Occupational stress among farmers is not a new issue. Researchers have assessed stress using standard survey techniques beginning in the 1980s [1
]. In the 1990s, with the reduction in the number of farmers and full-time employees and the increase in part-time employment in farmers, researchers recognized the challenges that were occurring in rural areas [3
]. Issues studied in association with farm stress included paperwork, new legislation, loss of family traditions, finances, isolation, media criticism, ill health, and the future of the farm [3
]. That farmers experience on-going and complex stress is a fact acknowledged worldwide [7
]. Farmers continuously experience feelings of a lack of control in response to chronic, unpredictable stressors [8
]. Depressive and anxiety disorders can result from exposure to chronic, unpredictable environmental stress [9
]; see [14
] for a recent review on stressors and mental health in farmers.
Indicators of stress in farmers were first examined by Walker and Walker in 1986 in their initial development of the Farm Stress Inventory (FSI). Respondents were asked to report the frequency of experiencing 19 stress symptoms adapted in part from the Hopkins Symptom Checklist [15
]. The items included increases in alcohol consumption and smoking, weight gain or loss, change in health, trouble relaxing, nightmares, chronic fatigue, sleep disruptions, frequent illness, headaches, forgetfulness, trouble concentrating, increase in arguments, behavior problems in children, marriage problems, back pain, losing one’s temper, and avoiding decisions.
Thu and colleagues (1997) used seven stress-related Center for Epidemiological Studies-Depression (CES-D) items [16
] and 10 additional items from a neurological symptoms scale and summed them to create a stress score [17
]. Items selected were comparable to the Perceived Stress Scale developed by Cohen and colleagues [18
]. The CES-D included seven items: felt bothered by things that don’t usually bother you, had trouble keeping your mind on what you were doing, slept restlessly, did not enjoy life, not felt like eating or had a poor appetite, felt like everything you did was an effort, and felt unhappy. The neurological symptoms scale included tire easily, light-headed and dizzy, difficulty concentrating, trouble remembering things, relatives notice your memory problems, had to make notes to remember, hard to read written materials, felt irritable, difficulty falling asleep, and headaches.
In 1986, Walker and Walker initiated development of the first farm-specific stress survey, the Farming Stress Inventory (FSI), by asking 140 Canadian farmers to rank the top five farm stressors [8
]. Financial stress was most common (83%), followed by unfavorable or unpredictable weather (75%), and government agricultural policies (75%). Further development of the scale involved interviewing 808 Canadian farmers and showing positive associations between stressors in the FSI and symptoms of stress [1
]. Over half of the stressors most strongly predicting stress symptoms were unique to farming, demonstrating that using a stress scale designed for the general population may be missing important stressors in farmers.
Additional measures specific to farmers followed in the 1990s [19
] with a call for the development of a theory specifically linked to farming as a basis for the development of farm stress measures [21
]. However, research on farm-related occupational stress has lagged behind other industries [22
]. The current approach to workplace stress focuses on the mismatch between work demands and pressures and the knowledge of the workers, their capabilities to do the work, and their ability to cope [22
]. Using this perspective, the stressors associated with farming are left out of the approaches used for interventions, which have focused on stigma and attitudes towards employees suffering from stress and mental illness [22
] but have neglected the role of the external environment (e.g., weather), regulations, and market forces.
Eberhardt and Pooyan (1990) interviewed six farmers to develop constructs related to areas of their work that concerned them [19
]. Qualitative analysis of the interview data produced 28 items with a 7-category Likert scale to capture responses. The scale was tested in 11 farmers at a farm show and, after review by a farm stress counselor, was found to adequately represent farmer concerns. Subsequently, 362 of 1300 possible farmers responded by mail to the 28-item instrument (response rate 28%). Principal components analysis indicated that six constructs adequately captured the 28 items. The six constructs were: economics, geographic isolation, time pressures, climatic conditions, personal finances, and hazardous working conditions. The constructs formed the Farm Stress Survey (FSS) and were validated using a Life Satisfaction Scale, Emotional Strain Symptoms Scale, and illness frequency. The six factors accounted for 61.8% of the variance in the 28 survey items. The survey did not measure stress related to health or injury.
Deary and colleagues (1997) developed a survey based on the literature and items from the FSS and the FSI, modified for farmers in the United Kingdom [20
]. Data from this study of 318 farmers from a variety of farm types (mixed, dairy, cereals, cattle, sheep) became the basis of the Edinburgh Farming Stress Inventory, which identified six domains they reported as causing high stress levels. The domains were described as farming bureaucracy, finances, isolation, uncontrollable natural forces, personal hazards, and time pressures. Using their scale, they found differences in stress levels by type of farming and that younger farmers and women had higher levels of farm stress. The constructs were similar to the FSS, again lacking heath and injury.
Several farm stress scales were developed after 2000 [24
]. A 2008 study of 1343 Iowa farmers asked about a list of farm-related stressors (45% response rate) [24
]. The stress items used were based on an unpublished pilot study. Respondents were asked to compare each of 62 items to the stress of being married. Farmers rated 44 items as being more stressful than marriage; 18 items were less so. Personal adversity (loss of spouse or children), disabling injuries, and crop losses were highest on the stress comparison. Farmers between the ages of 40–79 and females showed the highest stress levels.
The Farming Family Stressor Scale addressed stress in Australian farmers [25
]. Based on a survey of 278 farm family members in 2010 and consisting of 29 items, the scale showed good reliability and validity. The domains of stressors included hazardous working conditions, geographic isolation, personal finances, time pressures, climate conditions, and general economic conditions. Health was not included.
Using a modified version of Welke’s Farm Ranch Stress Inventory [26
] in a sample of 128 private pesticide applicators in North Carolina in 2012, several additional stressors were identified [27
]. Substantial numbers of respondents said weather (60.2%), the future of the farm (29.7%), outsiders not understanding the nature of farming (25.2%), machinery problems (23.4%), commodity prices (45.3%), taxes (38.3%), health care costs (32.5%), and lack of family recreation time (13.3%) were “very stressful”. The farmers were mostly 40–59 years of age, had farmed at least 20 years, and worked more than 40 h per week on the farm. Welke’s survey included health care concerns and the future of the farm, which had not been reported previously.
Risk factors for stress are highly correlated; a factor can be both a source of stress and an outcome of stress, e.g., sleep deprivation, illness, injury. Variables associated with stress such as negative life events, substance use, farm workload, chemical exposures, social support, and pre-existing physical conditions are related and may be rare events. The presence of collinearity can result in small singular values in the design matrix, causing instability in estimators or non-convergence of the model [28
]. Previous studies on correlated stressors have chosen to include only a single stressor and to remove stressors correlated with it, however, that can result in removing an important stressor with more proximal causal associations with stress. In addition, interactions among stressors have been ignored entirely. Stressors do not occur in isolation, especially on a farm. Classical statistical models cannot address these challenges.
A neural network is a learning algorithm that originated independently in psychology, statistics, and artificial intelligence. It is a nonlinear statistical model that uses a hidden layer and back propagation to minimize an error function [29
]. Modeled on the functioning of the human brain, a neural network contains a hidden layer with nodes that represent neurons [30
]. Each connection between nodes represents a synapse. A node (neuron) fires when the signal it receives exceeds a threshold value, resulting in being in an “on” or “off” state. The threshold value is typically modeled using a sigmoid activation function. The hidden layer, usually called “Z”, is not directly observed. This is analogous to latent variable models such as a structural equation model. The Z layer is an expansion of a linear function formed from a transformation of the original independent variables. The transformed vector of predictors is used in the activation function to produce the hidden units and then a linear transformation is applied to predict the outcome variable. Each hidden unit represents a different interaction term. For example, a model with two hidden units (neurons) can model a three-way interaction between predictors. The difference between the observed outcome and the predicted outcome is captured using the mean squared error (MSE) with the goal of minimizing the MSE. The parameters of the model are “learned” from the data using a back propagation algorithm in a forward and backward sweep [31
Sources of stress in the agricultural community have remained unchanged for decades and remain multidimensional and complex. Asking farmers directly what they think is contributing to their stress levels might miss important risk factors for stress. It is unlikely that farmers understand the interplay and complexity of all possible contributors to stress they experience in a larger context. There have been no studies that have addressed stress in a sample of farmers without asking them directly about what they perceive their stressors are. The development of stress theory specific to farmers would benefit from an extrinsic approach with novel statistical methods linking correlated stressors to the stress response, allowing for complex interactions. The purpose of this study is to identify important farm stressors that could be used in a stress measure. Using data collected in 1992–1997 in eight counties in Colorado and in 1993 in a statewide survey in Colorado, we used the novel method of a neural network to ask two questions: (1) Can we identify a set of stressors out of 31 possible indicators that are most important in predicting stress based on the stress-related items in the CES-D and (2) can we validate the model by using a second set of test data to see how well we can reproduce the model with minimal errors.
Using an approach designed to manage complexity, we identified predictors of stress that have not previously been considered. Our approach to better elucidate factors that increase stress in farmers used previously collected data but did not ask farmers directly about what they believe causes stress. Our results indicate that debt, healthcare, high pesticide exposures, and physical wellness are of highest importance to predicting stress in farmers. The relative importance of predictors appeared to group into those with magnitude of less than 100 (range 24.8–59.5) and those of magnitude greater than 100 (range 150–922) with a gap between the two groups. Several surprising results were seen in predictors with high importance compared to what has been reported in the literature. First, having had a pesticide-related illness had similar effects on stress levels as having a chronic disease. Second, many health-related risk factors showed very strong effects. In terms of access to health care, being able to see a doctor when it is needed, having a usual place to go for medical care, and having health insurance were strongly stress-reducing. Health status measured by having to stop activity due to illness, having a chronic disease, taking prescription medications, and being a current smoker also showed strong contributions to stress. Third, having a high school education reduced stress levels. Fourth, working on another farm strongly predicted higher stress levels and was more important than having farming as the primary occupation. As seen in every published study of stress in farmers, increase in debt was the strongest predictor of stress. Debt was far more important than income.
The second group of predictors were not as important, but removing them increased the MSE in the model. The more families on the farm, the lower the stress, may be providing a measure of social support. The number of alcoholic drinks consumed when drinking was associated with increased stress levels. Increased stress is known to increase alcohol consumption, which in turn, could increase stress due to reduced productivity. Having a hospitalization, increasing number of visits to the doctor in the past year, decreased income, and experiencing a work-related injury contributed to increased stress.
Two counter-intuitive findings were observed. A higher sales value of crops and being involved in more clubs increased stress levels, but their importance values were low. The directionality might be an unexpected product of interaction terms, reflecting the complexity of relationships between predictors. Possibly crop value is related to other economic factors such as debt. Debt might be so strong as to suppress the importance of crop sales. Maybe the number of clubs increases time away from the farm and elevates stress and does not act as a social support in farmers. The complex nature of these models makes understanding these findings difficult, although a great deal of work is currently being done to make these models more interpretable.
Aside from the strength of the importance of a high school education to protect against stress, the other demographic factors were not surprising and were commonly observed in previous studies. Females had higher stress levels and being married reduced stress. Adverse life events strongly increased stress levels. In contrast to previous studies, age and years in agriculture were not important in predicting stress levels. These are not characteristics that lend themselves to interventions and possibly are better accounted for by other risk and protective factors in the model. The neural network may be the better tool for identifying the underlying reasons that age and years in agriculture have been related to stress in previous studies, such as health measures and debt.
As is often the case when comparing regression models to neural networks, the neural network outperforms regression models [39
] but not always [41
]. Using a regularized regression model identified only seven of thirty-one covariates that significantly predicted stress level, however, the neural network identified different and a greater number predictors that were influencing stress. Since neural networks do not produce p
values, it is the importance of the predictor in the model that provides additional information. Whether neural networks perform better than classical regression approaches may be a matter of how the variables interact with one another. Simply put, it may be a matter of how closely the process is best captured by a signal-processing model that reflects what happens in the brain. When environmental signals reach a threshold, the neuron is turned on, and the signaling pathway is activated. In the case of stressors, the result is the glucocorticoid cascade resulting in increased cortisol flooding the body. In the case of psychological traits such as stress, the allostatic load is best measured by high level interactions among a variety of inputs. In our model, two hidden nodes produced the lowest MSE, indicating that a model containing a third-degree polynomial adequately fit the data. This translates to a three-way interaction.
The ridge regression identified poorer self-perceived health status as increasing stress levels. In the neural network, self-perceived health status showed a weight near zero, indicating it was not contributing anything in explaining stress levels, however, many other health related predictors became highly influential. As we were able to reproduce these findings in the test data, this is a meaningful result and reveals possible targets of intervention in farm families.
It is interesting that an obvious difference in the CES-D items and the CES-D and neurological items combined is that being light-headed or dizzy and having headaches are not included in the CES-D symptoms, but each of these could easily be related to alcohol consumption. Half of the neurological symptoms were related to memory or concentration and could be strongly correlated with “having trouble keeping your mind on what you were doing” in the CES-D scale. Spearman’s correlation between the CES-D and neurological scale scores was 0.57 (p < 0.0001) so although statistically significant, they were only moderately correlated. Given the widespread use of the CES-D scale, if certain items in the scale form a stress subscale, it would be a readily available tool to measure stress. It might be that certain items in the CES-D scale capture stress and are also related to depression since these constructs overlap. Future work should compare the CES-D stress items with other validated stress scales in a general population sample. The CES-D stress items should also be compared with stress scales previously developed in farming samples to test their validity.
Designing interventions to address stress in farmers is challenging because so many of the important factors are a product of national and international economic policies. Commodity prices and weather are not amenable to intervention, but improving health care accessibility is. The results of this study provide greater detail and insight into the sources of stress in farm residents. As a first step, we need to improve access to health care in rural areas and prevent high pesticide exposure by continuing to provide safety training and promote the use of personal protective equipment. We should allow farmers to spend their time farming as their primary occupation without working off their farm. Keeping farmers healthy would reduce the overall burden of stressors they have been experiencing for decades and continue to experience. This study forms a foundation for future development of a farm-specific stress scale by identifying new risk factors and providing a better understanding of the importance of these factors in increasing stress levels and activating the stress response.
This study has several limitations. The data were collected in the 1990s and do not include extreme weather events, issues related to farm labor shortages, the future of the farm, changes in regulations, and low commodity prices resulting from trade policies, which are issues that have exacerbated the constant stress that farmers have been experiencing since the 1970s. Due to the lack of data, important stressors such as trade policies and extreme weather events were not included in this study, resulting in reduced generalizability to the current situation. However, these stressors have affected nearly all farmers equally in recent years. In addition, several conceptual models of stress include coping strategies and the only ones that were available in the data used were related to social support [42
]. Future work should address stressors not included in this study and additional coping strategies in addition to the stressors identified in the present study. Additionally, more work is needed on using the subset of CES-D items as a measure of the stress response in other farm samples and in the general population.