Use of generalized additive models and cokriging of spatial residuals to improve land-use regression estimates of nitrogen oxides in Southern California

Development of Spatio-Temporal Land Use Regression Models for Fine Particulate Matter and Wood-Burning Tracers in Temuco, Chile

Biomass burning is common in much of the world, and in some areas, residential wood-burning has increased. However, air pollution resulting from biomass burning is an important public health problem. A sampling campaign was carried out between May 2017 and July 2018 in over 64 sites in four sessions, to develop a spatio-temporal land use regression (LUR) model for fine particulate matter (PM) and wood-burning tracers levoglucosan and soluble potassium (Ksol) in a city heavily impacted by wood-burning. The mean (sd) was 46.5 (37.4) μg m-3 for PM2.5, 0.607 (0.538) μg m-3 for levoglucosan, and 0.635 (0.489) μg m-3 for Ksol. LUR models for PM2.5, levoglucosan, and Ksol had a satisfactory performance (LOSOCV R2), explaining 88.8%, 87.4%, and 87.3% of the total variance, respectively. All models included sociodemographic predictors consistent with the pattern of use of wood-burning in homes. The models were applied to predict concentrations surfaces and to estimate exposures for an epidemiological study.

A pruned feed-forward neural network (pruned-FNN) approach to measure air pollution exposure

Environmental epidemiology studies require accurate estimations of exposure intensities to air pollution. The process from air pollutant emission to individual exposure is however complex and nonlinear, which poses significant modeling challenges. This study aims to develop an exposure assessment model that can strike a balance between accuracy, complexity, and usability. In this regard, neural networks offer one possible approach. This study employed a custom-designed pruned feed-forward neural network (pruned-FNN) approach to calculate the air pollution exposure index based on emission time and rates, terrain factors, meteorological conditions, and proximity measurements. The model's performance was evaluated by cross-validating the estimated exposure indexes with ground-based monitoring records. The pruned FNN can predict pollution exposure indexes (PEIs) that are highly and stably correlated with the monitored air pollutant concentrations (Spearman's rank correlation coefficients for tenfold cross-validation (mean ± standard deviation: 0.906 ± 0.028) and for random cross-validation (0.913 ± 0.024)). The predicted values are also close to the ground truth in most cases (95.5% of the predicted PEIs have relative errors smaller than 10%) when the training datasets are sufficiently large and well-covered. The pruned-FNN method can make accurate exposure estimations using a flexible number of variables and less extensive data in a less money/time-consuming manner. Compared to other exposure assessment models, the pruned FNN is an appropriate and effective approach for exposure assessment that covers a large geographic area over a long period of time.

Co-kriging with a low-cost sensor network to estimate spatial variation of brake and tire-wear metals and oxidative stress potential in Southern California

Due to regulations and technological advancements reducing tailpipe emissions, an increasing proportion of emissions arise from brake and tire wear particulate matter (PM). PM from these non-tailpipe sources contains heavy metals capable of generating oxidative stress in the lung. Although important, these particles remain understudied because the high cost of actively collecting filter samples. Improvements in electrical engineering, internet connectivity, and an increased public concern over air pollution have led to a proliferation of dense low-cost air sensor networks such as the PurpleAir monitors, which primarily measure unspeciated fine particulate matter (PM2.5). In this study, we model the concentrations of Ba, Zn, black carbon, reactive oxygen species concentration in the epithelial lining fluid, dithiothreitol (DTT) loss, and OH formation. We use a co-kriging approach, incorporating data from the PurpleAir network as a secondary predictor variable and a land-use regression (LUR) as an external drift. For most pollutant species, co-kriging models produced more accurate predictions than an LUR model, which did not incorporate data from the PurpleAir monitors. This finding suggests that low-cost sensors can enhance predictions of pollutants that are costly to measure extensively in the field.

National Empirical Models of Air Pollution Using Microscale Measures of the Urban Environment

National-scale empirical models of air pollution (e.g., Land Use Regression) rely on predictor variables (e.g., population density, land cover) at different geographic scales. These models typically lack microscale variables (e.g., street level), which may improve prediction with fine-spatial gradients. We developed microscale variables of the urban environment including Point of Interest (POI) data, Google Street View (GSV) imagery, and satellite-based measures of urban form. We developed United States national models for six criteria pollutants (NO₂, PM_2.5, O₃, CO, PM₁₀, SO₂) using various modeling approaches: Stepwise Regression + kriging (SW-K), Partial Least Squares + kriging (PLS-K), and Machine Learning + kriging (ML-K). We compared predictor variables (e.g., traditional vs microscale) and emerging modeling approaches (ML-K) to well-established approaches (i.e., traditional variables in a PLS-K or SW-K framework). We found that combined predictor variables (traditional + microscale) in the ML-K models outperformed the well-established approaches (10-fold spatial cross-validation (CV) R² increased 0.02-0.42 [average: 0.19] among six criteria pollutants). Comparing all model types using microscale variables to models with traditional variables, the performance is similar (average difference of 10-fold spatial CV R² = 0.05) suggesting microscale variables are a suitable substitute for traditional variables. ML-K and microscale variables show promise for improving national empirical models.

"Spatial heterogeneity of environmental risk in randomized prevention trials: consequences and modeling"

Background

In the context of environmentally influenced communicable diseases, proximity to environmental sources results in spatial heterogeneity of risk, which is sometimes difficult to measure in the field. Most prevention trials use randomization to achieve comparability between groups, thus failing to account for heterogeneity.

This study aimed to determine under what conditions spatial heterogeneity biases the results of randomized prevention trials, and to compare different approaches to modeling this heterogeneity.

Methods

Using the example of a malaria prevention trial, simulations were performed to quantify the impact of spatial heterogeneity and to compare different models.

Simulated scenarios combined variation in baseline risk, a continuous protective factor (age), a non-related factor (sex), and a binary protective factor (preventive treatment). Simulated spatial heterogeneity scenarios combined variation in breeding site density and effect, location, and population density.

The performances of the following five statistical models were assessed: a non-spatial Cox Proportional Hazard (Cox-PH) model and four models accounting for spatial heterogeneity—i.e., a Data-Generating Model, a Generalized Additive Model (GAM), and two Stochastic Partial Differential Equation (SPDE) models, one modeling survival time and the other the number of events. Using a Bayesian approach, we estimated the SPDE models with an Integrated Nested Laplace Approximation algorithm.

For each factor (age, sex, treatment), model performances were assessed by quantifying parameter estimation biases, mean square errors, confidence interval coverage rates (CRs), and significance rates. The four models were applied to data from a malaria transmission blocking vaccine candidate.

Results

The level of baseline risk did not affect our estimates. However, with a high breeding site density and a strong breeding site effect, the Cox-PH and GAM models underestimated the age and treatment effects (but not the sex effect) with a low CR.

When population density was low, the Cox-SPDE model slightly overestimated the effect of related factors (age, treatment). The two SPDE models corrected the impact of spatial heterogeneity, thus providing the best estimates.

Conclusion

Our results show that when spatial heterogeneity is important but not measured, randomization alone cannot achieve comparability between groups. In such cases, prevention trials should model spatial heterogeneity with an adapted method.

Trial registration

The dataset used for the application example was extracted from Vaccine Trial #NCT02334462 (ClinicalTrials.gov registry).

Electronic supplementary material

The online version of this article (10.1186/s12874-019-0759-z) contains supplementary material, which is available to authorized users.

Cluster-based bagging of constrained mixed-effects models for high spatiotemporal resolution nitrogen oxides prediction over large regions

BACKGROUND

Accurate estimation of nitrogen dioxide (NO₂) and nitrogen oxide (NO_x) concentrations at high spatiotemporal resolutions is crucial for improving evaluation of their health effects, particularly with respect to short-term exposures and acute health outcomes. For estimation over large regions like California, high spatial density field campaign measurements can be combined with more sparse routine monitoring network measurements to capture spatiotemporal variability of NO₂ and NO_x concentrations. However, monitors in spatially dense field sampling are often highly clustered and their uneven distribution creates a challenge for such combined use. Furthermore, heterogeneities due to seasonal patterns of meteorology and source mixtures between sub-regions (e.g. southern vs. northern California) need to be addressed.

OBJECTIVES

In this study, we aim to develop highly accurate and adaptive machine learning models to predict high-resolution NO₂ and NO_x concentrations over large geographic regions using measurements from different sources that contain samples with heterogeneous spatiotemporal distributions and clustering patterns.

METHODS

We used a comprehensive Kruskal-K-means method to cluster the measurement samples from multiple heterogeneous sources. Spatiotemporal cluster-based bootstrap aggregating (bagging) of the base mixed-effects models was then applied, leveraging the clusters to obtain balanced and less correlated training samples for less bias and improvement in generalization. Further, we used the machine learning technique of grid search to find the optimal interaction of temporal basis functions and the scale of spatial effects, which, together with spatiotemporal covariates, adequately captured spatiotemporal variability in NO₂ and NO_x at the state and local levels.

RESULTS

We found an optimal combination of four temporal basis functions and 200 m scale spatial effects for the base mixed-effects models. With the cluster-based bagging of the base models, we obtained robust predictions with an ensemble cross validation R² of 0.88 for both NO₂ and NO_x [RMSE (RMSEIQR): 3.62 ppb (0.28) and 9.63 ppb (0.37) respectively]. In independent tests of random sampling, our models achieved similarly strong performance (R² of 0.87-0.90; RMSE of 3.97-9.69 ppb; RMSEIQR of 0.21-0.27), illustrating minimal over-fitting.

CONCLUSIONS

Our approach has important implications for fusing data from highly clustered and heterogeneous measurement samples from multiple data sources to produce highly accurate concentration estimates of air pollutants such as NO₂ and NO_x at high resolution over a large region.

Association between ambient air pollution and breast cancer risk: The multiethnic cohort study

Previous studies using different exposure methods to assess air pollution and breast cancer risk among primarily whites have been inconclusive. Air pollutant exposures of particulate matter and oxides of nitrogen were estimated by kriging (NO_x , NO₂ , PM₁₀ , PM_2.5 ), land use regression (LUR, NO_x , NO₂ ) and California Line Source Dispersion model (CALINE4, NO_x , PM_2.5 ) for 57,589 females from the Multiethnic Cohort, residing largely in Los Angeles County from recruitment (1993-1996) through 2010. Cox proportional hazards models were used to examine the associations between time-varying air pollution and breast cancer incidence adjusting for confounding factors. Stratified analyses were conducted by race/ethnicity and distance to major roads. Among all women, breast cancer risk was positively but not significantly associated with NO_x (per 50 parts per billion [ppb]) and NO₂ (per 20 ppb) determined by kriging and LUR and with PM_2.5 and PM₁₀ (per 10 μg/m³ ) determined by kriging. However, among women who lived within 500 m of major roads, significantly increased risks were observed with NO_x (hazard ratio [HR] = 1.35, 95% confidence interval [95% CI]: 1.02-1.79), NO₂ (HR = 1.44, 95% CI: 1.04-1.99), PM₁₀ (HR = 1.29, 95% CI: 1.07-1.55) and PM_2.5 (HR = 1.85, 95% CI: 1.15-2.99) determined by kriging and NO_x (HR = 1.21, 95% CI:1.01-1.45) and NO₂ (HR = 1.26, 95% CI: 1.00-1.59) determined by LUR. No overall associations were observed with exposures assessed by CALINE4. Subgroup analyses suggested stronger associations of NO_x and NO₂ among African Americans and Japanese Americans. Further studies of multiethnic populations to confirm the effects of air pollution, particularly near-roadway exposures, on the risk of breast cancer is warranted.

Air Pollution Dispersion Modelling Using Spatial Analyses

Air pollution dispersion modelling via spatial analyses (Land Use Regression—LUR) is an alternative approach to the standard air pollution dispersion modelling techniques in air quality assessment. Its advantages are mainly a much simpler mathematical apparatus, quicker and simpler calculations and a possibility to incorporate more factors affecting pollutant’s concentration than standard dispersion models. The goal of the study was to model the PM10 particles dispersion via spatial analyses in the Czech–Polish border area of the Upper Silesian industrial agglomeration and compare the results with the results of the standard Gaussian dispersion model SYMOS’97. The results show that standard Gaussian model with the same data as the LUR model gives better results (determination coefficient 71% for Gaussian model to 48% for LUR model). When factors of the land cover were included in the LUR model, the LUR model results improved significantly (65% determination coefficient) to a level comparable with the Gaussian model. A hybrid approach of combining the Gaussian model with the LUR gives superior quality of results (86% determination coefficient).

Spatiotemporal land use random forest model for estimating metropolitan NO2 exposure in Japan

Adequate spatial and temporal estimates of NO₂ concentrations are essential for proper prenatal exposure assessment. Here, we develop a spatiotemporal land use random forest (LURF) model of the monthly mean NO₂ over four years in a metropolitan area of Japan. The overall objective is to obtain accurate NO₂ estimates for use in prenatal exposure assessments. We use random forests to convey the non-linear relationship between NO₂ concentrations and predictor variables, and compare the prediction accuracy with that of a linear regression. In addition, we include the distance decay effect of emission sources on NO₂ concentrations for more efficient model construction. The prediction accuracy of the LURF model is evaluated through a leave-one-monitor-out cross validation. We obtain a high R² value of 0.79, which is better than that of the conventional land use regression model using linear regression (R² of 0.73). We also evaluate the LURF model via a temporal and overall cross validation and obtain R² values of 0.84 and 0.92, respectively. We successfully integrate temporal and spatial components into our model, which exhibits higher accuracy than spatial models constructed individually for each month. Our findings illustrate the advantage of using a LURF to model the spatiotemporal variability of NO₂ concentrations.

Source Characterization and Exposure Modeling of Gas-Phase Polycyclic Aromatic Hydrocarbon (PAH) Concentrations in Southern California

Airborne exposures to polycyclic aromatic hydrocarbons (PAHs) are associated with adverse health outcomes. Because personal air measurements of PAHs are labor intensive and costly, spatial PAH exposure models are useful for epidemiological studies. However, few studies provide adequate spatial coverage to reflect intra-urban variability of ambient PAHs. In this study, we collected 39-40 weekly gas-phase PAH samples in southern California twice in summer and twice in winter, 2009, in order to characterize PAH source contributions and develop spatial models that can estimate gas-phase PAH concentrations at a high resolution. A spatial mixed regression model was constructed, including such variables as roadway, traffic, land-use, vegetation index, commercial cooking facilities, meteorology, and population density. Cross validation of the model resulted in an R² of 0.66 for summer and 0.77 for winter. Results showed higher total PAH concentrations in winter. Pyrogenic sources, such as fossil fuels and diesel exhaust, were the most dominant contributors to total PAHs. PAH sources varied by season, with a higher fossil fuel and wood burning contribution in winter. Spatial autocorrelation accounted for a substantial amount of the variance in total PAH concentrations for both winter (56%) and summer (19%). In summer, other key variables explaining the variance included meteorological factors (9%), population density (15%), and roadway length (21%). In winter, the variance was also explained by traffic density (16%). In this study, source characterization confirmed the dominance of traffic and other fossil fuel sources to total measured gas-phase PAH concentrations while a spatial exposure model identified key predictors of PAH concentrations. Gas-phase PAH source characterization and exposure estimation is of high utility to epidemiologist and policy makers interested in understanding the health impacts of gas-phase PAHs and strategies to reduce emissions.

Constrained Mixed-Effect Models with Ensemble Learning for Prediction of Nitrogen Oxides Concentrations at High Spatiotemporal Resolution

Spatiotemporal models to estimate ambient exposures at high spatiotemporal resolutions are crucial in large-scale air pollution epidemiological studies that follow participants over extended periods. Previous models typically rely on central-site monitoring data and/or covered short periods, limiting their applications to long-term cohort studies. Here we developed a spatiotemporal model that can reliably predict nitrogen oxide concentrations with a high spatiotemporal resolution over a long time span (>20 years). Leveraging the spatially extensive highly clustered exposure data from short-term measurement campaigns across 1-2 years and long-term central site monitoring in 1992-2013, we developed an integrated mixed-effect model with uncertainty estimates. Our statistical model incorporated nonlinear and spatial effects to reduce bias. Identified important predictors included temporal basis predictors, traffic indicators, population density, and subcounty-level mean pollutant concentrations. Substantial spatial autocorrelation (11-13%) was observed between neighboring communities. Ensemble learning and constrained optimization were used to enhance reliability of estimation over a large metropolitan area and a long period. The ensemble predictions of biweekly concentrations resulted in an R2 of 0.85 (RMSE: 4.7 ppb) for NO2 and 0.86 (RMSE: 13.4 ppb) for NOx. Ensemble learning and constrained optimization generated stable time series, which notably improved the results compared with those from initial mixed-effects models.

A spatial multicriteria model for determining air pollution at sample locations

Atmospheric pollution in urban centers has been one of the main causes of human illness related to the respiratory and circulatory system. Efficient monitoring of air quality is a source of information for environmental management and public health. This study investigates the spatial patterns of atmospheric pollution using a spatial multicriteria model that helps target locations for air pollution monitoring sites. The main objective was to identify high-priority areas for measuring human exposures to air pollutants as they relate to emission sources. The method proved to be viable and flexible in its application to various areas.

Asthma morbidity and ambient air pollution: effect modification by residential traffic-related air pollution

BACKGROUND

Ambient air pollution has been associated with asthma-related hospital admissions and emergency department visits (hospital encounters). We hypothesized that higher individual exposure to residential traffic-related air pollutants would enhance these associations.

METHODS

We studied 11,390 asthma-related hospital encounters among 7492 subjects 0-18 years of age living in Orange County, California. Ambient exposures were measured at regional air monitoring stations. Seasonal average traffic-related exposures (PM2.5, ultrafine particles, NOx, and CO) were estimated near subjects' geocoded residences for 6-month warm and cool seasonal periods, using dispersion models based on local traffic within 500 m radii. Associations were tested in case-crossover conditional logistic regression models adjusted for temperature and humidity. We assessed effect modification by seasonal residential traffic-related air pollution exposures above and below median dispersion-modeled exposures. Secondary analyses considered effect modification by traffic exposures within race/ethnicity and insurance group strata.

RESULTS

Asthma morbidity was positively associated with daily ambient O3 and PM2.5 in warm seasons and with CO, NOx, and PM2.5 in cool seasons. Associations with CO, NOx, and PM2.5 were stronger among subjects living at residences with above-median traffic-related exposures, especially in cool seasons. Secondary analyses showed no consistent differences in association, and 95% confidence intervals were wide, indicating a lack of precision for estimating these highly stratified associations.

CONCLUSIONS

Associations of asthma with ambient air pollution were enhanced among subjects living in homes with high traffic-related air pollution. This may be because of increased susceptibility (greater asthma severity) or increased vulnerability (meteorologic amplification of local vs. correlated ambient exposures).

Modeling the concentrations of on-road air pollutants in southern California

High concentrations of air pollutants on roadways, relative to ambient concentrations, contribute significantly to total personal exposure. Estimation of these exposures requires measurements or prediction of roadway concentrations. Our study develops, compares, and evaluates linear regression and nonlinear generalized additive models (GAMs) to estimate on-road concentrations of four key air pollutants, particle-bound polycyclic aromatic hydrocarbons (PB-PAH), particle number count (PNC), nitrogen oxides (NOx), and particulate matter with diameter <2.5 μm (PM2.5) using traffic, meteorology, and elevation variables. Critical predictors included wind speed and direction for all the pollutants, traffic-related variables for PB-PAH, PNC, and NOx, and air temperatures and relative humidity for PM2.5. GAMs explained 50%, 55%, 46%, and 71% of the variance for log or square-root transformed concentrations of PB-PAH, PNC, NOx, and PM2.5, respectively, an improvement of 5% to over 15% over the linear models. Accounting for temporal autocorrelation in the GAMs further improved the prediction, explaining 57-89% of the variance. We concluded that traffic and meteorological data are good predictors in estimating on-road traffic-related air pollutant concentrations and GAMs perform better for nonlinear variables, such as meteorological parameters.

Estimating Spatiotemporal Variability of Ambient Air Pollutant Concentrations with A Hierarchical Model

Studies have linked exposure to air pollutants to short-term and sub-chronic health outcomes. However, individual-level air pollution exposure is difficult to measure at a high spatial and temporal resolution and for larger populations due to limitations in sampling techniques. We presented a hierarchical model to capture spatiotemporal variability of nitrogen dioxide (NO2) and nitrogen oxides (NOx) concentrations in Southern California by combining high temporal resolution data from routine monitoring stations with high spatial resolution data from investigator-initiated episodic measurements. In this model, the spatiotemporal field of concentrations was first decomposed into a mean and residual and the mean representing the seasonal trend was further decomposed into a constant and varying temporal basis functions. The mean of the spatially varying coefficients of temporal basis functions were modeled by local covariates using non-linear generalized additive model and least square fitting using measurements from both routine monitoring and additional episodic sampling locations, while the spatially-correlated residuals of the coefficients were co-kriged. We found traffic, land-use and wind accounted for a large portion of the variance (beyond 35%) for the long-term average trend of concentrations. Spatial residuals accounted for a large portion of the variance of the temporal components (about 30% for NO2 and 20% for NOx). Leave-one-out cross validation produced an R2 of 0.84 for NO2 and 0.81 for NOx when comparing the modeled weekly concentration with the observed trends at all routine monitoring stations.