Dietary microplastic exposure and lip–oral cavity cancer: a global ecological analysis with pre-specified spatial econometric sensitivity Authors: Shankargouda Patil, Shilpa Bhandi, Frank W. Licari Published: 2026 ABSTRACT To assess whether estimated dietary microplastic (MP) intake is linked to age-standardized lip and oral cavity cancer rates at the country level, independent of smoking, alcohol, socioeconomic status, and urbanization, we conducted a cross-sectional ecological study across 106 countries (2020 anchor year). While testing for effect modification by smoking, dietary MP intake showed a non-linear association with lip–oral cavity cancer incidence in unadjusted spline models. Smoking prevalence and urbanization were also positive predictors, although the evidence was inconsistent. Placebo/permutation testing, spatial residual diagnostics, and dietary-structure sensitivity analysis using fish/seafood supply as a proxy supported the stability and specificity of the association. We additionally performed a pre-specified spatial econometric sensitivity analysis: residual Moran’s I on the refitted OLS model (n = 79) was 0.099 (p_sim = 0.082); Anselin’s robust LM-lag test was significant (p = 0.003), identifying a Spatial Lag Model (SLM) as the appropriate primary spatial specification (ρ = 0.35; post-fit Moran’s I = − 0.006, p_sim = 0.466 autocorrelation absorbed). Under SLM, the MP p90-vs-p10 exposure contrast was − 0.41 (p = 0.85); the pre-specified decision rule thereby classifies this result as null/hypothesis-generating (Path C). These findings generate hypotheses and do not establish causality, but encourage further individual-level exposure assessments and mechanistic studies to better understand the potential role of microplastics in oral carcinogenesis.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-58951-8. Introduction Primary microplastics (MPs) are intentionally produced at small sizes for commercial and industrial uses, whereas secondary MPs arise from the weathering and fragmentation of larger plastic products during environmental degradation. MPs are commonly defined as plastic particles < 5 mm in size and are broadly classified as primary or secondary microplastics1,2. MPs are now ubiquitous across environmental compartments, including oceans and freshwater systems, soils, the atmosphere, and biota, creating multiple pathways for human exposure. Although the global literature on MP contamination is expanding, evidence remains uneven across regions, with comparatively fewer studies from India despite increasing concern about MP pollution in local environments3,4. Because MPs degrade slowly and can be transported over long distances by wind and water currents, they can exert persistent effects on ecosystems and biota5. They are detected across food webs and can enter organisms via ingestion and inhalation6. In humans, MPs have been reported in multiple matrices and tissues, including placenta, lung, liver, sputum, breast milk, faeces, urine, and blood7 prompting scrutiny of potential health impacts. Reported and hypothesized consequences include inflammation and oxidative stress, immune dysregulation, carcinogenesis-related pathways, and perturbations of reproductive and neurocognitive processes7,8. A study emphasized that plastic particles are widely distributed in air, soil, water, plants, and animals, and are also present in the human diet. Reviews have summarized the major sources and routes of human exposure to MPs, including contaminated food and water, inhalation, and dermal contact9,10. Cancers of the oral cavity, lip, and pharynx account for roughly 300,000 new diagnoses annually and more than 145,000 deaths worldwide. Oral squamous cell carcinoma (OSCC) constitutes > 90% of oral cancers11 and frequently metastasizes to regional lymph nodes; the risk of cervical nodal spread increases with advancing T stage, depth of invasion, and tumour thickness12. Other malignancies affecting the oral cavity include melanomas, lymphomas, and tumours arising from the minor salivary glands. Micro- and Nano-plastics may act as carriers of mutagenic or carcinogenic chemicals, with the potential to perturb genomic integrity and contribute to malignant transformation13,14. Their accumulation may be particularly concerning in biologically critical tissues, including the bone marrow1,2. However, Deng et al.1 noted that epidemiologic evidence directly linking microplastic exposure to specific malignancies remains limited, despite experimental findings that micro(nano)plastics can accumulate in organisms and induce cellular injury consistent with pro-carcinogenic processes1,15. Established risk factors for oral cancer include betel quid chewing, alcohol consumption, use of smokeless and smoked tobacco, and chronic mucosal inflammation11. Against this background, it remains unclear whether higher dietary microplastic intake shows a measurable dose–response relationship with age-standardized lip and oral cavity cancer (LOCC) incidence, independent of alcohol, smoking, and broader structural determinants. To address this gap, we evaluated the association between estimated dietary microplastic exposure and LOCC incidence in a global ecological analysis, accounting for major behavioral and socioeconomic covariates. Materials and methods We investigated the association between national dietary microplastic (MP) intake and lip and oral cavity cancer (LOCC) incidence using a cross-sectional ecological design. The 2020 calendar year served as the analytic anchor, and all variables were harmonised to this reference period. The primary exposure was estimated dietary MP intake at the country level, and the primary outcome was the age-standardized LOCC incidence rate (both sexes). We considered adult smoking prevalence, per-capita alcohol consumption, GDP per capita, and degree of urbanisation as key confounders. For sensitivity analyses, we additionally evaluated a dietary-structure proxy based on per-capita fish and seafood energy supply. After harmonisation and complete-case restriction for covariates, 81 of 106 countries contributed to the fully adjusted model. We modeled the natural logarithm of LOCC incidence as a function of dietary MP exposure using restricted cubic splines (RCS) to allow for non-linear exposure–response relationships commonly observed in environmental epidemiology. The spline specification used four knots and omitted a spline intercept. Models were adjusted for smoking prevalence, per-capita alcohol consumption, GDP per capita, and urbanisation. To assess potential effect modification, we included an interaction term between dietary MP exposure and smoking prevalence. We evaluated model adequacy using observed-versus-fitted plots, residual distribution diagnostics (Omnibus and Jarque–Bera tests), and influence diagnostics (leverage and Cook’s distance). Multicollinearity was assessed using condition numbers and by examining correlations between the spline basis terms and structural covariates. To further interrogate specificity, we conducted a placebo (permutation) test by randomly permuting dietary MP exposure 2000 times, refitting the model, and deriving an empirical p-value from the resulting null distribution of the model fit statistic. Spatial dependence was assessed by computing Moran’s I on model residuals using country centroids and a k-nearest-neighbours spatial weights matrix (k = 5). We performed three sensitivity analyses to evaluate robustness. First, we introduced per-capita fish and seafood energy supply as a proxy for dietary structure, while explicitly considering collinearity with GDP and urbanisation. Second, we assessed temporal sensitivity by expanding the harmonisation window from the primary ± 2-year frame to a ± 5-year window around the 2020 anchor. Third, we visualised spatial patterns and concordance between exposure and outcome using global choropleth maps and bivariate quantile maps. For contextual interpretation, we overlaid marine MP density. Spatial econometric sensitivity analysis. We implemented a pre-specified spatial econometric sensitivity analysis16,17; the model-selection procedure and decision rule were fixed before the spatial models were fitted, and the full analysis code is openly available (see Code availability). We constructed three spatial weights matrices on country centroids: primary k-nearest-neighbours (k = 5, row-standardized, W1); queen-contiguity (W2); and inverse-distance capped at 2,000 km (W3). Following a refit of the adjusted OLS model, we computed Anselin’s classical and robust Lagrange multiplier tests (LM-lag, LM-error, robust LM-lag, robust LM-error)18. A Spatial Lag Model (SLM) or Spatial Error Model (SEM) was then selected as the primary spatial specification per Anselin’s decision procedure, fitted via maximum likelihood using pysal/spreg19. A Spatial Durbin Model (SDM) was fitted as a generalization robustness check. Post-fit Moran’s I was recomputed on spatial-model residuals to confirm absorption of autocorrelation. Pre-specified three-path decision rule. The magnitude, direction, and significance of the microplastic exposure effect at the p90-vs-p10 exposure contrast under the chosen spatial model dictated the interpretation: Path A (same sign as OLS, p ≤ 0.10) would classify the finding as primary; Path B (same sign, 0.10 < p ≤ 0.20) as hypothesis-generating after spatial adjustment; Path C (sign flip or p > 0.20) as a null/sensitive-to-confounding result with honest null reporting. The decision rule was fixed before any spatial model was fitted and is implemented in the openly available analysis code. Biological confounder proxies (HPV, smokeless tobacco). We programmatically acquired two additional country-level covariates: (i) HPV 16/18 prevalence among women with normal cytology from ICO/IARC HPV Information Centre country reports20 (n = 106 extracted; n = 45 country-specific estimates vs. n = 61 regional/global pooled fallback substitutions, flagged transparently), and (ii) adult current smokeless tobacco prevalence from the WHO Global Health Observatory Adult_curr_smokeless indicator21 (n = 112 country-years, latest-available year ≤ 2022; proxy for areca-nut/betel-quid chewing exposure; joined to n = 61 of our panel). Complete data provenance (source organization, URL, access date, license, citation) for every variable is summarized in Table 1. Reproducibility. Random seeds were fixed (seed = 20260423), and the Python dependencies are version-pinned. The analysis code and the harmonised country-level dataset are openly available (see Code Availability and Data Availability). Results We analyzed 106 countries with harmonised estimates of age-standardized lip and oral cavity cancer (LOCC) incidence and dietary microplastic (MP) intake. The fully adjusted primary analysis was restricted to countries with complete covariate data (n = 81). Across countries, LOCC incidence tracked positively with smoking prevalence and the degree of urbanisation, consistent with established epidemiologic patterns. To probe robustness to broader dietary structure, we additionally incorporated average per-capita fish and seafood energy supply (2018–2022) as a proxy dietary covariate in sensitivity analyses. In the primary adjusted model, dietary MP intake was associated with LOCC incidence in a non-linear pattern modeled using restricted cubic splines (Figs. 1, 2 and 3). The adjusted model explained a modest proportion of between-country variation in LOCC incidence (adjusted R² = 0.301). LOCC incidence was relatively stable at lower MP intake levels. It increased across the mid-to-higher exposure range, with wider uncertainty at the extreme upper end, where fewer countries contributed data (Fig. 3). Smoking prevalence remained a strong positive predictor of LOCC incidence, while urbanisation showed a modest positive association that may reflect correlated behavioral, social, and health system factors. After adjustment for these covariates, GDP per capita showed no clear independent association, and per-capita alcohol consumption contributed minimally in the primary model. Fig. 1Country-level association between dietary microplastic exposure and lip–oral cavity cancer incidence. (a) Scatterplot of estimated dietary microplastic exposure (µg/kg/day) versus age-standardized incidence rate (ASR; per 100,000) of lip and oral cavity cancer (LOCC) across countries. The solid line indicates the fitted linear regression, and the shaded band denotes the 95% confidence interval (R² = [0.020], P = [0.145]). The regression coefficient (β) corresponds to the change in LOCC ASR (per 100,000) per unit increase in microplastic exposure. (b) Log–log relationship between microplastic exposure and LOCC incidence, with both variables natural log-transformed. The solid line indicates the fitted linear regression on transformed data, and the shaded band denotes the 95% confidence interval (R² = [0.061], P = [0.011]). The slope represents the elasticity, i.e., the percent change in LOCC incidence associated with a 1% increase in microplastic exposure. Country-level association between dietary microplastic exposure and lip–oral cavity cancer incidence. (a) Scatterplot of estimated dietary microplastic exposure (µg/kg/day) versus age-standardized incidence rate (ASR; per 100,000) of lip and oral cavity cancer (LOCC) across countries. The solid line indicates the fitted linear regression, and the shaded band denotes the 95% confidence interval (R² = [0.020], P = [0.145]). The regression coefficient (β) corresponds to the change in LOCC ASR (per 100,000) per unit increase in microplastic exposure. (b) Log–log relationship between microplastic exposure and LOCC incidence, with both variables natural log-transformed. The solid line indicates the fitted linear regression on transformed data, and the shaded band denotes the 95% confidence interval (R² = [0.061], P = [0.011]). The slope represents the elasticity, i.e., the percent change in LOCC incidence associated with a 1% increase in microplastic exposure. Fig. 2Diagnostic plots for the adjusted regression model. (a) Quantile–quantile (Q–Q) plot of standardized residuals versus theoretical normal quantiles; alignment with the dashed 45° line indicates approximate normality. (b) Residuals versus fitted values plot to assess homoscedasticity; an approximately random scatter around zero is consistent with constant variance. (c) Cook’s distance for identification of influential observations; the dashed red line denotes the conventional threshold (4/n = 0.049). All diagnostics are shown for the multivariable regression model, including restricted cubic splines for dietary microplastic exposure and adjustment for GDP per capita and percent urbanisation. Diagnostic plots for the adjusted regression model. (a) Quantile–quantile (Q–Q) plot of standardized residuals versus theoretical normal quantiles; alignment with the dashed 45° line indicates approximate normality. (b) Residuals versus fitted values plot to assess homoscedasticity; an approximately random scatter around zero is consistent with constant variance. (c) Cook’s distance for identification of influential observations; the dashed red line denotes the conventional threshold (4/n = 0.049). All diagnostics are shown for the multivariable regression model, including restricted cubic splines for dietary microplastic exposure and adjustment for GDP per capita and percent urbanisation. Fig. 3Non-linear association between dietary microplastic exposure and LOCC incidence and exploratory assessment of interaction with smoking. (a) Restricted cubic spline (RCS) model of the association between estimated dietary microplastic exposure and age-standardized incidence rate (ASR; per 100,000) of lip and oral cavity cancer (LOCC). The solid line shows the model-predicted ASR from a multivariable regression including RCS terms for microplastic exposure (4 knots; df = 4), and the shaded band indicates the 95% confidence interval. Points show country-level observations. Model adjusted for GDP per capita and percent urbanisation; n = 81. (b) Predicted LOCC ASR across microplastic exposure at two levels of smoking prevalence (10th percentile: 8.0% and 90th percentile: 34.5%). The interaction between microplastic exposure and smoking prevalence was not statistically supported in this model (P_interaction = 0.360); curves are shown to illustrate the fitted model predictions. All models are based on countries with complete data (n = 81) and include the same covariate adjustments as in panel A. Non-linear association between dietary microplastic exposure and LOCC incidence and exploratory assessment of interaction with smoking. (a) Restricted cubic spline (RCS) model of the association between estimated dietary microplastic exposure and age-standardized incidence rate (ASR; per 100,000) of lip and oral cavity cancer (LOCC). The solid line shows the model-predicted ASR from a multivariable regression including RCS terms for microplastic exposure (4 knots; df = 4), and the shaded band indicates the 95% confidence interval. Points show country-level observations. Model adjusted for GDP per capita and percent urbanisation; n = 81. (b) Predicted LOCC ASR across microplastic exposure at two levels of smoking prevalence (10th percentile: 8.0% and 90th percentile: 34.5%). The interaction between microplastic exposure and smoking prevalence was not statistically supported in this model (P_interaction = 0.360); curves are shown to illustrate the fitted model predictions. All models are based on countries with complete data (n = 81) and include the same covariate adjustments as in panel A. We evaluated whether the association between dietary MP exposure and LOCC differed by smoking prevalence (Figs. 4 and 5). Evidence for interaction was not consistent across model specifications; therefore, the MP-by-smoking interaction was formally non-significant (P_interaction = 0.360), and no synergistic effect should be inferred. Fig. 4Global distributions of dietary microplastic exposure and LOCC incidence. (a) Choropleth map of estimated dietary microplastic exposure across n = 101 countries. (b) Choropleth map of age-standardized incidence rate (ASR; per 100,000) of lip and oral cavity cancer (LOCC) across the same n = 101 countries. Both maps use percentile-based colour scaling (5th–95th percentiles) to reduce the influence of extreme values and facilitate visual comparison of geographic patterns. Global distributions of dietary microplastic exposure and LOCC incidence. (a) Choropleth map of estimated dietary microplastic exposure across n = 101 countries. (b) Choropleth map of age-standardized incidence rate (ASR; per 100,000) of lip and oral cavity cancer (LOCC) across the same n = 101 countries. Both maps use percentile-based colour scaling (5th–95th percentiles) to reduce the influence of extreme values and facilitate visual comparison of geographic patterns. Fig. 5Distribution of dietary microplastic exposure, LOCC incidence, and their bivariate relationship across countries. (a) Top 20 countries ranked by estimated dietary microplastic exposure (g/person/week). Austria shows the highest estimated exposure (0.0024 g/person/week), followed by Denmark, Sweden, and Switzerland; high-exposure countries are predominantly in Europe. (b) Distribution of dietary microplastic exposure across all countries (mean = 0.0016 g/person/week; median = 0.0016 g/person/week). (c) Bivariate association between dietary microplastic exposure and age-standardized lip and oral cavity cancer incidence (ASR; per 100,000) across n = 106 countries (Pearson r = 0.142; P = 0.145; R² = 0.020). (d) Distribution of LOCC incidence rates across countries (mean ASR = 2.9 per 100,000), showing a right-skewed pattern with most countries exhibiting low incidence and a subset exhibiting higher burden. (e) Comparative ranking of countries by microplastic exposure and LOCC incidence. Several countries (e.g., Austria, Denmark, Sweden) exhibit concordantly high exposure and incidence, whereas other countries show discordant patterns, consistent with the influence of additional factors beyond exposure alone. Distribution of dietary microplastic exposure, LOCC incidence, and their bivariate relationship across countries. (a) Top 20 countries ranked by estimated dietary microplastic exposure (g/person/week). Austria shows the highest estimated exposure (0.0024 g/person/week), followed by Denmark, Sweden, and Switzerland; high-exposure countries are predominantly in Europe. (b) Distribution of dietary microplastic exposure across all countries (mean = 0.0016 g/person/week; median = 0.0016 g/person/week). (c) Bivariate association between dietary microplastic exposure and age-standardized lip and oral cavity cancer incidence (ASR; per 100,000) across n = 106 countries (Pearson r = 0.142; P = 0.145; R² = 0.020). (d) Distribution of LOCC incidence rates across countries (mean ASR = 2.9 per 100,000), showing a right-skewed pattern with most countries exhibiting low incidence and a subset exhibiting higher burden. (e) Comparative ranking of countries by microplastic exposure and LOCC incidence. Several countries (e.g., Austria, Denmark, Sweden) exhibit concordantly high exposure and incidence, whereas other countries show discordant patterns, consistent with the influence of additional factors beyond exposure alone. Standard diagnostic checks did not indicate major violations of model assumptions, and influence diagnostics did not suggest that results were driven by a small number of highly influential countries (Fig. 6). In a placebo (permutation) analysis that randomly reassigned MP exposure values 2,000 times, the observed association was unlikely to arise under a null distribution of no relationship between exposure and outcome (empirical p = 0.001). We also assessed residual spatial dependence using Moran’s I computed on model residuals with a k-nearest-neighbours spatial weights matrix (k = 5) (Supplementary Fig. 1). Residual spatial autocorrelation was explicitly evaluated, and where present, it indicates that some geographically structured variation remains unexplained; Accordingly, we interpret the country-level association as hypothesis-generating and potentially sensitive to unmeasured, spatially patterned confounding. Importantly, the permutation test assesses whether the observed association could arise under random label reassignment of the exposure; it does NOT adjust for spatial confounding. The spatial econometric sensitivity analysis (Supplementary Figs. 1, 3) addresses that latter, distinct concern. Fig. 6Robustness of the microplastic–LOCC association across model specifications. (a) Forest plot of regression coefficients (points) and 95% confidence intervals (horizontal lines) for the association between log-transformed dietary microplastic exposure and log-transformed LOCC incidence across multiple model specifications. The vertical dashed line denotes the null (β = 0). Models are ordered from an unadjusted/bivariate specification to progressively adjusted models, including available covariates; sample size (n) for each specification is shown. Asterisks indicate nominal significance (*P < 0.05, **P < 0.01, ***P < 0.001). (b) Model fit across specifications, shown as R² (variance explained). Bars indicate R² for each model, with corresponding n displayed within bars. Robustness of the microplastic–LOCC association across model specifications. (a) Forest plot of regression coefficients (points) and 95% confidence intervals (horizontal lines) for the association between log-transformed dietary microplastic exposure and log-transformed LOCC incidence across multiple model specifications. The vertical dashed line denotes the null (β = 0). Models are ordered from an unadjusted/bivariate specification to progressively adjusted models, including available covariates; sample size (n) for each specification is shown. Asterisks indicate nominal significance (*P < 0.05, **P < 0.01, ***P < 0.001). (b) Model fit across specifications, shown as R² (variance explained). Bars indicate R² for each model, with corresponding n displayed within bars. Including fish and seafood energy supply as a dietary-structure proxy did not materially change the estimated association between dietary MP exposure and LOCC incidence, arguing against the MP signal being solely a surrogate for fish consumption patterns (Fig. 7). Results were also robust to the temporal harmonisation choice: expanding the inclusion window from 2020 ± 2 years to 2020 ± 5 years yielded qualitatively similar inferences (Fig. 7). Geospatial visualisation provided additional context: countries with lower estimated MP exposure generally exhibited lower LOCC incidence, whereas higher-exposure countries more frequently coincided with higher LOCC incidence (Supplementary Fig. 2). A global marine MP density overlay highlighted major shipping corridors and ocean gyres, providing contextual support for plausible environmental-to-dietary pathways (Supplementary Fig. 2). Fig. 7Spearman correlation matrix showing relationships between microplastic exposure, oral cancer incidence, and behavioral risk factors across n = 97 countries. Correlation coefficients: MP Exposure vs. Oral Cancer (ρ = 0.28), MP Exposure vs. Smoking (ρ = 0.36), MP Exposure vs. Alcohol (ρ = 0.50), Oral Cancer vs. Smoking (ρ = 0.51), Oral Cancer vs. Alcohol (ρ = 0.46), Smoking vs. Alcohol (ρ = 0.25). All correlations were statistically significant (p < 0.001). The matrix reveals moderate to strong associations between environmental microplastic exposure and established behavioral risk factors for oral cavity cancer. Spearman correlation matrix showing relationships between microplastic exposure, oral cancer incidence, and behavioral risk factors across n = 97 countries. Correlation coefficients: MP Exposure vs. Oral Cancer (ρ = 0.28), MP Exposure vs. Smoking (ρ = 0.36), MP Exposure vs. Alcohol (ρ = 0.50), Oral Cancer vs. Smoking (ρ = 0.51), Oral Cancer vs. Alcohol (ρ = 0.46), Smoking vs. Alcohol (ρ = 0.25). All correlations were statistically significant (p < 0.001). The matrix reveals moderate to strong associations between environmental microplastic exposure and established behavioral risk factors for oral cavity cancer. Taken together, in unadjusted spline models (before formal spatial econometric adjustment), higher estimated dietary MP intake appeared associated with higher LOCC incidence at the ecological (country) level after adjustment for smoking prevalence, per-capita alcohol consumption, GDP per capita, and urbanisation. The exposure–response relationship was non-linear, with the strongest increases occurring across the mid-to-upper exposure range (Figs. 3 and 8). The association was supported by falsification testing (empirical p = 0.001) and remained qualitatively stable across dietary and temporal sensitivity analyses (Figs. 6 and 7), while spatial diagnostics indicate that residual geographic structure should be considered when interpreting effect estimates (Supplementary Fig. 1). As detailed below, this apparent exposure-response relationship was NOT sustained under the pre-specified Spatial Lag Model sensitivity analysis (Supplementary Figs. 1, 3); the decision rule triggered Path C (null/hypothesis-generating reporting) (Fig. 8). Fig. 8Distribution of placebo effect sizes from 2000 permutations of microplastic exposure data. The vertical magenta line shows the observed effect size (β = − 0.530) from the actual data. Dashed blue lines indicate the 95% confidence interval of the observed effect. The proportion of permutation effects exceeding the observed effect magnitude provides the permutation p-value (p = 0.0), testing the null hypothesis that the observed association could occur by chance. Distribution of placebo effect sizes from 2000 permutations of microplastic exposure data. The vertical magenta line shows the observed effect size (β = − 0.530) from the actual data. Dashed blue lines indicate the 95% confidence interval of the observed effect. The proportion of permutation effects exceeding the observed effect magnitude provides the permutation p-value (p = 0.0), testing the null hypothesis that the observed association could occur by chance. Spatial econometric results. The refitted OLS model (complete-case n = 79) explained 32.2% of variance in log-LOCC ASR (adjusted R²). Moran’s I on OLS residuals under W1 (k-NN k = 5) was 0.099 (p_sim = 0.082). Anselin’s classical LM-lag (p = 0.28) and LM-error (p = 0.21) were non-significant, but the robust LM-lag was significant (p = 0.003) while the robust LM-error was not (p = 0.51), indicating a spatial-lag data-generating process. Per the pre-specified rule, a Spatial Lag Model (SLM) was selected as the primary spatial specification. The SLM spatial autoregressive parameter was ρ = 0.346; SLM AIC = 115.9 improved on OLS AIC = 121.0 and was comparable to SDM AIC = 116.3. Post-fit Moran’s I on SLM residuals was − 0.006 (p_sim = 0.466), indicating complete absorption of adjacency-based spatial autocorrelation (Supplementary Figs. 1, 3). Under the SLM, the microplastic exposure p90-vs-p10 contrast was − 0.409 (p = 0.846); the linear (non-spline) adjusted benchmark was β = − 321.3 (p = 0.078). Per the pre-specified three-path decision rule, these results trigger PATH C: the apparent MP–LOCC association observed in OLS spline models is NOT robust to formal spatial adjustment, and is reported as a null/hypothesis-generating finding rather than as supportive evidence of a causal pathway. Smoking prevalence (β = + 0.015 per %, p = 0.015) and per-capita alcohol consumption (β=+0.044 per L, p = 0.008) remained positive predictors under SLM, consistent with established oral-cancer literature. The four contrasts (p10→p50, p50→p90, median ± IQR/2, and p10→p90) and the AIC comparison across OLS, SLM, SEM, and SDM are summarized in Supplementary Fig. 3. Sensitivity analyses. Across alternative spatial weights matrices (W2 queen; W3 inverse-distance, 2000 km cap), the SLM-selected MP effect remained non-significant (p ≥ 0.83). In the HPV-country-specific subgroup (n = 37), adjusting additionally for HPV 16/18 prevalence, and in the smokeless-tobacco subgroup (n = 48), adjusting for current smokeless tobacco prevalence, neither produced a significant MP association (both p ≥ 0.79). The combined HPV + smokeless subgroup (n = 21) similarly showed no robust association. These sensitivity results do not overturn the Path C classification. Discussion Here we report a hypothesis-generating ecological pattern between estimated dietary microplastic intake and LOCC incidence in unadjusted/spline models, but this pattern is NOT robust to formal spatial econometric adjustment (pre-specified Spatial Lag Model, decision rule Path C). Plastics are complex materials composed of polymer backbones combined with numerous chemical additives. In addition, plastic particles can carry environmental contaminants that adsorb to their surfaces, including polycyclic aromatic hydrocarbons, phenolic compounds, and other substances arising from residual monomers and manufacturing by-products. The additive burden varies substantially by polymer type, with reported counts generally highest for PVC and polyurethane materials and lower for polymers such as PET, polyamides, ABS, and polystyrene22. Micro- and Nano-plastics may persist in tissues because of their small size and limited biodegradability, with experimental studies reporting inflammatory responses consistent with chronic exposure23,24. The impact of micro(nano)plastics on intracellular redox balance and the ensuing rise in DNA damage and mutation rates is one of the key methods to raise cancer risk, according to a review of the literature by Deng et al.1 In addition, common plastic additives, including plasticisers and antioxidants, have been reported to exhibit endocrine-disrupting activity, which could influence hormone signaling and proliferation pathways relevant to cancer development23,24. Microplastic contamination has been reported in a range of commonly consumed items, including salt, honey, beer, and drinking water. As these products are used routinely, they represent plausible contributors to chronic human exposure. Current evidence indicates that the main exposure routes include inhalation of microplastic-containing airborne particles, ingestion through contaminated food and beverages, and dermal contact with microplastic-containing products or contaminated surfaces25,26. A growing body of work has focused on biomonitoring across diverse populations and exposure settings, and has raised questions about potential health consequences of environmental MP exposure. However, despite the apparent ubiquity of MPs across environmental matrices, workplaces, and the food supply, substantial gaps remain in quantifying human tissue burdens and in establishing the clinical relevance of chronic accumulation, underscoring the need for well-designed epidemiologic and mechanistic studies27,28. Microplastics have been detected in a wide range of dietary sources, including drinking water and other beverages, fruits and vegetables, salt, honey, sugar, and marine foods such as fish and shellfish29. In aquatic ecosystems, MPs can move through food webs via trophic transfer, from lower trophic levels (e.g., zooplankton) to higher trophic organisms such as fish, crustaceans, and molluscs, creating a pathway for human exposure through seafood consumption. Consistent with this, Van Cauwenberghe and Janssen30 estimated that shellfish intake alone could contribute up to ~ 11,000 ingested plastic particles per year. Airborne microplastics may also contribute to exposure during food preparation and consumption, even when the food itself is not measurably contaminated. Cantarino et al. estimated that this pathway could result in ingestion of approximately 13,731–68,415 particles per year31. In parallel, microplastics in drinking water and other beverages have received substantial attention, and multiple reviews have synthesized evidence of MPs across water sources and beverage matrices32,33. Microplastics can be transported by atmospheric processes over tens of kilometers, and airborne concentrations are influenced by local anthropogenic activity. Because particles can remain suspended in indoor air, inhalation represents an additional exposure route alongside ingestion and dermal contact. Human exposure may therefore occur directly through inhalation or contact with airborne MPs, and through ingestion of contaminated media (e.g., water, salt, soil-derived dust) or indirectly via trophic transfer through seafood and plant-based foods. Experimental studies also indicate that MPs can interact with marine and soil organisms through ingestion or entanglement and can induce adverse effects; these impacts depend on particle type, concentration, and size, and can extend to changes in microbial communities in soil and to effects in animals and plants34. The observed non-linear exposure–response pattern is biologically plausible, as dietary MPs may perturb host–microbiome interactions and activate mucosal immune pathways, in addition to serving as carriers for additives and sorbed environmental pollutants. Given that our pre-specified spatial econometric sensitivity analysis triggered the null-reporting Path C, we explicitly do NOT claim that our country-level findings support a causal role for dietary MP exposure in oral carcinogenesis. Any mechanistic plausibility described above remains speculative in the absence of individual-level exposure data and experimental evidence. Ecological inference is inherently limited, and associations observed at the national level may not translate to individual risk (ecological fallacy). Residual confounding is also plausible, including from factors not captured uniformly across countries, such as HPV prevalence, oral hygiene, betel quid/areca use, screening intensity, and cancer registry completeness. In addition, correlations among structural covariates and the spline basis terms can complicate the interpretation of partial effects. Although we harmonised variables to an anchor year, cross-source differences in reporting years may introduce measurement error. Finally, dietary MP exposure was modeled rather than directly measured, and population-scale biomarker validation is not yet available. Cross-country differences in cancer surveillance and registry completeness may contribute to heterogeneity in reported LOCC incidence. In addition, the fully adjusted analysis was restricted to countries with complete covariate data, which could introduce selection bias if missingness is related to exposure, outcome, or health-system capacity. Ecological fallacy. Inferring individual-level associations from country-level aggregates (the ecological fallacy) is a primary limitation of this analysis. National-level averages of dietary microplastic exposure cannot speak to individual exposure patterns, the mechanism of ingestion, or interaction with individual behavioural risk factors. Our findings generate hypotheses at the population level and cannot establish causation or individual risk. We explicitly caution against translating these country-level associations to clinical or public-health recommendations for individuals. Unmeasured biological confounders. Two of the globally most important risk factors for oral cancer—human papillomavirus (HPV) infection20 and areca-nut (betel-quid) chewing35,36—were not directly available at country level. We incorporated country-level proxies (ICO/IARC HPV 16/18 prevalence; WHO Adult_curr_smokeless prevalence), but coverage and precision are limited. IARC publishes country-specific HPV estimates only when sufficient primary studies exist; for countries without such studies, regional or global pooled values are substituted, and these were flagged in our analysis (n = 61 regional-fallback vs. n = 45 country-specific). Smokeless-tobacco prevalence is an imperfect proxy for areca-nut/betel-quid use. Sensitivity analyses including these proxies did not produce a robust MP association, but residual confounding by these factors or by other unmeasured spatially-patterned drivers of oral-cancer incidence cannot be excluded. Residual spatial confounding. Although our spatial econometric sensitivity analysis absorbed adjacency-based spatial autocorrelation (post-fit Moran’s I = − 0.006, p_sim = 0.466), this addresses only the observed spatial structure captured by k-nearest-neighbour, queen-contiguity, and inverse-distance weights16,17. Unmeasured, spatially-patterned confounders including regional variation in oral-health-care access, diagnostic sensitivity, cancer-registry completeness, and dietary traditions, may still drive apparent country-level patterns. Conclusion In this global ecological analysis of 106 countries, we did NOT find robust evidence of an independent association between dietary microplastic exposure and age-standardized lip and oral cavity cancer incidence after adjustment for established behavioural and socio-economic covariates AND formal spatial structure. The apparent non-linear association observed in OLS spline models (Moran’s I = 0.099 on OLS residuals) was NOT sustained under a pre-specified Spatial Lag Model sensitivity analysis (post-fit Moran’s I = − 0.006; SLM p90–p10 contrast = − 0.409, p = 0.846). The pre-specified three-path decision rule identifies this result as Path C, a null/hypothesis-generating finding rather than evidence of a causal pathway. Smoking prevalence and per-capita alcohol consumption remained positive predictors across specifications. These findings do NOT support a population-level causal role for dietary microplastic exposure in lip–oral cavity carcinogenesis. Individual-level exposure assessment, mechanistic studies, and improved global coverage of HPV and areca-nut/betel-quid data remain necessary before firmer inferences can be drawn. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Fig. 1. Residual diagnostics and spatial dependence for the adjusted spline model and the spatial lag model. a. Choropleth map of residuals from the multivariable restricted cubic spline (OLS) model, with colours indicating residual magnitude and sign (positive: over-prediction; negative: under-prediction). b. Choropleth map of residuals from the spatial lag model (SLM) after spatial adjustment; panels a and b share a common colour scale for direct comparison. c. Moran’s I scatterplot of the OLS residuals against their spatial lag (k-nearest-neighbours, k = 5): Moran’s I = 0.099 (p_sim = 0.082). d. Moran’s I scatterplot of the SLM residuals: Moran’s I = -0.006 (p_sim = 0.466), showing that the adjacency-based spatial autocorrelation present in the OLS residuals is absorbed under the spatial model. Supplementary Material 1. Supplementary Fig. 1. Residual diagnostics and spatial dependence for the adjusted spline model and the spatial lag model. a. Choropleth map of residuals from the multivariable restricted cubic spline (OLS) model, with colours indicating residual magnitude and sign (positive: over-prediction; negative: under-prediction). b. Choropleth map of residuals from the spatial lag model (SLM) after spatial adjustment; panels a and b share a common colour scale for direct comparison. c. Moran’s I scatterplot of the OLS residuals against their spatial lag (k-nearest-neighbours, k = 5): Moran’s I = 0.099 (p_sim = 0.082). d. Moran’s I scatterplot of the SLM residuals: Moran’s I = -0.006 (p_sim = 0.466), showing that the adjacency-based spatial autocorrelation present in the OLS residuals is absorbed under the spatial model. Supplementary Material 2. Supplementary Fig. 2. Joint geographic patterns of dietary microplastic exposure and LOCC incidence. Bivariate choropleth map showing the combined distribution of estimated dietary microplastic exposure (tertiles) and LOCC age-standardized incidence (tertiles). Colours represent the nine possible exposure–incidence combinations, highlighting regions with concordantly high exposure and high incidence as well as areas with discordant patterns. The map includes countries with complete data for both variables. Supplementary Material 2. Supplementary Fig. 2. Joint geographic patterns of dietary microplastic exposure and LOCC incidence. Bivariate choropleth map showing the combined distribution of estimated dietary microplastic exposure (tertiles) and LOCC age-standardized incidence (tertiles). Colours represent the nine possible exposure–incidence combinations, highlighting regions with concordantly high exposure and high incidence as well as areas with discordant patterns. The map includes countries with complete data for both variables. Supplementary Material 3. Supplementary Fig. 3. Pre-specified spatial econometric sensitivity analysis. a. Akaike information criterion (AIC; lower = better) for four candidate models fit to log-LOCC ASR on the complete-case panel (n = 79) using the primary k-nearest-neighbour spatial weights matrix (k = 5, row-standardized): Ordinary Least Squares (OLS), Spatial Lag Model (SLM; primary), Spatial Error Model (SEM), and Spatial Durbin Model (SDM). SLM (AIC = 115.9) improves on OLS (AIC = 121.0) and is comparable to SDM (AIC = 116.3). b. Forest plot of four MP exposure effect contrasts under the primary SLM: p10→p50, p50→p90, median ± IQR/2, and the pre-specified p10→p90 headline contrast. The dashed line marks the null. None of the contrasts is statistically distinguishable from zero. Footer line records the pre-specified decision-rule path (C = null/hypothesis-generating) together with the SLM spatial autoregressive parameter (ρ = 0.346) and the post-fit Moran’s I on SLM residuals (-0.006). Supplementary Material 3. Supplementary Fig. 3. Pre-specified spatial econometric sensitivity analysis. a. Akaike information criterion (AIC; lower = better) for four candidate models fit to log-LOCC ASR on the complete-case panel (n = 79) using the primary k-nearest-neighbour spatial weights matrix (k = 5, row-standardized): Ordinary Least Squares (OLS), Spatial Lag Model (SLM; primary), Spatial Error Model (SEM), and Spatial Durbin Model (SDM). SLM (AIC = 115.9) improves on OLS (AIC = 121.0) and is comparable to SDM (AIC = 116.3). b. Forest plot of four MP exposure effect contrasts under the primary SLM: p10→p50, p50→p90, median ± IQR/2, and the pre-specified p10→p90 headline contrast. The dashed line marks the null. None of the contrasts is statistically distinguishable from zero. Footer line records the pre-specified decision-rule path (C = null/hypothesis-generating) together with the SLM spatial autoregressive parameter (ρ = 0.346) and the post-fit Moran’s I on SLM residuals (-0.006). Supplementary Material 4. Supplementary Table 1. Data provenance for all variables used in the analysis: For each variable, we report the source organization, the specific dataset or publication from which it was drawn, the year(s) used, country coverage, canonical URL, license, access date, and citation. The table is generated programmatically from provenance.json sidecar files attached to every downloaded artefact, guaranteeing consistency between the in-text descriptions and the archived raw data. Supplementary Material 4. Supplementary Table 1. Data provenance for all variables used in the analysis: For each variable, we report the source organization, the specific dataset or publication from which it was drawn, the year(s) used, country coverage, canonical URL, license, access date, and citation. The table is generated programmatically from provenance.json sidecar files attached to every downloaded artefact, guaranteeing consistency between the in-text descriptions and the archived raw data.