Policy Research Working Paper 10393 The Unintended Consequences of Curfews on Road Safety Guadalupe Bedoya Amy Dolinger Caitlin Dolkart Arianna Legovini Sveta Milusheva Robert Marty Peter Taniform Development Economics Development Impact Evaluation Group April 2023 Policy Research Working Paper 10393 Abstract During COVID-19, curfews spread like wildfire. Although the curfew hours when cars are off the road, but that these their impact on curbing the spread of disease remains to reductions can be fully offset by an increase in crashes be proven, curfews have the potential to bring about costs during the hours before the curfew when people rush to to society in multiple domains. This paper investigates the get home. These findings forewarn that the use of curfews in impact of curfews on road safety in an urban setting. It future crises and pandemics should be carefully scrutinized shows that they lead to large reductions in crashes during and designed to minimize unintended negative effects. This paper is a product of the Development Impact Evaluation Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at smilusheva@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team The Unintended Consequences of Curfews on Road Safety* Guadalupe Bedoya†1 , Amy Dolinger‡1 , Caitlin Dolkart§2 , Arianna Legovini¶1 , Sveta Milusheva||1 , Robert Marty** 1 and Peter Taniform††1 1 The World Bank Group 2 Flare JEL Codes: R41, R42, I18 Keywords: Congestion Externalities, Curfew, Road Safety, Pandemic, Vehicle Crash * The authors are grateful for the long-term collaboration with transport authorities in Kenya, including Kenya’s National Police Service (NPS) and the Kenya Urban Roads Authority (KURA), which made this work possible. We thank the NPS, KURA and Flare for sharing data used in this paper. Kelvin Gakuo provided excellent data analysis and support, and Christine Okeyo provided excellent field support on the ground. We are grateful to the anonymous reviewers for their useful feedback. The research benefited from financial contributions from the United Kingdom Foreign, Commonwealth and Development Office (FCDO), the European Union (EU), the World Bank’s Umbrella Facility for Impact Evaluation (i2i) and the World Bank’s Knowledge for Change Program (KCP). The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations or those of the Executive Directors of the World Bank or the governments they represent. The list of authors is in alphabetical order. Computational reproducibility verified by DIME Analytics. † gbedoya@worldbank.org ‡ adolinger@worldbank.org § caitlin@rescue.co ¶ alegovini@worldbank.org || smilusheva@worldbank.org ** rmarty@worldbank.org †† ptaniform@worldbank.org 1 Introduction Historically, curfews have been used to manage social unrest and control crime, looting and violence. Often, they have been used to contain specific populations, especially young people. Their effectiveness in achieving public safety goals remains ambiguous, with some research finding decreases in crime, while others find curfews lead to increases in gun violence (Carr & Doleac, 2018; Kline, 2012). Curfews can have significant direct costs on freedom of movement, assembly and expression, especially when enforcement resorts to violence (Ravindran & Shah, 2023; Katana et al., 2021). Their indirect costs are even less well understood. The use of curfews must thus be carefully scrutinized and their effects, especially unintended ones, must be rigorously measured. During COVID, more than 100 countries resorted to lockdowns and curfews to manage the spread of the virus. However, their justification did not have an empirical or theoretical underpinning. First, was the use of curfews during the onset of the pandemic justified? Did curfews reduce assembly or simply displace and concentrate assembly from night to daytime, potentially increasing transmission? Curfews appear to have had little impact on slowing the pandemic (de Haas et al., 2022; Huber & Langen, 2020) and may have even increased the spread of disease as individuals shifted activities to pre-curfew hours, increasing contact-density during these times (Sprengholz et al., 2021). Second, what were the effects of curfews on mobility and what was the effect of that change on other economic outcomes? For instance, did curfews reduce the volume of cars and road crashes, or did they lead to an increase in speeds and car crashes? Providing evidence on these questions is important for preventing the use of policy instruments in future pandemics and crises that are inherently costly as well as ineffective. In this study, we shed some light on the question of whether curfews affect road safety. This is important because the cost of road safety is measured in people’s lives just as is done in pandemics. The context we use is Nairobi, a capital in a developing country with high rates of car crashes, and where we have made a multiyear investment in data infrastructure for urban management that turned a data-poor into a relatively data-rich environment. A small pre-COVID literature looks at curfews used in various contexts to curb night driving by inexperienced drivers. This is effective in reducing crashes (Williams & Preusser, 1997). Even when inexperienced drivers fully substitute night with day driving, crashes are reduced (Bolduc et al., 2013). Here we investigate what happens when these restrictions are extended to all drivers. The question we ask is whether the reduction of cars on the road me- chanically decreases the number of car crashes, or whether the curfew displaces driving, increases speeds and has a rebound effect on car crashes. In our context, the dusk-to-dawn curfew was implemented in March 2020 purportedly to prevent the spread of COVID-19. We use a difference-in-differences strategy, harnessing detailed high-frequency data on road traffic crashes, to look at crashes before and after the curfew implementation and relative to the same period a year earlier. In so doing, we study the impact of curfews on road traffic crashes. We also examine negative externalities during 2 non-curfew hours and the incidence of spillovers in the hours right before the curfew. From the literature we know that the probability of a crash depends on the density of cars on the road (Edlin & Karaca-Mandic, 2006). We also expect driver behavior, especially speed, to affect the likelihood of a crash (Bauern- schuster & Rekers, 2022). Speed, in turn, is a function of car density and the cost of time for drivers. The higher the cost of time, the faster drivers will drive to reach their destination in a shorter amount of time. Therefore, crashes are a function of density, D, and speed, S , Crash = f (D, S (D, C )), where speed is a function of density and cost, C . We expect curfews to impact both parameters D and C . During the hours of the curfew, we expect few cars and people on the road; therefore, D will approximate zero. Simultaneously, as density falls, lack of congestion induces increased speeds, which in turn can increase the likelihood of a crash and its severity. Therefore, the effect of the curfew on road safety during the curfew hours is ambiguous. During the hours right before the curfew, density is expected to increase as people must complete their work, chores or other activities and get home before the curfew. The cost of time is also affected. Right before the curfew, time is scarce and more valuable. Speeds might increase as a result. As density and the cost of time rise, the effects of the curfew in the hours before the curfew are also ambiguous. Density will reduce speed but people will respond to the increased cost of time by weaving through traffic to get to their destinations on time. Our results show that the curfew leads to large reductions in crashes during the curfew hours, when cars are off the road. The spillover effects are large and significant. These reductions are fully offset by an increase in crashes in the hours right before the curfew. Importantly, analyzing the change in crashes as compared to the change in vehicles on the road, the probability of a crash increases both during the curfew period and during the hours right before the curfew. We look at a second round of curfews to understand whether these spillovers are inevitable. We find that the spillovers can be managed by delaying the onset of the curfew. When the government revised the curfew from 19:00 to 21:00, crashes were reduced in the night hours without generating offsetting increases in the hours prior to the curfew. There is also no longer an increase in the probability of a crash. As curfew hours become less binding, spillovers and externalities are reduced. Our contribution to the literature is twofold. First, the literature on congestion focuses on pricing policies (Green et al., 2016; Tang & van Ommeren, 2022). We add to the literature on the externalities of congestion on road safety by evaluating a policy that affects congestion indirectly. We also extend the analysis of road safety beyond high-income unnings & Schiele, 2021; Edlin & Karaca-Mandic, 2006; Bauernschuster & Rekers, 2022; Van Benthem, contexts (B¨ 2015). In those settings, good infrastructure and traffic enforcement may explain the decrease in crashes associated with congestion pricing or speeding policies. In lower and middle-income countries, with uneven speed enforcement and poorer infrastructure, policies that reduce congestion, such as congestion pricing or curfews, could potentially increase crashes via the speed channel. Second, we demonstrate an additional important negative externality of curfew 3 policies. Rigorous research in this space has been limited, with existing papers demonstrating curfews leading to increases in gun violence (Carr & Doleac, 2018) and negative mental health consequences (Altindag et al., 2022). We add to this literature on the externalities of curfews by showing that not only are curfews ineffective in curbing disease transmission but worse, they create large negative spillovers on road safety, increasing the probability of a crash during non-curfew hours. The remainder of the paper is structured as follows. Section 2 describes the background on road safety and the policies implemented in Kenya. Section 3 describes the data and empirical strategy. In Section 4, we present our results and robustness checks. Conclusions are presented in Section 5. 2 Background Road traffic crashes (RTCs) are the leading cause of death for children and young adults age 5-29; 20 million to 50 million people suffer non-fatal injuries each year; therefore, any policies that affect road safety can have important health consequences (WHO, 2018). Kenya’s rate of estimated fatalities per 100,000 is 27.8, higher than the already highest regional average of 26.6 fatalities per 100,000 for Africa (WHO, 2018). Policies that affect road safety, whether intentionally or unintentionally, can therefore have important consequences for human life and public health. The lockdowns and other mobility policies widely implemented across the world to curb COVID-19 transmission have also significantly affected road traffic crashes. NSC (2021) published a report detailing the rise in fatal road crashes in the United States in 2021 after a year of lockdowns, and Lin et al. (2021) found road crashes decreased for all age groups, races, and genders in Los Angeles and New York City in 2021. As with most other countries, Kenya implemented a number of measures to limit mobility. On March 15, 2020, two days after its first confirmed case, Kenya closed schools, restricted travel, and enacted work from home measures, among other policies (Ministry of Health, 2020). On March 22nd, the Kenyan government announced that bars would close and that restaurants would remain open only for take-out (Kagwe, 2020). Shortly after, on March 25th, the government announced that a dusk-to-dawn (19:00-5:00) curfew would take effect on March 27th (OSAC, 2020a). The curfew was strictly enforced, with reports indicating that Kenyan Security forces detained those out after the curfew (Bearak & Ombuor, n.d.). On June 7th, the government eased some movement restrictions and revised the curfew hours to 21:00-4:00 (OSAC, 2020b). This curfew continued to remain in place until November 4, 2020, when curfew hours were reduced to 22:00-4:00. 4 3 Empirical Framework 3.1 Data There are three main types of data that we use in this paper. First, to study road traffic crashes (RTCs), we use data from the National Police Service that we digitized from paper situation reports. Police officers generate these reports for each crash that they record where there was an injury, fatality, or significant property damage. Second, we use data from Flare, the largest emergency response dispatch platform in Kenya with over 800 ambulances available on call, 200 of which are in Nairobi. Finally, we use several big data sources to study mobility and congestion including video-generated traffic counts, Waze, Uber and Google Maps. Our main source of data on road traffic crashes comes from the Kenya National Police Service (NPS). Detailed data on the time and location of RTCs did not exist for Kenya; therefore, we worked with NPS to digitize all the Situation Reports for the 14 traffic police stations in Nairobi. These are paper records that are produced for each crash that involves an injury or notable property damage. We use 5,075 crash reports from January 1, 2017-December 31, 2020 that we have manually digitized. While this data does not include minor crashes reported to the police or crashes that were not reported to the police, it is the most comprehensive set of data on the time and location of crashes for the city of Nairobi. We aggregate total crashes occurring per hour for Nairobi. We cross-reference this dataset with data on road traffic crashes from Flare, an emergency response dispatch platform that aggregates emergency response fleets into a single platform and runs a 24/7 dispatch center much like 911. The majority of emergency calls made to Flare are for RTCs. Of the calls coming in, almost half come from good Samaritans who witness a situation requiring an ambulance. The other half of the calls are evenly split between calls coming from the police who show up on the scene of an emergency and calls coming from those who are part of the membership service that Flare offers.1 Similarly to the data from the police, we aggregate data on the number of RTCs recorded by Flare per hour for January 1, 2019-December 31, 2020. We rely on four data sources to measure changes in vehicle density and speed. We use traffic jam information from the Waze Connected Cities Program (Waze, 2021). Waze provides information on the location, speed, and delay time of traffic jams; data is updated at 2 minute intervals. We compute the total traffic delay time due to traffic jams in Nairobi hourly. We use this delay time of traffic jams as a proxy for congestion on the roads. We use data on number of vehicles on the road measured from cameras installed in six major intersections in Nairobi by the Kenya Urban Roads Authority. Data for four of the intersections is available for January and April 2020, for one of the intersections data is available for February and April in 2019 and 2020 and July 2020, and for one intersection data is available for January-April 2020. We use the relationship between vehicles on the road and 1 Businesses can sign up through Flare’s membership service, rescue.co, to have their employees or customers receive access to Flare services free of charge. This service has been especially popular with ride sharing companies. 5 congestion as measured by Waze to predict vehicles on the road during the months when we do not have data for some of the intersections.2 In a context like Nairobi, where there is not a good measure of vehicles on the road or vehicle miles traveled, this high-frequency data provides one of the best measures for how the number of vehicles traveling throughout the day might be changing during this period.3 Analyzing the change in vehicle counts after the curfew is implemented across the different intersections demonstrates very similar magnitudes of change. This makes us more confident that the changes we measure in these intersections are likely representative of vehicles on the road more broadly in the city. We generate an index for each intersection for vehicle counts using vehicle counts in the pre-curfew period of 2020 by day of week as the baseline. We use an average of this index across intersections as a proxy for the change in total vehicles counts for the city. To measure speed, we query real-time travel times from Google Maps. We query travel times along segments of major corridors in Nairobi, including Mombasa Road, Uhuru Highway and Waiyaki Way, that we selected due to their importance for travel within Nairobi. We obtain information on the time it takes to travel from the start to the end of the segment using a vehicle given current traffic conditions. The speed is calculated by dividing the distance of the segment by the estimated travel time. We query this information every 30 minutes. We verify these measures using speed information from Uber Movement data. Uber provides speed data aggregated from Uber trips at the road segment level across Nairobi. Average speed is available hourly, and data is made available through March 2020, after which time Uber stopped producing these statistics. Given data is not available beyond this time, the Uber data is only used for cross-validating speed data from Google and is not used for the analysis. When working with data derived from smartphones, it is important to consider the potential for biased measures (Milusheva et al., 2021). The traffic and congestion data from Waze and Google Maps are likely to be minimally affected by bias arising from smartphone use because they collect measures of how quickly cars are able to travel based on the traffic on the road, which would be determined by both smartphone and non-smartphone owners on the road. Nevertheless, the data may have quality issues when few smartphone owners are on the road. By combining and comparing datasets, we can triangulate our measures to ensure consistency. 3.2 Compliance The curfew plausibly affects road traffic crashes by changing the number of vehicles on the road. Additionally, driving speeds may change, either because the lower density of vehicles enables higher speeds or because the cost of being caught breaking the curfew leads people to drive faster in order to arrive at their destination before the curfew. If the curfew were not to have any impact on vehicle density or speeds, there would be no expected impact on road safety. Therefore, we first examine whether the curfew affected mobility and therefore vehicles and speeds. 2 See Appendix B for a description of how these predictions were generated. 3 Forexample, while there are data on vehicle registrations, data is not collected on vehicles that are no longer in commission, which hinders an accurate picture of the volume of active vehicles. 6 There is a large decrease in the number of vehicles on the road based on the index for predicted vehicles on the road (Figure 1a). On average the drop in vehicles during the curfew hours is 88%. During the daytime hours, the decrease is 31%. During the pre-curfew hours, the drop is smaller than the drop seen during the daytime hours—only 25%. As expected, the decrease in vehicles on the road allows drivers to increase the speed at which they drive, particularly during the daytime and early evening hours (Figure 1b). The increase is especially pronounced for weekdays when traffic is usually extremely heavy prior to the COVID-19 policies.4 Figure 1: Change in Number of Vehicles and Speed in 2020 3.3 Identification and Estimation Our empirical strategy is a modified difference-in-differences specification, using the previous year as our comparison group. We compare crashes in the period before any COVID-19 policies were implemented (January-March 15) to the period after the first curfew was implemented (March 27-June 6) and look at the difference between these two periods in 2019 and in 2020. We use 2019 as the comparison given that it is the year closest to our year of interest and 4 Note that during the curfew hours, there seems to be a decrease in speed as measured by Google Maps. The number of vehicles on the road at this time drops to close to 0; therefore, the measured decrease in speed may be a function of how google calculates driving time in the absence of vehicles. To confirm the data, we compare the percent change in speeds from Google with the percent change in speeds from Uber. While the Uber and Google data are fully aligned for the daytime hours and the hours right before the curfew, there is a deviation in the percent change between Uber and the Google speed data for the curfew hours. Therefore, we do not consider these hours when conducting any analysis related to speed (Appendix Figure A.1). 7 therefore most comparable in terms of road infrastructure and policies. We conduct a robustness check where we also include 2017 and 2018 as additional comparison years. Figure 2a shows the comparison of crashes between 2019 and 2020, before and after the curfew policy.5 Our main specification is: yt = β0 + β1 yeart + β2 pt + β3 yeart × pt + Xt γ + δt + t (1) where yt is the outcome of interest, either number of crashes or a proxy for rate of crashes on day t. year is either 2019 or 2020, p represents the period before the COVID-19 policies (Jan 1-March 15) or after the curfew policy (March 27-June 6), Xt includes weather variables and indicator variables for the day of week, δt represents week fixed effects and t is the error term. We run this analysis focused on crashes during curfew hours (20:00-5:00) as well as during the hours right before the curfew to test for spillovers (17:00-20:00). In the absence of the curfew, mobility and the number of cars on the road would still be lower due to the policies that closed down restaurants, bars and schools, and mandated work from home for government officials (Figure 1). In order to try to difference out the effect of lower mobility due to other policies, we use the 5:00 to 17:00 hours as a comparison. This time interval may have been affected by the curfew as well if individuals shift trips they would have performed in the evening to the daytime, but this would bias us towards finding smaller spillover effects during the pre-curfew hours. We run a triple difference specification where we interact the year and period variables with a variable for the hourly interval. Crashes for these different hour intervals are also visualized in Figure 2 at the weekly level. We estimate Equation 2 where we interact the curfew period and the year with the hour interval, hti , with the comparison being 5:00-17:00 and the two periods of interest are the pre-curfew hours of 17:00-20:00 and the curfew hours of 20:00-5:00. The outcome of interest yti is the number of crashes on day t during hour interval i. yti = β0 + β1 yeart + β2 pt + β3 hti + β4 yt × hti + β5 pt × hti + β6 yeart × pt + β7 yeart × pt × hti + Xt γ + δt + t (2) To further estimate externalities of the curfew and the change in vehicle density, we examine whether the likelihood of a crash changed for those still using the road. Specifically, we examine whether the curfew caused a change in the crash rate. Changes in crash rates have been observed in response to other policies that impact mobility, where some researchers have found a U-shaped association between traffic and crash rates—with the highest crash rates occurring at the lowest and highest levels of congestion (Green et al., 2016). Changes in the rate of crashes are an important indication of the additional externality of vehicle density and driver behaviors for other users as an additional car on the road or more aggressive behavior could lead to an increase in the likelihood of a crash that goes beyond the proportional increase (Edlin & Karaca-Mandic, 2006). We use the index for vehicles on the road during each time 5 Wedo not include the days from March 16-March 26 in the analysis because some policies had been implemented to limit mobility but the curfew was not yet implemented. 8 Figure 2: Weekly Crashes in 2019 and 2020 interval as a proxy for changes in total vehicles on the road in Nairobi. We divide crashes by the index to measure the proportional change in crashes. We use bootstrap standard errors given the estimated nature of the outcome variable. 4 Results We first look at the overall impact of the policies put in place to limit mobility on daily crashes (Table 1). During the initial curfew period, there is a significant overall decrease of 1.04 crashes per day (Column 1). This is over a mean of 3.27 crashes per day, which represents a 32% decrease. This effect includes the impact of the curfew but also other policies that limited mobility such as the closing of schools, bars and restaurants, and the work from home mandate for government workers and others. Given the count nature of our outcome of interest, we also estimate both Poisson and negative binomial models of crashes (Columns 2 and 3). The results similarly show large and significant reductions in crashes of around 34%, which are comparable to the magnitude from the OLS estimate. We reject the null of no over-dispersion of the dependent variable (p=.057); therefore, the model is more correctly estimated using the negative binomial compared to the Poisson. Similar to Green et al. (2016), we compare the mean squared residuals for both the negative binomial and the linear specification to determine whether to continue with the negative binomial. We find the results are very similar, but slightly better for the OLS specification; therefore, we continue with the OLS specification for the rest of the paper. 9 We look specifically at the curfew hours, starting from 20:006 until 5:00. Crashes decrease by 1.17 crashes per day during the curfew (Table 1 Column 4), which is around a 100% reduction in crashes. This reduction aligns with the raw data in Figure 2b, which showed weekly crashes going almost down to zero during the curfew period. With almost no cars on the road, this large reduction in crashes is to be expected. Overall, COVID policies led to a decrease in crashes, particularly during curfew hours; however, curfews can cause spillovers—affecting the number of crashes during non-curfew hours. To examine spillovers, we estimate the impacts of the curfew for each hour of the day. Figure 3 shows regressions comparing crashes before the curfew or any other mobility policies were implemented (January 1-March 15) with crashes after the first curfew was implemented (March 27-June 6), and compares crashes in those time periods between 2020 and 2019. Crashes increase in the hours right before and at the start of the curfew, while crashes decrease in the hours after the curfew starts and during the entire time of the curfew. Table 1: Impact of Mobility Reducing Policies on Daily Crashes OLS Poisson NegBi Only Curfew Hours OLS (1) (2) (3) (4) Curfew Period 1.888 0.566* 0.565* 0.947* (1.340) (0.335) (0.336) (0.519) Year 2020 1.397*** 0.442*** 0.445*** 0.585*** (0.484) (0.142) (0.142) (0.194) Curfew x 2020 -1.041** -0.340** -0.348** -1.166*** (0.480) (0.141) (0.141) (0.221) Constant -149.844*** -46.290*** -46.784*** -31.433 (52.882) (15.033) (15.076) (25.185) Observations 295 295 295 295 R2 0.172 . . 0.198 Notes. Robust standard errors are reported in parentheses. *** (**) (*) denotes significance at 1% (5%) (10%) level. Regressions include fixed effects for week and day of the week, and con- trols for weather (temperature, precipitation and two measures of wind). Given that hourly crash rates are noisy, we group the hours right before the curfew time, 17:00-20:00.7 Column 1 of Table 2 again shows the results for the curfew period. There is a significant positive increase in crashes during the three hours leading up to and at the beginning of the curfew (Column 2). This result indicates a large negative spillover of the curfew in the form of more crashes occurring right before the curfew hours. The effect measured in Table 2, columns 1 and 2, results from both the curfew and other mobility restricting poli- cies (e.g., work from home policies). To isolate the impact of the curfew—irrespective of other mobility restricting 6 While the curfew began at 19:00, we allow for a period of delay in the reporting of crashes by police that may lead to crashes happening right before the curfew being incorrectly attributed to the curfew period and exclude 19:00-20:00 from the curfew time. 7 While the curfew started at 19:00, we include the 19:00-20:00 hour in the pre-curfew analysis given the likely delay in the reporting of crashes by police. 10 Figure 3: Coefficients from Regression for Each Hour policies—we compare our results to changes in crashes that occurred during the daytime hours of 5:00-17:00. When analyzing only daytime hours, there is a negative coefficient of -0.465, though it is not significant (Table 2 Column 3). We use a triple difference specification to compare crashes in the pre-curfew hours and during the curfew hours to crashes during the daytime (again comparing 2019 to 2020 and comparing the period before the curfew was imple- mented and after the curfew was implemented). The coefficient on the curfew hours is still negative and significant, but it is now smaller at only -0.67 (Table 2 Column 4). On the other hand, the coefficient on the pre-curfew hours is much larger, at 0.88. Thus, controlling for the other policies that also affected mobility throughout the day, the negative spillovers during the pre-curfew hours fully offset any positive gains for road safety during the curfew hours.8 8 As discussed in the methods section, it is possible the daytime mobility was also affected by the curfew in the form of more movement happening during the daytime hours than otherwise would have happened during the curfew hours in the absence of a curfew. If this were the case, then in the absence of a curfew but with all the other policies in place, there would have been even fewer vehicles on the road during the daytime, which would mean even fewer crashes would have happened during this time interval if there were no curfew. In that case, the current estimates are biased so that the spillovers are biased downward and the effect during the curfew period is biased upward. 11 Table 2: Regressions for Different Daily Time Intervals Curfew Hours Pre-Curfew Hours Daytime Triple (20:00-5:00) (17:00-20:00) (5:00-17:00) Difference (1) (2) (3) (4) Curfew Period 0.947* -0.319 1.161 0.587 (0.519) (0.456) (1.085) (0.491) Year 2020 0.585*** -0.040 0.447 0.309 (0.194) (0.173) (0.328) (0.226) Curfew x 2020 -1.166*** 0.504** -0.465 -0.418 (0.221) (0.221) (0.322) (0.315) Nighttime Hours -0.471** (0.226) Pre-curfew Hours -0.892*** (0.183) Curfew x Nighttime 0.168 (0.282) Curfew x Pre-Curfew -0.001 (0.248) 2020 x Nighttime -0.073 (0.266) 2020 x Pre-Curfew -0.186 (0.248) Curfew x 2020 x Nighttime -0.667* (0.391) Curfew x 2020 x Pre-Curfew 0.876** (0.387) Constant -31.433 1.661 -38.352 -5.798 (25.185) (12.905) (36.521) (11.186) Observations 295 295 295 885 R2 0.198 0.196 0.135 0.176 Notes. Robust standard errors are reported in parentheses. *** (**) (*) denotes significance at 1% (5%) (10%) level. Regressions include fixed effects for week and day of the week, and controls for weather (temperature, precipitation and two measures of wind). 4.1 Rate The estimates in the previous two sections demonstrate a decrease in the number of crashes during the curfew hours, but the decrease is offset by spillovers during the hours right before the curfew begins. Especially for the curfew period, these results still do not speak to the externality of congestion since the decreases seen may be proportional to the decreases in vehicles. Focusing on the probability of a crash during curfew hours, we find that the curfew led to a significant increase in the crash rate despite the decrease in crash numbers (Table 3). We also find a large and significant increase in the probability of a crash during pre-curfew hours. This result is not surprising as the pre-curfew hours experienced a decrease in congestion (indicating a decrease in cars on the road), and an increase in crashes. Therefore, the curfew not only led to a spillover of crashes to pre-curfew hours, but the curfew also led to an additional externality of an increase in the probability of being in a crash during both curfew hours and the hours right before the curfew. There is no change in the crash rate during the daytime hours (Column 3). 12 Table 3: Rate of Crashes Curfew Hours Pre-Curfew Hours Daytime (20:00-5:00) (17:00-20:00) (5:00-17:00) (1) (2) (3) Curfew Period 5.630 -0.770 0.388 (3.734) (0.676) (1.193) Year 2020 -0.242 -0.172 0.133 (0.584) (0.201) (0.386) Curfew x 2020 1.787** 0.962*** 0.203 (0.798) (0.306) (0.389) Constant 27.947 4.884 -42.219 (89.915) (15.592) (47.028) Observations 293 293 293 R2 0.224 0.246 0.140 Notes. Bootstrap standard errors are reported in parentheses and are estimated after performing 200 replications. *** (**) (*) denotes significance at 1% (5%) (10%) level. Regressions include fixed effects for week and day of the week, and controls for weather (temperature, precipitation and two measures of wind). 4.2 Spillover Mechanisms The increase in crashes in the pre-curfew hours indicates that there could be a few things happening. First, as the number of vehicles on the road decreases and congestion decreases, there is an opportunity for vehicles to driver faster, which may increase the likelihood of a crash. During the pre-curfew hours, people may also drive even faster and more aggressively than they typically would under the same density circumstances because the time cost is higher because of the risk of punishment for being caught on the street past curfew. We explore this by looking at how speed changes in relation to the curfew during different times of day (see Appendix Figure A.2 for a comparison of speed versus congestion before and after the curfew for each hour of the day). There are increases in speed both during the daytime hours and during the pre-curfew hours, though the increase is twice as large during pre-curfew hours (Table 4). This demonstrates that the combined effect of changes in density and behavior is an increase in the speed that individuals drive. Additionally, given that we see no effect on crashes during the daytime hours (if anything the coefficient is negative, indicating a decrease in crashes), it seems that the increase in speed from the drop in congestion does not lead to increases in crashes. This is in line with the findings from Green et al. (2016). When we control for vehicle density using the congestion delay data from Waze, speed decreases during the daytime hours after the COVID-19 policies are implemented. During the pre-curfew hours there is still a significant increase in speed. This demonstrates that beyond the increase in speed that can occur when the number of vehicles decreases, individuals are driving at even higher speeds during the pre-curfew hours. We also look at this in a triple difference framework, comparing the increases in speed between the pre-curfew hours and the daytime hours, before 13 and after the curfew period in 2019 and in 2020, controlling for the level of congestion. We see a large positive increase in speed during the pre-curfew period as compared to the other daytime hours. This increase in speed that goes beyond the mechanical change in speed that occurs when congestion decreases could help to explain the increase in the probability of a crash during the pre-curfew time interval. Table 4: Change in Speed Daytime Pre-Curfew Hours Daytime Pre-Curfew Hours Triple (5:00-17:00) (17:00-20:00) Control Congestion Control Congestion Difference (1) (2) (3) (4) (5) Curfew Period 1.395*** 1.557*** 0.574*** 1.013** 0.682*** (0.226) (0.593) (0.190) (0.403) (0.196) Year 2020 0.114 0.379 1.416*** 3.662*** 1.931*** (0.228) (0.614) (0.194) (0.435) (0.198) Curfew x 2020 4.492*** 10.201*** -1.128*** 1.464** -1.641*** (0.271) (0.763) (0.235) (0.646) (0.241) Pre-Curfew Hours 1.614** (0.786) Curfew x Pre-Curfew -0.228 (0.360) 2020 x Pre-Curfew 0.178 (0.356) Curfew x 2020 x Pre-Curfew 4.993*** (0.527) Waze Delay (hr) -7.157*** -5.985*** -6.502*** (0.217) (0.393) (0.174) Waze Delay (hr)2 1.045*** 0.519*** 0.797*** (0.059) (0.081) (0.041) Waze Delay (hr)3 -0.051*** -0.015*** -0.034*** (0.005) (0.005) (0.003) Constant 13.698 -25.808 54.842*** 19.236 46.879*** (13.278) (38.932) (8.078) (20.416) (7.799) Observations 2981 744 2981 744 3725 R2 0.771 0.755 0.907 0.898 0.905 Notes. Robust standard errors are reported in parentheses. *** (**) (*) denotes significance at 1% (5%) (10%) level. Regressions include fixed effects for week and day of the week, and controls for weather (temperature, precipitation and two measures of wind). 4.3 Timing of the Curfew When the curfew is revised on June 6, 2020, we can look at how the timing of the curfew can affect the externality on road safety.9 There is a significant but smaller drop in the number of crashes during the revised curfew period (Table 5 Column 1). Importantly, there is no longer a spillover effect of more crashes occurring during the pre-curfew hours from 19:00 to 22:00 (Column 2). Studying additional externalities by using the proxy for rate of crashes, there is no significant impact on the probability of a crash either during the curfew hours or the pre-curfew hours (Columns 3 and 4). These findings suggest that the exact timing of the curfew can have important implications for road safety. Setting the start time of the curfew after the evening rush hour can minimize negative externalities of increased crashes due to 9 We focus on the first two months after the curfew is revised in order to match the length of time for which we observe the first curfew. 14 people speeding to get home before the curfew. Depending on the purpose of the curfew, there may be less flexibility with the specific hours. If the main goal, though, is having people stay home during the night hours, a curfew start time at 21:00 seems to be a better alternative for minimizing the externalities on road safety. Table 5: Difference-in-Differences Studying the Curfew Revision after June 6 2020 Crashes Crash Rate Curfew Hours Pre-Curfew Hours Curfew Hours Pre-Curfew Hours (22:00-4:00) (19:00-22:00) (22:00-4:00) (17:00-20:00) (1) (2) (3) (4) Curfew Period 0.764 1.176 1.003 0.766 (0.492) (0.767) (1.142) (0.574) Year 2020 0.241 0.413** 0.146 0.009 (0.177) (0.170) (0.329) (0.223) Curfew x 2020 -0.657*** -0.058 0.475 0.033 (0.208) (0.264) (0.550) (0.299) Constant -24.674 2.403 30.660 8.586 (22.705) (17.486) (53.752) (17.278) Observations 259 259 258 258 R2 0.229 0.178 0.302 0.148 Notes. Robust standard errors are reported in parentheses for columns (1) and (2) and bootstrap standard errors are re- ported in parentheses for columns (4) and (5). *** (**) (*) denotes significance at 1% (5%) (10%) level. Regressions include fixed effects for week and day of the week, and controls for weather (temperature, precipitation and two measures of wind). 4.4 Robustness Checks We conduct a number of different robustness checks (Table 6). First, there is the possibility that the likelihood of a crash being recorded by the police may have been affected by the curfew policy if it changed the likelihood of reporting a crash to the police or if the curfew led to increased police presence that increased reporting. In order to check if this may be driving the results, we use a different source of data on crashes coming from Flare to analyze the impacts of the curfew on crashes. We see very similar results for the curfew hours and pre-curfew hours as with the police data (Panel A of Table 6). The coefficient for the decrease in crashes during the curfew hours is smaller; therefore, in the triple difference, we do not see a significant drop in crashes during the curfew period hours in comparison to the daytime hours. We see a very large and significant increase during the pre-curfew hours. Since we have data on road traffic crashes from the police, we go further back in time to 2017 and use data from 2017, 2018 and 2019 as a comparison to 2020 and comparing the pre-COVID-19 policies period of Jan 1-March 15 with the post-curfew implementation period of March 27-June 6 for each year. When adding in the historical data, the results remain consistent and significant (Table 6 Panel B). For the main analysis, the crash data is aggregated daily for each time period, but we also explore running the analysis at the individual hourly level and also aggregating further at the weekly level. In both cases, we find the results to remain consistent and significant (Table 6 Panel C and Panel D). 15 Given some extreme outliers in the number of crashes in a given daily time interval, we also test removing the top 1% of time intervals with the most crashes (out of all time intervals with at least one crash). This means we remove all time intervals that have 5 crashes or more, and we see that the results remain stable and are not driven by these extreme outliers (Table 6 Panel E). Finally, we conduct placebo tests where we compare 2018 and 2019, using 2019 as the treatment year and similarly we compare 2017 and 2018, where we treat 2018 as the treatment year. We do not find any significant effects when comparing the pre and post periods in each of the pairs (Table 7). 5 Conclusion Rigorous evidence on the use of curfew policies for public safety is close to non-existent, Yet, curfews were enacted ubiquitously during the COVID-19 pandemic, in the hope of curbing the spread of the disease.10 The limited evidence that surfaced suggests that curfews have had limited effectiveness in reducing the burden of infectious disease and could have even increased the spread of it. Curfews resonate of war and conflict, crowd management and public safety, and the question of whether curfews are an effective tool to protect public safety is important, especially because curfews can impinge directly on personal freedom. Here we have studied rigorously one aspect of public safety: safety on the road. Indeed, this policy instru- ment has important externalities for road safety that might directly juxtapose the goal of improving public safety. In the context of a developing country metropolis, with high mortality on the road, and leveraging high-frequency data from multiple sources, we study the impact of curfews on road safety. We find that there is a direct decrease in crashes during the curfew hours when a dusk-to-dawn curfew is imple- mented, and a concomitant increase in the number of crashes during the hours before the curfew. The spillover effect is large enough to cancel out the post curfew reduction in crashes, and it occurs even though the number of vehicles on the road is lower, signifying an increase in the probability of a crash per car. Controlling for congestion on the road, we observe that speed increases during the hours right before the curfew, as people rush to reach their final destination just before the curfew. This is especially true when the curfew begins during rush hour. When it begins after rush hour, spillovers are mitigated. We conclude that the evidence is at best too thin to justify the enactment of curfews, especially to manage pandemics, and that governments should exercise caution and consider other instruments in their toolbox. 10 Curfews were still being implemented as late as end of 2021 due to the omicron variant. 16 Table 6: Robustness Tests Curfew Hours Pre-Curfew Hours Daytime Triple (20:00-5:00) (17:00-20:00) (5:00-17:00) Difference (1) (2) (3) (4) Panel A: Crashes Measured Using Data from Flare Curfew x 2020 x Nighttime Hours -0.226 (0.332) Curfew x 2020 x Pre-Curfew Hours 0.910*** (0.352) Curfew x 2020 -0.783*** 0.424** -0.423 (0.158) (0.203) (0.293) Observations 295 295 295 885 R2 0.301 0.281 0.253 0.254 Panel B: Crashes from 2017-2019 Used as Controls Curfew x 2020 x Nighttime Hours -0.595* (0.321) Curfew x 2020 x Pre-Curfew Hours 0.818** (0.332) Curfew x 2020 -0.897*** 0.603*** -0.298 (0.181) (0.195) (0.268) Observations 589 589 589 1767 R2 0.157 0.119 0.080 0.135 Panel C: Analysis at Hourly Level Curfew x 2020 x Nighttime Hours -0.091** (0.036) Curfew x 2020 x Pre-Curfew Hours 0.191** (0.075) Curfew x 2020 -0.127*** 0.191*** -0.030 (0.026) (0.071) (0.025) Observations 2655 885 3540 7080 R2 0.046 0.079 0.014 0.036 Panel D: Analysis at Weekly Level Curfew x 2020 x Nighttime Hours -4.190 (3.053) Curfew x 2020 x Pre-Curfew Hours 6.508** (3.020) Curfew x 2020 -7.814*** 3.251** -3.221 (1.469) (1.555) (2.717) Observations 46 46 46 138 R2 0.520 0.442 0.305 0.500 Panel E: Outlier Observations (5 or more crashes) Removed Curfew x 2020 x Nighttime Hours -0.532 (0.348) Curfew x 2020 x Pre-Curfew Hours 0.966*** (0.352) Curfew x 2020 -1.217*** 0.449** -0.521* (0.210) (0.215) (0.290) Observations 293 294 285 872 R2 0.216 0.173 0.113 0.166 Notes. Robust standard errors are reported in parentheses. *** (**) (*) denotes significance at 1% (5%) (10%) level. Regressions include fixed effects for week and day of the week, and controls for weather (temperature, precipitation and two measures of wind). 17 Table 7: Placebo test using 2017/2018 and 2018/2019 2017 to 2018 Comparison 2018 to 2019 Comparison Curfew Hours Pre-Curfew Hours Curfew Hours Pre-Curfew Hours (20:00-5:00) (17:00-20:00) (20:00-5:00) (17:00-20:00) (1) (2) (3) (4) Curfew Period -0.610 -0.377 -0.344 -0.182 (0.645) (0.382) (0.515) (0.451) Year 2020 -0.354* -0.210 -0.165 -0.046 (0.186) (0.163) (0.174) (0.136) Curfew x 2020 0.116 0.276 0.104 0.083 (0.275) (0.218) (0.271) (0.215) Constant 13.950 -2.765 -39.654 -4.962 (29.707) (13.325) (26.007) (11.627) Observations 294 294 294 294 R2 0.208 0.152 0.191 0.087 Notes. Robust standard errors are reported in parentheses. *** (**) (*) denotes significance at 1% (5%) (10%) level. Re- gressions include fixed effects for week and day of the week, and controls for weather (temperature, precipitation and two measures of wind). 18 References Altindag, O., Erten, B., & Keskin, P. (2022). Mental health costs of lockdowns: Evidence from age-specific curfews in turkey. American Economic Journal: Applied Economics, 14(2), 320–43. Bauernschuster, S., & Rekers, R. (2022, June). Speed limit enforcement and road safety. Journal of Public Economics, 210, 104663. Retrieved 2023-01-16, from https://www.sciencedirect.com/science/article/ pii/S0047272722000652 doi: 10.1016/j.jpubeco.2022.104663 Bearak, M., & Ombuor, R. (n.d.). Kenya’s coronavirus curfew begins with wave of police crackdowns. The Washington Post. Retrieved 2020-03-28, from https://www.washingtonpost.com/world/africa/ kenyas-coronavirus-curfew-begins-with-wave-of-police-crackdowns/2020/03/28/ 358327aa-7064-11ea-a156-0048b62cdb51 story.html Bolduc, D., Bonin, S., & Lee-Gosselin, M. (2013). A disaggregated tool for evaluation of road safety policies. Re- search in Transportation Economics, 37 (1), 79-98. Retrieved from https://www.sciencedirect.com/ science/article/pii/S0739885911000394 (Modelling National Road Safety Performance) doi: https://doi.org/10.1016/j.retrec.2011.08.009 unnings, C., & Schiele, V. (2021). Spring forward, don’t fall back: The effect of daylight saving time on road safety. B¨ Review of Economics and Statistics, 103(1), 165–176. Carr, J. B., & Doleac, J. L. (2018, October). Keep the Kids Inside? Juvenile Curfews and Urban Gun Violence. The Review of Economics and Statistics, 100(4), 609–618. Retrieved 2023-01-21, from https://direct.mit .edu/rest/article/100/4/609-618/58496 doi: 10.1162/rest a 00720 otz, G., & Heim, S. (2022, Nov 17). Measuring the effect of covid-19-related night curfews in a bundled de Haas, S., G¨ intervention within germany. Scientific Reports, 12(1), 19732. Retrieved from https://doi.org/10.1038/ s41598-022-24086-9 doi: 10.1038/s41598-022-24086-9 Edlin, A., & Karaca-Mandic, P. (2006). The accident externality from driving. Journal of Political Economy, 114(5), 931-955. Retrieved from https://doi.org/10.1086/508030 doi: 10.1086/508030 Green, C. P., Heywood, J. S., & Navarro, M. (2016). Traffic accidents and the london congestion charge. Jour- nal of Public Economics, 133, 11-22. Retrieved from https://www.sciencedirect.com/science/ article/pii/S0047272715001929 doi: https://doi.org/10.1016/j.jpubeco.2015.10.005 Huber, M., & Langen, H. (2020, Aug 26). Timing matters: the impact of response measures on covid-19-related hospitalization and death rates in germany and switzerland. Swiss Journal of Economics and Statistics, 156(1), 10. 19 Retrieved from https://doi.org/10.1186/s41937-020-00054-w doi: 10.1186/s41937-020-00054 -w Kagwe, H. M. E. (2020). National emergency response committee press statement on the update of coro- navirus in the country and response measures as at 22nd march 2020. (https://www.health.go.ke/wp- content/uploads/2020/03/CORONA-PRESS-STATEMENT-MARCH-22.pdf Accessed: 07/28/2021) Katana, E., Amodan, B. O., Bulage, L., Ario, A. R., Fodjo, J. N. S., Colebunders, R., & Wanyenze, R. K. (2021, March). Violence and discrimination among Ugandan residents during the COVID-19 lockdown. BMC Public Health, 21(1), 467. Retrieved 2023-01-21, from https://doi.org/10.1186/s12889-021-10532-2 doi: 10.1186/s12889-021-10532-2 Kline, P. (2012). The impact of juvenile curfew laws on arrests of youth and adults. American Law and Economics Review, 14(1), 44–67. Lin, L., Shi, F., & Li, W. (2021). Assessing inequality, irregularity, and severity regarding road traffic safety during covid-19. Scientific Reports, 11(1), 1–7. orkegren, D., & Viotti, L. (2021). Assessing bias in smartphone mobility estimates in low income Milusheva, S., Bj¨ countries. Proceedings of the 4th ACM SIGCAS Conference on Computing and Sustainable Societies. Ministry of Health. (2020). Kenya confirms two more cases of covid-19. nairobi sunday march 15, 2019. (https://www.health.go.ke/kenya-confirms-two-more-cases-of-covid-19-nairobi-sunday-march-15-2019/ Accessed: 07/28/2021) NSC. (2021). Nsc injury facts. Retrieved from https://injuryfacts.nsc.org/motor-vehicle/ overview/preliminary-monthly-estimates/ OSAC. (2020a). Health alert: Government of kenya announces curfew effective march 27. (https://www.osac.gov/Content/Report/fdaa6d90-4a90-4f8a-8eab-184b375bb291 Accessed: 07/28/2021) OSAC. (2020b). Health alert: Kenya, government announces changes to curfew and travel restrictions. (https://www.osac.gov/Content/Report/eb77c55d-fbca-4670-9c2d-18e154a19666/ Accessed: 07/28/2021) Ravindran, S., & Shah, M. (2023, January). Unintended consequences of lockdowns, COVID-19 and the Shadow Pandemic in India. Nature Human Behaviour, 1–9. Retrieved 2023-01-21, from https://www.nature.com/ articles/s41562-022-01513-5 doi: 10.1038/s41562-022-01513-5 Sprengholz, P., Siegers, R., Goldhahn, L., Eitze, S., & Betsch, C. (2021). Good night: Experimental ev- idence that nighttime curfews may fuel disease dynamics by increasing contact density. Social Science 20 Medicine, 286, 114324. Retrieved from https://www.sciencedirect.com/science/article/pii/ S0277953621006560 doi: https://doi.org/10.1016/j.socscimed.2021.114324 Tang, C., & van Ommeren, J. (2022). Accident externality of driving: evidence from the london congestion charge. Journal of Economic Geography, 22(3), 547-580. Retrieved from https://EconPapers.repec.org/ RePEc:oup:jecgeo:v:22:y:2022:i:3:p:547-580. Van Benthem, A. (2015). What is the optimal speed limit on freeways? Journal of Public Economics, 124, 44–62. Waze. (2021). Connected citizens program. (https://www.waze.com/wazeforcities Accessed: 07/28/2021) WHO. (2018). Global status report on road safety 2018. World Health Organization. Williams, A. F., & Preusser, D. F. (1997, Sep 01). Night driving restrictions for youthful drivers: A literature review and commentary. Journal of Public Health Policy, 18(3), 334-345. Retrieved from https://doi.org/ 10.2307/3343314 doi: 10.2307/3343314 21 Appendix A Additional Figures Figure A.1: Percent Change in Speed based on Google Data and Uber Data Figure A.2: Speed versus Congestion Hourly, Before and After the Curfew Policy in 2020 22 Appendix B Predicting Vehicles on the Road We predict the vehicles on the road for six major intersections with available data.11 The main predictor is congestion measured using Waze data (Figure B.1), and we include both linear and quadratic terms for this. We also control for the year and for indicators of weather (temperature, precipitation and two measures of wind). We estimate regressions separately for the different time intervals of the day and predict vehicle counts for each interval separately based on each regression. Table B.1 predicts vehicle counts for a single intersection as an illustration and Figure B.2 presents actual and predicted values for the six intersections. The correlations between the actual vehicle counts and the predicted counts are .97, .966, .974, .973, .97, and 0.961, respectively, for each intersection. Figure B.1: Trends in Waze Data 11 Datafor four of the intersections is available for January and April 2020, for one of the intersections data is available for February and April in 2019 and 2020, and for one intersection data is available for January, February, March and April 2020. 23 Table B.1: Predicting Vehicle Counts from Waze Congestion Data Curfew Hours Pre-Curfew Hours Daytime (20:00-5:00) (17:00-20:00) (5:00-17:00) (1) (2) (3) Waze Delay (hr) 0.251*** 0.043*** 0.032*** (0.001) (0.001) (0.000) Waze Delay (hr)2 -0.008*** -0.001*** -0.000*** (0.000) (0.000) (0.000) Year -0.600*** -0.031*** -0.120*** (0.004) (0.003) (0.002) Precipitation (m) -365.048*** -6.875*** 54.321*** (11.167) (1.980) (5.266) Temperature (K) 0.061*** 0.020*** 0.016*** (0.002) (0.001) (0.001) Eastward Wind (m/s) 0.081*** -0.003* -0.015*** (0.003) (0.002) (0.001) Northward Wind (m/s) -0.238*** -0.046*** -0.033*** (0.003) (0.003) (0.002) Constant 1200.971*** 65.707*** 246.420*** (9.135) (6.752) (4.867) Observations 112 110 112 Notes. Robust standard errors are reported in parentheses. *** (**) (*) denotes significance at 1% (5%) (10%) level. Temperature is measured in Kelvins (K), precipitation is measured in meters (m), and wind is measured in meters/second (m/s). 24 Figure B.2: Total Vehicle Counts Compared to Vehicle Counts Predicted Based on Congestion Data from Waze 25