Bootstrap Confidence Interval for Nonparametric Regression - repository civitas UGM

Sponsoring societies

Bernoulli Society
fr;r N4;rLiierrrrtieiii \i:ir;'',r.
and Prob:rliilitv

8th World Congress in Probability and Statistics

lstan b ul, 9 -1 4 July, 2012

fointly sponsored by the Bernoulli Society and the lnstitute of Mathematical Statistics

PROGRAM COMMITTEE
ADRIAN BADDELEY, University of Western Australia
VLADIMIR BOGACHEV, Moscow State University
KRZYSZTOF BURDZY University oJ Washington
T. TONY CAl, University of Pennsylvania
ELVAN CEYHAN, Kog University
P

RO BA L


CHAU

D H U Rl, I n d i a n

Stati sti ca I I n stitute

MINE qAe LA& Kot lJniversity
ERHAN qlNLAR, Princeton University

ANTHONY DAVISON,

Ecole Polytechnique F\ddrale de Lausanne
RICK DURRETT, Duke University
ARNOLDO FRlcESSi, l.)niversity of Oslo (Chair)

ALICE GUIONNET, Ecole Normale Supdrieure de Lyon
PETER CLYNN, Stanford University
PETER CUTTORB University of Washington
ON


ES

I

MO

H

ERNAN DE Z, I n stituto

Po

litecn i co

N a ci o n a

I

DMITRY IOFFE Technion-lsrael Institute of Technology

SUSHMITA MITRA, lndian Statistical lnstitute
SHf GE PENG,

ShandongUniversity

DOMINIQUE PICARD Universite Paris Vll
KAVITA RAMANAN, Brown University
SYLVIA RICHARDSON, lm perial College

VLADAS SIDORAVICIUS, Centrum Wiskunde & Informatica lnstituto Nacional de Matematica Pura
Aplicada
MICHAEL S@RENSEN University of Copenhagen
MATTHEW STEPHENT University of Chicago
GU ENTH ER WALTH ER, Stanford U niversity

VICTOR YOHAL Universidad de Buenos Aires
NAKAHIRO YOSHIDA" University oJ Tokyo
THALEIA ZARIPHOPOULOU The university of Texas at Austin

LOCAL ORGANIZI NG COMMITTEE

Fl KRI AKDENiZ Qukurov a U niv ersity
ELVAN CEYHAN Kog University (Co-chair)
MINE qAe LAR Kog LJniversity (Co-chair)
UXU CUnlER" Bitkent L)niversity
$ENNUR ONUR, Turkish Statistical Institute
SULEYMAN OZefiCi, Kog LJniversity
SEMIH SEZER, Sabanu University
FETI H Yl LDI Rl/W Qankay a U niv ersity

SCI ENTI FIC

SECRETARIAT

MEHMET 62, Xog University

ORGAN IZATIONAL SECRETARIAT
CAN OZHAN, Figilr Kongre & Organizasyon

e


Contents

Program Overview
Program

5

11

Abstracts

117

lndex

295

Abbreviations
IS


Invited Session

CS

Contributed Session

Name in boldface
BS GA BS
BS EC&CM
BS ERC BS

Speaker
General AssemblY

IMS BM
IMS CM
IMS ED
IMS EX
IMS PA


BS Executive Committee & Council Meeting
European Regional Committee Meeting
IMS Business Meeting
IMS Council Meeting

IMS Editors'Luncheon
IMS Executive Committee Meeting
IMS Presidential Address
.

f-

Program Overview

Monday July 9
08:00-09:00 Registration

13

09:00:1 0:00 Opening Ceremony


13

1

0:00-1 1 :00 Tukey Lecture

11:30-12:300 Wald Lecture 'l
14:00-15:45
IS 3 Geometric Perspectives on High-Dimensional Data
IS 33 Stochastic Methods for Equilibrium In Financial Markets
IS
Statistical Methods for Neuroscience
IS 30 High-Throughput Genomic Assays
IS l8 Probability and Computer Science
IS l7 MCMC: Adaptive, Likelihood Free, Particle Filters

3l

14:00-15:40

CS Applications to Biology and Medicine
CS 2 Asymptotic Distribution Theory
CS 3 Design of Experiments

I

CS
CS
CS

4 Air and Oceans
5 Likelihood and Her Cousins
6 Inference and Markov Chains

13
13

t3

l4

t4
l4
l5
15

t6

l7
18

l9
t9
20

<
CS

PROCRAM OVERVIEW

7 Measurement


>

Error and Missing Data

fiS 8 Queueing Theory and Traffic
(iS 9 Causality
l6:15-18:00
IS l4 Machine Learning

2l
22
22

23

8 Functional Data Analysis
IS 9 Graphical Models, Networks and Causalify
IS 40 Stochastic Population Genetics and Evolution

24
24

IS 34 Dynamics and Cubature for Stochastic Non-Linear Systems

25

IS

25

16:15-17:55

CS 10 Bayesian Statistics I
CS l1 Copula Models
CS l2 Classificationand Clustering
CS 13 Image Analysis
CS 14 Model Selection I
CS 15 Stochastic Differential Equations
CS 16 Biostatistics I
CS l7 Inference for Stochastic Processes I
CS 18 Theory I
CS 19 Time Series I
CS 20 Theory 2
1

9:30 -22:O0

Welcome Reception

25

26
27
28
28
29
30
30

3l
32
33
33

Tuesday 10July
09:00-10:00 Medallion Lecrure

1

10:00-11:00 Wald Lecrure 2
"l 1

:3O-12:30 Kol mogorov Lecture

34
34
34

12:30-14:O0 Meeting IMS ED

34

12:30-14:00 Meering

34

BS ES

& CM

13:30-17:30 Meeting IMS Ex

14:00-15.45
IS 39 Stochastic Partial Differential Equations with Applications
IS l5 Modem Applications of Malliavin Calculus
IS l0 High-Dimensional Inference
IS
Random Matrices and Applications
IS 24 Critical2-D Systems and SLE

2l

34

34
35
35

36
36

'14:00-"15:40

CS
CS
CS

21 Event History Analysis
22 Commodity Markets and Finance
23 Graphical Models and Random Graphs

36
37
38

<
CS
CS
CS
CS
CS
CS

PROCRAMOV€RVIEW

>

24 Thresholding and Variable Selection
25 Nonparametric Inference
26 Mode{ Selection 2
27 Geometric Probability and Stochastic Gcottrclrr
28 Gaussian Processes and Sample Path Properties
29 Theory 3

14:00-1 8:00 Poster Session

16:15-18:00
IS 13 Long-Range Dependence and Self-Similarity
IS 37 Stochastic Models of Cancer
IS 5 Data Depth
IS I Bayesian Nonparametrics
IS 26 Sparse Signals

!i

i-'

li
44

,16
,1

li

49
,19

50

16:15-17:55
CS 30 Statistical Genomics
CS 3l Central Limit and Other Weak Theorems I
CS
CS
CS
CS
CS
CS
CS

32 Markov Processes I
33 Factor Analysis and Principal Components
34 Finance
35 Statistical lndustrial Process Control
36 Stochastic Integrals
37 Processes with Independent Increments
38 Distribution Theory I

18:10-19:00

BS

Awards Ceremony and General Assembly

57

Wednesday 11 July
09:00-10:00 Wald Lecture 3

58

0:00-1 1:00 Laplace Lecture

58

1

11:3O-12:30 Medallion Lecture 2

58

14:00-1 5:30 IMS Presidential Address and IMS Awards Ceremony

58

16215-17:55
CS 39 Algorithms
CS 40 Biostatistics 2
Central Limit and other Weak Theorems 2
CS
CS 42 Models for Risk I
CS 43 Detection and Identification

4l

CS
CS
CS
CS
CS
CS

44 Markov Processes 2
45 High Dimensional Data
46 Networks
47 Re.gression I
48 Stochastic Calculus of Variations and Malliavin Calculus
49 Stein's Method and Applications

58
59

60

6l
61

62
63

63

64
6s
65

(S
(S
(-\
(S

i0

ln

ir

[)r stri

ir
il

linrtc Drr. rsrbrlrty' :rnil [-cvv Processes
bution f ltcor"I'uirc Scric. l

66
67
67

Titcorr

68

-1

Thursday "l2luly
70

09:00-1 0:00 Doob Lecture

10:00-11:00 World Congress Public Lecture
1:30-1 2:30 Bernoulli Lecture

70

17:3O-14:OO Meeting BS ERC

70

12:3O-14:3O Meeting IMS CM

70

14:00-1545
IS 27 Spatial Stochastic Models
IS 36 Stochastic Methods and Finance

70

1

l2

7l

Large-Scale Multiple Comparisons

7t

Composite Likelihood Inference
IS 23 Rare-Event Monte Carlo Methods

72
72

IS
IS

2

14:OO-15:40

CS
CS
CS
CS
CS
CS
CS
CS
CS
CS

54
55
56
57
58
59
60

Theory

72

5

Bayesian Statistics 2

73

Models for Risk 2
Robust Methods
Order Statistics

74
75

76
76
77

Quantile Regression

Reliability

6l Stationary Processes

and Gaussian Processes

62 Stable Laws and Fractional Diffusions
63 Theory6

78

79
79

14:00-18:00 Poster Session 2

80

15:45-16:15 Meeting IMS BM

85

16:15-18:00
IS 20 Random Combinatorial Structures
IS 6 Decision Theory, Control Theory and Games
IS 25 Space Time Data

85
85
86

IS 19 Quantile Regression

86

IS 38 Stochastic Networks with Applications

87

16:15-17:55
CS
CS
CS

64 Applications in Biology, Agriculture and Forestry
65 Functional Limit Theorems and Invariance Principles
66 Pair Copula Constructi,ons

87

88
89

ll.'L >

( \ ,r iiyrr4tltcsrs Testing
( \ r'\ l'\'r':rlisetl rcglession.p t l
i \ ,i\r \tr)!ir.rsric Ordinary Difterential Equations
( \ '-{) Il.rindr)n) lle lds
( S I I Rundt)ni Processes with lnteractions
( S 7l Sanrpling and Survey Methods
f

90
90
91

92
93
93

Friday t I July
09:00-10:00 Medallion Lecture

3

95

1:00 Medallion Lecture 4

95

1.1:30-12:30 Medallion Lecture 5

95

123O-'14:OO Meeting BS EC & CM

95

.10:00-1

14:OO-'15:45

IS

IS

i5

Stochastic Geometry and Spatial Point Processes

l6

Markov Processes
IS 28 Statistical Climatology
IS 29 Statistical Inference for Stochastic Differential Equations
IS 22 Random Media

95
96
96
97
97

14:00-15t4O

73 Semiparametric and Nonparametric Bayesian Methods
74 Stochastic Processes I
75 Industrial Applications and Clinical Trials
76 Portfolio Optimisation and Volatility
77 Markov Processes 3
CS 78 Random Matrices
CS 79 Regression 2
CS 80 Stochastic Processes 2
CS
CS
CS
CS
CS

CS

8l

Foundations

16:15-18:00
IS 4 Copula models
IS 32 Statistical methods in finance
IS 7 Extremes for Complex Phenomena
IS 11 Interacting Particle Systems

98

99
99
100
101

102

102
103

104

105
105
106
106

"16:15-17t55
CS
CS
CS
CS
CS

82
83
84
85

Limit Theorems and Random Walks

106

Weather and Climate

101

Mixfures

108

Wavelets

109

86 Stochastic Partial Differential Equations
CS 87 Spatial Models
CS 88 Inference for Stochastic Processes 2
CS 89 Stopping Times and Optirnal Stopping Problems
CS 90 Theory 7
CS 91 Time Series 3 and Longitudinal Data

109
110
111

tt2
113

114

<

PROCRAMOVERVIEW

>

Saturday 14July
09:00-10:00 Le Cam Lecture

115

10:00-1 1:00 L6vy Lecture

r15

1 1 :3O -

12:30

Cl osi n g

Ceremony

115

r< )li(x,RA\,1 p

,+"1*nclay July

I

cs 20 Theory 2
Vrruur: Kristal

CHAIR: YUVAL BENJAMINI, Department of Statistics, University af
Cahfornia Berkeley, Berkeley, CA
16:15 Shuffle estimators: measuring the explainable variance in fMRl experiments

Assrnncr 4zSPACE 235
YUVAL BENJAMINI, Department

oJ

Statistics, University of Cahfornia

Berkeley, Berkeley, CA

BIN YU, Department of Statistics, University oJ Caltfornia Berkeley, Berkeley,
CA

1

6:35 Modified goodness of flt test under selective order statistics
AssrRecr 336 PACE 2o9
FARIDATULAZNA AHMAD SHAHABUDDIN, School of Mathematical
Sciences,

Universiti Kebangsaan, Malaysia

KAMARULZAMAN IBRAHIM, School

oJ

Mathematical

Sciences,

Universiti

Kebangsaan, Malaysia
ABDU L AZIZ JEMAIN , School of Mathematical Sciences, Universiti
Kebangsaan, Malaysia
16:55 Delta method for deriving the consistency of bootstrap estimator for
parameter of autoregressive model
AssrRecr 146 PACE i53

Palembang lndonesia
SURYO GURITNO, Mathematics Department, Gadjahmada University,
Yogyakarta, lndonesia
Rl HA RYAT Ml, M ath e m ati cs
Yogyakarta, lndonesia

S

17:1 5

i'

D ep a

rtment,

Ga di ah m a d a U niv er sity,

Bootstrap confidence interval for nonparametric regression
AssrRecr 3B4Pecrzzz
SRI HARYATMI KARTIKO Department of Mathematics, Cadjah Mada

University
17:35 Test of arch effect based on second-order least square estimation

l.

Assrnect 169 PACE't60
HERNI UTAMI, Department

oJ

Mathematics, Gadjah Mada University,

Yogtakarta, lndonesia
SUBANAR SUBANAR, Department of Mathematics, Gadjah Mada
U niv ersity, Y o gy aka rta, n d o n esi a
DEDI ROSADI, Department of Mathematics, Gadjah Mada University,
I

Yogyakarta, lndonesia
19:3O -22:OO We lcome

reception

VrNue: Kog University

73

WORLDCONG2012 9-14 July 2012 ISTAMBUL, TURKtrY

BOOTSTRAP CONFIDENCE INTERVAL
FOR NONPARAMETRIC REGRESSION
Sri Haryatmi Kartiko
Mathematics Department
Gad.i ahmada University, Yogyakarta, Indonesia
Abstract

In

regression, pairs

objcctive is to estimate

of data (Xt,Yt),.'.,(X,",Y")

EIYIX: r]'
^(") -

a,re observed, and the

Nonparametric techniques allow m to

bc estimatcd without any assumptions on its underlying form. Unfortunately, this

flexibility

ca,n rrra,ke

it difficult to a,scertairt wlretlrer observed leatules are physica,lly

meaningful or the resull, of random va,riabilitv in the data. Confidence inl,ervals gives
a mcchanism to asscss the impact of this variability on the cstimate.
'I'he bootstrap a,pproach to constructing confidence intervals for irernel regression
cstimates is proposcd. First, hernel Nada.raya Watson rcgrcssion estimator is uscd

to calculatc thc residuals, which approximate thc tlue nnol:sclved errors.

Second,

thc r-csiduals arc re sampled, to create "ncw" erroLs. Third, thc re sa.mplcd errors
arc aclded to an estimate of thc regrcssion function, creating a new bootstrap data
set (X1

,yi),.. ,(X,.,Y,i).Finallv,

the bootstrap data set is smoothed, taking into

accolult the bia.s ald used for constructing qua,ntile bootstla,p cortfidertce irrterva.l.
Simulation study is conducted to see that the covcrage probability approximates

its level of significance.

Key wolds : kcrncl firnction. biam. bootstlap, quantilc, covcragc probabilit.v
l

I.

INrRoorrcrroN

I'he goal of regression fitting is to fincl the relationship of inclependent(s) r.ariable

7 and dependent

variable

fronr pair valiables

Y. If

('l,Y),

we havc n indepcndent obscrvations

{(",,y,)}i:1

regression equation can be wlitten as

Y:SQ)*e;i,:1,...,rL

(1 1)

whcre e is a random variable denoting the variation of
regrcssion cwve

Y around .q(f), the

mean

EIYIT : t].

The problem is to approximate the mean rcsponse function g,

s(t)
where

7.

: n(vlr : r, -

f(l,g) is the joint density of (T,Y)

J ttf(t'Y)da

j

f(t)
and

(1.2)

/(l) is thc marginal density of

Thc nonparametric approach fol cstimating thc regression curve, gives morc

Ilexibility since the modcl is no1, spccified such

a,s

in

1,hc

paramel,ric model. From

this point of view, Hardle (1991) said that in this approach we,'let the data speak

for themselves". This regression model

assumcs

that g belongs to sorne infinite

dimensional collection of functions. The choice of g is motivatcd by smoothness
propertics the regression function can be assumed to possess. In this research.g is
smooth of ordcr p.

II.

Esrrrur.q.rroN

The estimator of g is the wcighted avcrage of Y; rvith thc rvcight dcpcnding on
thc distancc of

Z

from

l. formulizc

: r tiw,,i(t;l1.-.,7,,)Y,,

(2 3)

is thc wcight function clcpencling on bandwiclth

ol smoothing paramctcr

o.Q)
r'vhcle

llf";

as

h and samplc Tt . . . ,7,, from thc inclcpcndcnt varitrblc. For

sinplicity W,,;(t;Ty

...

,7,,)

is u'r'itten asl,V,,;(.t). Thc abovc cstimator is a r.vcightcd lcast sqnarcs cstimator as
provcd in thc rrcxt thcorcm.

Theorem

2,L For sent'ipara,m,e,t,r,ic:
lt,,tt)

uhere LDl,'-r!4r.,,r1/)

: I,

'is

rertress'iort,,model 7.1.. th,t: t:st,i,m,ator

-,, 't
-

ll',,,(/)Y'.

(2 +)

a uLeighted least sq,uares estirnator for g(t)

Nrrrlala\a arrd \\'atson(196.1) choose I,I,',,i(t)

- ;+*;i-J.

Iil,

(1)

: iri(il,

1( is a kcrncl function, and g,.(l) is thcn called kerncl estimator', wlittcn
(2.5)

u,hclc lr is bandr,r,idth or smoothing pararneter u'hich defines the degree of smoothncss

of the estimator. The most uscd kerncl functions arc Uniform, Tliangle,

Epancchrrikov and Gaussian.

Bv taking u,,,i(t)

: t+trfi

thc cstimator of 9(l) statect abovc is cquivalcnt

to

g,,(t) - D::tw-i(t)Y.
Sevcral properties of wcight function I'tr/,;(l) (which is also

truc for.r',;(l) with

scvcral adjustment) are:

(1) The Weights tr44",(1) depend on the samplc {Ti}'i':, through the hcrncl dcnsity estimatot ir,(t).

(2) Obscrvations g; obtained morc
ing
(3)

fi

r'vcight in thosc aLeas

t'hclc thc collcspond-

are spalsc.

If /i - 0, thcn I41,;(l) - n for t - Tt.

(4) Fol h

-

oo, the weight function W,,i(t)

-

1 for all

l, hcrcc A(l)

convergcs

rn lltc collstarr frrDctiorr y.
(5) llhc bandrviclth h dctcrmincs thc lcvcl of snoothncss. clccrcasing h lcacls to

a lcss smooth estimatc.

Tlreorem 2.2.

^rsElL,

Xt[ean Squarcd,

Error' .for g.,,(t)

zs

(t)): h581r( l3+ 'i (n" (r) + 2
)') LL3Q{ )
^tft;Q
lo(rth t) + o(tLr)

(2.6)

4

Frorn thc abor.c theolen, lt,ISElQ,,(l)] is of olclcl O(rz l/5). fbr
'l'he fir'st srrrnrna,rrd in'-['heolern 2.2 t]csclibcs

- 11,-1/5.
thc zrsvniptotic valia.rcc ol (t,,(t1.

This fbrnmla for the variancc is a fiurction of the ma.r'ginal dcnsitv
conditior-ral r.ariancc

a2(l) and

11

l(z)

and thc

agrecs u'ith thc assr.rnrption that the lcgrcssiol cnrvc

is more stable in thosc arcas rvhere we havc plcnty of obscrvations. The second
sunrnrand corresponds

to thc squared bias of g.,,(t) and is cithcl doniinated by

the second derivative g" (t) ifwe arc ncar thc local extreme of g(l) or bv the first
dcrivativc A'(l) when we ilre nezrr the inflection point of 9(i).
From Eubank(1988), the use ofseveral kernel will not give abig difference
fbr tlie estirnator. The choicc of bandwidth gives rnorc irifluctrcc. Irr otllcr wolds,

if

the choice of bandwidth is optimized, the different hernel will give almost the same

result. Sevcral distance measurencnts arc:
(1) I\{can Squared

Errol
'lL

MSE(h): ASE(h):
(2) Integrated Squarccl
I

"-'fi:1' .(r')(s.(T") - g(.ri))2

(2-7)

Error

SE(h)

:

fn
J _*utlt)(!1,,(t)

-

s\t))" .f (t)dt

(2.8)

(3) Condition Squared Error
A,,r

AS

E(h)

:

E(AS E(.h)

lTt,. . .,7,,)

(2.e)

(r) Nlcan Integra,ted Squarcd Enor

NrrsE(h): E(rsE(h))
The nonncgatir,-e rveight function
of

X.

u.r

(2.10)

is a tool to dlop observa.tions at the boundaly

Thc abovc distance neasLll'cmcnt lcads asyrnptotica,lly to thc sirmc lcvcl of

smoothing.

Theorem 2.3. (Hardle dan Al[artrn. 1996)

Suppose that

: Al. E(yk'ft : t) < c:t < cc, k - 7.2,.. . ,
: A2. ,f (t) kortt'inu,

Hol,dcr dan positiJ'pado,

su:1tort, d,o.rt

u.

I

A3.

Ii

Ilolder

kontin'Lt,

then. .for kernel

c:st;int,at.or-s.

-I ::3 0 untrr,k l'L
supld(h)AH- AIISEth\l
(2.11)
- cx:.
S E(h)
t, u,, I
wlrere H,,: [,.'r-1,n i]. 0 < 6 < Il2 a,rt d(Lt) is n,e r.f the th,ee d,,istance,meusu,rement except fuI I

S

E.

The theoren implics that cach sequcnce of bandwidths minimizing any of thcsc
threc distance measuremcnts AStr(h), ISE(h), N,IAStr(h) is also asymptotically opti-

mal in rclation to the optimizing bandwidth of the \,Iean Intcgrated Squarcd Error.
Fol convcniencc. u'e will tal1,ive

Amer'ican Stati,sti.cal

A

ss

suroothilg :rlrl
ocio,ti,on 83 (401