Glyphosate exposure and GM seed rollout unequally reduced perinatal health

Significance

While the herbicide glyphosate is the most commonly used herbicide globally, the effects of glyphosate exposure on human health and the environment remain unclear—particularly in more developed countries, where glyphosate exposure is often considered low. Using spatiotemporal variation in the adoption of glyphosate-resistant crops, we document significant adverse perinatal health effects due to increased glyphosate exposure in the rural United States. Further, historically disadvantaged groups disproportionately bear these health effects. These results conflict with current regulatory guidance, suggest current regulations may be inadequate, and highlight the need to improve pesticide use and exposure monitoring.

Abstract

The advent of herbicide-tolerant genetically modified (GM) crops spurred rapid and widespread use of the herbicide glyphosate throughout US agriculture. In the two decades following GM-seeds’ introduction, the volume of glyphosate applied in the United States increased by more than 750%. Despite this breadth and scale, science and policy remain unresolved regarding the effects of glyphosate on human health. We identify the causal effect of glyphosate exposure on perinatal health by combining 1) county-level variation in glyphosate use driven by 2) the timing of the GM technology and 3) differential geographic suitability for GM crops. Our results suggest the introduction of GM seeds and glyphosate significantly reduced average birthweight and gestational length. While we find effects throughout the birthweight distribution, low expected-weight births experienced the largest reductions: Glyphosate’s birthweight effect for births in the lowest decile is 12 times larger than that in the highest decile. Together, these estimates suggest that glyphosate exposure caused previously undocumented and unequal health costs for rural US communities over the last 20 years.

While the introduction of genetically modified (GM) crops profoundly transformed the agricultural landscape of the United States, our understanding of the implications for human health remains limited. At the heart of GM technology lies its resistance to the herbicide glyphosate, which enabled farmers to directly spray glyphosate onto GM crops, eliminating weeds while sparing the crops themselves. Recent research advertises the potential of this technology to improve farm productivity (1–3), but the pairing of GM seeds with glyphosate introduced a complex array of health externalities—with substantial uncertainty regarding the total effect. On the one hand, if glyphosate replaced more toxic herbicides common to non-GM cultivation, its introduction could yield positive health effects. However, the ensuing liberal application of glyphosate—enabled by the central innovation of herbicide resistance—has led to substantially higher volumes of chemicals sprayed, potentially worsening health (4–6).

In this study, we use spatiotemporal variation in GM-crop adoption to quantify the sign and magnitude of the human health externality that resulted from the widespread adoption of GM crops and the rapid increase of glyphosate. We then document heterogeneity in the effects of the GM weed management regime on perinatal health. Our results suggest that glyphosate has adverse effects on perinatal health, net of any benefits associated with reductions in the use of other herbicides. These adverse effects concentrate among the most at-risk births.

The United States first approved glyphosate for agricultural use in 1974, and the chemical subsequently gained prominence as a broad-spectrum herbicide used extensively in agriculture. Its effectiveness in weed control and relatively low toxicity contributed to its widespread adoption, becoming a critical component of weed-management practices. Pairing glyphosate with glyphosate-resistant GM crops removed a natural limiter in glyphosate use—glyphosate kills non–genetically modified crops and targeted weeds. Relaxing this constraint resulted in dramatic increases in glyphosate application intensity. Before the release of GM seeds in 1996, US farmers applied 0.1 kg of glyphosate per hectare of cropland; since GM seeds, the intensity has risen to over 1.3 kg/hectare (7). Meanwhile, in the EU, which never approved GM seeds, the glyphosate application rate remains near the United States’ pre-GM levels: approximately 0.2 kg/hectare (8). While potentially affected by other factors, this US-EU gap in glyphosate intensity illustrates how GM technology enabled substantially higher levels glyphosate intensity than otherwise available.

Since its approval, US regulators have maintained “there are no risks to human health from the current registered uses of glyphosate” (9)—despite a dearth of population-wide, causally founded studies. However, two recent studies document negative health impacts of glyphosate exposure in Brazil. Both Dias et al. (10) and Skidmore et al. (11) find that glyphosate exposure—driven by the expansion of GM seeds and transported through rivers—increased infant mortality and pediatric cancer deaths in Brazil. These two studies offer the first large-scale, population-level, plausibly causal estimates of the health costs of glyphosate exposure. Our study complements these analyses by considering glyphosate impacts in a substantively different socioeconomic setting—the United States’ gross domestic product per capita is approximately per capita is approximately nine times greater than Brazil’s (12)—with a potentially different exposure mechanism stemming from differences in intensity of glyphosate use, hydrology, geology, and meteorology. Brazil applies nearly twice the amount of glyphosate per hectare of cropland as the United States (10). Thus, while the United States and Brazil overlap as leaders in glyphosate application, key differences between the contexts warrant new study in the United States and abroad. Finally, given the Brazilian context, the papers focus on effects driven by exposure through upstream glyphosate application; we find that local exposure drives negative health impacts in the US context.

Background and Motivation.

Glyphosate and other herbicides pervade the environment, giving rise to multiple pathways of human exposure (13) including water (14), dust blown by the wind (15), aerial drift (16), direct contact (17), and food residue (18). Following application, glyphosate exhibits a relatively short breakdown period, with a half-life ranging from 2 to 215 d (19). Although a significant portion of the herbicide binds to the soil, reducing runoff, its high water solubility allows the unbound remnant to enter both surface and groundwater (20, 21). A comprehensive study across US waterways from 2015 to 2017 revealed the presence of glyphosate or its degradate, aminomethylphosphonic acid (AMPA), in 90% of samples (14). Additionally, wind-dispersed dust particles containing soil-bound herbicide residues can contribute to air pollution (15). While food residue is suspected to be another source of population-wide exposure to glyphosate (22), we cannot capture its effect in this study.

Exposure.

The multiple exposure mechanisms—coupled with the breadth and volume of glyphosate application—have resulted in widespread detection of glyphosate in the urine and blood of US residents. The US Center for Disease Control (CDC) detected glyphosate in 81% of urine samples from a nationally representative cohort (23). Multiple studies with pregnant women found glyphosate present in the urine of nearly every tested mother-to-be (24, 25). This ubiquity of glyphosate exposure in the US population highlights the importance of understanding the impacts of glyphosate exposure at a national scale—particularly within populations likely exposed to higher levels. We focus on rural populations’ exposures to local glyphosate sources—glyphosate exposure through dust, drift, direct contact, or water originating within the county of residence. We explore the potential effects of upstream spraying in SI Appendix, section N.

Health impacts.

A growing body of literature suggests that glyphosate has the potential to impact human health through a variety of biological mechanisms. Existing evidence typically comes from either laboratory studies on nonhuman animals or human-focused observational studies. While laboratory-based studies offer well-identified causal effects, they often suffer challenges of external validity. Previous observational studies (24–26) primarily focus on associations between self-reported exposure and health outcomes and typically avoid causal claims. The studies of Dias et al. (10), Skidmore et al. (11), and Camacho and Mejía (27) are exceptions—employing quasi-experimental methods to make causal statements.

Lab studies have established links between glyphosate and congenital anomalies in rats, developmental issues in frogs and chickens, and endocrine disruption for male reproduction in mice (28–30). Multiple additional studies link glyphosate exposure to endocrine disruption, which can affect developmental and reproductive health (31) and often have nonmonotonic effects (32). Research also documents glyphosate toxicity for placental cells—raising concerns for adverse effects in fetal development (33).

Consistent with lab-based concerns for the effect of glyphosate on development and reproductive health, several observational studies report associations between glyphosate exposure and miscarriage (26), gestational length (24), and birthweight (25)—related to a growing literature on health risks from occupational exposure (34) to glyphosate (35) and other pesticides (36). Considering these established mechanisms and documented associations, we evaluate the evidence for a causal effect of glyphosate on perinatal health outcomes—particularly birthweight and gestational length.

Beyond glyphosate-specific studies, a growing body of literature documents adverse health effects associated with pesticide exposure, even at low doses (37–43). One of the closest studies to the current paper—by Larsen, Gaines, and Deschenes (37)—finds pesticide exposure increases adverse birth outcomes (weight, gestation, and abnormalities) for California mothers. The authors highlight that this effect is driven by the subpopulation exposed to the highest levels of pesticides. These results and the growing body of literature motivate three points for our paper. First, birthweight and gestation are plausible outcomes to test for the health effects of glyphosate. Second, if glyphosate affects health, rural populations are likely the most impacted. Third, the effect of glyphosate may be heterogeneous—varying within the exposed population. These three observations lay the foundation of our analysis.

Empirical Approach.

We employ two complementary empirical approaches–reduced form difference-in-differences (DID) and two-stage least squares and two-stage–to estimate the perinatal health impacts from the rollout of GM crops and the ensuing intensification of glyphosate. Both approaches employ similar strategies to isolate plausibly exogenous variation in county-level GM adoption and glyphosate exposure by combining temporal variation in the commercial release of GM seeds in the United States with spatial variation in the suitability for growing the main crops for which GM seeds were available—corn, soy, and cotton. The temporal dimension of this strategy utilizes the arbitrary timing of the release of GM seeds. GM seeds became commercially available in the United States in 1996, and farmers rapidly adopted GM seeds and intensified glyphosate applications in the following years. The strategy’s spatial dimension uses the fact that these changes, on average, had larger impacts in places that were relatively more suitable for the crops—typically where the crops were already grown. Consequently, each approach leverages two comparisons to identify glyphosate’s effect on perinatal health outcomes: 1) before vs. after GM-induced glyphosate expansion, and 2) areas more suitable for GM crops vs. less-suitable areas.

The two empirical approaches provide complementary strengths. The reduced-form DID approach requires weaker assumptions for identification (discussed in Empirical Strategy) but only provides comparisons between more and less GM-suitable regions—rather than direct quantification of glyphosate’s health impacts. The two-stage approach provides this quantification but comes with stronger assumptions.

Data.

We measure perinatal health outcomes using the universe of individual-level birth data from the National Vital Statistics System (NVSS) during 1990–2013—the restricted-access natality files that allow us to match each birth to the counties of occurrence and the mother’s residence (44). These natality data include our primary perinatal health outcomes—birthweight and gestational length—and demographic and residence information from both parents and birth-location information. Our primary analyses focus on the over 10 million births that occurred between 1990 and 2013 in rural US counties or involved mothers residing in rural counties—as defined by the US Department of Agriculture (USDA). We focus on this subset as it represents the births most likely to be impacted by the increase in glyphosate intensity and exposure induced by the release of GM seeds.

Using each infant’s mother’s county of residence, we match the birth to a measure of glyphosate exposure—glyphosate intensity, defined as the volume of annual, county-level estimates of agricultural glyphosate applications per square kilometer of total county area—from the US Geological Survey (USGS) National Pesticide Synthesis Project spanning 1992–2017 (7). Our measures of crop suitability—attainable yield estimates for corn, soy, and cotton—come from the Food and Agriculture Organization of the United Nations Global Agro-Ecological Zones modeling framework (FAO-GAEZ). These time-invariant predicted yields result from modeling crop responses to environmental conditions such as soil type and climate—holding management practices constant (45). We calculate a continuous GM suitability percentile from the GAEZ data using a process described in Attainable Yield.

Fig. 1 illustrates both the spatial and temporal variation that we use to identify the effect of glyphosate on infant health. Fig. 1A maps the spatial variation in GM-crop suitability. This suitability strongly correlates with the increase in glyphosate intensity (kg/km²) after the introduction of GM seed—shown in Fig. 1 B and C highlights how the commercial release of GM seeds (dashed vertical line in 1996) drove marked increases in glyphosate use—particularly for higher GM-suitability counties. While the figures show all counties, we only use rural counties in our main analysis; SI Appendix, Fig. S1 maps these rural counties. Additionally, SI Appendix, Fig. S2 shows state-level adoption rates of GM seeds over time—adoption was rapid among all states.

Fig. 1.

GM-crop suitability predicts glyphosate increases after GM-seed introduction. (A) Percentile of attainable yield for GM crops equals the difference in attainable yield between high- and low-input scenarios from FAO GAEZ (45) for corn, soy, and cotton. We rescale each crop to be a national percentile, take the maximum over the three crops, and finally scale again to be a national percentile. (B) Change in glyphosate censored at the 1st and 99th percentiles. (C) Total glyphosate application by GM crop suitability quartile. We restrict our analysis to rural counties. See SI Appendix, Fig. S1 for a map of the rural designation and SI Appendix, Fig. S2 for state-level GM adoption over time.

Estimation.

We now discuss our two empirical approaches. Empirical Strategy elaborates on these strategies’ assumptions and implementations.

Reduced-form DID approach.

Our reduced-form DID approach estimates how infant-health outcomes in higher- and lower-GM-suitable counties diverged after the introduction of GM seeds and the subsequent glyphosate intensification (conditional on controls). Accordingly, this approach employs a DID estimator with continuous “treatment” (GM-crop suitability)—mainly relying upon a parallel-trends assumption, i.e., if GM seeds had not been introduced, more and less GM-suitable counties would have continued on similar trajectories (46).

Two-stage approach.

Our second approach uses a similar set of comparisons to the reduced-form approach—high vs. low GM suitability, before vs. after the GM-seed rollout. However, our two-stage approach—implemented as two-stage least squares (2SLS)—scales reduced-form estimates by a first stage: the increase in glyphosate intensity (kg/km²), estimated by the same double difference. We capture the pattern of adoption by interacting the instrument (GM-crop suitability) with indicators for year. 2SLS’s central assumption for a causal interpretation is the exclusion restriction (46), i.e., conditional on our controls, any divergence in perinatal health between high- and low-GM suitability counties after the GM-seed rollout only occurred through the channel of increased glyphosate exposure (discussed further in Empirical Strategy).

Event studies.

Because the instrument in our 2SLS approach interacts GM-crop suitability with year indicators, its reduced-form and first-stage results provide convenient graphical summaries, i.e., event studies. We define suitability as a county’s maximum percentile of attainable yield across the GM crops corn, soy, and cotton—scaled between 0 and 1 (further discussed in Attainable Yield). One can interpret the event study’s yearly coefficients as summarizing the difference between the average health outcome in high-GM-crop suitability counties relative to low-GM-crop suitability counties each year (relative to the 1995 difference). The event study coefficients depict how infant-health differences between high- and low-suitability counties evolved through time.

These event studies also provide evidence on the plausibility of the parallel-trends assumption upon which the reduced-form DID results rely. Because the event studies are also reduced-form results, causal interpretation only relies upon the parallel-trends assumption, rather than the stronger exclusion restriction required by 2SLS.

Separating glyphosate’s direct and policy effects.

Interpreting these causal effects requires some nuance. Glyphosate-tolerant GM seeds allowed farmers to change their weed management practices—reducing their usage of nonglyphosate pesticides and mechanical tilling (6). These—and other—changes could potentially affect perinatal health, violating the exclusion restriction. Therefore, we present two effects in our results—a policy effect and a glyphosate effect.

The policy effect does not control for these other changes and thus captures the total (net) effect of the introduction of GM seeds. For our reduced-form DID results, the policy effect is potentially the more policy-relevant parameter as it aggregates across all of the health effects brought about by GM seeds and glyphosate intensification relative to alternative practices in place before the introduction of GM seeds. For this reason, the exclusion restriction for the 2SLS-estimated policy effect represents a fairly strong assumption—violated if the GM-seed rollout induced employment or nonglyphosate agricultural changes that affected infant health.

Alternatively, by directly controlling for changes induced by the GM rollout—nonglyphosate pesticides, fertilizer, population composition, and various local economic conditions—the glyphosate effect provides an estimate of the direct causal effect of glyphosate on perinatal health. This direct glyphosate effect still requires an exclusion restriction: Except for the pesticides, fertilizers, population composition, and economic outcomes for which we control, no other mechanisms affected perinatal that also correlated with both suitability for GM crops and the timing of the GM-seed rollout. Accordingly, the glyphosate effect is likely more appropriate for the two-stage approach—permitting a per unit of glyphosate interpretation with a plausible exclusion restriction.

Finally, glyphosate-based herbicides like Roundup typically contain other chemicals, e.g., surfactants that reduce surface tension. Our measured effects include impacts from other ingredients mixed with glyphosate in commercial herbicide formulations.

Heterogeneity.

In addition to estimating the average effect of glyphosate exposure (methods described above), we also estimate heterogeneous effects. Understanding heterogeneity in this setting is critical: Health and policy implications can differ depending on whom glyphosate impacts and how diffuse the impacts are. Reductions in birthweight among infants already at risk for low birthweight are likely more costly than decreases in birthweight among higher-weight infants.

We estimate glyphosate’s impact as a function of the infant’s expected birthweight, which we estimate using a random forest trained to predict infants’ expected birthweights. Specifically, this learning model predicts birthweight as a function of the infant’s and parents’ information—using data from all pre-1996 births, before glyphosate-resistant seeds were widely available. (See SI Appendix, section H for detailed methodology.) The resulting predictions enable us to test whether glyphosate’s adverse health impacts affected infants equally or whether they concentrated in low- or high-weight births. Using predicted birthweight allows us to avoid bias from splitting the sample on the outcome (47). Subsequently, we estimate our reduced-form and 2SLS results by predicted-birthweight percentiles.

Results.

Reduced-form evidence of GM rollout’s health effects.

The event study in Fig. 2A demonstrates that GM crop suitability strongly predicts post-1996 increases in glyphosate application levels. Because GM crop suitability is a continuous percentile, the plotted coefficients represent the difference in glyphosate intensity between the highest and lowest GM suitability counties relative to their difference in 1995. Before 1996, lower- and higher-GM-crop-suitable counties followed similar glyphosate trajectories: The event study remains relatively flat and near zero. However, after the 1996 introduction of GM seeds, glyphosate intensity in counties with higher attainable yields for corn, soy, and cotton quickly outpaced glyphosate intensity in less-suitable counties. As adoption accelerated, this glyphosate intensity gap between high- and low-attainable yield counties widened. The event study confirms the strength of our instrument (percentile of GM-crop suitability) and illustrates the first stage of our 2SLS approach.

Fig. 2.

GM-seed introduction increased glyphosate intensity in GM-crop suitable areas; birthweight reductions were also higher in GM-suitable counties and match GM-seed timing. (A) Estimated event-study coefficients for the effect of local GM max attainable yield percentile on glyphosate by year relative to 1995. Pesticide data only go back to 1992 (no coefficients in 1990–1991). (B) Similar event study for birthweight. Both regressions represent the policy-effect specification: county and year-by-month fixed effects and controls for parent demographics (mother’s age, race, education, marital status, birth facility, resident status, previous births, fathers age and race, and sex of infant). SEs cluster by state and year. Sample restricted to births occurring in rural counties or to mothers residing in rural counties.

The event-study results mirror the reduced-form DID estimates for the policy effect on glyphosate in the first row of Panel A of Fig. 3 (Left columns). The introduction of GM seeds significantly increased the glyphosate intensity in high-GM-crop suitability counties by 0.023 kg/km² relative to low-suitability counties. This effect is large—approximately the same level as the 2012 mean glyphosate intensity.

Fig. 3.

Glyphosate’s direct and policy effects reduced birth outcomes for the average rural birth. All reported estimates are the effect at the weighted mean of glyphosate in 2012, where we weight by total births. Panel (A) reports difference-in-difference estimates comparing high- and low-GM suitability counties before and after the release of GM seeds in 1996. We use 1990–1995 as our preperiod and 2005–2010 as the postperiod to drop the transition phase. Panel (B) reports 2SLS estimates from 2SLS regression of perinatal health on glyphosate, instrumenting for glyphosate with GM suitability interacted with year dummies. The Policy effect controls for family demographics. The glyphosate effect controls for other pesticides and unemployment. See text for details. LBW and VLBW give the probabilities of low birthweight ($<$2,500 g) and very low birthweight ($<$1,500 g) in percentage points (0 to 100). 95% CIs calculated using SEs clustered by year and state. Sample restricted to births occurring in rural counties or to mothers residing in rural counties.

These first-stage findings align with a plausible mechanism: Higher attainable yields fostered greater GM crop adoption, increasing glyphosate application. SI Appendix, Fig. S4 shows applications of several prevalent herbicides—alachlor, cyanazine, fluazifop, and metolachlor—decreased following the release of GM seeds, suggesting farmers substituted away from these herbicides and toward glyphosate. Consequently, we control for these pesticides—along with fertilizer, several employment measures, and population composition—when estimating the glyphosate effect and omit them when estimating the policy effect.

Fig. 2B depicts a similar event study for birthweight—plotting each year’s average birthweight difference between births in high and low GM-crop suitability counties (relative to the 1995 difference) after controlling for maternal demographics. Prior to the introduction of GM crops, the birthweight gap between births in higher- and lower-attainable-yield counties remained stable. However, beginning in 1996—coinciding with the release of glyphosate-tolerant seeds and the intensification of glyphosate application—birthweights in higher-GM-suitability counties began declining relative to lower-yield counties. In 2005, a decade after the introduction of glyphosate-tolerant seeds, the average birthweight in the highest-yield county had fallen approximately 30 g relative to the lowest-yield county.

Again, the reduced-form DID estimates for the policy effect on birthweight in Panel A of Fig. 3 match the event study. Relative to low-suitability counties, the average infant in high-suitability counties lost 29.4 g from the introduction of GM seeds and the ensuing intensification of glyphosate.

The event study in Fig. 2B also visualizes the reduced-form estimates of our 2SLS estimator: the effect of high attainable yield for GM crops on birthweight over time. The flat trend prior to 1996 in the event study coefficients of Fig. 2B also supports the parallel-trends assumption that underpins our empirical approach: Before the introduction of GM-tolerant seeds, counties followed similar trajectories independent of GM-crop suitability.

The remainder of Panel A in Fig. 3 reproduces similar reduced-form DID policy-effect estimates for five additional health outcomes. The results depict similar, statistically significant adverse health effects for perinatal health. In higher-suitability counties, the average birth’s gestation length declined 1.03 d, probabilities of low birthweight (LBW) and very low birthweight (VLBW) increased 0.51 pp and 0.11 pp, probability of preterm increased 1.64 pp, and health index declined by 0.011. Reassuringly, the probability of C-section was unaffected. SI Appendix, Fig. S5 provides reduced-form event studies for each of these results.

The reduced-form DID estimates for the glyphosate effect on the Right side of Panel A (Fig. 3) tell a very similar, slightly attenuated, story: The introduction of GM seed significantly reduced perinatal health in high-suitability counties (LBW and VLBW are significant at the 10% level). The fact that directly controlling for potential confounders of glyphosate does not substantially alter the results lends plausibility to the exclusion restriction required by our 2SLS. Together, the reduced-form event-study and DID results bear considerable evidence that the release of GM seeds in 1996 led to significant deterioration of perinatal health in high GM-attainable-yield counties relative to low-yield counties.

The effect of glyphosate on perinatal health.

Our two-stage results integrate the variation depicted in Fig. 2 into a 2SLS framework—effectively dividing the health-outcome event study in Fig. 2B by the glyphosate event study in Fig. 2A. In fact, scaling the reduced-form DID by the corresponding glyphosate effects generates very similar estimates to 2SLS (SI Appendix, Table S4). Panel B in Fig. 3 reports the estimated effect (and 95% CI) of glyphosate exposure at the 2012 mean level of glyphosate intensity (0.023 kg/km²). SI Appendix, Table S3 contains the corresponding regression coefficients.

As discussed above, the glyphosate effect (i.e., controlling for potential violations of the exclusion restriction) is likely the more appropriate parameter to consider for the two-stage results. Accordingly, we focus our discussion of the 2SLS results on the glyphosate effect. However, Fig. 3 provides 2SLS policy-effect estimates for comparison.

The 2SLS glyphosate-effect results in Panel B of Fig. 3 (Right side) indicate that for the average rural birth, exposure to glyphosate at its 2012 mean intensity significantly reduced birthweight by 29.8 g; reduced gestation 1.49 d; increased the probabilities of LBW, VLBW, and preterm by 0.65, 0.20, and 2.14 pp; and reduced the composite health index by 0.013. Comparing these effects to each outcome’s mean value suggests that the effects are indeed meaningful: Relative to the outcomes’ means, the effects represent 0.9% of mean birthweight, 0.6% of mean gestation, 8.1% of LBW, 14.3% of VLBW, 10.3% of preterm probability, and 87.7% of the health index. Again, the effect on the probability of a C-section is positive but not statistically significant and small in magnitude. Together, these results indicate that even at the mean level of intensity in the United States, glyphosate exposure significantly deteriorates infant health.

These glyphosate effects result from local (own-county) exposure. Unlike Dias et al. (10) and Skidmore et al. (11), we do not find any significant effects of glyphosate sprayed upstream—SI Appendix, section N.2 contains the results of our water-based analysis.

These findings are robust to a number of modeling alternatives. SI Appendix, Fig. S6 provides 2SLS estimates for the marginal effect of glyphosate on birthweight, where we vary the inclusion of controls, fixed effects, and the definition of attainable yield. Event studies for birthweight under these alternative models are also robust (SI Appendix, section F), and specification charts for other outcomes are in SI Appendix, section G. We also estimate OLS results without instruments in SI Appendix, section C and a “shift-share” model with slightly different identifying assumptions in SI Appendix, section E. Finally, SI Appendix, Fig. S21 shows results from estimating the model on different geographic subsets of the United States. In each case, the results are qualitatively unchanged.

We also investigate various threats to identification. A primary concern for our approach is whether the release of GM seeds coincided with nonglyphosate socioeconomic effects which also affected birthweight—e.g., employment or income. Notably, there is no evidence that the introduction of GM seeds significantly affected average farm or nonfarm income in the study counties. The event study for the unemployment rate does show a trend—suggesting the potential for bias if excluded (SI Appendix, Fig. S28). However, controlling for unemployment does not meaningfully change our results, suggesting employment changes are not driving our results.

Heterogeneity in glyphosate’s health impacts.

The results in the previous sections estimate the average effect of glyphosate within a heterogeneous population—where the effect of glyphosate may differ across individuals. Fig. 4 shows the two-stage estimates for policy and glyphosate effects on birthweight vary considerably by decile of expected birthweight. SI Appendix, section H explains our prediction of birthweight. Both effects exhibit considerable heterogeneity that follows the same pattern: The magnitude of glyphosate’s estimated effect on birthweight is largest in the lowest decile (i.e., for births with the lowest expected birthweight), and the effect declines as the percentile of predicted birthweight increases. SI Appendix, Fig. S9 shows first-stage and reduced form results for birthweight by predicted birthweight quintile. The glyphosate effect on birthweight in the first decile is 12 times larger than in the tenth decile. Accordingly, glyphosate’s largest impact concentrates among the most vulnerable births: The estimated glyphosate effect in the first decile at mean 2012 glyphosate exposure is a loss of 75 g, relative to just 6 g in the tenth decile of predicted birthweight.

Fig. 4.

Birthweight losses due to glyphosate and GM are largest for births with the lowest expected birthweights. Estimates of the policy and glyphosate effects on birthweight by predicted birthweight deciles. Each estimate is from a separate regression. All regressions include controls for family demographics, county fixed effects, and year by month fixed effects. Sample restricted to births occurring in a rural county or to mothers residing in a rural county. SEs are clustered by year and state. Glyphosate instrumented with GM max attainable yield percentile.

We also find significant evidence that glyphosate’s effects vary with expected birthweights for several other health outcomes—gestation length and the probabilities of low and very-low birthweight. The increase in the probabilities of low birthweight and very-low birthweight are more than 60 times larger in the lowest decile than in the top decile of expected birthweight. SI Appendix, Fig. S10 illustrates heterogeneity in the policy effect for all outcomes by predicted birthweight quintiles, deciles, and ventiles. As with birthweight, the most acute effects for gestation occur in lower percentiles, declining in magnitude with predicted birthweight percentile. Unlike our birthweight results, where we find no significant effect among higher percentiles, we find statistically significant evidence that glyphosate reduces gestation length in every decile. At the 2012 mean level of glyphosate exposure, the glyphosate’s effect on gestation ranges from $-$2 to $-$0.8 d for the average birth. Again, these results suggest the most vulnerable infants bear the largest perinatal health impacts.

Discussion

This study finds significant evidence that glyphosate adversely affected births across several measures of perinatal health throughout the rural United States in the last twenty years. To our knowledge, these findings are the first quasi-experimental evidence of glyphosate’s adverse health effects at a population scale in the United States. However, they are consistent with a growing body of literature documenting glyphosate’s negative impact on development and reproduction. Using machine learning to predict expected birthweights, we find significant heterogeneity underlying the effect of glyphosate on perinatal health: Glyphosate’s adverse effects are largest for infants with the lowest expected birthweights.

Other Effects.

While this estimated effect of GM seeds and glyphosate on perinatal health highlights a critical consequence of herbicide application, it represents only part of the total potential externality associated with increased chemical usage. The previously discussed toxicology literature suggests several additional mechanisms and effects on human health that we do not measure here. Lab experiments (28–30) and recent observational work suggest additional ecological costs, such as biodiversity loss (48).

Although it does not address all of glyphosate’s effects, using birthweight as our primary outcome offers several advantages. As discussed above, prior research establishes adverse effects of glyphosate on reproduction/development and finds glyphosate in the urine of nearly every expectant mother. Numerous studies then link birthweight to other later-life outcomes (49, 50). Additionally, data on birthweight are widely available and reliable—the United States has accurately and objectively measured the birthweight of nearly every child born for decades. Infants are also less prone to endogenous responses to health and environmental shocks. Similarly, infants have short histories over which they can accumulate exposure—simplifying exposure measurement. Birthweight thus provides both a key health outcome and a canary-in-a-coalmine-like indicator of glyphosate toxicity. Ample opportunities exist for future work to test additional costs and benefits of glyphosate and GM technology to more broadly assess welfare effects.

Heterogeneity, Mechanisms, and Inequity.

Several mechanisms potentially explain the observed heterogeneity. While we cannot fully disentangle these mechanisms, we use the dimension along which we observe heterogeneity—predicted birthweight—to understand which attributes correlate with the most-affected births. As predicted birthweight results from a predictive model trained on the demographics of infants’ parents, we observe which parental attributes correlate with lower predicted birthweight and, consequently, more adverse glyphosate effects. Fig. 5 reveals that lower predicted-birthweight infants are more likely to be female, Black, and/or children of unmarried parents. The predictive model places nearly all births to Black parents into the first quintile, where we detect the most adverse impacts of glyphosate. If we estimate our main model separately for births to White and non-White mothers, we find that the glyphosate effect is 1.8 times larger for births to non-White mothers relative to White mothers (SI Appendix, Fig. S7). This finding links to an extensive environmental justice literature that documents Black households’ unequal pollutant exposure (51).

Fig. 5.

Female infants, children of Black and non-White parents, and children of unmarried parents have lower predicted birthweights. Each line represents the percent of births (y-axis) within the demographic group (line color) at each predicted birthweight percentile (x-axis) for births to mothers with rural residences. Predicted birthweight percentile is calculated relative to the distribution of rural-residence births in years prior to 1996. Averages cover two-percentile bins.

We observe similar effects on male and female births within birthweight quintiles (SI Appendix, Fig. S11). Therefore, the increased effect among low-predicted birthweights is not driven by a larger effect among females. However, female births still bear a larger health burden since they make up a larger proportion of low-predicted births.

The fact that the adversity of glyphosate’s effects correlates with race and sex highlights a potentially serious issue for equity. Further, we find the most adverse effects among the lowest expected-weight births—potentially magnifying short-term healthcare costs (52), later-life outcomes/welfare (50), and epigenetic/intergenerational consequences (49). Consequently, our results have potentially important implications for equity and justice in the United States.

Unfortunately, data limitations restrict us from further testing potential mechanisms. While our empirical approach recovers the causal effect of glyphosate on perinatal health, it does not provide causal estimates for the drivers of heterogeneity. One potential mechanism for the heterogeneity is differential exposure. Because the natality and glyphosate data are only resolved at the county, we cannot test whether infants with lower predicted birthweights face higher levels of glyphosate relative to other infants in their county. Differential baseline health (or healthcare access) could also contribute to the observed heterogeneity. Further, the shape of glyphosate’s damage function is unknown. If glyphosate has nonlinear effects on health—nonlinear in glyphosate exposure or due to interactions with other health risks/complications—differential exposure or heterogeneous nonglyphosate health risks could also produce these heterogeneous impacts. Finally, our examination of heterogeneity focuses on only one of many possible dimensions. Future work could contribute to many of these issues with more resolved data.

Magnitude of Results.

The results in this paper indicate glyphosate exposure adversely affects perinatal health. Whether the estimated effects are large in magnitude depends upon several dimensions of heterogeneity and potentially on additional nuance.

It is important to consider that the effects presented in Fig. 3 are average effects, in two senses, namely, averaged across all rural infants and calculated for exposure to the average glyphosate intensity in 2012 (approximately 0.023 kg/km²). For birthweight, this average effect of glyphosate is a reduction of 20 to 40 g (1%). The corresponding implied increases in probabilities of low birthweight and preterm birth (0.4 to 0.6 pp and 1.4 to 2.1 pp, respectively) represent percent increases of approximately 5 to 9% and 7 to 10%. If all infants faced a 20 to 40 g reduction in birthweight (or 1%), there may be minor concern for public health, with the exception of the most vulnerable infants—e.g., the lowest-weight infants for whom marginal benefits of weight may be high. Indeed, Currie et al. consider a similar birthweight effect small (53). However, our estimated average effect on birthweight is larger than impacts estimated by several recent studies on regulated pollutants—e.g., 13.2-g reduction from a one-standard-deviation increase in late-term CO exposure (54) and 6 to 7 g reduction for each “cheating” diesel Volkswagen (per 1,000 cars) (55). The magnitude of our average effect is quite close (but opposite sign) to estimates for the effect of the Food Stamp Program (56).

Our estimates for the average effects on the probabilities of low birthweight and preterm births are very similar (both in level and percent) to recent work on exposure to leaking underground petroleum storage tanks (8.7% LBW; 7.4% preterm) (57), automobile-based air pollution (11.8% LBW; 10.8% preterm) (58), late-term exposure to CO (8.6% LBW) (54), and in utero exposure to chemicals via drinking water (6.5% LBW) (59). Our estimated average effect of glyphosate on the probability of low birthweight is approximately one third of estimated effects for infants born to families within 1 km of a fracking site (53).

However, Fig. 4 highlights significant heterogeneity underlying the average effects discussed above. Considering both the heterogeneity depicted in Fig. 4 and additional heterogeneity in glyphosate intensity is critical in assessing the magnitude of the estimated effects.

As Fig. 4 illustrates, the estimated impact of glyphosate on birthweight for infants in the lowest expected-birthweight decile is 5 to 12 times larger than the highest decile—and approximately 2.5 times larger than the average effect. At the mean level of 2012 glyphosate exposure, the estimated effects for infants in the first decile are 46 to 77 g—suggesting infants in this already-low-weight group lost 2 to 4% of their expected birthweight due to glyphosate. We find even larger heterogeneity on the probability of low-weight birth: The effect (0.7 to 1.1 pp increase) for the lowest decile is 4 times larger than the average effect and 50 to 60 times larger than the effect for the top decile.

The paper presents results as effects from exposure to the mean level of 2012 glyphosate intensity—e.g., Fig. 3 (approximately 0.023 kg/km²). However, glyphosate intensity varies substantially in the United States: Exposure is 50% higher at the 75th percentile, 200% higher at the 90th percentile, and 300% higher at the 95th percentile. Accordingly, at the 90th percentile (i.e., for 10% of rural US births), we estimate that the average (pooling across expected birthweight) loss in birthweight is 62 to 95 g—i.e., 2 to 3% of the mean birthweight. Exposure to glyphosate at the 90th percentile of glyphosate intensity increases the probabilities of low birthweight and preterm birth by 1 to 2 pp (LBW) and 4 to 7 pp, respectively. At the 90th percentile, our estimated effects for the probability of low birthweight are on the same scale as the effect of living within 1 km of a fracking site (53). The relatively large size of effects implied at higher percentiles of pesticide intensity is consistent with the results of Larsen, Gaines, and Deschênes, who find adverse infant health effects from pesticides concentrate in the top 5 percent of exposure (37). It is important to note that this discussion assumes glyphosate’s damage function is linear (SI Appendix, section I discusses further). Regardless of the damage function’s shape, only considering average exposure may obscure key insights into glyphosate’s health impacts within the US population.

The discussion above separately considers the two dimensions of heterogeneity. However, low-expected birthweight infants are not all exposed to the average glyphosate intensity. In fact, the correlation between expected birthweight and glyphosate intensity is quite low (0.02). For a first-decile expected birthweight infant, exposed at the 90th percentile of glyphosate (affecting approximately one percent of rural births), the estimated weight lost due to glyphosate is 146 to 243 g—a loss of 5 to 9% of the infant’s expected weight. The corresponding calculation for the probability of a low-weight birth implies an increase of 5 to 8 percentage points. These estimates suggest a small subset of the population may bear particularly large health burdens from glyphosate exposure.

Finally, the paper’s results likely present a lower bound for the true impact of glyphosate due to aggregation bias (ecological fallacy) inherent to the data—understating glyphosate’s actual health burden for exposed individuals. Because the match between birth outcomes and glyphosate exposure is only possible at the county level, it is likely that some infants with little true exposure to glyphosate received high glyphosate exposure in the data (the reverse is also likely). The resulting aggregation bias attenuates the differences between true high- and low-exposure individuals—likely attenuating our estimates (51). Future work with better measurements of mothers’ exposures could resolve this issue as recent work has done in other settings (53, 54).

Broader Considerations.

GM crops and the resulting glyphosate intensification profoundly changed agriculture. In this study, we quantify one health externality caused by these technological innovations, which reduced average birthweight by 23 to 32 g at the average level of glyphosate exposure.

To put the estimated glyphosate health damages in perspective, we convert them to dollars. Waitzman, Jalali, and Grosse estimate that a preterm birth costs an additional 82 thousand USD (2023 dollars) relative to a full-term birth (60). This estimate includes additional medical care following the birth, special education expenses, and lost labor market earnings later in the child’s life. These estimates may omit additional costs from glyphosate. We combine our estimates from Fig. 3 on the increased probability of preterm birth and the number of total births in 2012. This calculation implies the economic costs of the reduced-form policy effect were 800 million USD annually and just over 1 billion USD annually for the 2SLS glyphosate effect (both in 2023 dollars).

Our findings, combined with other recent work (10, 11, 27), challenge the prevailing regulatory position that GM crops and their associated agricultural practices are safe—and even beneficial—for health. Advocacy historically argued that glyphosate is less toxic than the herbicides it replaced. The mounting evidence of the negative health externalities associated with the rollout GM crops and the ensuing glyphosate intensification warrant new policy discussions about informed, efficient, and equitable regulation of these technologies. Efficient policy must carefully weigh these practices’ economic benefits against the adverse health effects we identify and other human/ecological costs. Further, policymakers must consider the unequal burden glyphosate appears to levy.

Additional work is needed to better understand the benefits and costs of GM crops, glyphosate, and the specific exposure mechanisms underlying their effects. For instance, unlike Dias et al. (10) and Skidmore et al. (11), we do not find evidence of health effects from upstream glyphosate use (see SI Appendix, section N). This difference could result from differences in water treatment, measurement error in upstream-glyphosate exposure, or other meteorologic/geologic/hydrologic factors.

The US EPA’s current approval process for agricultural chemicals is rife with opportunities for regulated entities to affect the review’s outcome—including a reliance upon non-peer-reviewed, applicant-generated studies (ignoring conflicts of interest) (61–63); selective reporting/testing (61, 64); flawed research designs (65), implementation (66), and analyses (63); limited postapproval monitoring (61); and a general lack of transparency/oversight (67). (See SI Appendix, Regulatory oversight for an extended discussion.) Uncertainty around glyphosate and GM health impacts has not slowed the spread of these technologies. Neither has this uncertainty led to substantive monitoring that would enable regulators or researchers to precisely estimate damages, disentangle exposure mechanisms, or understand heterogeneity in glyphosate’s damage function. Despite the relative dearth of data, consistent evidence is emerging that GM-spurred glyphosate intensification adversely affects health. Nevertheless, without further research, improved monitoring/data, and reoptimized policies, the public will likely continue to bear the health burden of glyphosate inefficiently and unequally.

Materials and Methods

Infant Health Data.

We use the universe of births in the United States between 1990 and 2013 from the NVSS. The birth data contain information recorded on the birth certificates, including month and year of birth, sex, birthweight, APGAR score, live birth order, total birth order, whether the birth was a C-section, gestation length, and dummies for a battery of birth defects. The birth certificates also contain certain demographics of the mother and father—namely race, ethnicity, age, education, marital status, and residence status. We have access to the restricted human-subjects exempt (68) versions of these files, which identify the county of birth and county of mother’s residence for all births, compared to the publicly available data, which hides geographic identifiers for counties with less than 100,000 residents.

Pesticide Use Estimates.

Our measure of glyphosate use comes from the United States Geological Survey’s Pesticide National Synthesis Project (7, 69). The USGS surveyed farmers to calculate pesticide use rates per acre of different crops planted at the crop reporting district level. They then multiply these usage rates by the total acreage of each crop within the crop reporting district to estimate the total amount of each pesticide used, measured in kilograms. Each pesticide has two estimates, high and low, where the high value assumes crop reporting districts with a missing usage rate for a pesticide applied the pesticide at the same rate as their neighbors on each crop. The low value assumes a missing usage rate for a pesticide means that farmers did not apply that pesticide. We use the high value throughout our analysis. Additionally, we normalize by the total area of each county, thus our measure of glyphosate and other pesticides are in kilograms of active ingredient per square kilometer (kg/km²).

Attainable Yield.

We use the FAO-GAEZ attainable yield for soybeans, corn, and cotton to measure the suitability of a county for genetically modified crops (45). These data assign potential yield values to one square kilometer pixels based on environmental factors such as soil type, slope, and climate. We aggregate the pixels to the county by taking the average of all pixels within a county. We take the difference between the high-input attainable yield and the low-input attainable yield to focus on the counties with the largest incentive to adopt GM crops. The underlying model calculates the high-input scenario assuming that farmers have access to modern technology for crop management, including GM seeds. Whereas the model calculates the low-input scenario assuming more traditional farming methods. We aggregate the three GM crops by standardizing the attainable yield difference so that each crop has a mean of zero and a SD of one. We then take the maximum across the standardized yield differences and then rescale this average into national percentiles. SI Appendix, section B discusses the validity of the GAEZ data in light of new work comparing historical yields to GAEZ predictions (70).

Additional Data.

We supplement our analysis with several additional data sources. First, we classify counties as rural using the 2003 Rural–Urban Continuum codes from the USDA (mapped in SI Appendix, Fig. S1). A rural county is any nonmetro county, where the USDA defines a metro county as, “broad labor-market areas that include central counties with one or more urban areas with populations of 50,000 or more people. They also include outlying counties that are economically tied to the core counties as measured by labor-force commuting” (71).

We use workforce data from the Bureau of Labor Statistics Local Area Unemployment Statistics. These data report the annual average number of employed and unemployed workers going back to 1990. We supplement these data with annual, county-level income and employment data split by farm and nonfarm from the Bureau of Economic Analysis. They report these data going back to 1969 and source the data from the US Bureau of Labor Statistics and Internal Revenue Service.

Finally, we use annual county-level acreage and yield of various crops—including the main GM crops, corn, soy, and cotton—from the USDA National Agricultural Statistics Service. Additionally, we collect data on county-level fertilizer use from the USGS (72). We employ an interpolation process described in SI Appendix, section B to generate annual fertilizer estimates for each county.

Empirical Strategy.

We aim to estimate the causal effect of glyphosate exposure—induced by the roll-out of GM seeds—on infant health. To isolate exogenous variation in glyphosate use, we leverage temporal variation from the commercial release of GM crops in 1996 and spatial variation in GM adoption due to differences in environmental suitability for growing those crops.

Define a linear model for the effect of local glyphosate $\left(\beta \right)$

$${\text{Health}}_{\mathit{ijt}}=\beta {\text{GLY}}_{\mathit{jt}}+\mathrm{\Gamma }{X}_{\mathit{ijt}}+{\alpha }_j+{\lambda }_t+{\epsilon }_{\mathit{ijt}}$$ [1]

for individual $i$, in county $j$, in year $t$. ${\text{GLY}}_{\mathit{jt}}$ represents local glyphosate exposure, measured as the total mass of glyphosate sprayed per square kilometer of total area in county $j$ in year $t$ (kg/km²). $X_{\mathit{ijt}}$ provides a vector of controls. ${\alpha }_j$ and ${\lambda }_t$ are county and month-of-sample (e.g., January 2012) fixed effects. Estimating Eq. 1 with OLS is unlikely to identify the true effect of glyphosate on health due to endogeneity and measurement error.

The main endogeneity concern is that the adoption of GM technology and glyphosate may correlate with unobservable factors that affect perinatal health. To rectify the measurement and endogeneity issues, we use instruments—in a reduced-form DID and a 2SLS estimator—that isolate exogenous variation in local glyphosate exposure. Our instruments are the maximum percentile of attainable yield for GM crops for county $j$, denoted ${\text{GM}}_j$, interacted with dummy variables for each year (described in Attainable yield). The first stage in our 2SLS estimator is

$${\text{GLY}}_{\mathit{ijt}}=\sum _{\tau \ne 1995}{\gamma }_{\tau }{\text{GM}}_j\times \mathbb{1}(t=\tau )+\mathrm{\Gamma }{X}_{\mathit{ijt}}+{\alpha }_j+{\lambda }_t+{\epsilon }_{\mathit{ijt}}$$ [2]

2SLS uses predictions from Eq. 2 in place of ${\text{GLY}}_{\mathit{jt}}$ in Eq. 1 to estimate a causal effect of glyphosate on perinatal health.

While we have microdata (birth-level) on birth outcomes, we do not have precisely measured glyphosate exposure leading to measurement error. We expect some mothers within a county are highly exposed to glyphosate while others are not. However, data limitations force us to assign the same level of exposure to all births within a county. This measurement concern relates to the ecological fallacy and likely leads us to underestimate the magnitude of glyphosate’s damages to an individual’s health. As described by Banzhaf, Ma, and Timmins (51) “When measuring the correlation between pollution and demographics, the ‘ecological fallacy’ can arise when inferring relationships between individual units (like households) from larger, more aggregated units (like counties) that contain those units.”

Identification.

Throughout the paper, we employ two complementary empirical approaches: a reduced-form DID and 2SLS. Both approaches use similar instruments that combine spatial variation in the suitability for GM crops with temporal variation from the introduction of glyphosate-tolerant seeds. While thus related, the two approaches rely upon different assumptions for causal identification.

Reduced-form DID.

The reduced-form DID results rely upon a parallel-trends assumption: Had GM seeds not been released, then the relative difference between more and less GM-crop-suitable counties (high- and low-attainable yield counties) would have remained constant in terms of infant health—conditional on fixed effects and controls.

Differential trends in higher and lower attainable-yield counties, prior to the release of GM crops, would violate this assumption. A single outlier year for higher- or lower-yield counties could also affect our results despite a smooth increase in glyphosate through time. To assess these concerns, we estimate the event-study model

$${\text{Health}}_{\mathit{ijt}}=\sum _{\tau \ne 1995}{\gamma }_{\tau }{\mathrm{GM}}_j\times \mathbb{1}(t=\tau )+\mathrm{\Gamma }{X}_{\mathit{ijt}}+{\alpha }_j+{\lambda }_t+{\epsilon }_{\mathit{ijt}}$$

interacting $G{M}_j$ (continuous percentile of GM crop suitability) with year indicators. The measured effects $\left({\gamma }_{\tau }\right)$ provide the difference in perinatal health between the highest and lowest attainable yield counties relative to their 1995 difference (the year before the GM-seed rollout and ensuing glyphosate intensification). County fixed effects absorb average differences between higher and lower yield counties. Consequently, an event study with no trend in the ${\gamma }_{\tau }$ prior to treatment (pre-1996) supports our identifying assumption for the reduced-form DID.

Event studies in Fig. 2 and SI Appendix, Fig. S5 support this assumption: Prior to 1996, the event studies are relatively flat and near zero—bearing no strong evidence of differential trends in infant health or (glyphosate intensity) related to counties’ GM suitability. After the GM-seed rollout higher-suitability areas accelerate in glyphosate intensity and decline in perinatal health outcomes.

The panel variation in our instrument is critical for identification, as it permits a comparison of pre-GM vs. post-GM outcomes across higher- and lower-suitability areas. Without this panel variation (and the shock of GM seed’s rollout), identification would exclusively come from cross-sectional pesticides’ differences and could easily conflate health effects from glyphosate exposure with cross-sectional health differences that correlate with the geophysical determinants of GM suitability. Instead, reduced-form estimates come from comparing the difference between higher- and lower-GM-suitability counties after the GM rollout to the difference between higher- and lower-GM-suitability counties before the GM rollout. Thus, even if the geophysical determinants of GM-crop suitability correlate with cross-sectional differences in health (e.g., healthcare access), our identification only relies upon more- and less-suitable areas following similar trajectories (i.e., parallel trends). Cross-sectional differences will be absorbed by our fixed effects. Again, the timing of the effects depicted in the event studies supports the idea that differences between high- and low-GM-suitable counties were stable until the introduction of GM seed and the glyphosate intensification. These points are also relevant for the 2SLS results discussed next.

2SLS.

Our two-stage approach estimates Eq. 1 with 2SLS, instrumenting glyphosate exposure with GM-crop suitability interacted with year indicators (Eq. 2). Interpreting these 2SLS estimates as the causal effect of glyphosate requires the instrument is relevant to glyphosate intensity, exogenous, and satisfies an exclusion restriction (46).

The relevance of the instrument is evident Fig. 2A: On average, after the GM rollout, glyphosate intensity increased more in higher-GM-crop-suitability counties than in lower-suitability counties.

The exogeneity of this instrument—and thus the variation it extracts—comes from the fact that its spatial variation (GM suitability) is a function of environmental factors unaffected by GM/glyphosate adoption and its temporal variation comes from the arbitrary timing of GM seed’s rollout. The event studies discussed above support this exogeneity.^*

Interpreting 2SLS estimates as the causal effect of glyphosate hinges on the instrument satisfying an exclusion restriction. In our context, this exclusion restriction is that conditional on our controls, GM-crop suitability (interacted with year) only affected infant health through changes in glyphosate use. In other words, GM-crop suitability (interacted with year) should only affect infant health via increased glyphosate exposure, conditional on controls. If the rollout of GM technology induced other, nonglyphosate changes in more-GM-suitable counties and these changes had infant-health effects (after conditioning on covariates and fixed effects), then the exclusion restriction would be violated.

Nonglyphosate pesticides replaced by glyphosate potentially violate this exclusion restriction. Consequently, 2SLS without controls for nonglyphosate pesticides could be biased (likely downward) for glyphosate’s actual health effects—failing to exclude health improvements from reduced exposure to nonglyphosate pesticides. Fertilizer, employment, and income changes present similar potential to violate the exclusion restriction. Thus, we emphasize that the glyphosate effect is the appropriate estimand for 2SLS—directly controlling for potential exclusion-restriction violations. With these controls and the instrument’s exogeneity, we believe the exclusion-restriction is plausible.

Finally, settings with potentially heterogeneous treatment effects require an additional assumption: monotonicity. In our context, monotonicity requires that increasing attainable yield in a county would result in that county using weakly more glyphosate—i.e., increasing suitability for GM crops would not reduce a county’s glyphosate intensity. This assumption appears quite reasonable in our setting, and we have no reason to believe that decreasing a county’s suitability for corn, soy, or cotton would increase glyphosate, all else equal.

Birth-Outcome Index.

Because we estimate the impact of glyphosate on multiple birth-health outcomes, we are testing multiple hypotheses. Accordingly, the stated error rate of each hypothesis test may not accurately reflect the joint error rate. Following Currie et al., to assuage multiple-hypothesis-testing concerns, we construct a single birth-health index from our five variables of interest—birthweight, gestation length and indicators for low birthweight ($<$2,500 g), very low birthweight ($<$1,500 g), and preterm ($\le$37 wk). (Because C-section functions as a falsification test, we exclude it from our index.) This single index provides a single test for composite health at birth upon which one may focus without concerns regarding multiple inference (73). Following Currie, Greenstone, and Meckel, the index is a weighted average of the variables, where the weights use inverse-covariance weighting—capturing variables’ covariance and differences in their variances (53).

Predicting Birthweight.

Models like Eq. 1 identify a single, average effect—as in Fig. 3. If individuals respond differently to glyphosate exposure, an average effect obscures this heterogeneity. Further, this heterogeneity is potentially critical to understanding the health impacts and policy remedies of the documented losses of birthweight—for example, if the lost birthweight came from lower-weight infants, as opposed to higher-weight infants. Thus, who lost weight could be key. We estimate heterogeneity in the effect of GM seeds and glyphosate on infant birthweight and gestation as a function of the infant’s birthweight percentile—with two nuances. The first nuance centers on an issue of causal identification; the second is operational.

First, rather than using an infant’s actual birthweight percentile, we use the percentile of the infant’s predicted birthweight. We train a random forest to predict birthweight using features from the NVSS that describe the infant, mother, father, and birth location.^† The training sample for this prediction is the universe births in the contiguous US before the mass introduction of GM crops and the ramp-up of glyphosate application (before 1996)—i.e., before our treatment began. We tune^‡ the model on 80% of the pre-1996 births using five-fold cross-validation. Finally, we train the selected (minimum RMSE) model in a five-fold pattern—ensuring each prediction comes from a model that has not seen the predicted individual. This hold-out approach in the prediction step and our predicted birthweight approach help avoid bias in the heterogeneity regressions. This bias could arise because birthweight—our heterogeneity dimension—is the outcome variable; conditioning on the outcome can introduce endogeneity (47). Instead, we condition on predicted birthweight, which is a function of 1) the infant’s family’s observable features and 2) other, pretreatment infants’ birthweights. Because our predictions are relatively accurate (predicted birthweight is, on average, quite close to actual birthweight as shown in SI Appendix, Fig. S8), we can estimate how glyphosate differentially affects lower- and higher-birthweight infants without introducing bias from conditioning on the outcome.

The second nuance relates to the structure of the heterogeneous treatment effect. Because the shape of the heterogeneity is unknown, we take a semiparametric approach that allows us to remain relatively agnostic. We split the sample using infants’ predicted birthweight percentiles (e.g., deciles) and then separately estimate the 2SLS model for each group. For example, Fig. 4 contains the estimated effect of glyphosate on birthweight for the first through tenth deciles. This bin-based approach is commonly applied to recover heterogeneous treatment effects, as it allows one to approximate arbitrary nonlinear functions without making stronger assumptions about functional form (74). This approach still returns an average treatment effect estimate within each bin/group. SI Appendix, Fig. S10 shows robustness to different bins in predicted birthweight across all of our main outcomes.

This approach—using ML-based counterfactuals in a causal-inference setting—relates to recent work by Burlig et al. (75). While we use the predictions to estimate heterogeneity in our treatment effect, Burlig et al. take advantage of high-frequency electricity-consumption data for each individual (school)—using ML-based counterfactuals to identify the causal effect of energy-efficiency upgrades (and heterogeneity thereof). Another related approach—Athey and Wager’s causal forests (76)—provides a more generalized approach for using random forests to estimate conditional average treatment effects in higher dimensions, permitting interactions and nonlinearities. Because the heterogeneity of interest for this paper is only along a single dimension (expected birthweight), our bin-based approach provides sufficient flexibility.

Exposure to Glyphosate Sprayed Upstream Through Water.

We use a spatial water model to estimate glyphosate exposure based on the amount of glyphosate sprayed upstream of each county using a methodology similar to that of Dias et al. (10). The HydroBASINS watershed shapes form the building blocks of this model (77). We summarize the process here but leave the details in SI Appendix, section N.

The amount of glyphosate that runs off into surface water will be affected by the erodibility of the soil, the slope of the land, and precipitation. We collect soil erodibility and slope data from the USGS gridded soil survey in each watershed, which we aggregate to the watershed level by taking the average over all grid cells within a watershed (78). We take the interaction of soil erodibility and slope and then convert that interaction into a percentile based on the distribution from all watersheds in the United States. These data are static and do not change over time.

Next, we use gridded monthly precipitation from the PRISM climate group to capture whether there was potential for glyphosate to run off (79). We aggregate rainfall during the growing season (April through September) for each watershed and again convert it into a percentile from the distribution of all watershed-month-years.

We then map our county-level attainable yield percentile to watersheds and take the interaction between high erodibility, high precipitation, and attainable yield to create an instrument for upstream glyphosate use. We expect there to be effects from upstream spraying only when there is both high soil erodibility and high precipitation. We also estimate the effect of high attainable yield upstream without the interaction with soil erodibility and precipitation. The HydroBASINS data allow for easy linking of upstream and downstream watersheds. In the linking process, we calculate the distance between two watersheds by tracing along centroids of all watersheds between those two watersheds. This measure allows us to aggregate variables over 50-km distance bins upstream and downstream from each watershed. The downstream variables serve as a nice placebo test since we do not expect glyphosate sprayed downstream to affect infant health.

Once we estimate values upstream of each watershed, we must aggregate to the county level to analyze them with our health metrics. We take the weighted average of the upstream variables for all watersheds in a county, where the weights are the portion of the county’s population that lives within that watershed. We use population estimates for one square kilometer pixels from NASA’s Socioeconomic Data and Applications Center to calculate the population weights for each watershed (80).

In addition to the binned approach described above, we use a machine-learning model to predict concentrations of glyphosate in surface water using a limited dataset of samples taken across the United States. We utilize the geographic structure and physical characteristics of land, along with spatially disaggregated herbicide use, to predict downstream concentrations of glyphosate in surface water. We then regress perinatal health outcomes on these predictions on glyphosate and AMPA exposure from water. The details are in SI Appendix, section N.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Code, publicly available data, and instructions for gaining access to the restricted birth certificate data have been deposited on GitHub (https://github.com/edrubin/glyphosate-birthweight) (81). Due to our data-use agreement with the CDC NCHS, we are not able to share the infant health data that form the backbone of our analysis. However, we have all of our code in a GitHub repository, which also provides instructions for gaining access to the restricted birth certificate data. This repository contains the rest of the data we use in our analysis, which is all publicly available).