Pesticides associated with incident diabetes among licensed private pesticide applicators in the Agricultural Health Study cohort (1993–2021)

Abstract

Background:

Growing evidence suggests pesticides may increase risk of type 2 diabetes, but data are limited on many specific chemicals.

Methods:

In 29,527 private pesticide applicators in the Agricultural Health Study cohort (enrolled 1993–1997 in Iowa and North Carolina), 3,847 incident diabetes cases were identified by self-report during follow-up surveys in 1999–2003, 2005–2010, 2013–2015, and 2019–2021. We examined 50 pesticides reported at enrollment, updated in 1999–2003 or 2005–2010, prior to diabetes diagnosis or end of follow-up, using log-binomial regression to calculate relative risks (RR) and 95% confidence intervals (CI) for ever-use and intensity-weighted lifetime days (IWLD) use (tertiles, T1-3), adjusting for covariates and correlated pesticides.

Findings:

Greater diabetes risk was associated with 7 organochlorine insecticides: ever-use of DDT, aldrin, dieldrin, chlordane, heptachlor, and toxaphene (RRs 1.08–1.31), without monotonic exposure–response trends, and lower IWLD of lindane use (_T1RR=1.32; 95%CI 1.12–1.57). Risk was associated with 5 organophosphate or carbamate insecticides: ever-use of diazinon and carbofuran, and exposure–response trends for malathion (_T3RR=1.13;95%CI 1.02–1.25, p-trend=0.025), phorate (_T3RR=1.22;95%CI 1.08–1.39, p-trend=0.001), and carbaryl (_T3RR=1.26;95%CI 1.11–1.43, p-trend=0.005). Risk was associated with 2 phenoxy herbicides, 2,4,5-T (ever-use RR=1.25;95%CI 1.14–1.37) and 2,4,5-TP (_T1RR=1.35;95%CI 1.04–1.76), and 3 other herbicides [butylate (_T3RR=1.26;95%CI 1.10–1.44, p-trend<0.001), metribuzin (_T3RR=1.16;95%CI 1.16–1.32, p-trend=0.022), chlorimuron ethyl (_T3RR=1.16;95%CI 1.02–1.31, p-trend=0.033)], and the fumigant carbon tetrachloride/disulfide (RR=1.16;95%CI 1.02–1.33). Associations were not confounded by BMI and weight gain.

Conclusions:

These results show greater diabetes risk associated with use of persistent organochlorine insecticides and banned phenoxy herbicides. Novel findings for widely used insecticides and other pesticides warrant further investigation.

1. Background

Type 2 diabetes mellitus (hereafter referred to as diabetes, typically adult onset) is a chronic metabolic disease of epidemic proportions (Magliano et al., 2021). Risk increases with aging, obesity, family history of diabetes, and socioeconomic factors (Hassan et al., 2023; Hill-Briggs et al., 2020). Although the diabetes epidemic follows increased obesity rates, growing evidence suggests exposure to certain pesticides may play a role (Evangelou et al., 2016; Thayer et al., 2012). Research has focused on persistent organic pollutants, such as the organochlorine (OC) insecticide dichlorodiphenyltrichloroethane (DDT) and metabolite dichlorodiphenyldichloroethylene (DDE) (Evangelou et al., 2016; Jaacks et al., 2019; Song et al., 2016), including measured OC levels, with inconsistent findings (Magliano et al., 2021; Lee et al., 2018; Tornevi et al., 2019). Less is known about other pesticides, such as organophosphorus (OP) insecticides, or dioxin-contaminated herbicides, which may also increase diabetes risk or metabolic risk factors (Czajka et al., 2019; Disease associated with exposure to certain herbicide agents, 2001). Recent cross-sectional studies provide mixed evidence on metabolic risk factors associated with contemporary insecticides and herbicides measured in spot samples (Otaru et al., 2024; Dong et al., 2024; Park et al., 2019; Zhu et al., 2025; Tang et al., 2024; Vuong et al., 2022).

Agricultural populations, with higher occupational exposures, offer further evidence on diabetes risk associated with specific pesticides. In a retrospective study of a Thai rice farming community, diabetes was associated with ever-use of 3 insecticides and one fungicide (out of 35 pesticides examined) (Juntarawijit and Juntarawijit, 2018). The Agricultural Health Study (AHS), a prospective cohort enrolled in 1993–1997, previously investigated pesticide use and incident diabetes in licensed private pesticide applicators (mostly male farmers) through 1999–2002, reporting associations with several OC and OP insecticides, and some herbicides (Montgomery et al., 2008). However, use of specific pesticides is often correlated and internal dose may be impacted by application methods or use of personal protective equipment, neither of which were previously considered.

While mechanisms linking pesticides to diabetes are likely diverse, endocrine disruption and obesity are recurring themes, especially for OC insecticides (Lee et al., 2006; Lee et al., 2010; Mendes et al., 2021). A review of experimental and human studies concluded that DDT was a “presumed” human obesogen (Cano-Sancho et al., 2017). Given the role of obesity and weight gain in metabolic dysfunction and diabetes, BMI may be a confounder, mediator, or susceptibility factor in epidemiologic studies of pesticides and diabetes. The prior study of AHS applicators did not evaluate confounding by BMI. In the current study, we examined pesticide use and diabetes among private applicators in the AHS over a longer follow-up period (up to 2021), including a larger number of incident cases, updated pesticide data (through 2010), and cumulative exposure variables weighted for exposure intensity. We also considered potential confounding by BMI and weight gain, and associations stratified by age and BMI category.

2. Methods

2.1. Study population and sample

The AHS is a prospective cohort including N=52,394 licensed pesticide applicators enrolled in 1993–1997 (Phase 1) (Alavanja et al., 1996). Participants were recontacted in 4 follow-up surveys (Phases 2–5: 1999–2003, 2005–2010, 2013–2015 and 2019–2021), and the eligible study sample included those who completed at least one follow-up survey response on diabetes (N=39,197 applicators). Informed consent was obtained following Institutional Review Board approved procedures at enrollment and each follow-up. Questionnaires collected data on pesticide use, sociodemographic factors, diabetes diagnosis, and covariates (https://aghealth.nih.gov/collaboration/questionnaires.html).

The analysis sample included those at risk of diabetes with complete data on covariates used for inverse probability weighting and main analysis models [enrollment age and weight (for calculated BMI), sex, race/ethnicity, enrollment smoking status, age and weight (for calculating weight change) prior to diabetes or end of follow-up, state, education, and family history of diabetes]. The sample (Supplemental Fig. 1) excluded 13,197 participants lacking all follow-up, 1,207 who reported diabetes at enrollment, 468 who later reported a diagnosis age suggesting prevalent diabetes, 244 with missing enrollment data on diabetes, 172 potential incident cases missing diagnosis age, 290 who refuted their diagnosis at their last follow-up, and 2 diagnosed before age 25 (presumed Type 1 diabetes). Covariate data were missing on 7,287 participants (7,006 at enrollment, 281 during follow-up), leaving an analysis sample of 29,527, including 3,847 (13%) with incident diabetes. Of these, 26,927 (802 incident cases) completed the survey in 1999–2003, 18,651 in 2005–2010 (998 new cases), 18,398 in 2013–2015 (1,163 new cases), and 14,364 in 2019–2021 (884 new cases).

2.2. Case ascertainment

Participants were asked at each phase whether a doctor ever told them they had diabetes and at what age. In those reporting diabetes at multiple phases, the first reported age was accepted. Responses were evaluated for consistency when multiple follow-up surveys were completed; a refuted diagnosis after a positive self-report were considered equivocal and excluded from the sample (above).

2.3. Exposure assessment

The enrollment questionnaire asked about ever-use of 50 specific pesticides, which was updated in the first (1999–2003) and second (2005–2010) follow-up surveys. Use was based on the phase prior to age at diabetes diagnosis or end of follow-up (non-cases). In the analysis sample, 91.2% of respondents had exposures updated in 1999–2003, 4.8% were missing exposures in 1999–2003 but were updated in 2005–2010, leaving 4.0% missing updated exposure data among those with incident diabetes or at end of follow-up in 2013–2015 or 2019–2021.

Intensity-weighted lifetime days (IWLD) of pesticide use were derived from cumulative lifetime days (days per year times the years of use) for 22 pesticides listed on the enrollment questionnaire and 28 pesticides with additional data on the take-home survey, and updated for pesticides still in use by the first two follow-up surveys, weighted by a score incorporating application practices and use of personal protective equipment impacting potential exposure levels (Coble et al., 2011). Exposure-response analyses were limited to 48 pesticides with at least 21 cases per tertile.

2.4. Covariates

Covariates included sex (male, female), race/ethnicity (non-Hispanic white, Black and other), age of diabetes diagnosis or final follow-up (continuous and age-squared), state (Iowa, North Carolina), education (≤high school, some college, ≥college degree), smoking prior to diabetes diagnosis or final follow-up (current, former, never), family history of diabetes, and use of correlated pesticides (Spearman rank correlation, ρ>0.40). BMI at enrollment (<25, 25–<30, 30+ kg/m²) was based on reported weight and height, and weight change (10% loss, stable, 10% gain) was derived based on weight at the last survey prior to diabetes diagnosis or final follow-up) versus enrollment. Cases reported in the first follow-up (1999–2001) or non-cases lacking later follow-up were assigned a stable weight prior to diagnosis/end of study.

2.5. Statistical analyses

To account for ineligible participants lacking all follow-up we used inverse probability weighting based on propensity scores of having follow-up information. Logistic regression models calculated participant observation weights based on enrollment age, sex, race/ethnicity, BMI, state, smoking status, and family history of diabetes. Models of diabetes risk used these weights to reflect the full AHS sample. Due to high missing rates of BMI at enrollment, a sensitivity analysis was conducted where BMI was multiply imputed with n=20 replicates to maintain a maximal sample size.

Log-binomial regression modeling was used to calculate relative risks (RR) and 95% confidence intervals (CI) to estimate associations of diabetes with pesticides, adjusting for age, sex, race/ethnicity, state, education, smoking, family history of diabetes, and correlated pesticides (ρ>0.40, Supplemental Table 1; i.e., Model 1). A second set of models additionally adjusted for BMI and weight gain on study (i.e., Model 2). We also evaluated exposure response across IWLD tertiles of use; trends were calculated using a 4-level categorical variable, centered on the median IWLD per category. Associations are reported for RR point estimates for which the 95%CI excluding the null (p<0.05). For IWLD, we define evidence of monotonic exposure response, i.e., RR for T3 or T2>T1, but also show findings when trends are flat or RR are elevated for the lower (T1) tertiles, which may reflect threshold effects such as seen for endocrine disrupting chemicals (Vandenberg et al., 2012).

We performed multiple sensitivity analyses. Given the number of comparisons performed, we used the Benjamin-Hochberg method to calculate False Discovery Rate (FDR)-adjusted p-values with an overall alpha level of 0.05 for ever-use associations. Analyses were also conducted on ever-use of pesticides in different samples (Fig. 1): (1) those without weight change during follow-up; (3) imputing enrollment BMI data in those missing enrollment BMI; (4) those without a family history of diabetes; (5) phase-specific associations in those with complete follow-up. We considered associations, stratified by enrollment age (<50 vs. 50+ years) and BMI (<25, 25+, <25, 25-<30, 30+ kg/m²). Our study sample included participants completing at least one follow-up survey. However, nearly half were lost to follow-up prior to the last survey (2018–2021): 9,433 (32%) were presumed living and 4,256 (14%) deceased according to linkage records from the National Death Index. Both groups had an average of 12 years active follow-up (IQR 6,18), while deceased participants were older at enrollment [age 61 years (IQR 53, 67), vs. 40 years (IQR 33, 44) among the living], were more often from North Carolina (42% vs. 33%) and less likely to be non-smokers (40% vs. 58%) (Supplemental Table 2). We used Fine-Grey models to calculate hazard ratios (HR) and 95% CI adjusting for Model 1 covariates, to address potential bias due to informative censoring (i.e., mortality) influencing the observed exposure response patterns observed in RR across IWLD tertiles use of OC insecticides and phenoxy herbicides 2,4,5-T and 2,4,5-TP.

Fig. 1.

Diabetes risk associated with cumulative pesticide use in licensed private pesticide applicators in the Agricultural Health Study: tertiles of intensity-weighted lifetime days (IWLD) reported at enrollment, 1999–2003, or 2005–2010, using the most recent available data prior to reported diabetes or end of follow-up (non-cases).Fig. 2. Diabetes risk in licensed private pesticide applicators in the Agricultural Health Study: ever use of select pesticides at enrollment, 1999–2003, or 2005–2010 prior to reported diabetes or end of follow-up (non-cases), stratified by median enrollment age of cases (50), and BMI (overweight/obese). Relative risk (RR) and 95% confidence intervals (CI) were calculated in log-binomial models adjusting for sex/gender, race/ethnicity, age at end of study (+age-squared), state, educational attainment, smoking status, family history of diabetes, and correlated pesticides (R²>0.40). Models are weighted inversely proportional to propensity scores for baseline only information. IWLD data limited among those not completing the take-home questionnaire indicated with an asterix (*).

This study used data releases AHSREL202309.00, P1REL202210.00, P2REL202210.00.V2, P3REL202210.00.V2, P4REL201909.00, P5Pre-REL202309. Analyses were conducted in SAS version 9.4 (Cary, NC).

3. Results

Cases were older at enrollment (median 49 years; interquartile range, IQR 42–57) than non-cases (44 years, IQR 37–55), and had shorter median follow-up time (17 years, IQR 11–20 vs. 20, IQR 12–25) (Table 1). Median diagnosis age was 66 years (IQR 58–73). Most were male and non-Hispanic white (98%). Cases were more often from North Carolina (38% vs. 30% non-cases), obese (46% vs. 19% non-cases), past smokers (41% vs. 33% non-cases), and had a family history of diabetes (31% vs. 18% non-cases). Weight gain on study was similar among cases and non-cases.

Table 1.

Characteristics of the study sample, including diabetes cases and non-cases through Phase 5 (1999–2021) in licensed private pesticide applicators in the Agricultural Health Study.

	Overall (N=29,527)	Diabetes cases (N=3,847)	Non-cases (N=25,680)
Sample characteristics enrollment or as specified	N (%) or Median (IQR)a	N (%) or Median (IQR)a	N (%) or Median (IQR)a
Age at enrollment	45 (37, 56)	49 (42, 57)	44 (37, 55)
Age at end of study/diabetesb	65 (55, 75)	66 (58, 73)	65 (54, 75)
Time on study (years)	19 (12, 25)	17 (11, 20)	20 (12, 25)
Gender
Male	28,824 (98)	3,768 (98)	25,056 (98)
Female	703 (2)	79 (2)	624 (2)
Race/ethnicity
White, non-Hispanic	28,817 (98)	3,720 (97)	25,097 (98)
Black/Other	710 (2)	127 (3)	583 (2)
State
Iowa	20,394 (69)	2,384 (62)	18,010 (70)
North Carolina	9,133 (31)	1,463 (38)	7,670 (30)
BMI at enrollment (kg/mb)
<25	7,730 (26)	307 (8)	7,423 (29)
25–<30	15,111 (51)	1,745 (46)	13,366 (52)
30+	6,686 (22)	1,795 (46)	4,891 (19)
Diabetes family history
None	23,737 (80)	2,649 (69)	21,088 (82)
Any	5,790 (20)	1,198 (31)	4,592 (18)
Missing
Educational attainment
High school or lower	15,849 (54)	2,210 (58)	13,639 (53)
Some college	7,785 (26)	966 (25)	6,819 (26)
College degree or higher	5,893 (20)	671 (17)	5,222 (20)
Smoking Status
Never	16,617 (56)	1,886 (49)	14,731 (57)
Past	9,791 (34)	1,548 (41)	8,243 (33)
Current	3,119 (11)	413 (11)	2,706 (11)
Weight change on studyc
Loss	739 (3)	54 (1)	685 (3)
Stable	24,422 (83)	3,215 (84)	21,207 (83)
Gain	4,366 (15)	578 (15)	3,788 (15)

IQR, interquartile ratio; BMI, body mass index.

^aUnweighted frequencies.

^bAge at reported diabetes diagnosis or completed follow-up (year last completed survey).

^cWeight change calculated based on reported weight prior to reported diabetes diagnosis or final completed follow-up (greater than 10% loss or gain, or stable within 10% of baseline weight).

3.1. Ever-use

Of 7 OC insecticides examined, 5 were associated with greater diabetes risk: aldrin (RR=1.16; 95%CI 1.05–1.28), chlordane (RR 1.08; 95%CI 1.00–1.17), dieldrin (RR=1.17; 95%CI 1.02–1.35), DDT (RR=1.31; 95%CI 1.20–1.43), heptachlor (RR=1.20; 95%CI 1.08–1.33), and toxaphene (RR=1.18; 95%CI 1.08–1.29) (Table 2). Risk was positively associated with 2 insecticides, phorate (RR=1.10; 95%CI 1.02–1.19) and carbofuran (RR=1.12; 95%CI 1.05–1.20), but inversely associated with permethrin use on livestock (RR=0.84; 0.76–0.92). Diabetes was also positively associated with 2,4,5-T (RR=1.25; 95%CI 1.14–1.37) and butylate (RR=1.12; 95%CI 1.03–1.21), and the fumigant, carbon tetrachloride/disulfide (RR=1.16; 95%CI 1.02–1.33) (Table 3). Use of glyphosate (RR=0.88; 95%CI 0.80–0.96) and petroleum oil/distillates (RR=0.90; 95%CI 0.84–0.97) were inversely associated with diabetes. No substantial confounding by BMI and weight change was seen.

Table 2.

Associations of ever use of insecticides with diabetes risk during follow-up of AHS private pesticide applicators.

	Diabetes cases	Non-cases	Relative risk and 95% confidence intervals
History of pesticide use (ever/never)a	(N=3,847) N (%)b	(N=25,680) N (%)b	Model 1 covariatesc	+ BMI and weight gaind
Organochlorines
Aldrin	811 (24)	4,425 (20)	1.16 (1.05–1.28)	1.11 (1.01–1.22)
Chlordane	1,039 (30)	5,770 (26)	1.08 (1.00–1.17)	1.07 (0.99–1.15)
DDT	1,060 (31)	5,511 (25)	1.31 (1.20–1.43)	1.22 (1.13–1.33)
Dieldrin	284 (8)	1,501 (7)	1.17 (1.02–1.35)	1.16 (1.02–1.32)
Heptachlor	672 (20)	3,747 (17)	1.20 (1.08–1.33)	1.12 (1.02–1.24)
Lindane	789 (23)	5,023 (23)	1.04 (0.96–1.13)	1.05 (0.97–1.13)
Toxaphene	588 (17)	3,129 (14)	1.18 (1.08–1.29)	1.13 (1.04–1.24)
Organophosphates
Chlorpyrifos	1,709 (44)	11,754 (46)	0.95 (0.89–1.01)	0.94 (0.88–1.00)
Coumaphos	347 (9)	2,273 (9)	1.02 (0.91–1.14)	1.03 (0.93–1.15)
Diazinon	1,244 (36)	7,450 (34)	1.00 (0.93–1.08)	0.98 (0.91–1.06)
Dichlorvos	433 (11)	2,885 (11)	1.07 (0.97–1.19)	1.08 (0.97–1.19)
Fonofos	840 (22)	5,662 (22)	1.04 (0.96–1.13)	1.02 (0.95–1.11)
Malathion	2,553 (74)	16,544 (74)	1.01 (0.93–1.09)	0.99 (0.92–1.07)
Parathion	561 (16)	3,254 (15)	1.01 (0.92–1.11)	0.96 (0.87–1.05)
Phorate	1,258 (36)	7,990 (36)	1.10 (1.02–1.19)	1.05 (0.98–1.13)
Terbufos	1,528 (40)	10,257 (40)	1.06 (0.99–1.14)	1.02 (0.95–1.09)
Trichlorfon	23 (1)	168 (1)	0.86 (0.57–1.31)	0.92 (0.61–1.37)
Carbamates
Aldicarb	381 (11)	2,212 (10)	0.91 (0.81–1.02)	0.88 (0.79–0.98)
Carbaryl	2,116 (62)	12,659 (57)	1.07 (0.99–1.16)	1.07 (0.99–1.15)
Carbofuran	1,215 (32)	7,088 (28)	1.12 (1.05–1.20)	1.07 (1.00–1.14)
Pyrethroids
Permethrin (crops)	562 (15)	3,919 (15)	0.94 (0.86–1.03)	0.93 (0.86–1.02)
Permethrin (animals)	511 (13)	4,227 (17)	0.84 (0.76–0.92)	0.89 (0.81–0.98)

^aReported at enrollment, 1999–2003, or 2005–2010, using the most recent data prior to diabetes diagnosis or end of follow-up (non-cases).

^bObservations weighted inversely proportional to propensity scores using baseline data.

^cRelative risk (RR) and 95% confidence interval (CI) calculated in log-binomial models adjusting for sex/gender, race/ethnicity, age at end of study (+age-squared), state, educational attainment, smoking status, family history of diabetes, and correlated pesticides (R²>0.40).

^dRelative risk (RR) and 95% confidence interval (CI) further adjusting for BMI at enrollment and weight gain during follow-up prior to diabetes diagnosis.

Table 3.

Associations of ever use of herbicides, fungicides, and fumigants with diabetes risk during follow-up of AHS private pesticide applicators.

	Diabetes cases	Non-cases	Relative risk and 95% confidence intervals
History of pesticide use (ever/never)a	(N=3,847) N (%)b	(N=25,680) N (%)b	Model 1 covariatesc	+ BMI and weight gaind
Herbicides
Anilids/Anilines
Alachlor	2,091 (55)	13,628 (54)	1.04 (0.97–1.11)	0.98 (0.92–1.04)
Metolachlor	1,910 (50)	12,856 (50)	1.04 (0.97–1.11)	1.01 (0.95–1.08)
Pendimethalin	1,632 (47)	10,759 (48)	0.95 (0.88–1.01)	0.92 (0.86–0.98)
Trifluralin	1,984 (52)	13,787 (54)	1.00 (0.92–1.08)	0.97 (0.90–1.04)
Phenoxy
2,4-D	3,077 (80)	20,834 (81)	1.01 (0.93–1.10)	0.96 (0.88–1.03)
2,4,5-T	938 (27)	5,182 (23)	1.25 (1.14–1.37)	1.19 (1.10–1.30)
2,4,5-TP	400 (12)	2,081 (9)	1.06 (0.94–1.19)	1.01 (0.90–1.13)
Thiocarbamate
Butylate	1,226 (36)	7,467 (33)	1.12 (1.03–1.21)	1.05 (0.98–1.13)
EPTC	769 (20)	5,167 (20)	1.08 (0.99–1.17)	1.04 (0.96–1.13)
Triazine/Triazinone
Atrazine	2,915 (76)	19,610 (76)	1.02 (0.94–1.10)	0.96 (0.89–1.03)
Cyanazine	1,575 (41)	10,967 (43)	1.03 (0.95–1.10)	1.00 (0.93–1.07)
Metribuzin	1,616 (47)	10,756 (48)	0.99 (0.91–1.08)	0.97 (0.90–1.05)
Other Herbicides
Chlorimuron Ethyl	1,323 (38)	8,903 (40)	0.98 (0.91–1.05)	0.94 (0.88–1.00)
Dicamba	1,963 (52)	14,103 (55)	1.00 (0.93–1.09)	0.97 (0.90–1.05)
Glyphosate	3,259 (85)	22,042 (86)	0.88 (0.80–0.96)	0.88 (0.80–0.95)
Imazethapyr	1,592 (42)	11,936 (47)	0.96 (0.88–1.04)	0.93 (0.86–1.00)
Paraquat	888 (26)	4,871 (22)	1.05 (0.97–1.15)	0.98 (0.91–1.07)
Petroleum Oil/distillates	1,525 (44)	10,584 (47)	0.90 (0.84–0.97)	0.89 (0.83–0.95)
Fungicides
Benomyl	384 (11)	2,124 (10)	1.04 (0.91–1.19)	0.99 (0.87–1.11)
Captan	463 (12)	3,306 (13)	0.94 (0.85–1.04)	0.95 (0.86–1.04)
Chlorothalonil	328 (9)	1,944 (8)	0.91 (0.79–1.04)	0.91 (0.80–1.04)
Maneb	381 (11)	2,124 (10)	0.97 (0.85–1.10)	1.01 (0.89–1.13)
Metalaxyl	902 (26)	4,948 (22)	1.03 (0.94–1.13)	1.02 (0.93–1.11)
Ziram	61 (2)	298 (1)	1.18 (0.92–1.52)	1.17 (0.92–1.48)
Fumigants
Aluminum Phosphide	193 (6)	1,225 (5)	0.99 (0.86–1.15)	0.94 (0.81–1.08)
Carbon tetrachloride/disulfide	234 (7)	1,400 (6)	1.16 (1.02–1.33)	1.11 (0.98–1.26)
Ethylene Dibromide	199 (6)	1,080 (5)	1.00 (0.86–1.15)	0.97 (0.84–1.11)
Methyl bromide	665 (17)	3,584 (14)	0.95 (0.85–1.06)	0.96 (0.86–1.07)

^aReported at enrollment, 1999–2003, or 2005–2010, using the most recent data prior to diabetes diagnosis or end of follow-up (non-cases).

^bObservations weighted inversely proportional to propensity scores for baseline data.

^cRelative risk (RR) and 95% confidence interval (CI) calculated in log-binomial models adjusting for sex/gender, race/ethnicity, age at end of study (+age-squared), state, educational attainment, smoking status, family history of diabetes, and correlated pesticides (R²>0.40).

^dRelative risk (RR) and 95% confidence interval (CI) further adjusting for BMI at enrollment and weight gain during follow-up prior to diabetes diagnosis.

3.2. Exposure-response

Frequencies and RR across tertiles of IWLD use are shown in Supplemental Table 3, including models adjusting for BMI and weight change, which showed no substantial confounding. Those with ≥ 1 tertile with a 95%CI excluding the null are shown in Fig. 1.

No OC insecticides showed a monotonic exposure–response; RR were largest in the lowest tertile for DDT (_T1RR=1.41; 95%CI 1.21–1.63), dieldrin (_T1RR=1.47; 95%CI 1.06–2.04), lindane (_T1RR=1.32; 95%CI 1.12–1.57), and toxaphene (_T1RR=1.40; 95%CI 1.16–1.69). For the OPs and carbamates, risk was greater for the lowest tertile of diazinon use (_T1RR=1.24; 95%CI 1.08–1.43), while positive trends were seen for malathion (_T3RR=1.13; 95%CI 1.02–1.25; p-trend=0.025), phorate (_T3RR=1.22; 95%CI 1.09–1.39; p-trend=0.001), and carbaryl (_T3RR=1.26; 95%CI 1.11–1.43; p-trend=0.005). Negative exposure–response trends were seen for use of permethrin/pyrethroids (_T3RR=0.82; 95%CI 0.70–0.96 for permethrin use on livestock, p-trend=0.005, and _T3RR=0.81; 95%CI 0.69–0.95 for permethrin/pyrethroid use on crops, p-trend=0.012).

No exposure–response was seen for 2,4,5-T (_T1RR=1.28; 95%CI 1.08–1.52) and 2,4,5-TP (_T1RR=1.35; 95%CI 1.04–1.76). Positive trends were seen for butylate (_T3RR=1.26; 95%CI 1.10–1.44; p-trend=0.001), metribuzin (_T3RR=1.16; 95%CI 1.02–1.32; p-trend=0.022), and chlorimuron ethyl (_T3RR=1.16; 95%CI 1.02–1.31; p=0.033). A negative trend was seen for glyphosate (_T3RR=0.77; 95%CI 0.70–0.86, p-trend<0.001). While none of the tertiles excluded the null for alachlor and atrazine, we saw positive trends (p=0.045 and 0.009, respectively).

3.3. Sensitivity analyses

Most associations in Tables 2 and 3, also met an FDR-adjusted p-value of <0.05, except for chlordane (Supplemental Table 4). We saw no substantial differences among those without weight gain (Supplemental Table 5, adding those with imputed BMI (Supplemental Table 6), nor excluding those with a diabetes family history (Supplemental Table 7). Among those with complete follow-up (Supplemental Table 8), RRs were elevated for DDT through the 2013–2015 follow-up and the RR for 2,4,5-TP was greatest in 2005–2010, while in 2013–2015 and 2018–2021, associations emerged for butylate (positive) and glyphosate (negative). Calculated HR using the Fine-Grey models accounting for competing causes due to mortality, for IWLD OC insecticides and phenoxy herbicides, 2,4,5-T and 2,4,5-TP, showed similar patterns of non-monotonic exposure response or stronger HR in the lowest tertile of use (Supplemental Table 9).

We explored differences stratified by median enrollment age of diabetes cases (50 years) and overweight/obese BMI (Supplemental Tables 10 and 11; Fig. 2). Most associations were more apparent among those <50 years, except for DDT, 2,4,5-T, and butylate, with similar RR in by age, and dieldrin, with an elevated RR for ages ≥50 years. Some RR were greater among overweight/obese participants, including aldrin, chlordane, heptachlor, carbaryl, and carbofuran, while RRs for DDT were the same regardless of BMI. The inverse association for glyphosate was only seen in those with BMI<25 mg/kg² (RR=0.52; 95%CI 0.39–0.68). Further stratification of overweight and obese BMI was uninformative (Supplemental Table 12).

Fig. 2.

Diabetes risk in licensed private pesticide applicators in the Agricultural Health Study: ever use of select pesticides at enrollment, 1999–2003, or 2005–2010 using the most recent available data prior to reported diabetes or end of follow-up (non-cases), stratified by median enrollment age of cases (50), and BMI (overweight/obese). Relative risk (RR) and 95% confidence intervals (CI) were calculated in log-binomial models adjusting for sex/gender, race/ethnicity, age at end of study (+age-squared), state, educational attainment, smoking status, family history of diabetes, and correlated pesticides (R²>0.40). Models are weighted inversely proportional to propensity scores for baseline only information.

4. Discussion

4.1. Overview and specific findings

Pesticides, especially persistent organic compounds, may contribute to greater diabetes risk (Taylor et al., 2013). In this analysis of licensed private pesticide applicators in the AHS cohort, with >3,800 incident cases, we found consistent evidence that greater diabetes risk was associated with using most OC insecticides. Risk was also elevated for ever-use of phorate and carborfuran, and higher IWLD use of malathion and carbaryl. Greater diabetes risk was associated with phenoxy herbicides, 2,4,5-T and lower IWLD use of 2,4,5-TP, and higher IWLD use of other herbicide types, butylate, metribuzin, and chlorimuron ethyl, and ever-use of the fumigant carbon tetrachloride/disulfide. While confirming some previously identified associations with diabetes in early follow-up (Montgomery et al., 2008), our findings provide new insights based on a larger number of cases and longer follow-up, cumulative exposure data weighted for factors impacting exposure intensity, adjustment for correlated pesticides, and propensity weights to account for loss to follow-up. We also saw no evidence of substantial confounding (or potential mediation) by BMI or weight gain.

Diabetes risk was associated with all OC insecticides examined, based on ever use or at least one of the cumulative IWLD tertiles, including in sensitivity analyses accounting for survival bias. The greatest risk appeared to be associated with DDT use, across all exposure levels. Findings on diabetes and measured DDT and DDE levels have been inconsistent across populations and study designs (Magliano et al., 2021; Jaacks et al., 2019; Mendes et al., 2021; Turyk et al., 2009; Son et al., 2010; Grant-Alfieri et al., 2024). Widely used in the U.S. from the late 1940’s until 1972, DDT exposure persists through environment media and in malaria prevention. We confirmed prior findings for aldrin, chlordane, and heptachlor, and saw new associations with dieldrin, lindane, and toxaphene.

Dieldrin is one of 4 cyclodiene OC insecticides (aldrin, dieldrin, chlordane, and heptachlor- all banned for agricultural use in the U.S. by the late 1980’s) that have been associated with diabetes or metabolic risk in other studies (Evangelou et al., 2016; Mendes et al., 2021). In the U.S. from 1999 to 2006, diabetes was associated with higher measured levels of heptachlor epoxide (an oxidation product of heptachlor), but not dieldrin (Everett and Matheson, 2010; Patel et al., 2010). We also saw a novel association of diabetes with lindane [γ-Hexachlorocyclohexane (HCH)]. In other studies, adipose γ-HCH levels were associated with higher fasting glucose levels in males (Reina-Pérez et al., 2023), and serum HCHs (including biproducts found in γ-HCG) were prospectively associated with diabetes risk in women (Zong et al., 2018). Agricultural lindane use, starting in the 1940’s, was restricted in 1970’s and cancelled by 2006. Until recently, lindane remained a clinical treatment for lice and scabies, and HCG biproducts from lindane production are a major environmental contaminant (Vijgen et al., 2011). Toxaphene is a complex mixture of >1000 chlorinated terpenes widely used in the 1970’s to 1980’s until banned in 1990, which remains a contaminant in superfund sites throughout the U.S. (Martyniuk et al., 2020). We identified no human studies on toxaphene-associated diabetes risk. Many OC insecticides remain in use in other countries, have long half-lives in the body, are stored in fat and magnify up the food chain, and otherwise persist in the environment, exposing populations through various pathways (Vasseghian et al., 2021; Xie et al., 2024; van den Berg et al., 2025).

Greater diabetes risk was associated with of 3 OP insecticides (diazinon, malathion, and phorate). Previously in AHS applicators, diabetes was associated with both diazinon and phorate, with a non-significantly elevated OR for malathion (Montgomery et al., 2008). Phorate, registered in 1959, is still used in U.S. agriculture, despite being considered highly hazardous, while diazinon, registered since 1956, was widely used until canceled in 2004. Malathion, registered in 1956, is considered to have low acute toxicity and is used widely in agricultural, residential, and public health settings. We only saw an association for higher IWLD malathion use. Measured malathion levels were associated with insulin resistance in non-diabetic Egyptian farmers and malathion metabolites have been associated with insulin resistance in non-diabetic U.S. adults (Zhu et al., 2025; Raafat et al., 2012). We did not replicate prior AHS findings for other OP insecticides (chlorpyrifos, coumaphos, and terbufos) (Montgomery et al., 2008). We observed new findings of greater diabetes risk for use of carbamate insecticides, carbofuran, and carbaryl. Carbaryl and carbofuran were approved in 1959 and 1969, respectively; carbofuran use banned in the U.S. by 2009, while carbaryl, considered moderately toxic, remains widely used in agricultural, residential, and other settings.

We also saw evidence of diabetes associations with the phenoxy herbicides 2,4,5-T and 2, 4, 5-TP. Developed during the late 1940 s, both may be contaminated during production with low levels with dioxins – a potent endocrine disruptor (Vandenberg et al., 2012). Both are banned for use in the U.S., first 2,4,5-T in 1970 (for all uses except rice) and then both by 1985. Findings on dioxins and diabetes risk vary (Goodman et al., 2015; Seo et al., 2024). Diabetes was also associated with the herbicide butylate; Introduced in 1967, butylate has low mammalian toxicity and continues to be used in agricultural settings. Diabetes also was associated with higher tertiles IWLD use of metribuzin and chlorimuron ethyl. Both are considered to have low toxicity, and were introduced in 1973 and 1985, respectively; neither were associated with diabetes early in follow-up (Montgomery et al., 2008), and we found no studies on diabetes or metabolic risk factors in humans. Finally, diabetes risk was associated with ever use of the fumigant carbon tetrachloride/carbon disulfide, used as an insecticide starting in the 1930’s through the 1970’s and phased out as a fungicide by 1980 (Fishbein, 1976; Bakke et al., 2007). Carbon tetrachloride, historically used as a solvent, remains a widespread contaminant in hazardous waste sites. Limited evidence suggests occupational exposure to carbon disulfide may impact glucose tolerance and metabolism (Rich et al., 2016; Franco et al., 1978). We did not replicate prior AHS findings for alachlor or cyanazine (Montgomery et al., 2008).

We did not hypothesize decreased pesticide-associated diabetes risk but saw inverse associations and negative exposure–response trends for permethrin/pyrethroids and glyphosate. Glyphosate was previously inversely associated with diabetes in early follow-up (Montgomery et al., 2008), but there is little other evidence in humans. Introduced in 1979, permethrin and other pyrethroids are considered moderately toxic. A cross-sectional study of U.S. adults showed that having diabetes was associated with higher levels of the pyrethroid metabolite 3-phenoxybenzoic acid, but not others (Park et al., 2019). Approved for use in the U.S. in 1974, and having low acute toxicity, agricultural glyphosate use rapidly increased through the mid-1990’s, with the introduction of genetically modified glyphosate-resistant crops. Cross-sectional analyses in the U.S. have shown that urinary glyphosate levels were positively associated with metabolic syndrome and insulin resistance (Otaru et al., 2024; Dong et al., 2024; Feng et al., 2025). We also saw an inverse association for petroleum oil/distillates use as an herbicide, but no exposure response. Interpreting findings for petroleum oil/distillates is complicated by its diverse historic and contemporary uses as an insecticide or herbicide or mixed as an adjuvant with other pesticides (Jungers et al., 2022). Notably, these inverse associations were seen primarily among applicators <50 years of age at enrollment, who may not have had sufficient time to develop diabetes during follow-up. Glyphosate was inversely associated with diabetes only among those who were not overweight/obese, also at lower risk of developing diabetes, and later in follow-up among those with complete data. Exposure misclassification is possible, with likely use in 2013–15 and 2019–21, when specific pesticide data were not collected. Also, we did not examine other newer pyrethroids reported in 1999–2003 or 2005–2010. Reverse causality is possible if older/overweight participants used less of these pesticides, though we saw no difference adjusting for BMI and weight gain. Thus, these results warrant cautious interpretation.

4.2. Potential mechanisms and susceptibility

Besides the neurotoxic effects of insecticides, OC exposure may increase oxidative stress, reduce the ability of pancreatic cells to secrete insulin, induce mitochondrial dysfunction or, for toxaphene, down-regulate insulin-related glycogenesis gene-expression pathways in the liver (Magliano et al., 2021; Martyniuk et al., 2020; Bresson and Ruzzin, 2024; Ge et al., 2026). Our findings for OCs and 2,4,5-T and – TP are consistent with endocrine disrupting effects and the low-threshold characteristic of endocrine disruptors (Vandenberg et al., 2012). The lack of an exposure response may also be due to mixtures of pesticides with opposing-effects (Lee et al., 2018; Taylor et al., 2013; Mrema et al., 2013; Park et al., 2025).

Both OP and carbamate insecticides inhibit acetylcholinesterase (ACh) activity (irreversible and reversible, respectively). In a prospective study of Ugandan farmers, less erythrocyte ACh inhibition was associated with lower average glucose levels (Hansen et al., 2020). Experimental studies reported diazinon effects on glucose tolerance and homeostasis in rats (Pakzad et al., 2013; Ueyama et al., 2008). Phorate induced hypoglycemia in mice (Cao et al., 2022), and a meta-analysis of experimental studies suggested chronic malathion exposure could lead to increased blood glucose levels (Ramirez-Vargas et al., 2018). Carbofuran and carbaryl may also impact diabetes risk through pathways involving the neurohormone melatonin. Both can bind to melatonin receptors and share an aromatic ring system (Popovska-Gorevski et al., 2017), which is missing in aldicarb (not associated with diabetes); and variants in the melatonin receptor 1b gene have been shown to effect fasting glucose levels and diabetes risk (Karamitri and Jockers, 2019). Carbon tetrachloride/disulfide has been used in preclinical models of liver damage, which may impact peripheral insulin levels and glucose regulation (Guerra and Gastaldelli, 2020). We found no experimental evidence consistent with the observed inverse associations of diabetes with glyphosate and permethrin/pyrethroids. Instead, evidence suggests potential elevated risk due to oxidative stress, hyperglycemia, insulin resistance, and adiposity (Riechelmann-Casarin et al., 2025; Jayaraman et al., 2023; Lin et al., 2025; Kim et al., 2014).

5. Strengths and limitations

In this prospective study, self-reported exposure data were collected prior to diagnosis, and analyses included cumulative exposure based on IWLD, reflecting exposure intensity by incorporating application techniques and use of personal protective equipment. Studies show relatively high accuracy and reliability of self-reported pesticide data in the AHS (Blair et al., 2002; Hoppin et al., 2002), and good performance of the IWLD algorithms relative to measured exposures (Thomas et al., 2010; Blair et al., 2011; Hines et al., 2011). Non-differential exposure misclassification due to recall or reporting errors remain plausible, biasing findings toward the null. For 28 pesticides, including pesticides banned at enrollment (e.g., OCs and phenoxy herbicide 2,4,5-T) and others (e.g., malathion and carbaryl), IWLD were based on data collected in the take-home questionnaire completed by less than half of enrolled applicators. Bias is possible if specific pesticide were strongly related to both missing IWLD data and diabetes. Our exposure estimates only reflect potential for internal dose. Nonetheless, compared to studies with measured levels, cumulative IWLD reflect exposure variation that might be missed in spot samples, which limit interpretation due to a lack of data on the timing of sampling relative to external exposures, pesticide half-life, and stage of disease pathogenesis. Data on pesticides reported at enrollment were updated through 2010, but we did not evaluate exposures during follow-up or newer pesticides, some of which may influence glucose metabolism and insulin resistance, e.g., neonicotinoids (Vuong et al., 2022). Still, exposures decades earlier remain relevant, as diabetes typically develops slowly, with years of undetected metabolic disfunction and pre-diabetes (Owora et al., 2022).

Incident diabetes was self-reported, which was shown to have high sensitivity and specificity compared to medical records in post-menopausal women (Jackson et al., 2014). Potential cases who later refuted their diagnosis (i.e., reported “no diabetes” in a subsequent follow-up survey) were excluded from the study sample, but were a small fraction compared to those who did not. About 28% of diabetes cases in the general U.S. population are not diagnosed (based on self-report) and would have been included as non-cases. Adult Type 2 diabetes is a heterogenous condition, reflecting major contributing causes, such as obesity and aging, genetic or environmental factors, and latent autoimmunity (Buzzetti et al., 2017; Schrader et al., 2022; Suzuki et al., 2024). When stratified by enrollment age and BMI, many observed RR were more apparent among younger participants, and in those who were overweight/obese. While uncommon, latent adult autoimmune diabetes may be over-represented among younger cases who were not over-weight/obese. Cases in young adults (<25 years, possible Type 1 diabetes) were excluded, but we lacked complete data on insulin dependence at older ages. Diabetes incidence in our study may be lower than in males in the general population due to healthy lifestyle, supported by prior findings of lower mortality overall and due to diabetes in the cohort (Shrestha et al., 2019).

Compared to prior analyses in AHS applicators (Montgomery et al., 2008), our analyses adjusted for correlated pesticides, as well as state, which is related to type of farming and diversity of pesticides used, and education, as a proxy for socioeconomic status and pesticide use practices. We also explicitly ruled out confounding by BMI and weight gain; however, we cannot rule out mediation effects. A prior AHS analysis showed limited evidence that specific pesticide use was associated with weight gain in the first 5 years of follow-up, except for atrazine (LaVerda et al., 2015). In the current study, the association of atrazine with diabetes was null, and tertile analyses did not reach the threshold for reporting. Future research on the pesticides, BMI and weight gain, and potential mediation effects on diabetes risk in the AHS may be warranted but is beyond the scope of the current study. While unlikely confounders of the association between pesticides and diabetes, diet and physical activity may influence obesity and could contribute to misclassification of overweight BMI in those with greater muscle mass, but we lacked these data on a majority of the study sample.

We did not examine mixtures, as temporal relationships of pesticide use were not known. Recent prospective data suggested mixtures of measured herbicides and OP insecticides were associated with diabetes risk in a Chinese rural population, while another study of mid-life women in the U.S. reported no association of diabetes risk with mixtures of persistent organic pollutants (Grant-Alfieri et al., 2024; Ma et al., 2023). Multiple comparisons can result in chance findings, but most of our results were upheld after correcting for multiple testing using the FDR. We lacked power to conduct some sensitivity analyses, especially among those who were not overweight or obese. Some participants were lost to follow-up in later phases; but bias cannot be assessed due a lack of data on diabetes among non-respondents. Most participants lost to follow-up tended to be younger, while those who died during follow-up were older and more likely to smoke. Older participants would have had greater opportunity for long-term use of banned pesticides. However, Fine-Grey analyses suggested limited bias due to mortality during follow-up or competing causes as an explanation for non-monotonic or low threshold exposure–response patterns seen with organochlorine insecticides and phenoxy herbicides 2,4,5-T and 2,4,5-TP. Exposures in AHS applicators (mostly males) are greater than the general population, limiting generalizability to other non-occupationally exposed populations and females. Notably, several of the current findings (e.g., dieldrin, carbofuran, phenoxy herbicides, and butylate) are consistent with prior findings in AHS spouses for Type 2 or gestational diabetes (Saldana et al., 2007; Starling et al., 2014).

6. Conclusions

Our findings support the idea that diabetes risk is greater for those exposed to persistent OC insecticides and phenoxy herbicides contaminated with dioxins during manufacturing. Novel findings for other pesticides warrant replication in other populations, especially those pesticides with ongoing agricultural use and widespread residential or public health uses.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.envint.2026.110082.

Data availability

The authors do not have permission to share data.