Content uploaded by Venus Khim-Sen Liew

Author content

All content in this area was uploaded by Venus Khim-Sen Liew on Apr 21, 2016

Content may be subject to copyright.

1

An overview on various ways of bootstrap methods

Venus Khim-Sen Liew

Department of Economics, Faculty of Economics and Business, Universiti

Malaysia Sarawak, 94300 Kota Samarahan, Sarawak, Malaysia.

Abstract

The introduction of the bootstrap methods by Efron (1979) enables many

empirical researches, which would otherwise be difficult if not totally impossible.

Nowadays, bootstrapping has become an important aspect in research. This

paper reviews various ways of bootstrapping data for cross-sectional and time

series samples. Various ways of bootstrapping confidence intervals for

estimators, an important application of bootstrap methods, are also discussed in

this paper. Several other applications of bootstrap methods are briefly mentioned

preceding to the concluding remarks this paper.

1. Introduction

Statistical inference plays a dominant role in empirical research. Researchers use to draw

general conclusions and recommend policy suggestions based on the inference they made

from their statistical estimates of available sample data, which are usually limited. Thus,

getting accurate statistical estimates as well as reliable statistical inference inevitably is

an obligatory requirement for fine empirical research. Specifically, in the field of

economic research, many test-statistics for hypothesis testing intention, including the

commonly exploited t-statistics and F-statistics, imposed normality assumption. This

2

strong assumption is easily violated in reality as most economic data series are found to

possess non-normal empirical distribution functions (EPF), let alone their unknown

population probability distribution functions (PDF). As such, be oblivious to its

importance, statistical inference can be a formidable wall of empirical research. By the

introduction of the bootstrap method (Efron, 1979) and the dissemination of a series of

related study (Efron, 1981; Efron, 1985; Efron, 1987; Efron, 1990; Efron and Tibshirani,

1985; Efron and Tibshirani, 1986, just to name a few), Bradley Efron not only bulldozes

this wall, but also opens up alternative road to explore many of the otherwise inaccessible

area of research1.

Bootstrap is a computer-intensive method of statistical inference that can answer many

real statistical questions. Briefly, bootstrap enables researchers to estimate the precision

of sample statistics (such as medians, variances, percentiles) by drawing randomly with

replacement from a set of available data. Bootstrap method is very useful in

circumstances where the asymptotic distributions of the test statistics of interest are

unknown or statistically too complicated to derive. Besides, in cases where normality

assumption have been violated (as mentioned above) thereby invalidate the use of

conventional standard errors, confidence intervals as well as t-statistics, bootstrap method

offers us an alternative procedure in estimating these statistics of interest by resampling

with replacement the original finite sample, which has been branded as bootstrapping

nowadays. Practically, bootstrapping is similar to that of a Monte Carlo simulation, with

an essential difference in the generation of the random variables (stochastic errors). The

1 Indeed, the contributions of Bradley Efron are so significance that it is not unreasonable for one to predict

that he will one day be awarded the Nobel Prize.

3

latter generates the random variables from a given distribution (assumed known in priori)

such as the Normal, chi-squared, Student-t and F-distributions, whereas the bootstrap

drawn the random variables from the empirical distribution function. Essentially, it is this

simple method but powerful so-called plug-in principle (Efron and Tibshirani, 1993) –

the use of empirical distribution function as an estimate of their actual distribution – that

enables the estimation of reliable parameters from finite samples. See also Davidson and

MacKinnon (2004, Section 4.6) for more discussion that is reasonably accessible on the

differences between Monte Carlo simulation and bootstrapping.

Bootstrap methods are not only useful in cases whereby certain assumptions (on the

distribution of disturbance terms, for instance) are clearly not met or even unknown, but

also crucial for the applications of new test statistics (for instance, nonlinear unit roots

tests, panel unit root tests) in which the complete set of critical values is unavailable yet.

As bootstrapping is almost unavoidable in today’s research that involves statistics, it is

important to understand how bootstrap works. In the light of this, this paper reviews the

various ways of bootstrapping sample data in the literature. From this bootstrap sample,

one may proceed to compute the statistics of interest, and whenever necessarily (such as

in the case of computing bootstrap confidence intervals, critical values and the like), the

bootstrap procedures discussed below may be replicated as many times as needed.

Various ways of bootstrapping confidence intervals will be discussed in this paper, as

illustration on the usefulness of bootstrap methods. To keep the idea simple and to ensure

a smooth flow of discussion, this paper focuses on the illustration of computing the

bootstrap linear regression coefficients.

4

The remainder of this paper is structured as follows. Section 2 describes the regression

model that spins the bootstrap sample data. Two ways of bootstrapping cross-sectional

data are respectively discussed in Sections 3 and 4. Time series model that generates

time series data is discussed in Section 5. This is followed by Section 6, which deals with

bootstrapping residuals method for time series data. Two alternatives, namely

bootstrapping block method for time series data and bootstrapping moving block method

for time series data are mentioned in Sections 7 and 8 respectively. One common

application of bootstrapped sample is to compute bootstrap confidence intervals. Various

ways of bootstrapping confidence intervals are discussed in Section 9. The final section

contained concluding remarks.

2. The regression model

Consider the regression model:

i

y

=

jji

xf

,(

) +

i

(1)

where

i

y

is the dependent variable,

jji

xf

,(

) may be either linear or a nonlinear

function of

ij

x

and

j

, where

ji

x

(

pj ,...1

and

Ni ,...,1

where

N

is in turn the

sample size) in turn is the

thi

observation of the

thj

explanatory variable which is fixed

in repeating sampling and

j

is the unknown parameters to be estimated.

i

is the error

terms, which is traditionally assumed to be identically and independently distributed with

commonly distribution G with zero mean and finite variance (

2

).

5

One form of linear regression model is given by

jji

xf

,(

) =

ji

p

jjx

1

+

i

,

Ni ,...,1

(2)

which sometimes may be more conveniently expressed in the following matrix form:

XY

(3)

where Y =

'

1),...,(N

YY

is a

1N

column vector of dependent variable,

X

denotes

pN

matrix of explanatory variables,

=

'

1),...,(p

is a

1p

column vector of unknown

parameters and

=

'

1),...,(N

is a

1N

column vector of

N

error terms.

The conventional least square (OLS) estimate

ˆ

for

is given by

ˆ

=

YXXX '

1

'

(4)

It is a textbook case that under the classical assumptions

ˆ

has mean

and variance-

covariance matrix

1

'2

XX

. Further, it can be shown that (see Freedman, 1981, for

example) the asymptotical distribution of

)

ˆ

(

N

is normal with mean zero and

variance covariance given by

1

'2

XXN

.

6

In practice, however, both G and

2

are unknown. In such case, Efron (1979) proposes

the use of empirical distribution function

G

ˆ

of the centered residuals of Equation (1) in

place of the unknown true G2. As such,

G

ˆ

will have the desire expectation zero. Note

that it is critical to center the residuals otherwise bootstrap may fail (Freedman, 1981,

Davidson and MacKinnon, 2004). Besides, it is also common to rescale the empirical

residuals by a factor of

)/( kNN

to give the desired homogeneous variance (Davison

and Hinkley, 1997, Section 6.2)3, 4.

3. Bootstrapping residuals method for cross-sectional data

Efron (1979) originally proposes to bootstrap a random sample of size n drawn with

replacement from the centered residuals such that places a probability of

N/1

on each of

the residuals by construction. Given

G

ˆ

, we can generate another sample of

Y

, known as

the bootstrap sample, denoted by

*

Y

for the estimation of

. The resulting estimated

is known as bootstrap

, denoted by

*

ˆ

. This bootstrapping procedure is known as

bootstrap residuals method and its algorithm is given as follows5:

2 This is especially applicable if no intercept is included in the regression model (Berkowit and Killian,

2000).

3 If there is reason to suggest that the residuals variance is proportional to

Xx

, then the

thi

empirical

residuals may be rescaled by the square root of the

thi

observation of

x

(Hinkley, 1998).

4 See also Masarotto (1990), Li and Maddala (1996), Kim (1999), Berkowitz and Killian (2000), Lam and

Veall (2002), Alonso et al. (2003) and Christopher and Zhou (2003) for additional information on

autoregression models.

5 In cases when

jji

xf

,(

) is called upon, all the bootstrap procedures discussed in this section are just

valid provided the OLS estimation process is substituted by Nonlinear Least Squares (NLS) or Maximum

Likelihood (ML) method.

7

Algorithm 1 (Bootstrapping Residuals Method for Cross-sectional Data)

1. From the observed sample of

Y

and

X

, obtain

ˆ

as in Equation (4).

2. Compute empirical residuals

)

ˆ

,(

ˆ

XfY

.

3. Transform the residuals

ˆ

of Step 2 into centered residuals

~

. Practically,

compute

N

iiii N1

ˆ

1

ˆ

~

, where

i

~

and

i

ˆ

are the

thi

element of vectors

~

and

ˆ

respectively.

4. Randomly draw a sample

*

, known as bootstrap residuals from

~

with

replacement.

5. Generate

*

Y

from the equation

*

Y

=

)

ˆ

,( jji

xf

+

*

=

*

ˆ

X

.

6. Regress

*

Y

on

X

to obtain OLS

*

ˆ

given by

*

ˆ

=

*'

1

'YXXX

.

It has been shown that the distribution of the bootstrap will give the same asymptotical

results as classical methods (Freedman, 1981). In particular,

)

ˆ

(*

N

which can be

computed directly from the sample data approximates the distribution of

)

ˆ

(*

N

,

provided

N

is large and

12 )(

XXtracep T

is small6.

4. Bootstrap cases method

In Algorithm 1,

*

Y

values are bootstrapped from the preliminary estimated model (4).

The idea is to fit a suitable model to the data, to recenter and resample the model’s

residuals and to generate new series by incorporating the resultant residuals into the fitted

model. Thus, this method is also known as model-based bootstrap method. One may

bootstrap cases of

Y

and

X

values using the bootstrapping cases method, instead of just

the

Y

values as in the bootstrapping residuals method. This method does not rely on the

prior information of the relationship between

Y

and

X

variables as well as assumption of

6 The trace of a square matrix is the sum of its diagonal elements.

8

residuals. The procedures are much more simplified and time-saving. This bootstrap

cases method involves resampling of cases of (

XY,

) with replacement from the original

series, and then computing the

*

from the resampled (

*

Y

,

*

X

) data. The algorithm is

given as:

Algorithm 2 (Bootstrap Cases Method)

1. From the observed sample of

Y

and

X

, bootstrap N cases of (

*

Y

,

*

X

)7. As such,

we have (

*1

*

21

*

11

*

1,...,,, p

xxxy

), (

*2

*

22

*

12

*

2,...,,, p

xxxy

),…,(

**

2

*

1

*,...,,, pNNNN xxxy

).

2. Regress

*

Y

on

*

X

to obtain OLS

*

ˆ

given by

*

ˆ

=

*

'

*

1

*

'

*YXXX

.

As in the case of model-based bootstrap, the OLS

*

ˆ

obtained in Algorithm 2 also

approximate the properties of the true

(Davison and Hinkley, 1997, Section 8.2.2).

5. Time series model

Both Algorithms 1 and 2 are valid in the context of cross-sectional data only. However, it

is widely known that most economic research involved the use of time series data. In this

regards, the model-based bootstrap may be slightly modified to suit the requirements of

time series, whereas the bootstrapping cases method is helpless as it destroys the serial

dependency property of the time series. To circumvent this problem, bootstrapping block

method has been introduced as an alternative the model-based bootstrap method.

7 If there is only one

X

variable, we have pairs of (

Y

,

X

) and the method is accordingly regarded as

bootstrap pairs method.

9

Suppose the series

Y

is determined endogeneously by its own lagged values, given by the

following autoregressive representation:

t

y

=

jjt

xf

,(

) =

it

p

iiyy

1

0

t42

pi ,...,1

,

Tt ,...,1

(5)

where

0

y

is an intercept parameter, which take the value of zero if the series

t

y

(

Tt ,...,1

where T is the sample size) has a zero mean, otherwise it will be a non-zero

constant.

i

’s (

pi ,...,1

where p is the optimal autoregressive lag length) are known as

autoregressive parameters.

Model (5) is called AR(p) model, which stands for autoregressive model of order p8. It

may be compactly rewritten in the matrix form as:

))((YYLA

(6)

where

Y

=

T

tt

y

T1

1

is the mean value of

Y

and

)(LA

)...1( 2

21 p

pLLL

is an

invertible polynomial9 in the lag operator with

i

L

(

pi ,...,1

) is the backshift operator

such that

i

B

t

y

=

it

y

.

8 See Berkowitz and Killian (2000) for the extension of bootstrap algorithm for AR(p) model into the

general autoregressive moving average ARMA(p, q) models of the form

)())((LBYYLA

, where

)(LA

and

)(LB

are invertible polynomials in the lag operator.

10

6. Bootstrapping residuals method for time series data

There are two methods of bootstrapping

Y

in the time series context. The simplest is

analogous to the model-based bootstrap method as in Algorithm 1. This method is

described in Algorithm 3 below:

Algorithm 3 (Bootstrapping Residuals Method for Time Series Data)

1. From the observed sample of

Y

, determine the order of the AR(p) model.

2. Estimate

i

y

ˆ

and

ˆ0

(

pi ,...,1

) as in Equation (5).

3. Compute empirical residuals

)

ˆ

,(

ˆ

XfY

.10

4. Transform the residuals

t

ˆ

of Step 2 into centered residuals

t

~

using the formula

T

tttt T1

ˆ

1

ˆ

~

.

5. Randomly draw a sample

*

t

, known as bootstrap residuals from

t

~

with

replacement.

6. Initialized

0,..., **

2

*

1 p

yyy

and generate

*

Y

from the relationship

*

Y

=

*

t

y

=

)

ˆ

,( jji

xf

+

*

t

=

it

p

iiyy

1

0ˆ

ˆ

+

*

t

, for

Tpt ,...,1

.

7. Using

*

Y

, reestimate Equation (5) to obtain OLS

*

ˆ

.

The series so generated may not be stationary due to initialization effect. To overcome

this problem, it is advisable to generate extra so-called ‘burn-in’ observations (Davison

and Hinkley, 1997). It is recommended that we should generate

Tm

observations and

keep only the last

T

observations for latter use by discarding the first

m

observations.

9 An invertible polynomial satisfies the condition

0)( LA

for all

1 || L

(Brockwell and Davis, 1996,

p. 84).

10 One may rescale the empirical residuals by a factor of

)/()( dpTpT

, where

d

stands for

the number of estimated coefficients to give the desired variance (Freedman and Peters, 1984; Stine, 1987).

11

For instance, for

T

ranging from 30 to 1000, Lam and Veall (2002) use

m

= 100. Mean

while Field and Zhou (2003) choose

m

= 100 for

T

ranging from 20 to 1000.

7. Bootstrapping block method for time series data

The second method bootstrapping a time series data is the bootstrapping block method. In

this method, the data is divided into

b

non-overlapping blocks of length

l

, such that

blT

, yielding blocks

b

zzz ,...,,21

, where

'

11 ),...,(l

yyz

,

'

212 ),...,( ll yyz

, …,

'

1)1( ),....,( Tlbb yyz

. The idea is to take a bootstrap sample with equal probabilities

b/1

from the

i

z

(

),...,1bi

, and then pool these end-to-end to form a new series.

Algorithm 4 (Bootstrapping Block Method for Time Series Data)

1. Divide observed sample of

Y

=

'

1),...,(T

yy

into blocks

'

11 ),...,(l

yyz

,

'

212 ),...,( ll yyz

, …,

'

1)1( ),....,( Tlbb yyz

. This yields

b

non-overlapping

blocks of length

l

, such that

blT

11.

2. Bootstrap

*

Y

=

'*** ),...,,( 21 b

zzz

from

Y

(

'

21 ),...,, b

zzz

.

3. Using

*

Y

, estimate Equation (5) to obtain OLS

*

ˆ

.

Bose (1998) shows that model-based method improves the asymptotic properties of the

estimate of the distribution of OLS estimates in the autoregressive model, whereas Smith

and Field (1993) show, through a simulation study, that the performance of the block

bootstrap is quite dependent on the window size. The latter finding is in contrast to

Künsch (1989) and Bühlman (1994) who have proven that the block bootstrap provides a

valid approximation to the unknown distribution of the statistics given by smooth

11 However, it is likely that

lT /

is not an integer. In such case the last block will be shorter than l.

12

functions of normalized sample mean. In this respect, Davison and Hinkley (1997,

Section 8.2.3) argue that if the length of the blocks in the bootstrapping block method is

long enough, the form of dependency in the original data will be preserved faithfully.

Thus the estimated OLS

*

ˆ

will have approximately the same distribution as those

obtained from the model-based bootstrap method. One may choose

pl

in order to

optimally preserve the dependency.

8. Bootstrapping moving block method for time series data

Another way to preserve the original dependency is to divide the original sample into

overlapping blocks. The resulting blocks will be something like

'

11 ),...,(l

yyz

,

'

122 ),...,(

l

yyz

, …,

'

111 ),....,( TNlN yyz

. Due to the way the sample is blocked, this

method is referred to moving blocks bootstrap. Algorithm 4 is still applicable with the

above modification incorporated in the first step. This modification ensures that each

original observation has a more equal chance of appearing in the simulated series. It also

has the advantage of removing the minor problem with the non-overlapping procedure

that the last block is shorter than the rest if

lT /

is not an integer. The moving blocks

bootstrap method is given as:

Algorithm 5 (Bootstrapping Moving Block Method for Time Series Data)

1. Divide observed sample of

Y

=

'

1),...,(T

yy

into blocks

'

11 ),...,(l

yyz

,

'

122 ),...,(

l

yyz

, …,

'

111 ),....,( TNlN yyz

. This yields

b

=

1 lN

overlapping blocks of length

l

.

2. Bootstrap

*

Y

=

'*** ),...,,( 21 b

zzz

from

Y

(

'

21 ),...,, b

zzz

.

13

3. Using

*

Y

, estimate Equation (5) to obtain OLS

*

ˆ

.

9. Bootstrapping of confidence intervals

Practically, bootstrap methods have a variety of applications in various research fields,

including economics. Besides being used to bootstrap regression models and OLS

estimators as described above, it is most commonly applied in the bootstrapping of

confidence intervals. Three ways of bootstrapping confidence intervals, namely (1)

studentized-t confidence intervals, (2) bootstrap-t confidence intervals, and (3) bootstrap

percentile confidence intervals are discussed as follows.

Suppose

ˆ

is a consistent estimator of

~

))(,( 2

N

. Then the studentized estimator

is given as:

t

)

ˆ

(

ˆ

/)

ˆ

(

(7)

where

ˆ

is the standard deviation of the estimator.

It is possible to approximate this studentized-t by:

*

t

=

)

ˆ

(

ˆ

/)

ˆ

(***

(8)

where parameters with * are obtained through bootstrap method.

14

As such, the

21

studentized-t confidence intervals (based on t table) for

may be

represented by:

(

ˆ

ˆ

,

ˆ

ˆ1tt

) (10)

where

t

denotes the

th) 100(

percentile of the t distribution (tabulated in most statistics

or econometrics textbook) on

1N

degrees of freedom.

Lam and Veall (2002) illustrates via a simulation study the confidence intervals as in (10)

is inaccurate when the disturbances are non-normally distributed and that this inaccuracy

may increase rather than diminish as sample size grows12. This finding is similar to that

of Godfrey and Orme (2000), who have demonstrated that non-normal distributions can

adversely affect the performance of prediction error in the case of regression models.

These empirical findings are in line with Efron and Tibshirani (1993), who point out that

studentized-t confidence intervals fail to account for skewness in the underlying

population or other error that can result when

is not in the sample mean. The latter

propose the bootstrap-t confidence intervals and bootstrap percentile confidence intervals,

which adjust for these errors. The

21

bootstrap-t confidence intervals are defined as:

(

****

1ˆ

ˆ

,

ˆ

ˆ

tt

) (11)

12 Another noteworthy issue addressed in Lam and Veall (2002) is that autoregressive forecast model is

important and in many cases more realistic than approaches that require known values for independent

variables.

15

where

*

t

is the

th

value of the ranked (in the ascending order)

*

t

s values, which are

obtained by repeatedly computing Equation (7) for

B

times. In order to yield robust

confidence intervals,

B

may be as large as 999 or more (MacKinnon, 1999; Davidson

and MacKinnon, 2004). Davidson and MacKinnon (2004, Section 4.6) mentioned that B

has to be sufficiently large; otherwise the power of tests (the ability of the tests to reject

the null hypothesis when it is false) will be negatively affected. Another problem that

may arise due to insufficiently large sample is that the sequence of random numbers used

to generate the bootstrap samples may influence the outcome of the tests. Thus, B=999 is

recommended by Davidson and MacKinnon (2004) as a rule of thumb. The authors also

suggested that if bootstrapping is inexpensive and the outcome is at all ambiguous, it is

desirable to use a large number like 9999. Otherwise, it may be safe to use a value as

small as 99.

In large sample, the coverage of bootstrap-t confidence intervals tends to be closer than

the desired level (for example 10%) than the coverage of the studentized confidence

intervals based on t table. The only shortcoming is that unlike the latter, which applies to

all samples, the former applies only to the given sample. However, with the advancement

in computer technology, it is possible to compute bootstrap-t confidence intervals for

each sample one encounters (Efron and Tibshirani, 1993).

The bootstrap approximation in Equation (11) is at least approximately asymptotically

pivotal (Efron and Tibshirani, 1993). A statistics is said to be asymptotically pivotal if its

16

limiting distribution does not depend on any nuisance parameters (unknowns). In this

regard, many statistics of interest based on AR(p) or ARMA(p, q) models are

asymptotically normal and can be studentized to make them asymptotically pivotal; see,

for example, Berkowitz and Killian (2000). Among others, Li and Maddala (1996) note

that bootstrapping asymptotical pivotal statistics produces more accurate finite-sample

confidence.

A faster way of obtaining bootstrap confidence intervals is to directly replicate

*

, the

bootstrap version of

for

B

times. From the ranked (in the ascending order)

*

, the

21

bootstrap confidence intervals is constructed as:

(

*

1

*,

) (12)

where

*

is the

th) 100(

empirical percentile of the

*

values.

The intervals obtained this way are known as bootstrap percentile confidence intervals. It

has the advantage of by-passing the process of computing

*

t

and thus is time-saving.

Efron and Tibshirani (1993) show that if the distribution of

*

is roughly normal, then

the bootstrap percentile confidence intervals will converge to the studentized confidence

intervals. However, as economic data are non-normal most of the time, the discrepancy

between the two confidence intervals will be substantial in practice. Under such

circumstance, it is obvious that the former is preferable as it follows the empirical

distribution faithfully. In this respect, it is worth pointing out that the two above-

17

mentioned bootstrap intervals may be obtained without having to make normal theory

assumptions. In fact, one does not even have to know the asymptotic distribution of the

statistics of interest!

Empirically, Lam and Veall (2002) find that bootstrap prediction intervals based on either

the percentile principle or the percentile-t principle performs substantially better than the

studentized confidence intervals. This finding is consistent with Kim (1999) who uses the

percentile-t bootstrap method in studying prediction intervals for vector autoregression

models. Apart from this, the small-sample effectiveness of bootstrap method in

constructing hypothesis tests and confidence intervals for the parameters of cointegrating

regressions with autocorrelated errors has also been demonstrated in Psaradakis (2001).

See Davidson and MacKinnon (2004, Section 5.3) for overview on bootstrap confidence

intervals, and Efron and Tibshirani (1993, Chapters 12 and 13) and Davison and Hinkley

(1997, Chapter 5) for a comprehensive discussion regarding the theoretical aspects and

algorithms of constructing these intervals.

10. Concluding remarks

Bootstrap is a computer-intensive method of statistical inference that can answer many

real statistical questions. This paper reviews the various ways of bootstrapping data for

cross-sectional and time series samples. From this bootstrap sample, one may proceed to

compute the statistics of interest, for instance, OLS estimators, t-test statistic and F-test

18

statistic. Perhaps, more importantly, the bootstrap sample may also be utilized to compute

bootstrap confidence intervals, critical values, marginal significance values, and etc.

Chong et al. (2008) and Liew and Yusuf (2008a), for instance, constructed the bootstrap

percentile confidence intervals for the estimated intercept and coefficient of trend term in

their study on income convergence. Other than constructing empirical confidence

intervals, bootstrap method has also been popularly applied in evaluating the performance

of existing test statistics such as unit root test in the works of Harris (1992), Angelis et al.

(1997), Ohtani (2000), Psaradakis (2002) and van Giersbergen (2003). Besides, bootstrap

methods are also commonly applied nowadays in newly proposed test statistics including

linearity tests (Andrson et al., 1999; Chong et al., 2008), nonlinear unit root tests (Sarno,

2001; Sekioua, 2003; Chong et al., 2008; Liew et al. 2008) and nonlinear cointegration

tests (Escribano and Mira, 2001; and Dufrénot et al., 2006). Another usefulness of the

bootstrap methods is to compute the marginal significance values of the test statistics of

interest (Henry, Olekalns and Summers, 2001; Basher and Haug, 2003; Sekioua, 2003;

Liew, 2004, 2008; Liew and Yusuf, 2007, 2008; Liew et al. 2008)13.

To conclude this section, bootstrap methods first introduced by Bradley Efron in 1979 is

very useful in empirical research, especially when the test statistics of interest possess

unknown asymptotical sample distribution. Note that modern researches, especially those

involved in newly developed statistical testing procedures, rely heavily on bootstrap

method for robustness of their findings. In fact, the usefulness of bootstrap has become

13 Just to mention a few examples. Researches, especially those in the field of economics, finance and

econometrics, that apply bootstrap methods are ample in recent literature.

19

very influential in research so much so that authors for textbooks on qualitative method

like Davidson and MacKinnon (2004) have to discuss bootstrap in depth. It is noteworthy

that economic researchers nowadays need to equip themselves with computer

programming technology so as to keep pace with the competitive and yet evolving

research environment. While bootstrapping may be familiar to some, it may be unfamiliar

to many others. Interested readers may refer to Efron and Tibshirani (1993) and Davison

and Hinkley (1997), among others, for comprehensive and reasonably accessible

discussion on bootstrap methods and their applications.

References

Alonso, A.M., Peña, D. and Romo, J. (2003) On Sieve Bootstrap Prediction Intervals,

Statistics and Probability Letters 65, 13 – 20.

Anderson, Michael K., Eklund, B. and Lyhagen, J. (1999) An ARCH Robust STAR Test,

Working Paper Series in Economics and Finance, No. 317, Stockholm School of

Economics.

Angelis, De D., Fachin, S. and Young, G.A.(1997) Bootstrapping Unit Root Tests,

Applied Economics 29, 1155 – 1166.

Basher, S.A. and Haug, A.A. (2003) Nonlinear Cointegration and Purchasing Power

Parity: Results from Rank Tests, Working Paper, Department of Economics, York

University.

Berkowitz, J. and Killian, L. (2000) Recent Developments in Bootstrapping in Time

Series, Econometric Review 19, 1 – 48.

Bose, A. (1998) Edgeworth Correction by Bootstrap in Autoregressions, Annals of

Statistics 16, 1709 – 1722.

Brockwell, P.J. and Davis, A.D. (1996) Introduction to Time Series and Forecasting.

New York: Springer-Verlag.

Bühlmann, P. (1994) Blockwise Bootstrapped Empirical Processes For Stationary

Sequences, Annals of Statistics, 22, 995-1012.

20

Chong, T.T.L., Hinich, M. Liew, V.K.S. and Lim, K.P. (2008) Time Series Test of

Nonlinear Convergence and Transitional Dynamics. Economics Letters, 100, 337-339.

Christopher, F. and Zhou, J. (2003) Confidence Intervals Based on Robust Regression,

Journal of Statistical Planning and Inference 115, 425 – 439.

Davison, A.C. and Hinkley, D.V. (1997) Bootstrap Methods and Their Applications,

Cambridge: Cambridge University Press.

Davidson, R. and MacKinnon, J.G. (2004) Econometric Theory and Methods. Oxford:

Oxford University Press.

Dufrénot, G., Mathieu, L., Mignon, V. and Pégun-Feissolle, A. (2006) Persistent

Misalignments of the European Exchange Rates: Some Evidence from Nonlinear

Cointegration, Applied Economics 38, 203 – 229.

Efron, B. (1979) Bootstrap Methods: Another Look at the Jackknife, The Annals of

Statistics 7, 1 – 26.

Efron, B. (1981) Nonparametric Standard Errors and Confidence Intervals (with

Discussion), Canadian Journal of Statistics 9, 139 – 172.

Efron, B. (1985) Bootstrap Confidence Intervals for a Class of Parametric Problems,

Biometrika 72, 45 – 58.

Efron, B. (1987) Better Bootstrap Confidence Intervals (with Discussion), Journal of

American Statistical Association 82, 171 – 200.

Efron, B. (1990) More Efficient Bootstrap Computations, Journal of American Statistical

Association 85, 79 – 89.

Efron, B. and Tibshirani, R.J. (1985) The Bootstrap Method for Assessing Statistical

Accuracy, Behaviormetrika 17, 1 – 35.

Efron, B. and Tibshirani, R.J. (1986) Bootstrap Measures for Standard Errors,

Confidence Intervals, and Other Measures of Statistical Accuracy, Statistical Science 1,

54 – 77.

Escribano, A. and Jordá, O. (2001) Testing Nonlinearity: Decision Rules for Selecting

Between Logistic and Exponential STAR Models, Spanish Economic Review, 3, 193-209.

Field, C. and Zhou, J. (2003) Confidence Intervals Based On Robust Regression, Journal

of Statistical Planning and Inference, 115, 425 – 439.

Freedman, D.A. (1981) Bootstrapping Regression Models, Annals of Statistics 9, 1218 –

1228.

21

Freedman, D.A. and Peters, S.C. (1984) Bootstrapping an Econometric Model: Some

Empirical Results, Journal of Business and Economic Statistics 2, 150 – 158.

Godfrey, L.G. and Orme, C.D. (2000) Controlling the Significance Levels of Prediction

Error Tests for Linear Regression Models, Econometrics Journal 3, 66 – 83.

Harris, R.I. (1992) Small Sample Testing for Unit Root, Oxford Bulletin of Economics

and Statistics 54, 615 – 625.

Henry, O.T., Olekalns, N. and Summers, P.M. (2001) Exchange Rate Instability: A

Threshold Autoregressive Approach, The Economic Record 77, 160 – 166.

Hinkley, D.V. (1988) Bootstrap Methods, Journal of Royal Statistical Society 50, 321 –

337.

Kim, J.H. (1999) Asymptotic and Bootstrap Prediction Regions for Vector

Autoregressions, International Journal of Forecasting 15, 392 – 403.

Künsch, H.R. (1989) The Jackknife and The Bootstrap for General Stationary

Observations, Annals of Statistics, 17, 1217–1241.

Lam, J.P. and Veall, M.R. (2002) Bootstrap Prediction Intervals for Single Period

Regression Forecasts, International Journal of Forecasting 18, 125 – 130.

Li, H. and Maddala, G.S. (1996) Bootstrapping Time Series Models, Econometric

Reviews, 15, 115 – 195.

Liew, V.K.S. (2004) Nonlinear Adjustment of ASEAN-5 Real Exchange Rates:

Symmetrical or Asymmetrical?. Economics Bulletin, 6(8), 1-19

Liew, V.K.S. (2008) Monetary Models of Exchange Rate: Nonlinearity Matters.

Germany: VDM Verlag.

Liew, V.K.S., Chong, T.T.L. and Lim, K.P. (2003) Inadequacy of Linear Autoregressive

Model for Real Exchange Rates: Empirical Evidence from Asian Economies, Applied

Economics 35, 1387 – 1392.

Liew, V.K.S. and Ahmad, Y. (2007) Long Run Convergence and Catching Up in Income

Convergence: Evidence from the Four Asian Dragons, Journal of International Economic

Review, Vol. 1, No. 2, 121-128.

Liew, V.K.S. and Ahmad, Y. (2008) Income Convergence: Fresh Evidence from the

Nordic Countries, Applied Economics Letters, iPrint, 1 – 4. DOI:

10.1080/13504850701367205.

22

Liew, V.K.S., Lee, H.A., Lim, K.P. and Lee, H.H. (2008) Linearity and Stationarity of

South Asian Real Exchange Rates, ICFAI Journal of Applied Economics, 7, 48-58.

MacKinnon, J.G. (1999) Bootstrap Testing in Econometrics, A paper presented at the

CEA Annual Meeting (29 May 1999). Queen’s University at Kingston.

Masarotto, G. (1990) Bootstrap Prediction Intervals for Autoregressions, International

Journal of Forecasting 6, 229 – 239.

Ohtani, K. (2000) Exact and Bootstrap Distributions of a Unit Root Test, Applied

Economics Letters, 7, 463 – 466.

Psaradakis, Z. (2001) On Bootstrap Inference in Cointegrating Regressions, Economics

Letters, 72, 1-10.

Psaradakis, Z. (2002). Bootstrap Tests for an Autoregressive Unit Root in the Presence of

Weakly Dependent Errors, Journal of Time Series Analysis, 22, 577 – 594.

Sarno, L. (2001) The behaviour of US Public Debt: A Nonlinear Perspective, Economics

Letters 74, 119 - 125.

Sekioua, S.H. (2003) The Nominal Exchange Rate and Monetary Fundamentals:

Evidence from Nonlinear Unit Root Tests, Economic Bulletin 6, 1 – 13.

Smith, B. and Field, C. (1993) Variance Estimation for Quadratic Statistics, Journal of

Time Series Analysis, 14, 381 – 395.

Stine, R.A. (1987) Estimating Properties of Autoregressive Forecasts, Journal of

American Statistical Association 82, 1072 – 1078.

Van Giersbergen, N.P.A. (2003) A Note on Bootstrapping Unit Root Tests in the

Presence of a Non-zero Drift, Economics Letters 78, 259 – 265.