Analytics to Determine Causal Drivers of Air Quality in Monterrey

Analytics to Determine Causal Drivers of Air Quality at the Obispado Station (Monterrey) PCA and Maximum-Likelihood Factor Analysis on SIMA Hourly Time Series Alejandro José Murcia Alfaro Tecnológico de Monterrey (ITESM) Data Science and Mathematics Engineering September 2022 Abstract Assessing urban air quality is a multicausal problem: anthropogenic emissions (mobility, industry, services) interact with meteorology (wind, temperature, solar radiation, humidity, pressure) and local urban form. This report summarizes an analysis of the Obispado monitoring station (Monterrey, Nuevo León) using hourly time series of criteria pollutants (PM 10 , PM 2.5 , O 3 , SO 2 , CO, NO, NO 2 , NO x ) and meteorological variables (TOUT, RH, SR, PRS, WSR, WDR) from the SIMA network. After restructuring, validity-flag filtering, outlier treatment, and multivariate imputation via Predictive Mean Matching (PMM), we apply Principal Component Analysis (PCA) and Maximum-Likelihood Factor Analysis (FA) with varimax rotation. Both methods converge to three interpretable latent structures: (i) NO x coupled with solar radiation (traffic combustion plus photochemical modulation), (ii) SO 2 coupled with wind and thermodynamic state (industrial emissions under dispersion regimes), and (iii) particulate matter with CO and ozone-related behavior (mixed sources and secondary formation/accumulation). A 2019–2020 contrast supports behavioral interpretation while highlighting meteorology as a primary modulator. Keywords: air quality; SIMA; Monterrey; PCA; factor analysis; PM; NO x ; SO 2 ; meteorology; imputation. 1 Context and Objective Air quality indices enable public-risk communication, but operational and scientific understanding requires separating source intensity from dispersion and chemistry. Monterrey is a relevant case given high industrial density, persistent mobility demand, and complex topography and wind regimes. Figure 1 provides the institutional and operational framing: the SIMA network coverage and the indicator scheme used for reporting. This context clarifies why station-level interpretation is necessary: even within the same city, local circulation and nearby activity corridors can generate markedly different pollutant signatures. Objective. Extract compact, interpretable latent drivers from correlated pollutant–meteorology time series at Obispado. Components/factors are interpreted as candidate source/process signatures (not intervention-based causal proofs). 1 (a) SIMA monitoring network context. (b) SIMA indicator scheme used for monitoring/communication. Figure 1: SIMA context supporting station-level analysis and interpretation. 2 Data, Variables, and Preprocessing 2.1 Variables We analyze: •Pollutants: PM10, PM2.5, O3, SO2, CO, NO, NO2, NOx. • Meteorology: temperature (TOUT), relative humidity (RH), solar radiation (SR), pressure (PRS), wind speed (WSR), wind direction (WDR). • Validity flags: operational markers for invalid readings (calibration, shutdowns, out-of-range, etc.). 2.2 Integration pipeline Raw spreadsheets are reshaped into timestamp-indexed series for Obispado (measurement + flag). Invalid flagged readings are removed. Remaining missingness is handled via multivariate imputation, and pollutants and meteorology are merged into one aligned hourly dataset. 2.3 Why multivariate methods are appropriate The system is inherently multivariate: pollutants co-vary due to shared sources and coupled chemistry, while meteorology controls dispersion and transport. This dependence is visible in the correlation structure (Figure 2), where pollutant groups and meteorological blocks exhibit coherent positive/negative relationships. This motivates PCA/FA as tools to summarize correlated patterns into a small number of interpretable dimensions. 2 Figure 2: Correlation heatmap showing structured dependence among pollutants and meteorological variables. This motivates dimensionality reduction and latent-driver extraction. 3 Missingness, Outliers, and Imputation Air-quality monitoring data commonly present structured missingness due to calibration, maintenance, or instrument downtime. Missing values are not necessarily random in time, which can bias covariance estimation if ignored. The missingness summary in the original analysis (e.g., Figure ??) supports treating missingness explicitly rather than dropping large blocks of data. Outliers are handled conservatively by marking them as missing prior to model fitting, preventing undue leverage on correlation and factor extraction. Missing values are then imputed using Predictive Mean Matching (PMM), a multivariate method designed to preserve empirical distributions and cross-variable relationships—particularly important for episodic pollutants such as particulate matter. 3 4 Methods: PCA and Maximum-Likelihood Factor Analysis 4.1 Dimension choice A scree plot supports retaining three dimensions (Figure 3). This choice balances information retention with interpretability and is consistent with a domain-plausible separation between (i) mobility combustion, (ii) industrial emissions under dispersion regimes, and (iii) mixed/secondary processes. Figure 3: Scree plot supporting a three-component / three-factor solution. The elbow indicates diminishing returns beyond three dimensions. 4.2 Interpretability: PCA map and FA loadings The PCA variable map (Figure 4) visually confirms clustering among variables: NO x -family variables align together, meteorological variables form dispersion-related structure, and PM/CO/O 3 occupy a mixed regime. Factor Analysis (Figure 5) provides a complementary view focused on shared covariance; varimax rotation yields clearer factor separation and supports stable interpretation across methods. 4 Figure 4: PCA variable map / biplot. Variable groupings anticipate the three-driver interpretation and show how pollutants relate to meteorological modulation. 5 Figure 5: Factor Analysis loadings (ML estimation with varimax rotation). The factor structure mirrors PCA groupings, strengthening robustness. 5 Results: Three Interpretable Drivers Both PCA and FA converge to three dominant latent structures consistent with plausible source/process regimes (Table 1). We interpret them as candidate drivers because they summarize repeated co-variation across time, while remaining consistent with known emissions and atmospheric mechanisms. Driver High-loading variables Interpretation D1 NO, NO2, NOx, SR Traffic combustion; photochemical modulation via solar radiation D2 SO2, WSR/WDR, TOUT, RH, PRS Industrial emissions + dispersion regime (wind/thermodynamics) D3 PM10, PM2.5, CO, O3 Mixed sources; secondary formation/accumulation Table 1: Dominant latent structures recovered by PCA/FA. 6 5.1 Driver 1: Mobility combustion and photochemistry The NO x family behaves as a combustion signature shaped by human activity patterns (commuting/workdays). The association with solar radiation suggests photochemical modulation (e.g., NO x partitioning and ozone-regime sensitivity), which is consistent with daytime chemistry and diurnal structure. 5.2 Driver 2: Industrial SO2under meteorological control SO 2 is commonly associated with industrial combustion and can be strongly shaped by transport and stability. Its association with wind and thermodynamic variables indicates that concentrations reflect both emissions intensity and dispersion conditions (ventilation vs. stagnation). 5.3 Driver 3: PM/CO/O3mixed and secondary processes PM reflects direct sources (resuspension, combustion) and accumulation under low dispersion; CO adds combustion context; O 3 reflects secondary formation. Their grouping indicates a mixed regime influenced by both local activity and atmospheric processing. 6 2019–2020 Behavioral Contrast The 2019–2020 contrast provides an observational plausibility check: if mobility intensity changes, NO x -dominated structure should respond. Figure 6 illustrates a change in NO x behavior between 2019 and 2020, consistent with shifts in human activity, while still showing that meteorology remains a key modulator. To separate industrial behavior from meteorological transport, we include SO 2 wind summaries (Figure 7). These highlight that dispersion regime (wind direction/speed) is integral to interpreting concentration differences across years. Finally, the mobility-associated component comparison (Figure 8) links the latent driver directly to year-to-year behavior, reinforcing that the extracted factor/component is not an abstract mathematical artifact but a stable, interpretable signature. 7 Figure 6: NO x comparison (2019 vs. 2020). Changes in amplitude/smoothness support the interpretation of Driver 1 as mobility-related, while meteorology remains a confounder/modulator. (a) SO2wind summary (2019). (b) SO2wind summary (2020). Figure 7: SO 2 patterns under different dispersion regimes (2019 vs. 2020). Wind context is necessary to interpret industrial-related concentration differences. 8 (a) Mobility-associated component (2019). (b) Mobility-associated component (2020). Figure 8: Driver 1 component behavior (2019 vs. 2020). The latent index summarizes NO x –SR structure and makes year-to-year comparison more interpretable. 7 Limitations and Practical Implications 7.1 Limitations • No strict causal identification: PCA/FA reveal correlated structures and candidate drivers, not interventions. • Meteorology as modulator/confounder: year-to-year differences require dispersion context (wind/thermodynamics). •Station specificity: generalization to all Monterrey requires multi-station validation. 7.2 Implications • Monitoring narratives: separate emissions intensity from dispersion regime for clearer explanations to stakeholders. • Dashboards: implement three driver indices (D1–D3) as compact signals for operations and communication. • Targeting: D1 supports mobility policy/traffic management; D2 supports industrial oversight with wind-aware alerts; D3 supports mixed-source and secondary-process mitigation. 8 Conclusions Using SIMA Obispado hourly series and a multivariate workflow, PCA and Maximum-Likelihood Factor Analysis converge to three stable, interpretable latent structures aligned with mobility combustion (NO x + SR), industrial/dispersion dynamics (SO 2 + wind/thermodynamics), and mixed/secondary processes (PM + CO + O 3 ). The included figures support each modeling decision and interpretation step, producing a compact, committee-friendly technical summary. References 1. SIMA Nuevo León. Sistema Integral de Monitoreo Ambiental (SIMA): open data and station documentation used in the project. 9