Saturday Driving Restrictions Fail to Improve Air Quality in Mexico City

Policymakers around the world are turning to license-plate based driving restrictions in an effort to address urban air pollution. The format differs across cities, but most programs restrict driving once or twice a week during weekdays. This paper focuses on Mexico City, home to one of the oldest and best-known driving restriction policies. For almost two decades Mexico City’s driving restrictions applied during weekdays only. This changed recently, however, when the program was expanded to include Saturdays. This paper uses hourly data from pollution monitoring stations to measure the effect of the Saturday expansion on air quality. Overall, there is little evidence that the program expansion improved air quality. Across eight major pollutants, the program expansion had virtually no discernible effect on pollution levels. These disappointing results stand in sharp contrast to estimates made before the expansion which predicted a 15%+ decrease in vehicle emissions on Saturdays. To understand why the program has been less effective than expected, the paper then turns to evidence from subway, bus, and light rail ridership, finding no evidence that the expansion was successful in getting drivers to switch to lower-emitting forms of transportation.

1 Air Quality Data

Histograms
As explained in the paper, the analysis uses hourly air pollution data from the Automated Environmental Monitoring Network (Red Automático de Monitoreo Atmosférico), also known as RAMA. Monitoring station locations were determined by Mexico City's Environmental Agency (Secretaría del Medio Ambiente) and are intended to reflect a representative sample of neighborhoods in Mexico City.
RAMA tracks hourly measures of carbon monoxide (CO), nitric oxide (N O), nitrogen dioxide (N O 2 ), nitrogen oxides (N O x ), ozone (O 3 ), large particulates (P M 10 ), small particulates (P M 2.5 ), and sulfur dioxide (SO 2 ), as well as hourly measures of temperature, humidity, and wind speed. Appendix Table 2 presents descriptive statistics and Appendix Figures 1,2,and 3 show histograms for all eight pollutants and all four meteorological measures.

Raw Time Series
Appendix Figure 4 plots average daily pollution levels during the period 2005-2011. This is a 7-year window including approximately 3.5 years before and after the introduction of Saturday driving restrictions. These figures include one observation per monitoring station for each Saturday. Mean daily pollution levels were constructed by averaging over all hours of the day. Overall, pollution levels demonstrate little trend over this period.
For most pollutants there is a pronounced seasonal pattern with higher pollution levels during the winter season due to thermal inversions. In the summer air pollution is less of a problem because the sun warms surface air causing it to rise, carrying pollutants up and out of the city. In the winter, however, the sun provides less warmth and cool surface air is trapped by warmer air above. These temperature inversions cause air quality to be lower during winter months [1,2,3].
The vertical line corresponds to July 5th 2008, the day the program was expanded to include Saturdays. There is no discernible decrease corresponding to the program expansion for any of the eight pollutants. Pollution levels are highly variable however, so it is difficult to draw strong conclusions on the basis of these raw data. The analysis in the paper refines this before-and-after comparison by introducing parametric controls for seasonality, weather, and other factors.

Hourly Pattern During Weekends
Appendix Figures 5 and 6 describe the Friday to Monday pattern for air pollution in Mexico City. These plots were constructed using all observation from 2006 and 2007, the two complete years prior to the introduction of Saturday driving restrictions. Each observation represents mean pollution averaged over all monitoring stations for a specific hour and day during this period. Several days are included in the same plot to facilitate comparison both within and across days.
Carbon monoxide and nitrogen oxides are clearly associated with driving, with peak daily pollution levels reached during the morning commute and then a second, more diffuse peak during the afternoon commute and evening. Carbon monoxide and nitrogen oxides exhibit markedly lower peaks on Saturdays and, in particular, on Sundays. This is consistent with emissions inventories that have been performed for Mexico City indicating that vehicles in Mexico City are responsible for 99% of carbon monoxide and 82% of nitrogen oxides, and 62% of small particulates but only 34% of VOCs and 23% of large particulates [4].
Ozone follows a different pattern, peaking once in the middle of the day when the sun is highest in the sky, and with similar peaks all days of the week. This suggests that ozone production in Mexico City's atmosphere is not limited by ambient nitrogen oxides. 1 That is, ozone levels are relatively insensitive to changes in emissions of nitrogen oxides, and thus are unlikely to have been meaningfully decreased by driving restrictions.
These figures show that air quality responds relatively quickly to changes in vehicle emissions. Exactly how long pollutants remain in the atmosphere depends on wind speed, thermal inversions, and other climatological factors but, overall, the rapid changes over the course of the day indicate that pollutants do not typically remain in the atmosphere for more than a couple of hours. This is important for the analysis that follows because it shows that it should be possible to make inference about changes in emissions by comparing air pollution levels within a narrow time window.

Empirical Strategy
Determining the causal impact of driving restrictions is difficult because air quality depends on numerous factors which are constantly evolving. For example, vehicle emissions are constantly changing as households and firms buy new vehicles and change driving patterns, and as the transportation network itself is upgraded and modified. Similarly, industrial emissions are constantly changing as plants open and close, and as firms adopt new production technologies and emissions control equipment. Consequently, naive before-and-after comparisons will be biased by these other factors that are changing over time.
The approach adopted in this paper is to use a regression discontinuity (RD) analysis. RD is a quasi-experimental research design that exploits the sharpness with which the Saturday driving restrictions were implemented and compares air quality immediately before and after the policy change. The main regression specification used in the analysis is equation (1) in the paper. As described in the paper, all specifications include a fifthorder polynomial in time to control for time-varying omitted variables.
An important advantage of RD is that it requires a considerably weaker identifying assumption than other approaches [5]. Nonparametric identification with RD requires only that the conditional mean function E[u t |x t ] be continuous at the threshold, t 0 . Under this assumption there may be unobserved factors that influence air quality, but their effect cannot change discontinuously at the threshold and the parameter of interest γ 1 is identified within an arbitrarily narrow window around the discontinuity, Increasing the interval around this threshold may introduce bias. However, if an assumption is made about the functional form of the relationship between air pollution and time then a wider interval can be used to identify the effect of the policy change. This is why the regressions include the flexible polynomial in time. The RD estimate of γ 1 is unbiased as long as this polynomial accurately captures the time-varying omitted variables.
A limitation of RD is that it only identifies the effect locally at the point in time when the policy change occurred. This is a significant weakness of RD because for policy evaluation one would like to know both the short-and long-run effects of policy changes. Still, the short-run impact is quite informative, particularly in this case. Although the relative size of short-and long-run impacts is theoretically ambiguous [6], in practice the tendency has been for restrictions to increase the vehicle stock, leading to worse air quality in the long-run [7,8]. If this is indeed the pattern of longer-run adaptations, then finding zero or small short-run impacts is very disappointing news from a policy perspective.
In all specifications the variance matrix is estimated taking into account that u t may be serially correlated. Standard diagnostic tests were used to assess the magnitude of serial correlation. In the preferred specification, the autocorrelation coefficients are statistically significant only for three or four days. Accordingly, variance matrices are estimated allowing for arbitrary correlation within week-of-sample, defined Thursday through Wednesday. Newey-West standard errors with a seven day lag are also reported as an alternative specification in Appendix Table 3.

Alternative Specifications
Appendix Figure 7 is identical to Figure 1 in the paper, but uses daily maximum pollution levels rather than daily means. As discussed in the paper, the health impacts of air pollution are highly nonlinear, so these results are of large intrinsic interest. Across pollutants, there is no visually discernible improvement in air quality when the program is expanded. Consistent with the regression estimates in Table 2 in the paper, there is no evidence of a downward shift in pollution levels coincident with the beginning of Saturday driving restrictions.
RD estimates can be sensitive to minor changes so it is important to evaluate alternative specifications. Appendix Table 4 reports estimates using alternative polynomials, ranging from third-order to seventh-order. Similarly, Appendix Table 5 reports estimates using shorter windows and less flexible polynomials [9]. As is clarified in the table notes, these specifications are otherwise identical to our baseline estimates.
Overall, the results are very similar with these alternative estimating equations. In many specifications the estimates for carbon monoxide are negative and statistically significant at the 5% level, but estimates for other pollutants are positive or very close to zero across specifications. Moreover, the estimates in the stacked specification are always very close to zero.
Finally, Appendix Table 6 tests for intertemporal substitution toward Fridays and Sundays. There is no evidence that the Saturday driving restrictions increased driving on these other days of the week. The estimate in the stacked specification for Sundays is slightly positive (0.6%) but not statistically significant.

Impacts by Hour
There are several reasons why the impacts of driving restrictions might differ by hour of the day. First, driving restrictions ban driving only between 5am and 10pm, so one might expect driving to increase outside these hours if drivers are able to intertemporally substitute. Second, the composition of drivers and trips may vary across hours, with Saturday afternoon and, in particular, Saturday evening driving including much more discretionary driving than other hours.
Appendix Figure 8 plots RD estimates and 95th percentile confidence intervals for the effect of Saturday driving restriction on air quality by hour-of-the day. Separate regressions are estimated for each pollutant and hour-of-the-day, all with fifth-order polynomials in time, weather variables, and indicator variables for week-of-year and day-of-week. That is, these results are identical to the main regression estimates but estimated hour-byhour.
Of the 192 total estimates, 107 (56%) are negative, thus only slightly more negative estimates than one would expect due to chance alone. There are negative and sta-tistically significant estimates for carbon monoxide, nitric oxide, and nitrogen oxides between about 2pm and 10pm, but other pollutants do not have this pattern. Moreover, none of the eight pollutants decrease on Saturday mornings. Thus, overall, the hourby-hour evidence points to very limited air quality improvements from the Saturday expansion.

Additional Evidence on Behavioral Responses
These supplementary materials also include plots of raw daily data for the different forms of public transportation and additional measures of economic activity such as Museum attendance. These plots are useful because they show how Saturday activity compares, in general, to the rest of the week, and because they show the overall level and trend in activity over this period. In addition to the measures reported in the paper, these materials include evidence from the Mexico City subway. As can be seen in Appendix Figure 19, there is an increase in subway ridership during all days of the week around July 2008. Because this increase is observed all seven days of the week, this is very unlikely to be related to the Saturday expansion and there is no differential increase in ridership on Saturdays relative to the other days of the week.
[8] Gallego F, Montero J, Salas C.      Appendix Appendix Note: These results are identical to the baseline estimates except that the coefficients and standard errors come from estimating a Newey-West specification with a 7-day lag.
Appendix Note: These results are identical to the baseline estimates except for the order of the polynomial time trend and sample period as indicated in the panel headings. In addition, the first two panels include week-of-year indicator variable but the last three panels do not. None of the estimates are statistically significant at the 5% level. Note: This table reports coefficient estimates and standard errors from 9 separate regressions, one per column. The specifications are identical to the baseline specification in the paper except instead of an indicator for Saturdays after July 5th, 2008, two separate indicators are included for Fridays and Sundays after July 5th, 2008. None of the estimates are statistically significant at the 5% level.