Several epidemiologic studies have shown associations between biomarkers of prenatal exposure to pesticides and poorer childhood neurodevelopment [1
]. We have also shown in the Center for the Health Assessment of Mothers and Children of Salinas (CHAMACOS) birth cohort study that the average of two measurements of organophosphate pesticides (OPs) dialkyl phosphate (DAPs) metabolites in maternal urine collected during pregnancy was associated with decrements in cognitive development of 7-year-old children living in an agricultural community [11
]. However, although this biomarker potentially represents exposure to a multitude of OPs, it does not represent all OPs used and it does not indicate exposure to specific pesticides. To address this limitation, we previously investigated prenatal residential proximity to agricultural pesticide-use using California’s unique database of Pesticide Use Data (PUR), which provides the poundage (kg), date, and location (to one-square-mile sections) of specific pesticides (active ingredient) applied; we found significant adverse associations between cognitive development in children at 7 years of age and specific agricultural pesticides or grouped pesticide classes, namely oxydemeton-methyl, acephate, pyrethroids, neonicotinoids, and Mn-fungicides [12
]. However, due to the inherent limitations with conventional multiple regression techniques—such as limited capacity to handle large numbers of highly correlated exposures—our previous analyses and others’ [1
] have relied on multiple tests of association with neurodevelopment for individual pesticides or pesticide groups, and we therefore have yet to evaluate the potential neurodevelopmental effects of joint exposure to several different individual pesticides with different mechanisms of toxicity.
Since agricultural populations are potentially exposed through a variety of routes to a combination of pesticides that exhibit varying toxicities and modes of action [14
], a modeling approach is needed which limits multiple tests of association (that enhance type 1 error rate) while also handling multicollinearity. Since many pesticides are neurotoxic, albeit of varying potency (even within chemical pesticide classes) [15
], and some have similar mechanisms of action [15
], there is also a potential for either additive or non-additive effects given different combinations of pesticide exposures. Statistical approaches have recently emerged to facilitate analysis of the combined health effects of joint exposure to multiple environmental chemical exposures [19
Agricultural communities present an opportunity to investigate the potential health effects of exposure to multiple pesticides. California presents a strong case study, due both to its relatively high volume of pesticide use—approximately 85 million kg in 2014 [28
]—and the fact that 100% of all agricultural pesticide use is reported to the state’s Department of Pesticide Regulation [29
]. Past environmental and biomonitoring studies that have utilized PUR data have shown positive correlations between nearby reported agricultural pesticide use and pesticide concentrations measured in outdoor air and house dust [30
]. We have also found that the CHAMACOS mothers—during pregnancy—had higher OP urinary metabolites levels compared to the general U.S. population [34
] and that living near agricultural fields is related to higher levels of urinary OP metabolites at younger ages in the children [35
]. To date, however, little is known about the spatial patterning of joint use near residents in agricultural communities and whether different joint use patterns may be more strongly associated with health outcomes.
In this study, we investigated the joint distribution of potentially neurotoxic pesticide-use profiles during pregnancy with measures of childhood cognition. We included in this analysis pesticides not previously considered because they do not devolve to DAPs and/or have no known biomarker. We also explored whether there was a distinct spatial patterning of some of the combinations of pesticide-use that showed stronger associations with child cognition. To accomplish this, we employed a novel statistical method known as Bayesian profile regression (BPR) [23
] to analyze pesticide use profiles that are estimated from prenatal residential proximity to reported agricultural pesticide use. The BPR approach is based on well-established Bayesian Dirichlet process mixture modeling techniques [36
] and is capable of: (1) accounting for the collinearity of the exposure data inherent with agricultural pesticide use; (2) appropriately handling model uncertainty in cluster assignment and number of clusters used; and (3) drawing inference on health associations by linking profiles of exposure with a health outcome of interest. In this way, we employ a joint exposures approach to identify and characterize the important clusters of prenatal pesticide profiles that are asociated with deficits in childhood cognition. We also aimed to identify individual pesticides that are most strongly associated with childhood cognition when considering joint exposure by applying Bayesian Kernel Machine Regression (BKMR), a data-adaptive method that allows for fitting multiple correlated exposures jointly into the same model, and evaluating each parameter’s relative importance across the model space.
Our primary aim was to use Bayesian profile regression (BPR) to characterize joint distributional patterns of correlated agricultural pesticide usage near pregnant women’s homes among a cohort residing in the Salinas Valley, CA and evaluate clustered pesticide profiles in relation to full scale IQ (FSIQ) in their children at seven years of age. The BPR partitioned the joint pesticide distributions effectively given that the analysis resulted in distinct pesticide profiles whether or not the outcome was included in the analysis. Including the outcome in the analysis, however, enabled us to characterize the association of different clusters of pesticide profiles with FSIQ while also accounting for model uncertainty in the Bayesian framework. We found that pesticide profiles with elevated joint distributions of multiple pesticides showed the strongest associations with deficits in childhood FSIQ (e.g., >1/2 SD from baseline group); conversely, pesticide profiles with joint pesticide-use distributions that were substantially lower (CP3 and CP6) were characterized by FSIQs that were above the study population’s overall mean FSIQ. Importantly, mapping of the pesticide use profiles suggested a distinct spatial dependency for the most hazardous pesticide profiles, along with evidence of positive spatial autocorrelation (clustering) in higher pesticide levels near maternal residences.
While previous studies suggest that prenatal proximity to agricultural pesticide use for individual pesticides is associated with neurodevelopmental outcomes [12
], ours is the first to explicitly show that it may be the combined exposure to multiple different pesticides and pesticide classes near maternal residences that may be of importance rather than nearby use of a single pesticide or single class of pesticides. Our findings also imply that studies investigating the neurodevelopmental effects of current-use agricultural pesticide exposures should analyze the pesticide mixtures as a whole, rather than using the more conventional single pollutant models typically applied [12
], as such methods are unable to control for the influence of other covarying pesticides.
Conventional multivariate regression struggles to reliably estimate combined effects for a large number of exposures within the context of highly correlated exposures. This observation is particularly true when the number of parameters are large and the number of observations are small, as is common in most cohort studies. For instance, in previous work by our group on this same cohort and using this same PUR data set using separate pesticide-specific regression models, we found adverse associations between single pesticides and FSIQ [12
]. With conventional regression, we were previously unable to analyze multiple pesticide exposures jointly and thus unable to evaluate the potential for combined effects. In addition, while several pesticides in our previous work showed linear exposure-response relationships with FSIQ, the highly correlated nature of the pesticides meant that we could not disentangle the true magnitude of associations for each pesticide, which likely over-attributed the effect of a single pesticide. These challenges strongly support the need for a more advanced and diverse set of statistical approaches to investigate health effects from multiple pesticide exposures [20
]. Implementing Bayesian clustering as a framework of analysis allowed us to evaluate potential joint patterns of multiple pesticide exposures, the spatial patterning of potential joint exposures, as well as adverse childhood cognitive associations with joint exposure to multiple correlated pesticides.
Our findings lend support to the idea that exposure to mixtures of pesticides may have unanticipated effects on neurodevelopment in children. Given the results from our previous study mentioned above [12
], and working under the assumption of additive effects of pesticides on FSIQ, we would expect to see substantially larger deficits in FSIQ between pesticide profiles with large differences in cumulative exposure levels, but this was not evidenced in our analysis. Pesticide profile cluster 1 (CP1) resulted in a cumulative pesticide-use profile that was more than two and half times higher than CP7 (670 kg vs. 235 kg), yet relative to the pesticide profile with the lowest cumulative pesticide-use (CP3) both CP1 and CP7 exhibited differences in ΔFSIQ (∆FSIQCP1
= −6.9 vs. ∆FSIQCP7
= −6.4, Table 3
) that did not correspond to the magnitude of the difference in cumulative pesticide use levels. Additionally, after weighting the two clusters by their respective OP toxicity based on relative potency factor, as done in our previous study [12
], we observe that CP1 is nearly two and half times higher (data not shown) in its estimated toxicity weighted use, yet this difference in apparent toxicity was also not reflected in our results. In addition, CP2 resulted in a cumulative (summed) pesticide use estimate that was slightly higher than that of CP7, yet the difference in FSIQ from the lowest pesticide profile cluster for CP2 was substantially different compared to that of CP7. A potentially important difference between these two clusters (CP7 and CP2) was that acephate and thiodicarb were classified as “very high” in CP7, whereas these two pesticides were only “moderately high” in CP2. This finding reveals the possibility that combined exposures to multiple pesticides may not result in assumed additive effects for each compound in a mixture [20
], as has been seen in some toxicity studies [64
]. This finding, however, should be taken with caution, since our study lacked a direct measure of prenatal exposure to pesticides. Also, unlike conventional regression models, profile regression is limited by the fact that it does not assume additivity of effects from multiple exposures because it partitions continuous joint exposure distribution into discrete clusters, which essentially represent latent categorical variables. This points towards an important limitation in using clustering-based methods, whereby some clusters may result in joint exposure distributions too wide to elucidate the effect of an individual chemical within a cluster on the outcome, especially where the signal is not particularly strong [22
Given some of the limitations with profile regression already discussed, supplementary to profile regression, we also implemented Bayesian kernel machine regression (BKMR). Even though BKMR has its own set of limitations, the results indicate that oxydemeton-methyl, acephate, and maneb were particularly important pesticides in the observed exposure profile associations with FSIQ, and these same pesticides were elevated in the pesticide profile clusters that exhibited the largest deficits in FSIQ. While there is no clear guidance with BKMR in how to group exposures with respect to hierarchical variable selection, our sensitivity analysis of exposure groupings showed consistency in terms of the importance of these three pesticides with respect to showing the strongest associations with FSIQ.
Despite the numerous potentially neurotoxic pesticides evaluated in the present study, our analysis revealed only eight distinct pesticide-use pesticide profiles when including the outcome and nine pesticide profiles when excluding the outcome in the BPR analysis. This suggests that pesticide exposures in agricultural communities may occur in a relatively small number of discrete patterns of exposure, which could reflect the small number of different types of crops near maternal residences that have only a certain combination of pesticides applied. This finding, however, is consistent with a recent French study, wherein a set of pesticide mixtures that the French population is potentially exposed to was similarly identified using a Dirichlet process mixture model [66
]. Using a combination of dietary and pesticide residue information on 79 different pesticides, researchers observed that just 25 pesticides contributed to the clustering, with a total of only seven exposure profiles observed from their analysis [66
]. In a follow-up study, Crépet et al. [65
] evaluated the toxicological effects of these seven different pesticide mixtures in vitro and found certain exposure mixtures effects went either beyond or below predicted additivity effects and that toxic effects were not readily predicted based on exposure to each individual compound within a pesticide profile. Our finding and those of Crépet et al. [65
] highlight the potential value in examining joint pesticide exposure patterns to help prioritize specific “real-world” exposure-response combinations that can be tested in toxicological studies [65
]. BPR could also be extended to take “real world” scenarios of exposures profiles to multiple pesticides to predict potentially adverse neurological effects [60
A promising aspect of our analytic approach that is worthy of further exploration is determining the particular biologic drivers in the clustering of the pesticide profiles. Even though the correlation structure of pesticide use patterns clearly played an important role in how these profiles clustered together, it is less clear the degree to which the disease sub-model in BPR determined the clustering or whether biologic factors related to neurotoxicity may be important as well. Our sensitivity analysis suggested an important role for the outcome in partially determining the clustering patterns observed in our data (Table 4
). Considering that the clustering of pesticide profiles was sensitive to the outcome and that pesticide-specific neurotoxic effects are inherently dependent on biologic pathways, the clustering observed in our study may be driven by similar or dissimilar mechanisms of toxicity, metabolism, or distribution for certain combinations of pesticides. For instance, the ordering of OP and carbamate anti-acetylcholinesterase activity, or possibly some other biologically plausible pathway, may counteract the cognitive effects of multiple neurotoxic pesticides to explain the potential for sub-additive effects [68
]. Unfortunately, these questions of biologic drivers cannot be readily elucidated in our study data because we lacked the appropriate biomarkers of exposure. This area represents an important avenue for future research.
An important implication of our study points toward the value in examining spatial patterns of exposure profiles related to agricultural applications. Our mapping of pesticide profiles demonstrated that the profiles associated with the largest deficits in childhood FSIQ exhibited a distinct spatial pattern suggestive of spatial clustering in the southern Salinas Valley and along the outer border of the City of Salinas. Importantly, a test for spatial autocorrelation using Moran’s I test failed to reveal evidence within our study population of spatial clustering for lower FSIQ scores, signifying that spatial residual confounding in our outcome due to unmeasured sub-population characteristics is unlikely to explain the spatial patterns observed for the highest risk clusters. We further observed that the highest risk pesticide profiles tended to be on the outskirts of the towns closer to agricultural fields, especially for the City of Salinas. Hence, cluster analysis of pesticide-use patterns combined with spatial information on participants’ residences can be leveraged to map communities that are most likely to be disproportionately impacted by hazardous environmental chemical mixtures [21
]. Such spatial information is potentially useful for stakeholders, including public policy makers, growers who apply pesticides, and members of the public potentially impacted by application of multiple pesticides to agricultural fields. This approach can also be extended for other purposes such as in evaluating the spatial patterns of hazardous air pollution mixtures [21
We did not evaluate potentially neurotoxic agricultural herbicides or fumigants in our analysis of pesticide mixtures, which is an important limitation. Future studies, with larger sample sizes, are thus needed to evaluate a broader class of neurotoxic pesticides and other potentially neurotoxic chemicals. Another important limitation is a lack of validation of our pesticide exposure measure (i.e., PUR data) either with biomarkers or environmental measures for all of the pesticides considered. The application of pesticides near maternal residences during pregnancy does not necessarily mean the women were exposed to these pesticides during their pregnancy, or that the relative proportion of exposure is represented by the relative use of the active ingredients. In addition, the potential exposure to the pesticides considered in our study vary by application method and their physicochemical properties including volatility, degradation rates, deposition rates, and other characteristics that will ultimately determine their spatiotemporal fate and transport in the environment; yet these factors were not considered in our exposure assessment [69
]. Thus, relying solely on residential proximity to reported pesticide applications can lead to exposure misclassification and potentially bias our findings towards the null. Several studies (including in the CHAMACOS cohort) demonstrate some positive correlations between nearby reported agricultural pesticide use based on PUR data and environmental contamination (i.e., house dust and outdoor air) or proximity to agricultural fields and pesticide metabolite levels in biological samples [30
]. We were unable to fully account for other potential sources of pesticide exposure such as residential use of pesticides and ingestion of pesticide residues from fruits and vegetables. Although we controlled for total prenatal DAP concentrations, which is a strength of our study since it likely represents other sources (i.e., diet and take-home from resident farmworkers) of prenatal OP pesticide exposures that cannot be ascertained solely from PUR data, these exposure biomarkers have their own limitations. DAPs are non-specific to a particular OP, they do not include exposure to certain OPs such as acephate [73
], they do not represent exposure to other non-OP pesticides, they do not represent long-term exposures [74
], and they may reflect exposure to preformed metabolites and not simply their parent compounds [75
]. Measured total prenatal DAPs were not correlated with any of the PUR pesticide use estimates, thus reinforcing the concept that total DAPs are likely to be representative of other OP pesticide exposure sources and that DAPs do not sufficiently represent long-term exposure levels. That being said, we removed DAPs from the model as a sensitivity analysis and did not see any substantive difference in our modelling results (data not shown). Furthermore, use of PUR data as a proxy of exposure allows us to evaluate health associations for pesticides that currently do not have a reliable biomarker of exposure. PUR data also allows us to evaluate which pesticides are driving the clustering and observed associations with health, as opposed to the non-specificity of DAPs, which obscures the variation in toxicity between OPs. There is a clear need to develop improved biomarkers with better specificity and sensitivity for a wider array of pesticides that represent long-term exposures.
Future work should attempt to overcome the limitations noted above, for instance, by improving characterization of pesticide exposure by developing predictive models based on measured pesticide concentrations in house dust, outdoor air, and personal samplers. Future work should also estimate proximity to neurotoxic pesticide use during the postnatal period using residential history information in addition to location of daycare facilities and schools that children attended. Furthermore, other chemical mixtures analytic frameworks may be applied to this or similar data sets to better characterize the possible contribution of individual pesticides to adverse neurologic effects while considering all exposures jointly [19
]. Finally, we encourage other researchers to attempt to replicate our findings in other studies, especially in studies containing larger sample sizes.