Journal list menu

Volume 49, Issue 1 p. 152-162
TECHNICAL REPORTS
Open Access

Causal factors for pesticide trends in streams of the United States: Atrazine and deethylatrazine

Karen R. Ryberg

Corresponding Author

Karen R. Ryberg

USGS, 821 East Interstate Ave., Bismarck, ND, 58503

Correspondence

Karen R. Ryberg, USGS, 821 East Interstate Ave., Bismarck, ND 58503.

Email: [email protected]

Search for more papers by this author
Wesley W. Stone

Wesley W. Stone

USGS, 5957 Lakeside Blvd., Indianapolis, IN, 46278

Search for more papers by this author
Nancy T. Baker

Nancy T. Baker

USGS, 5957 Lakeside Blvd., Indianapolis, IN, 46278

Search for more papers by this author
First published: 28 February 2020
Citations: 12

Assigned to Associate Editor Zhiming Qi.

Abstract

Pesticides are important for agriculture in the United States, and atrazine is one of the most widely used and widely detected pesticides in surface water. A better understanding of the mechanisms by which atrazine and its degradation product, deethylatrazine, increase and decrease in surface waters can help inform future decisions for water quality improvement. This study considers causal factors for trends in pesticide concentration in U.S. streams and models the causal factors, other than use, in structural equation models. The structural equation models use a concomitant trend in corn (Zea mays L.) and a latent variable model, indicating moisture supply and management. The moisture supply and management latent variable model incorporates long-term moisture conditions in the individual watersheds by using the Palmer hydrologic drought index, human influence on the hydrologic cycle through the percentage of the watershed drained by tile drains in 2012, and the base-flow contribution to streamflow, using the base-flow index. The structural equation models explain 77 and 38% of the variability in atrazine and deethylatrazine trends, respectively, across the conterminous United States. The models highlight future water quality challenges, particularly in tile-drained settings where fall precipitation and heavy precipitation are increasing.

Abbreviations

  • BFI
  • base-flow index
  • DEA
  • deethylatrazine
  • GC–MS
  • gas chromatography–mass spectrometry
  • LVM
  • latent variable model
  • MSM
  • moisture supply and management
  • MWHs
  • multiple working hypotheses
  • PCA
  • principal components analysis
  • PHDI
  • Palmer hydrologic drought index
  • SEAWAVE-Q
  • a trend analysis method with seasonal wave and adjustment for streamflow
  • SEM
  • structural equation modeling/model
  • 1 INTRODUCTION

    The attribution of trends in pesticide concentration in surface water is challenging because of the quality and quantity of data needed to support attribution and because pesticide chemical properties vary greatly and govern how pesticides move through the environment. However, better understanding these mechanisms is important because of water quality concerns and the importance of pesticides to society. United States expenditures for conventional pesticides (active ingredients other than biological pesticides and antimicrobial pesticides) were approximately U.S. $13 billion in 2009 and $14 billion in 2012 (Atwood & Paisley-Jones, 2017). Approximately two-thirds of total expenditures were for agricultural purposes in 2009 and 2012 and most of those agricultural expenditures were for herbicides (Atwood & Paisley-Jones, 2017). The most commonly used herbicides in 2005, 2007, 2009, and 2012 were glyphosate and atrazine (ranked first and second, respectively, in Table 3.4 of Atwood & Paisley-Jones, 2017), both of which are used extensively on corn (Zea mays L.) (USGS, 2018a).

    Atrazine is a selective pre- or postemergent herbicide that is used to control broadleaf weeds and some grassy weeds by inhibiting photosynthesis in susceptible plants (Fishel, 2006; USEPA, 2006). Although urban use of atrazine is less than agricultural use, atrazine was one of the most widely detected herbicides for the urban streams analyzed for trends in a 1992–2008 period (Ryberg, Vecchia, Martin, & Gilliom, 2010). Atrazine has moderate sorption to soil, whereas its primary degradation product, deethylatrazine (DEA), has low sorption; both compounds are moderately persistent in soil, and atrazine is moderately persistent in water (Lewis, Tzilivakis, Warner, & Green, 2016; USEPA, 2012). The primary conversion of atrazine to DEA is through metabolic activity of soil fungi and bacteria. Hence, atrazine is likely to degrade to DEA when atrazine has more contact with soil microorganisms, such as in nonpoint source contamination. Atrazine is less likely to degrade to DEA when there is less contact with soil microorganisms, such as in point-source contamination (Scribner et al., 2005).

    Atrazine use varies across the conterminous United States, but the major crop beneficiary is corn (Figure 1). On a national scale, use declined from 2002–2012, in part because of regulatory actions by the USEPA and an increase in the use of glyphosate on corn with a glyphosate-resistant trait (Coupe & Capel, 2015). However, changes in use and corn acreage vary among watersheds, with reported atrazine use increasing in areas that previously had little or no reported use (USGS, 2018b) as corn acreage expanded (USDA-NASS, 2018).

    Details are in the caption following the image
    Atrazine use, in million kilograms, in the conterminous United States, 2002–2012 (USGS, 2018a)

    From past studies, we know that most pesticide surface-water concentration trends are similar to pesticide use intensity trends, including atrazine trends (Ryberg & Gilliom, 2015; Sullivan, Vecchia, Lorenz, Gilliom, & Martin, 2009). For DEA though, past studies have shown that uptrends can happen during periods with significant downtrends in both the use and concentrations of atrazine (Ryberg & Gilliom, 2015; Ryberg, Vecchia, Gilliom, & Martin, 2014). Ryberg and Gilliom (2015) hypothesized that atrazine use and DEA concentration trends differ because of some factor, such as a management practice, that has over time increased the proportion of applied atrazine that runs off to streams as DEA, or by a transport pathway for DEA, such as groundwater, that has lags between use and arrival at a stream. Decreases in atrazine use would eventually lead to decreases in DEA, but there is a lag time during which the trends may be in opposing directions. Gilliom et al. (2006) reported that DEA/atrazine ratios were generally higher in groundwater than streams, reflecting the longer periods of time spent in contact with soil for the atrazine compounds detected in the groundwater system, relative to streams, because degradation takes place with the assistance of soil microorganisms. Once atrazine is degraded to DEA, it is much more soluble (Mackay, Shiu, & Ma, 1997), meaning it is more likely to move with water through the soil. Hence, the setting and environmental conditions that influence degradation can be important in determining the drivers of trends in pesticide concentration other than use.

    Past studies concluded that reductions in concentrations because of improved management practices (those unrelated to use reduction) might be difficult to discern (Ryberg & Gilliom, 2015; Sullivan et al., 2009) and that more precise estimates of uses and ancillary data on specific management practices would likely be needed to assess the large-scale effects. The National Water-Quality Assessment Project of the USGS National Water-Quality Program has been developing such datasets and methods for modeling the effects of factors that influence water quality. In a desire to better support statements about the causes of trends and move beyond speculation, simple bivariate correlation (such as that between use and concentration), or citation of such work by others when making those statements, structural equation modeling (SEM) has been used to test causal hypotheses about the drivers of changes in water quality (Ryberg, 2017; Ryberg, Blomquist, Sprague, Sekellick, & Keisman, 2018). This is a continuation of that effort to better understand the factors, other than use, that influence trends in pesticide concentration in streams.

    Oelsner et al. (2017) completed an assessment of water quality trends for US rivers and streams that included pesticide trends for the periods 1992–2012 and 2002–2012. The pesticide concentration and streamflow datasets used to evaluate the pesticide trends, as well as the trend results, were published in a data release (Ryberg et al., 2017). A subset of those results was used in this study to examine atrazine and DEA, the compounds with the most calculable trends across the United States.

    We developed multiple working hypotheses (MWHs) for the causal factors, or drivers, of pesticide trends in general and of atrazine and DEA in particular. We then acquired available data related to these hypotheses and used SEM to test these causal hypotheses. This allows us to better understand drivers of concentration trends beyond pesticide use.

    Core Ideas

    • Atrazine trends are mostly downward across the conterminous United States.
    • The major driver of atrazine trends is the concomitant trend in corn acreage.
    • Deethylatrazine trends are mostly upward across the conterminous United States.
    • Deethylatrazine trends are driven by moisture supply and management and corn acreage.

    2 MATERIALS AND METHODS

    2.1 Site and chemical selection and trend analysis

    A subset of the trend results presented in the assessment of national water quality trends by Oelsner et al. (2017) was used for this study. The sites were all USGS water quality monitoring sites, and the chemical compounds were analyzed at the USGS National Water Quality Laboratory using gas chromatography–mass spectrometry (GC–MS; Lindley, Stewart, & Sandstrom, 1996; Madsen, Sandstrom, & Zaugg, 2003; Zaugg, Sandstrom, Smith, & Fehlberg, 1995). The pesticide concentration data were prepared for trend analysis by adjusting concentrations to compensate for bias resulting from temporal changes in recovery of the GC–MS analytical method (Martin & Eberle, 2011) and by recensoring routine nondetections by site and pesticide. The screening process to determine the sites with sampling representative of the trend period and chemicals with sufficient detections for trend analysis, as well as the data processing steps, was that described in Oelsner et al. (2017, p. 22). The trend analysis method was the seasonal wave model with adjustment for streamflow, SEAWAVE-Q (Ryberg and Vecchia, 2013; Sullivan et al., 2009; Vecchia, Martin, & Gilliom, 2008). The SEAWAVE-Q model was specifically developed to address the challenges of pesticide trend analysis, including strong seasonality driven by use, a large percentage of concentrations below laboratory reporting levels, complex streamflow and concentration relations, and changing sampling frequencies (Vecchia et al., 2008). The model is a parametric regression model that incorporates flow-related variability in the form of streamflow anomalies (Ryberg and Vecchia, 2012; Vecchia, 2003), seasonality in concentration data, application season of pesticides, timing of maximum concentrations, and a decay rate representing the decline in concentration after the application season (Ryberg and Vecchia, 2013; Vecchia et al., 2008).

    Selected trend results were used for this study, based on which chemicals had trend results for at least 60 sites across the country. Sixty sites were deemed necessary to have a sufficient sample size for making inferences about the drivers of pesticide concentrations using SEM. This limited us to the period 2002–2012 and the two chemicals, atrazine and its degradation product DEA. Sixty-seven sites have trend results for atrazine, and 62 sites have trend results for DEA. The pesticide concentration and streamflow datasets used for this study were published in a USGS data release (Ryberg et al., 2017). Oelsner et al. (2017) also present results for 1992–2012, but few sites had a complete record for the longer period.

    2.2 Multiple working hypotheses framework and causal factor data

    Developing MWHs was an integral part of the SEM process, distinguishing this confirmatory process from more exploratory forms of multivariate analysis. We developed MWHs potentially describing the causes of changes in pesticide concentrations in surface water to systematically consider causal mechanisms and available data. Supplementary Material 1 of Ryberg et al. (2018) provides background on the MWHs framework, which was suggested >100 yr ago (Chamberlin, 1890) and has recently resurfaced in hydrology (Clark, Kavetski, & Fenicia, 2011; Harrigan, Murphy, Hall, Wilby, & Sweeney, 2014).

    After developing the MWHs (see Supplement 1) for pesticide concentration trends in surface water, we hypothesized that the most important factors driving trends in atrazine concentration, other than use, would be corn acreage (expressed as a concomitant trend in proportion of watershed harvest corn; Falcone, 2017), weather and climatic conditions (represented by the Palmer hydrologic drought index [PHDI] and available in Falcone, 2017) that might influence cropping and runoff patterns over the trend period, and tillage or conservation practices. We hypothesized that DEA trends might also be influenced by the same factors but might differ from atrazine in that groundwater contribution to surface water might transport DEA, and that this could be represented by the base-flow index (BFI). Base flow is the sustained, slowly varying component of streamflow, usually attributed to groundwater discharge to a stream, and the BFI is a ratio of base flow to total streamflow, expressed as a percentage and ranging from 0–100 (Falcone, 2017).

    We examined the potential for the working hypotheses using bivariate correlation analysis. Although correlation is not causal and bivariate correlations are not necessarily indicative of variable behavior in a multivariate model (e.g., once a series is adjusted for one variable, another relation may become apparent), the bivariate correlations alerted us to non-normalities in the data, outliers, and potentially useful relations. The climate, agricultural, and other watershed characteristic data used have all been published by Falcone (2017), and the code and correlation plots are available in Supplement 2.

    2.3 Testing causal hypotheses with structural equation modeling

    To test causal hypotheses, we used SEMs. This type of modeling has some similarities with multiple regression models (Grace, 2006). Both SEM and multiple regression models allow one to specify relations prior to statistical analysis based on known or hypothesized causal mechanisms and allow one to compare models. However, SEM offers more flexibility in how these models may be specified and allows for more complexity, given a sufficient sample size. For example, the response variable in one regression equation in an SEM may be a predictor in another regression equation within the same SEM. Additionally, variables in an SEM may influence one another directly or through other variables as intermediaries, or mediating variables. These structural equations are meant to represent causal relations among the variables (Fox, 2002). Another advantage of SEM is that covariances between variables, which might cause multicollinearity problems in multiple regression, can be explicitly modeled in SEMs.

    Structural equation models can have one or more latent variables that are modeled as latent variable models (LVMs). A reflective LVM is indicated (measured) by observed (manifest) variables (indicators) that in some cases may not be significant on their own but contribute to the latent influence. One could construct an LVM representing agricultural practices, for example, in which all agricultural practices are not measured, but multiple factors (indicators of practices), such as the percentage of land in agricultural production, proportion of watershed in harvested corn, and percentage of agricultural land in the Conservation Reserve Program, are observed and join in an overall agricultural effect on water quality. Latent variable models represent the underlying structure (covariance) that produced relations among the indicator variables and need enough nonredundant information to generate unique parameter estimates (Beaujean, 2015). An LVM is similar to a common factor in factor analysis. An LVM is also similar to principal components analysis (PCA; Hotelling, 1933) in that PCA can use many variables reduced to principal components. These principal components are ascribed a meaning in terms of processes or inputs and are used in an analogous manner to SEM LVMs to describe broad mechanisms. Principal component analysis does not, however, allow users to specify relations among the variables and many find the dimension reduction difficult to understand. Structural equation modeling has the advantage that it is based on causal pathways that can be visualized in a causal graph.

    The number of possible variables for an SEM was constrained in this study. Structural equation modeling is a large sample method (sample size considerations and “rules of thumb” are described further in Supplement 1 of Ryberg, 2017), and that meant that we needed to investigate the chemicals with the most trend results in Oelsner et al. (2017) and model causal hypotheses using relatively simple national models, rather than complex models for a specific state or region. Given the sample sizes of 67 and 62 for atrazine and DEA, respectively, the SEM models had to be constrained to <10 parameter estimates to maintain stability of the estimates.

    Maximum likelihood methods require multivariate normality of endogenous variables, which are internal or dependent variables in an SEM model, whereas no distributional assumptions are required for exogenous (or entirely explanatory) variables (Eliason, 1993; Kline, 2012). The chi-squared test statistic, used as a first test of overall model fit, can be biased and indicate rejections of satisfactory models when there are distributional violations (Rosseel, 2012). A nonsignificant result (p > .01) indicates the model fits relatively well because the null hypotheses is that the hypothesized model fits the data, whereas a significant result would indicate that the hypothesized model is not adequate (Hoyle, 2012). The models used in this study passed the chi-squared test and, therefore, we did not transform variables.

    We used the software package lavaan, latent variable analysis (Rosseel, 2012), for the statistical software R (R Core Team, 2018) to fit the SEM with maximum likelihood methods (the argument “estimator” set equal to “ML” in the sem function of lavaan). Additionally, in the sem function, the fixed.x argument was set to TRUE, fixing the means, variances, and covariances of the exogenous variables (those with no arrows pointing toward them in the causal graph, independent variables) to their sample statistics (Rosseel, 2012, 2016).

    We assessed model fit using four measures. The first was the abovementioned chi-squared test from the output of sem function in lavaan (Rosseel, 2012). The second measure of model fit used was the standardized root mean square residual (SRMR), which does not compare models or account for model complexity, but simply measures absolute fit and ranges from 0–1, with values closer to 0 indicative of a better fit (Rosseel, 2012). The third measure was the comparative fit index (CFI; Beaujean, 2015; Bentler, 1990), which ranges from 0–1, with values closer 1 indicative of a better fit (Bagozzi & Yi, 2012; Rosseel, 2012). The fourth measure was the root mean square error of approximation (RMSEA), which considers model complexity and penalizes models with more parameters and ranges from 0–1, with values closer to 0 being better (Beaujean, 2015). These measures of model quality and their ideal ranges are summarized in Supplemental Table S4 of Ryberg (2017).

    3 RESULTS AND DISCUSSION

    3.1 Trend overview

    Although the trends have been previously reported (Oelsner et al., 2017; USGS, 2017), the trend directions and geographic patterns are important for understanding the SEM models. Numerical trend results are available in a USGS data release (Ryberg et al., 2017). Maps showing the sites assessed and the geographic distribution of trends are shown in Figures 2 and 3 and are available in an interactive USGS map application of water quality changes at https://nawqatrends.wim.usgs.gov/swtrends/, along with changes in many other water quality constituents (USGS, 2017). The trends are presented as “likely” down or up, “somewhat” likely down or up, or “about as likely as not” to have a trend based on a likelihood approach as an alternative to a significance testing approach.

    Details are in the caption following the image
    Atrazine concentration trends for the conterminous United States, 2002–2012 (Oelsner et al., 2017; Ryberg et al., 2017)
    Details are in the caption following the image
    Deethylatrazine concentration trends for the conterminous United States, 2002–2012 (Oelsner et al., 2017; Ryberg et al., 2017)

    The likelihood approach is an effort to provide more intuitive information when summarizing trend results. Trend likelihood values for pesticides were determined using the p value reported from the SEAWAVE-Q trend model, using the equation 1 – (p/2). This approach is explained in detail in the user guide of the USGS map application (USGS, 2017), and that explanation is modified here to specifically refer to pesticides. Consider an example where the chance of an upward trend in atrazine concentrations at a site is 80 out of 100 (a trend likelihood value of .80). Using the significance testing approach and a traditional significance level of .05 or .01, the trend would be reported as nonsignificant. Using the likelihood approach, the trend would be reported instead as “somewhat likely up.” The likelihood approach indicates that it is somewhat likely that conditions in the stream are not improving, giving resource managers more information to use when making decisions about watersheds. For example, in Figure 2, there are numerous “likely down” trends in atrazine concentration in Nebraska, Iowa, and Illinois. There are also several “somewhat likely down” trends in these states. In many past trend studies using traditional trend significance, the “somewhat likely” trends would not be reported or might be reported as no trend. However, in Figure 2, we see that the trend direction for the “somewhat likely” trends in the three states match the pattern of “likely” trends and contribute to the overall picture of the decline in atrazine concentration in this part of the United States over the period 2002–2012.

    In contrast with the atrazine trends, most likely or somewhat likely DEA trends are upward, with downward trends predominantly in the Midwest. This pattern has been noted in trend periods starting in the late 1990s or early 2000s in other studies based on field observations (Ryberg & Gilliom, 2015; Ryberg et al., 2010, 2014). The opposing trends in atrazine and DEA are a result of the differing paths the two compounds take to streams and the many changes to atrazine regulation that began in the early 1990s. To address concerns about surface water contamination, risk reduction measures were instituted, including a decrease in application rates for corn and sorghum [Sorghum bicolor (L.) Moench], a decrease in maximum application rates for noncrop land use, discontinuation of uses for total vegetation control, well-head protection requiring 15-m setbacks around wells when workers mix and load atrazine-containing products, a 61-m application setback around lakes and reservoirs, and classification of all atrazine-containing products (except those for lawn and turf care and conifers) as restricted-use pesticides (USEPA, 2006). In 2003, the USEPA found that registered uses for atrazine were eligible for interim reregistration, with several label changes and risk management measures, further explained in the “Interim Reregistration Eligibility Decision for Atrazine” (USEPA, 2006).

    Well-head protection measures and setbacks from water bodies that began in the 1990s (USEPA, 2006), as well as agricultural management practices, such as no-till agriculture, may have resulted in longer residence time of atrazine in soil and a greater amount of transformation to DEA before runoff to a stream or transport to groundwater. Upward trends in DEA have happened in areas where both atrazine concentration trends and atrazine use trends are downward, with a possible explanation that the increased DEA in streams is from groundwater contributions reflective of higher past atrazine use (Ryberg & Gilliom, 2015).

    3.2 Multiple working hypotheses

    Supplemental Table S1 lists the MWHs and associated datasets for this study. In some cases, data were not sufficient to test potential hypotheses, and those cases are documented in Supplemental Table S1. Based on the MWH exercise and available data, we hypothesized that, in addition to use, atrazine concentrations would be driven by corn acreage in the watershed; weather and climate, with wetter conditions increasing the likelihood that atrazine would reach the stream as atrazine, rather than its degradation product, DEA; and potentially by conservation or tillage practices. We hypothesized that DEA trends would be driven by the same factors and base-flow contribution, expressed as BFI.

    Given the need to constrain the model to a small number of parameters, the correlation analysis allowed us to select the mean PHDI from among several measures of PHDI (annual maximum, annual minimum, annual mean, annual median, and annual standard deviation). We also examined correlation between the trends in atrazine and DEA and several variables related to agricultural conservation and management, including the concomitant trend in the percentage of agricultural land enrolled in the Conservation Reserve Program; the percentage of the watershed in 2012 drained by tile drains; the percentage of the watershed in 2012 drained by ditches; the percentage of the watershed in 2012 on which conventional, conservation, or no-till tillage practices were used; the percentage of watershed land planted to a cover crop; and the percentage of watershed land under a conservation easement. After examining correlations (some of which are shown in Supplement 2) of the concentration trends with conservation practices in Falcone (2017), we determined that the best candidate for including in the SEMs was the percentage of watershed drained by tile drains in 2012.

    3.3 Structural equation modeling to test causal hypotheses

    To test our causal hypotheses about drivers of atrazine and DEA trends other than use, we created a grouped SEM in which there were two groups, one the atrazine trends, and the other the DEA trends. This allowed us to model the pesticide and its degradate with the same variables, but the parameter estimates may vary between the two compounds. Figure 4 shows the resulting model for atrazine, and Figure 5 shows the DEA model. Standardized parameter estimates for the SEMs, as well as the measures of model quality, which all showed that these are excellent models, are provided in Supplement 2.

    Details are in the caption following the image
    Structural equation model explaining the factors that affect trends in atrazine concentration in streams across the conterminous United States, 2002–2012. Squares are observed variables, and circles are unobserved latent variables. Green arrows are positive coefficients, and red arrows are negative coefficients. The darker the color, the larger the coefficient. The numeric values are the standardized path coefficients (placed on the path lines) and error variances (placed on the lines with both arrows pointing to a single variable). Moisture supply and management is a latent variable derived from the latent variable model indicated by mean Palmer hydrologic drought index (PHDI) for the watershed, 2002–2012; tile drains, conservation practices in the form of proportion of watershed drained by tile drains in 2012; and the average base-flow index for the watershed. Corn trend is the change in proportion of watershed in harvested corn from 2002–2012. Concentration trend is the atrazine concentration trend expressed as a percentage over the period of record, 2002–2012. Concentration trend data are from Ryberg et al. (2017), other data are from Falcone (2017)
    Details are in the caption following the image
    Structural equation model explaining the factors that affect trends in deethylatrazine concentration in streams across the conterminous United States, 2002–2012. Squares are observed variables, and circles are unobserved latent variables. Green arrows are positive coefficients, and red arrows are negative coefficients. The darker the color, the larger the coefficient. The numeric values are the standardized path coefficients (placed on the path lines) and error variances (placed on the lines with both arrows pointing to a single variable). Moisture supply and management is a latent variable derived from the latent variable model indicated by mean Palmer hydrologic drought index (PHDI) for the watershed, 2002–2012; tile drains, conservation practices in the form of proportion of watershed drained by tile drains in 2012; and the average base-flow index for the watershed. Corn trend is the change in proportion of watershed in harvested corn from 2002–2012. Concentration trend is the atrazine concentration trend expressed as a percentage over the period of record, 2002–2012. Concentration trend data are from Ryberg et al. (2017), other data are from Falcone (2017)

    Both models have a latent variable derived from the LVM labeled “moisture supply and management” (MSM). This is a conceptual variable representing moisture supply and management that is measured by the mean PHDI, which is descriptive of the precipitation variation across the sites, conservation practices indicative of the percentage of the watershed drained by tile drains in 2012 (tile drains), and the average BFI for the watershed. This LVM directly influences the concentration trend in percentage over the period of record (concentration trend). Both models also have a variable, corn trend, that represents the change in the proportion of the watershed in harvested corn (hereafter referred to as corn) over the trend period.

    Figure 4 shows that the change in corn is the major driver of the atrazine concentration trend for the period 2002–2012. The standardized coefficient for this variable is .873 and represents a “large” effect (Cohen, 1977). The MSM latent variable was not statistically significant for the atrazine model. The model explains 77% of the variability in the atrazine trend, with corn explaining most of the concentration trend. Because corn is the main crop on which atrazine is used (Figure 1), changes in corn over time act as a surrogate for atrazine use trends.

    Figure 5 shows that the trend in corn is less important for the trend in DEA concentration than is the MSM latent variable (the absolute value of the standardized coefficient for corn is less than the absolute value of the standardized coefficient for the MSM LVM), although both variables have a “medium” effect on the concentration trend for DEA (standardized path coefficient >.30 and <.50; Cohen, 1977). The LVM, which represents the underlying correlation structure of the mean PHDI, tile drainage, and BFI, has a negative sign, whereas the PHDI and tile drain indicator variables have a positive sign and the BFI has a negative sign. This shows that the BFI has a positive correlation with the concentration trend, whereas the mean PHDI and tile drains have a negative correlation. Therefore, as BFI increases, the concentration of DEA increases. This supports our hypothesis that, because of the mechanism of degradation, the solubility of DEA, and the movement of DEA through the soil to groundwater, DEA would be influenced by BFI, whereas atrazine would not. This also supports the past, untested hypothesis of Ryberg et al. (2014) that the sign of atrazine and DEA concentration trends differ in some cases because of a groundwater transport pathway for DEA that lags between use and arrival at a stream. The mean PHDI has a negative correlation with the DEA concentration trends. This supports the hypothesis that wetter conditions would be preferential for atrazine runoff, rather than atrazine degradation to DEA in soils and subsequent transport via groundwater. Atrazine readily dissolves in water, and infiltration through soils to tile drains is a major pathway (Baker, Stone, Frey, & Wilson, 2007). The percentage of the watershed drained by tile drains in 2012 (the end of the trend period) has a negative correlation with the concentration trend, supporting that tile drains may transport atrazine to streams before it degrades to DEA. This SEM explains 38% of the variability in the DEA concentration trend.

    4 CONCLUSIONS

    The overall trend patterns and past investigation of use trends (Ryberg et al., 2014) indicate that regulatory changes may have been successful in encouraging atrazine degradation to DEA. The acute and chronic toxicity of DEA is less than that of atrazine for aquatic organisms (Ralston-Hooper et al., 2009); therefore, increasing DEA concentration trends would be preferential to increasing atrazine concentration trends, and in some cases, declines in DEA may follow declines in atrazine, with lag times depending on groundwater movement.

    The complexity of the SEMs was limited because of the number of observations available for model development. Even with this limitation, SEMs were developed that explained 77 and 38% of the variability in atrazine and DEA concentration trends, respectively. The SEMs quantitatively supported our hypotheses and past hypotheses about drivers of concentration trends for atrazine and DEA. More of the variability in atrazine trends across the conterminous United States was explained by SEM than was variability in DEA trends; however, mechanisms behind the DEA concentration trends are more complex because of lag time with groundwater, and, importantly, we have related the trends to a management practice—tile drains. To better understand the lag time, more site-specific investigation would need to be done that is outside the scope of the national perspective presented here.

    Atrazine use patterns on corn have changed over time with the adoption of genetically modified corn and the use of glyphosate. Overall, the atrazine stream concentrations trends were mainly downward between 2002 and 2012, with some sites showing upward trends. Nationally, the amount of harvested corn acres increased from 2002–2012 by >60,700 km2 (15 million acres) (USDA-NASS, 2018). The growth in corn acres was likely in areas not historically dominated by corn acres and is reflected in the scattered streams that showed upward trends in atrazine stream concentrations (Figure 2). These areas, like southern Idaho, western Missouri, and northern Alabama, experienced an increase in the amount of corn grown in the period 2007–2012 (Supplement 3; USDA-NASS, 2012). Therefore, the relation between corn acres and atrazine use over time is not constant, should not be extrapolated beyond the 2002–2012 period, and varies with watershed. In addition, regulatory changes (such as setbacks) may influence both the relation between corn acres and atrazine use, as well as the relation between atrazine use and stream concentration trends. It is possible that with a larger sample size and the resulting ability to develop a more complex model, additional factors could be identified as important for atrazine trends.

    The SEMs show that preferential degradation to DEA might happen in drier areas without tile drains. This highlights dual challenges for keeping atrazine from the stream because of human influence and climate change. Tile drains are negatively correlated with DEA trends and provide a pathway for atrazine to move more quickly to streams. Atrazine is sometimes applied to fields in the fall to control volunteer weeds and prepare for no-tillage planting in the spring (Comfort & Roeth, 1993). Fall has been considered a good application time because it tends to be dry in many areas; however, fall in the conterminous United States has experienced the most widespread increase in precipitation, exceeding 15% over the period 1901–2015 in much of the northern Great Plains (U.S. Global Change Research Program, 2017) where corn cultivation has been expanding (Supplement 3). In addition, the frequency and intensity of heavy precipitation events have increased across much of the United States over the period 1901–2016 and are projected to continue to increase (U.S. Global Change Research Program, 2017), and such increases could exacerbate atrazine runoff to streams. Although changes in regulation and use have benefited water quality with decreases in atrazine concentrations in surface water, climate changes and human-altered landscapes need to be considered in future water quality improvement efforts and in monitoring and analysis.

    ACKNOWLEDGMENTS

    This project was funded by the USGS National Water Quality Program's National Water-Quality Assessment Project. Thanks to Paul Stackelberg, USGS, for an early review of this manuscript, and thank you to three anonymous reviewers for their constructive comments. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.

      CONFLICT OF INTEREST

      The authors declare no conflict of interest.