Ozone as an environmental driver of influenza

Seasonal influenza causes substantial global health and economic burdens, with about half a million deaths annually. Understanding environmental drivers of influenza dynamics is a priority alongside factors such as antigenic drift, host susceptibility, and social mixing. Prior epidemiology and experimental work has focused mainly on absolute humidity (AH) and temperature (T). Emerging evidence from Hong Kong suggested a negative relationship between ambient ozone (O3) and influenza transmissibility, consistent with laboratory and clinical findings indicating O3’s virucidal potential and ability to prime host immunity. The present study investigates whether ambient O3 is an environmental driver of influenza in temperate US states, using weekly state-level data (2010–2015) and an evidence-triangulation framework across three distinct methods (CCM, PCMCI+, GLM) to test whether weekly changes in O3 lead to changes in community influenza activity within 0–2 weeks.

Population-level studies on ambient O3 and influenza have been limited and mixed. The USEPA Integrated Science Assessment cited reports of positive associations in Hong Kong and Brisbane; however, the Hong Kong report aggregated influenza and pneumonia, and the Brisbane pediatric study insufficiently controlled temporal confounding. More recent Hong Kong time-series work (1998–2013) found ambient O3 negatively associated with the effective reproduction number (Rt) of influenza. Laboratory evidence shows O3 can disrupt virions and inactivate influenza at higher concentrations, and clinical/experimental studies indicate O3 can prime innate and adaptive immunity. Mechanistically, O3 exposure induces epithelial and myeloid cytokines, notably IL-33, which modulates antiviral immunity (e.g., enhancing CD8+ T cell and NK responses) and tissue repair, potentially offering cross-protection against influenza. IL-33 has also demonstrated adjuvant-like effects in mucosal influenza vaccination contexts. These findings motivate evaluating O3’s role in influenza dynamics beyond subtropical settings, into temperate US states.

Study design and data: Weekly state-level data from the US CDC were used from Oct 3, 2010 to Sep 27, 2015. Influenza activity (Flu) was proxied as the product of the weekly proportion of influenza-like illness (ILI) consultations (ILINet) and the weekly proportion of specimens testing positive for influenza (WHO/NREVSS): Flu_proxy = (Flu_positive/Flu_total) × (ILI_outpatient/Outpatient_total). Environmental exposures included ambient ozone (daily maximum 8-h average O3 in ppb) from the Tropospheric Ozone Assessment Report (TOAR/IGAC), and meteorology (air temperature and dew point from NOAA/NCEI) used to derive daily absolute humidity (g/m³). A centering approach reduced station-level biases (centering measurements by long-term station means and adding the state-wide average). Daily series were aggregated to weekly to match influenza data. Analytic window and inclusion: Because off-season influenza had long zero stretches with little causal information, analyses were restricted to flu seasons (Oct–May). States with ≥3 consecutive years of data were included; 46 of 50 states qualified (VT, RI, NJ, FL excluded). Analytical framework: Three complementary methods tested lags 0–2 weeks. 1) CCM (Convergent Cross Mapping) within Empirical Dynamic Modeling (EDM) assessed causality under dynamical systems assumptions. Cross-mapping skill was quantified as Δρ_CCM (difference in prediction correlations between maximum and minimum library sizes). To guard against spurious seasonality-driven correlations, 1000 seasonal surrogates generated a null distribution; state-level p-values were combined via Fisher’s method, with a stringent P_meta < 1.0×10^-3 threshold. Effect magnitudes were then estimated by multivariate S-map (standardized coefficients approximating ∂Flu/∂Env). 2) PCMCI+ (graphical causal discovery for time series) used partial correlation tests to learn lagged and contemporaneous dependencies, with α_PC = 0.05 for state-level networks and 1.0×10^-3 for the nationwide concatenated analysis, yielding directed acyclic graphs of causal links (cross-MCI effect signs/strengths). 3) GLM time-series regression: Within each state, quasi-Binomial models with logit link regressed Flu on O3, AH, and T with lags 0–2, adjusting for year (secular trend), month (seasonality), and a “transmission term” (log of 1-week lagged outcome) to account for autocorrelation. Same-week AH and T were included when estimating O3 associations to control confounding. State-specific coefficients were pooled using random-effects meta-analysis (REML), with significance threshold p < 1.0×10^-3. Software: R 4.1.1 (rEDM 1.9.2, mgcv 1.8-37, usmap 0.6.1) and Python 3.8 (Tigramite 4.2).

• Across CCM, PCMCI+, and GLM, ambient O3 consistently showed a negative effect on influenza activity, most prominently at 1-week lag. - CCM: Nationwide meta-significance indicated O3 as a causal driver of influenza activity at lag 1 (P_meta < 1.0×10^-3). CCM skill for O3 exceeded that for AH (p < 2.8×10^-5) and T (p < 4.3×10^-5). Multivariate S-map median effect size for O3 at lag 1 was −0.106 (one SD increase in O3 associated with 0.106 SD decrease in logit-transformed influenza activity the following week), with most states showing negative effects. - PCMCI+: In the nationwide graph (α_PC = 1.0×10^-3), O3 had direct negative effects on influenza at lag 0 and lag 1; T influenced influenza indirectly via O3 at lag 0. State-level graphs (α_PC = 0.05) showed 18/46 states with direct O3→Flu links (17 negative, 1 positive in North Dakota); 15 states showed negative T→Flu links, 10 direct and 5 indirect via O3; AH effects were mixed, largely negative, with both direct and indirect pathways (including indirect positive effects via O3). - GLM meta-analysis: One SD increase in O3 was associated with a pooled coefficient of −0.102 in logit-transformed influenza activity at lag 1 (Table 1; p = 5.9×10^-10; reported CI −0.186 to −0.018). AH showed a negative association at lag 1 (e.g., GLM effect −0.310; p = 6.7×10^-10). Temperature showed a negative association at lag 2 in GLM (−0.158; p = 2.0×10^-10). - Table 1 (nationwide pooled results) highlights O3 effects consistently negative and statistically significant at lags 0–2 across all three methods (all p < 1.0×10^-3), with particularly strong evidence at lag 1.

Triangulating three distinct methodologies strengthens causal inference that ambient O3 inhibits community-level influenza activity on short time scales (same week to one week later). While some prior population studies suggested positive O3 associations, methodological limitations and outcome definitions differed, and more recent Hong Kong evidence aligns with a negative O3–influenza relationship. Mechanistically, at typical US ambient levels (average daily max 8-h O3 < 40 ppb), direct virucidal inactivation is less likely than immunological modulation. O3 exposure can prime pulmonary immunity and induce cytokines, notably IL-33, which modulates antiviral responses (enhancing CD8+ T cell and NK activity), balances inflammation and repair, and has demonstrated mucosal vaccine adjuvant-like effects—plausible pathways for reduced influenza transmission/activity. The convergent evidence across CCM (dynamical systems causality), PCMCI+ (graphical causal dependencies), and GLM (adjusted statistical associations) indicates robustness despite differing assumptions and potential biases, offering a coherent picture where O3 contributes to lowering influenza activity, and where T and AH have more context-dependent effects (AH often negative at short lags; T partly mediated through O3).

This study integrates dynamical systems causality testing (CCM), graphical causal discovery (PCMCI+), and regression/meta-analysis (GLM) on US state-level weekly data (2010–2015) to identify ambient ozone as a consistent, negative driver of influenza activity, especially at a one-week lag. Findings suggest that, beyond known roles of absolute humidity and temperature, O3 contributes to modulating influenza dynamics in temperate settings. Future research should: (1) elucidate biological mechanisms (e.g., IL-33–mediated immune priming) with laboratory and molecular studies; (2) assess heterogeneity by influenza subtype/lineage with richer virological data; (3) analyze finer spatial and temporal scales to capture local demography, mobility, tourism, and interventions; and (4) further develop causal inference frameworks that integrate multiple methods and data sources.

• Surveillance-based outcome data may involve measurement error and variability in testing and healthcare-seeking across states and years; the composite Flu proxy (lab positivity × ILI proportion) mitigates but does not eliminate this issue. - Limited subtype/lineage data necessitated aggregation across influenza A/B, potentially masking subtype-specific environmental sensitivities. - Analyses were at weekly, state-level resolution over five flu seasons; findings may not generalize to finer spatial/temporal scales. Local demographics, connectivity, tourism (e.g., Hawaii), and interventions may interact with environmental drivers. - Method-specific assumptions (e.g., CCM’s low-dimensional determinism; PCMCI+ conditional independence and causal sufficiency; GLM linearity on the logit scale) could influence results; unmeasured confounding cannot be completely excluded.