STUDY DESIGN AND POPULATION

We used data from the Generation R Study, a population-based birth cohort from pregnancy onward, set up in the urban area of Rotterdam, the Netherlands (Kooijman et al. 2016). A total of 8,879 women were enrolled during pregnancy, and 899 women were recruited shortly after the delivery. Children were born between April 2002 and January 2006. We included only singleton pregnancies for this research, resulting in 9,610 mother–child pairs. Children were examined at several time points after birth. We gathered data for the Generation R brain MRI substudy. A brain MRI protocol was developed that included brain morphology, DTI, and rs-fMRI.

A select group of children enrolled in the Generation R study, who were between the ages of 6 and 10 years, was recruited for the first wave of the brain MRI substudy (White et al. 2013). The whole cohort of children between the ages of 9 and 12 years was recruited for the second wave of the brain MRI substudy (White et al. 2018). Due to the logistics of the visits, there was some overlap on the age of the participants between waves.

The first wave included 1,932 children. This was an oversample based on certain maternal exposures during pregnancy, such as cannabis, nicotine, selective serotonin reuptake inhibitors, depressive symptoms, plasma folate levels, and child behavior problems, such as attention deficit hyperactivity disorder, pervasive developmental problems, dysregulation problems, and aggressive problems. Exclusion criteria comprised contradictions for the MRI procedure, severe motor or sensory disorders, neurological disorders, head injuries with loss of consciousness, and claustrophobia. Of the 1,932 children, the invitation call was not answered for 155, participation was refused for 447, and 5 could not participate due to contraindications for the MRI procedure. Of the 1,325 that attended the MRI visit, we excluded twins and children with poor brain MRI data quality and major abnormalities. We had brain MRI measurements available for 1,070 children. Finally, after excluding those without air pollution estimations, we included 783 children in the project.

In the second wave, the brain MRI measurements were offered to all children still involved in the Generation R study who were 9 to 12 years old (n = 8,548) (White et al. 2018). The invitation call was not answered for 911; participation was refused for 244; 1,531 did not visit the research center; and 1,875 did not undergo the MRI procedure. We excluded twins and children with poor brain MRI data quality and major abnormalities. We had brain MRI measurements available for 3,888 children. Finally, after excluding those without air pollution estimations, we included 3,857 children in the project. We obtained brain structural morphology data for 3,133 of them, structural connectivity data for 2,954, and functional connectivity data for 2,197. The reduction in the number of children for which data of each imaging modality were available was due to attrition along the MRI session.

For 387 children, we had information on air pollution and brain imaging data from both waves. However, we did not combine that information because of the small sample size and because the scanners differed between waves.

AIR POLLUTION ASSESSMENT

We used LUR models to estimate the levels of air pollution at all reported home addresses of each participant during the pregnancy and childhood following a standardized procedure in the context of the ESCAPE (European Study of Cohorts for Air Pollution Effects) and the TRANSPHORM (Transport related Air Pollution and Health impacts — Integrated Methodologies for Assessing Particulate Matter) projects (Beelen et al. 2013; Brunekreef 2008; de Hoogh et al. 2013; Eeftens et al. 2012a; Jedynska et al. 2014b; Yang et al. 2015). We performed 2-week measurements of NO_x and NO₂ in three different seasons (i.e., warm, cold, and intermediate). We performed these measurements between February 2009 and February 2010 at 80 sites spread across the Netherlands and Belgium (Appendix Figure A1; see Additional Materials on the HEI website) (Cyrys et al. 2012). Additionally, at 40 of those sites, we carried out measurements of PM₁₀ and PM_2.5 (Eeftens et al. 2012b). We calculated PM mass between 10 and 2.5 μm (PM_COARSE) by subtracting PM_2.5 from PM₁₀. We measured different parameters in the PM_2.5 filters: PM_2.5absorbance, the composition of PM_2.5 including PAHs, OC, Cu, Fe, Si, and Zn, and the oxidative potential of PM_2.5 (de Hoogh et al. 2013; Jedynska et al. 2014a; Yang et al. 2015). PM_2.5absorbance is a measure of the reflectance of the PM_2.5 filter, and it is highly correlated with elemental or black carbon, also referred to as soot (Cyrys et al. 2003). The oxidative potential of PM_2.5 is a quantification of the potentiality of fine particles to induce oxidative stress (Yang et al. 2015). We evaluated oxidative potential using two acellular methods: OP^DTT and OP^ESR (Yang et al. 2015).

Measurement sites were selected to represent the anticipated spatial variation of air pollution at the home addresses of subjects in the epidemiological studies set up in the study area. Measurement sites reflected a large diversity of potential sources of air pollution (e.g., population density, traffic intensity, industry, proximity to ports), selecting regional background, urban background, and street sites. Regional background sites were located outside major urban areas and were not directly influenced by traffic sources. Urban background sites were located inside an urban area, but at least 50 meters from major roads. Street sites were selected at building facades representative for homes, in streets with traffic intensities of 10,000 or more vehicles per day. Street sites were overrepresented compared with the fraction of addresses on major roads, as the goal was to describe spatial variation in the area of which traffic is a main source. Measurements were performed simultaneously in 20 NO_x–NO₂ sites and in 10 NO_x–NO₂–PM sites. Thus, we performed four rounds of 2-week measurements to complete a measurement period for one season (i.e., to complete the measurements in 80 sites for NO_x–NO₂ and 40 sites for PM). Each round included different types of sites, for example, regional, urban background, and traffic sites. To adjust for the temporal variability of concentrations, a centrally located reference site was chosen at a regional background location that was not directly influenced by local sources, and measurements were taken over the entire year. For each round, we applied a temporal correction in two steps: (1) calculating the difference between the concentration for a 2-week specific sampling period and the annual average of the continuous-reference site, and (2) subtracting that difference from each measurement (Beelen et al. 2013; Brunekreef 2010; de Hoogh et al. 2013; Eeftens et al. 2012a; Jedynska et al. 2014b; Yang et al. 2015). Next, for each pollutant, we averaged the concentrations of the three 2-week corrected measurements resulting in one annual mean concentration.

We assigned to each monitoring site a variety of potential land-use predictors, such as proximity to the nearest road, traffic intensity on the nearest road, and population density. Then, we applied LUR models by using linear regression models to determine which combination of predictors best explained the annual average concentration of each pollutant (Appendix Table A1). The cross-validation R² of the LUR model varied between pollutants and ranged between 0.31 for PAHs and 0.89 for PM_2.5absorbance (Appendix Table A2). We then applied LUR models to each geocoded address where participants lived during the period of interest (i.e., since conception until date of the MRI assessments) to estimate the exposure of each air pollutant at each address (Figures 1 and ​and22 and Appendix Figure A2). Addresses were geocoded using the cadaster of all buildings of the Netherlands. For the participants who were recruited shortly after birth, we used their birth address for the pregnancy period. We considered the number of days that the participant spent at each address and weighted the pollution exposures accordingly. Then, we calculated the mean exposure of each pollutant for each participant for several time periods: (1) for the entire pregnancy period, (2) from birth until 2 years of age, (3) from 2 until 5 years of age, (4) from 5 until 9 years of age, and (5) for the entire childhood period (i.e., from birth until the date of the second-wave MRI assessment).

Open in a separate window

Figure 1.

Spatial distribution of density of addresses per municipality across the study area: (A) At recruitment; and (B) At the MRI assessment at 9–12 years. To ensure anonymization, a municipality was not depicted if too few participants lived there (i.e., addresses at the MRI assessment at 9–12 years were also distributed in several other municipalities across the whole study area).

Open in a separate window

Figure 2.

Air pollutant concentrations in Rotterdam where most of the participants lived: (A) NO_x; (B) NO₂; (C) PM_2.5; and (D) PM_2.5absorbance.

No historical data were available for most of the pollutants. Therefore, we could not perform back and forward extrapolation of the concentrations to match the exact periods of interest. For most analyses, we assumed that the spatial contrast remained constant over time. This assumption was previously demonstrated in the Netherlands for a period up to 8 years (1999–2007) (Eeftens et al. 2011), and in Great Britain for a period up to 18 years (1991–2009) (Gulliver et al. 2013). Exceptionally, in a couple of analyses, we could perform back and forward extrapolation of the concentrations for some pollutants that had historical data available (i.e., NO_x, NO₂, PM₁₀, PM_COARSE, PM_2.5, and PM_2.5absorbance). To do this, we used daily data from the seven continuous-reference monitoring sites (Brunekreef 2012). We combined the estimated yearly air pollution concentration at each home address i (Cyearly, i) with time-specific measurements from the seven continuous -reference monitoring sites. This was done by averaging the daily concentrations during different periods: (1) the year corresponding to the LUR yearly concentration (Cyearly), and (2) each time period ti considered (Cti). The ratio Cti/ Cyearly constituted the temporal component of the model. For each pollutant, the concentration estimated at the home address i (Cti, i) during each time period was estimated as the product of the temporal (Cti/Cyearly) and spatial (Cyearly, i) components. When data from the continuous-reference monitoring sites were unavailable for a given pollutant, we used measurements for another one during the same time period as a replacement. We based the choice of the pollutant used to back and forward extrapolate another one on an extensive study of temporal correlations between pollutants that were simultaneously available. Specifically, we used NO_x when PM_2.5absorbance was missing and PM₁₀ when PM_2.5 was missing.

BRAIN MRI

In both waves, we wanted participants to be familiarized with the magnetic resonance environment to reduce the possibility of not being able to complete the scanning session. Therefore, each child underwent a half-hour mock scanning session prior to the actual MRI (White et al. 2013, 2018). To limit the movement of the head, we accommodated children by: (1) providing them with a thorough explanation before the scanning session, (2) offering to let them watch a movie or listen to music during the session, and (3) placing cushions around the head to fix it in a comfortable way.

The brain MRI protocol involved different measurements, such as structural morphology (i.e., brain volumes, cortical thickness, and cortical surface area) using T₁-weighted images, structural connectivity (i.e., white matter microstructure) using DTI, and functional connectivity (i.e., connectivity score between brain areas) using rs-fMRI. In both waves, we used a scanner specifically dedicated for the study. In the first wave, we used a 3 Tesla General Electric scanner (Discovery 750; GE Worldwide, Milwaukee, WI); in the second wave, we used a different 3 Tesla General Electric scanner (Discovery MR750W; GE Worldwide, Milwaukee, WI). Both scanners had an 8-channel head coil. At each wave, the entire session took about 26 minutes. The refined protocol minimized the burden on the participant while allowing for high-quality data collection.

COGNITIVE FUNCTION

We assessed the cognitive function of children on the day of the scanning, or shortly after, by using an array of tasks from the Dutch version of the NEPSY-II (Brooks et al. 2010). The tasks were chosen to tap into specific domains, such as: attention and executive functioning, language, memory and learning, sensorimotor function, and visuospatial processing (White et al. 2013). We individually tested the children in a quiet room by trained investigators. As described in the section Statistical Methods and Data Analysis, for the post hoc mediation analysis we selected two specific domains: the attention and executive functioning and the memory and learning domains.

NONRESPONSE ANALYSIS

Compared with the study population initially recruited (9,610 children), children who were included in the current analyses were more likely to have parents with a higher education level and a higher household income, from Dutch ethnicity, younger, with lower body mass index, higher height, and lower psychological distress, as well as more likely to have mothers with a partner, nulliparous, who did not smoke during pregnancy, and with higher IQ (data not shown). Selection bias potentially arises when the study includes only the population with available data on exposure and outcome instead of the full initial cohort recruited at pregnancy. To correct such bias, we used inverse probability weighting in each analysis (Weisskopf et al. 2015; Weuve et al. 2012). First, we imputed the missing covariates to all eligible participants recruited at pregnancy. Then, we used all the available information to predict the probability to participate in each analysis. Finally, we used the inverse of those probabilities as weights in the analyses. This way, the results would be representative for the initial population of the cohort.

EXPOSURE–OUTCOME ASSOCIATIONS

AIR POLLUTION EXPOSURE AND BRAIN STRUCTURAL MORPHOLOGY

Exposure During Pregnancy and Childhood — Cortical Thickness and Surface Area in School-Age Children and Pre-adolescents

Exposure during pregnancy — cortical thickness in school-age children

Higher exposure to PM during pregnancy was associated with thinner cortices in several brain regions in both hemispheres of school-age children (e.g., cerebral cortex of the precuneus region of the right hemisphere was 0.045 mm thinner [95% CI = -0.062 to -0.028] per 5-μg/m³ increase in PM_2.5) (Figure 4 and Table 6). These regions had a mean size between 532 and 2,995 mm² and a mean thickness between 2.31 and 3.17 mm² (with a minimum of 1.61 to 2.23 mm² and a maximum of 3.23 to 3.97 mm²) (data not shown).

Open in a separate window

Figure 4.

Brain regions identified on the adjusted associations between exposure to PM_2.5, PM_COARSE, and PM_2.5absorbance during pregnancy and cortical thickness (mm) in school-age children. Blue indicates negative coefficients; a darker blue color indicates a stronger association. Associations from linear regression models were adjusted for maternal and paternal education, ethnicity, age, height, body mass index, and psychological distress during pregnancy, maternal smoking and alcohol use during pregnancy, maternal parity, maternal intelligence quotient, family status, household income, and child’s genetic ancestry, sex, and age at the scanning session. All brain regions survived the correction for multiple testing (built-in Monte Carlo null-Z simulation with 10,000 iterations, P < 0.01). The expected direction of the association was higher air pollution exposure during pregnancy and smaller cortical thickness.

Table 6.

Adjusted Associations Between Exposure to Air Pollutants During Pregnancy and Cortical Thickness (mm) in School-Age Children^a

Brain Region – Hemisphere	Air Pollutant (per unit of change)	Region Size (mm2)	Coefficient (95%> CI)b (mm)
	Precuneus – Right	PM2.5 (5 (g/m3)	936	–0.045 (–0.062 to –0.028)
Pars opersularis – Right	PM2.5 (5 (g/m3)	753	–0.024 (–0.033 to –0.014)
Rars orbitalis – Right	PM2.5 (5 (g/m3)	651	–0.028 (–0.043 to –0.012)
Rostral middle frontal – Right	PM2.5 (5 (g/m3)	2995	–0.029 (–0.041 to –0.018)
Superior frontal – Right	PM2.5 (5 (g/m3)	722	–0.029 (–0.043 to –0.016)
Lateral orbitofrontal – Right	PMCOARSE (5 (g/m3)	565	–0.037 (–0.059 to –0.016)
Cuneus – Left	PM2.5 (5 (g/m3)	843	–0.022 (–0.035 to –0.009)
Fusiform – Left	PM2.5absorbance (10–5m–1)	532	–0.105 (–0.160 to –0.049)

Open in a separate window

^a The expected direction of the association was higher air pollution exposure during pregnancy and smaller cortical thickness.

^b Coefficients and 95% confidence intervals from linear regression models were adjusted for maternal and paternal education, ethnicity, age, height, body mass index, and psychological distress during pregnancy; maternal smoking and alcohol use during pregnancy; maternal parity, maternal intelligence quotient, family status, household income; and child’s genetic ancestry, sex, and age at the scanning session. All brain regions survived the correction for multiple teeing (built-in Monte Carlo null-Z simulation with 10,000 iteration, P value < 0.01).

Exposure during pregnancy and childhood — cortical thickness and surface area in pre-adolescents

Higher exposure to OC, PM_2.5absorbance, and Cu during pregnancy, and to OP^DTT during childhood was associated with thinner cortices in three brain regions in pre-adolescents (e.g., cerebral cortex of the postcentral gyrus of the right hemisphere was 0.06 mm thinner [95% CI = -0.11 to -0.01] per 1-μg/m³ increase in OC) (Figure 5 and Table 7). These regions had a mean size between 164 and 198 mm².

Open in a separate window

Figure 5.

Brain regions identified on the adjusted associations between exposure to OC, PM_2.5absorbance, and Cu during pregnancy and cortical thickness (mm), OP^DTT during childhood and cortical thickness (mm), and Zn, OP^ESR, and PM_COARSE and cortical surface area (mm²) in pre-adolescents. Blue indicates negative coefficients and green indicates positive coefficients. Associations from linear regression models were adjusted for maternal and paternal education, ethnicity, age, height, body mass index, and psychological distress during pregnancy, maternal smoking and alcohol use during pregnancy, maternal parity, maternal intelligence quotient, family status, household income, and child’s genetic ancestry, sex, and age at the scanning session. All brain regions survived the correction for multiple testing (built-in Monte Carlo null-Z simulation with 10,000 iterations, P < 0.001). The expected direction of the association was higher air pollution exposure during pregnancy and childhood and smaller cortical thickness and smaller cortical surface area.

Table 7.

Adjusted Associations Between Exposure to Air Pollutants During Pregnancy and Childhood and Cortical Thickness and Cortical Surface in Pre-adolescents^a

Brain Region – Hemisphere / Air Pollutant (per unit of change)	Pregnancy Exposureb		Childhood Exposureb
	Region Size (mm2)	Coefficient (95% CI)	Region Size (mm2)	Coefficient (95% CI)
Cortical thickness (mm)
Postcentral gyrus – Right
OC (1 Og/m3)	164	–0.06 (–0.11 to –0.01)	—	——
Rostral middle frontal gyrus — Right
PM2.5absorbance (10–5m–1)	174	–0.07 (–0.13 to –0.01)	—	——
Cu (5 ng/m3)	198	–0.12 (–0.21 to –0.03)	—	——
Lingual gyrus – Left
OPDTT (1 nmol DTT/min/m3)	—	——	177	–0.15 (–0.28 to –0.03)
Cortical surface (mm2)
Precentral gyrus – Right
Zn (10 ng/m3)	—	——	240	0.01 (0.00 to 0.02)
OPESR (1,000 units/m3)	—	——	157	0.02 (0.00 to 0.04)
Pars triangularis – Right
PMCOARSE (5 μg/m3)	—	——	180	–0.10 (–0.18 to –0.02)
Precuneus – Left
Zn (10 ng/m3)	—	——	177	0.02 (0.01 to 0.04)
Pericalcarine cortex–- Left
Zn (10 ng/m3)	—	——	289	0.02 (0.01 to 0.03)

Open in a separate window

^a The expected direction of the association was higher air pollution exposure during pregnancy and childhood and smaller cortical thickness and smaller cortical surface area. — = not significant.

^b Coefficients and 95% confidence intervals from linear regression models were adjusted for maternal and paternal education, ethnicity, age, height, body mass index, and psychological distress during pregnancy; maternal smoking and alcohol use during pregnancy; maternal parity, maternal intelligence quotient, family status, and household income; and child’s genetic ancestry, sex, and age at the scanning session. All brain regions survived the correction for multiple testing (built-in Monte Carlo null-Z simulation with 10,000 iteration, P value < 0.001).

Higher exposure to Zn and OP^ESR during childhood was associated with a larger cortical surface area in three brain regions (e.g., cortical surface area of the precentral gyrus of the right hemisphere was 0.01 mm² larger [95% CI = 0.00 to 0.02] per 10-ng/m³ increase in Zn). Higher exposure to PM_COARSE during childhood, instead, was associated with a smaller cortical surface area in one brain region (cortical surface area of the pars triangularis of the right hemisphere was 0.10 mm² smaller [95% CI -0.18 to -0.02] per 5-μg/m³ increase in PM_COARSE) (Figure 5 and Table 7). These regions had a mean surface area between 157 and 289 mm².

AIR POLLUTION EXPOSURE AND BRAIN STRUCTURAL CONNECTIVITY

Exposure During Pregnancy and Childhood — Specific FA and MD Tracts

Higher exposure to PM_2.5 during pregnancy was associated with four FA tracts in pre-adolescents (i.e., superior longitudinal fasciculus tract of the right hemisphere, forceps minor tract, and corticospinal tracts of the left and right hemispheres) (Figure 6 and Appendix Table A8). Higher exposure to NO_x during childhood was associated with five FA tracts (i.e., uncinated fasciculus tracts of the left and right hemispheres, superior longitudinal fasciculus tract of the right hemisphere, inferior longitudinal fasciculus tract of the right hemisphere, and corticospinal tract of the left hemisphere).

Open in a separate window

Figure 6.

Adjusted associations in the twelve individual white matter tracts in pre-adolescents. Blue indicates negative coefficients and green indicates positive coefficients. Associations from linear regression models were adjusted for maternal and paternal education, ethnicity, age, height, body mass index, and psychological distress during pregnancy, maternal smoking and alcohol use during pregnancy, maternal parity, maternal intelligence quotient, family status, household income, and child’s genetic ancestry, sex, and age at the scanning session. The expected direction of the association was higher air pollution exposure during pregnancy and childhood and lower global fractional anisotropy and higher global mean diffusivity. CGC-L = cingulum bundle left; CGC-R = cingulum bundle right; CST-L = corticospinal tract left; CST-R = corticospinal tract right; FMI = forceps minor; ILF-L = inferior longitudinal fasciculus left; ILF-R = inferior longitudinal fasciculus right SLF-L = superior longitudinal fasciculus left hemisphere; SLF-R = superior longitudinal fasciculus right hemisphere; UNC-L = uncinated fasciculus left hemisphere; UNC-R = uncinated fasciculus right hemisphere.

Higher exposure to Si during pregnancy was associated with five MD tracts (i.e., cingulate gyrus part of the cingulum tracts of the left and right hemispheres, superior longitudinal fasciculus tract of the left hemisphere, forceps minor tract, and inferior longitudinal fasciculus tract of the left hemisphere) (Figure 6 and Appendix Table A9). Higher exposure to Zn during childhood was associated with nine MD tracts (i.e., uncinated fasciculus tracts of the left and right hemispheres, cingulate gyrus part of the cingulum tracts of the left and right hemispheres, superior longitudinal fasciculus tracts of the left and right hemispheres, forceps minor tract, and inferior longitudinal fasciculus tracts of the left and right hemispheres). Higher exposure to OP^DTT during childhood was associated with one MD tract (i.e., cingulate gyrus part of the cingulum tract of the left hemisphere).

The Generation R Study is conducted by the Erasmus Medical Center in close collaboration with the School of Law and Faculty of Social Sciences of the Erasmus University Rotterdam, the Municipal Health Service Rotterdam area, Rotterdam, the Rotterdam Homecare Foundation, Rotterdam and the Stichting Trombosedienst & Artsenlaboratorium Rijnmond (STAR-MDC), Rotterdam. We gratefully acknowledge the contribution of children and parents, general practitioners, hospitals, midwives, and pharmacies in Rotterdam. The general design of Generation R Study is made possible by financial support from the Erasmus Medical Center, Rotterdam, the Erasmus University Rotterdam, the Netherlands Organization for Health Research and Development (ZonMw), the Netherlands Organization for Scientific Research (NWO), and the Ministry of Health, Welfare and Sport. Air pollution exposure assessment was possible by funding from the European Community’s Seventh Framework Program (GA#211250, GA#243406). In addition, the study was made possible by financial support from the Netherlands Organization for Health Research and Development (ZonMW Geestkracht Program 10.000.1003 & ZonMw TOP 40-00812-98-11021). The neuroimaging infrastructure was funded via TOP project number 91211021 to Tonya White and supercomputing computations for imaging processing were supported by the NWO Physical Sciences Division (Exacte Wetenschappen) and SURFsara (Lisa compute cluster, www.surfsara.nl). Henning Tiemeier received funding from the Netherlands Organization for Health Research and Development (NWO-grant 016.VICI.170.200). Hanan El Marroun received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 633595 (DynaHEALTH) and No. 733206 (LifeCycle). The Erasmus University Rotterdam granted Hanan El Marroun a personal fellowship (EUR Fellow 2014) and supported this work financially. Mònica Guxens is funded by a Miguel Servet fellowship (MS13/00054, CP13/00054, CPII18/00018) awarded by the Spanish Institute of Health Carlos III. We acknowledge support from the Spanish Ministry of Science and Innovation and State Research Agency through the “Centro de Excelencia Severo Ochoa 2019–2023” Program (CEX2018-000806-S) and support from the Generalitat de Catalunya through the CERCA Program.

APPROACH

The study by Dr. Guxens assessed the possible relationship of air pollution exposure during pregnancy and childhood with brain outcomes in children. Brain structural and functional measures were studied in Generation R — an existing birth cohort in Rotterdam, the Netherlands. Mother–child pairs were recruited during pregnancy or at birth from 2002–2006 and followed up until 2015.

Dr. Guxens and colleagues used air pollution data and high-resolution neuroimaging data collected in about 800 school-age children and in about 3,100 pre-adolescents. About 400 children underwent imaging at two time points. Early life exposure was estimated at the residential address level for various air pollutants using existing land-use regression (LUR) models, mainly from the European ESCAPE project. They applied single pollutant regression models to assess the association between early life air pollution exposure and brain structural and functional measures corrected for important confounders. Additionally, they used multipollutant models using a deletion/substitution/addition (DSA) approach.

METHODS

The aims of the study are detailed in Critique Table 1. These aims were to investigate the possible relationship of air pollution exposure during pregnancy and childhood with brain outcomes in children by assessing:

• (Aim 1) brain structural morphology (brain volumes, cortical thickness, and cortical surface area),

• (Aim 2) brain structural connectivity (white matter microstructure), and

• (Aim 3) brain functional connectivity (connectivity scores among brain areas).

Dr. Guxens and colleagues used high-resolution neuroimaging data collected in school-age children (age 6–10) and in pre-adolescents (age 9–12). MRI measurements of structural morphology were performed at both time points. Measurements of structural connectivity and of functional connectivity were performed at both time points as well, but for the Guxens study they used these data in pre-adolescents only. Thus, Aim 1 was pursued in both school-age children and pre-adolescents, whereas Aims 2 and 3 were investigated in pre-adolescents only. In both time points, they used a scanning platform specifically dedicated for the study, though a different scanner was used for each time point. The Investigators’ Report provides the detailed protocol that was followed for the measurements.

In addition to the MRI measurements, cognitive function of the school-age children was assessed by using an array of tasks from the Dutch version of the Developmental Neuropsychological Assessment test (NEPSY-II). For this analysis, tests were chosen in the attention and executive functioning domain and the memory and learning domain.

Exposure was estimated at the residential address level for 14 air pollutants using existing LUR models, mainly from the European ESCAPE project (e.g., Beelen et al. 2013; Eeftens et al. 2012). Those models were based on air pollution measurements between February 2009 and February 2010 at 40 to 80 sites spread across the Netherlands and Belgium. For some analyses, prenatal and childhood exposure was estimated using back and forward extrapolation methods based on a few continuous-reference monitoring sites to match the exact period of interest for six pollutants (nitrogen oxides [NO_x], NO₂, PM₁₀, PM_COARSE, PM_2.5, and PM_2.5absorbance [absorbance of the PM_2.5 fraction]). For the remaining pollutants (polycyclic aromatic hydrocarbons [PAHs], organic carbon [OC], copper [Cu], iron [Fe], silicon [Si], zinc [Zn], and two oxidative potential measures), assumptions were made that the contrast remained constant over time because there were no historical data available for those pollutants.

Box 2: Major Structures and Functions of the Brain^a

Open in a separate window

The brain can be divided into three basic structural units: the hindbrain, the midbrain, and the forebrain. The hindbrain includes the upper part of the spinal cord, most of the brainstem, and the cerebellum. This part of the brain controls the body’s vital functions such as respiration and heart rate. The cerebellum (little brain) coordinates movement, posture, and balance. The midbrain is the uppermost part of the brainstem and controls some reflex actions and is part of the circuit involved in the control of eye movements and other voluntary movements. The forebrain is by far the largest part of the human brain and contains the entire cerebrum and several structures directly nestled within it, such as the thalamus.

Cerebrum. The cerebrum is the largest part of the forebrain and accounts for about two thirds of the total brain’s mass. It contains the cerebral cortex and several subcortical structures, including the hippocampus, basal ganglia, and amygdala. Right beneath the cerebral cortex is the corpus callosum.

Corpus callosum. The most prominent white matter structure is the corpus callosum, which consists of about 200 million myelinated fibers, most of which connect homologous areas of the left and right cortex. The functions of the corpus callosum can generally be thought of as integrating the activities of the left and right cerebral hemispheres, including functions related to the unification of sensory fields, memory storage and retrieval, attention and arousal, enhancing language, and auditory functions.

Hippocampus. Located deep within the brain, the hippocampus has a major role in learning and memory.

Basal ganglia. The basal ganglia consist of the caudate nucleus, putamen, globus pallidus (palladium), subthalamic nucleus, nucleus accumbens, and substantia nigra. The basal ganglia have long been known to play a role in the control of movement and muscle tone but more recently have been shown to be involved in circuits that mediate higher cognitive functions, attention, and affective states.

Amygdala. The amygdala is a small, almond-shaped complex of nerve cells that receive input from both the olfactory system and the cerebral cortex. It performs a primary role in the processing of memory, decision-making, and emotional responses, including fear, anxiety, and aggression. The amygdala might be best known as the part of the brain that drives the fight or flight response. Although it is often associated with the body’s fear and stress responses, it also plays a pivotal role in memory.

Thalamus. The thalamus is a mostly a gray matter structure, located at the top of the brain stem, with many essential relaying roles that range from perception, movement, and the body’s vital functions. It acts as a two-way relay station, sorting, processing, and directing signals from the spinal cord and midbrain structures up to the cerebrum in the forebrain and, conversely, from the cerebrum down the spinal cord to the nervous system.

^a Largely based on Lenroot and Giedd 2006. From: https://openstax.org/books/biology/pages/1-introduction. OpenStax. Available under Creative Commons license CC BY 4.0.

Guxens and colleagues applied single pollutant regression models to assess the association between early life air pollution and brain structural and functional measures corrected for important potential confounders, such as maternal smoking, prepregnancy body mass index and socioeconomic status. Missing confounder data — for example, up to 41% of data on paternal psychological distress — were imputed using standard techniques. To increase statistical power in the first round, a potential sampling bias was introduced by deliberately selecting more school-age children with child behavior problems and with mothers who reported certain exposures during pregnancy (e.g., exposure to drugs, nicotine, alcohol, and psychiatric medication). All analyses used inverse probability weighting to increase the chance the results would be representative for the study population initially recruited (about 9,600). In a sensitivity analysis for the global white matter microstructure outcomes, they applied a correction for exposure measurement error using the bootstrap method from Szpiro and colleagues (2011).

Critique Table 1 summarizes the data and analyses conducted by aim. Single-pollutant analyses were used for each brain outcome separately, with one pollutant from one exposure window added to the model at a time. The number of outcome measurements was very large — for example, for the cortical thickness and connectivity scores measures — about 140,000 to 320,000 suboutcomes were analyzed depending on the brain measure. Hence, analyses of those brain outcomes were corrected for multiple comparisons using a Monte Carlo approach. In contrast, analyses of other brain outcomes, including brain volumes and white matter microstructure outcomes, were not corrected for multiple testing; for those outcomes, multipollutant models were run instead. Those multipollutant models consider the correlation matrix of air pollutants and simultaneously correct for multiple testing. The investigators chose multipollutant analyses using the DSA approach (Sinisi and van der Laan 2004). DSA uses an iterative selection algorithm to fit a generalized linear model predicated on the power of cross-validation to select the best predictive model for each brain outcome. They picked the pollutants in the multipollutant model using the criterion that any exposure variable selected in at least 10% of the runs would go into a final model; each DSA model was run 200 times.

SUBSTANTIAL TEMPORAL AND SPATIAL MISALIGNMENT OF THE EXPOSURE DATA

The Committee had concerns about the exposure assessment because of the substantial temporal and spatial misalignment of the data. That issue is particularly important when studying the developing brain, which is exceptionally complex with potentially critical time windows of development. The study relies on an historical exposure assessment that is temporally misaligned with the health data by about 4–8 years later than the collection of the brain outcome data. Several issues of concern with the exposure assessment were noted by the Committee. First, the limited number of sites for the PM and PM species (40) does not afford strong data support for the modeling; this might have led to over-smoothing in the pollution surfaces, particularly for some of the metals that likely have fine-scale variations, such as Cu, Fe, and Zn. Second, PM measurements were not measured simultaneously (with only 10 sites per round); this seems to have necessitated temporal adjustments that can introduce error because of differential accuracy in the temporal adjustment, depending on the spatial relationship between the research sites and the single continuous-reference monitor site. Third, the forward and back extrapolations used in some analyses to match the exact period of interest could introduce additional exposure error. Fourth, the Committee noted several general concerns regarding the use of standard linear regression for the exposure model development. For example, the algorithm may overfit the data when there are relatively few monitoring sites to train a model and a large number of potential predictor variables offered. Also, the algorithm could fail to capture potentially complex relationships within the data because it assumes the relationship between air pollution and a predictor is linear over the whole range of the predictor values and includes potential interactions among covariates to only a limited extent. Moreover, the algorithm could result in unstable and uninterpretable coefficient estimates when highly correlated predictors are included in one model. To try to alleviate those concerns, other modeling techniques have increasingly been applied to develop LUR models for air pollution exposure assessment, such as machine learning techniques. The few studies, however, that compared performance of different modeling techniques for exposure assessment report similar performance for the models, although those results may differ with the study setting, hampering a generic recommendation for one algorithm (e.g., Chen et al. 2019; Kerkhoffs et al. 2019; Weichenthal et al. 2016). Some, but not all, of these limitations in the exposure assessment were noted in the discussion section of the Investigators’ Report and acknowledged by the study team. Although the Committee understands that Dr. Guxens and her team made best use of the exposure models available, the substantial temporal and spatial misalignment of the exposure data might have influenced the analysis of brain outcomes in unpredictable ways. An effort was made by the study team to adjust the analysis for exposure measurements error by using the bootstrap method from Szpiro et al. 2011, which was laudable, but was limited to a few brain outcomes only, and it is unclear whether the method has sufficient data support to derive reliable corrections.

STATISTICAL ANALYSES

The Committee thought the statistical analyses pursued was comprehensive, straightforward, and clearly presented. The analysis was hypothesis driven rather than an exploratory discovery-based analysis, and the main analyses were based on simple regression. The Committee appreciated the hypothesis provided after each aim, and the expected direction of the relationship for the many brain outcomes indicated under each table in the Investigators’ Report. The Committee noted that the study was exploratory in many aspects, without much prior knowledge on which to build. For example, investigators had no expected direction for the relationship for some brain measure outcomes, such as for the connectivity scores for the different brain areas. In those cases, it was clearly stated by the investigators that there was no a priori hypothesis on the direction of the relationship.

The Committee noted two potential areas for future work to compliment the current analyses. They would recommend a broader exploration that uses some of the untargeted analyses. Those analyses could address some of the main challenges more fully, such as simultaneous testing of multiple hypotheses and consideration of (a) multiple correlated exposures, (b) exposure interactions and nonlinear exposure–response relations, and (c) temporal factors in exposures. The Committee thought that one could borrow from some of the approaches currently used in omics analyses in studies of the exposome (Agier et al. 2016; Chadeau-Hyam et al. 2013; Staffogia et al. 2017). Other valuable studies would be future longitudinal analyses with repeated MRI measurements to assess the effect of air pollution on the developmental trajectories of the brain outcomes included in the current cross-sectional analysis. The brain is exceptionally complex and has a dynamic structure that is constantly evolving throughout life and influenced by many factors (Lenroot and Giedd 2006; Tau and Peterson 2010). The investigators are well positioned to pursue such longitudinal analyses in the near future. The ongoing Generation R just finished a third round of MRI measurements on the same children who are now 13–17 years old. They used the same scanner as in the second time point, and it is expected that the investigators will now have sufficient statistical power for such an analysis. Unfortunately, there are fewer children (about 400) that underwent imaging at both the first and second time points, and the use of different scanners has prevented a longitudinal analysis using those data.

Although the Committee thought the analyses were generally well done, all brain outcomes should have been corrected for multiple comparisons because of the large number of analyses. Multiple comparison corrections are an imperative step in reducing Type I error in MRI brain research (Bennett et al. 2009; Lindquist and Mejia 2015). For example, for the brain volume outcomes, the investigators ran 364 regression models (14 pollutants × 13 brain volume outcomes × 2 exposure windows). The investigators justified their decision with their hypothesis-based approach to selecting only volumes in specific brain regions and argued that the use of the multipollutant modeling alleviates multiple testing issues.

MULTIPOLLUTANT MODELING

The Committee was not convinced that the multipollutant approach added much. The use of the DSA algorithm in modeling the effects of multiple exposures on brain outcomes is said to provide a compromise between sensitivity and false discovery proportion compared with other methods. How stable the identified specific exposure associations are, however, remains open to question. Each DSA model was run 200 times, and the final model was identified as the one in which any exposure variable was selected in at least 10% of the runs — suggesting strongly that these findings are not particularly stable.

In addition, the Committee thought that the DSA approach for multipollutant modeling was not reliable because of the vastly different performance and data support for the various exposure models. Optimizing the prediction could favor those pollutants that are better supported and predicted than others. The Committee preferred some further analysis of multipollutant models, for example, using a manually supervised selection and combinations of variables to the DSA approach implemented, given the limitations of the data and exposure models.

Multipollutant modeling in general remains difficult in the presence of high correlations among pollutants, irrespective of the method (Dominici et al. 2010). In particular PM₁₀, PM_2.5, NO_x, NO₂, and PM_2.5absorbance were highly correlated (>0.7), as were Cu, Fe, and oxidative potential using electron spin resonance (>0.8), and Fe and Si (0.8). The correlation between PAHs and Cu was similarly high (0.7). Typically, somewhat lower correlations were reported for the other pollutant pairs. Likewise, the often-high correlation between prenatal and childhood exposure hampered the identification of a susceptible exposure window to tease out effects of pregnancy versus childhood exposure and an approach for how to allow for effects during both time periods. The correlation was typically around 0.60 between prenatal and childhood exposure. Thus, the Committee would like to emphasize that the individual pollutant results should be considered as indicators of the ambient air pollution mixture, and prenatal and childhood exposure should be viewed as indicators of early life exposure. Thus, when summarizing the results in this Critique, the Committee has focused on single pollutant results and did not distinguish between pregnancy and childhood exposures.

GENERALIZABILITY OF THE FINDINGS

The Committee had some concerns about the generalizability of the findings. On the one hand, some strengths are that in Generation R the overall response rate at recruitment was 61%, which is relatively high and laudable (Kooijman et al. 2016). In addition, all analyses used inverse probability weighting to increase the chance that the results would be representative for the study population initially recruited (about 9,600). The Committee appreciated this correction approach because indeed MRI data were available from only a subset of the cohort, which was due in part to the high cost of performing an MRI. Also, a potential sampling bias was introduced by deliberately selecting more school-age children with child behavior problems and with mothers who reported certain exposures during pregnancy (e.g., exposure to drugs, nicotine, alcohol, and psychiatric medication). Nevertheless, the Committee thought the generalizability of the results was still reduced because the current study population is a convenience sample. Particularly noteworthy is that 26% of mothers smoked and about 50% drank alcohol during pregnancy at recruitment — much higher rates than the average consumption during pregnancy in the Netherlands (Lanting et al. 2015). Moreover, participants were recruited from Rotterdam, which hosts one of the largest seaports in the world and has associated refineries located upwind of the population studied. Although those issues might not affect the internal validity, they could affect the generalizability of the findings.

The Review Committee thanks the ad hoc reviewers for their help in evaluating the scientific merit of the Investigators’ Report. The Committee is also grateful to Hanna Boogaard and Annemoon van Erp for their oversight of the study, to Hanna Boogaard for her assistance in preparing its Critique, to Carol Moyer for editing of this Report and its Critique, and to Kristin Eckles and Hope Green for their roles in preparing this Research Report for publication.