Epidemiologic studies of glyphosate and non-Hodgkin's lymphoma: A review with consideration of exposure frequency, systemic dose, and study quality

Abstract

I reviewed the epidemiologic literature for glyphosate and non-Hodgkin's lymphoma (NHL) in the context of the frequency of exposure in each epidemiologic study, systemic dose from biomonitoring studies of applicators, and aspects of study quality. Nine studies were identified, 7 case control and 2 cohort, by a literature search and a review of reference lists from published studies and recent regulatory evaluations. All but one study involved exposure scenarios that were so infrequent that they are not credible for cancer causation. Most studies failed to address potential confounding from other pesticides. Only one study – the US Agricultural Health Study (AHS) – included individuals with relatively frequent exposure to glyphosate and involved comprehensive statistical analyses to address potential confounding by personal factors and other pesticide exposures. The AHS did not find an association between glyphosate and NHL, even among the most frequently exposed participants (≥ 109 days of use) (RR = 0.80, 95% CI 0.60, 1.06). These findings are consistent with observations that glyphosate systemic doses from agricultural applications are many orders of magnitude less than daily lifetime doses considered by regulatory agencies to impart no excess risk of deleterious health effects, even for sensitive subpopulations.

Introduction

Glyphosate [N-(phosphonomethyl)glycine] is currently the world's most widely used herbicidal agent. It was first approved for use in the United States (US) in 1974. As a non-selective herbicide introduced into an industry built around selective herbicides, it took time for glyphosate to be integrated into agricultural practice. By 1987, glyphosate ranked 17th in pounds applied in the United States and it ranked 5th in that category by 1997. [1].

The focus of this literature review is on the epidemiologic studies for glyphosate and NHL. Meta-analyses on this topic will be considered briefly, but not as primary data. The review will draw context from biomonitoring assessments of the systemic dose received from real world glyphosate applications, the frequency of exposure in the available epidemiologic studies, and epidemiologic study quality considerations.

Glyphosate regulatory evaluations

On an acute basis, glyphosate is relatively non-toxic. The acute oral lethal dose (LD) 50 in rodents is >4000 mg/kg. [48] Chronic rodent carcinogenicity studies and genotoxicity studies have been conducted for registration purposes by 8 registrants at dose levels up to and exceeding 1000 mg/kg/day. Prior to 2015, regulatory reviews worldwide concluded that glyphosate was not mutagenic, carcinogenic, teratogenic, or a reproductive or developmental toxicant [e.g., [63,68]]. In 2015, the International Agency for Research on Cancer [36] classified glyphosate as a probable human carcinogen based on the judgment of the relevant working groups that there was sufficient evidence of cancer in experimental animals and strong evidence of genotoxicity. The epidemiology evidence was considered limited, based on some case-control studies showing positive associations for non-Hodgkin's lymphoma (NHL). The IARC review stimulated numerous regulatory reviews worldwide to reevaluate their previous glyphosate assessments. The reevaluations did not agree with the IARC toxicology assessment. [8,[26], [27], [28],31,51,54,65,69,70].

Integrating results from a range of toxicology studies, regulatory agencies concluded that glyphosate is not carcinogenic or mutagenic and established an acceptable daily intake or chronic reference dose for glyphosate ranging from 0.3 to 1 mg/kg/day. [28,37,54,64] The acceptable daily intake or chronic reference dose refers to a daily internalized amount of glyphosate for individuals, including those in sensitive subpopulations, that is likely to be without an appreciable risk of deleterious effects during a lifetime.

Glyphosate applications and systemic dose

There is a difference between reported pesticide use (viz., exposure) and systemic dose (the quantity of a substance that is internalized). Exposure is a poor surrogate for systemic dose for many pesticides, including glyphosate. [3,9] Systemic dose is much more relevant for epidemiologic research than exposure and is a consequence of the physical/chemical properties of a pesticide and the circumstances of the agricultural or other application.

Glyphosate is usually formulated as the isopropylamine salt, which has an extremely low vapor pressure of 1.6 × 10⁻⁸ mm Hg. [62] In regulatory evaluations, dermal contact has been considered to be the primary route of exposure that may be internalized. Dermal penetration experiments, where glyphosate was left undisturbed on skin surfaces of experimental animals and on human skin in vitro, indicate a percutaneous absorption of <2%. [67].

In humans, glyphosate once internalized has a short half-life, estimated in a recent human study to be between 5 and 10 h. [20] It is excreted almost completely in urine as the parent compound. It is well established that glyphosate systemic dose from an application can be reliably estimated through urine biomonitoring of the parent compound. [49,60].

The most comprehensive biomonitoring data for glyphosate were collected as part of the Farm Family Exposure Study (FFES). [44] The FFES included 48 farmers who applied glyphosate on their farms in South Carolina and Minnesota. The farmers provided 24-h urine samples 24 h before the on-study application, 24 h the day of application (starting with the initial handling of glyphosate) and 24-h a day for 3 subsequent days after the application.

Urine samples were evaluated for daily concentration with a 1 part per billion (ppb) limit of detection. The glyphosate applications ranged from 10 acres to >400 acres: 16 were on 10 to 45 acres, 16 were on 46 to 124 acres, and 16 were on 125 to 439 acres. Sixty percent of the applicators had quantifiable glyphosate in their urines, while 40% did not have detectable urinary glyphosate. [4] The 40% included 9 farmers who made applications of >100 acres. The distribution of glyphosate urinary concentrations was highly positively skewed with a 3 ppb geometric mean (GM). The GM systemic dose was approximately 0.0001 mg/kg and approximately 90% of the values were not appreciably different than the GM. [5].

Niemann et al. and Solomon reviewed the glyphosate biomonitoring literature. [49,60]. Across all glyphosate biomonitoring studies, the highest estimated systemic dose from an agricultural application was 0.004 mg/kg - an applicator from the FFES. The median systemic dose per application was 0.0001 mg/kg.

Study quality considerations

Regulatory agencies in the US and Europe have specified quality considerations for pesticide epidemiologic studies in recent reviews. Criteria from the recent review of glyphosate by the US Environmental Protection Agency [65] are paraphrased in Table 1. These considerations focus on the likely accuracy of the exposure and outcome data and the potential for selection bias, information bias and confounding.

Table 1.

Quality factors adapted from USEPA's 2017 Glyphosate Review.

Factor	High	Medium	Low
Exposure	Firsthand chemical specific questionnaire information from study subjects	Chemical specific questionnaire information from subjects or proxy respondents	Low quality interview; information is not specific for individual pesticides or not collected at the individual level; focus on ever/never exposure
Outcome	Diagnoses from high quality cancer registries or hospitals with pathologic confirmation	Medical sources without histopathologic confirmation	Other sources without histopathologic verification
Statistical analysis	Clearly reported accepted methods. Comprehensive analysis approach. Sample size is adequate for informative analyses.	Less comprehensive approach. Sparse data for key analyses. Analytic choices that lose information, unclear reporting of methods.	Minimal attention to statistical analyses presented; comparisons not performed or described clearly.
Confounder control	Good control for personal factors and other relevant exposures.	Control for a limited number of personal factors and relevant factors.	Multivariate analyses not performed or not implemented correctly.
Other bias	Major sources of bias unlikely or addressed through study design or analysis.	Potential sources of bias present that may influence the magnitude of risk estimates, but not the direction of estimates.	Major study biases present with high potential to impact effect estimates.

One other quality consideration that is extremely relevant is whether the population(s) chosen for study had frequent enough exposure to be informative. In chemical carcinogenesis, researchers usually seek to study populations with frequent exposure. For example, Linet et al., in a study of benzene workers in China, specified: “… statistical analysis excluded workers exposed to benzene for less than 6 months … to reduce the potential for including serious health outcomes unlikely to be due to benzene exposure.”[43] Studies that are focused on a very few days of glyphsate use over a lifetime are not likely to be informative.

Methods

I conducted a literature search to identify all epidemiologic studies that addressed the possible relationship between glyphosate exposure and NHL with individual level direct exposure and outcome data. The exposure data could come from self-reports or assignments of exposures by research personnel based on individual level occupational information from study participants. I excluded studies where glyphosate exposure was presumed for those who did not work directly on a glyphosate application.

I searched PubMed on September 28, 2022 using the key words “glyphosate” and “cancer” and “glyphosate” and “lymphoma.” The former yielded 209 citations and the latter yielded 19 citations. I also reviewed the reference lists of recent reviews, both published in journals and from regulatory bodies. In total, focusing on the latest iteration for specific study populations, I identified 7 case control studies and 2 cohort studies (Table 2, Table 3). The cohort studies and the case control studies are reviewed in separate sections in this article because the quality considerations are fundamentally different by study design as implemented in the glyphosate/NHL literature. In some instances, to fill in analysis gaps, I calculated a weighted average of a study's results across exposure quartiles or across studies using the meta-analysis module in Stata 17. [61].

Table 2.

Relevant primary glyphosate NHL epidemiologic studies.

Author(s)(year)	Study design	Country	Case identification	Prior Related research
McDuffie et al. [45]	Case control	Canada	1991–1994	n/a
Hardell et al. [33]	Case control	Sweden	1987–1990	[50]; [74]
De Roos et al. [24]	Case control	U.S.	1979–1981 (Kansas) 1981–1983 (Iowa, MN) 1983–1986 (Nebraska)	[35] [14] [71]
Eriksson et al. [29]	Case control	Sweden	1999–2002
Orsi et al. [52]	Case control	France	2000–2004
Cocco et al. [18]	Case control	6 EU countries	1998–2004
Meloni et al. [46]	Case control	Italy	2012–2016
Andreotti et al. [7]	Cohort	U.S.	1993–2012/3	[22]
Leon et al. [42]	Cohort	US France, Norway	1993–2010 US 2005–2009 France 1993–2012 Norway	[7]

Table 3.

Relevant details glyphosate NHL epidemiologic studies.

Authors	Subjects	Participation	Proxy respondents	% exposed; days glyphosate use among exposed	Exposure metric	RR	95% CI
McDuffie et al.	517 cases	Cases 67%	Cases 21%	Cases 10%	≥ 1 day/lifetime	1.2	0.8, 1.7
	1506 controls	Controls 48%	Controls 15%	Controls 9%
				Median ~ 10 days	≤ 2 days/year	1.0	0.6, 1.6
					> 2 days/year	2.1	1.3, 2.7
Hardell et al.	515 cases	Cases 91%	Cases 44%	Cases 1.6%	≥ 1 day	3.0	1.1, 8.5
	1141 controls	Controls 84%	Controls 44%	Controls 0.7%		1.9⁎	0.6, 6.2
				Unknown frequency of use, likely very few days
De Roos et al.	650 cases	91% Nebraska	Cases 31%	Cases 5.5%	≥ 1 day/lifetime	2.1⁎, !	1.1, 4.0
		89% Iowa	Controls 40%	Controls 3.2%		1.6⁎, !!	0.9, 2.8
		89% Minnesota
		96% Kansas
				< 7 days median use
	1933 controls	85% Nebraska
		78% Iowa
		78% Minnesota
		93% Kansas
Eriksson et al.	910 cases	Cases 81%	none	Cases 3%	≥ 1 day/lifetime	2.0	1.1, 3.7
	1016 controls	Controls 65%		Controls 2%		1.5⁎	0.8,2.9
				> 10 days median use	≥ 1 day/lifetime
					≤ 10 years latency	1.1	0.2, 5.1
					> 10 years latency	2.3	1.2, 4.4
					< 10 days of use	1.7	0.7, 4.1
					≥ 10 days of use	2.4	1.0, 5.4
Orsi et al.	244 cases	Cases 96%	None	Cases 5%	≥ 1 day/lifetime	1	0.5, 2.2
	436 controls	Controls 91%		Controls 6%
				Unknown frequency of use, likely very few days
Cocco et al.	2348 cases	Cases 88%	None	Cases < 1%	≥ 1 day/lifetime	3.1⁎⁎	0.6, 17.1
	2462 controls	Hospital controls 81%		Controls < 1%
		Population controls 52%
				Unknown frequency of use, likely very few days
Meloni et al.	867 cases	Cases 93%	None	Cases 2.4%	≥ 1 day/lifetime
	774 controls	Controls 62%		Controls 1.9%	High, medium, low confidence
						1.4	0.6, 2.9
				Medium & high confidence	Medium/high confidence
				Cases 1.0%		1.2	0.4, 1.5
				Control 0.8%
				Unknown frequency of use, likely very few days
Andreotti et al.	54,251 licensed pesticide applicators	Not applicable	none	83%	Intensity weighted days of use
					Quartile 1
				Median exposure 48 days	Quartile 2	0.8⁎	0.6, 1.2
					Quartile 3	0.8⁎	0.6, 1.1
				IQR 20 to 166 days	Quartile 4	0.9⁎	0.7, 1.2
						0.9⁎	0.6, 1.2
					Days of use:
					< 1 to 14 days
					14 to 38 days	0.8⁎	0.6, 1.0
					29 to 108 days	0.9⁎	0.7, 1.1
					≥ 109 days	0.9⁎	0.6, 1.1
						0.8⁎	0.6, 1.1
Leon et al.	316,210 agricultural workers	Not applicable	None	AHS 81%	≥ 1 day all subtypes	1.0^, ⁎	0.8, 1.2
				AGRICAN 37%	CLL/SLL	0.9⁎	0.7, 1.2
				NCAP 39%	FL	0.8⁎	0.5, 1.2
					MM	0.9⁎	0.7, 1.2
				Unknown frequency of glyphosate use	DLBCL	1.4⁎	1.0, 1.9

^⁎adjusted for other pesticides.

^!logistic regression estimate.

^!!hierarchical regression estimate.

^⁎⁎B-cell lymphoma.

^{^}In order: all NHL, chronic lymphocytic lymphoma/small lymphocytic lymphoma, follicular lymphoma, multiple myeloma/plasma cell leukemia, and diffuse large B-cell lymphoma.

Glyphosate/NHL case control studies

For time efficiency and cost reasons, most of the studies in the glyphosate/NHL literature are case control studies. It is a common misconception to think that case control studies are inferior to cohort studies. [56] A case control study can be thought of as a cohort study where the cases are identified during a specified time period and the controls are a sample of the concurrent source population that gave rise to the cases.[57] Properly designed and executed case control studies will produce valid results as will properly designed and executed cohort studies. However, it can be difficult in case control studies to select a control population that is representative of the population that gave rise to the cases and to collect self-reported pesticide and other exposure information that is unaffected by the cases' natural self-examination about what might have caused their grievous illness. In addition, most individuals exposed to glyphosate have had exposure to other pesticides and chronic exposures typical of agricultural work, so assessment of other agricultural exposures and control of confounding needs to be comprehensive in both case control and cohort studies.

McDuffie et al. conducted a multi-center case control study in Canada to evaluate the relationships between pesticide exposures, personal factors, and NHL. [45] Cases (n = 517) were identified during the period 1991–94 from provincial Cancer Registries, except in Quebec, where hospital ascertainment was used. Controls (n = 1506) were selected at random from the provincial Health Insurance records (Alberta, Saskatchewan, Manitoba, Quebec), computerized telephone listings (Ontario) or voters' lists (British Columbia). Participation was 67% for cases and 48% for controls. Pesticide exposure was determined through telephone interviews of study participants or their proxies (21% of cases, 15% of controls). Ten percent of cases reported ≥1 day of glyphosate use versus 9% of controls. The authors used conditional logistic regression to estimate odds ratios (ORs) and 95% confidence intervals (CIs). The OR for ≥1 day of reported glyphosate use during a lifetime was 1.2 (95% CI 0.8–1.7) controlling for age, province, and medical variables associated with NHL. Analysis by days of glyphosate use per year (none, ≤ 2 days/year, >2 days/year) showed ORs of 1.0, 1.0 (95% CI 0.6–1.6), and 2.1 (95% CI 1.3–2.7), respectively, controlling for age and province. The authors concluded that their findings by days of use provided evidence of a dose-response relationship for glyphosate and NHL.

Evaluation: This was a large case control study with reliable identification of NHL cases. Reported glyphosate use was relatively infrequent for cases (10%) and controls (9%). It is difficult to determine how many lifetime days of glyphosate use were experienced by exposed participants. Pahwa et al. reanalyzed this study, pooled with several US case control studies as previously reported by DeRoos et al., and found a median of 7 days of lifetime exposure among those exposed across both populations. [24,53].

Among the limitations of this study, participation was much lower for controls (48%) than cases (67%) and exposure information was provided by secondhand (proxy) sources for 1/5th of cases and 1/6th of controls. The analysis by days of use per year did not consider the number of years of glyphosate use, so there was no assessment of results with more frequent cumulative use of glyphosate. The authors concluded that their analysis by days of use per year demonstrated a dose-response relationship, but they did not actually do a statistical trend test and did not control for medical variables or for other pesticide exposures in this analysis. The subsequent pooled reanalysis by Pahwa et al. found that control for 2,4-D, dicamba, and malathion appreciably attenuated the OR estimate for 3 or more days per year of use (OR = 2.4, 95% CI 1.5 to 4.0) to OR = 1.7, (95% CI 1.0 to 2.9, p trend 0.2). [53] In McDuffie et al.'s overall statistical model, NHL was associated positively with a personal history of cancer, a history of cancer in first-degree relatives, and exposure to dicamba, mecoprop, and aldrin. [45] A personal history of measles and of allergy desensitization treatments was associated with a lowered risk. Glyphosate was not reported as a risk factor in the overall statistical model.

Crump conducted an analysis of the available glyphosate NHL studies to discern indirectly whether recall bias and/or selection bias were likely operative. [21] His thesis was that an indication of these biases was when appreciably >50% of ORs for non-glyphosate exposures were greater than the null value of 1.0. Crump found that 93% of the ORs were > 1.0 for the non-glyphosate pesticide analyses by McDuffie et al. [45], a marked skew toward positive results across pesticides in their analyses.

Hardell et al. [33] conducted a pooled analysis of two Swedish case control studies; one of NHL and the other of hairy cell leukemia (HCL) [33,50], respectively. The rationale for combining these two study populations was that modern classification for lymphomas includes HCL among NHL subtypes. The 404 NHL cases were males aged 25 and older, diagnosed in 1987–1990, and living in mid- and northern Sweden, drawn from regional cancer registries. Each case was matched on age and sex to two controls drawn from the National Population Registry. The 111 HCL cases were males diagnosed in 1987–1990, identified from the Swedish National Cancer Registry. Each HCL case was matched on age, sex and county to four controls drawn from the National Population Registry. A total of 515 cases and 1141 controls were included in pooled analyses of NHL and HCL. A questionnaire was completed by study subjects or next-of-kin regarding complete working history and exposure to various chemicals. Exposure to each chemical was dichotomized, with at least one lifetime day of pesticide use occurring a year or more before diagnosis being regarded as a qualifying exposure. Eight cases (1.6%) and 8 controls (0.7%) reported glyphosate exposure. Conditional logistic regression was used to estimate ORs and 95% CIs, adjusted for study (NHL versus HCL), study area, and vital status. In the analyses for specific pesticides, only subjects with no pesticide exposure were regarded as unexposed. Analysis for glyphosate exposure of ≥1 day in a lifetime, unadjusted for other pesticides, showed a positive association (OR 3.0, 95% CI 1.1–8.5). The multivariate OR adjusted for other pesticides was 1.9 (95% CI 0.6, 6.2).

Evaluation: This was a large case control study with reliable identification of NHL cases. Participation was reported to be 91% for cases and 84% for controls. Exposure frequency was extremely low, and the cumulative frequency of glyphosate use by cases and controls was not reported. The cumulative days of glyphosate use seem likely to be extremely low given the infrequency of any use for cases and controls (1.6% and 0.7%, respectively) and a comment in a later publication by these authors that glyphosate use was not common at the time of this earlier study. [29].

Forty four percent of the exposure information for cases and controls came from proxy respondents. Accordingly, the investigators should have presented results stratified by type of respondent. ORs based on proxy respondents can differ markedly from those based on self-respondents [38,41] and secondhand information is usually less accurate than primary information. Another limitation is the lack of clarity about how the authors sought to control confounding by other agricultural exposures. In fact, confounding by other pesticides could not have been controlled adequately because, as noted by Chang & Delzell, analyses for specific pesticides in this study excluded those with other pesticide exposures from the unexposed group. [16] This practice of “fine tuning” the unexposed group has been shown to distort exposure prevalences. [2] The numbers of cases and controls excluded from the glyphosate multivariate analysis were not discernable from the tables in this paper, but judging, for example, by the overall exposure frequencies for insecticides (22% of cases, 16% of controls), impregnating agents (20% of cases, 14% of controls) and 2,4-D (9% of cases, 6% of controls) the numbers of exclusions would have been substantial. As a practical matter, there were likely too few subjects reporting glyphosate exposure to support a multivariate pesticide analysis. Crump determined that 90% of the non-glyphosate ORs were >1.0, indicating a marked skew toward positive results, perhaps due to recall bias or uncontrolled confounding.

De Roos et al. [24] conducted a pooled analysis of several midwestern US NHL case-control studies of pesticides. [14,35,71] Cases from the Kansas study by Hoar et al. (n = 153) represented a random sample of cases diagnosed between 1979 and 1981 and selected from the Kansas Cancer Data Service. [35] Cases from the Nebraska study by Zahm et al. (n = 187) were diagnosed between July 1983 and June 1986 and were identified using the Nebraska Lymphoma Study Group as well as from area hospitals. [71] Cases from the studies in Iowa and Minnesota by Cantor et al. (n = 520) were diagnosed between 1981 and 1983 and were identified from the Iowa State Health Registry along with a surveillance system established in Minnesota. [14] Controls for these studies were randomly selected from population databases (e.g. Medicare, random digit dialing, and state mortality files for deceased cases) and frequency matched to cases on race, sex, age and vital status at time of interview. Cases and controls were interviewed, including next-of-kin when necessary, regarding use of pesticides and other known or suspected risk factors for NHL. Forty-seven pesticides were included in the analysis. The final analysis dataset included 650 (of 870, 75%) cases and 1933 (of 2569, 75%) controls after excluding study subjects with missing information or a “don't know” response for any one of the 47 pesticides. Exposure information was provided by proxy respondents for 31% of cases and 40% of controls. The exposure metric used in the analysis was ≥1 day of use during a lifetime for glyphosate and other pesticides. Two types of statistical models were used to estimate ORs and 95% CIs: (1) standard logistic regression and (2) hierarchical regression, wherein logistic regression estimates were adjusted in a second stage model based on a priori carcinogenic probability for specific pesticides as determined by external review bodies. For pesticides like glyphosate that were presumed to have a low probability of being carcinogenic, this second stage adjustment tended to draw positive associations toward the null. All analyses were adjusted for age and for the use of 46 other pesticides. Results for ≥1 day of glyphosate use during a lifetime showed an OR of 2.1 (95% CI: 1.1, 4.0) in the logistic regression controlling for other pesticides and a lesser association (OR 1.6, 95% CI: 0.9, 2.8) in the hierarchical regression.

Evaluation: This pooled analysis included a very large number of cases and controls. Cases were identified from reliable sources and controls were selected in a way to be representative of the population that gave rise to the cases. Reported exposure frequencies for ≥1 day of glyphosate use were 5.5% for cases and 3.2% for controls. No information was provided about the number of days of glyphosate use.

This study had a large amount of information from proxy respondents (31% for cases and 40% for controls). Results should have been presented separately by type of respondent. A limitation acknowledged by the authors was their reliance on a crude indicator of exposure – ≥ 1 day of glyphosate use during a lifetime with no consideration of the extent of use. It is noteworthy that the cases included in this pooled analysis were diagnosed very soon after glyphosate's initial approval in 1974: 16% of cases were identified during the period 1979–81, 67% of cases were identified between 1980 and 83, and 17% of the cases dated from the period 1983–86. The actual date of first glyphosate use for cases was not considered in the analysis, but the interval between first use and NHL diagnosis was very short for a high proportion of the cases and likely insufficient for a plausible induction-latent period [55]. Identifying cases so early in the glyphosate life cycle would also suggest very few days of exposure for exposed study participants.

A 2019 reanalysis by Pahwa et al. pooled the data from this study with data from the study by the McDuffie et al. [45,53]. The authors imputed missing data to include more study subjects in the analyses and they considered a variety of exposure metrics and addressed potential confounding by other pesticides in the pooled data set. The OR for ≥1 day of lifetime glyphosate use and NHL overall, adjusted for age, sex, state/province, lymphatic or hematopoietic cancer in a first-degree relative, proxy response, use of personal protective equipment, 2,4-D, dicamba, and malathion, was 1.1 (95% CI 0.8, 1.5). This logistic regression result differs appreciably from the logistic regression OR of 2.1 (95% CI 1.1, 4.0) in DeRoos et al. and differs minimally from the OR for any glyphosate use in McDuffie et al. of 1.2 (95% CI 0.8–1.7). [24,45] Weighted averages of ORs from the studies by McDuffie et al. [45] and De Roos et al. [22], using fixed and random effects models [24,45,61], were 1.4 (95% CI 1.0, 1.9) and 1.5 (95% CI 0.9, 2.5), respectively (my calculations, [61]). The disparity with the OR of 2.1 from DeRoos et al. is important since the DeRoos et al. OR of 2.1 has been used in meta-analyses. Results for NHL varied across exposure metrics in the pooled analysis. Longer duration of use (> 3.5 years) and the higher cumulative days of use (> 7 days) showed near null ORs of 0.9 (95% CI 0.6, 1.4) and 1.1 (95% CI 0.7, 1.8), respectively, while use >2 days/year was associated with an elevated OR of 1.7 (95% CI 1.0, 2.9). Results also varied across NHL subtypes, though the results were statistically imprecise due to small numbers.

Eriksson et al. conducted a population-based case control study of NHL and pesticide exposures in Sweden. [29] Cases and controls were recruited from among those enrolled by another research group working in Sweden. Cases were identified during the period 1999–2002 through physicians who diagnosed and treated NHL. All cases were histologically verified. Controls were randomly chosen from regional population registries and were frequency matched in 10-year age and sex groups. A total of 910 NHL cases and 1016 controls were included in the analyses. Participation was 81% for cases and 65% for controls (92% of controls from the parent case control study that had 71% control participation per Chang and Delzell). [16] All subjects received a mailed questionnaire focusing on work history and exposure to pesticides, solvents, and other chemicals. In univariate analyses for specific pesticides, only subjects with no pesticide exposure at all were regarded as unexposed, Unconditional logistic regression was used to calculate ORs and 95% CIs, adjusted for age, sex, and year of diagnosis. In the univariate analysis for ≥1 day of glyphosate use in a lifetime and NHL, the OR was 2.0 (95% CI 1.1–3.7) based on 29 (3%) exposed cases and 18 (2%) exposed controls. In a multivariate analysis for ≥1 day of glyphosate use, with adjustment for agents that had a statistically significant OR or with an OR > 1.5 and at least 10 exposed subjects, the glyphosate OR was 1.5 (95% CI 0.8, 2.9). The authors conducted additional analyses to assess the glyphosate/NHL relationship by days of use and latency. For those with <10 days of use (12 cases and 9 controls) the OR was 1.7 (95% CI 0.7, 4.1) and for those with ≥10 days of use (17 cases and 9 controls) the OR was 2.4 (95% CI 1.0, 5.4). Any exposure to glyphosate with a latency period of <10 years was associated with an OR of 1.1 (95% CI 0.2, 5.1) and the OR was 2.3 (95% CI 1.2, 4.4) for those with >10 years latency.

Evaluation: This was a large case control study with reliable identification of NHL cases. Participation was greater for cases (81%) than controls (65%). Ever use of glyphosate was infrequent for cases (3%) and controls (2%). Based on the analyses by days of use, the median exposure for the study subjects who reported glyphosate use appears to be slightly >10 lifetime days.

It is unclear from this publication whether the results for glyphosate by days of use and latency were adjusted for other pesticide exposures. There was a conflict between the language in the methods section and the footnote in the table where the results were detailed. EPA concluded that these analyses did not control for other exposures. [65] Control for other exposures in the multivariate analysis for any glyphosate use reduced the OR (2.0 to 1.5). Crump noted that there were positive associations for 90% of individual non-glyphosate pesticides in this study, indicative of a skew toward positive results in pesticide specific analyses. [21].

Orsi et al. reported a French hospital-based case control study of occupational exposure to pesticides and lymphoid neoplasms (including but not limited to NHL). [52] Incident histologically confirmed cases of NHL (n = 244) were identified from six French hospital center catchment areas between 2000 and 2004. Controls (n = 436) were selected from the same hospitals as cases, primarily from rheumatology and orthopedic departments. Controls were matched to cases by center, age (±3 years) and gender. Participation was 90% for cases and controls. Information from cases and controls was collected via a standardized self-administered questionnaire on socioeconomic status, family medical history, and lifelong residential and occupational histories. For those who had worked as a farmer or gardener for at least 6 months, a trained interviewer collected information about use of pesticides (insecticides, fungicides, and herbicides) and work on farms. ORs and 95% CIs were calculated using logistic regression. Results for ≥1 day of glyphosate use in a lifetime and NHL, controlling for age, study center and blue versus white collar employment, showed no association (OR.1.0, 95% CI: 0.5–2.2) based on 12 exposed cases (5%) and 24 exposed controls (6%).

Evaluation: This is a comparatively small case control study with reliable identification of NHL cases. The frequency of any reported exposure was low and there is no information about cumulative exposure to glyphosate.

The authors noted that there were problems with the pesticide questionnaire information. Most cases and controls who were administered the agricultural questionnaire were judged to need a re-interview and, of those, only 56% could be re-interviewed. Expert opinion at times filled in the gaps. Another important limitation is the likelihood that the control population was not representative of the population that gave rise to the cases. Controls were selected mainly from orthopedic and rheumatology departments and excluded those admitted for cancer or a disease directly related to occupation, smoking, or alcohol abuse. Crump determined that 76% of the non-glyphosate ORs were > 1.0, an appreciable skew toward positive results. Lastly, the glyphosate OR was not adjusted for other pesticides or other occupational exposures.

Cocco et al. reported results from the EPILYMPH case control study in six European countries, conducted in 1998–2004. [18] The study included 2348 incident lymphoma cases and 2462 controls. Controls were population-based in Germany and Italy, matched on gender, age (within five years) and residence area. Hospital controls were used in the Czech Republic, France, Ireland and Spain. The participation rate was 88% in cases, 81% in hospital controls, and 52% in population controls in Germany and Italy. [19] Trained interviewers conducted in-person interviews with a structured questionnaire regarding full time jobs held for a year or longer.

Subjects who reported having worked in agriculture were given a job-specific module inquiring in detail about tasks, kinds of crops, size of cultivated area, pests being treated, pesticides used, procedures of crop treatment, use of personal protective equipment, reentry after application and frequency of treatment in days/year. Specific pesticide exposures were determined based on a crop exposure matrix. Unconditional logistic regression was used to calculate ORs and 95% CIs, adjusted for age, gender, education and study center. Subjects unexposed to any pesticide were the referent category for all analyses. The authors reported an OR for glyphosate (ever/never exposure) and B-cell NHL of 3.1 (95% CI 0.6–17.1) based on four exposed cases and two exposed controls.

Evaluation: This was a very large case control study with accurate ascertainment of lymphoma cases. However, as the authors noted, there were too few exposed cases for any meaningful inference to be drawn. For glyphosate, < 1% of cases and controls were ever exposed. The number of days of glyphosate use was not specified, but presumably days of exposure were very few.

Other limitations include the possibility of selection bias (esp. the 52% participation for population controls versus 88% for cases) and confounding due to other pesticides not being controlled in the statistical analysis.

Meloni et al. investigated the association between occupational exposure to glyphosate and risk of lymphoma in a multicenter case-control study conducted in Italy. [46] Overall, 867 incident lymphoma cases and 774 controls participated in the study. Participation was 93% for cases and 62% for controls. Based on questionnaire information and the development of a crop exposure matrix, occupational experts classified their confidence that there was glyphosate exposure, duration, frequency, and intensity of exposure for each study subject. Exposure was not assessed for any pesticide other than glyphosate. Using unconditional logistic regression analysis, the authors estimated the OR for major lymphoma subtypes associated with exposure to glyphosate adjusted by age, gender, education, and study center. Thirty-six study subjects (2.2%) were classified as ever exposed to glyphosate, only 15 subjects (0.9%) were classified as ever exposed with median or high confidence. There was no association between glyphosate and NHL (OR = 1.4, 95% 0.6, 2.9 for low, medium, or high confidence exposures; OR = 1.2, 95% CI 0.40, 3.46 for medium or high confidence exposures). Analyses using various exposure metrics were statistically imprecise due to the low number of exposed cases and controls.

Evaluation: This is a large case control study with cases diagnosed during the period 2012–2016. Diagnoses were reviewed by pathologists in each of 6 study centers. [58] The frequency of ever using glyphosate was very low (< 1% with medium or high confidence). It is uncertain how many annual or lifetime days of glyphosate exposure were experienced by cases and controls, but presumably days of exposure were few.

Among the limitations of this study, the authors did not assess exposure to pesticides other than glyphosate, precluding even a superficial assessment of confounding, and there was a marked disparity in participation between cases (93%) and controls (62%).

Glyphosate cohort studies

There are 2 cohort studies in the glyphosate/NHL literature. [7,42] By design, both studies largely obviate concerns about recall bias because health outcomes cannot influence self-reported exposure information, a key advantage over the case control studies. [6] Also, the within cohort comparisons of users and non-users of specific pesticides should make for more comparable comparisons than in the case control studies where control participation was low or limited to sources that were unlikely to be representative of the population that gave rise to the cases.

Andreotti et al. updated the initial AHS glyphosate publication by DeRoos et al. [7,22] The updated analysis included 54,251 individuals who had sought licenses to apply restricted-use pesticides and were enrolled in the AHS between 1993 and 1997. Incident cancer diagnoses were identified via linkage to cancer registries in Iowa (through 2013) and North Carolina (through 2012). Subtypes of lymphoid malignancies were defined according to the Surveillance, Epidemiology, and End Results Program Lymphoma Subtype Recodes. According to this updated classification, multiple myeloma was included in the analyses as a subtype of NHL. Vital status was determined through state mortality registries and the National Death Index. State of residence was regularly updated using various government databases and participants were censored when they left either Iowa or North Carolina. In general, the cohort had low residential mobility and lost-to-follow-up was minimal. [12].

Lifetime use of glyphosate and 49 other pesticides was ascertained at enrollment and in follow-up questionnaires. At enrollment, applicators reported the number of years and days per year each pesticide was used, while at follow-up applicators reported the number of days each pesticide was used in the most recent year farmed. Using this information, three metrics of cumulative lifetime exposure were created for each pesticide: ever versus never use, lifetime days of use (days per year multiplied by the number of years), and intensity-weighted lifetime days of use (lifetime days multiplied by an intensity score). The intensity score was derived from an algorithm based on whether the participant personally mixed or applied pesticides, repaired pesticide-related equipment, used personal protective equipment, and the application method used – practices known to influence absorbed dose for pesticides. [17] Participation in the follow-up questionnaires was 64% for phase 2 and 46% for phase 3 (aghealth.nih.gov/collaboration/questionnaires.html accessed 10/17/22).

Poisson regression was used to calculate incidence rate ratios (RRs) and 95% CIs. RRs were adjusted for attained age (continuous), cigarette smoking status (never, former, current), alcohol drinks per month (none, ≤ 6 per month, ≥ 7 per month), family history of any cancer (yes, no), state of recruitment (North Carolina, Iowa), and the five pesticides most highly correlated with glyphosate. Lagged exposure was also evaluated allowing for 5, 10, 15, or 20 years to address the induction-latency period for specific cancers. Other potential confounding factors evaluated were body mass index (BMI; <25, 25–<30, _30 kg/m²) and pack-years of cigarettes smoked (tertiles of use among former and current smokers). For NHL and other lymphohematopoietic cancers, RRs were additionally adjusted for occupational exposure to solvents, gasoline, x-ray radiation, and engine exhaust, and pesticides linked to lymphohematopoietic malignancies in previous AHS analyses (lindane, DDT, diazinon, terbufos, and permethrin). Lastly, sensitivity analyses were conducted to evaluate conventions/assumptions in the various analyses performed.

Among the 44,932 participants who ever used glyphosate (83% of the cohort), the median lifetime days of use was 48 days (interquartile range [IQR] = 20–166 days), and the median lifetime years of use was 8.5 years (IQR = 5–14 years). Overall, glyphosate use was not associated with NHL. The NHL RRs, adjusted for confounding factors, by lifetime days of use categories were: ≤ 14 days RR 0.76 (95% CI 0.57, 1.01), 14 to 38 days RR 0.87 (95% CI 0.66. 1.14), 39 to 108 days 0.85 (95% CI 0.64, 1.13), ≥ 109 days RR 0.80 (95% CI 0.6, 1.06) ([7], supplementary table 1). In analyses by exposure intensity weighted lifetime days of use findings were: quartile 1 RR 0.83 (95% CI 0.59, 1.18), quartile 2 RR 0.83 (95% CI 0.61, 1.12), quartile 3 0.88 (95% CI 0.65, 1.19), quartile 4 RR 0.87 (95% CI 0.64, 1.20). The results for the various lagged analyses were consistent with those for the unlagged analyses.

Evaluation: This study provides follow-up data on a large cohort of licensed pesticide applicators with high quality ascertainment of NHL. The median days of glyphosate use for study participants who used glyphosate was 48 days and 83% of the cohort reported glyphosate use at enrollment or during follow-up. Analyses were provided by increasing days of use categories up to ≥109 days of use. Based on the IQR, the median exposure in the highest days of use category was >166 days.

The results across quintiles of exposure intensity or days of use were consistent and did not provide evidence of a positive relationship between glyphosate use and NHL. The authors did not report a RR for ≥1 day of lifetime use as had been done in all the other studies. The weighted average RR across days of use and exposure intensity quartiles – essentially equivalent to the RR for ≥1 day of glyphosate use - was 0.85 (95% CI 0.73, 1.0) for both fixed and random effects models (my calculation [61]).

Due to the prospective cohort design, there is minimal concern about exposure recall bias. Participants in the AHS are licensed pesticide applicators who should be able to report pesticide use accurately compared with other study populations. The frequency of glyphosate use in this study greatly exceeds that from all other studies in the literature. The statistical analyses were well described and comprehensive in controlling for a myriad of personal factors, other pesticides, and other agricultural exposures. In addition, the authors conducted analyses that allowed for 5, 10, 15, and 20+ years induction-latency periods.

A limitation of this study is that the follow-up questionnaires, intended to gain additional information about pesticide use after enrollment, had limited participation. To address this limitation, the authors imputed missing pesticide use data after enrollment based on demographics, medical history, farm characteristics, and reported pesticide use at enrollment. They also conducted analyses based only on self-reported information from phases 2 and 3 and self-reported information at enrollment. There was no appreciable change in results from these analyses. As a practical matter, frequent glyphosate users are highly likely to continue being frequent glyphosate users, so their cumulative days of use should be able to be imputed reliably.

Leon et al. evaluated pesticide use and risk of non-Hodgkin's lymphoma (NHL) and subtypes based on a combined analysis of data from three agricultural cohorts: the AHS, France's Agriculture and Cancer cohort (AGRICAN), and Norway's Cancer in the Norwegian Agricultural Population (CNAP). [42] Taken together, these three cohorts included 316,210 participants: 51,167 participants from the AHS, 127,282 participants from AGRICAN, and 137,821 participants from NCAP. The three cohorts included in this analysis had follow-up as follows:

• -AHS - first date of enrollment between 1993 and 1997 through December 31, 2010 (North Carolina participants) or December 31, 2011 (Iowa participants);
• -AGRICAN - first date of enrollment between 2005 and 2007 through December 31, 2009;
• -CNAP – January 1, 1993 through December 31, 2012.

For all three cohorts, NHL was identified through linkage with cancer registries.

Pesticide exposure assessment methodologies were fundamentally different for the European cohorts and the US AHS cohort. In the AHS, the investigators collected self-reports of the use of specific pesticides directly from individual cohort members prior to follow-up for health outcomes. For AGRICAN and NCAP, exposure assessment was determined using a crop-exposure matric (CEM):

• -AGRICAN cohort members were considered exposed to a pesticide if they reported growing a crop, indicated that they used pesticides on that crop, and a pesticide was registered for use on that crop in France. For example, cohort members who farmed grains were attributed exposure to every pesticide registered for grains for each calendar year during their years of agricultural work with grains.
• -CNAP cohort members were considered exposed to a pesticide if: they reported having cultivated a given crop, they possessed spraying equipment or spent money on pesticides, and the active ingredient was registered for use on the crop. CNAP cohort members were attributed exposure for all the pesticides registered for specific crops for each calendar year during their working years.

Cox proportional hazards modeling was used separately for each cohort to estimate RRs and 95% CIs for 14 chemical groups and 32 active ingredients with NHL and for 4 NHL subtypes: chronic lymphocytic lymphoma/small lymphocytic lymphoma (CLL/SLL), diffuse large B-cell lymphoma (DLBCL), follicular lymphoma (FL), and multiple myeloma/plasma cell leukemia (MM/PCL). All cohort specific models were adjusted for age, sex, and animal production. The AGRICAN model was also adjusted for retirement status and the number of crops personally treated with pesticides. For the CNAP cohort model, adjustment was also made for specific pesticides. The cohort specific RRs were combined using a random effects model to calculate a meta-RR.

There was no association between ≥1 day of lifetime glyphosate use and overall risk of NHL (RR 0.95, 95% CI 0.8, 1.2). Results for the NHL subtypes did not show positive associations for CLL/SLL (RR = 0.92, 95% CI 0.69 1.24), FL (0.79, 95% CI 0.52, 1.21), and MM (RR = 0.87, 95% CI 0.66, 1.15). However, the authors reported a positive association between ≥1 day of glyphosate use and DLBCL (RR = 1.36, 95% CI 1.00, 1.85). DLBCL findings for the individual cohorts were: AHS (HR = 1.20, 95% CI 0.72, 1.98), CNAP (HR = 1.67, 95% CI 1.05, 2.65) and AGRICAN (HR = 1.06, 95% CI 0.51, 2.19). The authors concluded that there was no association between glyphosate and NHL overall or for most subtypes, but that there was an association with DLBCL.

The authors provided some insights about the limitations of their study methodology. Foremost was exposure misclassification as evidenced by the very low correlations when the CEM methodology was compared to the self-reported pesticide use data for AHS cohort members (median correlation coefficient for specific pesticides r = 0.07). They also mentioned the inadequacy of a ≥ 1 day of lifetime glyphosate use metric for characterizing cancer risk from exposure. Lastly, they noted that in their analysis of AHS members that, in contrast to the most recent publication from the AHS [7], they did not control for cigarette smoking, alcohol consumption or family history of cancer, while they did control for animal production and different pesticide active ingredients than in the AHS analyses.

Evaluation: This pooled analysis included an enormous study population with high quality NHL ascertainment. Exposure to ≥1 lifetime day of glyphosate use varied across the cohorts: AHS 81%, AGRICAN 37%, CNAP 39% (Leon et al. [42] supplemental fig. 1). Overall, the authors found no association between glyphosate and NHL. There was, however, an association reported between glyphosate exposure ≥1 day and the DLBCL NHL subtype.

There are noteworthy aspects of the association for DLBCL that deserve examination. First, the estimate of the glyphosate RR for the AHS cohort differs from the recently published DLBCL result from the AHS by Andreotti et al. [7] that includes 2 more years of follow-up (unlagged RR = 1.0, 95% CI 0.8, 1.4). [Table 2 Andreotti et al. [7] did not present a DLBCL result for ≥1 day of lifetime use, but it can be estimated by a weighted averaging of RRs for the exposure quartiles in that study (my calculation [61])]. Also, Leon et al. [42] excluded >4000 commercial applicators from the AHS cohort, so the AHS population is different than that reported by Andreotti et al. [7] On balance, it appears that there is no association between ≥1 day of glyphosate use and DLBCL in the AHS cohort and that the positive association between glyphosate and DLBCL is restricted to the Norwegian cohort.

The CEMs used to assign exposure for French and Norwegian workers were based on a small number of crops. In France, the CEM was based on grassland, corn, grains, potatoes, tobacco, orchard crops, and vineyards. For Norway, the CEM was based on grassland, potatoes, grains, orchard crops, and greenhouses. As the authors noted in a previous publication [13], the CEM methodology misclassifies an unknown number of cohort members as falsely having exposure to specific pesticides because cohort members are assigned exposure for every pesticide that was registered for the crops they farmed. Further, Brouwer et al. noted that a recent survey in Norway indicated that pesticides were applied to potatoes on only 66% of farms and only a fraction of farmers applied any herbicides to grassland. [13] This would have added to the exposure misclassification characteristic of the CEM methodology.

For glyphosate and all the specific pesticides evaluated in this study, exposure classification via the CEM is not specific. Those truly exposed to glyphosate may have been in the minority among those classified as having had glyphosate exposure. Consider, for example, a pesticide that had <50% market share across the crops included in the CEM. In that instance, more than half of the workers attributed exposure would be unexposed. Imagine a pesticide that had 25% market share, etc. Such widespread misclassification precludes a valid assessment for most, if not all, pesticides. It also precludes controlling for the potential confounding effects of other pesticide exposures. One needs to know exposure specifically to estimate risk for glyphosate and to control for confounding from other pesticides. Exposure misclassification on this scale and the concurrent inability to control confounding call into question the validity of results for every specific pesticide in this study.

Meta-analyses

A number of meta-analyses have been published regarding glyphosate and NHL including Schniasi & Leon, Chang & Delzell, Zhang et al., Donato et al., Boffetta et al., and Kabat et al. [11,16,25,39,59,72] The different authors used different conventions in their analyses. These meta-analyses are not primary data, but efforts to take a weighted average of the primary data to increase statistical precision and learn something about between study variability. Given the variability across studies in number of days of glyphosate use, possible recall bias in the case control studies, possible selection bias in the case control studies, variable amounts of proxy information in the case control studies, incorrect analytic approaches (defining unexposed as not having exposure to any pesticide), approach to exposure assessment, and little to no control for confounding by other pesticide exposures, these meta-analyses are averaging results of questionable similarity and validity. Fixed and random effects models, as implemented by these authors, cannot compensate for the underlying variability in exposure scenarios and the limitations of many of the primary studies. As Greenland and O'Rourke counsel: averaging results should be limited to results that could be reasonably expected to be similar across studies.[32] Further, they counsel that if systematic variation exists between studies, as it clearly does in this literature, it violates the assumptions of random and fixed effects models.

Meta-analyses can also be out of date as the literature progresses. The pooled reanalysis of the North American case-control studies by Pahwa et al. estimated the OR for any use of glyphosate and NHL overall to be 1.1 combining the studies by McDuffie et al. and DeRoos et al. [24,45,53] Several meta-analyses incorporated the OR of 2.1 from DeRoos et al. that was not likely a valid input. [24].

Summary & assessment

Ideally, one would study the possible relationship between glyphosate and NHL in populations with long and frequent exposure to glyphosate, available objective information on exposures to all pesticides and chronic agricultural exposures, representative control groups in case control studies, and well executed and clearly specified statistical approaches to control confounding. Several of the studies in the literature involve exposure scenarios that are so infrequent (a few days during a lifetime) or uncertain that they would not be given credence in other areas of epidemiology (Table 4). Included among these studies are Hardell et al., Cocco et al., and Meloni et al. [18,33,46] These studies also did not control confounding acceptably. The pooled case control analysis by DeRoos et al. [24] also seems questionable because of very infrequent exposure, the very short interval between glyphosate's initial registration in 1974 and the identification of NHL cases, the extensive information from proxy respondents, and the difference between the OR for ≥1 day of glyphosate use and NHL in the subsequent pooled reanalysis by Pahwa et al. [53]. The study by Leon et al. is also likely to be uninformative because of the major uncertainty about whether many or most of those considered to be exposed to glyphosate were truly exposed (totally dependent on market share by crop in the CEM) and the inability to control confounding with such widespread misclassification across all pesticides. [42] For perspective, consider a study of a potential side effect of a blood pressure medication where each patient with high blood pressure was given credit for exposure to every blood pressure medication on the formulary in his/her health plan and all other medications on formulary for documented comorbidities. Such an analysis would not be considered credible.

Table 4.

Summary evaluation of glyphosate/NHL studies.

Author(s)	Exposure assessment	Ever exposure frequency; median days of glyphosate use	Control for other pesticides	Other considerations	Quality
McDuffie et al.	proxy information for 21% cases, 15% controls	10% cases, 9% controls; median days ∼10 days	No	Marked skew toward positive results	Poor
Hardell et al.	Proxy information for 44% cases and controls	1.6% cases; 0.7% controls; median days unknown, but likely very few days	No	Marked skew toward positive results	Poor
De Roos et al.	Proxy information for 31% cases, 40% controls	5.5% cases, 3.2% controls; median days unknown, likely few	Yes	Cases identified very soon after glyphosate was available for use	Poor
Eriksson et al.	No proxy information	3% cases, 2% controls; median days ∼10 days	Yes in ≥1 day analysis; no otherwise	Marked skew toward positive results	Fair
Orsi et al.	No proxy information.	5% cases, 6% controls; median days unknown	No	Significant problems with interviews	Poor
Cocco et al.	No proxy information	< 1% cases, < 1% controls; median days unknown, but likely very few	No		Poor
Meloni et al.	No proxy information	< 1% with medium or high confidence for cases and controls; median days unknown, but likely very few	No	Did not collect information about other pesticides	Poor
Andreotti et al.	No proxy information; some imputation during follow-up period	83% for cohort members; median days 48 with an IQR of 20 to 166 days	Yes		Good
Leon et al.	No proxy information. No firsthand reporting of pesticide uses for EU cohorts. Exposures assumed based on crops farmed.	AHS 81% AGRICAN 37% CNAP 39% [Leon et al., suppl fig. 1]; median days unknown	Yes	Glyphosate exposure not specific. Questionable control for confounding in EU cohorts due to lack of specificity in assigning exposures	Poor

The remaining studies have the most frequent exposure scenarios in the glyphosate/NHL literature. [7,29,45] The median exposure in the McDuffie et al. study and the Eriksson et al. study appear to be about 10 days of lifetime use. [29,45] Ten days over a lifetime is a very questionable extent of exposure for cancer causation, excluding extremely potent genotoxins like intravenous cytotoxic chemotherapy and high dose radiation exposure. The study by Andreotti et al. is the clearest about exposure frequency and is the only study where exposure might be described as relatively frequent. [7] Participants had a median exposure of 48 days of glyphosate use with an interquartile range of 20 to 166 days. The authors provided NHL results for glyphosate exposure categories of 14 to 38 days, 39 to 108 days, and ≥ 109 days of exposure and for several categories of intensity weighted days of use.

Other than the multivariate analysis for ≥1 day of glyphosate use in Eriksson et al. [29], the glyphosate analyses in McDuffie et al. [45] and Eriksson et al. [29] do not seem to have controlled statistically for other pesticide exposures. There also seems to be uncontrolled bias in these studies based on Crump's finding that these studies had positive associations with NHL for ≥90% of pesticides other than glyphosate. [21].

Confounding seems to have been addressed well by Andreotti et al. [7] The authors were comprehensive in their glyphosate modeling to adjust for personal factors and other exposures. They had consistent results across increasing categories of days of glyphosate use and the categorical results were statistically precise. There is some question about the accuracy of the missing data they imputed from the phase 2 and 3 questionnaires, but their sensitivity analyses supported the imputation results. It also seems likely that continuing use for frequent glyphosate users can be predicted. In terms of amount of exposure, the study by Andreotti et al. [7] focuses on the most frequent users of glyphosate by far.

From a multidisciplinary perspective, several regulatory agencies have established 0.3 to 1.0 mg/kg/day as an acceptable daily dose for glyphosate over a lifetime. Glyphosate and other pesticide use is intermittent [10] and most of the studies in the literature involve very infrequent exposure scenarios. Biomonitoring data show that systemic doses from uses of glyphosate in agriculture have a median of approximately 0.0001 mg/kg per application day, orders of magnitude below what regulatory agencies consider to impart no excess risk on a daily basis. These regulatory assessments support the validity of the NHL findings by [7] and the likelihood that positive findings in the literature for populations with very infrequent exposure were likely due to systematic error.

Postscript

I am indebted to one of the peer reviewers for pointing out a potentially relevant publication that appeared online soon after my literature search was completed. In that article, DeRoos et al. pooled data from 10 case-control studies participating in the International Lymphoma Epidemiology Consortium to assess possible relationships between numerous pesticides and NHL and the various NHL subtypes. [23] Eight of the studies had some information for glyphosate. [15,18,30,34,40,47,52,66,73] De Roos et al. concluded that they found no increased risk for glyphosate and NHL overall (ever use OR = 1.03 95% CI 0.83, 1.29; >15.5 years of use OR = 0.90 95% CI 0.59, 1.37), but they highlighted an association for glyphosate and follicular lymphoma that was limited to shorter-term use (ever use OR = 1.42 95% CI 0.98, 2.05; use for <8 years OR = 1.66 95% CI 1.12, 2.45; use for >15.5 years OR = 1.06 95% CI 0.52, 2.17). Taken at face value, these results would not normally be interpreted to support a causal relationship between glyphosate and NHL overall or for any subtypes. But it is not infrequent in the pesticide literature to overinterpret selected findings in a milieu of findings and to ignore potential systematic error and biologic plausibility.

Arguments about interpretation of this recent pooled analysis aside, consider whether the pooled studies are of high enough quality to be informative about glyphosate. Two of the 8 studies were included in my review [18,52] and found to be poor quality and to involve very infrequent glyphosate use for exposed individuals. Of the 6 studies that were not included in my review, based on a text search for “glyphosate” or “gly”, only 1 study as initially published mentioned the word glyphosate at all. That 1 study had an analysis that included glyphosate as part of an “other herbicides” category including carbamates and other pesticides [30] and very few individuals classified by an occupational hygienist as having probable substantial exposure to the category (0.4% of controls, 1.3% of cases). Across the 6 studies the exposure assessment methods utilized, and the exposure scenarios described, varied markedly. One study focused primarily on hair dyes, but use of pesticides was self-reported on the study questionnaire, though not considered in the initial analysis. [40,73] One study focused on participants' self-reports of pesticide use on lawns and gardens and defined exposure as use by the respondent, a pesticide service, or someone else.[34] It is worth noting that the amount of glyphosate that would be used in a typical residential application is miniscule compared to the amount that would be used in a typical agricultural scenario where large numbers of acres are usually treated. This study is clearly an outlier since the remaining 7 studies were described as focusing on agricultural exposures. Three of the remaining 4 studies had exposure assigned based on crop-exposure matrices, from which it is impossible to know whether individual study participants actually used glyphosate or their frequency of use.

The inherent assumptions behind pooled analyses and meta-analyses are that the studies being combined are estimating the same effect parameter for a similar exposure (or dose) scenario and that the studies being combined are fit for purpose: viz., that a meaningful amount of exposure or dose is being studied and that the studies have been executed in such a way as to allow for valid combined analyses. To the extent that these assumptions do not hold, the pooled result will be uninformative. Also, pooling cannot resolve the limitations of the individual case-control studies including possible recall bias, potential for selection bias, and exposure misclassification. It was an enormous effort by De Roos et al. [23] to pool these 8 studies and their analyses were comprehensive in many ways. But it is questionable whether the studies included in the pooled analysis were of a high enough quality, considered a biologically plausible frequency of use or amount of internalized dose, and were similar enough in design and execution to yield informative pooled results for glyphosate and NHL.

Sponsorship

This review was funded by Bayer US - Crop Sciences, Monsanto Company, Chesterfield, MO, USA. The content was developed solely by the author and was not shared with the sponsor prior to acceptance for publication.

Disclosure

The author worked for Monsanto from 1989 to 2004 and has consulted for them on glyphosate-related epidemiology issues, including litigation, since January 2015.