Key Findings

Results from 15 analytical methods applied to 3,248 US counties. Two cancer types — kidney and colorectal — show robust, converging evidence of association with pesticide exposure.

Surviving Associations

Of eight cancer types tested in BYM2 Bayesian spatial models, only kidney and colorectal cancer show statistically credible associations with total pesticide application that survive all model specifications.

BYM2 pesticide rate ratios across 8 cancer types
Figure 1. BYM2 pesticide rate ratios (RR) with 94% highest density intervals across eight cancer types. Only kidney (RR=1.025) and colorectal (RR=1.015) have credible intervals entirely above 1.0. Six other types are null.
BYM2 forest plot for all-site cancer showing all covariate effects
Figure 2. BYM2 covariate effects for all-site cancer. Smoking (RR=1.044) and obesity (RR=1.008) are the dominant predictors. Pesticide is null for all-site, consistent with the signal being limited to specific cancer types.
Publication-quality forest plot of kidney cancer associations across all methods
Figure 2. Kidney cancer: pesticide effect estimates across all analytical methods showing convergent evidence from BYM2, long-difference, compound-specific, and gauntlet analyses.
Publication-quality forest plot of compound-specific BYM2 results
Figure 3. Compound-specific BYM2 rate ratios for kidney and colorectal cancer. Herbicides (green) show consistent significant associations; insecticides (orange) and fungicides (purple) are largely null for colorectal.

Confounding Robustness

The pesticide–kidney and pesticide–colorectal associations were tested across five progressively inclusive model specifications. Rate ratios remain stable or strengthen as confounders are added.

Model Covariates Added Kidney RR P(>0) ρ Colorectal RR P(>0) ρ
BYM2 v1 (9 cov) Baseline: demographics, health, SES 1.025 0.997 0.986 1.015 0.991 0.989
BYM2 v2A (12 cov) + food insecurity, food access, binge drinking 1.026 0.997 0.982 1.015 0.987 0.988
BYM2 v2B (13 cov) + nitrate water contamination 1.034 0.997 0.971 1.018 0.991 0.939
+ Livestock (NB 15) + hog/cattle/chicken density (all NS) 1.025 0.997 0.986 1.015 0.991 0.989
+ Diabetes (NB 15) + diabetes prevalence (NS) 1.025 0.997 0.986 1.015 0.991 0.989

Key: Nitrate Does Not Confound

Adding nitrate water contamination (a plausible agricultural confounder) to BYM2 v2B actually strengthens the kidney RR from 1.025 to 1.034, while nitrate itself is null for all cancer types. This rules out the hypothesis that the pesticide signal is merely a proxy for agricultural water pollution.

Compound Specificity

Twelve individual pesticide compounds were tested in separate BYM2 models (each with 12 covariates). The results show clear chemical-class specificity: herbicides are consistently associated with both cancer types, while insecticides are uniformly null.

Forest plot of compound-specific BYM2 results for kidney cancer
Figure 3a. Compound-specific BYM2 rate ratios for kidney cancer (original 6 compounds).
Forest plot of compound-specific BYM2 results for colorectal cancer
Figure 3b. Compound-specific BYM2 rate ratios for colorectal cancer (original 6 compounds).
Forest plot of expanded compound BYM2 results for kidney cancer
Figure 4a. Expanded compound BYM2 rate ratios for kidney cancer (6 additional compounds).
Forest plot of expanded compound BYM2 results for colorectal cancer
Figure 4b. Expanded compound BYM2 rate ratios for colorectal cancer (6 additional compounds).

All 12 Compounds: BYM2 Results

Compound Class Cancer Rate Ratio 94% HDI P(>0) Sig?
GlyphosateHerbicideKidney1.027[1.011, 1.045]0.999Yes
GlyphosateHerbicideColorectal1.009[0.998, 1.020]0.928No
AtrazineHerbicideKidney1.017[0.998, 1.035]0.954No
AtrazineHerbicideColorectal1.016[1.004, 1.029]0.991Yes
2,4-DHerbicideKidney1.015[1.001, 1.030]0.977Yes
2,4-DHerbicideColorectal1.012[1.002, 1.022]0.988Yes
DicambaHerbicideKidney1.026[1.009, 1.043]0.998Yes
DicambaHerbicideColorectal1.015[1.005, 1.026]0.997Yes
AcetochlorHerbicideKidney1.012[0.991, 1.032]0.854No
AcetochlorHerbicideColorectal1.016[1.001, 1.030]0.981Yes
Metolachlor-SHerbicideKidney1.018[1.001, 1.033]0.985Yes
Metolachlor-SHerbicideColorectal1.013[1.001, 1.024]0.983Yes
ChlorpyrifosInsecticideKidney1.011[0.996, 1.024]0.926No
ChlorpyrifosInsecticideColorectal1.000[0.991, 1.009]0.525No
MalathionInsecticideKidney0.999[0.981, 1.018]0.493No
MalathionInsecticideColorectal0.991[0.981, 1.001]0.038No
DiazinonInsecticideKidney1.000[0.988, 1.012]0.528No
DiazinonInsecticideColorectal1.004[0.996, 1.010]0.846No
AzoxystrobinFungicideKidney1.021[1.005, 1.039]0.991Yes
AzoxystrobinFungicideColorectal1.006[0.997, 1.016]0.891No
PropiconazoleFungicideKidney1.026[1.009, 1.044]0.997Yes
PropiconazoleFungicideColorectal1.005[0.994, 1.015]0.811No
TebuconazoleFungicideKidney1.008[0.989, 1.025]0.788No
TebuconazoleFungicideColorectal1.003[0.992, 1.013]0.677No

Chemical Class Summary

Colorectal cancer shows the cleanest pattern: 5 of 6 herbicides significant, 0 of 3 insecticides, 0 of 3 fungicides. Kidney cancer is muddier—2 fungicides (azoxystrobin, propiconazole) are also significant for kidney, likely reflecting spatial collinearity with herbicide application in agricultural regions.

Causal Identification

Two quasi-causal methods provide evidence beyond cross-sectional association:

Instrumental Variables (IV/2SLS)

Using total crop acreage as an instrument for pesticide application (first-stage F=377.8, far exceeding the weak-instrument threshold of 10), the IV estimate for pesticide on all-site cancer (0.035) exceeds the OLS estimate (0.019). This pattern—IV > OLS—suggests OLS is attenuated by measurement error in pesticide exposure, not inflated by confounding.

IV first stage: crop acreage strongly predicts pesticide use
Figure 5a. First-stage relationship: crop acreage (instrument) strongly predicts total pesticide application (F=377.8).
IV vs OLS comparison showing IV estimate exceeds OLS
Figure 5b. IV estimate (0.035) exceeds OLS (0.019), consistent with measurement-error attenuation rather than confounding inflation.

Long-Difference Estimator

Within-county changes in pesticide application from 1997 to 2012 predict changes in cancer incidence. This eliminates all time-invariant confounders (geography, demographics, healthcare access).

Kidney (Long-Diff)
β = 0.068
p = 0.003
Δpesticide 1997→2012
Colorectal (Long-Diff)
β = 0.079
p = 0.0004
Δpesticide 1997→2012
All-Site (Long-Diff)
NS
Consistent with BYM2
Δpesticide 1997→2012

Exposure Pathways

To distinguish occupational from dietary/environmental exposure, BYM2 models were run separately on urban (RUCC 1–3) and rural (RUCC 4–9) counties.

Urban vs rural stratified BYM2 results
Figure 6. Stratified BYM2 results. Colorectal cancer shows significant pesticide associations in both urban (RR=1.040, n=135) and rural (RR=1.027, n=1190) counties. Similar RRs across strata suggest a dietary or ubiquitous environmental pathway rather than purely occupational exposure.

Negative Controls

Several negative control analyses confirm the signal is specific to pesticides and not a proxy for general agricultural intensity or metabolic risk factors:

Sensitivity Analyses

LASSO and OLS coefficient comparison
Figure 7a. LASSO selects pesticide among top 10 of 29 predictors, confirming it adds predictive value beyond confounders.
Leave-one-state-out sensitivity analysis
Figure 7b. Leave-one-state-out: pesticide coefficient remains positive across all 50 jackknife iterations, indicating no single state drives the result.

Continue Exploring

Risk Factor Gauntlets — See how pesticide associations survive when smoking, obesity, alcohol, and inactivity are modeled as primary exposures.
Exploratory Screening — Hypothesis-free scan of all predictors across 26 cancer types confirms pesticide specificity for kidney and colorectal.
Temporal Trends — Animated maps showing how cancer incidence and pesticide use evolve over time.

Full data tables for all results above are available on the Downloads page. Interactive maps of model outputs are on the Maps page.