key: cord-0989289-j9vtjczn
authors: Burns, Jacob; Hoffmann, Sabine; Kurz, Christoph; Laxy, Michael; Polus, Stephanie; Rehfuess, Eva
title: COVID-19 mitigation measures and nitrogen dioxide – A quasi-experimental study of air quality in Munich, Germany
date: 2020-11-22
journal: Atmos Environ (1994)
DOI: 10.1016/j.atmosenv.2020.118089
sha: 88c09ab2072778483dfd3bbfe20a664648151bc2
doc_id: 989289
cord_uid: j9vtjczn

BACKGROUND: In response to the COVID-19 pandemic, the Bavarian State government announced several COVID-19 mitigation measures beginning on March 16, 2020, which likely led to a reduction in traffic and a subsequent improvement in air quality. In this study, we evaluated the short-term effect of COVID-19 mitigation measures on NO(2) concentrations in Munich, Germany. METHODS: We applied two quasi-experimental approaches, a controlled interrupted time-series (c-ITS) approach and a synthetic control (SC) approach. Each approach compared changes occurring in 2020 to changes occurring in 2014–2019, and accounted for weather-related and other potential confounders. We hypothesized that the largest reductions in NO(2) concentrations would be observed at traffic sites, with smaller reductions at urban background sites, and even small reductions, if any, at background sites. All hypotheses, as well as the main and additional analyses were defined a priori. We also conducted post-hoc analyses to ensure that observed effects were not due to factors other than the intervention. RESULTS: Main analyses largely supported our hypotheses. Specifically, at the two traffic sites, using the c-ITS approach we observed reductions of 9.34 μg/m(3) (95% confidence interval: −23.58; 4.90) and 10.02 μg/m(3) (−19.25; −0.79). Using the SC approach we observed reductions of 15.65 μg/m(3) (−27.58; −4.09) and 15.1 μg/m(3) (−24.82; −9.83) at these same sites. We observed effects ranging from smaller in magnitude to no effect at urban background and background sites. Additional analyses showed that the effect was largest in the first two weeks following introduction of measures, and that a 3-day lagged intervention time also showed a larger effect. Post-hoc analyses suggested that at least some of the observed effects may have been attributable to changes in air quality occurring before the intervention, as well as unusually high concentrations in January 2020. CONCLUSION: We applied two quasi-experimental approaches in assessing the impact of the COVID-19 mitigation measures on NO(2) concentrations in Munich. Taking the 2020 pre-intervention average concentrations as a reference, we observed reductions in NO(2) concentrations of approximately 15–25% and 24–36% at traffic sites, suggesting that reducing traffic may be an effective measure to reduce NO(2) concentrations in heavily trafficked areas by margins which could translate to public health benefits.

Main analyses largely supported our hypotheses. Specifically, at the two traffic sites, using the c-22

ITS approach we observed reductions of 9.34 µg/m 3 (95% confidence interval: -23.58; 4.90) and 10.02 23 µg/m 3 (-19.25; -0.79). Using the SC approach we observed reductions of 15.65 µg/m 3 (-27.58; -4.09) and 24

15.1 µg/m 3 (-24.82; -9.83) at these same sites. We observed effects ranging from smaller in magnitude to 25 no effect at urban background and background sites. Additional analyses showed that the effect was 26 largest in the first two weeks following introduction of measures, and that a 3-day lagged intervention 27 time also showed a larger effect. Post-hoc analyses suggested that at least some of the observed effects 28 may have been attributable to changes in air quality occurring before the intervention, as well as 29 unusually high concentrations in January 2020. 30

We applied two quasi-experimental approaches in assessing the impact of the COVID-19 31 mitigation measures on NO 2 concentrations in Munich. Taking the 2020 pre-intervention average  32 concentrations as a reference, we observed reductions in NO 2 concentrations of approximately 15-25% 33 and 24-36% at traffic sites, suggesting that reducing traffic may be an effective measure to reduce NO 2 34 concentrations in heavily trafficked areas by margins which could translate to public health benefits. 35

In December 2019 the first cases of the novel coronavirus, SARS-CoV-2, were observed in Wuhan, China. 41 Over the next days and weeks the virus, and the associated respiratory disease referred to as 42 spread further into China and by mid-January cases were documented in Thailand, Japan and South 43

Korea (WHO, 2020a). By 11 March 2020, when the World Health Organization declared COVID-19 a 44 global pandemic, cases had been observed in over 100 countries and territories across the globe (WHO, 45 2020b). 46

To slow the spread of this viral respiratory infection, the effects of which range from limited or no 47 symptoms to death, national and subnational governments have implemented numerous mitigation 48 measures (Health System Response Monitor, 2020). These mitigation measures differ between 49

countries, but include, for example, social distancing recommendations and requirements, school 50 closures, border closures, non-essential business closures and required wearing of masks. the suspension of the public transportation system due to a strike in Ottawa, Canada (Ding et al., 2014) , 57 the suspension of trucking operations due to a nationwide strike in India (Latha et al., 2004) , political 58 demonstrations in Nepal (Fransen et al., 2013) traffic related to grocery shopping or outdoor recreational activity, likely did not decrease or decreased 100 to a lesser extent, so any change was likely driven by a reduction in driving by those commuting to work 101

and/or driving their children to school. We assume that this reduction in traffic likely also subsequently 102 led to reduced concentrations of automobile-related pollutants like NO 2 . 103

J o u r n a l P r e -p r o o f

The study uses an approach that compares the trend in NO 2 concentrations in 2020, i.e. the intervention 105

year, with the trend in several years in which no mitigation measures for COVID-19 control were 106 implemented, i.e. the control years. allowed us to ensure that any effect observed in 2020 is neither due to the current trend in NO 2 119

concentrations nor due to yearly seasonal fluctuations. 120

The main difference between the two approaches, however, relates to how data from control units are 121

utilized. The c-ITS study utilizes data from all control units in full. Specifically, we compared the change in 122 NO 2 concentrations between the pre-and post-intervention periods in 2020, the intervention year, to 123 changes in concentrations between the pre-and post-intervention periods in 2014-2019, the control 124 years (Lopez Bernal et al., 2018) . The SC study can be utilized when there are multiple controls to draw 125 from, but no clear rationale for choosing which is the most appropriate. Specifically, we compared the 126 change in NO 2 concentrations between the pre-and post-intervention periods in 2020 to changes 127 between the pre-and post-intervention periods in a weighted average of 2014-2019. This data-driven 128 weighted average is calculated to provide the most similar comparison, with respect to the pre-129

intervention outcome trend and a pre-defined set of covariates (Bouttell et al., 2018) . 130

Outcome 132

The Bavarian Environmental Administration (Bayerisches Landesamt für Umwelt) is charged with the 133 monitoring of air quality in Bavaria, and data for the 50 monitoring stations are freely available (LfU 134

Bayern, 2020). We obtained NO 2 data for the five stations located in Munich, which included two 135 classified as urban traffic monitors -Landshuter Allee (LAN) and Stachus (STA), one as urban background 136 -Lothstrasse (LOT), and two as background -Allach (ALL) and Johanneskirchen (JOH). Hourly data were 137 provided, which we converted to daily averages. 138

We obtained data for other factors that are associated with NO 2 concentrations, including several 140

weather-related variables -daily averages of temperature, rain fall, air pressure, humidity, and wind 141 speed (Peel 2010). These data were freely available from the German Weather Service (Deutscher 142 J o u r n a l P r e -p r o o f Wetterdienst) (DWD, 2020). We also used publicly available information indicating when school holidays 143

were in place -these included the Christmas, winter and Easter holidays. Within these time periods, 144 relevant days were defined as either holiday high travel days (i.e. specific holidays or holiday weekends -145

Friday and Saturday, on which people tend to travel more) or holiday low travel days (i.e. during the 146

week when people tend to travel less -Sunday through Thursday). 147

We registered a study protocol on 3 May 2020 through OSF (https://osf.io/7vkfc); all hypotheses and 149 methods for main and additional analyses were defined a priori in the protocol. We designed and piloted 150 these analyses using data from 2014-2019. The data for the intervention year, 2020, were downloaded 151 and analyzed only after registration of the protocol. 152

As part of the main analyses we applied a c-ITS and SC approach. for a level change. This level change represents an immediate change, which is sustained across the post-162 intervention period. 163

As described above in section 3.3, we obtained data from five air quality monitoring stations. Our a priori 164 hypothesis was that the observed effect would be greatest at the two traffic monitors LAN and STA, a 165 smaller effect at the urban background monitor LOT, and the smallest effect, if any, at the two 166 background monitors ALL and JOH. 167

For the c-ITS approach, we fitted a linear model using the general least squares method. The model took 168 the following form: 169 includes the potentially important covariates, including temperature, rain fall, air  179 pressure, humidity, wind speed, holiday high and low travel days and day of the week. β 5 , the change in 180 NO 2 concentrations between the pre-and post-intervention periods in 2020 relative to the change in 181 2014-2019, represents the level change described above in section 3.3, and is thus the effect estimate of 182

interest. Given the serially correlated nature of the data, we used auto-correlation and partial auto-183 correlation plots to determine an appropriate correlation structure for each model. For the site ALL, 184 substantial data were missing for the year 2014 (23%); because of this, 2014 was excluded from the c-ITS 185 analysis for ALL only. 186

The SC approach was structured similarly. However, instead of comparing changes in 2020 to changes in 187 2014-2019, the method allows for the construction of a synthetic control, ensuring that the intervention 188

year and synthetic control year were similar with regard to the pre-intervention outcome trend and 189 potentially important covariates. Specifically, this synthetic control was constructed using input data 190 from the pre-intervention NO 2 concentrations, as well as the covariates listed above, from 2014-2019. Additional analyses specified a priori 199

We conducted a series of sensitivity analyses to evaluate the extent to which our results were robust to 200 changes to our assumptions, and to further explore how the intervention effect developed and changed 201 over time. 202

The mitigation measures were dependent on individuals changing their behavior, and this behavior may 203 have been adapted over time. We suspected that the effect in the two weeks immediately following the 204 intervention may have been larger than the effect in the subsequent two weeks. We investigated this 205 using both the c-ITS and SC approaches. For the c-ITS approach, we modelled two intervention effects 206 separately, one specifically for the first two-week period, and the other for the second two-week period. 207

For the SC approach, we shortened the post-intervention time period to two weeks. 208

It is also plausible that individuals did not immediately change their behavior on 16 March 2020, but 209 instead slowly adapted as further mitigation measures were announced. To investigate this possibility, 210

we mimicked the main analyses, treating 19 March 2020 as the first day of the post-intervention period, 211

under the assumption that behaviors changed measurably after a lag of three days. 212

After conducting the a priori specified main and additional analyses, we further conducted three sets of 214 analyses to ensure that observed changes were not due to factors other than the mitigation measures. 215

To ensure that concentration changes occurring prior to the intervention were not driving observed 216 changes, we conducted all analyses with a series of backdated intervention start points 2, 4 and 6 weeks 217 prior to 16 March 2020. Each of these 'placebo analyses' assessed whether changes occurred within two 218 weeks of the respective intervention point, although no intervention actually occurred. Next, to assess 219 whether high concentrations in January 2020 may have biased the pre-intervention trend and thus the 220 calculated effects, we conducted all analyses with a shortened pre-intervention period lasting 6 weeks. 221

Finally, to ensure that the noisy nature of daily air quality data, characterized by serial correlation as well 222

as random noise, was not driving observed concentration changes, we repeated all analyses with 223 smoothed NO 2 data. To do so, we analyzed only the trend component of the decomposed data. 224

All data processing and analyses were conducted using R version 3.6.3. The c-ITS approach was 225 conducted using the Fit Linear Model Using Generalized Least Squares (nlme) (Weisberg and Fox, 2015) 226 and the SC approach was conducted using the Generalized Synthetic Control Method (gsynth) package 227 (Xu, 2017) . 228 

Regarding the effect of the COVID-19 mitigation measures on NO 2 concentrations across the post-241 intervention period, our main analyses are summarized in Figure 2 (panel A) and Table 1 . after approximately 1.5-2 weeks. Additional analyses below explore how the intervention effect 262 developed and changed over time. 263 Figure 3 shows that the SC approach was not able to calculate an optimal counterfactual -for an optimal 264 counterfactual, the pre-intervention ATT would lie very close to 0 at all points along the time series. 265

Additionally, one can see that the average NO 2 concentration approximately 4 weeks prior to the 266 intervention appears to lie below the 0 ATT line, meaning that the observed effects may in part be 267 attributable to changes occurring before the intervention. Post hoc analyses, described below, explore 268

whether these aspects may have biased observed effects. 269 

Regarding the timing of the effect, we further investigated whether the effect in the first two-week post-275 intervention period was larger in magnitude than the effect over the entire four weeks. These results are 276 summarized in Figure 2 (panel B) and Table 1 . As hypothesized, across sites effects were slightly larger 277 when considering a two-week post-intervention period rather than a four-week period. Regarding the 278 second two-week post-intervention period, which we assessed using the c-ITS approach, observed 279 effects were smaller at all sites than in the first two-week period. Confidence intervals for all estimates 280

should be noted; for the c-ITS approach, a significant effect was observed only at STA, while for the SC 281 approach significant effects were observed at LAN, STA and LOT. For all other estimates, confidence 282 intervals contained 0, indicating some uncertainty regarding the direction of these effects. 283

Additionally, we investigated whether the effect differed if the intervention start was delayed for three 284 days from 16 March to 19 March 2020. These results are summarized in Figure 2 (panel C) and Table 1.  285 As hypothesized, a lagged intervention start resulted in a slightly larger effect at traffic sites. At urban 286 background and background sites similar to slightly larger effects were observed. Confidence intervals 287 for all estimates should be noted; for the c-ITS approach, a significant effect was observed only at STA, 288

while for the SC approach significant effects were observed at LAN, STA and LOT. For all other estimates, 289

confidence intervals contained 0. 290 summarized in Figure 4 panels A-C, respectively, and Appendix Table 1 in the supplementary material. 301

Compared to the main analyses, effects at traffic sites are smaller when either 3 February or 2 March is 302 taken as the intervention point. However, for 17 February, observed effects are actually larger than 303 those observed in the main analyses. At urban background and background sites, where we expected a 304 small or no effect due to the COVID-19 measures, larger effects were observed for almost all backdated 305 analyses compared to main analyses. Taken together, this suggests that the effect observed in main 306

analyses may be at least partially attributable to changes in air quality across Munich (i.e. not only in 307 heavily trafficked areas) already occurring prior to 16 March 2020. 308

Analyses of a shortened pre-intervention period allowed us to assess whether the high concentrations 309 observed in January 2020 influenced the observed effect; these results are summarized in Figure 5 (panel 310

A) and Appendix Table 2 in the Supplementary material. Smaller effects at traffic sites were observed for 311 the shortened pre-intervention period than for main analyses, potentially suggesting that observed 312 effects are at least partially attributable to high concentrations observed in January 2020. Analyses of 313

smoothed NO 2 data are summarized in Figure 5 (panel B) and Appendix Table 2 in the Supplementary  314 material. The smoothed data allowed for the calculation of a better counterfactual than the raw data 315 (Appendix Figure 1 ). Compared to results from the main analyses, a slightly smaller effect across sites 316 was observed. This suggests that some of the effect observed in main analysis may be attributable to 317 random noise or serial correlation, although at the same time, it is possible that the smoothing of the 318 data smoothed away part of an actual effect. 319 J o u r n a l P r e -p r o o f 

In this study, we applied a c-ITS and SC approach to evaluate the short-term effect of COVID-19 327 mitigation measures on NO 2 in Munich. Main and additional analyses suggest a consistent pattern -after 328

introduction of the mitigation measures decreases in NO 2 concentrations were observed at traffic sites, 329 while little to no change was observed at urban background and background sites. As expected, 330 reductions were largest in magnitude in the two weeks immediately following the introduction; a lagged 331 intervention start suggests that the effect became more pronounced as additional measures were 332

implemented. Specifically, the c-ITS approach allows comparison of trends in 2020 to the average of trends over the 373 time period of 2014-2019 so that the comparison will not be heavily skewed by any one year that does 374 not fit the true long-term trend. The SC study complements this approach by creating a control condition 375 from 2014-2019 that most closely matches the intervention time trend. We further accounted for 376 potentially important confounders in both approaches: the c-ITS model was adjusted for temperature, 377 rainfall, air pressure, humidity, wind speed, day of the week and holidays; the SC approach used these 378 factors in creating an appropriately weighted synthetic control. We defined most hypotheses and 379 analyses a priori and registered a study protocol, before downloading the data for 2020. Only the 380 analyses of a backdated intervention point, a shortened pre-intervention period and smoothed NO 2 data 381 were defined post hoc; these were added to ensure that observed effects were not attributable to other 382 factors. 383

Nevertheless, there are limitations to this study. We assume that the COVID-19 mitigation measures led 384

to reductions in traffic, which subsequently led to reductions in NO 2 concentrations. Lacking reliable data 385 on traffic, however, we cannot assess to what extent this assumption of effects along the causal chain 386 are appropriate. Mobility data from smartphones made available by Google (Google, 2020) and Apple 387 (Apple, 2020) suggest that mobility was starkly reduced during these weeks; however the current study 388 would have benefited from the incorporation of long-term, representative routine traffic data. Post hoc 389

analyses suggest that effects observed in main analyses may at least partly stem from factors other than 390 the mitigation measures, including reductions in NO2 concentrations occurring prior to 16 March, high 391 concentrations observed in January and the noisy nature of the data. However, the large decrease 392

immediately after 16 March 2020 is observable across all main and additional analyses, meaning it is 393 unlikely that observed effects are due only to factors other than the mitigation measures. This large 394 decrease is consistent with the reduction in traffic reported in the mobility data described above. While 395 the monitoring sites assessed represent all regulatory sites available for Munich during the study period, 396

it is possible that these are not fully representative of air quality across Munich. Additionally, for the ALL 397 site, the year 2014 was excluded from the c-ITS approach because much of the data from that year were 398 missing. However, we consider it unlikely that this substantially influenced our results. We assessed 399 changes only in NO 2 concentrations, as this allowed us to most closely assess whether changes to air 400 quality were likely due to changes in traffic reductions. Nevertheless, a more comprehensive assessment 401 of the impact of a specific intervention, measure or event would entail the assessment of multiple 402 pollutants. To the best of our knowledge, this is the first use of historical controls within a SC study, and 403

we feel that this is an appropriate use of the available data. Nevertheless, our study highlights challenges 404 associated with calculating an optimal counterfactual using a SC study in the context of air quality data. 405 However, given that the c-ITS approach, which can better account for time-varying confounders, and the 406 analyses of smoothed data yielded similar results, if slightly smaller in magnitude, we think it unlikely 407 that our results are biased by this limitation. 408

Given that traffic is only one source of NO 2 and other air pollutants, continuing to improve air quality will 409 likely require multiple control measures targeting multiple sources. However, this study suggests that 410 reducing traffic may be an effective measure to reduce NO 2 concentrations in heavily trafficked areas by 411 margins which could translate to public health benefits. 412

Lufthygienisches Landesüberwachungssystem Bayern. Bayerisches Landesamt für 415 Umwelt

Interrupted time series regression for the evaluation 417 of public health interventions: a tutorial

Accountability Studies on Air 419 Pollution and Health: the HEI Experience

Synthetic control methodology 421 as a tool for evaluating population-level health interventions

Effect of road blockages on local air pollution during the Hong Kong 424 protests and its implications for air quality management. Science of The Total Environment

Interventions to reduce ambient particulate matter air pollution and their 428 effect on health

Interventions to reduce ambient air pollution and their effects on 431 health: An abridged Cochrane systematic review

Natural Experiments: An Overview of 433

Approaches, and Contributions to Public Health Intervention Research. Annual Review 434 of Public Health

The impact of COVID-19 436 partial lockdown on the air quality of the city of Rio de Janeiro, Brazil. Science of The Total 437 Environment

Characterization of chemical composition and concentration 439 of fine particulate matter during a transit strike in Ottawa, Canada

Impact of vehicular strike on particulate 447 matter air quality: Results from a natural intervention study

NO2 Konzentrationen während Covid-19 Lock-Downs in München. HawaDawa. 450 GOOGLE 2020. COVID-19 Community Mobility Report: Munich. Google

The short-term impacts of COVID-19 lockdown on urban air pollution 452 in China

Evaluating the effectiveness of 454 air quality regulations: A review of accountability studies and frameworks

Impact of Temporary Freeway Closure on 457 Regional Air Quality: A Lesson from Carmageddon in Los Angeles, United States

Estimate of the development of the epidemic 461 reproduction number Rt from Coronavirus SARS-CoV-2 case data and implications for political 462 measures based on prognostics

Impact of diesel vehicular emission on ambient black 464 carbon concentration at an urban location in India

LEOPOLDINA 2019. Saubere Luft. Stickstoffoxide und Feinstaub in der Atemluft: Grundlagen und

Nationale Akademie der Wissenschaften Leopoldina

The air quality impacts of road closures associated with the 468 2004 Democratic National Convention in Boston

The use of controls in interrupted time series 470 studies of public health interventions

Effect of lockdown amid COVID-19 pandemic on air quality of 472 the megacity Delhi

Social distancing measures following COVID-19 474 epidemics had positive environmental consequences. GREEN Working Paper Series

COVID-19 Health System Response Monitor

COVID-19 pandemic and environmental pollution: A 479 blessing in disguise? Science of The Total Environment

11,000 air pollution-related deaths avoided in Europe as coal, oil 481 consumption plummet

Mortality effects of a copper smelter strike and reduced 483 ambient sulfate particulate matter air pollution

Accountability studies of air pollution and health effects: lessons learned and 486 recommendations for future natural experiment opportunities

Bericht aus der Kabinettssitzung

Bericht aus der Kabinettssitzung

An R Companion to Applied Regression, Appendix: Mixed Effects Models in 491 R, SAGE. 492 WHO 2020a. Novel Coronavirus (2019-nCoV) Situation Report -1. World Health Organization. 493 WHO 2020b. Novel Coronavirus (2019-nCoV) Situation Report -51

Generalized Synthetic Control Method: Causal Inference with Interactive Fixed Effects 495 Models

Sabine Hoffmann: none known Christoph Kurz: none known

none known Stephani Polus: none known Eva Rehfuess: none known