Environmental Mixtures and Body Mass Index in Two Prospective US-Based Cohorts of Female Nurses

Abstract

We estimated the joint effect of particulate matter ≤2.5 μm in diameter (PM_2.5), nitrogen dioxide (NO₂), seasonal temperature, noise, greenness, light at night, and neighborhood socioeconomic status (NSES) on body mass index (BMI) in a mixture context among 194,966 participants from the Nurses’ Health Study (NHS) and Nurses’ Health Study II (NHSII) over 30 years. BMI was calculated from self-reported weight and height. Single- and multi-exposure generalized estimating equations models were used to estimate the difference in BMI per interquartile range (IQR) increase of environmental factors, and quantile g-computation methods were used to estimate joint associations. In both cohorts, we consistently observed positive associations of BMI with PM_2.5 and NO₂ concentrations as well as negative associations with light at night and NSES regardless modeling approach. A positive association with noise was only observed in NHS. Negative associations with greenness and winter temperature were only observed in NHSII. Overall, the changes in BMI per quintile increase in all eight exposures were −0.11 (−0.13, −0.08) in NHS and −0.39 (−0.41, −0.37) in NHSII, which were largely driven by air pollution and nighttime noise (18–45%) in the positive direction and NSES (>70%) in the negative direction. Future intervention on environmental factors, especially reducing PM_2.5, NO₂ and noise or improving the NSES, might be helpful to lower BMI.

Graphical abstract

Introduction

Worldwide, more than 1 billion people are obese, and this number is still increasing¹. Similar trends have also been observed in the US with an obesity prevalence of 41.9% between 2017 and 2020². If current trends persist and prevention or treatment does not improve, it has been predicted that 86.3% of US adults will be overweight or obese by 2030³. Obesity has been identified as an important risk factor for numerous health conditions, such as changes in physiological parameters (e.g. electrocardiogram) that could be monitored by soft bioelectronics devices⁴, cardiovascular diseases⁵, diabetes⁶ and even certain types of cancer⁷. To address the obesity epidemic and reduce the burden of premature death caused by chronic disease, many studies have explored potential individual-level risk factors of obesity, including individual demographic characteristics, diet habits and sedentary lifestyles^8–10. Compared to other risk factors, environmental factors can be controlled at the population level, which might be more efficient in slowing the obesity epidemic.

Several environmental factors have been linked to body mass index (BMI) in previous studies, but the findings have been inconclusive^11–14. A recent systematic review and meta-analysis concluded that increased BMI was associated with higher long-term exposure to particulate matter less than 2.5 micrometers in diameter (PM_2.5), while associations with nitrogen dioxide (NO₂) were mixed¹⁵. Higher ambient temperatures were linked to higher obesity prevalence only within specific temperature ranges (5–20℃)¹⁶. Environmental noise from road traffic and aircraft but not railways was also associated with increased waist circumference and central obesity in several studies conducted in Europe^17,18. In a systematic review of surrounding greenness with overweight and obesity, more than half of the included 57 studies observed an inverse association between greenness and overweight or obesity¹⁹. While light at night (LAN) was reported as a risk factor for overweight and obesity in some studies, most only assessed indoor nightlight exposure rather than outdoor exposure¹⁴. The existing limited evidence for cross-sectional associations between outdoor LAN and obesity-related outcomes was inconsistent across studies, possibly due to the modifying effect of age or sleep duration^20,21. Lower neighborhood socioeconomic status (NSES) has also been associated with higher BMI in a review of 10 studies²². However, there is a significant heterogeneity in the magnitude of the associations across studies²².

Notably, most previous studies only investigate these environmental factors in isolation, though no single exposure can exist in a void. Given exposures are often correlated, associations of BMI with these environmental factors may share similar mechanisms (e.g., oxidative stress¹⁵, inflammation¹⁵, metabolic disorders^15,16, stress^14,22,23, sleep deprivation^14,23, diet and physical activity patterns^16,22,23), and there may be confounding or interaction among exposures.²⁴ Therefore, studies only focusing on one of these correlated environmental factors may fail to accurately estimate the association of BMI attributed to the examined environmental factor. Taking account of the complex mixtures of exposures individuals experience is crucial for implementing population-level interventions and identifying vulnerable subpopulations. However, to date, few studies have assessed environmental exposure mixtures in relation to BMI.

In this study, we aimed to examine the association of exposure to multiple environmental factors with BMI among women from two large, prospective, nationwide US cohorts. In addition to investigating the association of each exposure independently, we also estimated the overall association of environmental mixtures simultaneously, with the goal of identifying the key factors using mixture methods.

Method

Study Population

Both the Nurses’ Health Study (NHS) and the NHSII are prospective cohorts of female registered nurses in the US. In 1976, NHS enrolled 121,700 women aged 30–55 years who resided in 11 US states, and in 1989, NHSII enrolled 116,429 women aged 25–42 years who resided in 14 US states. Once recruited, participants completed a mailed questionnaire asking detailed questions about demographic characteristics, lifestyles and history of chronic diseases. Follow-up questionnaires are mailed every two years to the participants’ residential addresses to update lifestyle and health status information. The residential addresses of both NHS and NHSII participants are distributed throughout the contiguous US. All residential addresses have been geocoded²⁵.

In the present analyses, we used 1988 as the baseline for NHS and 1989 for NHSII to align with the time period most exposure information was available. Given that age-related loss of muscle mass tends to begin around 75 years of age ^26,27, we excluded person-years when participants were older than 75 years. We also excluded person-years after participants were diagnosed with diabetes or cancer. The study protocol was approved by the Institutional Review Boards of the Brigham and Women’s Hospital and the Harvard T.H. Chan School of Public Health.

Outcome measurement

Self-reported data on height was collected on baseline questionnaires and information on weight was collected on the baseline and all follow-up questionnaires. We calculated BMI as body weight in kilograms divided by squared height in meters. The self-reported measure of body weight used to calculate BMI has been shown to be highly correlated with measured weight (r=0.96) in the cohorts²⁸. We then defined obesity as BMI≥30kg/m³ according to the World Health Organization (WHO) criteria for adults and used it as the secondary outcome in our analyses.

Exposure Assessment

We examined exposure to ambient air pollution, ambient temperature, environmental noise, surrounding greenness, LAN, and NSES. All exposures were assigned based on geocoded residential addresses updated over follow-up.

Ambient Air Pollution

Ambient PM_2.5 concentrations were predicted by a spatiotemporal model²⁹. This model incorporated i) monitor data from various government and private sources, ii) geospatial covariates, and iii) other time-varying covariates (e.g., meteorological parameters) to predict monthly PM_2.5 concentrations at each residential address. We then calculated 2-year averages prior to each questionnaire return date. The 2-year average concentrations of NO₂ were also estimated with a similar spatiotemporal model³⁰.

Ambient Temperature

Ambient temperatures were estimated using data from the PRISM climate group at Oregon State University at 800-meter resolution³¹. For the development of this model, temperature data were obtained from over 10,000 surface stations throughout the US and Canada. Monthly moving average temperatures were calculated from daily temperature predictions and assigned to participants’ residential addresses based on the 800m grid cell of the address. The 2-year summer and winter average temperatures were then calculated using monthly average temperatures during summer (June, July, August) and winter (December, January, February) over the 2 years before the questionnaire return date.

Environmental Noise

Outdoor sound levels were predicted from a geospatial noise model created by the National Park Service researchers³². This model used acoustic data from 1.5 million hours of long-term measurements from 492 urban and rural monitoring sites located across the US to predict noise levels for daytime (7am-7pm) and nighttime (7pm-7am in the next morning) separately at a 270m resolution. Given the increased likelihood that participants occupy their residence during nighttime, we selected the anthropogenic nighttime L50 sound pressure level metric as our main exposure for noise. The L50 is the A-weighted sound level that exceeded 50% of the time during the measurement period. Since the noise estimates were time-invariant over 2000–2014, we assigned them to all residential addresses throughout the follow-up, assuming the relative noise levels were temporally comparable throughout the follow-up. According to repeated noise measurement campaigns in 10 U.S. cities between 1974, 1998, 2008, and 2009, the sound levels in the US were similar over the study period, suggesting our assumption for stable noise levels was reasonable³³.

Surrounding Greenness

Exposure to greenness around each participant’s residential address was estimated using the Normalized Difference Vegetation Index (NDVI), based on Landsat satellite imagery. This index is the most widely used indicator of the quantity of vegetation on the ground ranging between −1 to 1³⁴, where a higher positive value between 0 to 1 corresponds to greater vegetation coverage, a value around 0 corresponds to barren areas of rock or sand, and a negative value corresponds to water. Negative values were set to 0 given greenness is only captured on the positive scale. We assessed buffer sizes around each address of 270m, representing the visible area directly around the home, and 1230m, representing the walkable area around the home. Given the high correlation between greenness in 270m buffer and greenness in 1230m buffer, we only included the greenness in 270m buffer in our primary analyses.

Outdoor LAN

The average annual outdoor LAN was derived using satellite data from the US Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS), which is maintained by the National Oceanic and Atmospheric Administration’s (NOAA’s) Earth Observation Group^35,36. The high dynamic range measurement was available for 1996, 1999, 2000, 2002, 2004, 2005, and 2015 and was converted to unit of radiance (nW/cm²/sr) at a resolution of 1km². We assigned a LAN value to each residential address. For addresses before 1998, we carried assigned the LAN value based on data from 1996. For addresses after 1998, we assigned exposure based on the most temporally recent past LAN measures.

NSES

The NSES was assessed with a series of variables at the census tract level, including educational attainment (% over 25 with college or higher education), income (median household income), wealth (i.e., median household home value, % families receiving interest dividends or rent income, % occupied housing units), employment status (% population 16 + years old unemployed), and racial composition (i.e., % White, % Black, % foreign-born), from the US Decennial Census or the American Community Survey closest to the questionnaire return date. We then calculated a composite score by z-scaling each component measure and summing across the indicators³⁷.

Covariates

Information on potential time-varying confounders was obtained from biennial follow-up questionnaires. We selected covariates that were associated with environmental factors and have been associated with BMI, but were unlikely to be mediators between the two, including survey calendar year, self-reported age, race (White, Black, American Indian, Asian, and other), menopause status, postmenopausal hormone use, smoking status (never, past, and current smokers), alcohol intake (1–4, 5–9, 0–14, 15–29, >30 gm/day). We conceptualized race as a social construct in our analyses. We also included information on SES at the individual level, including spouse’s education (less than high school, high school, more than high school), marital status, and living alone. Population density and census region (Northeast, Midwest, West, South) of each address throughout the follow-up were also included in our main models.

Due to the potential mediation or effect modification role of physical activity and diet quality in the associations between environmental factors and BMI, we evaluated the physical activity based on self-reported total physical activity during the follow-up. The reported time spent weekly at each activity (walking, jogging, running, bicycling, lap swimming, playing tennis, playing squash or racquetball, using a rowing machine, and engaging in calisthenics, aerobics, or aerobic dance) was multiplied by its typical energy expenditure requirements expressed in metabolic equivalents (METs), then summed all the activity figures to yield a MET hours per week score³⁸. We measured the overall diet quality with the Alternate Healthy Eating Index (AHEI). Briefly, AHEI scores were calculated using the food frequency questionnaire based on nine components: vegetables (excluding potatoes), fruits, nuts and soy, white to red meat ratio, trans fat, polyunsaturated to saturated fat ratio, cereal fiber, alcohol, and multivitamin use³⁹.

Statistical Analyses

Because statistical models for mixture analyses can be unstable and difficult to interpret, partially due to the correlation among different environmental factors, we calculated pair-wise Spearman correlation coefficients across the exposures prior to implementation of different statistical modeling approaches to examine the association of BMI and obesity with environmental factors, as described below. To adjust for potential confounders, we have included the survey calendar year, age, race, lifestyle factors (smoking status, alcohol intake), individual SES (spouse’s education, marital status, and living alone), menopause status, post-menopausal hormone use, NSES (except models considering NSES as an exposure), population density and census region in all models.

We first used single-exposure generalized estimating equations (GEE) models to estimate the change in BMI associated with an interquartile range (IQR) increase in the level of each environmental factor while accounting for repeated measures within participants. To further account for possible correlations between multiple environmental factors, we also fit a multi-exposure model with all eight exposure measures (i.e., PM_2.5, NO₂, summer temperature, winter temperature, nighttime noise, greenness in 270m buffer, LAN and NSES) included simultaneously. Given that obesity is a more typical risk factor for chronic disease than continuous BMI, we also estimated the risk ratio (RR) of obesity associated with the exposures using single-exposure and multi-exposure Poisson GEE models⁴⁰. Finally, we estimated the joint association of simultaneous exposure to environmental mixtures with BMI using a quantile-based g-computation approach⁴¹. Briefly, we first transformed the exposure into quintiles and then fitted a prespecified model that accounts for individual effects of exposures on BMI, including interaction and non-linear terms. Next, we made predictions from the prespecified model and fitted a marginal structural model to these predictions to estimate the joint association of modifying all environmental factors in the mixtures by one quintile simultaneously ⁴¹. This approach enables us to focus on the effect of the environmental mixtures as a whole rather than the effects of individual exposure when holding others constant. In addition to the joint association, this approach also provided the weights representing the proportion of partial association contributed by each environmental factor in the mixtures. These weights can be positive or negative, indicating the directions of the partial association for each environmental factor. The weights can help us to identify key individual environmental factors that contribute to the joint associations. We tested for potential non-linear associations between each environmental factor and the outcomes by fitting each environmental factor with restricted cubic splines in the multi-exposure model. Given there were no deviations from linearity, we presented the continuous exposure-response results.

To assess whether the associations differed across susceptible subpopulations, we examined possible effect modification by age (<55 years, 55–75 years), menopause status, quintile population density, quintile NSES score, self-reported physical activity (< 3, 3–8.9, 9– 17.9, 18–26.9, 27–41.9 and ≥ 42 MET hours per week), and overall diet quality measured by the AHEI⁴² by adding interaction terms between each exposure and the potential effect modifier in the multi-exposure models. We also examined the potential effect modification of the joint association using quantile g-computation models.

In sensitivity analyses, we excluded person-years with outliers of BMI, defined by any data point that’s more than 1.5 IQR points below the first quartile data or more than 1.5 IQR points above the third quartile. We also evaluated the impact of additionally adjusting for physical activity and overall diet quality measured by the AHEI⁴² in fully adjusted muti-exposure models as we believed physical activity and diet quality may potentially lie along the causal pathway between environmental factors and BMI. All analyses were conducted in SAS 9.4 (SAS Institute, Cary, NC) or R (version 3.6.3) (R Foundation for Statistical Computing, Vienna, Austria).

Results

Table 1 illustrates the characteristics of our 194,966 participants overall and by obesity status throughout follow-up (1988–2016 in NHS and 1989–2017 in NHSII). We included 1,199,906 and 1,894,362 person-years of follow-up among 88,368 and 106,598 participants respectively in NHS and NHSII. The mean age over follow-up was 61.47 years in NHS, which is about 15 years older than that for NHSII. Participants were predominantly white and lived in the Northeastern US. Compared with those classified as non-obese, those with obesity were older and more likely to be post-menopausal, currently receive hormone replacement therapy, have lower alcohol consumption, and have lower NSES scores. We found moderate correlations between NO₂, nighttime noise, and LAN (Spearman correlation coefficient: 0.41–0.68), possibly due to their common correlation with urbanicity. Greenness (NDVI) was negatively correlated with all other exposures except for NSES (Figure 1).

Table 1.

Participant characteristics at the time of each visit from 1988 to 2016 among women in the Nurses’ Health Study and from 1989 to 2017 among women in the Nurses’ Health Study II^a.

	NHS (N/nb=88,368/642,524)		NHSII (N/n=106,598/948,012)
	No obesity (N/n=78,563/514,941)	With obesity (N/n=25,960/127,583)	No obesity (N/n=96,509/744,083)	With obesity (N/n=36,120/203,929)
Age, mean, y	61.48 (7.85)	61.45 (7.48)	46.23 (9.68)	48.84 (8.92)
White, %(n)	97.77 (503,478)	97.23 (124,047)	96.52 (718,171)	96.54 (196,868)
Husband’s Education, %
< High school, %(n)	3.49 (17,984)	5.62 (7,175)	0.49 (3,644)	0.95 (1,945)
High school, %(n)	28.55 (147,025)	33.88 (43,230)	13.85 (103,042)	18.16 (37,030)
> High School, %(n)	44.72 (230,266)	35.40 (45,162)	71.31 (530,627)	61.25 (124,908)
Missing, %(n)	23.24 (119,666)	25.09 (32,016)	14.35 (106,770)	19.64 (40,046)
Marital Status, Married, %(n)	57.23 (294,723)	59.30 (75,659)	80.26 (597,190)	71.22 (145,231)
Currently Live Alone, %(n)	12.54 (64,575)	12.37 (15,787)	8.60 (64,013)	11.95 (24,377)
Post-menopausal, %(n)	90.15 (464,213)	90.90 (115,968)	32.86 (244,523)	42.03 (85,710)
Hormone Replacement Therapy
Never, %(n)	24.03 (123,729)	29.17 (37,218)	67.14 (499,559)	57.97 (118,219)
Former, %(n)	35.19 (181,194)	27.67 (35,297)	10.85 (80,721)	10.60 (21,621)
Current, %(n)	27.10 (139,558)	30.12 (38,431)	21.69 (161,370)	31.03 (63,279)
Missing, %(n)	13.68 (70460)	13.04 (16637)	0.33 (2,433)	0.40 (810)
Smoking Status
Never, %(n)	44.08 (226,972)	46.06 (58,763)	65.69 (488,801)	65.23 (133,032)
Former, %(n)	42.29 (217,760)	45.71 (58,318)	25.05 (186,429)	26.81 (54,683)
Current, %(n)	13.47 (69,384)	8.03 (10,250)	9.14 (67,987)	7.84 (15,985)
Missing, %(n)	0.16 (825)	0.20 (252)	0.12 (866)	0.11 (229)
Alcohol Consumption (gm/day)
None, %(n)	24.70 (127,209)	32.30 (41,213)	34.73 (258,432)	47.54 (96,941)
1–9, %(n)	30.02 (154,594)	26.78 (34,165)	48.16 (358,340)	42.62 (86,922)
10–29, %(n)	12.30 (63,357)	5.95 (7,596)	14.40 (107,098)	7.95 (16,209)
30+, %(n)	3.38 (17,416)	1.76 (2,251)	2.64 (19,614)	1.84 (3,748)
Missing, %(n)	29.59 (152,365)	33.20 (42,358)	0.08 (599)	0.05 (109)
Physical Activity (hours/week)
<3, %(n)	15.53 (79,955)	26.88 (34,299)	12.21 (90,838)	24.37 (49,693)
3–9, %(n)	20.26 (104,351)	23.88 (30,467)	18.74 (139,428)	24.39 (49,735)
9–18, %(n)	19.82 (102,071)	17.63 (22,499)	19.82 (147,479)	18.38 (37,474)
18–27, %(n)	13.19 (67,927)	9.55 (12,178)	13.51 (100,538)	9.82 (20,031)
27–42, %(n)	11.82 (60,870)	7.36 (9,388)	13.20 (98,225)	8.12 (16,557)
>42, %(n)	11.93 (61,442)	5.56 (7,094)	16.30 (121,277)	7.32 (14,927)
Missing, %(n)	7.44 (38,325)	9.14 (11,658)	6.22 (46,298)	7.61 (15,512)
Region of residence
Northeast, %(n)	53.07 (273,271)	54.90 (70,037)	33.66 (250,446)	33.30 (67,907)
Midwest, %(n)	16.57 (85,351)	19.17 (24,455)	31.56 (234,857)	35.40 (72,200)
South, %(n)	16.78 (86,418)	15.25 (19,458)	18.65 (138,791)	19.09 (38,933)
West, %(n)	13.57 (69,901)	10.69 (13,633)	16.13 (119,989)	12.20 (24,889)
Population density (people per mile2)	1232.50 (2732.20)	1223.22 (2434.86)	1425.94 (4089.33)	1226.65 (3116.34)
PM2.5 (μg/m3)	13.37 (3.75)	13.19 (3.67)	12.56 (3.90)	11.97 (3.62)
NO2 (ppb)	11.44 (6.97)	11.20 (6.98)	11.05 (6.97)	10.22 (6.33)
Summer temperature	22.42 (2.65)	22.37 (2.56)	22.78 (2.66)	22.85 (2.64)
Winter temperature	2.15 (6.25)	1.65 (5.89)	2.22 (5.82)	1.71 (5.65)
Median nighttime noise (dB)	42.78 (3.58)	42.74 (3.70)	43.05 (3.69)	42.96 (3.74)
NDVI (greenness)	0.36 (0.10)	0.36 (0.10)	0.36 (0.10)	0.36 (0.10)
Light at night (nW/cm 2 /sr )	25.39 (24.11)	24.65 (23.99)	25.39 (26.40)	23.23 (24.43)
NSES (z-score)	0.23 (3.90)	−0.43 (3.53)	0.23 (3.86)	−0.75 (3.43)

^a.Unless otherwise indicated, data are expressed as the percentage of (the number of observations from biennial surveys) for categorical variables and the mean (SD) for the continuous variables.

^b.N/n refers to the number of participants (the number of observations from biennial surveys).

Figure 1.

Spearman correlation coefficients among measures of air pollution, temperature, ambient noise, surrounding greenness, LAN and NSES in the US-based Nationwide Nurses’ Health Study and Nurses’ Health Study II.

In the single-exposure models (Figure 2), we found positive associations of BMI with PM_2.5 and NO₂ after adjustment for potential confounders in both NHS and NHSII. Specifically, the strength of the associations with BMI for PM_2.5 (0.04, 95%CI: 0.02, 0.06 in NHS; 0.07, 95%CI: 0.05, 0.09 in NHSII) was approximately twice as strong as that for NO₂ (0.02, 95%CI: −0.01, 0.04 in NHS; 0.03, 95%CI: 0.01, 0.05 in NHSII). On the other hand, we found negative associations of BMI with winter temperature, LAN, and NSES in both cohorts. Specifically, an IQR increment in winter temperature, LAN and NSES were respectively associated with a 0.03 (95%CI: 0.02, 0.04), 0.02 (95%CI: 0.00, 0.03), and 0.09 (95%CI: 0.07, 0.10) kg/m² lower in BMI in NHS. We also observed similar associations in NHSII. However, we found negative associations with summer temperature, nighttime noise, and greenness only in NHSII with a 0.03 (95%CI: 0.01, 0.04), 0.01 (95%CI: 0.00, 0.02) and 0.02 (95%CI: 0.01, 0.03) kg/m² lower in BMI per IQR increment in each exposure.

Figure 2.

Difference (95% confidence intervals) in BMI associated with an interquartile range increase of each exposure in single and multi-exposure models in the US-based Nationwide Nurses’ Health Study and Nurses’ Health Study II.

Compared to single-exposure models, in multi-exposure models, the observed associations with PM_2.5, NO₂, summer temperature, winter temperature, greenness and NSES remained consistent but decreased in magnitude, while the associations with LAN in NHS and nighttime noise in NHSII were not robust to the adjustment of other environmental factors. The association with summer temperature and nighttime noise turned from null to positive in NHS. When using obesity as our outcome (Figure S1), the directions of observed associations were similar to what we observed for BMI in multi-exposure models.

When we used the quantile g-computation mixture approach (Figure 3 and Table S4), we observed a negative association for the mixtures of all eight exposures combined in both cohorts. The mean change in BMI per quintile increase was −0.11 (95%CI: −0.13, −0.08) in NHS and −0.39 (95%CI: −0.41, −0.37) in NHSII. The overall associations of all eight exposures were largely driven by the strong negative weights of NSES, accounting for over 70 percent of the overall negative associations. When removing NSES from the overall mixtures, the mean change in BMI for increasing all seven exposures by one quintile was changed to 0.16 (0.14, 0.18) in NHS and 0.06 (0.04, 0.08) in NHSII. These overall positive associations were mainly driven by positive weights of PM_2.5 and NO₂. When using the mixtures of all eight exposures as exposure, we consistently observed positive association with BMI for PM_2.5, NO₂ and nighttime noise in both cohorts with weights ranging from 18 to 45 percent of the overall positive association. Besides NSES, greenness and LAN were also consistently assigned negative weights in both cohorts. The directions of the weights for the summer and winter temperatures, however, were inconsistent between cohorts or when excluding NSES from the exposure mixtures in the same cohort.

Figure 3.

The proportion of the positive or negative partial contribution for each environmental factor to the joint association of environmental mixtures with BMI in the US-based Nationwide Nurses’ Health Study and Nurses’ Health Study II.

In stratified analyses, the associations between environmental factors and BMI were generally consistent across age groups except for stronger negative associations with LAN and NSES in the older age group (55–75 years) in NHSII (Figure S2). We observed that the association between nighttime noise and BMI varied by menopausal status in NHS. This association appeared to be positive among post-menopausal participants but was negative in pre-menopausal women (Figure S3). We also found stronger negative association with greenness and positive association with PM_2.5 among those living in lower NSES in NHS. We found no evidence for population density to be an effect modifier for the association between environmental factors and BMI.

When stratified by the levels of physical activity, we observed stronger negative associations between greenness and BMI among those with higher levels of physical activity in both cohorts. In NHSII, the negative association of summer temperature was only found among participants with physical activity lower than 18 MET hours per week. However, the negative association of winter temperature was only found for those with physical higher than 3 MET hours per week (Table S2). When stratified by diet quality measured by AHEI, stronger positive associations of BMI with PM_2.5, NO₂, summer temperature, and stronger negative associations with winter temperature were observed among those within the lower quintile of AHEI compared with those with higher AHEI. In contrast, stronger positive associations of BMI with nighttime noise were observed among those with a higher diet quality (Table S3). The joint negative association of BMI with the mixtures of all eight exposures was respectively stronger among those younger than 55 years and not reported post-menopause compared with those aged 55–75 years and already post-menopause. We also observed stronger joint positive associations of BMI with the mixtures of all exposures excluding NSES among those living in areas with lower NSES (Table S4).

In sensitivity analyses (Table S1), changes in BMI association with each exposure in the multi-pollutant model were robust to the additional adjustments of physical activities or diet quality, which might act as a mediator between environmental factors and BMI. Though slightly attenuated or became less accurate with wider CIs, our results were also robust after excluding outliers in BMI or limiting our analyses to person-years when participants reported retired.

Discussion

In this nationwide analysis among almost 200,000 female adults, we explored associations of BMI with long-term exposures to air pollution, ambient temperature, noise, surrounding greenness, LAN, and NSES. Results from single-, multi-exposure and quantile g-computation mixture models consistently suggested that exposure to higher levels of PM_2.5, NO₂ and nighttime noise were associated with higher BMI in NHS, while only exposure to higher levels of PM_2.5 and NO₂ were consistently associated with higher BMI in NHSII. Furthermore, higher NSES and LAN were found to be associated with lower BMI in both cohorts regardless of the modeling approach. The evidence of the negative associations of BMI with greenness and winter temperature, however, was only consistently observed in NHSII. In general, our results demonstrate that air pollution and noise might be important risk factors for higher BMI or obesity, while higher temperatures in winter, surrounding greenness, higher LAN and NSES might decrease the risk of higher BMI or obesity. The overall association between the eight environmental exposure mixtures and BMI was negative but turned to be positive after excluding NSES from the mixtures.

Air pollution has been extensively identified as one of the risk factors for obesity in previous study. Our study added to the previous evidence of the positive association between long-term PM_2.5 and BMI among adults^15,43,44. However, our associations with PM_2.5 and NO₂ were relatively lower compared with pooled estimates in a systematic review of three papers¹⁵. This may be due to the inconsistent results among the three selected papers and the limited number of studies in this meta-analysis. Furthermore, though an increasing number of studies evaluate the association between air pollution and BMI, this evidence remains inconclusive. According to a review of 66 related papers, only 44% of them reported at least one positive association with BMI or outcomes that defined by BMI¹¹. This inconsistency may be partially due to the variation in the study regions or the age distribution of the study populations. Several possible mechanisms for the associations between air pollution and BMI have been proposed in recent studies, mainly including inflammation and oxidative stress^45,46. Given that air pollution may discourage outdoor physical activity⁴⁷, physical activity may mediate the association between air pollution and BMI. Though observed associations with air pollution were robust to the further adjustment of physical activity in our study, formal causal mediation analysis was needed to examine the mediation role of physical in the associations in future studies. Compared with participants with relatively high diet quality, participants with poor diet quality (within the first three quintiles of AHEI) might be more susceptible to the influence of air pollution.

Our findings on the positive association between nighttime noise and BMI in the NHS were generally consistent with previous analyses¹⁷. A 2018 WHO review paper reported a 0.03kg/m³ increase in BMI for each 10dB increase in road noise within a range of 40–65dB based on three cross-sectional studies⁴⁸. An updated meta-analysis in 2020, incorporating five studies, strengthened this evidence, though with larger but more uncertain estimates of associations²³. Similarly, a large study of 500,000 participants in three European cohorts observed mixed findings on the impacts of long-term exposure to road traffic on obesity defined by BMI. This inconsistency may be due to varied levels of exposure misclassification resulting from the urban-rural differences across the cohorts ²³. Unlike previous studies that focused primarily on traffic-related noise, our study is the first to provide evidence of an association of anthropogenic noise (from multiple sources including traffic, military aircraft, industry, etc.) with BMI or obesity defined by BMI in the US. This association may be explained by several potential mechanisms. One is the activation of the hypothalamic-pituitary-adrenal (HPA) by noise exposure⁴⁹, which can lead to the overproduction of stress hormones linked to increased food consumption⁴⁸. Another possible mechanism involves sleep depreciation and endocrine system disruption resulting from long-term exposure to nighttime noise, which could also affect food consumption, energy expenditure or stress hormone regulation⁵⁰. We found that participants who were older (55–75 years), post-menopause, or with higher diet quality were more susceptible to the impacts of environmental noise. The null associations observed among NHSII might be explained by the younger age distribution of its participants relative to those in NHS.

The observed link between higher surrounding greenness and lower BMI in our study is consistent with most of the previous studies¹⁹. This association, however, could be heterogeneous when using greenness measurements other than NDVI (e.g., residential proximity to greenspace, proportion of greenspace, number of parks in an area)¹⁹. Several mechanisms have been suggested to explain the association between greenness and BMI^44,51. In contrast to PM_2.5 and noise, exposure to greenness may relieve stress and motivate physical activity, which may result in lower BMI. This mechanism was in line with the stronger negative associations of greenness we observed among participants with high-level physical activity in both cohorts. In addition, higher surrounding greenness could reduce exposure to air pollution and noise, which, as previously mentioned, have been identified as risk factors for obesity. Among a nationwide cohort of 7424 middle-aged and older adults in China, PM_2.5 was found to mediate 8.85% of the association between greenness and incident overweight⁴⁴.

Though not accounting for physical and chemical environmental factors as ours, lower NSES has been linked to higher BMI or increased risk of obesity after adjusting for participants’ individual socioeconomic status⁵². In a meta-analysis of 21 studies among 1,244,438 individuals, a 1.09kg/m² increase in the mean BMI was observed among those with low NSES compared to those with high NSES²². Possible reasons include 1) disparities in health education and opportunity, 2) lower access to grocery store and healthy food and 3) less time spent on physical activity caused by the limited walkability of a neighborhood. We also found evidence that NSES could modify the association of BMI with PM_2.5, greenness as well as the mixtures of all exposures excluding NSES. Participants living in areas with lower NSES were less likely to limit PM_2.5 exposure using avoidance behaviors like regularly replacing indoor filters and may be more sensitive to the harmful effect of PM_2.5 due to generally poorer health status⁵³. The stronger negative association between greenness for those with higher NSES might be due to the high walkability of the neighborhood, which can promote physical activity and result in lower BMI⁵⁴.

The negative association between LAN and BMI observed in our study was inconsistent with previous studies, where LAN has been identified as a risk factor for higher BMI^20,21,55,56. One of the reasons for this inconsistency might be the high prevalence of night shift work among our study population. Since night-shift work is considered a key contributor to LAN exposure⁵⁷, our outdoor LAN metric may be a poor proxy for personal LAN exposure. Our observed LAN associations may also be affected by potential collinearity issues due to the relatively high correlation among LAN, NO₂ and NSES.

Finally, only a few cross-sectional studies explored the potential effect of long-term ambient temperature on BMI or obesity defined by BMI with mixed results ^58–60. All three studies assessed the ambient temperature by mean annual temperature (MAT) and found associations between MAT and BMI-related outcomes. In one nationally representative study of 422,603 adults in the US, a parabolic relationship between ambient MAT and obesity-related outcomes was observed with a positive odds ratio of obesity observed within the 5 to 20°C range⁵⁸. However, another study conducted in Korea reported a positive association between MAT and BMI without a range⁶⁰. These mixed results may be attributed to regional variations in SES, which can influence home characteristics and access to air conditioning. Additionally, the variation in the thermoneutral zone (TNZ, approximately 20°C) may also partially explain the inconclusive findings. Specifically, changes in BMI may result from the chronic imbalance between energy intake and energy consumption^61,62. Individuals exposed to the temperature within TNZ usually have basal metabolic energy consumption, while exposure to temperatures either lower or higher than the TNZ will increase energy consumption⁶³. Therefore, it is reasonable to explain the negative association of winter temperature with BMI observed in our study by the increase in energy consumption due to a decrease in winter temperature (2±6°C), which was much lower than the TNZ. The stronger negative association of winter temperature for participants with physical activity higher than 3 MET hours per week can further be explained by the higher energy consumption during physical activity as well as the higher likelihood of exposure to ambient cool temperatures during winter. Given the range of summer temperatures in our study (23±3°C) may overlap with the TNZ, the associations between summer temperature and BMI were divergent between cohorts and less consistent compared to those of winter temperature.

Our study had several limitations. Firstly, given the lack of individual time-activity information, we assigned measures of exposure at the residential address without considering time spent at home or the use of mitigations like air purifiers and indoor heating/cooling or sound insulation, which may lead to exposure measurement error. Although such measurement error was likely to be non-differential and would likely bias associations towards the null, it could still affect the estimation in the mixture analyses. We also cannot fully rule out the possibility of residual spatial confounding, especially in such national cohorts, though we have included population density and census region in our models. Although we have examined the role of some potential mediators or effect modifiers, such as physical activity and diet quality, in the associations between environmental mixtures and BMI, we were unable to test the role of psychological stress in these associations due to the unavailability of the data. In addition, we only used BMI and general obesity defined by BMI as our outcome, while waist circumference or waist-to-hip ratio has been suggested as a better indicator of obesity in relation to chronic diseases⁶⁴. However, this viewpoint remains controversial and some researchers have identified BMI, waist, and waist-to-hip ratio as similar predictors for diabetes, which is one of the most common chronic diseases related to obesity^65,66. Last, our study population was comprised of females who had a nursing degree, which potentially reduces the generalizability of our results. However, there is no evidence suggesting that the distribution of health status or health behaviors among nurses significantly differs from that of the general population⁶⁷. Additionally, their nursing degrees may enable them to provide more accurate self-reported information (e.g., height and weight for BMI calculation) compared to other cohorts and reduce the potential information bias in our analyses⁶⁸.

Nevertheless, our study has several strengths compared to previous studies. We are the first to estimate mixtures of environmental factors and their associations with BMI or obesity. We applied a quantile g-computation approach to assess the joint association of environmental mixtures⁴¹, which is robust to the exposure correlation when evaluating the joint associations. The availability of over 20 years of environmental exposure data enables us to examine the individual and joint impact of long-term exposure to environmental factors on BMI, especially for long-term ambient temperature and environmental noise, which have seldom been explored before. Finally, this nationwide study included almost 200,000 participants over a broad geographic range with a relatively wide range of levels of environmental factors across the US, which enhances the reliability of our findings.

Conclusion

In this mixture analysis, we comprehensively examined the associations of eight environmental factors with BMI, an indicator of adiposity in both the NHS and NHSII. To the best of our knowledge, this is the first study exploring the potential effect of environmental mixtures on BMI. Our findings suggest that the mixtures of PM_2.5, NO₂ and nighttime noise were associated with increased risk of higher BMI, while such increased risk might be mitigated by exposure to higher levels of surrounding greenness, LAN, warmer temperatures in winter, or living in Census tracts with higher NSES. The overall effect of the eight environmental factors was mainly driven by air pollution and nighttime noise in the positive directions and NSES in the negative directions, suggesting intervention on specific environmental factors might be helpful when trying to lower BMI and reduce the burden of obesity at the population levels.

Highlights.

• We performed a longitudinal study of 200,000 adults in two nationwide US cohorts.

• Associations between mixtures of eight environmental factors and BMI were assessed.

• The joint positive association with BMI was largely driven by air pollution and noise.

• The joint negative association with BMI was largely driven by NSES.

• The direction of the joint association of all eight environmental factors changed from negative to positive when removing NSES from the overall mixtures.

Environmental Implication.

People experience a mixture of environmental factors daily including air pollution, climate factors, noise, built environment, and neighborhood socioeconomic status. To our knowledge, this is the first exposome-wide association study investigating the effect of environmental mixtures on BMI. This is also the first study that employed mixture analysis approach to identify environmental factors that are most relevant to BMI. Our findings provided public health implications that intervention on specific environmental factors, especially reducing air pollution and noise exposure or improving the NSES, might be helpful when trying to lower BMI and address the obesity epidemic at the population level.