Trimester-specific ambient PM_2.5 exposures and risk of intellectual disability in Utah

Abstract

Prenatal fine particulate matter (PM_2.5) exposure is an understudied risk factor for neurodevelopmental outcomes, including intellectual disability (ID). Associations among prenatal exposures and neurodevelopmental outcomes may vary depending on the timing of exposure. Limited numbers of studies examining PM_2.5 and neurodevelopmental outcomes have considered exposures occurring during the preconception period. To address these gaps, we conducted a case-control study of children born in Utah between 2002-2008 (n=1032). Cases were identified using methods developed by the Centers for Disease Control and Prevention’s Autism and Developmental Disabilities Monitoring Network and matched with controls on birth year, sex, and birth county. We estimated the daily average PM_2.5 concentration during a period spanning 12 weeks before the estimated conception date, as well as during each of the three trimesters at the maternal residential address listed on the child’s birth certificate. In a multivariable model, the third (OR: 2.119, CI: 1.123-3.998, p=0.021) and fourth (OR: 2.631, CI: 1.750-3.956, p <.001) quartiles for preconception average PM_2.5 demonstrated significantly increased risk of ID relative to the first quartile. Second quartile preconception exposure was also associated with increased risk, though it did not reach significance (OR: 1.385, CI: 0.979-1.959, p=0.07). The fourth quartile of first trimester average PM_2.5 was positive and significant (OR: 2.278, CI: 1.522-3.411, p <0.001); the third quartile was positive, but not significant (OR: 1.159, CI: 0.870-1.544, p=0.312). Quartiles of second and third trimester were not associated with higher risk of ID. These findings from Utah, which were robust to a variety of sensitivity analyses, provide initial evidence that preconception and prenatal PM_2.5 exposure may be associated with ID. Future studies are needed across other geographic locations and populations.

1. Introduction

The most recognized health-harming ambient air pollutant is particulate matter 2.5 micrometers or less in diameter (PM_2.5). In the US, where air quality has improved in recent decades, PM_2.5 is still responsible for 3% of deaths from all causes and 22% of deaths from environmental causes (Institute for Health Metrics and Evaluation 2020). PM_2.5 also causes substantial morbidities with major social and economic costs (Trasande, Malecha, and Attina 2016; Shea, Perera, and Mills 2020).

Despite mounting evidence linking many prenatal air pollutants (e.g., polycyclic aromatic hydrocarbons) to adverse neurodevelopment (Volk et al. 2021), the literature on prenatal PM_2.5 exposure remains underdeveloped in several specific regards, including a neglect of specific conditions (e.g., intellectual disability) and examination of specific developmental windows of risk. Overall, the literature on prenatal PM_2.5 and neurocognitive outcomes exhibits mixed findings (Volk et al. 2021). Some studies find no effect of PM_2.5 on neurocognitive outcomes, e.g., studies examining annual average concentrations during birth year and risk of ADHD and dyslexia (Fuertes et al. 2016); third trimester concentrations and executive function, behavior problems (Harris et al. 2016) and IQ (Harris et al. 2015); and pregnancy average concentrations and infant cognitive development at 15 months (Lertxundi et al. 2015). Fewer report significant effects: pregnancy average PM_2.5 exposure was associated with decreased motor scores at 15 months (Lertxundi et al. 2015), and prenatal exposure to PM_2.5 was associated with reduced mental and psychomotor development at age 2 years (Wang et al. 2022).

A limitation of these studies (Lertxundi et al. 2015; Harris et al. 2015; Fuertes et al. 2016; Harris et al. 2016; Wang et al. 2022) is that they examined only one window of prenatal exposure (e.g., entire pregnancy average or third trimester average) in relation to neurodevelopmental outcomes. This is a serious limitation because associations between prenatal exposures and neurodevelopmental outcomes may vary depending on timing of exposures (Chiu et al. 2016), although such knowledge is currently limited (Bansal et al. 2021; Girardi et al. 2021). The literature suggests that sensitive windows may vary based on the neurodevelopmental outcome being examined (Rahman et al. 2022; Su et al. 2022; Chun et al. 2019; Dutheil et al. 2021), making research on a variety of outcomes essential to identifying general trends and outcome-specific patterns.

The preconception period (e.g., 12 weeks before conception) may be another critical window of exposure to PM_2.5, but it has received little attention (Li et al. 2021). While studied less frequently than gestational exposures, preconception exposures to PM_2.5 have been linked to several neurodevelopmental outcomes, including neural tube defects (Zhang et al. 2020), delayed mental and psychomotor development (Li et al. 2021), and autism spectrum disorder (ASD) (Dutheil et al. 2021).

While there is growing evidence for the effects of PM_2.5 exposure on the risk of adverse neurodevelopmental outcomes, including ASD (Clifford et al. 2016; Volk et al. 2021; Dutheil et al. 2021), no studies have examined the effects of ambient PM_2.5 exposure on risks of intellectual disability (ID). ID involves problems with general mental abilities in two areas: intellectual functioning (e.g., learning, problem solving, and judgement) and adaptive functioning (e.g., communicating, independent living, and other activities of daily living). ID affects about 1% of the US population, and 85% of affected individuals have mild ID (American Psychiatric Association 2020; Patrick et al. 2021).

ID is a condition that is generally understudied in terms of environmental correlates, apart from research on heavy metals like lead (Pb). Lead is associated with neurological impacts in children and has been linked with reductions in IQ scores (Needleman 2004; Rothstein, Harrell, and Marchant 2017) and increased prevalence of ID (Carrington et al. 2019). Higher concentrations of heavy metals in the soil at pregnant mothers’ residences have been linked to ID in their children (McDermott et al. 2011; Onicescu et al. 2014). Airborne industrial lead emissions and urinary arsenic metabolites are associated with poorer executive functioning in children (Gatzke-Kopp et al. 2021; Desai et al. 2020). Reductions in children’s IQ are correlated with higher concentrations of manganese in drinking water and hair (Bouchard et al. 2011). Apart from heavy metals, research has associated prenatal exposures to pesticides (Lyall et al. 2017) and industrial air pollution (Grineski et al. 2022) with increased ID risk. Among a nationally representative UK sample, children with ID were 33% more likely to live in areas with high diesel particulate matter than those without ID (Emerson et al. 2018).

No studies have yet investigated if prenatal ambient PM_2.5 exposure is associated with risk of ID, despite mounting evidence of linkages between prenatal air pollution and adverse neurodevelopment (Volk et al. 2021). It is currently unknown if preconception or trimester-specific environmental exposures are associated with ID. We address these gaps using a retrospective case-control study design on a sample of Utah children to answer the following questions: 1) Is the second, third or fourth quartile of daily average exposure to ambient PM_2.5 across the entire pregnancy (i.e., from conception to birth) at the maternal residential address associated with a child’s odds of ID relative to the first quartile of exposure? 2) Are the second, third or fourth quartiles of daily average preconception and trimester-specific exposures to ambient PM_2.5 at the maternal residential address associated with a child’s odds of ID relative to the first quartile of exposure?

2. Material and Methods

2.1. Study Population

Children with ID were identified in a three-county catchment in Utah, which included Davis, Salt Lake, and Tooele counties, located along the populous Wasatch Front. Identification of ID was undertaken by the Utah Registry of Autism and Developmental Disabilities (URADD) using a population-based records screening methodology developed by the Autism and Developmental Disabilities Monitoring (ADDM) Network that relies primarily on IQ testing (Bilder, Pinborough-Zimmerman, et al. 2013). We used data collected in 2010 and 2012 from children with available health source records who were born in 2002, 2004, 2006, and 2008. In brief, URADD utilized a passive, population-based screening system based on community medical diagnoses.

Case ascertainment involved multiple source screening and records abstraction. Screening included pulling and analyzing records from major health sites in the three-county catchment, i.e., state health clinics, hospitals, private clinics, diagnostic centers, and individual providers specializing in services for children with disabilities. All cases were identified as having ID based on medical records; some children also had information from educational records, which was used in the screening process. ID cases were determined based on the most recent IQ score being ≤ 70 or expert clinician opinion. Here, 93% of children with ID (n=274) had IQ scores. To determine ID case status in the absence of IQ scores (n=22), the records were reviewed by either a child psychologist or a child psychiatrist specializing in neurodevelopmental disabilities. Children were classified as having ID in the absence of an IQ score when the records showed that psychological testing was attempted or considered, but the child was described by the psychologist or psychiatrist as being untestable because of severe cognitive deficits.

While adaptive functioning is a component of establishing a diagnosis of ID, ADDM methodology relies on IQ scores and clinician designation (when a child is untestable because of severe cognitive deficits) to classify ID (Bilder, Bakian, et al. 2013; Obi et al. 2011; Year gin-Allsopp et al. 1992; McDonnell et al. 2019; Patrick et al. 2021). Children in poverty and those with mild ID are more likely to lack adaptive functioning scores in their records as compared to those not in poverty and those with more severe impairment, respectively (Obi et al. 2011; Yeargin-Allsopp et al. 1992). Through its reliance on IQ scores and clinician designations, ADDM method avoids the bias that can accompany the use of adaptive functioning testing.

We then utilized the Utah Population Database (UPDB) to match each child with ID with children lacking evidence of ID. The UPDB collates birth, death, and medical records for the State of Utah. Records from URADD derived from health sources can be linked to UPDB with permission from the Utah Department of Health and Human Services (UDOHHS), URADD Oversight Committee, and the Institutional Review Boards of the UDOHHS, the University of Utah, and the Utah Resource for Genetic and Epidemiologic Research (RGE). Only children with ID who were linked to their Utah birth certificates within the UPDB were included in the analysis, thus limiting this study to those children born in Utah.

We matched each linked ID case with three non-ID affected controls based on county of birth, sex, and birth year. We matched on sex as it is a known neonatal risk factor for ID (Huang et al. 2016). We used birth year to account for potential variations over time. We used county, as opposed to a finer geographic unit, to allow for variability on the exposure estimates. Non-ID affected controls are “neurotypical,” meaning they had no medical records in UPDB indicating a mental health diagnosis (e.g., depression) or developmental disability (e.g., ASD) as well as no records in the URADD indicating a developmental disability. After matching, we excluded cases and controls without maternal residential birth address listed on their birth certificates, which we needed to assign PM_2.5 exposure estimates to each child. We also excluded children born before 27 weeks gestation, due to our interest in third trimester exposures. As a final step, we excluded non-ID affected controls whose associated case had been excluded due to prematurity or missing address. This left 296 cases and 736 controls (n of 1032). Of those 296 cases, 151 were ascertained at age 8 (born in 2002 and 2004) and 145 were ascertained at age 4 (born in 2006 and 2008). Children who were ascertained at age 4 vs. 8 did not significantly differ in terms of low birthweight, prematurity, mother’s education, mother’s age, maternal smoking, or racial/ethnic status. IQ scores tend to be quite stable over time, especially for children who initially score “low”, as compared to those with “normal” or “high” IQ scores (Schneider, Niklas, and Schmiedeler 2014). This indicates that using two ascertainment ages has minimal or no influence on inferences from this study. We use the presence or absence of ID (1=yes [cases], 0=no [controls]) as our primary dependent variable.

The project received IRB approval from the University of Utah (#00130218), as well as from the URADD Oversight Committee, the Utah Resource for Genetic and Epidemiologic Research (RGE) board (#00004589), and Utah Department of Health (#599).

2.2. PM_2.5 Exposure Estimates

We used the USEPA’s Downscaler product to create our PM_2.5 estimates (United States Environmental Protection Agency 2019). Downscaler PM_2.5 concentration estimates were generated using a Bayesian space-time downscaling fusion model that integrated data from a gridded atmospheric model (i.e., Community Multi-Scale Air Quality Model or CMAQ with point air pollution measurements from the National Air Monitoring Stations/State and Local Air Monitoring Stations (NAMS/SLAMS) (Berrocal, Gelfand, and Holland 2012). CMAQ estimates gridded averages with no missing values, but is subject to calibration error. That error is accounted for through fusion modeling with monitoring data from NAMS/SLAMS that provide direct, accurate measurements of PM_2.5. Downscaler daily PM_2.5 concentration estimates are available at 2010 census tract centroid locations (United States Environmental Protection Agency 2019) and have increasingly been employed in air pollution epidemiology studies (e.g., Bravo et al. 2017). Additional details on the Downscaler modeling approach, validation, and performance are available elsewhere (Berrocal, Gelfand, and Holland 2012). We downloaded daily average concentration estimates in μg/m³ for all census tracts in the State of Utah from 2002-2015 (United States Environmental Protection Agency 2019).

Based on the host census tract of each maternal residential address, we calculated five exposure estimates utilizing estimated birth date (i.e., we assumed each child was born on the 15^th of the month, since the dataset provided only month and year) and gestational age at birth (in weeks). First, we calculated daily average PM_2.5 between estimated conception date and birth date (i.e., entire pregnancy). Second, we calculated preconception average as daily average over the 12 weeks before estimated conception date. Third, to represent the first trimester, we calculated the daily average over weeks 0-12 based on estimated conception date. Fourth, we calculated the daily average over weeks 13-26 to represent the second trimester. Fifth, pertaining to the third trimester, we calculated the daily average over week 27 through birth. Finally, we transformed each variable into quartiles for entry into the statistical model with the first quartile serving as the reference category. For pregnancy average, the mean value for PM_2.5 in each of the quartiles was 8.262, 10.147, 11.334 and 14.472. For preconception, the mean values were: 6.563, 8.043, 10.134 and 15.950. For first trimester average exposure (TAE), the mean values per quartile were: 7.000, 8.540, 10.550, and 16.889 and for the second TAE they were 7.259, 8.745, 10.650 and 17.794. The mean quartile values were 7.120, 8.463, 10.967 and 18.722 for the third trimester TAE. Our use of quartiles follows other studies of prenatal air pollution exposure that employ them as the sole strategy (Tanner et al. 2014) or as one strategy (Guo et al. 2018; Vinikoor-Imler et al. 2013; Huynh et al. 2006). We use them as the primary strategy here due to concerns about non-linearity (Tamayo-Ortiz et al. 2021) and to avoid false precision, but we present other strategies (e.g., treating the variables as continuous and as dichotomous based on regulatory standards) as sensitivity analyses.

2.3. Other Covariates

We controlled for known ID risk factors, obtained from birth certificate data. This included the child’s race/ethnicity (Maenner et al. 2016; Patrick et al. 2021) (i.e., BIPOC [Black, Indigenous and People of Color] vs. non-Latina/x/o White [reference]); 86% of the BIPOC children in the study are Hispanic/Latino. It also included child sex (Huang et al. 2016; Patrick et al. 2021) (i.e., male vs. female), maternal age (i.e., ≤21 years (Leonard et al. 2005), ≥35 years (Huang et al. 2016; Bilder, Pinborough-Zimmerman, et al. 2013), and 22-34 years [reference]), any maternal prenatal smoking (Huang et al. 2016) (i.e., yes vs. no), and maternal education (Huang et al. 2016) (i.e., four categories of mother’s educational attainment: less than high school diploma, high school diploma, some college, or 2-year degree and bachelor’s degree or more [reference]). We also controlled for conception season (i.e., December-February, March-May, September-November, and June-August [reference]) following another study of prenatal PM_2.5 exposure (Guo et al. 2018).

Low birth weight (Huang et al. 2016) (i.e., <2500 g vs. ≥2500 g) and prematurity (Heuvelman et al. 2018) (i.e., <37 weeks) are established risk factors for ID and may be thought of as confounders in some studies. However, in the context of PM_2.5, they may be on the causal pathway between exposure to PM_2.5 and outcome as PM_2.5 is understood to be a risk factor for these birth outcomes (Coker et al. 2016; Yuan, Zhang, and Tian 2019). For that reason, we did not include these two variables in the models.

2.4. Statistical Methods

We first calculated univariate statistics for the analysis variables. Then, we conducted bivariate analyses (i.e., chi-square) to compare ID cases and non-ID affected controls in terms of PM_2.5 exposure and covariates. Next, we applied multiple imputation (MI) to estimate the missing values of each variable (Enders 2010). We used a regression-based approach to impute 20 datasets, each of which includes imputed values for each missing observation, with each imputed dataset selected at the 200th iteration (Enders 2010). We used all variables included in the main and sensitivity analyses in the MI procedure, including those not missing any values. Those variables are the five PM_2.5 variables, mother’s education code, father’s education code, birthweight, gestational age, any maternal smoking, multiple birth, C-section delivery, vaginal birth delivery with interventions, mother’s age at birth, father’s age at birth, sex, Hispanic/Latinx, American Indian, Asian, Pacific Islander, Black, Multiple race, White, conception month, birth year, birth month, and ID. Table 1 reports percent missing for each variable; three variables used in our analyses had missing values, i.e., maternal education, race/ethnicity and preconception PM_2.5. Unlike the other PM_2.5 variables, preconception PM_2.5 has missing data because the preconception period extends into 2001 for some children born in 2002 and EPA Downscaler data were not available for 2001.

Table 1.

Descriptive statistics for original data (n=1032)

Variables	N (% missing)	Min	Max.	Mean	Median	SD	Yes (%)
ID	1032 (0)						297 (28.7)
Mother’s age
≤21 years	1032 (0)						139 (13.5)
22-34 years	1032 (0)						773 (74.9)
≥35 years	1032 (0)						120 (11.6)
Mother Smoked During Pregnancy	1032 (0)						76 (7.4)
Mother’s Education	999 (3.2)
Less than high school	999 (3.2)						176 (17.6)
High school diploma	999 (3.2)						371 (37.1)
Some college/Associates	999 (3.2)						243 (24.3)
College degree +	999 (3.2)						209 (20.9)
BIPOC Child	790 (23.5)						298 (30.7)
Male	1032 (0)						749 (72.6)
Conception Season
Summer	1032 (0)						234 (22.7)
Fall	1032 (0)						259 (25.1)
Winter	1032 (0)						264 (25.6)
Spring	1032 (0)						275 (26.6)
PM2.5 Exposure Estimates
Entire Pregnancy Average Exposure	1032 (0)	5.103	16.184	11.054	10.702	2.352
Quartile 1	1032 (0)						258 (25.0)
Quartile 2	1032 (0)						258 (25.0)
Quartile 3	1032 (0)						258 (25.0)
Quartile 4	1032 (0)						258 (25.0)
Preconception Average Exposure	962 (6.8)	3.631	28.233	10.170	8.920	4.397
Quartile 1	962 (6.8)						240 (24.9)
Quartile 2	962 (6.8)						241 (25.1)
Quartile 3	962 (6.8)						241 (25.1)
Quartile 4	962 (6.8)						240 (24.9)
First TAE	1032 (0)	4.907	27.328	10.745	9.556	4.500
Quartile 1	1032 (0)						258 (25.0)
Quartile 2	1032 (0)						258 (25.0)
Quartile 3	1032 (0)						258 (25.0)
Quartile 4	1032 (0)						258 (25.0)
Second TAE	1032 (0)	4.432	27.827	11.112	9.598	4.715
Quartile 1	1032 (0)						258 (25.0)
Quartile 2	1032 (0)						258 (25.0)
Quartile 3	1032 (0)						258 (25.0)
Quartile 4	1032 (0)						258 (25.0)
Third TAE	1032 (0)	4.242	35.607	11.318	9.443	5.226
Quartile 1	1032 (0)						258 (25.0)
Quartile 2	1032 (0)						258 (25.0)
Quartile 3	1032 (0)						258 (25.0)
Quartile 4	1032 (0)						258 (25.0)

Note: TAE=trimester average exposure

We used generalized estimating equations (GEEs) to model associations between PM_2.5 and ID. GEEs build from the generalized linear model to adjust for clustering (Garson 2012) and have been used in other studies of prenatal pollution exposures to account for clustering and nesting, such as by clinic, pollution monitor catchment, birth year, or census tract (Grineski et al. 2022; Percy et al. 2019; Faiz et al. 2012; Fleischer et al. 2014). Here, we defined clusters based on birth year (n=5) and birth county (n=11), which resulted in 25 clusters with 1-325 children per cluster. Using year and county accounted for potential differences in ID surveillance across time and across space; a similar approach was used by Grineski et al. (2022). GEE models require an intracluster dependency correlation matrix and we used exchangeable, which assumes compound symmetry or constant intracluster dependency (Garson 2012). We used a binomial distribution and a logit link function because the dependent variable was dichotomous. We used SPSS v. 28 to conduct all analyses.

Model 1 examined associations between PM_2.5 quartiles and ID. We adjusted for conception season in this base model as PM_2.5 exhibits strong seasonal trends in Utah. The Spearman’s correlation coefficients between the preconception average and three TAEs were 0.211 for preconception and first trimester, 0.034 for first and second trimesters, 0.068 for second and third trimesters, −0.242 for preconception and second trimester, −0.391 for preconception and third trimester, and −0.354 for first and third trimesters. Inverse correlations between trimesters have also occurred in other locations with temporally varying patterns of PM_2.5 (Kalkbrenner et al. 2015).

Model 2 adjusted for all covariates. The two models were repeated twice, once when using entire pregnancy average PM_2.5 quartiles and once with preconception PM_2.5 along with first, second, and third TAEs instead. We included the four pollutant variables in the same model as per the joint TAE approach (Wilson et al. 2017). We entered the preconception and TAE variables into the model simultaneously to avoid the bias that can accompany the separate TAE approach (i.e., where each TAE is included in a separate model). A simulation study found that estimates from the separate TAE approach were biased in the trimesters when there was no true exposure effect (Wilson et al. 2017). A joint TAE approach was possible because none of the models suffered from multicollinearity as per the tolerance and variance inflation factor statistics. We did not conduct a paired matched analysis because matching cases with controls does not necessarily adjust for confounding by the matched variables (Pearce 2016). Instead, Pearce (2016) recommends accounting for matching factors through the analysis. As such, we included sex in the GEEs and adjusted for clustering by birth year and birth county.

2.5. Sensitivity Analyses

We conducted eight sensitivity analyses to explore how robust findings were for PM_2.5. We ran Model 1 and Model 2 for each sensitivity analysis and we re-calculated the PM_2.5 quartiles when the sample size changed. First, we included only cases that have ID without any chromosomal anomalies, including Down syndrome, Edwards syndrome, autosomal deletion syndromes, Klinefelter, and Trisomy 13, and their controls. Second and third, we used threshold variables for PM_2.5 instead of continuous values, i.e., average ≥ 10.0 μg/m³ (which was the World Health Organization [WHO] standard from 2006-2021) and average ≥ 12.0 μg/m³ (which is generally equivalent to the current US NAAQS) (following Percy et al. 2019). Fourth, we included children born ≥ 30 weeks instead of ≥ 27 weeks to allow for a longer minimum third trimester exposure period. Fifth, we tested an alternative clustering definition by substituting census tract for county, along with birth year. Sixth, we used an ordinal dependent variable pertaining to the severity of impairment (i.e., 0 = no impairment [controls], 1 = mild ID, i.e., IQ 50-70; and 2 = moderate to severe ID, i.e., IQ 20-49 or clinician estimate of ID without available IQ data) in an ordinal logistic model with a multinomial distribution and cumulative logit link.

Seventh, we expanded beyond the ADDM ID cases/controls to include additional ID cases/controls. This expanded sample included the ADDM ID cases and non-ID affected controls as well as additional cases of ID (and their non-ID affected controls) identified by URADD using non-ADDM methodology. These URADD ID cases (n=857) and non-ID affected controls (n=2264) were born from 2002-2014. The ID cases were comprehensively identified by ICD-9/−10 codes in medical records, however the records were not reviewed and abstracted. Sources of medical records included in this study included University of Utah Health, Intermountain Healthcare, and other regional behavioral health clinics. While this approach has not yet been validated, it is inherently conservative as it would likely under-ascertain cases because health insurance may not cover services that were billed using ID as a primary diagnosis. Eighth, we used continuous versions of the preconception and TAE variables as opposed to using quartiles.

3. Results

Table 1 provides univariate statistics for the analysis variables. While we only use continuous versions of PM_2.5 in the sensitivity analyses, we report those descriptive statistics in Table 1 for informational purposes, and they show that the averages for the entire pregnancy, preconception, and each trimester are under the NAAQS threshold (i.e., ≥12.0 μg/m³), but over 10.0 μg/m³, which was the WHO standard from 2006-2021. The median values for preconception and each trimester are just under 10.0 μg/m³, while the median for the entire pregnancy is over 10.0 μg/m³, but under 12.0 μg/m³.

Table 2 presents results from bivariate analyses. Table 2 shows that there were minimal differences in pregnancy average PM_2.5 between cases and controls, with controls having slightly higher values (p<.10). Cases had significantly higher daily PM_2.5 concentrations during the preconception period and first trimester (p<.001) while controls had significantly higher PM_2.5 concentrations during the second and third trimesters (p<.01) (Table 2). The distribution of cases vs. controls in each of the quartiles is also presented in Table 2.

Table 2.

Comparing cases with ID (n=296) to their neurotypical controls (n=736): Descriptive statistics and statistical testing

	Cases				Controls
Variables	N	Mean	Yes (%)	No (%)	N	Mean	Yes (%)	No (%)	p- value 1
Low birthweight	297		46 (15.4)	250 (93.6)	735		32 (4.4)	703 (95.6)	<0.001
Premature	296		52 (17.6)	244 (82.4)	736		58 (7.9)	678 (92.1)	<0.001
Mother’s age: 22-34 years	296		214 (72.3)	82 (27.7)	736		559 (76.0)	177 (24.0)	0.221
Mother’s age: ≤21 years	296		47 (15.9)	249 (84.1)	736		92 (12.5)	644 (87.5)	0.150
Mother’s age: ≥35 years	296		35 (11.8)	261 (88.2)	736		85 (11.5)	651 (88.5)	0.901
Any Smoking	296		24 (8.1)	272 (91.9)	736		52 (7.1)	684 (92.9)	0.562
Mother’s Education1
Less than high school	282		62 (22.0)	220 (78.0)	717		114 (15.9)	603 (84.1)	0.023
High school diploma	282		126 (44.7)	156 (55.3)	717		245 (34.2)	472 (65.8(	0.002
Some college/Associates	282		63 (22.3)	219 (77.7(	717		180 (25.1)	537 (74.9)	0.359
College degree +	282		31 (11.0)	251 (89.0)	717		178 (24.8)	539 (75.2)	<0.001
BIPOC Child	290		103 (33.6)	187 (63.2)	680		195 (28.7)	485 (71.2)	0.034
Male	296		215 (72.6)	81 (27.4)	736		534 (72.6)	202 (27.4)	0.979
Conception Season
Summer	296		55 (18.6)	241 (81.4)	736		222 (30.2)	514 (69.8)	<0.001
Fall	296		62 (20.9)	234 (79.1)	736		173 (23.5)	563 (76.5)	0.375
Winter	296		107 (36.1)	189 (63.9)	736		147 (20.0)	589 (80.0)	<0.001
Spring	296		72 (24.3)	224 (75.7)	736		194 (26.4)	542 (72.6)	0.499
PM2.5 Exposure Estimates
Entire Pregnancy	296	10.83			736	11.14			0.051
Quartile 1	296		85 (28.7)	211 (71.3)	736		173 (23.5)	563 (76.5)	0.080
Quartile 2	296		67 (22.6)	229 (77.4)	736		191 (26.0)	545 (74.0)	0.266
Quartile 3	296		80 (27.0)	216 (73.0)	736		178 (24.2)	558 (75.8)	0.340
Quartile 4	296		64 (21.6)	232 (78.4)	736		194 (26.4)	542 (73.6)	0.112
Preconception	251	11.16			711	9.82			<0.001
Quartile 1	251		40 (15.9)	211 (84.1)	711		200 (28.1)	511 (71.9)	<0.001
Quartile 2	251		53 (21.1)	198 (78.9)	711		188 (26.4)	523 (73.6)	0.094
Quartile 3	251		78 (31.1)	173 (68.9)	711		163 (22.9)	548 (77.1)	0.010
Quartile 4	251		80 (31.9)	171 (68.1)	711		160 (22.5)	551 (77.5)	0.003
First Trimester	296	11.67			736	10.37			<0.001
Quartile 1	296		70 (23.7)	226 (76.3)	736		188 (25.5)	548 (74.5)	0.525
Quartile 2	296		63 (21.3)	233 (78.7)	736		195 (26.5)	541 (73.5)	0.080
Quartile 3	296		56 (18.9)	240 (81.1)	736		202 (27.4)	534 (72.6)	0.004
Quartile 4	296		107 (36.2)	189 (63.8)	736		151 (20.5)	585 (79.5)	<0.001
Second Trimester	296	10.12			736	11.51			<0.001
Quartile 1	296		94 (31.7)	202 (68.2)	736		164 (22.3)	572 (77.7)	0.001
Quartile 2	296		85 (28.7)	211 (71.3)	736		173 (23.5)	563 (76.5)	0.080
Quartile 3	296		61 (20.6)	235 (79.4)	736		197 (26.8)	539 (73.2)	0.039
Quartile 4	296		56 (18.9)	240 (81.1)	736		202 (27.5)	534 (72.5)	0.004
Third Trimester	296	10.60			736	11.61			0.005
Quartile 1	296		82 (27.7)	214 (72.3)	736		176 (23.9)	560 (76.1)	0.204
Quartile 2	296		83 (28.0)	213 (72.0)	736		175 (23.8)	561 (76.2)	0.153
Quartile 3	296		71 (24.0)	225 (76.0)	736		187 (25.4)	549 (74.6)	0.633
Quartile 4	296		60 (20.3)	236 (79.7)	736		198 (26.9)	538 (73.1)	0.026

¹Statistical significance derives from an independent samples t-test for continuous variables and a chi square test for dichotomous variables.

In terms of other covariates, cases were significantly over-represented among those conceived in the winter and underrepresented among those conceived in the summer (both p<.001) (Table 2). Cases were also overrepresented among those with low birthweight (p<.001), prematurity (p<.001), and BIPOC identities (p<.05). Children with ID were underrepresented among mothers with college degrees or more and overrepresented among mothers with less than high school degree and those with a high school diploma (p<.05) (Table 2).

As shown in Table 3, the third quartile of pregnancy average PM_2.5 was associated with the odds of ID relative to the first quartile (OR: 1.674, CI: 1.056-2.656, p=0.03) in Model 1, but the finding did not hold in the fully adjusted model (Models 2). For preconception average exposure and the three TAEs, findings in terms of odds ratio, confidence interval and p-value were quite stable between Model 1 and 2. As per Model 2 (Table 3), for preconception average exposure, the third (OR: 2.119, CI: 1.123-3.998, p=0.021) and fourth (OR: 2.631, CI: 1.750-3.956, p <.001) quartiles demonstrated significantly increased risk of ID relative to the first quartile. Second quartile preconception exposure was also associated with increased risk (1.385, 0.979-1.959, p=.07). These results suggest a dose-response relationship with the odds ratio increasing with each subsequent quartile. For first TAE, the odds ratios increased between the third and fourth quartile, although only the fourth quartile of exposure is significant (OR: 2.278, CI: 1.522-3.411, p<.001). Second and third TAE were not associated with higher risk of ID (Table 3).

Table 3.

Pooled results of GEEs¹: Associations between PM_2.5 and odds of intellectual disability (n=1032)

	Model 12				Model 23
	Odds Ratio	Lower 95% CI	Upper 95% CI	p	Odds Ratio	Lower 95% CI	Upper 95% CI	p
Entire Pregnancy
Quartile 2	0.814	0.504	1.315	0.400	0.811	0.476	1.380	0.440
Quartile 3	1.674	1.056	2.656	0.029	1.606	0.970	2.659	0.066
Quartile 4	1.133	0.855	1.501	0.386	1.147	0.811	1.623	0.437
Preconception Average Exposure
Quartile 2	1.390	0.949	2.036	0.091	1.385	0.979	1.959	0.066
Quartile 3	2.081	1.105	3.919	0.023	2.119	1.123	3.998	0.021
Quartile 4	2.696	1.770	4.106	<0.001	2.631	1.750	3.956	<0.001
First TAE
Quartile 2	1.026	0.617	1.704	0.923	0.976	0.593	1.607	0.925
Quartile 3	1.127	0.859	1.479	0.389	1.159	0.870	1.544	0.312
Quartile 4	2.226	1.429	3.466	<0.001	2.278	1.522	3.411	<0.001
Second TAE
Quartile 2	1.123	0.825	1.530	0.461	1.149	0.836	1.580	0.392
Quartile 3	1.014	0.735	1.400	0.931	1.173	0.843	1.632	0.344
Quartile 4	0.981	0.679	1.417	0.918	1.028	0.714	1.480	0.883
Third TAE
Quartile 2	0.875	0.637	1.202	0.409	0.853	0.620	1.174	0.330
Quartile 3	1.012	0.735	1.394	0.940	1.047	0.731	1.500	0.801
Quartile 4	1.303	0.943	1.800	0.109	1.257	0.912	1.731	0.162

¹All models use a binomial distribution, logit link function, and an exchangeable correlation matrix and account for clustering based on child’s birth year and maternal residential county. Pooled results from 20 multiple imputed datasets are presented. TAE = trimester average exposure.

²Model 1 includes only conception season (i.e., December-February, March-May, June-August (reference), and September-November)

³Model 2 adjusts for child’s race/ethnicity (i.e., BIPOC [Black, Indigenous and People of Color] vs. non-Latina/x/o White [reference]), child sex (i.e., male vs. female), maternal age (i.e., ≤21, ≥35, and 22-34 [reference]), any maternal prenatal smoking (i.e., yes vs. no), and maternal education (i.e., four categories of mother’s educational attainment), and conception season.

3.1. Sensitivity Results

The PM_2.5 results were robust across a range of sensitivity analyses summarized in Table 4. Across the eight analyses, there were few divergent results, with the rest of the models showing the same positive and significant (p<.05) associations with ID for the third and fourth quartiles during the preconception window and the fourth quartile of the first TAE. The third quartile for pregnancy average was also positive and significant when looking only at children without chromosomal abnormalities (p<.01) (S1) and when including only children with >30 weeks gestation (p<.05) (S4). The fourth quartile for first TAE was not significant when clustering by tract (S5) nor in the expanded sample (S7), although the odds ratios were positive and p<.15. The second quartile for preconception average was also positive and significantly (p<.05) associated with increased severity (S6). The third quartile for second TAE was negative and significant (p<.05) when using the expanded sample (S7). Third TAE was positive and significant (p<.05) when using the threshold of ≥ 10.0 μg/m³ (S2) and when using the continuous variable (S8). Finally, when using the continuous versions of the variables (S8), preconception average exposure and first TAE were positive and significant (p<.01), matching Table 3 findings.

Table 4.

Summary of PM_2.5 results across Eight Sensitivity Analyses

Analysis	N	Entire Pregnancy Odds Ratio	p	Preconception Odds Ratio	p	First Trimester Odds Ratio	p	Second Trimester Odds Ratio	p	Third Trimester Odds Ratio	p
Main Analyses (from Table 3, Model 3)
Quartile 2	1032	0.811 (0.476-1.380)	0.440	1.385 (0.979-1.959)	0.066	0.976 (0.593-1.607)	0.925	1.149 (0.836-1.580)	0.392	0.853 (0.620-1.174)	0.330
Quartile 3	1032	1.606 (0.970-2.659)	0.066	2.119 (1.123-3.998)	0.021	1.159 (0.870-1.544)	0.312	1.173 (0.843-1.632)	0.344	1.047 (0.731-1.500)	0.801
Quartile 4	1032	1.147 (0.811-1.623)	0.437	2.631 (1.750-3.965)	<0.001	2.278 (1.522-3.411)	<0.001	1.028 (0.714-1.480)	0.883	1.257 (0.912-1.731)	0.162
S1: Excluded cases with chromosomal abnormalities and their controls
Quartile 2	938	0.840 (0.506-1.392)	0.498	1.318 (0.904-1.922)	0.151	0.984 (0.593-1.632)	0.949	1.175 (0.820-1.684)	0.379	0.855 (0.571-1.280)	0.446
Quartile 3	938	1.812 (1.205-2.724)	0.004	2.097 (1.161-3.787)	0.014	1.178 (0.825-1.683)	0.367	1.104 (0.779-1.566)	0.578	0.867 (0.587-1.281)	0.474
Quartile 4	938	1.039 (0.711-1.519)	0.842	2.555 (1.650-3.957)	<0.001	2.419 (1.495-3.915)	<0.001	0.857 (0.589-1.248)	0.422	1.015 (0.716-1.441)	0.931
S2: Used ≥ 10.0 μg/m3 threshold
	1032	1.241 (0.896-1.718)	0.193	1.534 (0.961-2.451)	0.073	1.336 (1.002-1.782)	0.048	0.851 (0.661-1.096)	0.211	1.286 (1.041-1.587)	0.019
S3: Used ≥ 12.0 μg/m3 threshold
	1032	0.909 (0.753-1.098)	0.321	1.819 (1.346-2.458)	<0.001	2.436 (1.755-3.382)	<0.001	1.109 (0.653-1.884)	0.701	1.251 (0.870-1.799)	0.226
S4: Included children with >30 weeks gestation
Quartile 2	1026	0.851 (0.509-1.424)	0.539	1.353 (0.972-1.884)	0.073	0.958 (0.589-1.557)	0.862	1.156 (0.829-1.612)	0.392	0.901 (0.662-1.227)	0.509
Quartile 3	1026	1.652 (1.009-2.707)	0.046	2.141 (1.166-3.932)	0.014	1.234 (0.930-1.638)	0.145	1,188 (0.826-1.709)	0.352	1.104 (0.789-1.546)	0.562
Quartile 4	1026	1.174 (0.824-1.672)	0.374	2.521 (1.716-3.703)	<0.001	2.389 (1.567-3.641)	0.000	1.028 (0.721-1.466)	0.880	1.255 (0.914-1.723)	0.160
S5: Used birth year and census tract to define clusters
Quartile 2	1032	0.692 (0.456-1.050)	0.084	1.390 (0.873-2.213)	0.165	0.808 (0.527-1.239)	0.329	1.042 (0.710-1.529)	0.834	0.881 (0.594-1.307)	0.530
Quartile 3	1032	1.275 (0.834-1.952)	0.262	2.122 (1.295-3.477)	0.003	0.775 (0.484-1.241)	0.289	0.770 (0.479-1.236)	0.279	0.840 (0.535-1.318)	0.447
Quartile 4	1032	0.925 (0.590-1.452)	0.736	3.201 (1.872-5.473)	<0.001	1.605 (0.928-2.777)	0.091	0.680 (0.381-1.212)	0.191	1.084 (0.624-1.883)	0.775
S6: ID severity in three categories as dependent variable
Quartile 2	1032	0.789 (0.460-1.351)	0.387	1.472 (1.010-2.144)	0.044	0.884 (0.523-1.495)	0.646	1.104 (0.835-1.460)	0.488	0.856 (0.599-1.224)	0.395
Quartile 3	1032	1.528 (0.939-2.489)	0.088	2.076 (1.038-4.153)	0.039	0.941 (0.668-1.325)	0.727	0.991 (0.730-1.344)	0.952	0.943 (0.651-1.365)	0.754
Quartile 4	1032	1.200 (0.887-1.623)	0.237	2.774 (1.773-4.341)	<0.001	1.930 (1.294-2.879)	0.001	0.915 (0.620-1.351)	0.655	1.188 (0.827-1.708)	0.351
S7: Used expanded sample
Quartile 2	3121	0.952 (0.760-1.194)	0.672	0.898 (0.693-1.164)	0.417	0.892 (0.711-1.119)	0.324	0.951 (0.748-1.209)	0.680	1.105 (0.883-1.384)	0.382
Quartile 3	3121	0.878 (0.700-1.102)	0.263	1.328 (1.040-1.696)	0.023	0.930 (0.706-1.225)	0.606	0.775 (0.606-0.991)	0.042	0.933 (0.723-1.203)	0.592
Quartile 4	3121	1.123 (0.905-1.394)	0.291	1.597 (1.091-2.338)	0.016	1.260 (0.934-1.701)	0.130	0.802 (0.597-1.076)	0.141	0.907 (0.640-1.286)	0.584
S8: Used continuous versions of PM2.5 variables
	1032	1.036 (0.982-1.093)	0.192	1.087 (1.049-1.127)	<0.001	1.097 (1.061-1.135)	<0.001	1.041 (0.992-1.037)	0.220	1.056 (1.021-1.092)	0.002

Note: All models adjust for child’s race/ethnicity (i.e., BIPOC [Black, Indigenous and People of Color] vs. non-Latina/x/o White [reference]), child sex (i.e., male vs. female), maternal age (i.e., ≤21, ≥35, and 22-34 [reference]), any maternal prenatal smoking (i.e., yes vs. no), and maternal education (i.e., four categories, with BA and higher as reference), and conception season (i.e., December-February, March-May, June-August (reference), and September-November). S1-S5 and S7-8 models use a binomial distribution, logit link function, and an exchangeable correlation matrix. S6 are ordinal logistic models with a multinomial distribution and cumulative logit link and an exchangeable correlation matrix. S1-S4 and S6-S8 account for clustering based on child’s birth year, maternal residential county, and birth season. In S1 and S4-S7, the first quartile of PM_2.5 is the reference group. Pooled results from 20 multiple imputed datasets are presented.

4. Discussion

Results suggest that ID risk is most associated with higher quartiles of preconception and first trimester PM_2.5 exposures at the maternal home address in this Utah sample. While these are less commonly identified windows of risk than second and third trimesters (as applicable to lower cognitive scores as well as ASD (Chun et al. 2019; Dutheil et al. 2021; Bansal et al. 2021; Zhang et al. 2022), among other outcomes), both preconception and first trimester windows have been linked to neurodevelopmental outcomes in other studies. Preconception exposures have been associated with neural tube defects (Zhang et al. 2020), delayed mental and psychomotor development (Li et al. 2021), and autism spectrum disorder (ASD) (Dutheil et al. 2021). First trimester windows have been associated with delayed mental and psychomotor development (Li et al. 2021), developmental delays (Su et al. 2022), and ASD (Rahman et al. 2022). In our sample, the preconception findings were especially robust, persisting across all sensitivity analyses. Sensitivity results also showed that the risks associated with having daily average PM_2.5 concentrations over the current US NAAQS threshold (i.e., ≥12.0 μg/m³) during the preconception and first trimester windows were substantial; the odds ratios were 1.8 and 2.4, respectively.

The mechanisms hypothetically linking prenatal PM_2.5 exposure and neurodevelopmental problems, including ID, are not completely understood but it is likely that prenatal PM_2.5 may alter the developing fetal brain morphology by increasing oxidative stress and inflammation in the central nervous system (CNS) (Calderón-Garcidueñas et al. 2016; Calderón-Garcidueñas et al. 2008; Chiu et al. 2016). Prenatally, the CNS develops in a timed cascade with different anatomic regions forming sequentially. The developing neurons stretch across these different anatomic regions and connect to other tissues forming a network that is differentially vulnerable to toxins depending on the timing of the exposure and the anatomic region affected (Chiu et al. 2016; Rodier 2004). Even minute and subtle disruptions to this developmental process can cascade to affect later health and development (Weiss 2000; Volk et al. 2021).

In terms of why third and fourth quartiles of preconception PM_2.5 was associated with ID specifically, it may be because PM_2.5 exposure affects the oocytes and sperm, which mature over several months preceding conception. The ovaries contain oocytes at multiple stages of development. A group of oocytes mature from preantral follicles (0.05 mm in size) to antral follicles (2.00 mm in size) over a two-month period. Then, a small proportion of that group develops into larger antral follicles. Finally, after one antral follicle is selected, it takes another 14 days for it to gain preovulatory follicle status and be released (Broekmans et al. 2010). Similarly, it takes around 74-90 days for sperm to mature, and they do so in twelve-stage multiday cycle with sperm maturing continuously (Johnson 1995; Li et al. 2021; Griswold 2016). The relatively long maturation process for oocytes and sperm make the three months before conception a salient window of exposure. Environmental exposures occurring while eggs and sperm are forming can cause epigenetic changes that may affect the health of the future child (Marcho, Oluwayiose, and Pilsner 2020). In addition, exposure to particulate matter affects the health of the uterus, altering endometrial receptivity to implantation (Castro et al. 2013) and causing inflammation of the uterine lining (Kim et al. 2021).

In terms of why the fourth quartile of first-trimester PM_2.5 and ID may be linked, exposure to toxins during gestational periods when cells are rapidly differentiating can change organ and tissue formation, which may predispose people to impaired functioning and elevated risks of disease (Kwon and Kim 2017). In addition, particulates affect the health of the developing placenta, which starts to form early in the first trimester (Michikawa et al. 2016); placental pathologies/defective placentation may be linked to later neurodevelopmental problems (Altshuler 1993), including ID (Villamor, Susser, and Cnattingius 2022).

We found limited evidence of risks for second and third TAE in this study. While there was very little signal for the second TAE, third TAE was significant and positive in two sensitivity models (i.e., when using the threshold of ≥ 10.0 μg/m³ (S2) and when using the continuous variable (S8)). While there is limited evidence to date, our findings for second and third TAE and ID contradicts some other studies looking at different neurodevelopmental outcomes, e.g., third trimester PM_2.5 and lower cognition scores (Zhang et al. 2022), second and third trimester PM_2.5 and vigilance and inhibitory control ability tasks (Bansal et al. 2021), and second (Dutheil et al. 2021) and third (Chun et al. 2019) trimester PM_2.5 and ASD. As nearly 40% of persons with ID also have ASD (Patrick et al. 2021), it is interesting that the pattern we uncovered here for windows of risk for ID (e.g., preconception and first trimester) differs from findings of two meta-analyses on ASD, which found risks during the second and third trimester (Dutheil et al. 2021; Chun et al. 2019). This research is still emerging though; a recent study not included in those reviews found first and second trimester PM_2.5 to be risk factors for ASD (Rahman et al. 2022). It is the case that “increased exposure to prenatal particulate air pollution may have… time-dependent neurotoxic effects that may vary for different cognitive or behavioral domains, which likely reflect different underlying pathways” (Chiu et al. 2016, p. 64). For ID, this study suggests that early exposures were more harmful as the preconception and first trimester windows were significantly associated with risk of ID.

4.1. Limitations

The analyses are limited to some extent by exposure misclassification. While the calculation of our PM_2.5 exposure estimates assumes that the mother resided at the maternal residential address for her entire pregnancy and her preconception period, this is not always the case. Some women move during pregnancy, which could introduce error into the PM_2.5 exposure estimates. In a Massachusetts study conducted between 1988-2008, 12% of women moved while pregnant (Bell, Banerjee, and Pereira 2018). We also were unable to account for variations in exposures from indoor sources of particulate matter and from mothers’ spending time at locations other than their census tract of residence. As this study focuses on prenatal exposures at the maternal residential address, we are unable to account for exposures that occur after birth.

We did not use a distributed lag model (DLM). While acceptable, the joint TAE approach is less accurate than a DLM, if the critical windows of exposure do not align within the trimester bounds (Wilson et al. 2017). While Wilson et al. (2017) recommend using the DLM when researchers have access to daily or weekly pollution data (which we do have), we do not have access to the last menstrual period date. We have estimated the date of conception and trimester windows based on gestational age at birth and estimated birth date. We have assumed that each child was born on the 15^th of their birth month, meaning that their actual birthdate is +/− two weeks of that date. When looking at trimesters (12-week periods), our data could be off by +/− 2 weeks, with the majority of the weeks in each trimester being correct. If we were looking at week-by-week pollution without access to data on the last menstrual period date, the pollution values assigned to any given gestational week may not even encompass that actual week in the pregnancy. We therefore lack data with the temporal precision to accurately implement a DLM. Future research on ID should use DLM with weekly pollution levels linked to last menstrual period date. With its use of joint TAEs, this study provides an acceptable first step.

We did not include all possible risk factors, e.g., parity and birth injury. This is likely a negligible limitation as we have included primary risk factors and results are stable across the three models and the sensitivity analyses. Lastly, our Utah focus may affect the generalizability of the findings as the sample is 69% non-Hispanic White. Future studies using similar data should be conducted in other places.

5. Conclusion

These findings from Utah provide initial evidence that prenatal PM_2.5 exposure may be a risk factor for ID. Our findings related to PM_2.5 add to the growing body of literature documenting associations between prenatal exposure to various environmental toxins (e.g., heavy metals in soil, industrial emissions, pesticides) and ID (McDermott et al. 2011; Onicescu et al. 2014; Carrington et al. 2019; Grineski et al. 2022; Lyall et al. 2017). Future studies need to test these associations across other geographic locations and populations to assess generalizability. More generally, findings underscore the importance of looking at trimester-specific pollution estimates and the preconception period in studies of neurodevelopmental outcomes.

HIGHIGHTS.

• Prenatal PM_2.5 is an understudied risk factor for neurodevelopmental outcomes.

• Prenatal PM_2.5 exposure and intellectual disability (ID) have yet to be examined.

• Preconception PM_2.5 was associated with greater odds of ID.

• First trimester PM_2.5 was associated with greater odds of ID.

• PM_2.5 during the second and third trimesters was not associated with ID.