Nonparametric curve estimation with time series errors

Nonparametric curve estimation with time series errors

Journal of Statistical Planning and Inference 167 28 (1991) 167-183 North-Holland Nonparametric errors* curve estimation with time Young K. ...

812KB Sizes 10 Downloads 235 Views

Journal

of Statistical

Planning

and Inference

167

28 (1991) 167-183

North-Holland

Nonparametric errors*

curve estimation

with time

Young K. Truong Department Received

of Biostatistics,

University

11 May 1989; revised

Recommended

Abstract:

a model

a parametric estimators

for the flexibility

Key

time series error.

Under

regression

part of a stationary

process

then these parameters

Subject words

average;

Chapel Hill, NC 27514, USA

21 June

1990

and the serial correlation is a smooth

appropriate

can be chosen to achieve the pointwise

processes), AMS

received

in which the mean function

by Stone in nonparametric forms

manuscript

Carolina,

by M. Hallin

To account

we propose

of North

Classification: and phrases:

optimal

Natve

regularity

function

estimation.

can be estimated

kernel

rate of convergence;

62605;

conditions,

or the uniform

with a finite number

Primary

in fitting

estimator;

optimal

Moreover,

of parameters

measurement

function a sequence

data,

and the noise is of local average

rate of convergence

as defined

if the time series error

structure

(e.g.,

with the usual root-n

secondary

repeated

nonparametric

finite order autoregressive

rate of convergence.

62E20.

nonparametric

time series; autoregressive

regression;

curve

estimation;

local

process.

1. Introduction In analysis

of human

longitudinal

x,,=ti(t;)+z;,n,

growth

curves,

data are often

modeled

as

t;=i/n,

where 19(. ) is a smooth function on [0, l] and Z,, are random errors with mean 0 and variance cr*, i=O , 1, . . . , n. See, for example, Gasser et al. (1984), Seber and Wild (1989) and the references given therein for many interesting applications of curve estimation. In the above model, the function f?(. ) is said to be parametric if it is defined in terms of a finite number of unknown parameters. Otherwise, it is called nonparametric. An advantage of the parametric approach is its interpretation of the model. However, as indicated in Gasser et al. (1984), parametric modelings are often not very flexible in fitting data. They provided an example on human growth * Research the University

was supported of North

0378-3758/91/$03.50

by Junior

Carolina

Faculty

at Chapel

@? 1991-Elsevier

Development

Award

and a Research

Hill.

Science

Publishers

B.V. (North-Holland)

Council

grant

from

168

Y.K. Truong / Semiparametric repeated measurement models

in which many parametric growth spurt (MS).

analyses

fail to account

for a phenomenon

known

as mid-

When data are collected sequentially over time, the errors Z; often have substantial correlation. This happens because measurements are being taken over time from the same subject or there may be other environmental factor such as sickness (Seber and Wild (1989, p. 272)). To model this correlation in practice, the error process is often

assumed

to be part of a stationary

process.

Followings

are two examples:

Example 1.1. AR(p) errors. In many applications, the correlation may expect to be decreasing as the delay k between Zj and Zi+k increases. The simplest case is Corr(Z,,Zi+,)=&, This is the covariance

lel cl.

structure

Z;=@Z;_,+&;,

of the AR(l)

process

e;‘~qo,c2).

Here (~~),~z is an innovation process. In general, let p denote a positive integer and (b,), 5 ;_ be constants called parameters. The process Z; is said to be a p-th order autoregressive time series if it satisfies Z;=b,Z;_,

i.i.d. E; (0,a2).

+ . ..+b.Z;_,+E;,

1.2. ARMA(p, q) errors. Let p and q be positive integers. Also, denote the The process Zj is said to be an autoparameters by (bi)i
In the above

examples,

l-b,z-...

suppose

-b/#O,

the following 1z1 51,

condition:

ZEC,

and there exists a sequence of real numbers holds. Then (Z,)iCL is stationary (~7~)~~~ such that (see Theorem 3.1.1 of Brockwell and Davis (1987)) Zj= C auEimu, u

C la,1
For example, the AR(l) process Zj=~Zi_i+ei (1~1
0(. ) is a smooth

c, i.d(o,02),

+ C a,.&,_., U function

with

bounded

t,=i/n, first

i=O,l,...,

derivative

on

n,

(1.1)

(0,l)

and

Y. K. Truong / Semiparametric

(a,.),.z

are unknown

constants.

repeated measurement

Note that this model

models

has the flexibility

169

to fit a

wide variety of data sets because of the nonparametric function 19(.) and it has the ability to account for serial correlation in the measurements. A similar model was also considered by Hardle and Pham (1986) in time series trend estimation. The nonparametric estimator to be considered in this paper is the naive kernel Under some appropriate conditions, this estimator based on local averages. estimator will be shown to possess the optimal rate of convergence nm”3 both pointwise and in the L2 norm; and the optimal rate of convergence (K’ log n)“3 in the L” norm. These results generalize partially the corresponding results established by Stone (1980, 1982) to include correlated observations. Moreover, they can also be used to fit the error process. For example, if the error process is an AR(p), then the parameters can be estimated with nm”2 rate of convergence. In fact, this result extends the corresponding ones in time series regression in which 0(. ) is assumed to be a parametric mean function. See Fuller (1976) and Hannan (1970). The rest of the paper is organized as follows: Section 2 describes the nonparametric estimator of Q( .) and optimal rates of convergence. The fitting of the error process is given in Section 3. Section 4 presents two numerical examples. Some concluding remarks are given in Section 5 and proofs are given in Section 6. We conclude this section with a brief review of some related work. For uncorrelated error models, besides the previously mentioned Gasser et al. (1984), Stone (1980, 1982) and the references given therein, Wong (1983) considered the problem on bandwidth selection. For correlated errors, Roussas (1988), Truong and Stone (1988) established rates of convergence for nonparametric estimators of the mean function based on the a-mixing, which is a nonparametric description of the dependence structure. One difficulty in this approach is the achievability of the L” rates for unbounded correlated observations, although optimal rates for bounded series are obtained by Truong and Stone (1988). Muller and Stadtmuller (1987, of correlation based on m1988) considered the L” rate and the estimation dependent error processes. Hart and Wehrly (1986), Hart (1989) and Chiu (1989) considered the problem on optimal bandwidth selection in the L2 sense. Hardle and Pham (1986) obtained pointwise central limit results for a class of robust smoothers. Except the work by Muller and Stadtmiiller (1988) on fitting finite order moving average processes, none of the above papers addressed the problem of nonparametric estimation of the mean function Q( .) and the fitting of error process.

2. Statement

of results

The kernel estimators of the function 0( .) will now be described. Let the observations y be given by (1.1). Let (a,,), k 1 be a sequence of positive numbers tending to zero. Given t E [0, 11, set I,(t)={i:Osisn

and

It;-tl

sS,l

and

N,,(t)=

#I,(t).

Y. K.

170

Truong / Semiparametric repeated measurement models

Also set &(t)=N,(t)-’

c I$, I!,(l)

which is a kernel estimator of e(t) based on local averages. Note that although this is a simple estimator, the local mean estimator has the advantage to provide justification

of the usual root-n

Section 3. The rates of convergence the following conditions. Condition

2.1.

about

of the nonparametric

There is a positive

constant

the parametric estimator

error process.

treated

here depend

See on

M, such that

for s, tE [0, 11.

~e(S)-ee
inference

2.2. sup c l%,ul cm. ” U

Condition

2.3. i.i.d.

hi -

N(0,~2),

u=o,

&l, +2 )....

The normality in Condition 2.3 is imposed only to avoid technical details. It can be weakened by using, for example, Assumption 2.6.3 of Brillinger (1981) which assumes basically that the innovation process (~;)~~z has moments (or cumulants) of any order so that its characteristic function admits a Taylor expansion. Given positive numbers a, and b,, n L 1, let a, - b, mean that a,/b, is bounded away from zero and infinity. Given random variables I$, n ~1, let V, = OJb,) mean that the random variables 6,’ V,, n L 1 are bounded in probability or, equivalently,

that lim

lim sup P( 1V,) > cb,)

Y. K.

Truong

/ Semiparametric

Let g( . ) be a real-valued

Theorem

2.2.

function

repeated

measurement

models

171

on [0, 11. Set

Suppose 6, - n ~’ and that Conditions 2.1 and 2.2 hold. Then

IIe,(.)-e(.>II,=O,(n~“>.

2.3. Suppose Conditions 2.1-2.3 hold. If 6, -(n-l ists a positive constant c such that

Theorem

lim P(lle,( n Proofs

.) - 19(.)I/mlc[n-’

of these theorems

log n)‘, then there ex-

log(n)]‘) =O.

will be given in Section

6.

Remark 2.1. The above theorems generalize results of Stone (1982) to cover dependent observations Y. Furthermore, Theorem 2.3 does not require interpolation in order to achieve the L” rate of convergence. Remark 2.2. Note that model (1.1) is a semiparametric model. An advantage of the present approach is the feasibility to carry out statistical inference about the correlation structure. Details are given in the next section. A completely nonparametric formulation of the curve estimation is given in Truong and Stone (1988) where the t, are allowed to be random. For this approach, the o-mixing condition is used in place of Condition 2.2 of the present paper.

3. Fitting of error processes In practice, the serially correlated error process is unknown and it may be of interest to estimate it. In this section, this process will be treated as a p-th order autogressive model AR(p). Specifically, let y=e(t;)+Z,,

(3.1)

Z;=b,Z;~,+~~~+b,Z;~,+E;,

(3.2)

- b,zP #O for all z E C such that (z( 5 1 and E; are i.i.d. N(0, a’), {Z;} is observable, then the AR(p) model can be as usually done in ordinary time series fitted by regressing Zj on Z,_,, . . ..Zj_. analysis. Since the actual observations are Yo, . . . , &, the parameters will have to be estimated from the residuals. To this end, denote the residuals by

where

1-b,z-...

i=O, 19*.., n. If the error process

p;=

r;-

Q&j) =

z;-

(&(t;) - d(f;)),

172

Y.K.

Truong

/ Semiparametric

repeated

and b :=(b,, . . . . 6,)‘. Also, let b^ be the estimator Pi-t, ...,Z;_P.

measurement

model5

of b obtained

by regressing

Zj on

Theorem 3.1. Suppose Condition 2.1, (3.1) and (3.2) hold. If gfl( .) is constructed based on 6, -(IT’ log n)“, then I/(6-b)

= N(0, 02r,-‘),

where the p XP matrix r, := [COV(Z,, Z,)]fj= The proof

of this theorem

1.

will be given in Section

6.

Remark 3.1. In classical time series analysis, the conclusion of the above theorem holds by fitting a parametric model to the mean function 0(. ). See Fuller (1976, Chapter 9) and Hannan (1970, Chapter 8). The above theorem shows that such a parametric assumption is not necessary. A practical implication of this is that the model has greater flexibility in fitting data. Remark 3.2. Let e2 and e2 denote respectively the estimator of cr2 based on Z ,, . . . , Z, and Z,, . . . , 2,. Then the above theorem shows b2+ o2 in probability. Furthermore, it also shows that fi(d2a’) and fi(82-o2) have the same limiting distributions. Remark 3.3. Note that the covariance matrix of b^ is identical to that of the estimator obtained by regressing Zj on Zj_ t, . . . , Z,_,. This implies that there is no loss of efficiency in parametric estimation under the reconstruction of the error process through smoothing.

4. Numerical

results

This section illustrates the methodology discussed in Sections 2 and 3 by giving two examples: The first is a simulation study and the second is based on actual data. Example 4.1.

The observations

y=sin(2rt,)+Z;,

were generated

Z,=0.5Z,-,+E;,

from 1.i.d. E, - N(O,O.S’),

i=O, 1) . ...200.

Figure 1 shows the scatterplot for these observations. A local average estimate of 0(t)=sin(2at) is obtained by choosing a smoothing parameter 6,=0.17. This is also presented in Figure 1. (In fact, a sequence of local average estimates were constructed with bandwidths ranging from 0.12 to 0.20. 6, = 0.17 was chosen because it provided the most suitable smooth estimate. One may also try to select the bandwidth via the method considered by Chiu (1989), which, however, requires the knowledge of time series component.) After the data were smoothed, the residuals y-@Jr,) (Figure 2) were used to

Y.K. Truong / Semiparametric repeated measurement models

0.2

0.0

Fig. 1. Scatter

plot of Y, =

0.4

0.6

0.8

sin(a?rt,)+Z,, i=O, 1, . . . . 200, where Z,=O.~Z,_I

The solid line is the estimate

based on local mean with bandwidth

173

=0.17.

1.0

and E,“G N(O,0.5’).

+E,

The broken

line is the curve

Q(t) = sin(2at).

0.2

0.0

0.4

Fig. 2. Residual

provide

parametric

estimates

0.6

t

plot of y-

0.8

1 .o

f?,,(f,), i=O, 1, . . . . 200.

for the time series Z,. The estimated

by

model

is given

1.l.d.

Z; = 0.463052;_,

+ E;,

E; -

N(O,O.23 136).

Indeed, this was obtained by fitting four autoregressive models (AR(p), p = 1, . . . ,4) to the residuals using Brillinger’s DRBLIB. These results are given in Table 1. Here, HQ(p) is defined by HQ(p) = log@*) + 2prC ’ log log(n) (Hannan and Quinn, 1979) with s* denoting the residual variance. The order of the fitted process is chosen to be the value of p at which HQ(p) is minimum. In this case, the minimum value is HQ(l). (The Hannan-Quinn criterion is preferred over other model selection criteria because (1) it is a consistent rule, and (2) in a simulation study, Liitkepohl (1985) has shown that this criterion leads most often to correct estimates for model order and has the smallest prediction error.) Table

1

Parametric

time series model

fitting

for the error

6, = 0.46305

1

6, = 0.43035,

2 4

s2

6

P

3

structure

6, ~0.43124, 6,=0.43051,

& = 0.07063

&,=0.07605,

&2=0.08041,

&=0.01260

&=0.01212,

&,=0.05733

HQ(P)

0.23136

~ 1.4304

0.23021

- 1.4187

0.23017

- 1.4022

0.22942

~ 1.3888

174

Y.K.

Truong

/ Semiparametric

repeated

measurement

models

Remark 4.1. Different values of 6, were used to study the effect of bandwidth on fitting parametric model to the residuals. The value of b^, changed from 0.44784 to 0.47061 when the bandwidth 6, was chosen from 0.15 to 0.18. Here any 6, E (0.15,O. 18) would yield a reasonable 0.20 produced under and oversmooth HQ was consistent-it 6,

smooth

estimates,

estimate

of 19(. ), as

respectively. picked first

the

0.12 hand,

AR

for

(0.12,0.20). 4.2.

random of males, 35-59, selected parlowering with center in ment Biostatistics University North (LRC-CPPT, In study, cholesterol were over period 7.6 so effectiveness cholestyramine drug lowering can evaluated. set measurements also by (1989) cancer Figure shows bimonthly of participant, ted the interval. observations then with bandwidth 0.10. was chosen eyes.) examine correlated the described Section was to residual (Figure by AR(l), . , AR(4). Results are summarized in Table 2, which indicates that serial correlation is present in the measurements. In fact, based on the Hannan-Quinn criterion, Table 2 suggests that the residuals may be best described by an AR(2) model. a

0.4

0.0

Fig. 3. Scatter

0.0

plot of 46 cholesterol

measurements.

0.2

Fig. 4. Residuals

0.4

plot of r;-

0.8

The data were smoothed

0.6

B,(t;), i=O, 1,

with bandwidth

0.8

,45.

1.0

= 0.10.

1.0

175

Y. K. Truong / Semiparametric repeated measurement models Table 2 Parametric

residuals

analysis

of cholesterol

measurements

6

P

b^, = ~ 0.28356

1

b^, = -0.41485,

2

t;z = -0.46303

6, = -0.50007, &= -0.53938, &= -0.18405 6, = -0.52286, 62= -0.60615, & = -0.24595, l& = PO.12379

3 4

s2

HQ(P)

312.71

5.8620

245.67

5.6791

237.35

5.7030

233.71

5.7459

5. Conclusion In this paper, a semiparametric model is proposed for analyzing repeated measurement data. The mean function is nonparametric so that it has greater flexibility in fitting various data sets. A parametric error process is imposed in response to the fact that not only serially correlated data arise quite often in growth curve analysis, it is also an important issue to estimate the correlation structure. Under some reasonable conditions, the proposed procedures for estimating the parametric and nonparametric components are shown to possess the desired optimal properties. In particular, the limiting distribution of the parametric estimators-parametric time series inference-may be useful for examining correlated structures in repeated measurement analyses. However, for analyses based on finite samples, one has to provide a bandwidth for the local average estimator. This can be accomplished by (1) trying a few bandwidths and select the ‘best’ one by eyes; (2) using a method proposed by Chiu (1989). An interesting question is: If a sequence of local average estimators is constructed with bandwidth chosen by Chiu’s method, do the optimal properties described in Section 2 still hold? (Chiu (1989) viewed the nonparametric mean function as a nuisance parameter and hence did not address rates of convergence in nonparametric estimation.)

6. Proofs

&, la,l*. For tE[O, 11, set K,(t)=

In the

following

an+ will be denoted ll,t,_,,56,,l, i=O, l,...,

Lemma

2.2 holds.

6.1. Suppose

Var

Proof.

discussion,

Condition

simply by a,,. Set Ilall*:= n. Then C, K,(t)=O(n&,).

Then

C K,(t)[ I$- t9(t,)] = O(n6,). > c I

Set U,= I’-O(t;)=

C, ai-u~,.

By Condition

2.2,

176

Y. K. Truong / Semiparametric repeated measurement models

Var

C K.(t)U-

(LL

+ 2 C C COV(Kj(t)ui,Kj+j(t)u,,j)

= C E[K;(t)U;12 = O(nd,)a2

+2 C C rO(nfi,)6: =

ij

llaj12

C C ai-ua;+~-uCOV(~u~~~)

Ki(t>K;+j(t)

,,a,,2 +2a2

c i;(J, i

c

lajl c

J

u

laJ

O(n&)

as desired. Lemma 6.2. Suppose that Conditions 2.2 and 2.3 hold. Then there is a positive constant c, such that

P

C K,(t)[I:-B(t,)] (I i

52exp(-c2c,ndi),

c>O.

The proof

k random cumulant,

depends on the following notion. Let (Z,, . . . , Z,) denote a vector of variables with EIZJlk
. . . , Z,) = c (-1y-l(p-

l)!E

n (;G”,

zJ)‘*-E(Jiipq)

where the sum extends over all partitions { vl, . . . , vp}, p = 1,2, . . . , k, of (1, . . . , k}. of iks, “..sk in the Taylor Equivalently, cum(Z,, . . . , Z,) is given by the coefficient series expansion of log E[exp(i C: ZjSJ)] about the origin (Brillinger, 1981). An important special case of this definition occurs when ZJ = Z, j = 1, . . . , k. This then gives the cumulant of order k of a univariate random variable. Examples: (a) cum(Z,) = EZ,. (b) cum(Z,, Z,) = Cov(Z,, Z,) = E(Z, Z,) - EZ,EZ,. (c) cum(Z1, Z,) = Var(Z,). - EZ,Z,EZ, - EZ,Z,EZ, + (d) cum(Zt, Z2, Z,) = EZt Z2 Z, - EZ,Z,EZ, 2EZ, EZ, EZ, . (e) For Gaussian stationary processes, cumk =0, for k23. See Brillinger (1981) for additional methods on computing cumulants. Proof of Lemma 6.2. As in Lemma 6.1, Ui := q- B(t,), i= 1, . . . , n. By Condition 2.3 and Theorem 2.3.l(iv) of Brillinger (1981), the absolute value of the k-th cumulant of C, K,(f)CI, has the form

Y.K.

= Therefore

Truong

/ Semiparametric

K,(r)Li,)j

6.1, there is a positive E [ exp ( sCK(t)U ; ,

Set s=cS,/c,.

Then P

=.s”Var(+ constant

models

117

there is a positive

7 K,(i)U,).

c,, such that

,)] =exp(s’Var

C K,(t)U,rcndfj

(

measurement

:

logE[exp(cC By Lemma

repeated

constant

(+ F K,(t)L’,)) c1 such that


i

~exp(&s2n~,).

7 Ki(r)U;jl

> 5 exp( - scnf3t, + +s2c@,,) 5 exp( - c2c,n8:).

Similarly, P

C K,(t)U,s

-cnf_St,

i

(

Iexp(-c2crn8i). )

Proof of Theorem Condition 2.1,

2.1.

je(t;)-e(t)l

t e [0, 11, set Z,,= I,,(t) and N, =N,,(t).

Given

sA4()Bn,

According

to

iEI,.

Thus N,’

F [@t;)-@(t)] ,I

Note that there is a positive Chebyshev’s inequality,

=O(K’).

constant

rVar

(6.1)

c2 such that N,, 1 c2nJn. By Lemma

(cn~‘n~,)2=0(1)

6.1 and

as c-w.

i (Note that 6, - n -‘.) Therefore N,’

;

[I’-Qt;)]

=0&P).

(6.2)

II The conclusion

of the theorem

follows

from (6.1) and (6.2).

Y. K. Truong / Semiparametric

178

Proof of Theorem

2.2.

le(tj)-e(t)l

By Condition SM,

If;-tl

repeated measurement

models

2.1, IMa&,

iEZ,@),

tE[O, I].

Thus N,(t))’ Set z,(t)

c [W;)-e(t)] I,,(0

= CiE,n(f’ [I:-e(ti)]. E[Zi(t)]

IM()6,,

By Lemma

= O(nd,)

te [O, 11.

ieZ,(t),

(6.3)

6.1, over t e [O, 11.

uniformly

Consequently, dr=O(n&). for t E [0, 11. Hence

Recall that N,(t)rc&,

(6.4)

by (6.4) and Markov’s

inequality

P

(6.5) It follows

from (6.3) and (6.5) that lim limP(I~8,-el~2~c(~n+(n-‘6,1)“2))=0. c+rn n

The conclusion or equivalently,

of the theorem 6, = n Pr.

Proof of Theorem 2.3. L, 1 [,;@+r’ log n]. Let whose coordinates is of [0, I] can be written as and all of its vertices in w such that t E Q,v. Let

now follows by choosing

6, so that 6, = (nm’s,‘)“2,

Let T be a positive constant such that 0< t< 1 and set W, be the collection of (L, + 1) points in [O, l] each of the form j/L,, for some integer j such that 01j~ L,. Then the union of L, subintervals, each having length A,, = L,’ W,. For each t E [0, l] there is a subinterval Q, with center C, denote the collection of centers of these subintervals.

Then P (

Izbpl, 1$(t)-e(t)1 >

=P ( It follows large)

ZC(H-'

i0gn)'

>

~~7 ,“~“t,p&(t)-e(t)1 n n

from An- 6:+‘/log le(t)-e(bq

n = o(6,)

rM,lt-wl

Therefore, to prove the theorem, tant c such that

IC(K'

>

and Condition

244,6,, it is sufficient

.

i0gn)r

tEQw,

2.1 that (for n sufficiently WEC,.

to show that there is a positive

cons-

Y. K. Truong / Semiparametric

repeated measurement

FEa; I”EU~,10n(r)-O(w)l 1imP n n ,* ( Set r,=In(w)= 8(w)=ave(

179

models

=O.

2c(n-‘logn)’

(6.6)

>

{i: 1
Na=Nn(w)= from

I$: iota},

#f,(w)

=0

and

(6.7)

i and ~,“c” /On(w)-d(w)l limp n 1, (1

)

To verify (6.7) and (6.8), set yn=iJIfl(w)= are positive constants cs and c, such that N~(w)-~,,(w)~cs~,~+~ By Lemma

and

6.2, there is a positive

(6.8)

=0.

~c(n-~logn)'

constant

#{i:

It,- WI 16,-A,}.

Nn(w)~c4ndn

Then

there

for all WEC,.

cs such that

P (I Note that there is a positive

constant

K

such that

#c, 5 nK. Hence

P

5 2n” exp( - c2c,na i). Since nS3, -log P Consequently,

n, we conclude NR(w)-’

that,

for c sufficiently

C [ x-O(t,)] i,(w)

large, 52nKexp(-c2

log n).

for c2 > K, (6.9)

Observe

that (6.8) follows

from (6.3) and (6.9).

Given t E [0, 11, set N, = N,(t) y,,
Thus

and Z, = I,(t)

and choose

w such that t E Q,,. Then

Y.K. Truong / Semiparametric

180

repeated measurement

models

and hence

Since

# C, I nK, it follows

from Chebyshev-type

5

inequality

P

involving

max max ( WEC,, I s,sn

=O(l)n’+KE/Y,I for k sufficiently theorem.

large.

Hence

(6.7) is valid.

This

E j Y, 1k that

2CsS;‘+r k

>

‘Lk (log n)k = O(l)

completes

the proof

of the

for all zE@ such that 1~15 1, Proof of Theorem 3.1. Since 1-6rz... -b,zp#O it follows from Theorem 3.1 .l of Brockwell and Davis (1987) that there exists a sequence (ai)ick such that C, Ia,1 < 00 and z;=

c a,_&,,. U

Thus by Theorem

2.3,

II&C.) -

et.)lI, = qm-’

(6.10)

log(n)I

Note that z;=

Y--&Jr;)

=z;-

(e,(t,) - &t;)).

Thus -~i+h~8n(~i)-~(~;)l-~;~B,(f,+~)--(~,+,)l

4Z+h=Z,Zi+h

+]e,(~,)-~(t,)l]&(~;+,)-O(t;+h)l, By Condition

(Proof

h=O,l,...,p-1.

2.1, (3.1) and (3.2),

n-’

c z;+,[&(t,) - fw,)]

.-’

c z;[&(t;+,)-W;+/J]

=

of (6.12) will be given shortly.)

O,((n~’ log n)lr), =0&KI

(6.12)

1ogn)“r).

Set

n-h

c

&(h)=(n-h)_’

h C .Z?,.Z!i+h, i=l

II

Z;Z;+h

and

i=l

9,(h)=(n-h))’

h=O,l,_.., p-l. Then

(6.11)

by (6.10)-(6.12),

n-h p);,(h) n

n-h = ~~~(h)+0,((n-110gn)2r), n

h=O,l,...,p-1.

Y. K. Truong / Semiparametric

That

repeated measurement

181

models

is, 8,(h)-~,(h)=o,(n~“2),

h=O,l,...,p-1.

If 6 denote the estimator of b obtained Yule-Walker estimator of b. Then

by regressing

Z, on Zi_,, . . ..Z._,,

i.e. the

b^-6=o,(n-“2). Consequently,

by Proposition

8.10.1 of Brockwell

and Davis (1987),

fi(b-b)d(O,o’T,‘). Proof of (6.12). Note that

W,.hN

n-l c Z;[&(t;+,)-

+n-’

c z,

N

n

(i;.,,, k jxt _) te(fk)- Wi+/?>l. E ,I / ii

(6.13)

Let m = m, = InS,,]. Then 2m-l5N,(t;)l2m+l, Put y(i)=E(Z,Z,+j),

i=O,l,...,

n.

(i=O, i: l,... ). By Theorem

qn-1p(

l

2.3.2 of Brillinger

(1981),

Zkj1_

c

~n@i+h)kEI,tU,+h)

+

Idi-0

W-A

=$ (,j;n(l-nml Ii11 lY(i)l)2=O($), since

Ci ly(i)l < 03. By m = [nS,] - log(n)(n-’

‘-’

7 “(&(:i+i,)

k&+/J

‘k

>

log(n)))2’

and Markov’s

inequality,

Y.K.

182

=

Truong

/ Semiparametric

O,((K’ log n)2’),

repeated

measurement

models

h = 0, 1, . . . ,p - 1.

(6.14)

This takes care of the first term of (6.13). The second term will now be considered. According to Condition 2.1,

M;Jt,

c EZ;+A4;6;

c ;

c lE(Z;ZJ J

= O(ns&

(6.15)

as 1; C, lE(ZfZj)l 5 Ci C, laiPul Cj laj_ulo*Sn ll~ll’o*. It now follows from (6.13)~(6.15) and 6, - (n-l log n)’ that (6.12) holds. This completes the proof of the theorem.

Acknowledgements The author is grateful to thank Chuck Stone,

to David Brillinger for the use of DRBLIB. Also, he wishes Dennis Cox, Editors and referees for helpful comments.

References Brillinger,

D.R.

Brockwell,

P.J.

(1981).

Time Series: Data Analysis

and R.A.

Davis (1987).

Chiu, S.T. (1989). Bandwidth

selection

Time Series:

and Theory. Theory

for kernel estimate

Holden-Day,

and Methods.

with correlated

San Francisco,

CA.

Springer,

New York.

noise. Statist.

Probab.

Lett.

8, 347-354. Gasser, T., H.G. Miiller, W. Kchler, L. Molinari and A. Prader analysis of growth curve. Ann. Statist. 12, 210-229. Fuller,

W.A.

(1976). Introduction

to Statistical

Time Series Analysis.

Hannan,

E.T.

(1970). Mulfiple

Time Series. Wiley,

Hannan,

E.T.

and B.G. Quinn

(1979). The determination

Statist. Sot. Ser. B 41, 190-195. HBrdle, W. and D.T. Pham (1986). Some theory

(1984).

Nonparametric

Wiley,

New York.

New York. of the order

on M-smoothing

of an autoregression.

Stark-r. Assoc.

measurements

7,

data.

J.

81, 1080-1088.

Kritchevsky, S. (1989). Lipid changes ment of Epidemiology, University

preceding the diagnosis of cancer. Doctoral of North Carolina, Chapel Hill, NC.

LRC-CPPT

Clinics

(1984).

J. Roy.

of time series. J. Time Ser. Anal.

191-204. Hart, J.D. (1989). Kernel regression with time series errors. Preprint. Hart, J.D. and T.E. Wehrly (1986). Kernel regression estimation using repeated Amer.

regression

Lipid

Research

Program.

Lipid

research

clinics

Dissertation,

primary

Depart-

prevention

trial

results: I. reduction in incidence of coronary heart disease. J. Amer. Medical Assoc. 251, 351-364. Liitkepohl, H. (1985). Comparison of criteria for estimating the order of a vector autoregressive process. J. Time Ser. Anal.

6, 35-52.

Y.K.

Miiller,

H.G.

Statist. Miiller,

Truong

and U. Stadtmiiller

/ Semiparametric

(1987). Estimation

measurement

of heteroscedasticity

183

models

in regression

analysis.

Ann.

15, 610-625. H.G.

and

Biometrika

U.

Stadtmiiller

(1988).

Detecting

C.J.

dependencies

in smooth

regression

models.

75, 639-650.

Roussas, G.G. (1988). Nonparametric regression Seber, G.A.F. and D.J. Wild (1989). Nonlinear Stone,

repeated

(1980).

Optimal

rates

estimation Regression.

of convergence

for

under mixing conditions. Wiley, New York.

nonparametric

estimators.

Preprint. Ann.

Statist.

8,

Ann.

Staiist.

10,

1348-1360. Stone,

C.J. (1982). Optimal

global

rates of convergence

for nonparametric

regression.

1040-1053. Truong,

Y.K. and C.J.

pear in Ann. Wong,

W.H.

Stat&r.

Stone (1988). Nonparametric

function

estimation

involving

time series. To ap-

Statist. (1983). On the consistency

11, 1036-1041.

of cross-validation

in kernel nonparametric

regression.

Ann.