Short term association between ozone and mortality: global two stage time series study in 406 locations in 20 countries

Abstract Objective To assess short term mortality risks and excess mortality associated with exposure to ozone in several cities worldwide. Design Two stage time series analysis. Setting 406 cities in 20 countries, with overlapping periods between 1985 and 2015, collected from the database of Multi-City Multi-Country Collaborative Research Network. Population Deaths for all causes or for external causes only registered in each city within the study period. Main outcome measures Daily total mortality (all or non-external causes only). Results A total of 45 165 171 deaths were analysed in the 406 cities. On average, a 10 µg/m3 increase in ozone during the current and previous day was associated with an overall relative risk of mortality of 1.0018 (95% confidence interval 1.0012 to 1.0024). Some heterogeneity was found across countries, with estimates ranging from greater than 1.0020 in the United Kingdom, South Africa, Estonia, and Canada to less than 1.0008 in Mexico and Spain. Short term excess mortality in association with exposure to ozone higher than maximum background levels (70 µg/m3) was 0.26% (95% confidence interval 0.24% to 0.28%), corresponding to 8203 annual excess deaths (95% confidence interval 3525 to 12 840) across the 406 cities studied. The excess remained at 0.20% (0.18% to 0.22%) when restricting to days above the WHO guideline (100 µg/m3), corresponding to 6262 annual excess deaths (1413 to 11 065). Above more lenient thresholds for air quality standards in Europe, America, and China, excess mortality was 0.14%, 0.09%, and 0.05%, respectively. Conclusions Results suggest that ozone related mortality could be potentially reduced under stricter air quality standards. These findings have relevance for the implementation of efficient clean air interventions and mitigation strategies designed within national and international climate policies.


eMethods 1 -Description of the MCC data
Here we provide a detailed description of the mortality and environmental data used in the present study. As mentioned in the method section of the main manuscript, this data is included in a large dataset collected through the Multi-Country Multi-city (MCC) Collaborative Research Network (http://mccstudy.lshtm.ac.uk/). The dataset has been used in previous publications on temperature-related mortality and related topics (e.g. Gasparrini et al. 2015 Lancet) and a recent study on mortality associated to inhalable particles (Liu et al. 2019 NEJM). A detailed description of the air pollution data is provided in the latter. Below we describe for each country the specific data sources, the definition of the variables and the quality checks that were applied.

Selection of the study locations
We initially included 434 locations available in the MCC database at the time of the study, with available data on daily mortality, mean temperature and ozone (computed as 8-hour maximum). The selection was restricted to cities or metropolitan areas included in the MCC dataset (i.e., regions or provinces were excluded). We then selected 423 that provided data for more than 3 years between 1985 and 2015. Note that for most of the cities, air pollutants were available during specific time intervals which might not be homogeneous within country and not consistent with the mortality and temperature data. The main study period reported in Table 1 is defined as the time between the earliest year and latest year for which ozone data was available in each city within country and by air pollutant. Then, an exploratory analysis of the ozone data was performed as a quality check. From the 423 locations initially selected, we excluded 6 locations (1 in Italy and 5 in the US) with a percentage of missing values above 75% in the 30-days moving average of ozone. We then excluded 6 locations for which the temporal patterns in daily ozone did not follow the expected seasonal trends for this air pollutant, for instance showing abrupt steps, potentially due to changes in the monitors (e.g. relocation, change in the device) or other technical problems. Finally, 5 locations were excluded (1 in Mexico, 1 in South Africa, and 3 in the US) where mortality series presented more than 50% of missing counts. We specify below the locations that were excluded and the reason for exclusion in each country. For each sensitivity analysis using other pollutants, additional exclusions were performed depending on data availability of each air pollutant and weather variable.
the UK Air Quality Archive, which reports results from the network of monitoring stations operated by the UK government. Hourly measurements of PM10 and O3 were available in the same period from urban and sub-urban monitoring stations. Daily PM10 levels were computed as the 24-hour average from urban monitoring stations, and as 8hour-maximum for O3. In total, missing data amount for 0.0, 0.0, 13.6, 7.3 and 7.9% of the mortality, mean temperature, relative humidity, O3 and PM10 series, respectively.
United States (188 cities, 1985States (188 cities, -2006 Daily mortality is represented by counts of deaths for all causes. Mean daily temperature (in ˚C) and relative humidity (%), computed as the 24-hour average based on hourly measurements, were obtained from the National Climatic Data Center (NCDC) of the National Oceanic and Atmospheric Administration (NOAA). Hourly measurements of PM10, PM2.5, NO2 and O3 were gathered from the U.S. Environmental Protection Agency (EPA) Air Quality System (AQS), from urban and sub-urban monitoring stations. Daily PM10, PM2.5 and NO2 levels were computed as the 24-hour average from urban monitoring stations (only for 176 and 137 cities), and as 8-hour-maximum for O3 from monitors located in the county or set of contiguous counties in which the city is located. 91 out the 182 US cities registered daily O3 levels only during the summer period. Data from 205 cities were initially collected, but 8 locations with a short study period (less than 3 years) were excluded, along with other 9 locations with a poor quality of O3 data. PM10 and PM2.5 were collected between 2005and 2006, and 2004and 2006 In total, missing data amount for 0. 8, 0.8, 5.2, 23.4, 64.7, 41.2 and 14.5% of the mortality, mean temperature, relative humidity, O3, PM10, PM2.5 and NO2 series, respectively.

Control for temperature in the main model
We control for temperature with distributed lag non-linear model (DLNM) of mean daily temperature. According to a previous study of MCC collaborative network (Gasparrini et al. Lancet 2015), the non-linear and delayed temperature-mortality association was modelled through a quadratic b-spline with internal knots in the 10 th , 75 th , 90 th percentiles in the exposure-response dimension, and natural spline with 3 equally-spaced internal knots in the log scale of the lag dimension, up to 21 days of lag.

Description of the additional analysis
Here we provide a detailed description of specific features of the additional analyses.
As described in the main manuscript, we explored potential non-linearity and delayed associations between ozone and mortality. The modelling choices selected (specifications of modelling function of ozone) for each sub-analysis were chosen based on q-AIC among different combinations of smoothing functions and placement and number of knots for the exposure-response and lag dimension, respectively.
-Modelling choices for the non-linear concentration-response function: natural cubic spline and quadratic b-spline with 1 or 2 internal knots placed in 50 and 60 µg/m 3 . -Modelling choices for the lag-association function: natural cubic spline with 1, 2 or 3 internal knots placed equally-spaced in the natural or log scale of the lag dimension up to 30 days.
City-specific non-linear exposure-response associations and lag-response associations obtained in these two additional analyses were pooled following the same method described in the main manuscript for the final model. More details on this methodology can be found in Gasparrini et al. 2013 BMC Med Res Methodol.

Description of the sensitivity analyses
Different control for time trends: natural spline with 4 and 10 degrees of freedom per year.
Control for PM10, PM2.5 and NO2: air pollutants were included one by one in the main model as linear unconstrained distributed-lag linear models (DLMs) for the same and previous day of the exposure (lags 0 and 1).
Control for relative humidity: we controlled for relative humidity in the model as a linear term of lag0.
Modelling approaches used to control for temperature: -Main model: quadratic b-spline with internal knots in the 10 th , 75 th , 90 th percentiles in the exposure-response dimension, and natural spline with 3 equally-spaced internal knots in the log scale of the lag dimension, up to 21 days of lag. -Approach 1: quadratic b-spline with internal knots in the 10 th , 75 th , 90 th percentiles in the exposure-response dimension, and natural spline with 3 equally-spaced internal knots in the log scale of the lag dimension, up to 13 days of lag. -Approach 2: quadratic b-spline with 3 internal knots equally-spaced quantiles in the exposure-response dimension, and natural spline with 1 internal knot equally-spaced in the log scale of the lag dimension, up to 13 days of lag. -Approach 3: natural cubic spline with 3 degrees of freedom of the moving average lag013. -Approach 4: natural cubic spline with 3 degrees of freedom of the moving average lag03. eTable 3. Overall and country-specific excess mortality fractions (%, 95% confidence interval) associated to ozone for the total (above 70 µg/m 3 ) and above specific thresholds consistent with current air quality standards (AQS).
*Total refers to ozone-related deaths when levels above 70 µg/m 3 (defined as maximum background levels). **No mortality fractions associated to ozone were found in Australia, as daily ozone levels were below the maximum background level set up at 70 µg/m 3 .