Our study was conducted in pollinator-dependent cucurbit (Cucurbitaceae) (cucumber [Cucumis sativus], melon [Citrullus lanatus], squash [Cucurbita moschara], pumpkin [Cucurbita pepo]) crops of the Midwest, USA (Michigan, Ohio, Indiana, Illinois) (Figure 1). Midwest farms are dominated by row crops such as corn and soybean (Meehan and Gratton, 2016) with pollinator-dependent specialty crops geographically restricted to match suitable abiotic conditions for crop production (MWVG, 2020). Illinois, Indiana, and Ohio are part of the “corn belt” with ~35% of the landscape in corn and soybean production and <0.05% of land in cucurbit crops (USDA NASS, 2019). Farm size is variable across cucurbit crops and the states within our study region (Figure 1). For example, >50% of pumpkins across all states in our study region are grown on small farms with <15 acres in production. Overall, though, large scale producers (>100 acres), ~3% of cucurbit farmers, grow cucurbits on approximately two-thirds of all acres in production (USDA NASS, 2017). Cucurbit farmers rely on pesticides and IPM measures (e.g., crop rotation) to manage pests that reduce yields (Sánchez et al., 2018; MWVG, 2020). However, cucurbits also rely on bees for fruit production (McGregor, 1976) and these pollinators are at risk from pesticide exposure.

A survey of cucurbit farmers was developed to elicit information regarding farmers' knowledge, perceptions, and behaviors regarding pest management, pollination strategies, and potential tradeoffs among these two key inputs to pollinator-dependent crop production (see Supplementary Tables 1–4 for survey questions). The multi-step design process included individual interviews, focus groups, and survey instrument pre-tests with farmers, extension personnel, and members of the research team (Mitchell and Carson, 1989). The final survey instrument had multiple sections with questions related to farm and farmer characteristics, current pest and pollination management practices, views on pollinator health and risks, and willingness to adopt pollinator best management practices. This paper reports on a subset of the survey (see Supplementary Tables 1–4); other questions are analyzed elsewhere.

The survey was mailed in early 2019 to a list of 2,543 cucurbit farmers across our study area (Figure 1) purchased from FarmMarketID (www.farmmarketid.com), an agri-business market research firm. We followed a multi-step mailing process (January through March) that included an introductory letter, a survey instrument, a reminder postcard, a second survey instrument to non-respondents, and a final reminder postcard to non-respondents (Dillman et al., 2008). By the end of May, we were contacted (via phone, email, or written correspondence) by 978 survey recipients (38% of the mailing sample), most frequently via a prepaid postcard included in the survey package to maintain respondent anonymity. Only 74 postcards specified that growers had returned a completed survey, although we ultimately received 106 mostly complete and 15 partially complete surveys. We subsequently removed 900 names from our potential respondent pool of 2,543 because the survey was undeliverable or, after hearing directly from the mail recipient, they did not meet our sampling criteria (i.e., the recipient was retired or deceased, had never grown cucurbits, had not grown cucurbits in the previous 5 years, only grew cucurbits in a home garden for personal use, or had never farmed [e.g., resided in a new residential subdivision on former farmland]). It is impossible for us to calculate a legitimate response rate, because more than 68% of the survey recipients that we heard from did not grow cucurbit crops. We ultimately received 106 complete or mostly complete, usable surveys from cucurbit farmers.

Farmers who perceive severe threats to pollinator populations are more likely to engage in pollinator-supporting practices (Nalepa et al., 2020). To assess farmers perceptions of the status of pollinators, we first asked farmers if they thought wild pollinator populations were in decline (yes, no, not sure). We then questioned farmers perceptions of changes to the availability and quality of managed bees over the past 5 years (increased, decreased, no change, or not sure). Declines in pollinators are thought to be driven by diverse and numerous stressors including habitat loss and degradation, agricultural and residential pesticide use, increases in pathogens, changes in weather patterns such as increases in extreme weather events, and pollution, among others (Winfree et al., 2009; Potts et al., 2010; Cameron et al., 2011; Goulson et al., 2015). Agricultural pesticides, in particular, are a major stressor on pollinator populations and pollinators can be exposed to pesticides directly, through exposure during spraying or contact with contaminated planting dust, or indirectly, via exposure to residues in pollen, translocation through plants, and drinking water contaminated with pesticide runoff (van der Sluijs et al., 2013; Sanchez-Bayo and Goka, 2014). Farmers who perceive multiple stressors and multiple pesticide exposure pathways as important may be more likely to engage in pollinator supporting practices. To assess farmers views about factors potentially impacting pollinator health and potential pesticide exposure pathways, farmers were asked to rank the importance of 10 factors contributing to pollinator health and eight factors contributing to pesticide exposure (Supplementary Tables 1, 2). Responses were recorded by the farmer on a Likert scale, with scores ranging from 1 to 5 (1 = Not at all important; 2 = Slightly important; 3 = Somewhat important; 4 = Very important; 5 = Extremely important).

Farmer classifications can be useful for targeting those farmers most likely to adopt new technologies or participate in conservation programs (Daloglu et al., 2014). We classified farmers by their responses to the aforementioned perception questions using a two-step process, a principal component analysis (PCA) followed by a cluster analysis. To begin, we examined correlations between our variables (i.e., question responses) using the generic cor function in R (R Core Team, 2020). No strong correlations were observed. We then used Bartlett's test to inspect the correlation matrix using the cortest.barlett function in the psych package (Revelle, 2020). Barlett's test was highly significant (χ² = 959.06, p < 0.00001, df = 153) indicating that PCA was appropriate for our dataset. We also evaluated the determinant of the correlation matrix, which did not indicate that PCA would be a problematic approach (Field et al., 2012). PCA was then conducted on the variables using the principal function in the psych package (Revelle, 2020). Scree plots of eigenvalues were visually examined to find the point of inflection. This approach suggested a three-component solution. We then reran the PCA with the three-component solution with varimax rotation and examined the component loading matrix using the print.psych function in the psych package (Revelle, 2020). Variables with component loadings above 0.6 were retained for the cluster analysis (Balukas et al., 2019).

To classify farmers, we then used the 14 variables retained from the PCA in a k-means cluster analysis. We used the generic kmeans function in R and varied the number of cluster solutions (R Core Team, 2020). A three-cluster solution was found to maximize within-cluster homogeneity. We used the generic kruskal.test and pairwise.wilcox.test functions in R to conduct a Kruskal-Wallis rank sum test followed by pairwise comparisons using Wilcoxon rank sum tests with continuity correction to confirm that input variable values differed between the farmer clusters (R Core Team, 2020). We also used this series of tests (Kruskal-Wallis rank sum followed by Wilcoxon rank sum tests) to evaluate a subset of input variables that were not retained by the cluster analysis (see Results for variables tested). We used descriptive statistics of the perception variables for each cluster to better understand the heterogeneity of views of pollinator health and pesticide exposure risks among the different farmer classes and additional descriptive statistics on farm and farmer characteristics to better understand the common attributes of farmers in each class (Balukas et al., 2019).

To evaluate the impact farm and farmer characteristics have on the adoption of strategies that farmers can use to manage pests and pollinators, we quantified 12 agricultural production practices using survey data. To determine the frequency of pesticide use, farmers were asked to record how many applications of insecticides, herbicides, and fungicides they applied per year, on average, over the last 5 years to their cucurbit fields. We also totaled across pesticide groupings to evaluate overall pesticide use. To quantify the use of IPM, farmers were asked if they engaged in 20 different IPM practices within the past 5 years. Practices were grouped in one of four categories (prevention, avoidance, monitoring, and suppression) and multiple selections from each category were allowed. For example, prevention IPM practices included: cultivated for pest control, removed or burned crop residues, cleaned equipment after field work, and chopped, sprayed, mowed, plowed or burned field edges, ditches or fence lines. Selected practices were totaled for each category (e.g., across the four options for prevention) and summed across all 20 practices. Lastly, to quantify pollination management practices, we asked farmers if they used managed bees to pollinate their crops (Yes or No) or installed pollinator habitat on their farm (Yes or No). We also asked how likely they were to participate in a pollinator habitat conservation incentive program (e.g., cost share programs), administered at five levels (federal, state, county, regional, non-governmental). This question was implemented using a Likert scale, and the values reported for each administration level were summed. Additional details on the survey questions used to quantify these 12 agricultural practices can be found in the Supplementary Table 3.

Farmers were asked a series of questions that characterized their farming system and demographics. To characterize their farming systems, we asked for the number of field and pollinator dependent crops grown within the previous 5 years. Farmers could select multiple options, which were given, and allowed room to detail other field crops that they grew. We also asked farmers to estimate the number of acres in cucurbit crops and total acres of cropland, managed in the year of the survey. To assess insect pests, farmers were given a list of common arthropod pest species, from which they could select multiple options, and also allowed to state other pests that they manage within their farm. To characterize farmers, we asked questions regarding farm ownership, if the farm was the main source of household income, the number of years farming, the length of time in family farming, and farmer age. Additional demographical information was collected (e.g., gender), though, these variables were highly skewed or sparse and could not be included in our statistical analysis (see below). Information sources regarding pest management were collected across nine categories (crop scout, university extension resource, National Resource Conservation Service staff, neighbor or friend, session at a grower meeting, newsletter [paper], email listserv, text message [automatic alert], social networking site [Twitter, Facebook, etc.]) and four response levels (used in past, use now, would use, never have, and never will use) from which farmers could select one option. For our analysis we summed the number of responses for “use now” across the nine categories to determine the number of current information sources for each farmer (e.g., a value of 9 would indicate the farmer used all sources of information). Additional details on the questions used to evaluate farmer and farm characteristics can be found in the Supplementary Table 4; see also Table 1 for summary statistics.

To evaluate factors influencing pesticide use, IPM, and pollination management, we constructed generalized linear models using the 12 agricultural practices described earlier as dependent variables: sprays per year for insecticides, herbicides, fungicides, and all pesticide sprays summed together; number of IPM practices for prevention, avoidance, monitoring, suppression, and all practices summed together; use of managed bees; creation of pollinator habitat; and future participation in pollinator habitat conservation programs. Each model used the same set of explanatory variables, including the farmer classifications generated by the cluster analysis described earlier, as well as farm and farmer characteristics (Table 1). The assumed error distributions varied by model depending on the response variable (gaussian for continuous variables [e.g., the number of pesticide sprays] or binomial for dichotomous variables [e.g., a Yes or No response]).

We used an exploratory statistical analysis approach starting with a global model constructed for each dependent variable with the same set of explanatory variables:

where β₀ is the model intercept, Y_i is the estimated value for the dependent variable being modeled (e.g., number of pesticide applications), X_i are the explanatory variables (farm characteristics, farmer demographics, and clusters), β_i are the partial regression coefficients for each explanatory variable, and ε_i is the error term which varied by the dependent variable being modeled (see above) (Zuur et al., 2009). The global model was constructed with the generic glm function in R (R Core Team, 2020). We used the dredge function in the MuMIn package to find model permutations (4,096 models for each dependent variable) (Bartón, 2020). Thus, dredge was used to generate models using all additive combinations of the explanatory variables from the global model, including a null model (intercept only). We then used the get.models function to extract a list of fitted model objects which was passed to the model.avg function to find estimates for all parameters (Bartón, 2020). In other terms, we used a model averaging approach (multi model inference) where parameter estimates were averaged across competing models to generate estimates for all explanatory variables. Therefore, we drew inferences from multiple models, rather than conducting comparisons of models. This approach was exploratory because it was not hypothesis driven. That is, we were interested in evaluating the relationship between all explanatory and response variables. By using model averaging, we found estimates for all explanatory variables and avoided uncertainty common to information theoretic (model selection) approaches (Burnham and Anderson, 2002). We considered explanatory variables important when the model averaged p-value estimates were below an α = 0.05. All statistical analyses were conducted in R 4.0.2 (R Core Team, 2020).

Farmers, on average, perceived no change or were not sure regarding the availability and quality of honey bee hives and bumble bee colonies (Table 1). The result regarding managed bumble bee colonies is not surprising as only six survey respondents actually buy bumble bees for pollination and the remaining respondents would likely be unsure. On average, farmers also tended to be not sure about whether pollinators were in decline (Table 1). However, as described below, there are differences among some farmer classes regarding pollinator status.

The principal component analysis identified a three-component solution which we further ascribed as variables associated with: (1) pesticide exposure, (2) bee diet/diseases and habitat loss, and (3) changes in weather (Figure 2; Table 2; Supplementary Tables 1, 2). Three observations are worth noting. First, some variables—most notably, the pollinator status (i.e., wild pollinator declines, changes in availability, and quality of managed bees) and the agricultural and residential pesticide use variables—were dropped from the cluster analysis by the PCA process either because they didn't contribute significantly to the first three principal components or because of missing data. However, as described below, some of these variables differ among the farmer classes. Second, while the survey presented the pollinator health and pesticide exposure ranking questions in two groups (Supplementary Tables 1, 2), the questions related to weather changes resulted in their own principal component indicating different views regarding weather vs. other potentially more controllable factors. Third, even though the question regarding pesticide residues remaining in the soil after a previous crop rotation was grouped with the set of questions on pollinator health (see Supplementary Table 1), the PCA included that question in the component with the group of questions related to pesticide exposure routes (Table 2). In hindsight, that grouping makes more sense than the original survey design and survey respondents' views confirmed this.

Using cluster analysis, we identified three classes of farmers according to how they rated the importance of the factors potentially impacting pollinator health including potential routes of pesticide exposure on their farms (Figure 2; Table 2). We found that farmer classes represented a spectrum of the Likert scale. Farmers classified as “proponents” typically responded with Likert scores ranging from 3 to 5 (n = 16), farmers classified as “neutral” occupied the middle of the scale with scores ranging from 2 to 4 (n = 41), and farmers classified as “skeptics” responded with Likert scale scores ranging from 1 to 3 (n = 36) (Figure 2; Table 2). Via Kruskal-Wallis rank sum test (χ² = 423.79, p < 0.001, df = 2) and pairwise comparisons of the Wilcoxon rank sum test (neutral-proponent: p < 0.001; neutral-skeptic: p < 0.001; proponent-skeptic: p < 0.001) we confirmed that scores for the statements regarding the factors impacting pollinator health and routes of pesticide exposure differed between the three farmer classes, as qualitatively described above (Supplementary Figure 1).

Our additional tests on the subset of variables dropped in the cluster analysis (see above) suggested skeptics were less likely to think wild pollinator populations were declining than the proponent and neutral classes (Kruskal-Wallis rank sum test [χ² = 12.04, p = 0.00243, df = 2], Wilcoxon rank sum test [neutral-proponent: p = 0.67; neutral-skeptic: p = 0.0029; proponent-skeptic: p = 0.039]). That is, skeptics were less likely to perceive pollinator populations as threatened (Supplementary Figure 2). Skeptics were also less likely to rate agricultural and residential pesticide use as a very or extremely important factor impacting pollinator health than the proponent and neutral classes (agricultural: Kruskal-Wallis rank sum test [χ² = 20.12, p < 0.001, df = 2], Wilcoxon rank sum test [neutral-proponent: p = 0.26; neutral-skeptic: p < 0.001; proponent-skeptic: p < 0.001]; residential: Kruskal-Wallis rank sum test [χ² = 22.76, p < 0.001, df = 2], Wilcoxon rank sum test [neutral-proponent: p = 0.009; neutral-skeptic: p = 0.0045; proponent-skeptic: p < 0.001]), despite there being substantial evidence of the impacts of these pesticides on pollinators (Supplementary Table 1; Supplementary Figures 3, 4).

Regarding the variables retained in the cluster analysis, proponents were more likely to rate (on a 1–5 Likert scale) factors impacting pollinator health as extremely important or very important (means from 3.81 to 4.63), while skeptics were more likely to rate these factors as slightly or not important (means from 2.19 to 3.69) (Table 2). On average, neutral farmers rated pollinator health factors as somewhat important (means from 2.95 to 4.02). Despite the differences in overall ratings, all three farmer classes ranked “the spread of pathogens and parasites harmful to bees” the most important factor for pollinator health (means of 4.6, 4.0, and 3.7 for the proponent, neutral, and skeptic classes, respectively; Table 2).

Proponents were also more likely to rate the different routes of pollinator exposure to pesticides as extremely important or very important (means from 3.88 to 4.81), while skeptics were more likely to rate the different routes of exposure as slightly or not important (means from 1.25 to 2.72) (Table 2). Neutral farmers on average rated the different pesticide exposure routes as somewhat important (means from 2.44 to 4.12; Table 2). Of the possible routes, proponents on average ranked “pesticide drift from neighboring fields” as the most important contributor to pollinator pesticide exposure (mean of 4.81; Table 2). In comparison, skeptics and neutral farmers on average ranked “direct spraying” as the most important contributor of pollinator pesticide exposure (means of 2.72 and 4.12 for the skeptic and neutral classes, respectively; Table 2).

Qualitatively, the three farmer classes were similar in their demographics and farm characteristics (Table 1). However, compared to the neutrals and skeptics, proponents were younger (mean of 52.6 years old), had fewer years farming (mean of 29.6 years) but more years in family farming (mean of 96.1 years), and had sources of income outside the farming system (31% had off-farm income) (Table 1). Proponents also reported smaller farms (mean farm size of 1037.5 acres), higher pest richness (mean of 3.0 pests), a greater proportion of their farm in cucurbits (17% of land in cucurbit production), and greater richness of crops that were not pollinator dependent (mean of 3.0 crops) (Table 1).

Pest management was mediated by the physical environment and economic conditions of the farming system (Figures 3, 4; Tables 3, 4). The use of all pesticides summed together, and insecticides, herbicides, and fungicides individually, were positively correlated with the pest richness of the farming system (all pesticides: estimate = 1.4, SE = 0.23, z-value = 5.9, p-value <0.001; insecticides: estimate = 0.51, SE = 0.098, z-value = 5.1, p-value <0.001; herbicides: estimate = 0.12, SE = 0.053, z-value = 2.2, p-value = 0.026; fungicides: estimate = 0.79, SE = 0.17, z-value = 4.7, p-value <0.001; Table 3; Figures 3A,C–E). All pesticide use, and particularly, fungicide use, was higher when farming was the main source of household income (all pesticides: estimate = 2.5, SE = 0.94, z-value = 2.7, p-value = 0.0079; fungicides: estimate = 2.1, SE = 0.67, z-value = 3.00, p-value = 0.0026) (Figures 3B,F). No explanatory variables predicted individual IPM practice categories (e.g., prevention) (Table 4). However, when selections for all 20 IPM practices were summed, there was a positive correlation with pest richness (estimate = 0.75, SE = 0.23, z-value = 3.2, p-value = 0.0015) and farm size (estimate = 0.00071, SE = 0.00033, z-value = 2.2, p-value = 0.031) (Figures 4A,B; Table 4).

The use of managed bee colonies was positively associated with farming as the main source of household income (estimate = 1.90, SE = 0.77, z-value = 2.40, p-value = 0.017) (Figure 5A; Table 5). The creation of pollinator habitat was positively related to the richness of pollinator-dependent crops grown by the farmer (estimate = 0.39, SE = 0.19, z-value = 2.00, p-value = 0.044) and negatively associated with farming as the main source of household income (estimate = −1.80, SE = 0.73, z-value = 2.40, p-value = 0.017) (Figures 5B,C; Table 5).

Farmer classes (proponents, neutrals, and skeptics; Table 2) were not related to pesticide use, IPM, the use of managed bees, or the installation of pollinator habitat. However, proponents were more likely than neutrals and skeptics to indicate future participation in pollinator habitat conservation incentive programs (e.g., cost-share programs) when Likert scores were summed across all administrative levels (federal; state; county; regional; non-governmental) (proponent: estimate = 4.60, SE = 2.00, z-value = 2.30, p-value = 0.023) (Figure 5D; Table 5).