This article presents results from the
first statistically significant study of
cost escalation in transportation infrastructure projects. Based on a sample of 258 transportation infrastructure projects worth US$90 billion and
representing different project types,
geographical regions, and historical
periods, it is found with overwhelming
statistical significance that the cost estimates used to decide whether such
projects should be built are highly and
systematically misleading. Underestimation cannot be explained by error
and is best explained by strategic misrepresentation, that is, lying. The policy implications are clear: legislators,
administrators, investors, media representatives, and members of the public who value honest numbers should
not trust cost estimates and cost-benefit analyses produced by project promoters and their analysts.
Flyvbjerg is a professor of planning with
the Department of Development and Planning, Aalborg University, Denmark. He is
founder and director of the university’s research program on transportation infrastructure planning and was twice a Visiting
Fulbright Scholar to the U.S. His latest
books are Rationality and Power (University
of Chicago Press, 1998) and Making Social
Science Matter (Cambridge University Press,
2001). He is currently working on a book
about megaprojects and risk (Cambridge
University Press). Holm is an assistant professor of planning with the Department of
Development and Planning, Aalborg University, and a research associate with the
university’s research program on transportation infrastructure planning. Her main interest is economic appraisal of projects.
Buhl is an associate professor with the Department of Mathematics, Aalborg University, and an associate statistician with the
university’s research program on transportation infrastructure planning.

Underestimating
Costs in Public
Works Projects
Error or Lie?
Bent Flyvbjerg, Mette Skamris Holm, and Søren Buhl

C

omparative studies of actual and estimated costs in transportation
infrastructure development are few. Where such studies exist, they are
typically single-case studies or they cover a sample of projects too
small to allow systematic, statistical analyses (Bruzelius et al., 1998; Fouracre
et al., 1990; Hall, 1980; Nijkamp & Ubbels, 1999; Pickrell, 1990; Skamris &
Flyvbjerg, 1997; Szyliowicz & Goetz, 1995; Walmsley & Pickett, 1992). To
our knowledge, only one study exists that, with a sample of 66 transportation projects, approaches a large-sample study and takes a first step toward
valid statistical analysis (Merewitz, 1973a, 1973b).1 Despite their many merits in other respects, these studies have not produced statistically valid answers regarding the question of whether one can trust the cost estimates
used by decision makers and investors in deciding whether or not to build
new transportation infrastructure. Because of the small and uneven samples used in existing studies, different studies even point in opposite directions, and researchers consequently disagree regarding the credibility of cost
estimates. Pickrell (1990), for instance, concludes that cost estimates are
highly inaccurate, with actual costs being typically much higher than estimated costs, while Nijkamp and Ubbels (1999) claim that cost estimates are
rather correct. Below we will see who is right.
The objective of the study reported here was to answer the following
questions in a statistically valid manner: How common and how large are
differences between actual and estimated costs in transportation infrastructure projects? Are the differences significant? Are they simply random
errors? Or is there a statistical pattern to the differences that suggests other
explanations? What are the implications for policy and decision making regarding transportation infrastructure development?

Journal of the American Planning Association,
Vol. 68, No. 3, Summer 2002. © American
Planning Association, Chicago, IL.

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 279

 BENT FLYVBJERG, METTE SKAMRIS HOLM, AND SØREN BUHL

Four Steps to Understanding
Deceptive Cost Estimation
We see four steps in the evolution of a body of scholarly research aimed at understanding practices of cost
underestimation and deception in decision making for
transportation infrastructure. The first step was taken
by Pickrell (1990) and Fouracre, Allport, and Thomson
(1990), who provided sound evidence for a small number
of urban rail projects that substantial cost underestimation is a problem, and who implied that such underestimation may be caused by deception on the part of project promoters and forecasters. The second step was
taken by Wachs (1990), who established—again for a
small sample of urban rail projects—that lying, understood as intentional deception, is, in fact, an important
cause of cost underestimation. Wachs began the difficult task of charting who does the lying, why it occurs,
what the ethical implications are, etc.
The problem with the research in the first two steps
is that it is based on too few cases to be statistically significant; the pattern found may be due to random properties of the small samples involved. This problem is
solved in the third step, taken with the work reported in
this article. Based on a large sample of transportation infrastructure projects, we show that (1) the pattern of cost
underestimation uncovered by Pickrell and others is of
general import and is statistically significant, and (2) the
pattern holds for different project types, different geographical regions, and different historical periods. We
also show that the large-sample pattern of cost underestimation uncovered by us lends statistical support to the
conclusions about lying and cost underestimation arrived at by Wachs for his small sample.
The fourth and final step in understanding cost underestimation and deception would be to do for a large
sample of different transportation infrastructure projects what Wachs did for his small sample of urban rail
projects: establish whether systematic deception actually takes place, who does the deception, why it occurs,
etc. This may be done by having a large number of forecasters and project promoters, representing a large number of projects, directly express, in interviews or surveys,
their intentions with and reasons for underestimating
costs. This is a key topic for further research.
In sum, then, we do not claim with this article to
have provided final proof that lying is the main cause of
cost underestimation in transportation infrastructure
projects. We claim, however, to have taken one significant step in a cumulative research process for testing
whether this is the case by establishing the best and
largest set of data about cost underestimation in transportation infrastructure planning so far seen, by carry-

280 APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3

ing out the first statistically significant study of the issues involved, and by establishing that our data support
and give statistical significance to theses about lying developed in other research for smaller, statistically nonsignificant samples.
As part of further developing our understanding of
cost underestimation, it would also be interesting to
study the differences between projects that are approved
on a competitive basis, by voters at an election, and those
that are funded through formula-based allocations. One
may speculate that there is an obvious incentive to make
a project look better, and hence to underestimate costs,
in the campaign leading up to an election. A good single-case study of this is Kain’s (1990) article about a rail
transit project in Dallas. Votes are cast more often for
large rail, bridge, and tunnel projects than for road projects. For example, most U.S. highway funds are distributed to states based on a formula (i.e., there is no competitive process). A state department of transportation
(DOT) is likely to have a fixed annual budget for construction. The DOT leadership would presumably want
fairly accurate cost estimates before allocating the budget. One may speculate that large cost underestimation
is less likely in this situation. There are exceptions to this
scenario. Sometimes DOT officials want to persuade
state legislators to increase their budget. And states occasionally submit bond issue proposals to voters. In Europe, the situation is similar on important points, although differences also exist. This may explain the result
found below, that cost underestimation is substantially
lower for roads than for rail, bridges, and tunnels, and
that this is the case both in the U.S. and Europe. Needless to say, more research is necessary to substantiate this
observation.
Finally, we want to emphasize that although the
project sample used in this study is the largest of its kind,
it is still too small to allow more than a few subdivisions,
if comparative statistical analyses must still be possible.
Therefore, in further work on understanding cost underestimation, the sample should be enlarged to better
represent different types of projects and different geographical locations. As to project types, data for more
private projects would be particularly useful in allowing
statistically valid comparisons between public and private sector projects. Such comparisons do not exist
today, and nobody knows whether private projects perform better or worse than public ones regarding cost underestimation. The sample should also be enlarged to
contain data for more fixed-link and rail projects. Such
data would allow a better (i.e., a statistically corroborated) comparative understanding of cost underestimation for more specific subtypes of projects such as
bridges, tunnels, high-speed rail, urban rail, and conven-

 UNDERESTIMATING COSTS IN PUBLIC WORKS PROJECTS

tional rail. Such an understanding is nonexistent today.
As to geography, immediate rewards would be gained
from data for projects outside Europe and North America, especially for fixed links and roads. But even for Europe and North America, data on more projects are
needed to allow better comparative analysis.

Measuring Cost Inaccuracy
The methods used in our study are described in the
Appendix. All costs are construction costs. We follow international convention and measure the inaccuracy of
cost estimates as so-called “cost escalation” (often also
called “cost overrun”; i.e., actual costs minus estimated
costs in percent of estimated costs). Actual costs are defined as real, accounted construction costs determined
at the time of project completion. Estimated costs are defined as budgeted, or forecasted, construction costs at the
time of decision to build. Although the project planning
process varies with project type, country, and time, it is
typically possible for any given project to identify a specific point in the process as the time of decision to build.
Usually a cost estimate was available at this point in time
for the decision makers. If not, then the closest available
estimate was used, typically a later estimate resulting in a
conservative bias in our measure for inaccuracy (see the
Appendix). All costs are calculated in fixed prices in Euros
by using the appropriate historical, sectoral, and geographical indices for discounting and the appropriate exchange rates for conversion between currencies.
Project promoters and their analysts sometimes object to this way of measuring cost inaccuracy (Flyvbjerg
et al., in press). Various cost estimates are made at different stages of the process: project planning, decision
to build, tendering, contracting, and later renegotiations. Cost estimates at each successive stage typically
progress toward a smaller number of options, greater detail of designs, greater accuracy of quantities, and better
information about unit price. Thus, cost estimates become more accurate over time, and the cost estimate at
the time of making the decision to build is far from final.
It is only to be expected, therefore, that such an early estimate would be highly inaccurate. And this estimate
would be unfair as the basis for assessing the accuracy
of cost forecasting, or so the objection against using the
time-of-decision-to-build estimate goes (Simon, 1991).
We defend this method, however, because when the
focus is on decision making, and hence on the accuracy
of the information available to decision makers, then it
is exactly the cost estimate at the time of making the decision to build that is of primary interest. Otherwise it
would be impossible to evaluate whether decisions are
informed or not. Estimates made after the decision to

build are by definition irrelevant to this decision. Whatever the reasons are for cost increases after decision makers give the go-ahead to build a project, or however large
such increases are, legislators and citizens—or private investors in the case of privately funded projects—are entitled to know the uncertainty of budgets. Otherwise
transparency and accountability suffer. We furthermore
observe that if the inaccuracy of early cost estimates were
simply a matter of incomplete information and inherent difficulties in predicting a distant future, as project
promoters often say it is, then we would expect inaccuracies to be random or close to random. Inaccuracies,
however, have a striking and highly interesting bias, as
we will see below.
Another objection to using cost at the time of decision to build as a basis of comparison is that this supposedly would entail the classical error of comparing apples and oranges. Projects change over the planning and
implementation process. When, for instance, the physical configuration of the original Los Angeles Blue Line
Light Rail project was altered at substantial cost to comprise grade-crossing improvements, upgrading of adjacent streets, better sidewalks, new fences, etc., the project was no longer the same. It was, instead, a new and
safer project, and comparing the costs of this project
with the costs of the older, less safe one would supposedly entail the apples-and-oranges error. A problem with
this argument is that existing research indicates that
project promoters routinely ignore, hide, or otherwise
leave out important project costs and risks in order to
make total costs appear low (Flyvbjerg et al., in press;
Wachs, 1989, 1990). For instance, environmental and
safety concerns may initially be ignored, even though
they will have to be taken into account later in the project cycle if the project lives on, and the project is more
likely to live on if environmental and safety concerns are
initially ignored. Similarly, ignoring or underplaying
geological risks may be helpful in getting projects approved, and no other risk is more likely to boomerang
back and haunt projects during construction. “Salami
tactics” is the popular name used to describe the practice of introducing project components and risks one
slice at a time in order to make costs appear low as long
as possible. If such tactics are indeed a main mechanism
in cost underestimation, as existing research indicates,
then, clearly, comparing actual project costs with estimated costs at the time of decision to build does not entail the error of comparing apples and oranges but is
simply a way of tracking how what was said to be a small,
inexpensive apple turned out to actually be a big, expensive one.
Finally, we observe that if we were to follow the objections against using the cost estimate at the time of de-

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 281

 BENT FLYVBJERG, METTE SKAMRIS HOLM, AND SØREN BUHL

cision to build as the basis of tracking cost escalation, it
would be impossible to make meaningful comparisons
of costs because no common standard of comparison
would be available. We also observe that this method is
the international standard for measuring inaccuracy of
cost estimates (Fouracre et al., 1990; Leavitt et al., 1993;
National Audit Office & Department of Transport,
1992; Nijkamp & Ubbels, 1999; Pickrell, 1990; Walmsley & Pickett, 1992; World Bank, 1994). This standard
conveniently allows meaningful and consistent comparisons within individual projects and across projects,
project types, and geographical areas. This standard,
then, is employed below to measure the inaccuracy of
cost estimates in 258 transportation infrastructure projects worth US$90 billion.

Inaccuracy of Cost Estimates
Figure 1 shows a histogram with the distribution of
inaccuracies of cost estimates. If errors in estimating
costs were small, the histogram would be narrowly concentrated around zero. If errors in overestimating costs
were of the same size and frequency as errors in underestimating costs, the histogram would be symmetrically
distributed around zero. Neither is the case. We make
the following observations regarding the distribution of
inaccuracies of construction cost estimates:
• Costs are underestimated in almost 9 out of 10
projects. For a randomly selected project, the likelihood of actual costs being larger than estimated
costs is 86%. The likelihood of actual costs being
lower than or equal to estimated costs is 14%.
• Actual costs are on average 28% higher than
estimated costs (sd=39).
• We reject with overwhelming significance the
thesis that the error of overestimating costs is as
common as the error of underestimating costs
(p<0.001; two-sided test, using the binomial
distribution). Estimated costs are biased, and the
bias is caused by systematic underestimation.
• We reject with overwhelming significance the
thesis that the numerical size of the error of
underestimating costs is the same as the
numerical size of the error of overestimating costs
(p<0.001; nonparametric Mann-Whitney test).
Costs are not only underestimated much more
often than they are overestimated or correct, costs
that have been underestimated are also wrong by a
substantially larger margin than costs that have
been overestimated.
We conclude that the error of underestimating costs is
significantly much more common and much larger than

282 APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3

the error of overestimating costs. Underestimation of
costs at the time of decision to build is the rule rather
than the exception for transportation infrastructure projects. Frequent and substantial cost escalation is the
result.

Cost Underestimation by Project
Type
In this section, we discuss whether different types of
projects perform differently with respect to cost underestimation. Figure 2 shows histograms with inaccuracies of cost estimates for each of the following project
types: (1) rail (high-speed; urban; and conventional,
inter-city rail), (2) fixed link (bridges and tunnels), and
(3) road (highways and freeways). Table 1 shows the expected (average) inaccuracy and standard deviation for
each type of project.
Statistical analyses of the data in Table 1 show both
means and standard deviations to be different with a
high level of significance. Rail projects incur the highest
difference between actual and estimated costs, with an
average of no less than 44.7%, followed by fixed-link projects averaging 33.8% and roads at 20.4%. An F-test falsifies the null hypothesis at a very high level of statistical
significance that type of project has no effect on percentage cost escalation (p<0.001). Project type matters.
The substantial and significant differences among project types indicate that pooling the three types of projects
in statistical analyses, as we did above, is strictly not appropriate. Therefore, in the analyses that follow, each
type of project will be considered separately.
Based on the available evidence, we conclude that
rail promoters appear to be particularly prone to cost
underestimation, followed by promoters of fixed-link
projects. Promoters of road projects appear to be relatively less inclined to underestimate costs, although actual costs are higher than estimated costs much more
often than not for road projects as well.
Further subdivisions of the sample indicate that
high-speed rail tops the list of cost underestimation, followed by urban and conventional rail, in that order. Similarly, cost underestimation appears to be larger for tunnels than for bridges. These results suggest that the
complexities of technology and geology might have an
effect on cost underestimation. These results are not statistically significant, however. Even if the sample is the
largest of its kind, it is too small to allow repeated subdivisions and still produce significant results. This problem can be solved only by further data collection from
more projects.
We conclude that the question of whether there are
significant differences in the practice of cost underesti-

 UNDERESTIMATING COSTS IN PUBLIC WORKS PROJECTS

FIGURE 1. Inaccuracy of cost estimates in 258 transportation infrastructure projects (fixed prices).

TABLE 1. Inaccuracy of transportation project cost estimates by type of project
(fixed prices).

Project type

Number of
cases (N)

Average cost
escalation (%)

Standard
deviation

Level of
significance (p)

Rail
Fixed-link
Road
All projects

58
33
167
258

44.7
33.8
20.4
27.6

38.4
62.4
29.9
38.7

<0.001
<0.004
<0.001
<0.001

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 283

 BENT FLYVBJERG, METTE SKAMRIS HOLM, AND SØREN BUHL

FIGURE 2. Inaccuracy of cost estimates in rail, fixed-link, and road projects (fixed prices).

284 APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3

 UNDERESTIMATING COSTS IN PUBLIC WORKS PROJECTS

mation among rail, fixed-link, and road projects must be
answered in the affirmative. The average difference between actual and estimated costs for rail projects is substantially and significantly higher than that for roads,
with fixed-link projects in a statistically nonsignificant
middle position. The average inaccuracy for rail projects
is more than twice that for roads, resulting in average
cost escalations for rail more than double that for roads.
For all three project types, the evidence shows that it is
sound advice for policy and decision makers as well as
investors, bankers, media, and the public to take any
estimate of construction costs with a grain of salt, especially for rail and fixed-link projects.

Cost Underestimation by
Geographical Location
In addition to testing whether cost underestimation
differs for different kinds of projects, we also tested
whether it varies with geographical location among Europe, North America, and “other geographical areas” (a
group of 10 developing nations plus Japan). Table 2
shows the differences between actual and estimated
costs in these three areas for rail, fixed-link, and road projects. There is no indication of statistical interaction between geographical area and type of project. We therefore consider the effects from these variables on cost
underestimation separately. For all projects, we find that
the difference between geographical areas in terms of underestimation is highly significant (p<0.001). Geography
matters to cost underestimation.
If Europe and North America are compared separately, which is compulsory for fixed links and roads because no observations exist for these projects in other geographical areas, comparisons can be made by t-tests (as
the standard deviations are rather different, the Welch
version is used). For fixed-link projects, the average difference between actual and estimated costs is 43.4% in

Europe versus 25.7% North America, but the difference
between the two geographical areas is nonsignificant
(p=0.414). Given the limited number of observations and
the large standard deviations for fixed-link projects, we
would need to enlarge the sample with more fixed-link
projects in Europe and North America in order to test
whether the differences might be significant for more
observations. For rail projects, the average difference between actual and estimated costs is 34.2% in Europe versus 40.8% in North America. For road projects, the similar numbers are 22.4% versus 8.4%. Again, the differences
between geographical areas are nonsignificant (p=0.510
and p=0.184, respectively).
We conclude, accordingly, that the highly significant
differences we found above for geographical location
come from projects in the “other geographical areas” category. The average difference between actual and estimated costs in this category is a hefty 64.6%.

Have Estimates Improved Over
Time?
In the previous two sections, we saw how cost underestimation varies with project type and geography. In
this section, we conclude the statistical analyses by
studying how underestimation has varied over time. We
ask and answer the question of whether project promoters and forecasters have become more or less inclined
over time to underestimate the costs of transportation
infrastructure projects. If underestimation were unintentional and related to lack of experience or faulty
methods in estimating and forecasting costs, then, a priori, we would expect underestimation to decrease over
time as better methods were developed and more experience gained through the planning and implementation
of more infrastructure projects.
Figure 3 shows a plot of the differences between actual and estimated costs against year of decision to build

TABLE 2. Inaccuracy of transportation project cost estimates by geographical location (fixed prices).
Europe

Project
type
Rail
Fixed-link
Road
All projects

Average
Number
cost
of projects escalation
(N)
(%)
23
15
143
181

34.2
43.4
22.4
25.7

North America

Standard
deviation
25.1
52.0
24.9
28.7

Average
Number
cost
of projects escalation
(N)
(%)
19
18
24
61

40.8
25.7
8.4
23.6

Other geographical areas

Standard
deviation
36.8
70.5
49.4
54.2

Average
Number
cost
of projects escalation
(N)
(%)
16
0
0
16

64.6
—
—
64.6

Standard
deviation
49.5
—
—
49.5

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 285

 BENT FLYVBJERG, METTE SKAMRIS HOLM, AND SØREN BUHL

for the 111 projects in the sample for which these data
are available. The diagram does not seem to indicate an
effect from time on cost underestimation. Statistical
analyses corroborate this impression. The null hypothesis that year of decision has no effect on the difference
between actual and estimated costs cannot be rejected
(p=0.22, F-test). A test using year of completion instead
of year of decision (with data for 246 projects) gives a
similar result (p=0.28, F-test).
We therefore conclude that cost underestimation
has not decreased over time. Underestimation today is
in the same order of magnitude as it was 10, 30, and 70
years ago. If techniques and skills for estimating and
forecasting costs of transportation infrastructure projects have improved over time, this does not show in the
data. No learning seems to take place in this important
and highly costly sector of public and private decision
making. This seems strange and invites speculation that
the persistent existence over time, location, and project
type of significant and widespread cost underestimation
is a sign that an equilibrium has been reached: Strong
incentives and weak disincentives for underestimation
may have taught project promoters what there is to
learn, namely, that cost underestimation pays off. If this
is the case, underestimation must be expected and it
must be expected to be intentional. We examine such
speculation below. Before doing so, we compare cost underestimation in transportation projects with that in
other projects.

Cost Underestimation in Other
Infrastructure Projects
In addition to cost data for transportation infrastructure projects, we have reviewed cost data for several
hundred other projects including power plants, dams,
water distribution, oil and gas extraction, information
technology systems, aerospace systems, and weapons
systems (Arditi et al., 1985; Blake et al., 1976; Canaday,
1980; Department of Energy Study Group, 1975;
Dlakwa & Culpin, 1990; Fraser, 1990; Hall, 1980; Healey,
1964; Henderson, 1977; Hufschmidt & Gerin, 1970;
Merewitz, 1973b; Merrow, 1988; Morris & Hough, 1987;
World Bank, 1994, n.d.). The data indicate that other
types of projects are at least as, if not more, prone to cost
underestimation as are transportation infrastructure
projects.
Among the more spectacular examples of cost underestimation are the Sydney Opera House, with actual
costs approximately 15 times higher than those projected, and the Concorde supersonic airplane, with a cost
12 times higher than predicted (Hall, n.d., p. 3). The data
also indicate that cost underestimations for other proj-

286 APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3

ects have neither increased nor decreased historically,
and that underestimation is common in both First- and
Third-World countries. When the Suez canal was completed in 1869, actual construction costs were 20 times
higher than the earliest estimated costs and 3 times
higher than the cost estimate for the year before construction began. The Panama Canal, which was completed in 1914, had cost escalations in the range of 70 to
200% (Summers, 1967, p. 148).
In sum, the phenomena of cost underestimation
and escalation appear to be characteristic not only of
transportation projects but of other types of infrastructure projects as well.

Explanations of Underestimation:
Error or Lie?
Explanations of cost underestimation come in four
types: technical, economic, psychological, and political.
In this section, we examine which explanations best fit
our data.

Technical Explanations
Most studies that compare actual and estimated
costs of infrastructure projects explain what they call
“forecasting errors” in technical terms, such as imperfect
techniques, inadequate data, honest mistakes, inherent
problems in predicting the future, lack of experience on
the part of forecasters, etc. (Ascher, 1978; Flyvbjerg et al.,
in press; Morris & Hough, 1987; Wachs, 1990). Few
would dispute that such factors may be important
sources of uncertainty and may result in misleading forecasts. And for small-sample studies, which are typical of
this research field, technical explanations have gained
credence because samples have been too small to allow
tests by statistical methods. However, the data and tests
presented above, which come from the first large-sample study in the field, lead us to reject technical explanations of forecasting errors. Such explanations simply do
not fit the data.
First, if misleading forecasts were truly caused by
technical inadequacies, simple mistakes, and inherent
problems with predicting the future, we would expect a
less biased distribution of errors in cost estimates
around zero. In fact, we have found with overwhelming
statistical significance (p<0.001) that the distribution of
such errors has a nonzero mean. Second, if imperfect
techniques, inadequate data, and lack of experience were
main explanations of the underestimations, we would
expect an improvement in forecasting accuracy over
time, since errors and their sources would be recognized
and addressed through the refinement of data collection,
forecasting methods, etc. Substantial resources have

 UNDERESTIMATING COSTS IN PUBLIC WORKS PROJECTS

FIGURE 3. Inaccuracy of cost estimates in transportation projects over time, 1910–1998 (fixed prices, 111
projects).

been spent over several decades on improving data and
methods. Still our data show that this has had no effect
on the accuracy of forecasts. Technical factors, therefore,
do not appear to explain the data. It is not so-called forecasting “errors” or cost “escalation” or their causes that
need explaining. It is the fact that in 9 out of 10 cases,
costs are underestimated.
We may agree with proponents of technical explanations that it is, for example, impossible to predict for
the individual project exactly which geological, environmental, or safety problems will appear and make costs
soar. But we maintain that it is possible to predict the

risk, based on experience from other projects, that some
such problems will haunt a project and how this will affect costs. We also maintain that such risk can and
should be accounted for in forecasts of costs, but typically is not. For technical explanations to be valid, they
would have to explain why forecasts are so consistent in
ignoring cost risks over time, location, and project type.

Economic Explanations
Economic explanations conceive of cost underestimation in terms of economic rationality. Two types of
economic explanation exist; one explains in terms of eco-

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 287

 BENT FLYVBJERG, METTE SKAMRIS HOLM, AND SØREN BUHL

nomic self-interest, the other in terms of the public interest. As regards self-interest, when a project goes forward, it creates work for engineers and construction
firms, and many stakeholders make money. If these
stakeholders are involved in or indirectly influence the
forecasting process, then this may influence outcomes
in ways that make it more likely that the project will be
built. Having costs underestimated and benefits overestimated would be economically rational for such stakeholders because it would increase the likelihood of revenues and profits. Economic self-interest also exists at the
level of cities and states. Here, too, it may explain cost
underestimation. Pickrell (1990, 1992) pointed out that
transit capital investment projects in the U.S. compete
for discretionary grants from a limited federal budget
each year. This creates an incentive for cities to make
their projects look better, or else some other city may get
the money.
As regards the public interest, project promoters and
forecasters may deliberately underestimate costs in order
to provide public officials with an incentive to cut costs
and thereby to save the public’s money. According to this
type of explanation, higher cost estimates would be an
incentive for wasteful contractors to spend more of the
taxpayer’s money. Empirical studies have identified promoters and forecasters who say they underestimate costs
in this manner and with this purpose (i.e., to save public
money; Wachs, 1990). The argument has also been
adopted by scholars, for instance Merewitz (1973b), who
explicitly concludes that “keeping costs low is more important than estimating costs correctly” (p. 280).
Both types of economic explanation account well for
the systematic underestimation of costs found in our
data. Both depict such underestimation as deliberate,
and as economically rational. If we now define a lie in the
conventional fashion as making a statement intended to
deceive others (Bok, 1979, p. 14; Cliffe et al., 2000, p. 3),
we see that deliberate cost underestimation is lying, and
we arrive at one of the most basic explanations of lying,
and of cost underestimation, that exists: Lying pays off,
or at least economic agents believe it does. Moreover, if
such lying is done for the public good (e.g., to save taxpayers’ money), political theory would classify it in that
special category of lying called the “noble lie,” the lie motivated by altruism. According to Bok (1979), this is the
“most dangerous body of deceit of all” (p. 175).
In the case of cost underestimation in public works
projects, proponents of the noble lie overlook an important fact: Their core argument—that taxpayers’ money is
saved by cost underestimation—is seriously flawed. Anyone with even the slightest trust in cost-benefit analysis
and welfare economics must reject this argument. Underestimating the costs of a given project leads to a

288 APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3

falsely high benefit-cost ratio for that project, which in
turn leads to two problems. First, the project may be
started despite the fact that it is not economically viable.
Or, second, it may be started instead of another project
that would have yielded higher returns had the actual
costs of both projects been known. Both cases result in
the inefficient use of resources and therefore in waste of
taxpayers’ money. Thus, for reasons of economic efficiency alone, the argument that cost underestimation
saves money must be rejected; underestimation is more
likely to result in waste of taxpayers’ money. But the argument must also be rejected for ethical and legal reasons. In most democracies, for project promoters and
forecasters to deliberately misinform legislators, administrators, bankers, the public, and the media would not
only be considered unethical but in some instances also
illegal, for instance where civil servants would misinform
cabinet members or cabinet members would misinform
the parliament. There is a formal “obligation to truth”
built into most democratic constitutions on this point.
This obligation would be violated by deliberate underestimation of costs, whatever the reasons may be. Hence,
even though economic explanations fit the data and help
us understand important aspects of cost underestimation, such explanations cannot be used to justify it.

Psychological Explanations
Psychological explanations attempt to explain biases in forecasts by a bias in the mental makeup of project promoters and forecasters. Politicians may have a
“monument complex,” engineers like to build things,
and local transportation officials sometimes have the
mentality of empire builders. The most common psychological explanation is probably “appraisal optimism.” According to this explanation, promoters and
forecasters are held to be overly optimistic about project
outcomes in the appraisal phase, when projects are
planned and decided (Fouracre et al., 1990, p. 10; Mackie
& Preston, 1998; Walmsley & Pickett, 1992, p. 11; World
Bank, 1994, p. 86). An optimistic cost estimate is clearly
a low one. The existence of appraisal optimism in promoters and forecasters would result in actual costs being
higher than estimated costs. Consequently, the existence
of appraisal optimism would be able to account, in
whole or in part, for the peculiar bias of cost estimates
found in our data, where costs are systematically underestimated. Such optimism, and associated cost underestimation, would not be lying, needless to say, because
the deception involved is self-deception and therefore
not deliberate. Cost underestimation would be error according to this explanation.
There is a problem with psychological explanations,
however. Appraisal optimism would be an important

 UNDERESTIMATING COSTS IN PUBLIC WORKS PROJECTS

and credible explanation of underestimated costs if estimates were produced by inexperienced promoters and
forecasters, i.e., persons who were estimating costs for
the first or second time and who were thus unknowing
about the realities of infrastructure building and were
not drawing on the knowledge and skills of more experienced colleagues. Such situations may exist and may
explain individual cases of cost underestimation. But
given the fact that the human psyche is distinguished by
a significant ability to learn from experience, it seems unlikely that promoters and forecasters would continue to
make the same mistakes decade after decade instead of
learning from their actions. It seems even more unlikely
that a whole profession of forecasters and promoters
would collectively be subject to such a bias and would
not learn over time. Learning would result in the reduction, if not elimination, of appraisal optimism, which
would then result in cost estimates becoming more accurate over time. But our data clearly shows that this has
not happened.
The profession of forecasters would indeed have to
be an optimistic group to keep their appraisal optimism
throughout the 70-year period our study covers and not
learn that they were deceiving themselves and others by
underestimating costs. This would account for the data,
but is not a credible explanation. As observed elsewhere,
the incentive to publish and justify optimistic estimates
is very strong, and the penalties for having been overoptimistic are generally insignificant (Davidson & Huot,
1989, p. 137; Flyvbjerg et al., in press). This is a better explanation of the pervasive existence of optimistic estimates than an inherent bias for optimism in the psyche
of promoters and forecasters. And “optimism” calculated on the basis of incentives is not optimism, of
course; it is deliberate deception. Therefore, on the basis
of our data, we reject appraisal optimism as a primary
cause of cost underestimation.

Political Explanations
Political explanations construe cost underestimation in terms of interests and power (Flyvbjerg, 1998).
Surprisingly little work has been done that explains the
pattern of misleading forecasts in such terms (Wachs,
1990, p. 145). A key question for political explanations is
whether forecasts are intentionally biased to serve the interests of project promoters in getting projects started.
This question again raises the difficult issue of lying.
Questions of lying are notoriously hard to answer, because in order to establish whether lying has taken place,
one must know the intentions of actors. For legal, economic, moral, and other reasons, if promoters and forecasters have intentionally fabricated a deceptive cost
estimate for a project to get it started, they are unlikely to

tell researchers or others that this is the case (Flyvbjerg,
1996; Wachs, 1989).
When Eurotunnel, the private company that owns
the tunnel under the English Channel, went public
in 1987 to raise funds for the project, investors were
told that building the tunnel would be relatively straightforward. Regarding risks of cost escalation, the prospectus read:
Whilst the undertaking of a tunneling project of
this nature necessarily involves certain construction risks, the techniques to be used are well
proven. . . . The Directors, having consulted the
Mâitre d’Oeuvre, believe that 10% . . . would be a
reasonable allowance for the possible impact of unforeseen circumstances on construction costs.2
(“Under Water,” 1989, p. 37)
Two hundred banks communicated these figures for
cost and risk to investors, including a large number of
small investors. As observed by The Economist (“Under
Water,” 1989), anyone persuaded in this way to buy
shares in Eurotunnel in the belief that the cost estimate
was the mean of possible outcomes was, in effect, deceived. The cost estimate of the prospectus was a best
possible outcome, and the deception consisted in making investors believe in the highly unlikely assumption—
disproved in one major construction project after another—that everything would go according to plan, with
no delays; no changes in safety and environmental performance specifications; no management problems; no
problems with contractual arrangements, new technologies, or geology; no major conflicts; no political
promises not kept; etc. The assumptions were, in other
words, those of an ideal world. The real risks of cost escalation for the Channel tunnel were many times higher
than those communicated to potential investors, as evidenced by the fact that once built, the real costs of the
project were higher by a factor of two compared with
forecasts.
Flyvbjerg, Bruzelius, and Rothengatter (in press)
document for a large number of projects that the Everything-Goes-According-to-Plan type of deception used
for the Channel tunnel is common. Such deception is,
in fact, so widespread that in a report on infrastructure
and development, the World Bank (1994, pp. ii, 22)
found reason to coin a special term for it: the “EGAPprinciple.” Cost estimation following the EGAP-principle simply disregards the risk of cost escalation resulting from delays, accidents, project changes, etc. This is a
major problem in project development and appraisal, according to the World Bank.
It is one thing, however, to point out that investors,
public or private, were deceived in particular cases. It is

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 289

 BENT FLYVBJERG, METTE SKAMRIS HOLM, AND SØREN BUHL

quite another to get those involved in the deceptions to
talk about this and to possibly admit that deception was
intentional, i.e., that it was lying. We are aware of only
one study that actually succeeded in getting those involved in underestimating costs to talk about such issues (Wachs, 1986, 1989, 1990). Wachs interviewed public officials, consultants, and planners who had been
involved in transit planning cases in the U.S. He found
that a pattern of highly misleading forecasts of costs and
patronage could not be explained by technical issues and
were best explained by lying. In case after case, planners,
engineers, and economists told Wachs that they had had
to “cook” forecasts in order to produce numbers that
would satisfy their superiors and get projects started,
whether or not the numbers could be justified on technical grounds (Wachs, 1990, p. 144). One typical planner admitted that he had repeatedly adjusted the cost
figures for a certain project downward and the patronage
figures upward to satisfy a local elected official who
wanted to maximize the chances of getting the project
in question started. Wachs’ work is unusually penetrating for a work on forecasting. But again, it is small-sample research, and Wachs acknowledges that most of his
evidence is circumstantial (Wachs, 1986, p. 28). The evidence does not allow conclusions regarding the project
population. Nevertheless, based on the strong pattern
of misrepresentation and lying found in his case studies, Wachs goes on to hypothesize that the type of abuse
he has uncovered is “nearly universal” (1990, p. 146;
1986, p. 28) and that it takes place not only in transit
planning but also in other sectors of the economy where
forecasting routinely plays an important role in policy
debates.
Our data give support to Wachs’ claim. The pattern
of highly underestimated costs is found not only in the
small sample of projects Wachs studied; the pattern is
statistically significant and holds for the project population mean (i.e., for the majority of transportation infrastructure projects). However, on one point, Wachs
(1986) seems to draw a conclusion somewhat stronger
than is warranted. “[F]orecasted costs always seem to be
lower than actual costs” (p. 24) he says (emphasis in original). Our data show that although “always” (100%) may
cover the small sample of projects Wachs chose to study,
when the sample is enlarged by a factor of 20–30 to a
more representative one, “only” in 86% of all cases are
forecasted costs lower than actual costs. Such trifles—14
percentage points—apart, the pattern identified by Wachs
is a general one, and his explanation of cost underestimation in terms of lying to get projects started fit our
data particularly well. Of the existing explanations of
cost development in transportation infrastructure projects, we therefore opt for political and economic expla-

290 APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3

nations. The use of deception and lying as tactics in
power struggles aimed at getting projects started and at
making a profit appear to best explain why costs are
highly and systematically underestimated in transportation infrastructure projects.

Summary and Conclusions
The main findings from the study reported in this
article—all highly significant and most likely conservative—are as follows:
• In 9 out of 10 transportation infrastructure
projects, costs are underestimated.
• For rail projects, actual costs are on average 45%
higher than estimated costs (sd=38).
• For fixed-link projects (tunnels and bridges),
actual costs are on average 34% higher than
estimated costs (sd=62).
• For road projects, actual costs are on average 20%
higher than estimated costs (sd=30).
• For all project types, actual costs are on average
28% higher than estimated costs (sd=39).
• Cost underestimation exists across 20 nations
and 5 continents; it appears to be a global
phenomenon.
• Cost underestimation appears to be more
pronounced in developing nations than in
North America and Europe (data for rail projects
only).
• Cost underestimation has not decreased over the
past 70 years. No learning that would improve
cost estimate accuracy seems to take place.
• Cost underestimation cannot be explained by
error and seems to be best explained by strategic
misrepresentation, i.e., lying.
• Transportation infrastructure projects do not
appear to be more prone to cost underestimation
than are other types of large projects.
We conclude that the cost estimates used in public
debates, media coverage, and decision making for transportation infrastructure development are highly, systematically, and significantly deceptive. So are the costbenefit analyses into which cost estimates are routinely
fed to calculate the viability and ranking of projects. The
misrepresentation of costs is likely to lead to the misallocation of scarce resources, which, in turn, will produce
losers among those financing and using infrastructure,
be they taxpayers or private investors.
We emphasize that these conclusions should not be
interpreted as an attack on public (vs. private) spending
on infrastructure, since the data are insufficient to decide whether private projects perform better or worse

 UNDERESTIMATING COSTS IN PUBLIC WORKS PROJECTS

than public ones regarding cost underestimation. Nor
do the conclusions warrant an attack on spending on
transportation vs. spending on other projects, since
other projects appear to be as liable to cost underestimation and escalation as are transportation projects.
With transportation projects as an in-depth case study,
the conclusions simply establish that significant cost underestimation is a widespread practice in project development and implementation, and that this practice
forms a substantial barrier to the effective allocation of
scarce resources for building important infrastructure.
The key policy implication for this consequential
and highly expensive field of public policy is that those
legislators, administrators, bankers, media representatives, and members of the public who value honest numbers should not trust the cost estimates presented by
infrastructure promoters and forecasters. Another important implication is that institutional checks and balances—including financial, professional, or even criminal penalties for consistent or foreseeable estimation
errors—should be developed to ensure the production
of less deceptive cost estimates. The work of designing
such checks and balances has been begun elsewhere,
with a focus on four basic instruments of accountability:
(1) increased transparency, (2) the use of performance
specifications, (3) explicit formulation of the regulatory
regimes that apply to project development and implementation, and (4) the involvement of private risk capital, even in public projects (Bruzelius et al., 1998; Flyvbjerg et al., in press).
ACKNOWLEDGMENTS
The authors wish to thank Martin Wachs, Don Pickrell, and
three anonymous JAPA referees for valuable comments on an
earlier draft of the article. Research for the article was supported by the Danish Transport Council and Aalborg University, Denmark.

NOTES
1. Merewitz’s (1973a, 1973b) study compared cost overrun
in urban rapid transit projects, especially the San Francisco Bay Area Rapid Transit (BART) system, with overrun in other types of public works projects. Merewitz’s
aims were thus different from ours, and his sample of
transportation projects was substantially smaller: 17
rapid transit projects and 49 highway projects, compared
with our 58 rail projects, 167 highway projects, and 33
bridge or tunnel projects. In addition to issues of a small
sample, in our attempt to replicate Merewitz’s analysis we
found that his handling of data raises a number of other
issues. First, Merewitz did not correct his cost data for inflation, i.e., current prices were used instead of fixed ones.

This is known to be a major source of error due to varying
inflation rates between projects and varying duration of
construction periods. Second, in statistical tests, Merewitz compared the mean cost overrun of subgroups of
projects (e.g., rapid transit) with the grand mean of overrun for all projects, thus making the error of comparing
projects with themselves. Subgroups should be tested directly against other subgroups in deciding whether they
differ at all and, if so, which ones differ. Third, Merewitz’s
two reports (1973a, 1973b) are inconsistent. One (1973a)
calculates the grand mean of cost overrun as the average
of means for subgroups; that is, the grand mean is unweighted, where common practice is to use the weighted
mean, as appears to be the approach taken in the other
(1973b). Fourth, due to insufficient information, the pvalues calculated by Merewitz are difficult to verify; most
likely they are flawed, however, and Merewitz’s one-sided
p-values are misleading. Finally, Merewitz used a debatable assumption about symmetry, which has more impact
for the nonparametric test used than nonnormality has
for parametric methods. Despite these shortcomings, the
approach taken in Merewitz’s study was innovative for its
time and in principle pointed in the right direction regarding how to analyze cost escalation in public works
projects. The study cannot be said to be a true large-sample study for transportation infrastructure, however, and
its statistical significance is unclear.
2. The Mâitre d’Oeuvre was an organization established to
monitor project planning and implementation for the
Channel tunnel. It was established in 1985, and until 1988
it represented the owners. In 1988 it was reverted to an
impartial position (Major Projects Association, 1994, pp.
151–153).

REFERENCES
Arditi, D., Akan, G. T., & Gurdamar, S. (1985). Cost overruns in
public projects. International Journal of Project Management,
3(4), 218–225.
Ascher, W. (1978). Forecasting: An appraisal for policy-makers and
planners. Baltimore: Johns Hopkins University Press.
Blake, C., Cox, D., & Fraize, W. (1976). Analysis of projected vs. actual costs for nuclear and coal-fired power plants (Report repared for the United States Energy Research and Development Administration). McLean, VA: Mitre Corporation.
Bok, S. (1979). Lying: Moral choice in public and private life. New
York: Vintage.
Bruzelius, N., Flyvbjerg, B., & Rothengatter, W. (1998). Big decisions, big risks: Improving accountability in mega projects. International Review of Administrative Sciences, 64(3),
423–440.
Canaday, H. T. (1980). Construction cost overruns in electric utilities:
Some trends and implications (Occasional Paper No. 3).
Columbus: National Regulatory Research Institute, Ohio
State University.
Cliffe, L., Ramsey, M., & Bartlett, D. (2000). The politics of lying:
Implications for democracy. London: Macmillan.

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 291

 BENT FLYVBJERG, METTE SKAMRIS HOLM, AND SØREN BUHL

Commission of the European Union. (1993). Growth, competitiveness, employment: The challenges and ways forward into the
21st century (White Paper). Brussels: Author.
Davidson, F. P., & Huot, J.-C. (1989). Management trends for
major projects. Project Appraisal, 4(3), 133–142.
Department of Energy Study Group. (1975). North Sea costs escalation study (Energy Paper No. 8). London: Department
of Energy.
Dlakwa, M. M., & Culpin, M. F. (1990). Reasons for overrun in
public sector construction projects in Nigeria. International Journal of Project Management, 8(4), 237–240.
Flyvbjerg, B. (1996). The dark side of planning: Rationality and
Realrationalität. In S. Mandelbaum, L. Mazza, & R. Burchell (Eds.), Explorations in planning theory (pp. 383–394).
New Brunswick, NJ: Center for Urban Policy Research
Press.
Flyvbjerg, B. (1998). Rationality and power: Democracy in practice.
Chicago: University of Chicago Press.
Flyvbjerg, B., Bruzelius, N., & Rothengatter, W. (in press).
Megaprojects and risk: An anatomy of ambition. Cambridge,
UK: Cambridge University Press.
Fouracre, P. R., Allport, R. J., & Thomson, J. M. (1990). The performance and impact of rail mass transit in developing countries
(TRRL Research Report 278). Crowthorne, UK: Transport
and Road Research Laboratory.
Fraser, R. M. (1990). Compensation for extra preliminary
and general (P & G) costs arising from delays, variations
and disruptions: The palmiet pumped storage scheme.
Tunneling and Underground Space Technology, 5(3), 205–
216.
Hall, P. (1980). Great planning disasters. Harmondsworth, UK:
Penguin Books.
Hall, P. (n.d). Great planning disasters revisited. Unpublished
manuscript, Bartlett School, University College, London.
Healey, J. M. (1964). Errors in project cost estimates. Indian Economic Journal, 12(1), 44–52.
Henderson, P. D. (1977). Two British errors: Their probable
size and some possible lessons. Oxford Economic Papers,
29(2), 159–205.
Holm, M. K. S. (1999). Inaccuracy of traffic forecasts and cost estimates in Swedish road and rail projects. Unpublished manuscript, Aalborg University, Department of Development
and Planning.
Hufschmidt, M. M., & Gerin, J. (1970). Systematic errors in
cost estimates for public investment projects. In J. Margolis (Ed.), The analysis of public output (pp. 267–315). New
York: Columbia University Press.
Kain, J. F. (1990). Deception in Dallas: Strategic misrepresentation in rail transit promotion and evaluation. Journal of
the American Planning Association, 56(2), 184–196.
Leavitt, D., Ennis, S., & McGovern, P. (1993). The cost escalation
of rail projects: Using previous experience to re-evaluate the calspeed estimates (Working Paper No. 567). Berkeley: Institute of Urban and Regional Development, University of
California.
Lewis, H. (1986). The metro report: The impact of metro and public
transport integration in Tyne and Wear. Newcastle, UK: Tyne
and Wear Passenger Transport Executive.

292 APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3

Mackie, P., & Preston, J. (1998). Twenty-one sources of error
and bias in transport project appraisal. Transport Policy,
5(1), 1–7.
Major Projects Association. (1994). Beyond 2000: A source book
for major projects. Oxford, UK: Author.
Merewitz, L. (1973a). How do urban rapid transit projects compare
in cost estimate experience? (Reprint No. 104). Berkeley: Institute of Urban and Regional Development, University
of California.
Merewitz, L. (1973b). Cost overruns in public works. In W.
Niskanen, A. C. Hansen, R. H. Havemann, R. Turvey, & R.
Zeckhauser (Eds.), Benefit cost and policy analysis (pp. 277–
295). Chicago: Aldine.
Merrow, E. W. (1988). Understanding the outcomes of megaprojects:
A quantitative analysis of very large civilian projects. Santa
Monica, CA: RAND Corporation.
Morris, P. W. G., & Hough, G. H. (1987). The anatomy of major
projects: A study of the reality of project management. New York:
John Wiley and Sons.
National Audit Office, Department of Transport. (1985). Expenditure on motorways and trunk roads. London: National
Audit Office.
National Audit Office, Department of Transport. (1992). Contracting for roads. London: National Audit Office.
National Audit Office, Department of Transport; Scottish Development Department; & Welsh Office. (1988). Road
planning. London: Her Majesty’s Stationary Office.
Nijkamp, P., & Ubbels, B. (1999). How reliable are estimates of
infrastructure costs? A comparative analysis. International
Journal of Transport Economics, 26(1), 23–53.
Pickrell, D. H. (1990). Urban rail transit projects: Forecast versus
actual ridership and cost. Washington, DC: U.S. Department
of Transportation.
Pickrell, D. H. (1992). A desire named streetcar: Fantasy and
fact in rail transit planning. Journal of the American Planning Association, 58(2), 158–176.
Riksrevisionsverket. (1994). Infrastrukturinvesteringar: En kostnadsjämförelse mellan plan och utfall i 15 större projekt inom
Vägverket och Banverket. Stockholm: Author.
Simon, J. (1991). Let’s make forecast and actual comparisons
fair. TR News, 156, 6–9.
Skamris, M. K., & Flyvbjerg, B. (1997). Inaccuracy of traffic
forecasts and cost estimates on large transport projects.
Transport Policy, 4(3), 141–146.
Summers, R. (1967). Cost estimates as predictors of actual
costs: A statistical study of military developments. In T.
Marschak, T. K. Glennan, & R. Summers (Eds.), Strategy
for R&D: Studies in the microeconomics of development (pp.
140–189). Berlin: Springer-Verlag.
Szyliowicz, J. S., & Goetz, A. R. (1995). Getting realistic about
megaproject planning: The case of the new Denver International Airport. Policy Sciences, 28(4), 347–367.
Under water, over budget. (1989, October 7). The Economist, pp.
37–38.
Vejdirektoratet. (1995). Notat om anlægsregnskaber. Copenhagen: Danish Road Directorate.
Wachs, M. (1986). Technique vs. advocacy in forecasting: A
study of rail rapid transit. Urban Resources, 4(1), 23–30.

 UNDERESTIMATING COSTS IN PUBLIC WORKS PROJECTS

Wachs, M. (1989). When planners lie with numbers. Journal of
the American Planning Association, 55(4), 476–479.
Wachs, M. (1990). Ethics and advocacy in forecasting for public policy. Business and Professional Ethics Journal, 9(1–2),
141–157.
Walmsley, D. A., & Pickett, M. W. (1992). The cost and patronage of rapid transit systems compared with forecasts (Research

Report 352). Crowthorne, UK: Transport Research
Laboratory.
World Bank. (1994). World development report 1994: Infrastructure for development. Oxford, UK: Oxford University Press.
World Bank. (n. d.). Economic analysis of projects: Towards a results-oriented approach to evaluation (ECON Report). Washington, DC: Author.

APPENDIX
The first task of the research reported in this paper
was to establish a sample of infrastructure projects substantially larger than what is common in this area of research, a sample large enough to allow statistical analyses of costs. Here a first problem was that data on actual
costs in transportation infrastructure projects are relatively difficult to come by. One reason is that it is quite
time consuming to produce such data. For public sector
projects, funding and accounting procedures are typically unfit for keeping track of the multiple and complex
changes that occur in total project costs over time. For
large projects, the relevant time frame may cover 5, 10, or
more fiscal years from decision to build, until construction starts, until the project is completed and operations
begin. Reconstructing the actual total costs of a public
project, therefore, typically entails long and difficult archival work and complex accounting. For private projects, even if funding and accounting practices may be
more conducive to producing data on actual total costs,
such data are often classified to keep them from the
hands of competitors. Unfortunately, this also tends to
keep data from the hands of scholars. And for both public and private projects, data on actual costs may be held
back by project owners because more often than not, actual costs reveal substantial cost escalation, and cost escalation is normally considered somewhat of an embarrassment to promoters and owners. In sum, establishing
reliable data on actual costs for even a single transportation infrastructure project is often highly time consuming or simply impossible.
This state of affairs explains why small-sample studies dominate scholarship in this field of research. But despite the problems mentioned, after 4 years of data collection and refinement, we were able to establish a sample
of 258 transportation infrastructure projects with data
on both actual construction costs and estimated costs at
the time of decision to build. The project portfolio is
worth approximately US$90 billion (1995 prices). The
project types are bridges, tunnels, highways, freeways,
high-speed rail, urban rail, and conventional (interurban)
rail. The projects are located in 20 countries on 5 continents, including both developed and developing nations.
The projects were completed between 1927 and 1998.

Older projects were included in the sample in order to
test whether the accuracy of estimated costs improved
over time. The construction costs of projects range from
US$1.5 million to US$8.5 billion (1995 prices), with the
smallest projects typically being stretches of roads in
larger road schemes, and the largest projects being rail
links, tunnels, and bridges. As far as we know, this is the
largest sample of projects with data on cost development
that has been established in this field of research.
In statistical analysis, data should be a sample from
a larger population, and the sample should represent the
population properly. These requirements are ideally satisfied by drawing the sample by randomized lot. Randomization ensures with high probability that factors
that cannot be controlled are equalized. A sample should
also be designed such that the representation of subgroups corresponds to their occurrence and importance
in the population. In studies of human affairs, however,
where controlled laboratory experiments often cannot
be conducted, it is frequently impossible to meet these
ideal conditions. This is also the case for the current
study, and we therefore had to take a different approach
to sampling and statistical analysis.
We selected the projects for the sample on the basis
of data availability. All projects that we knew of for which
data on construction cost development were obtainable
were considered for inclusion in the sample. Cost development is defined as the difference between actual and
estimated costs in percentage of estimated costs, with all
costs measured in fixed prices. Actual costs are defined
as real, accounted costs determined at the time of completing a project. Estimated costs are defined as budgeted, or forecasted, costs at the time of decision to build.
Even if the project planning process varies with project
type, country, and time, it is typically possible to locate
for any given project a specific point in the process that
can be identified as the time of decision to build. Usually
a cost estimate was available for this point in time. If not,
the closest available estimate was used, typically a later
estimate resulting in a conservative bias in our measurement of cost development. Cost data were collected from
a variety of sources, including annual project accounts,
questionnaires, interviews, and other studies.

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 293

 BENT FLYVBJERG, METTE SKAMRIS HOLM, AND SØREN BUHL

Data on cost development were available for 343
projects. We then rejected 85 projects because of insufficient data quality. For instance, for some projects we
could not obtain a clear answer regarding what was included in costs, or whether cost data were given in current or fixed prices, or which price level (year) had been
used in estimating and discounting costs. More specifically, of those 85 projects, we rejected 27 because we
could not establish whether or not cost data were valid
and reliable. We rejected 12 projects because they had
been completed before 1915 and no reliable indices were
available for discounting costs to the present. Finally, we
excluded 46 projects because cost development for them
turned out to have been calculated before construction
was completed and operations begun; therefore, the actual final costs for these projects may be different from
the cost estimates used to calculate cost development,
and no information was available on actual final costs. In
addition to the 85 rejected projects mentioned here, we
also rejected a number of projects to avoid double counting of projects. This typically involved projects from
other studies that appeared in more than one study or
where we had a strong suspicion that this might be the
case. In sum, all projects for which data were considered
valid and reliable were included in the sample. This covers both projects for which we ourselves collected the
data and projects for which other researchers in other
studies did the data collection (Fouracre et al., 1990;
Hall, 1980; Leavitt et al., 1993; Lewis, 1986; Merewitz,
1973a; National Audit Office, Department of Transport,
1985, 1992; National Audit Office, Department of
Transport, Scottish Development Department, & Welsh
Office, 1988; Pickrell, 1990; Riksrevisionsverket, 1994;
Vejdirektoratet, 1995; Walmsley & Pickett, 1992). Cost
data were made comparable across projects by discounting prices to the 1995 level and calculating them in
Euros, using the appropriate geographical, sectoral, and
historical indices for discounting and the appropriate
exchange rates for conversion between currencies.
Our own data collection concentrated on large European projects because too few data existed for this type
of project to allow comparative studies. For instance, for
projects with actual construction costs larger than 500
million Euros (1995 prices; EUR1=US$1.29 in 1995), we
were initially able to identify from other studies only two
European projects for which data were available on both
actual and estimated costs. If we lowered the project size
and looked at projects larger than 100 million Euros, we
were able to identify such data for eight European projects. We saw the lack of reliable cost data for European
projects as particularly problematic since the Commission of the European Union had just launched its policy
for establishing the so-called trans-European transport

294 APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3

networks, which would involve the construction of a
large number of major transportation infrastructure
projects across Europe at an initial cost of 220 billion
Euros (Commission of the European Union, 1993, p.
75). As regards costs, we concluded that the knowledge
base for the Commission’s policy was less than well developed, and we hoped to help remedy this situation
through our data collection. Our efforts on this point
proved successful. We collected primary data on cost for
37 projects in Denmark, France, Germany, Sweden, and
the U.K. and were thus able to greatly increase the number of large European projects with reliable data for both
actual and estimated costs, allowing for the first time a
comparative study for this type of project in which statistical methods could be applied.
As for any sample, a key question is whether the sample is representative of the population. Here the question is whether the projects included in the sample are
representative of the population of transportation infrastructure projects. Since the criterion for sampling
was data availability, this question translates into one of
whether projects with available data are representative.
There are four reasons why this is probably not the case.
First, it may be speculated that projects that are managed well with respect to data availability may also be
managed well in other respects, resulting in better than
average (i.e., nonrepresentative) performance for such
projects. Second, it has been argued that the very existence of data that make the evaluation of performance
possible may contribute to improved performance when
such data are used by project management to monitor
projects (World Bank, 1994, p. 17). Again, such projects
would not be representative of the project population.
Third, we might speculate that managers of projects
with a particularly bad track record regarding cost escalation have an interest in not making cost data available,
which would then result in underrepresentation of such
projects in the sample. Conversely, managers of projects
with a good track record for costs might be interested in
making this public, resulting in overrepresentation of
these projects. Fourth, and finally, even where managers
have made cost data available, they may have chosen to
give out data that present their projects in as favorable a
light as possible. Often there are several estimates of
costs to choose from and several calculations of actual
costs for a given project at a given time. If researchers collect data by means of survey questionnaires, as is often
the case, there might be a temptation for managers to
choose the combination of actual and estimated costs
that suits them best, possibly a combination that makes
their projects look good.
The available data do not allow an exact, empirical
assessment of the magnitude of the problem of misrep-

 UNDERESTIMATING COSTS IN PUBLIC WORKS PROJECTS

resentation. But the few data that exist that shed light on
this problem support the thesis that data are biased.
When we compared data from the Swedish Auditor General for a subsample of road projects, for which the problems of misrepresentation did not seem to be an issue,
with data for all road projects in our sample, we found
that cost escalation in the Swedish subsample is significantly higher than for all projects (Holm, 1999, pp. 11–
15). We conclude, for the reasons given above, that most
likely the sample is biased and the bias is conservative. In
other words, the difference between actual and estimated
costs derived from the sample is likely to be lower than
the difference in the project population. This should be
kept in mind when interpreting the results from statistical analyses of the sample. The sample is not perfect by
any means. Still it is the best obtainable sample given the
current state of the art in this field of research.
In the statistical analyses, percentage cost development in the sample is considered normally distributed
unless otherwise stated. Residual plots, not shown here,
indicate that normal distribution might not be completely satisfied, the distributions being somewhat
skewed with larger upper tails. However, transformations (e.g., the logarithmic one) do not improve this significantly. For simplicity, therefore, no transformation
has been made, unless otherwise stated.

The subdivisions of the sample implemented as part
of analyses entail methodological problems of their own.
Thus the representation of observations in different
combinations of subgroups is quite skewed for the data
considered. The analysis would be improved considerably if the representation were more even. Partial and
complete confounding occur; that is, if a combination
of two or more effects is significant, it is sometimes difficult to decide whether one, the other, or both cause the
difference. For interactions, often not all the combinations are represented, or the representations can be quite
scarce. We have adapted our interpretations of the data
to these limitations, needless to say. If better data could
be gathered, sharper conclusions could be made.
The statistical models used are linear normal models
(i.e., analysis of variance and regression analysis with the
appropriate F-tests and t-tests). The tests of hypotheses
concerning mean values are known to be robust to deviations from normality. Also, chi-square tests for independence have been used for count data. For each test,
the p-value has been reported. This value is a measure
for rareness if identity of groups is assumed. Traditionally, a p-value less than 0.01 is considered highly significant and less than 0.05 significant, whereas a larger pvalue means that the deviation could be due to chance.

APA Journal ◆ Summer 2002 ◆ Vol. 68, No. 3 295