Raw signal
Phones record acceleration on the vertical axis roughly 5 times per second during Uber trips.
infrastructure
costs
inequality
Every pothole has a price. One study computed it to the cent.
Accelerometers in Uber trip phones measured roughness across more than five million urban US road segments. In the authors’ November 26, 2024 revision, one standard deviation more roughness than the median is associated with 33 cents per mile in the full model. This is a statistical association, not a driver’s bill or proof of causation.
The authors’ current pages list the November 26, 2024 revision as “revise & resubmit” at the Quarterly Journal of Economics. It is not a final publication. NBER Working Paper 31981 from 2023 is an earlier version.
5M+
urban road segments measured in the US
$0.33
per mile for one standard deviation above median roughness in the full model
$450
authors’ rounded annual example for the Black population-share contrast at 3,000 local miles
1,285
locality pairs in the revised border sample
What your road costs you, per year
Choose a result and change the mileage. The authors’ 2024 revision estimates 33 cents per mile driven for one standard deviation more roughness in the full model and 31 cents in the simple model. The third result shows the contrast associated with the Black population share, not an individual bill.
At 3,000 local miles per year (4,800 km), the full model associates one standard deviation above median roughness with this amount. It is not an actual bill.
$990
$0.33/mi
3,000
≈ 4,800 km
Formula: annual estimate = local miles × per-mile coefficient. At 3,000 miles, the full model gives $990. These are statistical estimates, not individual bills.
Over 10 years, at the same mileage: $9,900.
How the phone became a road sensor
The researchers used the vertical acceleration recorded by Uber drivers' phones, then separated driving style from the physical quality of the asphalt. Validation came from official indices, railroad crossings, and Chicago resurfacing records.
Standardization
For each segment, roughness is predicted at a fixed speed, so the measurement reflects the road rather than the driver.
Validation
Correlation with the official IRI index: 0.5-0.7 on highways, 0.7 on arterials. Segments with railroad crossings come out, predictably, much rougher.
Pricing
Cost is inferred from the speeds drivers freely choose when crossing locality borders or entering freshly resurfaced stretches.
A rough road slows you down and taxes you
The authors’ November 2024 revision estimates 33 cents per mile for one standard deviation above median roughness in the full model (95% confidence interval: 28–40 cents). The simple model estimates 31 cents (27–34 cents).
At 3,000 local miles per year, the estimates are $990 and $930. They are modelled road-cost associations, not amounts billed directly to drivers. The 2023 NBER Working Paper 31981 separately estimated 31 cents for a median-roughness road; that is a different measure from the 2024 simple model’s 31 cents for one standard deviation more roughness.
Bad infrastructure is unevenly distributed
Across the 100 largest US metro areas, local roads are rougher where incomes are lower. The gap shows up mostly between localities: they own most local roads and have very different tax bases.
Table 5, column 4, with climate controls, estimates 15.1 cents per mile for a 100-percentage-point difference (0–100%) in the Black population share. The authors report the 3,000-mile example as $450; multiplying the rounded coefficient gives $453. About 85% of the association lies between localities. Column 5, with metro fixed effects, estimates 14.7 cents; with locality fixed effects, it falls to 2.1 cents, or about $63 at the same mileage.
Table 9, column 3, associates a one-unit increase in log median income with 4.3 cents less per mile, across 1,235 towns with more than 5,000 residents. One log unit means income is about 2.72 times higher. The model includes metro fixed effects and controls for expenditure, distance from the metro centre, population, employment, land area, commuting share, and Uber use; it does not include racial shares. At 3,000 miles, the calculation is −$129 per year. This is an association, not a causal effect.
In Table 6, column 6, the Black-share coefficient is −0.5 cents per mile (standard error: 0.6 cents) after locality fixed effects and controls for income, poverty, and unemployment. It is not statistically significant. These models show associations, not causation.
Repairs rarely reach the worst roads
The 2024 revision studies resurfacing decisions in eleven cities. It separates observed roughness from a modelled repair ranking based on costs borne by drivers.
New York: estimated weight of 9–16%
New York decisions are consistent with a 9–16% weight on the modelled prioritization criterion based on roughness and traffic. This estimates decision weights; it is not an official city rule.
Two measures, different results across cities
Observed roughness is modestly correlated with work in New York, Denver, and Philadelphia. In the cost-based prioritization model, the association is positive in at least one panel for Columbus, Denver, Philadelphia, Charlotte, and Chicago; it is not statistically significant in the other five cities.
Two locality samples
The authors received responses from 25 Massachusetts localities and 96 from the rest of the US. In the rest-of-US sample, 62% say they resurface less than 30% of roads that need it.
Utility work changes the order
Upcoming utility work affects prioritization for 44% of Massachusetts localities and 46% in the rest of the US. When assessing resurfacing need, the shares are 76% and 60%, respectively.
The public question
If roughness can be measured at segment level, the criterion "worst roads first" becomes testable. Any city can compare its actual resurfacing list against a list ordered by the cost borne by drivers.
Budgets, utility work, safety, and political balance remain separate decisions. The test still puts a shared unit on the table: dollars per local mile, paid by the people who drive there.
The method travels. Any city with ride-sharing traffic, Bucharest included, is already generating the raw data. What's missing is the will to turn it into a public priority list.
Sources
Current figures come from the authors’ November 26, 2024 PDF and the cited tables. NBER Working Paper 31981 is the earlier 2023 version; its figures remain only where labelled as historical results.
Lindsey Currier, Edward L. Glaeser, and Gabriel E. Kreindler, “Infrastructure Inequality: Who Pays the Cost of Road Roughness?”, November 26, 2024
The PDF is printed with the date November 26, 2024. The authors’ current pages list this version as “revise & resubmit” at the Quarterly Journal of Economics; it is not a final publication. References used: Tables 2, 3, 5, 6, and 9–12; Sections 5–6.
Lindsey Currier’s Harvard research page
Lindsey Currier’s Harvard page lists the same version and status; co-author Gabriel Kreindler’s page provides a matching listing.
NBER Working Paper 31981 · December 2023
This is the earlier version of the study. Its 2023 results, including the 31-cent estimate for a median-roughness local road, are not estimates from the 2024 revision.