Effect of pop-up bike lanes on cycling in European cities
EEffect of pop-up bike lanes on cycling in European cities
Sebastian Kraus ∗ and Nicolas Koch Mercator Research Center on Global Commons and Climate Change, Torgauer Str. 19, 10829 Berlin, Germany Potsdam Institute for Climate Impact Research, Telegrafenberg, 14473 Potsdam, Germany Technical University of Berlin, Straße des 17. Juni 135, 10623 Berlin, Germany
September 8, 2020
Abstract
The bicycle is a low-cost means of transport linked to low risk of COVID-19 trans-mission. Governments have incentivised cycling by redistributing street space as partof their post-lockdown strategies. We evaluated the impact of provisional bicycle infras-tructure on cycling traffic in European cities using a generalised difference-in-differencesdesign. We scraped daily bicycle counts spanning over a decade from 736 bicycle coun-ters in 106 European cities. We combined this with data on announced and completedpop-up bike lane road work projects. On average 11.5 kilometres of provisional pop-upbike lanes have been built per city. Each kilometre has increased cycling in a city by 0.6%.We calculate that the new infrastructure will generate $2.3 billion in health benefits peryear, if cycling habits are sticky.
As social and economic activity resume after a period of social distancing to curb COVID-19, policy-makers are seeking mitigation measures with favourable cost-benefit ratios thatcan be implemented in the short-run. While overall mobility is almost back to pre-crisis lev-els in many European countries, the use of public transport is still lagging behind . Earlyevidence points to shifts from public transport to car use as users react to the pandemic .Governments have started incentivising cycling as a low-cost, sustainable, equitable, andspace-saving mode of transport that reduces the risk of COVID-19 transmission. A keymeasure has been the redistribution of street space in cities to create provisional bike in-frastructure often provisionally marked and protected by materials readily available fromroad construction companies. As of 8 July 2020, 2000 kilometres of these infrastructurechanges had been announced .In Europe, typically more than 50% of overall trips measured in transport surveys areshorter than 5 kilometres . In 2019, 3 million electric bicycles were sold in the EU likelymaking cycling more demographically diverse and increasing the distances travelled . Thisspeaks to an important short-term potential for shifts in transport mode choice that couldreduce crowds in public transport and help avoid traffic congestion in response to increasedcar use out of fears of infection. ∗ Corresponding author. Email address: [email protected]. ORCID: https://orcid.org/0000-0003-1161-2988 a r X i v : . [ phy s i c s . s o c - ph ] S e p ode choices are subject to behavioural effects, such as status quo bias, default effects, andtime-inconsistent preferences . This complicates the task of policy-makers to encouragepeople to cycle, particularly in the short-run. However, major disruptions to public trans-port, such as strikes, cause people to reconsider their habits . Furthermore, highly visible,large-scale expansions in the provision of bicycle amenities, such as bike sharing or a city-wide network of 120 kilometres of separated bike lanes built within four years in Sevilla ,have increased cycling and reduced congestion .Here, we provide causal estimates of the effect of the post-COVID-19-lockdown roll-outof provisional (“pop-up”) bike lanes in European cities. We compile new data on dailybike counts in 110 cities. We connect to the open data application programming interfaces(APIs) of these cities to download bike counts from a total of 736 counters spanning overa decade. We combine this data with information on day-to-day changes in the numberof kilometres of pop-up bike lanes, which is collected by the European Cyclists’ Federationbased on official documents and media reports. Our sample consists of large and medium-sized cities in 20 European countries.
We estimate a Poisson regression model at the counter level with daily counts of cyclists asthe outcome variable and the number of kilometres of pop-up bike lanes in service in a cityon a given day as the treatment.Since the roll-out of pop-up bike lanes is not a controlled experiment, our main empiricalconcern is that both the implementation of bike lanes and bicycle counts are driven by athird factor that cannot be measured (omitted variable bias). We may also worry that bikelanes are built as a reaction to increased cycling traffic (reverse causality). We address theseconcerns using quasi-experimental variation in the roll-out of pop-up bike lanes in differentEuropean cities.Planning for provisional cycling infrastructure in Europe has started early in the pandemicas a reaction to civil society pressure after announcements by the City of Bogota on 16and 17 March to create 76 km of provisional bike lanes that was widely reported in theinternational media. Similar plans were assembled in several European cities and the roll-out of these plans has started during lock downs as a means to allow necessary travel underhigh safety standards particularly for “key workers” (see Fig. 1).Officials have stated in interviews and personal conversations, that the geographic place-ment of pop-up bike lanes has mainly been driven by the availability of street space thatcan be redistributed without restricting car traffic to only one direction and the existenceof “shovel-ready” construction plans. The exact timing of pop-up bike lane construction isdriven by administrative idiosyncrasies and the availability and schedules of constructionfirms. Therefore, we argue that the timing of the roll-out of pop-up bike lanes has been asgood as random.Our regression analysis is based on comparisons between treatment and control groups be-fore and after treatment around each cohort of new bike lanes (differences-in-differences).We use a set of indicator variables (fixed effects) that remove variation from our estima-tion sample that could be biasing our estimates. Our study design allows for systematicdifferences in the level of bike traffic between treatment and control group, but relies ona common trends assumption, that bike traffic in treated and control cities would have2 ig. 1. Intensity of pop-up bike lane treatment over time This Figure shows treated cities and their treatment intensities inimplemented kilometres of public bike lanes at a given day between March and July 2020. Control cities are not plotted butare included in Fig. S1 (see Supplementary Materials). London, Milan, Rome, and Lisbon are missing from the sample due toa lack of daily bicycle counter data. Bars that do not cover the whole study period to July 8 2020 indicate missing bicyclecount data for the most recent dates due to updating time lags of the counter APIs. Information on individual pop-up bikelanes with their street location, announcement date, and implementation date is from the European Cyclists’ Federation. Thenewest data can be found at: https://ecf.com/dashboard evolved on a parallel trend in the absence of treatment. Since we cannot observe treatedunits in their untreated state after treatment (potential outcome), we cannot test the com-mon trends assumption formally. However, we can investigate pre-treatment trends andcheck the sensitivity of our estimates to changes in the control group definition.As a baseline our difference-in-differences model includes fixed effects at the unit (counter)and time (day) level. We thereby control for time-invariant factors at the level of eachcounter and city, such as public transport and population density, topography, and prefer-ences for green lifestyles. With our counter fixed effect we also rule out that our effect isdriven by new counters that get placed next to provisional bike lanes. The day fixed effectremoves trends from the treatment and outcome variation that are common to the wholesample. These could be overall trends in cycling, seasonality, and the overall evolution ofthe COVID-19 pandemic in Europe. Fixed effects at the country-day level remove variationin cycling infrastructure and behaviour that is driven by state- or national-level COVID-19policies.For potentially biasing factors that vary at the city-level over time, such as local mobilityor weather, we cannot include fixed effects since this is the geographical level at whichour treatment is measured. We therefore include control variables in our regressions thatmeasure overall changes in mobility at the state-level. This variable is based on the aggre-gated movements of Facebook users. This is to rule out that our effect is driven by localauthorities reacting to increased traffic volumes. We also control for local temperatures,sunshine, wind and precipitation. Weather could for instance create bias, when construc-3ion firms decide to create new bike lanes in weeks with good weather that will also havemore cycling.
We compare bike traffic in treated cities in the days before and after they get treated com-pared to control cities and find that one kilometre of popup bike lane increases cycling by0.6% (see right one of coefficients marked in blue in Fig. 2). When we multiply this estimatefor a kilometre of bike lane with the average number of kilometres (11.5), we find that theaverage effect of bike lane programs is a 7% increase in city-wide cycling.
Model:Control group:Fixed effects:Control variables:
PoissonOLSTo be treatedTreated onlyTiming variation onlyEvent studyCounter FECounter−week FEDay FECity−week FECity−calendar week FECountry−day FEOverall mobilityTemperatureSunshineWindPrecipitationNumber of counters0.00.51.01.52.02.53.0 % c hange pe r k m b i k e l ane Fig. 2. Effect of pop-up bike lanes on cycling in different model specifications This figure shows estimates of regressions ofthe daily cyclist count on the number of kilometres of pop-up bike lane implemented at a given day in a city. The unit ofobservation is the bike counter. Baseline specifications are marked in blue. Darker colours in the bottom panel indicate thetype of specification used. The 95% confidence interval is shown in darker colour and the 90% confidence interval in lightercolour. One estimate is from an OLS specification and uses the natural logarithm of the bicycle count as the outcome. Allother specifications are Poisson regressions. The estimates can be interpreted as the average increase in the level of cyclingcaused by one kilometre of pop-up bike lane. Control variables are from Facebook (mobility index measured by usermovements) and the ERA5 climate model (weather variables). The variable
Number of counters indicates the total of countersper city.
The bike count data spans over a decade. We can compare changes in cycling in the weeksafter the introduction of pop-up bike lanes with the same calendar weeks in previous years.Fig. 2 shows a regression estimate based on this comparison (see left one of coefficientshighlighted in blue). Comparisons between weeks in 2020 and weeks in previous yearsmay be biased by differential trends between treatment and control group. Fig. 3 showsthe estimated difference between treatment and control group in the 12 months before and3 months after the begin of the pop-up bike lane roll-out in March. The baseline category inthis event study specification is −
13. This means that all estimated coefficients for monthsbefore and after treatment are relative to February 2019.4 .50.51 E s t i m a t e -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 Months relative to begin of pop-up rollout
Fig. 3. Treatment effect in months before and after beginning of pop-up bike lane treatment This figure shows the treatmenteffect of treated cities compared to untreated cities. Observations are binned into months before and after treatment. Thetreatment is hard-coded to March 2020. The baseline category and the begin of the sample is February 2019. Estimates arefrom Poisson regressions that include city and country-day fixed effects. The shaded area shows the 95% confidence interval.
We can see that a treatment effect becomes only apparent after the treatment sets in. Before,treatment and control group have been on the same trend. There is a slight but statisticallyinsignificant downward trend before treatment (Ashenfelter’s dip ), hinting at the possi-bility of stronger mobility reductions due to COVID-19 in cities that have decided to buildpop-up bike lanes. This could be the case because local and national governments aremore likely to take wide-ranging action, if their country is hit by a more intense outbreak.It could also be due to governments acting upon idiosyncratic risk-aversion of their popu-lations towards cycling in the context of emptier roads and increased speeding during thelockdown. We rule out that these potential selection into treatment effects are driving ourresults by controlling for COVID-19 related dynamics with fixed effects and a variable thatcaptures human mobility at the sub-national level based on Facebook user movements.The treatment effect magnitude in Fig. 3 is higher than our baseline estimate. This dif-ference stems from hard-coding the treatment in March 2020 and therefore discardingvariation in treatment timing at the day and week levels. This creates a more standarddifference-in-difference setting, that avoids the issue of already treated cities acting as con-trols for later cohorts, while they are still on a different trend because of prior treatment .Therefore our main estimates tend to be attenuated compared to the setup shown in Fig. 3.We check the sensitivity of our results to reshaping our regression-based treatment andcontrol group comparisons. Fig. 2 shows fairly stable estimates for comparisons between(i) treated and untreated cities, (ii) cities that are already treated and those that have onlyannounced pop-up bike lanes, (iii) between treated cities only using their variation in treat-ment dose (km of bike lane built) and treatment timing or (iv) treatment timing only (eventstudy).Estimates based on days before and after treatment within the same week have highermagnitudes when country-day fixed effects are excluded from the model. This suggeststhat within a narrow time window around treatment, national policy events and cyclingbehaviour are correlated. Estimates that either include country-day fixed effects or uselonger pre- and post-treatment windows (no city-week fixed effect or city-calendar weekfixed effect only) mitigate this bias. We find robust evidence for substantial short-run increases in cycling in European citiesdue to new provisional cycling infrastructure. An average pop-up bike lane program5as led to a 7% increase in city-wide cycling. The effects of this cycling infrastructure onCOVID-19 transmission should be investigated with high-resolution case data for a largeenough number of cities.Independent of its impacts on COVID-19 transmission, the net benefits of the interventionare likely to be large. The direct cost of cycling infrastructure including planning is low.For the Sevilla network one kilometre of bike lane cost A C250000 . Iterative planning withprovisional infrastructure reduces costs further. In Berlin, one kilometre of pop-up bikelane has so far cost A C9500 . Previous research has found that every kilometre of cyclinggenerates health benefits of $0.45 . We calculate baseline values for total cycling in a citybased on data on daily kilometres cycled in German cities in 2018 and extrapolate thesenumbers to the rest of our sample based on city-level data on modal splits and population(see Supplementary Material). We calculate that the additional cycling caused by the pop-up bike lane treatment during its first three months of operation has generated at least $580million in health benefits . The new infrastructure will generate $2.3 billion per year inhealth benefits, if the new bike lanes become permanent and if cycling habits are sticky.The magnitude of our estimate is large compared to previous evaluations of cycling in-frastructure improvements that have found statistically unclear or modest effects, typicallybecause of the limited scale of the interventions . Further research could investigate thenon-linearities in cycling adoption in terms of scale and timing of an infrastructure roll-out. It remains to be evaluated, if cycling behaviour is sticky and how similar treatmentsinfluence behaviour outside of the pandemic environment.Research based on surveys indicates that separated, protected infrastructure is a key ele-ment to incentivise up-take of cycling . Cities have experimented with a range of mea-sures to create new spaces for cycling, ranging from painted to provisionally protected bikelanes and from traffic calming with signs to built “modal filters” that only let bicycles andpedestrians pass. In our data, we do not see which share of increased bike counts is fromnew cyclists and which is from existing cyclists, who decide to cycle more often or farther.Large representative individual level samples, for instance based on transport mode detec-tion by smartphone sensors, may help to investigate changes in modal split at a sufficientlyhigh geographical resolution. GPS traces of individual trips could also help understand,how new infrastructure changes the route choices of cyclists and to measure the willing-ness to take detours for better infrastructure in terms of value of time. Contributions
S.K. ran the analyses. S.K. and N.K. designed the analysis, interpretedresults, designed figures and wrote the paper.
Declaration of interests
We declare no competing interests
Acknowledgements
We thank Ben Thies and Lennard Naumann for their excellent re-search assistance. We thank Jill Warren and Aleksander Buczy ´nski at the European Cy-clists’ Federation (ECF) for their data, that can also be browsed at their dashboard . Wethank numerous volunteers, that have contributed to this data collection effort. We thankEco-Counter for their technology allowing cities to share their cycling counts publicly. Wealso thank Ariel Ortiz-Bobea for sharing his code to produce Specification Charts. Note, that our sample does not include infrastructure built after the July, 8th and excludes a small numberof important cities, for which adequate open bike counter data is missing. ata and materials availability: Data and code used in our analysis are available at .7 eferences
1. Apple.
Mobility Trends Reports
NBERWorking Paper. Working Paper Series doi: . < > (2020) (July 2020).3. ECF.
COVID-19 Cycling Measures Tracker https://ecf.com/dashboard. Brussels, 2020.4. Ahern, A. et al. Analysis of National Travel Statistics in Europe: OPTIMISM WP2 : Harmonisation of National Travel Statistics inEurope. tech. rep. (Luxembourg, 2013). < http://dx.publications.europa.eu/10.2788/59474 > (2020).5. LEVA.
European Light Electric Vehicle (LEV) Market Shows Continued Growth https://leva-eu.com/2020/04/07/european-light-electric-vehicle-lev-market-shows-continued-growth/. Apr. 7, 2020. (2020).6. Moser, C., Blumer, Y. & Hille, S. L. E-Bike Trials’ Potential to Promote Sustained Changes in Car Owners Mobility Habits.
Environmental Research Letters
ISSN : 1748-9326 (Apr. 2018).7. Mattauch, L., Ridgway, M. & Creutzig, F. Happy or Liberal? Making Sense of Behavior in Transport Policy Design.
Trans-portation Research Part D: Transport and Environment
ISSN : 13619209 (June 2016).8. Larcom, S., Rauch, F. & Willems, T. The Benefits of Forced Experimentation: Striking Evidence from the London Under-ground Network.
The Quarterly Journal of Economics
ISSN : 0033-5533 (Nov. 1, 2017).9. Hamilton, T. L. & Wichman, C. J. Bicycle Infrastructure and Traffic Congestion: Evidence from DC’s Capital Bikeshare.
Journal of Environmental Economics and Management
ISSN : 0095-0696 (Jan. 1, 2018).10. Marqués, R., Hernández-Herrador, V., Calvo-Salazar, M. & García-Cebrián, J. A. How Infrastructure Can Promote Cyclingin Cities: Lessons from Seville.
Research in Transportation Economics. Bicycles and Cycleways
ISSN : 0739-8859 (Nov. 1,2015).11. Ashenfelter, O. Estimating the Effect of Training Programs on Earnings.
The Review of Economics and Statistics
ISSN :00346535 (Feb. 1978).12. Goodman-Bacon, A. Difference-in-Differences with Variation in Treatment Timing.
NBER Working Paper. doi:
10 . 3386 /w25018 . < > (2018) (Sept. 2018).13. Bezirksamt Friedrichshain-Kreuzberg.
Einrichtung von pandemieresilienter Infrastruktur in Form von temporären Radverkehrsan-lagen
Preventive Medicine
ISSN : 00917435 (Jan. 2018).15. Yang, L., Sahlqvist, S., McMinn, A., Griffin, S. J. & Ogilvie, D. Interventions to Promote Cycling: Systematic Review.
BMJ
ISSN : 0959-8138, 1468-5833. doi: . pmid: . < >(2020) (Oct. 18, 2010).16. Winters, M., Buehler, R. & Götschi, T. Policies to Promote Active Travel: Evidence from Reviews of the Literature.
CurrentEnvironmental Health Reports ISSN : 2196-5412 (Sept. 2017).17. Aldred, R. Built Environment Interventions to Increase Active Travel: A Critical Review and Discussion.
Current Environ-mental Health Reports ISSN : 2196-5412 (Dec. 1, 2019).18. Manaugh, K., Boisjoly, G. & El-Geneidy, A. Overcoming Barriers to Cycling: Understanding Frequency of Cycling in aUniversity Setting and the Factors Preventing Commuters from Cycling on a Regular Basis.
Transportation
ISSN :1572-9435 (July 1, 2017).19. Aldred, R., Elliott, B., Woodcock, J. & Goodman, A. Cycling Provision Separated from Motor Traffic: A Systematic ReviewExploring Whether Stated Preferences Vary by Gender and Age.
Transport Reviews
ISSN : 0144-1647 (Jan. 2, 2017).20. Hong, J., McArthur, D. P. & Livingston, M. The Evaluation of Large Cycling Infrastructure Investments in Glasgow UsingCrowdsourced Cycle Data.
Transportation.
ISSN : 1572-9435. doi: . < https://doi.org/10.1007/s11116-019-09988-4 > (2020) (Mar. 14, 2019).21. Kraus, S. & Koch, N.
Effect of Pop-up Bike Lanes on Cycling in European Cities (Code and Data) https://doi.org/10.5281/zenodo.4015974 > (2020). upplementary Material Materials and Methods
Bicycle count data
We assemble a new data set of daily bicycle counts from municipalbicycle counters. We connect to national and municipal open data portals for bike counterdata sets and connect directly to the API of those cities that use the Eco-Counter standard(see . We also obtain longer time series of bike counts going back to 2012 directly from theMayor’s staff for road planning and data in Paris.Our raw data set contains roughly a million daily counts starting in 2007. We drop thelower and upper percentiles from this raw sample since counters can record very low val-ues, when they are not functioning properly or very high values, when there is a cyclingevent that drives up counts. We drop the counter 100041252 from Bergen that varies be-tween very low values and some of the highest daily counts in the sample. Our resultsare robust to keeping these extreme values in the sample. The bulk of the bike counts arefrom most recent years (see Table S1) and we focus most of the comparisons made in ourregressions on the years 2019 and 2020. Figures S3 and S4. show the variation in weeklyaverage bike counts for cities in our study sample. Fig. S3. shows treated cities and Fig.S4. control cities. For certain cities, such as Paris and Berlin the raw data already indicatesthat increase in peak in June 2020 compared to June 2019. Many of the control cities showa similar pattern. Our regression analyses find a robust effect of new infrastructures, bothwhen taking the difference in these differences between treatment and control cities, butalso when focusing on variation in treatment timing exclusively.Table S1 shows summary statistics for the main variables included in our analysis. Theunit of observation in our analyses is the bike counter and counts vary daily. An averagecounter detects 1457 cyclists per day. The average number of counters per city is 22.9. Theaverage size of cities in our sample is 33000 ha. European cities tend to be denser thanAmerican cities. Thus, our study areas can be thought of small commuting zones ratherthan city cores.
Pop-up infrastructure data
We use project-level data on provisional infrastructure in Eu-ropean cities as a reaction to the COVID-19 pandemic collected by the European Cyclists’Federation . In the data we see the street, where the project is implemented, its size mea-sured in kilometres, the date of announcement, and the date of implementation. The dataalso contains the type of project. 80% are categorised as bike lanes and 16% as traffic calm-ing. Our data includes all projects recorded until 8 July 2020. We aggregate this data atthe city-day level to construct a variable of daily implemented kilometres of pop-up bikelane. We use the city definition and corresponding polygons from the European UrbanAudit 2020 . Typically areas defined by the European Urban Audit include suburbs. Forinstance, the Paris polygon includes many areas beyond the ring highway that surroundsthe municipality of Paris ("Ville de Paris"). This allows us to capture commuting enabledby new bike lanes from the suburbs into the city centre, which make up an important shareof projects (see Fig. S1). However, this also means that for infrastructure projects, which areconcentrated in one part of a city, such as in Berlin’s district of Friedrichshain-Kreuzberg,we tend to underestimate the effect.Our estimation sample contains 22 treated cities and 84 control cities, both of which someare dropped from our Poisson regressions depending on the specification because of a lack9f variation after removing fixed effects or because we do not have observations for ourcontrol variables. Fig. S2. shows the 20 treatment cities, for which the size of pop-upinfrastructure projects has been recorded in kilometre. We can see that Dublin and Berlinhave been the earliest adopters of pop-up bike lanes in the sample and Paris has beenthe city with the largest program. We use this variation in both timing and the extentof the implemented infrastructure to estimate our effects. We also include control citiesin the chart to illustrate the distribution of control cities across European countries. Wehave a large sample from both France and Germany. This allows us to estimate our effectbased on within-country variation removing time-varying factors related to the pandemicthat could create bias in our estimates. Note that, while important cities such as London,Milan, Lisbon and Rome had either announced or already implemented a pop-up bike laneprogram at the time of the analysis, they are missing from the sample due to insufficientspatial or temporal coverage of the bike count data. The average length by city of all bikeinfrastructures in our sample combined is 11.5 kilometres, the length of bike lanes is 8.2and the number of measures implemented 19.8.We check the sensitivity of our results to different specifications of the treatment, for in-stance as an indicator variable that is 1, if there is any cycling related infrastructure changein a city and 0 otherwise. The average effect of having any pop-up infrastructure treatmentin a city is 6% (column 4 in Table S3). The effect per individual measure taken by cities is0.4% (column 3). Our findings are robust, when we define treatment based exclusively onthose projects that are clearly marked as bike lanes in the data (column 2) rather than basedon all types of cycling and traffic calming measures combined (column 1). Mobility and weather controls
Our identification strategy relies on the use of differentcontrol groups that we expect to be on a common trend around individual daily cohorts ofpop-up infrastructure projects. As a baseline we remove and therefore control for time-invariant differences between cities and the locations of the individual counters in ourdata. Therefore any additional time-invariant control variables would be redundant inour analysis. We also use fixed effect interacting different spatial levels with time dimen-sions, thereby controlling for many time-varying factors. We use additional data that variesat a high spatial and temporal resolution to rule out any bias that may be introduced bytime-varying factors below our fixed effect levels.We use weather data from the ERA5 climate model, that provides hourly reanalysis mea-sures of surface temperature, UV radiation, precipitation and wind at a 0.25 ◦ × ◦ reso-lution . We use the ecwmfr package to aggregate this to the EU Urban Audit city polygonsat the daily level. We capture average human mobility throughout the phase of the COVID-19 pandemic starting in March with a human mobility index based on Facebook data . Theindex is from a data set called "movement range maps" that Facebook shares after aggregat-ing individual user movements for humanitarian and research purposes with a referenceto the principles outlined by epidemiologists and public health researchers . It measuresthe number of daily 600 meter grid cells visited by Facebook users compared to a baselinein February. For most of our sample the index is aggregated to the state-level, where weuse the data. Table S1 shows that on average in our sample period daily mobility has beenbelow the February baseline. Regression model
We model the relationship between cycling traffic and the pop-up bikelane treatment as: 10og Bike Count icd = β Bike Lane (km) cd + X id + λ i + σ cw + ϕ nd + ε id (1)where i indexes a counter, c indexes a city, n indexes a country, d indexes a day, and w indexes as week. λ i is a counter fixed effect that controls for time invariant factors at a high spatial resolution. σ cw is a city-week fixed effect that controls for week-specific time-varying factors effectivelyrestricting identifying variation to days before and after treatment within the same week inthe same city. ϕ nd is a country-day fixed effect that captures any daily changes common toall cities in a country.The coefficient of interest is β . It captures the effect of the pop-up bike lane treatment onaverage bicycle counts in a city. Our baseline treatment variable is defined as the numberof kilometre of pop-up bike lanes implemented on a given day. Multiplied by a 100 theestimate can be interpreted as the change in bicycle count for a unit change in the treatmentvariable. X id is a vector of control variables including the mobility index based on Facebook data,weather variables (temperature, UV radiation, wind, precipitation) and the number ofcounters per city.We use Poisson pseudo-maximum likelihood regressions (PPML) to estimate this model .As a robustness check we also use ordinary least squares (OLS) with the natural logarithmof the bicycle count as the outcome (see Fig. 1). We cluster standard errors at the city-level,where treatment is assigned . Calculating the health benefits of the policy
We calculate the health benefits by com-bining our estimates of cycling increases for each kilometre of pop-up bike lane with anestimate of the average health benefits of a kilometre cycled ($0.45 converted from 0.62Australian Dollars) , which is lower than typical values from the grey literature . Ourregression estimates only provide us with a percentage increase in cycling (0.6%) per kilo-metre of bike lane. We convert this result into additional kilometres cycled in a city basedon baseline values of kilometres cycled per person in a city from a detailed transport be-haviour survey in 135 German cities . We impute values of kilometres cycled for otherEuropean cities based on information on the modal split (trips) of commutes and a city’spopulation both taken from the European Urban Audit .Our estimate only counts benefits from cycling but not the saved costs of a potential modalshift from car use to cycling, that we cannot measure with our data. It also does not takeinto account shifts from walking or jogging and cycling for exercise to cycling in the city,where counters in our sample are typically placed. Since the external costs of car use arehigh , we interpret our calculation as a lower bound.11 ean Std. Dev. 25% 50% 75% 95% Min. Max.Daily number of cyclists 1457.2 1895.7 255 744 1923 5151 1 13339City size (ha) 32893.8 42393.6 14163.3 22018.5 40659.9 89180.2 455.6 251517Year 2017 2 2016 2018 2019 2020 2007 2020Number of counters in the same city 22.9 23.2 4 14 32 82 1 90Facebook mobility index -0.16 0.21 -0.27 -0.11 -0.0044 0.088 -0.81 0.51Observations 995818 Table S1. Summary statistics at the counter-day level The unit of observation of our analysis is the counter and data variesdaily. Count data is from municipal bike counters and is obtained from different APIs. Treatment and control variables areassigned to counters based on their city attribute. City definitions are from the EU Urban Audit. The Facebook mobilityindex is only available from March 2020. It measures aggregate movement activity by Facebook users in a givenadministrative area (districts or states). ean Std. Dev. 25% 50% 75% 95% Max.Total length of bike infrastructures 11.5 20.0 1.39 2.57 16.6 57.9 85.1Total lenght of bike lanes 8.24 18.3 0.24 2.05 7.35 24.8 84.3Number of measures 19.8 48.1 1 4 17 52 226Observations 22 Table S2. Summary statistics of most recent state of infrastructure at the city level show We use data from the EuropeanCyclists’ Federation. The raw data includes information on individual infrastructure projects announced or implemented.We aggregate it to the city-day level using city definitions from the EU Urban Audit. Our analysis includes data up to the 8July 2020. The newest data can be found at: https://datastudio.google.com/u/0/reporting/ba90a08c-9841-4beb-9e26-7d4f7d002709/page/yMRTB utcome: Cyclist count(1) (2) (3) (4)All km Bike lane km Num of measures Any treatmentPop-up treatment 0 · ∗∗ · ∗∗ · ∗ · ∗ ( · ) ( · ) ( · ) ( · ) City clusters 78 78 78 78N 59904 59904 59904 59904
Table S3. Different treatment specifications Each column shows the effect of treatment with pop-up infrastructure on a city’scycling count compiled from city APIs. The data on daily pop-up bike lane additions are from the European Cyclists’Federation . The unit of observation is the cycling counter. Time variation is daily. Coefficients are from Poisson regressions.Column (1) shows the effect of a kilometre of any bike infrastructure, (2) shows the effect of bike lanes, (3) the effect of anysingle measure in a city, and (4) the overall treatment of an implemented pop-up infrastructure program in a city. Allregressions include counter and day fixed effects and controls for overall mobility (measured with Facebook usermovements), weather (temperature, wind, sunshine, precipitation), and the number of counters in a city. We cluster standarderrors at the city level, where treatment is assigned. Significance levels are ∗ p < ∗∗ p < ∗∗∗ p < ig. S1. Pop-up bike lanes and bicycle counters in Paris The map shows pop-up bike lanes implemented in Paris up to 3 July2020 (green lines) and the location of bike counters (dots) in our data set. The detailed infrastructure data has been collectedby a consortium of French NGOs and researchers. It is available at: https://carto.parlons-velo.fr/ ig. S2. Intensity of pop-up bike lane treatment over time in treatment cities and control cities This Figure shows treatedcities and their treatment intensities in implemented kilometres (colouring on a log scale) of public bike lanes at a given daybetween March and July 2020. Control cities are plotted in white. London, Milan, Lisbon and Rome are missing from thesample due to insufficient spatial or temporal coverage of the data. Information on individual pop-up bike lanes with theirstreet location, announcement implementation is from the European Cyclists’ Federation. The newest data can be found at: https://ecf.com/dashboard ig. S3. Average bike count per week in treated cities Daily bike counts are aggregated by city and averaged over the week.Bike counts are assembled from municipal open data feeds. The lower and upper percentiles from the initial sample (treatedand control cities combined) are removed from the sample. Only measurements from 2019 and 2020 are shown. Citydefinitions are chosen according to EU Urban Audit. ig. S4. Average bike count per week in control cities Daily bike counts are aggregated by city and averaged over the week.Bike counts are assembled from municipal open data feeds. The lower and upper percentiles from the initial sample (treatedand control cities combined) are removed from the sample. Only measurements from 2019 and 2020 are shown. Citydefinitions are chosen according to EU Urban Audit. eferences
1. ECF.
COVID-19 Cycling Measures Tracker https://ecf.com/dashboard. Brussels, 2020.2. Eurostat.
Urban Audit Geodata (GISCO) https://ec.europa.eu/eurostat/web/gisco/geodata/reference-data/administrative-units-statistical-units/urban-audit > (2020).3. Hersbach, H. et al.
The ERA5 Global Reanalysis.
Quarterly Journal of the Royal Meteorological Society, qj.3803.
ISSN : 0035-9009,1477-870X (June 15, 2020).4. Hufkens, K., Stauffer, R. & Campitelli, E.
The ecwmfr package: an interface to ECMWF API endpoints . < https://bluegreen-labs.github.io/ecmwfr/ >.5. Facebook.
Movement Range Maps https://data.humdata.org/dataset/movement-range-maps. 2020. (2020).6. Buckee, C. O. et al.
Aggregated Mobility Data Could Help Fight COVID-19.
Science
ISSN : 0036-8075, 1095-9203(Apr. 10, 2020).7. Correia, S., Guimarães, P. & Zylkin, T. Fast Poisson Estimation with High-Dimensional Fixed Effects.
The Stata Journal
ISSN : 1536-867X (Mar. 1, 2020).8. Abadie, A., Athey, S., Imbens, G. W. & Wooldridge, J. When Should You Adjust Standard Errors for Clustering?
NBERWorking Paper. Working Paper Series doi: . < > (2020) (Nov. 2017).9. Zapata-Diomedi, B., Gunn, L., Giles-Corti, B., Shiell, A. & Lennert Veerman, J. A Method for the Inclusion of PhysicalActivity-Related Health Benefits in Cost-Benefit Analysis of Built Environment Initiatives.
Preventive Medicine
ISSN : 00917435 (Jan. 2018).10. Hubrich, S., Ließke, F., Wittwer, R., Wittig, S. & Gerike, R.
Methodenbericht Zum Forschungsprojekt - Mobilität in Städten–SrV2018
Transport - Cities and Greater Cities (Urb_ctran) https://appsso.eurostat.ec.europa.eu/nui/show.do?dataset=urb_ctran&lang=en > (2020).12. Eurostat.
Population on 1 January by Age Groups and Sex - Cities and Greater Cities (Urb_cpop1) https://appsso.eurostat.ec.europa.eu/nui/show.do?dataset=urb_cpop1&lang=en > (2020).13. Gössling, S., Choi, A., Dekker, K. & Metzler, D. The Social Cost of Automobility, Cycling and Walking in the EuropeanUnion.
Ecological Economics
ISSN : 09218009 (Apr. 2019).: 09218009 (Apr. 2019).