EPA Vectors: Spatial Predictive Pollution Mixtures

• EPA Vectors are multivariate statistical objects that represent pollutant mixtures and state-dependent air pollution profiles, enabling dimensionality reduction and spatial predictability.
• They are computed via predictive sparse PCA and universal kriging, leveraging geographic covariates to optimize both interpretability and spatial interpolation.
• Additionally, EPA Vectors underpin dynamic regime-switching models that capture temporal autocorrelation and cross-pollutant dependencies for enhanced epidemiological risk analysis.

EPA vectors are multivariate statistical objects representing pollutant mixtures or state-dependent profiles of air pollution, motivated by analyses of data from the U.S. Environmental Protection Agency (EPA) monitoring networks. These vectors arise in distinct but related contexts: as sparse principal component loadings constructed to maximize both variability explained and spatial predictability, and as dynamic conditional mean vectors and covariances in hidden semi-Markov vector autoregressive (VAR) models. The principal aims are interpretability, spatial predictability, and effective dimensionality reduction for downstream epidemiological and environmental risk analyses, particularly when the pollutant data are spatially misaligned or temporally autocorrelated.

1. Mathematical Formulation of EPA Vectors: Predictive Sparse PCA

The EPA vector framework introduced by Jandarov et al. formalizes the extraction of pollutant mixtures from high-dimensional, spatially-misaligned data via predictive sparse principal component analysis (PCA) (Jandarov et al., 2015). Given a data matrix $X\in \mathbb{R}^{n\times p}$ of transformed and standardized pollutant concentrations measured at $n$ monitoring sites for $p$ pollutants, and a matrix $\widetilde Z\in \mathbb{R}^{n\times m}$ of geographic covariates at those sites, the objective is to identify $k\ll p$ sparse loading vectors $\{v_\ell\}_{\ell=1}^k$ and corresponding score vectors $u_\ell = X v_\ell$. The vectors $v_\ell$ encode interpretable pollutant mixtures, while the scores $u_\ell$ are required to be highly predictable by geostatistical models based on $\widetilde Z$.

For each component $n$0, the method solves

$n$1

where $n$2 is a tuning parameter promoting sparsity in $n$3 and $n$4 is constrained to the column space of $n$5 (Jandarov et al., 2015). The components are extracted sequentially with a deflation step.

2. Spatial Prediction via Universal Kriging

EPA vectors operationalize spatial prediction of population-level exposure mixtures. After obtaining scores $n$6 at monitor locations, universal kriging is used to interpolate these scores at unmeasured locations. The model

$n$7

is specified for each site $n$8, with $n$9 following a stationary Gaussian process with covariance parameters $p$0. The kriging predictor at a new location $p$1 is given by

$p$2

This spatial anchoring, enforced at the PCA stage, improves out-of-sample prediction accuracy relative to unconstrained PCA (Jandarov et al., 2015).

3. Algorithmic Workflow and Model Tuning

The EPA vector construction follows a rigorous workflow:

• Data standardization: Pollutant data $p$3 are centered and scaled; covariate matrix $p$4 incorporates preprocessed GIS predictors and spatial basis functions.
• Sequential predictive sparse PCA: For each component, alternate updates are performed for $p$5 (closed form) and for $p$6 (component-wise soft thresholding) until convergence.
• Cross-validation: The sparsity parameter $p$7 is tuned to optimize either out-of-sample kriging $p$8 or Frobenius norm reconstruction error on held-out data.
• Score kriging and exposure estimation: Universal kriging is fitted for each score, and exposures for cohort individuals are predicted at specific locations.
• Health effect estimation: Predicted mixture exposures serve as covariates in regression models assessing health outcomes (Jandarov et al., 2015).

4. EPA Vectors in Probabilistic Regime-Switching Models

EPA vectors also denote the conditional mean vectors and covariances of multivariate pollutant concentrations in dynamic, regime-switching models based on nonhomogeneous hidden semi-Markov chains with penalized VAR observation models (Mingione et al., 17 Sep 2025). Formally, for each hidden regime $p$9, the expected pollutant vector

$\widetilde Z\in \mathbb{R}^{n\times m}$0

defines the EPA pollutant vector for state $\widetilde Z\in \mathbb{R}^{n\times m}$1 at time $\widetilde Z\in \mathbb{R}^{n\times m}$2. This framework captures temporal autocorrelation, cross-pollutant dependence, and nonstationary regime behavior, with LASSO regularization ensuring automatic lag and variable selection (Mingione et al., 17 Sep 2025).

EPA pollutant vectors at time $\widetilde Z\in \mathbb{R}^{n\times m}$3 are characterized in terms of: $\widetilde Z\in \mathbb{R}^{n\times m}$4 where $\widetilde Z\in \mathbb{R}^{n\times m}$5 are filtered regime probabilities, $\widetilde Z\in \mathbb{R}^{n\times m}$6 the conditional means, and $\widetilde Z\in \mathbb{R}^{n\times m}$7 the state-specific covariances, reconstructing the statistical profile of multivariate air quality at each time point.

5. Applied Illustration: U.S. EPA Speciation Networks

An applied case study uses annual averages for $\widetilde Z\in \mathbb{R}^{n\times m}$8 PM$\widetilde Z\in \mathbb{R}^{n\times m}$9 species at $k\ll p$0 monitors, with $k\ll p$1 GIS covariates and thin-plate spline spatial basis. Three EPA vectors are extracted, each interpretable as a mixture:

EPA Vector	Dominant Loadings	Kriging $k\ll p$2
$k\ll p$3	S, OC, SO$k\ll p$4 surrogate, Zn	$k\ll p$5
$k\ll p$6	EC, OC, Ni, V	$k\ll p$7
$k\ll p$8	Ca, Si, Al, K	$k\ll p$9

$\{v_\ell\}_{\ell=1}^k$0 identifies a sulfur-rich mixture prevalent in industrial regions; $\{v_\ell\}_{\ell=1}^k$1 corresponds to traffic/combustion sources; $\{v_\ell\}_{\ell=1}^k$2 reflects soil/dust loadings. Spatial predictions of these vectors display interpretable gradients, and their predicted exposures are used for epidemiological regression modeling, where coefficients $\{v_\ell\}_{\ell=1}^k$3 quantify health effects of exposure to each mixture (Jandarov et al., 2015).

6. Diagnostics, Episode Detection, and Risk Attribution

In hidden semi-Markov VAR models, high-pollution episodes manifest as time intervals assigned to regimes $\{v_\ell\}_{\ell=1}^k$4 with elevated EPA pollutant vectors (identified by higher $\{v_\ell\}_{\ell=1}^k$5 and marginal means). Episodes are detected via posterior decoding (Viterbi/MAP) or by thresholding $\{v_\ell\}_{\ell=1}^k$6 (Mingione et al., 17 Sep 2025). Time-resolved risk attribution is achieved via Shapley value decompositions of dynamic multivariate Value-at-Risk (MCoVaR) or Expected Shortfall (MCoES) measures (Mingione et al., 17 Sep 2025). Diagnostics recommended include:

• Partial autocorrelation plots of original time series for order selection and lag inclusion.
• State-specific companion matrix eigenvalue analysis for VAR stability.
• LASSO path plots visualizing lag selection in each regime.
• Bootstrap confidence intervals for all inferred parameters.
• State-specific correlation matrices for visualizing shifting co-pollutant dependencies.
• Dwell-time histograms for validation of semi-Markov sojourn structure.

7. Significance and Application Domains

EPA vectors provide low-dimensional, interpretable summaries of pollutant mixtures and regime-dependent exposure profiles, optimizing for both explainable variance and predictability at unmeasured sites or time points. Their use enables improved spatial interpolation, supports causal inference in health studies accounting for exposure misalignment, and offers a structured approach to episode detection and risk decomposition in environmental monitoring (Jandarov et al., 2015, Mingione et al., 17 Sep 2025). The frameworks are extensible to other environmental, epidemiological, or financial multivariate time series where latent mixtures and state-dependent correlation are of primary interest.