Air Pollution, Social Deprivation, and Mortality

A Multilevel Cohort Study

Abstract

Background: 

It is becoming increasingly evident that exposure to air pollution and its adverse effects are not equitably distributed. Our goal was to investigate the role of social deprivation in explaining the effect of neighborhood differences in level of air pollution fine particulates (PM_2.5) on mortality when the indicators of social deprivation are measured at both individual level and at neighborhood level.

Methods: 

All inhabitants registered in Oslo, Norway on 1 January 1992 in the age group 50–74 years (n = 105,359) constitute the study base. We used an air dispersion model (AirQUIS) to estimate levels of exposure in the period 1992–1995 in all 470 administrative neighborhoods. These data were linked to Census, educational, and death registries. Deaths were recorded in the period 1992–1998.

Main Results: 

PM_2.5 was associated with most neighborhood-level indicators of deprivation, as was most clearly seen for type of dwelling and ownership of dwelling. The effect of PM_2.5 on mortality was to some extent explained by these indicators independently of the corresponding individual-level indicators.

Conclusions: 

Findings from this study suggest that socially deprived neighborhoods have higher exposure to air pollution. Deprivation at both the individual and neighborhood level is associated with air pollution, accounting for some of the excess mortality associated with air pollution in these neighborhoods.

A number of studies have in recent years provided evidence of adverse health effects attributable to air pollution. Studies using time-series analysis have reported positive associations between short-term air pollution exposure and daily mortality.^1–4 A few cohort studies have extended this evidence.^5–8 Increasingly, it has come to the attention of researchers and policy makers that the distribution of exposure to air pollution is not equitable,⁹ but this inequity has until recently received little formal epidemiologic attention. Typically, studies of air pollution have adjusted for socioeconomic position by any available indicator in an attempt to achieve estimates that are claimed to be independent of the social environment. Other studies have investigated the associations between outdoor air pollution and mortality after adjustment for socioeconomic position at the individual level or at the district level. In the first type of study,^6,10,11 it has been customary to account for socioeconomic measurements such as occupational class, income, or education. In the latter, each person is assigned a value based on aggregate characteristics of all individuals in this person's place of residence.^12–19 In a recent report from Los Angeles, Jerrett et al¹⁹ were able to include a number of contextual variables at zip code level in addition to individual-level socioeconomic position.

Few multilevel studies have taken both individual and neighborhood level into account. The influence of the social environment is likely to be multidimensional, depending on the specific causal model for the disease of interest and the theoretical relevance of a socioeconomic indicator in any given local context.²⁰ As important environmental exposures such as outdoor air pollution are spatially located, the degree to which these exposures are socially patterned is likely to be located at this level as well. In this study we investigated the role of social deprivation at both individual and neighborhood level in explaining the effect of air pollution on mortality.

METHODS

Population

Our study population was a cohort of all inhabitants age 50–74 years who lived in the municipality of Oslo, Norway in 1992, comprising a total of 105,359 persons. Among these, 18,531 (18%) had missing data either because they had not responded to all the Census questions or did not live in Norway before 1992. Data on housing conditions were available only from the 1980 Census; to make this information valid for the period of air pollution exposure, those 27,898 (26%) who had moved between 1980 and 1992 were excluded. Data on air pollution were linked with the tax, educational, and the death registers, facilitated by the Norwegian identity code and provided by Statistics Norway. Deaths were recorded in the period 1992–1998. All death certificates are registered with Statistics Norway, with no missing cases. There were 10,055 deaths in the analyzed cohort. We investigated the effect of fine particulates (PM_2.5) on causes of death grouped into those plausibly related to air pollution, such as cardiovascular causes (ICD-9: 390–459, ICD-10: I00-I119), lung cancer (ICD-9: 162, ICD-10: C34), lung cancer (ICD-9: 162, ICD-10: C34), and chronic obstructive pulmonary disease (COPD) (ICD-9: 490–496 and ICD-10: J40-J47), and those causes not plausibly related to air pollution, such as psychiatric causes of death (ICD-9: 290, 292–302, 304, 306–319; ICD-10: F00–F09, F11–F69), stomach cancer (ICD-9: 151; ICD-10: C16), and violence (ICD-9: 800–999; ICD-10: S00–Y90, Y95–Y98). These causes were chosen a priori because they were shown in a previous study of the same population to be socially patterned by different pathways.²⁰

Exposure Assessment

The Norwegian Institute for Air Research has estimated average concentrations of air pollutants monthly in the period 1992–1995. Concentrations (μg/m³) of particulate matter pollutants (PM_2.5 and PM₁₀) and nitrogen oxides (NO₂) were computed hourly. An AirQUIS system developed at the Norwegian Institute was implemented to estimate the concentration of pollutants in 468 neighborhoods in the city.²¹ This air pollution dispersion model uses emission, geographic information (such as proximity to roads and topography), and meteorologic information to calculate hourly estimates. These concentrations describe the level of air pollution to which the average population in every neighborhood is exposed. Input data for this AirQUIS model are provided hourly and geographically determined. This includes emissions from road particles, vehicles, and home heating. The meteorologic model is based on hourly measured meteorologic values. Person-weighted median concentrations for each neighborhood are calculated by using concentration values and number of inhabitants living in buildings defined by grids within each neighborhood and additional grids from other points within each neighborhood. These additional points were homogeneously and uniformly generated (“random draw”). Average exposure values for each individual residing in any of these neighborhoods were then calculated over this 4-year period. These averaged values for each individual were divided into quartiles for each air pollutant.

The AirQuis model has been evaluated to assess the long-term (month-years) estimation of local ambient air pollution by comparing it with available measurements from monitoring stations in Oslo.²² Measured and modeled long-term mean concentrations were correlated for all pollutants, with the best correlation and lowest bias for nitrogen dioxide (NO₂) and PM_2.5. For PM_2.5 the correlation was 0.79 during winter and 0.57 during summer. The correlations increased with the length of averaging period, and all pollutants performed better at the urban background stations compared with the traffic stations. Modeling is probably less demanding for background levels than for levels affected by strong line sources leading to steeper spatial concentration gradients, which to some extent explains the lower agreement between the model and measurements at the traffic stations.

Indicators of Deprivation

The following indicators of deprivation were recorded: education (1 = primary education only, 0 = more than primary education), equalized household income below median (below = 1, above = 0), occupational class (1 = manual, 0 = nonmanual), status of ownership of dwelling (no = 1, yes = 0), type of dwelling (flat = 1, attached house = 0), and crowded household (1 = less than 1 room per capita, 0 = more than 1 room per capita). Education and income were recorded in 1990. Household income was the taxable income in the household minus claimed tax according to the revenue authorities, divided by the number of consumer units in the household. Consumer units were 1.0 for the first adult, 0.5 for each child, and 0.7 for a married adult and each child of 16 and older living at home without income. Zero or negative household income was treated as missing. Occupational class, ownership of dwelling, type of dwelling, and crowded household were recorded in the 1980 Census. Each individual-level indicator was aggregated to neighborhood level by calculating the percentage in each neighborhood with primary education only, being in manual class, not owning their dwelling, having household income below median, living in a flat, and having less than 1 room per capita. In the multilevel logistic regression, these aggregated values of the deprivation indicators were standardized by generating Z-scores, which makes it possible to compare these effect estimates across neighborhoods as one standard deviation increase in the neighborhood-level effect of social deprivation.

Statistical Analysis

The models were stratified by sex and age-adjusted by 5-year age bands. We fitted 5 models for each deprivation measure: a model including just the individual covariate, one including just the contextual covariate, one including the individual covariate together with neighborhood air pollution, one including the contextual covariate together with air pollution, and the final model including both the individual and contextual covariates together with neighborhood air pollution. Spatial models were fitted using WinBUGS with individuals (level 1) nested in neighborhoods (level 2).²³ The dependent variable y_ij was a dichotomous outcome (death), with a logit link used to link the probability of an event to the covariates. To model the spatial correlation between the small areas we assumed a convolution prior for the random effects; the random effect for each area was taken to be the sum of a spatially structured component v_j and an unstructured or heterogenous component u_j. A conditional autoregressive prior was assigned to the v_j and a normal prior to the u_j.^24,25

where Xβ is the fixed part of the model (including age and any other individual or neighborhood covariates), v_−j denotes the spatial effects for all areas other than area j, JOURNAL/epide/04.03/00001648-200711000-00007/ENTITY_OV0456/v/2021-02-05T040115Z/r/image-png_j is the mean of the v_j for the areas bordering area j, and n_j is the number of neighbors of area j. Flat (improper) priors were assigned to the fixed parameters β_p and uniform U (0,2) priors were assigned to the square root of the variances σ_u² and σ_v². All parameter means and credible intervals were estimated from 2 chains of length 20,000 after burn-ins of length 5000.

RESULTS

There were fewer individuals in the deprived category for most indicators except ownership of dwelling and type of dwelling (Table 1). More individuals had missing data for occupational class than for other covariates. Those with missing data on housing conditions generally had higher mortality rates. A substantial part of the excluded population had moved between 1980 and 1993; these individuals had mortality rates similar to the total population. Men included in the analysis had somewhat lower mortality than the men not included, whereas for women the 2 groups had similar mortality.

TABLE 1: Distribution of Individual-Level Independent Variables and Age-Adjusted Mortality Rates (per 10,000 Person-Years) for Each Indicator of Deprivation in Men and Women

The mean (±SD) concentration (μg/m³) of PM_2.5 in the study population was 14.2 ± 3.6, with a median of 13.3. The interquartile ranges were 1st, 6.6–11.3; 2nd, 11.3–13.3; 3rd, 13.3–17.3; and 4th, 17.3–22.3. Mean ± SD values for the proportion in each neighborhood being deprived were 0.35 ± 0.19 for primary education only; 0.34 ± 0.17 for manual class; 0.46 ± 0.14 for income below median; 0.70 ± 0.32 for not owning their dwelling; 0.56 ± 0.38 for living in a flat; and 0.44 ± 0.16 for living in crowded household. Average number of inhabitants per neighborhood was 1010. The average size of neighborhood was 283,000 m².

Figure 1 presents concentrations of PM_2.5 (range, 25–75%) by 6 equal groups of neighborhood-level deprivation presented separately for each indicator. For neighborhood-level proportion in manual class, household income below median, and primary education only, those with values below the lowest quantile lived in neighborhoods with higher concentrations of PM_2.5. For the proportion not owning their dwelling, there were clearer gradients of exposure to PM_2.5 with best discrimination in the most deprived neighborhoods for education and in the least deprived neighborhoods for ownership of dwelling. Neighborhood-level living in a flat seemed to discriminate best, with a clear dose-response relationship across levels of PM_2.5 (mean values ranging from 12.1 in the lowest percentile category to 17.0 in the highest). There was a slightly increased exposure in the most deprived quantile for proportion in crowded households, but this covered a much narrower range of concentrations of PM_2.5.

FIGURE 1.: Concentration of PM_2.5 at place of residence (estimated at neighborhood level) by 6 quantiles of neighborhood level proportion in manual class (A), with income below median (B), with primary education only (C), not owning their dwelling (D), living in flat (E), and in crowded household (F).

The correlation between PM_2.5 and the neighborhood-level indicators ranged from low for proportion in manual class (r = 0.04), proportion with primary education (r = 0.14), and proportion with crowded households (r = 0.13), to moderate with proportion not owning their dwelling (r = 0.31), proportion having income below median (r = 0.28), and proportion living in flat (r = 0.48). PM_2.5 was highly correlated with NO₂ (r = 0.87) and PM₁₀ (r = 0.95).

Tables 2 and 3 present age-adjusted effects of PM_2.5 (by one quartile increase in μg/m³), individual-level deprivation, and neighborhood-level deprivation, analyzed separately for each indicator in men and women. In men the individual-level effects of each indicator were rather similar except that crowded household had a smaller effect size. Individual-level housing tenure and type of dwelling explained more of the neighborhood-level variance than the other indicators. Crowded household explained less than the others. The neighborhood-level effect of most indicators of deprivation was of similar size and explained a substantial part of all the neighborhood-level variances in mortality; the exception was overcrowding, which showed no effect on all-cause mortality. None of the individual deprivation measures had much influence on the effect of PM_2.5. Neighborhood-level deprivation attenuated this effect of PM_2.5 to a larger degree, especially for income below median, housing tenure, and type of dwelling.

TABLE 2: Age-Adjusted Fixed Effect on All-Cause Mortality (Odds Ratio and 95% Coverage Interval) of PM_2.5, Indicators of Deprivation at Individual and Neighborhood Level and Heterogeneous and Spatial Neighborhood Level Variances (95% Coverage Intervals) in Men Age 50–74 Years

TABLE 3: Age-Adjusted Fixed Effect on All-Cause Mortality (Odds Ratio and 95% Coverage Interval) of PM_2.5, Indicators of Deprivation at Individual and Neighborhood Level and Heterogeneous and Spatial Neighborhood Level Variances (95% Coverage Intervals) in Women Age 50–74 Years

In women the effect of individual- and neighborhood-level deprivation was generally similar to that seen in men. The neighborhood-level variances dropped in a similar pattern to men. The neighborhood-level spatial variance fell by an average of 51% when the indicators of neighborhood-level deprivation were added one at a time compared with the variance in a model including only age. The average decrease was greater (62%) among the deprivation indicators excluding overcrowding. The effect of PM_2.5 was attenuated to only a minor degree for most indicators at the individual level, except for housing tenure and type of dwelling. There was a larger attenuation when adding indicators of deprivation at neighborhood level, especially for education, income below median, and type of dwelling. The independent effect of neighborhood-level deprivation in the fully adjusted model could to some degree be explained by individual-level deprivation. When we used individual-level education, occupational class, and household income with 5 categories rather than the dichotomized variables, there was a minimal impact on the results. We saw no significant cross-level interaction between indicators of deprivation. There was a statistically significant negative interaction between levels of PM_2.5 and the proportion with household income below median, and a positive interaction of PM_2.5 with the proportion living in a flat and owning their dwelling among women. The effect of PM_2.5 on both sexes combined dropped to 1.05 (1.02–1.08) after including all indicators at individual and neighborhood level.

In the cause-specific analysis, there was an increased risk in both sexes for causes plausibly related to air pollution, with a particularly strong effect for chronic obstructive pulmonary disease (Table 4). The attenuation in effects was strongest when the neighborhood-level indicators of deprivation were added. For the group of causes not plausibly related to air pollution, there was an association with violent causes but not psychiatric diseases and gastric cancer. For violent causes, adding indicators of deprivation at both levels had much less impact than seen for the causes plausibly related to air pollution.

TABLE 4: Hazard Ratio (95% Confidence Interval) of Interquartile Increase in PM_2.5 for Causes of Death Adjusting for Age and Sex Only and Then Including Each Indicator of Deprivation at Individual and Neighborhood Level Separately in Men and Women Age 50–74 Years

DISCUSSION

Neighborhood-level PM_2.5 was related to a number of neighborhood-level indicators of deprivation, particularly for neighborhoods where a large proportion of people were living in flats and not owning their dwelling. The effect of air pollution on mortality was to some extent explained by several neighborhood-level indicators of deprivation independently of individual-level deprivation, suggesting that air pollution has a role in the contextual effect of neighborhood-level deprivation on mortality. The results seemed consistent in both sexes.

There are several strengths of this study. We were able to investigate the role of social deprivation at both individual and neighborhood levels in explaining the effect of air pollution on mortality. Compared with other cohort studies, we used neighborhoods as the level of aggregation. The effect of air pollution using this level has been shown to be less biased than when higher levels of aggregation are used.^6,26 The estimates of the effect and the change found were rather similar in both sexes. Additionally, we were able to include indicators of deprivation, such as housing conditions, that are plausibly more closely linked to determinants that are amenable to interventions addressing environmental inequity through urban policy.

In recent years, several epidemiologic studies have found effects of social deprivation at the area level on ill health, independent of the composition of individual characteristics in the areas.²⁷ There is sparse evidence as to which level of aggregation leads to the strongest effects being observed.²⁸ Whether this is at the county or neighborhood level probably depends on the causal model of interest. We think in many respects that neighborhood level is a relevant level for the social clustering of air pollution risk. Air pollution varies locally and so does social deprivation. Indicators of housing conditions are particularly important for this purpose because they are meaningful at both the individual and neighborhood level.^29,30 Housing conditions are powerful indicators of socioeconomic position at the individual level, measuring both exposures experienced within the house and, indirectly, through indicating the availability of resources to buy property. Housing conditions may be directly related to other factors at the neighborhood level, such as air pollution as shown here. Household income may have interpretations along similar lines.²⁸ In this sense, neighborhoods with poor household income (where few own their house and most live in flats) may experience increased risk of air pollution because low social investment in such places (such as fewer measures to reduce air pollution) coincide with these places being more affordable.³¹ The study by Jerrett et al¹⁹ has several similarities with ours: It was an intraurban study giving larger effect estimates of exposure to PM_2.5 than previously reported across metropolitan areas. The study also used neighborhoods. Some of the ecologic variables were similar. They did not look at the same ecological variables, such as aspects of housing conditions. As indicators of neighborhood-level deprivation, housing conditions may be particularly useful because they may tap into development of the urban structure, including proximity to roads.

It has been a common place in environmental epidemiology to adjust for socioeconomic position by available indicators, on the understanding that the adjusted estimates are less confounded and provide support for a causal effect.⁹ In this study, we have seen that social deprivation at the neighborhood level has an independent effect on mortality and attenuates the effect of air pollution further than seen by social deprivation at an individual level. We do not think this necessarily means there is less evidence of a genuine air pollution effect. However, social deprivation and air pollution could be related to mortality in different ways: area-level social deprivation may be a distal cause mediated by more proximate factors such as air pollution. On the other hand, if the effect of air pollution on mortality is noncausal, adding social indicators at the area level may pick up important confounders that may explain this relationship. We are not able, based on this study alone, to judge the relative importance of social confounding or causation. Both social deprivation and air pollution are clustered at neighborhood level. As Oakes³² suggests, a number of issues remain unresolved when making causal inferences about the contextual effects of neighborhoods and health. It may be problematic to estimate contextual neighborhood effects that are independent from the individual people residing in them. And even for apparently exogenous factors such as aspects of the physical environment, in this case air pollution, these may be endogenous to the local communities as people have ultimately selected or been selected to live in these places. Future research may take advantage of new multilevel models where residential history through people's life course is taken into account.^33,34

The study used Census data on the whole population, which obviously did not include smoking history. Thus, it was not possible to assess whether smoking is a confounder in the analysis. We checked the proportion of smokers in each neighborhood measured in a survey of 14,495 individuals in 2001, and correlated this with the estimates of air pollutants.³⁵ The correlation was 0.18, suggesting that smokers only to a small degree reside in areas with high levels of air pollution. In another cohort study of air pollution and lung cancer from Oslo in the same age group,⁷ there was no attenuation in effect after adjusting for smoking. We were unable to analyze this cohort with more recent exposure data.

The AirQUIS model employs information on topography, road density, housing density, and meteorology to calculate hourly concentrations of air pollutants in geographic grids. These data may be socially patterned, making the air pollution exposure effectively proxies for socioeconomic circumstances and not valid indicators of actual air pollution. Places within each neighborhood that are close to roads may have higher levels of air pollution but are also more likely to be deprived, as they may be less popular. The AirQUIS model applies weights according to number of people who live close to roads within each neighborhood. It was our intention here to study neighborhoods deprivation, and we considered neighborhood-level air pollution to be the most relevant level as well. We were not able to look at where people spent their work time. The probability of dying is highest in the oldest age groups who are most likely to spend more of their time at their home place, making the impact of this potential misclassification less pronounced.

The ability of this model to predict long-term levels has been evaluated and was found to be a valid instrument in epidemiologic studies, giving somewhat better prediction for PM_2.5 and NO₂ than for PM₁₀.²² The cause-specific results suggest that the AirQUIS model was reasonably able to discriminate between causes that are and are not plausibly related to air pollution. The fact that violent causes with a short induction time were related to air pollution might suggest that spatial clustering of social disadvantage at the neighborhood level (responsible for this outcome) may concur with residential air pollution. This is in contrast with gastric cancer, which is linked to social position in childhood mediated through infection with Helicobacter pylori, but was not associated with air pollution in the present study.³⁶ The concentration-response relationship in various causes has been reported elsewhere.³⁷

In conclusion, neighborhood-level PM_2.5 and its effect on mortality was related to a number of neighborhood-level indicators of deprivation, particularly for neighborhoods where a large proportion of people were living in flats and not owning their dwelling.

ACKNOWLEDGMENTS

We thank Statistics Norway and the Norwegian Institute of Air Research for preparing the data.