Secondary organic aerosols from anthropogenic volatile organic compounds contribute substantially to air pollution mortality

Anthropogenic secondary organic aerosol (ASOA), formed from anthropogenic emissions of organic compounds, constitutes a substantial fraction of the mass of submicron aerosol in populated areas around the world and contributes to poor air quality and premature mortality. However, the precursor sources of ASOA are poorly understood, and there are large uncertainties in the health benefits that might accrue from reducing anthropogenic organic emissions. We show that the production of ASOA in 11 urban areas on three continents is strongly correlated with the reactivity of specific anthropogenic volatile organic compounds. The differences in ASOA production across different cities can be explained by differences in the emissions of aromatics and intermediateand semi-volatile organic compounds, indicating the importance of controlling these ASOA precursors. With an improved model representation of ASOA driven by the observations, we attribute 340 000 PM2.5-related premature deaths per year to ASOA, which is over an order of magnitude higher than prior studies. A sensitivity case with a more recently proposed model for attributing mortality to PM2.5 (the Global Exposure Mortality Model) results in up to 900 000 deaths. A limitation of this study is the extrapolation from cities with detailed studies and regions where detailed emission inventories are available to other regions where uncertainties in emissions are larger. In addition to further development of institutional air quality management infrastructure, comprehensive air quality campaigns in the countries in South and Central America, Africa, South Asia, and the Middle East are needed for further progress in this area.

210 the observations to improve the modeled representation of ASOA. The results provide insight 211 into the importance of ASOA to global premature mortality due to PM 2.5 and further 212 understanding of the precursors and sources of ASOA in urban regions.  For values not previously reported in the literature ( Table S4 ), observations taken 225 between 11:00 -16:00 local time were used to determine the slopes of SOA versus 226 formaldehyde (HCHO) ( Fig. S1 ), peroxy acetyl nitrate (PAN) ( Fig. S2 ), and O x (O x = O 3 + NO 2 ) 227 ( Fig. S3 ). For CalNex, there was an approximate 48% difference between the two HCHO 228 measurements ( Fig. S4 ). Therefore, the average between the two measurements were used in this 229 study, similar to what has been done in other studies for other gas-phase species (Bertram et al., 230 2007) . All linear fits, unless otherwise noted, use the orthogonal distance regression fitting 231 method (ODR). 232 For values in Table S4 through Table S8 not previously reported in the literature, the 233 following procedure was applied to determine the emissions ratios, similar to the methods of 234 Nault et al. (2018) . An OH exposure (OH exp = [OH]×Δt), which is also the photochemical age 235 (PA), was estimated by using the ratio of NO x /NO y ( Eq. 1 ) or the ratio of 236 m+p-xylene/ethylbenzene ( Eq. 2 ). For the m+p-xylene/ethylbenzene, the emission ratio 237 ( Table S5 ) was determined by determining the average ratio during minimal photochemistry, 238 similar to prior studies (de Gouw et al., 2017) . This was done for only one study, TexAQS 2000. 239 This method could be applied in that case as it was a ground campaign that operated both day 240 and night; therefore, a ratio at night could be determined when there was minimal loss of both 241 VOCs. The average emission ratio for the other VOCs was determined using Eq. 3 after the 242 OH exp was calculated in Eq. 1 or Eq. 2 . The rate constants used for determining OH exp and 243 emission ratios are found in Table S12 . With the combination of the new dataset, which expands across urban areas on three 252 continents, the SIMPLE parameterization for ASOA (Hodzic and Jimenez, 2011) is updated in 253 the standard GEOS-Chem model to reproduce observed ASOA in Fig. 2  Eq. 5 265 Where E and k stand for the emission rate and reaction rate coefficient with OH, respectively, for 266 benzene (B), toluene (T), and xylenes (X). Ethylbenzene was not included in this calculation 267 because its emission was not available in HTAPv2 emission inventory. However, ethylbenzene 268 contributed a minor fraction of the mixing ratio (~ 7%, Table S5 ) and reactivity (~6%) of the 269 total BTEX across the campaigns. Reaction rate constants used in this study were 1.22×10 -12 , 270 5.63×10 -12 , and 1.72×10 -11 cm 3 molec. -1 s -1 for benzene, toluene, and xylene, respectively 271 (Atkinson and Arey, 2003; Atkinson et al., 2006) . The R aromatics /ΔCO allows a dynamic 272 calculation of the E(VOC)/E(CO) = SOA/ΔCO. Hodzic and Jimenez (2011) and Hayes et al. 273 (2015) used a constant value of 0.069 g g -1 , which worked well for the two cities investigated, 274 but not for the expanded dataset studied here. Thus, both the aromatic emissions and CO 275 emissions are used in this study to better represent the variable emissions of ASOA precursors 276 (Fig. S5).

277
Second, E SOAP /E CO can be obtained from the result of Eq. 6 , using slope and intercept in 278 Fig. 2 306 We performed an additional sensitivity analysis using the Global Exposure Mortality 307 Model (GEMM) (Burnett et al., 2018) . For the GEMM analysis, we also used age stratified 308 population data from GWPv3. Premature death is calculated the same as shown in Eq. 8 ; 309 however, the relative risk differs. For the GEMM model, the relative risk can be calculated as 310 shown in Eq. 10.

316
For GBD, we do not consider age-specific mortality rates or risks. For GEMM, we 317 calculate age-specific health impacts with age-specific parameters in the exposure response 318 function ( Table S1 5). We combine the age-specific results of the exposure-response function 319 with age distributed population data from GPW (CIESIN, 2017) and a national mortality rate 320 across all ages to assess age-specific mortality. 321 We calculated total premature deaths using annual average total PM 2.5 concentrations 322 derived from satellite-based estimates at the resolution of 0.1°×0.1° from van Donkelaar et al.  To the best of the 349 authors' knowledge, this variability has not been explored and its physical meaning has not been 350 interpreted. As shown in Fig. 3 ( Table S1 2); however, their emissions and concentration can be higher than BTEX 376 ( Table S7 ). Thus, alkenes would dominate R Total , leading to O x , HCHO, and PAN being produced 377 more rapidly than ASOA ( Fig. 2 b-d). When R BTEX becomes more important for R Total , the emitted 378 VOCs are more efficient in producing ASOA. Thus, the ratio of ASOA to gas-phase 379 photochemical products shows a strong correlation with R BTEX /R Total ( Fig. 2 b-d).

380
An important aspect of this study is that most of these observations occurred during  To investigate the correlation between ASOA and R BTEX , a box model using the emission 400 ratios from BTEX (Table S5), other aromatics (Table S8), IVOCs (Sect. S1), and SVOCs (Sect.  This investigation shows that the bottom-up calculated ASOA agrees with observed 473 top-down ASOA within 15%. As highlighted above, this ratio is explained by the co-emissions 474 of IVOCs with BTEX from traditional sources (diesel, gasoline, and other fossil fuel emissions) 475 and VCPs (Fig. 5) along with similar rate constants for these ASOA precursors (Table S12). 476 Thus, the ASOA/R BTEX ratio obtained from Fig. 2 ( Fig. 2 and Fig. 3), it is not surprising that the SIMPLE model  The "improved" SIMPLE shows higher ASOA compared to the default VBS 502 GEOS-Chem (Fig. 6a,b). In areas strongly impacted by urban emissions (e.g., Europe, East Asia, 503 India, east and west coast US, and regions impacted by Santiago, Chile, Buenos Aires, 504 Argentina, Sao Paulo, Brazil, Durban and Cape Town, South Africa, and Melbourne and Sydney, 505 Australia), the "improved" SIMPLE model predicts up to 14 µg m -3 more ASOA, or ~30 to 60 506 times more ASOA than the default scheme (Fig. 6c,d). As shown in Fig. 1, during intensive 507 measurements, the ASOA composed 17-39% of PM 1 , with an average contribution of ~25%. The 508 default ASOA scheme in GEOS-Chem greatly underestimates the fractional contribution of 509 ASOA to total PM 2.5 (<2%; Fig. 6e). The "improved" SIMPLE model greatly improves the 510 predicted fractional contribution, showing that ASOA in the urban regions ranges from 15-30%, 511 with an average of ~15% for the grid cells corresponding to the urban areas investigated here 512 (Fig. 6f). Thus, the "improved" SIMPLE predicts the fractional contribution of ASOA to total 513 PM 2.5 far more realistically, compared to observations. As discussed in Sect.

531
The procedure for this analysis is described in Fig. 7 and Sect. 2.3 and S3. Briefly, we 532 combine high-resolution satellite-based PM 2.5 estimates (for exposure) and a chemical transport  Additional recent work (Burnett et al., 2018) has suggested less reduction in the 548 premature deaths versus PM 2.5 concentration relationship at higher PM 2.5 concentrations, and 549 lower concentration limits for the threshold below which this relationship is negligible, both of 550 which lead to much higher estimates of PM 2.5 associated premature deaths. This is generally 551 termed the Global Exposure Mortality Model (GEMM). Using the two attribution methods 552 described above (a and b), the ASOA PM 2.5 premature deaths are estimated to be ~640,000 553 (method a) and ~900,000 (method b) ( Fig. S9 and Fig. S12 and Table S17 ).

554
Compared to prior studies using chemical transport models to estimate premature deaths 555 associated with ASOA (e.g., Silva et al., 2016;Ridley et al., 2018) , which assumed non-volatile 556 POA and "traditional" ASOA precursors, the attribution of premature mortality due to ASOA is 557 over an order of magnitude higher in this study ( Fig. 9 ). This occurs using either the IER and 558 GEMM approach for estimating premature mortality ( Fig. 9 ). For regions with larger populations 559 and more PM 2.5 pollution, the attribution is between a factor of 40 to 80 higher. This stems from  Fig. 2 and Fig. 4, see Sect. 4) leads to a more accurate than earlier 566 estimation of the contribution of photochemically-produced ASOA to PM 2.5 associated 567 premature mortality that has not been possible in prior studies. We note that ozone concentrations 568 change little as we change the ASOA simulation (see Sect. S4 and Fig. S14 ).

569
A limitation in this study is the lack of sufficient measurements in South and Southeast 570 Asia, Eastern Europe, Africa, and South America ( Fig. 1 ), though these areas account for 44% of 571 the predicted reduction in premature mortality for the world ( Table S1 6). However, as 572 highlighted in Table S1 8, these regions likely still consume both transportation fuels and VCPs, 573 although in lower per capita amounts than more industrialized countries. This consumption is 574 expected to lead to the same types of emissions as for the cities studied here, though more field 575 measurements are needed to validate global inventories of VOCs and resulting oxidation 576 products in the developing world. Transportation emissions of VOCs are expected to be more 577 dominant in the developing world due to higher VOC emission factors associated with inefficient 578 combustion engines, such as two-stroke scooters (Platt et al., 2014) and auto-rickshaws (e.g., 579 Goel and Guttikunda, 2015) .    Table 1 for further information on 685 measurements, studies, and apportionment of SOA into ASOA and BSOA.