Professional Documents
Culture Documents
Title: Should We Build More Large Dams? The Actual Costs of Hydropower Megaproject Development
Title: Should We Build More Large Dams? The Actual Costs of Hydropower Megaproject Development
Affiliations:
1Sad Business School, University of Oxford, OX1 1HP, UK.
2Blavatnik School of Government, University of Oxford, OX1 4JJ, UK.
3Department of Statistics, University of Oxford, OX1 3GT, UK.
Full reference: Ansar, Atif, Bent Flyvbjerg, Alexander Budzier, and Daniel
Lunn, 2014, "Should We Build More Large Dams? The Actual Costs of Hy-
dropower Megaproject Development," Energy Policy, March, pp. 1-14, DOI:
10.1016/j.enpol.2013.10.069, URL: http://bit.ly/1ekyL7Q.
from China to Brazil. Whether benefits of new dams will outweigh costs re-
ing evidence that budgets are systematically biased below actual costs of
vironmental, and social costs. Using the largest and most reliable reference
data of its kind and multilevel statistical techniques applied to large dams
for the first time, we were successful in fitting parsimonious models to pre-
dict cost and schedule overruns. The outside view suggests that in most
1
countries large hydropower dams will be too costly in absolute terms and
take too long to build to deliver a positive risk-adjusted return unless suita-
ble risk management measures outlined in this paper can be affordably pro-
prefer agile energy alternatives that can be built over shorter time horizons
to energy megaprojects.
2
Main Text:
The 21st Century faces significant energy challenges on a global scale. Popu-
lation and economic growth underpin increasing demand for energy from
tion and mitigation of climate change, and energy security present policy
makers and business leaders with difficult decisions and critical trade-offs in
implementing sound energy policies. Demand for electricity is, for example,
slated to almost double between 2010 and 2035 requiring global electricity
capacity to increase from 5.2 terawatt (TW) to 9.3 TW over the same period
(IEA, 2011). Currently, the de facto strategic response to these big energy
challenges is big solutions such as large hydropower dams. Are such big
solutions in general and large hydropower dams in particular the most effec-
Might more numerous small interventions be more prudent from the per-
spective of risk management and maximizing net present value even when
dropower capacity along with a long and varied list of corollary benefits: re-
ducing fossil fuel consumption, flood control, irrigation, urban water supply,
stan, Jinsha river dams in China, Myitsone dam in Myanmar, or the Gilgel
Gibe III dam in Ethiopia, all in various stages of development, are unprece-
dented in scale.
Large dams are, however, controversial because they exert substantial finan-
cial costs (World Bank, 1996; World Commission on Dams, 2000). Beyond the
Scudder, 2005; Stone, 2011), ecological (Nilsson et al., 2005; Ziv et al., 2012),
and social (Bakker, 1999; Duflo and Pande, 2007; Richter et al., 2010; Sovacool
and Bulan, 2011) impacts. Stone (2011, p. 817) reports in Science that the
Three Gorges dam in China is an environmental bane that will cost over
USD 26.45 billion over the next 10-years in environmental mitigation ef-
forts. Despite their outsized financial and environmental costs, the pur-
World Commission of Dams (2000, p. 30) reported that for large hydropower
eration is 80% of the targeted valuea trend of which the recently complet-
and Bulan, 2011). Similarly, Duflo and Pande (2007) find adverse distribu-
with losers upstream yielding a more modest, if any, net economic benefit.
The scale of contemporary dams is so vast that even for a large economy
such as Chinas the negative economic ramifications could likely hinder the
4
economic viability of the country as a whole if the risks inherent to these
projects are not well managed (Salazar, 2000). Similarly, Merrow et al. (1988,
pp. 2-3) warn that such enormous sums of money ride on the success of
megaprojects [such as large dams] that company balance sheets and even
outcomes. Such warnings are not idle alarmism. There is mounting evi-
social and environmental impact, and public support (McCully, 2001; Scud-
der, 2005; Singh, 2002; Sovacool and Bulan, 2011; WCD, 2000). There are ac-
society about whether large dams are a boon or a curse. Should we build
more large hydropower dams? How confident can planners be that a large
that won Princeton psychologist Daniel Kahneman the Nobel Prize in eco-
nomics in 2002 (Kahneman and Tversky, 1979a, 1979b; Kahneman, 1994) ex-
jects (Flyvbjerg et al., 2003; Flyvbjerg, 2009). We present statistical and com-
parative evidence from the largest reference class of actual costs of large hy-
find that even before accounting for negative impacts on human society and
environment, the actual construction costs of large dams are too high to yield
5
a positive return. Large dams also take inordinately long periods of time to
build, making them ineffective in resolving urgent energy crises. Our evi-
dence pertains primarily to large dams and the results cannot be applied ei-
ther to smaller dams or other large energy solutions such as nuclear power
without first conducting a separate reference class for other type of power
that require less upfront outlays and that can be built very quickly.
There is no doubt that harnessing and managing the power of water is criti-
cal for economies but large dams are not the way to do so unless suitable risk
psychology, and planning research, this paper further provides public agen-
cies (e.g. national planning and finance ministries, power and water authori-
reference class forecasting methods proposed here can improve the selection
Our approach to address the debates about whether or not to build dams is
among alternative options are actually made and on what basis. Theoretical
6
and empirical literature on decision-making under uncertainty proposes two
systematically and predictably too optimistic about the time, costs, and bene-
Buehler et al., 1994) stems from actors taking an inside view focusing on
the constituents of the specific planned action rather than on the outcomes of
similar actions already completed (Kahneman and Lovallo, 1993). Thus, for
example, the estimated costs put forward by cities competing to hold the
Olympic Games have consistently been underestimated yet every four years
finds strong evidence that misplaced political incentives and agency prob-
al. (2009, p. 180) further discuss that delusion and deception are complemen-
tary rather than alternative explanations for why megaprojects typically face
7
deception in practice. Using quasi-experimental evidence from China, Ansar
et al. (2013) suggest that while better incentive alignment can help lower the
frequency and to a lesser extent the magnitude of biases, it does not entirely
cure biases.
is the risk that costs might outweigh benefits for a proposed dam? While the
concluded, actions (or events). The outside view has three advantages: First,
to test and fit models to explain why the outcomes of a reference class of past
actions follow the observed distribution. Third, it allows to predict the uncer-
outside view were first described by Kahneman and Tversky (1979b) and
later by Kahneman and Lovallo (1993) and Lovallo and Kahneman (2003) as
a means to detect and cure biases in human judgment. The methodology and
data needed for employing the outside view, or reference class forecasting,
The outside view, applied to large dams for the first time here, involves three
the selected reference class of the parameter that is being forecasted; iii)
compare the specific case with the reference class distribution. We take a fur-
in the methodology of the outside view that can be generalized and applied
biased forecasts managers can make statistically grounded, rather than op-
timistic, judgments (Dawes et al., 1989; Buehler et al., 1994; Gilovich et al.,
2002).
its limitations (see Sovacool and Cooper, 2013 for a discussion specifically
about energy megaprojects). For example, RCF focuses on generic risk inher-
tional RCF methods in the result section below. Sovacool and Cooper (2013,
p. 63) further suggest that RCF may not provide sufficiently accurate indica-
tion of the risks of rare megaprojects the likes of which have never been built
before. Such out of the sample problems are well noted in probability the-
ory. They do not, however, deny the fundamental usefulness of RCF. If any-
9
thing our results err towards conservative estimates of actual cost overruns
Following literature on the planning fallacy (op. cit.), the parameters central
ratio of estimated costs1; cost overruns can also be thought as the underesti-
mation of actual costs (Bacon and Besant-Jones, 1998; Flyvbjerg et al., 2002).
Schedule slippage, called schedule overrun, is the ratio of the actual project
the main financiers and the key decision makers, and the end is the date of
proxies for the underlying risk factors that led to the inaccuracies. For exam-
ple, cost overruns reduce the attractiveness of an investment and if they be-
all require upward revision in the nominal electricity tariffs. Financial costs
ety. Projects with a poor cost and schedule performance are also likely to
In taking the outside view on the cost and schedule under/overruns, our
first step was to establish a valid and reliable reference class of previously
dams with a wall height > 15 m as large. The total global population of large
dams with a wall height > 15 m is 45,000. There are 300 dams in the world of
monumental scale; these major dams meet one of three criteria on height
(>150 m), dam volume (>15 million m3), or reservoir storage (>25 km3)
11
From this population of large dams, our reference class drew a representa-
tive sample of 245 dams (including 26 major dams) built between 1934 to
ble data of its kind. The portfolio is worth USD 353 billion. All large dams for
which valid and reliable cost and schedule data could be found were includ-
ed in the sample. Of the 245 large dams, 186 were hydropower projects (in-
cluding 25 major dams) and the remaining 59 were irrigation, flood control,
hydropower dam projects in our reference class to test whether project types
overview of the sample by regional location, wall height, project type, vin-
12
Location Size - Dam Wall Height (m)
Multipurpose 1970-1979
(with hydropower)
12.5%
10.0%
7.5%
5.0%
2.5%
0.0%
10 1,000 40,000
Millions of 2010 US dollars
timated versus actual costs of dams. Primary documents were collected from
13
2. World Bank, also see World Bank (1996) and Bacon and Besant-Jones
(1998);
6. U.S. Bureau of Reclamation, also see Hufschmidt and Gerin (1970)3 and
The procedures applied to the cost and schedule data here are consistent
ministration (2003); Pickrell (1989, 1992); World Bank (1996) and Bacon and
Besant-Jones (1998) with which our data are consistent. All costs are total
struction of all civil works and facilities; equipment purchases. Actual out-
turn costs are defined as real, accounted construction costs determined at the
casted, construction costs at the time of decision to build. The year of the
date of the decision to build a project is the base year of prices in which all
estimated and actual constant costs have been expressed in real (i.e. with the
effects of inflation removed) local currency terms of the country in which the
14
post environmental remedial works, and opportunity cost of submerging
land to form reservoirs. This makes comparison of estimated and actual costs
2.3. Analyses
Multilevel or hierarchically structured data are the norm in the social, medi-
ses are taught within schools; and schools may be administered within local
authorities or school boards. The units in such a system lie at four different
boards to level 4. Units at one level are recognized as being grouped, or nest-
ed, within units at the next higher level. Such a hierarchy is often described
in terms of clusters of level 1 units within each level 2 unit etc. and the term
that the dependent variable is at the lowest level of the nested structure.
Multilevel models are necessary for research designs where data for observa-
tions are organized at more than one level (i.e., nested data) (Gelman and
15
Hill, 2007). Failing to use multilevel models in such instances would result in
With respect to our data on dams, projects are nested in the countries of their
domicile. Like test scores of pupils from the same school tend exhibit within-
Our second step was to establish an empirical distribution for the cost fore-
ables, listed in Table 1, for the 245 large dams in our reference class.
16
Time
Year of final decision to build
Estimated implementation schedule (months)
Year of start of full commercial operation
Actual implementation schedule (months)
Procurement
Estimated project foreign exchange costs as a proportion of estimated total
project costs (percentage)
Competitiveness of procurement process, international competitive bidding
amount as a proportion of estimated total project costs (percentage)*
Main contractor is from the host country (dummy variable)
Country variables
Country (second level to control for within country correlation)
Political regime of host country is a democracy (dummy variable)
GDP of host country (current U.S. dollars)
Per capita income of host country in year of loan approval (constant USD)
Average actual cost growth rate in host country over the implementation
periodthe GDP deflator (percentage)
MUV Index of actual average cost growth rate for imported project compo-
nents between year of loan approval and year of project completion
Long-term inflation rate of the host country (percentage)
Actual average exchange rate depreciation or appreciation between year of
formal-decision-to-build and year of full commercial operation (percentage)
South Asian projects (dummy variable)
North American projects (dummy variable)
* Denotes variables with a large number of missing values not used for regression analysis
1. Three out of every four large dams suffered a cost overrun in constant
2. Actual costs were on average 96% higher than estimated costs; the
median was 27% (IQR 0.86). The evidence is overwhelming that costs
17
(i.e. cost overrun) is larger than the error of cost overestimation (p <
0.01). The skew is towards adverse outcomes (i.e. going over budget).
3. Graphing the dams cost overruns reveals a fat tail as shown in Figure
2; the actual costs more than double for 2 out of every 10 large dams
and more than triple for 1 out of every 10 dams. The fat tail suggests
Median
Mean
4. Large dams built in every region of the world suffer systematic cost
every region. Figure 3 shows the geographical spread and cost over-
runs of large dams in our reference class. Large dams built in North
18
America (n = 40) have considerably lower cost overrun (M = 11%)
ance on the North American experience with large dams may bias
choring their forecasts in familiar cases from North America and ap-
19
planners expected the net present benefits to exceed the net present
costs by about 40%. Nearly half the dams suffered a cost overrun ratio
of 1.4 or greater breaching this threshold after which the asset can be
covered. This is assuming, of course, that the benefits did not also fall
short of targets, even though there is strong evidence that actual bene-
fits of dams are also likely to fall short of targets (WCD, 2000; McCul-
Mean
large hydropower dam on average costs 1,800 million in 2010 USD with an
makes no attempts to take into account any covariates, shows that increase in
the scale of a dam, e.g., measured as height of the dam wall, increases the ab-
21
solute investment required exponentially, e.g. a 100m high dam wall is four
times more costly than a 50m wall (R2 = 0.27, F = 92.5, p < 0.01). An even
Furthermore, the rate of cost overrun outliers increases with increase in dam
tween dam height and hydropower installed capacity (r = 0.47, p < 0.01), ev-
idence suggests that larger scale in general is prone to outlying cost over-
Not only are large dams costly and prone to systematic and severe budget
overruns, they also take a long time to build. Large dams on average take 8.6
tions:
higher than the estimate with a median of 27% (or 1.7 years) as shown
10. Graphing the dams schedule overruns also reveals a fat tail as shown
in Figure 5, albeit not as fat as the tail of cost overruns. Costs are at a
Median
Mean
11. There is less variation in schedule overruns across regions than cost
time.
count any covariates, shows that increase in the scale of a dam, e.g., meas-
Large scale is intimately linked with the long-term (see model 2 below). The
actual implementation schedule, reported here, does not take into the ac-
count lengthy lead times in preparing the projects. Dams require extensive
studies, and political negotiations. The actual implementation cycles are far
longer than the average of about 8.6 years, as shown in our data, that it takes
that the benefits of large dams (even assuming that large dam generate bene-
fits as forecasted) do not come online quickly enough. The temporal mis-
match between when users need specific benefits and when these benefits
ments that can bridge needs quickly, without tremendous time lags, are
24
3.3 Multilevel regression analysis of cost and schedule performance
ted using the lme procedure in the nlme package in R software. This
Laird and Ware (1982) but allowing for nested random effects. The within-
as noted in Table 2. Using stepwise variable selection, we are not only able to
fit explanatory models for cost and overruns and estimated duration but also
schedule and the long-term inflation rate in the country in which the project
pending on the country whilst holding the inflation rate constant (see figure
Table 3. Model 1 - Significant variables for cost accuracy for large dam pro-
jects (constant local currency)
Regression Standard 2-tailed
Variable t-stat
coefficient error significance
Intercept 1.402 0.185 7.560 0.000
Log estimated duration
-0.100 0.041 -2.424 0.016
(months)
26
Log of country's long-
-0.085 0.029 -2.930 0.005
term inflation rate (%)
Note: Dependent variable is the estimated/actual cost ratio (i.e. 1/x of the cost over-
run to remove excessive skewness), based on 239 observations. Since the de-
pendent variable in Model 1 is the inverse of the cost overrun a negative sign on the
coefficients of both significant variables suggests that an increase in the estimated
duration or long-term inflation rate increases the cost overrun.
The first finding in Model 1 is that the larger the estimated implementation
schedule the higher the cost overrun (p = 0.016), all other things being equal,
longer in the future they are asked to project the risks facing a large
= 0.04) of the dam wall are significant variables associated with the es-
27
be seen from the coefficients in Table 4: a greater height, installed ca-
not only acts as a temporal variable but also as a surrogate for scalar
variables such as wall height (which is also highly correlated with in-
stalled capacity). The larger the dam, the longer the estimated imple-
Taken together, the multilevel models for cost overruns and estimated
schedule suggest that longer time horizons and increasing scale are underly-
The second finding in Model 1 is that higher the long-term inflation rate of
the host country the higher the cost overrun suffered by a dam (p = 0.02). The
long-term inflation rate was calculated by fitting a linear model to the log of
28
the time series of the GDP deflator index of each country. The slope of this
fitted line can be interpreted as the annual average growth rate of the log in-
flation for each country. This slope is a different constant for each country
with some countries such as Brazil with a considerably higher long-term in-
flation rate, and hence greater propensity to cost overruns, than China or the
United States. Moreover, this slope is stable in the short-run (it takes years of
high or low inflation to change this slope) and hence our estimate can be as-
sumed to be a reliable predictor. Recall that the cost overrun is being meas-
ured in constant terms (i.e. with the effects of inflation removed); yet Model
when making durable investments. The multilevel model thus suggests that
once country specific factors have been taken into account the factor that
or not a country is a democracy; the per capita income of the country in 2000
constant USD in the year of the decision to build; the planned installed ca-
pacity (MW); and planned length of the dam wall (meters). Avid dam build-
trolled for this fact by including a dummy variable for South Asia in the
Table 5. Model 3 - Significant variables for schedule slippage for large dam
projects
Regression Standard 2-tailed
Variable t-stat
coefficient error significance
*Dummy based on the Polity2 variably of Polity IV regime index. Score of +10 to +6 = democracy;
score of +5 to -10 = autocracy.
Note: Dependent variable is 1/x of the actual/estimated schedule ratio, based on
239 observations6.
are systematically more optimistic than autocracies even after controlling for
the coefficient is large suggesting that political process has profound impact
autocracies to build large dams by fitting a model to explain the actual im-
30
that effects of political regime on the actual schedule are not significant. In
other words, while democracies do not take longer to build large dams than
Given its vast scope, we defer a further investigation of this important result
cally elected politicians to present a rosier picture about large dams than
Second, countries with a higher per capita income in constant 2000 USD in
the year of decision to build tend to have lower schedule overruns than
countries with lower per capita income. We concur with the interpretation of
Bacon and Besant-Jones (1998, p. 325) that the best available proxy for most
nomic support that a country can provide for the construction of complex
spite seemingly the most in need of complex facilities such as large dams,
ought to stay away from bites bigger than they can chew.
Note: Dependent variable is log of the actual construction schedule, based on 239 observations.
MW installed capacity has the opposite effect. Model 3 in Table 5 shows that
the size of coefficients for the two significant variables related to physical
scalei.e. Log of dam wall length (m) and Log of hydropower installed ca-
pacity (MW)is approximately the same but with the opposite sign6.
are bespoke constructions tied to the geological and other site-specific char-
lar fashion. For example, the 690 MW installed capacity of the recently com-
that require onsite construction, e.g. dam wall, are more prone to schedule
ject designs that seek to reduce the bespoke and onsite components in favor
certainty.
This conjecture is supported by Model 4 in Table 6, which shows that the ac-
pacity does not necessarily equate with a smaller dam. For example, it is not
when, for instance, the dam is primarily being used for irrigation or flood
management purposes.
The statistical results reported in the preceding sections show that cost and
schedule estimates of large dams are severely and systematically biased be-
low their actual values. While it is beyond the scope of this paper to discuss
previous findings that point to twin problems that cause adverse outcomes
in the planning and construction of large and complex facilities such as large
gue that large scale and longer planning time horizons exacerbate the impact
large dams typically face to illustrate the statistical results reported above.
33
Globally, experts optimism about several risk factors contribute to cost over-
runs in large dams. For example, the planning documents for the Itumbiara
hydroelectric project in Brazil recognized that the site chosen for the project
was geologically unfavorable. The plan optimistically declared, the cost es-
for the removal of larger amounts [of compressible, weak, rock] if further in-
vestigations show the need (World Bank, 1973). This weak geology ended
up costing +96% of the base cost in real terms. Itumbiaras case is illustrative
little planners can do to hedge against it. For example, exhaustive geological
investigation for a large dam can cost as much as a third of the total cost
(Hoek and Palmieri, 1998); at which point still remains a considerable chance
ument was upbeat that there will be no changes in the exchange rate be-
tween the Colombian Peso and the U.S. dollar during the construction peri-
ence in Colombia where the Government has been pursuing the enlightened
34
(World Bank, 1970). In fact, the Colombian currency depreciated nearly 90%
Since over half the projects costs covers imported inputs, this currency de-
preciation caused a 32% cost overrun in real Colombian Peso terms. Curren-
cy exposure arises when the inputs required to build a project are denomi-
nated in one currency but the outputs in another, or vice versa. The outputs
any increases in tax receipts a dam may enable for the host government also
Since the USD liabilities also have to eventually be paid in local currency,
35
currency exposure consistently proves to be a fiscal hemorrhage for large
projects.
match the currency of liabilities with the currency of future revenue are
preferable.
Although, following convention, our cost analysis excludes the effects of in-
initial budget for Brazils Estreito dam (1965-1974), or 110-times initial budg-
Tarbela dam, it was assumed that inflation would not have a signification
impact on the projects costs. The appraisal report wrote: A general contin-
gency of 7% has been added in accordance with normal practice for works
of this size and duration (World Bank, 1968). The project, launched in 1968,
was meant to start full commercial operation in 1976, but the opening was
was 380%; the actual cost of the dam in nominal terms nearly four times the
flation rate (Pickrell, 1992, p. 164). For our reference class, 8 out of 10 large
36
dams came in late with an average delay of 2.3 years. Moreover, forecasters
expected the annual inflation rate to be 2.5% but it turned out to be 18.9%
(averages for the entire sample). Large dams have a high propensity to face
unanticipated inflation.
ergy alternatives that can be built sooner and with lower risk of schedule
Large dams are typically financed from public borrowing. While our calcula-
tions exclude debt-servicing, cost overruns increase the stock of debt but also
the recurring financing costs that can further escalate if interest rates go up.
The optimistic risk assessments of the costs of large dams are consistent with
explosive growth of Third World debt (Bulow and Rogoff, 1990; Mold,
2012). For example, the actual cost of Tarbela dam, the majority of which was
stans external public debt stock between 1968-1984; or 12% for Colombias
37
Table 7. Total stock of public net external debt (US$
current, MM)
Year Colombia Pakistan
1968 3,252.4
1970 1,296.6
1977 2,699.6
1984 9,692.8
Cost of mega-dam
Chivor dam Tarbela dam
(US$ current MM)
168.7 1,497.9
12.0% 23.0%
sults: bigger projects entail uncontrollable risks, which even when an-
age cost per unit as output increases. Instead our argument is that any
sorb adverse outcomes of big bets gone awry often face financial ruin.
38
Similarly, policymakers, particularly in countries at lower levels of eco-
forecasting (RCF)
As discussed in the methods section, the third step of the outside view or
RCF techniques is to compare a specific venture with the reference class dis-
tribution, in order to establish the most likely outcome for the specific ven-
ture. Thus if systematic errors in the forecasts generated using the inside
biased outside view forecast. For example, empirical literature has estab-
lished that rail projects suffer a cost overrun of 45% on average (Flyvbjerg,
2008; also see Table 8). The 50th percentile cost overrun for rail projects is 40%
and the 80th percentile is 57%. Based on these findings, RCF techniques sug-
gest that decision-makers ought to apply a 57% uplift to the initial estimated
budget in order to obtain 80% certainty that the final cost of the project
more risk tolerant then they could apply a 40% uplift to the initial estimated
budget but then there will remain a 50% chance that the proposed project
In line with the RCF techniques, the third and final step of our investigation
on dams was to derive a good predictor of cost and schedule overruns for
39
proposed large dams based on the distributional information of the reference
class. This predictor serves to correct the systematically biased ex ante cost
schedule overrun (see Kahneman and Tverksy, 1979b; Flyvbjerg, 2006, 2008).
First, using traditional RCF (Flyvbjerg, 2006, 2008), we traced the empirical
multilevel Models 1 and 3, described above, for predicting cost and schedule
two reasons: First both models are fitted with variables known ex ante. Se-
cond, both models were successfully fitted with only a few significant varia-
For example, Model 1 on cost overruns has only two significant variables
estimate schedule and the long-term inflation rate of the host country. Data
on both these variables is readily available for any proposed large dam mak-
ing it possible to predict the cost overrun before construction begins. We il-
With respect to cost overruns, using traditional RCF (Flyvbjerg, 2006, 2008),
we find that if planners are willing to accept a 20% risk of a cost overrun, the
uplift required for large dams is +99% (i.e. ~ double experts estimates) as
willing to accept a 50-50 chance of a cost overrun, the uplift required is 26%
40
Figure 7: Required uplift for large dam projects as function of the max-
imum acceptable level of risk for cost overrun, constant local currency
terms (N = 245)
!
In terms of cost overruns, Figure 7 also illustrates that large dams are one of
the riskiest asset classes for which valid and reliable are available. Compare,
for example, Figure 7 with reference class forecasts previously conducted for
rail, road, tunnel, or bridge projects (Flyvbjerg, 2006, 2008) also summarized
in Table 8.
Second, using our multilevel Model 1 we were able to derive predictions for
cost overrun (in constant local currency) and schedule overrun respectively.
42
Experts estimate, for instance, that Pakistans Diamer-Bhasha dam, whose
construction began shortly after the 2010 floods, will cost PKR 894 billion
2008 GDP) (WAPDA, 2011). The dam is forecasted to take ten years from
2011 and become operational in 2021. Using our first approach, the reference
class forecast for cost overruns suggests that planners need to budget PKR
1,788B (USD25.4B) in real terms to obtain 80% certainty of not exceeding the
possibility.
Using our second approach, our multilevel Model 1 predicts that given the
10 year estimated duration and a long-term inflation rate of about 8% the ex-
pected (average) cost overrun of a large dam in Pakistan will be 44% (PKR
for the cost overrun on the Diamer-Bhasha dam is 44% at which point there
the US, Model 1 predicts that it would only suffer a cost overrun of 16%,
which the much larger US economy could absorb without any lasting dam-
age.
first approach, the reference class forecast for schedule slippage suggests that
43
planners for large dams around the world need to allow for a 66% schedule
overrun to achieve 80% certainty that the project will be completed within
take 10 years to build (2011-2021), planners need to adjust their schedule es-
timate upwards to nearly 17 years (i.e. an actual opening date of 2028). Using
our second approach, our multilevel Model 3 predicts that given that the
ed government, when the per capita income was USD 497 in 2000 constant
dollars, a dam wall length of 998 meters, and an installed capacity of 4,500
MW, the expected outcome is a 60% schedule overrun. Thus, using either
remains a 20% risk of further delay. Pakistan is facing an energy crisis today
(Kessides, 2011). A dam that brings electricity is 2027 will be a little late in
coming.
being built in the US (with its high per capita income of approximately USD
0.05%. Recall that per capita income is a useful proxy for the economic sup-
port that a country can provide for the construction of complex facilities.
This suggests that rich and not developing countries best attempt very large
energy projects, such as large dams. Even so, richer countries should adopt
the risk management measures of the outside view illustrated here to choose
44
Using their inside cost estimates, the net present benefits to cost ratio of
the dam according to experts is 1.43 (WAPDA, 2011). Even assuming ex-
99% in constant prices suggest that the benefits to cost ratio will be below
discussing potential effects of inflation and interest rates, potential social and
Our reference class forecasting techniques suggests that other proposed large
dam projects such as the Belo Monte, Myitsone, or the Gilgel Gibe III among
many others in early planning stages are likely to face large cost and sched-
ule overruns seriously undermining their economic viability. Large dams al-
deployed to better uses, sinking vast amounts of land that could have yield-
ed cash flows and jobs from agricultural, timber, or mineral resources. Risks
of large dams. Decision-makers are advised to carefully stress test their pro-
posed projects using the risk management techniques of the outside view
The outside view techniques applied to large dams have broader application
45
investors a framework to improve the upfront selection among alternatives.
The problems of cost and schedule overrun are not unique to large hydro-
power dams. Preliminary research suggests that other large-scale power pro-
issues. Our research of large hydropower projects reveals that there is a seri-
ous dearth of valid and reliable data on the risk profiles of actually complet-
ed energy projects across the board. Much of the data in existing literature
are drawn from surveys and interviews of dubious validity. At times, inter-
assets of large, medium, and small scales across production technologies. For
example, comparing the likely actual cost, schedule, and production vol-
We propose that prior to making any energy investment, policy makers con-
sult a valid and reliable outside view or reference class forecast (RCF)
46
not only from the perspective of financial cost and benefit but also en-
profiles of energy alternatives does not yet exist. We have sought to bridge
projects. As a venue for further research we hope valid and reliable data on
the actual cost, schedules, benefits, and impacts of other production technol-
ogies will become available to enable comparative analysis with novel impli-
References:
Ansar, A., Flyvbjerg, B., Budzier, A., 2013. Do Hard States Make Better Decisions? Incentives
versus Biases. Unpublished manuscript, 42 pp., Sad Business School, University of Oxford
Ansar, A., Flyvbjerg, B., Budzier, A., 2012. Big is Fragile: The Effects of Temporal Uncertainty and
Sunk Costs on the Outcomes Of Big Ventures. Unpublished manuscript, 58 pp., Sad Busi-
ness School, University of Oxford.
Bacon, R.W., Besant-Jones, J.E., 1998. Estimating construction costs and schedules: Experi-
ence with power generation projects in developing countries. Energy Policy 26, 317
333.
Bakker, K., 1999. The politics of hydropower: Developing the Mekong. Political Geography
18, 209232.
Billington, D.P., Jackson, D.C., 2006. Big dams of the New Deal era: a confluence of engineer-
ing and politics. University of Oklahoma Press.
Buehler, R., Griffin, D., Ross, M., 1994. Exploring the planning fallacy: Why people under-
estimate their task completion times. Journal of Personality and Social Psychology
67, 366.
Bulow, J., Rogoff, K., 1990. Cleaning up third world debt without getting taken to the clean-
ers. The Journal of Economic Perspectives 4, 3142.
47
Copeland, T., Tufano, P., 2004. A real-world way to manage real options. Harvard business
review 82, 9099.
Dawes, R.M., Faust, D., Meehl, P.E., 1989. Clinical versus actuarial judgment. Science 243,
16681674.
Duflo, E., Pande, R., 2007. Dams. The Quarterly Journal of Economics 122, 601646.
Federal Transit Administration (FTA), 2003. Predicted and actual impacts of new starts pro-
jects: Capital Cost, Operating Cost and Ridership Data. U.S. Department of Trans-
portation, Washignton, DC.
Flyvbjerg, B., 2003. Delusions of success: comment on Dan Lovallo and Daniel Kahneman.
Harvard Business Review 81, 121122.
Flyvbjerg, B., 2006. From Nobel Prize to project management: Getting risks right. Project
Management Journal 37, 5.
Flyvbjerg, B., 2008. Curbing optimism bias and strategic misrepresentation in planning: Ref-
erence class forecasting in practice. European Planning Studies 16, 321.
Flyvbjerg, B., 2009. Survival of the unfittest: why the worst infrastructure gets built--and
what we can do about it. Oxford Review of Economic Policy 25, 344.
Flyvbjerg, B., Bruzelius, N., Rothengatter, W., 2003. Megaprojects and risk: An anatomy of
ambition. Cambridge Univ Press, Cambridge.
Flyvbjerg, B., Budzier, A., 2011. Why your IT project may be riskier than you think. Harvard
Business Review 89, 2325.
Flyvbjerg, B., COWI, 2004. Procedures for dealing with optimism bias in transport planning:
Guidance document 2006-09 [WWW Document]. URL goo.gl/lLiNj
Flyvbjerg, B., Holm, M., Buhl, S., 2005. How (in) accurate are demand forecasts in public
works projects. Journal of the American Planning Association 71, 131146.
Flyvbjerg, B., Holm, M.S., Buhl, S., 2002. Underestimating costs in public works projects:
error or lie? Journal of the American Planning Association 68, 279295.
Flyvbjerg, B., Massimo, G., Lovallo, D., 2009. Delusion and deception in large infrastructure
projects. California Management Review 51.
Gelman, A., Hill, J., 2006. Data Analysis Using Regression and Multilevel/Hierarchical
Models. Cambridge University Press, Cambridge.
Gilovich, T., Griffin, D., Kahneman, D., 2002. Heuristics and biases: The psychology of intui-
tive judgment. Cambridge University Press.
Goel, R.K., Singh, B., Zhao, J., 2012. Underground infrastructures: planning, design, and
construction. Butterworth-Heinemann, Waltham, MA.
Goldstein, H., 2010. Multilevel Statistical Models, 4th ed. Wiley.
Hoek, E., Palmieri, A., 1998. Geotechnical risks on large civil engineering projects, in: Pro-
ceedings 8th Congres IAEG. pp. 7988.
Hufschmidt, M.M., Gerin, J., 1970. Systematic errors in cost estimates for public investment
projects, in: Margolis, J. (Ed.), The Analysis of Public Output. Columbia University
Press, New York.
ICOLD, 2010. International Commission on Large Dams [WWW Document]. URL
http://www.icold-cigb.org/
IEA, 2011. Policies and Scenarios [WWW Document]. URL
http://www.iea.org/publications/scenariosandprojections/
Kahneman, D., 1994. New challenges to the rationality assumption. Journal of institutional
and theoretical economics 150, 1836.
Kahneman, D., 2011. Thinking, fast and slow, 1st ed. Farrar, Straus and Giroux, New York.
Kahneman, D., Lovallo, D., 1993. Timid choices and bold forecasts: A cognitive perspective
on risk taking. Management Science 1731.
Kahneman, D., Tversky, A., 1979a. Prospect theory: An analysis of decision under risk.
Econometrica 263291.
48
Kahneman, D., Tversky, A., 1979b. Intuitive Prediction: Biases and Corrective Procedures,
in: Makridakis, S., Wheelwright, S.C. (Eds.), Studies in the Management Sciences:
Forecasting. North Holland, Amsterdam.
Kessides, I.N., 2011. Chaos in Power: Pakistans Electricity Crisis.
Kreft, I.G.G., Leeuw, J. de, 1998. Introducing Multilevel Modeling. Sage Publications Ltd.
Laird, N.M., Ware, J.H., 1982. Random-effects models for longitudinal data. Biometrics 963
974.
Lovallo, D., Kahneman, D., 2003. Delusions of success. Harvard Business Review 81, 5663.
Luehrman, T.A., 1998. Strategy as a portfolio of real options. Harvard Business Review 76,
89101.
McCully, P., 2001. Silenced rivers: the ecology and politics of large dams. Zed Books, Lon-
don.
Merewitz, L., 1973. How do urban rapid transit projects compare in cost estimating experi-
ence. Institute of Urban & Regional Development, University of California.
Merrow, E.W., McDonnell, L., Argden, R.Y., Corporation, R., 1988. Understanding the out-
comes of megaprojects: A quantitative analysis of very large civilian projects. Rand
Corp.
Mold, A., 2012. Will it all end in tears? Infrastructure spending and African development in
historical perspective. Journal of International Development 24, 237254.
Nilsson, C., Reidy, C.A., Dynesius, M., Revenga, C., 2005. Fragmentation and flow regula-
tion of the worlds large river systems. Science 308, 405408.
Osborne, J. W., Overbay, A. 2004. The power of outliers (and why researchers should always
check for them). Practical Assessment, Research & Evaluation 9(6). Retrieved Sep-
tember 1, 2013 from http://PAREonline.net/getvn.asp?v=9&n=6
Pickrell, D.H., 1989. Urban rail transit projects: forecast versus actual ridership and costs (
No. PB90148693), Office of Grants Management, Urban Mass Transportation Ad-
ministration. Transportation System Center, U.S. Department of Transportation,
Cambridge, MA.
Pickrell, D.H., 1992. A desire named streetcar: Fantasy and fact in rail transit planning.
American Planning Association. Journal of the American Planning Association 58,
158176.
Rasbash, J., Browne, W., Goldstein, H., Yang, M., Plewis, I., Healy, M., Woodhouse, G.,
Draper, D., Langford, I., Lewis, T., 2009. A users guide to MLwiN. University of
London, Institute of Education, Centre for Multilevel Modelling.
Recchia, A., 2010. R-squared measures for two-level hierarchical linear models using SAS.
Journal of Statistical Software 32.
Richter, B.D.a , Postel, S.b , Revenga, C.c , Scudder, T.d , Lehner, B.e , Churchill, A.f , Chow,
M.., 2010. Lost in developments shadow: The downstream human consequences of
dams. Water Alternatives 3, 1442.
Salazar, J.G., 2000. Damming the child of the ocean: The Three Gorges project. The Journal of
Environment & Development 9, 160174.
Scudder, T., 2005. The future of large dams: Dealing with social, environmental, institutional
and political costs. Earthscan, London.
Schlissel, D., Biewald, B., 2008. Nuclear Power Plant Construction Costs. Synapse Energy
Economics, Inc.
Singh, S., 2002. Taming the Waters: The Political Economy of Large Dams in India. New
Delhi: Oxford University Press.
Sovacool, B.K., Bulan, L.C., 2011. Behind an ambitious megaproject in Asia: The history and
implications of the Bakun hydroelectric dam in Borneo. Energy Policy 39, 48424859.
Sovacool, B.K., Cooper, C., 2013. The governance of energy megaprojects: politics, hubris
and energy security. Edward Elgar, Cheltenham, UK.
49
Stone, R., 2011. Hydropower. The legacy of the Three Gorges Dam. Science 333, 817.
Taleb, N.N., 2010. The Black Swan: The Impact of the Highly Improbable with a new section:
On Robustness and Fragility, 2nd ed. Random House Trade Paperbacks.
Tversky, A., Kahneman, D., 1974. Judgment under uncertainty: heuristics and biases. Sci-
ence 185, 11241131.
Wachs, M., 1989. When planners lie with numbers. Journal of the American Planning Asso-
ciation 55, 476479.
WAPDA, 2011. Diamer-Basha dam project status as on 20 june 2011.
World Bank, 1968. Report and recommendation of the president to the executive directors
on a proposed loan for the Tarbela project and on proposed extension of closing
dates of loan 266-PAK (Indus basin project), Pakistan ( No. P-616). The World Bank,
Washington, DC.
World Bank, 1970. Staff appraisal report (pu-31a) of the Chivor hydroelectric project, Inter-
conexion Electrico S.A., Colombia ( No. PU-31a). The World Bank, Washington, DC.
World Bank, 1973. Staff Appraisal Report (150a-BR) of the Itumbiara hydroelectric power
project, FURNAS, Brazil ( No. 150a-BR). The World Bank, Washington, DC.
World Bank, 1996. Estimating construction costs and schedules: Experience with power gen-
eration projects in developing countries ( No. WTP325). The World Bank, Washing-
ton, DC.
World Commission on Dams, 2000. Cross-check survey: final report. Cape Town, South Af-
rica.
Ziv, G., Baran, E., Nam, S., Rodrguez-Iturbe, I., Levin, S.A., 2012. Trading-off fish biodiver-
sity, food security, and hydropower in the Mekong River Basin. Proceedings of the
National Academy of Sciences 109, 56095614.
ENDNOTES
1 Cost overruns can also be expressed as the actual outturn costs minus estimated costs in
2 Note that the World Bank, Asian Development Bank, and the WCD typically report cost
data in nominal U.S. dollars. We, however, converted these data, adapting methods from
3 Hufschmidt and Gerin (1970) report data on over 100 dams built in the United States be-
tween 1933-1967. The salient results of the study were that in nominal USD terms dams built
by TVA suffered a 22 per cent cost overrun; U.S. Corps of Engineers overrun was 124 per
cent for projects built or building prior to 1951, and 36 per cent for projects completed be-
tween 1951 and 1964; while U.S. Bureau of Reclamation overrun was 177 per cent for pro-
jects built or building prior to 1955 and 72 per cent for all projects built or building in 1960
(ibid: 277). Despite its large sample, Hufschmidt and Gerin (ibid) do not report data broken
50
down project-by-project. The validity and reliability of these data could not thus be estab-
4 A more comprehensive enquiry into planned versus actual benefits of dams is postponed
until a future occasion but data available on 84 of the 186 large hydroelectric dam projects
thus far suggests that they suffer a mean benefits shortfall of 11%.
5 Recchia (2010, p. 2) explain further for why a R-squared measure cannot be used for a mul-
tilevel model because it unlike a single-level model which includes an underlying assump-
tion of residuals that are independent and identically distributed. Such an assumption could
easily be inappropriate in the two[or multi]-level case since there is likely to be dependence
among the individuals that belong to a given group. For instance, it would be difficult to
imagine that the academic achievements of students in the same class were not somehow
related to one another. Also see Kreft and Leeuw (1998) and Goldstein (2010).
6 Note further that like in Model 1, the dependent variable in Model 3 is the inverse of
schedule overrun (i.e. 1/x of the schedule overrun or Estimated/Actual schedule). Thus a
negative sign on the Log of dam wall length (m) suggests that an increase in wall length de-
creases the inverse of the schedule overrun. In other words, increase in wall length increases
schedule overrun.
51
Appendix 1: Visual Representation of Model 1 (reported in Table 3)
52