tr(DiFji )−

ni

σ−1j

]

r F ji = P iJ j,iRi s nir×nir ♦♣rtt♦♥ ♠tr① t ni×ni ♠trs

(f

ji

)k,l∈1,...,r

.

s

−2δQ

δσ−1j

= 2r∑

k=1

N∑

i=1

(σ−1k tr

[(f

ji

)kk

]− ni

σ−1j

)

♥ t r♠♥s t♦ s♦ t s②st♠ ♦ qt♦♥s δQ

δσ−1

j

= 0, j = 1, . . . , r t♦ t t ♠①♠♠ ♦

Q t rs♣t t♦ σ−11 , . . . , σ−1

r ♦s② ts s②st♠ s ♥♦t ♥r ♦r ♥ ♥♦ttt t rts s x′B = 1/x r

x′ =(σ−11 , σ−1

2 , . . . , σ−1r

), 1/x = (σ1, σ2, . . . , σr) ♥ B =

(∑Ni=1 tr

[(f

ji

)kk

]∑N

i=1 ni

)

j,k∈1,...,r

,

♥ s♦ trt② s♥ t rrs♦♥ ♦r♠ xmB = 1/xm+1s ①♣♥ ♥ t♦♥ αjk ♥ ρjk s ♥ [−1, 1] ∀j, k ∈ 1, . . . , r s ♥ s②♦♣t♠③ ♥♠r② Q ♦♥ ts ♥tr t♦ ♦t♥ t st♠ts ♦ αjk ♥ ρjk ♦♦ ♦ ♦ strt♥ s s♣s ♣ t ♦♥r♥ ♦ ts ♦rt♠ s strt♥s ♥ s② r r♦♠ ♥rt ♥②ss ♥②ss ♦ t jth r s♥♠♦ s βj ♥ Φij ζij ♦r i ∈ 1, . . . , N ♠♣r r♥ ♦ Φij ♥

s s strt♥ ♦ Ω βj ♥ s ♦r βj trt♥ s ♦r ♥tr♥r♥ ♦♠♣♦♥♥ts ♥ ♦t♥ s ♦♦s

Σjj′ =N∑

i=1

ni∑

k=1

ζijkζij′k/(niN), σj =√

Σjj, αjj′ = Σjj′/(σjσj′).

strt♥ s ♦r ρjj′ r ♦♠♣t ② ♠①♠③♥ t ♦♦ ♦t♥ ② ss♠♥tt ζi r strt ♦r♥ t♦ N (0,Λi(ρ,Σ)) r♥ ts ♠①♠③t♦♥ Σ sst t♦ ts ♥t ♦♠♣tr t♠ ♥ ♦r ♣r♠tr st♠t♦♥ s ss t♥ ♦♥s♦♥ s♥ ♥ ♦r♥r② ♣t♦♣

Page 8: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♥ ♣rt♦♥ r♦♥s

♠♥ tt ss♠ tt ♦srt♦♥s W ♦r t r rs r ♥ ♥♥ ❲ r ♦♥ t♦ ♣rt♦♥ r♦♥ ♦r t ♥①t ♦srt♦♥ U ♦r ts ♥♥ r♦♠ t ♠♦ ♥ ♥

U = Auθ +Zuψu + εu.

♦r♥ t♦ t♦♥ ss♠ ♥ ts st♦♥ tt t ♠♦ ♣r♠trs r ♥♦♥❲ ♥♦t

E = vec(εw, εu) ∼ N

(0;

(Λw(ρ,Σ) Mwu(ρ,Σ)′

Mwu(ρ,Σ) Λu(ρ,Σ)

))

r Λw(ρ,Σ) ♥ Λu(ρ,Σ) r ♥ ♥ t rst st♦♥ ♥ Mwu(ρ,Σ) s r × (rnw)♠tr① t

Mwu(ρ,Σ) = ♦ (εw; εu) =

♦(ε1u; ε

1k=1,...,nw

). . . ♦

(ε1u; ε

rk=1,...,nw

) . . .

♦(εru; ε

1k=1,...,nw

). . . ♦

(εru; ε

rk=1,...,nw

)

r εiu s t it tr♠ ♦ εu ♥ εjk=1,...,nws nw ♠♥s♦♥ t♦r ♦r r j ♥

♥ W ♥

♦(εiu; ε

jk=1,...,nw

)=(Σijρ

|tu−t1|ij , . . . ,Σijρ

|tu−tnw |ij

).

❯s♥ ts ♥♦tt♦♥s ♥ r ♠♠ ♦t♥ t ♦♦♥ ♣r♦♣♦st♦♥

Pr♦♣♦st♦♥ t α ♥② r ♥♠r ♥ [0; 1] ♥ χ2r,1−α t 1 − α q♥t ♦

sqr strt♦♥ t r rs ♦ r♦♠ t s ♦♥sr t t♦r m(ξ,W ) ♥

t ♠tr① V (ξ) ♥ ②

m(ξ,W ) = Auθ + (ZuΩZ′w +Mwu(ρ,Σ)) (ZwΩZ

′w +Λw(ρ,Σ))

−1(W −Awθ) ,

V (ξ) = (ZuΩZ′u +Λu(ρ,Σ))

− (ZuΩZ′w +Mwu(ρ,Σ)) (ZwΩZ

′w +Λw(ρ,Σ))

−1(ZuΩZ

′w +Mwu(ρ,Σ))

′.

(1− α) ♣rt♦♥ r♦♥ ♦ U ♦♥t♦♥② t♦ W s t st

S =u ∈ R

r; ‖V (ξ)−1/2 (u−m(ξ,W )) ‖2 ≤ χ2r,1−α

r V (ξ)−1/2 s t ♥rs ♦ t ♦s② tr♥s♦r♠t♦♥ ♦ V (ξ).

❲♥ r > 1 t ♣rt♦♥ r♦♥ ♦r U s ts ♥ ♣s♦ ♥tr ♦♥ m(ξ,W ) s

♣s♦ ♥rts t♦ t ♥tr[m(ξ,W )− τ(1−α/2)

√V (ξ);m(ξ,W ) + τ(1−α/2)

√V (ξ)

]

r τ(1−α/2) s t (1−α/2) q♥t ♦ t st♥r ss♥ strt♦♥ ♥ ♦♥ ♥ts

Page 9: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

t♦ ♣rt t ♥①t U ♦ s♥ r r = 1♥ ts s ρ = 0 ♥ ss♠♥ tt tr s ♥♦ ♦r t jt ♦srt♦♥ ♥ t it

♥ rtsXij = Yij = θ + ψi + εij

t ψi ∼ N (0, ω2) ♥ εij ∼ N (0, σ2) . ❯s♥ r ♦♠♣♠♥t t 100(1−α) ♣rt♦♥♥tr ♦r t tr ♥ k − 1 ♦srt♦♥s r r② ♥ ♥ ♥s t ♦♦♥ ①♣rss♦♥

1 + γ2(k − 1)+

γ2(k − 1)

1 + γ2(k − 1)W k−1 − τ(1−α/2)

√1 + γ2k

1 + γ2(k − 1)σ2,

θ

1 + γ2(k − 1)+

γ2(k − 1)

1 + γ2(k − 1)W k−1 + τ(1−α/2)

√1 + γ2k

1 + γ2(k − 1)σ2

],

r γ = ω/σ ♥ W k−1 s t r ♦ t k − 1 ♦srt♦♥s ♦t tt γ♠srs t ♥t ♦ t ♥③t♦♥ ♦♠♣r t♦ t s rr♥ ♥tr tt s♥ ♣r ♥ ❲♥ γ s t ♣rt♦♥ ♥trs ♦s t♦

[W ± τ(1−α/2)σ

]♥ t ♥③t♦♥ s ♥ ❲♥ k = 1 ♥

♥♦ ♦srt♦♥ s ♦r t ♥ ♥ t ♥tr ♥rts t♦ t s[θ ± τ(1−α/2)

√σ2 + ω2

] s ts ♠♦ ♦s ♥♦t ♥ ♥② ♦rrt♦♥ t♥ ♦srt♦♥s

t♥ ♥ t t♠ ♥tr t♥ ♥ ♦srt♦♥ ♥ tr ♦s ♥♦t♠♦② t t ♦ t ♣rt♦♥ ♥tr ♣♥ st♠t♦r ♦ S s S ♦t♥ ② r♣♥ ♥ qt♦♥ t ♣r♠tr ξ ② tsst♠t ♦t♥ ♣rt♦♥ r♦♥ s t♦ r♥t ♥ ①t ♦r ♦ (1−α) s♦rs ♥ t s♠♣ s③ n s ♦ ♥ t♦ s♦♠ ♠♣rs st♠t ♦r ξ ♦ ♣r♥tts ♣r♦♠ ♣r♦♣♦s r♥t strts t♦ ♦rrt t ♦r ♦ ♦r ♣rt♦♥ r♦♥

P♥ ♦rrt♦♥s

t s ♥ KW (ξ0, ξ) = ||V −1/2(ξ) (U −m(ξ,W )) ||2 r (U |W ) ∼N (m(ξ0,W ),V (ξ0)) t s ♦♦s tt ♦r ξ KW (ξ, ξ) s r♥♦♠ rstrt ♦♥t♦♥② t♦W ♦r♥ t♦ χ2 strt♦♥ s s ♥♦ ♦♥r t s ♦rKW (ξ0, ξ) t ξ n ♦♥sst♥t st♠t ♦ ξ t ♥♦♥ strt♦♥ L s (W ′U ′)

s ss♠ t♦ ♥♣♥♥t ♦ t ♦srt♦♥s Y i (W ,U ) ♥ ξ r ♥♣♥♥t ❲

♥t t♦ st♠t xα,ξ0,ξ,Ws tt P

(KW

(ξ0, ξ

)≤ xα,ξ0,ξ,W

∣∣∣W)

= 1 − α ❲ ♥

♥ t ♦♦♥ sst♦♥s tr r♥t ♦rrt♦♥s ♥♦t ② x1α,ξ0,ξ,W

x2α,ξ0,ξ,W

♥ x3α,ξ0,ξ,W

♦rt ♣r♦♣rts ♦ ts ♦rrt♦♥s r ♥♦t sss r

♥ rr t ♥trst rr t♦ t rr♥s r♥ t ❯ ♥ ❱♦♥ ♦♥s t

Page 10: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

rst ♠t♦

r♦ st♠t♦♥ x4α,ξ,ξ

♦ xα,ξ0,ξ,W♥ ♦t♥ ② ss♠♥ tt t t r r♥

r♦♠ strt♦♥ t ♣r♠trs ξ ♥st ♦ ξ0 s KW

(ξ, ξ)

s strt ♦♥

t♦♥② t♦ W ♦r♥ t♦ χ2r strt♦♥ x4

α,ξ,ξ= χ2

r,1−α t (1 − α)q♥t ♦ ts

strt♦♥ s ξ s t ♠①♠♠♦♦ st♠t ♦ ξ0 t s rst ♦rrr② ❯ ♥ ❱♦♥ ♦♥s t ♥

1− α+ (ξ0,W ) = Pξ0

(KW

(ξ0, ξ

)≤ x4

α,ξ,ξ|W)= 1− α+ δ (ξ0,W ) /N +O

(N−3/2

).

t ♦♦s tt

Pξ0

(KW

(ξ0, ξ

)≤ x4

α−δ(ξ0,W )/N,ξ,ξ|W)= 1− α +O

(N−3/2

).

❲ ts ♦t♥ tr ♦rr r② ② ♦rrt♥ α ② δ(ξ0,W )/N t

δ(ξ0,W )/N =(1− α+(ξ0,W )

)− (1− α) +O

(N−3/2

).

s ξ0 s ♥♥♦♥ ts ♦rrt♥ tr♠ s s♦ ♥♥♦♥ ♦r ♥ st♠t t ②

s♥ ♣r♠tr ♦♦tstr♣ ♠t♦ ♥ t♦(1− α+(ξ,W )

) ♦

δ(ξ,W )/N =(1− α+(ξ,W )

)− (1− α) +O

(N−3/2

)

♥ s δ(ξ,W )− δ(ξ,W ) = O(1/√N) ♦t♥ δ(ξ,W )/N = δ(ξ,W )/N +O

(N−3/2

).

♦ t r♠♥s t♦ t 1 − α+(ξ,W ) ♠♥ tt ξ∗ s ♦♦tstr♣ s♠♣ ♦ ξ

s②♠♣t♦t② L

(ξ|ξ0

)≈ L

(ξ∗|ξ

) r ♦♣t♦♥s rst ♥♦

tt ξ − ξ0 ∼ N(0, I−1/2(ξ0)/N

)s ♥tr ♠t t♦r♠ ♥ s (ξ∗ − ξ)|ξ ∼

N(0, I−1/2(ξ)/N) r ξ∗ ♥ N(ξ, I−1/2(ξ)/N) strt♦♥ ♥ t s♦♥ ♦♣t♦♥ s♠t (Y ∗

i )1≤i≤N s♥ t strt♦♥ ♦ (Y i)1≤i≤N t ♣r♠tr ξ ♥ t ξ∗ ②st♠t♥ t ♣r♠tr ♦ t ♠♦ s♥ (Y ∗

i )1≤i≤N ♥st ♦ (Y i)1≤i≤N

❯s♥ qt♦♥s ♥ ♥ s KW

(ξ, ξ∗

)= || (V (ξ∗))−1/2 (U −m(ξ∗,W )) ||2 r

U ∼ N(m(lξ,W );V (ξ)

) st ♦ (ξ∗h)1≤h≤H ♥ tt

1− α+(ξ,W ) =1

H

H∑

h=1

1

L

L∑

l=1

1[||(V (ξ∗h))

−1/2(U l−m(ξ∗h,W ))||2≤x4

α,ξ,ξ

]

r H ♥ L r rs♣t② t ♦♦tstr♣ s♠♣ s③s ♦ (ξ∗) ♥ U ❲ ♦♥ tt

x1α,ξ0,ξ,W

= x4α−δ(ξ,W )/N,ξ,ξ

+O(N−3/2

).

Page 11: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♦♥ ♠t♦

t Fξ0,W (x) = P(KW (ξ0, ξ) ≤ x|W

) ❲ ♥t x2

α,ξ0,ξ,Ws tt Fξ0,W

(x2α,ξ0,ξ,W

)=

1− α t Fξ0,W s ♥♥♦♥ s s♦♥ ♥

1− α+(ξ0,W ) = Fξ0,W

(x4α,ξ,ξ

)= 1− α +

δ(ξ0,W )

N+O

(N−3/2

).

♦♥sq♥t② ♦t♥

x2α,ξ0,ξ,W

= F−1ξ0,W

(1− α) = F−1ξ0,W

1− α+(ξ0,W )− δ(ξ0,W )/N +O

(N−3/2

)

♦s s②♠♣t♦t ①♣♥s♦♥ s

x2α,ξ0,ξ,W

= F−1ξ0,W

1− α+(ξ0,W )

− δ(ξ0,W )

N[F−1

ξ0,W]′1− α+(ξ0,W )

+O

(N−3/2

).

t s s t♦

x2α,ξ0,ξ,W

= x4α,ξ,ξ

− δ(ξ0,W )

Nfξ0,W (x4α,ξ,ξ

)+O

(N−3/2

)

r fξ0,W s t ♣r♦t② ♥st② ♥t♦♥ ♦ KW

(ξ0, ξ

) ♦rs fξ0,W s ♥♥♦♥

t s t ♠♦ s rr ♦r x(fξ0,W (x)− f

ξ,W (x))= O (1/

√n) . ❲

s♥ ♥ t ♣r♦s st♦♥ tt KW

(ξ, ξ)s r♥♦♠ r strt ♦r♥

t♦ χ2 strt♦♥ ♦♥sq♥t② t ♣r♦t② ♥st② ♥t♦♥ ♦ t χ2 strt♦♥♥ ♣♣r♦①♠t fξ0,W ♥ ♣ ♥t♦ t♦

x2α,ξ0,ξ,W

= x4α,ξ,ξ

− δ(ξ0,W )

Nfξ,W (x4

α,ξ,ξ)+O

(N−3/2

).

t r♠♥s t♦ st♠t δ(ξ0,W ) tt ♥ ♦t♥ s t t ♣r♦s ♠t♦

r ♠t♦

rst t♦ ♠t♦s ♥ r s t♠t♦s tr ♠t♦ s ♥ ♣♣t♦♥ ♦ s♠♣ ♣r♠tr ♦♦tstr♣ ♠t♦ s ①♣♥ ♥ t rst ♠t♦ ♦r x

Fξ0,W (x) = P(KW

(ξ0, ξ

)≤ x|W

)≈ P

(KW

(ξ, ξ∗

)≤ x|W

).

r♦r s♥ ♥♦tt♦♥s ♦ ♣r♦s sst♦♥s Fξ0,W (x) ♥ st♠t ②

Fξ0,W (x) =1

H

H∑

h=1

1

L

L∑

l=1

1[||(V (ξ∗h))

−1/2(U l−m(ξ∗h,W ))||2≤x

].

Page 12: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♥♦t♥ F−1(y) = infx : F (x) ≥ y

x3α,ξ0,ξ,W

≈ F−1ξ0,W

(1− α).

s sss ♥ ♦♥s t ts ♥ s ♥ ♥ t Pr♦♣♦st♦♥ ♦ r♥ tt s♦s tt

P(KW (ξ0, ξ) ≤ x3

α,ξ0,ξ,W|W)= 1− α +O(N−2).

strt♦♥s

tst

t ♦♠ r♦♠ ♣r♦s♣t st② ♠ t t♥ rt♦♥s ♦r t♠ ♦ sr♦♠ rs ♥ t② ts s st② s rr ♥ t ♥s ♦ t ❱tr♥r②♦ s② r s ♥♠s ♦r t② ♥♠s ♦r str③t♦♥ ♦♦s②♦♥② ♦♥ s s t rs♦♥ ② ♦♥② N = 20 t② ts ♦ ♥ ♥ ♥ts st② ♠♥ rs ♦r r♥ ♦♦♣ r t r X1 t rt♥♥ X2 ♥ t♣r♦t♥ X3 r ♣♦tt ♥ r ♦r t 20 t② ts r s ♥♦ rs♦♥ t♦ t♥tt ts rs r ♥♦t st ♦r t♠ ♥ t② ts ②♥♦s t r ❯♥rt ♥②ss r ♣r♦r♠ ♥ t t ♦ t♠ s ♦♥ ♥♦t s♥♥tr② t t tr r ♠sr t♠s 0 3 6 12 ♥ 24 ♠♦♥ts tr ♥s♦♥ r♠♥♥ tr r s♠♣ ♦♥② ♦r t rst ♦r t♠s ♦t tt t ♥tr st②s ♣r♦r♠ ♦♥ t ♦ tr♥s♦r♠t♦♥ ♦ t rs s s♦ ♦r♥ t♦ t♦♥ ♣r♦♣♦s t ♦♦♥ ♠♦

X i = Biβ + T iΦi + ζi

r X i s ni × 3 ♠tr① t ni t ♥♠r ♦ ♦srt♦♥ ♦r t it t ♦r Bi ♥ T i r t♦rs ♦ ♥t ni s tt Bi = T i = (1, . . . , 1) β = (β1, β2, β3) ♥φi = (φi1, φi2, φi3) ♥ ζi s ni × 3 ♠tr① ♦r♥ t♦ t♦♥ ♥ ts ♣♣t♦♥ ni = 4 ♦r 5 r = 3 p = 1 q = 1 ♥ N = 20 s ♥♦ ♦r s① t ♠tr① Bi ♦s ♥♦t ♥♦r♣♦rt ♥② ♥♦r♠t♦♥ t ts ♥ ♦ ♥♦r♠t♦♥ ♥s② ♥srt ♥ ♦r ♠♦ s ♥ ♦tts t

st♠t♦♥ ♦ t ♣r♠trs

❯s♥ t ♠t♦ ♣r♦♣♦s ♥ t♦♥ st♠t t ♣r♠trs ♥ ♦♥ t ♦♦♥s σ1 = 0.083 σ2 = 0.062 σ3 = 0.032 θ1 = 2.054 θ2 = 4.834 θ3 = 4.368

α =

10.306 10.074 0.211 1

ρ =

0.411 0.001 0.0100.013 0.005 0.0100.009 0.896 0.005

♥ Ω =

0.0220.011 0.015−0.001 −0.001 0.003

Page 13: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♦ strt t s ♦ ts st♠ts r r st♠ts ♦ ♦rrt♦♥s t♥ r 2♥ 3 t t♠s t ♥ ♠♦♥ts ♥ t′ > t rr ♦t ♥ t s♠ ♥

ω23 =

1

ρ t′−t22 1

α23 α23ρt′−t32 1

α23ρt′−t23 α23 ρ t′−t

33 1

=

10.005t

′−t 10.211 0.211 ∗ 0.896t′−t 1

0.211 ∗ 0.010t′−t 0.211 0.005t′−t 1

.

♥ ts ①♠♣ t ♦rrt♦♥ t♥ t♦ sss ♠sr♠♥ts rr ♦t ♥ t s♠♥ s rtr ♦ ♦r ♣rt s t t′− t > 1 ♠♦♥t ♦r sr♣rs♥② t ♣♣rstt ♥♦ r s ♥ rr ♠rr t♥ t ♦trs t♦ tt ♥② ♥s♥② ♥ ♦tr♦rs tr s ♥♦ ♠♦r ♦rrt♦♥ t♥ t♦ r♥t rs t t♦ r♥t t♠ss rst ♦ ♥♦t ♥t♣t ❲t ts rst t ♥t ♦ t ♥③t♦♥♥ r♦② ♠sr ② t rt♦ γ = ω/σ s s q t♦ 1.8 2.0 ♥ 1.7♦r r rt♥♥ ♥ ♣r♦t♥ rs♣t② s ts rt♦s r rtr t♥ ♦♥ ♦♥ ♥①♣t t ♥③ r♦♥ t♦ ♥rr♦r t♥ t ♣♦♣t♦♥ ♦♥tr♣rt

♦♠♣rs♦♥ ♦ t tr ♦rrt♦♥s

s ♦r rst ♠♦tt♦♥ s t♦ ♣r♦♣♦s rr♥ r♦♥ ♦r rs ♠sr♥ ♥② ♥s♥② ♥ ts ♦♠♣r t tr ♣r♦♣♦s ♦rrt♦♥s ♥ ts ♦♥t①t ♦♦♠♣r t tr r♥t ♦rrt♦♥s ♦t♥ t t tr ♣r♦♣♦s ♠t♦s ♥t st♥r χ2 trs♦ x4

α,ξ,ξt♦t ♦rrt♦♥ rr ♦t s♠t♦♥ st②

♦ ♦ t st② s t♦ ♦♠♣r P(KW

(ξ0, ξ

)≤ x

∣∣∣W)

♦r x ♥ t stx1α,ξ0,ξ,W

,x2α,ξ0,ξ,W

,x3α,ξ0,ξ,W

,x4α,ξ,ξ

♦ rst ♥ t♦ ♥♦ ξ0 = (θ0,Ω0,Σ0,ρ0) ♥

t ξ0 s t ♠①♠♠ ♦♦ st♠t ♦r ξ ♥ t ts tst ❯s♥ ξ0

s♠t 500 ♥ tsts((Y k

i

)i=1,...,20

)k=1,...,500

, r(Y k

i

)s t t♦r ♦ s③ rni = 15

♦♥t♥♥ t s ♦ t tr rs t t s♠♣♥ t♠s ♦r t it t ♦ t kt

tst tr ♦rrt♦♥s ♥t t♦ ♦♠♣r ♣♥s ♦♥ t ♣st s ♦sr ♥t ♥ ♦ ♥trstW s ♥t t♦ ♦♠♣r t r ♣r♦r♠♥s ♦ ts♦rrt♦♥s ♦s t♦ ①♣♦r t ♣♦ssW s tt ♦ ♦sr ♥ ♥ ♥ s s t rs♦♥ ② ♦r tst ♦s t♦ ♣rt ♦r s♥ t t sU k r♦♠W k ♦♥t♥s t nw = 4 ♠sr♠♥ts t 0 3 6 ♥ 12 ♠♦♥ts ♦r t trrs ❲ s♠t 5000 ♣rs (W k,U k) r♥ r♦♠ N (Awθ0,ZwΩ0Z

′w +Λ(ρ0,Σ0))

rU k ♦♥t♥s t ♦rrs♣♦♥♥ ♠sr♠♥ts t t♠ tu = 24♠♦♥ts ♦r tst((Y k

i

)i=1,...,20

)k rst ♦♠♣t t ♠①♠♠ ♦♦ st♠t ξk ❲ t♦♦ H = 200

♥ L = 1 t♦ ♦♠♣t t tr ♦rrt♦♥s ♥ ♦r tst k ♦♠♣t ♦rr

t♦♥(xlα,ξ0,ξk,W k

)l=1,2,3

♥ t ss x4α,ξk,ξk

tt ♦s ♥♦t ♣♥ ♦♥ W k ♦t

tt ♦♥② ♣r♦r♠ ♦♥ st♠t ♦ t ♦rrt♦♥s ② ♦♦tstr♣ s♠♣ ♦♥t♦♥② t♦W k t ♣rt♦♥ r♦♥s r t♥ t ♥ tr ♦r ♥♦t U k ♦♥t♦ ts r♦♥s

Page 14: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♥② st♠t P(KW (ξ0, ξk) ≤ x

)② ts ♠♣r r rs♦♥

1

5000

5000∑

k=1

1[KW (ξ0,ξk)≤x]

♦r x ♥ t stx1α,ξ0,ξk,W k

,x2α,ξ0,ξk,W k

,x3α,ξ0,ξk,W k

,x4α,ξk,ξk

rsts ♦ ts

♦rrt♦♥s r s♦♥ ♥

♦♠♣rs♦♥ ♦ r ♦r ♣r♦ts st♥r t♦♥s ♦r t tr♦rrt♦♥s ♥ t st♥r χ2 trs♦ ♦r α = 0.95

t♦ 1st ♦rrt♦♥ 2♥ ♦rrt♦♥ 3r ♦rrt♦♥ t♥r χ2

♦r ♣r♦t②

❲ ♥ ♥♦t tt t tr ♦rrt♦♥ s t ♥rr♦r ♣rt♦♥ r♦♥ ♥ t s ♦x4α,ξ,ξ

t ♠♣rs♦♥ ♦ t st♠t♦♥ s ♥♦t t♥ ♥t♦ ♦♥t ♥ t ♦♠♣tt♦♥ ♦ t♣rt♦♥ r♦♥ ♥ t♦ s♠ ♣rt♦♥ r♦♥ s s ♥♦t t s ♦r t rst ♥t s♦♥ ♦rrt♦♥ tt ♦s rsts ♦♥trr② t♦ t rsts ♦t♥ ② ❱♦♥ t② ♦t ♦rst♠t t ♦r ♥ t♦ ♣rt♦♥ r♦♥s ♦t tts♠r rsts r ♦t♥ ♦r ♦tr ♦♥♥ s ♦♠♣r♥ t r ♦r♣r♦ts s rt♥② sr t t ♦s ♥♦t ♦ t s③ ♦ t ♣rt♦♥ r♦♥st♦ ♦♠♣r s ② r♣rs♥t t strt♦♥s ♦ t ♦rrt♦♥s ♥ r ♦r ♥ ♦ ξk t s③ ♦ t ♣rt♦♥ r♦♥ s ♦r♥ ② t r♥t xl♥ ♦rrt♦♥ ♦ ♥♦t ♥ssr② ξ0 r ♥♦♥ ♥ ♥♦t st♠t r t

♥♦ ♦ ξ0 ♥s t trt q♥t t♦ t ♦r (W k, ξk

) ♥

♦♥sq♥t② t ♦♣t♠ s③ ♦r t ♣rt♦♥ r♦♥ ♦♥t♦♥♥ ♦♥(W , ξ

)s

t strt♦♥ ♦ t t q♥t ♥ r ♥ r s strt♦♥ s st♠t

♠♣r② t 5000 s ♦(W k, ξk

) s ♥ s t strt♦♥ ♦ x3 s ♦r

r♥ t♥ t ♦trs tt ♥srs s♠ rt♦♥ ♦ t ♣rt♦♥ r♦♥ s③ r♦♠ ♦♥s♠♣ t♦ ♥♦tr tr♠♦r ts strt♦♥ s t ♦sst t♦ tt ♦ t t q♥t♦r♥ t♦ ts s♠t♦♥ st② ♣r♦ rsts ♥ t ♥①t st♦♥ s♥ t tr♦rrt♦♥ x3

α,ξ0,ξ,W♥ t t s t r ♦ 5000 ♦♦tstr♣ s♠♣s ♦t

tt t st ♣r♦r♠♥s ♦ t tr ♠t♦ r ♦♥② sts r ♦r t s tst♥ ♦ ♥♦t tr ♥ ♥r ♣♦ss ①♣♥t♦♥ ♦ tt t rst ♦rrt♦♥ s

s ♦♥ χ2 ♣♣r♦①♠t♦♥ ♦r t strt♦♥ ♦ KW

(ξ0, ξ

)♥ ♦♦tstr♣ st♠♥t

♦r t trt ♦r rtr♠♦r t s♦♥ ♦♥ s tr♥t s②♠♣t♦t ①♣♥s♦♥ ♦t rst ♦♥ ♥ ♦♥sq♥t② s ss rt ② ts r② ♦♥strt♦♥ t tr ♦rrt♦♥s ①♣t t♦ t st ♦♥ s t ♦s ♥♦t ♣r♦r ss♠ χ2 strt♦♥ ♦r

KW

(ξ0, ξ

) ts ♦♥② ♣♣r♦①♠t♦♥ s t♦ ssttt t r strt♦♥ ♦ ξ ② ts ♦♦tstr♣

♦♥tr♣rt s ①♣t t s ♦r ♣r♦t② r② ♦s t♦ t trt ♦♥

Page 15: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

Prt♦♥ r♦♥

❲ ♥ t ♦r ♣♦ssss ♦r ♠sr♠♥ts t 0 3 6 ♥ 12 ♠♦♥ts ♦r r ❯s♥ t ♣r♦♣♦s ♠t♦ ♥ ♥ ♥ rr♥ r♦♥ ♥♣s♦ ♦r tr s ♦r ts rs ts tr ♠srs ♦ts ts r♦♥ts t s ♦ ♣r♦t② ♦ ♥ t② rsts ♦♥ ts ♥ t r ♣♦tt ♥r s ♥♥s r ♥♦t st♦♠ t♦ ♠tr① s t s ♥♦t s② t♦ tr ♦r ♥♦t ♥ ♣♦♥t ♦♥ t ♥ t ♦♥s t♦ ts ♣rt♦♥ r♦♥ s s t

rs♦♥ ② ♣r♦♣♦s t♦ r♣rs♥t t ♣r♦t♦♥ ♦ t ♣s♦ KW

(ξ0, ξ

)≤ x3

α,ξ0,ξ,W

♦r r s s ♥ ♥tr ♦ ♣rt♦♥ ♦r r ♥ tr t♠♦ ♠sr♠♥t ♦t tt ts ♥trs r ♣rs♥t t♦ r♣ r♣rs♥tt♦♥s t② r ♦t♥ ② ♣r♦t♦♥ t② ♦ ♥♦t r♥t t rt ♦r ♦♥trr② t♦t ♣s♦s ♥ ② Pr♦♣♦st♦♥ ♦ t② ♥ ♥♦t s s♣rt② t♦ ♥♦s t s t tr rs r str♦♥② rt s s♦♦♥ s ♦ r s ♦ts t♣rt♦♥ r♦♥ t t ♥ ♦♥sr s ♣r♦② ♥♦t t②❲ ♥ r♠r tt ♦r ♣rt♦♥ ♥trs r r② r♥t ♥ ♥rr♦r t♥ t s♦ rr♥ ♥trs ♥ tr♦r t♦ r♥t ♥ s♦♥s s ♥ ①♠♣ ♦rt♥♥ ♦ t t♥ ♠♦♥ts ♦ tt s ss♣♦s ♦r t ♥ ts♥ t st♥r rr♥ ♥trs t ♥③t♦♥ ♦s ♥♦t trr s s r♠ ♥ t ♦tr ♥ ♦❯r ♦ ♦ tt s ♥♦r♠② ♦r ♠t♦ t ♥♦t ② t s rr♥ ♥trs rt♦♥ ♦ t ♦r t♣rt♦♥ r♦♥ rss t ♣r♦t② ♦r ♥ ♦ ♥ tt s s♥t s♣t t ♦♥sr r♥ t♥ t χ2 trs♦ x4

α,ξ,ξ♥ x3

α,ξ0,ξ,W

t ♦rrs♣♦♥♥ ♣rt♦♥ r♦♥s r r② ♦s ♥ ts s ts ♥ ①♣♥② s♠ r♥ ♥ tr ♦♥t♦♥ ♦♥ t ♦srt♦♥s s ♥♥♦t s② ♥t♣t ② s♠♣ ♥ ♦♥ t ♣r♠tr st♠t♦♥s s ts ♦♥t♦♥ r♥♣♥s ♦♥ ♦♠♣t ♥t♦♥ ♦ t r♥ ♣r♠trs s Pr♦♣♦st♦♥

sss♦♥

♥ ts ♣♣r ♥tr♦ ♥ ♠♦ t♦ ♣rt♦♥ r♦♥s s ♣♣r♦ s ♥s t s ♥③ ♥ ♠t♠♥s♦♥ ♥ r② ♥ ts ts ♦♥♣rt♦♥ r♦♥ ts ♥t♦ ♦♥t t ♣♦ss ♦rrt♦♥s t♥ t rst t r♥t t♠s s ♥ts ♥ s t♦ ♥rr♦r ♣rt♦♥ r♦♥st♥ t s rr♥ ♥trs ♠t♦ ❯s♥ ♦r ♠t♦♦♦② ♥♥s rtt ♠♦r ♣rs♦♥ t♦ ♣♦t♥t ♥t② ♥♠ ♦r ♣rs♦♥rtss ♦r ♠♦ s s ♦♥ str♦♥ ss♥ ss♠♣t♦♥ ♥ s ♥tr♥t ♦ t s ♦ ♥♦♥♣r♠tr r♠♦r t t ♦ ♥ ♠♦r ♥s♥ ② ♦♥sq♥ t ♥ ♥♦t ♣♣ t♦ ♦r ♣rt ♣r♦♠ s ss♠♣t♦♥ s s♦ ss ♦♥ t st ♣ t♦ ♦①♦① tr♥s♦r♠t♦♥ ♥ t ♦tr ♥ ♥ ss♠♣t♦♥ s s♦ ♠ ♦♥ t ①♣♦♥♥t rs ♦ t ♦rrt♦♥ ♦r t♠ tt ♥ ♣♣r rstrt ♦ t st ♦ ♦r ♥♦ ts ♥ ♦

Page 16: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

♣r♦♠ s r② ♥ ♠♦ ② ♥ AR(p)♣r♦ss ❲♥ ♥ ♥ ♥ ss♠♣t♦♥ t t♦ ♥ ts rs♣t t ♠♦ ♣r♦♣♦s ♥ s♥ ♥ ♦♥t♥♦s t♠s rst ♦rr ♣♣r♦①♠t♦♥ ♦ s ♥sr ♣r♠tr r♠♦r ♦s s t♦ s ♣♥ ♣♣r♦ ♦r t ♠♦ st♠t♦♥t t ♣♥ ♣♣r♦ s rr♦rs ♥ t st♠t♦♥ ♦ t ♣rt♦♥ r♦♥ ♥ st♦ r♥t ♥ ①t ♦r rt ♦ r♠♥t ts ♣r♦♠ s tr r♥t♦rrt♦♥s ♥ t ♠t♦ s♠s t♦ ttr rsts ♦♥ ♦r tst♥t② s♦♠ r♦st ①t♥s♦♥s ♦♥r♥♥ t ♠trt ♥r ♠① ♠♦ ♥♣r♦♣♦s s ①t♥s♦♥s r s ♦♥ t ♠trt t ♥ s ♥♦r♠ strt♦♥s♥ ♣r♦ ♣♣r♦s t ♠♦r ♥r ♠② ♦ strt♦♥s ❲♥ ♥ ♥ ❲♥ ♥ ♥ ❲♥ ♥ ♥trst♥ tr ♦r ♦ ♥ ①t♥s♦♥♦ ♦r rsts ♦r ts ♠♦s

r♥s

r♥♦rs♥ ♥ ♦① Prt♦♥ ♥ s②♠♣t♦ts r♥♦

r♥ rt♥ ♣rt♦♥ r♦♥s ♦r♥ ♦ t ♠r♥ ttst ss♦

t♦♥

t P♦st ♥t trs Pr♥t♦♥ rs ♥ ♣♣ t♠tsPr♥t♦♥ ❯♥rst② Prss

♥ ♥s ♦s ♦r ♦♥t♥ t t r♥♦♠ ts ♥ rr♦rs ♦r♥ ♦ t ♠r♥ ttst ss♦t♦♥

♥♥ sts♥ ♥ ❱r②♥ r♥ ♥trs ♥ t ♥ ♦

rt♦r② ♣♣r♦ ♥ ♦♠♥t ❲②♥ P

♥ ♥ t♥♥ ♦♥♥r ♠♦s ♦r r♣t ♠sr♠♥t t♦♥♦r♣s ♦♥ sttsts ♥ ♣♣ ♣r♦t② ♣♠♥ ♦♥♦♥ ❨♦r

♠♣str P r ♥ ♥ ①♠♠ ♦♦ r♦♠ ♥♦♠♣tt t ♦rt♠ ♦r♥ ♦ t ♦② ttst ♦t② rs

♦♥s ♠♠♦è ♥ ❱♦♥ P ♥♦t ♦t rt ♣rt♦♥ r♦♥s♥ strt♦♥s ♦r♥ ♦ ttst P♥♥♥ ♥ ♥r♥

♦♥s ♠♠♦è ♥ ❱♦♥ P rt♥ ♣rt strt♦♥s♦r♥ ♦ ttst ♦♠♣tt♦♥ ♥ ♠t♦♥

P P♥ ♥ ♥ ♣rt♦♥ ♥trs s ♦♥ ♣rt♦♦ ♦r ♦♦tstr♣ ♠t♦s ♦♠tr

Page 17: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

r P ♥ ♥t♦♥ tst♥ ♥ ♥ P t♦rs ♣r♦♦②♥ r♦♦② ♦ s♠ ♥♠s ♣s Ps♥ ♠s ❯

♥ ♥ ❲♥ ❲ trt s♥♦r♠ t ♥r ♠① ♠♦s ♦r♠t♦t♦♠ ♦♥t♥ t ttst ♦♥

②♥♦s ♦♥♦rt r♠♥ st ♦t ♥ r r ♣♥♥② ♦ rr♥ ♥trs ♦r ♣s♠ ♦♠ s ♥ ts ♦r♥ ♦❱tr♥r② ♥tr♥ ♥

♦tts P ♠ ♥ ③r ♠r ♥ ② ②s♥ tt♦♥ ♦ ♥♦r♠ s ♥ ♦♥t♥ ♦♠rrs t ♥ ♣♣t♦♥ t♦t rt♦ ♦sttsts

❯ ♥ st♥ st♠t ♣rt♦♥ ♠ts ♦♠tr

❱r ♥ ♦♥rs ♥r ♠① ♠♦s ♦r ♦♥t♥ t ♣r♥rrs ♥ ttsts ♣r♥r❱r ❨♦r

❱♦♥ P ♠♣r♦ ♣rt♦♥ ♥trs ♥ strt♦♥ ♥t♦♥s ♥♥♥

♦r♥ ♦ ttsts

❲♥ ❲ trt t ♥r ♠① ♠♦s ♦r rrr② ♦sr ♠t♣r♣t ♠srs t ♠ss♥ ♦t♦♠s ♦♠tr ♦r♥

❲♥ ❲ ♥ ♥ s ♠①♠♠ ♦♦ ♥r♥ ♦r ♠trt ♥r ♠① ♠♦s t t♦rrss rr♦rs ♦♠♣tt♦♥ ttsts ♥ t

♥②ss

❲♥ ❲ ♥ ♥ st♠t♦♥ ♥ ♠trt t ♥r ♠① ♠♦s ♦r♠t♣ ♦♥t♥ t ttst ♥

❲♥ ❲ ♥ ♥ ②s♥ ♥②ss ♦ ♠trt t ♥r ♠①♠♦s s♥ ♦♠♥t♦♥ ♦ ♥ s s♠♣rs ♦r♥ ♦ trt ♥②

ss

❩♥ r ♦♠♣♠♥t ♥ ts ♣♣t♦♥s ♠r t♦s ♥♦rt♠s ♣r♥r

❩♦r③♦ ♥ ♦ss ♠♣♠♥tt♦♥ ♦ t ♦♦ ♣ss♣♦rt ①♣r♥♦ t ♥tr♥t♦♥ ②♥ ♥♦♥ r st♥ ♥ ♥②ss

Page 18: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

r tr rs ♦ ♥trst r ♣♦tt ♦r t ♦ tst ♥ t ♥ t♥ ♦ s rr♥ ♥trs ♥ ♦tts ♥s r r t♥ t ♥♦♥s ♥ s ♥s ♣rt♦♥ r♦♥ s♥ t st♥r χ2 trs♦ x4

0.95,ξ,ξs ♦s

t♦ t♦s ♦t♥ s♥ x30.95,ξ0,ξ,W

s r ♣♣rs ♥ ♦♦r ♥ t tr♦♥ rs♦♥ ♦

ts rt

Page 19: Individual prediction regions for multivariate ... · better detect abnormal alues.v Recentl,y Wang and anF (2010) proposed a method to build prediction regions. They used a porder

7 8 9 10 11

Distribution of thetargeted quantileDistribution of x1

Distribution of x2

Distribution of x3

x4

r strt♦♥s ♦ t trt q♥t ♥ ♦ t r♥t ♦rrt♦♥s r r♣rs♥ts♥ r♥ st♠t xl st♥♥ ♦r xl

0.95,ξ0,ξ,W♦r l = 1, 2, 3 ♥ x4 ♦r x4

0.95,ξ,ξ

s♦ ♥ strt♦♥ s♦s ♦ t t q♥t ♥s t W strt♦♥ ♦x30.95,ξ0,ξ,W

s t ♦sst t♦ t t strt♦♥ r ♠t♦s ♥ ♦rst♠t

t t q♥ts ♠t♦ ♥rst♠ts t♠ s r ♣♣rs ♥ ♦♦r ♥ ttr♦♥ rs♦♥ ♦ ts rt