Paper
Davis, Lucas W. (2008). Journal of Political Economy, 116(1): 38–81
Read the original paper → Opens at the publisher; use the DU library / JSTOR login if it asks for access.
An evaluation of Mexico City’s 1989 “Hoy No Circula” driving ban using high-frequency pollution-monitor data and a regression-discontinuity-in-time design. The explainer below shows how to spot a policy effect in a noisy time series — and what it means when there isn’t one.
An interactive, section-by-section guide to The Effect of Driving Restrictions on Air Quality in Mexico City by Lucas W. Davis, Journal of Political Economy (2008)
Hoy NoCircula
Spanish for “today, it doesn’t circulate.” From November 1989, most cars in Mexico City had to stay home one weekday a week.
Pick the last digit of your license plate
Plates ending in 5 or 6 carry a yellow sticker and stay home on Mondays, from 5 a.m. to 10 p.m.
Drive anyway and the car could be impounded for 48 hours, plus a fine of about US$200. Every weekday, one in five of the 2.3 million vehicles covered, about 460,000, stays off the road.
That was the plan to clean up the city’s badly polluted air. Using hourly readings from Mexico City’s air-monitoring stations, Lucas Davis asks whether it worked, and what drivers did instead.
A day without a car
Covers the paper’s Section I, the introduction
Mexico City’s air was, and still is, a serious health problem. Between 1986 and 2005, ozone levels exceeded the World Health Organization’s guideline on 92% of all days, and most residents avoided spending time outdoors when the air was bad. Air pollution is linked to respiratory infections, heart and lung disease, and infant deaths.
On 20 November 1989, the city introduced Hoy No Circula (HNC). Each vehicle was banned from the roads on one weekday, from 5 a.m. to 10 p.m., according to the last digit of its license plate. The rule covered the whole metropolitan area and most private and commercial vehicles. Taxis, buses, police cars, ambulances and fire trucks were exempt, as were commercial vehicles running on liquid propane gas or carrying perishable goods.
Enforcement was strict from the first day. Plates make violators easy to spot, and although police could be bribed, paying several bribes on a single short trip was expensive. Compliance was close to universal.
One in five cars leaves the road every weekday. How much would you expect average air pollution to change?
Why study this? Driving restrictions are cheap to enforce and need little public investment, and they spread quickly: Bogotá’s pico y placa, Santiago’s restricción vehicular and São Paulo’s rodízio all followed. By the time of the paper, more than 50 million people lived in cities that restricted driving by license plate. Yet there was almost no evidence on whether such rules work. Eskeland and Feyzioglu (1997) had studied gasoline sales and car registrations; this paper is among the first to test the effect on air quality directly.
Bottom line. HNC is a clean natural experiment: a sharp policy change on a known date, affecting hundreds of thousands of vehicles a day, in a city with hourly air-quality records.
Measuring the air
Covers Section II, the air-quality data
The city’s Automated Environmental Monitoring Network, set up in 1986, records hourly levels of five pollutants at stations across the metropolitan area. The US Environmental Protection Agency certifies the stations every year, and ozone readings are accurate to within about 3%.
The analysis covers 1986–93, the widest window that is symmetric around November 1989. It uses only stations that were already operating in 1986, so that stations added later can’t change the mix; none closed or moved during these years.
Meet the five pollutants
Averages and station counts from Table 1; WHO guidelines and shares of days above them from Table 6 and its footnote (1986–93); peak times read approximately from Figure 2; vehicle shares from the city’s 2004 emissions inventory, as cited in the paper. Levels are in parts per million (ppm).
How much of each pollutant comes from vehicles?
Share of emissions from vehicles in the city’s emissions inventory (Secretaría del Medio Ambiente 2004), as cited in the paper. Ozone isn’t emitted directly, so the relevant figure is for volatile organic compounds, one of its ingredients.
Two features make this a good setting to detect a policy effect. Pollution responds quickly: levels swing through the day with traffic, and at the city’s average wind speed of about 6 km/h, pollutants rarely stay in the air for more than a day. And vehicles dominate emissions, above all of carbon monoxide.
The main confounders are season and weather. Mexico City sits in a valley ringed by mountains that rise about 1,000 meters above it. In winter, cool air near the ground gets trapped under warmer air above, a temperature inversion that holds pollution in the valley. Every model therefore controls for month, weekday, hour and detailed weather.
The city’s environmental agency credited HNC with cutting emissions by 30 million tons a month. But that figure simply assumed that weekday vehicle emissions fell by 20%. If drivers adapted, the assumption fails. Measuring the air directly avoids it.
Bottom line. Hourly records from a fixed set of stations give a direct test of the policy, without assuming how drivers behave.
How to spot a policy effect
Covers Section III.A, the empirical strategy
The basic idea is to compare pollution after 20 November 1989 with pollution before, holding fixed everything else that moves it: season, weekday, hour and weather.
The model, term by term
Tap any part of the equation.
Pick a term to see what it does.
The danger is that other things changed over time as well: the economy, fuel quality, the car fleet. From 1994, for example, all new cars had to meet US emission standards. Davis tackles this in two ways.
Narrow windows. Estimate the model by least squares on shrinking windows around the start date, from 1986–93 down to 1989–90. The narrower the window, the less room for other changes to creep in. Windows shorter than two years are ruled out, because seasons could no longer be controlled for credibly.
Regression discontinuity. Use all of 1986–93, but add a very flexible polynomial in time, of the seventh to ninth order. The polynomial soaks up slow-moving changes, so the only thing left to explain a sudden jump on the start date is the policy itself.
What a regression discontinuity looks for
A schematic with simulated weekly data, for illustration only. In the paper’s Figure 3, the dots are real weekly pollution levels after removing seasonal, weekday, hourly and weather effects.
Why not compare Mexico City with another city? Its valley geography, transport system and sheer size mean that no other city makes a credible stand-in.
For economists: identification and inference
The regression discontinuity identifies the effect if everything else that affects pollution changes smoothly through 20 November 1989; formally, the conditional mean of the error must be continuous at the cutoff (Hahn, Todd and Van der Klaauw 2001). Unobserved factors can still move pollution, as long as they don’t jump on that date.
Diagnostic tests found serial correlation lasting 2–12 weeks in the least-squares models and 2–5 weeks in the RD models, so standard errors allow arbitrary correlation within 5-week blocks. Newey–West errors with a 5-week lag give similar answers.
Bottom line. If HNC had cut pollution, it should show up as a sudden drop in November 1989 that a smooth trend can’t explain.
Did average pollution fall?
Covers Section III.B, mean pollution levels
The first test uses all hours of all days. Pick a method to see its estimate of HNC’s effect on each pollutant, and on all five pooled together.
The estimated effect of HNC on pollution levels
Tables 2 and 3, with 95% confidence intervals (±1.96 standard errors). Coefficient × 100 is roughly the % change. The pooled (“stacked”) estimate lets everything except the HNC effect differ by pollutant.
Why do the least-squares estimates come out so much higher? Carbon monoxide rose during 1990 and ozone during 1991, before both fell in 1992–93. A before-and-after comparison attributes part of that rise to HNC, while the flexible trend absorbs it. Either way, neither method finds a fall.
Bottom line. Across pollutants and specifications, there is no evidence that average air quality improved.
Nights and weekends
Covers Section III.C, pollution by time of day and week
HNC applied only on weekdays between 5 a.m. and 10 p.m. If drivers moved banned trips to other times, pollution should rise at night and at weekends relative to the restricted hours.
When did pollution change?
Tables 4 and 5, with 95% confidence intervals. Each estimate compares the same period of the week before and after HNC.
This points to intertemporal substitution: people moved driving into hours the rule didn’t cover. And if they shifted across hours, they probably also shifted across days, driving more on the days they were allowed to, which helps explain why even the restricted hours show no absolute improvement.
For economists: the congestion angle
Moving trips out of the rush hour isn’t worthless. In Vickrey-style congestion models, the social cost of driving is highest at the peak, so shifting trips away from it can bring large welfare gains from lower congestion; the paper points to this literature on congestion pricing. It just doesn’t clean the air.
Bottom line. Relative to the hours it covered, HNC left nights and weekends dirtier. In absolute terms, no period got cleaner.
The worst days, and stress tests
Covers Sections III.D and III.E, extreme pollution and alternative specifications
Much of the harm to health may come from pollution spikes, so averages could hide a benefit. Two more outcomes check for that: each day’s highest hourly level, and whether the day broke a WHO guideline.
Did the worst pollution get any better?
Table 6, with 95% confidence intervals. Both outcomes are estimated for 1986–93 with a polynomial time trend.
Next comes a battery of alternative specifications, each aimed at a possible flaw in the main estimates.
Does the result survive other specifications?
Table 7: regression discontinuity with a seventh-order time trend, 1986–93, with 95% confidence intervals.
Bottom line. No decline in daily peaks, no fewer bad-air days, and the result holds up under every alternative specification.
Fuel, metro and buses
Covers Sections IV.A and IV.B, gasoline sales and public transport
Why didn’t the air improve? If drivers had left their cars at home and taken the metro or the bus, gasoline sales should have fallen and ridership risen. Davis checks both, applying the same discontinuity approach to monthly data.
What happened to fuel sales and public transport?
Tables 8 and 9, with 95% confidence intervals from Newey–West standard errors. Each row uses a different polynomial time trend; “main” marks the one discussed in the text.
Why no switch to public transport? Around 1990 there was roughly one car for every eight residents, so drivers came mostly from middle- and upper-income households with a high value of time. Buses and the metro were cheap but slower and less convenient. For people used to driving, the attractive substitutes were other private options: a second car, or a taxi.
For scale: a 1994 survey found that 64% of trips in the city were by bus, 17% by private car and 13% by metro.
Bottom line. HNC didn’t push drivers onto cleaner transport: gasoline sales held steady, bus use didn’t change, and metro use actually fell.
More cars, older cars
Covers Section IV.C, vehicle registrations and sales
A household with two cars whose plates end in different digits can drive every day of the week. So HNC gave families a reason to buy a second vehicle, or to keep an old one they would otherwise have scrapped.
Registrations and new-car sales
Table 10, with 95% confidence intervals from Newey–West standard errors; “main” marks the specification discussed in the text.
Mexico City’s fleet, in icons
Each icon is about 23,000 of the 2.3 million vehicles covered. The parked color follows the plate digit you picked at the top of the page. The slider spans the paper’s 95% confidence interval for the increase in registered vehicles.
Were the extra vehicles new or used? New-car sales rose too, but they were small relative to the fleet, so a little arithmetic settles it.
How many of the extra vehicles could have been new?
of the extra vehicles could have been new
The paper’s arithmetic: in 1990, new-car sales equaled 7.5% of the registered stock, and registered vehicles rose by about 19 log points (roughly 21%) when HNC began.
That matters because older cars pollute more. They lack modern emission controls, the controls they do have wear out, and in Mexico City many were poorly maintained or deliberately tuned for power. A few “super-emitters” did outsized damage: in remote-sensing tests of 30,000 vehicles in the summer of 1990, half of all carbon monoxide came from 24% of the fleet, and half of hydrocarbons from 14%.
How much more carbon monoxide does an older car emit?
A rough illustration that compounds the estimate, cited in the paper, that each extra year of age raises a vehicle’s carbon monoxide emissions by about 16% (Beaton, Bishop and Stedman 1992).
Couldn’t people just fake a second plate?
In Santiago, after similar restrictions, some drivers obtained extra plates illegally and swapped them on restricted days. That was unlikely to happen at scale in Mexico City: plates were tightly controlled, and every vehicle had to display a sticker matching its plate inside the rear window.
Bottom line. HNC grew the fleet by roughly 325,000 vehicles, overwhelmingly used ones: more cars, and dirtier ones.
What about taxis?
Covers Section IV.D, substitution to taxis
Taxis were exempt from HNC, and Mexico City had a lot of them: about 75,000 in 1989, one for every 100 residents, against roughly one per 600 in New York and one per 175 in Beijing. The fleet had grown by 7.8% a year during the 1980s, partly because taxi permits were handed out as political favors, so it was unusually large just when HNC began.
Could taxis absorb the banned driving?
extra driving per taxi per day
The paper’s arithmetic: 460,000 banned vehicles × about 36 km a day (from a 2004 government study) = 16.6 million km, or 221 km for each of the 75,000 taxis if all of it moved to taxis.
Is there direct evidence of a shift to taxis?
Tables 11 and 12, with 95% confidence intervals; “main” marks the specification discussed in the text.
Even a modest rise in taxi use could have hurt the air. Most taxis were air-cooled Volkswagen Beetles with no catalytic converters, a model whose new sales the US had banned from 1977 because of its emissions. Taxis averaged 11 years old in 1990, against 8 for private cars, and were driven much harder. Many were tuned for power, a practice that can double carbon monoxide emissions, and the 44-horsepower Beetle was a prime candidate.
Bottom line. There is no direct evidence of a shift to taxis, but the taxi fleet was large enough to absorb extra trips and dirty enough that even a small shift would have undercut any gains.
Costs without benefits
Covers Section V, the cost-benefit analysis
Cleaner air would have been worth a lot. The World Bank valued a 10% cut in ozone and fine particles (PM10) in Mexico City at about $882 million a year. Using US evidence, a 10% cut in suspended particulates would prevent roughly 800 infant deaths a year in the city, worth about $1.48 billion at a value of a statistical life of $1.85 million. All dollar figures are in 2006 US dollars.
But HNC produced no measurable improvement in the air, so none of those benefits can be counted. The costs are harder to see. Here Davis uses revealed preference: people who bought an extra car to get around the ban showed that avoiding it was worth at least that much to them.
What did the extra cars cost?
Cost against what cleaner air would have been worth, $ million a year
Section V of the paper. It rounds its total down to “more than $300 million”, which works out to $130 per vehicle a year and $2.50 per banned day. The two benefit bars are what a 10% cut would have been worth; HNC delivered no measurable cut.
Why this could overstate or understate the cost
It may overstate the cost: a second car has other uses, such as letting two people drive at once, and some households were already close to buying one.
It probably understates it: someone who bought a car valued avoiding the ban at least at the price, and possibly far more; the many drivers who didn’t buy a car were still inconvenienced; and police time spent enforcing HNC is left out entirely.
Bottom line. HNC imposed costs likely above $300 million a year with no detectable benefit, which makes it hard to justify on cost-effectiveness grounds.
Lessons for other cities
Covers Section VI, the conclusion
Put the pieces together and a consistent story emerges.
- Average air pollution
- Hoped: down. Found: no detectable change (+3.7% pooled, not significant).
- Nights and weekends
- Found: dirtier relative to ban hours, consistent with drivers shifting trips.
- Days above WHO guidelines
- Hoped: fewer. Found: no fewer, and more for carbon monoxide and sulfur dioxide.
- Gasoline sales
- Hoped: down. Found: no change (+1.8%).
- Metro ridership
- Hoped: up. Found: down about 8%.
- Bus ridership
- Hoped: up. Found: no change.
- Registered vehicles
- Found: up about 325,000, overwhelmingly used and older.
- Taxis
- Found: no change in their number or price, though extra taxi use can’t be ruled out.
The lesson is about substitution. Rationing one way of getting around helps only if people switch to cleaner alternatives. In Mexico City they switched hours, bought more cars, mostly used ones, and may have taken more taxis, undoing any gain.
Plate-based restrictions have spread well beyond Mexico City, and at the time of the paper Monterrey and Beijing were considering them. Davis argues that the Mexico City pattern is likely to repeat wherever people value fast, convenient travel, so a city should work out in advance what drivers will do instead.
Test yourself
Eight questions on the paper’s key results
Glossary
Terms used in the paper and this guide
- Catalytic converter
- An exhaust device that cuts emissions of carbon monoxide, hydrocarbons and nitrogen oxides.
- Clustered standard errors
- Standard errors that allow the errors within a group, here a 5-week block, to be correlated, because pollution shocks persist.
- Concession
- A government permit to operate a taxi or minibus. In Mexico City it was sold together with the vehicle and its plates.
- Emissions inventory
- An account of how much of each pollutant each source, such as vehicles or industry, emits.
- Intertemporal substitution
- Shifting an activity to other times, here moving driving to nights and weekends when the ban didn’t apply.
- Linear probability model
- A regression with a yes-or-no outcome, such as whether a day broke a WHO guideline, so coefficients are changes in probability.
- Log points
- Changes in the log of a variable. For small values, 100 × the coefficient is approximately the percentage change.
- Newey–West standard errors
- Standard errors that remain valid when errors are correlated over time, up to a chosen lag.
- Polynomial time trend
- A smooth curve in time, here up to the ninth or higher order, that absorbs gradual changes so they aren’t mistaken for a policy effect.
- Regression discontinuity
- A design that compares outcomes just before and just after a sharp cutoff, here a date, and attributes any sudden jump to the policy.
- Revealed preference
- Inferring what people value from what they choose. Buying a second car shows that avoiding the ban was worth at least its cost.
- Stacked specification
- A regression pooling all five pollutants, letting everything except the HNC effect differ by pollutant, to estimate an average effect.
- Temperature inversion
- A layer of warm air above cooler air that traps pollution near the ground, common in Mexico City’s winter.
- Value of a statistical life
- What people are collectively willing to pay for small reductions in the risk of death, scaled to one expected life saved.
Davis, L. W. (2008). The Effect of Driving Restrictions on Air Quality in Mexico City. Journal of Political Economy, 116(1), 38–81. https://doi.org/10.1086/529398
This guide paraphrases the published paper and redraws its results from its tables. The diagram in Section III.A uses simulated data, peak times are read approximately from Figure 2, and the weekday sticker colors follow the program’s calendar, of which the paper gives the Monday example.