Spatial and time spillovers of driving restrictions: Causal evidence from Lima's Pico y Placa policy

Salgado, Edgar; Mitnik, Oscar A. Working Paper Spatial and time spillovers of driving restrictions: Causal evidence from Lima's Pico y Placa policy IDB Working Paper Series, No. IDB-WP-01278 Provided in Cooperation with: Inter-American Development Bank (IDB), Washington, DC Suggested Citation: Salgado, Edgar; Mitnik, Oscar A. (2021) : Spatial and time spillovers of driving restrictions: Causal evidence from Lima's Pico y Placa policy, IDB Working Paper Series, No. IDBWP-01278, Inter-American Development Bank (IDB), Washington, DC, https://doi.org/10.18235/0003849 This Version is available at: https://hdl.handle.net/10419/290086 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc-nd/3.0/igo/legalcode Spatial and Time Spillovers of Driving Restrictions: Causal Evidence From Lima’s Pico y Placa Policy Edgar Salgado Oscar A. Mitnik IDB WORKING PAPER SERIES No IDB-WP01278 Inter-American Development Bank Office of Strategic Planning and Development Effectiveness Institutions for Development Sector December 2021 Spatial and Time Spillovers of Driving Restrictions: Causal Evidence From Lima’s Pico y Placa Policy Edgar Salgado Oscar A. Mitnik Inter-American Development Bank Office of Strategic Planning and Development Effectiveness Institutions for Development Sector December 2021 Cataloging-in-Publication data provided by the Inter-American Development Bank Felipe Herrera Library Salgado, Edgar. Spatial and time spillovers of driving restrictions: causal evidence from Lima's Pico y Placa Policy / Edgar Salgado, Oscar A. Mitnik. p. cm. — (IDB Working Paper Series ; 1278) Includes bibliographic references. 1. Traffic regulations-Peru. 2. Traffic congestion-Peru. 3. Urban transportation policyPeru. I. Mitnik, Oscar Alberto. II. Inter-American Development Bank. Office of Strategic Planning and Development Effectiveness. III. Inter-American Development Bank. Institutions for Development Sector. IV. Title. V. Series. IDB-WP-1278 http://www.iadb.org Copyright © 2021 Inter-American Development Bank. This work is licensed under a Creative Commons IGO 3.0 AttributionNonCommercial-NoDerivatives (CC-IGO BY-NC-ND 3.0 IGO) license (http://creativecommons.org/licenses/by-ncnd/3.0/igo/legalcode) and may be reproduced with attribution to the IDB and for any non-commercial purpose, as provided below. No derivative work is allowed. Any dispute related to the use of the works of the IDB that cannot be settled amicably shall be submitted to arbitration pursuant to the UNCITRAL rules. The use of the IDB’s name for any purpose other than for attribution, and the use of IDB’s logo shall be subject to a separate written license agreement between the IDB and the user and is not authorized as part of this CC-IGO license. Following a peer review process, and with previous written consent by the Inter-American Development Bank (IDB), a revised version of this work may also be reproduced in any academic journal, including those indexed by the American Economic Association’s EconLit, provided that the IDB is credited and that the author(s) receive no income from the publication. Therefore, the restriction to receive income from such publication shall only extend to the publication’s author(s). With regard to such restriction, in case of any inconsistency between the Creative Commons IGO 3.0 Attribution-NonCommercial-NoDerivatives license and these statements, the latter shall prevail. Note that link provided above includes additional terms and conditions of the license. The opinions expressed in this publication are those of the authors and do not necessarily reflect the views of the Inter-American Development Bank, its Board of Directors, or the countries they represent. Spatial and Time Spillovers of Driving Restrictions: Causal Evidence from Lima’s Pico y Placa Policy* Edgar Salgado†Oscar A. Mitnik‡ December, 2021 Abstract Driving restrictions are popular interventions in rapidly urbanizing developing countries. Their relatively inexpensive implementation appeals to the pressing need to reduce traffic congestion and pollution. Their effectiveness however, remains contested. Using high frequency data from the community-based driving directions app Waze, we evaluate the causal effect on traffic congestion of Lima’s Pico y Placa driving restriction policy introduced in 2019. We find small improvements in traffic congestion for the policy’s directly targeted areas. However, those improvements are offset by time and spatial spillovers in the opposite direction in the aggregate. Speed improved by 2 percent during the early weeks of the intervention, but this effect disappeared 16 weeks after the start of the policy. Moreover, traffic conditions worsened in adjacent areas and in hours outside the time schedule of the policy. In the aggregate, accounting for time and spatial spillovers, a simulation exercise suggests that overall welfare declined by 2 percent, mostly driven by the extensive margin (more roads becoming congested) outside the direct areas and hours targeted by the policy. The policy seems not only to have failed to achieve its intended benefits in terms of congestion, but also probably caused increases in traffic-related pollution. These results highlight the need for policy makers to take into account the overall impacts of driving restrictions policies before implementing them. Keywords: driving restrictions, congestion spillovers, welfare impacts JEL codes: H41; Q58; R41; R48 *We would like to thank Waze for providing access to its data through the Waze for Cities Program ( https:// www.waze.com/ccp/ ). We are also grateful for helpful comments from Allen Blackman, Francisco Gallego, Alessandro Maffioli, Weihua Zhao, an anonymous Inter-American Development Bank (IDB) reviewer, and seminar participants at the IDB, the annual meeting of the Impact Evaluation Network, and the European Meeting of the Urban Economics Association. We also thank João Carabetta for producing the code to process Waze traffic jam datasets and convert them to segment-level datasets. The opinions expressed in this paper are those of the authors and do not necessarily reflect the views of the Inter-American Development Bank, its Board of Directors, or the countries they represent. †Inter-American Development Bank, [email protected]. ‡Inter-American Development Bank and IZA, [email protected]. 1 Introduction With the acceleration of urbanization in developing countries, traffic congestion and its effect on pollution and economic activity remains a major concern. The United Nations (2018) estimates that by 2018, 55 percent of the world population lived in cities, and it forecasts that in the coming decades 90 percent of urban expansion will take place in developing countries. Latin America, in particular, finds itself vulnerable to the detrimental spillovers of unplanned urbanization (YañezPagans et al.,2019). In this context, the city of Lima, Peru, provides a stark reminder of the consequences of uncontrolled urbanization for the quality of life of its inhabitants. Lima is ranked as the seventh most congested city in the world (TomTom,2020). According to a 2018 survey by Lima Cómo Vamos (2019), the inhabitants of Lima consider problems with public transport as the second most pressing issue (46 percent) in the city, just after crime (82 percent). In addition, survey respondents cite pollution as the fifth most pressing problem (28 percent), and when asked about the causes of pollution, 72 percent of the respondents said vehicle pollution was the main factor. Traffic congestion in Lima is typical of the congestion problem across major Latin American cities.1The policy responses have also been similar, with the imposition of driving restrictions being one of the preferred responses. Several cities have imposed such restrictions to combat traffic congestion and its associated ailments in the last 30 years (Blackman et al.,2018b). Despite their popularity, however, the effectiveness of driving restrictions remains controversial. The literature has found mixed evidence on the impact of such restrictions, and in some cases there are signs of some perverse effects on pollution after drivers adjust to the specifics of the policy. While the existing literature has concentrated on the impacts of driving restriction on pollution, to our knowledge no prior studies have examined the impacts of these policies on traffic congestion. Therefore, this paper focuses on the traffic congestion margin — which can arguably be considered a sine qua non condition for observing any impacts on the pollution margin — and evaluate the causal effect on traffic congestion of Lima’s Pico y Placa driving restriction policy introduced in 2019. Thanks to the detailed nature of our data, as will be explained below, we are able to examine novel within-city and across-hours impacts. The earliest study in the driving restrictions literature is Eskeland and Feyzioglu (1997), who evaluated Mexico City’s Hoy No Circula, a program introduced in 1989 that restricted vehicle circulation based on the last digit of license plates. The study found that, as a result of the policy, households bought an additional car to get additional driving privileges, which led to an increase in the number of cars in the city. In a subsequent evaluation of the same program, Davis (2008) showed that it did not improve air quality, while leading to more vehicles in circulation and more purchases of high-emission vehicles. De Grange and Troncoso (2011) evaluated the imposition of a restriction on vehicles without catalytic converters – a device to reduce pollutant gases – and found no effects on the use of private cars, except when the restriction was temporarily extended to all vehicles for certain hours of the day. When all vehicles faced a temporary restriction based on the last digit of the license plate, car usage dropped by 5.5 percent while ridership in the Metro system rose by 3 percent. Troncoso et al. (2012) evaluated the effect of the same temporary vehicle restriction on pollution and found reductions in carbon monoxide (CO) of between 7.6 and 9.4 percent. Gallego et al. (2013a) modeled households’ transport use decisions allowing for public and private modes. Studying the driving restrictions in Mexico City, the authors estimated that this 1According to TomTom (2020), five cities in Latin America are among the 20 most traffic congested cities in the world: Bogota (5th), Lima(7th), Mexico City (13th), Recife (15th), and Rio de Janeiro (20th). 1 type of policy had the unintended impact of increasing the number of circulating cars. In a similar study Gallego et al. (2013b) confirmed these results and found additional detrimental effects in terms of increased pollution. Other authors also find negative impacts. Ye (2017) found that driving restrictions in Lanzhou were ineffective in improving its air quality because drivers shifted their travel schedules, took detours, and acquired more cars. Bonilla (2019) found that Bogota’s Pico y Placa generated a light increase in CO during the morning peak hours and higher vehicle ownership and gasoline consumption. A small number of studies find positive impacts of driving restrictions policies, though some times paired with negative impacts on other outcomes. Viard and Fu (2015) estimated a 21 percent drop in air pollution during one-day-per-week restrictions in Beijing, together with a reduction in labor supply among workers with discretionary work time. Carrillo et al. (2016,2018) found evidence of reductions in pollution levels resulting from Quito’s driving restriction policy, but at the cost of higher crime rates following its implementation. In a study of a 1992 program in Santiago (Chile) that targeted old cars in an attempt to rid the city from high polluting vehicles, Barahona et al. (2020) found compelling evidence that vintage-specific driving restrictions incited fleet renewal towards cleaner cars. Other studies suggest that incorporating all costs related to the restrictions, may diminish any benefits achieved by the policy. This become particularly relevant for policies with meager results . Blackman et al. (2018a,b) suggest that despite its growing popularity worldwide, licenseplate–based driving restrictions do not always make good economic sense based on the availability of public transportation modes or a market for used cars. While prior studies have been able to capture city-wide impacts, they could not identify impacts on smaller geographical areas within cities. In this paper, owing to highly refined data, we can explore in detail how the impact of driving restrictions propagates through the city. Thus, a major contribution of our analysis is the quantification of spatial and time spillovers of the policy. By using high-frequency and geocoded data on traffic jams we are able to disentangle these two types of spillovers associated with driving restrictions, which most of the prior literature speculated upon but was not able to estimate due to lack of adequate data. The analysis relies on high-frequency data on traffic jams from the community-based driving directions app Waze. Through the use of a generalized propensity score, we select the streets and road segments that are comparable across different rings around the Pico y Placa intervention areas. Then, we estimate inverse probability weighting flexible difference-in-differences regressions, weighted by the inverse of the generalized propensity score, to obtain the causal impacts of the Pico y Placa policy. Methodologically, our paper is related to recent studies that use high-frequency data. Hanna et al. (2017) used traffic speed data from Android phones collected through Google Maps to investigate whether high-occupancy vehicle policies reduce traffic congestion. Kreindler (2016) used information collected through an app with precise GPS coordinates for over 100,000 commuter trips in Bangalore, India to examine the welfare effects of congestion pricing. We find initial improvements in traffic congestion after the imposition of the Pico y Placa policy in Lima in the directly affected areas, with mostly non-significant negative spillovers on other areas or times, both during a transition period and in the first six weeks of full implementation of the policy. By week 16 of full policy implementation small positive impacts remain only for some areas directly affected by the policy and for certain times of day. However, these positive impacts are offset by time and spatial spillovers in the opposite direction in the aggregate. The number of minutes of a severe traffic jam on high-capacity roads increased by 117 percent in the morning in the area immediately adjacent to the intervened area. The probability of a severe traffic jam also increased by 5pp in nearby areas in the hours in between the morning and afternoon, which represents 71 percent of the pre-treatment average on high-capacity roads. Aggregate data on fines imposed on 2 drivers suggest a sustained enforcement effort of driving restrictions throughout the analysis period, implying that observed changes are likely due to behavioral changes by drivers. A simulation exercise that accounts for all the effects suggests an overall welfare loss of about 2 percent in the final period of implementation of the policy. The welfare analysis, furthermore, suggests that the direct congestion benefits of the Pico y Placa policy during 2019 were at best small and localized, and that there were negative welfare impacts associated with time and spatial spillovers. A welfare decomposition suggests that most of those negative impacts were caused by the extensive margin (i.e. more roads becoming congested). The results highlight the need for policy makers to take into account the overall impacts of driving restrictions policies before implementing them. The next section of this paper presents the context that motivated the Pico y Placa policy. Section 3presents the data, while Section 4describes the empirical methodology, including the study design, the selection of comparable streets, and the difference-in-differences approach used for estimation. Section 5presents the results of the analysis, and Section 6puts forth a welfare analysis to understand the overall impacts of the policy. Section 7concludes. 2 Lima, Traffic Congestion, and Pico y Placa The capital of Peru, Lima is a coastal city located by the shores of the Pacific Ocean. It is home to 10 million people, a third of the country’s population. The city grew rapidly from a small cosmopolitan area in the early years of the 20th century to a robust economic center that has attracted large waves of migrants from the rest of the country since the 1950s. In 1991 the national government decreed that any resident with a motor vehicle could operate in the city as a provider of public transport services.2The rationale was to provide income-earning opportunities for the urban population, but the cost of the measure was a severe deterioration of the provision of public transit (Jauregui-Fung et al.,2019). The 21st century brought efforts to reorganize the public transit problem with the construction of an elevated light rail system (Linea 1) and a Bus Rapid Transit system (BRT), known as the Metropolitano, which were completed between 2010 and 2011. However, traffic congestion in Lima remains a severe problem. Calatayud et al. (2021) estimate that the total cost of congestion in Lima in 2019 accounted for 0.7 percent of the city’s per capita GDP. Calculations by TomTom (2020) for the same year suggest that the inhabitants of Lima lose 8.7 days a year stuck in traffic, below Bogotá where people lose 9.6 days a year, but above Santiago where people lose about 7.6 days a year, or Buenos Aires where the loss is about 5.5 days a year. Time lost in large metropolises from richer countries fall below Latin American standards: 6.2 days in London, 5.9 in New York, and 4.8 in Madrid. In an influential annual opinion survey in 2018 only 37.5 percent of the respondents in Lima expressed satisfaction with their city (Lima Cómo Vamos,2019). And, as previously mentioned, while 82.2 percent of respondents listed crime as the city’s main problem, the second-most frequently cited problem was public transport (46.2 percent), with pollution cited by 28.5 percent of respondents. In 2019, the Metropolitan Authority in Lima introduced a plan to restrict circulation in certain areas of the city in an effort to reduce air pollution and relieve traffic congestion, and as a way to ease the expected additional congestion associated with the 2019 Pan American and Parapan American Games to be held in the city in late July and August. The measure was piloted during the games, and based on initial results, the authorities decided to keep it after the conclusion of the Games.3 2Legislative Decree 651 of 1991. 3The Metropolitan Authority reported that in the week of 5 to 8 August 2019 speed increased by 11 percent in the 3 Figure 1shows the area with the intervened roads in red, as well as six adjacent rings. The policy restricted circulation four days a week, from Monday to Thursday. Vehicles with an oddnumbered last digit on their license plate were restricted from circulation on Mondays and Wednesdays, while vehicles with an even-numbered last digit were restricted on Tuesdays and Thursdays. The municipality implemented two time schedules for the restriction. For the morning rush hour, circulation was restricted initially from 7:30 to 10 a.m., while for the afternoon rush hour circulation was restricted from 5 to 9 p.m. Later, the morning schedule was extended to start from 6:30 a.m. In the empirical estimations we consider morning schedules from 6 to 10 a.m. and afternoon schedules from 5 to 9 p.m.4 The restricted area included major thoroughfares within the red core depicted in Figure 1.5The figure shows a buffer of 250 meters around the restricted streets, roads, or highways in red. This red area is deemed “the direct intervention area,” or Pico y Placa area. The figure also shows six adjacent rings of 250 to 500 meters each that define the areas of analysis in the empirical section, as explained below. Figure 2shows the evolution of Lima’s traffic conditions throughout the day from January to June 2019, prior to the implementation of the restrictions. The figure shows three variables: speed, probability of a severe traffic jam and the number of minutes in a severe traffic jam. It also distinguishes by road type: local (panel a) and non-local (panel b) roads. Speed is measured in kilometers per hour, while the definition of a “severe” traffic jam is constructed based on the Waze classification. Waze considers a hierarchy of four types of traffic jams with categories from 1 to 4, where 4 is the most severe type. Our classification of “severe” corresponds to types 3 and 4; that is, when speed is less than 40 percent of a reference speed calculated by Waze as the non-trafficjam speed.6As explained below, the unit of observation in this study is the road segment. We calculate speed for each segment, in every hour in the data, whether it is part of a severe traffic jam within that hour, and the number of minutes within that hour that the road segment has been reported as having a severe traffic jam. Waze does not provide the reference speed used for traffic jam classification types. However, as explained below, we can use information provided by Waze to calculate a time-invariant freeflow speed (i.e. under non-traffic-jam conditions) for each road segment in the sample. While Section 3provides more details, here we simply clarify that we define local roads as those road segments whose free-flow speed is less than 30 km per hour, and non-local roads are defined as road segments with free-flow speed of 30 km per hour or higher. With this definition we aim to proxy for road capacity: local roads are mainly small streets, while non-local roads could include highways or freeways. Based on this classification, we explore traffic conditions in Lima in the pre-Pico y Placa period in Figure 2. Beyond the distinction between local and non-local roads, Figure 2distinguishes by each distance ring depicted in Figure 1. Regardless of the road type, the Pico y Placa area under-performs all the other six adjacent areas, with the exception of speed on local roads (top left panel). Every public transport blue and red corridors. It also reported increases of 19 percent in private transport speed in one of the main corridors, Javier Prado. See Municipalidad de Lima (undated). 4In both cases the ending hour of the restriction implies no restriction from that time; that is, the restriction from 6 to 10 a.m. indicates that the restriction is in place until 9:59, and as of 10:00 there is no restriction. Similar logic applies for the afternoon. 5These included Panamericana Sur and Panamericana Norte, from north to south in the east of the city; Javier Prado and Avenida La Marina from west to east; and two parallel key thoroughfares, Via Expresa and Arequipa corridors. 6To be classified as type 4, the reported speed must be less than 20 percent of the reference speed; for type 3, the reported speed must be less than 40 percent but higher than 20 percent of the reference speed; for type 2, the reported speed must be less than 60 percent but higher than 40 percent of the reference speed; and for type 1, the reported speed must be less than 80 percent but higher than 60 percent of the reference speed. 4 Table 2. Pre-treatment Summary Statistics of Dependent Variables by Schedule and Distance Ring Distance to Speed Pr(Severe Traffic Jam) Minutes in Severe Traffic Jam Pico y Placa Total Local Non-Local Total Local Non-Local Total Local Non-Local (meters): Roads Roads Roads Roads Roads Roads (1) (2) (3) (4) (5) (6) (7) (8) (9) Panel a. Morning 0-250 23.67 18.32 42.76 0.04 0.03 0.06 0.65 0.45 1.40 250-500 19.77 17.75 35.78 0.03 0.03 0.05 0.44 0.38 0.93 500-1000 20.25 18.00 35.09 0.03 0.03 0.06 0.47 0.37 1.13 1000-1500 20.57 17.66 36.97 0.02 0.02 0.05 0.36 0.26 0.93 1500-2000 20.99 17.61 38.18 0.02 0.02 0.05 0.33 0.22 0.91 2000-2500 20.23 17.21 36.85 0.02 0.01 0.04 0.23 0.17 0.56 2500-3000 21.58 18.18 36.64 0.01 0.01 0.02 0.14 0.10 0.31 Panel b. Midday 0-250 22.93 17.43 40.92 0.07 0.09 0.10 1.19 1.56 2.29 250-500 19.32 17.06 34.68 0.06 0.07 0.09 0.94 1.28 1.78 500-1000 19.62 17.22 33.42 0.07 0.08 0.12 1.16 1.42 2.37 1000-1500 19.98 17.05 35.27 0.06 0.06 0.11 0.93 0.99 2.25 1500-2000 20.52 17.13 36.76 0.05 0.05 0.10 0.73 0.75 1.97 2000-2500 19.83 16.87 35.50 0.03 0.03 0.08 0.50 0.49 1.23 2500-3000 21.39 17.89 36.14 0.02 0.03 0.03 0.20 0.27 0.47 Panel c. Afternoon 0-250 22.02 17.43 38.39 0.11 0.09 0.17 2.20 1.56 4.50 250-500 18.87 17.06 33.20 0.08 0.07 0.15 1.54 1.28 3.59 500-1000 19.16 17.22 31.93 0.09 0.08 0.18 1.79 1.42 4.25 1000-1500 19.55 17.05 33.63 0.07 0.06 0.16 1.43 0.99 3.91 1500-2000 20.17 17.13 35.61 0.06 0.05 0.14 1.10 0.75 2.89 2000-2500 19.53 16.87 34.15 0.05 0.03 0.12 0.78 0.49 2.38 2500-3000 21.16 17.89 35.64 0.03 0.03 0.05 0.37 0.27 0.81 Source: Prepared by the authors. Note: Averages calculated using pre-treatment data on segments that satisfy overlap, from 7 January 2019 to 30 June 2019. Morning: 6:00 to 9:59 a.m., Midday: 10:00 to 4:59 p.m., Afternoon: 5:00 to 8:59 p.m. 4 Estimation As the Pico y Placa policy was implemented to ease traffic congestion before and during the Pan American and Parapan American Games that were due to start 26 July and 23 August, respectively, the “full” effect of Pico y Placa is considered as starting as of the week of 2 September, right after the closure of the Parapan American Games. The difference-in-differences empirical strategy is based on using the road segments in the outermost ring (2500 to 3000 meters away from the restricted streets) as a comparison group while we the effect of the policy on the other rings around the intervened zone is evaluated. The estimating equation is: yith =µi+ 6 ∑ k=1 βkTki ×PyPt+ 6 ∑ k=1 θkTki ×PanGt +τt+γh+f(t)z+εith (1) where yith is the outcome variable for road segment ion day tat hour h. The three outcomes are (i) Ln(Speed), (ii) a dummy variable that activates when the road segment has a severe traffic 11 jam during the hour, and (iii) the number of minutes within that hour that the road segment has a severe traffic jam. The coefficients of interest are as follows: each βkis associated with a distance ring, Tki;µiis a road segment fixed effect; PyPtis a dummy variable for the period after the Parapan American Games; θkcaptures the effect of the Pan American and Parapan American Games that coincided with the first weeks of Pico y Placa (July 22 to September 1); τtis a day fixed effect to account for aggregate shocks in the city; γhis an hour of the day fixed effect; f(t)zis a set of linear trends by district zbefore and after the first day of the driving restrictions (July 22); and εith is the error term. Standard errors are clustered by road segment. Clustering at the road segment level allows for arbitrary correlation across time within segments. To allow for both time and spatial correlation between segments we explored an alternative clustering strategy, based on allowing correlation between all segments within an enclosed area.13 As discussed in Section 5, this alternative strategy does not change the statistical significance of most coefficients. However, it is quite costly in terms of computational time. Thus, we use the road segment level clustering for our main results. A difference-in-differences strategy does not require that units in different treatment arms are equal in levels prior to treatment, only that they follow parallel trends pre-treatment. However, ensuring that units are similar in levels in the pre-treatment periods, and re-weighting units to ensure that the parallel trends assumption holds, lends robustness to the identifying assumptions (Ryan et al.,2019;Callaway and Sant’ Anna,2021). Thus, to ensure road segment comparability across different distance rings, we estimate a generalized propensity score (GPS), the probability of a road segment being in any of the rings (Imbens,2000;Hirano and Imbens,2004).14 We follow the strategy proposed by Flores and Mitnik (2013) to identify road segments satisfying simultaneous overlap across distance rings.15 Simultaneous overlap is attained by defining the overlap region for each distance ring: based on the probability of belonging to a particular ring T=Tk, only those road segments with probability above a certain quantile qthreshold are deemed to be part of the overlap region for that distance ring. Simultaneous overlap deems as satisfying the overlap condition all the road segments that are part of the overlap region simultaneously for all distance rings. Intuitively, it implies that road segments are comparable in terms of covariates in each of the distance rings. We estimate the GPS with a multinomial logit model using as covariates historical pre-intervention data and road segment characteristics. In particular, we model such probability based on average speed and minutes in severe traffic jam in morning and afternoon peak hours calculated for each of the 28 weeks preceding the intervention. We also use time invariant road segment characteristics such as free-flow speed (in logs), length, and number of appearances in the dataset before the start of the restrictions. Additionally, we include other time variant pre-intervention variables at the road segment level such as the average of the share of the segment length over the length of the traffic jam to which the segment belongs, in the morning and afternoon time periods. We also include variables from the 2017 population census calculated at the traffic-zone level:16 proportion 13We define those areas by using resolution 10 hexagonal H3 cells (with a side length of around 66 meters and an area of approximately 11,300 square meters). H3 cells are a hexagonal hierarchical geospatial grid system originally developed by Uber to analyze sub-areas of the world at different grid sizes (“resolutions”). For more details on H3 cells, see https://h3geo.org/ . 14The GPS is the probability of a particular treatment group conditional on covariates, Pr(T=Tk|X=x), where k=1, ..., 7 (the seven rings) and Xrefers to pre-treatment covariates. 15Following Flores and Mitnik (2013) simultaneous overlap is defined as 0<ξ<Pr(T=Tk|X=x)for all Tkand xeX. 16We include these variables to attempt to control for any zone-level characteristics that may affect congestion patterns in each road segment. We rely on the 427 traffic zones defined in the 2004 and 2011 Origin-Destination surveys for the Lima metropolitan area (JICA,2013). These traffic zones are constructed to capture homogeneous transport 12 of males, average age, proportion of individuals with primary, secondary, and tertiary education, proportion of permanent residents in the district, proportion of residents with more than five years in the district, share of Spanish speakers, share of individuals insured in the health system, proportion of people who study in a school outside their district of residence, and the proportion of working individuals whose job requires them to commute to another district. After estimating the GPS, we impose simultaneous overlap across all seven rings, following Flores and Mitnik (2013), as explained above, and using the value of q=1(i.e. percentile 1).17 As mentioned in Section 3, we drop close to 18 percent of the road segments that do not satisfy the overlap condition. In all regressions we only use data associated with the road segments that satisfy this condition. Furthermore, to ensure good balancing between road segments across different distance rings, all regressions are estimated using inverse probability weighting (IPW) by the inverse of the GPS. In Appendix Table A1 we show how the imposition of simultaneous overlap plus weighting by the inverse of the GPS substantially improves balancing in time-invariant and pre-intervention time-variant covariates, across distance rings.18 5 Results In this section we discuss different estimation results. First, we present an analysis to validate our empirical strategy. Second, we present the impacts of Pico y Placa for the intervention period (following the transition period during the 2019 Pan American and Parapan American Games), as well for subperiods of it. Third we present detailed impacts by hour of the day. 5.1 Validity of the Empirical Strategy The validity of the difference-in-differences methodology depends on the “parallel trends” assumption (Angrist and Pischke,2009) between the treatment and comparison groups before the start of the policy on July 22. The assumption is that unobservable characteristics associated with selection into treatment (defined by the proximity to the intervened area) remain constant and there are no observed differences in trends before the onset of the policy. To examine this, we estimate a variation of equation (1) that considers only the 28 weeks before the policy started, and divides it into two groups: the “pre-period” of 16 weeks prior to the implementation of Pico y Placa and the remaining 12 weeks as the excluded (base) period. The equation we estimate is: yith =µi+ 6 ∑ k=1 αkTki ×Pret+τt+γh+f(t)z+εith (2) where Pretis a dummy variable for the “pre-period” (the “placebo” treatment period). The validity of the strategy requires that the coefficients associated with αare not statistically significant. Table 3shows the results of this test for all road segments in panel a, local road segments in panel b, and non-local segment roads in panel c. We group the results into three schedules: morning and characteristics of the population within each zone. 17The results are robust to alternative values of qand are available upon request. 18Appendix Table A1 shows the raw means prior to imposing overlap and the IPW weighted means after imposing overlap, using the inverse of the GPS as the weight. In addition, for both types of means, it shows the p-value associated with the test of the joint hypothesis that the mean values for all distance rings are equal and the root mean square distance (RMSD) associated with each set of means. The RMSD is a normalized overall measure of distance among the estimated means (see Flores and Mitnik (2013) for details). Better balancing is captured by the RMSD values decreasing. 13 afternoon coinciding with Pico y Placa restriction hours, and the hours in between grouped under “midday”. All regressions consider only road segments satisfying the overlap condition, and use the inverse of the GPS as weight. Reassuringly, no coefficient is statistically significant, which validates the identification strategy. Appendix Figures B1 to B3 take a more flexible approach to test for parallel trends and report the results of a lead-and-lags exercise with biweekly coefficients, which confirms that trends in all variables before the start of the restrictions were parallel.19 19The estimated equation is yith =µi+∑6 k=1αkTki ×BWt+τt+γh+f(t)z+εith,, where BWtis the set of biweekly dummies reported in the figures. 14 Table 3. Pre-treatment (Weeks -16 to -1) “Placebo” Impacts Distance to Ln(Speed) Pr(Severe Traffic Jam) Minutes in Severe Traffic Jam Pico y Placa Morning Midday Afternoon Morning Midday Afternoon Morning Midday Afternoon (meters): (1) (2) (3) (4) (5) (6) (7) (8) (9) Panel a. Overall 0-250 -0.01 0.01 0.00 0.00 -0.01 -0.00 0.15 -0.19 -0.01 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.20) (0.21) (0.20) 250-500 -0.00 0.01 0.00 0.00 -0.01 -0.00 0.11 -0.25 -0.05 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.20) (0.21) (0.20) 500-1000 -0.00 0.01 0.00 0.00 -0.01 -0.00 0.11 -0.14 0.00 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.20) (0.21) (0.19) 1000-1500 -0.00 0.00 0.00 0.00 -0.00 -0.00 0.13 -0.08 0.06 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.20) (0.21) (0.20) 1500-2000 -0.00 0.01 0.00 -0.00 -0.01 -0.00 0.09 -0.16 -0.02 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.20) (0.21) (0.20) 2000-2500 0.00 0.01 0.00 -0.00 -0.01 -0.00 0.05 -0.20 -0.04 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.20) (0.21) (0.20) Panel b. Local Roads 0-250 -0.00 0.01 0.01 0.00 -0.01 -0.00 0.07 -0.15 0.02 (0.02) (0.01) (0.02) (0.01) (0.01) (0.01) (0.23) (0.26) (0.24) 250-500 0.00 0.01 0.01 -0.00 -0.01 -0.01 0.02 -0.20 -0.02 (0.02) (0.01) (0.02) (0.01) (0.01) (0.01) (0.23) (0.26) (0.23) 500-1000 0.00 0.01 0.01 -0.00 -0.01 -0.00 0.02 -0.16 -0.02 (0.02) (0.01) (0.02) (0.01) (0.01) (0.01) (0.23) (0.26) (0.23) 1000-1500 0.00 0.01 0.01 -0.00 -0.01 -0.00 0.07 -0.11 0.01 (0.02) (0.01) (0.02) (0.01) (0.01) (0.01) (0.23) (0.26) (0.23) 1500-2000 0.00 0.01 0.01 -0.00 -0.01 -0.00 0.02 -0.12 0.04 (0.02) (0.01) (0.02) (0.01) (0.01) (0.01) (0.23) (0.26) (0.23) 2000-2500 0.01 0.01 0.01 -0.01 -0.01 -0.01 -0.06 -0.20 -0.12 (0.02) (0.01) (0.02) (0.01) (0.01) (0.01) (0.24) (0.26) (0.24) Panel c. Non-Local Roads 0-250 -0.02 0.00 -0.00 -0.00 -0.01 0.00 0.38 -0.30 -0.12 (0.01) (0.02) (0.02) (0.01) (0.01) (0.01) (0.33) (0.36) (0.36) 250-500 -0.02 0.01 -0.00 0.01 -0.01 -0.00 0.38 -0.41 -0.24 (0.01) (0.02) (0.02) (0.01) (0.01) (0.01) (0.33) (0.35) (0.36) 500-1000 -0.02* -0.01 -0.02 0.01 0.00 0.01 0.45 0.01 0.13 (0.01) (0.02) (0.02) (0.01) (0.01) (0.01) (0.33) (0.36) (0.36) 1000-1500 -0.02 -0.02 -0.02 0.00 0.01 0.01 0.30 0.15 0.36 (0.01) (0.02) (0.02) (0.01) (0.01) (0.01) (0.33) (0.36) (0.38) 1500-2000 -0.02* -0.01 -0.02 0.01 -0.01 0.00 0.40 -0.25 -0.08 (0.01) (0.02) (0.02) (0.01) (0.01) (0.01) (0.33) (0.37) (0.37) 2000-2500 -0.02 -0.01 -0.03 0.01 0.01 0.02 0.42 -0.07 0.42 (0.01) (0.02) (0.02) (0.01) (0.01) (0.01) (0.34) (0.36) (0.40) Source: Prepared by the authors. Note: Standard errors clustered at the road segment level. *, ** and *** denote statistical significance at the 10, 5 and 1 percent level, respectively. This table shows the coefficients associated with the Pre dummy in equation (2). The excluded category covers weeks -28 to -17. Morning: 6:00 to 9:59 a.m., Midday: 10:00 to 4:59 p.m., Afternoon: 5:00 to 8:59 p.m. 15 5.2 Overall and subperiods Impacts of Pico y Placa Having established that the empirical strategy is credible, in Table 4we report the baseline impacts of the Pico y Placa intervention, that is, the βkcoefficients from equation (1). Panel a presents the results for all roads pooled, while panels b and c report impacts for local and non-local roads respectively. We show results for all distance rings: the first 0-250 meter ring indicates the effect of the policy in the area of direct influence, while the remaining rings offer insights on the spatial spillovers. The results indicate positive impacts on the Pico y Placa ring in the morning and for the local road segments, both in terms of speed and minutes in severe traffic jam. The impacts on speed are very small, just 3 log points. However, the impacts on minutes in a severe traffic jam are larger: the decrease of 0.45 minutes20 is equivalent to an almost 70 percent decrease compared to the pre-treatment mean. There is some evidence of increases in the number of minutes in a severe traffic jam during the midday hour, but none of the coefficients are statistically significant at the 5 percent level. Panel b, for local road segments, confirms the increase in speed in the 0-250 meter ring as well as the reduction in the number of minutes in a severe traffic jam. In this case, the 0.38 reduction in the number of minutes in severe traffic jam during the morning represents an 84 percent reduction of the pre-treatment mean. Interestingly, spatial spillovers are observed in the morning that reinforce the scope of the policy. In terms of speed, there is some evidence of improvements in speed of 2 log points in the morning, but the coefficient is not statistically significant. Also in the morning there is evidence of reductions in the number of minutes in a severe traffic jam for the second and third distance rings. With respect to pre-treatment averages, the reduction of 0.29 minutes in the second ring and 0.26 minutes in the third ring represent 76 percent and 70 percent of pre-treatment levels respectively. Panel c shows that non-local road segments also improved their traffic congestion. Speed increased by 6 log points in the 0-250 meter ring both in the morning and afternoon. The number of minutes in a severe traffic jam also declined by 0.74 in the morning and 1.08 in the afternoon. These reductions represent 53 percent and 24 percent of pre-treatment averages, respectively. However, there is some evidence of increases in the number of minutes in a severe traffic jam during the mornings for the remaining five distance rings. The 0.87 minute increase for the second ring represents 93 percent of the pre-treatment average, while the increase of 0.8 minutes in the 1500-2000 meter ring represents 87 percent of the pre-treatment average. To investigate further whether the impacts of the driving restrictions varied over time, in Table 5we split the post-treatment period into two subperiods: (i) an early impact from September 2 to October 13 2019 and (ii) a late impact from October 14 to December 22 2019. Though we do not observe drivers directly, this distinction between early and late impacts aims to understand possible changes in behavior in line with adaptations to the policy. As shown in Figure 3, the number of traffic infractions related to Pico y Placa remained high and fairly constant from August to December 2019, suggesting that there were no major changes in enforcement intensity of the policy in the analysis period. Thus, we believe the results in Table 5capture drivers’ behavioral changes. In Table 5we show results only for local and non-local roads in panels a and b. Each panel is subdivided into subpanels for these subperiods. This distinction helps to uncover important early impacts on traffic congestion on local roads that later disappear. Subpanel a.1 shows that morning speed improved across all distance rings. The probability of a severe traffic jam and the number 20That is, 27 seconds. As the unit of measure for this variable is minutes, values below one minute are proportions of a minute. For example, 0.5 minutes equals 30 seconds, and 2.5 minutes equals 150 seconds. 16 of minutes in a severe traffic jam also declined in the morning. For the afternoon, we only observe reductions in the number of minutes in severe traffic jam across all distance rings. Subpanel a.2 indicates that all these gains disappeared in the period from October 14 to December 22. Regarding non-local roads, subpanel b.1 shows improvements in speed both in the morning and afternoon, and reductions in the probability of a severe traffic jam and in the number of minutes in a severe traffic jam both in the morning and afternoon. However in the following period (October 14 to December 22, subpanel b.2) these gains disappeared and in some cases we start to see a decline in the quality of traffic for some rings at different times of day. For instance, there is some evidence of a reduction in speed during midday, although only the reduction for the 20002500 meter ring of 0.07 log points is statistically significant at the 5 percent level. The number of minutes in a severe traffic jam increased in many distance rings away from the policy area during the three times of day, morning, midday and afternoon. We investigate this further in the next section. Appendix Table A2 is similar to Table 5but uses the alternative standard errors clustering approach to allow for both time and spatial correlation, as discussed in Section 4. While standard errors are generally larger, all statistically significant results remain significant. Another issue of interest is whether there are heterogeneous impacts of Pico y Placa within the first treatment ring. In particular, this first ring includes road segments not directly affected by the policy in addition to the ones that are directly affected. In Appendix Table A3 we split the first ring in two: those road segments within a distance buffer of 20 meters around the centroids of the segments directly affected by the policy (58 percent of the ring segments), and the rest of road segments in the ring. Although the impacts in general appear as larger for the 0-20 meter ring, overall results appear as qualitatively the same. 17 Table 4. Pico y Placa Impacts Distance to Ln(Speed) Pr(Severe Traffic Jam) Minutes in Severe Traffic Jam Pico y Placa Morning Midday Afternoon Morning Midday Afternoon Morning Midday Afternoon (meters): (1) (2) (3) (4) (5) (6) (7) (8) (9) Panel a. Overall 0-250 0.03*** -0.02 0.01 -0.01* 0.02* -0.00 -0.45*** 0.29* -0.13 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.13) (0.16) (0.13) 250-500 0.01 -0.02 -0.01 -0.00 0.02* 0.01 -0.11 0.31* 0.21 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.13) (0.16) (0.13) 500-1000 0.01 -0.02 -0.01 -0.00 0.02* 0.01 -0.09 0.27* 0.19 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.13) (0.16) (0.13) 1000-1500 0.01 -0.02 -0.01 0.00 0.02* 0.01 -0.04 0.28* 0.16 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.14) (0.16) (0.13) 1500-2000 0.01 -0.02 -0.00 0.00 0.02 0.00 -0.02 0.24 0.04 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.13) (0.16) (0.13) 200-2500 0.00 -0.01 -0.00 0.00 0.01 0.00 -0.04 0.26 0.10 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.14) (0.16) (0.13) Panel b. Local Roads 0-250 0.02** -0.01 -0.01 -0.01 0.01 0.00 -0.38*** 0.22 0.07 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.12) (0.16) (0.09) 250-500 0.02* -0.01 -0.01 -0.01 0.01 0.01 -0.29** 0.22 0.10 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.12) (0.16) (0.09) 500-1000 0.01 -0.01 -0.01 -0.01 0.01 0.01 -0.26** 0.14 0.01 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.12) (0.16) (0.09) 1000-1500 0.01 -0.01 -0.01 -0.00 0.01 0.01 -0.18 0.19 0.07 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.13) (0.16) (0.09) 1500-2000 0.01 -0.01 -0.00 -0.00 0.01 0.00 -0.18 0.19 -0.04 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.12) (0.16) (0.09) 200-2500 0.01 -0.01 0.00 -0.00 0.01 0.00 -0.13 0.13 -0.02 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.12) (0.16) (0.09) Panel c. Non-Local Roads 0-250 0.06** -0.02 0.06** -0.03 0.03 -0.04** -0.74* 0.50 -1.08** (0.03) (0.03) (0.03) (0.02) (0.02) (0.02) (0.41) (0.50) (0.52) 250-500 -0.03 -0.04 -0.02 0.02 0.04 0.02 0.87** 0.68 0.80 (0.03) (0.03) (0.02) (0.02) (0.02) (0.02) (0.42) (0.50) (0.53) 500-1000 -0.03 -0.04 -0.04* 0.03 0.04* 0.03* 0.76* 0.84* 1.08** (0.03) (0.03) (0.02) (0.02) (0.02) (0.02) (0.41) (0.50) (0.52) 1000-1500 -0.01 -0.03 -0.02 0.02 0.03 0.02 0.58 0.60 0.59 (0.03) (0.03) (0.02) (0.02) (0.02) (0.02) (0.41) (0.50) (0.54) 1500-2000 -0.03 -0.03 -0.01 0.03 0.03 0.01 0.80** 0.45 0.61 (0.03) (0.03) (0.02) (0.02) (0.02) (0.02) (0.40) (0.50) (0.52) 200-2500 -0.03 -0.03 -0.01 0.03* 0.03 0.02 0.34 0.67 0.57 (0.03) (0.03) (0.02) (0.02) (0.02) (0.02) (0.46) (0.50) (0.53) Source: Prepared by the authors. Note: Standard errors clustered at the road segment level. *, ** and *** denote statistical significance at the 10, 5 and 1 percent level, respectively. This table shows the coefficients associated with the PyP dummy in equation (1). Morning: 6:00 to 9:59 a.m., Midday: 10:00 to 4:59 p.m., Afternoon: 5:00 to 8:59 p.m. 18 Table 5. Early and Late Pico y Placa Impacts Distance to Ln(Speed) Pr(Severe Traffic Jam) Minutes in Severe Traffic Jam Pico y Placa Morning Midday Afternoon Morning Midday Afternoon Morning Midday Afternoon (meters): (1) (2) (3) (4) (5) (6) (7) (8) (9) Panel a. Local Roads Panel a.1 Early Pico y Placa (September 2 to October 13 2019) 0-250 0.03*** -0.00 0.01** -0.02*** 0.00 -0.01** -0.54*** 0.06 -0.28** (0.01) (0.01) (0.00) (0.00) (0.01) (0.00) (0.09) (0.10) (0.13) 250-500 0.03*** -0.00 0.01 -0.01*** 0.00 -0.00 -0.45*** 0.06 -0.24* (0.01) (0.01) (0.00) (0.00) (0.01) (0.00) (0.09) (0.11) (0.13) 500-1000 0.02*** -0.00 0.01 -0.01*** 0.00 -0.00 -0.39*** -0.01 -0.29** (0.01) (0.01) (0.00) (0.00) (0.01) (0.00) (0.09) (0.10) (0.13) 1000-1500 0.02*** -0.00 0.01 -0.01* 0.00 -0.00 -0.32*** 0.03 -0.22* (0.01) (0.01) (0.01) (0.00) (0.01) (0.00) (0.10) (0.10) (0.14) 1500-2000 0.02*** -0.00 0.01** -0.01** 0.00 -0.01* -0.33*** 0.04 -0.35*** (0.00) (0.01) (0.00) (0.00) (0.01) (0.00) (0.09) (0.11) (0.13) 200-2500 0.01*** 0.00 0.01** -0.01 0.00 -0.01* -0.26*** 0.04 -0.24* (0.00) (0.01) (0.00) (0.00) (0.01) (0.00) (0.10) (0.11) (0.14) Panel a.2 Late Pico y Placa (October 14 to December 22 2019) 0-250 0.01 -0.03 -0.02 -0.00 0.02 0.02 -0.22 0.38* 0.44** (0.02) (0.02) (0.02) (0.01) (0.02) (0.01) (0.19) (0.22) (0.20) 250-500 0.01 -0.03 -0.02 -0.00 0.02 0.02 -0.14 0.38* 0.45** (0.02) (0.02) (0.02) (0.01) (0.02) (0.01) (0.19) (0.23) (0.20) 500-1000 0.00 -0.02 -0.02 -0.00 0.02 0.02 -0.11 0.30 0.33* (0.02) (0.02) (0.02) (0.01) (0.02) (0.01) (0.19) (0.22) (0.20) 1000-1500 0.00 -0.03 -0.02 0.00 0.02 0.02 -0.04 0.36 0.38* (0.02) (0.02) (0.02) (0.01) (0.02) (0.01) (0.20) (0.23) (0.20) 1500-2000 0.00 -0.02 -0.01 0.00 0.02 0.01 -0.03 0.34 0.28 (0.02) (0.02) (0.02) (0.01) (0.02) (0.01) (0.19) (0.23) (0.20) 200-2500 -0.00 -0.02 -0.01 0.00 0.02 0.01 0.01 0.23 0.21 (0.02) (0.02) (0.02) (0.01) (0.02) (0.01) (0.19) (0.23) (0.20) Panel b. Non-Loca Roads Panel b.1 Early Pico y Placa (September 2 to October 13 2019) 0-250 0.06** -0.00 0.08*** -0.03* 0.01 -0.05*** -0.80** 0.27 -1.40*** (0.02) (0.03) (0.02) (0.02) (0.02) (0.02) (0.38) (0.52) (0.49) 250-500 -0.03 -0.01 -0.01 0.02 0.02 0.01 0.66* 0.32 0.28 (0.02) (0.03) (0.02) (0.02) (0.02) (0.02) (0.39) (0.52) (0.49) 500-1000 -0.03 -0.02 -0.02 0.03 0.03 0.02 0.56 0.49 0.52 (0.02) (0.03) (0.02) (0.02) (0.02) (0.02) (0.37) (0.52) (0.48) 1000-1500 -0.01 -0.01 -0.02 0.02 0.01 0.01 0.46 0.21 0.28 (0.02) (0.03) (0.02) (0.02) (0.02) (0.02) (0.37) (0.52) (0.51) 1500-2000 -0.02 -0.01 -0.01 0.02 0.01 0.00 0.51 0.05 0.22 (0.02) (0.03) (0.02) (0.02) (0.02) (0.02) (0.37) (0.52) (0.49) 200-2500 -0.03 0.01 0.01 0.03 0.00 0.00 0.37 0.15 -0.13 (0.02) (0.03) (0.02) (0.02) (0.02) (0.02) (0.44) (0.54) (0.53) Panel b.2 Late Pico y Placa (October 14 to December 22 2019) 0-250 0.05* -0.04 0.04 -0.03 0.04 -0.03 -0.67 0.73 -0.78 (0.03) (0.03) (0.03) (0.02) (0.03) (0.02) (0.46) (0.51) (0.58) 250-500 -0.03 -0.06* -0.04 0.03 0.05** 0.03 1.09** 1.03** 1.30** (0.03) (0.03) (0.03) (0.02) (0.03) (0.02) (0.47) (0.51) (0.59) 500-1000 -0.03 -0.06* -0.06** 0.03 0.06** 0.05** 0.97** 1.19** 1.63*** (0.03) (0.03) (0.03) (0.02) (0.03) (0.02) (0.45) (0.51) (0.57) 1000-1500 -0.01 -0.05 -0.03 0.01 0.04 0.02 0.72 0.99* 0.88 (0.03) (0.03) (0.03) (0.02) (0.03) (0.02) (0.46) (0.51) (0.59) 1500-2000 -0.03 -0.05 -0.02 0.03 0.04 0.02 1.10** 0.86* 1.00* (0.03) (0.03) (0.03) (0.02) (0.03) (0.02) (0.45) (0.52) (0.58) 200-2500 -0.03 -0.07** -0.03 0.04* 0.06** 0.03 0.30 1.20** 1.28** (0.03) (0.03) (0.03) (0.02) (0.03) (0.02) (0.51) (0.56) (0.63) Source: Prepared by the authors. Note: Standard errors clustered at the road segment level. *, ** and *** denote statistical significance at the 10, 5 and 1 percent level, respectively. This table shows coefficients associated with a version of equation (1) where the PyP dummy is split into two subperiods (early and late). Morning: 6:00 to 9:59 a.m., Midday: 10:00 to 4:59 p.m., Afternoon: 5:00 to 8:59 p.m. 19 5.3 Hourly Impacts of Pico y Placa To study in more detail the impacts of Pico y Placa, and to also analyze potential time spillover effects, we re-estimate equation (1) by each hour of the day and report in Figures 4and 5the associated βkcoefficients for both local and non-local roads and for the two post-treatment periods, as in Table 5. We discuss here the results for speed; those for the probability of severe traffic jam and minutes in severe traffic jam can be found in Appendix Figures B4 to B7. Figure 4focuses on local roads. Panel a shows the early impacts while panel b shows late impacts. Results in panel a show improvements in speed in some hours of the morning peak times. These increases in log speed were not higher than 0.05 log points. Panel b indicates that those hourly gains are no longer statistically significant and close to zero in the later period. It also shows some evidence of reductions in speed at midday, though no coefficient is statistically significant. Figure 5indicates that early impacts on non-local roads are higher in magnitude. For the 0-250 meter ring, increases in morning speed are close to 0.10 log points for 7, 8 and 9 a.m. A similar pattern emerges for the afternoon with important increases in speed for 6 and 7 p.m. (hours 18 and 19, respectively, as shown in the figure). For the rest of the distance rings the effects are not statistically significant. In terms of late impacts, panel b shows that some of the morning and afternoon increases in speed for the 0-250 meter ring remain, but for the remaining rings there is evidence of reductions at different hours of the day. Taken together, the results imply that an account of the full impact of the driving restrictions intervention requires properly considering both types of spillovers (spatial and time), while also considering the heterogeneous impact by type of road. We attempt such analysis in the next section. 20 Figure 6. Overall Welfare Effects Source: Prepared by the authors. Note: For each subperiod the welfare change is estimated as a weighted average of ητ. Weights are the interaction of the initial ratio of average free-flow speed over average circulation speed by hour, distance ring and road type; and total kilometers by distance ring and road type. The aggregation of kilometers weights by free-flow speed to consider differences in road capacity. 27 Figure 7. Welfare Effects by Time of Day Schedule Source: Prepared by the authors. Note: For each period the welfare change is estimated as a weighted average of ητ. Weights are the interaction of the initial ratio of average free-flow speed over average circulation speed by hour, distance ring and road type; and total kilometers by distance ring and road type. The aggregation of kilometers weights by free-flow speed to consider differences in road capacity. 28 Figure 8. Welfare Effects by Distance Ring Source: Prepared by the authors. Note: For each period the welfare change is estimated as a weighted average of ητ. Weights are the interaction of the initial ratio of average free-flow speed over average circulation speed by hour, distance ring and road type; and total kilometers by distance ring and road type. The aggregation of kilometers weights by free-flow speed to consider differences in road capacity. 29 6.3 Distributional Welfare Impacts Based on Commuter Routes As mentioned above, we can only very imperfectly try to approximate the distributional welfare impacts of the driving restriction policy, by relying on broad characterizations of commuter routes. An argument could be made that welfare losses for individual vehicle riders could be offset by welfare gains for public transport users, who represent around 71 percent of all commuters (Lima Cómo Vamos,2019). While we cannot observe public transport users directly, we can attempt to approximate the welfare impacts of the policy on public transport users by evaluating changes in travel time along city roads that are heavily used by public transport. Lima’s public transport system is heavily informal, making it hard to pinpoint exact routes. However, we can focus on the route covered by one of the few formal modes available in the city, the Metropolitano Bus Rapid Transit (BRT) system. There is evidence linking the opening and operation of this BRT system to higher usage of public transport, specially among women (Martinez et al.,2020). We use all the road segments that fall within the BRT route and its feeder buses to estimate changes in welfare for the late-impact subperiod (from October 14 to December 22 2019), only for those road segments. The results for this reduced group of road segments, shown in Figure 9, are quite similar to those for all the road segments. The total change in estimated welfare is essentially the same, a 2.05 percent loss. However, the welfare loss associated with the extensive margin is slightly larger at 1.65 percent, representing over 80 percent of the total loss, compared to 73 percent explained using all the segments. The evidence therefore suggests that the negative impacts of the policy are not different for those roads used by the BRT system when compared to the overall negative impacts, and that the extensive margin appears to be even more important in explaining those negative welfare impacts, for BRT system-related roads. An alternative attempt to quantifying distributional impacts is to compare welfare changes across all commuters, not only public transit users, classified by socioeconomic status (SES). For this, we classify different areas of the city according to the SES of the majority of commuters they attract, using data from Lima’s 2012 Origin-Destination survey (JICA,2013). The survey has five SES categories ranging from A (rich) to E (very poor). We reclassify these five categories into three where categories A and B are classified as upper SES, C as middle SES, and D and E as low SES. A traffic zone is classified by SES using a “majority rule” based on the category with the highest share across the three categories. Using this approach, 17 percent of the traffic zones are classified as upper SES, 39 percent as middle SES, and the remaining 44 percent as low SES. Appendix Figure B8 presents the total welfare change for each of the SES categories and overall. Compared to the total overall welfare impacts, we observe only marginal differences across the three SES categories; qualitatively there are no substantive differences. The decomposition into the extensive and intensive margins (not presented) shows similar results. Based on these two analyses we conclude that there is no evidence that the Pico y Placa policy had any differential welfare impacts by commuter-characteristics, at least based on the limited characterization conducted in this subsection. 30 Figure 9. Total Welfare Effects for BRT-System-Related Road Segments Source: Prepared by the authors. Note: For each period the welfare change is estimated as a weighted average of ητ. Weights are the interaction of the initial ratio of average free-flow speed over average circulation speed by hour, distance ring and road type; and total kilometers by distance ring and road type. The aggregation of kilometers weights by free-flow speed to consider differences in road capacity. 7 Conclusion This paper has evaluated the effects of Lima’s driving restriction policy known as Pico y Placa. The policy restricted vehicle circulation on several important roads with the aim of reducing traffic congestion and pollution. The analysis exploited high-frequency road-segment-level congestion data from the communitybased driving directions app Waze. We restricted the analysis to road segments within three kilometers of the area of influence of the restrictions and imposed a generalized propensity scorebased overlap condition, to guarantee comparability of segments across seven distance rings from roads directly affected by the policy. We estimated difference-in-differences models using the segments that satisfy the overlap condition and weighting by the inverse of the generalized propensity score. The outcomes of interest were circulation speed, probability of a severe traffic jam and the number of minutes in a severe traffic jam. Our results suggest small gains in the intervened area combined with small losses in nearby areas and at hours outside the time schedule of the policy. Although the direct area of influence improved speed by 2 percent, further analysis that distinguishes by road types and timing of the policy suggests that these gains disappeared weeks after the start of the policy. 31 The highly detailed data allowed for quantifying impacts by hour, road type and distance to the intervened area, which ultimately helped to quantify spatial and time spillovers that appear negative in areas and hours not directly affected by the policy. We conducted a welfare analysis and estimated welfare impacts in a transition period and in two (early and late) impact subperiods. While during the transition and early-impact subperiods there were small welfare gains, by the late-impact subperiod we estimate an overall welfare loss from the policy of 2 percent. Most of those losses, 73 percent, are explained by the extensive margin, that is, more roads becoming severely congested. Most of the welfare loss took place in the midday hours outside the target hours of the policy. While the area directly targeted showed a net small welfare gain, all other areas showed clear welfare losses. Both, the time and spatial spillover-driven welfare losses seem to be mostly explained by the extensive margin. Finally, welfare changes in areas with heavy public transport use, or with a majority of commuters with different socioeconomic status seem to be very similar to the overall welfare changes; that is, they suffered very similar welfare losses. Our results indicate that while the very localized direct effects of the Pico y Placa policy may appear to have been slightly positive, the overall impact was clearly negative. Overall, congestion increased in the areas analyzed, which goes against one of the stated objectives of the policy (to reduce traffic congestion). Furthermore, if cars are taking longer to drive the same routes (or are taking less-optimal routes to avoid the driving restrictions), both gasoline consumption and pollution associated with driven miles and idling time increases. That is, the policy appears to have failed in its second stated objective (to reduce pollution), while increasing other environmental costs through increased gasoline consumption. These results highlight the need for policy makers to take into account the overall impacts of driving restrictions policies before implementing them. Indeed, the analysis in this paper suggests that these types of policies can only be justified if the authorities put a very high weight on the very localized areas that unambiguously benefit from the policy. Otherwise, these policies seem difficult to justify, at least in the setting studied. 32 References Angrist, J. and J.-S. Pischke (2009): Mostly Harmless Econometrics, New York: Princeton University Press. Barahona, N., F. A. Gallego, and J.-P. Montero (2020): “Vintage-Specific Driving Restrictions,” The Review of Economic Studies, 87, 1646–1682. Blackman, A., F. Alpízar, F. Carlsson, and M. R. Planter (2018a): “A Contingent Valuation Approach to Estimating Regulatory Costs: Mexico’s Day without Driving Program,” Journal of the Association of Environmental and Resource Economists, 5, 607–641. Blackman, A., Z. Li, and A. A. Liu (2018b): “Efficacy of Command-and-Control and Market-Based Environmental Regulation in Developing Countries,” Annual Review of Resource Economics, 10, 381–404. Bonilla, J. A. (2019): “The More Stringent, the Better? Rationing Car Use in Bogotá with Moderate and Drastic Restrictions,” World Bank Economic Review, 33, 516–534. Calatayud, A., S. Sánchez González, F. Bedoya Maya, F. Giraldez Zúñiga, and J. M. Márquez (2021): Congestión urbana en América Latina y el Caribe: Características, costos y mitigación, Inter-American Development Bank. Callaway, B. and P. H. C. Sant’ Anna (2021): “Difference-in-Differences with multiple time periods,” Journal of Econometrics, 225, 200–230. Carrillo, P. E., A. Lopez-Luzuriaga, and A. S. Malik (2018): “Pollution or crime: The effect of driving restrictions on criminal activity,” Journal of Public Economics, 164, 50 – 69. Carrillo, P. E., A. S. Malik, and Y. Yoo (2016): “Driving restrictions that work? Quito’s "Pico y Placa" Program,” The Canadian Journal of Economics / Revue canadienne d’Economique, 49, 1536–1568. Davis, L. W. (2008): “The Effect of Driving Restrictions on Air Quality in Mexico City,” Journal of Political Economy, 116, 38–81. De Grange, L. and R. Troncoso (2011): “Impacts of vehicle restrictions on urban transport flows: The case of Santiago, Chile,” Transport Policy, 18, 862 – 869. Eskeland, G. S. and T. Feyzioglu (1997): “Rationing Can Backfire: The "Day without a Car" in Mexico City,” The World Bank Economic Review, 11, 383–408. Flores, C. A. and O. A. Mitnik (2013): “Comparing Treatments across Labor Markets: An Assessment of Nonexperimental Multiple-Treatment Strategies,” Review of Economics and Statistics, 95, 1691–1707, 00000. Gallego, F., J.-P. Montero, and C. Salas (2013a): “The effect of transport policies on car use: A bundling model with applications,” Energy Economics, 40, S85 – S97, supplement Issue: Fifth Atlantic Workshop in Energy and Environmental Economics. ——— (2013b): “The effect of transport policies on car use: Evidence from Latin American cities,” Journal of Public Economics, 107, 47 – 62. 33 Hall, J. D. (2021a): “Can Tolling Help Everyone? Estimating the Aggregate and Distributional Consequences of Congestion Pricing,” Journal of the European Economic Association, 19, 441– 474. ——— (2021b): “Inframarginal Travelers and Transportation Policy,” SSRN Electronic Journal, https://www.ssrn.com/abstract=3424097 . Hanna, R., B. Olken, and G. Kreindler (2017): “Citywide effects of high-occupancy vehicle restrictions: Evidence from “three-in-one” in Jakarta,” Science, 357, 89–93. Hirano, K. and G. W. Imbens (2004): “The Propensity Score with Continuous Treatments,” in Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives, ed. by A. Gelman and X.-L. Meng, Hoboken, NJ: John Wiley and Sons, Wiley Series in Probability and Statistics, 73–84. Imbens, G. W. (2000): “The Role of the Propensity Score in Estimating Dose-Response Functions,” Biometrika, 87, 706–710. INEI (2020): “ENAHO - Ingreso Promedio Proveniente del Trabajo,” Tech. rep., https://www. inei.gob.pe/media/MenuRecursivo/indices_tematicos/ing-cuad-1_1.xlsx , accessed August 30, 2021. Jauregui-Fung, F., J. Kenworthy, S. Almaaroufi, N. Pulido-Castro, S. Pereira, and K. GoldaPongratz (2019): “Anatomy of an Informal Transit City: Mobility Analysis of the Metropolitan Area of Lima,” Urban Science, 3. JICA (2013): “Encuesta de Recolección de Información Básica del Transporte Urbano en el Área Metropolitana de Lima Y Callao. Informe Final.” Tech. rep., Japan International Cooperation Agency, https://openjicareport.jica.go.jp/pdf/12087532_01.pdf . Kreindler, G. (2016): “Driving Delhi? Behavioural Responses to Driving Restrictions,” Mimeo. Lima Cómo Vamos (2019): “¿Cómo Vamos en Lima y Callao? Noveno Informe de Indicadores sobre Calidad de Vida,” Tech. rep., http://www.limacomovamos.org/wp-content/uploads/2019/ 11/Encuesta-2019_web.pdf , accessed January 17, 2021. Martinez, D. F., O. A. Mitnik, E. Salgado, L. Scholl, and P. Yañez-Pagans (2020): “Connecting to Economic Opportunity: the Role of Public Transport in Promoting Women’s Employment in Lima,” Journal of Economics, Race, and Policy, 3, 1–23. Municipalidad de Lima (undated): “Pico y Placa Lima - Vehiculos en general,” https:// aplicativos.munlima.gob.pe/pico-y-placa , accessed November 24, 2020. Ryan, A. M., E. Kontopantelis, A. Linden, and J. F. Burgess (2019): “Now trending: Coping with non-parallel trends in difference-in-differences analysis,” Statistical Methods in Medical Research, 28, 3697–3711. SAT (undated): “Tabla de Infracciones - Reglamento Nacional de Tránsito,” Tech. rep., Servicio de Administración Tributaria de Lima, https://www.sat.gob.pe/websitev8/modulos/ contenidos/mult_papeletas_ti_rntv2.aspx , accessed August 27, 2021. TomTom (2020): “Traffic Index 2019,” Tech. rep., https://www.tomtom.com/traffic-index/ ranking/ , accessed January 10, 2021. 34 Troncoso, R., L. de Grange, and L. A. Cifuentes (2012): “Effects of environmental alerts and preemergencies on pollutant concentrations in Santiago, Chile,” Atmospheric Environment, 61, 550 – 557. United Nations (2018): World Urbanization Prospects, New York: United Nations. U.S. Treasury (undated): “U.S. Treasury Reporting Rates of Exchange - December 31, 2019,” Tech. rep., https://www.fiscal.treasury.gov/files/reports-statements/ treasury-reporting-rates-exchange/ratesofexchangeasofdecember312019.pdf , accessed August 27,2021. Viard, V. B. and S. Fu (2015): “The effect of Beijing’s driving restrictions on pollution and economic activity,” Journal of Public Economics, 125, 98 – 115. Yañez-Pagans, P., D. Martinez, O. A. Mitnik, L. Scholl, and A. Vazquez (2019): “Urban transport systems in Latin America and the Caribbean: lessons and challenges,” Latin American Economic Review, 28, 1–25. Ye, J. (2017): “Better safe than sorry? Evidence from Lanzhou’s driving restriction policy,” China Economic Review, 45, 1 – 21. 35 Online Appendix - Not for publication A Appendix Tables 36 B Appendix Figures 43 Figure B1. Leads and lags: Ln(Speed) Panel a. Morning Panel b. Afternoon Source: Prepared by the authors. Note: Horizontal axis represents two-weeks periods. Period 0 are the two weeks when Pico y Placa sarted. Period 1 is the two-weeks following, and so on. The two dashed vertical lines mark the start and end of the transition period that coincided with the Panamerican and Parapanamerican Games. The reference period includes weeks -28 to -21 before the start of the restrictions. All figures show 95 percent CI. 44 Figure B2. Leads and lags: Probability of severe traffic jam Panel a. Morning Panel b. Afternoon Source: Prepared by the authors. Note: Horizontal axis represents two-weeks periods. Period 0 are the two weeks when Pico y Placa sarted. Period 1 is the two-weeks following, and so on. The two dashed vertical lines mark the start and end of the transition period that coincided with the Panamerican and Parapanamerican Games. The reference period includes weeks -28 to -21 before the start of the restrictions. All figures show 95 percent CI. 45 Figure B3. Leads and lags: Number of minutes in severe traffic jam Panel a. Morning Panel b. Afternoon Source: Prepared by the authors. Note: Horizontal axis represents two-weeks periods. Period 0 are the two weeks when Pico y Placa sarted. Period 1 is the two-weeks following, and so on. The two dashed vertical lines mark the start and end of the transition period that coincided with the Panamerican and Parapanamerican Games. The reference period includes weeks -28 to -21 before the start of the restrictions. All figures show 95 percent CI. 46 Figure B4. Hourly impacts of Pico y Placa on Pr(severe traffic jam) - Local roads Panel a. Early Impact (September 2 to October 13 2019) Panel b. Late Impact (October 14 to December 22 2019) Source: Prepared by the authors. Note: The horizontal axis is the hour of the day. Morning and afternoon Pico y Placa hours are represented by the shaded areas. The vertical lines represent 95 percent confidence intervals. 47 Figure B5. Hourly impacts of Pico y Placa on Pr(severe traffic jam) - Non-local roads Panel a. Early Impact (September 2 to October 13 2019) Panel b. Late Impact (October 14 to December 22 2019) Source: Prepared by the authors. Note: The horizontal axis is the hour of the day. Morning and afternoon Pico y Placa hours are represented by the shaded areas. The vertical lines represent 95 percent confidence intervals. 48 Figure B6. Hourly impacts of Pico y Placa on minutes in severe traffic jam - Local roads Panel a. Early Impact (September 2 to October 13 2019) Panel b. Late Impact (October 14 to December 22 2019) Source: Prepared by the authors. Note: The horizontal axis is the hour of the day. Morning and afternoon Pico y Placa hours are represented by the shaded areas. The vertical lines represent 95 percent confidence intervals. 49 Figure B7. Hourly impacts of Pico y Placa on minutes in severe traffic jam - Non-local roads Panel a. Early Impact (September 2 to October 13 2019) Panel b. Late Impact (October 14 to December 22 2019) Source: Prepared by the authors. Note: The horizontal axis is the hour of the day. Morning and afternoon Pico y Placa hours are represented by the shaded areas. The vertical lines represent 95 percent confidence intervals. 50 Figure B8. Total Welfare Effects by Socioeconomic Status Source: Prepared by the authors. Note: For each period the estimated welfare change is estimated as a weighted average of ητ. Weights are the interaction of the initial ratio of average free-flow speed over average circulation speed by hour, distance ring and road type; and total kilometers by distance ring and road type. The aggregation of kilometers weights by free-flow speed to consider differences in road capacity. Weights are computed for the group of segments that fall within the corresponding SES boundaries defined by the traffic zone. The SES classification of the traffic zones was based on the majority SES category in the traffic zone according to Lima’s 2012 Origin-Destination survey (JICA,2013) 51