Flyvbjerg 2002: The Study Behind "Nine in Ten Projects Run Over Budget"
The first statistically significant study of cost overrun in transport infrastructure. This page gives the sample, the headline numbers by mode, the methodology, and the canonical citation so you can use the figures with confidence.
What is the Flyvbjerg 2002 study?
"Underestimating Costs in Public Works Projects: Error or Lie?" (Flyvbjerg, Holm & Buhl, 2002) was the first statistically significant study of cost escalation in transport infrastructure. It examined 258 projects worth about US$90 billion, built across 20 nations between 1927 and 1998. Actual costs exceeded the estimate in 86% of cases: rail ran 44.7% over on average, fixed links (bridges and tunnels) 33.8%, and roads 20.4%. Accuracy showed no improvement across the seventy years, which led the authors to conclude that underestimation is driven by strategic misrepresentation, not honest error.
Average cost escalation by transport mode
| Project type | Mean cost escalation | Standard deviation |
|---|---|---|
| Rail | +44.7% | 38.4 |
| Fixed links (bridges and tunnels) | +33.8% | 62.4 |
| Roads | +20.4% | 29.9 |
| All projects (average actual vs estimate) | +28% | — |
From Flyvbjerg, Holm & Buhl (2002), Table 1. Escalation is measured in real (inflation-adjusted) terms against the cost estimated at the decision to build. The standard deviations are as striking as the averages: they show that cost-overrun risk is not only large but highly variable, so an average figure understates the danger for any single project. Overruns ran roughly 65% higher for projects outside Europe and North America.
Sample and methodology
The sample is 258 transport infrastructure projects: rail, fixed links (bridges and tunnels) and roads, worth a combined US$90 billion in constant 1995 prices. The projects were built in 20 nations across Europe, North America and elsewhere, and completed between 1927 and 1998, giving roughly seventy years of coverage. It was, at publication, the first study of cost overrun in the sector large enough to draw statistically significant conclusions.
Cost overrun is defined as actual outturn cost minus estimated cost, in real terms, measured against the estimate made at the time of the decision to build. This is the same definition Flyvbjerg carried into his later 16,000-plus project database: it deliberately anchors to the figure used to justify the project, not to a later re-baselined budget, because re-baselining is one of the main ways overruns are hidden.
Because the estimate is taken at the decision to build, the finding is not simply that projects get more expensive. It is that the numbers used to approve them are systematically too low, and by a wide margin.
"Error or lie?" the study's central argument
The paper's title poses the question directly. If cost underestimation were honest error, you would expect it to be random (some projects over, some under) and to shrink over time as forecasting techniques improved. The data show neither. Underestimation is one-directional (costs are almost always too low, not too high) and it does not improve across the seventy-year window.
The authors conclude that the pattern is best explained by strategic misrepresentation: the political and organisational incentive to make a project look affordable enough to win approval. That argument is what later became the behavioural core of Flyvbjerg's work, from optimism bias and the planning fallacy through to the Iron Law of Megaprojects.
How this differs from the 16,000-project database
These are two different Flyvbjerg datasets that are often conflated. The 2002 study on this page is the original 258-project transport sample. The much larger figure people cite today, more than 16,000 projects across many sectors, comes from the database assembled later and summarised in How Big Things Get Done (Flyvbjerg & Gardner, 2023). If you need per-sector base rates (IT, nuclear, Olympics and so on), use the database page; if you need the seminal transport-cost-overrun finding, cite the 2002 study here.
How to cite
Open-access author copy: SSRN. Author's site: bentflyvbjerg.com.
The paper is behind a publisher paywall at Taylor & Francis, but the author's accepted manuscript is freely downloadable from SSRN, so the figures here can be checked against the primary source.