Privatization and Economic Growth
-
Upload
anisur-rahman-faroque -
Category
Documents
-
view
261 -
download
2
Transcript of Privatization and Economic Growth
This article was downloaded by:[Hiroshima University]On: 25 July 2008Access Details: [subscription number 789277225]Publisher: RoutledgeInforma Ltd Registered in England and Wales Registered Number: 1072954Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
Journal of Development StudiesPublication details, including instructions for authors and subscription information:http://www.informaworld.com/smpp/title~content=t713395137
Privatisation and economic growth in developingcountriesPaul Cook a; Yuichiro Uchida aa Centre on Regulation and Competition, Institute for Development Policy andManagement, University of Manchester, UK.
Online Publication Date: 01 August 2003
To cite this Article: Cook, Paul and Uchida, Yuichiro (2003) 'Privatisation andeconomic growth in developing countries', Journal of Development Studies, 39:6,121 — 154
To link to this article: DOI: 10.1080/00220380312331293607URL: http://dx.doi.org/10.1080/00220380312331293607
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article maybe used for research, teaching and private study purposes. Any substantial or systematic reproduction,re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expresslyforbidden.
The publisher does not give any warranty express or implied or make any representation that the contents will becomplete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should beindependently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings,demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with orarising out of the use of this material.
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
Privatisation and Economic Growth inDeveloping Countries
PAUL COOK and YUICHIRO UCHIDA
This article re-examines the relation between privatisation andeconomic growth. Previous studies that have attempted to measurethis relationship have concluded that privatisation has had asizeable positive effect on economic growth. Our study uses datafor 63 developing countries over the time period 1988–97. It usesthe framework of an extreme-bounds analysis (EBA) to conduct across-country growth regression analysis. Our findings contradictearlier results, but reaffirm the view that effective competition andits regulation may need to accompany privatisation to make apositive impact on economic growth.
I . INTRODUCTION
Privatisation accelerated in the 1990s. In Europe this was sparked by theliberalisation of markets at the European Union level and budgetaryconstraints faced by government [Parker, 1999]. Privatisation particularlygained momentum in the late 1980s and spread to a wide range ofdeveloping economies. Over the last decade a significant proportion ofprivatisation transactions have been in developing economies and haveentailed sales of public utilities. Privatisation transactions for the utilitiessector have accounted for over a third of all transactions in developingeconomies since 1988. In the period 1988–93 the value of sales forinfrastructure industries amounted to US$30 billion, compared to US$78billion for all privatisation transactions in developing economies [WorldBank, 1995a, Kikeri, 1998].
There is now sufficient time lapse since privatisation was firstimplemented in developing countries for the results to be evaluated at themacroeconomic level. Admittedly, privatisation proceeded in the 1980s
Paul Cook and Yuichiro Uchida, Centre on Regulation and Competition, Institute forDevelopment Policy and Management, University of Manchester, UK; correspondence: [email protected].
The Journal of Development Studies, Vol.39, No.6, August 2003, pp.121–154PUBLISHED BY FRANK CASS, LONDON
396jds05.qxd 29/08/03 14:13 Page 121
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
without much knowledge of either its impact or contribution to economicgrowth. Instead greater attention was paid to measuring the extent ofprivatisation [Bennell, 1997]. This situation is changing and more recentcontributions assessing post-privatisation performance are reviewed inMegginson and Netter [2001]. Among those most frequently cited are:Galal, Jones, Tandon and Vogelsang [1994]; Megginson, Nash and VanRandenborgh [1994] and Boubakri and Cosset [1998]. However, thesestudies have largely concentrated on financial measures of performance,and have incorporated industrialised countries and higher incomedeveloping countries in their relatively small databases. In recent yearsthere have also been a few attempts to measure the direct impact ofprivatisation on economic growth in a cross-country context [Plane, 1997;Barnett, 2000]. These studies have concluded that privatisation has had asizeable positive effect on economic growth.
The purpose of this analysis is to re-examine the relation betweenprivatisation and economic growth, using data for 63 developing countriesover the period 1988–97. The next section provides a brief review of thetheoretical links between a change in ownership from public to private andeconomic growth, and of the empirical literature that has examined thisrelationship. The third section discusses the methods adopted to explore therelation between privatisation and economic growth and the data used. Thefourth section discusses the results of the analysis. The final sectionprovides a summary and draws some broad conclusions.
II . THEORY AND EMPIRICAL ANALYSIS
Policy-makers in developing economies have often set a broader agenda forprivatisation than the efficiency and resource allocation objectives that wereimplicit in the policy conditions of structural adjustment programmes. Themotives for privatisation have encompassed improved fiscal equity anddistributional performance, although the importance attached to each hasvaried between and within countries over time [UNCTAD, 1996; Yarrow,1999]. Nevertheless, the link between privatisation and economic growthrelates most directly to the microeconomic theories used to justifyprivatisation. At the heart of the debate are theoretical perspectives on theownership issue drawn from property rights theory, public choice theoryand principal agent analysis [Alchian, 1965; Tullock, 1965; Jensen andMeckling, 1976]. By the end of the 1970s these theories were influencingattitudes towards public ownership among policy-makers in developed anddeveloping countries [Cook and Kirkpatrick, 1988; Martin and Parker,1997]. The key theoretical elements underpinning the argument for achange in ownership from public to private related, first, to the view that
122 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:13 Page 122
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
public ownership led to the pursuit of objectives that detracted fromeconomic welfare maximisation [Boycko, Shleifer and Vishny, 1996].Secondly, an ownership change could improve economic performance bychanging the mechanisms through which different institutionalarrangements affect the incentives for managing enterprises [Vickers andYarrow, 1988; Laffont and Tirole, 1991; Cook and Fabella, 2002].
These arguments are linked to presumptions concerning the condition ofpublicly-owned enterprises before they are privatised. A typical viewpresented publicly-owned enterprises as overextended and poor performers[Kikeri, Nellis and Shirley, 1994]. In this situation publicly-ownedenterprises crowded out private enterprises in their access to credit anderected statutory barriers to preserve the monopoly status of publicly-ownedenterprises. It was argued that the net effect of a change in ownership frompublic to private would be improved economic efficiency, and over time anincrease in investment.
If privatisation was sufficiently extensive and had efficiency-inducingeffects, the contribution of improved performance could be detected at themacroeconomic level. Privatisation would reduce crowding out and providemore credit to the private sector. It would increase the opportunities forinvestment in newly privatised enterprises by releasing them from thecapital constraints previously faced under public ownership. A change inownership would increase efficiency by introducing changes to thegovernance mechanisms and structure of incentives facing employees.
Attempting to measure the contribution of an ownership change oneconomic growth is complicated by the fact that economic performance islikely to be affected by factors that affect the wider economic environmentin which privatised enterprises operate. Privatisation is often accompaniedin developing countries by changes in economic policies that affecteconomic growth. Significant attention has focussed on the process ofderegulation and the importance of competition and its relation to economicefficiency. Vickers and Yarrow [1988] have argued that competition and itsregulation are crucial for the improvement of efficiency in privatisedenterprises. Unravelling the separate effects of policy changes and degreesof competition is difficult, and partly explains the relative deficiency ofempirical analysis in this area. The other major constraint to thedevelopment of empirical investigations has obviously related to the timeperiod that has elapsed since privatisation. Until recently insufficient datawas available to carry out studies capable of measuring the dynamic effectsof privatisation.
Two recent studies have attempted to measure the impact of privatisationon economic growth in developing countries. The first, by Plane [1997],was used for 35 developing countries covering the period 1984–92. The
123PRIVATISATION AND ECONOMIC GROWTH
396jds05.qxd 29/08/03 14:13 Page 123
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
study, using Probit and Tobit models, found that privatisation positivelyaffected GDP growth and that the effect on growth was more significant foractivities of a public goods type than for other sectors. The study concludesthat on average institutional reform increased economic growth from 0.8 percent to 1.5 per cent between the sub-periods 1984–88 and 1988–92.
The second study was published by the IMF in 2000 [Barnett, 2000].This study investigated the effect of privatisation on real GDP growth,unemployment, and investment. Only twelve developing countries wereused in this 18-country study. The rest consisted of transitional economies.The empirical analysis from this study strongly supported the hypothesisthat privatisation was positively correlated with real GDP growth. Theyfound that privatisation of a one per cent of GDP was associated with anincrease in the real growth rate of 0.5 per cent in period one and 0.4 per centin period two. These periods vary for each country to reflect periods ofactive privatisation, but the precise span of years for the study is notspecified. It has also been suggested that the privatisation variable used inBarnett’s is likely to capture the positive impact of a general regime changetowards better economic policies [Davis, Ossowski, Richardson andBarnett, 2000].
III . METHODOLOGY AND DATA
In most developing countries privatisation has been a new policy instrumentand has typically been implemented sporadically rather than annually. As aresult, this study applies a cross-country regression analysis to examinewhether or not privatisation affects economic growth. Cross-countryregression analysis has been widely used to scrutinise the determinants oflong-run growth. These studies have identified various policy and socio-demographic factors that may contribute to long-run growth. Studies ofthese types and their interpretation have led to debates over the validity ofthese approaches.
It has become apparent that cross-country growth regressions potentiallysuffer a number of statistical and conceptual problems. Harberger [1987]points out that one of the most fundamental statistical problems could becaused by the inclusion of countries that have little in common into the sameregression. In principle, regression analysis necessitates observations thatare drawn from a distinct population. Thus, the inclusion of the countrieswhich are intrinsically different may result in unacceptable levels ofstatistical bias [Levine and Zervos, 1993].
An important conceptual problem relates to the way in which cross-country growth regression analysis is conducted. A typical cross-countrygrowth regression analysis uses data that are averaged over a significantly
124 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:13 Page 124
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
long period, to permit the coefficients on various variables to be interpretedas elasticities. In turn, these coefficients are used to indicate the percentagegrowth that results from a one percent change in a policy variable. None theless, Levine and Zervos [1993] warn that this type of analysis ought to betreated cautiously since a cross-country regression itself does not providethe complete answer to any causal relationships between policy variablesand growth. Accordingly, they argue that ‘cross-country regressions shouldbe viewed as evaluating the strength of partial correlations, and not asbehavioural relationships that suggest how much growth will change whenpolicies change’ [1993: 426].
In practice cross-country regression analysis has been one of the mostwidely used methods in research on the relationship between policy changesand growth. In order to take advantage of its usefulness, it is imperative tominimise any possibility that distorts results, and to check for the robustnessof results. With these reservations in mind, this paper carries out a cross-country growth regression analysis using the framework of the so-calledextreme-bounds analysis (EBA). The EBA was developed by Leamer[1978, 1983, 1985] and Leamer and Herman [1983] and has been appliedpractically in a number of empirical studies [Cooley and Leroy, 1981; Derviet al., 1990; Levine and Renelt, 1992; Levine and Zervos, 1993].
The EBA can be conducted by applying the following linear ordinaryleast squares (OLS) regression:
Y = β1I + β2M + β3Z + u,
where Y is the GDP per capita growth rate, I is a set of variables that arecommonly included in the regression, M is the policy variable of particularinterest, and Z is a set of up to three variables chosen from a pool of policyvariables. This method follows Levine and Renelt [1992] and is differentfrom the original EBA developed by Leamer [1978], which includes all theZ variables in a regression. It is more practical and has been referred to as amodified version of EBA [Doppelhofer, Miller, and Sala-i-Martin, 2000].
In accordance with the EBA procedure, the combinations of a set of theZ variables are changed to examine whether or not the coefficient on M isstatistically significant and maintains the same sign throughout the process.To conduct the EBA, all possible combinations of regressions are estimatedto obtain β2 and the corresponding standard deviation, σ, for eachregression. Then, the upper or maximum extreme bound is estimated as themaximum value of β2 plus σ2. The lower or minimum extreme bound isdefined as the minimum value of β2 minus σ2. If these maximum andminimum extreme bounds keep the same sign, the result is interpreted asrobust rather than ‘fragile’ (see Appendix 1 for a more detailed account ofthe EBA).
125PRIVATISATION AND ECONOMIC GROWTH
396jds05.qxd 29/08/03 14:13 Page 125
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
Several potential problems relating to EBA have been identified [Sala-i-Martin, 1994]. Significant among these is the so-called reverse datamining problem where the control variables in a regression are drawn fromthe true population. This implies that all data contain some element of error.By seeking different combinations of control variables, it is highly likelythen that there exists a combination that results in a variable of interestbeing statistically insignificant or changing its coefficient sign. As a result,the EBA ‘may be too strong’ [Sala-i-Martin, 1994: 743].
Accordingly, Sala-i-Martin [1997] has attempted to estimate the entiredistribution of β2. He also suggests an alternative method to conduct asensitivity analysis, although the statistical rationale for the approach hasbeen shown to be uncertain [Folster and Henrekson, 2001]. More recently,Doppelhofer, Miller, and Sala-i-Martin [2000] have developed analternative method based on a Bayesian approach. Our analysis, however,continues to use the modified EBA developed by Levine and Renelt [1992]as these later contributions are still in their experimental stage.
Privatisation Variable
The most crucial question is how to measure the magnitude of privatisationin a given country. Plane [1997] has used the cumulative proceeds fromprivatisation during the period 1988–92 as a share of GDP in 1990 and adummy variable based on the information obtained from this indicator. Asimilar indicator is used in this study, although the average GDP during thesample period is used as a weight instead of choosing a particular year. Thisreduces the arbitrariness from choosing a particular year for which thesample countries might experience a variety of external or internal shocks.In addition, the use of a dummy variable for privatisation is droppedsince it cannot convey information on the magnitude of privatisation intothe regression.
The dataset for privatisation is summarised in Table 1. The availabilityof privatisation data largely determines the number of sample countries andthe sample period. As a result, 63 developing countries and the period1988–97 are selected. The dataset shows the highly concentrated pattern ofprivatisation, despite its adoption in a wide range of countries. LatinAmerican countries dominate the table in terms of privatisation as a shareof GDP. The dataset and full list of other variables corresponding to thesecountries are shown in Appendix 2.
To reduce the risk that the privatisation variable reflects other policy andstructural reforms at the macroeconomic or aggregate level, it is necessaryto check if the effect of this variable can be isolated. This can be achievedby examining the relationship between some of the policy variables thatmight be captured by the privatisation variable. The effects of the
126 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:13 Page 126
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
TA
BL
E 1
CU
MU
LA
TIV
E P
RO
CE
ED
S F
RO
M P
RIV
AT
ISA
TIO
N (
CP
P)
AN
D A
VE
RA
GE
GD
P (
AG
DP
) D
UR
ING
TH
E P
ER
IOD
19
88
–9
7(I
N U
S$
MIL
LIO
NS
)
Cou
ntry
CPP
AG
DP
CPP
/C
ount
ryC
PPA
GD
PC
PP/
AG
DP
(%)
AG
DP
(%)
1A
rgen
tina
2792
1.0
2826
7.3
9.9
33Ta
nzan
ia14
0.8
1503
1.3
0.9
2Pe
ru74
77.5
8705
1.1
8.6
34U
gand
a15
1.6
1642
9.3
0.9
3M
alay
sia
1007
6.3
1186
95.4
8.5
35Z
imba
bwe
197.
322
240.
30.
94
Sing
apor
e45
78.0
5966
9.2
7.7
36H
ondu
ras
93.5
1100
3.2
0.8
5B
eliz
e54
.379
3.0
6.8
37K
enya
235.
128
329.
60.
86
Jam
aica
532.
981
22.0
6.6
38To
go38
.852
20.3
0.7
7Pa
nam
a83
2.8
1336
3.5
6.2
39In
dia
7073
.011
7216
6.0
0.6
8T
rini
dad
& T
obag
o44
8.4
7599
.45.
940
Kor
ea, R
epub
lic25
46.0
4422
76.5
0.6
9Z
ambi
a41
7.2
7865
.95.
341
Gui
nea
45.0
9928
.10.
510
Mex
ico
3335
3.1
6479
00.5
5.1
42Jo
rdan
58.7
1195
9.8
0.5
11Pa
pua
New
Gui
nea
223.
643
97.3
5.1
43T
haila
nd13
78.4
2969
25.5
0.5
12B
oliv
ia88
4.2
1772
0.7
5.0
44T
unis
ia17
1.0
3738
0.8
0.5
13B
razi
l34
559.
487
7419
.53.
945
Om
an60
.113
370.
70.
414
Gha
na87
2.6
2309
8.0
3.8
46E
cuad
or16
9.4
4833
6.9
0.4
15V
enez
uela
5914
.116
8554
.83.
547
Mal
i21
.960
15.3
0.4
16C
olom
bia
5685
.121
5739
.92.
648
Cos
ta R
ica
56.7
1873
5.7
0.3
17B
arba
dos
51.0
2039
.52.
549
Para
guay
42.0
1576
7.2
0.3
18M
oroc
co18
46.7
7730
7.4
2.4
50C
amer
oon
41.1
2377
9.9
0.2
19So
uth
Afr
ica
6064
.425
7829
.92.
451
Gui
nea-
Bis
sau
0.5
235.
20.
220
Cot
e d’
Ivoi
re47
6.3
2087
4.1
2.3
52M
alaw
i10
.855
49.1
0.2
21Sr
i Lan
ka72
5.9
3464
0.0
2.1
53B
angl
ades
h60
.398
712.
60.
122
Nic
arag
ua13
0.2
6520
.62.
054
Bur
kina
Fas
o6.
484
11.6
0.1
23E
gypt
2777
.714
2925
.91.
955
Bur
undi
4.2
4332
.20.
124
Phili
ppin
es38
10.0
2046
78.8
1.9
56G
uate
mal
a43
.433
999.
60.
125
Sene
gal
191.
412
344.
01.
657
Nep
al15
.018
530.
00.
126
Tur
key
3843
.430
6790
.81.
358
Sier
ra L
eone
1.6
2396
.30.
127
Moz
ambi
que
110.
689
32.9
1.2
59U
rugu
ay17
.023
925.
70.
128
Paki
stan
1951
.016
1154
.21.
260
Iran
18.1
2094
17.6
0.0
29C
hile
1484
.112
9306
.11.
161
Mau
rita
nia
1.1
3357
.20.
030
Ben
in56
.556
92.5
1.0
62V
ietn
am2.
672
856.
00.
031
Indo
nesi
a51
62.8
4944
40.0
1.0
63Y
emen
0.8
8544
.50.
032
Nig
eria
763.
587
263.
80.
9
Sour
ce: P
rim
ary
data
fro
m W
orld
Ban
k Pr
ivat
izat
ion
Dat
abas
e an
d Pr
ivat
isat
ion
Inte
rnat
iona
l (v
ario
us i
ssue
s),
and
Wor
ld B
ank
Glo
bal
Dev
elop
men
tN
etw
ork
Gro
wth
Dat
abas
e.
396jds05.qxd 29/08/03 14:13 Page 127
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
privatisation variable may be linked to the size of the government budgetdeficit, since privatisation may have been implemented as a quick solutionto a budgetary problem in developing countries [Yarrow, 1999]. In order torule out this possibility, a simple correlation between the privatisationindicator and the average ratio of government budget deficit to GDP duringthe sample period 1988–97 was examined.
The result, shown in Figure 1, indicates that there is no correlationbetween the two variables (the correlation coefficient is 0.23). Argentina,Peru, and Malaysia implemented sizeable privatisation programmes whilemaintaining relatively low budget deficits. Guinea-Bissau and Mozambiquerecorded considerably high budget deficits with relatively smallprivatisation programmes. It should be emphasised that this simplecorrelation analysis, conducted without lags, does not eliminate thepossibility that some countries might have reduced their budgetary deficitsas a result of privatisation.
Similarly, the privatisation variable may capture the effects of policyreforms pursued through World Bank adjustment programmes. In oursample, 56 of the 63 countries received various adjustment loans withpolicy reform conditions during the period 1985–97. In particular,Argentina, Bangladesh, Bolivia, Ghana, Jamaica, Mauritania, and Zambia
128 THE JOURNAL OF DEVELOPMENT STUDIES
FIGURE 1
THE CORRELATION BETWEEN PRIVATISATION ANDGOVERNMENT BUDGET DEFICIT
Note: PRIV/GDP and BUD/GDP are the average percentages during the period 1988–97.Primary data from World Bank Privatisation Database and World Bank GlobalDevelopment Network Growth Database.
0123456789
10
-30 -25 -20 -15 -10 -5 0 5 10 15
BUD/GDP
PRIV
/GD
P
396jds05.qxd 29/08/03 14:13 Page 128
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
received adjustment loans with explicit conditions to implementprivatisation programmes.
The World Bank states that ‘privatisation became an important feature inthe Bank’s adjustment lending from the mid-1980s. This reflected both theconcurrent ideological shift in favour of private ownership as well as theBank’s long, and painful, experience with failed attempts at reformingpublic enterprises’ [World Bank, 1995b: 56]. Before 1985, eight per cent ofall public enterprise reform-related adjustment loans attached conditionsregarding the implementation of privatisation. Between 1985 and 1989, theproportion increased to 14 per cent and to 32 per cent in the period between1990 and 1994 [World Bank, 1994].
A simple correlation was conducted between privatisation and WorldBank adjustment loans in order to eliminate the possibility that theprivatisation variable captures the macroeconomic effects of adjustmentprogrammes. The magnitude of World Bank adjustment loans is calculatedin the same manner as the privatisation indicator (i.e. the cumulative amountof World Bank adjustment loans during the period 1985–97 divided by theaverage GDP during the same period). The result reported in Figure 2clearly indicates that there is no apparent relationship between privatisationand adjustment loans (the correlation coefficient is 0.03).
129PRIVATISATION AND ECONOMIC GROWTH
FIGURE 2
THE CORRELATION BETWEEN PRIVATISATION ANDWORLD BANK ADJUSTMENT LOANS
Note: PRIV/GDP and SAL/GDP are the average percentage during the period 1988–97. Primarydata from World Bank Privatisation Database, World Bank Global Development NetworkGrowth Database, and World Bank SAL Database.
0
2
4
6
8
10
12
0 2 4 6 8 10 12 14 16
SAL/GDP
PRIV
/GD
P
396jds05.qxd 29/08/03 14:13 Page 129
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
Of course, these results do not automatically confirm that theprivatisation indicator is able to capture the effects of privatisation at themacroeconomic level. It only assures us that the privatisation indicator isnot picking up the effects of other policy reforms. In addition, since ourstudy has significantly increased the number of sample countries comparedwith the previous studies, then the possibility that the privatisation indicatorcaptures the effects of other macroeconomic reforms is considerablyreduced.
IV. EXTREME BOUNDS ANALYSIS (EBA)
The analysis is conducted in two parts. In the first, the control variables areexamined to find appropriate specifications for the I variables. In the secondpart, the results of introducing privatisation are reported.
The Basic Test for the Control Variables (the I variables)
In principle, the I variables are always included in an EBA. These variableshave to be selected with great care. In the context of a cross-country growthregression, the control variables ought to control for initial economic,political, and social conditions of countries. There are several variables thatare routinely included in these types of regressions. They are: (1) log ofinitial GDP per capita; (2) initial life expectancy at birth; (3) averagepopulation growth rate; (4) the ratio of government consumption to GDP;(5) the ratio of gross domestic investment to GDP; and (6) the rate ofsecondary school enrolment or attainment. These variables are sometimescalled ‘Barro-regressors’ and are used widely in cross-country growthregressions [Barro, 1991].
Earlier, Levine and Zervos [1993] have argued that the inclusion of afiscal variable (4) does not necessarily control for initial conditions orpolitical stability but is a contemporaneous economic policy indicator.However, Barro and Sala-i-Martin [1995] have argued that a fiscal variablecan be used as a proxy ‘for political corruption or other aspects of badgovernment, as well as for the direct effects of non-productive publicexpenditures and taxation’ [1995: 434].
The inclusion of investment as a control variable also remainscontroversial. Levine and Renelt [1992] have included the ratio ofinvestment to GDP in their I variables, although a number of studies haveshown that causality runs from growth to investment and not vice versa[Blomstrom, Lipsey and Zejan, 1993; Barro and Sala-i-Martin, 1995]. Morerecently, Temple [1999] argues that the correlation is reverse and robust.Besides the concern over the direction of causality, there is also a questionover what measure of investment to use. In general, gross domestic
130 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:13 Page 130
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
investment is often used as a proxy for investment. However, Barro andSala-i-Martin [1995] argue that since gross domestic investment includesboth private and public spending, it may be an inappropriate measure ofinvestment. DeLong and Summers [1991] also argue that the use of grossdomestic investment can be ruled out since producer durables are thesignificant component of investment. Since we are unable to resolve thecausality problem, we test for different combinations of the I variables usingboth gross domestic investment and government consumption.
Another specification problem we have encountered relates to theinclusion of the secondary school enrolment ratio. This was not included inour study since it does not capture the quality of education. Barro and Sala-i-Martin [1995] have argued that it is more appropriate to use the secondaryschool attainment ratio, although if we use this variable, the number ofcountries will have to be severely reduced owing to a lack of data. Instead,the initial life expectancy at birth is used. This variable captures variousaspects of human capital development, such as investment in education andincreases in schooling, thereby reflecting the quality of education in general[Ram and Schulz, 1979; Sala-i-Martin, 1997; Kalemili-Ozcan, Ryder andWeil, 2000].
Combinations of variables were used to find the I variables relevant forthis study. This was achieved by examining four different combinations ofselected I variables. The first specification included the log of initial GDPper capita for 1988 (hereafter LGYP), the initial life expectancy at birth for1987 (LIFE), the average population growth during the sample period1988–97 (POP), and the average ratio of GDI to GDP during the sampleperiod (GDI). The second specification included LGYP, LIFE, POP, GDI,and the average ratio of government consumption to GDP during the sampleperiod (GOVC). The third specification included LGYP, POP, LIFE, andGOVC. Finally, the fourth specification consisted of LGYP, LIFE, and POP.In all cases the dependent variable is the average growth rate of GDP percapita during the sample period (GYP).
In order to check the robustness of the basic results, a series of EBAswas conducted by including the Z variables. The following Z variables wereselected: (1) openness, which is calculated as (export + import)/GDP(hereafter OPEN); (2) informal market premium (IMP); (3) the standarddeviation of IMP (SDIMP); (4) the average ratio of FDI to GDP (FDI); (5)the average number of major political crises, revolutions, and coups as aproxy for political instability (STAB); (6) the average inflation rate (INFL);(7) the standard deviation of average inflation rates (SDINF); (8) theaverage ratio of liquid liabilities (M3) to GDP used as a proxy for the levelof financial sector development (M3); (9) the standard deviation of M3(SDM3); (10) the average ratio of total external debt to GDP (DEBT); (11)
131PRIVATISATION AND ECONOMIC GROWTH
396jds05.qxd 29/08/03 14:13 Page 131
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
the average ratio of government budget surplus/deficit to GDP (BUD); and(12) three regional dummies for East Asia, Latin America and Sub-SaharanAfrica (EA, LA, AF, respectively). Some of the I variables are also includedas Z variables. They are GOVC (in the first and fourth specifications) andGDI (in the third and fourth specifications).
These variables, with the exception of (4), are frequently used in cross-country growth regressions. The ratio of FDI to GDP is included on theassumption that FDI may play a significant role in generating positivespillover effects, bringing new technologies and management andmarketing skills that contribute to economic growth in developingcountries. In addition, STAB is included as a Z variable, although it couldbe used as a control variable since there may exist a reverse causality thatruns from political instability to economic growth. In this respect,Londregan and Poole [1990] argue that higher political instability is likelyto be caused by lower economic growth, and Barro [1991] and Alesina andPetrotii [1993] have also found evidence supporting this reverse causality.
The EBAs for the control variables in the four specifications wereconducted by computing a total of 7,140 regressions, and the results areshown in Table 2. Note that any result obtained from a regression thatexhibits a significant level of multicollinearity (that is, a regression showingthat a variance inflation factor (VIF) is greater than 10) was discarded fromthese results.
In general, the inclusion of LGYP is expected to represent the rate ofconvergence across the sample countries. Other studies that have includeddeveloped countries have found evidence of conditional convergence [e.g.,Mankiw, Romer and Weil, 1992; Barro and Sala-i-Martin, 1995]. Anegative and statistically significant coefficient for this variable signifiesevidence of conditional convergence. In our study, the results for LGYPobtained from the first, third, and fourth specifications appear to be negativeand fragile (see the notes in Table 2 for the criterion used to interpret EBAresults).
With respect to the results for the first and second specifications, LIFEappears to be fragile, although robust results were obtained in the third andfourth specifications. Throughout, the sign of the coefficient is positive, andthe robust results imply that this variable captures aspects of human capitaldevelopment reasonably well. A robust negative result was obtained forPOP from the first and forth specifications. In contrast, the result obtainedfrom the second specification was completely insignificant and was fragilefor the third. These contrasting results may suggest that this variable isextremely sensitive to the inclusion or exclusion of other variables, asindicated by Levine and Renelt [1992]. Robust results were obtained inrelation to GDI and GOVC. The signs of the coefficients for GDI are
132 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:13 Page 132
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
133PRIVATISATION AND ECONOMIC GROWTH
TABLE 2
THE EBA RESULTS OF THE I VARIABLES (AT 0.05 LEVEL)
Specification I
Variable Coefficient t-statistic Z-variables included EBA result
LGYP Min –0.012 –2.05 STAB, EA, AF FragileMax –0.011 –2.02 STAB, FDI, EABase –0.005 –0.94 –
LIFE Min 0.052 2.01 GOVC, SDIMP, LA FragileMax 0.061 2.23 GOVC, BUD, LABase 0.025 0.98 –
POP Min –1.065 –3.25 FDI, IMP, LA RobustMax –0.650 –2.01 INFL, M3, EABase –0.782** –2.34 –
GDI Min 0.104 2.17 FDI, EA, LA RobustMax 0.219 5.20 IMF, SDM3, DEBTBase 0.203*** 5.03 –
Specification II
Variable Coefficient t-statistic Z-variables included EBA result
LGYP Min – – – InsignificantMax – – –Base –0.003 –0.56 –
LIFE Min 0.051 2.02 FDI, IMP, LA FragileMax 0.066 2.37 STAB, BUD, LABase 0.028 1.12 –
POP Min – – – InsignificantMax – – –Base –0.390 –1.07 –
GDI Min 0.108 2.25 STAB, EA, LA RobustMax 0.209 5.03 BUD, SDM3, DEBTBase 0.196*** 5.03 –
GOVC Min –0.160 –2.98 BUD, SDIMP, LA RobustMax –0.107 –2.02 STAB, IMP, EABase –0.120** –2.29 –
Specification III
Variable Coefficient t-statistic Z-variables included EBA result
LGYP Min –0.013 –2.34 STAB, EA, AF FragileMax –0.011 –2.04 STAB, SDIMP, EABase –0.005 –0.82 –
LIFE Min 0.052 2.01 GDI, SDIMP, LA Robust Max 0.109 3.95 STAB, BUD, LABase 0.068** 2.43 –
POP Min –0.760 –2.04 STAB, FDI, LA FragileMax –0.742 –2.01 FDI, SDIMP, LABase –0.354 –0.82 –
396jds05.qxd 29/08/03 14:13 Page 133
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
positive, whereas they are negative for GOVC, confirming previous studies[e.g., Barro, 1991].
In summary, LGYP exhibits fragile results, and LIFE and POP appear tobe sensitive to different combinations of other variables, resulting inambiguous results. On the other hand, the results for GDI and GOVC arealways robust. Nevertheless, it is difficult to choose a particularspecification for the control variables from these results. Since a continuedsearch for the most suitable combination of the control variables is beyondthe scope of this study and may prove impossible, as some past studies havedemonstrated, then all four specifications are used to conduct an EBA forthe privatisation variable.
The EBA for Privatisation
The EBA proceeded in two steps. First, basic tests were carried out toestimate baselines before a formal EBA was conducted for privatisation.These regressions included the control variables in the four specifications
134 THE JOURNAL OF DEVELOPMENT STUDIES
GOVC Min –0.203 –3.17 STAB, BUD, M3 RobustMax –0.109 –2.04 GDI, SDINF, SDM3Base –0.139** –2.24 –
Specification IV
Variable Coefficient t-statistic Z-variables included EBA result
LGYP Min –0.015 –2.69 STAB, EA, AF FragileMax –0.011 –2.01 MDM3, EA, AFBase –0.008 –1.20 –
LIFE Min 0.054 2.05 GOVC, FDI, M3 Robust Max 0.103 3.77 GOVC, BUD, LABase 0.067** 2.30 –
POP Min –1.216 –3.46 FDI, SDM3, LA RobustMax –0.660 –2.01 GDI, M3, EABase –0.810** –2.04 –
Notes: Dependent variable: GYP. The base results are calculated excluding the M and Zvariables. Note that following Levine and Renelt [1992] (see Table 1, notes (a), p.947),where an EBA indicates a robust result, but the base result is statistically insignificant, theoverall result is interpreted as fragile. ** = statistically significant at 0.05 level, and ***= statistically significant at 0.01 level.
TABLE 2 (cont.)
Specification III (cont.)
Variable Coefficient t-statistic Z-variables included EBA result
396jds05.qxd 29/08/03 14:13 Page 134
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
examined previously and the privatisation variable. The Z variables wereexcluded. The tests showed that the base results for privatisation werestatistically insignificant. Second, the EBA for the privatisation variablewas conducted by this time including three regressors from a pool of the 15Z variables. The results are shown in Table 3. Although the EBA indicatedthat the results were robust, they still have to be interpreted as fragile sincethe earlier basic tests indicated that they were statistically insignificant.Accordingly, at this stage, we have concluded that only a weak or fragilenegative partial correlation between privatisation and economic growthexists.
135PRIVATISATION AND ECONOMIC GROWTH
TABLE 3
THE EBA RESULTS (AT 0.05 LEVEL) FOR THE PRIVATISATION VARIABLE
The I variables based on Specification I (LGYP + LIFE + POP + GDI)
Variable Coefficient t-statistic Z-variables included EBA result
PRIV Min –0.243 –2.25 BUD, FDI, DEBT Fragile Max –0.219 –2.01 GOVC, FDI, IMPBase –0.056 –0.55 – –
The I variables based on Specification II (LGYP + LIFE + POP + GDI + GOVC)
Variable Coefficient t-statistic Z-variables included EBA result
PRIV Min –0.249 –2.35 BUD, FDI, DEBT Fragile Max –0.216 –2.04 BUD, FDI, EABase –0.077 –0.77 – –
The I variables based on Specification III (LGYP + LIFE + POP + GOVC)
Variable Coefficient t-statistic Z-variables included EBA result
PRIV Min –0.347 –2.82 FDI, SDM3, DEBT FragileMax –0.219 –2.01 GDI, FDI, IMPBase –0.090 –0.76 – –
The I variables based on Specification IV (LGYP + LIFE + POP)
Variable Coefficient t-statistic Z-variables included EBA result
PRIV Min –0.345 –2.76 FDI, SDM3, DEBT FragileMax –0.228 –2.02 FDI, SDIMP, LABase –0.067 –0.55 – –
Notes: Dependent variable: GYP. Note that following Levine and Renelt [1992] (see Table 1,notes (a), p.947), in the case that the EBA indicates a robust result, but the base result isstatistically insignificant, the overall result is interpreted as fragile.
396jds05.qxd 29/08/03 14:13 Page 135
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
Two cautionary notes are required at this point. First, it is known that theinclusion or exclusion of certain control variables sometimes eliminates abivariate relationship [Easterly and Rebelo, 1993]. It is thus plausible toargue that our inconclusive results are due to an inadequate choice orcombinations of the control variables. In order to examine this possibility,all possible combinations of the control variables and the privatisationvariable were scrutinised by applying the EBA technique. This exercisereconfirmed our conclusion by showing that no combination of controlvariables exists that would establish a robust partial correlation betweenprivatisation and economic growth.
Second, it is known that cross-country regression analysis is sensitive tothe inclusion or exclusion of a particular country. This may arise due to theexistence of influential observations and/or outliers in the regressions. Thispossibility is also examined by checking whether or not there are anyinfluential observations or outliers in our basic regression models. It should,however, be pointed out that there is no universally accepted indicator thatis able to identify influential observations and outliers. For instance, Donaldand Maddala [1993] argue that the examination of studentised residuals isthe most appropriate method to identify influential observations, and todetect outliers leverage should also be examined. On the other hand,Beckman and Cook [1983] and Fiebig [1992] argue that large studentisedresiduals may not necessarily point to influential observations, and Fiebig[1992] specifically points out that influential observations may actually berelated to small residuals.
In addition, even if influential observations and outliers are identified, itmay prove difficult to deal with them. Influential observations and outliersmay arise as a result of data errors and model misspecifications. Suitablecorrections may be hampered because of poor data, especially fordeveloping countries, and simply deleting the countries in question may stillbe appropriate. On the other hand, influential observations may representnatural variation which provides valuable information, and eliminatingthese observations would clearly be inappropriate. Even if influentialobservations are kept in the regression, however, the potential effect on theresults of eliminating them ought to be noted [Fiebig, 1992]. One way toaccommodate influential observations in the regression is to apply robustregression estimators instead of OLS (see Rousseeuw and Leroy [1987] fora detailed account of robust regression estimators). It should also be notedthat even these detection techniques and robust estimators may fail to detectoutliers [Luke and Schaffer, 2000].
Accepting these limitations, studentised residuals were examined in anattempt to detect influential observations in our basic regression models. Wealso applied some of the influential observation and outlier detection
136 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:13 Page 136
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
techniques, DFITS [Welsch and Kuh, 1977], Cook’s Distance [Cook, 1977],and Welsch Distance [Welsch, 1982]. These techniques, in essence,encapsulate the information in the leverage versus residual-squared plot intoa single statistic, creating an index that is influenced by the size of theresiduals or outliers and the size of leverage (see Appendix 3 for thecalculation procedures). In addition, we tested for influential observationsand outliers specifically to examine the privatisation variable by applyingDFBETAs [Belsley, Kuh and Welsch, 1980]. The application of DFBETAspermits us to examine the difference between the coefficients of a specificvariable when a particular observation is included or excluded
Table 4 reports the results of the tests for influential observations.Although it is difficult to compare the indices across the differentspecifications and techniques, these tests identified 15 countries as potentialoutliers or a total of 95 observations as influential. Singapore and Malaysiacontributed significantly to the influential observations, 17.9 per cent and14.7 per cent respectively, followed by Oman 13.7 per cent, and to a lesserextent Sierra Leone 8.4 per cent. These four countries generated more than50 percent of the influential observations. In particular, Welsch Distanceindicates that only Malaysia, Oman, and Singapore are potential outliers.More specifically, of those 20 test results shown in Table 4, Singaporeappears as a potential outlier in 17 of them, Malaysia appears in 14 tests,and Sierra Leone is identified in eight tests. Oman was detected asinfluential by 13 test results only in the first, second, and thirdspecifications. In addition, the combination of Malaysia and Singaporeappear in 14 test results, including all the test results in the third and fourthspecifications. Furthermore, the results of DFBETAs suggest that these twocountries have substantial positive effects on the privatisation variable, inparticular in the third and fourth specifications.
After identifying the influential observations and thus the potentialoutliers in our basic regression models, we attempted to re-estimate thebaselines for privatisation in two steps. First, we re-estimated the baselinesfor privatisation by applying two robust regression estimators, the medianleast squares (MLS) and a robust regression estimator using iterativelyreweighted least squares, including all of the influential observations (seeRousseeuw and Leroy [1986] and Luke and Schaffer [2000] for details ofthese estimators). We then re-estimated the baselines using OLS byeliminating all the influential observations. Second, we deleted theinfluential observations one by one according to the higher absolute valueof each of the indicators for influential observations. Following these steps,we could not find any statistically significant results for privatisation in thefirst and second specifications but found a number of statisticallysignificant results for privatisation in the remaining specifications.
137PRIVATISATION AND ECONOMIC GROWTH
396jds05.qxd 29/08/03 14:13 Page 137
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
138 THE JOURNAL OF DEVELOPMENT STUDIES
TABLE 4
TESTS FOR INFLUENTIAL OBSERVATIONS AND OUTLIERS
Specification I
Studentized DFBETAs DFITS Cook’s Welsch
1 Uganda 2.563 Jamaica –0.543 Oman 1.065 Jordan 0.305 Oman 10.092 Jamaica –2.559 Malaysia 0.437 Singapore 0.862 Sierra Leone 0.231 Singapore 7.6483 Cameroon –2.514 Singapore 0.364 Vietnam 0.789 Jamaica 0.1944 Jordan –2.341 Peru –0.356 Uganda 0.709 Oman 0.1845 Sierra Leone –2.041 Oman –0.290 Malaysia 0.672 Singapore 0.1206 Jordan 0.255 Vietnam 0.0997 Uganda 0.0778 Malaysia 0.075
Specification II
Studentized DFBETAs DFITS Cook’s Welsch
1 Cameroon –2.959 Jamaica –0.510 Oman 2.327 Oman 0.687 Oman 23.692 Oman 2.837 Oman –0.459 Singapore 0.742 Sierra Leone 0.2473 Uganda 2.321 Peru –0.456 Vietnam 0.733 Jordan 0.2354 Sierra Leone –2.287 Malaysia 0.409 Uganda 0.727 Jamaica 0.1565 Jamaica –2.196 Singapore 0.279 Singapore 0.0776 Jordan –2.148 Cameroon 0.254 Vietnam 0.0747 Uganda 0.0708
Specification III
Studentized DFBETAs DFITS Cook’s Welsch
1 Cameroon –2.810 Malaysia 0.756 Oman 1.327 Sierra Leone 0.323 Oman 13.382 Sierra Leone –2.549 Singapore 0.443 Singapore 1.048 Oman 0.285 Singapore 9.0743 Malaysia 2.324 Peru –0.434 Malaysia 0.971 Singapore 0.170 Malaysia 8.2894 Singapore 2.296 Korea –0.366 Korea 0.641 Malaysia 0.1465 Thailand –0.302 Barbados 0.0666 Argentina –0.294 Korea 0.0657 Oman –0.260
Specification IV
Studentized DFBETAs DFITS Cook’s Welsch
1 Singapore 2.616 Malaysia 0.794 Singapore 1.081 Sierra Leone 0.326 Singapore 9.2132 Cameroon –2.411 Singapore 0.537 Malaysia 0.987 Singapore 0.212 Malaysia 8.4023 Malaysia 2.398 Korea –0.366 Korea 0.646 Malaysia 0.1804 Sierra Leone –2.325 Peru –0.340 Barbados 0.1005 Thailand –0.303 Korea 0.0806 Jamaica –0.295 Jamaica 0.0687 Jordan 0.065
396jds05.qxd 29/08/03 14:13 Page 138
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
Accordingly, the first and second specifications were dropped from anyfurther examination.
Correspondingly, we re-estimated the baselines for privatisation byapplying the two robust regression estimation techniques which enabled usto incorporate the influence of the influential observations or potentialoutliers in the regressions without deleting them. The results are reported inTable 5 in the third and fourth columns of each specification. In the thirdspecification, we found statistically significant results for privatisation fromMLS (at the 0.10 level) and Robust Regressions (at the 0.01 level). In thecase of the fourth specification, we could not find statistically significantresults for privatisation from the MLS but found the statistically significantresults from Robust Regression (at 0.01 level). As for the MLS results fromspecification four, pseudo R2 is too low, and the signs of the coefficients,except those for PRIV and POP, changed. In essence, we have found that arobust partial correlation between privatisation and economic growth can beestablished in these specifications, in particular in the third specification, byusing the robust estimators instead of OLS and without deleting theinfluential observations.
The results of excluding all the influential observations using OLS (atotal of 10 models were estimated) are shown from the fifth to the lastcolumn in Table 5. In these specifications, we found statistically significantresults for privatisation at the 0.01 level (shown in column (1) in the thirdspecification and column (6) in the fourth specification) by deleting all theinfluential observations identified by studentised residuals. Similarly, wefound statistically significant results for privatisation in the fourthspecification (column 10) at the 0.05 level and in the third specification(column 5) at the 0.10 level by eliminating all the influential observationsidentified by Welsch Distance. As for Cook’s Distance, we also foundstatistically significant results for privatisation at the 0.05 level (the thirdspecification (column 4)) and at the 0.10 level (the fourth specification(column 9)).
In contrast, the elimination of all the influential observations identifiedby DFBETAs resulted in statistically insignificant results for privatisation inboth specifications. In terms of the deleted influential observationsidentified by DFITS, the results are somewhat mixed. We found astatistically significant result for privatisation at the 0.10 level in the fourthspecification (column 8) and statistically insignificant result in the thirdspecification (column 3). In sum, we tested ten models using OLS in thethird and fourth specifications and found statistically significant results inseven models.
The results of eliminating the influential observations to examinewhether or not these have led to increases in R2 and/or the statistical
139PRIVATISATION AND ECONOMIC GROWTH
396jds05.qxd 29/08/03 14:13 Page 139
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
TA
BL
E 5
TH
E B
AS
EL
INE
RE
SU
LT
S F
OR
PR
IVA
TIS
AT
ION
Sp
ecif
icat
ion
III
OL
SM
LS
R
obus
t R
egre
ssio
n(1
) S
tude
ntis
ed(2
) D
FB
ET
As
(3)
DF
ITS
(4)
Coo
k’s
(5)
Wel
sch
PR
IV–0
.090
–0
.192
*–0
.254
***
–0.2
92**
*–0
.063
–0.1
78–0
.214
**–0
.223
*(–
0.76
)(–
1.89
)(–
2.94
)(–
2.91
)(–
0.46
)(–
1.58
)(–
2.01
)(–
1.96
)L
GY
P–0
.004
–0.0
10*
–0.0
06–0
.002
–0.0
16**
–0.0
16**
–0.0
09–0
.013
**(–
0.64
)(–
1.72
)(–
1.34
)(–
0.37
)(–
2.66
)(–
2.56
)(–
1.50
)(–
2.18
)L
IFE
0.06
9**
0.06
8***
0.03
20.
036
0.08
5***
0.08
9***
0.05
1*0.
085*
**(2
.46)
(2.7
1)(1
.55)
(1.4
4)(3
.57)
(3.6
3)(1
.95)
(3.3
7)P
OP
–0.3
72–0
.648
–0.8
86**
*–0
.727
**–0
.749
*–0
.897
**–1
.156
***
–1.0
12**
(–0.
85)
(–1.
64)
(–2.
79)
(–2.
06)
(–1.
87)
(–2.
22)
(–2.
85)
(–2.
45)
GO
VC
–0
.143
**–0
.133
**–0
.140
***
–0.1
30**
–0.1
56**
–0.1
45**
–0.1
29**
–0.1
49**
(–2.
28)
(–2.
44)
(–3.
08)
(–2.
60)
(–2.
50)
(–2.
48)
(–2.
25)
(–2.
48)
Con
stan
t–0
.207
**–0
.156
**–0
.025
–0.0
77–0
.179
**–0
.191
**–0
.079
–0.1
86**
(–2.
40)
(–1.
99)
(–0.
40)
(–1.
01)
(–2.
48)
(–2.
57)
(–1.
00)
(–2.
43)
Adj
uste
d R
20.
22–
–0.
310.
300.
320.
310.
34P
seud
o R
2–
0.13
––
––
––
No.
of
infl
uent
ial
––
–A
ll (
4)A
ll (
7)A
ll (
4)A
ll (
6)A
ll (
3)ob
s. e
xclu
ded
Sp
ecif
icat
ion
IV
OL
SM
LS
R
obus
t R
egre
ssio
n(6
) S
tude
ntis
ed(7
) D
FB
ET
As
(8)
DF
ITS
(9)
Coo
k’s
(10)
Wel
sch
PR
IV–0
.067
–0.1
06–0
.250
***
–0.2
85**
*–0
.094
–0.2
04*
–0.2
21*
–0.2
49**
(–0.
55)
(–1.
24)
(–2.
61)
(–2.
71)
(–0.
74)
(–1.
73)
(–1.
95)
(–2.
10)
LG
YP
–0.0
070.
002
–0.0
05–0
.006
–0.0
16**
*–0
.014
**–0
.008
–0.0
12*
(–1.
06)
(0.3
2)(–
1.05
)(–
0.99
)(–
2.67
)(–
2.31
)(–
1.27
)(–
1.95
)L
IFE
0.06
7**
–0.0
110.
023
0.03
80.
085*
**0.
082*
**0.
051*
0.07
8***
(2.3
1)(–
0.54
)(0
.99)
(1.4
4)(3
.37)
(3.1
9)(1
.83)
(2.9
7)P
OP
–0.8
34**
–1.3
40**
*–1
.185
***
–1.1
71**
*–0
.934
**–0
.975
***
–1.2
83**
*–1
.112
***
(–2.
08)
(–4.
69)
(–3.
79)
(–3.
60)
(–2.
60)
(–2.
69)
(–3.
15)
(–3.
04)
Con
stan
t–0
.188
**0.
083
–0.0
05–0
.066
–0.1
93**
–0.1
93**
–0.1
00–0
.187
**(–
2.11
)(1
.31)
(–0.
08)
(–0.
83)
(–2.
57)
(–2.
49)
(–1.
18)
(–2.
35)
Adj
uste
d R
20.
17–
–0.
240.
230.
230.
220.
25P
seud
o R
2–
0.09
––
––
––
No.
of
infl
uent
ial
––
–A
ll (
4)A
ll (
6)A
ll (
3)A
ll (
7)A
ll (
2)ob
s. e
xclu
ded
Note
s:D
epen
dent
var
iabl
e: G
YP.
The
mod
els
(1)
- (1
0) w
ere
esti
mat
ed u
sing
OL
S.
* =
sta
tist
ical
ly s
igni
fica
nt a
t 0.
10 l
evel
, *
* =
sta
tist
ical
ly s
igni
fica
nt a
t0.
05 l
evel
, an
d **
* =
sta
tist
ical
ly s
igni
fica
nt a
t 0.
01 l
evel
. t
-sta
tist
ics
are
in p
aren
thes
es.
396jds05.qxd 29/08/03 14:13 Page 140
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
significance of the privatisation variable are shown in Table 6. Statisticallysignificant results were obtained for privatisation in all six models estimatedin the third and fourth specifications. In these models, higher adjusted R2and/or increased statistical significance for privatisation were obtained bydeleting a few countries from a pool consisting of Barbados, Cameroon,Malaysia, Oman, Sierra Leone and Singapore. Malaysia and Singaporeappear to be the most influential observations. In two models (column 11)and (column 16), higher R2 and increased statistical significance forprivatisation were obtained by eliminating these two countries, and in theremaining 4 models (shown in columns 12, 13, 14 and 15), higher R2 andincreased statistical significance were obtained by eliminating a few of thecountries including Malaysia and Singapore. Accordingly, Malaysia andSingapore appear to bias our regression results (see Stokey [1994] andLorgelly and Owen [1999] for a discussion of biased regression resultscaused by the inclusion of certain East Asian countries).
In summary, we obtained a total of 14 new estimates for privatisation,using OLS and deleting the influential observations identified by fivedifferent outlier detection techniques (note that some of the results areidentical: (column 5) and (column 12) in the third specification and (column10) and (column 16) in the fourth specification). In order to conduct theEBAs for these new estimates, we have selected from Tables 5 and 6baseline results for privatisation, which are statistically significant at 0.05level, for each specification according to the highest adjusted R2. This wasdone because the elimination of a subset of some influential observationsfrom our dataset resulted in the largest reduction in the residual sum ofsquares. It is likely that the eliminated observations are outliers [Gentlemanand Wilk, 1975].
All the baseline results were adopted from Table 6. We chose thebaseline results obtained by Cook’s Distance (the third and fourthspecifications) and the studentised residuals (the fourth specification). Themost likely candidates for outliers turned out to be Barbados, Cameroon,Oman, Malaysia, Sierra Leone, and Singapore. On this basis, Table 7 showsEBAs for privatisation by deleting the influential observations or the mostlikely outliers. This time, robust results were obtained from all twospecifications. We also conducted the EBAs for the other statisticallysignificant base results for privatisation at the 0.05 level by deleting theinfluential observations in Tables 5 and 6. The EBA results for privatisationare negative and robust in all the models.
Consequently, a negative but fragile partial correlation betweenprivatisation and economic growth has been found using the full 63-countrysample with OLS estimators, and a robust negative partial correlation hasbeen found by eliminating the possible outliers. We have also found a robust
141PRIVATISATION AND ECONOMIC GROWTH
396jds05.qxd 29/08/03 14:13 Page 141
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
142 THE JOURNAL OF DEVELOPMENT STUDIES
TABLE 6
THE BASELINE RESULTS FOR PRIVATISATION EXCLUDINGSELECTED INFLUENTIAL OBSERVATIONS
Specification III
Full sample (11) DFBETAs (12) DFITS (13) Cook’s
PRIV –0.090 –0.252** –0.223* –0.256**(–0.76) (–2.16) (–1.96) (–2.40)
LGYP –0.004 –0.009 –0.013** –0.007(–0.64) (–1.52) (–2.18) (–1.14)
LIFE 0.069** 0.078*** 0.085*** 0.047*(2.46) (3.03) (3.37) (1.72)
POP –0.372 –0.761* –1.012** –1.290***(–0.85) (–1.86) (–2.45) (–3.16)
GOVC –0.143** –0.103* –0.149** –0.128**(–2.28) (–1.77) (–2.48) (–2.18)
Constant –0.207** –0.201** –0.186** –0.069(–2.40) (–2.56) (–2.43) (–0.85)
Adjusted R2 0.22 0.28 0.34 0.34Likely outliers – Malaysia, Oman, Sierra Leone,
excluded Singapore Singapore, Oman,Malaysia Singapore,
Malaysia,Barbados
Specification IV
Full sample (14) DFBETAs (15) DFITS (16) Cook’s
PRIV –0.067 –0.271** –0.288** –0.249**(–0.55) (–2.41) (–2.61) (–2.10)
LGYP –0.007 –0.011* –0.005 –0.012*(–1.06) (–1.95) (–0.77) (–1.95)
LIFE 0.067** 0.075*** 0.038* 0.078***(2.31) (3.01) (1.40) (2.97)
POP –0.834** –1.063*** –1.414*** –1.112***(–2.08) (–3.07) (–4.03) (–3.04)
Constant –0.188** –0.179** –0.069 –0.187**(–2.11) (–2.38) (–0.84) (–2.35)
Adjusted R2 0.17 0.27 0.27 0.25Likely outliers – Singapore, Sierra Leone, Singapore,
excluded Cameroon, Singapore, MalaysiaMalaysia Malaysia,
Barbados
Notes: Dependent variable: GYP. These models were estimated using OLS. * = statisticallysignificant at 0.10 level , ** = statistically significant at 0.05 level, and *** = statisticallysignificant at 0.01 level. . t-statistics are in parentheses. The likely outliers excluded areshown in the order of higher absolute values.
396jds05.qxd 29/08/03 14:13 Page 142
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
negative partial correlation between these two variables using the full 63-country sample with the robust regression estimators. These results indicatethat the privatisation variable is sensitive to combinations of controlvariables and the inclusion of certain countries. In other words, in order todetect a significant result for the privatisation variable, great care is requiredin the selection of a specific conditioning information set and countries.
V. CONCLUSION
The analysis has shown that there is a robust partial correlation betweenprivatisation and economic growth, suggesting that privatisation hascontributed negatively to economic growth. This conclusion is contrary tothe results obtained by Plane [1997] and Barnett [2000]. The analysis in thisstudy has used a different approach to that adopted in Plane’s study. Byusing an extreme bound analysis and a larger sample of countries we havebeen able to undertake a more comprehensive and rigorous analysis of the
143PRIVATISATION AND ECONOMIC GROWTH
TABLE 7
THE EBA RESULTS (AT 0.05 LEVEL) FOR THE PRIVATISATION VARIABLEEXCLUDING OUTLIERS
Specification III (LGYP + LIFE + POP + GOVC)
Coeffi- t-sta- Z-variables included Likely outliers EBA resultcient tistics excluded
Cook’s (13) Min –0.357 –3.23 FDI, SDM3, DEBT Sierra Leone, Max –0.209 –2.11 GDI, FDI, LA Oman, RobustBase –0.256** –2.40 – Singapore,
Barbados
Specification IV (LGYP + LIFE + POP)
Coeffi- t-sta- Z-variables included Likely outliers EBA resultcient tistics excluded
Studentised Min –0.373 –3.21 OPEN, FDI, DEBT Singapore, (14) Max –0.203 –2.01 GOVC, GDI, BUD Cameroon, Robust
Base –0.271** –2.41 – Malaysia
Cook’s (15) Min –0.383 –3.34 OPEN, FDI, DEBT Sierra Leone,Max –0.214 –2.02 FDI, EA, LA Sinagpore, RobustBase –0.288** –2.61 – Barbados
Notes: Dependent variable: GYP. The EBA results for PRIV are obtained by including the controlvariables and three regressors from the pool of the Z variables. ** = statistically significantat 0.05 level. The likely outliers excluded are in the order of higher absolute values.
396jds05.qxd 29/08/03 14:13 Page 143
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
relationship between privatisation and economic growth. The analysis hasalso been strengthened by the application of outlier detection techniques.Despite these improvements in methodology, the results obtained from thistype of analysis must continue to be treated with a fair degree of caution.
Are there policy implementations arising from this analysis? Clearly, thefindings do not refute the notion that the wider economic and socio-politicalenvironment may have important effects on economic growth, and on thesuccess of privatisation. The IMF study, using a limited number ofcountries, has suggested that the positive relationship between privatisationand economic growth could be explained by privatisation acting as a proxyin their analysis for a range of structural measures signifying changes in theeconomic environment. We have attempted, with our measure ofprivatisation, to rule out the possibility that this variable captures the effectsof wider policy reforms.
If, as is argued in the theoretical literature, the change in ownership aloneat the microeconomic level is not sufficient to guarantee greater enterpriseefficiency, then other reforms, more directly related to enterprisedevelopment, may indeed play a crucial role. If the success of privatisationis linked to competition and the regulation of competition, then weaknessesin these fields may explain why privatisation is negatively related toeconomic growth in developing countries. Recent reviews of competitionpolicy in developing countries indicate fundamental weaknesses inimplementation [Cook, 2001]. Similarly, regimes for regulating competitionin developing countries have not developed uniformly and with the samedegree of effectiveness across developing countries. The share of utilities inprivatisation has increased significantly in the 1990s in developing countries,resulting in the creation of numerous private sector monopolies and the needfor better regulation. Even in countries in Latin America where regulatorysystems have developed they may not be working effectively [Cook, 1999].In addition, the IMF report evidence of weak regulatory systems in thetelecommunications sector in some developing countries, which may havecontributed to enhancing the proceeds from privatisation [Davis, Ossowski,Richardson and Barnett, 2000]. The combination of weak regulation andhigh proceeds from privatisation, if accompanied by low gains in economicefficiency, may also provide an explanation for the disappointing results asfar as economic growth is concerned.
The policy implications arising from this analysis squarely point to theneed for further research into the policy environment, in particular the rolesplayed by competition and its regulation if the relation betweenprivatisation and economic growth is to be better understood.
final revision accepted July 2002
144 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:13 Page 144
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
REFERENCES
Alchian, A., 1965, ‘Some Economics of Property Rights’, II Politico, Vol.30, pp.816–29.Alesina, A. and R. Petrotii, 1993, ‘Income Distribution, Political Instability and Investment’,
NBER Working Paper, No.4486.Barnett, S., 2000, ‘Evidence on the Fiscal and Macroeconomic Impact of Privatisation’, IMF
Working Paper, July, Washington, DC: IMF.Barro, R., 1991, ‘Economic Growth in a Cross-Section of Countries’, Quarterly Journal of
Economics, Vol.106, No.2, pp.407–43.Barro, R. and X. Sala-i-Martin, 1995, Economic Growth, New York: McGraw-Hill.Beckman, R. and R. Cook, 1983, ‘Outliers’, Technometrics, Vol.25, pp.119–49.Belsey, D., Kuh, E. and R. Welsch, 1980, Regression Diagnostics, New York: John Wiley.Bennell, P., 1997, ‘Privatisation in Sub-Saharan Africa: Progress and Prospects during the
1990s’, World Development, Vol.25, No.11, pp.1785–803.Blomstrom, M., Lipsey, R. and M. Zejan, 1993, ‘Is Fixed Investment the Key to Economic
Growth?’ NBER Working Paper, No.4436.Boubakri, N. and J. Cosset, 1998, ‘The Financial and Operating Performance of Newly Privatised
Firms: Evidence from Developing Countries’, unpublished draft, Universite Laval.Boycko, M., Shleifer, A. and R. Vichny, 1996, ‘A Theory of Privatisation’, Economic Journal,
Vol.106, pp.309–19.Cook, P., 1999, ‘Privatisation and Utility Regulation in Developing Countries: The Lessons So
Far’, Annals of Public and Cooperative Economics, Vol.70, No.4, pp.549–87.Cook, P., 2001, ‘Competition and its Regulation: Key Issues’, paper presented at an international
workshop on ‘Competition and Regulation’, University of Manchester, 9 Feb. 2001Cook, P. and C. Kirkpatrick, 1988, Privatisation in Less Developed Countries, Brighton:
Weatsheaf.Cook, P. and R. Fabella, 2002, ‘The Welfare and Political Economy Dimensions of Private vs.
State Enterprise’, The Manchester School, Vol.70, No.2, pp.246–61.Cook, R., 1977, ‘Detection of Influential Observations in Linear Regression’, Technometrics,
Vol.19, pp.15–18.Cooley, T. and S. Leroy, 1981, ‘Identification and Estimation of Money Demand’, American
Economic Review, Vol.71, No.5, pp.825–44.Davis, J., Ossowski, R., Richardson, T. and S. Barnett, 2000, ‘Fiscal and Macroeconomic Impact
of Privatisation’, Occasional Paper, No.194, Washington, DC: IMF.DeLong, J. and L. Summers, 1991, ‘Equipment Investment and Economic Growth’, Quarterly
Journal of Economics, Vol.106, No.2, pp.445–502.Deravi, K., Hegji, C. and D. Moberly, 1990, ‘Government Debt and the Demand for Money: An
Extreme Bound Analysis’, Economic Inquiry, Vol.28, No.2, pp.390–401.Donald, S. and G. Maddala, 1993, ‘Identifying Outliers and Influential Observations in
Econometric Models’, in G. Maddala, C. Rao and H. Vinod (eds.), Handbook of Statistics,Vol.2, Amsterdam: Elsevier.
Doppelhofer, G., Miller, R. and X. Sala-i-Martin, 2000, ‘Determinants of Long-Term Growth: ABayesian Averaging of Classical Estimates (BACE) Approach’, OECD Working Paper, No.266.
Easterly, W. and S. Rebelo, 1993, ‘Fiscal Policy and Economic Growth: An EmpiricalInvestigation’, Journal of Monetary Economics, Vol.32, No.2, pp.417–58.
Fiebig, D., 1992, ‘Diagnostic Checking in Practice: The Case of Outliers and Influential Data’,mimeo, University of Sydney.
Folster, S. and M. Henrekson, 2001, ‘Growth Effects of Government Expenditure and Taxationin Rich Countries’, European Economic Review, Vol.45, No.8, pp.1501–20.
Galal, A., Jones, L., Tandon, P. and I. Vogelsang, 1994, Welfare Consequences of Selling PublicEnterprises: An Empirical Analysis, New York: Oxford University Press.
Gentleman, J. and M. Wilk, 1975, ‘Detecting Outliers II: Supplementing the Direct Analysis ofResiduals’, Biometrics, Vol.31, No.2, pp.387–410.
Harberger, A., 1987, ‘Comment’, NBER Macroeconomics Annual, pp.255–8.Jensen, M. and W. Meckling, 1976, ‘Theory of the Firm: Managerial Behavior, Agency Costs,
and Ownership Structure’, Journal of Financial Economics, Vol.3, No.4, pp.305–60.
145PRIVATISATION AND ECONOMIC GROWTH
396jds05.qxd 29/08/03 14:13 Page 145
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
Kalemili-Ozcan, S., Ryder, H. and D. Weil, 2000, ‘Mortality Decline, Human Capital Investment,and Economic Growth’, Journal of Development Economics, Vol.62, No.1, pp.1–23.
Kikeri, S., 1998, ‘Privatisation and Labor: What Happens to Countries When GovernmentsDivest’, World Bank Technical Paper, No.396, Washington, DC: World Bank.
Kikeri, S., Nellis, J. and M. Shirley, 1994, Privatisation: The Lessons of Experience, Washington,DC: World Bank.
Laffont, J. and J. Tirole, 1991, ‘Privatisation and Incentives’, Journal of Law, Economics andOrganisation, Vol.7, No.1, pp.84–105.
Leamer, E., 1978, Specification Searches: Ad Hoc Inference from Non-Experimental Data, NewYork: Wiley.
Leamer, E., 1983, ‘Let’s Take the Con Out of Econometrics’, American Economic Review,Vol.73, No.1, pp.31–43.
Leamer, E., 1985, ‘Sensitivity Analysis Would Help’, American Economic Review, Vol.75, No.3,pp.308–13.
Leamer, E. and L. Herman, 1983, ‘Reporting the Fragility of Regression Estimates’, Review ofEconomics and Statistics, Vol.65, No.2, pp.306–17.
Levine, R. and D. Renelt, 1992, ‘A Sensitivity Analysis of Cross-Country Growth Regressions’,American Economic Review, Vol.82, No.4, pp.942–63.
Levine, R. and S. Zervos, 1993, ‘What We Have Learned About Policy and Growth from Cross-Country Regressions?’ American Economic Review, Vol.83, No.2, pp.426–30.
Londregan, J. and K. Poole, 1990, ‘Poverty, the Coup Trap, and the Seizure of Executive Power’,World Politics, Vol.423, No.2, pp.151–83.
Lorgelly, P. and P. Owen, 1999, ‘The Effect of Female and Male Schooling on Economic Growthin the Barro-Lee Model’, Empirical Economics, Vol.24, No.3, pp.537–57.
Luke, P. and M. Schaffer, 2000, ‘Wage Discrimination in Russia: An Econometric Investigation’,William Davidson Institute Working Paper Series, No.295, University of Michigan BusinessSchool.
Mankiw, N., Romer, D. and D. Weil, 1992, ‘A Contribution to the Empirics of EconomicGrowth’, Quarterly Journal of Economics, Vol.107, No.2, pp.407–37.
Martin, S. and D. Parker, 1997, The Impact of Privatisation: Ownership and CorporatePerformance in the UK, London: Routledge.
Megginson, W., Nash, R., and M. Van Randenborgh, 1994, ‘The Financial and OperatingPerformance of Newly Privatised Firms: An International Empirical Analysis’, Journal ofFinance, Vol.49, No.2, pp.403–52.
Megginson, W and J. Netter, 2001, ‘From State to Market: A Survey of Empirical Studies onPrivatisation’, Journal of Economic Literature, Vol.39, No.2, pp.321–89.
Parker, D., 1999, ‘Privatisation in the European Union: A Critical Assessment of itsDevelopment, Rationale and Consequences’, Economic and Industrial Democracy, Vol.20,pp.9–38.
Plane, P., 1997, ‘Privatisation and Economic Growth: An Empirical Investigation from a Sampleof Developing Market Economies’, Applied Economics, Vol.29, No.2, pp.161–78.
Ram, R and T. Schulz, 1979, ‘Life Span, Health, Savings, and Productivity’, EconomicsDevelopment and Cultural Change, Vol.27, No.3, pp.399–421.
Rousseeuw, P. and A. Leroy, 1987, Robust Regression and Outlier Detection, New York: JohnWiley.
Sala-i-Martin, X., 1994, ‘Cross-Sectional Regressions and the Empirics of Economic Growth’,European Economic Review, Vol.38, No.3/4, pp.739–47.
Sala-i-Martin, X., 1997, ‘I Just Run Two Million Regressions’, American Economic Review,Vol.87, No.2, pp.176–83.
Stokey, N., 1994, ‘Comments on Barro and Lee’, Carnegie-Rochester Conference Series onPublic Policy, Vol.40, pp.47–57.
Temple, J., 1999, ‘The New Growth Evidence’, Journal of Economic Literature, Vol.XXXVII,pp.112–56.
Tullock, G., 1965, The Politics of Bureaucracy, Washington, DC: Public Affairs Press.UNCTAD, 1996, Comparative Experiences with Privatisation: Policy Insights and Lessons
Learned, New York: United Nations.
146 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:13 Page 146
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
Vickers, J. and G. Yarrow, 1988, Privatisation: An Economic Analysis, London: MIT PressYarrow, G., 1999, ‘Theory of Privatisation, or Why Bureaucrats are still in Business’, World
Development, Vol.27, No.1, pp.157–68.Welsch, R., 1982, ‘Influence Functions and Regression Diagnostics’, in R. Launer and A. Siegel
(eds.), Modern Data Analysis, New York: Academic Press, pp.149–69. Welsch, R. and E. Kuh, 1977, ‘Technical Report’, in Linear Regression Diagnostics, 923-77,
Cambridge, MA: Sloan School of Management, MIT.World Bank, 1994, World Bank Assistance to Privatization in Developing Countries, OED
Report, No.3273, Aug. World Bank, 1995a, Bureaucrats in Business: The Economics and Politics of Reform, Oxford:
Oxford University Press, for World Bank.World Bank, 1995b, Structural and Sectoral Adjustment Lending: World Bank Experience
(1980–1992), Operations Evaluation Department, Washington, DC: World Bank.
APPENDIX 1
Consider the following three variable linear regression:
yt = xtβ + z1tγ1 + z2tγ2 + ut (1)
where β is the coefficient of particular interest to be estimated, x is the variable of particularinterest or the variable that is always included in a regression, z denotes the ‘doubtful’ variables,and γ is the corresponding coefficient to be estimated. Here, the z variables represent specificationuncertainty, and cannot be excluded a priori.
A Bayesian approach assumes that it is possible to define a prior distribution with the priorlocation of such doubtful variables and to specify their covariance matrix, thereby obtaining aposterior distribution for β. In practice, however, it is difficult to specify such a matrix. Leamer(1978) has developed a theorem in which specification of the prior covariance matrix is notrequired, assuming that specification of the prior location and the sample covariance matrix isonly necessary to constrain the posterior means to lie within a certain ellipsoid or the locus ofconstrained estimates.
Suppose that the object is to show that β is large or that the estimation of γ1 and γ2 is difficult.In a conventional approach, four different regressions would be computed using differentcombinations of the control variables, z1 and z2. Then, among the four different results obtained,the most favourable result would be reported. What Leamer has developed is an alternativeprocedure in which the specification search is enlarged and reports consist of the most favourableand least favourable results.
To enlarge the search for more specifications, a composite variable is defined as follows:
wt(θ) = z1t + θz2t, (2)
where θ is a number to be selected and is allowed to take any value.Then, equation (1) can be rewritten as:
yt = xtβ + wt(θ)γ + ut. (3)
The search can be enlarged by including all values of θ. For each value of θ, a differentconstraint of the form, γ2 = θγ1 is imposed, and there exits a least squares estimate of β, β(θ). Bymaximising β(θ) with respect to θ, the most favourable value of θ can be obtained, and the leastfavourable value can be obtained by minimising the value.
Figure A1 shows the theorem graphically. In a conventional approach, four differentregressions are estimated. These are indicated by the four points, (γ1*,γ2*), (γ1*,0), (0,γ2*), (0,0)in the figure. However, there is no reason why the search should be restricted by only estimatingthese four regressions. To enlarge the search, any values for θ can be used and a differentregression is estimated and the corresponding focus coefficient β* (θ) for each value of θ. Thedished line in the figure depicts the ellipse of constrained estimates defined by the set of allpossible values of (γ1,γ2) that are obtained by changing the value of over the real line. Thus, theellipse contains all posterior means for the distribution of (γ1, γ2).
147PRIVATISATION AND ECONOMIC GROWTH
396jds05.qxd 29/08/03 14:13 Page 147
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
148 THE JOURNAL OF DEVELOPMENT STUDIES
However, the extreme values, max β* (θ) and min β* (θ), may be located in an unlikelyparameter space on the locus of constrained estimates in relation to the data. To resolve thisproblem, define a statistically feasible likelihood ellipse, say at 0.05 level or 95 per cent, whichis depicted in the dashed lines in the figure, within the likelihood ellipse. This is defined by theassumption that all doubtful variables are included in the regression. The intersection of thepoints in the interior of the ellipse of constrained estimates and this likelihood ellipse is the set ofpoints (γ1
*, γ2*). The area defined by the set is shown by the shaded area in the figure.
Thus, the set of all posterior means for the distribution of (γ1*, γ2
*) is here subject to furtherconstraint in this area, eliminating the unlikely parameter space. The parameter pairs in the areacan be obtained by posterior estimates from prior distribution centred at the origin, and themeasure of specification uncertainty is the difference between the extreme values of β* (θ) overthe area.
FIGURE A1
ELLIPSOIDS OF CONSTRAINED LEAST SQUARES POINTS
396jds05.qxd 29/08/03 14:13 Page 148
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
AP
PE
ND
IX 2
PRIV
GY
PL
GY
PL
IFE
POP
CO
NST
AB
OPE
NG
DI
BU
DFD
IIM
PSD
IMP
INFL
SDIN
FM
3SD
M3
DE
BT
Arg
entin
a0.
100.
028.
5871
.01
0.01
0.04
6.00
0.16
0.17
–0.0
10.
010.
220.
245.
6510
.78
0.18
0.05
0.04
Ban
glad
esh
0.00
0.03
7.17
53.3
00.
020.
041.
000.
230.
19–0
.06
0.00
0.80
0.46
0.05
0.03
0.26
0.03
0.02
Bar
bado
s0.
030.
008.
9174
.34
0.00
0.20
0.00
0.98
0.15
–0.0
20.
010.
080.
070.
020.
020.
620.
060.
07B
eliz
e0.
070.
038.
0972
.57
0.03
0.15
0.00
1.16
0.26
–0.0
30.
030.
150.
100.
030.
020.
430.
030.
05B
enin
0.01
0.01
6.90
50.8
50.
030.
111.
000.
580.
16–0
.07
0.00
0.03
0.01
0.07
0.11
0.26
0.03
0.02
Bol
ivia
0.05
0.02
7.42
56.8
10.
020.
133.
000.
480.
15–0
.03
0.03
0.02
0.02
0.12
0.04
0.35
0.11
0.06
Bra
zil
0.04
0.01
8.34
64.7
60.
020.
174.
000.
170.
21–0
.08
0.01
0.29
0.32
10.2
19.
530.
440.
180.
03B
urki
na F
aso
0.00
0.01
6.26
45.8
70.
020.
146.
000.
380.
220.
000.
000.
030.
010.
050.
090.
230.
010.
02B
urun
di0.
00–0
.03
6.33
46.7
10.
020.
126.
000.
350.
13–0
.06
0.00
0.27
0.13
0.10
0.07
0.19
0.02
0.04
Cam
eroo
n0.
00–0
.04
7.19
53.3
10.
030.
112.
000.
400.
16–0
.03
0.00
0.03
0.01
0.04
0.06
0.19
0.04
0.05
Chi
le0.
010.
068.
2972
.67
0.02
0.10
0.00
0.61
0.25
0.03
0.04
0.13
0.06
0.13
0.07
0.41
0.02
0.08
Col
ombi
a0.
030.
028.
0868
.23
0.02
0.13
3.00
0.34
0.20
–0.0
10.
020.
090.
050.
240.
030.
320.
070.
08C
osta
Ric
a0.
000.
028.
1274
.84
0.02
0.16
0.00
0.82
0.26
–0.0
10.
030.
020.
070.
180.
050.
410.
030.
08C
ote
d’Iv
oire
0.02
–0.0
17.
2650
.96
0.03
0.15
1.00
0.68
0.11
–0.1
40.
010.
030.
010.
050.
130.
280.
010.
13E
cuad
or0.
000.
017.
9567
.04
0.02
0.10
2.00
0.57
0.20
–0.0
20.
020.
130.
100.
420.
160.
270.
060.
09E
gypt
, Ara
b R
ep.
0.02
0.02
7.55
61.0
70.
020.
112.
000.
530.
22–0
.04
0.02
0.07
0.19
0.13
0.05
0.87
0.03
0.06
Gha
na0.
040.
026.
7055
.96
0.03
0.11
0.00
0.50
0.18
–0.0
40.
010.
090.
100.
290.
100.
170.
020.
07G
uate
mal
a0.
000.
017.
6659
.69
0.03
0.06
12.0
00.
420.
15–0
.01
0.01
0.10
0.09
0.16
0.11
0.25
0.02
0.03
Gui
nea
0.00
0.02
6.65
42.5
50.
030.
100.
000.
490.
19–0
.07
0.00
0.07
0.05
0.12
0.11
0.07
0.04
0.04
Gui
nea-
Bis
sau
0.00
0.02
6.48
41.5
60.
020.
081.
000.
510.
31–0
.26
0.00
0.45
0.20
0.54
0.24
0.16
0.02
0.05
Hon
dura
s0.
010.
017.
2665
.40
0.03
0.13
3.00
0.75
0.28
–0.0
60.
020.
260.
300.
180.
080.
330.
030.
11In
dia
0.01
0.04
7.09
57.8
50.
020.
1111
.00
0.21
0.23
–0.0
60.
000.
100.
040.
090.
030.
480.
020.
03In
done
sia
0.01
0.06
7.46
60.2
10.
020.
081.
000.
510.
300.
000.
010.
070.
080.
090.
020.
430.
080.
09Ir
an, I
slam
ic R
ep.
0.00
0.02
8.08
64.7
30.
020.
120.
000.
390.
28–0
.02
0.00
2.29
1.19
0.27
0.09
0.53
0.09
0.04
Jam
aica
0.07
0.01
7.80
72.4
10.
010.
142.
001.
190.
31–0
.03
0.03
0.18
0.09
0.28
0.16
0.53
0.03
0.18
Jord
an0.
00–0
.02
8.16
66.7
10.
040.
250.
001.
320.
31–0
.02
0.00
0.05
0.03
0.06
0.06
1.18
0.15
0.13
Ken
ya0.
010.
006.
8057
.55
0.03
0.16
3.00
0.62
0.21
–0.0
50.
000.
140.
160.
130.
070.
470.
050.
10K
orea
, Sou
th0.
010.
078.
6369
.27
0.01
0.10
0.00
0.64
0.36
–0.0
10.
000.
010.
050.
060.
020.
670.
140.
03M
alaw
i0.
000.
016.
2044
.88
0.03
0.16
0.00
0.60
0.19
–0.0
70.
000.
290.
160.
290.
240.
230.
040.
06M
alay
sia
0.08
0.06
8.36
69.5
00.
030.
131.
001.
650.
36–0
.03
0.06
0.01
0.01
0.04
0.01
1.05
0.27
0.10
396jds05.qxd 29/08/03 14:13 Page 149
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
Mal
i0.
000.
016.
2546
.51
0.03
0.13
2.00
0.54
0.20
–0.0
50.
000.
030.
010.
060.
100.
220.
010.
03M
auri
tani
a0.
000.
016.
6749
.46
0.03
0.13
0.00
1.02
0.20
–0.0
20.
010.
700.
570.
060.
030.
230.
050.
12M
exic
o0.
050.
018.
5869
.82
0.02
0.10
0.00
0.44
0.23
0.00
0.02
0.05
0.05
0.31
0.29
0.25
0.07
0.08
Mor
occo
0.02
0.01
7.66
62.0
10.
020.
172.
000.
570.
22–0
.04
0.01
0.05
0.04
0.04
0.02
0.66
0.08
0.10
Moz
ambi
que
0.01
0.03
6.63
43.4
70.
020.
136.
000.
520.
17–0
.27
0.01
0.25
0.20
0.43
0.14
0.33
0.05
0.06
Nep
al0.
000.
036.
9152
.02
0.03
0.09
1.00
0.46
0.22
–0.0
80.
000.
300.
110.
100.
040.
350.
030.
02N
icar
agua
0.02
–0.0
17.
2762
.17
0.03
0.22
7.00
0.80
0.23
–0.1
50.
020.
410.
6331
.11
45.4
40.
360.
160.
13N
iger
ia0.
010.
026.
8548
.31
0.03
0.13
3.00
0.74
0.18
–0.1
00.
040.
640.
500.
360.
240.
200.
050.
09O
man
0.00
0.01
8.89
67.9
60.
040.
330.
000.
860.
16–0
.09
0.01
0.02
0.01
0.00
0.09
0.31
0.02
0.07
Paki
stan
0.01
0.02
7.22
58.2
40.
030.
146.
000.
360.
16–0
.08
0.01
0.06
0.03
0.10
0.02
0.44
0.03
0.05
Pana
ma
0.06
0.01
7.94
71.7
40.
020.
1711
.00
0.73
0.19
0.02
0.02
0.00
0.00
0.02
0.02
0.52
0.13
0.07
Papu
a N
ew G
uine
a0.
050.
017.
3953
.93
0.02
0.20
9.00
0.97
0.25
–0.0
30.
040.
100.
060.
050.
060.
340.
030.
14Pa
ragu
ay0.
000.
017.
6167
.60
0.03
0.08
0.00
0.82
0.23
0.01
0.02
0.23
0.25
0.20
0.09
0.26
0.05
0.05
Peru
0.09
0.00
7.91
64.3
90.
020.
0812
.00
0.26
0.21
–0.0
20.
020.
360.
4910
.13
20.1
10.
190.
040.
03Ph
ilipp
ines
0.02
0.01
7.42
64.0
00.
020.
1116
.00
0.72
0.23
–0.0
40.
020.
050.
020.
090.
030.
460.
120.
07Se
nega
l0.
020.
007.
0748
.25
0.03
0.14
0.00
0.62
0.15
–0.0
10.
010.
030.
010.
040.
240.
230.
010.
06Si
erra
Leo
ne0.
00–0
.05
6.79
36.8
10.
020.
101.
000.
410.
07–0
.06
0.00
0.88
0.93
0.48
0.27
0.13
0.03
0.06
Sing
apor
e0.
080.
079.
2473
.59
0.02
0.09
0.00
3.68
0.35
0.12
0.10
0.02
0.01
0.04
0.02
1.18
0.04
0.11
Sout
h A
fric
a0.
02–0
.01
8.12
60.4
30.
020.
201.
000.
480.
17–0
.05
0.01
0.04
0.02
0.12
0.03
0.46
0.02
0.03
Sri L
anka
0.02
0.04
7.61
70.7
00.
010.
1012
.00
0.73
0.24
–0.1
00.
010.
130.
100.
110.
030.
400.
020.
04Ta
nzan
ia0.
010.
006.
2851
.01
0.03
0.17
0.00
0.51
0.22
–0.0
90.
010.
300.
230.
260.
080.
240.
030.
04T
haila
nd0.
000.
078.
0067
.49
0.01
0.10
5.00
0.80
0.39
0.02
0.02
0.03
0.01
0.05
0.01
0.79
0.07
0.06
Togo
0.01
0.00
6.46
51.4
10.
030.
138.
000.
750.
16–0
.05
0.00
0.03
0.01
0.07
0.11
0.32
0.05
0.04
Tri
nida
d an
d 0.
060.
009.
0070
.39
0.01
0.13
2.00
0.83
0.16
–0.0
20.
050.
230.
180.
070.
070.
530.
050.
10To
bago
Tun
isia
0.00
0.02
7.90
65.6
30.
020.
160.
000.
890.
26–0
.04
0.02
0.05
0.05
0.06
0.02
0.50
0.02
0.10
Tur
key
0.01
0.02
8.14
64.2
50.
020.
115.
000.
390.
24–0
.05
0.00
0.03
0.04
0.75
0.15
0.29
0.04
0.06
Uga
nda
0.01
0.04
6.27
48.3
40.
030.
092.
000.
300.
14–0
.07
0.01
0.38
0.40
0.48
0.60
0.10
0.01
0.04
Uru
guay
0.00
0.02
8.45
72.1
10.
010.
132.
000.
430.
13–0
.03
0.00
0.12
0.08
0.58
0.28
0.49
0.09
0.06
Ven
ezue
la0.
040.
008.
8270
.53
0.02
0.08
8.00
0.54
0.18
–0.0
20.
020.
190.
360.
500.
310.
350.
080.
08
AP
PE
ND
IX 2
(con
t.)
PRIV
GY
PL
GY
PL
IFE
POP
CO
NST
AB
OPE
NG
DI
BU
DFD
IIM
PSD
IMP
INFL
SDIN
FM
3SD
M3
DE
BT
396jds05.qxd 29/08/03 14:13 Page 150
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
AP
PE
ND
IX 2
(con
t.)
PRIV
GY
PL
GY
PL
IFE
POP
CO
NST
AB
OPE
NG
DI
BU
DFD
IIM
PSD
IMP
INFL
SDIN
FM
3SD
M3
DE
BT
Vie
tnam
0.00
0.05
6.79
65.8
00.
020.
080.
000.
690.
210.
000.
040.
460.
730.
701.
230.
220.
010.
03Y
emen
0.00
0.00
7.54
51.4
90.
040.
190.
000.
610.
20–0
.08
0.04
0.17
0.07
0.25
0.16
0.51
0.07
0.51
Zam
bia
0.05
–0.0
16.
6349
.63
0.03
0.17
1.00
0.72
0.13
–0.1
10.
020.
690.
680.
770.
500.
220.
070.
17Z
imba
bwe
0.01
0.01
6.99
56.9
10.
030.
190.
000.
620.
21–0
.09
0.00
0.29
0.16
0.21
0.06
0.40
0.05
0.08
Mea
n0.
020.
027.
5159
.83
0.02
0.13
3.06
0.67
0.21
–0.0
50.
020.
220.
181.
091.
480.
390.
060.
08St
anda
rd D
evia
tion
0.03
0.02
0.80
9.95
0.01
0.05
3.88
0.47
0.06
0.06
0.02
0.33
0.24
4.28
6.40
0.23
0.05
0.07
Not
es:
PRIV
= t
he m
agni
tude
of
priv
atis
atio
n as
the
rat
io o
f cu
mul
ativ
e pr
ocee
ds f
rom
pri
vatis
atio
n du
ring
the
sam
ple
peri
od 1
988-
97 t
o av
erag
e G
DP
duri
ng t
he s
ame
peri
od;
prim
ary
data
for
the
pro
ceed
s ar
e ta
ken
from
Wor
ld B
ank
Priv
atiz
atio
n D
atab
ase
and
Priv
atis
atio
n In
tern
atio
nal
(var
ious
issu
es);
and
the
data
for
GD
P ar
e ta
ken
from
Wor
ld B
ank
Glo
bal D
evel
opm
ent N
etw
ork
Gro
wth
Dat
abas
e.G
YP
= t
he a
vera
ge r
eal
GD
P pe
r ca
pita
(in
con
stan
t U
S do
llars
, int
erna
tiona
l pr
ices
, bas
e ye
ar 1
985)
gro
wth
rat
e du
ring
the
sam
ple
peri
od;
and
the
data
are
take
n fr
om W
orld
Ban
k G
loba
l Dev
elop
men
t Net
wor
k G
row
th D
atab
ase.
LG
YP
=th
e lo
g of
ini
tial
real
GD
P pe
r ca
pita
in
1988
cap
ita (
in c
onst
ant
US
dolla
rs, i
nter
natio
nal
pric
es, b
ase
year
198
5);
and
prim
ary
data
are
take
n fr
om W
orld
Ban
k G
loba
l Dev
elop
men
t Net
wor
k G
row
th D
atab
ase
POP
= t
he a
vera
ge p
opul
atio
n gr
owth
rat
e du
ring
the
sam
ple
peri
od;
and
the
data
are
tak
en f
rom
Wor
ld B
ank
Glo
bal
Dev
elop
men
t N
etw
ork
Gro
wth
Dat
abas
eL
IFE
= t
he i
nitia
l lif
e ex
pect
ancy
at
birt
h in
198
7 (t
he d
ata
in 1
988
are
not
avai
labl
e);
and
the
prim
ary
data
are
tak
en f
rom
Wor
ld B
ank
Glo
bal
Dev
elop
men
t Net
wor
k G
row
th D
atab
ase
STA
B=
pol
itica
l in
stab
ility
cal
cula
ted
as t
he a
vera
ge n
umbe
r of
maj
or p
oliti
cal
cris
es, r
evol
utio
ns, a
nd c
oups
dur
ing
the
sam
ple
peri
od;
and
the
prim
ary
data
are
take
n fr
om W
orld
Ban
k G
loba
l Dev
elop
men
t Net
wor
k G
row
th D
atab
ase
BU
D=
the
aver
age
ratio
of
cent
ral g
over
nmen
t bud
get s
urpl
us/d
efic
it to
GD
P (c
alcu
late
d in
nat
iona
l cur
renc
y, c
onst
ant p
rice
) du
ring
the
sam
ple
peri
od; a
nd th
e pr
imar
y da
ta a
re ta
ken
from
Wor
ld B
ank
Glo
bal D
evel
opm
ent N
etw
ork
Gro
wth
Dat
abas
eO
PEN
= o
penn
ess
of a
cou
ntry
cal
cula
ted
as (
expo
rt-i
mpo
rt)/
GD
P; a
nd th
e pr
imar
y da
ta a
re ta
ken
from
Wor
ld B
ank
Glo
bal D
evel
opm
ent N
etw
ork
Gro
wth
Dat
abas
eC
ON
=th
e av
erag
e ra
tio o
f go
vern
men
t con
sum
ptio
n to
GD
P (c
alcu
late
d in
nat
iona
l cur
renc
y, c
onst
ant p
rice
) du
ring
the
sam
ple
peri
od; a
nd th
epr
imar
y da
ta a
re ta
ken
from
Wor
ld B
ank
Wor
ld D
evel
opm
ent I
ndic
ator
sG
DI
= t
he a
vera
ge r
atio
of
gros
s do
mes
tic in
vest
men
t to
GD
P (c
alcu
late
d in
nat
iona
l cur
renc
y, c
onst
ant p
rice
) du
ring
the
sam
ple
peri
od; a
nd th
epr
imar
y da
ta a
re ta
ken
from
Wor
ld B
ank
Wor
ld G
loba
l Dev
elop
men
t Net
wor
k G
row
th D
atab
ase
396jds05.qxd 29/08/03 14:13 Page 151
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
152 THE JOURNAL OF DEVELOPMENT STUDIES
AP
PE
ND
IX 2
(con
t.)
FDI
= t
he a
vera
ge r
atio
of
fore
ign
dire
ct i
nves
tmen
t (n
et i
nflo
ws)
to
GD
P (c
alcu
late
d in
cur
rent
US
dolla
rs)
duri
ng t
he s
ampl
e pe
riod
; an
d th
epr
imar
y da
ta a
re ta
ken
from
Wor
ld B
ank
Wor
ld D
evel
opm
ent I
ndic
ator
sIN
FL=
the
ave
rage
infl
atio
n ra
tes
(GD
P de
flat
or)
duri
ng th
e sa
mpl
e pe
riod
; and
the
prim
ary
data
are
take
n fr
om W
orld
Ban
k W
orld
Dev
elop
men
tIn
dica
tors
SDIN
F=
the
sta
ndar
d de
viat
ion
of I
NFL
IMP
= l
og(1
+ in
form
al m
arke
t pre
miu
m);
info
rmal
mar
ket p
rem
ium
is c
alcu
late
d us
ing
the
follo
win
g fo
rmul
a: (
Para
llel E
xcha
nge
Rat
e/O
ffic
ial
Exc
hang
e R
ate
-1)*
100;
the
prim
ary
data
are
take
n fr
om W
orld
Ban
k G
loba
l Dev
elop
men
t Net
wor
k G
row
th D
atab
ase
SDIM
P=
the
sta
ndar
d de
viat
ion
of I
MP
M3
= t
he r
atio
of
liqui
d lia
bilit
ies
(M3)
to G
DP,
the
prim
ary
data
are
take
n fr
om W
orld
Ban
k G
loba
l Dev
elop
men
t Net
wor
k G
row
th D
atab
ase
SDM
3=
the
stan
dard
dev
iatio
n of
M3
DE
BT
=th
e ra
tio o
f to
tal e
xter
nal d
ebt t
o G
DP,
the
prim
ary
data
are
take
n fr
om W
orld
Ban
k G
loba
l Dev
elop
men
t Net
wor
k G
row
th D
atab
ase.
* W
orld
Ban
k G
loba
l Dev
elop
men
t Net
wor
k G
row
th D
atab
ase
is a
vaila
ble
from
ww
w.w
orld
bank
.org
/res
earc
h/gr
owth
/GD
Nda
ta.h
tm
396jds05.qxd 29/08/03 14:13 Page 152
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
153PRIVATISATION AND ECONOMIC GROWTH
APPENDIX 3
Influential Observations/Outlier Detection Techniques:In general outliers are categorised into two types. The first is characterised as the
substantially large residual arising from the difference between the actual Yi and the fitted Y∧
i. Thesecond is that Xi is far from the cluster of Xi’s, resulting in the direction of the fitted line varying,although this is not indicated by a high residual. Detecting more than one outlier in a dataset is,in practice, difficult owing to the so-called masking and swamping effects [Luke and Wilk, 2000].The masking effect occurs when a subset of outliers is not identified due to the presence ofanother subset of observations. The swamping effect occurs when normal observations aremisidentified as outliers due to the presence of another subset of observations.
One way to detect influential observations/outliers is to analyse the size of residuals from theregression estimation. This can be done by computing studentised residuals. The studentisedresidual can be computed as:
SRi
= ei
/ (s(i)
),
where hi is the ith diagonal element of the hat matrix H, s is the variance from sample residuals,and ei is the ordinary residual. (i) is the standard error obtained by excluding the ith observation,and then s(i) is calculated as:
s(i)2 = ,
where k is the number of explanatory variables and n is the number of observations in theregression.
DFITS, Cook’s Distance, and Welsch Distance summarise the information in the leverageversus residual-squared plot into a single statistics by creating an index that is influenced by thesize of the residuals or outliers and the size of leverage.
DEFITS can be calculated by:
DFITSi = ,
where ri are the studentised residuals and hi are the size of leverage. Cook’s Distance is calculated by:
CDi = 1/k • si2/s2 • DFITSi
2 ,
where s is the root mean square error of the regression, and si is the root mean square error whenthe ith observation is eliminated.
Welsch Distance can be estimated by:
WDi = DFITSi ,
Then, outliers or the observations that ought to be examined further are defined as:
DFITS > 2 , , Cook’s Distance > 4/n, and Welsch Distance > 3 .knk
ih
n
−−
1
1
i
ii
h
hr
−1
−−
−−1
ˆ22
kn
ekns i
1
)1/()( 22
−−−−−
kn
heskn ii
ih−1
396jds05.qxd 29/08/03 14:14 Page 153
Dow
nloa
ded
By:
[Hiro
shim
a U
nive
rsity
] At:
07:1
6 25
Jul
y 20
08
DFBETAs measure the effect on each coefficient by testing β∧
j – β∧
(i)j , which is defined as:
DFBETAsij = ,
where is the (X′ X)–1jj is the jth diagonal element of (X′ X)–1. The observations that ought to be
examined further are defined as DFBETAs > .n/2
1')( )(
)(ˆˆ
−
−
jji
jj
XXs
iββ
154 THE JOURNAL OF DEVELOPMENT STUDIES
396jds05.qxd 29/08/03 14:14 Page 154