Sensitivity of estimated NO₂-attributable pediatric asthma incidence to grid resolution and urbanicity

We calculate the fraction of PA that is attributable to NO₂ using an epidemiologically-derived health impact function. This approach is used widely throughout the literature estimating the health impacts of air pollution exposure, including the GBD, as well as the first two studies to estimate global NO₂-attributable asthma burdens (Anenberg et al 2018, Achakulwisut et al 2019). The health impact function is:

$$\begin{align*}{\text{Burde}}{{\text{n}}_{{\text{C,AG}}}} & = \sum\limits_{{\text{Grid cells in C}}} {\text{I}} {\text{n}}{{\text{c}}_{{\text{C,AG}}}} \times {\text{Po}}{{\text{p}}_{{\text{Grid}}\left( {i,j} \right),{\text{AG}}}} \\ & \quad \times \left( {1 - {e^{ - {\beta _{{\text{AG}}}}{X_{{\text{Grid cell}}\left( {i,j} \right)}}}}} \right)\end{align*}$$

where C is the country, AG is the age group, Inc is the baseline asthma incidence rate for the age group and country, Pop is the grid-cell population, β is the concentration response factor relating the NO₂ concentration with increased risk of asthma incidence, and X is the grid-cell annual average NO₂ concentration. We calculate the burden of NO₂ on PA incidence for each grid-cell, and then sum the estimated NO₂-attributable PA cases for each country.

Following Achakulwisut et al (2019), we apply a log-linear relationship between relative risk (RR) and NO₂ concentration and use a theoretical minimum risk exposure level (TMREL) of 2 ppm NO₂. We apply an RR of 1.26 per 10 ppb increase in annual average NO₂, drawn from Khreis et al (2017), the largest and most recent multi-national meta-analysis of epidemiological studies examining associations between traffic-related air pollution and asthma. The TMREL of 2 ppm is the 5th percentile of the lowest concentrations reported in the individual epidemiological studies analyzed by Khreis et al (2017). Our use of a log-linear concentration-response function may overestimate risk at low concentrations due to the steeper slope of the relationship between NO₂ and PA incidence at low concentrations, as has been shown for PM_2.5 (Anenberg et al 2012). While we did not include any health impact of concentrations lower than 2 ppb, it is possible that we have overestimated NO₂-attributable PA impacts in grid-cells with low concentrations (e.g. <5 ppb). More epidemiological research will provide insights as to the shape of the concentration-response curve at low concentrations. We calculate the asthma burden attributable to NO₂ as the difference between the burden at the total concentration (X) and the TMREL (see also figure S1, which is available online at https://stacks.iop.org/ERL/16/014019/mmedia):

$$\begin{equation*}{\text{Burde}}{{\text{n}}_{{\text{TMREL}}}}\left( X \right) = {\text{Burden}}\left( X \right) - {\text{Burden}}\left( {{\text{TMREL}}} \right).\end{equation*}$$

We used national baseline asthma incidence rates from the GBD 2017 (Soriano et al 2020) Study. Sub-national asthma incidence rates are not currently available for most countries. For population, we use the 'World Pop Age & Sex structure' dataset at 100 m resolution for the year 2015 (Tatem 2017). We use data for age groups 1–5, 5–10, 10–15, and 15–20 years, and break out the last age group into 15–18 and 18–19 years, assuming a linear distribution between ages. We sum attributable burdens across age groups for each country, since age structure does not vary substantially at the grid-cell sizes under exploration.

Spatial boundaries of urban setting categories are from the Global Human Settlement (GHS) Settlement Model grid (SMOD) dataset (Pesaresi et al 2019). Grid-cells are categorized as: (a) minimum population of 50 000 and at least 1500 inhabitants km⁻² or built up area of at least 50% (urban), (b) minimum of 5000 population and 300 inhabitants km⁻² (suburban), (c) other inhabited areas (rural), and (d) undeveloped areas. We aggregate urban settings to coarser resolutions using the category of the majority of finer cells.

Surface annual average NO₂ concentrations are from a land use regression model based on ground observations, satellite observations of NO₂ columns (from GOME2 and SCIAMACHY sensors) scaled to surface concentrations using the GEOS-Chem chemical transport model (CTM), land use properties such as road density and normalized difference vegetation index, and locations of power plants (Larkin et al 2017). This dataset provides global coverage at 100 m resolution for the 2010–2012 average (figure S2) and was also used by Achakulwisut et al (2019). To our knowledge, this is the only publicly available dataset that provides surface NO₂ concentration at high resolution (<1 km) on a global scale. Even though NO₂ has seasonal variability, the annual average provided by this dataset matches the exposure used in epidemiological studies that we follow (Khreis et al 2017). The model has a global Pearson coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error of 0.54, 5.0 ppb, and 3.7 ppb, respectively (0.52, 5.7 ppb, 4.4 ppb for North American and 0.52, 5.3 ppb, 4.7 ppb for Asia).

To investigate the sensitivity of estimated NO₂-attributable asthma incidence to spatial resolution, we summed the input population and averaged the concentration dataset (both originally available at 100 m), to 500 m, 1 km, 10 km, and 100 km. The 100 m, 500 m, and 1 km resolutions represent the higher and lower end resolutions of available land use regression and neural network model exposure datasets, respectively (e.g. Larkin et al 2017, Di et al 2020). The 10 km, and 100 km resolutions represent typical resolutions of regional (typically 10 km or smaller) and global CTMs (ranging ∼25–200 km) or satellite-derived concentration datasets, such as the Ozone Monitoring Instrument satellite product, at ∼10 km resolution (Boersma et al 2007), and Modern-Era Retrospective analysis for Research and Applications reanalysis product, at ∼0.5° resolution (Gelaro et al 2017).

We performed the simulations for each age group for two countries that have large populations with contrasting patterns of urbanicity: the U.S. and India. While nearly 82% of the U.S. population lives in urban areas, India is an agrarian country with only ∼34% of its population in urban areas and a much larger fraction of the population living in undeveloped areas (United Nations, Department of Economic and Social Affairs, Population Division 2019; figure S2). While the U.S. and India have approximately the same area of urban land, India has many more high-density urban areas, which are also more scattered. India has 3775 high-density urban clusters with a total area of 85 991 km² compared to 321 clusters in the U.S. with a total area of 85 773 km². The overall population in India is much larger than in the U.S.: the population in urban, suburban, rural, and undeveloped zones is 104 million, 46 million, 68 million, 216 million for India and 37 million, 20 million, 21 million, 584 000 for U.S., respectively (figure S3). Only 0.7% of the U.S. population lives in undeveloped areas, compared with nearly half (49.6%) for India. An additional 26% of the U.S. population lives in rural areas, compared with 16% for India in this dataset. As NO₂ is a highly spatially heterogeneous pollutant and is much higher in urban compared with rural areas, we expect that the differences in urbanicity between the U.S. and India will lead to different influences of spatial resolution on NO₂ disease burden assessment.

We hypothesize that estimated NO₂-attributable asthma cases will decrease with coarser spatial resolution. We also hypothesize that this effect is amplified in India versus in the U.S., since the urban clusters in India are smaller in area (though more numerous than in the U.S.). These hypotheses are based on the expectation that, for smaller urban clusters, averaging from smaller to larger grid boxes will yield more artificial dilution because high concentrations in a small number of urban grid-cells are averaged with lower concentrations in surrounding suburban and rural grid-cells. In contrast, big urban clusters will be less affected because most of the smaller grid boxes are categorized as urban and have high NO₂ concentrations.

Beyond the U.S. and India, we also investigate this effect in the 500 most populated cities globally. We ranked urban clusters based on 2015 population, and investigated the effect of spatial resolution by comparing estimated NO₂-attributable PA incidence calculated at two spatial resolutions: 1 km and 10 km.

Using the native 100 m resolution of the NO₂ concentration dataset, national average surface annual average NO₂ concentrations are similar for the U.S. (7.5 ppb) and India (7.0 ppb; figure S4). When aggregated to 100 km resolution, national average concentrations stay about the same for both countries, but high concentrations in urban areas are smoothed out as they are averaged with lower suburban and rural concentrations surrounding the urban centers.

To illustrate this further, we calculated city-average NO₂ surface concentrations for three U.S. cities (Washington, DC; New York, NY; and Boston, MA) and New Delhi in India at three different spatial resolution of 100 m, 10 km, and 100 km (figure S5). Similar to the national scale, the city-average concentration stays the same, but the high concentration in urban areas that coincides with large and dense populations are averaged together with low concentrations in rural areas. This artificial dilution due to aggregation will result in the entire population in the coarser grid-cell to be exposed to similar levels of concentration, as opposed to the fine spatial resolution where the high population urban areas are co-located with the high levels of concentration.

We first estimate NO₂-attributable PA cases for the U.S. and India at 100 m grid resolution using 2015 population and baseline disease rates and 2010–2012 average NO₂ concentrations. We estimate that the number of NO₂-attributable asthma cases are of similar magnitude in both countries (209 000 in the U.S. and 240 000 in India; figure 1). While India has a much larger pediatric population (486 million people in India vs 78 million in U.S.), its baseline PA incidence rate is much lower. The overall number of asthma cases per year (not limited to those attributable to air pollution) for ages 1–18 years is ∼2.1 million in India, which is 0.42% of the pediatric population. That fraction is the lowest in GBD dataset by far (potentially caused by higher uncertainties in India due to poorer access to healthcare), with a median national fraction of 1.4% of the population across all countries globally. This fraction for U.S. is 1.4%, approximately equal to the global median. Due to the large difference in incidence rates between the two countries, we base our comparison between different spatial resolutions on percentage changes, rather than absolute numbers of new asthma cases. Since the health impact function is linear regarding the incidence rate parameter, and given that we also use the same incidence rate for different resolutions, comparing results of different spatial resolutions based on relative changes is unaffected by the differences in national incidence rates.

Zoom In Zoom Out Reset image size

We next compare estimated NO₂-attributable PA incidence for both countries at the different spatial resolutions from 100 m to 100 km (figure 1). We find that the effect of spatial resolution on NO₂-attributable asthma cases is insignificant up to 1 km resolution, with <1% difference between results using 100 m and 1 km resolutions. However, we observe a much larger decrease in the total NO₂-attributable asthma burden as the grid-cell size is broadened from 1 km to 100 km (figure 1). Estimated NO₂-attributable asthma burdens calculated using 100 km resolution are 17% and 32% lower compared to the results of the analysis performed at 100 m resolution for India and the U.S., respectively. When comparing the results for calculations performed at 10 km resolution versus 100 m, estimated attributable burdens are 5% and 6% lower for India and the U.S., respectively, indicating a much larger effect when moving from 10 km to 100 km than when moving from 100 m to 10 km.

In addition to the total NO₂-attributable disease burdens, the spatial distribution of estimated NO₂-attributable asthma incidence is also affected by grid resolution, driven largely by heterogeneity in population size, NO₂ concentrations, and their co-location. Comparing NO₂-attributable asthma rates per 100 000 people calculated at 100 m and 100 km resolution, we observe large differences in areas with higher concentrations, as well as in areas with large population (figure 2). The coarse 100 km resolution smooths out the urban features in the 100 m maps, with some cities no longer visible with the coarser resolution.

Zoom In Zoom Out Reset image size

For each country and spatial resolution, we summed estimated NO₂-attributable PA cases within urbanicity categories to explore how spatial resolution influences each of these categories individually (figure 3). As the estimates vary by only a small amount for 100 m, 500 m, and 1 km resolutions, we focus on the comparison between 1 km, 10 km, and 100 km resolutions. For both the U.S. and India, NO₂-attributable asthma cases decreased in magnitude in the urban and, to a lesser degree, suburban areas when moving from finer to coarser resolution, due to the artificial dilution effect. Contrastingly, NO₂-attributable asthma cases increased in undeveloped areas with coarser resolutions, as grid-cells categorized as urban and suburban at high resolutions are aggregated to rural due to the majority rule used to aggregate urbanicity. For rural areas, NO₂-attributable asthma cases were higher in the U.S. at coarser resolutions, but were not heavily influenced by spatial resolution in India.

Zoom In Zoom Out Reset image size

Here we investigate the influence of spatial resolution on estimated NO₂-attributable asthma incidence in cities globally, since NO₂ is largely an urban pollutant with large intra-urban variation (figure 4). Previous studies have shown that NO₂ may contribute 6%–48% of PA incidence among the 125 most populated cities globally (Achakulwisut et al 2019). We focus on an expanded group of 500 most populous cities around the world to identify patterns and trends in the influence of spatial resolution on estimated NO₂-attributable asthma burdens among cities with different spatial footprints. We find a consistent increase in the ratio of estimated NO₂-attributable asthma cases calculated at 10 km versus 1 km as city size increases. The ratio is <1 for almost all the cities, indicating lower estimated NO₂-attributable asthma cases using the coarser versus finer resolution. The ratio rapidly approaches 1 (representing equal results when the analysis is performed at both resolutions) as the area of the city increases. For cities larger than 1000 km² (which is equivalent to ten 10 km cells, or one thousand 1 km cells), nearly all cities have ratios >0.8. The only cities with values <0.6 are those with area <200 km². We also found that cities located near coastlines had ratios >1. In these cities, the opposite of 'artificial dilution' occurs, since there are no NO₂ values in our dataset over water. With artificial dilution, the population in an urban area is exposed to the lower NO₂ concentration resulting from averaging higher urban concentrations with lower concentrations in suburban, rural, and undeveloped areas. In coastal cities, the suburban populations are exposed to higher NO₂ concentrations resulting from averaging concentrations in small suburban area with larger urban areas.

Zoom In Zoom Out Reset image size