Changes in the Operational Efﬁciency of National Oil …...9. The intercept term thus measures the...

27

* Corresponding author. George and Cynthia Mitchell Professor of Economics and Rice Scholar,James A. Baker III Institute for Public Policy, Rice University MS22, 6100 Street Houston, TX,77005-1892 and Professor at Large and BHP Billiton Chair in Economics, University of WesternAustralia. E-mail: [email protected].

** James A. Baker, III, and Susan G. Baker Fellow in Energy and Resource Economics & DeputyDirector, Baker Institute Energy Forum and Adjunct Professor, Economics Department, RiceUniversity.

The authors thank James D. Coan, Sara El Hakim, Jane Kliakhandler and Likeleli Seitlheko for veryvaluable research assistance. We also thank Ken Clements at the University of Western Australia,participants in the 4th International Workshop on Empirical Methods in Energy Economics (EMEE)held in Dallas in July, 2011 and especially the discussant at EMEE of an earlier version of this paper,Clifton Jones from Stephen F. Austin University, for valuable comments.

The Energy Journal, Vol. 34, No. 2. Copyright � 2013 by the IAEE. All rights reserved.

Changes in the Operational Efficiency of National OilCompanies

Peter R. Hartley* and Kenneth B. Medlock III**

Using data on 61 oil companies from 2001–09, we examine the evolutionof revenue efficiency of National Oil Companies (NOCs) and shareholder-ownedoil companies (SOCs). We find that NOCs generally are less efficient than SOCs,but their efficiency increased faster over the last decade. We also find evidencethat partial privatizations increase operational efficiency, and (weaker) evidencethat mergers and acquisitions during the decade tended to increase the efficiencyof the merging firms. Finally, we find evidence that much of the inefficiency ofNOCs is consistent with the hypothesis that government ownership leads to dif-ferent firm objectives.

Keywords: National Oil Company, Government ownership, Revenueefficiency, Data envelopment, Stochastic Frontier

http://dx.doi.org/10.5547/01956574.34.2.2

1. INTRODUCTION

Using data on 61 oil and gas firms we assess how the revenue efficiencyof national oil companies (NOCs) and shareholder-owned oil companies (SOCs)changed over 2001–09. The analysis reaffirms findings in Eller et al (2011) that

28 / The Energy Journal

Copyright � 2013 by the IAEE. All rights reserved.

1. In our sample, NOCs held more than 82% of the crude oil reserves in 2009, and the six largest,and eight of the top ten largest holders were NOCs. ExxonMobil is the only SOC in this group, atrank of nine in ten.

2. In the future, new technologies may enable more production from unconventional resources orintroduce new oil substitutes that lessen the dominance of NOCs in the markets that oil products nowserve.

NOCs tend to be less revenue efficient than SOCs. The longer time period ex-amined in this paper, however, helps explain what types of changes may increaseefficiency.

We examine revenue efficiency for several reasons. Revenue is a keyobjective for both public and private firms. Furthermore, if the political overseersof an NOC force it to sell products to domestic consumers at subsidized prices,lost revenue would capture the effects in a way that physical output measureswould not. In addition, almost all firms in the industry produce a range of productsfrom crude oil and natural gas to refined products. The natural way to aggregatethese outputs is to value them at market prices, and hence take revenue as theoutcome measure. Finally, revenue figures are more readily available for manyfirms than physical outputs of different commodities.

Investigating the efficiency of NOCs relative to SOCs is of interest inpart because NOCs are the top holders of crude oil reserves.1 As a result, NOCsdominate global oil production and could be expected to do so for some time.2

If NOCs are less efficient than SOCs, we are likely to see lower oil productionand higher oil prices than would be the case had SOCs exploited the same re-sources.

It may also be useful to understand why NOCs tend to be less efficientlymanaged than SOCs. Hartley and Medlock (2008) argued that political pressureis likely to force NOCs to inefficiently distribute rents to domestic consumers byselling products at below-market prices, and to workers by overemploying labor.Once this is understood, governments or international organizations such as theWorld Bank may be in a better position to devise policies that allow resourcerents to be shared more efficiently.

In addition, panel data allows us to examine how the relative efficiencyof firms has changed over time. We find that while oil and gas firms as a wholetended to become more efficient at producing revenue over the decade 2001–09,NOCs on average gained more than SOCs largely because they started furtherfrom the efficient frontier. We also find evidence that partial privatization, andmergers and acquisitions, likely increased efficiency, which may explain theseactions in the first place.

Finally, for any given inefficient firm, the analysis reveals which efficientfirms are most representative of the inefficient firm in terms of the operationalvariables. This, in turn, allows us to identify which firms may be suitable modelsto emulate. Some NOCs that are found to be as efficient as the major SOCs mayalso be suitable role models for governments wishing to improve the performanceof their NOCs.

Changes in the Operational Efficiency of National Oil Companies / 29


3. The PIW Top 50 is the precursor to the more comprehensive Energy Intelligence “Top 100:Ranking The World’s Oil Companies,” which is the source we used for the current paper.

4. SFA is a special type of random effects multivariate panel regression estimator. Wolf commentsthat he prefers the fixed effects panel estimator because the random effects estimator yields incon-sistent results unless the firm-specific error terms are uncorrelated with the included measured vari-ables. He rejects the latter hypothesis using a Hausman specification test.

We use two methods to estimate revenue efficiency and changes in rev-enue efficiency: non-parametric data envelopment analysis (DEA) and parametricstochastic frontier analysis (SFA). The fact that we find similar results using verydifferent techniques adds to the confidence that the results reflect genuine differ-ences between firms rather than artifacts of the estimation methodologies.

2. RELATED LITERATURE

Despite the importance of NOCs in the world oil market, very few au-thors have examined the relative efficiencies of NOC’s using formal econometrictechniques such as DEA or SFA. Al-Obaidan and Scully (1991) used data for 44firms in a single year, 1981, to construct a production frontier using both deter-ministic and stochastic methods. Specifically, they examined the ability of firmsto use assets and employees to produce and process crude oil or earn revenue. Inparticular, they found that NOCs are only 63% to 65% as efficient as private firmsin generating revenue. Although our results are generally consistent with thoseof Al-Obaidan and Scully, our study differs in many respects, particularly in thedata used. We have panel data and include a broader set of oil companies thanAl-Obaidan and Scully. They omit all OPEC nations arguing that the demon-strated efficiency of those firms is “related more to the accident of geographythan to the allocation of resources within the firm.”

Our analysis also is related to two recent papers from the ElectricityPolicy Research Group at the University of Cambridge (Wolf (2009) and Wolfand Pollitt (2008)). Like Eller et al (2011), Wolf (2009) uses the Petroleum In-telligence Weekly “Ranking the World’s 50 Top Oil Companies” (PIW Top 50)as his data source.3 He conducts multivariate regression analyses with differentdependent variables and two different estimators. He considers a panel modelwith firm-specific intercept terms, and a total (or pooled ordinary least squares)estimator that ignores firm-specific heterogeneity. Wolf notes that the firm-specificintercept term in the fixed effects estimator captures “all (observed and unob-served) time-invariant variables” that affect the dependent variable. Thus, all firmsthat do not change ownership over the sample period are treated identically re-gardless of their extent of government or shareholder ownership. On the otherhand, while the total estimator permits an estimation of the (cross-sectional) effectof ownership, it cannot control for any firm-specific unobserved variables.

The SFA analyses conducted in this paper and Eller et al (2011) are alsomultivariate panel regression analyses,4 but with a special assumed structure on



5. Greene (2005) proposed modifying SFA to allow for fixed or random effects to measure firmheterogeneity apart from the one-sided random inefficiency term. In Eller et al (2011) and the analysisconducted below we estimate a parametric model where the firm-specific error component containstwo terms (the time variable and government ownership share) that can be regarded as “inefficiency”effects and one term (a vertical integration measure) more appropriately regarded as measuring firmheterogeneity.

6. For example, including 2000 data would have eliminated an additional 10 firms from our sampleincluding some important NOCs. Wolf examines data from 1987–2006, but many firms in his samplehave missing data for some variables in many years.

7. The performance metrics included returns on sales, assets and equity; output, revenue and netprofit per employee; finding and development, and production, costs per barrel of oil equivalent;reserve replacement ratio; capital expenditure; financial leverage and dividend payments.

8. In many cases, this produced a continuous sample from three years prior to the first offeringto three years after the final offering.

the error terms. In the simplest form of SFA, the error terms are assumed to havetime-invariant firm-specific components, drawn from a distribution that is strictlynonnegative, which represent deviations in firm efficiency from the efficient fron-tier. Another component of the error, representing measurement error or omittedexplanatory variables for example, is assumed to have a symmetric distribution.5

By contrast, the standard random effects panel estimator assumes error terms aresymmetric and thus ignores efficiency deviations.

Another difference between the analysis in Wolf (2009) and SFA is thattheory imposes more structure on the estimated equation in SFA. As we discussin more detail below, right hand side variables in SFA arise from a productionfunction. By contrast, none of the equations Wolf (2009) examines has a structuralinterpretation, making it difficult to specify the appropriate estimating equation.

Imposing more structure could distort conclusions if the additional as-sumptions are inappropriate. That is a major motivation for also using non-para-metric DEA. DEA constructs the revenue efficient frontier as a piecewise-linearoutermost limit of the set of observed input-output bundles in each year and thenmeasures the distance of firms from that frontier. While DEA avoids makingdetailed assumptions about the underlying production function, it does not allowfor measurement error or other sources of variation across firms that are unrelatedto differences in inputs or relative efficiency.

Different from Wolf’s analysis, but similar to that in Eller et al (2011),the methods we employ require a balanced panel (that is, all firms included inthe analysis require observations for all variables for all the years). This ultimatelyreduces the number of years and firms in our data set.6

Wolf and Pollitt (2008) focus on 60 share-issue privatizations by 28former NOCs (from 20 different countries) from 1977 to 2004. Many of thesewere follow-on offerings as government ownership shares were reduced in severalsteps. The NOCs included in their sample are predominantly from the more de-veloped world (17 of the 28 are from OECD countries and none are from OPEC).The authors collected firm performance metrics7 for seven-year periods surround-ing each offering.8 They then compared mean performance three years prior to



9. The intercept term thus measures the change in average performance as a result of privatizationwhile the year coefficients measure changes in performance trends. Unlike the simple comparison ofbefore and after means, the panel regression models control for other firm-specific influences andchanges in real oil prices.

10. She examined revenue per employee, return on assets, liquids and natural gas productionrelative to liquids and natural gas reserves (respectively), and revenue relative to production

the asset sale to mean performance three years after the asset sale. They generallyfound that privatization is associated with higher profitability, improved operatingefficiency, greater output and lower employment.

The authors also estimated a fixed effects panel data model allowing theintercept and the coefficient on “year” (a discrete variable ranging from 1–7 be-ginning 3 years prior to the privatization and ending 3 years after the asset sale)to differ before and after the privatization.9 They found that initial share-issueprivatizations improved average performance by all measures, but the effect wasstatistically significant at the 10% level only for increased return on sales or assetsand reduced employment per unit of assets. The trend in performance was fa-vorable for all metrics, and statistically significantly so at the 1% level for in-creased return on sales or assets, output per employee, output, and reduced em-ployment per unit of assets. On the other hand, the trend after privatization wasless favorable than the trend prior to privatization for seven out of ten indicators(although statistically significant only for returns on sales or assets). Wolf andPollitt conclude that performance generally improves as a result of privatizationbut the improvements begin in anticipation of the subsequent share sale and, ifanything, tend to slow down after the shares are sold.

The authors found much less conclusive effects of share offerings sub-sequent to the initial privatizations. The only strong result was that continuedreductions in government ownership were associated with continued reductionsin employment, which is consistent with some of the results we report below.

Victor (2007) analyzed data from the 1999–2006 editions of EnergyIntelligence “Top 100: Ranking The World’s Oil Companies.” Like Wolf (2009)and Wolf and Pollitt (2008), she examines a range of performance indicators.10

However, these analyses are each one dimensional, raising the possibility of omit-ted variable bias in the estimated coefficients. They also are all conducted assimple regressions without any allowance for the asymmetric error terms thatcould arise from constraints on maximum efficiency.

3. OVERVIEW OF METHODS

Consider an oil and gas firm i producing a revenue of yit millions ofcurrent $US in year t using inputs, xikt where k = 1. . .K, are oil (millions of barrels)reserves, natural gas (billions of cubic feet) reserves, distillation capacity (thou-sands of barrels per day), employees (end of year head count), and oil (current$US per barrel) and natural gas (current $US per MMBTU) prices. These input



11. Geological factors such as reservoir pressure and porosity bound the maximum annual outputQ from given proved reserves and likely vary with field age. We do not have data on the geologicalcharacteristics or average age of the fields exploited by each firm. These factors, therefore, are likelyto be a significant component of the error terms in the models subsequently estimated using SFA.

12. Recent econometric literature has bridged DEA and SFA by introducing statistical noise intoDEA or formulating inefficiency in SFA non-parametrically. Grosskopf (1996) is an early survey of

choices were motivated by the theoretical analysis of NOC behavior in Hartleyand Medlock (2008). Specifically, Hartley and Medlock (2008) assumed that cur-rent output Q of an oil-producing firm is given by

Q = F(L)� Rsv� G(E) (1)

where L is labor input, Rsv is proved reserves, and G(E) represents geologicallimitations on field productivity that depend on the level of cumulative past ex-ploitation, E.11 In its downstream operations, the firm uses labor and capital (es-pecially refining capacity) along with crude and wellhead natural gas to producemarketable products q, such that

q = H(K,L,Q) (2)

Revenue will then be given by

p(1– s) ⋅ q (3)

for a vector of product prices p and corresponding percentage subsidies s on eachmarketed product. While the prices of the products will vary, they will be coin-tegrated with crude prices, so we used the latter as a proxy for all relevant oilproduct prices.

As already noted, we use both DEA and SFA to determine the techno-logical frontier of the most efficient firms in each year and the distance of firmsfrom that frontier. An advantage of SFA is that the assumption that yit is revenueand the xikt are productive inputs restricts the functional relationship between yi

and the xik variables. On the other hand, DEA could be preferable if the assump-tions about the functional form of the technology are inaccurate. DEA also avoidspotential problems from assuming an inappropriate distribution of the error term.However, since DEA does not account for statistical noise, estimates of efficiencywill be biased when stochastic elements are important or variables are measuredwith error. In addition, SFA provides a statistical measure of how well the pro-posed model explains the data.

Both DEA and SFA have been used extensively to analyze productiveefficiency. Comparison of methods is available in Gong and Sickles (1992),Banker (1993), Cooper and Tone (1997), and Ruggiero (2007), to name a few.Our intention is not to examine the relative merits of different methods for mea-suring efficiency. Rather, we use both approaches to ensure our conclusions arerobust to choice of method.12



literature on statistical inference in DEA models. Desai et al. (2005) assumed the constraints in DEAonly held probabilistically. Tsionas (2003) discussed using DEA estimates of efficiency as priors forthe one-sided errors in an SFA model, which then are derived as posterior estimates in a Monte CarloBayesian analysis.

13. Where there were discrepancies between the PIW and annual report data, we assumed thatthe annual report data was correct (this was clearly the case for some obvious typographical errors).

14. Most, but not all, of the eliminated firms with missing data were NOCs or pNOCs and usuallythere was no annual report available for the year in question, although sometimes just one variable(like employment) might not have been reported.

15. We subsequently dropped Nippon Oil because its very heavy concentration in downstreamproduct markets made it an extreme outlier. Nippon Oil product sales were on average almost 20times its liquids production, compared with a mean across all firms of slightly less than 1.5 and meanfor the next highest firm of slightly more than 8.5. Including Nippon Oil substantially altered theeffect of this measure of vertical integration, but most of the remaining coefficients were not muchaffected by its inclusion.

16. Results were not materially affected by using just one “representative” price series for eachcommodity, no doubt because different oil and natural gas prices are very highly correlated with eachother.

4. DATA

As in Eller et al (2011), Victor (2007) and Wolf (2009), the primary datasource was the Energy Intelligence annual publication “Ranking the World’s OilCompanies”. We also consulted annual reports to check and revise data and toprovide missing data.13

Although we began with almost 150 firms, a firm missing just one vari-able in one year had to be dropped because the methods we use need a balancedpanel. This constrained both the number of years and the number of firms in thesample.14

In addition, to obtain measures for all variables for each firm, for firmsthat merged we combined the inputs and revenues of merger partners in yearsprior to the merger, further reducing the number of separate firms. To keep trackof the “synthetic” firms so formed, we defined an indicator variable (Premerge)that was set to 1 in years before the merger and to zero in years following themerger.

We were left with 61 firms covering 2001–2009, which compares witha sample of 78 firms over three years (2002–2004) in Eller et al (2011).15 Table1 lists the averages of key variables for each company in the sample.

We also used data on oil and natural gas prices from the US EnergyInformation Administration (EIA) and the International Energy Agency (IEA).We used the average annual US import oil price for North American firms, theaverage annual OPEC oil price for OPEC members, and the average non-OPECoil price for other firms. For natural gas prices, we used the average annual pricesat the Henry Hub for North American firms, average annual Japanese LNG importprices for firms in Asia, the Pacific and the Middle East, EU pipeline importprices for firms predominantly selling in Europe, and EU LNG import prices forLNG exporting firms in the Atlantic Basin.16



Tab

le1:

Mea

nsof

Key

Var

iabl

es

Fir

mH

eadq

uart

erC

ount

ry

Gov

ernm

ent

owne

rshi

psh

are

Rev

enue

($m

illio

n)

Oil

rese

rves

(mill

ion

barr

els)

Nat

.gas

rese

rves

(bill

ion

ft3 )

Refi

ning

Cap

acit

y(t

hous

ands

barr

els/

day)

Em

ploy

ees

Rev

enue

per

empl

oyee

Ver

tica

lin

tegr

atio

n(p

rodu

cts/

oil

prod

)

Adn

ocU

AE

127

260.

354

590.

613

8424

.063

3.1

8500

00.

321

0.23

0A

nada

rko

US

010

208.

113

66.2

9650

.20

5812

2.00

00

Apa

che

US

068

65.9

925.

763

23.4

028

262.

295

0B

GU

K0

1141

2.4

591.

191

64.3

050

752.

166

0B

HPB

illito

nA

ustr

alia

032

934.

958

6.7

4917

.80

3628

70.

882

0B

PU

K0

2458

61.0

9817

.846

815.

329

63.9

1003

542.

511

2.66

2C

NO

OC

Chi

na0.

6810

896.

914

83.7

5009

.80

2735

3.55

10.

047

CN

PCC

hina

0.93

9098

7.8

1588

7.0

6264

6.4

2391

.445

0057

0.19

40.

687

CN

RC

anad

a0

7565

.919

43.2

3024

.90

2661

2.84

50

Che

sape

ake

US

050

28.3

87.2

6778

.60

3792

1.37

90

Che

vron

US

016

7549

.083

20.0

2473

0.4

2197

.266

478

2.56

61.

967

Con

ocoP

hilli

psU

S0

1327

33.0

6070

.624

388.

326

90.7

4124

53.

663

1.99

0D

evon

US

088

96.7

903.

880

10.9

044

181.

981

0E

NI

Ital

y0.

3092

411.

836

94.2

1826

1.1

719.

374

978

1.22

71.

015

EO

GU

S0

3233

.114

3.7

5793

.00

1476

2.07

40

Eco

petr

olC

olom

bia

0.96

8126

.712

32.7

2928

.232

0.4

6516

1.27

10.

787

EnC

ana

Can

ada

013

724.

976

5.1

1028

0.7

50.2

4394

2.93

30.

376

Exx

onM

obil

US

029

5090

.011

851.

062

626.

463

00.9

8635

63.

496

2.94

8G

azpr

omR

ussi

a0.

4866

477.

913

370.

785

5000

.063

1.3

3865

800.

170

0.59

4H

ess

US

023

420.

981

9.4

2580

.425

5.3

1188

21.

938

1.61

3H

usky

Can

ada

010

138.

951

3.6

1749

.010

6.8

3525

2.68

60.

641

KPC

Kuw

ait

147

983.

898

771.

758

474.

110

77.6

1942

62.

686

0.44

6L

ukoi

lR

ussi

a0.

0247

998.

713

974.

421

893.

310

91.6

1470

560.

321

0.81

0M

aers

kD

enm

ark

060

05.2

728.

615

72.2

035

001.

716

0

(con

tinu

ed)



Tab

le1:

Mea

nsof

Key

Var

iabl

es(c

onti

nued

)

Fir

mH

eadq

uart

erC

ount

ry

Gov

ernm

ent

owne

rshi

psh

are

Rev

enue

($m

illio

n)

Oil

rese

rves

(mill

ion

barr

els)

Nat

.gas

rese

rves

(bill

ion

ft3 )

Refi

ning

Cap

acit

y(t

hous

ands

barr

els/

day)

Em

ploy

ees

Rev

enue

per

empl

oyee

Ver

tica

lin

tegr

atio

n(p

rodu

cts/

oil

prod

)

Mar

atho

nU

S0

4884

2.1

792.

132

30.3

991.

227

395

1.76

56.

588

Mol

Hun

gary

0.11

1158

5.6

119.

794

3.1

303.

817

279

0.69

26.

307

Mur

phyO

ilU

S0

1249

0.6

330.

951

9.7

211.

862

371.

824

3.78

3N

IOC

Iran

143

765.

012

7229

.095

0214

.015

17.9

1149

560.

381

0.42

6N

exen

Can

ada

040

96.4

682.

059

7.9

033

711.

165

0N

ippo

nJa

pan

050

212.

711

0.9

1334

.012

52.6

1372

93.

684

19.7

51N

oble

US

021

71.0

258.

125

28.7

010

432.

064

0O

MV

Aus

tria

0.36

1897

2.7

555.

423

59.3

458.

230

627

0.83

93.

217

ON

GC

Indi

a0.

7711

553.

937

68.9

1569

2.6

155.

935

893

0.33

20.

329

Occ

iden

tal

US

014

626.

422

54.0

3839

.60

8863

1.62

40

PDO

Om

an0.

697

93.3

3194

.418

958.

165

.044

462.

167

0.10

7PD

VV

enez

uela

176

195.

783

753.

915

8427

.030

73.9

5339

11.

506

0.96

8PT

TT

haila

nd0.

7928

706.

114

4.1

4795

.131

5.9

7530

3.46

38.

558

Pem

exM

exic

o1

8218

9.8

1434

1.1

1412

6.8

1705

.313

9787

0.58

50.

467

Pert

amin

aIn

done

sia

129

054.

227

65.0

4675

3.2

1035

.328

578

1.03

75.

168

Petr

obra

sB

razi

l0.

4263

511.

694

10.8

1121

9.8

2105

.757

381

1.03

81.

434

Petr

oecu

ador

Ecu

ador

0.89

4609

.030

63.7

611.

017

5.9

4011

1.15

00.

715

Petr

onas

Mal

aysi

a1

4475

2.2

7036

.210

4663

.038

6.3

3087

41.

383

0.85

6/ Pi

onee

rU

S0

1475

.742

2.1

2756

.40

1467

1.00

00

Plai

nsU

S0

1630

.036

7.0

1156

.30

816

1.95

90

QP

Qat

ar1

2194

6.7

1270

3.1

6031

47.0

177.

869

763.

140

0.14

8R

epso

lY

PFSp

ain

059

294.

714

15.7

1268

6.6

1226

.835

208

1.66

52.

412

Ros

neft

Rus

sia

0.83

2639

8.1

1109

1.3

5132

6.7

493.

888

611

0.25

30.

645

Sant

osA

ustr

alia

017

41.8

147.

425

22.3

017

430.

986

0

(con

tinu

ed)



Tab

le1:

Mea

nsof

Key

Var

iabl

es(c

onti

nued

)

Fir

mH

eadq

uart

erC

ount

ry

Gov

ernm

ent

owne

rshi

psh

are

Rev

enue

($m

illio

n)

Oil

rese

rves

(mill

ion

barr

els)

Nat

.gas

rese

rves

(bill

ion

ft3 )

Refi

ning

Cap

acit

y(t

hous

ands

barr

els/

day)

Em

ploy

ees

Rev

enue

per

empl

oyee

Ver

tica

lin

tegr

atio

n(p

rodu

cts/

oil

prod

)

Saud

iAra

mco

Saud

iAra

bia

117

5363

.026

3055

.024

6052

.023

09.7

5347

53.

282

0.30

1Sh

ell

Net

herl

ands

027

7173

.064

24.2

4143

4.7

3993

.010

7828

2.59

13.

315

Sino

pec

Chi

na0.

6811

3067

.031

47.9

4286

.133

16.0

3875

850.

309

2.59

4So

nang

olA

ngol

a1

7810

.329

78.3

1158

.434

.156

111.

409

0.08

2So

natr

ach

Alg

eria

143

917.

811

003.

615

1148

.047

1.1

4592

20.

996

0.42

9St

atoi

lH

ydro

Nor

way

0.62

7773

6.0

2636

.820

784.

334

0.3

4943

91.

852

0.49

7Su

ncor

Can

ada

0.02

2428

2.9

2075

.124

56.7

407.

110

274

2.22

81.

159

Surg

utne

fteg

asR

ussi

a0

1571

2.8

7289

.613

584.

335

8.6

9534

60.

160

0.28

7T

NK

-BP

Rus

sia

025

509.

955

24.2

6002

.941

5.7

8328

50.

380

0.35

8Ta

lism

anC

anad

a0

5361

.158

9.8

3970

.70

2178

2.38

90

Tota

lFr

ance

014

8744

.065

27.0

2412

3.4

2648

.010

7115

1.43

52.

221

Win

ters

hall

Ger

man

y0

1081

9.2

484.

633

29.8

9.4

1736

5.95

50.

116

Woo

dsid

eA

ustr

alia

023

07.7

224.

338

17.0

026

980.

824

0X

TO

US

039

06.1

222.

166

94.6

018

241.

850

0



17. Although it is not relevant for our study, the concept also is defined for multiple outputs aswell as multiple inputs.

18. If we assume the technology has constant returns to scale then the composite firm can becompared with a re-scaled version of firm i and the k weights need not sum to one. If we cannotassume constant returns to scale, the composite firm needs to match a full-sized copy of firm i andthe weights k need to sum to 1.

5. ANALYSIS

5.1 Data Envelopment Analysis

We used the program DEAP 2.1 written by Tim Coelli to calculate theDEA measures. We assumed that the technology displays constant returns to scale.Variable returns to scale would allow for wider variations in technologies withmore firms being measured as efficient, but it also renders firms more difficult tocompare.

Suppose we have N firms each using K inputs to produce a single “out-put”, which in our case is revenue.17 Define X as the matrix of inputs, YK� Nas the vector of revenues from all the firms, and let xi and yi denote the1� Ninputs and revenue, respectively, of firm i. The constant returns to scale output-oriented distance function D(yi, xi) of firm i is then calculated by solving the linearprogramming problem:

–1[D(y ,x )] ≡maxh (4)i ih,k

subject to

–hy + Yk≥0, x –Xk≥0 and k≥0i i

where 0�D(yi, xi)≤1 is the technical efficiency score and k is an vectorN�1of constants. This problem is solved for each firm i and thus N times each year.

We briefly explain the idea behind this linear programming problem.Imagine forming a weighted average of existing firms, referred to below as a“composite firm,” to be compared with firm i with kj being the weight of firm jin this composite.18 The composite firm uses inputs Xk at most equal to the inputsxi used by firm i while producing Yk at a minimum equal yi of firm i. A firm ithat is efficient would have h = 1 and all components of k except for the i-th oneequal to zero while ki = 1. If a value h�1 can be found for firm i, however, thatfirm is not efficient and h measures how much firm i has to increase its revenueto become efficient. In addition, if the composite firm would use strictly less ofsome inputs than firm i (that is, xi–Xk≥0 is non-binding for some inputs), firmi also is inefficient in the sense that it could produce the same revenue using lessof these particular inputs.

Figure 1 graphs the DEA efficiency scores for each firm and year in oursample. The firms have been grouped into three broad categories—national oil



Fig

ure

1:D

EA

Effi

cien

cySc

ores



Table 2: Average DEA Scores by Ownership (standard deviations inparentheses)

Year SOC NOC pNOC

pNOC meangovernment

share

2001 0.729 (0.289) 0.518 (0.271) 0.610 (0.276) 0.4942002 0.804 (0.251) 0.553 (0.269) 0.672 (0.306) 0.4922003 0.776 (0.253) 0.514 (0.241) 0.672 (0.299) 0.5342004 0.785 (0.265) 0.544 (0.294) 0.610 (0.329) 0.5582005 0.765 (0.282) 0.614 (0.264) 0.602 (0.285) 0.5822006 0.764 (0.270) 0.663 (0.272) 0.645 (0.282) 0.6032007 0.777 (0.266) 0.677 (0.277) 0.716 (0.281) 0.5852008 0.797 (0.258) 0.611 (0.225) 0.742 (0.279) 0.5852009 0.746 (0.274) 0.616 (0.291) 0.749 (0.275) 0.589

Average 0.771 (0.265) 0.586 (0.262) 0.668 (0.285) 0.557

companies (NOCs), shareholder-owned oil companies (SOCs) and part national-owned oil companies (pNOCs). The latter group includes two sub-groups—fourfirms (PTT, Rosneft, Ecopetrol and CNPC) that were fully government-ownedfor some years and four (Lukoil, Suncor, Mol and Petroecuador) that were fullyshareholder-owned for some years. Within each group or sub-group, firms havebeen ordered according to their average DEA score over the full nine years.

The movements in efficiency scores in Figure 1 from one year to thenext for each firm reveal that most firms became more efficient over the nine-year sample period. There are, however, some notable exceptions. Among privatefirms, Occidental, Chesapeake, BG, EOG, CNR, Devon, Talisman, Noble andPlains (this is around one-third of the SOCs) have lower relative efficiency scoresin 2009 than in earlier years.

Among the NOCs, only PDV from Venezuela has a lower score in 2009than in any earlier year. After beginning at slightly above 0.5 in 2001, PDV’sscore rose to a maximum of slightly above 0.6 in 2004 before falling back tobelow 0.3 in 2009. Partially privatized PDO (from Oman) also displays a similarrise and fall pattern, starting at around 0.3 in 2001, rising to a maximum efficiencyscore of almost 0.8 in 2007 and then falling back to around 0.44 by 2009.

The two standout NOCs are Saudi Aramco and Sonangol. Saudi Aramcois on the frontier from 2005. Sonangol is on the frontier from 2004 through 2007but falls back to a DEA score of 0.885 in 2008, rising again to 0.977 in 2009.Qatar Petroleum almost attains the frontier in 2009 with a score of 0.999, isslightly further below in 2007 with a score of 0.992, and has a higher efficiencyscore than Saudi Aramco in 2001 and 2002. Kuwait Petroleum Corporation attainsthe frontier in just one year (2007), and almost scores 0.8 in 2008, but otherwisehas efficiency scores between 0.4 and 0.67.

Table 2 summarizes average DEA efficiency scores for each group offirms for each of the years 2001–2009. We placed pNOCs in the NOC group in



19. The DEA score is truncated at 1 for 167 out of 549 observations.20. In a model that includes the amount of government ownership and its interaction with year

along with categorical variables for GovShare = 1 (NOC) and 0�GovShare�1 (pNOC) and their

any years they were fully government owned and in the SOC group in any yearsthey were fully privatized. The remaining pNOCs in each year therefore hadgovernment ownership shares strictly between 0 and 1. We have also includedthe average government ownership share in these firms for each year and theaverage across all years.

In most years, the average DEA efficiency score for pNOCs lies betweenthat of the NOCs and the SOCs. The exceptions are 2005 and 2006, when theNOCs had a higher average score than the pNOCs and 2009 when the pNOCshad a slightly higher average score than the SOCs. In the latter case, the averagescore for SOCs in 2009 was lower than any year since 2001, while that of thepNOCs was higher than in any other year. Looking at the trends in the averagesover the nine years, one cannot reject the hypothesis (p-value = 0.89) that theSOC average oscillates around a constant value of 0.769, while the averages forthe NOCs rise at about 0.017 per year (p-value = 0.016) and of the pNOCs atabout 0.015 per year (p-value = 0.031). In addition, the NOC and pNOC averagesoscillate more about their trend from year to year. Part of the explanation for gainin NOC and pNOC efficiency relative to SOCs is that more of the SOCs are onthe frontier, and hence have DEA scores at the upper bound of 1, every year.

We can obtain more systematic evidence on the relationship betweengovernment ownership and the DEA efficiency score by regressing the annualscores against government ownership share. In doing so, however, we need toaccount for the fact that the DEA efficiency score is, by definition, bounded aboveby 1.19 In addition, unmeasured firm-specific effects, such as geological or marketconditions, may be constant or nearly constant over the sample period. We there-fore estimate a Tobit regression model, which assumes that the dependent variableDit satisfies

D = x β + v + e (5)it it i it

for panels and periods with observedi = 1,2, . . . n t = 1,2, . . . ,T v + e = D –x βi it it it

if yit�1 and otherwise. The random firm-specific effects vi arev + e ≥1–x βi it it

assumed to be independently identically distributed (i.i.d.), independent of thei.i.d. error terms and independent of the regressors xit.eit

We examined several models. In the basic specification, we allowed boththe intercept and time trend to depend on the actual government ownership share(GovShare). We also allowed the intercept and trend to take just three differentvalues, one when GovShare = 1 (NOCs), one when GovShare = 0 (SOCs) and athird when GovShare is strictly between 0 and 1 (pNOCs). The best modelgrouped NOCs and pNOCs together with a common trend in DEA efficiencyscore and did not yield a trend in the DEA score for SOCs that was significantlydifferent from zero (here and in all subsequent regression equations, standarderrors are reported in parentheses below the estimated coefficients):20



interactions with year, the coefficients on GovShare, year and GovShare*year are individually notstatistically significantly different from zero and a test for the joint significance has a p-value of0.7904. Furthermore, after dropping these variables, a test for equality of the coefficients on NOCand pNOC had a p-value of 0.4466, while a test for equality of the coefficients on the interactionsNOC*year and pNOC*year had a p-value of 0.8963. Finally, the log likelihood for the estimatedmodel (6) is –28.6599 compared to –31.9477 for a model that includes GovShare and GovShare*yearin place of the regressors in (6).

21. Retail gasoline and diesel fuel prices were obtained from the Metschies surveys of internationalfuel prices. Since the Metschies data is biennial, we used the average of the two percentage subsidiesfrom the two adjacent years to proxy the percentage subsidy in the missing years. Eller et al (2011)used the same data but coded the subsidy variable 0–1 to indicate whether the country had an averagegasoline and diesel price below the corresponding average of the two prices in the US.

22. Estimating a model that includes GovShare, year and their interactions along with NOC andpNOC and their interactions with year, the coefficients on GovShare, year and GovShare*year are

DEA + +RevEff = 0.8517–0.2683GovSh + 0.0221GovSh *year (6)it it it t(0.0499) (0.0571) (0.0045)

where (which equals NOC + pNOC) takes the value 1 if GovShare�0+GovShit

and zero otherwise and year takes the values 1–9 for 2001–2009. Each of thecoefficients in (6) is significantly different from zero at a better than 0.001 level.Equation (6) implies that any amount of government ownership share decreasesthe average DEA score by more than 30% in 2001 (0.5833 compared to 0.8517).However, such firms also gain an average of 0.022 per year in DEA score so thatby 2009, they have an average DEA score that is only about 8% below the averagescore for SOCs. The estimated values of σv = 0.3307 (standard error 0.0350) andσe = 0.1648 (standard error 0.0065) indicate that more than 80% of the estimatedvariance is due to the firm-specific error components.

Systematic factors apart from the government ownership share couldalso affect the efficiency score. If mergers increase efficiency we would expectthe variable Premerge (defined in section 4 above) to negatively affect the DEAscore. We would also expect retail fuel subsidies to increase measured revenueinefficiencies. We tested for this by including a variable RetSubs equal to zerofor countries with average retail prices of gasoline and diesel at least equal thoseof the US while, for countries with average retail prices below those in the US,RetSubst equals the percent deviation below the US average in year t.21 Finally,firms with large retail operations may not only have refinery capacity, which isin the included inputs, but also substantial unmeasured capital invested in down-stream operations. Such firms might then artificially appear more efficient. Wetherefore also defined a variable VertInt equal to the ratio of product sales (inthousands of barrels per day) to annual liquids production (also in thousands ofbarrels per day) to measure the extent to which the firm is involved in downstreammarkets.

The best model again grouped NOCs and pNOCs with a common trendin DEA efficiency score and did not yield a significant trend in the DEA scorefor SOCs:22



individually and jointly insignificantly different from zero (p-value 0.8845). Furthermore, after drop-ping these variables, a test for equality of the coefficients on NOC and pNOC had a p-value of 0.8487,while a test for equality of the coefficients on the interactions NOC*year and pNOC*year had a p-value of 0.6031. Finally, the log likelihood for the estimated model (7) is –18.4300 compared to–21.0310 for a model that includes GovShare and GovShare*year as regressors in place of the firsttwo regressors in (7).

23. As we note below, however, other variables are unlikely to be the same for SOCs, NOCs andpNOCs.

DEA + +RevEff = 0.8640–0.2093GovSh + 0.0169GovSh *year (7)it it it t(0.0511) (0.0590) (0.0046)

–0.3224RetSubs –0.1351Premerge + 0.0157VertIntit it it(0.1246) (0.0455) (0.0112)

The coefficients on and in (7) are significantly different+ +GovSh GovSh *yearit it t

from zero at a better than 1% level, and have similar magnitudes to the corre-sponding coefficients in (6), suggesting that the strong negative effects of gov-ernment ownership on efficiency cannot be explained by the other factors on theright-hand side of (7). The estimated values imply that any non-zero governmentownership share decreases the average DEA score by more than 24% below theaverage score of 0.8640 for SOCs in 2001 (holding other variables fixed), but by2009 such firms have an average DEA score that is only about 6.6% below theaverage for corresponding SOCs.23

The coefficient on RetSubs in (7), which is significantly different fromzero at the 1% level, also implies that operating in a country where retail pricesare subsidized, which overwhelmingly applies to NOCs and pNOCs, also sub-stantially reduces estimated revenue efficiency. The mean retail price in thesecountries is about 40% below US retail price. On average, therefore, such sub-sidies reduce the estimated revenue efficiency for firms headquartered in subsi-dizing nations by almost 15% below the average shareholder firm operating inan environment without such subsidies.

The coefficient on Premerge in (7) also is significantly different fromzero at a better than 1% level. It implies firms that undertake mergers have a jointDEA efficiency score in the years before they merge that is more than 15% belowthe average score of 0.8640 for SOCs. Thus, mergers tend to be efficiency im-proving.

Finally, at the mean of strictly positive values of VertInt of 1.74, theeffect on the estimated DEA score is about 0.027. However, the coefficient is notsignificantly different from zero at the 10% level. Perhaps including refiningcapacity in the inputs substantially adjusts for the effects of relatively high par-ticipation in downstream markets.

Apart from the efficiency scores h the DEA linear program also producesa matrix of coefficients k. Recall that the non-zero values of kijt give the weightson efficient firms j in year t that would result in a composite efficient firm usingno more inputs while producing at least as much revenue as inefficient firm i.



Fig

ure

2:C

ontr

ibut

ions

ofE

ffici

ent

Fir

ms

toD

omin

atin

gC

ompo

site

Fir

ms



Figure 2 summarizes the extent to which firms on the efficiency frontiercontribute to composite firms that dominate inefficient firms in the same year.The values graphed are for each efficient firm j and where i indexes thek∑ ijti

inefficient firms in year t. The sums therefore reflect not only how often eachefficient firm appears in a dominating composite firm but also its contribution tosuch firms.

In most years, the German firm Wintershall makes a major contributionto the dominating composite firms on the efficient frontier. It appears in manydominating composite firms and frequently has the largest k weight, making itthe most important firm in the composite. Evidently, Wintershall has an inputcomposition mix similar to many inefficient firms in the sample. It may thereforebe the most representative model for those inefficient firms to emulate. Otherfirms featuring in dominating composite firms include StatoilHydro, EnCana, BP,BHPBilliton and Marathon.

Some firms contributing to dominating composite firms in some yearsdo not do so in other years because they are not on the frontier in those years(CNOOC is an example). In other cases, such as ExxonMobil, the firm is on thefrontier in every year but does not play prominent role contributing to dominatingcomposite firms perhaps because its input mix differs too substantially from thatof the inefficient firms.

5.2 Malmquist Index Measures of Productivity Change

Instead of comparing firm i with other firms operating in the same yearone could ask how efficient firm i would have been had it used xit– 1 to produceyit – 1 while the comparison firms used year t technologies in year t–1. Modifyingnotation slightly, let Dt(yit – 1, xit – 1) denote the latter quantity and Dt– 1(yit –1, xit – 1)the original DEA measure in year t–1. The ratio of these distance functions

D (y ,x )t it it1M = (8)it D (y ,x )t it–1 it–1

measures productivity growth in firm i from year t–1 to year t viewed from theperspective of the set of technologies available in year t.

We could also measure the productivity gain from year t–1 to year tfrom the perspective of the technologies available in year t–1. Thus, the DEAmeasure of firm i efficiency if it had generated revenue yit from inputs xit whilethe comparison firms only had access to year t–1 technologies would be writtenDt –1(yit, xit). The productivity gain from year t–1 to year t then also could bemeasured as:

D (y ,x )t–1 it it2M = . (9)it D (y ,x )t–1 it–1 it–1

The Malmquist index is then defined as the geometric mean of these two mea-sures:



24. Firms on the frontier in one of the two successive years will also generally influence the shapeof the frontier in that year, but will have no effect on the shape of the frontier in the year they areoff it.

25. Since we are measuring revenue efficiency, higher prices for refined products, for example,would result in measured “technological progress” even if “technology” narrowly interpreted is un-changed.

26. The “shape” of the convex hull that describes the production frontier will then change overtime.

1/2D (y ,x ) D (y ,x )t it it t–1 it itM = . (10)� �D (y ,x ) D (y ,x )t it–1 it–1 t–1 it–1 it–1

Equation (10) can also be written as the product of two components:

1/2D (y ,x ) D (y ,x ) D (y ,x )t it it t–1 it–1 it–1 t–1 it itM = . (11)� �D (y ,x ) D (y ,x ) D (y ,x )t–1 it–1 it–1 t it–1 it–1 t it it

The first ratio (outside the square brackets) is the so-called efficiency change, andmeasures movements towards the frontier from year t–1 to year t by firm i. Thiscan be obtained from the annual DEA measures graphed in Figure 1 above.

The first ratio inside the square brackets in (11) measures the propor-tional change in the efficient frontier at the data observed for firm i in period t–1,while the second ratio measures the change in the frontier at the data observedfor firm i in period t. The geometric average of these two ratios, called a measureof technical change, thus measures the change in frontier technology between thetwo periods for the parts of the frontier relevant for firm i. Figure 3 graphs thetechnical change measures for each firm and pair of successive years.

In interpreting Figure 3, it is useful to focus first on the firms that areon the frontier every year (Wintershall, Marathon, ExxonMobil, BP, BHPBilliton,StatoilHydro, Sinopec, PTT). They will play the major role in shifting the frontierfrom t–1 to t.24 Observe that the technical change measure can be written as:

1/2D (y ,x )/D (y ,x )t–1 it it t it–1 it–1 . (12)� �D (y ,x )/D (y ,x )t it it t–1 it–1 it–1

For firms on the frontier in both periods, the ratio in the denominator in (12) willbe 1 and the deviation of the technical change measure from 1 will reflect theratio in the numerator of (12). If the industry is undergoing positive technicalprogress,25 a frontier firm i in year t–1 allowed to use xit to produce yit whenother firms are using year t–1 technologies ought to remain on the frontier, thatis, we should find Dt –1(yit,xit) = 1. Even if firm i is also on the frontier in year t,however, it might not be able to stay there using xit– 1 to produce yit – 1 when otherfirms are using year t technologies and we could find Dt(yit– 1,xit –1)�1.26 Thus,the ratio in the numerator in (12) could exceed 1.



Figure 3: Technical Change Measures Relevant for Each Firm inSuccessive Years



27. Since Wintershall and BHPBilliton are often in dominating composite frontier firms, theymust have input combinations comparable to many other firms. Thus, when their inputs and outputare lagged one year it also is easier to find a composite of other firms that can dominate them.

28. The ratio of distance measures for firm i in t–1 and t equals the ratio of distance to the frontierat t to the distance to the frontier at t–1 along a ray determined by firm i input proportions. Thus, atechnical change measure exceeding 1 will indicate an expanding frontier at the ray relevant for firmi, implying that if the DEA scores remained the same in t–1 and t then firm i actually became moreproductive and thus M�1.

29. This is separate from the results found above that the DEA scores also increased on average,implying that firms got closer to the (shifting) frontier, over the decade.

30. The weaker performance in 2001–02 is evident in Figure 3 where 42 of the technical changescores for 2001–02 are below 1 and those above 1 average only 1.103.

Among the firms that remain on the frontier every year, ExxonMobilstands out as having technical change measures very close to 1 in all years. Thus,ExxonMobil would remain very close to the frontier even if its data were laggedone year relative to the other firms. The technical change measures for BP areonly slightly more dispersed than those for ExxonMobil. Among the five SOCfirms on the frontier every year, Wintershall and BHPBilliton have the largestdispersions in technical change measures.27 Marathon has an intermediate dis-persion in technical change measures.

The remaining three firms that are on the frontier every year (Statoil-Hydro, Sinopec and PTT) are all pNOCs. The dispersions in StatoilHydro andPTT technical change measures are quite similar to those of Marathon. Exceptfor 2006–07, the technical change measures for Sinopec are not much more dis-persed than those for ExxonMobil, but their average is substantially above 1. Theimplication is that, although Sinopec was on the frontier each year it had to makechanges to remain there. If it had not changed its inputs and revenue it wouldhave fallen below the frontier.

For firms that are not on the frontier in both t–1 and t, it makes moresense to view the technical change measures in Figure 3 using the original ex-pression in (11), which gives the technical change measure as the product of tworatios. The first ratio measures Dt –1 relative to Dt at (yit –1,xit – 1) while the secondmeasures Dt –1 relative to Dt at (yit,xit). If either of these ratios exceeds 1, thefrontier at t shifts out relative to the frontier t–1 at the observed input and outputproportions of firm i in either t–1 or t.28

The average of all the technical change measures in Figure 3 across bothyears and firms is 1.129, implying that on average the frontier expanded over thedecade.29 The average technical change measure across firms also exceeded 1 forall pairs of years except from 2001–02, when the average was 0.928.30 The fron-tier shifted in for almost as many firms in 2008–09, but was above 1.3 for enoughfirms that the average remained above 1 at 1.069. The year with the third largestnumber of technical change scores below 1 was 2006–07, when the average wasjust 1.017. The year with highest average technical change measure was 2002–03, when the average was 1.323. In 2004–05, all technical progress measures



31. The change in merger variable came closest to being so, with a coefficient of 0.1017 and a p-value of 0.129. This variable would be zero except for the year when a merger occurred when itwould take the value –1. A positive sign of the coefficient would imply that mergers tend to take the“synthetic” pre-merged firms to a part of the production space where the frontier is closer.

32. The coefficient on Retsubs has a p-value of 0.014, while that on Premerge has a p-value of0.004. When we also included changes in GovSh, and VertInt as regressors, the coefficients on Retsubsand PreMerge remained significant at the 3% and 1% levels respectively, while neither of the othervariables was significant at even the 30% level. Equation (13) was estimated as a panel regressionwith random effects. A Hausman test for random versus fixed effects resulted in a chi-square statisticwith two degrees of freedom of 2.58 with a p-value of 0.275.

33. Thus, DGovSh + would only be non-zero in the very small number of instances where gov-ernment ownership was eliminated entirely.

exceeded 1, but there were fewer large scores than in 2002–03 and the averagewas 1.237. In 2003–04, 2005–06, and 2007–08 the averages were 1.164, 1.130and 1.165 respectively.

Figure 3 also shows that the technical change measures for SOCs tendto vary much more than the corresponding measures for pNOCs and especiallyNOCs. Across all years, the average technical change measure for SOCs is 1.141compared to 1.123 for NOCs and 1.111 for pNOCs, but the standard deviations(0.181 for SOCs, 0.151 for pNOCs and only 0.113 for NOCs) are more different.With the exceptions of CNOOC and PDO in 2008–09, all the large technicalchange expansions, and the majority of the frontier contractions, occur for SOCs.

We also examined some panel regression relationships for the technicalchange and overall Malmquist productivity change measures. Since these mea-sures relate to changes from one year to the next we checked for systematicrelationships with changes in the ownership, retail subsidy, merger and verticalintegration variables.

For the technical change measure, none of the proposed explanatoryvariables was statistically significantly different from zero at the 10% level.31 Apanel regression for the Malmquist total productivity change measures yielded:32

Malmq = 1.163–0.9929DRetsubs –0.3677DPremerge (13)it it it(0.018) (0.4053) (0.1268)

where D is the time difference operator. The coefficients in (13) imply that merg-ers and a reduction in retail fuel subsidies both raise total productivity. This isconsistent with the results for the annual DEA measures in (7) that mergers raiseDEA scores while higher subsidies lower them. The fact that DVertInt and noneof the DGovSh measures proved significant (and were thus dropped from (13)) isalso consistent with the earlier results. In particular, note that although GovSh +

was significant in the DEA regression, a partial privatization would changeGovSh = 1 to 1�GovSh�0, so the value of GovSh + would not change. Of course,DGovSh + would also be zero for firms that stayed as NOCs or SOCs.33 The lackof significance of DGovSh in (13) therefore supports the conclusion from theDEA measures that partial government ownership has the same effect on effi-ciency as full government ownership.



34. We primarily used the program FRONT 4.1 also written by Tim Coelli for the stochasticfrontier analysis.

35. In the analysis presented below, we assume that vi is normally distributed and ui is a truncatednormal distribution multiplied by a specific function of time.

36. Schmidt and Sickles (1984) also proposed using one-sided fixed-effects and random-effectsto measure time-invariant producer-specific technical efficiency. See Kumbhakar and Lovell (2000)for a thorough survey of panel stochastic frontier analysis.

37. The time-varying specifications we shall examine were first proposed by Cornwell, Schmidtand Sickles (1990), and Battese and Coelli (1992, 1995).

5.3 Stochastic Frontier Analysis

The second approach that we use to measure relative revenue efficiencyis a parametric stochastic frontier analysis (SFA).34 Suppose we can use a pro-duction function qi = f (x1i, . . . ,xKi), i = 1,2, . . . ,N for a cross-section of N firms eachwith outputs qi and K inputs. Suppose the production technology can be repre-sented as Cobb-Douglas, and noting that the logarithm of revenue can be ap-proximated as the share-weighted sum of the logarithm of revenue from eachoutput, taking natural logarithms yields the linearized form

K

lny = β + β lnx + v –u (14)∑i 0 k ik i ik = 1

where yi is revenue, vi is a stochastic component reflecting measurement error oromitted variables (such as geological conditions in our example) and ui capturesthe nonnegative technical efficiency component. Define ei as the composed errorsuch that ei = vi–ui. Once distributions have been proposed for both vi and ui, onecan estimate (14) using maximum likelihood.35 Firm-specific technical efficiencycan then be calculated as the expected value of conditional on ei. Sinceie–uui ≥0, technical efficiency will be bounded between zero and one.

Pitt and Lee (1981) extended the maximum likelihood approach for es-timating firm-specific technical efficiency to panel data.36 The log-linear revenuefunction for panel data corresponding to (14) is

K

lny = β + β lnx + v –u (15)∑it 0 k ikt it itk = 1

for years t = 1, . . . ,T and where yit is the revenue of firm i in year t, and theefficiency component ui can depend in a parametric way on time or other mea-sured covariates, but for ui to be identified it cannot have an arbitrary time de-pendence.37 In the simplest specification, we assume:

(i) ,2v �N(0,σ )it v

(ii) with and N + is the truncated-nor-– (t– T) + 2gu = e u u �N (l,σ )it i i u

mal distribution,



38. Since some firms had no distillation capacity we actually used ln(1 + capacity), which is zerofor capacity equal to zero and barely changed from ln(capacity) when capacity is positive since theaverage of the strictly positive values is around 1000 (in thousands of barrels per day).

(iii) vit and ui are distributed independently of each other and the re-gressors.

This allows the efficiency of the panel of firms to change over time atthe uniform rate g per year, although the deviations ui of individual firms fromthe panel average do not change over time. Based on equations (1)–(3), the re-gressors xkit are oil and natural gas reserves, distillation capacity,38 employees,and oil and natural gas prices.

The estimated equation is then given as

lnRev = 4.3384 + 0.3180lnEmp + 0.1919lnOilRsv (16)it it it(0.5504) (0.0421) (0.0319)

–0.0001lnNGRsv + 0.0775lnRefcap + 0.4150lnOilpit it it(0.0272) (0.0162) (0.0618)

+ 0.4229lnNGp + v –uit it it(0.0550)

with estimated error structure parameters l = 1.506 (0.2657), g = 0.0291 (0.0047),(0.0888) and (0.0031).2 2σ = 0.4022 σ = 0.0484u v

Since each input should increase revenue, the coefficients in (16) shouldall be positive. The results are consistent with this expectation except for the caseof natural gas, whose coefficient is negative but not significantly different fromzero. The point estimate of g implies, as with the DEA estimates discussed above,that firms in the sample on average became more efficient at generating revenuefrom the inputs over the period 2001–2009. Finally, the much larger value of

relative to implies that more than 89% of the variation in the composite2 2σ σu v

error term is due to the one-sided systematic efficiency differences between thefirms rather than statistical noise or measurement error.

Although the DEA and SFA are very different methods, they give similarefficiency rankings for most of the firms. Estimating a panel Tobit regression withfirm-specific random effects we obtained

DEA SFARevEff = 0.4419 + 1.5124RevEff . (17)it it(0.0507) (0.1884)

The strong positive correlation between the measures for each firm isalso illustrated in Figure 4. This plots the average of the DEA scores over thenine years against the average from the stochastic frontier model.– uitE[e v –u ]⎪ it it

The shading of the data points reflects the extent of government ownership.



Figure 4: Average SFA versus Average DEA Efficiency Scores



39. The coefficients of GovSh + and GovSh + *year almost halve in value and the p-values increasefrom below 0.001 to 0.040 and 0.075 respectively. The magnitude of the coefficient of RetSubsdecreases by about 17% and its p-value increases from 0.010 to 0.017. The coefficient of Premergedecreases slightly and its p-value increases from 0.003 to 0.015. VertInt is not statistically significantin either model.

The five major SOCs (BP, Chevron, ConocoPhilips, ExxonMobil andShell) in the upper right corner of Figure 4 have high DEA and SFA scores.Conversely, the NOCs all tend to have low estimated SFA scores, although theirDEA scores are not nearly as concentrated among the low values.

An indication of which components of the DEA inefficiency measuresare also present in the SFA measures can be obtained by re-estimating (7) withthe SFA efficiency score as an additional variable. If a formerly significant vari-able is no longer significant when SFA scores are included, the SFA score mustaccount for the effects of that variable on the DEA inefficiency measure. Theresults indicated that the one-sided SFA efficiency scores capture a substantialfraction of the systematic effects of GovSh + , GovSh + *year, Premerge andRetSubs on the DEA efficiency scores.39

Some components of the DEA inefficiency scores may not be present inthe SFA scores because they could instead affect the estimated revenue function.This is particularly so for the retail subsidy variable, which was calculated as thepercentage discount on retail fuel prices below the corresponding US prices, andset to zero when prices were greater than or equal to the US prices. We wouldexpect (1–RetSubs) to multiply the oil price in producing revenue and thereforeln(1–RetSubs) should be included in the estimated revenue function rather thanthe error term.

Similarly, a major conclusion of the theoretical analysis of NOC behav-ior by Hartley and Medlock (2008) was that the relative inefficiency of NOCs islikely to be manifest in substantial over-employment. Equivalently, this shouldshow up as a reduced productivity of labor in NOCs. If this hypothesis is validwe should also expect to find an interaction term between GovSh and lnEmp inthe estimated revenue function (16).

We therefore estimated a revised SFA model that included Gov-Share*lnEmp and ln(1–RetSubs) in the estimated revenue function and allowedfor a structural model of the one-sided inefficiency component of the error termusing the specification of Battese and Coelli (1995). This specification allows themean of the firm-specific inefficiency measures uit to depend on firm-specificcovariates zlit, where l = 1, . . . ,L, such that

u = z d + w . (18)it it it

The random variable wit in (18) is a truncated normal distribution with zero meanand variance such that the point of truncation is –zitd that is, wit ≥ –zitd. As2σw

Battese and Coelli remark, these assumptions are consistent with uit being a non-



40. We also estimated models where the actual ownership share entered (20), or the indicator forpositive government ownership entered (19). These produced less significant values for the corre-sponding coefficients and a lower maximized log likelihood, but did not greatly affect remaining

negative truncation of the -distribution. To identify the model, Battese2N(z d,σ )it w

and Coelli assume that the non-efficiency related error component vit in (15) isiid and independently distributed from the uit. The latter are also assumed2N(0,σ )v

to be independently distributed from each other. Maximum likelihood estimatesof the parameters of the model were obtained using the program FRONTIER 4.1(Coelli, 1996b).

We allowed government ownership share, the pre-merger indicator vari-able and the vertical integration measure as potential z variables. In addition, thestrong statistical significance of the time trend coefficient g in the basic SFAmodel, and the significance of the time trend variables in the DEA panel Tobitregression models, suggested that we include time trend or time trend interactionsamong the set of potential z variables. The best estimated model was:

lnRev = 1.2441 + 0.3227lnEmp –0.0546GovSh *lnEmp (19)it it it it(0.2850) (0.0300) (0.0078)

+ 0.1506lnOilRsv + 0.2443lnNGRsvit it(0.0320) (0.0261)

+ 0.1226lnRefcap + 0.5281lnOilpit it(0.0134) (0.0811)

+ 0.3177ln(1–RetSubs ) + 0.3058lnNGp + v –uit it it it(0.0606) (0.1071)

with the auxiliary equation for the one-sided inefficiency error term given by:

+ +u = 0.5612 + 0.9471GovSh –0.1127GovSh *year (20)it it it t(0.2176) (0.2525) (0.0365)

–0.5555VertInt + wit it(0.1032)

In addition, the estimated overall variance of the error term was 0.66582 2σ + σv w

(with an estimated standard error of 0.0937), while the proportion due to the one-sided error was 0.9189 (0.0227).

All the coefficients in (19) except for the one on the natural gas priceare significantly different from zero at the 0.0001 level. The natural gas pricecoefficient has a p-value of 0.0043. In the model (20) for the one-sided ineffi-ciency error term, the coefficient on VertInt is significantly different from zero atthe 0.0001 level, the coefficient on GovSh + has a p-value of 0.0002, and thecoefficient on GovSh + *year, has a p-value of 0.0020.

Since the indicator for any amount of government ownership GovSh +

appears in (20),40 as with the DEA panel Tobit regression models examined pre-



coefficient estimates. We also estimated models with Premerge in (20) or interacting with ln Emp in(19). In neither case were the coefficients statistically significantly different from zero at even the20% level.

41. If the revenue function (19) represents a Cobb-Douglas production function times output pricesas hypothesized, we would expect the right hand side to be homogeneous of degree 1 in the nominalprices. Although the coefficients on lnOilp and lnNGP do not sum to 1, these variables measure theactual output prices with error, so the coefficient estimates may be biased toward zero.

42. At the mean of the variable of 1.1524, the effect is around 14% higher than the constant termin (20).

viously, any government ownership tends to reduce revenue efficiency. However,the extent of the reduced efficiency measured in (20) tends to decline over timeand is completely gone by 2009. Nevertheless, since the actual government own-ership share GovSh appears in the revenue frontier (19), the marginal product oflabor in a fully government owned firm is around 17% smaller than in a SOCand this effect does not decline over time. Consistent with the results of Wolf andPollitt (2008), the fact that actual government share enters the revenue function(19) implies that partial privatization reduces the gap in employment productivitybetween NOCs and SOCs.

The estimated revenue function (19) also implies that oil prices have alarger effect on revenue than natural gas prices,41 but natural gas reserves have alarger effect than oil reserves. The latter result may suggest that firms on averagetend to hold more excess oil than natural gas reserves. The efficiency differencesdue to retail price subsidies, like the reduced productivity of labor as a result ofgovernment ownership, remain throughout the sample period.

The positive coefficient on refining capacity in (19) implies that pro-ducing refined products enables the firm to generate more revenue. Even if thefirm does not have refining capacity, the negative, and quite large,42 coefficienton VertInt (the ratio of liquid product sales to liquids production) in (20) impliesthat firms with more downstream operations will make more revenue relative tothe inputs that have been measured in (19). This result could represent firm het-erogeneity rather than measured inefficiencies per se.

6. CONCLUSION

We examined the revenue efficiency of 61 oil and gas firms over theperiod 2001–09. The sample included national oil companies (NOCs), partiallyprivatized national oil companies (pNOCs) and shareholder owned oil companies(SOCs).

Despite using two very different methods, namely non-parametric dataenvelopment analysis (DEA) and the associated Malmquist measure of annualproductivity changes and parametric stochastic frontier analysis (SFA), the resultswere very similar. The two sets of relative inefficiency measures were highlycorrelated. Both methods produced strong evidence that the efficiency of thegroup of firms increased over the decade, but that NOCs and pNOCs tended to



improve at a faster rate. Nevertheless, a gap between the SOCs and the otherfirms was still present at the end of the decade. Both methods also showed thatretail subsidies were a major source of reduced revenue efficiency for many NOCsand pNOCs.

The SFA also revealed that government ownership tends to encourageover-employment, or equivalently a reduced productivity of labor. This effect wasreduced by partial privatization. On the other hand, the SFA analysis also revealedan additional residual negative impact of government ownership that was notaffected by partial privatization. This latter unidentified source of relative ineffi-ciency tended to disappear over the decade.

The DEA and Malmquist analyses found evidence that mergers tendedto raise the efficiency of the merging firms, but we did not find a similar effectin the SFA. The SFA, but not the DEA, analysis revealed that a larger ratio ofproduct sales to liquids production tends to increase revenue. This may be becausewe did not include any capital assets associated with retail operations among theinputs. Perhaps the DEA analysis did not reveal as strong an effect from thisvariable because refinery capacity proved a better proxy for downstream opera-tions in the non-parametric DEA than in the parametric SFA.

According to the Malmquist index analysis, the lowest technical changemeasures occurred in 2001–02 and the highest two years later in 2003–04. Gen-erally, the technical change measures for SOCs were more variable, perhaps sug-gesting that more innovation occurs among these firms. Such information is notevident from SFA.

The DEA and Malmquist analyses also provided some interesting detailson the relative efficiencies and efficiency changes of particular firms over thedecade. TNK’s efficiency jumped after the formation of TNK-BP in 2003. Afterbeing on the frontier in 2005–07, however, it fell in 2008 and 2009 after theinfluence of the Russian partner in the firm increased. The relative efficiency ofthe Venezuelan NOC also increased from 2001–04 before falling below its 2001value by 2009. The pNOC PDO from Oman also displays an increasing thendecreasing pattern of efficiency changes. The most efficient pNOCs, Statoil-Hy-dro, Sinopec and PTT, produced DEA scores similar to the most efficient privatefirms. Several others, especially ENI, CNOOC and Petrobras, also improved sub-stantially over the decade to end up on or close to the frontier in 2009. SaudiAramco and Sonangol were the standout NOCs.

Among the five large international SOCs, ExxonMobil stood out notonly by being on the frontier in every year but also by being on or close to thefrontier even if its inputs and revenue were lagged one year relative to all otherfirms. Of the five large SOCs, only BP plays a prominent role in the compositeefficient firms dominating the inefficient firms. In that sense, BP makes a morefeasible firm for the inefficient firms to emulate. The most common member ofsuch dominating composite peers, however, was Wintershall. Other firms fre-quently present in such composites included StatoilHydro, EnCana, BHPBillitonand Marathon. Among the few NOCs that appear on the frontier, only Saudi



Aramco has a noticeable presence, albeit a small one, in any dominating com-posite peer firms that inefficient firms should try to emulate.

Finally, the results support the notion that the political overseers of NOCstend to redistribute resource rents toward both domestic consumers and domesticemployees of the firm. Thus, as NOCs control more global oil and gas resourcesit is reasonable to expect that an increasing majority of oil and gas developmentswill be undertaken with political objectives in mind. This may result in less pro-duction and higher prices than would occur under commercial development. Theresults also support previous findings that even partial privatizations can increaseNOC efficiency, presumably by limiting the ability of politicians to interfere withoperating decisions and by requiring the firms to adhere to the accounting andother commercial practices of SOCs.

REFERENCES

Al-Obaidan, A.M. and G. W. Scully (1991). “Effificiency differences between private and state-ownedenterprises in the international petroleum industry,” Applied Economics, 23:237–246.

Banker, R.D. (1993). “Maximum Likelihood, Consistency and Data Envelopment Analysis: A Sta-tistical Foundation,” Management Science, 39(10): 1256–1273. http://dx.doi.org/10.1287/mnsc.39.10.1265.

Battese, G.E., and T.J. Coelli (1992). “Frontier Production Functions, Technical Efficiency and PanelData: With Application to Paddy Farms in India,” Journal of Productivity Analysis, 3(1–2): 153–169. http://dx.doi.org/10.1007/BF00158774.

Battese, G.E., and T.J. Coelli (1995). “A Model for Technical Inefficiency Effects in a StochasticFrontier Production Function for Panel Data,” Empirical Economics, 20:325–332. http://dx.doi.org/10.1007/BF01205442.

Coelli, T.J. (1996a). “A Guide to DEAP Version 2.1: A Data Envelopment Analysis (Computer)Program”, CEPA Working Paper 96/8, Department of Econometrics, University of New England,Armidale NSW Australia.

Coelli, T.J. (1996b), “A Guide to FRONTIER Version 4.1: A Computer Program for StochasticFrontier Production and Cost Function Estimation”, CEPA Working Paper 96/7, Department ofEconometrics, University of New England, Armidale NSW Australia.

Cornwell, C., P. Schmidt, and R.C. Sickles (1990). “Production Frontiers with Cross-Sectional andTime-Series Variation in Efficiency Levels,” Journal of Econometrics, 46(1–2): 185–200. http://dx.doi.org/10.1016/0304-4076(90)90054-W.

Cooper, W.W. and K. Tone (1997). “Measures of Inefficiency in Data Envelopment Analysis andStochastic Frontier Estimation,” European Journal of Operational Research, 99(1): 72–88. http://dx.doi.org/10.1016/S0377-2217(96)00384-0.

Desai, A., S.J. Ratick and A. P. Schinnar (2005). “Data envelopment analysis with stochastic variationsin data,” Socio-Economic Planning Sciences, 39: 147–164. http://dx.doi.org/10.1016/j.seps.2004.01.005.

Eller, Stacy L., P. R. Hartley and K. B. Medlock III (2011). “Empirical evidence on the operationalefficiency of National Oil Companies,” Empirical Economics, 40: 623–643. http://dx.doi.org/10.1007/s00181-010-0349-8.

Gong, B.H. and R.C. Sickles (1992). “Finite Sample Evidence on the Performance of StochasticFrontiers and Data Envelopment Analysis Using Panel Data,” Journal of Econometrics, 51(1–2):259–284. http://dx.doi.org/10.1016/0304-4076(92)90038-S.

Greene, W. (2005). “Fixed and Random Effects in Stochastic Frontier Models,” Journal of Produc-tivity Analysis, 23: 7–32.



Grosskopf, S. (1996). “Statistical inference and nonparametric efficiency: A selective survey,” Journalof Productivity Analysis, 7:161–76. http://dx.doi.org/10.1007/BF00157039.

Hartley, P.R. and K.B. Medlock III (2008). “A Model of the Operation and Development of a NationalOil Company,” Energy Economics, 30(5): 2459–2485. http://dx.doi.org/10.1016/j.eneco.2007.12.017.

Kumbhakar, S.C. and C.A.K. Lovell (2000). Stochastic Frontier Analysis. Cambridge: CambridgeUniversity Press. http://dx.doi.org/10.1017/CBO9781139174411.

Metschies, G.P. (2003–2009). International Fuel Prices. Available at http://www.metschies.com/Pitt, M. and L.F. Lee (1981). “The Measurement and Sources of Technical Inefficiency in the Indo-

nesian Weaving Industry,” Journal of Development Economics, 9:715–723. http://dx.doi.org/10.1016/0304-3878(81)90004-3.

Ruggiero, J. (2007). “A Comparison of DEA and Stochastic Frontier Model Using Panel Data,”International Transactions in Operational Research, 14(3): 259–266. http://dx.doi.org/10.1111/j.1475-3995.2007.00585.x.

Schmidt, K.M. (1996). “The Cost and Benefits of Privatization: an Incomplete Contracts Approach,”Journal of Law, Economics and Organization, 12(1): 1–24. http://dx.doi.org/10.1093/oxfordjour-nals.jleo.a023354.

Schmidt, P. and R.C. Sickles (1984). “Production Frontiers and Panel Data,” Journal of Business andEconomic Statistics, 2(4): 367–374.

Tsionas, E. G. (2003). “Combining DEA and stochastic frontier models: An empirical Bayes ap-proach,” European Journal of Operational Research, 147: 499–510. http://dx.doi.org/10.1016/S0377-2217(02)00248-5.

Victor, N. M. (2007). “On Measuring the Performance of National Oil Companies (NOCs),” WorkingPaper #64, PESD, Stanford University.

Wolf, C.O.H. (2009). “Does Ownership Matter? The Performance and Effificiency of State Oil vs.Private Oil (1987–2006),” Energy Policy, 37: 2642–2652.

Wolf, C.O.H. and M.G. Pollitt (2008). “Privatising National Oil Companies: Assessing the Impact onFirm Performance,” Cambridge Judge Business School Working Paper No. 02/2008, University ofCambridge.

Changes in the Operational Efﬁciency of National Oil …...9. The intercept term thus measures the...

Documents

Transcript of Changes in the Operational Efﬁciency of National Oil …...9. The intercept term thus measures the...