Weeding out inaccurate information on glyphosate-based herbicides and risk of non-Hodgkin lymphoma

Controversy of glyphosate-based herbicides

Glyphosate-based herbicides (GBH) are the most widely used herbicides in the world. Recently, the safety of GBH has come under intense public and scientific scrutiny given its ubiquitous presence in agriculture and reports of increased relative risk of non-Hodgkin lymphoma (NHL) among occupationally exposed individuals. Differing conclusions about glyphosate’s carcinogenic potential among national and international regulatory agencies have further incited controversy among stakeholders with competing interests. Given this controversy, it is even more important to critically analyze the available empirical data to resolve the uncertainty surrounding the herbicide so that it could be safely used and further regulated.

Association between GBH and NHL

To assess whether a link truly exists, our group (Zhang 2019 [1]) previously conducted a meta-analysis of occupational exposures to GBH and risk of NHL in epidemiological studies and detected a 41% increased meta-relative risk (meta-RR). We also conducted a number of sensitivity analyses that validated the robustness of this association.

Since our meta-analysis was released [1], a similar meta-analysis of epidemiologic studies of occupational exposure to glyphosate and NHL was published by Donato 2020 [2] where they reported no evidence of an increased risk. This finding deviates substantially our article [1], which was absent among previous meta-analyses cited by Donato 2020 [2].

To understand what factors contributed to the discrepancy between Donato 2020 [2] and our previous findings, we carefully reviewed each paper’s methods and determined that Donato’s approach was essentially very similar to ours [1]. In this commentary, we aim to address three key issues: 1) whether we could reproduce Donato 2020’s meta-analysis using the exact data they selected; 2) how we could better understand the major differences between Zhang 2019 [1] and Donato 2020 [2] regarding study selection and exposure categories; and 3) what are the inaccurate or misunderstood aspects surrounding the association of GBH and NHL which need to be weeded out. The answers are briefly described below.

Reproducibility of meta-relative risk of NHL

Table 1 summarizes the original meta-analysis findings reported by Donato 2020, the re-calculated meta-RRs we obtained using the same data they used, and the corresponding results from Zhang 2019. Donato 2020 reported a meta-RR of 1.03 (95% CI = 0.86-1.21) for ever-never exposure to glyphosate and NHL risk. When we replicated this analysis, the meta-RR we obtained was 1.14 (95% CI = 0.94-1.39, Table 1). Of note we produced the same funnel plot of these studies, indicating that we used the same datasets as Donato, which revealed asymmetry consistent with publication bias (Donato 2020, Figure 4).

Table 1.

Replicated major findings from the Donato 2020 meta-analysis of glyphosate exposure and non-Hodgkin lymphoma (NHL) model and comparison with Zhang 2019 using random-effects model.

Meta-Analysis	N	Donato 2020			Replicated Meta-Analysis			Zhang 2019
		meta-RR	CIL	CIU	meta-RR	CIL	CIU	N	meta-RR	CIL	CIU
Exposure category
Ever exposure (main analysis)	7	1.03	0.86	1.21	1.14	0.94	1.39	6	1.30	1.03	1.64
Highest (if available)a	-	-	-	-	-	-	-	6	1.56	1.12	2.16
Highest onlyb	3	1.49	0.37	2.61	1.49	0.67	3.34	3	1.63	0.97	2.76
Remove Leon 2019 (all case-control)c	6	1.27 d	0.92	1.61	1.34d	1.04	1.73	5d,e	1.84	1.33	2.48
Cell-type specific
DLBCL (ever)	3	1.31d	0.93	1.7	1.32d	0.99	1.76	-	-	-	-
MM (ever)	3	1.04	0.67	1.41	1.15	0.76	1.74	-	-	-	-

Abbreviations: CLL, chronic lymphocytic leukemia; CI, confidence interval; DLBCL, diffuse large b-cell lymphoma; N, number of studies; meta-RR, meta-analysis relative risk.

^aZhang 2019 used high exposure category when reported, and ever-exposure for all other studies.

^bOnly three studies that reported high exposure categories were used. For Andreotti 2018, Zhang 2019 selected highest intensity-weighted lifetime days lagged by 20 years or more, whereas Donato 2020 selected highest days per lifetime.

^cThe remaining studies are all case-control. In their analysis of all case-control studies, Zhang 2019 follows the a priori selection criteria and has (N=6) because Cocco 2013 was not included in the analysis.

^dFixed-effects model was used because between-study heterogeneity, defined as the X²-test statistic for heterogeneity being greater than its degrees of freedom (number of studies minus one), was not detected. Use of fixed-effects model was not reported in Donato 2020.

^eLeon 2019 was not used in Zhang 2019, though it included data from Andreotti 2018. Thus, Andreotti 2018 was removed to conduct an analysis of only case-control studies.

Another source of differences was the weights of original studies selected. According to the authors, Leon 2019 [3] was the most highly weighted study at 74.11%, though it was 48.03% according to our calculations. As a sensitivity analysis to determine the impact of including such a large study, Leon 2019 was removed, and Donato 2020 reported a meta-RR of 1.27 (95% CI = 0.92-1.61). However, when we removed this study, we found the meta-RR was both increased and statistically significant (meta-RR = 1.34, 95% CI = 1.04-1.73). Of note, inclusion of this one study changed the overall findings of the meta-analysis. Further sensitivity analyses should have been conducted to determine the sources of heterogeneity. The data discrepancies are clearly reflected in Table 1.

Major differences between both meta-analyses

Study selection difference: Andreotti 2018 vs. Leon 2019

The study selection varied between the two meta-analyses. Zhang 2019 used Andreotti 2018 which reported a high exposure estimate [4], whereas Donato 2020 used the recently published pooled analysis by Leon 2019 that reported only ever-never exposures [3]. Donato 2020 also included Cocco 2013 [5] in their analysis, a study that was excluded from our meta-analysis because it reported results for all B-cell lymphomas combined (two cases of NHL, one case of multiple myeloma, and one unspecified B-cell lymphoma; N = 4).

Statistical analysis model selection difference

Further, data selection from De Roos et al. [6] also differed. For example, Donato 2020 used the hierarchical regression model, whereas Zhang 2019 used logistic regression. When we did a sensitivity analysis using the hierarchical regression model in this study, we found the meta-RR to be 1.46 (95% CI: 1.08-1.96), as shown in Table 6 of Zhang 2019 [1]. Although we reported both the fixed-effects and random-effects models, and Donato 2020 reported the random-effects model, none of Donato meta-analysis results were able to be replicated.

Exposure category selection: high vs. ever

Zhang 2019 followed an a priori criteria to select the highest exposure category when it was available, whereas Donato 2020 used ever-never exposures. When they attempted a dose-response analysis, the exact criteria of how their highest exposure category was selected were not described in detail. While it appears that exposure frequency was prioritized, it is unclear why this metric was selected, as it does not factor in exposure intensity. In contrast, Zhang 2019 clearly listed the order for selection of the most highly exposed category based on the a priori hypothesis and the current scientific understanding of NHL risk in Section 2.6 [1]. If the a priori hypothesis detailed in Zhang 2019 were followed, the meta-RR from the same studies used in Donato 2020 would increase to 1.63 (95% CI = 0.97-2.76, Table 1). To determine whether high exposures were linked to higher risks of NHL, we compared risks for ever-never exposures to the high exposure groups and found the meta-RR increased by 35% in our replicated analysis of Donato 2020 and 33% in Zhang 2019 (Table 1), indicating presence of an exposure-response relationship in both analyses.

Overall, fundamental errors in methodological reasoning such as differences in study selection, statistical model use, and imprecise definitions of high exposure categories ultimately contributed to the flawed conclusions reported in Donato 2020.

Is the NHL risk present only for "highly" exposed subjects?

Given glyphosate’s ubiquitous use in the world, delineating an unexposed control population is virtually impossible, since almost everyone has been exposed to the residues of the chemical to some extent. Hence, to ensure sufficient exposure contrast and optimize our ability to detect a potential association, our a priori hypothesis focused on highly exposed groups with respect to their lower exposed counterparts. We defined “high” as a group exposed at a level relatively greater than the other exposed categories. For example, greater than two days of exposure per year (>2 d/y) group was selected as highly exposure group in the McDuffie et al. [7] study (Table 4, Zhang 2019). Being exposed to glyphosate for greater than two days every year is traditionally not considered high exposure, but we were constrained by the nature of the meta-analysis to rely on available data, and determined that individuals in the >2 d/y group were more highly exposed relative to other groups in that study. Importantly, these individuals were exposed to GBHs at routinely used agricultural spray concentrations, and not extremely high levels. In fact, we believe that the 41% increased relative risk of NHL reported in Zhang 2019 still underestimated the true risk.

Is our previously reported 41% increase in NHL meta-RR underestimated?

Due to the timing of GBH exposures and study subjects recruited in original individual studies, and potential long latency of NHL, our calculated meta-RR may have been actually underestimated. Most of the studies included in our meta-analysis were conducted prior to the exponential increase in glyphosate use and consequent widespread exposure. The available studies included in our meta-analysis evaluated cancers that developed prior to 2013, though the practices of spraying down crops with glyphosate to speed up desiccation (also known as the “Green Burndown”) started occurring in the mid-2000s (Figure 1, Zhang 2019). For example, the study Eriksson et al. [8] captured exposures well before the exponential increase during the Green Burndown (1999-2002); despite that, it still detected a positive dose-response relationship (increase of odds ratio from 1.69 to 2.36, 95% CI = 1.04-5.37). Hence, given exposures have drastically increased in recent years [9], our risk estimate is likely underestimated.

Figure 1.

Forest plot for ever-exposure to glyphosate and NHL using random effects model.

Lastly, although Leon 2019 [3] was published most recently, the follow-up periods of each of the cohorts was still limited to approximately ten years ago, indicating the current true NHL risk has yet to be uncovered.

A major NHL subtype, DLBCL, was reportedly linked to GBH

When taking a close look, we further identified discrepancies among the Donato 2020 analysis of diffuse large B-cell lymphoma (DLBCL) and multiple myeloma (MM) in our recalculations using the same studies and data (Table 1).

One interesting question posed by this endeavor is whether or not GBH contributes to risk of all NHL, or a particular subtype. DLBCL is the major subtype (~25%) of NHL that is heterogeneous with multiple phenotypic lymphoma subtypes [10]. Notably, the Leon 2019 study included in Donato 2020 detected a positive statistically significant association with DLBCL (mHR=1.36, 95% CI=1.00-1.85). This finding potentially indicates: 1) GBH may be more strongly associated with DLBCL than other NHL subtypes; 2) if true, analyzing NHL as a whole would attenuate potential associations, which is exactly what was reported [3]; and, 3) as a result, our meta-RR of NHL would be further underestimated.

Estimating NHL cases attributable to GBH exposure

Right after Zhang 2019 study was accepted for publication (February 2019), some blogs pointed out that our finding of a 41% increased meta-RR simply equated to an additional eight cases of NHL per 100,000 men and women resulting from exposure to glyphosate.^1,2 This calculation is clearly based on the US National Cancer Institute Surveillance, Epidemiology, and End Results (SEER) Program age-adjusted rate of new cases of non-Hodgkin lymphoma is 19.6 per 100,000 men and women per year [11]. Multiplying this rate by our increased meta-relative risk of 41% would yield roughly eight additional NHL cases per 100,000 people attributable to GBH exposure. There are a few issues with this inaccurate calculation.

First, the SEER rates span from 1972 to 2017, and include a sample of only 28% of US states and territories. Importantly, the data do not adjust for any potential confounders or other factors, such as exposures to GBH (or any pesticides/chemicals) linked to lymphomagenesis. In other words, the SEER rate for NHL (19.6 per 100,000) is a general measure of all new cases, including those with or without occupational exposure to GBH, within a specified population during a time period.

Second, the calculation is not valid, as it applies the increase risk of NHL with exposure to the current rates of NHL in the country. This measure does not exist in current epidemiologic methods, and has no interpretation. A better estimate is to assess how many NHL diagnosed each year are likely to be attributed to exposure to GBH. A measure of that is the population attributable fraction (PAF), which quantifies the contribution of a risk factor to a disease or a death. Another useful measure would be the population attributable risk (PAR), the proportion of the incidence of a disease in the population that is due to exposure. This measure indicates that the incidence of a disease in the population would be diminished if exposure were eliminated.

Lastly, from a public health perspective, an additional eight NHL cases from GBH exposure is substantial even though it is derived from an invalid calculation. For example, the California EPA Safe Drinking Water and Toxic Enforcement Act of 1986 defines the “no significant risk level” for known chemical carcinogens to be not more than one excess cancer case in 100,000 individuals exposed to the chemical over a 70 year lifetime [12]. Indeed, a 700% excess³ of that guideline value is concerning.

Conclusion

In this commentary, we discuss the importance of truly understanding the nature of the association between GBH exposure and NHL risk. The current evidence with our reanalysis does not support the conclusions drawn by Donato 2020 [2]. The fact that nearly none of the calculations were replicable (except the funnel plot), lack of transparent details regarding how high exposure categories were defined, and that one study contributes to nearly all the study weight (according to Donato 2020), raises serious concerns about the widespread misinformation surrounding the controversial herbicide. A closer look at the data has revealed that our calculated meta-RR of 41% increase in NHL among GBH-exposed workers is very likely underestimated, with recent studies indicating that subtype (DLBCL) specific associations may be stronger than the heterogeneous class of disease (NHL) as a whole. Despite these findings, no new studies have evaluated cancers beyond 2011, thus, our current understanding of the risk of NHL associated with widespread exposure to glyphosate is extremely limited. Given the magnitude of the public health issue at hand, we felt compelled to comment.