Environmental Inequality and Pollution Advantage among Immigrants in the United States

Abstract

Environmental inequality scholarship has paid little attention to the disproportional exposure of immigrants in the United States (U.S.) to unfavorable environmental conditions. This study investigates whether new international migrants in the U.S. are exposed to environmental hazards and how this pattern varies among immigrant subpopulations (e.g., Hispanics, Asian, European). We combine sociodemographic information from the American Community Survey with toxicity-weighted chemical concentrations (Toxics Release Inventory) to model the relationship between toxin exposure and the relative population of recent immigrants across Public Use Microdata Areas (PUMAs, n=2,054) during 2005–2011. Results from spatial panel models show that immigrants tend to be less exposed to toxins, suggesting resilience instead of vulnerability. This pattern was pronounced among immigrants from Europe and Latin America (excluding Mexico). However, our results revealed that Mexican immigrants are disproportionately exposed to environmental hazards in wealthy regions.

1. Introduction

Environmental inequality and its adverse health effects have received substantial public attention worldwide (Goodnough, 2016; Laurian & Funderburg, 2014; Nichols, 2001). Particularly the disproportionate exposure to unfavorable environmental conditions is of interest to researchers and policy makers alike. Unfavorable environmental conditions, or hazards, include natural and anthropogenic extreme events (e.g., earthquakes, chemical spills) or elements in the built environment (e.g., nuclear power plants) that pose a threat to the wellbeing of individuals (Burton et al., 1978; Hunter, 2005). Researchers have traditionally viewed immigrants as vulnerable (Bustamante, 2002) due to the following reasons: (1) weak legal standing, (2) low socioeconomic status, (3) limited English-language proficiency (Derose et al., 2007; Kim et al., 2011), and (4) racial and ethnic minority status, based on which immigrants might face institutional and individual discrimination similar to native-born minority groups (Noh et al., 2007; Viruell-Fuentes, 2007). While these factors do not automatically translate into exposure, they create social vulnerability and undermine immigrants’ resilience and capacity to withstand environmental hazards. Immigrants may find it harder to respond to a disaster, which can undermine their capacity to cope with future impacts and amplify their vulnerability (Maantay & Maroko, 2009; Wisner et al., 2003). Immigrants’ legal, socioeconomic and ethnic/racial status affects their knowledge of and access to environmental hazards awareness and assistance programs (Wisner et al., 2003). This study is guided by the following questions: How does the level of industrial pollution in an area affect the immigration rates in that area? How does this effect vary by the country/region of origin?

The paper is organized as follows: the background section provides an overview of environmental inequality research and identifies gaps relating to immigrants’ exposure to environmental hazards, that are further tested in the paper. The background section is followed by the description of data and methods used in the study, followed by the discussion of findings.

2. Background

Environmental inequality studies have shown that environmental risks are distributed unequally, with racial and ethnic minorities, low-income households, and other vulnerable populations disproportionally exposed to environmental hazards (Anderton et al., 1994; Bowen, 2002). Individual socioeconomic characteristics are important determinants of exposure to environmental hazards among the native-born U.S. population (Crowder & Downey, 2010; Mohai & Bryant, 1992). Wealthier, more educated persons are more likely to reside in environmentally safe areas (Mohai et al., 2009), while discriminatory housing market practices may expose minorities to environmental hazards regardless of their educational level (Crowder & Downey, 2010). However, these associations have been mostly tested for the general U.S. population and the relationship among socioeconomic factors, immigrant status, and exposure to environmental hazards remains unclear.

Immigrants in the U.S. are highly diverse and show varying levels of human capital (e.g., education) and social capital (e.g., access to networks) affecting economic success, residential patterns (Garip, 2008; Vigdor, 2002), and associated environmental hazard exposure at the place of residence (Crowder & Downey, 2010; Lievanos, 2015). Heterogeneity in immigrants’ education is well-recorded: immigrants from Mexico on average have 9 years of schooling as compared to Chinese immigrants with 13.6 years of schooling (Chiswick & DebBurman, 2004). Highly educated immigrants may find employment in medical, finance, or technology sectors with little exposure to environmental hazards at work and with sufficient income to afford housing in environmentally safe neighborhoods. By contrast, poorly educated immigrants may be predominantly employed in environmentally hazardous occupations in industrial manufacturing and construction with limited income, forcing these migrants to settle in inexpensive, environmentally unsafe neighborhoods.

Few studies have investigated the differential environmental inequality among the foreign-born population of the U.S. (Collins et al., 2011; Hernandez et al., 2015; Hunter, 2000). These studies reveal that immigrants’ English-language skills predict possible exposure to environmental hazards, and that U.S. counties with larger proportions of immigrants frequently contain environmental hazards. However, these studies do not distinguish between recent immigrants and the foreign-born population at large.

Social vulnerability emerges because social inequalities affect people’s exposure to risks and their ability to respond to and withstand risks (Wisner et al., 2003). Even though immigrants might be similar to native-born minorities in their socio-demographic characteristics (Hunter, 2000) and have better overall health (Gee, et al., 2006), they warrant special attention because of their legal standing and wide prevalence of limited English proficiency (Zeng & Batalova, 2016). The consequences of exposure for immigrants can be extremely debilitating because of their limited economic resources. As such, repairing damages and seeking health care can take a substantial toll on their resources and disrupt employment (Maantay & Maroko, 2009; Wisner et al., 2003). Immigrants may also lack support systems to protect themselves from the discriminatory attitudes of the majority population. Notable case studies document that environmental inequality can be exacerbated by the disproportionate siting of toxic facilities in immigrant communities (Cole & Foster, 2001; Laurian & Funderburg, 2014) and restrictive immigration policies on the grounds that excessive immigration depletes America’s ecosystems (Park & Pellow, 2011). All of these factors, combined with evidence from previous research, indicate that immigrants are potentially more susceptible to exposure to environmental hazards. To further elucidate the issue of environmental inequality among immigrants, it is beneficial to adopt a detailed approach and differentiate immigrants by country/region of origin.

In sum, while there has been some research on ethnic/racial minority groups and environmental inequality, we know little about environmental inequality among immigrants. This study contributes to the literature through a detailed investigation of environmental inequality among immigrants with an explicit focus on the following three knowledge gaps and associated research questions:

1. Existing research has paid little attention to environmental inequality among immigrants even though they represent a potentially vulnerable population and their share in the total U.S. population is increasing. We therefore ask: What is the effect of industrial pollution levels on international migration rates to PUMAs?

2. No work has investigated differential exposure to environmental hazards among immigrants originating from different regions/countries. Knowledge about differential exposure among immigrant groups will improve the scope of environmental hazards awareness and coordination programs, housing and resettlement support. Thus, we expand our first research question to ask: What is the effect of industrial pollution levels on immigration rates conditional upon immigrants’ country/region of origin?

3. We know little about the interaction among socioeconomic factors, immigrant status and exposure to unfavorable environmental conditions and consequently ask: Do immigration rates differ by the level of wealth and industrial pollution in a given area?

3. Data and Methods

3.1. Data

Sociodemographic data were obtained from IPUMS-USA (Ruggles et al., 2015) and represent single-year samples from the American Community Survey (ACS) from 2005 to 2011. The selection of the time period from 2005 to 2011 was guided by the availability of ACS data at the Public Use Microdata Area (PUMA) level, which is the lowest geographic level provided by IPUMS-USA (Ruggles et al., 2015). According to the U.S. Census Bureau definition (2015b), PUMAs represent statistical geographic areas that are nested within states and contain a minimum of 100,000 people for confidentiality purposes. The U.S. Census Bureau redraws PUMA boundaries for every decennial census. As such, the most recent year of our data sample was 2011, after which new 2010 PUMA boundaries were implemented into the ACS. The earliest year in which the PUMA variable was available was 2005. This study relies on a total of 2,054 PUMAs in the continental U.S. The ACS distinguishes international immigrants from internal migrants and the native-born population. The data obtained from IPUMS-USA were aggregated to the PUMA level. We chose to employ individual microdata aggregated at the PUMA level (in contrast to county-level aggregates) because it: (1) provides the flexibility to perform additional analyses at the household level in future research; and (2) allowed us to link important geographical and environmental variables from the Terra Populus data extract engine.

We used the Risk-Screening Environmental Indicators (RSEI) model data provided by the United States Environmental Protection Agency (EPA, 2015) as a source of environmental hazards data. The RSEI is a model representing hazard-based and risk-related impacts of chemical releases (in pounds) reported to the Toxics Release Inventory (TRI)

3.2. Variable construction

3.2.1. Outcome variables

Our six outcome variables represent: (1) the number of all international migrants per 10,000 people per PUMA; (2–6) the numbers of international migrants from the following regions/countries per 10,000 people per PUMA¹: Europe, Latin America (excluding Mexicans, referred to as Latin America in the rest of this paper), Mexico, Asia (excluding Chinese, referred to as Asia in the rest of the paper), and China. We focus our analysis on these selected subpopulations because they contribute the largest proportions to the foreign-born population of the U.S. (MPI, 2015b). Individuals were classified as international immigrants if they resided in a different country other than the U.S. one year ago. Figure 1 depicts the spatial distribution of immigrants across the United States. While the largest proportion of Mexican migrants can be found in the southwestern border regions, Chinese migrants tend to cluster in the Midwest. Immigrants from Europe, Latin America, and Asia do not exhibit clear spatial patterns.

Figure 1.

Spatial distribution of immigrants in the United States, averaged across 2005–2011.

Note: numbers reflect counts of immigrants per 10,000 people per PUMA.

3.2.2. Primary predictor

Our measure of the exposure to environmental hazards represents toxicity-weighted releases of chemicals in milligrams per cubic meter (mg/m³) per PUMA averaged for 2010–2012. We relied on the RSEI (Risk-Screening Environmental Indicators) model, produced by EPA with TRI data as input. RSEI provides estimates of the amount of toxins released and their risk to human health. The input TRI data includes air and surface water releases from industrial facilities but does not contain information about mobile sources of pollution. Reporting chemical releases and transfers to off-site facilities is mandated by the Emergency Planning and Community Right-to-Know Act of 1986 (EPA, 2015). In 2012, 623 chemicals and chemical categories were subject to mandatory reporting. RSEI weights reported emissions to produce risk estimates of exposure to certain chemicals. The constructed risk-based indicators take into account chemicals’ toxicity, sources of release into the environment, rate of decay, and the pathway to human exposure (EPA, 2015). Based on oral and inhalation exposure pathways, the EPA assigns toxicity weights to chemical releases, with higher weights for increasing levels of toxicity and associated human health risks. TRI does not include all toxic chemicals but only those with mandated reporting, assumes that the chemical concentrations are the same for outdoor and indoor exposure, and does not estimate dermal and food ingestion pathways of exposure.

The RSEI Geographic Microdata used in this paper are represented as a grid of 810 m² (28 × 28 meters) cells covering the continental U.S. We aggregated the data to the PUMA level by overlaying PUMA boundaries on the grid, summing concentrations from all the cells within each PUMA, and dividing the sum by the number of cells falling inside the PUMA. These steps were performed in PostGIS and R Statistical Environment (PostGIS, 2015; RCoreTeam, 2015). To account for skewed distribution and approximate normality, the continuous measure of toxin release was log-transformed. The variable was used as time-invariant average of 2010–2012 because the EPA only provides RSEI Geographic Microdata for these years. While EPA makes single year data for earlier periods available on request, merging different years of data is unreliable and erroneous because of frequent improvements in algorithms and toxicity weights contributing to the RSEI (Wayne Davis, Senior Environmental Scientist at EPA, personal communications). According to the TRI National Analysis conducted by the EPA (2017), total releases from the TRI facilities changed from 4,362 million pounds in 2005 to 4,114 million pounds in 2011, which makes it appropriate to use the toxin release measure as time-invariant.

A clear spatial pattern in toxin concentration is evident at the PUMA level (Figure 2). The highest concentrations are observed in the Rust Belt and near the Gulf of Mexico and are linked to steel and oil production, respectively (Chay & Greenstone, 2003).

Figure 2.

Average toxicity-weighted chemical concentrations for U.S. PUMAs during 2010–2012

3.2.3. Control variables

Prior research suggests a positive association between socioeconomic conditions in an area and exposure to environmental hazards (Been, 1994; Crowder & Downey, 2010; Freeman, 1972; Mohai & Bryant, 1992; Mohai et al., 2009). To account for socioeconomic conditions, we computed the average income and average proportion of adults employed for each PUMA. To account for the skewed distribution, the income measure was log transformed. Areas’ gender composition has been shown to be an important determinant of migration (Davis & Winters, 2001; De Jong, 2000; Shauman, 2010), and we control for PUMA’s gender composition as the proportion of males in the population. We also account for employment in manufacturing, mining and construction as well as the proportion of urban area in a PUMA, all of which are important factors in predicting the presence of environmental risks (Lievanos, 2015; Ard, 2015; Hunter, 2000). The socioeconomic measures were computed as time-varying controls based on ACS data from IPUMS-USA (Ruggles et al., 2015), while the area urban was derived from MODIS Land Cover data obtained from Terra Populus (Kugler et al., 2015; MPC, 2013) and merged with the ACS data using PUMA identifiers.

Amenity migration caused by the attractiveness of area’s natural or cultural features has received increasing attention in recent research and has been associated with unique migration patterns to the American West and Sun Belt (Gober & Zonn, 1983; Massey, 2010; Partridge, 2010). One of the distinctive features of amenity migration is the movement of people because of the draw of weather, landscapes and ambience (Glaeser, 2007; Partridge, 2010), which is often undertaken by retirees. We rely on the average maximum temperature and precipitation for the years 2010–2012 as well as retiree dependency ratio (proportion of the population ages 64+) to account for the effects of amenity migration. Temperature and precipitation measures were obtained from Terra Populus and originate from the Climate Research Unit Time Series (CRU-TS) data set (Harris et al., 2014). Summary statistics for all the variables can be found in Table 1 below.

Table 1.

Descriptive statistics for the variables used in the analysis exploring toxin exposure among various immigrant groups residing in the U.S. during 2005–2011

	Unit	TVd	Min	Max	SD	Sample mean
						2005	2006	2007	2008	2009	2010	2011
Outcome variables
International immigrants
All a	1/10000	Yes	0	662.1	51.5	51.26	52.1	50.82	53.91	47.14	50.23	53.61
Europe	1/10000	Yes	0	288.6	17.15	10.91	10.44	9.76	10.57	8.89	9.29	10.15
Latin America b	1/10000	Yes	0	220.1	16.69	7.95	7.64	7.22	7.52	6.91	7.78	7.77
Mexico	1/10000	Yes	0	220.2	16.16	10.28	8.86	7.63	7.2	5.41	4.8	5.82
Asia c	1/10000	Yes	0	273.3	21.34	10.85	12.23	12.95	13.96	12.55	13.4	15.06
China	1/10000	Yes	0	139.8	8.75	1.99	2.61	2.91	3.1	3.12	3.59	4.5
Environmental hazard
Toxin release	mg/m3	No	0	2242	56.36	9.31	9.31	9.31	9.31	9.31	9.31	9.31
Control variables
Males	%	Yes	40	62.53	1.86	48.08	48.45	48.41	48.36	48.39	48.49	48.62
Unemployed	%	Yes	0.9	19.46	2.06	4.25	3.83	3.78	3.8	5.89	6.48	5.96
Employed in manufacturing	%	Yes	0.6	38.85	5.5	11.78	11.39	11.19	11.09	10.79	10.61	10.4
Employed in mining	%	Yes	0	16.15	1.3	0.42	0.45	0.47	0.51	0.5	0.5	0.53
Employed in construction	%	Yes	0.6	22.54	2.33	6.91	7.26	7.2	6.96	6.81	6.51	6.35
Household income	USD 1,000	Yes	23	201.8	21.73	61.39	62.01	63.71	63.27	62.49	59.98	58.37
Retiree dependency ratio	%	Yes	5.5	191	8.58	22.51	22.72	23.18	23.94	24.35	24.64	25.75
Area urban	%	No	0	100	36.78	33.97	33.97	33.97	33.97	33.97	33.97	33.97
Average precipitation	mm/month	No	6.3	206	30.16	82.57	82.57	82.57	82.57	82.57	82.57	82.57
Average max temperature	deg. C	No	9.9	31.06	4.54	20.1	20.1	20.1	20.1	20.1	20.1	20.1

Notes:

^aAll immigrants combined;

^bexcluding Mexico;

^cexcluding China;

^dTV column indicates whether variable’ s values were time-varying (Yes) or time-constant (No); Values for Min, Max, and Standard Deviation (SD) refer to the complete sample across all years (2005–2011); Sample size: PUMA-years: n=14,378, PUMAs: n=2,054; In a preliminary analysis we included variables reflecting English-language skills and educational attainment as important predictors of residence in environmentally unsafe neighborhoods (Lievanos, 2015; Hunter, 2000). However, these variables showed high correlation with other variables of interest and were excluded from the final analysis to avoid issues of multicollinearity. With these adjustments, the Variance Inflation Factor (VIF) diagnostic remained below 1.67, demonstrating that multicollinearity does not bias our estimates.

3.3. Estimation strategy

Both toxin release and immigration show distinct spatial patterns (Figures 1 and 2). A formal test (Morans I, Anselin, 1995) revealed significant spatial autocorrelation in the residuals of regular OLS models. Ignoring spatial autocorrelation may lead to biased inferential statistics and increases the chance of committing a Type I error (rejecting the null hypothesis if it is true) (Anselin et al., 2000). As such, we employed spatial panel models that effectively account for spatial autocorrelation as well as heterogeneity across time (Call & Voss, 2016; Millo & Piras, 2012). We fitted spatial panel models using the package splm within the R Statistical Environment (Call & Voss, 2016; Millo & Piras, 2012). The splm package has been frequently used in the analysis of time-space dependent phenomena (Call & Voss, 2016; Millo & Piras, 2012).

We estimated spatial lag models with time period fixed effects. Spatial lag specification refers to a model that creates a spatially lagged dependent variable based on a spatial weights matrix. Spatial lag models account for spatial spillover effects, acknowledging that the values of the dependent variable in one unit depend on the values of the dependent variable in neighboring units. As such, spatial lag models account for network effects and the presence of migrant systems (DeWaard et al., 2012) by statistically accounting for the fact that places with high concentrations of migrants tend to attract the largest migrant inflows.

Within our spatial models, a neighborhood was defined as comprised of the eight nearest PUMAs in line with prior research (Franco-Lopez et al., 2001; McRoberts et al., 2002; Schelling, 1969). In contrast to other neighborhood definitions, this specification guarantees that all units have the same number of neighbors, even when located in border regions or spatially isolated. We performed a robustness test with various neighborhood definitions (e.g., 12-nearest neighbors, queen, rook) and found similar effects. The spatial matrix was constructed using a polygon shapefile of PUMA boundaries obtained from Terra Populus (Kugler et al., 2015; MPC, 2013)

Our models predict the number of immigrants relative to the total population (y_it) for each location (i) and year (t) and can be written as:

$$y_{it}=\alpha +{\beta }_0+{\beta }_n(x{ }_n)+{\beta }_1(to{x}_i)+\rho W{y}_{it}+u_{it}$$ (1)

In equation 1, year fixed effects are represented by coefficient α. The parameter β₀ represents the intercept while β_n constitutes a vector of regression coefficients of various sociodemographic and geographical characteristics (x_n). The control variables are included as time-varying (it) and time-constant (i) measures, as indicated by the generic subscript n. The parameter β₁ reflects the effect of the time-invariant measure of toxin release (tox_i) on migrant prevalence as the central focus of this analysis. The coefficient ρ is the spatial lag of the dependent variable and W is the spatial weights matrix, while u_it constitutes the conventional error term.

In the spatial lag model the value of the dependent variable y at location i depends on the values of predictor variable x at location i and the values of x at other locations, which, in turn, depend on the values on x in other neighboring locations (Anselin & Rey, 2014). As such, the value of y depends on all the values of x. The indirect effect (also known as spatial multiplier effect) takes into account the effect of x on y from neighboring locations. As a result, the total effect of x on y in spatial lag models is comprised of the direct and indirect effects (Anselin, 2003; Anselin & Rey, 2014; Kim et al., 2003). The direct effect of a predictor variable x on the dependent variable y is reflected by the conventional regression coefficient (β(x)). The indirect (β(x′)) and total (β(x″)) effect, due to the spatial multiplier effect, can be estimated using the following expressions:

$$\beta (x\text{'})=\beta (x)\rho /(1-\rho )$$ (2)

$$\beta (x")=\beta (x\text{'})+\beta (x)$$ (3)

In equation 2,β(x) is the regression coefficient for variable x, and p is the spatial lag. The total effect (β(x″)) consists of the sum of indirect (β(x′)) and direct (β(x)) effects (Equation 3). We estimate the indirect, direct and total effects of toxin release on the immigration rates across PUMAs.

Finally, to explore spatial patterns in the association of toxin release and the presence of immigrants, we employed a geographically weighted regression (GWR). We employed GWR for those outcome variables for which we observe a significant interaction effect between industrial pollution and income. Unlike the traditional regression approach that produces one estimate for the entire study area, GWR relies on a user-specified distance/lradius to produce estimates that vary across the study area. We relied on a bi-square adaptive kernel to fit a local regression model for each PUMA. The results of a GWR allow for an exploration of spatial patterns that are otherwise hidden in global estimates (Fotheringham et al., 2002; Gollini et al., 2015). GWR has been effectively used to explore spatial patterns of interactions between environmental and social factors (Gilbert & Chakraborty, 2011; Mennis & Jordan, 2005; Nawrotzki & Bakhtsiyarava, 2016).

4. Results and Discussion

4.1. General trends

Average migration rates for each study year reveal general trends with respect to the timing of migration streams from each country/region (Table 1). From 2005 to 2011 there were on average 51 international immigrants per 10,000 people in PUMAs. The average migration rate varied little across time, reaching 54 per 10,000 in 2011 and showing a small dip to 47 immigrants per 10,000 people in 2009, likely related to the economic recession and the U.S. housing bust (Ellis et al., 2014; Johnson, 2013).

Table 2 shows results from spatial lag models with time fixed effects estimating PUMA-level presence of various immigrant groups. The coefficient estimates for the various control variables in the model for all migrant groups combined (Model “All” in Table 2) confirm patterns observed in prior research. For example, highest immigration rates were observed for PUMAs with higher proportions of males, low levels of unemployment and employment in the primary (mining) and secondary (manufacturing, construction) sectors (Zeng & Batalova, 2016), characterized by a less affluent, younger population (Hunter, 2000), higher levels of urbanization, and a warmer and slightly more humid climate (Poston et al., 2009) Although largely similar, some of these trends differ in magnitude and direction by immigrant groups, likely as a result of occupational preferences and differences in immigrants’ socioeconomic conditions (MPI, 2015a, 2016a).

Table 2.

Spatial panel models predicting the proportions of various immigrant groups relative to the native-born population among U.S. PUMAs during 2005–2011

	International immigrants
	Allc		Europe		Latin Americad		Mexico		Asiae		China
	b	sig	b	sig	b	sig	b	sig	b	sig	b	sig
Intercept	−77 72	***	−18.19	***	−10.73	**	−3.53		−39.34	***	−3.84
Malesb	31.45	***	4.2	***	2.45	***	4.68	***	11.3	***	2.36	***
Unemployedb	−20.47	***	−5.44	***	−0.47		−5.99	***	−4.83	***	−1.01	*
Employed in manufacturingb	−13.47	***	−3.78	***	−1.7	***	−0.1		−2.8	***	−0.97	***
Employed in miningb	−25.8	***	−5.89	***	0.43		−3.89	***	−7.74	***	−1.04
Employed in constructionb	−38.42	***	−11.53	***	−0.39		8.6	***	−17.79	***	−5.77	***
Household income	−3.57	*	4.69	***	−1.78	***	−4.98	***	2.01	**	0.17
Retiree dep. ratiob	−5.68	***	−0.96	***	−1.16	***	−1.55	***	−1.56	***	−0.2	*
Area urbanb	2.41	***	0.12	*	0.84	***	0.69	***	0.48	***	0.14	***
Average precipitationb	0.41	**	0.12	*	0.9	***	−0.78	***	0		0.01
Average max temperatureb	9.78	***	0.3		2.17	***	5.27	***	1.88	***	−0.43	*
Effects of toxin release
Toxin release (direct effect)	−1.17	***	−0.28	***	−0.51	***	−0.11		−0.15		−0.01
Toxin release (indirect effect)a	−1.22		−0.19		−0.65		−0.07		−0.12		−0.01
Toxin release (total effect)a	−2.39		−0.47		−1.16		−0.18		−0.27		−0.02
Model Statistics
Spatial lag	0.51	***	0.41	***	0.56	***	0.4	***	0.45	***	0.38	***
AIC	244,296		216,425		212,762		213,278		222,011		198,214
R2	0.4		0.21		0.37		0.28		0.26		0.13
Pseudo R2 Spatial	0.27		0.14		0.22		0.22		0.16		0.07

Notes: Sample size for all models: PUMA-years: n=14,378, PUMAs: n=2,054;

^athe computation of indirect and total effect of the spatial multiplier does not permit obtaining significance test statistics and values should be interpreted in terms of effect size;

^bparameter estimates reflect an incremental change of 10% in the predictor variable;

^cAll immigrants combined;

^dexcluding Mexico;

^eexcluding China. As a more conservative goodness of fit measure, we report the Pseudo R² Spatial, which accounts for spatial dependence within the data. Pseudo R² Spatial is computed by squaring the correlation between the observed and fitted values of models’ outcome variables (Anselin and Rey, 2014).

^*p<0.05;

^**p<0.01;

^***p<0.001

4.2. Toxin release effects

Including the toxin release measure in all the models allows us to answer our first research question (What is the effect of industrial pollution levels on international migration to PUMAs conditional upon PUMAs’ socioeconomic characteristics?). In contrast to the assumption that immigrants in the U.S. are socially vulnerable and thus more likely to migrate to environmentally unsafe places (Hunter, 2000; Lievanos, 2015), our results indicate that international immigrants tend to migrate to less polluted areas, and therefore demonstrate a pollution advantage. A one unit (log mg/m³) increase in the toxin concentration in a PUMA is associated with a decrease in the number of international immigrants by 1.17 per 10,000 people.

In addition to the direct toxin release effect, the spatial lag models permit estimation of an indirect effect, mediated through spatial dependence on neighboring PUMAs. This indirect effect (1.22 per 10,000 people) is similar in size to the direct effect, leading to a doubling of the total effect with a decrease in the total immigrant population of 2.39 per 10,000 people for each one-unit increase in toxin release. Our finding of a pollution advantage reveals a possible pathway for the dynamics underlying the healthy immigrant paradox: lower exposure to environmental toxins may translate into better health outcomes among immigrant populations (Evans & Kantrowitz, 2002), which in turn explains the observed health advantage of immigrants over the native-born population in the U.S. (Hunt et al., 2003; McGlade et al., 2004; Wingate & Alexander, 2006).

As a robustness check, we performed a jackknife procedure, randomly excluding 10% of PUMAs from the sample and re-estimating the models across 100 iterations (Nawrotzki et al., 2015). We also tested for influential time periods by omitting one year at a time from the sample and refitting the models (Jennings & Gray, 2015). Both procedures indicated a high degree of robustness of our results.

Focusing on specific immigrant groups allows us to answer our second research question (What is the effect of industrial pollution levels on immigration rates conditional upon immigrants’ country/region of origin?). Our results reveal contrasting patterns of exposure to environmental hazards among various immigrant groups. For example, immigrants from Europe and Latin America show a pollution advantage and tend to migrate to less polluted areas. Employment in a polluting sector (construction, mining, manufacturing) is an important predictor of exposure to environmental hazards, and European immigrants as a group are less likely to be employed in such industries than both the native-born population and other immigrant groups (USCB, 2015a). Moreover, as opposed to the native-born U.S. population and overall foreign-born populations, European immigrants are on average older and wealthier (MPI, 2015a; USCB, 2015a), with implications for residential choice and migration to environmentally safe areas. Generally, Latin Americans tend to favor service occupations and are less likely to be employed in industrial and transportation-related occupations compared to other immigrant groups in the U.S. (MPI, 2016b). A trend towards amenity migration and a documented desire for a higher quality of life as reflected in better schools, affordable housing, and low crime rates (Massey, 2010), may make Latin American immigrants more likely to reside in environmentally safe areas.

In contrast, for Mexican, Chinese, and Asian immigrants we found no relationship between toxin release and their concentration in PUMAs, indicating exposure patterns comparable to their native-born counterparts. Employment in polluting industries such as oil-production and mining (Bauer et al., 2002), as well as lower average socioeconomic status (USCB, 2015a) may explain Mexican immigrants’ lack of a pollution advantage. The confluence of below-average education and low income (USCB, 2015a) may force Mexicans to settle in areas of low rental prices but high environmental burden (cf. Crowder & Downey, 2010). However, as a group, Asian immigrants are often employed in the fields of science, engineering and technology (MPI, 2016a), and show overall higher levels of income (MPI, 2016a), making the lack of a pollution advantage a surprise.

In sum, our results reveal that international immigrants in the U.S. overall do not disproportionally migrate to environmentally unsafe areas.

4.3. Interactions

Area-level wealth and socioeconomic conditions have been identified as important determinants of exposure to environmental pollutants (Freeman, 1972; Hunter, 2000). In order to answer our third research question (Do immigration rates differ by the level of wealth and industrial pollution in a given area ?), we interacted the toxin release measure with average PUMA income. Table 3 reveals a positive directionality for most immigrant groups (except for European migrants), which implies that the association between toxin release and immigration rates increases in strength for wealthier PUMAs. However, the interaction coefficient only reached significance for Mexican migrants.

Table 3.

Interactions between toxin release and income on the expected proportion of immigrants within each PUMA

	Inc		Tox		Inc × Tox
	b	sig	b	sig	b	sig
International immigrants
Alla	−3.63	*	−1.16	***	0.56
Europe	4 72	***	−0.28	***	−0.32
Latin Americab	−1 79	***	−0.51	***	0.03
Mexico	−5.08	***	−0.09		0.86	***
Asiac	1.97	**	−0.14		0.41
China	0.16		0		0.07

Notes:

^aAll immigrants combined;

^bexcluding Mexico;

^cexcluding China; Interactions were obtained by estimating full models with all socioeconomic and geographic controls as per Table 2; Inc = estimates of average PUMA income (logged); Tox = estimates of toxin release (logged); Inc × Tox = estimates of interaction effects between the measures of income and toxin release. Toxin release and income variables were centered;

^*p<0.05;

^**p<0.01;

^***p<0.001.

Figure 3 reveals that in PUMAS of average wealth (50 %ile, green line) the association between toxin release and the presence of Mexican migrants is negative but weak. This negative trend (slope) becomes stronger among poor PUMAs (10 %ile, red line), where the association resembles the general relationship observed for the combined immigrant population as well as for European immigrants. In these economically marginalized regions, Mexican immigrants display the pollution advantage and tend to migrate to less polluted areas. However, in the wealthiest PUMAs (90%ile, blue line), this relationship is inverted. In groups of these high-income PUMAs, Mexican immigrants disproportionally migrate to areas of highest toxin release. This pattern may be explained by Mexican immigrants’ generally low socioeconomic status as well as discriminatory housing market practices (Crowder & Downey, 2010) and unique legal status (Rivera-Batiz, 2004), preventing them from residing in environmentally safe areas when housing prices are high and rental markets are controlled by discriminatory elites.

Figure 3.

Interaction effect between toxin release and income on the expected proportion of Mexican immigrants within each PUMA

We employed a GWR to explore local patterns of the interaction between toxin release and PUMA-level wealth on the presence of Mexican immigrants. The strongest positive interaction effects (colored red and orange) are observed along the Gulf Coast, in the West and Midwest (Figure 4), which are known for their oil-production and steel manufacturing (Mani & Wheeler, 1998). In comparatively high-income PUMAs in those areas, Mexican immigrants are disproportionately exposed to toxin release and do not exhibit a pollution advantage.

Figure 4.

Local estimates of the effect of interaction between toxin release and income on the presence of Mexican immigrants in a given PUMA.

Note: GWR point estimates for each PUMA are derived from a local sub-sample of n=144 neighboring units. The optimal bandwidth of n=144 was determined empirically via cross-validation.

5. Conclusion

In this paper we have investigated whether immigrants in the U.S. migrate to environmentally hazardous areas. This study builds on previous research on environmental inequality by expanding the definition of vulnerable populations to immigrants of various origins. One strength of our approach is the use of spatial panel models that help account for complex space-time dynamics in the relationship between immigration rates and toxin exposure. Our results reveal that international immigrants in the U.S. overall do not disproportionally migrate to environmentally unsafe areas. On the contrary, they tend to migrate to less polluted areas and thus demonstrate a high degree of resilience. However, this pollution advantage is not shared by all immigrant groups alike and only emerges for immigrants from Europe and Latin America (excluding Mexico), perhaps due to the unique occupational selection and socioeconomic characteristics of these immigrant groups. (MPI, 2015a, 2016b). Interaction models and results from GWR reveal that in groups of high-income PUMAs, Mexican immigrants are disproportionally exposed to environmental hazards, thus not exhibiting any pollution advantage. These findings suggest that immigrants cannot be directly equated with vulnerable minority groups when it comes to exposure to environmental hazards and that immigrants from different regions/countries exhibit different patterns of exposure. Future research may employ qualitative and quantitative approaches to investigate the underlying dynamics that differentially protect certain immigrant groups from toxin exposure. Future research should also investigate another understudied group-native-born ethnic minorities — and their experiences of environmental inequality.

Despite a careful modeling approach and multiple robustness checks, a number of limitations deserve mention. First, the RSEI data did not include pollution from mobile sources and do not include all known toxins but only those with mandated reporting. Second, a classic problem in geographic research is the modifiable areal unit problem in that there can be shifts in parameter estimates (Higgs & Langford, 2009; Maantay, 2007; Openshaw, 1984), so using a finer scale than PUMAs would result in more accurate estimates of exposure. We use individual microdata aggregated to the PUMA level (in contrast to county-level aggregates) to be able to investigate the observed relationships at the household-level in future research. Finally, we did not take into account shifts in immigration policies, which could have impacted immigration rates for certain immigrant groups (Massey, 2010), although our time fixed effects account for some of these changes.

Our study has important policy implications. The results show that Mexican immigrants are especially exposed to environmental hazards in high-income areas. Given Mexican immigrants’ overall low levels of education and limited English proficiency, we recommend that communities should publish relevant information for toxin exposure prevention and the introduction of hazardous facilities in their neighborhood not only in English but also in Spanish. Besides preventing disproportionate siting of hazardous facilities efforts should also be made to mitigate the existing consequences of such siting, including housing support, planting of trees and protecting water sources. Such policies may help ameliorate environmental inequality among vulnerable subpopulations, such as Mexican migrants, and will help create a society where every individual can exercise their human right to a safe and healthy environment (UNEP, 2016).

Highlights.

• Overall, immigrants in the US reside in environmentally safe areas

• Exposure to environmental hazards depends on immigrants’ country/region of origin

• Mexican immigrants migrate to environmentally unsafe areas in wealthier regions