Agricultural pesticide use and Alzheimer’s disease dementia prevalence across US counties in a mixed supervised–unsupervised analysis Authors: Leon Wang, Sara Knox, Andrew B. Lawson, Abe Mollalo Published: 2026 ABSTRACT Agricultural pesticide use represents one of the most geographically patterned environmental systems, yet most prior research has focused largely on individual compounds rather than correlated exposure regimes. We conducted a cross-sectional analysis using modeled pesticide application intensity and Alzheimer’s Disease (AD) dementia prevalence at the county-level across the United States. Stability-based Elastic Net screening and clustering were used to identify exposure groupings, and associations with AD prevalence were estimated using adjusted regression models. Out of 462 total pesticides screened, 112 demonstrated high selection stability and were grouped into 25 exposure clusters. Twenty clusters were significantly associated with AD dementia prevalence (p<0.05). The strongest positive associations were observed for a soil fumigation/nematicide system, an herbicide-dominant vegetation control regime, and a neuroactive insecticide system. Neuroactive insecticides and soil-intensive treatment systems were disproportionately represented among positively associated clusters, whereas systems dominated by phenoxy- and photosystem II inhibiting herbicides were more frequently aligned with inverse gradients. The fully adjusted model explained 59% of between-county variance (baseline R2 = 0.44). Findings suggest that pesticide mixtures are associated with geographic heterogeneity in AD dementia prevalence and warrant higher-resolution, longitudinal investigation.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-55678-4. Introduction Alzheimer’s disease (AD) dementia prevalence varies substantially across the United States (U.S.), with marked geographic differences observed at the county and regional levels1,2. These spatial patterns suggest that contextual and environmental determinants may contribute to disparities in dementia burden beyond established individual-level risk factors. Although AD dementia is the leading cause of dementia worldwide and currently affects more than seven million older adults in the U.S., its environmental drivers remain incompletely characterized3. Identifying potentially modifiable environmental contributors to geographic variation in AD dementia prevalence is therefore critical to informing prevention-oriented public health research. Growing evidence suggests that broader environmental factors may contribute to population-level patterns of cognitive decline. Environmental gradients including drinking water quality, heavy metal exposure, air pollution, and chronic noise have been associated with an elevated risk of AD dementia4. Conversely, greater green space coverage and robust built environments (i.e., public transportation, walkability, etc.) have been associated with reduced risk5,6. Within this environmental context, exposure to pesticides is of particular concern because of their widespread agricultural use and neurotoxicity. Pesticides are often designed to disrupt critical neurobiological processes such as neurotransmission, mitochondrial function, and cell division. Experimental evidence from in vivo and in vitro studies across multiple pesticide classes have demonstrated induction of oxidative stress, neuroinflammation, impaired proteostasis, and direct neuronal damage – processes consistently implicated in neurodegenerative pathways7,8. Importantly, studies examining pesticide mixtures have provided evidence that combined exposures can produce synergistic neurotoxicity beyond what is observed for individual compounds or classes in isolation9,10. Epidemiologic evidence supports this biological plausibility. Early population-based investigations, including the French PAQUID cohort and the U.S. Agricultural Health Study, reported significant associations between occupational pesticide exposure and neurodegenerative outcomes in extended follow-up analyses11–14. Specific to AD dementia risk, the Cache County Study demonstrated increased risk among individuals with occupational organophosphate and organochlorine pesticide exposure in an agricultural community cohort15, while biomarker-based analyses have linked elevated serum organochlorine pesticide levels to a higher likelihood of AD dementia diagnosis16. Similarly, systematic reviews and meta-analyses conclude that pesticide exposure is associated with increased risk of cognitive impairment and AD dementia, despite variance in study design and effect estimations4,17. A central limitation of the existing literature, however, lies in the way pesticide exposure is conceptualized and modeled. Agricultural pesticide application rarely occurs in isolation. Modern cropping systems typically involve the combined use of multiple herbicides, insecticides, fungicides, fumigants, and plant growth regulators (PGR), creating intrinsically correlated exposure patterns across space10. Although biomarker-based studies offer improved exposure specificity and reduced exposure misclassification risk relative to ecological estimates18,19, correlated exposure patterns and shared toxicologic pathways can still complicate inference when individual compounds are modeled in isolation20. Moreover, modern agricultural datasets include hundreds of active ingredients, creating an exposure space characterized by significant intercorrelation and many chemicals with zero or near zero application in a large number of counties21. As a result, associations attributed to individual compounds may instead be indicative of unaccounted for co-application schemes and regional agricultural practices22,23. Conventional regression approaches are therefore vulnerable to significant residual confounding as well as instability in high-dimensional settings. Recent advances in environmental modeling emphasize high-dimensional, mixture-aware methods capable of characterizing joint exposures rather than isolated compounds20,24,25. However, comprehensive chemical investigations of pesticide exposure patterns in relation to AD dementia prevalence remain limited. Several ecological and geospatial studies – most prominently in Parkinson’s disease – have used high-resolution pesticide application data and residential modeling to evaluate associations at regional or county scale, but typically focus on limited sets of individual active ingredients26–29. Moreover, AD dementia literature has largely relied on prospective cohort, case-control, or biomarker-based designs that evaluate individual chemicals or aggregate exposure measures, with relatively few ecological analyses examining large-scale spatial variation in pesticide intensity4,13,30,31. As a result, the structured and high-dimensional nature of agricultural pesticide systems remains incompletely characterized at the national scale. To address this gap, we evaluate cross-sectional, county-level pesticide exposure patterns across the U.S. in relation to AD dementia prevalence. To our knowledge, no prior study has evaluated AD dementia prevalence using a comprehensive, mixture-aware framework that accounts for correlated pesticide application patterns across the U.S. We leverage a comprehensive set of available pesticide application estimates rather than restricting analyses to predefined chemical classes or previously implicated compounds. Specifically, we address the following research questions:Which pesticide mixtures are consistently associated with county-level AD dementia prevalence after adjustment for environmental, socioeconomic, and regional factors?Do these retained pesticide associations align with established mode-of-action (MOA) groupings or co-application schemes?Can correlated pesticide exposure gradients be reduced into interpretable mixture-level components suitable for joint modeling? Which pesticide mixtures are consistently associated with county-level AD dementia prevalence after adjustment for environmental, socioeconomic, and regional factors? Which pesticide mixtures are consistently associated with county-level AD dementia prevalence after adjustment for environmental, socioeconomic, and regional factors? Do these retained pesticide associations align with established mode-of-action (MOA) groupings or co-application schemes? Do these retained pesticide associations align with established mode-of-action (MOA) groupings or co-application schemes? Can correlated pesticide exposure gradients be reduced into interpretable mixture-level components suitable for joint modeling? Can correlated pesticide exposure gradients be reduced into interpretable mixture-level components suitable for joint modeling? Through this mixture-aware approach, we seek to provide a structured understanding of how agricultural pesticide exposures may contribute to geographic variation in AD dementia prevalence across the U.S. and highlight potentially modifiable environmental risks relevant to public health policy and prevention strategies. Importantly, the outcome in this study is county-level modeled AD dementia prevalence for 2020 rather than incident disease. Accordingly, the analysis evaluates geographic variation in aggregate disease burden, which reflects a combination of disease risk, survival, migration, and broader regional context. Methods AD dementia data County-level (n = 3,033) estimates of AD dementia prevalence among adults aged ≥65 years for 2020 were obtained from Dhana et al32. These estimates were derived from the Chicago Health and Aging Project (CHAP), a population-based cohort of more than 10,000 older adults with detailed neuropsychological assessments. In the original study, generalized additive quasibinomial regression models adjusted for age, sex, race/ethnicity, and educational attainment were applied to bridged-race postcensal population estimates from the National Center for Health Statistics to generate demographically standardized county-level prevalence estimates nationwide. This approach yields a spatially complete outcome measure and reduces bias from geographic variation in diagnostic practices and healthcare access. Pesticide exposure data County-level agricultural pesticide use data were obtained from the U.S. Geological Survey (USGS) National Pesticide Synthesis Project, which integrates pesticide-use surveys, crop-specific application rates, and crop distribution data to estimate annual pesticide application at the county level21. For each pesticide–county–year combination, high and low annual mass estimates (kg) were averaged when both were available. Because 2019 pesticide estimates were preliminary and less complete, cross-sectional exposure was defined as the mean mass applied from 2010 through 2018. Pesticides were included in the analytic exposure pool if they exhibited nonzero county-level application in at least one county-year during 2010–2018. We conceptualized the 2010–2018 average as a proxy for relatively stable, contemporary county-level agricultural pesticide systems and regional application intensity, while also providing the most complete pre-2019 summary available from the USGS dataset. This measure is intended to capture persistent spatial exposure gradients rather than specific etiologic latency window for AD dementia onset. In total, 462 unique pesticides met this criterion and constituted the candidate exposure set for supervised screening. Pesticide density was calculated by dividing mean mass by county land area (km2). Densities were log 10-transformed to address right skewness, with a small pesticide-specific offset added prior to transformation to accommodate zero-application counties. Confounding variables Covariates were selected a priori based on prior literature and conceptual relevance as potential common causes of both pesticide exposure intensity and AD dementia prevalence17,33,34. Adjustments were limited to upstream socioeconomic, environmental, and structural factors to avoid overadjustment. Demographic variables incorporated into the construction of the modeled AD dementia prevalence outcome were not re-adjusted in regression models to prevent redundancy and over-control. County-level confounders were grouped into socioeconomic, population structure, environmental, and regional variables. Socioeconomic variables included percentage of the population below the poverty line, proportion of mobile homes, proportion of crowded housing, and proportion of the population residing in group quarters. These variables were obtained from the CDC/ATSDR Social Vulnerability Index (SVI) datasets35. Population structure was represented by population density, derived from annual county population estimates from the U.S. Census Population Estimates Program and 2020 county land area from U.S. Census TIGER/Line shapefiles36,37. Environmental variables included mean normalized difference vegetation index (NDVI), a satellite-derived measure of green vegetation density and photosynthetic activity at the land surface obtained from MODIS38, proportion of agricultural land, obtained from the USDA Cropland Data Layer39 and ambient particulate matter ≤ 2.5 µm (PM2.5) and ozone (O3) concentrations, obtained from the Center for Air, Climate, and Energy Solutions40. Finally, U.S. Census division fixed effects were included in all models to account for broad regional variation41. Confounding data preparation All covariates were constructed as mean cross-sectional estimates aligned temporally with the pesticide exposure window (2010–2018). Socioeconomic variables from the SVI were combined across 2010, 2014, 2016, and 2018 using a weighted temporal average. Population density was calculated as mean county population (2010–2018) divided by 2020 county land area and log10-transformed prior to modeling. NDVI was computed as the mean county-level April–October value aggregated at 250 m spatial resolution, 16-day composite MODIS products averaged across 2010–2018. The proportion of agricultural land was calculated as the mean 30 m Cropland Data Layer classifications aggregated to the county level and then averaged over the same period. PM2.5 and O3 concentrations were averaged across 2010–2018 prior to inclusion in regression models. Analytical overview This study followed a multi-stage analytical pipeline to examine associations between agricultural pesticide use and county-level AD dementia prevalence across the U.S. We first assembled and temporally aligned county-level data on AD dementia prevalence, pesticide exposures, and relevant confounders. Next, we applied stability-based Elastic Net regression to screen pesticide exposures and retain those consistently associated with the outcome. The selected pesticides were then grouped into correlated exposure clusters using hierarchical clustering, and cluster-level composite scores were derived via principal component analysis (PCA). We estimated associations between these composite exposure scores and AD dementia prevalence using multivariable linear regression with full adjustment for confounders. To support interpretation, we conducted MOA enrichment and directional allocation analyses. Model robustness and spatial dependence were evaluated through repeated train–test validation and spatial autocorrelation diagnostics. More detailed descriptions of each analytical step are provided in the sections that follow. Figure 1 provides an overview of the full analytical workflow, including feature screening, clustering, dimensional reduction, inferential modeling, and spatial diagnostics.Fig. 1Summary of the multi-stage analytical framework used for this study. Summary of the multi-stage analytical framework used for this study. Stability-based elastic net screening Cross-validated elastic net screening To address high dimensionality and multicollinearity among pesticides, initial screening of pesticide exposures was conducted using Elastic Net regression, which combines ℓ1 and ℓ2 penalties to enable selection of correlated feature while stabilizing coefficient estimates 42. Model tuning and evaluation were conducted using nested cross-validation to prevent information leakage43. Ten repeated 10-fold outer splits were used to generate independent county-level train–test partitions. Within each outer training fold, 5-fold inner cross-validation was applied to select the optimal regularization strength (α) and mixing parameter (ℓ1 ratio) based on mean squared error. Models were implemented using scikit-learn’s ElasticNetCV module44. To remove background covariance and to ensure consistency with fully adjusted linear models during feature screening, both pesticide exposures and outcome were residualized with respect to confounders prior to modeling45. Stability pesticide filtering A stability-based filtering procedure was applied to further reduce modeling instability and retain only reproducibly selected features following high-dimensional regularized screening. Pesticide selection frequency was defined as the proportion of outer folds in which a pesticide received a nonzero coefficient and was used as a proxy for stability. Prioritizing pesticides that are consistently selected across resampling reduces sensitivity to sampling variability and improving robustness of downstream analyses. This framework applies stability-based feature principles for regularized feature selection as described by Meinshausen and Bühlman46. A data-driven cutoff for retained features was determined using a geometric elbow criterion applied to the ranked selection-frequency curve, defined as the point of maximal perpendicular distance from the chord connecting the minimum and maximum selection frequencies47. This criterion identifies a data-driven change point in the selection frequency distribution, separating high-stability pesticides from those with lower and more variable selection frequencies. Hierarchical clustering of retained pesticides Hard threshold-based clustering solution To summarize shared applications patterns among retained pesticides, hierarchical agglomerative clustering using Ward linkage and pairwise Pearson correlation distances was performed48. Hierarchical clustering was selected because it provides a dendrogram representation of correlated pesticide structures, which enables flexible, data-driven determination of number of clusters instead of pre-specification. Ward linkage was selected to form compact clusters that minimize aggregate within-cluster variance. As an initial descriptive step, strongly correlated pesticide pairs and networks (mean |r| ≥ 0.70) were identified to visualize high-correlation groupings but were not used for inferential modeling. Variance-explained criterion-based clustering solution Dendrograms were cut using a data-driven optimization criterion based on within-cluster dimensionality reduction for formal exposure summarization. Within each candidate clustering solution, PCA was conducted separately for each cluster to decompose correlated pesticide exposures into orthogonal components ordered by explained variance49. The first principal component (PC1) was retained for each cluster, which represents the linear combination of pesticide signals that best captures shared variation across counties. PC1 provides a singular weighted composite exposure score that accurately summarizes cluster-level use patterns while simultaneously reducing dimensionality. Higher-order components were not retained because they capture lower-variance, orthogonal patterns that are less representative of the dominant co-application structure and may disproportionately weight more diffuse clusters. Following this, a weighted total PC1 variance-explained score was computed across clusters, with weights proportional to cluster size. The optimal clustering solution was selected from the score/cluster number curve using the same geometric elbow criterion applied previously, and the resulting clusters were used for downstream regression analyses. This criterion selects a clustering solution by balancing model complexity at which additional clusters yield diminishing returns in explained variance. Multivariable ordinary least squares modeling of AD dementia prevalence Associations between cluster-level pesticide exposures and county-level AD dementia prevalence were evaluated using multivariable ordinary least squares (OLS) regression. All cluster-level PC1 exposure scores were standardized and modeled simultaneously to estimate their independent associations. The full confounder set was included directly in regression models as well to obtain fully adjusted effect estimates. Model specification was identical across all inferential analyses. To account for potential heteroscedasticity in county-level residuals, confidence intervals were computed using heteroscedasticity-consistent covariance matrix estimator 3 (HC3) standard errors50. This approach reduces the risk of over-optimistic inference in the presence of unequal variance across observations. For comparison, an alternative confounder-adjusted OLS model was fit using individual pesticide exposures rather than cluster-level composite variables. Pesticides retained after stability filtering were further reduced using correlation-based pruning (|r| < 0.85) and iterative variance inflation factor (VIF) filtering (VIF < 10) to mitigate multicollinearity prior to model fitting51,52. MOA enrichment analysis on selected pesticides To assess whether supervised screening preferentially retained pesticides from specific mechanistic groupings, overrepresentation analysis was conducted comparing retained pesticides to the full screened set. Enrichment testing was selected to determine whether particular mode-of-action (MOA) categories were represented more (or less) frequently than expected under random selection, consistent with established overrepresentation analysis frameworks53. MOA classifications were assigned according to established Herbicide Resistance Action Committee (HRAC), Insecticide Resistance Action Committee (IRAC), and Fungicide Resistance Action Committee (FRAC) numerical group designations54–56. For each MOA group, the observed number of retained pesticides was compared to the expected count under random selection based on the overall retention proportion. Hypergeometric testing was used to model selection without replacement from the full pesticide pool. Complementary permutation testing was performed by randomly reassigning retained labels across pesticides while preserving the total number retained. Empirical p values were calculated using directional (one-tailed) tests. For enrichment ratios (ER) > 1, the upper-tail probability was estimated as the proportion of permutations yielding counts greater than or equal to the observed value, while for ER < 1, the lower-tail probability was estimated analogously. MOA groups containing fewer than three pesticides in the screened set were excluded from primary interpretation. Directional MOA allocation analysis Directional allocation analysis was conducted to evaluate whether specific pesticide MOA groups were preferentially concentrated among positively versus inversely associated exposure clusters53. Pesticides were assigned to clusters based on the PC1-based clustering solution. Only clusters with statistically significant PC1 coefficients (p < 0.05) were classified as positive (OLS coefficient [β] > 0) or negative (β < 0). For each MOA group, we calculated the proportion of pesticides assigned to positively associated clusters (Ppos) and the proportion assigned to negatively associated clusters (Pneg). The directional allocation (DS) score was defined as DS = Ppos − Pneg. This analysis was descriptive and was not intended as a formal statistical test of directional enrichment. Model validation The robustness of out-of-sample model performance and estimated cluster-level associations was assessed using repeated 5-fold train–test splits, where each fold served as the test set once per repetition, with 10 repetitions in total. For each split, PCA was fit exclusively on training data within each cluster to derive PC1 loadings, which were then applied to held-out test data to generate test-set PC1 scores. OLS coefficients were estimated using training-set PC1s and evaluated on test-set PC1s to preserve separation between dimensional reduction and model evaluation. Spatial autocorrelation diagnostics Residual spatial dependence was evaluated using global Moran’s I computed on residuals from the global cluster-level OLS model57. Local Getis–Ord Gi* statistics were calculated to identify spatial clustering of high- and low-magnitude residuals58. All analyses were conducted in Python (version 3.13.5). Data manipulation was performed using NumPy and pandas. Elastic Net regression and PCA were implemented using scikit-learn. Clustering solutions were conducted using SciPy, while OLS models were fit using statsmodels. Spatial autocorrelation diagnostics were computed using geopandas and PySAL. Figure generation was implemented using matplotlib. Random seeds were fixed across cross-validation and permutation procedures to ensure reproducibility. Results Descriptive characteristics The mean county-level AD dementia prevalence across the U.S. was 11.2% (SD 1.40). Counties exhibited substantial socioeconomic and demographic heterogeneity, including variations in poverty (mean 16.0% [SD 6.2]), mobile homes (mean 13.4% [SD 9.2]), rural–urban gradients (log10 population density mean 1.19 [SD 0.70]). Environmental covariates varied widely as well, with mean NDVI of 0.63 [SD 0.17]; PM2.5 of 7.59 µg/m3 [SD 1.42]; O3 of 44.2 ppb [SD 4.17]; agricultural land coverage of 23.7% [SD 24.5]) (Table 1).Table 1Summary of county-level outcome and covariates characteristics (2010–2018).VariableMeanSDMedianInterquartile rangeRangePrimary VariablesAD dementia prevalence (%)11.21.410.91.7(5.60, 18.4)Socioeconomic (SVI Components)% Below poverty line16.06.215.17.8(2.79, 48.64)% Mobile homes13.49.211.613.3(0.03, 59.80)% Crowded housing2.21.71.81.4(0.00, 16.08)% Group quarters3.44.21.92.6(0.00, 43.68)Population structureLog10 population density1.190.701.210.79(-1.21, 3.63)EnvironmentalcCovariatesNDVI (mean)0.630.170.670.24(0.12, 0.87)PM2.5 (µg/m3)7.591.427.721.89(2.81, 13.76)O3 (ppb)44.164.1744.414.85(28.86, 58.02)Agricultural land (%)23.724.513.836.6(0.00, 90.33) Table 1Summary of county-level outcome and covariates characteristics (2010–2018).VariableMeanSDMedianInterquartile rangeRangePrimary VariablesAD dementia prevalence (%)11.21.410.91.7(5.60, 18.4)Socioeconomic (SVI Components)% Below poverty line16.06.215.17.8(2.79, 48.64)% Mobile homes13.49.211.613.3(0.03, 59.80)% Crowded housing2.21.71.81.4(0.00, 16.08)% Group quarters3.44.21.92.6(0.00, 43.68)Population structureLog10 population density1.190.701.210.79(-1.21, 3.63)EnvironmentalcCovariatesNDVI (mean)0.630.170.670.24(0.12, 0.87)PM2.5 (µg/m3)7.591.427.721.89(2.81, 13.76)O3 (ppb)44.164.1744.414.85(28.86, 58.02)Agricultural land (%)23.724.513.836.6(0.00, 90.33) Summary of county-level outcome and covariates characteristics (2010–2018). Among the 462 candidate pesticides subjected to nested Elastic Net cross-validation, an average of 181 (SD 34) pesticides received nonzero coefficients within individual outer folds prior to stability filtering. Ranking pesticides by cross-fold selection frequency revealed a highly skewed distribution (Fig. 2A), with most pesticides selected infrequently and a smaller subset demonstrating high selection frequency. The logit-transformed representation (Fig. 2B) further supported this separation between low-frequency and high-frequency selections. The empirical cumulative distribution function (Fig. 2C) demonstrated a distinct geometric elbow. The stability threshold of ≥ 0.94 derived from this elbow criterion reduced the feature set to 112 reproducibly retained pesticides for downstream analysis. Retained pesticides spanned all major classes, including herbicides, insecticides, fungicides, fumigants, and plant growth regulators (PGRs). The complete lists of retained and excluded pesticides are provided in Supplement Table S1.Fig. 2Distribution of pesticide selection frequencies across cross-validation folds from nested Elastic Net screening: (A) linear-scale histogram, (B) logit-scale histogram, and (C) empirical cumulative distribution function. Distribution of pesticide selection frequencies across cross-validation folds from nested Elastic Net screening: (A) linear-scale histogram, (B) logit-scale histogram, and (C) empirical cumulative distribution function. Hyperparameter tuning demonstrated a stable penalization region across folds (Figure S1), with moderately consistent α and ℓ1 ratios. Cross-validated predictive performance was stable, with a mean test root mean squared error (RMSE) of 0.88 (SD 0.06) and minimal train–test divergence (mean Δ = 0.033), corresponding to a mean test R2 of 0.60. Detailed hyperparameter distributions for global and per-fold models are summarized in Supplemental Table S2 and Supplemental Figure S1. Hierarchical clustering of retained pesticides Using a descriptive Pearson correlation threshold (|r| ≥ 0.70), several tightly correlated pesticide submodules were identified alongside substantial independence among other retained features (Fig. 3B; Table S3). Correlation heatmaps and network visualizations demonstrated dense within-module structure and clear separation across modules (Figures S2–S3). As shown in Fig. 3C (red dashed curve), this strict threshold produced 91 total submodules consisting of many singletons and small clusters with only a few tightly bounded submodules.Fig. 3Determination of exposure clusters among retained pesticides: (A) cluster size-weighted PC1 score plotted versus number of clusters formed by cutting at different levels of dendrogram with elbow point noted at k* = 25, (B) hierarchical dendrogram with coarse and fine cut levels, (C) empirical cumulative distribution of cluster sizes between coarse and fine cut clustering solutions. Determination of exposure clusters among retained pesticides: (A) cluster size-weighted PC1 score plotted versus number of clusters formed by cutting at different levels of dendrogram with elbow point noted at k* = 25, (B) hierarchical dendrogram with coarse and fine cut levels, (C) empirical cumulative distribution of cluster sizes between coarse and fine cut clustering solutions. In contrast, dendrogram cutting based on the PC1-based variance criterion yielded a coarse solution of 25 exposure clusters. The PC1-fit score across candidate cluster numbers showed a clear elbow at k = 25 (Fig. 3A), and the corresponding dendrogram illustrates the selected cut (Fig. 3B). The cluster size distribution for this solution (Fig. 3C, blue curve) was more evenly distributed across cluster sizes, demonstrating aggregation of fine-scale submodules into broader exposure groupings. Correlation heatmaps and cluster silhouette plots demonstrated retainment of significant within-module structure comparable to the hard threshold clusters above (Figure S4-S5). Cluster sizes ranged from 1 to 11 pesticides (Table S4). Within-cluster variance explained by PC1 ranged from 0.45 to 0.90, indicating moderate to strong shared variation within clusters. The PC1 from each cluster was retained as the composite exposure score for downstream analyses. Cluster associations with AD dementia prevalence We next evaluated whether cluster-level exposure profiles were associated with county-level AD dementia prevalence. In the fully adjusted multivariable OLS model including all 25 cluster PC1 signals, 20 clusters were significantly associated with AD dementia prevalence (Table 2). Nine clusters were positively associated with AD dementia prevalence (β range: 0.047 to 0.176) while 11 clusters were negatively associated with AD dementia prevalence (β range: −0.059 to −0.181), spanning all pesticide classes including mixed-class exposure profiles. More detailed information is found in Table 2.Table 2Multivariable OLS regression coefficients for statistically significant cluster-derived first principal component (PC1) exposure profiles associated with AD dementia prevalence.Cluster\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\beta}}$$\end{document}(SE)95% CInPesticidesDominant exposure profilePositively associated clusters170.176 (0.021)(0.136, 0.217)6DIURON, ETOXAZOLE, OXYFLUORFEN, PYRAFLUFEN ETHYL, SODIUM CHLORATE, SULFOXAFLORMulti-class herbicide system with insecticide/miticide use190.154 (0.023)(0.109, 0.200)5AMETRYN, BENFLURALIN, FENAMIPHOS, METHYL BROMIDE, NAPTALAMSoil fumigation/herbicide system with organophosphate nematicide use40.154 (0.025)(0.105, 0.203)2FLUROXYPYR, METSULFURONBroadleaf herbicide regime (sulfonylurea/auxinic)20.130 (0.021)(0.088, 0.172)11AZOXYSTROBIN, CLOTHIANIDIN, ESFENVALERATE, METCONAZOLE, METRIBUZIN, PENDIMETHALIN, PROPICONAZOLE, TEBUCONAZOLE, TRIBENURON METHYL, TRIFLOXYSTROBIN, TRIFLURALINIntensive multi-class system with herbicides, fungicides, and neuroactive insecticides240.108 (0.017)(0.076, 0.141)4BENOMYL, ETRIDIAZOLE, POLYHEDROSIS VIRUS, TRALOMETHRINBiocontrol-centered pest management system with fungicide and pyrethroid use210.088 (0.015)(0.058, 0.118)6CHLORMEQUAT, FLUVALINATE-TAU, MEVINPHOS, PELARGONIC ACID, PROMETON, TRINEXAPACPlant growth regulator system with concurrent insecticide use60.059 (0.024)(0.012, 0.106)2HEXYTHIAZOX, SPIROMESIFENMiticide-dominant (acaricide) regime160.052 (0.022)(0.009, 0.096)3ACEPHATE, FLUBENDIAMIDE, INDOXACARBMixed insecticide system (organophosphate/diamide)90.047 (0.022)(0.004, 0.090)3BACILLUS THURINGIENSIS, METHIDATHION, PYRETHRINSBiological and organophosphate insecticide programNegatively Associated Clusters3−0.181 (0.027)(−0.235, −0.128)92,4-DB, DIMETHENAMID, DIMETHENAMID & DIMETHENAMID-P, FLUTHIACET-METHYL, LACTOFEN, NICOSULFURON, PENFLUFEN, PERMETHRIN, THIENCARBAZONE-METHYLCorn-production herbicide system with pyrethroid use5−0.163 (0.022)(−0.206, −0.121)3FLUCARBAZONE, PINOXADEN, TRALKOXYDIMGrass-control herbicide regime (lipid biosynthesis inhibitors)18−0.153 (0.021)(−0.194, −0.112)3DICOFOL, ETHABOXAM, PROFENOFOSOrganophosphate insecticide system with acaricide use20−0.092 (0.018)(−0.128, −0.057)3ASPERGILLUS FLAVUS, ENDOTHAL, TEBUTHIURONMixed herbicide system with algicide and biocontrol inputs7−0.089 (0.024)(−0.135, −0.043)4EPTC, HEXAZINONE, NALED, SULFURIC ACIDHerbicide system with organophosphate insecticide use13−0.088 (0.020)(−0.128, −0.048)4CARBARYL, FLUENSULFONE, PHOSPHORIC ACID, TERBACILMulti-class pest management system with herbicide, insecticide, and nematicide use12−0.074 (0.019)(−0.112, −0.037)2FLUMETRALIN, MALEIC HYDRAZIDEPost-harvest plant growth regulator regime (sprout inhibitors)23−0.067 (0.011)(−0.088, −0.045)6FLORPYRAUXIFEN, IMAZOSULFURON, MOLINATE, ORTHOSULFAMURON, PROPANIL, THIOBENCARBRice-production herbicide regime10−0.065 (0.022)(−0.107, −0.023)7CYANTRANILIPROLE, CYFLUFENAMID, DICLORAN, EMAMECTIN, FOSETYL, REYNOUTRIA SACHALINENSIS, SPIROTETRAMATMixed fungicide/insecticide program15−0.063 (0.019)(−0.101, −0.025)6COPPER SULFATE, DIAZINON, METIRAM, NORFLURAZON, PETROLEUM OIL, STREPTOMYCINMulti-class disease-control system with fungicide, insecticide, and bactericide use14−0.059 (0.016)(−0.090, −0.029)31-METHYL CYCLOPROPENE, CLOFENTEZINE, DICHLOBENILPost-harvest regulator system with miticide use Table 2Multivariable OLS regression coefficients for statistically significant cluster-derived first principal component (PC1) exposure profiles associated with AD dementia prevalence.Cluster\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\beta}}$$\end{document}(SE)95% CInPesticidesDominant exposure profilePositively associated clusters170.176 (0.021)(0.136, 0.217)6DIURON, ETOXAZOLE, OXYFLUORFEN, PYRAFLUFEN ETHYL, SODIUM CHLORATE, SULFOXAFLORMulti-class herbicide system with insecticide/miticide use190.154 (0.023)(0.109, 0.200)5AMETRYN, BENFLURALIN, FENAMIPHOS, METHYL BROMIDE, NAPTALAMSoil fumigation/herbicide system with organophosphate nematicide use40.154 (0.025)(0.105, 0.203)2FLUROXYPYR, METSULFURONBroadleaf herbicide regime (sulfonylurea/auxinic)20.130 (0.021)(0.088, 0.172)11AZOXYSTROBIN, CLOTHIANIDIN, ESFENVALERATE, METCONAZOLE, METRIBUZIN, PENDIMETHALIN, PROPICONAZOLE, TEBUCONAZOLE, TRIBENURON METHYL, TRIFLOXYSTROBIN, TRIFLURALINIntensive multi-class system with herbicides, fungicides, and neuroactive insecticides240.108 (0.017)(0.076, 0.141)4BENOMYL, ETRIDIAZOLE, POLYHEDROSIS VIRUS, TRALOMETHRINBiocontrol-centered pest management system with fungicide and pyrethroid use210.088 (0.015)(0.058, 0.118)6CHLORMEQUAT, FLUVALINATE-TAU, MEVINPHOS, PELARGONIC ACID, PROMETON, TRINEXAPACPlant growth regulator system with concurrent insecticide use60.059 (0.024)(0.012, 0.106)2HEXYTHIAZOX, SPIROMESIFENMiticide-dominant (acaricide) regime160.052 (0.022)(0.009, 0.096)3ACEPHATE, FLUBENDIAMIDE, INDOXACARBMixed insecticide system (organophosphate/diamide)90.047 (0.022)(0.004, 0.090)3BACILLUS THURINGIENSIS, METHIDATHION, PYRETHRINSBiological and organophosphate insecticide programNegatively Associated Clusters3−0.181 (0.027)(−0.235, −0.128)92,4-DB, DIMETHENAMID, DIMETHENAMID & DIMETHENAMID-P, FLUTHIACET-METHYL, LACTOFEN, NICOSULFURON, PENFLUFEN, PERMETHRIN, THIENCARBAZONE-METHYLCorn-production herbicide system with pyrethroid use5−0.163 (0.022)(−0.206, −0.121)3FLUCARBAZONE, PINOXADEN, TRALKOXYDIMGrass-control herbicide regime (lipid biosynthesis inhibitors)18−0.153 (0.021)(−0.194, −0.112)3DICOFOL, ETHABOXAM, PROFENOFOSOrganophosphate insecticide system with acaricide use20−0.092 (0.018)(−0.128, −0.057)3ASPERGILLUS FLAVUS, ENDOTHAL, TEBUTHIURONMixed herbicide system with algicide and biocontrol inputs7−0.089 (0.024)(−0.135, −0.043)4EPTC, HEXAZINONE, NALED, SULFURIC ACIDHerbicide system with organophosphate insecticide use13−0.088 (0.020)(−0.128, −0.048)4CARBARYL, FLUENSULFONE, PHOSPHORIC ACID, TERBACILMulti-class pest management system with herbicide, insecticide, and nematicide use12−0.074 (0.019)(−0.112, −0.037)2FLUMETRALIN, MALEIC HYDRAZIDEPost-harvest plant growth regulator regime (sprout inhibitors)23−0.067 (0.011)(−0.088, −0.045)6FLORPYRAUXIFEN, IMAZOSULFURON, MOLINATE, ORTHOSULFAMURON, PROPANIL, THIOBENCARBRice-production herbicide regime10−0.065 (0.022)(−0.107, −0.023)7CYANTRANILIPROLE, CYFLUFENAMID, DICLORAN, EMAMECTIN, FOSETYL, REYNOUTRIA SACHALINENSIS, SPIROTETRAMATMixed fungicide/insecticide program15−0.063 (0.019)(−0.101, −0.025)6COPPER SULFATE, DIAZINON, METIRAM, NORFLURAZON, PETROLEUM OIL, STREPTOMYCINMulti-class disease-control system with fungicide, insecticide, and bactericide use14−0.059 (0.016)(−0.090, −0.029)31-METHYL CYCLOPROPENE, CLOFENTEZINE, DICHLOBENILPost-harvest regulator system with miticide use Multivariable OLS regression coefficients for statistically significant cluster-derived first principal component (PC1) exposure profiles associated with AD dementia prevalence. Adjusted OLS model fit Inclusion of the cluster exposure block increased model R2 from 0.44 in the confounder-only model to 0.59 in the fully adjusted specification (ΔR2 = 0.15), corresponding to a block partial R2 of 0.27. Model fit was comparable to the alternative confounder-adjusted OLS specification using individual pesticide densities, despite substantial reduction in the number of exposure variables. Global MOA enrichment among retained pesticides MOA enrichment analysis restricted to classes with ≥3 retained pesticides demonstrated asymmetric concentration of specific mechanistic groups within exposures. Among enriched classes (ER > 1), photosystem II inhibiting herbicides (HRAC 7) exhibited the strongest and only statistically significant enrichment (ER = 3.45, p = 0.011), with four of five member pesticides classified as stable. Lipid biosynthesis inhibitors (IRAC 23; ER = 2.88) and ryanodine receptor–modulating diamide insecticides (IRAC 28; ER = 2.16) also demonstrated strong enrichment ratios, though these did not reach statistical significance. Two other major neuroactive insecticide groups also demonstrated moderate enrichment, specifically sodium channel inhibiting pyrethroids (IRAC 3A; ER = 1.52) and acetylcholinesterase inhibiting organophosphates (IRAC 1B; ER = 1.39). Moreover, two major herbicide groups – photosystem II inhibitors (HRAC 5; ER = 1.54) and Acetyl-Coa carboxylase (ACC) inhibitors (HRAC 1; ER = 1.30) – and one major fungicide group, Q-outside inhibitors (QoI; FRAC 11; ER = 1.61), also attained moderate enrichment. Several additional herbicide and fungicide classes exhibited enrichment ratios modestly above unity but without statistical evidence of deviation from the null. In comparison, multiple MOA groups demonstrated underrepresentation relative to global core prevalence (ER < 1). Multi-site fungicides (FRAC M01; ER = 0.27, p = 0.09) and biological fungicides (FRAC BM02; ER = 0.33, p = 0.16) showed the strongest concrete depletion signal, approaching statistical significance. Additionally, numerous fungicides – including succinate dehydrogenase inhibitors (SDHI; FRAC 7; ER = 0.39), dithiocarbamate fungicides (FRAC M03; ER = 0.48) and demethylation inhibiting fungicides (DMI; FRAC 3; ER = 0.72) – as well as herbicides – including 4-hydroxyphenylpyruvate dioxygenase inhibitors (HRAC 27; ER = 0.48) and synthetic auxins (HRAC 4; ER = 0.62) – were also underrepresented without statistically significant depletion. Two neuroactive insecticides, specifically carbamate acetylcholinesterase inhibitors (IRAC 1A; ER = 0.43) and neonicotinoid nicotinic acetylcholine receptor inhibitors (IRAC 4A; ER = 0.72), were also moderately under-enriched. Finally, some pesticide classes were fully absent from the core pesticide list, including cytokinin PGRs, methyl isothiocyanate generating insecticides (IRAC 8F), ecdysone receptor activating insecticides (IRAC 18), and α-tubulin targeting herbicides (HRAC 22). The complete list can be found in Supplementary Table S5. Directional allocation of MOA groups (MOA ≥ 3) Restricting to MOA groups represented by ≥3 retained pesticides showed specific directional asymmetry in their allocation across positively and negatively associated clusters (Fig. 4). Several MOA groups were strongly skewed toward positively associated clusters. Demethylase inhibitors (FRAC 3; Ppos = 1.00) and microtubule assembly inhibitors (HRAC 3; Ppos = 1.00) were entirely allocated to positive clusters. Q-outside inhibitors (FRAC 11; Ppos = 0.67) and protoporphyrinogen oxidase inhibitors demonstrated substantial positive skew (HRAC 14; Ppos = 0.67), while pyrethroids (IRAC 3A; direction score [DS] = 0.33) and photosystem II inhibitors (HRAC 5; DS = 0.40) showed moderate positive allocation. Organophosphate acetylcholinesterase inhibitors also exhibited a slight positive imbalance (IRAC 1B; DS 0.11). Conversely, herbicide MOA groups were the only pesticide classes that were directionally allocated to negatively associated clusters. HRAC 1 (DS = −0.50), HRAC 7 (DS = −0.50), and HRAC 15 (DS = −0.80) showed strong negative skew, with HRAC 2 and HRAC 4 also demonstrating net negative allocation. In aggregate, total insecticides demonstrated a slight positive skew (DS = 0.15), while fungicides (DS = −0.21) and herbicides (DS = −0.22) both demonstrated moderate negative skews (Fig. 4).Fig. 4Directional allocation of MOA groups across AD dementia–associated exposure clusters for MOA groups represented by ≥3 retained pesticides. Directional allocation of MOA groups across AD dementia–associated exposure clusters for MOA groups represented by ≥3 retained pesticides. Model validation and spatial residual structures Across repeated fold-wise validation splits, model performance was stable, with training R2 ranging from 0.58 – 0.62 and test R2 ranging from 0.51 – 0.66. Train–test RMSE gaps were small and centered near zero, indicating no systematic overfitting (Table S6). Residuals from the global cluster-level model exhibited moderate spatial autocorrelation (global Moran’s I = 0.265; permutation p = 0.001). Across train–test splits, Moran’s I remained positive in both training (~0.22 – 0.26) and test sets (~0.06 – 0.24), with permutation p ≤ 0.05. Local Getis–Ord Gi* analyses identified coherent spatial clusters of high- and low-magnitude residuals (Figure S6). Discussion In this ecological study, we identified structured agricultural pesticide-use gradients associated with county-level AD dementia prevalence across the continental U.S. Given that disease prevalence was assessed as the outcome measure, these findings should be interpreted in the context of disease occurrence, survival, migration, and broader regional context rather than disease risk alone. We used clustered pesticide signals instead of individual exposures to model correlated pesticide deployment patterns found in real agricultural practices. The resulting gradients included both positively and negatively associated exposure profiles, suggesting that relationships between pesticide use and dementia burden are non-uniform across agricultural settings. AD dementia prevalence may instead reflect regional differences in cropping systems, agricultural intensity, and seasonal field-preparation strategies. Together, these findings suggest that coordinated pesticide-use systems may be more informative than aggregate pesticide volume or isolated single compounds when evaluating regional AD dementia burden. Finally, because counties with similar overall application intensity may differ substantially in the types of chemicals used and cropping systems present, further geospatial mixture analyses may help clarify these latent exposure relationships. Our findings also suggest that environmental exposure systems may be relevant for understanding geographic patterns in AD dementia burden and could inform future prevention-oriented research. Current frameworks primarily focus on modifiable variables such as educational attainment, cardiovascular health, and lifestyle behaviors, with comparatively less attention dedicated towards geographically-patterned exposures17,34,59–62. While these individual-level determinants remain important, environmental systems may also represent potentially relevant environmental correlates from a public health perspective when they are measurable, potentially modifiable, and biologically plausible63. Agricultural pesticide systems meet many of these criteria. Our findings demonstrate that by evaluating pesticide exposures as coordinated mixtures instead of individual compounds, we can better resolve the agricultural and land-use contexts. Such evidence could help policymakers identify high-exposure regions, prioritize environmental monitoring, and may help inform hypothesis generation and guide future regulatory or monitoring strategies aimed at reducing harmful exposure patterns while maintaining agricultural productivity64. Supporting this interpretation, we examined the composition of positively associated clusters using mixture-level analyses, which yielded several distinct but overlapping agricultural management profiles (Table S7). One profile corresponded to soil-intensive systems typical of specialty crop and high-value horticultural production, where early-season fumigation and nematicide use are applied to suppress soil-borne pests prior to planting65. Other profiles were insecticide-dominant and enriched for neuroactive classes such as pyrethroids, organophosphates, and ryanodine receptor modulators commonly used in foliar pest-control programs66. A third profile was characteristic of intensively managed row-crop and specialty production systems, where integrated pest-management strategies are used concurrently to address weed, pest, and disease burden using multi-class pesticide application67,68. Across these profiles, neuroactive insecticide classes were repeatedly overrepresented, and other classes that indirectly act on core components of neurotransmission were sporadically enriched as well. This recurring concentration suggests that the observed associations may reflect structured and sustained deployment of these pesticide classes within specific agricultural contexts rather than effects of individual chemicals alone. Furthermore, despite differences in crop type and pest-management strategy, positively associated clusters shared another common structural feature: coordinated pesticide use within agricultural production systems characterized by repeated seasonal deployment and intensive multi-class management64,67,69. Soil fumigation systems involve high-intensity pre-plant applications with volatilization and off-site transport risks, potentially extending exposure beyond treated fields into surrounding communities70,71. Alternatively, insecticide-dominant systems rely on above-ground foliar treatments in which spray drift and inhalational exposure may represent more relevant environmental diffusion pathways72–75. Fungicide-intensive row-crop programs and specialty crop mite-control systems further illustrated how coordinated pesticide regimens often combine multiple chemical classes across the growing season. Specialty crop production, in particular, has been associated with higher per-acre pesticide application intensity relative to other crop systems68,76,77. Beyond direct drift processes, these agricultural applications may also contribute to broader community exposure through deposition into surrounding soils and household dust, transport into surface or drinking water systems, and residues entering regional food supply chains78. Accordingly, county-level pesticide intensity likely captures only one component of total human exposure burden and may function more broadly as an indicator of cumulative environmental loading across multiple pathways79. Thus, variation in AD dementia prevalence may reflect not only differences in pesticide use, but also how agricultural systems shape cumulative community exposure through distinct environmental pathways. Within positively associated exposure systems, pyrethroids and organophosphate insecticides were among the only insecticides that were both enriched within the set of core pesticides and disproportionately allocated to positive clusters. These two classes have been regularly linked to neurotoxicity and adverse cognitive outcomes in experimental and population-level research15,80–82 with biomonitoring studies documenting elevated metabolite levels of these compounds among residents living in agricultural communities83,84. Out of the remaining pesticides classes, QoI fungicides represented the only additional MOA that met both criteria. Similar to pyrethroids and organophosphates, QoI fungicides are one of the most widely applied pesticides in numerous intensive agricultural contexts and have neurotoxic effects demonstrated in experimental models85–87. However, compared to these neuroactive insecticides, they are relatively underexplored at the population-level for their neurocognitive impacts even as they have been increasingly implicated in biomonitoring and household exposure studies88–91. In all, these observations suggest that rural communities located near intensive crop production may experience sustained exposure to pesticide mixtures characteristic of these agricultural systems and may warrant greater attention in public health research on environmental variables associated with AD dementia92. Conversely, clusters that were negatively associated with AD dementia prevalence presented with substantially different pesticide class compositions (Table S7). Herbicide classes, both individually and in aggregate, were over-enriched within these inverse profiles, while insecticides and fungicide classes that characterized many of the positive clusters were comparatively depleted (i.e. pyrethroids, organophosphates, QoIs, and DMIs). Several of these exposure patterns corresponded to weed-management strategies found in large-scale field-crop production, which center coordinated herbicide deployment rather than multi-class pest-control programs21,93,94. Accordingly, the pesticide mixtures associated with these systems may differ significantly from those observed in more intensive pest-management regimes where multiple pesticide classes are used throughout the growing season to target insect or fungal pathogens. Other inverse profiles corresponded to crop-specific management systems with distinct pesticide use structures. For example, several profiles represented herbicide programs characteristic of weed management strategies found in cereal/grain and row-crop regimes95,96. Likewise, herbicide-dominant patterns characteristic of rice production also emerged, where flooded field conditions and unique hydrologics can influence pesticide composition and transport dynamics97–99. Additionally, a small subset of inverse exposure profiles were also suggestive of specialized management strategies that were not geared towards active pest suppression. These included pesticide mixtures involving PGRs or post-harvest crop management chemicals100,101, along with biological control agents and herbicides used in non-crop (i.e. natural) vegetation management102,103. Overall, the inverse associations observed in this model likely reflect differences in agricultural intensity, pesticide composition, and community exposure profiles rather than true neuroprotective effects. Such profiles may vary according to application practices, environmental dispersion, and proximity to residential populations, thereby shaping cumulative burden across agricultural settings104. This contrast reinforces that differences in how pesticide use is structured across agricultural systems, rather than pesticide presence alone, may underlie the divergent patterns observed in this study. This study has several limitations. The analysis is ecological, with associations estimated at the county level, and the outcome represents modeled prevalence rather than incident disease. Prevalence encodes disease occurrence, survival, as well as geographic differences, and only reflect variation in disease burden rather than true risk. Because AD dementia develops over decades, the current cross-sectional analysis may not fully capture the long-term neurodegenerative effects of pesticide exposure, and observed associations may reflect both recent and historical exposures. The present analytical design also cannot resolve the relative contribution of dietary, airborne, soil, waterborne, or occupational take-home pathways to the observed county-level associations. Moreover, county-level AD dementia prevalence estimates were derived from a demographically adjusted extrapolation of the CHAP rather than direct surveillance. This framework does not incorporate regional variation in several established dementia risk factors that are not consistently available at the county level32. These unmeasured factors may confound or modify geographic differences in prevalence. Additionally, because the published estimates were already adjusted for age, sex, race/ethnicity, and education, inclusion of overlapping demographic variables in subsequent models risked distortion of the outcome structure. This constraint limited the range of additional covariates that could be incorporated. Health-related covariates were also not included, as some cardiometabolic factors may lie along the pathway between environmental exposures and AD dementia, and comparable age-standardized county-level indicators were not consistently available for the study period105,106. Residual spatial autocorrelation in the global cluster-level model also indicates that geographic structure persisted beyond measured exposures and covariates. This pattern suggests that additional spatially patterned environmental, healthcare, or demographic factors likely remain unmeasured57. Agricultural exposure systems should therefore be interpreted as one component of broader geographic patterns rather than a complete explanation of variation in AD dementia prevalence. Mixture construction approach also imposes structural assumptions. Hierarchical clustering defines discrete exposure groups based on correlation structure, whereas pesticide co-application patterns may operate more continuously. Alternative approaches such as varimax-rotated principal components analysis, fuzzy clustering, weighted quantile sum regression, or Bayesian kernel machine regression may capture nonlinear, overlapping exposure systems differently107–110. Future work should incorporate higher-resolution exposure assessment, including the linkage of pesticide application data at finer spatial scales, integration of environmental monitoring data, and, where feasible, biomarker-based measurements of cumulative exposure111. Longitudinal cohort studies linking biomarker-based pesticide exposure to incident AD dementia outcomes would further help ascertain temporal relationships and characterize potential latency between exposure and neurodegenerative disease development112. Future studies should also integrate occupational and residential exposure information, particularly for populations residing or working near high-intensity agricultural production areas where environmental pesticide exposures may be elevated28,113. Spatial error modeling could further elucidate the residual geographic structure observed by permitting geographically-varying error structures114. Simultaneously, methods that incorporate spatially varying relationships, including geographically weighted regression and Bayesian conditional autoregressive modeling, may help determine whether certain pesticide clusters exert stronger associations within specific agricultural region115,116. Conclusion In this nationwide ecological analysis, we found that county-level AD dementia prevalence was more closely aligned with coordinated pesticide exposure patterns than with individual chemical applications. Multiple exposure clusters spanning diverse pesticide classes were identified, with both positive and inverse associations reflecting heterogeneity in agricultural systems across the U.S. Overall, neuroactive insecticide and soil-intensive treatment systems were more consistently associated with higher prevalence gradients, whereas several herbicide- and fungicide-dominant regimes tended to align with lower gradients. These findings suggest that mixture-level pesticide patterns may provide a more informative representation of real-world agricultural exposures than analyses focused on single compounds alone. Given the cross-sectional and ecological design, these associations should be interpreted as indicators of geographic variation in disease burden rather than evidence of disease causation. Further research using higher-resolution exposure assessment and longitudinal designs is needed to clarify temporal relationships and evaluate potential underlying mechanisms. Supplementary Information Supplementary Information. Supplementary Information.