Вы находитесь на странице: 1из 142

Public Disclosure Authorized

FI
r .7
CL

a2
Public Disclosure Authorized

A~~~~A
A-~~~~~~~~A
q a

PIP-~~~~~~~~~~~~~~~~~l

> t - s.,i N 2t c: 3 - z~~~~'I~


Public Disclosure Authorized

ls s X f S n | z~~~~~~~~~~~

Wi.-C 0
Z-4~~~~~~~~~~~~~~~~~~~~~~~.
sclosure Authorized
RECENT WORLD BANK TECHNICAL PAPERS

No. 250 Rangeley,Thiam, Andersen, and Lyle, InternationalRiver BasinOrganizationsin Sub-SaharanAfrica


No. 251 Sharma, Rietbergen, Heimo, and Patel, A Strategyfor tileForestSectorin Sub-SaharanAfrica
No. 252 The World Bank/FAO/UNIDO/Industry Fertilizer Working Group, Worldand RegionalSupplyand Demand
BalancesforNitrogen,Phiosphate,
and Potashi,1992/93-1998/99
No. 253 Jensen and Malter, ProtectedAgriculture:A GlobalReview
No. 254 Frischtak, GovernanceCapacityand EconomicReformin DevelopingCountries
No. 255 Mohan, editor, Bibliographyof Publications:Technrical
Department,AfricaRegion,July 1987 to April 1994
No. 256 Campbell, Designand Operationof SmaliholderIrrigationin South Asia
No. 258 De Geyndt, Managingthe Quality of HealthCarein DevelopingCountries
No. 259 Chaudry, Reid, and Malik, editors, Civil ServiceReformin LatinAmericaand the Caribbean:Proceedingsof a Conference
No. 260 Humphrey, PaymentSystems:Principles,Practice,and Improvements
No. 261 Lynch, Provisionfor Childrenwith SpecialEducationalNeeds in the Asia Region
No. 262 Lee and Bobadilla, HealthiStatisticsfor the Americas
No. 263 Le Moigne, Subramanian, Xie,and Giltner, editors, A Guideto the Formulationof Water ResourcesStrategy
No. 264 Miller and Jones, Organicand Compost-Based GrowingMediafor TreeSeedlingNurseries
No. 265 Viswanath, Building Partnershipsfor PovertyReductioni:The ParticipatoryProjectPlanningApproachiof the Women's
EnterpriseManagementTrainingOutreachProgram(WEMTOP)
No. 266 Hill and Bender, Developingthe RegulatoryEnvironmentforCompetitiveAgriculturalMarkets
No. 267 Valdes and Schaeffer,Surveillanceof AgriculturalPricesand Trade:A Handbookforthe DominicanRepublic
No. 268 Vald6s and Schaeffer,SurveillanceofAgriculturalPricesand Trade:A HandbookforColombia
No. 269 Scheierling, OvercomingAgriculturalPollutionof Water:The Cliallengeof IntegratingAgriculturaland
EnvironmentalPoliciesin tiheEuropeanUnion
No. 270 Banerjee, Rehabilitationi
of DegradedForestsin Asia
No. 271 Ahmed, Technological Developmentand PollutionAbatement:A Study of How EnterprisesAre FindingAlternativesto
Chlorofluorocarbons
No. 272 Greaney and Kellaghan, Equity Issuesin PublicExaminationsin DevelopingCouintries
No. 273 Grimshaw and Helfer, editors, VetiverGrassfor Soil and WaterConservation,Land Rehabilitation,and Embankment
Stabilization:A Collectionof Papersand NewslettersCompiledby the VetiverNetwork
No. 274 Govindaraj, Murray, and Chellaraj, HealthExpenditiresin Latin America
No. 275 Heggie, Managemenitand Financinigof Roads:An Agendafor Refor,n
No. 276 Johnson, Quality ReviewSchemesforAuditors:Their PotentialforSub-SaharanAfrica
No. 277 Convery, Applying EnvironmentalEconomicsin Africa
No. 278 Wijetilleke and Karunaratne, Air QualityManagement:ConsiderationsforDevelopingCountries
No. 279 Anderson and Ahmed, The CaseforSolar EnergyInvestments
No. 280 Rowat, Malik, and Dakolias, JudicialReformin Latin Americaand the Caribbean: Proceedingsof a WorldBank Conference
No. 281 Shen and Contreras-Hermosilla, Environmentaland EconomicIssuesin Forestry:SelectedCaseStudies
in Asia
ControlProgram(OCP)
No. 282 Kirn and Benton, Cost-BenefitAnalysis of the Onchocerciasis
No. 283 Jacobsen, Scobie and Duncan, Statutory Interventionin AgriculturalMarketing:A New ZealandPerspective
No. 284 Valdes and Schaeffer in collaboration with Roldos and Chiara, Surveillanceof AgriculturalPriceand TradePolicies:A
HandbookforUruguay
No. 285 Brehrnand Castro, The Marketfor Water Rights in Chile:MajorIssues
No. 286 Tavoulareas and Charpentier, CleanCoalTechnologies for DevelopingCountries
No. 287 Gillham, Bell,Arin, Matthews, Rumeur, and Hearn, Cotton ProductionProspectsfor the Next Decade

(List continues on the inside back cover)


WORLD BANK TECHNICAL PAPER NO. 325

EnergySeries

Estimating Construction
Costs and Schedules
Experiencewith Power Generation
Projectsin Developing Countries

RobertW Bacon,JohnE. Besant-Jones,


andJamshidHeidarian
The WorldBank
Washington,
D.C.
Copyright © 1996
The Intemational Bank for Reconstruction
and Development/THE WORLD BANK
1818H Street, N.W.
Washington, D.C. 20433,U.S.A.

All rights reserved


Manufactured in the United States of America
First printing August 1996

Technical Papers are published to communicate the results of the Bank's work to the development com-
munity with the least possible delay. The typescript of this paper therefore has not been prepared in ac-
cordance with the procedures appropriate to formal printed texts, and the World Bank accepts no
responsibility for errors. Some sources cited in this paper may be informal documents that are not readily
available.
The findings, interpretations, and conclusions expressed in this paper are entirely those of the
author(s) and should not be attributed in any manner to the World Bank, to its affiliated organizations, or
to members of its Board of Executive Directors or the countries they represent. The World Bank does not
guarantee the accuracy of the data included in this publication and accepts no responsibility whatsoever
for any consequence of their use. The boundaries, colors, denominations, and other information shown on
any map in this volume do not imply on the part of the World Bank Group any judgment on the legal sta-
tus of any territory or the endorsement or acceptance of such boundaries.
The material in this publication is copyrighted. Requests for permission to reproduce portions of it
should be sent to the Office of the Publisher at the address shown in the copyright notice above. The
World Bank encourages dissemination of its work and will normally give permission promptly and, when
the reproduction is for noncommercial purposes, without asking a fee. Permission to copy portions for
classroom use is granted through the Copyright Clearance Center, Inc., Suite 910, 222 Rosewood Drive,
Danvers, Massachusetts 01923, U.S.A.
The complete backlist of publications from the World Bank is shown in the annual Index of Publica-
tions, which contains an alphabetical title list (with full ordering information) and indexes of subjects, au-
thors, and countries and regions. The latest edition is available free of charge from the Distribution Unit,
Office of the Publisher, The World Bank, 1818 H Street, N.W., Washington, D.C. 20433, U.S.A., or from
Publications, The World Bank, 66, avenue d'Iena, 75116 Paris, France.

ISSN: 0253-7494

Cover: Detail from WilliamGropper, "Construction of the Dam" (mural study, Department of the Interior, Wash-
ington, DC., 1937).
Used by permission of the National Museum of American Art, Smithsonian Institution, transfer from the U.S.De-
partment of the Interior, National Park Service.

Robert W. Bacon is a professor of economics at Lincoln College, Oxford University, England, and a consultant to
the Industry and Energy Department at the World Bank. John Besant-Jonesis a principal economist in the Power De-
velopment, Efficiency and Household Fuels Division of the Industry and Energy Department at the World Bank.
Jamshid Heidarian is a professor of economics at the University of the District of Columbia, Washington, D.C., and
consultant to the Industry and Energy Department at the World Bank.

Library of Congress Cataloging-in-Publication Data

Bacon, Robert, 1942-


Estimating construction costs and schedules: experience with
power generation projects in developing countries / Robert W. Bacon,
John E. Besant-Jones, and Jamshid Heidarian.
p. cm. - (World Bank technical paper, ISSN 0253-7494 ; no.
325) (Energy series)
Includes bibliographical references.
ISBN0-8213-3670-3
1. Electric power plants-Design and construction-Estimates-
Developing countries. 2. Electric power plants-Developing
countries-Costs-Statistics. 1. Besant-Jones,John, 1941-
11. Heidarian, Jamshid. III. Title. IV. Series: World
Bank technical paper. Energy series.
TK1193.D44B33 1996 96-22845
621.31'21-dc2O CIP
ENERGY SERIES

No. 240 Ahmed, RenewableEnergyTechnologies:


A Reviewof the Status of Costsof SelectedTechnologies
No. 242 Barnes, Openshaw, Smith, and van der Plas, What MakesPeopleCookwith ImprovedBiomass
Stoves?
No. 243 Menke and Fazzari, ImprovingElectricPowerUtility Efficiency:Issuesand Recommendations
No. 244 Liebenthal, Mathur, and Wade, SolarEnergy:Lessonsfromthe PacificIslandExperience
No. 271 Ahrned, Technological Developmentand PoliticalAbatement:A Study of How Enterprisesare Finding
Alternatives to Chlorofluorocarbons
No. 278 Wijetillekeand Karunaratne, Air QualityManagement:ConsiderationsforDevelopingCountries
No. 279 Anderson and Ahmed, The CaseforSolarEnergyInvestments
No. 286 Tavoulareas and Charpentier, CleanCoalTechnologies
in DevelopingCountries
No. 296 Stassen, Small-ScaleBiomassGasifiersforHeatand Power:A GlobalReview
No. 304 Foley,PhotovoltaicApplicationsin Rural Areas of the DevelopingWorld
No. 308 Adamson, Bates, Laslett, and Pototschnig, EnergyUse,Air Pollution,and EnvironmentalPolicyin
Krakow:Can EconomicIncentivesReallyHelp?
Contents
Foreword............................................................... xi

Abstract ............................................................... xiii

Acknowledgments ............................................................... xiv

Abbreviations and Acronyms ............................................................... xv

Units of Measure............................................................... xv

Executive Summary................................................................ 1

1. Importance of Cost and Schedule Estimation for Power Generation


Projects ............................................................... 5

2. Framework for Evaluating the Performance of Cost and Schedule


Estimates ...........................................................................................................7

3. Data Base of Power Generation Projects ............................................................


11

4. Statistical Approach to Analyzing the Performance of Cost and Schedule


Estimates............................................................. 15

5. The Overall Performance of Power Project Cost and Schedule Estimates .... 23
Group Performance of All Power Generation Projects............................................. 23
Distinction between Thermal Power Projects and Hydropower Projects................. 27
Prevalence of Bias and Uncertainty in Cost and Schedule Estimates....................... 29
Ex Post Analysis of Responsibility for Schedule Slip .......................... .................... 31
Reliability of the World Bank's Methodology for Computing Price
Contingencies ............................................................. 33

6. Significant Project Characteristics and External Variables for Cost and


Schedule Estimates............................................................. 37
Thermal Power Project Costs............................................................. 40
Hydropower Project Costs .............................................................. 43
Thermal Power Project Schedules ................................ .............................. 45
Hydropower Project Schedules.............................................................. 46
Grouping of Significant Variables ............................. ................................ 47

v
7. Implicationsof the Analysisfor PowerSystemPlanning.............. ...................
51
Principal Findings .............................................................. 52
Using Regressions to Improve Predictions for Project Costs and Schedules ........ ... 53
Risk and Planning Issues ............................................................. 57
Measurement of Project Risk............................................................. 58
Distribution of Possible Project Outcomes......................................................... 59
Planning Issues Involving Choice between Sequences of Projects .......... .......... 60
Basic Recommendations ............................................................. 61

Annex 1: World Bank-SupportedPowerGenerationProjects1965-86 Used


for the Analysisof Costand ScheduleEstimatingPerformance....... 63

Annex 2: RegressionResultswith All VariablesIncluded...................................69

Annex3: Comparisonsof ActualRatiosand PredictedRatiosfrom


Regressionsfor Costs and Schedules...................................................
75

Annex4: Statisticsfor Single-VariateAnalysisof All Variables........... ..............


79

Annex5: Ex Post Attributionof FactorsResponsiblefor ScheduleSlip in


World Bank-SupportedPowerGenerationProjects............ ................
83

Annex6: Methodologyfor DerivingActualProjectCosts in ConstantPrice


Terms............................................................... 85

Annex7: Analysisof Relationshipsfor the Performanceof Price


Contingencies ............................................................. 89
1 Relation between Actual and Estimated Current Costs ................. .......................89
2 Link between Actual Current Costs and Actual Constant Costs ........... ...............90
3 Relation between Actual and Estimated Cost Escalation ................. ....................91
4 Relation between Cost Overrun in Current Price Terms, in Constant Price
Terms, and Errors in Predicting Inflation Rates and Project Schedules ......... .... 91
Appendix A7.1 Composition of the World Bank's Price Contingency
Formula for Predicting Project Cost Escalation.................................................. 93

Annex8: Computationsof Probabilitiesof ExceedingSpecific Project


Costs............................................................... 95
Appendix A8.1 The Variance of a Predicted Value from a Regression ......... .........98
Appendix A8.2 Probability that the Outcome of One Project Is Greater Than
That of the Other ............................................................. 98

vi
Annex9: AssigningProbabilitiesto Scenariosfor Risk Analysis........... ...........
101

Annex10: The Calculationof the Meanand Varianceof the Costof Two


Projects ............................................................ 105

Annex11: Applyingthe OptionApproachto ConstructionCosts and


Schedules.......................................................... 107
A. Investment Valuation Under Uncertainty Using the Options Approach............. 107
B. General Solution Methods .............................................................. 109
C. A Simple Model for the Investment Option and Optimal Timing ........... I
11...........
D. Application of the Simple Options Model to Construction Costs and
Schedules .............................................................. 112
Dl. General Assumptions.............................................................. 113
D2. General Formulation to Estimate Option Model Uncertainty..................... 113
D3. The Cases of Thermal and Hydropower Plants........................................... 115

Annex12: Performanceof PowerDemandForecastsand of World BankOil


Price Projections........................................................... 117

References.......................................................... 119

Tables
3.1 GeographicalDistributionof PowerGenerationProjectsin the Data Base.......... ...... 12
3.2 Distributionof PowerGenerationProjectsby Year of Approvalin the Data Base .... 12
3.3 Distributionof PowerGenerationProjectsby InstalledCapacityin the DataBase.... 13
3.4 Distributionof ThermalProjectsby ProductionTechnology,PrimaryFuel, and
Unit Size in the Data Base .......................................................... 13
4.1 Variablesand CharacteristicsUsed in Regressionson ConstructionCost Overrun
and ScheduleSlip.......................................................... 18
5.1 Cases Omittedfrom Analysis.......................................................... 26
5.2 OverallStatisticsfor Cost and SchedulePerformance.....................................
........... 26
5.3 Comparisonof SquaredCorrelationsbetweenActual and EstimatedCosts, and
betweenActualand EstimatedSchedulesfor WorldBank-SupportedPower
GenerationProjects.......................................................... 28
5.4 Comparisonof Cost Overrunand ScheduleSlip betweenWorldBank-Supported
Power GenerationProjectsand All Bank-SupportedProjects..................................... 30
5.5 Chancesof OverrunsExceeding20 PercentSensitivityLevel.................................... 30
5.6 SquaredCorrelationsbetweenCost Overrunsand ScheduleSlips for Thermal
Powerand HydropowerProjects.......................................................... 31

vil
5.7 Ex Post Attribution of Responsibility for Project Schedule Slip................................. 32
6.1 Significant Variables for Thermal Power Project Costs (current values).................... 40
6.2 Significant Variables for Thermal Power Project Costs (constant values).................. 42
6.3 Significant Variables for Hydropower Project Costs (current values) ............ ............ 43
6.4 Significant Variables for Hydropower Project Costs (constant values) ........... ........... 44
6.5 Significant Variables for Thermal Power Project Schedules....................................... 45
6.6 Significant Variables for Hydropower Project Schedules ........................................... 46
6.7 Grouping of Significant Variables for Project Cost and Schedule Estimates.............. 48
7.1 Sensitivity of the Levels of Predicted Values to Indicator Variables ............ .............. 56
A1.1 World Bank-Supported Power Generation Projects 1965-86 Used for the
Analysis of Cost and Schedule Estimating Performance.63
A2.1 Variables for Log of Thermal Power Project Costs (Current Values) .69
A2.2 Variables for Log of Thermal Power Project Costs (Constant Values).70
A2.3 Variables for Log of Hydropower Project Costs (Current Values) .71
A2.4 Variables for Log of Hydropower Project Costs (Constant Values) .72
A2.5 Variables for Log of Thermal Power Project Schedules .73
A2.6 Variables for Log of Hydropower Project Schedules .74
A4.1 Single-Variate Regression Correlations .80
A4.2 Comparison of Significant Variables at 90 Percent Confidence Level between
Multivariate Analysis and Single-Variate Analysis .81
A5.1 Ex Post Attribution of Factors Responsible for Schedule Slip in
World Bank-Supported Power Generation Projects .83
A6.1 Standard Disbursement Profiles for Project Cost in Current Price Terms.86
A6.2 Example of Project Cost Derivation in Constant Price Terms:
Algeria, Base Year 1973.87
A8.1 Probability that Project 1 Has Higher Cost (Schedule) than Project 2 .97
A9.1 Pairs of Parametric Values that Fit the Required Regression Variance of 0.035 for
Scenarios where the External Variance Is a Function of the Size of the
Middle Value .102
A9.2 Pairs of Parametric Values that Fit the Required Regression Variance of 0.01 for
Scenarios where the External Variance Is Not a Function of the Size of the
Middle Value .102
A9.3 Probabilities for Scenarios With Predetermined Variances for Costs and Demand.... 103

viii
Figures
5.1 Relationship between Actual Costs and Estimated Costs for World Bank-
Supported Thermal Power Projects and Hydropower Projects, 1965-1986 ................ 24
5.2 Relationship between Actual Schedules and Estimated Schedules for World
Bank-Supported Thermal Power Projects and Hydropower Projects, 1965-1986 ...... 25
5.3 Distribution of Cost Performance for World Bank-Supported Thernal Power
Projects and Hydropower Projects, 1965-1986 (current prices) ............... .................. 27
5.4 Distribution of Schedule Performance for World Bank-Supported Thermal Power
Projects and Hydropower Projects, 1965-1986 ........................................................... 28
5.5 Errors in Cost Escalation Estimates for World Bank-Supported Power Generation
Projects Approved between 1970 and 1986................................................................. 34
A3.1 Plot of Actual and Predicted Ratios for Thermal Costs (current) .............. .................. 76
A3.2 Plot of Actual and Predicted Ratios for Thermal Costs (constant).............................. 76
A3.3 Plot of Actual and Predicted Ratios for Thermal Schedules........................................ 77
A3.4 Plot of Actual and Predicted Ratios for Hydro Costs (current) ............... .................... 77
A3.5 Plot of Actual and Predicted Ratios for Hydro Costs (constant) .............. ................... 78
A3.6 Plot of Actual and Predicted Ratios for Hydro Schedules ........................................... 78
A7.1 Comparison of MUV Index Actual Values with Values from Forecasts Made
between 1974 and 1988 (Based on actual value in 1980 = 100).................................. 92
A12.1 Performance of Power Demand Forecasts for Developing Countries ........... .............. 117
A12.2 World Bank Oil Price Projections in Constant 1987 US$ per Barrel ............ .............. 117

ix
Foreword
One of the main reasons for the ongoing reforms to power sectors around the world is
the desire for better management of the economic and financial risks of investing in large
power supply projects. These risks arise from uncertainty about future power demand,
fuel prices, and-as shown in this paper-construction costs and schedules.
Conventional planning approaches based on deterministic scenarios of the future under
centralized decisionmaking have seldom given sufficient attention to, or reliable guidance
on, these risks. Now, however, under the more decentralized planning process coming
into use in reformed power sectors, both public and private decisionmakers will have to
respond to the concerns of shareholders, consumers, financiers, and the general public
about the risks of investing in large power projects.
Although the risks from construction cost overruns and schedule delays are often
serious for developing countries, they have been poorly understood to date. In response,
the paper puts forward a number of straightforward techniques to improve the analysis of
these construction cost and schedule risks. The paper also complements several other
Industry and Energy Department and Energy Sector Management Assistance Programme
(ESMAP) publications and seminars on structuring and financing power generation
projects.

Richard Stem
Director
Industry and Energy Department

xi
Abstract
This paper helps national planning and finance ministries, power utilities, and
financing agencies to improve the reliability of their estimates for construction costs and
schedules of power generation projects in developing countries and thereby to improve
the selection and implementation of these projects.
The paper examines estimates of construction costs and schedules that were made for
a group of power generation projects approved for financing by the World Bank between
1965 and 1986. This group of some 64 thermal power plants and 71 hydroelectric plants
is then subjected to a statistical analysis. From this analysis, the paper assesses the
reliability of the estimates and identifies factors that were significantly associated with
bias and uncertainty in them.
The paper draws the following conclusions. First, the average estimating error among
projects as a whole was too large to be ignored. Second, estimated values were
significantly biased below actual values, and the accuracy of estimated values had a large
variance. Third, the performance of cost estimates was much better for thermal power
projects than for hydropower projects, but schedule estimates performed similarly for
these two groups of projects. Fourth, the performance of estimated values can be related
to a number of indicator variables through regression analysis, and these regressions can
be used to derive expected values that carry less uncertainty than the corresponding
estimated values.
The paper then demonstrates how to improve the prediction of the actual construction
cost or schedule for a power generation project by deriving an unbiased expected value
from the estimated value and the appropriate regression equation for the project. There is
a proviso, however, that the project has similar technology and implementation
arrangements (i.e., public sector) to those that characterize the projects analyzed in the
paper. The paper recommends that analyses of power generation projects include a case
in which expected values are used for construction costs and schedules. This case would
supplement the standard analysis that is based on appraised estimates of these values.
Nevertheless, substantial uncertainty remains about the reliability of even the
expected values. The regressions given in the paper can be used to provide a measure of
project risk arising from this source. Consequently, the paper also recommends that the
economic and financial risks associated with the selection of a particular power project or
power development strategy should be explicitly considered during project appraisal.
This analysis would elicit valuable insights about the riskiness of power generation
projects under consideration, such as projects considered to be the least-cost option from
the customary deterministic approach to power system planning. The paper presents
straightforward techniques for evaluating frequently encountered questions of risk about
power projects and development programs.

xiii
Acknowledgments
The authors gratefully acknowledge the valuable comments on drafts of the paper
given by Dennis Anderson, William Buehring, Joseph Gilling, Vladimir Koritarov, Spiros
Martzoukos, Lucio Monari, Lant Pritchett, Mark Segal, Charles Siebenthal, and Odd
Ystgaard.
Thanks also go to Vonica Burroughs and Carole-Sue Castronuovo for word
processing assistance and to Paul Wolman for his editorial and production work.
The authors take responsibility for any errors and omissions in the paper.

xlv
Abbreviationsand Acronyms
CPI consumer price index
forex foreign exchange
G-5 France, Germany, Japan, United Kingdom, and United
States
G&T generation and transmission
GDP gross domestic product
ICB international competitive bidding
IDA International Development Association
MUV UN Unit Value Index of manufactured goods exported
from G-5 countries to developing countries
NPV Net Present Value
O&M operation and maintenance
RPI Retail Price Index
SAR Staff Appraisal Report
SD standard deviation
SEE standard error of estimate
SER standard error of regression
UN United Nations

Unitsof Measure
dollars US$
GW gigawatt
GWh gigawatt hour
MW megawatt

xv
ExecutiveSummary
This paper helps national planning and finance ministries, power utilities, and
financing agencies to improve the reliability of their estimates for construction costs and
schedules of power generation projects in developing countries and thereby to improve
the selection and implementation of these projects.
Project cost and schedule estimates can deviate from actual costs and schedules in
two ways. First, estimates may be generally biased, in that the mean of the estimates for a
group of projects differs significantly from the mean of the actual costs or schedules for
the group. Second, even when such typical project bias is allowed for, estimates are still
subject to uncertainty, in which the relationship between estimated and actual values
shows a large variance around their mean values. By identifying and allowing for facts
that lead to variations in the degree of bias in the estimates for particular types of projects,
it is possible to reduce the overall uncertainty for the financing of power projects and the
development of power systems.
The paper proceeds in three main stages. First, it examines estimates of construction
costs and schedules that were made for a group of power generation projects approved for
financing by the World Bank between 1965 and 1986. This group of some 64 thermal
power plants and 71 hydroelectric plants is then subjected to a statistical analysis. From
this analysis, the paper assesses the reliability of past estimates and identifies factors that
were significantly associated with bias and uncertainty in them. The paper concludes by
reviewing the implications of these findings for the treatment of bias and uncertainty in
estimates of construction costs and schedules for power system planning.
The paper has four important findings:
* Estimates of construction costs and schedules were fairly strongly correlated with the
actual outcomes, but the average estimating error among projects as a whole was too
large to be ignored.
* These estimated values were significantly biased below actual values, and the
accuracy of estimated values had a large variance.
* The performance of estimated costs was much better for thermal power projects than
for hydropower projects, but schedule estimates performed similarly for these two
groups of projects.
* The performance of estimated values can be related to a number of indicator variables
through regression analysis, and these regressions can be used to derive expected
values that carry less uncertainty than the corresponding estimated values.
In covering only World Bank-supported projects, this analysis does not give a reliable
impression of estimating performance for projects financed and implemented by private
sector enterprises.

1
2 EstimatingConstructionCosts and Schedules

For all power generation projects, appraisal estimates showed substantial optimistic
bias and major uncertainty. For this group, the actual project cost (in current prices and
excluding interest during construction) exceeded the estimated project costs on average
by 21 percent of the estimated project cost, with a standard deviation of 36 percent.
Likewise, the actual project implementation periods exceeded the estimated periods on
average by 36 percent of the estimated periods, with a standard deviation of 42 percent.
Cost overruns and schedule slips were weakly correlated, especially for hydropower
projects. A simple sensitivity test reflected inadequately the inherent uncertainty in the
estimates of project costs and schedules. Bias and uncertainty on such extensive scales
have major impacts on investment selection and financing.
In addition, the reliability of the World Bank's methodology for computing price
contingencies for project cost estimates was examined. The results show that the Bank's
methodology improved the prediction of cost escalation, but that substantial room for
improvement remains.
A multivariate regression analysis was undertaken to identify correlation between four
performance ratios and 29 possible explanatory variables. Ten projects (five thermal and
five hydro) were omitted from this analysis because of truly exceptional differences
between their estimated and actual values that would give rise to misleading regressions.
Thermal power projects were explicitly distinguished from hydropower projects because
cost estimates for these groups behaved differently (average cost underestimation for
thermal projects was 6 percent, for hydro projects, 27 percent). The six regressions that
were analyzed thus covered thermal power costs (in current terms and constant terms),
thermal power schedules, hydropower costs (current and constant), and hydropower
schedules. The dependent variables were the ratios of actual to estimated values. The
explanatory variables reflected project technology, project size, procurement method,
host-country features, and World Bank appraisal guidelines. Of these variables, 18 were
found to be correlated significantly with one or more of the performance ratios, as
follows: 7 variables were significant in one of the regressions, 4 variables in two of the
regressions, 5 variables in three of the regressions, 1 variable-estimated cost-in four of
the regressions, and I variable-station extension dummy-was significant for five
regressions. These variables were able to explain about half of the observed variations in
the ratios of actual to estimated costs and schedules. The significance of many individual
variables indicates that the use of an "average adjustment factor" (determined over all
projects) would itself be biased.
The analysis accepted fairly weak evidence of systematic correlation (10 percent
significance test) to be sure of picking up significant factors. The findings show that
further analysis should lead to better models of reducing risk in estimating project con-
struction cost and schedules, in which risk assessment is able to rely on lower measures
of risk.
The paper then demonstrates how to improve the prediction of the actual construction
cost or schedule for a power generation project by deriving an unbiased expected value
ExecutiveSummary 3

from the estimated value and the appropriate regression equation for the project. There is
a proviso, however, that the project has similar technology and implementation
arrangements (i.e., public sector) to those that characterize the projects analyzed in the
paper. The paper recommends that the analysis of power generation projects should
include a case in which expected values are used for construction costs and schedules.
This case would supplement the standard analysis that is based on appraised estimates of
these values.
Nevertheless, substantial uncertainty remains about the reliability of even the
expected values. The regressions given in the paper can be used to provide a measure of
project risk arising from this source. Consequently, the paper also recommends that the
economic and financial risks associated with the selection of a particular power project or
power development strategy are explicitly considered during project appraisal. This
analysis would elicit valuable insights about the riskiness of power generation projects
under consideration, such as projects considered to be the least-cost option from the
customary deterministic approach to power system planning.
The final section thus presents straightforward techniques for evaluating the following
frequently encountered questions of risk about power projects and development
programs. These techniques avoid undue analytical complexity yet overcome the
deficiencies of simplistic sensitivity analysis, and hence they are intended to supplement
standard least-cost analysis. Briefly stated, the techniques aim at
* Ascertaining the cost level for a given project that will be exceeded with a specific
probability and the probability that the project cost will exceed a specified value
(similarly with project schedule).
- Choosing between projects based on the probability of exceeding a cost limit and
evaluating which project has the lower probability of exceeding a given cost limit.
* Assigning probabilities to costs or schedule scenarios for risk analysis that are
consistent with the variance of cost or schedule outcomes derived from the
regressions.
* Making the choice between a large project and a set of two or more smaller projects,
where the reliability of estimates depends on the scale variables identified in the
regression and risks also depend on project size.
* Deciding whether to delay a project or not, where the crucial issue is the need for
more time to improve estimates of key planning parameters-such as site geology,
hydrology, or environmental impacts-for the project or its alternatives. An
application of the financial options approach to the optimal timing of projects under
uncertainty about construction costs and schedules is developed in the paper.
Finally, the paper recommends that these techniques be tested operationally and
developed in case studies, so that guideline/s can be formulated for using them in the
appraisal of power generation projects.
1

Importanceof Costand ScheduleEstimationfor


PowerGenerationProjects
Developing countries are planning massive investments to meet their power needs
(Moore and Smith 1990).1 The construction costs and schedules of power generation
projects thus affect national economic development and the financial viability of project
investors. Power projects are capital-intensive and require lengthy construction periods.
In total, they account for a substantial proportion-averaging about 10 percent-of a
developing country's total physical investment. Some hydroelectric and thermal power
complexes in the largest developing countries rank among global megaprojects costing
billions of dollars. Even modest-sized schemes-by global standards-in small
developing countries can be huge in relation to the size of their economies.
Reliable estimates of construction costs and schedules presented by power utilities
and their consultants at the time of project approval are important for justifying a project
on economic grounds and for planning the means of financing it. Faced with the huge
economic and financial costs of expanding power supply, governments of developing
countries are under pressure to ensure that power projects are selected with due
consideration for the economic and commercial risks that arise from the uncertainty in
these estimates.
The economic impact of a construction cost overrun is the possible loss of the
economic justification for the project. A cost overrun can also be critical to policies for
pricing electricity on the basis of economic costs, because such overruns lead to
underpricing. The financial impact of a cost overrun is the strain on the power utility and
on national financing capacity in terms of foreign borrowings and domestic credit. The
recourse in this situation is to reduce the scale of a project to a level commensurate with

1. It is estimatedthat developingcountriesare planningto install384 GW of new generatingcapacity


during the 1990s at a cost of about US$450billionin 1989 price terms, whichwould increasetheir total
installed generating capacity by about 80 percent over the amount installed in 1989. The principal
productiontechnologiesfor this newcapacityare coal-firedsteam plants(172 GW)and hydroelectricplants
(137 GW). Gas turbines, fueled with natural gas, comprise34 GW. Oil-firedsteam plants compriseonly
14GW.

5
6 EstimatingConstructionCostsand Schedules

the planned financial outlay, but this alternative is seldom rational for a power generation
project because power generation units can produce benefits (i.e., power) only if they are
fully installed.2
The delay of output caused by slippage in a construction project schedule imposes
economic costs when the power system is short of capacity or is supplying power from
plants with high variable costs. The financial impacts on the power utility of schedule
slippage are an increase in its financing charges and the possibility of incurring project
loan repayments before the project generates revenues.
This paper helps national planning and finance ministries, power utilities, and
financing agencies to improve the reliability of their estimates for construction costs and
schedules of power generation projects in developing countries and thereby to improve
the selection and implementation of these projects. From an analysis of World Bank
experience with estimating the costs and schedules for constructing power generation
projects, the paper identifies factors that were significantly associated with the reliability
of cost and schedule estimates.3 In particular, it investigates whether project
characteristics-size, technology and procurement conditions, and external variables
(notably, country conditions and changes in World Bank guidelines for project
appraisals)-were associated with differences in estimating reliability. From this case
study, the paper draws implications for improving power system planning.4
The paper is structured as follows. Chapter 2 lays out the framework for evaluating
the performance of estimates of construction costs and schedules for power generation
projects in developing countries. Chapter 3 summarizes the data base of power
generation projects assembled for this evaluation. Chapter 4 describes the statistical
approach used to analyze the performance of the cost and schedule estimates. Chapter 5
assesses the overall performance of estimates for the group of power generation projects
as a whole, particularly the prevalence of bias and uncertainty in this performance. It also
examines the reliability of the World Bank's methodology for computing price
contingencies for project cost estimates. Chapter 6 identifies project characteristics that
show a significant correlation with cost overrun and schedule slip. Chapter 7 concludes
the paper by examining the implications of the analysis for the treatment of bias and
uncertainty in power system planning.

2. The exceptionto this situationis to defer installationof one or more generationunits in a multi-unit
generation station, but this decision is usually made at the time of project approval rather than during
implementation.
3. The paper does not compareestimateswith actual valuesfor fuel costs of thermal power plants or
operationand maintenance(O&M)costs.
4. The resultsof this analysiscould provide informationfor innovativeplanningmethodologies. See,
for example,Crousillat(1989).
2
Frameworkfor Evaluatingthe Performanceof
Cost and ScheduleEstimates
Project cost and schedule estimates can deviate from actual costs and schedules in
two ways. First, estimates might be biased generally, in that the mean of the estimates for
a group of projects differs significantly from the mean of the actual costs or schedules for
the group. If the estimates of all options for increasing power generation capacity were
similarly biased, the net distortion to project selection would then be confined to
erroneous timing rather than to incorrect choice of project, since the bias could be
explicitly allowed for. Where the bias is not identical for all projects (as is actually the
case), then a consistent direction in the bias for particular types of project can indicate the
presence of a strong influence on estimates that, once identified, is often correctable in
future estimates. Second, even when typical project biases are allowed for, estimates are
still subject to uncertainty, in which the relationship between estimates and actual values
shows a large variance around their mean values. By identifying and allowing for factors
that lead to variations in the degree of bias in the estimates for particular types of projects,
it is possible to reduce the overall uncertainty for the financing of power projects and
development of power systems.5
Most causes of deviation between estimates and actual values of costs and schedules
fall into three categories: (a) poor development of estimates and supervision of projects
by the project sponsor (normally a power utility) and its engineers; (b) poor project
implementation by suppliers and contractors; and (c) changes to the external conditions
(economic and regulatory) for a project. The long time span covered by the projects in
this analysis encompasses a variety of economic and market conditions, as indicated by
periods of high and low inflation and by several boom-and-bust cycles in the international
market for power plant construction. Poor assessment of the prevailing pressures on the

5. Significant bias and uncertaintyappear to exist in forecastsand estimates of most of the major
planning parametersfor power systemdevelopment. For example,powerdemandforecasts in developing
countries are shownto have this feature in Sanghviand Vernstrom(1989). Oil price forecastsalso have
been highly inaccurate. The historic forecasting performancefor these two parameters is shown in
Annex 12.

7
8 EstimatingConstructionCostsand Schedules

suppliers and contractors at the time of bidding for projects can lead to inaccurate
estimates. Furthermore, projects that are marginally economic and politically motivated
may be more liable to have optimistic cost estimates relative to actual construction costs
than projects that show good economic returns because the sponsors of politically
motivated and marginal projects are seeking to obtain approval for financing and to avoid
criticism about high costs. Changes in project scope during implementation can have a
significant impact on the project costs and schedules. Such changes can arise, for
example, from the inability of design-stage investigation to eliminate risks from unknown
geological conditions for construction of underground works, particularly for many
hydropower projects. In addition, for first-of-a-kind projects in developing countries-
and many of the power projects in the data base fall into this category-project estimators
do not have a track record of similar projects as a basis for carefully analyzing major
construction risks and deriving reliable contingencies for them. Instead, they often rely
on unreliable rules of thumb for such contingencies. Some account must also be taken of
such unpredictable events as natural disasters and civil disturbances that severely disrupt
project implementation. In practice, however, it is virtually impossible to obtain a direct
and reliable quantification of the allocation of responsibilities for cost and schedule
deviations between these categories from the available information on project appraisal
and implementation. 6 The closest approximation to this analysis that could be attempted
was a broad allocation of responsibilities for slippage in project implementation
schedules (see, in chapter 5, the section on ex post analysis of responsibility for schedule
slip).
Because a direct analysis of project factors that lead to cost and schedule overruns
could not be undertaken, the paper evaluates the estimating performance of cost and
schedules for a group of completed World Bank-supported power generation projects by
means of a statistical analysis of the following relationships:
* The prevalence of bias and uncertainty within Bank-supported power generation
projects and in relation to all Bank-supported projects.
* On the basis that significant bias and uncertainty is found in the first step, the relation
between the estimated performance and various project characteristics, such as
production technology and project size.
* The relation between external factors associated with project implementation and
estimating performance, particularly procurement method, country economic
conditions, and pressure to complete the project because of demand for its output.
The analysis is based on a comparison of the estimated cost at the time of project
approval with the actual cost of implementing the project (as determined after project

6. This problemmayexplainthe lack of publishedstudiesof cost and scheduleestimatingperformance


for capital projects (as opposed to the publishedstudies that derive econometricmodels for predicting
constructioncostsand schedulesthat are basedon the actualcostsand schedulesfor completedprojects).
Frameworkfor Evaluatingthe Performanceof Cost and ScheduleEstimates 9

completion) and a similar comparison for the project implementation schedule. The
technique of analysis used is multiple regression (using the statistical package
MicroTSP), which allows for the simultaneous correlation of the overruns with several
indicators that may themselves be correlated.7
The analysis of cost performance is done in both current price terms and constant
price terms, since there is no prior basis for assuming that cost performance in the two
cases is affected identically by factors that cause deviations between actual and estimated
values. In other words, price effects on costs would have to be purely random to justify
such an assumption, and the analysis reported in this paper does not support this
assumption, even though the factors influencing cost performance were found to be
similar for the two cases. The actual current costs are directly observable, whereas the
actual constant costs have to be derived from the former. Estimates of project costs in
both current and constant price terms are routinely given in World Bank staff appraisal
reports. Both forms of costs exclude interest during construction for this analysis.
The analysis of costs in current price terms allows for the impact of failure to allow
correctly for price inflation. In the World Bank's appraisal of a project, the cost estimate
in current prices is derived by adding to the constant price estimate a contingency for
price inflation during the project construction period.8 This price contingency allows
particularly for the expected effect on the project cost of contractual price adjustment
clauses relating to materials, labor, and equipment.9 In view of the importance of this
price contingency in estimates of project costs, this paper also includes in chapter 5 an
analysis of the reliability of the Bank's methodology for computing price contingencies.
The constant price cost estimate is the estimated cost of the project at the time of
negotiating the project loan provided that there is no major alteration in project scope,

7. Where none of the indicators are correlated among themselves, a series of single-variable
correlationsbetween the overruns and the indicatorswould give identicalresults to a multiplecorrelation.
Where such variables are intercorrelated,as in this study, multiple correlation is able to reveal which
variablesare significantlyrelatedto the overruns,allowingfor the fact that other variablesare also included
in the explanation. Thus, some variables that appear significant in a single-variablecontext are not
significant when other more important variables are included. Other variables that may not appear
significantin a single-variablecontext can be revealedas significantin the multiple-variablecontext. This
feature is illustratedin Annex4 forthe projectsstudiedin this paper.
8. The World Bank's practice on contingenciesfor the projects under review is given in now-
superseded World Bank internal documents (Central Projects Note [CPN] 3.11 of February 25. 1982;
"Project Cost Estimates and Contingency Allowances." and its Operational Manual Statement (OMS] 2.21
of May 1980, "Economic Analysis of Projects"). The World Bank has used explicit price contingencies for
project cost estimates from 1970 onward. The present methodology for computing price contingencies was
introduced around 1976.
9. Another vintage effect on estimating performance is the request by the World Bank's Board of
Executive Directors in the mid-1970s that all projects presented for its consideration have completed
designs. Before then. most projects presented to the Board had cost estimates based on feasibility level
work.
10 EstimatingConstructionCosts and Schedules

quantities, or contract prices during project implementation. The World Bank includes a
contingency in this estimate for cost (not price) increases attributable to minor changes in
project scope and quantities that are expected to occur between the time of project
appraisal and project completion. This physical contingency forms part of the estimated
value of the project cost and is not intended to compensate for the possibility of bias
toward underestimation.
The paper also assesses the reliability of the World Bank's standard sensitivity test for
uncertainty in estimating project costs and schedules. It is not the World Bank's practice
to use contingency allowances as safety margins for bias and uncertainty. However, the
World Bank does test the sensitivity of its economic analysis for power development
programs to deviations from its estimates for project costs and schedules, typically for a
20 percent overrun from the expected value.'I
In covering only World Bank-supported projects, the analysis applies to projects that
were generally implemented by state-owned power enterprises with government financial
support. This analysis therefore does not give a reliable impression of estimating
performance for projects that were financed and implemented by private sector
enterprises.

10. The World Bank also performs sensitivity analysis on the robustness of the economic justification
by comparing the maximum discount rate at which the project forms part of the least-cost means of meeting
power demand, with an estimate of the opportunity cost of capital to the host country.
3
Data Base of PowerGeneration Projects
The data base assembled for this paper consists of 135 power generation projects in
developing countries financed with World Bank loans and International Development
Association (IDA) credits, of which 64 were thermal power plants and 71 were
hydroelectric plants. Information on these projects is taken from World Bank documents
(principally staff appraisal reports and project completion reports). These projects
constitute virtually all the power generation projects approved for financing by the World
Bank between 1965 and 1986. The analysis thus captures the World Bank's experience
with estimation of project costs and schedules over the long term. Issues related to
sampling were avoided in this case by including all, but only, World Bank-supported
projects of this type. The projects are listed in Annex 1.
World Bank-supported power generation projects are a suitable class for this type of
analysis because they are
* Based on classifiable technologies for providing the same product
* Planned, designed, and procured according to well-established and identifiable
practices
e Not prone to significant changes in scope during implementation, so that estimated
and actual outcomes are comparable
* Fully implemented because they have to be completed to provide any output and,
thus, project benefits
* Well represented throughout all types of developing countries and over the three
decades under study
* Well documented in Bank staff project appraisal reports and project completion
reports, in which the data on appraisal estimates and actual implementation costs and
schedules have been objectively checked by World Bank staff.
The actual project total costs cover a wide range-between $3.2 million and $1,782
million in current-price terms. The actual project implementation schedules also cover a
wide range-between 1.2 and 14.4 years. The projects were implemented in 52
developing countries and were distributed between regions in the manner shown in Table

11
12 EstimatingConstructionCostsand Schedules

3.1. The majority of the thermal power projects were located in Asia and in Europe, the
Middle East and North Africa, whereas the majority of the hydroelectric projects were
located in Latin America and the Caribbean and in Sub-Saharan Africa.
Table 3.1 GeographicalDistributionof
PowerGenerationProjectsin the Data Base

All
Thermal projects Hydroelectric projects projects
Regiona Number % Number % %
Africa (Sub-Saharan) 6 9 19 27 19
Asia 28 44 12 17 30
Europe,MiddleEast
and North Africa 18 28 11 15 21
Latin Americanand
Caribbean 12 19 29 41 30
TOTAL 64 100 71 100 100
aRegions in this table correspond to the prevailing World Bank organizational categories.

In terms of project vintage the data base is fairly well distributed over the 20-year
range as shown in Table 3.2.
Table 3.2 Distributionof PowerGenerationProjectsby
Year of Approvalin the Data Base

All
Thermal projects Hydroelectric projects projects
Period Number % Number % %
1965-69 8 12 19 27 20
1970-74 19 30 16 23 26
1975-79 23 36 18 25 30
1980-86 14 22 18 25 24
TOTAL 64 100 71 100 100

In terms of project size, the data base is well represented in all the capacity ranges for
these projects, and the two groups have similar features, as shown in Table 3.3.
Data Base of PowerGenerationProjects 13

Table 3.3 Distributionof PowerGenerationProjectsby


InstalledCapacityin the Data Base

All
Capacity range Thermal projects Hydroelectric projects projects
(MW) Number % Number % %
0-49 18 28 11 15 22
50-199 13 20 24 33 27
200-499 15 23 22 31 27
500-999 10 16 9 13 14
1,000-2,499 8 13 5 7 10
TOTAL 64 100 71 100 100

The distribution of the thermal projects by type of production technology, primary


fuel, and unit size is summarized in Table 3.4. The size of generating plant in the group
of projects varies from about 2 MW diesel units to 400 MW and 500 MW units in recent
coal-fired steam plants (India, Indonesia, and Korea) and a 550 MW dual-fired (fuel oil
and natural gas) steam plant in Thailand.
Table 3.4 Distributionof ThermalProjectsby ProductionTechnology,
PrimaryFuel,and Unit Size in the Data Base

Number of Range of unit size


Technology projects % (MW)
Steam turbine
Coal-fired 12 19 30-500
Fuel-oil-fired 24 37 25-500
Lignite-fired 3 5 300-330
Gas-fired 1 2 150
Multi-fueleda 6 9 125-550
Subtotalsteamturbine 46 72
Gas turbine 3 5 12-40
Combined-cycle 1 2 300
Diesel 14 22 2.2-20
TOTAL 64 100
aComprisingtwo projectswith coal and fuel oil; two projectswith natural gas and fuel oil; and
two projectswith coal, fuel oil, and natural gas as alternativefuels.
14 EstimatingConstructionCosts and Schedules

The hydropower projects encompass a wide range of hydraulic heads from 12 to


1,035 meters, dams from 11 to 230 meters in height, tunnels from zero to 34 kilometers in
length, and reservoirs from virtually zero to about 840,000 hectares in surface area.
Extensions to existing thermal power stations accounted for 34 (53 percent) of all the
thermal power projects, whereas for hydropower projects 19 (27 percent) were
extensions.
4
Statistical Approachto Analyzingthe Performance
of Cost and ScheduleEstimates
The performance of estimates is measured by the following ratios:
Cost performance. This is the ratio of the actual project cost to the estimated project
cost. It therefore measures construction cost overrun. The analysis is done for two
specifications of cost, one in current prices and the other in constant prices. Interest
during construction is excluded in both cases. The estimate is usually based on
substantial preparation work before the time of approval to lend by the main project
financiers. These costs cover the capital works for the power generators and
associated transmission facilities, engineering services, and local duties on imports. I'
* Schedule performance. This is the ratio of the actual project implementation period
to the estimated project implementation period. It therefore measures slippage of the
construction program. The start of the implementation period is taken to be the date
of project approval by the main financiers, and the end is the date of entry into service
of the completed generation plant (formal acceptance by the power utility).
These ratios standardize the measurement of performance for differences in project
cost and schedule. This feature enables the performance of estimates to be analyzed in
terms of statistical distributions.
A value of unity for a performance ratio implies that the impact of associated factors
on actual implementation has been correctly anticipated in the estimate. It does not
necessarily imply that actual implementation was the best feasible and that the estimate
was made correctly on this basis. Any deviation from unity represents the net effect of
two deviations: that of actual implementation from the best feasible implementation
performance and that of the estimate from the best estimate (equal to best feasible
implementation, by definition).

I1. Local importduties cannot be excludedbecausethe actualpaymentsfor these duties are not given
in the WorldBank's project completionreports.

15
16 EstimatingConstructionCosts and Schedules

The crucial point is that estimates should allow for all foreseeable events and
experience with similar projects, so that deviations between estimates and actual
outcomes should be unforeseen and random. This is the "null" hypothesis. The analysis
reported in this paper tests this hypothesis by finding patterns through correlations. The
analysis does not specifically examine the effects of unforeseen and unforeseeable
conditions during project construction. These effects can only be assessed from an ex
post analysis such as the assessment (reported in chapter 5 of this paper) of force majeure
events on project schedules.
The analysis proceeds in two main stages. In the first stage (reported in chapter 5),
the analysis assesses the prevalence among the group of generation power projects as a
whole of bias and uncertainty in estimating performance for costs and schedules in the
following steps:
* Among power generation projects themselves as a group.
* Against all audited World Bank projects since 1974 (and thus approved from 1968),
in all economic sectors for which reliable data are available on actual and estimated
performance, numbering 2,032 projects. 2
* Against a reference distribution for the performance ratios based on the World Bank's
standard sensitivity test for whether a project is still justified under a 20 percent
overrun from the estimate (see, in chapter 5, the section on prevalence of bias and
uncertainty).
i An ex post assessment of the relative impact on schedule slip of the actions of
clients/engineers, suppliers/contractors, and uncontrollable events.
• The reliability of the World Bank's methodology for computing price contingencies.
Once the overall performance of estimating costs and schedules for power generation
projects has been described, the second stage of the analysis focuses on identifying
external variables and project characteristics that show a significant correlation with cost
overrun and with schedule slip (as reported in chapter 6). The approach is to look for
variables and characteristics that are known at the inception of the project, and that are
plausibly correlated with actual costs or schedules, so that if these links are misestimated,
these variables would also turn out to be correlated with the degree of cost overrun or
schedule slip. If sufficiently strong correlations can be established, then experience of
this effect can lead to better estimates of the likely costs and schedules for future projects.
In such cases there would be less measured overrun or slip.
The statistical analysis is thus searching for factors that have not been fully taken into
account in constructing cost and schedule estimates and whose presence affect costs and
schedules and will thus be correlated with the degree of cost overrun and schedule slip.

12. The analysis of cost and scheduleperformancefor audited World Bank projectsis reported in a
World BankOperationsEvaluationDepartmentReport(WorldBank 1992).
StatisticalApproachto Analyzingthe Performanceof Estimates 17

Because it is not possible to identify these factors reliably in advance of the analysis for
this paper, it was decided to include a wide-ranging group of variables and characteristics
in the analysis to maximize the possibility of identifying factors that are significantly
associated with inaccurate cost and schedule estimates.
Information was extracted from World Bank project reports on 29 project variables
and external characteristics that might be expected to have some correlation with the
degree of costs and schedules for projects as a whole. These characteristics and variables
are listed in Table 4.1. They are organized to bring out some of the underlying factors
that influence the performance of project cost and schedule estimates. Project-specific
variables are categorized under technology, size, and procurement. External
characteristics are assigned either to country variables or to World Bank guidelines for
project appraisal.
Technology variables reflect complexity of project construction, and it might be
expected that the uncertainty of cost and schedule estimates increases with this
complexity. The basic distinction in technology is between thermal power and
hydropower, particularly in terms of the amount of plant and equipment that is
constructed under the suppliers' control on their own premises, which is greater for
thermal power projects. On the other hand, civil works at the project sites are more
prominent in hydropower projects and face the uncertainties of local conditions. On these
grounds alone, hydropower project estimates are expected to be subject to greater bias
and uncertainty. An extension to an existing station is expected to produce better
estimates than for a new station because estimators should be familiar with the specific
station design and do not have to face the uncertainties associated with opening up a new
site.
Greater project size also is expected to increase project complexity and, thus, bias and
uncertainty for estimates. Size is not only reflected in the obvious variables-generator
unit capacity, station capacity, total cost, and construction schedule. Rather, certain
project parameters carry their own estimating risks, such as the well-known uncertainty
associated with underground works. Length of tunnel is thus tested, although the
uncertainty often arises from difficult ground conditions (e.g., karstic limestone; see
World Bank 1984) or from inadequate site investigation. Dam height and hydraulic head
are features that reflect overall project size, not necessarily simultaneously (e.g., a project
with a high dam can have a relatively low hydraulic head, and vice-versa). Reservoir area
reflects another aspect of size-project impact on the immediate vicinity, especially
displacement of resident population. It is expected that hydropower project schedules are
affected by the time required to relocate people from the reservoir area and other project-
related areas, since this component has carried a high degree of uncertainty for many past
projects.
18 Estimating Construction Costs and Schedules

Table 4.1 Variablesand CharacteristicsUsed in Regressionson


ConstructionCost Overrunand ScheduleSlip
Project-specific
variables
Technology
Hydropower project or thermal power project (dummy variable)
New power station or station extension (dummy variable)
Civil works-estimated cost as a proportion of estimated total project cost
(percentage)
Thermal power fuel and technology:
Diesel-fueled combustion turbine (dummy variable)
Coal- or lignite-fueled steam turbine (dummy variable)
Fuel-oil-fired steam turbine (dummy variable).
Size
Generator unit capacity (MW)
Total project generation capacity (MW)
Estimated project cost in current price terms (US$ million)
Estimated construction schedule (years)
Hydropower project features:
Dam height for new hydropower station (meters)
Hydraulic head for new hydropower station (meters)
Reservoir area created by project (hectares)
Length of tunnels (kilometers).
Procurement
Year of World Bank loan agreement
Anticipated sources of suppliers and contractors, by estimated project foreign
exchange costs as a proportion of estimated total project costs (percentage)
Competitiveness of procurement process, by anticipated amount contracted
under international competitive bidding as a proportion of estimated total project
costs (percentage)
Number of financing agencies in the project
Main contractor is from the host country (dummy variable).
Countryvariables
Per capita income of host country in year of loan approval (constant US$)
StatisticalApproachto Analyzingthe Performanceof Estimates 19

Average actual cost growth rate for project components procured from host
countrybetweenyear of loan approvaland year of project completion-the GDP
deflator (percentage)
Actual growth in national (or state) power sales (GWh) between year of loan
approvaland year of projectcompletion(percentage)
Indian thermalpower projects(dummyvariable)
Brazilianhydropowerprojects(dummyvariable)
Colombianhydropowerprojects(dummyvariable)
Index of actual average cost growth rate for imported project components
betweenyear of loan approvaland year of project completion(UN Unit Value
Index of manufactured goods exported from G-5 countries to developing
countries-in constantUS$ terms).
WorldBankappraisalguidelines
Basis of project cost estimate-recent similar projects or tenders for major
componentsof the project itself (dummyvariable)
Pre-1970 loan agreement(dummyvariable)
Post-1976 loan agreement(dummyvariable).

Procurement methods are particularly important for estimating project costs and
schedules because they reflect the degree of competition in the tendering and contract
award process. Since international suppliers and contractors dominate the market for
constructing power stations in nearly all developing countries, the prevailing state of the
order book for these firms is an important indicator of competition for a project. In the
absence of a detailed analysis of this feature from the mid-1960s to the mid-1980s, the
best available proxy indicator is project vintage, taken to be the year of the World Bank
loan agreement, which usually corresponds closely to the time of award of major
contracts for a project. This variable also captures any long-term secular change in
estimating performance over the period of the project loan approvals. The specific degree
of competition for a project can be indicated by the amount of procurement that took
place under international competitive bidding (ICB), as opposed to other procurement
approaches for the project, and by the proportion of foreign procurement measured by the
proportion of total project costs that are incurred in foreign exchange. The origin of the
main project contractor (from the host country rather than another country) is included as
a dummy variable to provide an additional test for procurement. Finally, since official
financing agencies tend to follow their own procurement guidelines (ranging from ICB by
multilaterals to own-country preferences by bilaterals), procurement complexity, and
20 EstimatingConstructionCosts and Schedules

hence estimating uncertainty, tends to increase with the involvement of more financing
agencies in a project.

Three country-specific characteristics and one international characteristic are tested


for their impact on estimating performance. The first is the general level of economic
support that a country can provide for the construction of complex facilities such as a
power station, which relieves and mitigates the uncertainty from reliance on imports. The
best available proxy for most countries is country-per-capita income (although this is not
a perfect indicator because, for example, India and China have low per capita incomes but
substantial industrial capacity, whereas a small middle-income country may have to rely
virtually entirely on imported goods and services). The inflation rate in the host country
during project construction is an indicator of economic (and, in some cases, political)
stability and hence of an important element of uncertainty for constructing complex
projects. The inflation rate for imported goods and services used in a project also
influences the reliability of cost project estimates. The growth rate in power sales in the
country is an indicator of how quickly the new power capacity is required and therefore of
the keenness of the host power company to complete the project within the estimated time
and cost. The international characteristic is a measure of international inflation that
reflects cost growth for imported components of these construction projects. 13
In three countries, however-namely India, Brazil, and Colombia-the World Bank
has supported a sufficiently large number of power generation projects to enable a
country variable to be tested specifically for each of these countries. Country dummy is a
co-variate-it picks up an average effect of variables whose coefficients are not allowed
to be country specific. Where a country dummy is found to be significant, a formal
regression of projects in that country would be the ideal approach to identifying the
country-specific significant variables. In such a situation, the covariance of risk among
projects in a country could then be explicitly considered. Tests for country-specific
groupings for economic variables were not generally carried out because of lack of
degrees of freedom with the number of countries (52) covered by the group of projects.
Nevertheless, remarkably strong results are obtained given the huge variation in economic
conditions encountered among the large number countries covered in the study.
Finally, the World Bank's guidelines for project appraisal specify some key practices
for estimating project costs and schedules. From the mid-1970s onward, such estimates
were based, wherever possible, on bids for major project components. Otherwise, these
estimates were based on completed designs and actual costs for recent similar projects.
Likewise, a formal price contingency in the Bank's cost estimates has been required since
1970, and the current methodology was introduced in 1976. The analysis tests the
impacts of these requirements. (A detailed assessment of the reliability of this

13. In cases where the changein price index (foreignor local) or power sales was negative(3 cases),
the data on the variablesare omitted from the data base becausethey cannot be used in log form for the
regressionequations.
StatisticalApproachto Analvzingthe Performanceof Estimates 21

methodology is described in chapter 5, in the section entitled Reliability of the Bank's


Methodology for Computing Price Contingencies.)
Using multiple regression analysis, a sequential analysis is undertaken in which the
cost and schedule overruns are each correlated with all the above variables and
characteristics; then, variables and characteristics that are insignificant (using a t test) are
dropped one at a time; and, finally, the regression equation is re-estimated. The
procedure is iterated until all the remaining variables are significant. The final regression
equations are taken as the "best" available predictors of cost and schedule overruns, based
on knowledge of the set of variables and characteristics used for the analysis.
5
The OverallPerformanceof
PowerProjectCostand ScheduleEstimates
This chapter is organized in the following stages. First, the performance of cost and
schedule estimates is given for World Bank-supported power generation projects as a
whole. Second, estimating performance is analyzed separately for the group of thermal
power projects and the group of hydropower projects. Third, the prevalence of bias and
uncertainty in cost and schedule estimates is examined. The fourth section examines an
ex post attribution of responsibility for schedule slip. The final section assesses the
reliability of the World Bank's methodology for computing price contingencies in its
construction cost estimates.

Group Performanceof All Power GenerationProjects


The group of power generation projects as a whole has a correlation (squared) of 0.76
between actual costs and estimated costs (in current prices and excluding interest during
construction) and a correlation of 0.55 between actual schedules and estimated schedules
(see Figures 5.1 and 5.2). It is clear that the estimates are fairly strongly related to the
actual outcomes but that there is considerable inaccuracy among projects as a whole in
the estimation of costs and schedules. Preliminary screening of the data on the ratios of
actual to estimated values for both costs and schedules reveals a few cases with truly
exceptional differences between estimated and actual values that should be treated as
such and omitted from any statistical analysis that is looking for regularities. The nine
omnittedcases are shown in Table 5.1.
In addition, a hydropower project in Portugal (Seventh Power Project) consists of a
number of dams, so that it is impossible to relate overall cost and schedule slip to the
characteristics of a single construction, as is the case for all the other projects. Omitting
these exceptional cases leaves 59 thermal power projects and 66 hydropower projects for
detailed analysis.

23
24 EstimatingConstructionCostsand Schedules

Figure5.1 RelationshipbetweenActualCostsand EstimatedCostsfor World


Bank-SupportedThermalPower Projectsand HydropowerProjects,1965-1986
(in currentprices,US$ million)

Estimated
currentcost
10000
log scale
3 Thermal

3150 A Hydro N

x/

1000 - 33u/

E Sh
A

,E~~~~N NE A
ig ~ ~ A NEA A
3132
100 N NA

Xh x

YA~~~~~~
/tx~~~tA Xx
A

100N

/ ~~~~~~~~~~~~~~~~~~~~log
scale

0 3 10 32 100 313 1000 3150 10000

Actual current cost

The data on the variables for these projects are almost complete, with the exception of
informationon reservoirarea. Experiments with this variable on the subset of available
observations suggest that it does not have a significant correlation with the accuracy of
cost or schedule estimates for hydropowerprojects.
The Overall Performance of Estimates 25

Figure 5.2 Relationship between Actual Schedules and Estimated Schedules for
World Bank-Supported Thermal Power Projects and Hydropower Projects,
1965-1986

Estimated schedule (months)

log scale *K Thermal


250
Hydro
200

160

125

100

80 _

63

40

31 _ A

25 A

20 - *

16

13 log scale

10
10 13 16 20 25 31 40 50 63 80 100 125 160 200 250

Actual schedule (months)


26 Estimating Construction Costs and Schedules

Table 5.1 CasesOmittedfrom Analysis


Country Project
Brazil Hydro: Volta Grande
Colombia Hydro: Las Mesitas
Panama Thermal: San Francisco
Romania Thermal: Second Turceni
Sierra Leone Thernal: Third Power Project (King Tom station)
Turkey Thermal: Elbistan
Uruguay Thermal: Battle Unit 6
Yugoslavia Hydro: Middle Neretva project: Grabovica and Salakovac dams
Zambia Hydro: Kariba North
Note: These projectswere more proneto force majeureeventsthan other projectsand thus facedgenuinely
unpredictablemajor risks. (The impactof force majeureevents and schedulesare discussedin the section
belowon ex post analysisof responsibilityfor scheduleslip.) The criterionfor omittingobservations(ratios
of actuals to estimates)is that they were more than 4 standarddeviationsfrom the mean of the remaining
points. Includingsuch observationscan give rise to seriouslymisleadingregressions,since they force the
regressioninto explainingsuch large outliersrather than the bulk of the more centralobservations. All the
remainingobservationslie within2.5 standarddeviationsfromthe mean.
The overall statistics-mean and standard deviation-for the group of projects before
and after removing the exceptional cases are given in Table 5.2.

Table 5.2 OverallStatisticsfor Cost and SchedulePerformance(%)

Project group Mean SD


Withexceptionalcases
Cost overrun
Current costs 21.4 39.9
Constant costs 21.9 37.8
Schedule slip 36.5 41.5

Withoutexceptionalcases
Cost overrun
Current costs 17.4 33.9
Constant costs 18.1 31.9
Schedule slip 29.0 28.5

Note: SD = standarddeviation.
The OverallPerformanceof Estimates 27

DistinctionbetweenThermal PowerProjectsand HydropowerProjects


The first regression carried out examined cost overrun for the whole group of projects
in the data base. Twenty-one of the independent variables are relevant to all classes of
projects. The correlation coefficient for this regression was 0.56 without attempting to
strip out the insignificant variables. With so many variables, this is a moderate result. It
was therefore decided to examine the scope for obtaining a stronger correlation by
carrying out separate regressions for thermal power projects and hydropower projects.
Discriminating between these two project types allows variables that are significant to
only one or the other type to be identified and, thus, improves the regressions. This initial
separation can also be understood in physical terms because these groups of projects
involve qualitatively different construction techniques, and also because preliminary
indications showed that as groups they behave differently, especially as regards the
typical cost overrun (see Figures 5.3 and 5.4).14

Figure5.3 Distribution
of CostPerformancefor WorldBank-Supported Thermal
PowerProjectsandHydropower Projects,1965-1986(currentprices)
Percentof projects

40 Thermal Hydro -

30

20

10

< 0.80 0.80 - < 1.00 1.00- < 1.25 1.25- < 1.50 1.50- < 2.00 > 2.00

Ratio of actualcostto estimatedcost

14. One indication of this difference is the much higher proportion of costs that are accounted by civil
works in hydropower projects (53 percent average) than in thermal power projects (18 percentaverage).
28 EstimatingConstructionCostsand Schedules

Figure5.4 Distributionof SchedulePerformancefor World Bank-Supported


ThermalPowerProjectsand HydropowerProjects,1965-1986
Percent of projects

60
- Thermal Hydro

50

40-

30 -

20 -

10 -

0
< 0.80 0.80 - < 1.00 1.00-< 1.25 1.25-< 1 50 1.50-< 2.00 > 2.00

Ratio of actual cost to estimated cost

One concern about splitting the group of projects into two subgroups is whether there
are sufficient degrees of freedom (number of observations less number of independent
variables). Significance tests take account of both the number of observations and the
number of variables used in any of the regressions. In this case, it was possible to identify
several strongly significant variables in each regression.
The first step in this analysis was to find the correlations between actual and
estimated costs and schedules for all World Bank-supported power generation projects,
thermal power projects, and hydropower projects-excluding the 10 outliers. These
correlations are compared in Table 5.3.

Table 5.3 Comparisonof SquaredCorrelationsbetweenActualand Estimated


Costs, and betweenActualand EstimatedSchedulesfor World Bank-Supported
PowerGenerationProjects

Cost
Project group Current Constant Schedule
All power projects 0.82 - 0.63
Thermalpower projects 0.90 0.78 0.72
Hydropowerprojects 0.81 0.93 0.52
Note: Sampleexcludesthe 10outlierprojects.
The OverallPerformanceof Estimates 29

These results indicate (particularly for schedules) only a moderate correlation between
actual and estimated values. Thermal power projects had a high correlation for both costs
and schedules.

Prevalenceof Bias and Uncertaintyin Cost and ScheduleEstimates


Analysis of the basic data on cost overrun and schedule slip for the 59 thermal power
projects and 66 hydropower projects-excluding the 10 outliers-gives the following
results.
For thermal power projects as a whole there was a relatively small bias for costs,
with an average underestimation of 6 percent, but for schedule there was a large average
underestimation of 30 percent. Both the cost estimates and the schedule estimates
showed substantial variation around these mean values, with a standard deviation of 23
percent (around a mean of 106 percent for actual to estimated values), for costs and a
standard deviation of 30 percent (around a mean of 130 percent) for schedules.
For hydropower projects both the average cost overrun of 27 percent and the average
schedule slip of 28 percent were very substantial. The cost ratios showed an extremely
high standard deviation of 38 percent (around a mean of 127 percent), whereas that for
the schedule ratios was 28 percent (around a mean of 128 percent).
Thus, even after removing the quite exceptional cases, it is clear that forecasts of
schedules generally seriously underestimate the actual schedule for both thermal power
and hydropower projects and that cost estimates were seriously underestimated for
hydropower projects. Costs for thermal power projects were generally only slightly
underestimated. There was also a very large variation in the reliability of estimates for
costs and schedules for both thermal power and hydropower projects. Even with attempts
to correct estimates by adding on a "typical" slip factor of (say) 27 percent for
hydropower project cost estimates, deviations would still have been large.
Costs and schedules for power generation projects have been estimated more
accurately than for World Bank-supported projects in general, both with respect to the
mean error and the standard deviation of these errors. The single exception is that cost
overruns for hydropower projects have shown a substantially larger average error than the
average cost overrun for the totality of World Bank-supported projects. The comparison
with World Bank-supported projects as a whole is shown in Table 5.4.
30 Estimating Construction Costs and Schedules

Table 5.4 Comparisonof CostOverrunand ScheduleSlip betweenWorldBank-


SupportedPowerGenerationProjectsand All Bank-SupportedProjects(%)

Cost overrun Schedule slip


Project group Mean SD Mean SD
All Bank projects 11 45 117 80
Thermal power projects 6 23 30 30
Hydropower projects 27 38 28 28
Note: SD = standard deviation. The effective cost overrun for all World Bank-supported projects may
have been much higher than indicated here because the scope of many of these projects was substantially
reduced during implementation to keep total project expenditure within the available funding (this did not
happen with power generation projects).
Source. World Bank (1992).
A 20 percent sensitivity test for economic analysis is inadequate to allow for the
inherent bias and variation in the estimates of power project costs and schedules. This
sensitivity test can be expressed as a null hypothesis that the average forecast error is zero
and the standard deviation of the forecast error is 20 percent. Assuming errors to be
normally distributed, it follows from this hypothesis that 16 percent of all projects would
be expected to have an actual-to-estimated ratio exceeding 120 percent (one standard
deviation), and only 2 percent of projects would have this ratio exceeding 140 percent
(two standard deviations).15 That is, one in six of projects might be expected to have a
larger cost or schedule overrun than the one used for sensitivity analysis. In the light of
the larger actual means and standard deviations found above for overruns, the chances of
exceeding the sensitivity criterion are as given in Table 5.516

Table 5.5 Chancesof OverrunsExceeding20 PercentSensitivityLevel(%)

Cost
Project grouip Current Constant Schedule
Thermal power projects 27 32 63
Hydropowerprojects 57 59 61

15. Sensitivitytests are usuallyconductedonly for overruns,so that a right-skeweddistributionwould


be more accurate than a normaldistribution. Findingan appropriatedistributionwould requirean analysis
of residuals,and hence a normal distributionwas chosen just to illustratethe limitationof this type of
sensitivitytest.
16. For example,for the thermal power projectcost overrunthat has a normaldistributionwith mean
1.06 and standarddeviationof 0.23, this is equivalentto askingwhat the chance is that a random drawing
exceeds 1.2. The answeris equal to the probabilityof drawingfroma standardnormal distributiona value
exceeding(1.2 - 1.06)/0.23= 0.61. whichis 27 percent.
The OverallPerformanceof Estimates 31

Even allowing for 20 percent underestimation, the majority of hydropower projects


could be expected to have a larger cost overrun and schedule slip, and the majority of
thermal power projects could be expected to have a larger schedule slip, than the
sensitivity analysis value. In order to construct a sensitivity test at a level where only a
few projects could be expected to show larger cost overruns or schedule slip, a 40 percent
level would be more appropriate. With such a large test deviation, however, standard
deterministic approaches to project justification become virtually unworkable. This
finding thus underscores the need for proper risk analysis rather than simplistic sensitivity
testing.
Finally, the standard procedure of testing sensitivity to cost overruns and schedule
slips separately fails to capture the compounding effect on the benefit/cost ratio of a
project when both errors occur. In this case, the increase in cost arising from schedule
slip through the effect of time on the value of money would lead to higher chances of
overruns exceeding the 20 percent level than those given in Table 5.5.
The correlations between costs overruns and schedule slips for thermal power and
hydropower projects were found to be weak, as shown in Table 5.6.

Table 5.6 SquaredCorrelationsbetweenCost Overrunsand


ScheduleSlipsfor ThermalPowerand HydropowerProjects

Thermalpowercost overrun 0.24


Thermalpower scheduleslip I
Hydropowercost overrun 0.01
Hydropowerscheduleslip 0

The difference in magnitude of these correlations suggests that the unanticipated


factors may have some common factors for costs and schedules of the thermal power
plants but that for hydropower plants the unanticipated factors are completely different as
regards cost overruns and schedule slip.

Ex Post Analysisof Responsibilityfor ScheduleSlip


An ex post analysis of the causes of slippage in project implementation schedules was
undertaken to attribute responsibility between the two main parties for project and
implementation-namely client/engineer and contractor/supplier-as well as to
uncontrollable events. This analysis covered 103 of the power generation projects
approved for financing by the World Bank between 1965 and 1986. The relative
importance of the causes was identified by counting the number of times each cause was
cited in project completion reports.
A wide range of causes were cited (they are listed in Annex 5; 25 for thermal power
projects, 41 for hydropower projects). Many of the causes (transportation difficulties,
32 EstimatingConstructionCostsand Schedules

delays in contract award, equipment failure during testing, replacement of substandard


work, and shortages of materials) were cited in both classes of projects. Other frequently
cited causes for hydroelectric projects were geological problems, design changes, bad
weather, and poor project management. Thermal power projects were also subject to
delays caused by labor disputes and shortages of skilled labor.
A noticeable feature of the list of causes is the emphasis on implementation
difficulties, with the implication that they would not have been anticipated in schedule
estimates based on the assumption of good project implementation performance.
However, it can be argued that project clients and their engineers should have been able
to anticipate problems under their control (and have had the incentives to do so) in
preparing their estimates of project schedules, even if they had to assume that contractors
and suppliers would perform soundly and that uncontrollable events (such as natural
disasters and civil disturbance) were too unpredictable to factor into their estimates.
The citations were then attributed to categories-client/engineer, contractor/supplier,
and uncontrollable events (relative frequencies of the citations are shown in Table 5.7).

Table 5.7 Ex Post Attributionof Responsibilityfor ProjectScheduleSlip (%)

Party orfactor Type of project


responsible for slip Hydropower Thermal power
Client/engineer 43 34
Contractor/supplier 48 50
Uncontrollableevents 9 16

In order to test whether these factors are captured by the variables used in the
regression analysis (see chapter 6), a simple test is used. Three "dummy" variables are
created for the three causes of slip, and for any project the dummy is given a value of
unity if that factor was identified as having been a problem during implementation (and
zero if not). The best regressions for schedule slip, as given below, are then rerun with
the inclusion of the three dummy variables. For neither thermal project schedule nor
hydro project schedule are any of the dummy variables significant, suggesting that the
included variables, which are known in advance of project implementation, are as
successful as a crude ex post analysis in identifying sources of schedule slip. If the
factors used to construct the dummy variables could be given weightings, rather than
simply dichotomized as present or absent, then ex post analysis might be able to identify
some of the variation that is unexplained by the regression.
The OverallPerformanceof Estimates 33

Reliabilityof the World Bank's Methodologyfor


ComputingPrice Contingencies
To allow for escalation in its project cost estimates during project implementation, the
methodology introduced in 1976 by the World Bank computes a price contingency from
the following three basic parameters:
* Estimated project cost in constant price terms of year project approved ("base year")
* Forecast escalation factor for the costs of imported project components
* Forecast escalation factor for the costs of locally procured project components.
This apparently simple formulation in fact involves a complex set of relationships
between numerous variables, as shown by the fully expanded formula given in the
appendix to Annex 7. The World Bank's formula thus has the disadvantage of
compounding the uncertainties involved in forecasting many variables. The extent of this
uncertainty is shown by the wide range of errors in cost escalation estimates for World
Bank-supported power generation projects that is shown in Figure 5.5. It is noticeable
that the greatest errors occurred during the periods of highest international inflation
between 1970 and 1986.
The factors that determine the reliability of this cost escalation formula are thus the
accuracy of the projections for these variables together with the forecast pattern of
commitments to project expenditures over the estimated schedule of project
implementation.
If full ex post information on assumptions and actual values for all these variables
were available for the projects, the reliability of the formula could be gauged by a simple
disaggregation of the formula to derive the contributions to cost escalation. Some of this
information is not available, however, particularly on the proportions of actual costs that
are incurred in each year of implementation. For purposes of assessing the reliability of
this formula from the available project data, the analysis therefore proceeds in the
following manner.
The estimate of project cost in constant price terms is noted as CO(E). To this value,
a price contingency factor is added to give an estimate in current price terms, CU(E).
At the termination of project construction, the actual cost in current price terms,
CU(A), is revealed. From this cost, an implicit actual value in constant prices, CO(A), is
calculated using the disbursement formula described in Annex 6. The analysis proceeds
by examining the cost overrun in current price terms, O(CU) = CU(A)/CU(E), and
finding factors that are correlated with the extent to which actual costs exceed estimated
costs (both measured in current prices). The analysis is carried out on the record of the
group of power generation projects covered by this paper.
34 Estimating Construction Costs and Schedules

Figure5.5 Errors in Cost EscalationEstimatesfor World Bank-SupportedPower


GenerationProjectsApprovedbetween1970and 1986
Year of approval
1986 $
1985
1984

1983

1982

1981

1980

1979 _

1978

1977

1976

1975

1974

1973

1972

1971

1970 l_l_l

-4 -2 0 2 4 6 8 10
Ratio

Note: Ratio of (ActualCost Growth less EstimatedCost Growth)to EstimatedCost Growth, wherecost
growthis the differencebetweenprojectcost in currentpricesand projectcostin constant(baseyear)prices.
The OverallPerfonnanceof Estimates 35

Since the estimate of current costs is based on an estimate of constant costs together
with estimates of the price factors, errors in either or both of the latter terms will lead to
errors in cost estimates expressed in current prices.
The basis of this analysis is to examine a series of price relationships to find
interrelationships between these different measurements. Four separate links are
explored:
a. The link between actual costs in current prices and estimated costs in current prices
b. The link between current actual costs and constant actual costs
c. The link between actual cost escalation (current actual costs relative to constant actual
costs) to estimated cost escalation (current estimated costs relative to constant
estimated costs)
d. The link between the cost overrun (actual versus estimated) in current terms to cost
overrun in constant price terms and the errors in predicting inflation rates and project
schedule.
These four links are interconnected and follow a natural sequence. The first reveals
the degree of accuracy of the central value-the cost in current prices-whereas the
second indicates the importance of inflation in the actual cost that emerges. Given that
there is inflation (cost escalation), the third reveals the extent to which the estimated cost
escalation is an adequate predictor of actual cost escalation. The fourth examines the
relationship between the failure to produce unbiased estimates of actual costs (in current
terms) and the failures to predict costs in constant price terms (the physical dimension)
and to predict inflation (a function of inflation rates and construction schedule) correctly.
Once it is established that any failure to forecast costs well is due to failures to predict
costs in constant terms as well as to failure to predict inflation, then it is important to
search for factors that are known at the time of project preparation that tend to be
correlated with errors in cost estimation, so that an appropriate adjustment can be made to
the estimates or a warning given on possible bias.
Annex 7 presents the statistical analysis of these four links. The principal findings are
as follows:
a. The actual and estimated values, both measured in current costs, are strongly but far
from perfectly correlated. Moreover, estimated values are significantly below actual
values. Such a difference can be due to differences in predicting the physical costs
(constant prices) or the inflation rate.
b. In order to check the accuracy of the working of the price contingency formula, the
actual current cost was correlated against the actual constant cost value, as well as the
actual domestic inflation, the actual inflation of imported manufactures, and the actual
schedule length. The strong relation between these variables (especially for thermal
power projects) confirms that the "physical" aspect of the project can be separated
from the inflation component.
36 EstimatingConstructionCostsand Schedules

c. The actual cost escalation is correlated with the estimated cost escalation, and this
revealed a strong but not perfect relationship, leaving room for further improvements
in constructing this aspect of a project cost forecast. As expected, no difference in
forecasting performance was found between thermal and hydropower projects.
d. A regression was made of the cost overrun in current price terms on the cost overrun
in constant price terms, errors in forecasting domestic inflation and imported
inflation, and errors in estimating the schedule period. This showed that both the
physical and inflation aspects of forecasting errors could be identified, because errors
in estimating three of these variables-constant project costs, domestic inflation, and
schedule length-were all significantly related to the overall error in estimating
project cost in current terms. The relationship was weaker for hydropower than for
thermal power projects, confirming an earlier finding that forecasting for the physical
component for hydropower projects has been less successful than for thermal power
projects.
6
SignificantProjectCharacteristics
and External
Variablesfor Costand ScheduleEstimates
In view of the size and distribution of the errors in estimating project costs and
schedules, it is desirable to look for a method that identifies classes of projects that tend
to have high or low cost overruns or schedule slips. The approach is to recognize that a
number of project-specific factors will affect the actual cost and actual construction
schedule of any project. Many of these factors are quantifiable (such as project size,
measured either in costs or some physical unit), although some are only classifiable by
their presence or absence (whether a project is new or merely an extension of an earlier
completed project).
Chapter 4 described in detail some 29 variables that can plausibly be considered to be
correlated with actual costs and schedule. For these links, the sign of the correlation
might be predictable (e.g., that in higher income countries the actual costs of a given
project would be less than they would be if the same project were constructed in a lower-
income country). If such features are known and fully recognized, they would be taken
into account in constructing the estimates of costs and schedules. Any errors in
estimation are then associated with unforeseeable events or with incorrect assessments of
the importance of the project-specific factors. In other words, the results of the regression
analysis do not specifically identify the main risk factors but rather how well they are
treated. Thus, some risk factors may appear significant because estimators did not
consider them with due care, even though these factors would not generally be considered
as among the most significant factors. Conversely, factors known to be risky may be so
well treated that they do not appear among the significant factors in this analysis.
Regarding the group of projects as a whole, purely random factors would make
estimates diverge from actual, with some above and some below. Moreover, there would
be no pattern in the errors taken in their entirety: Not only would the average error for a
large enough group of projects be zero (or very near to zero) but the error and any
observable variable known at the time of constructing the estimate would not be
correlated. Where a factor has been allowed for but its impact is systematically under- or
overemphasized, the regression residual would then be correlated with that factor. Even

37
38 EstimatingConstructionCostsand Schedules

when a factor is positively (say) correlated with actual costs, the error can be negatively
or positively correlated with that factor depending on whether the implicit link is over- or
underestimated.' 7
When all the relevant systematic factors have been allowed for in constructing
estimated values, and when the strength of these links has been correctly identified, then
the ratio of the actual to estimated values should fluctuate randomly around a mean of
unity. This is then the null hypothesis-that no variables should have a significant
correlation with the ratio of actual to estimated values. If a significant correlation can be
identified, then the analysis has picked up tendencies to over- or undercompensate for
risk factors by the project analyst. In this situation, the estimates were capable of
systematic improvement taken for the group of projects as a whole.
The use of regression techniques allows both the identification of any such significant
factors and the quantification of their relative importance in terms of the amount of
variance explained by them. The factors with the largest t values are significant at the
highest confidence levels. For example. a t value of 1.67 gives a 90 percent confidence
limit, and a t value of 3 gives a 99 percent confidence limit, for the number of
observations that correspond to the number of power projects in the thermal group or the
hydro group. The unexplained variance after allowing for the effect of significant factors
has to be treated as the unpredictable risk for these estimates. As given by the formula in
Annex 4, the t ratio on each variable in a regression can be used to calculate its partial
correlation with the dependent variable. The squared partial correlation is the percentage
of total variation in a model not explained by the set of all other variables included in the
equation that is explained by the variable in question. The higher the t ratio, the higher
the partial correlation attached to that variable.
Although even preliminary inspection can reveal that certain variables are strongly
correlated with differences in the accuracy of estimating actual values, with so many
possible variables to check it is necessary to adopt a systematic search procedure,
particularly since there are no prior indications of which factors have been taken into
account imperfectly in arriving at the estimated values. Accordingly, multiple regressions
of cost overrun and schedule slip were carried out on the factors described in Table 4.1.
Trials with such regressions for quantitative variables, including the performance ratio

17. These ideas can be formalized:Let actual values (A) of costs or schedules be related to an
indicatorX and a random term u by the loglinear form: A = aXb * eU;while the estimated value (E) is
calculated on the basis of the level of factor X by the form: E = cXd. Hence the log of the ratio of actual to
expected values is given by log (A/E) = log (a/c) + (b - d) log X + u. which is a simple form of the model
used in the multiple regressions. The sign of (b - d) depends both on the sign of the actual relation and the
degree to which the schedule is misestimated.
SignificantProjectCharacteristicsand ExternalVariables 39

itself. indicated that it was most satisfactory to run a "double log" form of regression so
that individual parameters can be interpreted as elasticities.'8
Once the most "general" equation that included all variables had been estimated,
variables that were "insignificant" (based on a 90 percent confidence level) were removed
one at a time, and the equation re-estimated omitting that variable.'9 The systematic
elimination of insignificant variables not only reduces the degree of bias in the remaining
regression but also reduces the overall level of uncertainty (variance) in it because it
removes the variances associated with these variables.
Given that it is not possible to predict which variables will be correlated with failures
to estimate costs and schedules accurately before the analysis, it was considered
reasonable both to use a large size of significance test in order to pick up even weak
regularities in the data, and to use a two-sided test since the impact of a variable on
arriving at estimated values could be either under- or overvalued. The final selected
equation thus has only "significant" variables.
The multiple correlation (R-squared) coefficient measures the percentage of the total
variation of the performance ratio that is explained by the inclusion of the variables. A
correlation of zero would indicate that the explanation was no better than that achieved by
just using the sample mean of the dependent variable, whereas a correlation of 100
percent would indicate that a complete explanation for the variation had been achieved.
A linearized correlation is obtained by correlating the anti-logs of the fitted values of the
equation with the actual ratios. From this latter equation, the average residual (linearized
standard error of regression) is calculated that indicates the unexplained residual variation
in the ratio by the regression and that can be compared to the standard deviation of the
basic data.
The same approach is used for the analysis of both the cost overrun ratio and the
schedule slip ratio. The log of each ratio is correlated with all variables, and then
insignificant variables are sequentially removed. No attempt was made to impose the
same factors in the analysis of schedule slip that were found to be important for cost
overrun or vice-versa.

The analysis for significant variables was carried out separately for the following six
cases: thermal power project costs, hydropower project costs (costs in both current and

18. It should also be noted that in absoluteterms the ratio of actual to estimate must fall betweenzero
and infinity (a skewed distribution),whereasthe log form of the ratio will be more symmetricand fall
between plus and minusinfinity. Taking logs thus supportsthe presumptionthat regressionresidualswill
be normallydistributed.
19. A confidencelevel of 90 percent(test size, 10percent)indicatesthat were the null hypothesisof no
correlationto be true, such a regressionvaluewould be expectedto occur in randomsamplingin less than
10 percentof the time. It is possibleto see from the test that some of the variableswould have passed a
higher significancetest of 5 percent or even I percent(two-tailedsignificanceresults).
40 EstimatingConstructionCostsand Schedules

constant values), thermal power project schedules, and hydropower project schedules.
Annex 2 shows the regression results with all the tested variables included in the
equation. Annex 3 shows plots of the actual ratio (actual value to estimated value) and
predicted ratio (predicted value from one of the regression equations to estimated value)
for the cost and schedule of each power generation project. These plots show how well
the regressions are generally able to predict the value for each project over the wide
ranges of the actual ratios among these projects.
Thermal Power Project Costs
The performance of project cost estimates is analyzed for two cases: first for costs
expressed in current values (including the effect of price inflation on costs) and second
for costs expressed in constant values. The results of the analysis for thermal power
project costs (current values) are given in Table 6. 1.

Table 6.1 SignificantVariablesfor ThermalPowerProjectCosts (currentvalues)

Regression 2-tailed
Variable coefficient Standard error t-stat. significance
Intercept 0.525 0.124 4.226 0.000
Log estimatedcost -0.146 0.028 -5.193 0.000
Log estimated 0.287 0.106 2.708 0.010
schedule
Extension dummy -0.156 0.054 -2.862 0.007
Indiadummy 0.201 0.074 2.728 0.010
Log civils costs ratio 0.083 0.034 2.466 0.018

R-squared 0.523 Mean of dependent 0.018


SER 0.177 SD of dependent 0.242
LinearizedR-squared 0.573 LinearizedSER 0.169
Note: Dependentvariable is log of the thermalcost ratio, basedon 45 observationsand using 10
percent retention rule.
civils = civil works; SD = standard deviation; SER = standard error of regression.

In the case of thermal power project cost overruns, the multiple correlation based on
all explanatory variables was 80 percent (Annex 2). Once all the insignificant variables
had been removed, a strong (linearized) correlation of 57 percent was obtained with five
SignificantProjectCharacteristicsand ExternalVariables 41

significant variables. Since any regression that includes an intercept has the property that
its mean prediction is equal to the mean of the actual data, this regression has an average
prediction error (residual) of zero (compared to the bias of 6 percent in the raw data). By
allowing for the variations in the measured indicator variables between projects, this
regression is also able to reduce the variation around this unbiased value by about one-
quarter, from the standard deviation of 0.23 for the cost ratios of thermal power projects
as a whole (reported in chapter 5 at the beginning of the section on prevalence of bias and
uncertainty) to the 0.17 of the linearized standard error of regression.
Both estimated costs and station extension dummy had a negative association with
cost overruns. Thermal power projects in India, projects with long estimated schedules,
and projects with a large estimated civil works component all had a positive association
with cost overruns. The two variables related to project size (cost and schedule) thus
influenced estimating performance in opposite directions. Where the variables are
entered into the regressions in log form, the coefficients are elasticities. Thus, Table 6.1
shows that a 1 percent increase in the level of estimated costs is associated with a 0. 14
percent decrease in cost overruns, whereas a 1 percent increase in the share of civil works
costs is associated with a 0.08 percent increase in cost overrun.
The use of the regression equation to predict the actual fitted values is illustrated for
one of the observations on thermal costs where the estimate was substantially too low.
The India 911 project (loan date, 1978) had an estimated cost of $405.9 million, but the
actual cost turned out to be $491.1 million. The ratio of actual to estimated values is
therefore 1.21, rather than the unit value it should have had if the estimate had correctly
predicted the actual outcome. Given that the group of thermal projects as a whole had a 6
percent cost overrun, it can be seen that even applying this average correction to the
estimated costs would have still produced a substantial error. The regression equation in
Table 6.1 is a project-specific way of making a prediction of the ratio of actual to
estimated costs. For the India 911 thermal power project the estimated cost was 405.9
(natural log value 6.006); the estimated schedule was 5.33 years (log value 1.673); the
project was not an extension (dummy value 0.0); the project was in India (dummy value
1.0) and the civil cost ratio was 0.192 (log value -1.655). Multiplying these values by the
regression coefficients shown in Table 6.1 and then summing these values gives a fitted
value for the log of the cost ratio of 0.192, and the antilog of this ratio gives a predicted
value of the cost ratio of 1.211. In this case, the regression model is extremely accurate
and clearly outperforms the basic estimate (ratio unity) or the average bias correction
(ratio 1.06) as a predictor of the ratio of actual to estimated costs. Not all projects are
predicted as well as this, but the linearized correlation of 57 percent shows that more than
half the variation that would be left unexplained by the average bias correction approach
(equivalent to using a regression model with just an intercept as independent variable) is
explainable in terms of a few simple variables whose values were known at the time of
loan approval.
42 EstimatingConstructionCostsand Schedules

The results of the analysis for thermal power project costs in constant values are given
in Table 6.2.

Table 6.2 SignificantVariablesfor ThermalPowerProjectCosts(constantvalues)

Regression Standard 2-tailed


Variable coefficient error t-stat significance
Intercept -10.574 2.957 -3.575 0.000
Log of loanapprovaldate 2.626 0.705 3.721 0.000
Log of estimatedcost -0.274 0.054 -5.083 0.000
Log of estimatedschedule 0.174 0.090 1.932 0.059
Log of MUV index -0.081 0.043 -1.899 0.063
Log of unit size 0.115 0.041 2.745 0.008
Extensiondummy -0.126 0.055 -2.299 0.025
Dummyfor 1976 -0.264 0.110 -2.388 0.020
Indiadummy 0.171 0.070 2.448 0.018

R-squared 0.463 Mean of dependent 0.065


SER 0.173 SD of dependent 0.219
LinearizedR-squared 0.447 LinearizedSER 0.179
Note: Dependentvariableis log of thethermalconstantcostratio,basedon 57 observations
and using10
percentretentionrule.
SD = standarddeviation;SER= standarderrorof regression.

Lower correlations are obtained in the case of constant costs than in that of current
costs, although the standard errors of regression are virtually equal. The multiple
correlation based on all explanatory variables for constant cost overruns is 58 percent
(Annex 2), and on only the significant variables, 46.3 percent. Nevertheless, there are
more significant variables for the constant costs case (8) than for the current costs case
(5). Four variables-estimated cost, estimated schedule, extension dummy, and India
dummy-are significant in both cases. Project size seems to be highly significant for the
constant cost case, since estimated cost and estimated schedule, as well as unit size, are
significant variables for this case. The significance of the MUV index and the 1976
dummy could indicate that the World Bank's methodology for computing price
contingencies (see chapter 5) has a significant bearing on the performance of project cost
estimates in constant values.
SignificantProjectCharacteristicsand ExternalVariables 43

HydropowerProjectCosts
The results of the analysis for hydropowerproject costs (currentvalues) are given in
Table 6.3.

Table6.3 Significant
Variablesfor Hydropower
ProjectCosts(currentvalues)

Regression 2-tailed
Variable coefficient Standard error t-stat significance
Intercept 0.397 0.394 1.009 0.317
Log estimatedcost -0.193 0.043 -4.468 0.000
Log forex -0.224 0.073 -3.062 0.003
Log stationsize 0.117 0.042 2.828 0.006
Log GDP deflator 0.076 0.032 2.383 0.021
Extensiondummy 0.360 0.198 1.819 0.074
Log hydraulichead 0.110 0.035 3.131 0.003
Colombiadummy -0.423 0.102 -4.167 0.000
Log financing 0.234 0.083 2.832 0.006
agencies
R-squared 0.511 Mean of dependent 0.192
SER 0.224 SD of dependent 0.301
LinearizedR-squared 0.511 LinearizedSER 0.267
Note: Dependentvariableis log ofthe hydropower
costratio,basedon 66 observations
andusing
io percentretentionrule.
SD = standarddeviation;SER= standarderrorof regression.

For hydropowerprojects,wherethe mean and standarddeviationsof the cost overrun


are much larger than for thermalpower projects,and wherethe regressionon all variables
has a correlationof 62 percent (Annex 2), the final regression is fairly successful,since
the squared correlationis 51 percent. The linearized standarderror of regressionof 0.27
shows that the regressionreduces the uncertaintyaround the mean overrun by one-third
from the standarddeviationof 0.38 for the cost ratios of hydropowerprojects as a whole.
Again, the log of the estimatedcost variablehas a negative sign, indicatingthat a given
percentage cost overrun was less likely with a larger financing plan. The log of the
percentageof foreignexchangealso had a strong negativerelation. The station size had a
positive relation, as did the local inflationrate. The hydraulichead (for projects that are
not extensions) had an important positive relation with cost overrun, since this is a
variable that can vary greatly in magnitude among projects. A particularlyinteresting
44 EstimatingConstructionCosts and Schedules

finding is that the group of Colombian hydropower projects as a whole tended to have a
lower percentage cost overrun than would be expected for their size, hydraulic head, and
number of financing agencies (about six). A project-by-project analysis of this group
confirms that their mean cost overrun (measured in current prices) was only 1.5 percent,
compared with 27 percent for hydropower projects as a whole.2 0 No such country-
specific effect was found for the group of Brazilian hydropower projects.
The results of the analysis for hydropower project cost in constant values are given in
Table 6.4.

Table 6.4 SignificantVariablesfor HydropowerProjectCosts(constantvalues)

Regression Standard 2-tailed


Variable coefficient error t-stat significance
Intercept 0.117 0.383 0.305 0.761
Log estimatedcost -0.253 0.063 -3.971 0.000
Log forex -0.144 0.080 -1.792 0.080
Log stationsize 0.174 0.057 3.063 0.003
Extensiondummy 0.370 0.193 1.913 0.062
Log hydraulichead 0.110 0.033 3.265 0.002
Log financingagencies 0.336 0.092 3.654 0.000
Colombiadummy -0.364 0.143 -2.538 0.014

R-squared 0.472 Mean of dependent 0.216


SER 0.216 SD of dependent 0.276
LinearizedR-squared 0.499 LinearizedSER 0.259
Note: Dependentvariable is the log of the hydropowerconstantcost ratio, based on 51 observationsand
using 10percentretentionrule.
SD = standarddeviation;SER = standarderror of regression.

Slightly lower correlations are obtained in the constant costs case than in the current
costs case. The multiple correlation for constant cost overruns based on all explanatory
variables is 56 percent (Annex 2), and on only the significant variables, 47.2 percent. The
two cases share seven significant variables, and the exception is that GDP deflator is
significant for only the current costs case.

20. The Las MesitasHydropowerproject is omitted from the data set becauseof its schedule slip of
160 percent. Its cost overrunwas high at 60 percent,but even includingthis one raises the averagefor the
Colombianprojectsto only 8 percent.
SignificantProjectCharacteristicsand ExternalVariables 45

Thermal Power ProjectSchedules


The resultsof the analysisfor thermalpower projectschedulesare given in Table 6.5.

Table 6.5 SignificantVariablesfor ThermalPower ProjectSchedules

Regression 2-tailed
Variable coefficient Standard error t-stat significance
Intercept -7.989 3.844 -2.078 0.043
Log loan approvaldate 2.335 0.930 2.511 0.015
Log forex -0.277 0.078 -3.573 0.001
Log station size -0.081 0.021 -3.802 0.000
Post-1976dummy -0.283 0.101 -2.815 0.007
Post-1970dummy -0.105 0.101 -1.834 0.073
Indiadummy -0.231 0.100 -2.309 0.025

R-squared 0.405 Mean of dependent 0.242


SER 0.180 SD of dependent 0.221
LinearizedR-squared 0.479 LinearizedSER 0.220
Note: Dependentvariableis log of the thermalpowerscheduleratio,basedon 57 observations
and using
10percentretentionrule.
forex = foreignexchange;SD = standarddeviation;SER= standard
errorof regression.

For the schedule slip of thermal power projects, which have a substantialmean and
standard deviation,the regressionon all variableshas a correlationof 58 percent (Annex
2). After dropping insignificant variables, the regression is again fairly successful,
accounting for 40 percent of the total variance. The linearized standard error of
regression of 0.22 is again about two-thirds of the standard deviation of 0.30 for the
scheduleratios of thermalpower projectsas a whole.
The positive sign for the log of the loan approvaldate indicatesthat for a given set of
values of the other variables,thermalpower projectsapprovedlater tendedto have larger
schedule slips. The interactionsbetweenthe date of the loan approval and the post-1970
and post-1976 dummies indicate a general deterioration in the trend for estimating
performancefor thermalpower constructionschedules,both before 1970 and after 1976
but also show, that around those dates, steep improvementsin performance,so that the
trend decline was from a smaller level of underestimationfrom 1977onwardthan it had
been before 1970. The amount of project costs incurredin foreign exchange,the station
46 EstimatingConstructionCosts and Schedules

size, and thermal power projects in India were all negatively correlated with schedule slip
for thermal power projects.

HydropowerProjectSchedules
The results of the analysis for hydropower projects schedules are given in Table 6.6.

Table 6.6 SignificantVariablesfor HydropowerProjectSchedules

Regression Standard 2-tailed


Variable coefficient error t-stat significance
Intercept 0.457 0.314 1.457 0.151
Log estimated schedule -0.286 0.071 -4.031 0.000
Log per capitaincome 0.057 0.024 2.406 0.019
Log % ICB -0.132 0.054 -2.456 0.017
Extensiondummy 0.502 0.114 4.396 0.000
Log hydraulichead 0.070 0.021 3.324 0.002
R-squared 0.438 Mean of dependent 0.222
SER 0.153 SD of dependent 0.197
LinearizedR-squared 0.434 LinearizedSER 0.200
Note: Dependentvariableis the log of the hydropowerscheduleratio,based on 63 observations,
and using 10percentretentionrule.
SE = standarderror; SD = standarddeviation;SER= standarderror of regression.

The regression for hydropower project schedule slip is moderately successful,


accounting for 44 percent of the total variance (compared to the regression on all
variables that had a correlation of 70 percent, Annex 2). Again, the linearized standard
error of regression of 0.20 indicates that the uncertainty is reduced by about one-third
from the standard deviation of 0.28 for the schedule ratios of hydropower projects as a
whole. The mean prediction of the ratio is equal to the mean of the ratio of actual to
estimated schedules. The regression indicates that the schedule estimates were relatively
more reliable at high levels of estimated schedule but less reliable for extension projects
and for larger hydraulic heads on new projects. Schedule slip for hydropower projects
was also positively correlated with per capita income and negatively correlated with the
proportion of the project that was subject to international competitive bidding.
SignificantProjectCharacteristicsand ExternalVariables 47

Groupingof SignificantVariables
The multivariate analysis of the performance of estimates for project construction
costs (current values) and schedules reported in the previous sections identified a total of
18 significant variables at the 90 percent confidence level for the six regression equations
that cover thermal power and hydropower projects separately. Of these variables, 7
variables are significant in one of the four regressions, 4 variables in two of the
regressions, 5 variables in three of the regressions, I variable-estimated cost-in four of
the regressions, and 1-extension dummy-was significant for five regressions. Thus,
many factors are significant, but each has a relatively small impact on overall estimating
performance.
The 18 significant variables are distributed widely among the 29 tested variables
listed in Table 4. 1. In terms of the five categories used for these variables, 2 of the
significant variables (station extension and civil works costs) are technological; 5
(estimated cost, estimated schedule, station size, unit size, and hydraulic head) relate to
project size; 4 (loan approval date, foreign costs, number of financing agencies, and
proportion of ICB) reflect procurement; 5 (local inflation, per capita national income,
MUV index, Colombia dummy, and India dummy) are country-specific; and 2 (post-1976
dummy and post-1970 dummy) fall under the category of World Bank appraisal
guidelines. Variables from the same category, however, seldom reinforce each others'
impacts because they tend to be scattered among the regression equations and in some
cases are correlated in opposing directions. These results are summarized in Table 6.7.
The presence of many variables in the regressions confirms the anticipated
relationships postulated in the selection of variables for this analysis (chapter 4). To
begin with, the basic technological distinction between thermal and hydropower projects
is strongly confirmed. But the significance of civil works only for thermal costs is
surprising, which indicates that the impact of this factor in estimates of schedules has
been generally well handled. On the other hand, the finding of a greater underestimation
of costs and schedules for extensions to hydropower projects, but a lesser underestimation
of costs for extensions to thermal power projects. indicates that extension projects often
are not as straightforward as expected.2
In the case of thermal power projects, fuel type and production technology were not
found to be significant variables in multivariate regression, although the diesel dummy
was significant in the single variate analysis for both costs and schedules, and the coal
dummy was significant for costs.

21. Since plant extensionsconstituteabout 40 percent of the projectsin the data base, it could be
argued that this subcategoryof projectsshould be analyzedseparatelyto test whetherits regressionshave
significantlydifferentcoefficientsthan thosefor the wholedatabase. However,in this case (and also in the
cases of the other dummy variables),the interceptis likelyto capture the mean impact of any change of
slope in the regressionline, so it shouldgive a goodapproximationof the impactof this variable.
48 EstimatingConstructionCosts and Schedules

Table 6.7 Groupingof SignificantVariablesfor


ProjectCostand ScheduleEstimates

Thermal power projects Hydropower projects


Cost overrun Schedule Cost overrun Schedule
Significant variable Current Constant slip Current Constant slip
Extension dummy - - + + ++
Log civil costs ratio +
Log estimated cost - - - -

Log estimated schedule + + *


Logstation size * 0 - + +++
Log unit size *+ *
Log hydraulic head n.r. n.r. n.r. ++ ++ ++
Log loan approval date * ++ + *
Log forex - - -

Log financing agencies + ++


Log % ICB . . .
Log MUV index -

Log GDPdeflator +
Log per capita income . . +
Colombiadummy n.r. n.r. n.r. - -

India dummy + + - n.r. n.r. n.r.


Post-1976 dummy - - * *
Post-1970 dummy * * -

Key: + positivecorrelationat 90 percentconfidencelevel


- negativecorrelationat 90 percent confidencelevel
++ positivecorrelationat 99 percentconfidencelevel
- negativecorrelationat 99 percentconfidencelevel
* correlationis not significantat 90 percentconfidencelevel
n.r. not relevant

The relationship between estimating performance and variables related to project size
was mixed. For power generation projects as a whole, there appears to be a tendency for
the percentage overrun to decline with the size of the estimated cost in current values.
The fact that this variable appears for both hydropower and thermal power projects
indicates that whatever leads to a failure to take into account fully the level of estimated
SignificantProjectCharacteristicsand ExternalVariables 49

costs is common to both types of projects. On the other hand, estimated schedule was
positively correlated with thermal cost estimation performance and negatively correlated
with hydro schedule estimating performance. Similarly, station size was positively
correlated with the estimating performance for hydro costs but negatively correlated with
the estimating performance of thermal schedules. Hydraulic head had a strong positive
correlation with the performance of both hydro costs and hydro schedules.2 2
One particularly striking feature of the analysis is that many of the variables listed in
Table 4.1 are not correlated significantly with the cost or schedule ratios, either singly or
in combination with other variables in the multiple regression context. In the case of
hydropower projects, for example, the dam height and tunnel length for new projects have
single squared correlations of under 2 percent with the cost ratio and schedule ratio (a
value of 5 percent would be needed to indicate significance for a single variable), and
they are never significant in the multiple regression context. This suggests that the
factors associated with those variables that affect costs and schedules have been correctly
allowed for in constructing the estimates, but that there is some aspect caught by the size
of the hydraulic head that was not fully allowed for in either the cost or the schedule
estimates.23
Procurement and country variables had mixed significance. Estimates for
hydropower projects seem to be more sensitive than those for thermal power projects to
local income levels (for schedules) and to foreign inflation rates (for costs). The
Colombian hydropower projects tended to have lower costs overruns than generally after
allowing for the impacts of other significant variables on estimating performance;
likewise, Indian thermal power projects tended to have higher cost overruns but lower
schedule slips.
World Bank appraisal guidelines for computing price contingencies do not appear to
have been a significant factor for cost-estimating performance in current terms. The 1976
guideline was found to be significant in the case of thermal costs in constant terms, but
this could reflect a technicality because it was used to derive the actual costs in constant
terms for the analysis (Annex 6). The significance of both guidelines (1970 and 1976) on
thermal project schedule estimating performance is discussed in the section on thermal
project schedules above.

22. The use of explanatoryvariablesthat reflectproject cost (schedule)with a dependentvariablethat


is the ratio of actualto estimatedcosts (schedules)can introducea problemof spuriouscorrelation,such as
that residualsare heteroskedastic. In fact the use of actual costs (schedules)as a dependentvariable with
estimatedcosts (schedules)as an explanatoryvariableexhibitsvery strong heteroskedasticity(see Annex 7,
section 1). Using the ratio, however, produces homoskedastic residuals.
23. This observation might be related to the finding that the Colombia dummy is significant for
hydropower cost estimations performance, but the Brazil dummy is not significant. since the hydraulic
heads of the Colombian projects are among the highest, and those of Brazilian projects are among the
lowest, in the group of hydropower projects.
50 EstimatingConstructionCosts and Schedules

Finally, the difference in conclusions about significant variables between the


multivariate analysis and a single variate analysis are considerable, as shown in Annex 4.
In the four regressions, single-variate analysis predicted only 11 of the 24 variables that
were significant in the multivariate analysis at the 90 percent confidence level. None of
the 6 significant variables that occur in two or three regressions in multivariate analysis
were found significant in two or more regressions in single-variate analysis. The latter
also falsely indicated 19 variables as significant.
7
Implicationsof the Analysisfor
Power System Planning
The paper has focused so far on analyzing the performance of past estimates of
construction costs and schedules for World Bank-financed power generation projects in
developing countries. This concluding section examines the implications of the analysis
for dealing with future projects. Estimates of construction costs and schedules are, of
course, two of many parametric estimates and forecasts that characterize risk and
uncertainty in choosing power generation projects (see Sanghvi and Vernstrom 1989 and
the discussion in chapter 2 above, footnote 5). Most published analyses on this subject
tend to focus on uncertainty in predicting future power demand, fuel prices, and cost of
capital. This chapter thus fills a significant gap in the treatment of risk and uncertainty in
this class of projects.
The analysis identified two problems with historical estimates that need to be
recognized in future project evaluation. First, for the group of projects as a whole (which
covered virtually all Bank lending for completed generation projects since the mid-
1960s), there has been a very large downward bias in the estimation of costs and in the
estimation of schedules. Clearly it is necessary to recognize that standard methods of
estimating costs and schedules tend to be overoptimistic. If all projects had experienced
equal or nearly equal bias in the estimation of costs and schedules, then the remedy would
be simple: multiply the standard estimates by an "adjustment factor." Provided projects
in the future were similar to those financed during the period of analysis, and provided
that the methods of estimating costs and schedules were not changed substantially, then
this simple adjustment would, on average, produce accurate predictions for the actual
outcomes.
In fact, such a simple adjustment rule is unlikely to be optimal because of the second
feature of estimated costs and schedules identified in this paper in that, even allowing for
the average underestimation, there is very large variation around this value. The average
shortfall in cost estimates for all projects was 21 percent, but some projects came in on
budget, whereas others had cost overruns of 40 percent or more. This variation in the
accuracy of estimation presents a more substantial problem for project evaluation.

51
52 Estimating Construction Costs and Schedules

This paper has shown how the apparent variability of cost and schedule estimation
can be reduced by categorizing the projects by type, through regression analysis. It has
also revealed that despite an intensive search for indicator factors that are known ex ante
and that correlate strongly with actual cost and schedule overruns, there appears to be a
substantial unpredictable element in these estimates.
Once all predictable aspects of estimating performance for costs and schedules have
been taken into account, there will be no overall bias in the predictions for any
identifiable type of project. The residual inaccuracy can then be seen as the risk
associated with costs and schedules rather than a systematic element in expected costs or
schedules for a particular group of projects.
The first section below recapitulates the principal findings of the paper. The next
section then indicates how the regressions developed in the paper can be utilized to obtain
more accurate predictions of project construction costs and schedules. The last section
addresses the question of risk in predicting these costs and schedules. Various
approaches to incorporating risks, as estimated from the regressions, into project
decisionmaking are outlined. The chapter concludes by dealing with special issues that
arise in choosing between power development programs that involve more than one
project.

PrincipalFindings
A number of important implications for the planning of power generation projects in
developing countries emerge from the analysis of the actual performance of construction
cost and schedule estimates for such projects:
a. Estimates were fairly strong correlated with the actual outcomes, but the average
error among projects as a whole was too large to be ignored. The group of power
projects as a whole has a (squared) correlation of 0.76 between actual and estimated
costs (in current prices) and 0.55 between actual and estimated schedules. The
modest values of these correlations indicates that the estimation process is rather
inexact and that large errors have often been made. This is the first justification for
seeking a method of improving the predictions of actual costs and schedules.
b. The estimated values were significantly biased below actual values. The second
crucial finding is that the estimated costs and schedules were, on average, severely
biased downward. Estimated construction costs for projects as a whole were on
average 21 percent below actual costs, whereas estimated schedules were on average
36 percent below actual values. Even when cases with exceptionally large overruns
are omitted from the group of projects, estimated costs still averaged 17 percent
below actual costs, whereas estimated schedules were an average 29 percent below
actual values. Both of these findings indicate that, on average, an overoptimistic view
of project costs and schedule performance was taken.
Implicationsof the Analysisfor PowerSystem Planning 53

C. The performance of estimated values differed sharply between different types of


generating plants. The most striking difference found was between thermal and
hydropower projects. The average cost underestimation for thermal projects was only
6 percent, whereas for hydro projects the average was 27 percent. Schedule estimates
were rather similar, with thermal projects showing a 30 percent average
underestimation and hydro projects on a 28 percent average underestimation. These
results indicate that in attempting to make any adjustment to the estimates of costs
and schedules, it would certainly be necessary to treat the costs of thermal projects
differently from those of hydro projects.
d. The performance of estimated values can be related to a number of indicator
variables through regression analysis. Regression analysis, based on a large set of
possible variables, indicated that 18 project variables had a significant relationship (at
the 90 percent confidence level) in one or more of the six regression equations
analyzed (for thermal and hydro project costs-current and constant-and schedules).
Of the variables, 2 (station extension, and civil works cost) are technological; 5
(estimated cost, estimated schedule, station size, unit size, and hydraulic head) relate
to project size; 5 (local inflation, per capita national income, MUV index, Colombia
dummy, and India dummy) are country specific; and 2 (post-1976 dummy, post
1970-dummy) relate to World Bank guidelines. The residual variances from the
regressions can be used to provide a measure of risk.
These variables were used in the analysis of the ratio of actual to estimated costs and
the ratio of actual to estimated schedules and were able to explain between 48 percent and
57 percent of the variation in these ratios. In other words, about half the variation in these
percentage errors that would be left unexplained by using an average correction factor for
all projects is explicable in terms of a few simple variables whose values were known at
the time the project loan was approved. In addition, the significance of individual
variables indicates that for a class of projects that took a particular value of the variable,
the use of an "average adjustment factor" (determined over all projects) would itself be
biased.

Using Regressionsto ImprovePredictionsfor ProjectCosts and Schedules


The size of the bias in estimating project costs and schedules indicates that the
simplest way to make some improvement in predicting actual values would be to attach
average correction factors to such estimates. If, for the set of 135 projects analyzed, cost
estimates had been adjusted upward by a factor of 1.21 and schedule estimates by a factor
of 1.36, then overall the adjusted estimated value would have been unbiased (i.e., the
average ratio of predicted to actual values would have been unity). However, for
individual subgroups of projects there would still have been a bias in the predicted values.
Regression analysis also reveals that about half the variation in the differences among the
ratios of actual to adjusted values is capable of being predicted and hence should not be
treated as if it were a residual "risk."
54 EstimatingConstructionCostsand Schedules

Using the regressions and the known values of the relevant associated variables,
predictions can be made of the ratio of actual to estimated costs, and hence an improved
prediction of actual costs can be obtained. The approach is illustrated for the India 911
thermal project. For this project, the actual costs turned out to be $491.1 million, whereas
the estimate of costs used for appraisal was $405.9 million. The project cost was thus
underestimated by 21 percent. If it had been agreed to use the overall adjustment factor
for thermal power project costs, then the estimated value would have been increased by 6
percent, still leaving a 15 percent underestimate. On the other hand, using the regression,
as described in the first section of chapter 6, to produce a predicted cost would have given
a predicted ratio of actual costs to estimated costs, for this class of project, of 1.207.
Applying that factor to the estimated cost of $405.9 million yields a prediction of just
under $490 million, very close to the true value.
It is therefore recommended that the analysis of power generation projects includes a
case in which expected values are used for construction costs and schedules. These
expected values can be derived from estimated values by applying ratios obtained from
the regression equations given in chapter 6 in the manner illustrated above for the India
911 project. This case supplements the standard analysis that is based on appraised
estimates of costs and schedules.
The regression approach is designed to give the most accurate statistical prediction of
actual costs and schedules based on historic experience. Two very important limitations
must be noted. The very best regressions explain less than 60 percent of the variance of
the ratio of actual to estimated values, with the remaining variation attributable to
unidentified factors. If no substantial improvement in this goodness of fit can be found,
then indeed a substantial element of genuine risk is present in appraising the construction
costs and schedules of power generation projects. This implies that it is important that a
coherent view is taken of the treatment of risk in project appraisal. Although 60 percent
correlation is a low value relative to those obtained in many time series econometric
studies, it is important to remember that the present study is of a cross-section type
(where correlations are typically lower than in time series) and that the dependent variable
is the ratio of the actual value to the estimated value rather than simply the actual value.
Since the estimated value itself already encapsulates much specialist knowledge of the
particular project, the regression is in effect seeking for influences on costs or schedules
that have been systematically over- or undervalued by project appraisers. Remembering
this interpretation, it is not likely that substantially higher correlations can be found (i.e.,
what is unexplained is indeed largely risk). Regression analysis cannot be expected to
provide a perfect solution to the problem of forecasting actual values.
The second aspect of the use of regressions for prediction is that a relation estimated
on one set of data to predict outcomes for new data will give unbiased predictions only if
the same relationship continues to hold between the indicator variables and the ratio of
actual to estimated values as in the historic sample. If the relationship changes-because
a new type of plant (e.g., combined cycle) is being considered, where the indicator factors
Implicationsof the Analysisfor PowerSystemPlanning 55

affect cost and schedule overruns to a different degree than in the sample (e.g., diesels);
or because of new ownership and contractual arrangements introduced, such as projects
being undertaken by the private sector rather than by the past practice of the public sector;
or because the accuracy of estimation itself changes-then bias can result from using the
regressions. However, in such cases it is still likely to be more accurate to use the
regression than to make no adjustments to estimated values, unless huge differences
between the past and present relationships are expected. The analysis of the impact of the
dummy variables for changes in World Bank estimating procedures on the accuracy of the
estimates tends to confirm this view. For four of the six equations the dummy variables
were insignificant, suggesting that changes in World Bank procedures in 1970 and 1976
were not associated with a significant improvement in estimation. Only for thermal
schedules was there a measurable improvement in the accuracy of estimation. It thus
seems reasonable to assume that it is unlikely that dramatic improvements in the accuracy
of estimating costs and schedules will occur. However, where new technologies are
involved there can be less confidence that the estimation errors will be similar in nature to
those identified in the sample.
The regressions also throw some light on the nature of the systematic underestimation
of costs and schedules. Each significant variable in a regression points to a factor whose
impact on the estimate was undervalued (positive sign) or overvalued (negative sign), and
the size of the coefficients indicates the magnitude of such effects. For example, in the
case of thermal project cost estimates (Table 6.1), the larger the estimated cost of the
project itself, and for projects that were extensions of existing projects, the lower was the
tendency to underestimate costs. Alternatively, the longer the estimated construction
schedule, or the higher the ratio of civil to total costs, and if the project was in India, the
greater was the tendency to underestimate costs. Although the construction of the cost
estimates took into account the scale of the project, they tended to underplay economies
of scale to costs themselves but not to make enough allowance for the fact that projects
with a lengthy construction period would tend to have higher costs than projects of
similar physical size with shorter estimated construction periods. Not enough allowance
was made for the tendency for costs on thermal power projects in India to be higher than
for similar projects elsewhere.
The factors summarized in Table 6.7 highlight the significant variables. For any
project where a variable is known to have a large value (e.g., a large hydraulic head for a
hydropower project), it is sensible to be alert to the tendency to misestimate such projects.
Where the project is only marginally viable, this can indicate the need for extensive
sensitivity analysis to alternative cost scenarios or even for design modifications that
reduce the influence of the problematic variable.
The magnitude of the coefficients allows a quantification of the likely sensitivity of
the estimates to the presence of the significant factor. Since the equations are estimated
in double log form, the coefficients on the logs of the indicator variables measure
elasticities-for a I percentage point increase in the size of the variable, the coefficient
56 EstimatingConstructionCostsand Schedules

measures the percentage increase in the predicted actual cost levels for given values of the
other indicator variables. Where the indicators are dummy (zero/one) variables, for
which logs cannot be used, the exponent of the coefficient measures the percentage
impact on the level of predicted costs of the presence of the variable. Table 7.1 shows the
magnitudes of the elasticities for the factors found significant in the six regressions.

Table 7.1 Sensitivityof the Levelsof PredictedValuesto IndicatorVariables


Thermnal costs Thermal Hydro costs Hydro
Variable Current Constant Schedules Current Constant schedules
Extensiondummy* 0.856 0.882 1.433 1.448 1.652
Civil costratio 0.083
Estimated costa 0.854 0.726 0.807 0.747
Estimated schedule 0.287 0.174 0.714
Station size -0.081 0.117 0.174
Unit size 0.115
Hydraulichead 0.110 0.110 0.070
Loan approvaldate 2.626 2.335
Forex -0.277 -0.224 -0.144
Financingagencies 0.234 0.336
MUV index -0.081
% ICB -0.132
GDPdeflator 0.076
National income 0.057
Colombia dummy* 0.655 0.695
India dummy* 1.223 1.186 0.794
Post-1976 dummy* 0.768 0.753
Post- 1970 dummy 0.900
Note: Percentage changes of predicted values given a I percent change in a continuous variable and
the presence* versus absence of a dummy variable. Consider the stylized fitted regression
log (A*/E) = + M D + ylogX (X)
where A* is the predicted level of costs (or schedules) from regressions, E is the estimate of costs,
D is a dummy variable (value I when the certain factor is present and zero when absent), and X is the level
of some indicator variable. Greek letters denote estimated regression coefficients. Taking antilogs and
expressing the equation in terms of the desired variable (A*) yields:
A* = E exp(a) exp(OD) X. (2)
The elasticity of A* with respect to X is given by XoA/A6X and is equal to y. The impact on A*,
for a given value of X and E, of the presence (versus absence) of the dummy variable factor is measured by
the ratio exp(p)/ exp(0) = exp(1). A value of less than unity indicates that the presence of the factor lowers
the predicted outcome. The elasticity of the predicted value (A*) with respect to the estimated value (E) is
unity, except in the case where the variable X is itself the estimated value. In the latter case the elasticity is
(I + y), which measures the "scale" effect between E and A*.
'Where the estimated value is the explanatory variable, the sensitivity coefficient (elasticity) is 1+
regression coefficient (see the note above).
Implicationsof the Analysisfor PowerSystem Planning 57

For the dummy variables, the impact of their presence is proportional solely to the
coefficients given above. It can be seen that the most important factor leading to an
upward adjustment in the prediction is the impact on hydro schedules of the project being
an extension. Here the estimated value, adjusted by the other factors, needs to be
multiplied by a factor of exp(0.502), that is, by 1.652. The largest downward adjustment
is to estimated costs for hvdro projects in Colombia, after allowing for effects of other
variables, which need to be multiplied by exp(-0.423), that is, by 0.655.
For variables where elasticities are available, the importance of a factor depends not
only on the coefficient but on the scale. The equation expressed in form (2) in the note to
Table 7.1 shows clearly the nonlinear nature of the relationship. To determine which
factor is the most important, it is necessary to consider how much such a variable might
change between projects. Where a factor exhibits high percentage changes between
projects, such a factor will be more important (for a given value of the elasticity) in
explaining variations in the predicted values between projects. Hence, both the
coefficient and the ratio of the indicator values are important.

Risk and PlanningIssues


Besides showing the presence of systematic errors in estimating performance, the
regression analysis has revealed that it is not possible to make perfect predictions of the
outcomes of project costs and construction schedule from the factors used in the analysis
and that the degree of uncertainty is large. The standard deviation of the ratio of actual to
estimated costs was 36 percent, and for the ratio of actual to estimated schedules it was
42 percent. In predicting the ratio of actual to estimated values for projects categorized
through the regression analysis, the estimated standard deviations of the errors varied
between 16.9 percent for thermal costs (current) and 26.7 percent for thermal schedules.
Insofar as the regression models have captured all systematic variation, these magnitudes
can be regarded as measures of the risk associated with costs and schedules. They also
allow expressions to be constructed for the probability distribution of outcomes for a
particular project.
Such expressions can be adapted for use in proprietary power system planning models
that specifically analyze financial and economic risk for a sequence of investments in
long-term expansion programs for power systems.25 Where risk is to be explicitly

24. Comparing two projects, alike in all respects (including the estimated values) apart from indicator
value X, the ratio of predicted values will be from equation 2 in the note to Table 7. 1: Al/A2 = (X1/X2 )'-
25. Some planning models treat key operating variables-such as power demand, fuel prices, and
hydrology-stochastically to derive a distribution of possible outcomes for the net present value of a
selected power development program. employing such techniques as Monte Carlo simulations. Stochastic
representation of construction costs and schedules could conceivably be added to such models. The
stochastic outcomes of alternative programs could then be compared by a straightforward statistical
technique (see the subsection below on planning issues involving choices between sequences of projects).
58 EstimatingConstructionCostsand Schedules

factored into project analysis, it is possible either to use a discount factor appropriate to
the degree of risk involved or to use a riskless discount factor and add in adjustments for
the risks involved to the benefit/cost stream such as those derived in this paper.
The assumption that similar factors will affect the outcomes of costs and schedules
for future projects as they affected the projects in the sample allows the variance of errors
from the regression equation to provide a measure of risk for future projects (i.e., a
quantifiable probability distribution of outcomes). There will of course be genuine
uncertainty-eventualities for which no probability can be constructed from historic
experience-but for planning purposes it is valuable to have measures of risk available.
Traditional investment analysis, where the outcomes are risky and investors are risk
averse, tends to focus on the explicit tradeoff between mean return (or cost) and the
variance. Investors choose the highest available "equal value" contour between means
and variances that reflects their underlying attitude toward risk. (In effect, this technique
requires the introduction of an extra explicit evaluation criterion that allows the investor's
attitude toward risk to be encapsulated in the analysis.) This technique is feasible where
the composition of numerous investments in a portfolio can be continuously varied to
achieve the desired risk profile. It is less directly applicable to the selection of power
generation projects, where only a few alternatives are to be considered and where the
decision is faced only periodically and concerns a relatively large investment.
Nevertheless, the measures of mean cost and variance of costs (or schedules) derived in
this paper can be used to examine important planning issues for power generation
projects, as illustrated in the following subsections.

Measurement of Project Risk


If the project is such that risk is to be taken into account in the decision making
process, then the regressions provide a measure of project risk arising from variability in
construction cost and schedule estimates. This measure is the standard error of the
regression. Standard analysis of publicly financed projects under uncertainty (Arrow and
Lind 1994) shows that if returns from the investment are independent of other
components of national income, the government should choose the project that
maximizes the expected return when using a discount rate appropriate for investments
with certain returns. That is, because project risk can be spread thinly over a large
taxpaying population, the government should ignore uncertainty in evaluating public
projects.
Where the share of the risk borne by individual taxpayers is substantial relative to
their income, this conclusion has to be modified. Similarly, if some of the risk accrues
directly to individuals, this should be discounted for, just as if it were a private
Implicationsof the Analysisfor PowerSystemPlanning 59

investment:.26 Where the return of the project is likely to be correlated with national
income, Little and Mirrlees (1974) provide an adjusted criterion for determining whether
or not to undertake the project. This criterion also requires the evaluation of the risk of
the project and the correlation between the project return and the level of national income.

Distribution of Possible Project Outcomes


The regressions provide estimates of the distribution of the possible outcomes due to
uncertainA, in construction costs and schedules, which can be used in risk analysis. In
deciding whether to go ahead with a given project (or not), or which of two mutually
exclusive projects to select, the basic criterion focuses on the expected value of returns,
and hence on the expected values of costs and schedules. Nevertheless, in cases where
the financing of a major cost overrun or an overrun beyond a specific limit to available
funds would impose an unacceptable budgetary strain, it is important to have a high level
of confidence that the predicted value has been truly identified by the regression with the
best performance (i.e., the one least likely to have omitted any systematic factor).
However, it is recognized that the actual cost is a random variable as seen from the ex
ante standpoint, whose mean is the predicted value from the regression. Moreover, the
variance of this random variable can be estimated from the regression. If it is assumed
that the variable used in the regression (the log of the ratio of actual to estimated values)
is normally distributed, then the distribution of outcomes can be computed. With the aid
of this distribution it is possible to calculate information that can be used as a supplement
to the decision making process.
Examples of such calculations are described in Annexes 8 and 9 for the following
questions:
a. Finding the cost level for a given project that will be exceeded with a specific
probability and the probability of the project cost exceeding a specified value
(similarly with project schedule).
b. Choosing between projects on the basis of the probability of exceeding a cost limit,
evaluating the probability of exceedance that has the same cutoff value for the
projects. and evaluating which project has the lower probability of exceeding a given
cost limit.
c. Assigning probabilities for risk analysis to a low scenario, an expected scenario, and a
high scenario, where values are assigned to these scenarios so that these values and
probabilities are consistent with a prior view of the variance of cost or schedule
outcome, such as obtained from one of the regression equations derived in this paper.

26. This distributionalissue mayalso apply to the low-incomegroupsthat do not benefitdirectly from
the project (i.e., lack access to the public electricitysupply system)but that would be adverselyaffectedby
reducedgovernmentexpenditureson social servicescausedby a cost overrunfor a power project.
60 EstimatingConstructionCosts and Schedules

PlanningIssuesInvolvingChoicebetweenSequencesof Projects
Analysis of a program for power system expansion can involve choices that require
more selections than an alternative between single projects. Three cases that are of
considerable practical importance are frequently encountered:
a. The choice between one large project and a set of two or more smaller projects, where
the reliability of estimates of costs and schedules depends on the scale variables
identified in the regressions, and risks also depend on project size.
b. The choice between two power expansion programs, each made up of a sequence of
several projects, where the crucial issue is fuel diversity (e.g., between a hydropower-
dominated program and a mixed hydro/thermal power program).
c. The choice of whether or not to delay a project, where the crucial issue is often that
of improving the estimates for construction costs and schedules for a project or its
alternatives by providing more time for further investigations into such aspects as
project site conditions (e.g., geology, topography, and environmental impacts).27
The first two cases raise the technical issue of how to construct the prediction of the
total program cost and its variance from values for the individual projects derived from
the regression analysis. Annex 10 shows how the mean and variance of a sum of the cost
of two projects are derived, as well as how to obtain the mean and variance for the
predicted cost (or schedule) when the regression model provides a predicted value for the
log of the ratio of actual to estimated values.
The formulas in Annex 10 can be manipulated to explore the possibility that a series
of smaller projects with approximately the same total cost and output as a single large
project may have a rather lower variance of costs and hence be less risky. Where the risk
of the project is an important factor, a strategy of risk reduction through the use of a series
of smaller component projects may be an important planning option.
The decision to delay a project to obtain better information on costs and schedules can
be taken in the context of the types of risk analysis outlined in Annexes 8 and 9.28 If the

27. Delayinga project is sometimesproposedunderconcernsabout uncertaintyof futurefuel costs and


power marketdemand. An interestingcase is the use of financialoptionsanalyticalmethodsfor a power
generationproject whose output is sold under competitivebidding, wherethe main concernis uncertainty
about the power sales price. See EPRI (1995). Anotherinterestingapplicationof this approachis to the
case of extendinga transmissionsystem,for whichsee Martzoukosand Teplitz-Sembitzky(1992).
28. In additionto the uncertaintyaboutthe performanceof estimatesat projectappraisalthat are based
on feasibilitystudies,engineeringdesignwork,and contractbidding,as reportedin this paper,this issue can
also arise in the context of preappraisaldecisionsbecause it is generallypresumedthat the reliabilityof
estimates of construction costs and schedules improves as a project progresses through successive
preparation stages-identification, prefeasibility,feasibility, engineeringdesign, and contract bidding.
Intuitively this presumption appears realistic when just considering bias in the estimates. However, the
substantial variance found for the performance of appraisal estimates should caution against overconfidence
in obtaining better estimates by delaying projects to obtain more information about them.
Implicationsof the Analysisfor PowerSystemPlanning 61

project appears to have too high a risk or an excessive cost or construction schedule, then
it can be sensible to consider delaying the project while more evidence on estimated costs
and schedule is collected about the project or alternatives. The delay option can then be
evaluated in terms of the change in expected costs and schedules from present estimates
by means of decision-tree analysis or by adapting the financial options approach. An
application of the options approach to the optimal timing of projects under uncertainty
about construction costs and schedules is developed in Annex 11. Unless there is specific
evidence to the contrary, it is probably sensible to assume that there will be no change in
the variance of estimates by delaying a project.
Basic Recommendations
Two basic recommendations for operational analysis emerge from the analysis of
estimates for construction costs and schedules of power generation projects in developing
countries. First, because methods of estimating costs and schedules have been
overoptimistic, the robustness of the analysis should be tested by applying a correction to
the estimates of costs and schedules. An "expected" value should be used for this test,
and it can be calculated using the appropriate regression once the features of the project
have been identified.
Second, because the regression analysis shows that the uncertainty in predicting costs
and schedules is also too large to ignore, even when expected values are used, it is also
recommended that the economic and financial risks associated with the selection of a
particular power project or power development strategy are explicitly considered during
project appraisal. A measure of risk is provided by the product of the estimated value and
the standard error of the regression equation, as shown in Annex 10.
Risk should be considered in the context of specific questions rather than in an
abstract context. There are some basic questions that should be examined to elicit
valuable insights about the riskiness of power generation projects that are considered to
be the least-cost option from the customary deterministic approach to power system
planning. The paper proposes straightforward techniques for analyzing some of these
questions (see the section above on risk and planning issues, as well as Annexes 8, 9, and
10). These techniques should be tested operationally and developed in case studies, so
that guidelines can be formulated for using them in the appraisal of power generation
projects.
Annex1:WorldBank-SupportedPower
GenerationProjects1965-86 Usedfor the
Analysisof Costand ScheduleEstimating
Performance
Table A1.1 World Bank-SupportedPowerGenerationProjects1965-86 Usedfor
the Analysisof Cost and ScheduleEstimatingPerformance

Installed Loan
capacity approval
Country Project name (MW) year
Thermalpowerprojects
Afghanistan (Power I) Kabul 40 1976
Algeria Algiers 98 1974
Bangladesh Ashunganj Thermal 450 1982
Botswana Morupule 90 1983
Costa Rica (Power IV) San Antonia 38 1972
Costa Rica (Fifth Power Project) Moin 30 1974
Cyprus (Power III) Moni Station Unit No. 4 30 1969
Cyprus (Power IV) Moni Station Unit No. 6 60 1972
Ecuador (Third Power Project) Cumbaya 18 1971
Egypt (Power II & III) Shourbrah El Kheiam 900 1979
Guatemala Guacalate 99 1968
Guyana (G&T Project) Garden of Eden/Rotterdam 36 1972
Haiti (Power I) Varreux 21 1976
Haiti (Second Power Project) Varreux 25 1979
Haiti (Third Power Project) Carrefur 16 1982
Honduras Nispero Power 30 1965
Honduras (Fifth Power Project) La Ceiba 24 1972
India Second Kothagudem 120 1978
India Singrauli 600 1977
India Third Trombay 500 1978
India Korba 600 1978
India Ramagundam 600 1979
India Second Singrauli 1400 1980
India Farakka 600 1980
India Second Ramagundam 1500 1982

(continued on next page)

63
64 Estimating Construction Costs and Schedules

(Table Al. I continued)

Installed Loan
capacity approval
Country Project name (MW) year
Thermalpowerprojects(continued)
India Second Korba 1500 1982
India Fourth Trombay 500 1984
India Comb. Cycle: Kawas, Anta, & Auraiya 1500 1986
Indonesia Power IV: Muara Karang No. 3 100 1975
Indonesia Power VI: Muara Karang 400 1977
Indonesia Power VII: Semarang Harbor 200 1978
Indonesia Power VIII & IX: Suralaya Units I & 2 800 1979
Indonesia Power XII & XIV: Suralaya Units 3 & 4 800 1984
Ireland Power II: Tarbert 250 1971
Ireland Power III: Tarbert 250 1972
Jordan Power I: Zarqa 78 1973
Jordan Second Hussein Thermal 33 1975
Jordan Power V: Aqaba Power Station 260 1982
Korea Gojeong Power 1000 1979
Malaysia Port Dickinson & Johore Bahru Thermal 180 1966
Malaysia Power IV: Prai & Port Dickinson Thermal 150 1969
Malaysia Power V: Port Dickinson Thermal 360 1970
Malaysia Power VII: Prai Thermal Extension 360 1975
Malaysia Power VIII: Pasi Gudang 240 1977
Pakistan Karachi 'C' Thermal Power Station 125 1967
Panama Power III: San Francisco Thermal 25 1973
Philippines Power IV: Bataan Thermal Plant 75 1967
Philippines Power V: Bataan Thermal Electric No. 2 150 1972
Romania First Turceni Thermal 1320 1974
Romania Second Turceni Thermal 330 1979
Sierra Leone Power II: King Tom Thermal 7 1968
Sierra Leone Third Power Project: King Tom Thermal 9 1977
Sri Lanka Power VIII: Sapugaskanda 80 1982
Sudan Sudan Power II: Burri & Juba Thermal 15 1975
Syria First Mehardeh Power 150 1974
Syria Second Mehardeh Power 150 1975
Thailand South Bangkok Thermal No. 3 310 1970
Thailand Bang Pakong Thermal Power 550 1970
Annex 1: World Bank-Supported Power Generation Projects 65

Installed Loan
capacity approval
Country Project name (MW) year
Thermalpowerprojects(continued)
Thailand South Bangkok Thermal No. 4 310 1971
Turkey Elbistan 1200 1974
Uruguay Power IV: Battle Unit No. 6 125 1971
Yemen Wadi Hadramout Power Project 16 1978
Yemen Power II: Wadi Hadramawt Thermal 7 1982
Zimbabwe Power I: Hwange II 400 1983
Hydropowerprojects
Argentina El Chocon 600 1968
Bolivia Second Empresa Nacional de Electricidad 34 1969
Brazil Estreito 1050 1964
Brazil Xavantes 400 1965
Brazil Volta Grande 400 1967
Brazil Porto Colombia 320 1968
Brazil Marimbondo 1400 1969
Brazil Salto Osorio 700 1970
Brazil Sao Simao 1608 1971
Brazil Paulo Afonso IV 2460 1974
China Lubuge 600 1984
Colombia Third Medellin 280 1964
Colombia El Colegio & Conoas 200 1966
Colombia Chivor 500 1969
Colombia Guatape Second 280 1972
Colombia First San Carlos 620 1978
Colombia Las Mesitas 600 1978
Colombia Second San Carlos 620 1979
Colombia Guadalupe IV 213 1980
Colombia Playas 204 1981
Costa Rica Fifth Power: Rio Macho & Cachi 62 1974
Ecuador (Third Power Project) Nayon 30 1971
Ethiopia Finchaa 99 1968
Fiji Manasow-Wailou 40 1977
Ghana Second Volta River Authority 324 1968
Ghana Kpong Hydroelectric 160 1976

(continued on next page)


66 Estimating Construction Costs and Schedules

(Table Al. I continued)

Installed Loan
capaciry approval
Country Project name (MW) year
Hydropowerprojects(continued)
Guatemala Aguacapa 90 1977
Guatemala Chixoy 300 1977
Honduras Fifth & Sixth Power: Rio Lindo 40 1973
Honduras Nispero Power 22 1977
Honduras El Cajon 292 1980
Iceland Sigalda 100 1972
Indonesia Power X: Saguling 700 1981
Ireland Pumped Storage 292 1968
Kenya Kamburu 60 1970
Kenya Gitaru 134 1974
Kenya Kiambere 140 1984
Lao-PDR Nam Ngum 40 1981
Madagascar Andekaleka 56 1978
Malawi Tedzani Sate I 16 1969
Malawi Nkula Falls II 54 1976
Malaysia Power IX: Bersia & Kenering 192 1980
Morocco Sidi Cheho-AI Massira 120 1976
Myanmar Kinda (Nyaunggyat Multipurpose) 56 1980
Nepal Kulekhani 60 1974
Panama (Second Power) Bayano 150 1970
Panama Fortuna 300 1977
Papua New Guinea Upper Ramu 45 1970
Peru Matucana Power 120 1966
Portugal (Power Project VII) 8 Hydro Plants 1495 1983
Romania Riui Mare-Retezat 349 1975
Sudan Roseires 90 1966
Sudan (Second Power) Roseires Extension 42 1974
Sudan Roseires Extension 80 1980
Swaziland Third Power: Lupohlo-Ezulwini 20 1981
Tanzania Kidatu Hydroelectric Stage I 100 1969
Tanzania Kidatu Hydroelectric Stage II 200 1975
Tanzania Power IV: Mtera 80 1983
Thailand Ban Choa Nen Srinagarind 450 1973
Annex 1: World Bank-Supported Power Generation Projects 67

Installed Loan
capacitv approval
Countrv Project name (MW) year
Hydropower projects (continued)
Thailand Pattan I 72 1977
Thailand Khao Laem 300 1979
Thailand (Power Subsector Project) Lan Suan and Chiewlarn 240 1981
Turkey Third & Fourth Cukurova Power 56 1964
Turkey Karakoya 1800 1980
Turkey Sir 284 1986
Yugoslavia Middle Neretva Hydro: Grabovica and Salakovac Dams 322 1978
Yugoslavia Middle Neretva Hydro: Mostar Dam 65 1978
Yugoslavia Visegrad 315 1985
Zaire Ruzizi II 26 1984
Zambia Kariba North 600 1970
Zambia Kafue Hydroelectric Stage II 300 1972
Annex 2: RegressionResultswith
All VariablesIncluded
Table A2.1 Variablesfor Logof ThermalPower ProjectCosts (CurrentValues)
Basedon 42 Observations

Regression Standard 2-tailed


Variable coefficient error t-stat. significance
Intercept -18.756 6.647 -2.821 0.011
Log loan date 4.676 1.617 2.890 0.009
1970 dummy -0.186 0.157 -1.185 0.250
Log estimated cost -0.302 0.122 -2.463 0.023
Log estimated schedule 0.214 0.138 1.544 0.139
Log forex -0.097 0.120 -0.806 0.430
Log per capita income -0.015 0.050 -0.297 0.769
Log station size 0.072 0.147 0.487 0.632
Log % ICB 0.054 0.080 0.673 0.508
Log MUV growth -0.050 0.060 -0.826 0.419
Log GDP deflator -0.040 0.056 -0.713 0.484
Log unit size 0.064 0.105 0.617 0.544
Basis for costs dummy 0.015 0.091 0.167 0.869
Diesel dummy 0.241 0.246 0.977 0.340
Coaldummy -0.019 0.118 -0.165 0.871
Steam dummy 0.079 0.097 0.815 0.425
Extension dummy -0.177 0.089 -1.987 0.061
Log sales growth -0.046 0.039 -1.183 0.251
Log agencies -0.022 0.103 -0.216 0.831
Contractor dummy 0.030 0.137 0.223 0.825
1976 dummy -0.300 0.151 -1.982 0.062
India dummy 0.186 0.137 1.353 0.192
Log estimated civil 0.100 0.037 2.638 0.016
R-squared 0.803 Mean of dependent 0.024
SER 0.159 SD of dependent 0.244
Linearized R-squared 0.798

69
70 Estimating Construction Costs and Schedules

Table A2.2 Variablesfor Log of ThermalPowerProjectCosts (ConstantValues)


Basedon 55 Observations

Regression Standard 2-tailed


Variable coefficient error t-stat. significance
Intercept -16.286 5.201 -3.130 0.003
Log of loan date 4.052 1.270 3.189 0.003
Dummy for 1970 -0.202 0.158 -1.278 0.209
Log of estimated cost -0.310 0.100 -3.083 0.004
Log of estimated schedule 0.128 0.117 1.091 0.282
Log of forex -0.054 0.112 -0.480 0.633
Log per capita income 0.038 0.048 0.780 0.440
Log station size 0.052 0.110 0.470 0.640
Log % ICB -0.017 0.080 -0.220 0.826
Log of MUV growth -0.097 0.049 -1.954 0.059
Log GDP deflator -0.020 0.058 -0.346 0.731
Log of unit size 0.100 0.070 1.430 0.161
Basis for cost dummy -0.061 0.072 -0.844 0.404
Diesel dummy 0.010 0.150 0.071 0.943
Coal dummy -0.141 0.099 -1.421 0.164
Steam dummy 0.048 0.081 0.591 0.558
Extension dummy -0.158 0.076 -2.064 0.046
Log sales growth -0.026 0.035 -0.727 0.471
Log agencies 0.001 0.105 0.015 0.988
Dummy for 1976 -0.322 0.132 -2.432 0.020
India dummy 0.240 0.140 1.711 0.096
R-squared 0.579 Mean of dependent 0.066
SER 0.182 SD of dependent 0.223
Linearized R-squared 0.601
Annex 2: Regression Results with All Variables Included 71

Table A2.3 Variablesfor Log of HydropowerProjectCosts(CurrentValues)


Basedon 50 Observations

Regression Standard 2-tailed


Variable coefficient error t-stat. significance
Intercept 3.768 7.814 0.482 0.634
Log loan date -1.038 1.858 -0.558 0.581
1970 dummy 0.154 0.229 0.676 0.505
Log estimated cost -0.184 0.215 -0.858 0.398
Log estimated schedule -0.027 0.200 -0.136 0.893
Log forex -0.086 0.193 -0.444 0.660
Log per capita income 0.037 0.067 0.558 0.581
Log station size 0.267 0.189 1.407 0.171
Log % ICB 0.098 0.127 0.776 0.444
Log MUV growth 0.034 0.117 0.292 0.772
Log GDP deflator 0.056 0.066 0.842 0.407
Log unit size -0.228 0.172 -1.325 0.196
Basis for costs dummy 0.021 0.110 0.193 0.848
Extension dummy 0.447 0.500 0.894 0.379
Log sales growth -0.073 0.076 -0.966 0.343
Log height ( new) 0.059 0.078 0.763 0.452
Log head (new) 0.094 0.074 1.256 0.220
Log agencies 0.195 0.156 1.250 0.222
Colombia dummy -0.381 0.197 -1.929 0.065
Brazil dummy 0.231 0.289 0.799 0.431
Tunnel length 0.007 0.010 0.665 0.511
Contractor dummy 0.329 0.326 1.010 0.321
Log civil ratio 0.056 0.162 0.350 0.729
1976 dummy -0.014 0.234 -0.061 0.952
R-squared 0.616 Mean of dependent 0.185
SER 0.255 SD of dependent 0.301
Linearized R-squared 0.631
72 Estimating Construction Costs and Schedules

Table A2.4 Variablesfor Log of HydropowerProjectCosts(ConstantValues)


Basedon 45 Observations

Regression Standard 2-tailed


Variable coefficient error t-stat. significance
Intercept -3.778 10.881 -0.347 0.731
Log of loan date 0.933 2.573 0.362 0.719
Dummy for 1970 -0.057 0.254 -0.227 0.822
Log of estimated cost -0.201 0.107 -1.875 0.072
Log of estimated schedule -0.047 0.169 -0.278 0.783
Log of forex -0.211 0.199 -1.064 0.297
Log per capita income -0.071 0.074 -0.954 0.348
Log station size 0.142 0.167 0.852 0.402
Log % ICB 0.032 0.131 0.243 0.809
Log of MUV growth -0.017 0.108 -0.163 0.871
Log GDP deflator 0.022 0.088 0.249 0.805
Log of unit size 0.012 0.150 0.082 0.934
Extension dummy 0.789 0.372 2.118 0.044
Log sales -0.066 0.078 -0.844 0.406
Log dam height 0.008 0.074 0.116 0.908
Log head 0.177 0.060 2.934 0.007
Log agencies 0.445 0.178 2.489 0.019
Dummy for 1976 -0.084 0.235 -0.359 0.722
Brazil dummy 0.172 0.334 0.517 0.609
Colombia dummy -0.410 0.245 -1.674 0.106
R-squared 0.562 Mean of dependent 0.212
SER 0.255 SD of dependent 0.291
Linearized R-squared 0.577
Annex 2: Regression Results with All Variables Included 73

Table A2.5 Variablesfor Log of ThermalPowerProjectSchedules


Basedon 42 Observations

Regression Standard 2-tailed


Variable coefficient error t-stat. significance
Intercept 1.311 7.711 0.170 0.867
Log loan date -0.310 1.877 -0.165 0.870
1970 dummy -0.051 0.177 -0.293 0.772
Log estimated cost 0.117 0.136 0.856 0.402
Log estimated schedule 0.055 0.160 0.347 0.732
Log forex -0.044 0.134 -0.332 0.743
Log per capita income -0.013 0.055 -0.238 0.814
Log station size -0.303 0.160 -1.889 0.074
Log % ICB 0.160 0.088 1.816 0.085
Log MUV growth 0.115 0.066 1.744 0.097
Log GDP deflator -0.076 0.062 -1.222 0.237
Logunitsize 0.187 0.113 1.654 0.114
Basis for costs dummy -0.057 0.097 -0.589 0.562
Diesel dummy 0.146 0.268 0.544 0.592
Coal dummy -0.004 0.130 -0.034 0.973
Steamdummy -0.162 0.105 -1.544 0.139
Extension dummy -0.152 0.098 -1.550 0.137
Log sales growth 0.001 0.043 0.034 0.973
Log agencies 0.193 0.113 1.696 0.106
Contractor dummy 0.293 0.150 1.946 0.067
1976 dummy -0.099 0.166 -0.601 0.555
India dummy -0.035 0.152 -0.233 0.818
Log estimated civil -0.021 0.042 -0.498 0.624
R-squared 0.575 Mean of dependent 0.200
SER 0.174 SD of dependent 0.182
Linearized R-squared 0.659
74 Estimating Construction Costs and Schedules

Table A2.6 Variablesfor Logof HydropowerProjectSchedules


Basedon 50 Observations

Regression Standard 2-tailed


Variable coefficient error t-stat. significance
Intercept 2.706 4.767 0.567 0.575
Log loan date -0.464 1.134 -0.409 0.685
1970 dummy 0.057 0.139 0.408 0.686
Log estimated cost 0.004 0.131 0.036 0.971
Log estimated schedule -0.431 0.122 -3.538 0.002
Log forex 0.010 0.118 0.090 0.929
Log per capita income 0.113 0.041 2.740 0.011
Log station size 0.082 0.115 0.710 0.484
Log % ICB -0.175 0.077 -2.261 0.032
Log MUV growth 0.002 0.071 0.040 0.968
Log GDP deflator 0.007 0.040 0.186 0.853
Log unit size -0.105 0.105 -0.998 0.327
Basis for costs dummy 0.059 0.067 0.873 0.390
Extension dummy 0.256 0.305 0.840 0.408
Log sales growth 0.019 0.046 0.423 0.676
Log height ( new) 0.000 0.047 -0.016 0.987
Log head ( new) 0.020 0.045 0.453 0.654
Log agencies -0.032 0.095 -0.337 0.739
Colombia dummy 0.144 0.120 1.194 0.243
Brazil dummy -0.123 0.176 -0.701 0.489
Tunnel length 0.004 0.006 0.623 0.538
Contractor dummy 0.040 0.198 0.204 0.840
Log civil ratio 0.099 0.099 1.002 0.325
1976 dummy -0.080 0.143 -0.563 0.578
R-squared 0.698 Mean of dependent 0.211
SER 0.156 SD of dependent 0.206
Linearized R-squared 0.719
Annex 3: Comparisonsof Actual Ratios and
Predicted Ratiosfrom Regressionsfor
Costs and Schedules
Figures A3.1 to A3.6 are plots of actual ratio (actual value to estimated value) and
predicted ratio (predicted value from one of the regression equations given in chapter 6 to
estimated value) for each power generation project, for each of the following variables:

* Thermal power project costs (current values)

* Thermal power project cost (constant values)

* Hydropower project costs (current values)

* Hydropower project costs (constant values)


* Thermal power project schedules

* Hydropower project schedules


The numbers along the horizontal axis are references to individual projects. The
numbers along the vertical axis are values of the ratios.

75
76 EstimatingConstructionCosts and Schedules

Figure A3.1 Plot of Actual and Predicted Ratios for Thermal Costs (current)
Ratloof actualto predictedvalue
2.0

1.80'

1.60"

1.40'
Actual
A ratio
1.20
A Predictedratio
1.00

0.80

0.60)

0.40
1 10 20 30 40
Projectnumber(orderedby dateof loanagreement)

FigureA3.2 Plotof Actualand PredictedRatiosforThermalCosts(constant)


Ratioofactualto predicted
value
1.80__ __ _

1.60

1.40

1.20 -a- - Actual ratio


- -A-- Predictedratio
1.00

0.80

0.60

0.40 __ _ _

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55

Projectnumber(ordered
bydateof loanagreement)
Annex 3: Comparisonsof Actual and PredictedRatios 77

Figure A3.3 Plot of Actual and Predicted Ratios for Thermal Schedules

Ratioof actualto predictedvalue


2.40 __ _ _ ---

2.00

1.6 NE Actualratio
D Predictedratio
1.2

0.80

1 3 5 7 9 11 13 15 16 18 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57
Projectnumber(orderedby dateof loanagreement)

Figure A3.4 Plot of Actual and Predicted Ratios for Hydro Costs (current)

Ratioof actualto predictedvalue


2.50

2.00

1.50 Actual
ratio
-- Predicted
ratio

1.00

0.50
65 75 85 95 105 115 125 130
Projectnumber(orderedby dateof loanagreement)
78 EstimatingConstructionCosts and Schedules

Figure A3.5 Plot of Actual and Predicted Ratios for Hydro Costs (constant)
Ratio of actual to predicted value

2.50-

2.00

HT-Actualratio
1.50 - Predictedratio

1.00;0

0.50 _ _ _ _ _ _ _ _ _ _ _

65 75 85 95 105 115
Project number (ordered by date of Ioan agreement)

Figure A3.6 Plot of Actual and Predicted Ratios for Hydro Schedules
Ratioof actualto predictedvalue
2.20 -.

2.00-

1.80

1.60 O Actlal ratio


A Predicted ratio
1.40

1.20

1.00

0.80
65 75 85 95 105 115 125

Project number(orderedby date of loan agreement)


Annex 4: Statisticsfor Single-VariateAnalysisof
All Variables
This annex contains schedules of single-variate correlations for all the variables
included in the regression equations. These correlations are presented to show how the
variables that are found to be significant from single variate analysis differ from those
found to be significant from multivariate analysis.
There is an equation that links the (unsquared) correlation coefficient to a t statistic
that would be obtained in the single-variable regression:
20.5
r = t (t2 +N-2)

t = r{(N -2) (I - r2

(N is the number of observations)


Hence a critical value of the correlation can be obtained from a t table. For example,
with 54 observations, the critical value of t (two-sided test) at 90 percent is 1.67, and at
95 percent it is 2.01. The corresponding critical values for r are thus (plus or minus)
0.223 and 0.266. Larger values than 0.05 for r2 thus indicate a significant correlation
between the variables at the 90 percent confidence level. In other words, finding a small
r2 is equivalent to finding a nonsignificant t-stat.

79
80 Estimating Construction Costs and Schedules

Table A4.1 Single-VariateRegressionCorrelations

Thermal power projects Hydropower projects


Cost overrun Cost overrun
Variable (current) Schedule slip (current) Schedule slip
Log loan date -0.320* -0.216 -0.233* -0.049
Post 1970 dummy 0.167 -0.127 -0.136 0.061
Log estimated cost 0.508* -0.384* -0.191 -0.303*
Log estimated schedule -0.271* -0.270* 0.118 -0.415*
Log forex 0.117 0.055 -0.302* 0.043
Log per capita income -0.091 0.029 0.028 0.227*
Log station size -0.464* -0.045 0.082 -0.202
Log % ICB 0.050 0.080 -0.090 -0.062
Log MUV growth 0.247* 0.230* 0.372* -0.001
Log GDP deflator -0.181 -0.179 0.235* -0.057
Log unit size -0.404* -0.390* 0.122 -0.196
Basis for cost dummy 0.005 -0.032 0.285* 0.028
Diesel dummy 0.285* 0.300* - -
Coal dummy -0.306* -0.140
Steam dummy -0.018 -0.207 - -

Extension dummy -0.178 -0.042 -0.205 0.334*


Log sales growth -0.123 0.045 0.015 0.155
Log agencies -0.233* -0.034 -0.087 -0.081
Post 1976 dummy -0.402* -0.352* -0.203 -0.030
India dummy -0.064 -0.253* - -
Local contractor dummy -0.037 0.094 -0.050 -0.031
Log civil costs ratio" 0.366* -0.009 0.246* 0.014
Log dam height (new) - - 0.171 -0.327*
Log head (new) - - 0.234* -0.164
Colombia dummy - - -0.169 0.209
Brazil dummy - - 0.402* -0.210
Tunnel length - - 0.154 0.119
Note: Numberof observationsfor all the regressions: 54. Significantvalueof correlationr at 90 percent
confidencelevel: ±0.233.
* Variableis significantlycorrelatedat 90 percentconfidencelevel.
45 observations.
Annex 4: Statistics for Single-Variate Analysis of All Variables 81

Table A4.2 Comparisonof SignificantVariablesat 90 PercentConfidenceLevel


betweenMultivariateAnalysisand Single-VariateAnalysis
Relationship Thermal power projects Hydropower projects
between Cost overrun Cost overrun
variables (current values) Schedule slip (current values) Schedule slip
Common Log civil costs ratio (+) Post 1976 dummy (-) Log hydraulic head (+) Extension dummy (+)

vagriable Log estimated cost (-) India dummy (-) Log forex (-) Log estimated schedule (-)
with same Post 1976 dummy ( Log GDP deflator (+) Log national income (+)
signs (I I)

Common Log estimated schedule


variables (+ for multi. - for single)
but opposite
signs (I)

Significant Extension dummy (-) Log station size (-) Extension dummy (+) Log hydraulic head (+)
for multi- India dummy (+) Log loan approval date (+) Log estimated cost (-) Log % ICB (-)
variate, not
significant Log forex (-) Log station size (+)
for single Post 1970 dummy (-) Log financing agencies (+)
variate (13)
Colombia dummy (-)

Significant Log loan approval date (-) Log estimated cost (-) Log loan approval date (-) Log estimated cost (-)
for single Log station size LLogestimatedschedule(-) Log MUV growth(+) Log dam height(-)
variate, not
significant Log MUV growth (+) Log MUV growth (+) Basis for cost dummy (+)
for multi- Log unit size Log unit size (-) Brazil dummy (+)
variate (19)
Diesel dummy (-) Diesel dummy (-) Log civil costs ratio (+)
Coal dummy (-)
Log financing agencies (-)
Annex5: Ex PostAttributionof Factors
Responsiblefor ScheduleSlip in WorldBank-
SupportedPowerGenerationProjects
Table A5.1 Ex Post Attributionof FactorsResponsiblefor ScheduleSlip in
World Bank-SupportedPowerGenerationProjects
Responsible party orfactor Specificfactor or event
Thermalpowerprojects
Client/engineer Legal requirements/bureaucratic procedure for awarding contracts
Initial schedule was too optimistic
Bid evaluation difficulties
Delays in procurement/placement of orders
Change in project scope
Modifications to major equipment required
Disagreementbetween Bank and borrower over contract award
Site change
Contractor/supplier Labor disputes/strikes in manufacturer's country
Labor disputes/strikes in project country
Shipping delays due to oil crisis
Substandard work had to be redone
Equipment failure during testing
Skilled labor shortage
Manufacturing difficulties
Shortage of materials
Contractor inefficiency/lack of coordination
Technical problems with equipment
Contractor bankruptcy
Transportation difficulties
Uncontrollable events Damage/need to redesign civil works due to earthquake or other natural
Disaster
Unusually bad weather
Accident-damage to equipment
Political turmoil/coup/invasion
Civil disturbance

(continued on next page)

83
84 Estimating Construction Costs and Schedules

Responsible party orfactor Specific factor or event


Hydroelectricpowerprojects
Client/engineer Initial schedule too optimistic
Geological problems
Financial difficulties/tariff implementation problems
Inefficient project management/institutional weakness
Design changes
Change in project scope
Relocation problems
Delay in award of contracts/bid evaluation difficulties
Design faults
Procurement delays/difficulties
Site change
Land acquisition/site access problems
Communication problems
No bids received due to working conditions
Major currencv devaluation threatened project viability
Project sponsors backed out
Legal problems/delay in settling claims
World-wide inflation
Supplier/contractor Contractor bankruptcy
Contractor inefficiency/inexperience/incompetence
Delays in shipping/deliver of equipment, transportation difficulties
Substandard work had to be redone
Shortage of materials
Manufacturing difficulties
Dam or tunnel collapse
Damage to equipment (other than dam/tunnel)
Water infiltration/pressure damage
Fire
Equipment failureduring testing
Labor disputes/strikes in manufacturer's country
Fuel shortage
Labor disputes/strikes in project country
Communication problems
Skilled labor shortage
Change in contractors
Legal problems/delay in settling claims
Uncontrollable events Landslides/mudsl ides/rockfalls
Unusually bad weather
Political turmoil/coup/invasion/war
Flood damage
Earthquake

Note: Factors attributed in World Bank Project Completion Reports.


Annex6: Methodologyfor DerivingActualProject
Costsin ConstantPriceTerms
Since World Bank project completion reports do not provide data on project costs in
constant-price terms and the actual disbursement patterns over the project implementation
period, the methodology described below is used to derive equivalent actual constant-
price costs from the actual current-price costs given in these reports. This methodology is
the reverse of the World Bank's methodology for computing the price contingency for
project cost estimates.
a. Total actual project costs are allocated among the project implementation years
according to standard disbursement profiles related to the total project implementation
periods. These profiles are shown on the next page.
b. The actual annual disbursements are divided between foreign exchange and local
currency costs by the proportion of each in the total actual project cost.
c. The stream of annual foreign exchange costs in U.S. dollar (US$) terms is converted
to constant-price terms in the project start year with the actual MUV indices for the
implementation period.
d. The stream of annual local currency costs in local currency terms is converted to
constant price terms in the project start year with the actual country CPI or GDP
deflator for the implementation period. The local currency cost in constant price
terms is converted into equivalent US$ terms at the average exchange rate in the
project start year.
e. The total project cost in constant US$ terms for the project start year is the sum of
steps (c) and (d).
A worked example is given in this annex.

85
86 Estimating Construction Costs and Schedules

Table A6.1 StandardDisbursementProfilesfor ProjectCostin CurrentPrice


Terms

Annual
disbursement
(of total)
in years Implementation periods (years)
11 10 9 8 7 6 5 4 3 2
1 0.02 0.03 0.03 0.04 0.05 0.10 0.15 0.20 0.25 0.35
2 0.03 0.04 0.07 0.10 0.15 0.20 0.25 0.30 0.50 0.65
3 0.06 0.08 0.12 0.15 0.20 0.25 0.35 0.40 0.25
4 0.10 0.12 0.15 0.25 0.25 0.25 0.20 0.10
5 0.20 0.20 0.20 0.20 0.20 0.15 0.05
6 0.20 0.20 0.20 0.15 0.10 0.05
7 0.15 0.15 0.12 0.07 0.05
8 0.10 0.10 0.08 0.04
9 0.08 0.05 0.03
10 0.04 0.03
11 0.02
SUM 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
Source: These profiles are based on the generic formulafor expenditureflow patterns for large capital
projects given below. The formulawas used by IndependentProject Analysis,Inc. of Reston, Va., in a
report publishedby the Industryand EnergyDepartmentof the WorldBank (Merrowandothers 1990).

Proportionof total expenditurethat occursin year i of a total projectconstructionperiodof I years:

i [L (Cos )4.08 ]
Annex 6: Methodologyfor DerivingActualProjectCosts 87

Table A6.2 Exampleof ProjectCost Derivationin ConstantPrice Terms:


Algeria,BaseYear 1973
Local
Forex cost
percent percent
0.58 0.42
Foreign Annual
Annual MUVfor cost in Annual local Local cost
Annual forex foreign Year I local Average cost Local In Year I
cost cost costs price terms cost annual (million CPI price terms
Schedule (USS (US$ (Year] (US$ (USS exchange rate of (Year I = (million of
year million) million)a 1.00) million) million)t (Dinnar/US$) Dinnar) 1.00) Dinnar)

1 2.9 1.68 1.00 1.68 1.22 3.96 4.83 1.00 4.83


2 10.0 5.80 1.22 4.75 4.20 4.18 17.56 1.06 16.57
3 11.0 6.38 1.35 4.73 4.62 3.95 18.25 1.11 16.44
4 16.7 9.69 1.37 7.07 7.01 4.16 29.16 1.21 24.10
5 15.0 8.70 1.51 5.76 6.30 4.15 26.15 1.32 19.81
6 8.3 4.81 1.74 2.76 3.49 3.97 13.86 1.48 9.36
7 3.6 2.09 1.97 1.06 1.51 3.85 5.81 1.73 3.36

TOTAL 67.5 39.15 10.16 27.81 28.35 115.62 94.47


Local cost @ Year I Price and
Exchange Rate: US$ million 23.86

Forexpercent0.58. b Localcost percent0.42.

In Year I Price Terns:


USSmillion
Total foreigncost 27.81
Total localcost 23.86

Total cost 51.67

Forex= foreignexchange.
Annex 7: Analysis of Relationshipsfor the
Performanceof Price Contingencies
1 RelationbetweenActualand EstimatedCurrentCosts
The first relationship serves to highlight the nature and extent of the difference
between actual and estimated current costs. Were they always approximately equal,
project appraisal could rely confidently on the estimated costs. Where these series are
different, it is important first to be aware of the possible magnitude of the difference and
next to analyze it for predictable features to be incorporated into project appraisal.
For the group of projects as a whole, the regression of the actual cost in current price
terms (CU(A)) on the estimated cost in current price terms (CU(E)) gave the result shown
in equation A7. 1:
CU(A) = 25.8 + 0.985CU(E)
(1.7) (20.8) (A7.1)
R2= 0.76, Standard error of estimate (SEE) = 296; t statistics are in parentheses.
If the estimated cost is on average an unbiased estimate of the actual value, then
equation A7.1 should satisfy the hypothesis that the intercept is zero and the slope is
unity. An F test comparing equation A7. 1 with the equation:
CU(A) = CU(E)
gives an F value of 0.55, compared with the critical value of 3.07. However, the standard
error of estimate, which is the average difference between the actual current cost and the
value predicted by this relationship (and is the same as the standard error of the
regression), is extremely large at 296 when compared with the sample mean actual
current cost of 216 (in million US$). Moreover, the residuals are strongly heteroskedastic
(being larger in absolute values at larger values of estimated cost), so that valid inferences
cannot be based on the standard F test. The mean value of the ratio of actual to estimated
current costs for the whole sample is 1.21, and this is significantly greater than unity, as
shown by equation A7.2:
CU(A)/CU(E) = 1.21 (A7.2)
(35.5)
SEE = 0.40
Once the data is put into ratio terms, the absolute values of the residuals do not
exhibit heteroskedasticity, and so standard significance tests on equation A7.2 are valid.
From this analysis it can be concluded that the actual and estimated current values are
strongly, but far from perfectly, correlated. It appears that the ratio of actual to estimated
values is significantly greater than unity and that there is a significant percentage bias in

89
90 Estimating Construction Costs and Schedules

the estimation of current prices. There is a very large difference in the ratio for thermal
power projects (1.06) and hydro projects (1.27), which indicates that the bias in
estimating thermal power project costs in current price terms has been relatively small.
Both groups have substantial standard deviations, but again that for hydropower projects
(38 percent) is higher than that for thermal projects (23 percent). This difference can be
due to differences in predicting either the physical costs (constant prices) or the inflation
rates, but is likely to be primarily due to the former, since it would be expected that errors
in forecasting the latter would not depend on the type of project involved.

2 Link betweenActual CurrentCosts and ActualConstantCosts


In order to analyze the sources of error in current price forecasting, it is necessary to
look at the error in forecasting the constant price cost, as well as the errors in forecasting
inflation. In order to check this physical error, it is important to check that the value
derived for the actual constant cost from the disbursement formula described in Annex 6
is strongly related to the actual current value. If these two series are only weakly linked,
then doubt would be cast on the construction of the actual constant price series and, as a
result, analysis of the success in predicting the "physical" aspect of the project, by
comparing actual and estimated constant costs, would be of less value.
The regression includes as explanatory variables not only the actual constant price
(CO(A)) but also the actual (annual) rate of inflation (ID), the actual (annual) rate of
inflation of imported manufactures (IF), and the actual schedule length (SA). All these
variables are involved in the relation between actual costs in constant prices and actual
costs in current prices. The regression on the set of observations for which there are data
is given in equation A7.3:
CU(A) = 91.8 + 0.78 CO(A) + 0.25 ID
(14.8) (0.63) (A7.3)

-14.0 IF+ 14.7SA


(4.1) (2.0)

R2 = 0.74, SEE = 153


This equation shows that there is certainly a strong relation between the constructed
actual constant cost value and the actual current cost value, allowing for actual inflation
rates and schedule length. Separate regressions for thermal and for hydropower plants
show that the relation is extremely close for thermal power projects (R2 = 0.98) and is
weaker for hydropower projects (R2= 0.75). Since none of these "explanatory" variables
can be known in advance, the equation cannot be used in a direct, predictive fashion.
However, it does justify the attempt to separate out the inaccuracy in estimating the
"physical" or constant cost dimension of the project from errors in estimating inflation
rates and schedule lengths.
Annex 7: Analysisof Relationshipsfor the Performanceof PriceContingencies 91

3 RelationbetweenActualand EstimatedCost Escalation


In order to check the extent to which errors are made in predicting actual current
prices because of failures to anticipate cost inflation correctly (as opposed to the physical
aspects of the project), the actual cost escalation, RA = CU(A)/CO(A), is regressed on the
estimated cost escalation, RE = CU(E)/CO(E), in equation A7.4.
RA= 0.30+0.70RE (A7.4)
(7.3) (22.1)

R2 = 0.79, SEE = 0.20


This relation appeared stable over time, in that the introduction of extra variables such
as the date of the loan agreement or a post-1976 dummy3 0 make only very slight changes
to the fitted equation.
Separation of the thermal power and hydropower project subsamples showed little
difference in the relationship for types of plant, which is to be expected, since the failures
to predict the inflation component should not depend on the type of plant but rather on
assumptions that are common to all plant types. Failure to predict cost escalation
correctly suggests that there is an aspect of the current cost forecast that is always likely
to be in error, unless general rules for predicting domestic and international inflation can
be improved.
The results show that the World Bank's procedures were fairly accurate in predicting
cost escalation, but that with an average error of 0.20, compared with the sample mean
actual cost escalation of 1.14, there was still room for improvement in the construction of
this aspect of the project price forecast.

4 RelationbetweenCost Overrun in CurrentPriceTerms, in ConstantPrice


Terms, and Errors in PredictingInflationRates and ProjectSchedules
The themes of the three previous sections are pulled together by regressing the cost
overrun in current price terms O(CU) on the cost overrun in constant price terms (O(CO))
and the errors in predicting foreign inflation (O(F)), domestic inflation (0(D)), and
schedule (O(S)) in equation A7.5 to give:
O(CU)= -0.10 + 0.82 0(CO) + 0.02 O(D) + 0.090(F)
(1.10) (12.4) (2.0) (1.6) (A7.5)
+0.12 O(S)
(2.2)

R2= 0.73, SEE = 0.20

29. The estimated cost in constant price terms (CO(E)) includes the physical contingencyfor cost
increases that is mentioned in section 2.
30. See main text footnotes 8 and 9 about changes in the World Bank's guidelines.
92 EstimatingConstructionCostsand Schedules

Equation A7.5 shows that errors in estimating the project cost in constant price terms,
errors in estimating the domestic inflation rate, and errors in estimating the project
schedule were all significantly related to the overall error in estimating project costs in
current terms.
There was no significant relation with the error in estimating the imported rate of
inflation; this finding does not mean that if an error were made in estimating the imported
inflation rate (Figure A7.1), such an error would not be reflected in total project cost
errors, but rather that such errors were on average sufficiently small to be dominated by
the other sources of error. The overall goodness of fit is quite high and indicates that this
decomposition is fairly successful in identifying the sources of the cost overruns by type.
Disaggregation into thermal power and hydropower project subsamples again shows a
stronger fit for thermal power projects (R2 = 0.85) and a weaker fit for hydropower
projects (R2 = 0.67), which suggests that the attempt to split the cost overrun into its
components is rather less satisfactory for hydropower projects. This may well relate to
the basic problem identified earlier in predicting the physical aspects of costs for
hydropower projects correctly.

FigureA7.1 Comparisonof MUVIndex ActualValueswith Valuesfrom Forecasts


Madebetween1974and 1988 (Basedon actualvalue in 1980 = 100)
220

200 1979

180 . 1981
1988
160 -

140

120

100

60

40
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93

Year
Note: Forecasts used for World Bank Project Appraisals of the Unit Value Index (in US$
terms) of manufactured goods exported from G-5 countries (France, Germany, Japan,
UK., and U.S.) to developing countries.
Annex 7: Analysisof Relationshipsfor the Performanceof PriceContingencies 93

AppendixA7.1 Compositionof the World Bank's PriceContingency


Formulafor PredictingProjectCost Escalation
The World Bank's formula for computing price contingencies is:

EE= B *C*pf +Cd*pd *ed

where:
EE is the estimated price contingency for escalation in project costs.
i is a year in the project implementation schedule of n years.
B is the estimated total project cost expressed in base year prices (i=l).
Bi is the proportion of B that is committed in year i.

cf is the estimated proportion of Bi that consists of imported components.


d
Cd is the estimated proportion of B, that consists of domestically procured
components, and thus equals ( I - C-).
f is the projected price index in year i relative to the price index in the project base

year for imported project components, which is taken to be the forecast change in
UN Unit Value Index (in US$ terms) of manufactured goods exported from the
G-5 countries to developing countries (the "MUV Index").

p is the projected price index in year i relative to the project base year for the
domestic economy, which is usually taken to be the domestic Retail Price Index
("RPI").

ed is the forecast change in the exchange rate in year i relative to project base year
for the domestic currency in US$ terms (for estimating a price contingency in
US$ terms).
Annex8: Computationsof Probabilitiesof
ExceedingSpecificProjectCosts
Case 1: The cost level for a given project that will be exceeded with a specified
probability. If a variable (X) follows a normal distribution N(m,s), where m is the mean
and s the standard deviation, there is a 50 percent chance that the actual value will exceed
the mean. As a control of unlikely outcomes it is possible to calculate the "cut-off' at
which there is only a specified (usually small) probability of exceeding this value. For
example, the "cut-off' (K) value at which there is only a 10 percent chance of a higher
outcome is given by the solution to the equation

0.1 = fX(m,s)dx
K

which is equivalent to the equation expressed in standard normal terms:

0.1 = |X(O,l)dx
(K- m)/s

The ordinates that solve such an equation are given in standard tables of the normal
distribution. For example, the value of the standardized score that cuts off the top 25
percent of the distribution is 0.675. Hence, for known values of m and s, the cut-off
value for X itself is

K = m + 0.675*s
This approach can be applied to the Indian power project discussed in chapter 6. The
predicted log of the ratio of actual to estimated costs was 0.192, while the standard error
of the regression was 0.189 (Appendix A8.1 to this annex shows how to compute the
variance of a forecast). Hence, the cut-off value of the log of the ratio, which has only a
25 percent chance of being exceeded, is:

K=0.192+0.675 * 0.189=0.320
The cut-off value of the ratio itself is thus 1.377 and, given that the estimated value of
costs was $406 million, the cut-off predicted actual value is $559 million, as compared with
the central predicted value of $492 million. The regression thus takes the estimated value
($406.29) and produces a predicted value of $492 million, which is the appropriate figure
for standard econornic evaluation. It also indicates that the value at which there is only a 25
percent chance of a higher cost is $559 million, and this finding can be used in risk analysis.
This approach can be inverted to ask for a given cost limit, what is the probability of
the project cost exceeding such a value. The fornal equation is:

95
96 EstimatingConstructionCosts and Schedules

|X(m,s)dX
f3=
L

where L is the specified limit, and 1is the probability of exceedance. Again this is put
into standard normal form:

|X(O,l)dX
f=
(L-m)/s

and, for a given standardized limit, the value of 1can be derived from standard normal
tables. In the above example, if the cash limit were set to $575 million, this implies the
ratio of actual to estimated values would be 1.416, with a log of 0.348. The standardized
value would then be (0.348 - 0.192)/0.189 = 0.825, which has a 20 percent chance of
being exceeded. Hence with a cash limit of $575 million, on a project estimated to cost
$406 million, there is still a 20 percent chance of some overrun beyond the cash limit.

Case 2: Chzoosingbetween projects on the basis of the probability of exceeding a cost


limit. The approach indicated above for augmenting the analysis of a single project can be
extended to the choice between two (alternative) projects. The standard question is: which
project has the lower cut-off value of the distribution of costs that will be exceeded with a
50 percent chance in each case; and the answer is simply the lower of the predicted
outcomes. Alternatively, the question can be focused on the smaller probability level of
exceeding a higher cost level (K) than the predicted outcomes. Consider projects with
means ml and m2, and standard deviations s, and s2. With a cut-off probability of 25
percent the standardized score for each is, as before, 0.675. Hence the cut-off cost levels
are:

K, = ml + 0.675 * si

K2 = m2 + 0.675 * s,

It is then desired to choose that project with the lower value of K. If the values of the
standard deviations are different then it is possible that a project with the higher mean cost
ratio might nevertheless have a lower K value, because of its smaller standard deviation.

It also is possible to calculate the probability (P) of exceedance that has the same cut-
off value (H) for each project (H = Ki). At this value:

ml + H * sI = m2 + H * S2

H=(ml -m2)/(S2-sI)
The standard table for a normal distribution then indicates the probability of exceeding
the common cut-off value of H.
Annex 8: Computationsof Probabilitiesof ExceedingSpecificProjectCosts 97

This approach also can be inverted to ask which project has the lower probability of
exceeding a given cost limit R. Since the expected values will in general be different for
the two projects (El and E2 ), the logs of the ratios of the cash limit to expected values (r1
and r2 ) will also be different. Putting both ratios into standardized form then allows the
probability of exceedance of the absolute cash limit for project I to be compared with that
for project 2. If project I has the lower probability, then

fX(0, I)dX < f X(O,I)dX


(ri-mo)Isi (r 2 -m2)5S2

When there is indifference between projects (i.e., the same probability of exceeding R
applies to both projects), the limits of integration are equal, so that

(r, - ml) / s = (r 2 - Mi2 ) / S2

This expression can then be used to determine the specific value of R for indifference
between projects, since it yields

[ln(R/E1 ) - ml]/s, = [ln(R/E 2 ) - m2 1/S2

or

R = exp [( s2(0nEI + ml) - s1 (lnE2 + M2 ) 1/ts2 - si }]

A final criterion of interest is to ask what is the probability that project I will be more
expensive than project 2, even when the mean value for project 1 is lower. The answer
depends on the evaluation of a double integral (details are shown in Appendix A8.2 to
this annex). Table A8. 1 gives some selected values for different parameter combinations.

Table A8.1 Probability that Project 1 Has Higher Cost (Schedule) than Project 2

V
W 0 0.4 0.8 1.1 1.6 2.0
0.20 0.500 0.347 0.216 0.120 0.058 0.025
0.60 0.500 0.377 0.266 0.174 0.106 0.059
1.00 0.500 0.398 0.299 0.207 0.141 0.089
1.67 0.500 0.421 0.343 0.271 0.208 0.154
5.00 0.500 0.488 0.472 0.456 0.441 0.425
Where: V = (m2 - m)s, W = s2/sl.

An important result is that whenever m) is larger than ml, the probability that project
I is more expensive has as an upper limit 0.5 as the ratio of s2 to si increases. For low
values of this ratio, the chance that project 1 is more expensive can be very small.
98 EstimatingConstructionCostsand Schedules

Consider projects where the value of m2 is 0.2 and ml is 0.16, while S2 is 0.05 and s, is
0.125. The parameter V is 0.8, and W is 2.5. The chance that project 2 is more
expensive is 39 percent, despite the very much greater standard deviation. The table also
assumes that there is no correlation between the outcomes of the alternative projects.
Calculations showed that allowing for substantial positive or negative correlations had
little effect on the probabilities, so that this issue could be ignored.

AppendixA8.1 The Varianceof a PredictedValuefrom a Regression


The predicted value YF is based on regression of a sample size T between J known
variables (XI... XJ) including the constant (XI = I all t) and the dependent variable. The
predicted outcome is given by

YF
~F= EJ I1 jX;F

where the f3jdenote estimated parameter values, and XjFare the assumed known values of
the indicator variables. The variance of the forecast around the actual value is given by
the standard formula:

VarYF = a 2 + a 2 {Ei; (XjFXiF)COV(Pi 3j))

where Cov (fipj) is the estimated covariance or variance (i=j) of the regression parameters.

For large samples, such as are used in the regressions in this paper, the terms in braces
tend to be much smaller than unity, so that the variance of the forecast will be
approximately equal to the residual variance (square of the standard error of regression). In
comparing projects with similar mean outcomes, small differences in the variance can
become important, and in such a case it would be more important to estimate the variance
correctly. This facility is provided on modem regression programs. For example, the case
of the India project referred to in the text has a variance of the residuals of 0. 1772= 0.031
(for the log of the ratio) from the regression, while the true variance of the predicted value is
0.035.

AppendixA8.2 Probabilitythat the Outcomeof One ProjectIs GreaterThan


That of the Other
Let X and Y be the random outcomes of the two projects, which follow a bivariate
normal (correlated) distribution f(X,Y). The probability that a drawing X (project 1) is
larger than the Y value (project 2) depends on the integral:

f f f(X, Y) dX DY
0 Y
Annex 8: Computationsof Probabilitiesof ExceedingSpecificProjectCosts 99

This can be transformed to standard bivariate form:

J f g(X, Y) dX dY
- wY+V

where g denotes the standard bivariate normal distribution, and

V = (M 2 - m1 )/sI, W = S2/SI.

For particular values of V and W, the bivariate normal distribution is used to evaluate
the above double integral (by a numerical method), and the results are shown in Table
A8.1 for an uncorrelated distribution. Allowing the correlation coefficient to take values
of -0.5, 0.0, or +0.5 made very little difference to the results.
Annex9: AssigningProbabilitiesto
Scenariosfor RiskAnalysis
In the analysis of risky projects, it is necessary to choose alternative scenarios and to
attach probabilities to these scenarios. For example, a project will have an expected cost
outcome, and it is desired to investigate high and low outcomes and assign probabilities
to all three cases. This method can then be extended to other planning variables such as
demand growth, to give a multivariate probability of joint outcomes (e.g., high cost, low
demand).
The issue is how to combine the values chosen for the three scenarios with
probabilities to be attached to them. The combination of the range of values chosen and
the probabilities imply a variance to the set of outcomes. Since there is a prior view on
the variance of cost or schedule outcome from the regression equations in this paper (and
there may also be a view for the variance of the demand growth rate), it would be sensible
to ensure that the sets of values chosen are consistent with these variances. This annex
shows how this assignment can be done.
Consider an expected cost outcome of M (middle). This is the expected value derived
from the project analysis, adjusted if necessary from an estimated value by the regression
equation. The analysis suggests that the variance of cost levels is a 2 M2 , where a 2 is the
variance from the regression errors. Alternative scenario values L (low) and H (high) are
to be chosen, together with probabilities for all three cases.
To simplify the analysis it is assumed that the low and high outcomes are equidistant
from the mean (expected) outcome. This in turn implies that, for the mean of the
distribution to equal M, the probabilities of high and low cases must be equal. Let the
probability of the middle case be ir, so that the probabilities for the other two cases are
each (1 - ir )/2.
The key insight is that if the variance across these three outcomes is equal to the
known value or 2 M2 , there is a relation between the relative distance of L from M and the
probability that must be assigned to the middle value itself. Let the value of L be
expressed as AM ( where Awill be less than unity), then the relationship is given by the
formula (see the technical note at the end of this annex for derivation):
Ir= I-{r/(l-A)} 2 (A9.1)
An illustration is given using the India 911 example. The central cost estimate using
the regression approach is 491.5 (M) with the variance of the regression (C52) equal to
0.035. Table A9. 1 shows combinations of A and ir that would produce a set of L, M,
and H values with variance (a2 M2 ) equal to 8460 (the variance of the cost level).

101
102 EstimatingConstructionCosts and Schedules

Table A9.1 Pairs of ParametricValuesthat Fitthe RequiredRegressionVariance


of 0.035for Scenarioswherethe ExternalVarianceIs a Functionof the Size of the
MiddleValue

0.813 0.000
0.800 0.126
0.750 0.440
0.700 0.612

The above table shows, for example, that with a value of A of 0.75, which
corresponds to a low cost of 368 (491.5 x 0.75) and a high cost of 614 (491.5 x 1.25), the
probability that must be assigned to the middle value outcome is 44 percent, while both
low and high outcomes must each have a probability of 28 percent. It can also be seen
that if the high and low values are taken too near the medium value, there exists no set of
probabilities that would yield a variance over the three outcomes equal to 8460, as
predicted by external analysis.
In the case where the variance of outcomes is not proportional to the value of the
mean outcome, then a different formula is required. In this case the relation between the
probability of the central case and the range of cases (as expressed by A) is:
7r=l-{C/[(-_A)Mj) 2 (A9.2)
where M is the value of the middle estimate.
Consider a central growth rate of 0.04 (4 percent per annum), where the variance of
growth rates of demand has been established from other studies to be 0.01. Table A9.2
gives the corresponding tradeoffs.

Table A9.2 Pairs of ParametricValuesthat Fit the RequiredRegressionVariance


of 0.01 for Scenarioswherethe ExternalVarianceIs Not a Functionof the
Size of the MiddleValue

A 7
0.750 0.000
0.700 0.306
0.600 0.609

Hence when A = 0.600, a low-growth-rate scenario of 0.024 and a high-growth-rate


scenario of 0.056, combined with a probability for the middle-scenario-growth rate of
60.9 percent, will provide a set of growth rates and probabilities with variance 0.01.
Annex 9: AssigningProbabilitiesto Scenariosfor Risk Analysis 103

Probabilities can be simply multiplied to obtain joint events provided that the cost
distribution is independent of the demand distribution. Combining the two examples
gives the probability matrix in Table A9.3 that has a variance for costs of 8460 and a
variance for demand of 0.01.

Table A9.3 Probabilitiesfor Scenarioswith


PredeterminedVariancesfor Costsand Demand

Costs
368 491 614
0.024 0.0546 0.0858 0.0546
Demand 0.040 0.1708 0.2684a 0.1708
0.056 0.0546 0.0858 0.0546
aDerived from the correspondingvalues for it in Tables A9.1 and A9.2,
namely,0.440x 0.609.

The method could be extended to correlated distributions, but this is much more
complicated and it would also require an estimate of the correlation between demand
errors and cost errors. A different extension to nonsymmetric probabilities of high and
low cases is also possible, but again the formulas will be much more complex.

Technical Note: The Derivation of Equation A@9.1


For a symmetrical distribution of outcomes, the values are AM, M, and (2 - A )M,
which yields an expected value of M whatever the probabilities. Denoting the probability
of M by ir, it follows that the variance across these three outcomes is:

(I -i 2 +
(AmM-M) -27 )[(2-A)M-M]2 = (I-7r)[(l-A)M] 2

Equating this to the externally identified variance of a 2M 2 yields equation A9. 1.


Annex 10:The Calculationof the Meanand
Varianceof the Costof TwoProjects
Let mi be the expected (mean) cost predicted for project i (i=1,2), and v; (=si2 ) be the
predicted variance of project i.
From standard statistical theory, the mean and the variance of the sum of the costs of
the two projects is given by

m=ml +m 2

v = vl + v,- + 2cov1 2

where m is the mean of the total, v is the variance of the total, and coV12 is the covariance
of the distribution of costs for the two projects. Assuming that the distribution of costs
are independent, the variance of the sum is equal to the sum of the variances. Hence, the
standard deviation of the sum is given by

s = 4(vI + V2 )
These formulas3 ' can be applied to the expressions developed in the paper for
predicting mean costs and the variances of costs. The regression model predicts the log
of the ratio of actual costs to expected cost, where the latter are the values in the World
Bank Staff Appraisal Reports (SARs). Denote the predicted values from the regression
by P

P = log (A/E)
Now P is determined from the regression equation in terms of the characteristics of
projects, while E is available from the SAR. Hence, to obtain the prediction of the actual
value we have:

A = E exp (P)
This can be justified as the mean (or unbiased) value of A, where P is the mean
prediction of the log of the ratio, by the general statistical results that

if E(x) = 0
then E [g (x)] = g (e)

31. The formulascan be generalizedfor morethan two projects. Where the projectsare implemented
sequentially over time under a long-termdevelopmentprogram, the costs of the projects should be
expressed in terms of their present values in a common base year (usually year 0 of the development
program)accordingto whenthey wouldbe constructedunderthe program.

105
106 EstimatingConstructionCostsand Schedules

hence 0 = g-' {E [g (x)])


The variance of A also is required and here an approximation is used. Following the
above general notation:

Var [g (x)] = (dg/dx)2 Var[x]

where the derivative is evaluated at the mean of x (i.e., at 0).


Since the variance of P is known from the regression equation, as shown in Annex 8,
Appendix A8.1, and is denoted W:

W = Var [log (A/E)] = (1/A)2 Var A

Var A = A2 W

So the standard error of the predicted cost is approximately equal to the predicted cost
multiplied by the standard error of regression. For example, the India 911 project had a
variance for the log of the ratio of actual to estimated (W) of 0.035, while the predicted
actual value (A) was 491.5. Hence, the variance of the actual value is 8460, with a
standard deviation of 92.
For two projects with their predicted costs and variances of costs, the overall
predicted total cost and variance of the total cost can then be calculated.
It can be seen that if (a) project costs are independent; (b) the predicted value of the
ratio of actual to expected costs is independent of project size; and (c) the variance of the
ratio is independent of project size, then the variance of a sum of two smaller projects is
less than that of a single larger project with the same predicted cost as the sum of the
predicted costs of the two projects. For example, where the predicted costs of the small
projects are each 50 percent of the predicted cost of the one larger project, their mean
total cost is the same, while the variance of the large project would be twice that of the
sum of the variances of the two smaller projects.
Annex11: Applyingthe OptionApproachto
ConstructionCostsand Schedules
This annex illustrates the application of the option approach to investment timing for
capital projects that are subject to uncertainty about their construction costs and
schedules.32 It uses the results of the regression analysis in this paper for power
generation projects in developing countries to provide the estimates of unbiased values
and variances that are needed to apply the option approach to this class of projects.
The first two sections of the annex give a general introduction to investment valuation
under uncertainty using the option approach and to general solution methods for this
approach. The third section develops a simple model for finding the optimal timing for
investments, which is then applied in the final section to the general cases of thermal and
hydropower projects. These cases provide generic measures of the critical benefit/cost
(B/C) ratios and their variances for the optimal timing of investments in these two
subclasses of projects. The results confirm the findinas of the main analysis in the
paper-that investment risks from construction are greater for hydropower projects than
for thermal power projects. Under some plausible assumptions about uncertainty, it is
shown that the critical value of the B/C ratio has a premium over the value based on
expected (nonbiased values) equal to about 17 percent in the case of thermal power
projects, and likewise to 21 percent for hydropower projects. The variance for
hydropower projects is also greater than that for thermal power projects.

A. InvestmentValuationUnder UncertaintyUsingthe OptionsApproach


The so-called real option approach deals with the theory of investments under
uncertainty, where uncertainty is treated in a continuous time framework. It examines the
impact of irreversibility induced by the existence of fixed capital costs under conditions
of uncertainty. Capital costs are viewed as sunk; that is, little, if any, of these costs can be
recovered after the investment. The main assumption is that investment can always be
deferred. Future investment is viewed as a mutually exclusive alternative to investing
today. Even without uncertainty the problem is one of optimal timing, because although
the Net Present Value (NPV) of investing today may already be positive, deferring the
investment would be preferable if it increased the NPV. Keeping the investment option
alive is thus valuable. When the result from the option analysis is not to invest, the
investment prospect is not abandoned completely but is deferred to the future.
Option models are an improvement on simple NPV models because they capture the
economic risk that new information might be revealed by waiting, possibly by investing

32. This annex is based on text and analysisprovidedby SpirosMartzoukos,to whom the authors are
gratefulfor this contribution.

107
108 EstimatingConstructionCostsand Schedules

in further studies. Option models do not consider all kinds of uncertainty, and in that
sense they are a partial treatment of risk. They certainly cannot capture directly the risk
of changes to the external economic and regulatory conditions, if such conditions affect
the economics of the project only after initiation. Option models can capture such risk if
relevant information is revealed during the study period. They indirectly capture the ex
post uncertainty by requiring a premium before investing. Ideally, such uncertainty
should be captured by appropriately discounting the expected project payoffs, regardless
of the adoption of the option models.
Two premiums are considered by the option model. One is due to deterministic
growth trends that capture optimal timing in the deterministic sense. The other is a
premium for learning something by waiting under uncertainty. This does not necessarily
imply that uncertainty is reduced by waiting, although this could be also the case, for
example, in the case of technological uncertainty. It implies mainly that the estimate for
the mean changes. The timing (waiting) premium only has zero value when optimal
timing has been reached. Investing earlier kills this premium. The optimal time to invest
occurs when the value of the underlying asset (the benefits of the investment) exceeds its
cost by some predetermined margin. This margin is a function of the discount rates and
parameters of the stochastic processes of the underlying variables, such as growth trends
(if any) for the underlying investment benefits and costs, and uncertainty. For the
discount rates, a continuous time capital asset pricing model is assumed to hold.
Option models capture the value of the flexibility to invest in the future if it is
profitable to do so but not to invest if conditions worsen. This flexibility is captured in
the asymmetry of the distribution of the expected payoff function under the flexibility to
invest only if it is profitable to do so. This distribution is unaffected for positive future
NPVs but is truncated for negative NPVs, which are replaced by the value of zero. The
resulting asymmetry often values the option more highly than investing today, which
implies that optimal timing has not been reached yet. An even higher NPV is required to
justify a commitment to invest. Uncertainty enhances the effect of irreversibility, and
hence the value of flexibility, by increasing the required margin before investment should
take place. Thus the option theory revises the conclusion of the classical approach to
investment valuation that investment should take place immediately when NPV becomes
positive. Such an NPV valuation rule is applicable only when investment is not
deferrable.
Instead of looking at the NPV, the ratio of benefits over costs is often compared with
a critical ratio derived from the model. Only when the actual ratio reaches the critical
ratio has optimal timing been achieved. Using a B/C ratio allows option models to be
used for two uncertain parameters, namely uncertain benefits and uncertain costs, even
though they were originally designed to incorporate only one uncertain parameter. If
more than one large investment alternative exist, then the one-alternative analysis
underestimates the critical B/C ratio. Assuming two investments of similar scale, the
Annex 11: Applyingthe Option Approachto ConstructionCostsand Schedules 109

critical ratio for the best option depends on the ratio for the second-best one (Martzoukos
1995).
The theory of financial option valuation has been adapted to the economic analysis of
capital investments.3 3 For power projects a least-cost approach is used to select capital
investments. The preferred investment is the one that achieves the system goals at the
least economic cost. To incorporate this criterion into the option formulation, a do-
nothing or a minimal-investment scenario can be used as the alternative to the proposed
large investment, where expansion of supply is achieved (for a while at least) through
purchase or import of power or through investments in small and versatile thermal units.
The large investment alternative, such as a large hydropower or thermal plant, or a
transmission line to connect a load center to a central grid, which involves much higher
capital costs but lower operating costs than the alternative, would incur fixed (sunk)
costs. The do-nothing alternative has its own operating costs, which are avoided if the
large investment is implemented. Hence, in the option model the sunk capital costs are
considered to be the costs of exercising the investment option, and the avoided costs of
the do-nothing scenario to be the benefits of investing. Both the capital costs on one side
and price of power imports or fuel on the other side are effectively cost components but
represent different types of uncertainty. Capital costs represent technological or
institutional uncertainty, whereas the other costs represent import price or fuel price
uncertainty. In a least-cost approach (that is now obviously equivalent to the option B/C
analysis), the "benefits" should at least exceed the "costs." The question is, by how much?

B. General SolutionMethods
There are two alternative methodologies to find a solution to a specific investment
problem. The first is the contingent claims (option) approach, and the second is the
stochastic dynamic optimization approach. Both methods are effectively equivalent to a
stochastic optimization of the expected benefits minus the expected costs and differ only
in the method of approximation to the same problem specification.
The investment option valuation problem is expressed as a fundamental partial
differential equation for the option value, which can be solved by numerical methods (see
Brennan and Schwartz 1978), or, more rarely (except for simple problem formulations),
analytically. The numerical methods can be adjusted to treat options with discrete
changes in the benefits and costs if investment is deferred. They can also be adjusted to
treat options when the relevant parameters are not constant but are continuous and

33. In the real options literature,importantreviews are Pindyck(1991), Dixit (1992), and Dixit and
Pindyck(1994). The last is the most updatedcomprehensivereviewand is a highlyrecommendedtechnical
reference. A comparison of alternative treatments for uncertainty is given in Crousillat (1989).
Applicationsin the power sector have appearedin Paddockand others (1988),Crousillatand Martzoukos
(1991),and Martzoukosand Teplitz-Sembitzky (1992). A recentWorldBank applicationwasin an energy
optionsstudyof Pakistanin July 1995. The importanceof the deterministicoptimaltimingis highlightedin
Schramm(1989).
110 Estimating Construction Costs and Schedules

deterministic functions of time (see Martzoukos 1995a). The prolific financial economics
literature on option pricing can be used when it offers a model that has the same or
similar assumptions to the problem at hand.34
Another way to solve the same option problem, and of course derive the same result,
is the so-called lattice approach. The underlying continuous stochastic process can be
approximated by a discrete-time random walk in a decision-tree-like (lattice) fashion.
Dynamic programming can then be used to value the initial investment option. The
solution methodology starts at the terminal boundary conditions (option maturity, usually
a very large number of possible outcomes after many intermediate steps at which options
are faced over a long time horizon) by valuing each terminal option at the maximum of its
exercise price or zero. Then the analysis proceeds backward in time to derive the initial
option value. At every step, the option value is a probability weighted average of the
subsequent option values. The process at each step allows for the possibility of
exercising an option (at a node in the lattice) by calculating the maximum of the live
investment option value or its exercise value (the NPV).35 This method thus derives the
valuation of the investment option and the estimation of the critical B/C ratio through
properly discounting the asymmetric expected payoffs of the option. Although this
method uses the concept of dynamic programming in a solution methodology that evolves
backward, it is not explicitly formulated as a dynamic programming problem. It is rather
a hybrid of the option approach to the problem formulation and a backward decision-tree
solution methodology.
The standard deviation used in the option model (and captured indirectly in the
numerical solution for the partial differential equation or directly in the tree-type lattice)
effectively replicates a scenario approach. This scenario predicts increasing forecast error
in terms of the time frame and is given as a continuous (lognormal) distribution instead of
only a few discrete points with some probabilities attached to them. For example, a
standard deviation of 10 percent per year, which is equal to a variance of I percent per
year, would imply that the variance of the percent change in the underlying variable
would be 9 percent around the forecast in 9 years; this equals a standard deviation of 30
percent for that year. The method is an improvement over scenario methods since (a) a
continuous distribution captures more information than just a few selected points, (b) it
captures the asymmetry of the flexibility to invest in the future only if and when it is
optimal to do so, and (c) it inherently discounts properly under uncertainty. Scenario or
decision-tree methods cannot estimate the critical B/C ratio nor can they indicate how
close the scenario is to the optimal decision.
Option theory gives the correct discounting methodology for cash flows under
uncertainty. A feature of option models is that they allow different discount rates for

34. The first seminal papersof the financial options literatureinclude Samuelson(1967), Black and
Scholes(1973),and Merton (1973).
35. The methodologywas demonstratedby Coxand Ross(1976).
Annex I1l:Applyingthe Option Approachto ConstructionCosts and Schedules 111

benefits and the costs. Furthermore, this discount rate could be a function of time (in
models that are solved with numerical techniques). Stochastic dynamic programming is a
mathematical methodology that does not deal with the problem of correct discounting but
can benefit in this respect from insights gained from the option methodology.
Consequently, the two methods can be made strictly equivalent, and they should give the
same results.36

C. A SimpleModel for the InvestmentOptionand OptimalTiming


In the deterministic sense, optimal timing implies an optimization of the function
defined as benefits (B) minus costs (C). When investment is deferred, these components
grow continuously at gBand gc. The discount rate (risk-free rate for a risk neutral option)
is denoted as R. The terms 8B and °c are defined as the difference between the discount
rate and the actual expected growth rates for the investment benefits and costs. Both of
them are effective discount rates. The first is the opportunity cost of not investing, and
the second is the opportunity cost of investing.
The deterministic optimization problem solves
max[B*exponent(-5Bt) - C*exponent(-6ct)]
where maximization is in respect to time t. Equating the derivative with respect to time
to zero gives the optimal timing:
t = log[8cC/(8BB)]/(8B-6c)
In the absence of uncertainty, 8C/8B equals the critical B/C ratio.37
A case for the investment option under uncertainty with a known analytic solution is
the McDonald and Siegel (1986) model, which extends Samuelson's (1967) work to the
case of investment timing.
It can be easily solved with a calculator or a spreadsheet. The critical B/C ratio is
given by
T/(, - 1) (A 11.1)
where the parameter X equals
= (0.5 - (5B - C)/2) + [((6B - 5C)/G 2 - 0.5)2 + 28C/a 2 ].

The variance used in the option model equals


ay2 = (y2B + Cy2 c - 2cBYcrBC (Al 1.2)

36. The relationshipbetween the optionapproachand dynamicprogrammingis discussedin Dixit and


Pindyck(1994,chapter4).
37. At optimaltime, t* equalszero,so the criticalbenefit-costratioequals8c/°sa
112 EstimatingConstructionCosts and Schedules

and is in fact the variance of the actual B/C ratio given as a function of the individual
variances and the correlation between them. The negative sign implies that if benefits
and costs are very strongly and positively correlated, the effective variance that affects the
option price decreases.
The following simple example shows how this model works in the two parameter
case with growth rates of the deferred investment benefits and costs both equal to zero,
the continuous discount rate equal to 10 percent, and the cost uncertainty and the benefits
uncertainty both equal to 10 percent. It is also assumed that the correlation between costs
and benefits is 10 percent. If these assumptions about zero growth rates are not proper,
then the model that assumes constant model parameters would not hold. In that case it
would be necessary to test first whether the deterministic optimal timing has been
reached, which can be done by assuming that the investment is deferred for one or two
years. Then the economic analysis is repeated in which the benefits and costs are
discounted to the present. If the NPV of the deferred investment exceeds the NPV of
investing today, then clearly the investment should wait and it would not be necessary to
proceed with the option analysis. Otherwise, the analysis follows with the use of the
option model.
The effective uncertainty cs equals
\(.10*.10+.l10* 10 -2 .*.10*.10)=0.134
and the critical B/C ratio equals 1.348.
As expected the results are sensitive to the values chosen for uncertainty. If one
uncertainty equals 10 percent and the other equals 5 percent, then the effective
uncertainty equals 10.7 percent and the critical ratio equals 1.270. If one uncertainty
equals 10 percent and the other equals 15 percent, then the effective uncertainty equals
17.2 percent and the critical ratio equals 1.465. In an example similar to the base case
above for when two large investment alternatives of similar scale exist, Martzoukos
(1995b) shows that if the actual ratio of the second-best equals 1.0 (or 1.20 or 1.40), then
the critical ratio for the best investment equals 1.30 (or 1.41 or 1.57).

D. Applicationof the SimpleOptionsModelto ConstructionCostsand


Schedules
In this section, the options valuation approach to investment timing is applied to the
problem of uncertainty in construction costs and schedules, using the generic regression
equations for thermal and hydropower generation projects in developing countries that are
derived in this paper. The simple model given in section C of this annex is used for this
example.
Annex 11: Applyingthe OptionApproachto ConstructionCostsand Schedules 113

Dl. GeneralAssumptions
The following four assumptions apply to this example:
a. Uncertainty is measured around the expected value, so that a correction has been
applied for bias to the estimated value.
b. The linearized errors of the regressions are used as proxy for the ex ante uncertainty
used in the option model, although in fact they represent ex post tuncertainty. They
are specific to the time horizon (corrected for bias) for project completion, so in the
option models "option variance" = variance*t; this implies that the derived option
model uncertainty for the relevant parameter (schedule or cost) is a forecast error that
is specific to the time adjusted for bias, and is a forecast error increasing in respect to
time. The detailed derivation of option model standard deviations is more
complicated and follows in the next section for both the benefits (related to schedule
slip) and costs.
c. The relationship betwveennormal and lognormal distributions. When a variable is
normally distributed as N(g,c), the notation implies a normal distribution with mean
,u and standard deviation (.
If the logarithm of a variable is distributed N(I.,o), then the variable is said to be
lognormally distributed with mean equal to exp(ji+.5C2 ) and variance equal to
exp(2u+cy2 ){exp(a2 )-l 1. To derive a yearly option model a, a2 is replaced in the
equation above with a 2 t, where t is the best estimate for the time to project
completion and is adjusted for bias.
d. Definition of linearized error. The logarithm of the ratio of actual cost over predicted
(estimated) cost is the dependent regression variable:
Ln(C/Cpred) = Cregr+ error
Thus, by taking the exponent of both the left and right sides of the regression
equation:
C = exp(lnCpred + Cregr)+ (linearized error)J(Cpred)
The first expression on the right side equals the mean of a lognormal distribution and
has a standard deviation equal to the linearized error adjusted with 4(Cpred). A similar
result holds for schedule slippage.
D2. GeneralFormulationto EstimateOptionModel Uncertainty
This section derives the estimate of the uncertainty for the benefits and costs that can
be used in the option model. This estimate is case specific. It depends on the percent
correction for bias that is given by the regression model for both schedule and costs, and
the actual time predicted.
114 EstimatingConstructionCostsand Schedules

a. Approximating schedule uncertainty. Demand uncertainty and fuel price


uncertainty are usually considered to be the factors associated with benefit uncertainty.38
Construction slip uncertainty can augment these two factors, and can even replace them
on the assumption that:
1. Project timing given by the deterministic methodology is in general
optimal in terms of excess capacity, so that demand uncertainty for a
project of given capacity is small; and
2. Fuel price uncertainty is not reduced by waiting due to the long project
construction period.
By considering schedule slippage uncertainty as the relevant uncertainty for the
benefits (or, in any case, as one of the relevant uncertainties for the benefits) in the option
model, it is implicitly accepted that waiting will improve the estimate for the expected
time to complete the project. Since time to completion can only be a positive number,
this complies with the lognormal distribution assumption of the option model.
The benefits from the investment option are denoted as B, and they equal
exp(-tR)Num, where t equals the project completion time, R the continuous discount rate,
and Num a number that represents the expected benefits discounted to the end of the
project completion period.
First, the variance a2 of B is needed for the option model if the variance of t is known.
It is assumed that Num is not sensitive to the estimate of t. In the option model, the entity
B is lognormally distributed, thus InB equals -tR + ln(Num) and is normally distributed
with variance a 2t. So the variance of t equals a 2 t/R. Because t is the predicted time
corrected for bias, the variance equals a2 tpredexp(tregr)/R.
Second, the variance of t is estimated from the regression model. It is known that
thus "t/tpred" equals exp(tregr), with the standard deviation equal to the
ln(t/tpred) = tregr,
linearized error e multiplied by tp,d. So the standard deviation of t equals e'I(tpred).
From the above:
el(tpmd) = a'l[(tpredexp(tregr)/R]

from which:
a = e4[R/exp(tregr)] (Al 1.3)
(b) Approximating cost uncertainty. Cost uncertainty relates directly to the cost C of
exercising the investment option.
From the regression model ln(C/Cpred)= Cregr,

38. The considerableuncertaintythat exists in forecastsfor oil prices and powerdemandis confirmed
by the Bank's experience,as shownin Annex 12.
Annex I1: Applyingthe Option Approachto ConstructionCosts and Schedules 115

CCpred :=exp(Cregr)
with standard deviation equal to the linearized error e. Equivalently it is assumed that the
standard deviation of C equals eaCpred-
The option model gives the assumption that C is lognormally distributed, so InC
has a standard deviation of C2 t, where t equals the expected time and is adjusted for bias
to tpredexp(tregr).
The relation between the means of normal and lognormal distribution is used to
derive p:
C = Cpredexp(Cregr)= exp(Q+.5e 2 Cpred)
SO .L = InCpred + Cregr - .5e2 Cpred.

Finally, an approximate value of a is derived by solving iteratively the equation for


the relationship between the variances of the log-normal and the normal distributions:
e2 Cpred= exp(2g+a 2 t){exp(a 2 t)-l } (Al 1.4)
D3. TheCasesof ThermalandHydropower
Plants
The following linearized errors for the two main groups of power generation projects
were derived in chapter 6 of the paper:
Thermal Hydro
Schedule 0.220 0.200
Cost (constant) 0.179 0.259
The predicted and adjusted schedules are given by ln(t/tpred) = tregr, where t/tpred is the
average schedule slip derived from the regression equations (30 percent for thermal, 28
percent for hydro):
Thermal Hydro
tregr ln(1.30) = .262364 ln(1.28) = .24686
The predicted and adjusted costs are given by ln(C/Cpred)= Cgr, where C/Cp,d is the
average cost overrun derived from the regression equations (6 percent for thermal, 27
percent for hydro):
Thermal Hydro
Cregr ln(l.06) = .058269 In(1.27) = .239017
Assuming a continuous discount rate R of 10 percent, equation Al 1.3 is solved for
the investment option benefits CB:
Thermal Hydro
aB .061017 .055902
116 EstimatingConstructionCosts and Schedules

Equation A 11.4 is solved by numerical methods for the investment option cost ac:
Thermal Hydro
ac .0735 .0685
Assuming a squared correlation between cost overrun and schedule slip for thermal
power projects of .24 and for hydro power projects of .01 (Table 5.6 of the paper), which
imply correlations of .4899 and .1, the relevant standard deviations are computed from
equation A 11.2 and the critical B/C ratios are computed from equation AI 1.1 (section C),
as follows:
Thermal Hydro
CSB/C .069 .084
(B/C) 1.166 1.206
Thus the critical benefit-cost ratios that account just for uncertainty in construction
costs and schedules about expected (unbiased) values (assuming that the deterministic
optimal timing is already exceeded) are case specific. In general they are shown to be
higher for hydro power plants, mainly because of the lower correlation between costs and
schedules for hydropower projects.
To recapitulate the main insight from this application of the option approach, if
uncertainty changes in the way that is modeled, then the invest now option should be
exercised only when the estimated benefit/cost ratio exceeds 1.166 based on expected
(unbiased) values in the general case of thermal power generation projects, and exceeds
1.206 in a similar fashion for hydropower generation projects.
Annex 12: Performanceof PowerDemand Forecasts
and of World Bank Oil Price Projections
FigureA12.1 Performanceof PowerDemandForecastsfor DevelopingCountries
Deviation (percent)
60I

50 -

40

30 -- Mean-1 STD
- ' deviation
+Mean
20 -- Mean +1 STD

10

-10.
Year 1 Year3 Year 7 Year 10

Note: Deviationpercent is deviationof forecastdemand from actual demandas a proportion of actual


demandbasedon a sampleof about 200 powerdemandforecasts.Source: SanghviandVernstrom(1989).

FigureA12.2 World Bank Oil Price Projectionsin Constant1987 US$ per Barrel
Dollarsper barrel
70-
1980forecast 1982forecast
60-- 60 - 8' '- ' ~~1984
~~~~~~~ / forecast

50- -

40--

30 Actual /

20-
1978 forecast 1988forecast
10

1974 1979 1984 1989 1994 1999

Source:Crousillat(1989).

117
References
Arrow, K. J., and R. C. Lind. 1994. "Uncertainty and the Evaluation of Public Investment
Decisions." In Cost Benefit Analysis (2nd edition), ed. R. Layard and S. Glaister.
Cambridge: Cambridge University Press.
Black, Fischer, and Myron Scholes. 1973. "The Pricing of Options and Corporate Liabilities."
Journal of Political Economy 81: 637-59.
Brennan, Michael J., and Eduardo Schwartz. 1978. "Finite Difference Methods and Jump
Processes Arising in the Pricing of Contingent Claims: A Synthesis." Journal of
Financial and Quantitative Analysis 20:461-73.
1985. "Evaluating Natural Resource Investments." Journal of Business 58: 135-57.
Cox, John, and Stephen A. Ross. 1976. "The Valuation of Option for Alternative Stochastic
Processes." Journal of Financial Economics 3:145-166.
Crousillat, Enrique 0. 1989. "Incorporating Risk and Uncertainty in Power System Planning."
Energy Series Paper 17. World Bank, Industry and Energy Department, Washington,
D.C.
Crousillat, Enrique, and Spiros Martzoukos. 1991. "Decision Making Under Uncertainty: An
Option Valuation Approach to Power Planning." Energy Series Paper 39. World
Bank, Industry and Energy Department, Washington, D.C.
Dixit, Avinash. 1992. "Investment and Hysteresis." Journal of Economic Perspectives 6
(Winter): 107-32.
Dixit, Avinash, and Robert Pindyck. 1994. Investment under Uncertainty. Princeton, N.J.:
Princeton University Press.
EPRI (Electric Power Research Institute). 1995. Option Pricing for Project Evaluation: An
Introduction. TR-104755, Project 1920-05, Final Report, January.
Little, I. M. D., and J. A. Mirrlees. 1974. Project Appraisal and Planning for Developing
Countries. London: Heinemann.
Martzoukos, Spiros H., and Witold Teplitz-Sembitzky. 1992. "Optimal Timing of Transmission
Line Investments in the Face of Uncertain Demand." Energy Economics 14 (January):
3-10.
Martzoukos, Spiros H. 1995a. "Issues on Irreversibility and Investment using the Contingent
Claims (Real Options) Approach." Unpublished Dissertation, Finance Department,
The George Washington University, Washington, D.C.
1995b. "Rational Indecision and a Numerical Investigation of the Two-Dimensional
Free Boundary Problem." School of Business and Public Management Working Paper
95-16, The George Washington University, Washington, D.C.
McDonald, Robert, and Daniel Siegel. 1986. "The Value of Waiting to Invest," Quarterly
Journal of Economics 101:707-727.

119
120 Estimating Construction Costs and Schedules

Merrow, Edward W., and Ralph F. Shangraw, Jr., with Scott H. Kleinberg, Lisa A. Unterkofler,
Richard Madaleno, Edward J. Ziomkoski, and Brett R. Schroeder. 1990.
"Understanding the Costs and Schedules of World Bank Supported Hydroelectric
Projects." Energy Series Paper 31. World Bank, Industry and Energy Department,
Washington, D.C.
Merton, Robert. 1973. "The Theory of Rational Option Pricing." Bell Journal of Economics and
Management Science 4 (Spring): 141-83.
Moore, Edwin A., and George Smith. 1990. "Capital Expenditures for Electric Power in the
Developing Countries in the 1990s." Energy Series Paper 21. World Bank, Industry
and Energy Department, Washington, D.C.
Paddock, James L., Daniel R. Siegel, and James L. Smith. 1988. "Option Valuation of Claims
on Real Assets: The Case of Offshore Petroleum Leases." Quarterly Journal of
Economics 103 (August): 479-508.
Pindyck, Robert. 1991. "Irreversibility, Uncertainty and Investment." Journal of Economic
Literature 29 (September): 1110-52.
1995. "Sunk Costs and Benefits in Environmental Policy: I. Basic Theory." MIT
Energy Laboratory Working Paper MIT-CEEPR 95-003, Cambridge, Mass., March.
Samuelson, Paul. 1967. "Rational Theory of Warrant Pricing." Industrial Management Review
6 (Spring): 41-50.
Sanghvi, Arun, and Robert Vernstrom, with John Besant-Jones. 1989. "Review and Evaluation
of Historic Electricity Forecasting Experience (1960-1985)." Energy Series Paper 18.
World Bank, Industry and Energy Department, Washington, D.C.
Schramm, Gunter. 1989. "Optimal Timing of Transmission Line Investment." Energy
Economics 11 (July).
Teplitz-Sembitzky, Witold. 1989. "Option Pricing in Power Planning." Draft working paper,
Industry and Energy Department, The World Bank, Washington, D.C.
World Bank. 1984. Geological Problems and Cost Overruns: A Survey of Bank-Financed
Hydroelectric Projects, World Bank Energy Department Note 61, July.
Operations Evaluation Department. 1992. Annual Review of Evaluation Results 1991.
Washington, D.C.
Distributors
DiStributors of
of ~~~~~~CHINA
ChinaFinancial&Econom,c
GREECE
PapasotinouS.A
ITALY
LcosaCommissonana SansoinSPA
NORWAY
NarvesenInformation
Center SINGAPORE, TAIWAN, VanDienmen Edntons
Techniq
World Bank Publishing
House 35.Soumarna Sir ViaDucaDi Calabna,1/1 BookDepartment MYANMAR, BRUNEI Ch.de Lacuez41
8. DaFo SCDongJie 10682Athens CasellaPostale552 PO Box6125EttersLad AsahgatePublishing
Asia CH1807Blonay
Publications Beijing Tel (1)364-1826 50125Firenze N-0602Oslo6 PacificPe Ltd Tel (021)9432673
Pricesand credit terns c1arti Tel (11)
333-8257 Fax (1) 364-8254 Te. (55)645-415 Tel (22)57-3300 41 KallangPuddingRoad004-03 Fax. (021)9433605
front country to cozntry. Fan (1)401-7365 Fax (55) 641-257 Fax (22)68-1901 GoldenWheelBuilding
from country to country. HONGKONG,MACAO Singapore 349316 TANZANIA
Consult yo;r local distributor COLOMBIA Asia2000Ltd JAMAICA PAKISTAN Tel (65)741-5166 OxCordUniversity
Press
beforeplacingan order IntoenlaceLUda. Sales& Circulation
Department tanRandlePubiishers
Ltd. MirzaBookAgency Fax (65)742-9356 MaktabaStreet
Apartado Aereo34270 SeabirdHouse unal1101-02 206OldHopeRoad 65.Shahrah-e-Quaid-e-Azarn e-mailashgate@asianconnen.com PO Box5299
BogotaD E. 22-28Wyndham Street,Central Kiopston6 PO BoxNo.729 Dar es Salaam
ALBANIA Tel.(1) 285-2798 HongKong Tel 809-927-2085 Lahore54000 SLOVAK REPUBLIC Tel (51)29200
AdnonLtd Faxa(1)285-2798 Tel 8522530-1409 Fax 809-977-0243 Tel (42)7353601 SlovanGT.G.Ltd. Fax(51)46622
PedatRexhepStr Fax 8522526-1107 Fax. (42)7585263 Krupinska4
Pall9, Shk.1Ap 4 COTED'IVOIRE URL:http/Aw.salesOasa2000.com hk JAPAN POBox152 THAILAND
Tirana CentredEdinionel de Diftusion EastemBookServce OxfordUivenrsiryPress 85299 Bratislava
5 CentralBooksDistribution
Tel (42)27419;22172 Alricaines(CEDA) HUNGARY Hongo3-Chome. 5 BangaloreTown Tel.(7)839472 306SilomRoad
Fax. (42)27419 04 B.P 541 Foundation
for Market Bunkyo-ku113 SharaeFaisal Fax.(7)839485 Bangkok
Abidjan04 Plaleau Economy Tokyo PO Box13033 Tel. (21235-5400
ARGENTINA Tel 225-24-6510 DombovariUt 17-19 Tel. (03)3818-0661 Karachl-75350 SOUTHAFRICA, BOTSWANA Fax (2)237-8321
OficinadelLibrolntremacionat Fax. 225-25-0567 H-1117Budapest Fax- (03)3816-064 Tet. (21)446307 Forsngle lt7es:
Av.Cordoba 1877 Tel 361 2042951or URLhttpiJwww)bekkoame
or.Ip-svl-ebs Fax (21)454-7640 OxfordUniversityPress TRINIDAD & TOBAGO,JAM.
1120BuenosAnres CYPRUS 3612042948 E-mail.oup@oup khienum.com.pk Southem Alnca SystemalicsStudiesUnit
Tel' (1)815-8156 CenterofAppliedResearch Fax:361204 2953 KENYA PO Box1141 #9 WaftsSreet
Fax. (1) 615-8354 CypnusCollege AfricaBookService(E.A( Ld. PERU Cape Town60B0 Curepe
T
6,DiogenesStreet,Engomi INDIA OuaranHouse,MfanganoStreet EdnonalDesarrollo
SA el. (21)45-7266 Trinidad,WestIndies
AUSTRALIA, FIJI,PAPUANEWGUINEA, PO Box2006 AlliedPublishers
LId. PO Box45245 Apartado3824 Fax (21)45-7265 Tel 809-662-5654
SOLOMON ISLANDS, VANUATU,
AND Nicosa 751MountRoad Nairobi Lima1 Fax. 609-662-5854
WESTERN SAMOA Tel 244-1730 Madras-860002 Tel.(2)23641 Tel (14)285380 Forsubcnptironao@ers
DA. InfontoabonServices Fax. 246-2051 Tel (44)852-3938 Fax (2)330272 Fax (14)286628 International
Subscnption
Service UGANDA
648Whnehrxse Road Fax: (44)852-0649 PO. Box41095 GusfroLtd.
Mncham 3132 CZECHREPUBLIC KOREA,REPUBLIC OF PHILIPPINES CraighaWI Madhvani Building
Vrctona NationalInlormation
Center INDONESIA Da0qon TradingCo.Lid lntemational
Booksounce
CenterInc Johannesburg 2024 POBox9997
Tel. (61)392107777 prodelna.Konviktska5 Pt.IndiraLimited P.O.Box34 Sure 720,Catyand10 Tel' (11)680-1448 Plot16/4JinjaRd
Fax. (61)392107788 CS- 11357 PragueI JalanBorobudur 20 Yeoeida Condominium Tower2 Fax: (11)680-6248 Kampala
URL hntpiJ/ww.dadirect.com
au Tel (2) 2422-9433 PO.Box181 Seoul H.VdelaCosta,comer TeVFax:(41)254763
Fax. (2)2422-1484 Jakarta10320 Tet:(2) 785-1631/4 VateroSt. SPAIN
AUSTRIA URL'httpiJwww.nis.ci Tel.(21()390-4290 Fax.(2) 764-315 Makati,Meto Manila Mundi-PrensaLibos, S.A UNITEDKINGDOM
GeroldandCo. Fax. (21)4214289 Tel: (2)817-9676 Castello37 MicroafoLtd
Graben31 DENMARK MALAYSIA Fax (2)817-1741 28001Madnd PO. Box3
A-1011Wien SamfundsaLtteratur IRAN University
of MalayaCooperahve Tel-(1)431-3399 Anon,HampshireGU342PG
Tel' (1)533-50-14-0 RosenoemsAlld 11 KowkabPublishers BookshopLiUmied POLAND Fax (1)575-3996 England
Fax (1)512-47-31-29 DK-1970Fredenksberg
C PO Box 19575-511 PO.Box 1127 International
Publishing
Service http:/hwwtsai.es/mprensa Tel (1420)6848
Tel.(31)-351942 Tehran JalanPanlaiBaou Ul Piekna31137 Fax: (1420)89889
BANGLADESH Fax: (31)-35782 Tel (21)258-3723 59790 KualaLumpur 00-577Warzawa Mundb-PrensaBarcelona
T
Mico IidusinesDevelopmenl Fax 98 (21)258-3723 el: (3)756-5000 Tel (2)628-6089 ConselldeCent,391 ZAMBIA
AssislanceSwcety(MIDAS) EGYPT, ARABREPUBUC
OF Fax (3)755-4424 Fax: (2)621-7255 08009Barcelona UniversityBookshop
House5,Road16 AtAhrmm KetabSaraCo. Pubhshers Tel:(3)488-3009 GreatEastRoadCampus
Dhanmordi FlArea A/GaaalStreet KhaledEslambok Ave., MEXICO PORTUGAL Fax-(3)487-7659 PO. Box32379
Dhaka1209 Cairo 6thShreet INFOTEC LivranaPonugal Lusaka
Tel. (2)326427 Tel (2) 576-6083 KushehDelafmozNo 8 AparladoPostal22-860 RuaDoCarmo70-74 SRILANKA,THEMALDIVES Tel.(1I 213221
EAt.482
Fax: (2)81118B Fax. (2) 578-6833 Tehran 14060Tialpan. 1200Lisbon LakeHouseBookshop
Tel 8717819or8716104 MexicoD.F Tel: (1)347-4982 P.OBox244 ZIMBABWE
BELGIUM TheMiddleEastObserer Fax. 8862479 Tel (5)606-0011 Fax: (1)347-0264 190,SirChittampalam
A Lorngman Zimbabwe
(Pte.)ULd
JeanDe Lannoy 41, Shen Stre E-mail:ketab-aora@neda.set.ir Fax: (5)624-2822 Gardiner Mawatha ToudeRoad,Ardbonnie
Av.duRo 202 Cairo ROMANIA Colombo 2 PO. BoxoST125
1060Bnrsels Tel (2) 393-9732 IRELAND NETHERLANDS Compans De Lbrani BucuestiS.A. Tel:(1) 32105 Southenon
Tel: (2)538-5169 Fax. (2)393-9732 Govemment Supplies
Agency De LindeboonVtnOr-Publikaties Sir Lipsaonino 26, ctor 3 Fax (1)432104 Harare
Fax: (2)538-841 OdigantSolAthair P.O.Box 202 Bucharest Tel' (4)6216617
FINLAND 4-SHabnourIRoad 7480AEHaakslbenoe Tel. (1)6139645 SWEDEN Fax. (4)621670
BRAZIL Akateemiren Kuqakauppa DClin 2 Tel. (53)574-004 Fax.(1)3124000 Frites CustomerServce
PublcacbesTmcncas lntemaoonais PO.Box23 Tel. (1)461-3111 Fax: (53(572-9296 Regenngsgpon 12
Ltda. FIN-00371Helsanbl Fax: (1)475-2670 RUSSIANFEDERATION S-10647Stockhotm
RuaPa/xono Gomide,209 Tel.(0(12141 NEWZEALAND lsdalelsoto VesMir> Tel' (86(909090
01409SaoPalo, SP Fax: (0) 1214441 ISRAEL EBSCONZLtd. 9a,KelpachoryPereulok Fax: (8)2147 77
Tel: (11)259-6644 URL:IttIpJlboAoknetculbe.ttleka/ YoznotLiteralune
Ud. PrivateMai Bag99914 Moscow101631
Fa: (11)258-699 P.O.Box 56055 NewMarket Tel (95)91787 49 Wernnergren-Williams
AB
FRANCE TelAvrv61560 Aucklarid Fax (95( 91792 59 P O.Box 1305
CANADA WorldBankPubicadions Tel. (3)5285-397 Tel: (9)524-8119 S-17125Solna
RenoulPubittag Co.LUd 66,avenue d'lana Fax: (3(5285-397 Fax: (9)524-8067 SAUDIARABIA,fATAR Tel: (8)705-97-50
1294AlgomaRoad 75116Paris JeanBookStore Fax: (8)27-0W71
Ottawa,Ontw K1B3Y8 Tel (1) 40-69-30-56vS7 ROY.Ontkrmabonal NIGERIA PO.Box3196
Tel:613-7414=3 Fax: (1)40-69-30-68 POBox13056 UnversmityPressLnUmed Riyadh11471 SWTrZERLAND
Fax:613-741-5439 TelAovt 61130 ThreeCrnwnis BuildingJicho Tel:(1(477-3140 LbainePayno
GERMANY Tel:(3) 5461423 PrvateMailBag5095 Fax: (1) 477-2940 SenrvceInstitobonnel
UNO-Verblag Faa:(3) 5461442 Ibadan C6tes-d-Montbenon 30
PoppedorterA/lee55 Tel: (22)41-1356 1002Lausanrse
53115Bon Pasestbme
anAut1ontylMtidle
East Fax (22)41-2056 Tel: (021)-341-3229
Tel' (228)212940 IndexInformabon
Semices Fax: (0211-341-3235
Fax: (2281217492 PO.B 19502Jensalem
RECENT WORLD BANK TECHNICAL PAPERS (continued)

No. 288 Biggs,Shaw, and Srivastiva, Technological


Capabilitiesand Learningin AfricanEnterprises
No. 289 Dinar, Seidl, Olem, Jorden, Duda, and Johnson, Restoringand Protectingthe World'sLakesand Reservoirs
No. 290 Weijenberg, Dagg, Kampen Kalunda, Mailu, Ketema, Navarro, and Abdi Noor, StrengtheningNationalAgricultual
ResearchSystems in Easternand CentralAfrica:A Framework-for Action
No. 291 Valdes and Schaeffer in collaboration with Errazuriz and Francisco, Survleillaniceof Agricultural Price and Trade
Policies:A HandbookforChile
No. 292 Gorriz, Subramanian, and Simas, IrrigationManagementTransferin Mexico:Processand Progress
No. 293 Preker and Feachem, MarketMechanismsand the Health Sectorin Centraland EasternEurope
No. 294 Valdes and Schaeffer in collaboration with Sturzenegger and Bebczuk, Surveillanceof
AgriculturalPriceand TradePolicies:A HandbookforArgentina
No. 295 Pohl, Jedrzejczak, and Anderson, CreatingCapitalMarketsin Centraland EasternEurope
No. 296 Stassen, Small-ScaleBiomassGasifiersforHeat and Polwer:
A GlobalReviezw
No. 297 Bulatao, Key IndicatorsforFamilyPlanningProjects
No. 298 Odaga and Heneveld, Girlsand Schoolsin Sub-SaharanAfrica:FromAnalysisto Action
No. 299 Tamale, Jones, and Pswarayi-Riddihough, Technologies Relatedto ParticipatoryForestryin Tropical
and SubtropicalCountries
No. 300 Oram and de Haan, TeclhnologiesforRainifedAgriculturein MediterraneanClimates:A Reviewof WorldBank Experiences
No. 301 Edited by Mohan, Bibliographyof Publications:TechnicalDepartment,AfricaRegion,JulyI1987to April 1995
No. 302 Baldry,Calamari, and Yameogo, EnvironmentalImpactAssessmentof Settlementand Developmentin the UpperLeraba
Basin
No. 303 Heneveld and Craig, SchoolsCount:WorldBank ProjectDesignsand the Quality of PrimaryEducationin
Sub-SaharanAfrica
No. 304 Foley, Phiotovoltaic
Applicationsin RuralAreas of the DevelopingWorld
No. 305 Johnson, Educationand Trainingof Accountants in Sub-SaharanAnglophoneAfrica
No. 306 Muir and Saba, ImprovingState EnterprisePerformance: The Role of Internaland ExternalIncentives
No. 307 Narayan, TowardParticipatoryResearch
No. 308 Adamson and others, EnergyUse,Air Pollution,and EnvironmentalPolicyin Krakow:Can EconomicIncentivesReally
Help?
No. 309 The World Bank/FOA/UNIDO/Industry Fertilizer Working Group, Worldand RegionalSupply and DemandBalances
for Nitrogen,Phosphate,and Potash,1993/94-1999/2000
No. 310 Edited by Elder and Cooley, SustainableSettlementand Developmentof the Onchocerciasis Control
ProgrammeArea:Proceedingsof a MinisterialMeeting
No. 311 Webster, Riopelle and Chidzero, WorldBank LendinigforSrnallEnterprises1989-1993
No. 312 Benoit, ProjectFinanceat the WorldBank:An Overviewof Policiesand Instruments
No. 313 Kapur, Airport Infrastructure:The EmergingRoleof the Private Sector
No. 314 Valdes, Schaefferin collaboration with Ramos, Surveillanceof AgriculturalPriceand TradePolicies:A Handbookfor
Ecuador
No. 316 Schware and Kimberley, InformationTechnologyand NationalTradeFacilitation:Making thteMost of GlobalTrade
No. 317 Schware and Kimberley, InformationTechnologyand NationalTradeFacilitation:Guide to Best Practice
No. 318 Taylor, Boukambou, Dahniya, Ouayogode, Ayling, Abdi Noor, and Toure, StrengtheningNationalAgricuIltural
ResearchSystems in the Humid and Sub-humidZones of West and CentralAfrica:A FrameworkforAction
No. 320 Srivastava, Lambert and Vietmeyer,MedicinalPlants:AnI ExpandingRolein Development
No. 321 Srivastava, Smith, and Forno, Biodiversityand Agriculture:ImplicationsforConservationand Development
No. 322 Charles M. Peters, The Ecologyand Managementof Non-TimberForestResources
No. 323 Edited by Dominique Pannier. CorporateGovernanceof PublicEnterprisesin TransitionalEconomies
No. 324 Cabraal, Cosgrove-Davies, and Schaeffer.Best PracticesforPhotovoltaicHouseholdElectrificationPrograms
| ~~~~~~~~~~~~~~~~~~I
| J.i

I~~~~~~~~~~~~~~~~~~~~~~~702 336700

Вам также может понравиться