The association between prenatal greenspace exposure and Autism spectrum disorder, and the potentially mediating role of air pollution reduction: A population-based birth cohort study

etiology remains inconclusive, but research suggests genetic, epigenetic, and environmental contributing factors and likely pre-natal origins. Few studies have examined modifiable environmental risk factors for ASD, and far fewer have examined protective exposures. Greenspace has been associated with positive child development, but very limited greenspace research has examined ASD risk or prenatal exposures. Only one ecological study in 2017 has evaluated the association between greenspace and ASD, observing protective benefits. Greenspace may have direct effects on ASD risk and indirect effects by reducing air pollution exposure, a growing suspected ASD risk factor. Objectives: To measure the association between prenatal greenspace exposure and ASD risk and examine if reduced air pollution levels in areas of higher greenspace mediate this association. Methods: We linked a population-based birth cohort of all deliveries in Metro Vancouver, Canada, from 2004 to 2009, with follow-up to 2014. Diagnoses were based on Autism Diagnostic Observation Schedule and Autism Diagnostic Interview-Revised instruments. Greenspace was quantified as the average of the annual mean Normalized Difference Vegetation Index (NDVI) within a 250 m buffer of a residential postal code. Air pollutant exposures — particulate matter with a diameter less than 2.5 µ m (PM 2.5 ), nitric oxide (NO), and nitrogen dioxide (NO 2 ) — were derived from previously developed and temporally adjusted land use regression models. We estimated air pollutant exposures as the mean concentration per month during pregnancy. We calculated odds ratios (ORs) using logistic regression per NDVI interquartile range (IQR) increase, adjusting for child sex, birth month and year, maternal age and birthplace, and neighborhood-level urbanicity and income. To estimate the health impact of greenspace on ASD at the population level, we used the logistic regression model and marginal standardization to derive risk differences (RDs). Lastly, to quantify the mediating effect of greenspace on ASD

Background: Autism spectrum disorder (ASD) incidence has increased in past decades.ASD etiology remains inconclusive, but research suggests genetic, epigenetic, and environmental contributing factors and likely prenatal origins.Few studies have examined modifiable environmental risk factors for ASD, and far fewer have examined protective exposures.Greenspace has been associated with positive child development, but very limited greenspace research has examined ASD risk or prenatal exposures.Only one ecological study in 2017 has evaluated the association between greenspace and ASD, observing protective benefits.Greenspace may have direct effects on ASD risk and indirect effects by reducing air pollution exposure, a growing suspected ASD risk factor.Objectives: To measure the association between prenatal greenspace exposure and ASD risk and examine if reduced air pollution levels in areas of higher greenspace mediate this association.Methods: We linked a population-based birth cohort of all deliveries in Metro Vancouver, Canada, from 2004 to 2009, with follow-up to 2014.Diagnoses were based on Autism Diagnostic Observation Schedule and Autism Diagnostic Interview-Revised instruments.Greenspace was quantified as the average of the annual mean Normalized Difference Vegetation Index (NDVI) within a 250 m buffer of a residential postal code.Air pollutant exposures-particulate matter with a diameter less than 2.5 µm (PM 2.5 ), nitric oxide (NO), and nitrogen dioxide (NO 2 )-were derived from previously developed and temporally adjusted land use regression models.We estimated air pollutant exposures as the mean concentration per month during pregnancy.We calculated odds ratios (ORs) using logistic regression per NDVI interquartile range (IQR) increase, adjusting for child sex, birth month and year, maternal age and birthplace, and neighborhood-level urbanicity and income.To estimate the health impact of greenspace on ASD at the population level, we used the logistic regression model and marginal standardization to derive risk differences (RDs).Lastly, to quantify the mediating effect of greenspace on ASD
The search for associations between ASD and environmental exposures has focused on non-specific risk factors or toxic chemicals.Less attention has been paid to protective factors that may diminish risk during critical prenatal developmental windows.Within the framework of "Positive epidemiology," (VanderWeele et al., 2020) evidence of health benefits from exposure to natural environments is increasing.(van den Bosch and Ode, 2017) To date, studies have reported associations between maternal exposure to urban green spaces and improved birth outcomes.(Dadvand et al., 2012;Laurent et al., 2013;Dzhambov et al., 2014) Exposure to green spaces in children's living environments has also been associated with improved mental health, wellbeing, and academic achievement.(McCormick, 2017;Browning and Rigolon, 2019;Davis et al., 2021) In addition, several recent studies demonstrate a link between childhood greenspace exposure and reduced risk of schizophrenia.(Engemann et al., 2020;Engemann et al., 2018;Engemann et al., 2019;Engemann et al., 2020) To our knowledge, only one previous ecological study has analyzed the association between childhood exposure to greenspace and ASD, (Wu and Jackson, 2017) observing a lower prevalence of ASD in greener neighborhoods.But its ecological design was subject to exposure misclassification and residual confounding.
Various pathways for the beneficial effects of green spaces on children's health and development have been suggested, including providing physical activity and play opportunities, (Söderström et al., 2004) stress relief, (Markevych et al., 2014) and social interactions.(Seeland et al., 2009) As for prenatal exposure, living in green environments may reduce the risk of stress and depression among pregnant mothers, (Banay et al., 2017) which could contribute to preventing ASD in their offspring.(Hagberg et al., 2018;Oberlander and Zwaigenbaum, 2017) As well, areas with higher levels of greenspace tend to have reduced concentrations of air pollutants, so the beneficial effects of greenspace may be mediated by reduced levels of air pollution.(Markevych et al., 2017;Doiron et al., 2020;Diener and Mudu, 2021) Recently, a study suggested a significantly higher risk of attention deficit hyperactivity disorder in areas with little greenspace and high levels of air pollution compared to the opposite condition, (Yuchi et al., 2022) and another concluded that greenspace could considerably reduce adverse childhood development health effects from air pollution, (Jarvis et al., 2021) although the impact is highly context-dependent.(Zheng et al., 2021;Tomson et al., 2021;Hashad et al., 2021).
This study evaluates the potential association between prenatal greenspace exposure and ASD in a large population-based birth cohort of Canadian children.Beyond a possible direct association between prenatal greenspace exposure and ASD, we also hypothesized that reduced levels of particulate matter with a diameter less than 2.5 µm (PM 2.5 ), nitric oxide (NO), and nitrogen dioxide (NO 2 ) could partly mediate the protective effect of greenspace.

Study population
We developed a linked file of all single, twin, and triplet births from 2004 to 2009 to mothers 15 to 49 years of age who resided in British Columbia (BC) during their pregnancy and were registered with the provincial health insurance plan for 275 days or longer.(Pagalan et al., 2019;Bickford et al., 2020).
The BC provincial public health insurance program covers nearly all residents, (Chamberlayne et al., 1998) and the British Columbia Perinatal Data Registry contains data abstracted from obstetrical and neonatal medical records on almost 100 % of births in the province, regardless of delivery location.Demographic information, residential six-digit postal codes at the time of birth registration, medical visits paid for on a fee-for-service basis, and data on births, deaths, and hospital discharges were collected through the British Columbia Ministry of Health and the British Columbia Vital Statistics Agency.Neighborhoodlevel data on urbanicity and socioeconomic status were gathered from Statistics Canada.
We excluded births with unknown gestational age or missing sex information, stillbirths, and infants who died before one year of age.Lastly, using residential postal codes, we excluded births outside of the Metro Vancouver study area, for which we have no air pollution exposure estimates.Metro Vancouver comprises more than 2.5 million residents and has a population density of approximately 855 people per km 2 .(Statistics Canada, 2016).

Case ascertainment
We identified ASD-positive children prospectively through BCAA-N and the Ministry of Education and Child Care.In BC, BCAAN provides diagnostic assessments for all children and youth up to age 19 referred for an ASD assessment.(Provincial Health Services Authority, 2021) BCAAN diagnostic assessments are conducted by specially trained clinicians and offered at no cost through the provincial health plan.Additional ASD-positive diagnoses made by private specialists, which must be paid for privately and are not captured by BCAAN, were identified through Ministry of Education and Child Care data.The cohort was followed until the end of 2014, enabling a minimum of five years of follow-up for ASD diagnoses.
All ASD assessments in BC made through BCAAN or by qualified private specialists follow a standardized approach based on the Ministry of Health standards and guidelines for assessing and diagnosing ASD.(Dua, 2003) Since 2004, this requires the use of the standardized Autism Diagnostic Observation Schedule (ADOS or ADOS-II) and Autism Diagnostic Interview-Revised (ADI-R) instruments.Within BCAAN, the final diagnosis is based on direct clinical evaluations made by pediatricians, psychiatrists, or psychologists and on detailed clinical history, evaluation of developmental status, and reports from community health practitioners.(Provincial Health Services Authority, 2021; Government of British Columbia, 2021) Qualified private specialists outside of BCAAN can diagnose children with ASD, but these private assessments must meet the same stringent Ministry of Health requirements as BCAAN.

Greenspace and air pollution exposures
We linked greenspace and air pollution exposure estimates using sixdigit residential postal codes at birth registration, which occurs within 30 days from birth.Six-digit residential postal codes represent small geographic areas in dense urban areas, such as a high-rise building or one side of a city block.(Charlotte et al.) Prenatal greenspace exposures were temporally resolved based on conception date, and air pollution exposures based on pregnancy duration.
We quantified greenspace using the normalized difference vegetation index (NDVI), which measures greenness using satellite spectral data of the Earth's surface and the visible and near-infrared light reflected by vegetation.NDVI values range from − 1 to 1, with negative values corresponding to water bodies, close to zero to barren areas, and positive values to green vegetation.Annual NDVI data from Landsat satellites, with an original 30 m spatial resolution, were linked to postal code coordinates by the Canadian Urban Environmental Health Research Consortium (CANUE), providing the average of the annual mean NDVI within 100, 250, and 500 m buffers for all Canadian postal codes.(Canadian Urban Environmental Health Research Consortium, 2019;Gorelick et al., 2017;Google Earth Engine Explorer, 1984;Google Earth Engine Explorer, 2013;Google Earth Engine Explorer, 1984;Google Earth Engine Explorer, 2013;CanMap Postal Code Suite, 2015) NDVI data were linked to the cohort using maternal residential postal codes and conception dates, which we derived using birth date and gestational age.
Gestational age was calculated by the British Columbia Perinatal Data Registry based on early gestational ultrasonograms (less than20 weeks) or from the date of the last menstrual period if an early gestational ultrasound was not performed.If neither field was recorded, gestational age was estimated from newborn clinical exam and chart documentation.(Perinatal Services BC, 2015).
Air pollution exposures were estimated using previously developed spatiotemporal models of PM 2.5 , NO, and NO 2 for the Metro Vancouver area.(Pagalan et al., 2019;Henderson et al., 2007;Wang et al., 2013;Nethery et al., 2008) These models provide monthly estimates of air pollutant concentrations at the six-digit postal code, capturing temporal and small-scale, localized spatial variation.The spatiotemporal models were developed using air pollution surfaces, with 10 m spatial resolution, specifically made to reliably estimate air pollutant concentrations for epidemiological and risk assessment applications.These air pollution surfaces were temporally interpolated using continuous monitoring data from Metro Vancouver's Air Quality Monitoring Network to estimate monthly concentrations over time.Air pollutant exposure estimates were calculated as the mean concentration per month during pregnancy at the maternal residential postal code.

Statistical analyses 2.4.1. Primary analysis
To evaluate the association between prenatal exposure to greenspace and ASD, we calculated odds ratios (OR) and 95 % confidence intervals (CI) using logistic regression per interquartile range (IQR) increase in the average of the annual mean NDVI within several buffer zones.We initially selected a 250 m buffer zone for the analysis, corresponding to a five-to-ten-minute walk.A 250 m buffer has been the most commonly used buffer size in previous greenspace research, (Labib et al., 2020) and it is the distance typically applied in urban policies and recommendations.(WHO Regional Office for Europe, 2016) Nevertheless, there is no definite consensus as to what distance best represents greenspace exposure, which is dependent on various factors, such as type of health outcome studied, assumed pathway, spatial resolution, and zoning and scaling effects.(Labib et al., 2020) We, therefore, also tested NDVI buffers of 100 and 500 m.We adjusted for child sex (categorical), birth month (categorical), birth year (categorical), maternal age (continuous), maternal birthplace (categorical), and neighborhood-level urbanicity (categorical) and income band (continuous) to control for confounding.Figure S1 in the supplementary material shows the directed acyclic graph used for the analyses.Income band measures socioeconomic status and consists of 1,000 bands of equivalized disposable income for each postal code (1 represents the lowest income; 1,000, the highest).We also conducted sex-stratified analyses and tested for interaction between sex and greenspace exposure.We assessed the effect of adjusting for gestational age, status as small for gestational age, and birth weight for sensitivity analyses.We also tested the effect of adjusting for maternal parity, which may be a proxy of unmeasured confounding by socioeconomic status.
We calculated risk differences (RD) to assess the potential population health benefits of increased greenspace exposure.First, we categorized NDVI into quartiles to establish defined contrasts (Q1 represents the lowest exposure; Q4, the highest).Next, we fitted the same adjusted logistic regression models described above, except replacing NDVI from a continuous to categorical exposure.Using marginal standardization, (Austin, 2010;Naimi and Whitcomb, 2020) we calculated the marginal probabilities of ASD if every-one in the cohort had been exposed to each NDVI quartile.We then used these marginal probabilities to estimate population average RDs between Q2 vs Q1, Q3 vs Q1, Q4 vs Q1, and between Q3 vs Q2, and Q4 vs Q3 for potential non-linear responses.Ninety-five percent CIs were obtained using 500 bootstrap samples.

Mediation analysis
We conducted causal mediation analyses to evaluate if any and how much of the association between greenspace and ASD might be mediated by reduced air pollution.We used marginal structural models (MSM) and potential outcome methods described by Lange et al. (Lange et al., 2012) to calculate marginal RDs and decompose the total effect of greenspace on risk for ASD into the natural direct effect and the natural indirect effect through air pollution reduction.We developed three mediation models for NDVI exposure at a 250 m buffer size, one model each for PM 2.5 , NO, and NO 2 , following the procedures described below.We also tested the impact for NDVI buffer sizes of 100 and 500 m.In addition, we categorized greenspace and air pollutant exposures into quartiles for the mediation analyses to establish defined contrasts (Q1 represents the lowest exposure; Q4, the highest).
The mediation approach by Lange et al. allows for the decomposition of the natural direct and indirect effects through nested counterfactuals, in which we compare potential outcomes if we intervene on each pregnant woman's exposure and mediator level multiple times; and the mediator can take its natural level corresponding to its observed or counterfactual exposures.Thus, we first replicated our data set to create new counterfactual exposures A* for all possible values of A for each observation (i.e., duplicated each observation for all NDVI quartiles).Then, using the original data set, we developed a multinomial logistic regression for exposure A conditional on confounding variables C (i.e., child sex, birth month, birth year, maternal age, maternal birthplace, and neighborhood-level urbanicity and income band).This regression was applied to the replicated data set to generate predicted probabilities and create exposure weights to balance confounders.Covariate balance was assessed using standardized differences, with greater than 10 % standardized differences suggesting meaningful imbalance.(Austin and Stuart, 2015) Exposure weights were stabilized using marginal probabilities of exposure.Similarly, using the original data set, we developed a multinomial logistic regression for mediator M conditional on exposure A and confounding variables C. Using this regression on the replicated data set, we generated predictions to establish mediator weights, which we then stabilized by the marginal probability of the mediator conditional on exposure and confounding variables C.These mediator weights create a pseudo-population where the mediator is, on average, at the level predicted by exposure A*.
Lastly, we fit a logistic MSM on the replicated data set for outcome Y conditional on A, A*, and their interaction and weighted by the product of the exposure and mediator weights.This fitted model was used to calculate predicted probabilities of outcome Y at observed exposure A and the natural level of the mediator corresponding to counterfactual exposure A*.
We estimated the natural direct effect as the marginal risk difference between the predicted probabilities of outcome Y with exposure A versus counterfactual A* while maintaining the mediator at its natural level corresponding to counterfactual exposure A* (i.e., E[Y a,M a * − Y a * ,M a * ]).Similarly, we calculated the natural indirect effect as the marginal risk difference between the predicted probabilities of outcome Y with exposure A but with the mediator set to its natural values corresponding to exposure A versus counterfactual A* (i.e., E[Y a,Ma − Y a,M a * ]).The total effect was calculated as the marginal risk difference between the predicted probabilities of outcome Y with exposure A versus A* and the mediator set to the corresponding values of A and A*, respectively.The total effect is the sum of the natural direct and indirect effects ).We calculated marginal RDs between NDVI quartiles Q2 vs Q1, Q3 vs Q1, Q4 vs Q1, and between Q3 vs Q2, and Q4 vs Q3.To calculate 95 % CIs, we repeated these procedures across 500 bootstrap replications.
In addition to the assumptions of consistency and positivity, our analyses assumed that the specified confounders are sufficient for all exposure-outcome, exposure-mediator, and mediator-outcome relationships, and that no descendant variable of greenspace confounds the mediator-outcome relationship.(Lange et al., 2017;VanderWeele, 2016).
If no mediation were to be observed, we proposed a post-hoc sensitivity analysis of our primary logistic regression model that additionally controls for air pollutants as confounders.

Results
The linked data contained 132,256 children born in Metro Vancouver from 2004 to 2009.After excluding 3,034 (2.3 %) observations with missing data, our final study cohort consisted of 129,222 children.Within the cohort, 1,921 (1.5 %) children were diagnosed with ASD, and the median age of assessment was 4.8 years (IQR, 3.8-5.8years).Among 66,508 males, 1,618 (2.4 %) were diagnosed with ASD, and among 62,714 females, 303 (0.5 %).Areas with higher NDVI had a larger proportion of mothers born in Canada.Mothers residing in higher NDVI areas also tended to be multiparous and living in higher-income band neighborhoods with lower air pollution (Table 1).

Primary analysis
In the unadjusted logistic regression model, we estimated a decrease in odds of ASD per IQR (0.12) increase in the average of the annual mean NDVI within a 250 m buffer (OR, 0.93 [95 % CI: 0.88, 0.99]; Table 2).After adjusting for child sex, birth month, birth year, maternal age, maternal birthplace, and neighborhood-level urbanicity and income band, the association remained, but the OR was slightly attenuated to 0.96 with the 95 % CI crossing 1.00 (95 % CI: 0.90, 1.02).In sexstratified models, the adjusted OR for male children was lower, and the 95 % CI was more precise compared to female children (male children: OR, 0.95 [95 % CI: 0.89, 1.01]; female children: OR, 1.02 [95 % CI: 0.88, 1.19]).The interaction between NDVI and child sex was not statistically significant in a likelihood ratio test (P value = 0.17).
We tested the effect of including gestational age, small for gestational age, and birth weight in the adjusted model for sensitivity analyses, which resulted in nearly identical ORs and 95 % CIs.However, adding parity to the adjusted models attenuated the effect estimate towards the null (OR, 1.00 [95 % CI: 0.94, 1.07]).Changing the NDVI buffer to 100 and 500 m produced similar effect estimates as the 250 m buffer (Table 2; Figure S5 in the Supplement).
The greatest marginal risk differences were observed for Q4 vs Q1 (reference) NDVI increase, with an estimated unadjusted RD of − 219.73 (95 % CI: − 401.56, − 18.01) per 100,000 and after adjustment, RD of − 124.34 (95 % CI: − 311.73, 87.54) per 100,000 (Table 3).The 95 % CIs for the adjusted risk difference models crossed zero.In sex-stratified models, males had a greater risk reduction for ASD than females across NDVI quartile comparisons.Adding gestational age, small for gestational age, and birth weight as covariates to the adjusted models produced small risk difference changes, with absolute differences ranging from 0.1 to 11.3 per 100,000 across NDVI quartile comparisons.Adjusting for parity, however, attenuated the adjusted risk differences.Similar patterns were observed for the 100 and 500 m NDVI buffers (Tables S1 and S2, and Figure S6 in the Supplement).

Mediation analysis
Covariate balance was assessed before and after applying stabilized weights.For the 250 m NDVI buffer and PM 2.5 mediator model, standardized differences among covariates ranged from 0.5 to 27.3 % (Figure S7 in the Supplement).After applying stabilized exposure weights, these standardized differences decreased to 0.1-1.8%, suggesting good covariate balance.Results were similar between NO and NO 2 mediation models, with standard differences initially ranging from 0.5 to 27.3 % and decreasing to 0.1-1.8% after applying stabilized exposure weights.After applying the stabilized weights, covariate imbalance increased slightly, as expected, with standardized differences ranging from 0.2 to 3.5 %, 0.3-3.7 %, and 0.4-5.6 %, respectively for PM 2.5 , NO, and NO 2 mediation models.
Table 4 shows the natural direct, natural indirect, and total effects of NDVI exposure during pregnancy for risk of ASD, estimated as marginal risk differences per 100,000.We observed no mediation, with natural direct, natural indirect, and total effects crossing zero across all NDVI quartile comparisons for PM 2.5 , NO, and NO 2 mediation models.Interactions between NDVI exposure and air pollutants mediators were not significant based on likelihood ratio tests (PM 2.5 : P value = 0.94; NO: P value = 0.18; NO 2 : P value = 0.09).Similar results were observed for 100 and 500 m NDVI buffers (Tables S3 and S4, and Figures S8-S10 in the Supplement).

Discussions
Our study assessed the relationship between prenatal greenspace exposure in several buffer zones around residence and ASD diagnosis in a population-based Canadian birth cohort comprising almost 130,000 children.We also tested whether any potential association was mediated by a reduction in PM 2.5 , NO, and NO 2. In the unadjusted models of our primary analyses, we found an association between prenatal greenspace exposure and reduced risk of ASD.This association was slightly attenuated with the OR 95 % CIs crossing one after controlling for confounders in the 250 m buffer zone, but remained statistically significant in the 100 m buffer zone.No interaction was observed between NDVI exposure and sex.There were no significant marginal risk differences per quartiles of greenspace exposure in adjusted models.We found no evidence of an indirect effect through air pollution reduction.Our results were consistent across several sensitivity analyses assessing associations in 100 and 500 m buffer zones and after adding birth outcomes and parity as potential confounders.The risk reduction at the 100 m buffer could suggest that greenspace exposure during pregnancy may reduce ASD risk, for example, mediated through greenspace-related stress reduction in pregnant women viewing green spaces from their residence, but this needs to be confirmed with finer spatial resolution greenspace data.(Labib et al., 2020) Future analyses should also examine links between greenspace and maternal mental health in pregnancy.While our primary hypothesis was that air pollution reduction would be on the causal pathway, we adapted the models in a post hoc analysis treating air pollutants as confounders after observing no mediation.Controlling for air pollutants further attenuated the adjusted ORs.
Our study showed that prenatal greenspace exposure was associated with a small reduction in the odds of ASD, but no significant risk differences in the additive scale nor mediation through air pollution reduction.To our knowledge, this is the first study to evaluate this association.One previous American study by Wu and Jackson analyzed ASD prevalence in elementary school districts in California and found an inverse relationship to several greenspace metrics, including tree canopy and green land cover types.(Wu and Jackson, 2017) This study, however, did not evaluate prenatal exposure, and it used an ecological study design.Another difference, which complicates comparison, is that the study by Wu and Jackson was conducted in a highly urbanized area with substantially higher levels of air pollution and less greenspace exposure than Metro Vancouver.The relatively high amount of greenspace in our study area reduces the variation in exposure estimates, possibly biasing our results towards the null.It would be particularly important to test this association in different geographical and climate contexts, including areas with higher and lower urbanization, to evaluate whether the results are dependent on relative environmental exposures.

Table 2
Odds of autism spectrum disorder for greenspace exposure during pregnancy.Main model adjusted for child sex, birth month, birth year, maternal age, maternal birthplace, and neighborhood-level urbanicity and income band.Abbreviations: CI, confidence interval; IQR, interquartile range; NDVI, normalized difference vegetation index; and ASD, autism spectrum disorder.Although some previous studies have analyzed prenatal environmental exposures with mental and cognitive health outcomes in later life, (Xu et al., 1987;Essers et al., 2021;Johnson et al., 2021;Carter et al., 2022) similar research on greenspace exposure is scarce.Studies have found associations between greenspace exposure during pregnancy and gestational age, small for gestational age, and birth weight, (Torres Toda et al., 2020;Donovan et al., 2019) although the findings are inconsistent.(Dadvand et al., 2012;Glazer et al., 2018) In our study, adjusting for gestational age, small for gestational age, and birth weight as potential confounders did not meaningfully change effect estimates.
Numerous studies provide support for the impact of greenspace exposure on children's behavioral and mental health, such as reduced problems with attention (Markevych et al., 2014) and improved cognitive (Dadvand et al., 2015) and early development.(Jarvis et al., 2021) However, these studies have not examined prenatal exposures.The etiology of ASD is not entirely understood, and multiple studies suggest that the etiology may have prenatal origins.(Bauman and Kemper, 2005) While there are hypothesized pathways for improved mental health outcomes in offspring from prenatal greenspace exposure, such as reduced air pollution (Asta et al., 2021) and lower stress levels (Verheyen et al., 2021) or lower risk of depression in pregnant mothers, (McEachan et al., 2016) the effect of greenspace on ASD may be too small to detect in our study context, or there could indeed be no effect.Metro Vancouver has a high amount of greenspace and relatively low air pollution levels.
There are a few limitations to our study.First, greenspace exposure was measured by NDVI derived from 30 m spatial resolution remote sensing imagery, which cannot distinguish between different greenspace attributes or qualities (e.g., private versus public spaces, trees versus grass, or quality of amenities in public parks).Future studies should evaluate if certain types of natural environments, including blue spaces, may be inversely associated with developing ASD.We also recommend including greenspace metrics of higher granularity, such as percentage vegetation derived from spectral unmixing of satellite imagery.(Jarvis et al., 2021;Czekajlo et al., 2020) This would increase the exposure variation and the possibility to detect effects.Nevertheless, NDVI is among the most commonly used greenspace metrics in epidemiological research, (Davis et al., 2021) which contributes to the comparability of our study results.Second, there is a risk of exposure misclassification related to residential postal code data linkages.While postal codes typically correspond to a small area, approximately half a block in dense city environments, it does not provide the same specificity in exposure alignment as an individual address.As well, residential postal codes do not account for time-activity patterns.Third, limited data on individuallevel socioeconomic status is a limitation of administrative health data.The attenuation of results after adjusting the models for parity-a potential proxy for socioeconomic status-may suggest some residual confounding by individual-level socioeconomic status not accounted for by neighborhood-level income.Education and income are negatively associated with a woman's number of children.(Huber et al., 2010) In Metro Vancouver, areas with lower socioeconomic status tend to have lower greenspace and higher air pollution levels.(Doiron et al., 2020) More studies are needed to detangle the impacts of socioeconomic and physical environments, for example, through an exposome approach that includes biological determinants.(Julvez et al., 1987;Maitre et al., 2021;Robinson et al., 2018) It is also plausible that parents of lower socioeconomic status were less able to access ASD assessment for their children, and thus more children would be undiagnosed on the lower end of the socioeconomic spectrum.Given that green space is also correlated with increasing socioeconomic status, this would bias our results towards the null.We were also unable to control for all potentially relevant confounders, such as maternal mental health.With these considerations in mind, it is conceivable that greenspace exposure during pregnancy may be a proxy exposure for yet unmeasured factors that are more proximal to the early origins of ASD.
Our study also has several notable strengths, and to our knowledge, it is the first study to analyze the impact of prenatal exposure of greenspace on ASD.This work contributes new knowledge on the potential protective effects of modifiable environmental exposures to help prevent childhood developmental disabilities.We used data from a large population-based cohort and strict, clinically verified diagnostic criteria for case ascertainment.(Bickford et al., 2020) Our temporally adjusted land use regression models for air pollution have a high spatial and temporal resolution, (Henderson et al., 2007;Wang et al., 2013;Nethery et al., 2008) reducing measurement error.Although we were unable to conduct trimester-specific analyses, exposures over the pregnancy period tend to be correlated.Future studies should consider analyzing pregnant mothers who lived in a low NDVI area for one period of their pregnancy and then moved to a higher NDVI area (or in the oppositive

Table 4
Natural direct, natural indirect, and total effects of greenspace exposure during pregnancy on autism spectrum disorder with particulate matter with a diameter less than 2.5 µm (PM 2.5 ), nitric oxide (NO), and nitrogen dioxide (NO 2 ) as mediators.Effects estimated using marginal structural models, which included covariates for child sex, birth month, birth year, maternal age, maternal birthplace, and neighborhood-level urbanicity and income band.Q1 represents the lowest normalized difference vegetation index (NDVI) exposure; Q4, the highest.The 95 % confidence intervals (CI) were derived via bootstrapping with 500 replicates.direction).Capturing this discordance in exposure in the same mother would allow for improved exposure contrast.

Conclusion
In this pioneering hypothesis-driven study, prenatal greenspace exposure was associated with a small reduction in the odds of ASD, but no significant risk differences were found in the additive scale nor mediation through air pollution reduction.Further research is needed to evaluate prenatal and long-term exposure to greenspace and ASD in different geographical, social, and climate contexts and using greenspace metrics of higher variability and spatial resolution.Future directions include focusing on specific gestational time windows of vulnerability and identifying factors that may elucidate how and why prenatal greenspace shapes child development.

Table 1
Cohort characteristics by quartiles of average of annual mean normalized difference vegetation index.Median average of annual mean NDVI was 0.35 (IQR, 0.29-0.41).Abbreviations: NDVI, normalized difference vegetation index; SMD, standardized mean difference; IQR, interquartile range; PM 2.5 , particulate matter with a diameter less than 2.5 µm; NO, nitric oxide; NO 2 , nitrogen dioxide; and ppb, parts per billion.

Table 3
Marginal risk differences of autism spectrum disorder per 100,000.Main model adjusted for child sex, birth month, birth year, maternal age, maternal birthplace, and neighborhood-level urbanicity and income band.Q1 represents the lowest normalized difference vegetation index (NDVI) exposure; Q4, the highest.The 95% confidence intervals (CI) were derived via bootstrapping with 500 replicates.