BACKGROUND

Suicide prevention has been designated a ”national imperative” in the United States1 and addressing suicide among veterans is a high priority. Many veterans receiving treatment in the Veterans Health Administration (VA) system have risk factors for suicide, including older age, male gender, and multiple medical and psychiatric comorbidities; veterans treated in the VA have approximately 50 % greater risk than the general population for suicide.2

Primary care has been highlighted by experts as a key setting for suicide prevention3 for several reasons. A systematic review of published studies up to the year 2000 indicated that, on average, 45 % of suicide victims have contact with primary care clinicians within 1 month of suicide.4 Several studies have shown that a minority of these patients have contacts with mental health clinicians.4,5 Primary care clinicians, often working upstream from specialty mental health, may therefore be in a uniquely advantageous position to intervene with patients at risk for suicide.

There are a number of challenges to enhancing suicide prevention in primary care. Little information is available in the literature describing the characteristics of veterans who see primary care clinicians and subsequently die by suicide. Prior studies have been limited to one state,5 examined suicidal ideation rather than suicide,6,7 synthesized data pre-dating recent VA suicide prevention initiatives,8 or focused on particular mental health populations9,10 or on patients referred from primary care.11

The objective of this study was to describe demographic, clinical, and psychosocial characteristics of a multi-state sample of veterans who received VA primary care in the 6 months prior to death, and compare these characteristics to those of a group of control patients who received care over the same time period from the same clinicians. Our goal was to highlight salient patient characteristics for primary care clinicians to identify and address among their patients.

METHODS

This study was approved by the Institutional Review Board of the VA medical center where the study was conducted. Data sharing agreements were completed and appropriate permissions obtained in order to access and analyze all sources of data.

Collaboration Between Investigators and VA Operations Partners

In 2010, Secretary of Veterans Affairs Shinseki engaged the governors of all U.S. states, requesting support and collaboration to improve the timeliness and utility of suicide mortality reporting for veterans. This request led to the development of state-level data use agreements for the purpose of sharing identified mortality data with the VA for known suicides (as of 2012, data use agreements had been approved by 34 states, and data had been received from 21 states). At approximately the same time, the investigators began a dialogue with the VA Mental Health Services–Suicide Prevention Office regarding suicide prevention in primary care. The Office encouraged us to pursue collaboration with them in order to help VA identify opportunities to intervene with veterans seen in primary care who may be at risk. As a result of this dialogue, a data-sharing agreement was established between the Veterans Integrated Service Network (VISN)-2 Center of Excellence for Suicide Prevention and our investigator group, to provide extracts of state death certificate data spanning 1999 to 2009.

Sources of Data

For the current project, we used data from 11 states (first set of states to complete data use agreements with the VA). This data set contained information on all individuals determined on death certificates to have died by suicide, and whose Social Security number matched exactly to an individual listed in the VA’s National Patient Care Database. We then obtained data on veteran status and prior healthcare from the national VA Corporate Data Warehouse (CDW),12 a collection of demographic, clinical, utilization, and cost data from many databases accessed securely via the VA Informatics and Computing Infrastructure (VINCI) Workspace.12 Additional data describing psychosocial context variables were obtained through manual medical record review using the VA Compensation and Pension Records Interchange.

Settings

The 11 states were diverse in terms of geographic distribution across the U.S. and population (Fig. 1). Within these states, 41 unique VA facilities were represented. Our sample represented approximately one-third of national VA facilities and, as compared to the national VA patient population, was very similar across multiple patient and facility characteristics (Table 1).

Figure 1
figure 1

Death certificate data were obtained from Alabama, Idaho, Maine, Michigan, Minnesota, Missouri, North Carolina, Nebraska, Nevada, Vermont, and Washington.

Full size image
Table 1 Patient Characteristics and Complexity of Veterans Affairs Facilities in Study Year (2009)
Full size table

SUBJECTS

Cases

We included all veterans from 11 states who died by suicide in 2009 and who had contact (telephone or in person) with a VA primary care clinician (physician, nurse practitioner, physician assistant) in the 6 months prior to suicide. There were no exclusion criteria.

Controls

Using CDW, controls were identified and matched 2:1 to cases based on a predefined set of characteristics. Typically, cases are matched to controls on a limited number of variables to control for effects of variables that prior work has demonstrated affect the outcome variable.13 Our approach also allowed for estimation of the association of the non-matched variables on the outcome using multivariable analysis.13 For each case, we first identified the primary care clinician seen most often in the 6 months prior to death. Then, we identified potential controls as all patients who had received care from the same clinician in the same 3-month time period (quarter). We then matched cases to controls on sex, then age—rates and methods of suicide have been shown to vary by sex and by age.14,15 When multiple suitable controls were available, we selected the control patient who saw the matched primary care clinician most proximally to the case’s primary care visit. Exact age matches were obtained for 93 % of controls, and 86 % of controls had been seen within 2 weeks of case primary care visits. An index date for each case–control triad was defined as the case patient’s date of death.

MEASURES

Demographic Characteristics

Date of death and education level were obtained from state death certificates. Veteran age, sex, race, ethnicity, marital status, ZIP code, and VA service connected disability rating (categorized as greater or equal to vs. less than 50 % disabled) were extracted from CDW. Rural versus urban residence was determined by matching patient residence ZIP codes with Rural–urban Commuting Area Codes.16,17 We also used CDW to determine the facility in which primary care clinician visits occurred.

Clinical Characteristics

We used ICD-9-Clinical Modification (ICD-9-CM) diagnoses from CDW that had been recorded by clinicians to document clinical encounters in the 6 months prior to index date. We specifically identified any diagnosis of mood disorders, psychotic disorders, substance including alcohol use disorders, post-traumatic stress disorder (PTSD), other anxiety disorders, traumatic brain injury (TBI), and pain diagnoses (ICD-9-CM codes available from authors on request). The Selim comorbidity index18,19 was used to measure overall medical morbidity; it counts across 30 medical conditions and six mental conditions from ICD-9-CM diagnoses made in the prior year.

We used a systematic, iterative process to develop a medical record data collection tool to record documentation of previously identified suicide risk factor variables in the 6 months prior to the index date; we have used similar approaches in prior studies.9,20–24 Clinical variables determined using this method included sleep problems, functional decline, anger, and suicidal ideation endorsement; psychosocial context variables included housing instability, isolation, grief/loss of a loved one, recent move, and financial, legal, job/school, or relationship problems (Table 2). These items were selected based on prior research suggesting possible relationships among these variables and suicide risk, and recommended data elements from the Centers for Disease Control and Prevention.9,15,20,22,24–27 The majority of these variables were recorded dichotomously— counted as “yes” if the construct was documented in any progress note by clinicians or other medical staff (e.g., medical assistants or nurses) during the 6 months prior to the index date. For suicidal ideation and attempts, there were three levels: not asked (suicidal ideation not mentioned in notes); not endorsed (veteran asked, but did not endorse); and endorsed (veteran asked and endorsed). The investigators and experienced medical record reviewers developed the data collection tool over several iterations until they agreed on content and coding, and inter-rater reliability was excellent (individual item and overall Kappa ≥ 0.75).28 Each patient’s record was reviewed independently with monthly re-testing of inter-rater reliability (which remained excellent).

Table 2 Coding Definitions for Data Extracted Via Manual Medical Record Review
Full size table

ANALYSIS

First, using t-tests for linear variables and chi-square tests for categorical variables, we assessed bivariate relationships between demographic, clinical and psychosocial context variables and case/control status. PTSD was examined separately from other anxiety disorders. Next, we used multivariate conditional logistic regression to examine associations between candidate predictor variables and suicide. Women were excluded from the analyses, as our sample contained few women, and suicide behaviors and risks for men and women have been shown to differ in important ways.29,30 In order to account for possible within-facility correlation, regression models were clustered on facility. To limit the number of variables in regression models, we used all variables (except the composite variable, “any” psychiatric diagnosis) that were associated with suicide in bivariate tests at p < 0.10. Alcohol use disorder was collapsed with other substance use disorders. All analyses were conducted using Stata 13.1 with two-sided alpha = 0.05, unless otherwise stated.

RESULTS

Across the 11 state sample, the 269 cases and 538 controls had similar age and sex characteristics compared to the national veteran population (Table 1). During their study periods, cases and controls in our sample had visits with 236 clinicians in total, 65 (28 %) nurse practitioners or physician assistants, and 171 physicians (73 %).

Because males comprised 97 % of the veteran sample, females (n = 24) were excluded from the analyses; the final sample included 261 cases and 522 controls. Table 3 shows demographic and clinical characteristics of male cases and controls, as well as unadjusted bivariate comparisons between the two groups. Across the sample, mean age was 63 years, 84 % were white, and 66 % lived in urban areas. Cases were significantly more likely than controls to be unmarried, white, and to have major depression, bipolar disorder, anxiety disorder other than PTSD, alcohol or other substance use disorder diagnoses. Cases were also more likely than controls to have documented functional decline, sleep disturbance, or expressions of anger. Within the subgroup of cases and controls specifically asked about suicidal ideation or attempts by clinicians or staff (44 %), 59 % vs. 16 % endorsed ideation or attempts, respectively (p < 0.001). Finally, stressful or negative psychosocial context variables were significantly more prevalent among cases compared to controls, except for housing instability. PTSD diagnosis was not significantly associated with suicide, nor were pain diagnoses or general medical comorbidity.

Table 3 Characteristics of Male Cases and Controls over 6 months Prior to Index Date
Full size table

In our primary set of multivariate models of men in the sample (Table 4), non-white race (OR = 0.51; 95% CI = 0.27–0.98, p = 0.04) and VA service-connected disability rating (OR = 0.54; 95% CI = 0.36–0.80, p < 0.01) were associated with decreased odds of suicide. Diagnosed anxiety disorder other than PTSD (OR = 3.52; 95% CI = 1.79–6.92, p < 0.001), recent functional decline (OR = 2.52; 95% CI = 1.55–4.10, p < 0.001), depression diagnosis (OR = 1.82; 95 % CI = 1.07–3.10, p = 0.03), and endorsement of suicidal ideation (OR = 2.27; 95% CI = 1.07–4.83, p = 0.03), were associated with greater odds of suicide. We did not detect significant relationships between marital status, substance (including alcohol) use disorder diagnoses, the psychosocial context variables, and suicide.

Table 4 Multivariate Correlates of Suicide (n = 783)
Full size table

DISCUSSION

Many studies have examined risk factors for suicide. But to our knowledge, this is one of only a few case–control studies to examine correlates of suicide across multiple domains (demographic, clinical, psychosocial context) using a multi-state sample of veterans treated in VA primary care. We found that veterans seen in VA primary care who die by suicide frequently have diagnosed psychiatric disorders, as well as sleep problems, functional decline, expressed anger and endorsed suicidal ideation. These veterans also have high rates of psychosocial stressors; in particular, relationship problems, isolation, and financial problems. Based on our multivariate models, several characteristics may be especially relevant for veterans treated in primary care: anxiety disorder other than PTSD, recent functional decline, depression, and endorsement of suicidal ideation.

Our findings differ somewhat from those of prior research by suggesting that anxiety disorder may indicate greater risk for suicide among veterans compared to substance use disorder or depression; in our sample, the odds of suicide associated with anxiety diagnoses were approximately twice those of the odds associated with depression. Conner and colleagues, who recently studied suicide risk in a national sample of veterans receiving VA healthcare, also found anxiety to be uniquely associated with suicide.31 Prior work examining anxiety and suicide risk has perhaps been confounded by studying patients already receiving depression treatment or who have depression diagnoses. Further, while anxiety and depression share similar features, anxiety may be differentiated from depression by the relative prominence of positive affect;32,33 thus, individuals suffering from anxiety may have the ability and sufficient discomfort to plan and act, while those suffering from depression alone may have less ability to act. Given the relatively high comorbidity of anxiety and depression, it would be important for future work to prioritize elucidation of the effects of these conditions versus their combination on suicide risk.

We found that functional decline was also strongly associated with increased odds of suicide. Prior work has identified relationships between functional decline or functional status and suicide,34,35 and it has been noted to be predictive of suicide independent of physical illness. However, it remains unclear how functional decline relates to depression among suicide decedents; some studies show an independent association between functional status and suicide, while others suggest that mood disorders mediate this relationship.36 The issue of functional decline may be particularly relevant for older populations: while the average age of veterans in our sample was 63 years, this is only slightly older than those who die by suicide in VA nationwide37 or among the general population.38 Taken together, our findings suggest that clinicians may need to place greater emphasis on understanding and integrating information on functional decline and anxiety into risk assessment activities. The findings further suggest that we may wish to prioritize development of interventions that assist people in coping with functional decline and treating anxiety conditions.

It is considered standard of care for clinicians to inquire about suicidal ideation among patients who may be at risk. Our finding that endorsement of suicidal ideation is associated with suicide supports this standard, and is consistent with recent work done by Simon et al.,39 which found that endorsement of the ninth item of the Patient Health Questionnaire-9 (“thoughts that you would be better off dead or of hurting yourself in some way”) was also associated with increased risk of suicide behaviors. Our finding underscores that when a veteran endorses suicidal ideation, it must be taken seriously. Unfortunately, veterans often do not disclose suicidal ideation in healthcare settings prior to suicide, even when asked.9,24 Indeed, trust is a critical factor in disclosing suicidal ideation,40 and clinical discussions of suicidal ideation would be well-served by approaching the topic in the context of strong patient–clinician rapport.

Our finding that the odds of white veterans dying by suicide are approximately twice those of non-white veterans is commensurate with prior research.35,41 Similarly, though less well-studied, service connected disability status has been shown to be protective as well, presumably by increasing access to healthcare services by decreasing out of pocket costs.5,42,43

We found that PTSD diagnosis was not associated with suicide. Evidence syntheses have shown mixed findings for linking PTSD to suicide among veteran and military populations.15 Some studies have suggested that PTSD treatment may attenuate the relationship between PTSD diagnosis and risk of suicide.44 Methodological differences in assigning PTSD status may also explain some of these mixed findings (e.g., using PTSD screens rather than ICD diagnosis codes45). Neither did we find relationships between pain diagnoses and suicide. While Ilgen et al.46 recently showed a relationship between pain diagnoses and suicide, this relationship may manifest over years—in our study, we examined only the 6 months prior to death. We note that the models used in most of these prior studies have generally not adjusted for several variables included in our study, including functional decline and psychosocial context.

There are several important limitations. Differences in mortality-reporting processes across states and misclassification of veteran status have the potential to affect our estimates; however, our approach of specifically linking death certificate data to CDW to identify individuals who actually received care in VA should minimize misclassification effects. While we did not use a national sample, our sample does appear to be nationally representative with respect to VA healthcare users. We did not have access to patient-level measures. Instead, our data were highly dependent on the accuracy and extent of clinician and staff documentation in the medical record. We also note that there were too few women veterans (n = 24), veterans of Iraq and Afghanistan wars (n = 10), and veterans with diagnoses of traumatic brain injury (n = 5) to examine these factors in our multivariable analyses; however, data sets using expanded years are currently being analyzed to examine suicide risk among these groups in more depth. Finally, exploration for interaction effects among predictors was beyond the scope of this study—such analysis has potential to provide additional information on how various factors might compound to confer risk, which would be critical to emerging risk-profiling efforts.31

This project demonstrates the value of Health Services Research & Development investigators collaborating with VA Operations partners to address issues of key importance to the VA and the nation in a timely fashion. This collaboration enabled the investigators to study one of the first nationally representative case–control data sets of veterans who died by suicide after receiving VA healthcare. The original plan for this project involved the investigators establishing data use agreements with individual states, which would have been a very complex and lengthy process. Instead, VA Operations partners facilitated sharing of data in a way that greatly enhanced efficiency. In turn, the investigators are now able to contribute these findings to the field, which may help shape VA policy and priorities. Preliminary results of this and related work were discussed at a summit of VA suicide prevention researchers that took place last year. In addition, since this collaboration began, several of the investigators have worked with representatives from VA Mental Health Services, the VISN-2 Center of Excellence, and other VA operations leaders through additional meetings and presentations on suicide research, and workgroups dedicated to shaping strategic planning for VA mental health.