Suicide and Pesticide Use among Pesticide Applicators and Their Spouses in the Agricultural Health Study

Background: An association may exist between pesticide exposure and suicide. Objective: We sought to evaluate the existence of an association between pesticide use and suicide using data from the Agricultural Health Study (AHS), a prospective cohort study of licensed pesticide applicators and their spouses in Iowa and North Carolina. Methods: Via linkage to state mortality files and the National Death Index, we identified 110 suicides occurring between enrollment in the AHS (from 1993 to 1997) and 31 May 2009, among 81,998 cohort members contributing 1,092,943 person-years of follow-up. The average length of follow-up was 13.3 years. AHS participants provided data on pesticide use and potential confounders via self-administered questionnaires at enrollment. We evaluated several measures of pesticide use: use of any pesticide, ever use of 50 specific pesticides, cumulative lifetime days of use and intensity-adjusted cumulative lifetime days of use of 22 specific pesticides, and ever use of 10 functional and chemical classes of pesticides. We used Cox proportional hazards regression models to estimate adjusted hazard ratios and 95% confidence intervals. Results: After adjusting for age at enrollment, sex, number of children in family, frequency of alcohol consumption during the past 12 months, and smoking status, we found no association between prior pesticide use and suicide in applicators and their spouses. Results were the same for applicators and spouses together or for applicators alone and were consistent across several measures of pesticide use. Conclusions: Our findings do not support an association between moderate pesticide use and suicide.


Research
Several studies have reported higher suicide rates among farmers than the general popu lation (Blair et al. 1993;Browning et al. 2008;Gunderson et al. 1993;Lee et al. 2002;Meltzer et al. 2008;Miller and Burns 2008;Page and Fragar 2002;Stallones 1990), although two studies found lower rates among Canadian farmers (Pickett et al. 1993(Pickett et al. , 1999. A review noted higher suicide rates among farmers than any other occupational group in the United Kingdom (Gregoire 2002).
Other studies suggested associations between chronic exposure to pesticides and suicide among farmers and other agricultural populations. In Australia, pesticide applica tors had higher suicide rates than the general population (MacFarlane et al. 2009(MacFarlane et al. , 2010, whereas applicators in Italy had a lower rate of accidents and suicide (Torchio et al. 1994). Parrón et al. (1996) found higher suicide rates in an intensive agricultural area of southeast ern Spain than in other areas with similar demographic and socioeconomic composi tions; they tentatively attributed the increased rates to pesticide exposure. Colorado farm ers potentially exposed to pesticides had higher suicide rates than the general popu lation (Stallones 2006), and ecological and case studies suggested an association between organophosphate pesticide (OP) use and sui cide (London et al. 2005).
The Agricultural Health Study (AHS) is a large, prospective cohort study of private pes ticide applicators (mostly farmers), commercial applicators, and spouses of private applicators in Iowa and North Carolina. It was designed to study associations between cancer and other chronic diseases and farmrelated exposures (Alavanja et al. 1996). Previously, AHS par ticipants in the highest category of use of chlo rpyrifos, an OP, were reported to be twice as likely to commit suicide as those who never used chlorpyrifos (Lee et al. 2007). Because that study was based on few cases and evaluated only one pesticide, we wanted to evaluate more fully the possible associations between the use of pesticides, particularly OPs, and suicide among applicators and their spouses in the AHS.

Methods
Population and case definition. The AHS cohort, enrolled from 1993 to 1997, provided data on demographic and lifestyle factors, pes ticide use, and other agricultural exposures at enrollment (Alavanja et al. 1996). Most private and commercial applicators were men (97% and 96%, respectively), whereas most spouses were women (99%). Mortality data includ ing date and cause of death were obtained by linking the cohort to state mortality files and the National Death Index. We used the International Classification of Diseases, 9th Revision (ICD9) [World Health Organization (WHO) 1977] codes to identify suicides from 1993 to 1998 and 10th Revision (ICD10) (WHO 1992) codes for those from 1999 to 2009. ICD9 codes beginning with E95 or 95 or ICD10 codes X60-X84 identified suicides listed as underlying or contributing causes of death. We excluded 129 individuals < 18 years of age at enrollment, 1 individual from a couple in which the private applicator and spouse had both committed suicide (the excluded indi vidual was randomly chosen from the couple to avoid correlated death times), and an additional 7,528 cohort members missing covariate infor mation. Exposure assessment. Applicators and spouses provided information on pesticide use and other factors via questionnaires completed at enrollment (AHS 2011). This information included years of use (duration) and average days per year of use (frequency) for any pesticide in addition to ever use for 50 specific pesticides. For applicators, information was also collected on duration and frequency of use for 22 of the 50 pesticides. We evaluated suicide risk in rela tion to 10 pesticide categories (four functional: fumigants, fungicides, herbicides, and insec ticides; and six chemical: phenoxy herbicides, triazine herbicides, carbamates, organochlorine insecticides, organophosphate insecticides, and pyrethroid insecticides) as well as individual pesticides and overall pesticide use.
For both applicators and spouses, we evalu ated cumulative lifetime days of use of any pes ticide. For applicators only, we also evaluated cumulative lifetime days of use and intensity adjusted cumulative lifetime days of use for 22 individual pesticides, although we present results for only the 17 pesticides for which there were at least five exposed cases. Duration and frequency of use data were collected in seven and eight categories, respectively. Using values set to the midpoint of each category, or 50% greater than the lower bound in the highest category, we determined an applica tor's cumulative lifetime days of use for any pesticide and for each specific pesticide as the product of the duration and frequency values. We categorized cumulative lifetime days of use of any pesticide into quartiles based on the distribution of use for all applicators and spouses. We categorized specific pesticides into three levels based on the distribution of use for all applicators: a) none, b) used for less than or equal to the median lifetime days of use, and c) used for more than the median life time days. We also evaluated intensityadjusted cumulative lifetime days of use, calculated as previously described (Dosemeci et al. 2002) and again categorized in three levels: none, ≤ median, > median.
Statistical analysis. We employed Cox proportional hazards regression models to esti mate hazard ratios (HR) and 95% confidence intervals (CIs) for the association of suicide with each measure of pesticide use. Survival times were defined as the time (in days) between enrollment in the AHS and death for those who died from suicide (n = 110) or some other cause (n = 5,980) or the time between enrollment and 31 May 2009, for study par ticipants still alive (n = 75,908). We ran mod els for applicators only and for applicators and spouses combined. There were too few sui cides among spouses (n = 9) to analyze data on spouses alone. Possible confounding variables identified from prior reports (categorized as in Table 1) included age, sex, state of residence, race/ethnicity, education level, marital status, number of children in family (as a measure of social connection), size of farm worked last year, frequency of alcohol consumption dur ing the past 12 months, smoking status, and ever diagnosed with heart disease or diabetes (as measures of chronic disease). We did not consider depression as a potential confounder, as it may be an intervening variable. Although spouses were not asked about number of chil dren in family, farm size, or ever being diag nosed with heart disease, we inferred spouses' answers for the first two items from appli cators' responses and based the last item on spouses' responses to questions on myocardial infarction, angina, and arrhythmia.
We evaluated all covariates significantly associated with suicide in both unadjusted and age and sexadjusted models as potential con founders (α = 0.05). We then performed man ual backward selection using α = 0.05 to select a base model from the covariates identified in the first step. As an alternative model selec tion method, we performed manual forward selection using α = 0.05 and selected the same model as with the backward selection method. We evaluated whether additional covariates should be included in the base model with likelihood ratio tests (α = 0.05) and evalu ated model fit using the Akaike information criterion and the Schwarz Bayesian criterion. , number of children in family (≤ 1, > 1), frequency of alcohol consumption during the past 12 months (never, < every day, every day), and smoking status (never, past, current). The base model for applicators and spouses together also included sex.
Additional analyses included stratifying models by state of residence, number of chil dren in family, and use of chemicalresistant gloves (applicators only). We formally com pared HRs from the two strata using a 95% CI constructed from the standard error of the difference between HRs. To additional models, we added state, marital status, race/ethnicity, ever being diagnosed with diabetes, or cumula tive lifetime days of use of any pesticide. We evaluated all four functional classes of pesti cides together in a single model. Additionally, we evaluated carbamates, herbicides, organo chlorine insecticides, OPs, and pyrethroid insecticides together in another model. Some analyses were repeated, restricting the sample to male applicators, applicators who person ally used pesticides, individuals who did not report a physician diagnosis of depression, or cohort members ≤ 50 years of age at enrollment (because pesticide use among younger applica tors is probably less likely to change during fol lowup than use among older applicators). We separately evaluated suicides committed within 5 years of enrollment or more than 5 years past enrollment. We evaluated pesticide use dur ing the year prior to enrollment. Finally, we used withincategory medians to assess doseresponse trends in cumulative lifetime days of use and intensityadjusted cumulative lifetime days of use.
We used the P1REL090600 release of the Phase I data set, the REL090500.00 release of the demographic data set, and the REL201004.00 release of the mortality data set for this study and performed all analyses with SAS version 9.2 (SAS Institute Inc., Cary, NC).

Results
In models adjusted for age at enrollment and sex, risk for suicide was significantly greater if participants a) were > 65 years of age com pared with 36-45 years of age, b) were living in North Carolina, c) were of a race/ethnic ity other than white/nonHispanic, d) were divorced or separated compared with married or living as married, e) drank alcohol every day during the 12 months prior to enrollment compared with drank alcohol < every day, f ) were current smokers compared with never smokers, or g) had ever been diagnosed with diabetes or depression (Table 1). Suicide risk was significantly lower for women than men and for participants with > 1 child compared with ≤ 1 child (Table 1).
We found little evidence that suicide was associated with overall pesticide use among applicators and spouses combined (Table 2). We saw no significant dose-response relation ships between pesticide use (measured by dura tion, frequency, or cumulative lifetime days of use of any pesticide) and suicide. Only 37% of suicide cases (n = 41), all applicators, provided information on experiencing a high pesticide exposure event, and only seven cases experi enced one (HR = 1.13; 95% CI: 0.50, 2.57).
No cases reported being diagnosed with pes ticide poisoning, but medical visits related to pesticide use were not significantly associated with suicide among applicators (Table 2). For no specific functional or chemical class of pesticides, including OPs, was ever use sig nificantly associated with suicide (Table 2). Only use of pyrethroid insecticides had an HR of > 1.0 (HR = 1.09; 95% CI: 0.68, 1.74), but that was not significant.
Among applicators alone, none of the 22 individual pesticides showed a significant dose-response relationship between cumula tive lifetime days of use and suicide (Table 4). Repeating those analyses with intensityad justed cumulative lifetime days of use, we found HRs that were generally < 1.0 but not significant (data not shown). We found sig nificant inverse dose-response relationships between three intensityadjusted pesticides and suicide: (2,4dichlorophenoxy)acetic acid When we stratified by state of residence or number of children in family separately and used the everuse herbicide variables for applica tors and spouses, we observed the same results in each stratum: generally inverse associations between herbicides and suicide with a few being significant (data not shown). Formal compari son of the stratumspecific HRs showed no significant differences. Data on use of chemical resistant gloves were available only for appli cators. Stratification by this variable yielded similar results in both strata for general use and herbicide variables (data not shown).
Results remained unchanged when state, marital status, race/ethnicity, ever being diag nosed with diabetes, or cumulative lifetime days of use of any pesticide were added to the mod els one at a time (data not shown). Evaluating all four functional pesticide classes together in a single model or evaluating carbamates, her bicides, organochlorine insecticides, OPs, and pyrethroid insecticides together in another model did not change results (data not shown). Excluding cohort members who had been diag nosed with depression changed results slightly: a few more pesticides were inversely associated with suicide and a few more inverse associations were significant (data not shown). Restricting analyses to male applicators, to applicators who personally used pesticides, or to cohort mem bers ≤ 50 years of age at enrollment did not change results (data not shown). Likewise, eval uating pesticide use during the year prior to the enrollment of cohort members in the AHS did not change results (data not shown).

Discussion
We found no association between pesticide use up to enrollment in the AHS and subsequent incidence of suicide in pesticide applicators and their spouses. This finding was consistent for use of any pesticide or individual pesticides, for functional or chemical classes, and for cumula tive lifetime days of pesticide use. Results were the same for applicators and spouses together or for applicators only. Although many studies have reported higher suicide rates among farmers and pes ticide applicators than the general popula tion (Blair et al. 1993;Browning et al. 2008;Gunderson et al. 1993;Lee et al. 2002;MacFarlane et al. 2009MacFarlane et al. , 2010Meltzer et al. 2008;Miller and Burns 2008;Page and Fragar 2002;Stallones 1990), others have found lower rates (Pickett et al. 1993(Pickett et al. , 1999Torchio et al. 1994). A recent mortality analysis showed lower suicide rates among AHS participants than among the general populations of Iowa and North Carolina at least 15 years of age, although deaths due to unintentional inju ries were elevated (Waggoner et al. 2011). The lower suicide rate in the AHS could reflect misclassification of suicides as unintentional injuries (e.g., collision with objects).
Another possible explanation for the lower suicide rate in the AHS is that individuals at risk of suicide may be less likely to enroll ini tially. This selfselection, if nondifferential by exposure, should not bias estimated HRs, although it might reduce precision. To reduce the influence of the healthy worker effect on their standardized mortality ratios (SMRs), Waggoner et al. (2011) calculated relative SMRs (rSMRs), a ratio of a causespecific SMR and the SMR for all other causes except the cause of interest. The rSMR for suicide among applica tors was 1.06 (95% CI: 0.87, 1.28) (Waggoner JK, personal communication), suggesting that the deficit in suicides observed in the AHS may be due to the healthy worker effect.
Previous studies that reported higher suicide rates among farmers and pesticide applicators had lessdetailed pesticide exposure information than that available in the AHS (London et al.  2005; Parrón et al. 1996;Pickett et al. 1998;Stallones 2006). For example, one study inferred exposure from the fact that the study area had the highest density of greenhouses in the world (Parrón et al. 1996), and another inferred expo sure from the occupation listed on death cer tificates (Stallones 2006). Pickett et al. (1998) found no association between suicide among Canadian farmers and three exposure measures: a) acres sprayed with herbicides, b) acres sprayed with insecticides, and c) the costs of purchased agricultural chemicals. We did not replicate a previous finding from the AHS that suggested an association of chlorpyrifos with suicide (Lee et al. 2007). We had 65% more suicide cases among applicators using chlorpyrifos (n = 38) than in the previous study (n = 23), suggesting that the present results may be more reliable.
Other rural lifestyle factors besides pesticide use may explain the higher suicide rates among farmers observed in other studies. Farmers have ready access to lethal means such as guns, pes ticides, and other chemicals (Booth et al. 2000;Goldney 2002;Gregoire 2002;Hawton and van Heeringen 2009;Nock et al. 2008). The social change that occurs when small farms are incorporated into larger ones could also con tribute to higher suicide rates via higher unem ployment rates and a breakdown in family relationships (Goldney 2002;Stallones 1990).
Factors previously reported to increase suicide risk, such as old age, male sex, unmarried status, social isolation, frequent alcohol consumption, frequent smoking, and being diagnosed with depression, did so in our study also, suggesting that there is nothing unusual about suicide cases in the AHS. Further, AHS farmers are generally similar to other U.S. farmers (Lynch et al. 2005), so our results should be generalizable.
The tendency for herbicide use to be inversely associated with suicide, even if only a few associations were statistically significant, is perplexing. These inverse associations did not change meaningfully in a variety of analyses. Individuals who later committed suicide may have experienced mental problems, financial stress, stress in general, or other problems that might have caused them to farm less or use her bicides less than they otherwise would, leading to apparent inverse associations between herbi cide use and suicide. However, the prospective design of our study helps alleviate this concern, particularly because results were similar for sui cides committed within 5 years of enrollment and those committed later.
Our results may be surprising, given the association between pesticide use and physician diagnosed depression found previously in the AHS (Beseler et al. 2006(Beseler et al. , 2008. Suicide, how ever, is not equivalent to depression. Only 17% (n = 18) of suicide cases in our study had been diagnosed with depression (Table 1). Other risk factors such as alcohol/substance use, mood disorders, and personality traits/disorders (e.g., impulsivity, aggressiveness, hopelessness, high emotional reactivity) are also important risk factors for suicide (Goldney 2002;Hawton and van Heeringen 2009;Nock et al. 2008) that cannot be ignored. Although depression may be underreported in the AHS, underreporting is not likely to account for our results, and our results did not change meaningfully when we excluded cohort members who had been diag nosed with depression.
Our study has several strengths. Selection bias was likely to be minimal, as 84% of eli gible applicators and 75% of eligible spouses were enrolled in the AHS. The AHS collected detailed information on pesticide use before suicides occurred, including data on the use of individual pesticides and duration, frequency, and intensity of use. We were able to control for potential confounding factors. Finally, sui cide from death certificates is a relatively valid outcome (Moyer et al. 1989).
Limitations include the small number of suicides (n = 110), which meant low power for estimating HRs for some individual pes ticides. Further, participants were farmer owners/ operators and their spouses or they were commercial applicators, that is, not farm workers. Therefore, results may not be gen eralizable to farm workers who may be more highly exposed than AHS farmers. Some infor mation potentially useful for a suicide analysis was unavailable from AHS questionnaires. For example, information on financial situation or pertinent life events such as divorce or death of a family member was unavailable, as was information on personality or access to mental health care. Measures of acute highintensity pesticide exposure were unavailable for most of the cohort, particularly suicide cases, and doseresponse information was available for only 22 of 50 pesticides. Using data on past pesticide use instead of ongoing use could also misclassify pesticide use if ongoing use, and not past use, is what increases suicide risk.
In conclusion, we found little evidence that pesticide use increases suicide risk in pes ticide applicators and their spouses. This find ing, although based on relatively few suicides, was consistent across multiple measures of pes ticide use and was robust to varying analytic strategies. Although this null finding needs confirmation in different populations, it could be reassuring to farming populations.