The relationship between monthly air pollution and violent crime across the United States

ABSTRACT

Recent evidence suggests a relationship between short-term pollution exposure and crime, with a particular emphasis on aggressive behavior. However, the previous analyses are limited in geographic scope. In this paper, we estimate the effect of fine particulate air pollution (PM${}_{2.5}$) exposure on crime across 99% of counties in the contiguous United States. We combine monthly data on crime, PM${}_{2.5}$, and satellite-derived smoke plumes for a ten-year period. We use adjusted satellite-based landscape fire smoke plume data as an instrument for overall changes in ${\mathrm{P}\mathrm{M}}_{2.5}$. Our findings are consistent with previous research and suggest that increases in ${\mathrm{P}\mathrm{M}}_{2.5}$ raise violent crime rates, and specifically assaults. Our results indicate the effect is relatively homogeneous across the U.S. However, we find the effect is positively correlated with county median age, suggesting older populations are more susceptible to changes in air pollution. Our results indicate a need for more research on the physiological and social mechanisms behind the measured effects.

KEYWORDS:

• Crime
• air pollution
• fire smoke

JEL CLASSIFICATION CODE:

• Q53

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Jude Bayham http://orcid.org/0000-0001-5383-2714

Notes

1 Herrnstadt et al. (2016) study this relationship in Chicago and Los Angeles while Burkhardt et al. (2018) study the relationship in 391 counties in the United States.

2 ${\mathrm{P}\mathrm{M}}_{2.5}$ is defined as the mass concentration of particulate matter with aerodynamic diameters smaller than 2.5 μm.

3 The EPA pollution monitor data is limited to monitor locations and is not collected on all calendar days. Not all counties have a pollution monitor.

4 Although most fire smoke plumes are visible from space, correlating observed smoke plumes with surface-level ${\mathrm{P}\mathrm{M}}_{2.5}$ is a notoriously difficult problem in the atmospheric sciences (Lassman et al. 2017; Brey et al. 2018). Satellite-observed smoke plumes may not affect surface-level ${\mathrm{P}\mathrm{M}}_{2.5}$. We ultimately develop a method to correct this measurement error and we describe this methodology in Section 3.

5 The results are robust to county-by-year-by-season fixed effects.

6 We should note that the lack of statistical significance across regions could be due to a mismatch between the Census Bureau regions used in this paper and regional PM composition, or other measurement error.

7 Several other studies have used upwind pollution as a source of exogenous variation (e.g. Moeltner et al. 2013; Deryugina et al. 2016; Keiser, Lade, and Rudik 2018).

8 The data is publicly available from https://aqs.epa.gov/aqsweb/airdata/download_files.html.

9 Kriging is a geostatistical interpolation method. Kriged surfaces have been used in previous research to estimate air pollution exposures (Jerrett et al. 2005; Janssen et al. 2008; Lassman et al. 2017) and ordinary kriging has been shown to effectively predict air pollution across large-geographic areas (Beelen et al. 2009). Universal kriging or other predictive models that incorporate land use or satellite based covariates provide improved prediction performance and more spatially specific surfaces (Beelen et al. 2009; Mercer et al. 2011; Di et al. 2016; Young et al. 2016). For reference, we use all available federal reference method (FRM) and FRM-corroborated daily ${\mathrm{P}\mathrm{M}}_{2.5}$ observations in the EPA AQS monitor network. We then krige the observations to a 15 km grid, which produces a gridded estimate of daily-average ${\mathrm{P}\mathrm{M}}_{2.5}$ concentrations for each day during the study period. Kriging parameters are as follows: sill $=2.6$, range $=8.5$, nugget $=0.1$. The parameters were determined using a k-fold cross validation with 10 folds, optimizing R² while also maintaining minimal mean bias and mean absolute error.

10 For example, the city of Chicago, which spans 12 counties, had a total of 17,400 violent crimes in 2015. This implies an average of only 1450 violent crimes per month across the 12 counties in the nation's most violent city. These statistics were downloaded from the Chicago police department here.

11 https://www.claimsjournal.com/news/west/2012/11/07/216957.htm

12 To be sure, in Table 5, we present a robustness check in which we drop observations where a fire was burning in a county in a particular month.

13 This means that ${\mathrm{P}\mathrm{M}}_{2.5}$ on a particular day must be elevated above the 95th percentile of the county-specific background ${\mathrm{P}\mathrm{M}}_{2.5}$ mean, assuming a normal distribution, for the own-county adjusted HMS variable to equal one.

14 Surface-level ${\mathrm{P}\mathrm{M}}_{2.5}$ is less than 1.64 standard deviations of the within-county background mean on 83% of the days in which a smoke plume is present (HMS=1 days). This means that smoke plumes observed in satellite imagery only increase surface-level PM${}_{2.5}$ more than 1.64 standard deviations above the mean 17% of the time, highlighting the need for this adjustment procedure.

15 The spatial distribution of the adjusted data is displayed in Burkhardt et al. (2018). We also show the spatial distribution of the adjusted data relative to the unadjusted data in Burkhardt et al. (2018). We find no discernible patterns in the areas that are adjusted.

16 In Burkhardt et al. (2018), we test the latter assumption and find that the background means between county c and county k have a correlation coefficient of 0.32. Thus, the two are not perfectly correlated. Similar studies such as Keiser, Lade, and Rudik (2018) and Deryugina et al. (2016) have used upwind air quality as an instrument for within county air quality. In these papers, the authors instrument pollution in a given county i with pollution in an upwind neighboring county. Our IV strategy is slightly different since we observe whether a county was covered by a smoke plume but not the elevation of the plume. We exploit the proximity of the neighboring county for smoke exposure (requiring that both counties lie under the same plume) but assume that non-smoke factors impacting PM2.5 in the neighboring county are exogenous to county i. Our instrument thus depends on correlation in smoke plumes across counties, not correlation in PM2.5 across counties, as an upwind analysis would utilize. Smoke plumes travel hundreds to thousands of miles and span multiple neighboring counties. Thus, an analysis that limits the sample to upwind counties only would likely produce virtually identical results precisely because a given smoke plume is usually affecting many contiguous counties, both upwind and downwind. Moreover, Bondy, Roth, and Sager (2018) study the relationship between daily pollution exposure and crime in London for a two year period. Like ours, their primary identification strategy relies on a series of high-dimensional fixed effects. However, they instrument pollution with wind direction as a robustness check. While wind direction is likely uncorrelated with daily crime, in their preferred instrumental variables model, the first stage F-statistic is fairly small (around 13) indicating that wind direction is not the strongest of instruments for daily air pollution.

17 We tried other sets of fixed effects such as county-by-year-by-season but these absorbed too much of the variation in ${\mathrm{P}\mathrm{M}}_{2.5}$ and crimes.

18 The first stage control function estimates are presented in Table A.2 in the Appendix. Column 1 of Table A.2 shows the first stage using the raw HMS variable and Column 2 the first stage estimates using the neighbor-adjusted HMS variable. The HMS variables in both models is statistically significant, however, the coefficient in Column 2 is seven times larger than the coefficient in Column 1. Indeed, the model in Column 1 suggests that an additional HMS day increases monthly average ${\mathrm{P}\mathrm{M}}_{2.5}$ by only 0.10 μg/m³. This is a very small amount given the mean ${\mathrm{P}\mathrm{M}}_{2.5}$ level is 9.42 μg/m³. In contrast, the model in Column 2 suggests that an additional HMS day increases the monthly average ${\mathrm{P}\mathrm{M}}_{2.5}$ by 0.73 μg/m³. Given these findings and our discussion in Section 3, we are confident that the adjusted HMS variable produces a stronger instrument than the raw HMS variable. Estimates using an own-county adjusted HMS variable are similar. Two-stage least squares estimates of the model produce Cragg-Donald weak identification F-statistics of 9,930 for the raw HMS variable and 68,000 for the adjusted HMS variable, respectively.

19 The sample size differs between specifications due to the number of zeros or missing values in the dependent variables within fixed effects clusters.

20 Results using Bureau of Economic Analysis regions are similar. The effect is statistically insignificantly largest in New England, the omitted category, which is consistent with our demographic estimates below. One caveat of these results is that the effects may be heterogeneous at a different geographic scale than the Census Bureau regions or our data may lack sufficient power to identify regional heterogeneity.

The results presented in Table 7 suggest virtually no differences in treatment effects across the U.S. A possible explanation for the lack of regional heterogeneity is that pollution affects behavior through physiological mechanisms, which are uniform in geographic scope, but are modified by socioeconomic conditions. To test for heterogeneous effects due to socioeconomic differences, we generate dummy variables indicating quartiles of county-level demographics and interact these dummy quartiles with ${\mathrm{P}\mathrm{M}}_{2.5}$. The results of the demographic heterogeneity analysis are presented in Table 4. Column 1 suggests that the effects are largest (statistically insignificant) in areas with the highest median household income (F-statistic p-value of joint significance for four presented coefficients is 0.009). Column 2 provides evidence that the effects are largest (again, statistically insignificant) in the largest metropolitan areas. The coefficients on each of the population quartile interactions are negative and jointly statistically significant with the coefficient on ${\mathrm{P}\mathrm{M}}_{2.5}$ (F-statistic p-value $=0.007$).

This result is expected given there are more crimes in more populated areas. However, the results remain significant for regions with low populations. Column 3 indicates that the effect is stronger in counties with a higher fraction of the population that is white. The effect is 50% smaller in counties in the first quartile for fraction of the population that is white. Column 4 indicates that average educational attainment does not significantly impact the effect of pollution on aggressive behavior. Column 5 indicates that the effect is generally largest in the second quartile of poverty rates, or the effect is largest in relatively wealthy counties, consistent with the income effects presented in Column 1. Finally, column 6 indicates that the effect is statistically significantly largest in counties with older populations. The latter result indicates that older populations are more susceptible to physiological impacts of pollution.

21 Short (2017) reports the latitude and longitude of wildfire ignitions as well as the discovery and end dates. We exclude county-days with an active fire burning.

Additional information

Funding

Authors acknowledge support from NASA project number NNX15AG35G.