Logan J, Hall J, Karch D. Suicide Categories by Patterns of Known Risk FactorsA Latent Class Analysis. Arch Gen Psychiatry. 2011;68(9):935–941. doi:10.1001/archgenpsychiatry.2011.85
Author Affiliations: Division of Violence Prevention, Etiology and Surveillance Branch, National Center for Injury Prevention and Control, Centers for Disease Control and Prevention, Atlanta, Georgia.
Context Multiple risk factors contribute to suicides; however, patterns of co-occurrence among these factors have not been fully identified.
Objectives To assess patterns of known suicide-related risk factors, classify suicide decedents by these patterns, track class proportions during a 6-year period, and characterize decedents across the classes to help focus prevention strategies.
Design, Setting, and Participants Latent class analysis was conducted using 2003-2008 data from the National Violent Death Reporting System. The population included 28 703 suicide decedents from 12 US states.
Main Outcome Measures The known risk factors included having the following: mental health conditions; a sad or depressed mood; substance abuse problems; medical problems; recent crises; financial, job, and legal problems; intimate partner and other relationship problems; and perpetrated interpersonal violence.
Results Nine distinct patterns of risk factors emerged. Of these classes, 1 only endorsed mental health–related factors and 1 only endorsed alcohol- and substance abuse–related factors; however, 7 classes of decedents had distinct patterns of factors that spanned multiple domains. For example, 5 of these classes had mental health factors with other risks (eg, substance abuse, financial problems, relationship problems, a recent crisis, and medical problems). Two classes had recent crises with relationship problems; one of these classes also had high probabilities for criminal problems and interpersonal violence. Class proportions differed during the 6 years. Differences across classes by demographic and event characteristics were also found.
Conclusions Most suicide decedents could be classified by patterns of risk factors. Furthermore, most classes revealed a need for more connected services across medical, mental health/substance abuse, and court/social service systems. Reducing fragmentation across these agencies and recruiting family, friend, and community support for individuals experiencing mental health problems and/or other stress might significantly reduce suicides.
For years, suicide has been the fourth-leading cause of death in the United States for those between 10 and 65 years of age.1 To prevent suicide, one strategy has been to identify and prevent the contributing circumstances (ie, risk factors) common to those who take their own lives. Health-related risk factors have included mental health (MH) problems2- 6; certain medical conditions (eg, cancer)7- 9; prior suicide attempts2; and substance abuse.2,10 Other life-stress–related risk factors have included intimate partner (IP) problems11; job loss12- 14 or financial problems15; recent crisis or stress2; and experiences with other forms of violence (eg, child abuse or peer abuse).16,17 While any one of these factors increases the risk of suicide, most suicide decedents have multiple risk factors prior to death.2 Therefore, strategies designed to prevent suicide might be more effective if they address multiple risk factors.2 To provide direction for such strategies, we identified common patterns of health-related and life-stress–related factors (ie, the most common co-occurring factors) known to increase risk of suicide among a suicide decedent population, classified decedents by these patterns, and assessed annual class proportions during a 6-year period to determine which co-occurring factors need more urgent attention. We also characterized decedents (eg, demographic characteristics) and the incidents (eg, the mechanisms and weapons used, location of incident) by these patterns to further understand their surrounding circumstances.
The National Violent Death Reporting System (NVDRS) is an incident-based surveillance system that captures details on different types of violent deaths, including suicides.18,19 Data collection began in 2003 in 7 states (Alaska, Maryland, Massachusetts, New Jersey, Oregon, South Carolina, and Virginia). In 2004, 6 states were added (Colorado, Georgia, North Carolina, Oklahoma, Rhode Island, and Wisconsin). In 2005, 4 states were added (California, Kentucky, New Mexico, and Utah), bringing the total (as of 2008) to 17 states. Data collection is statewide with the exception of California, which collects data in only 4 counties (Los Angeles, Riverside, San Francisco, and Santa Clara).
The NVDRS collects details such as information on decedents, the mechanisms and weapons involved in the incidents, and the circumstances leading up to the violent deaths.18 Data are collected from coroner or medical examiner reports, toxicology reports, law enforcement records, and death certificates and are linked by incident into a single repository.
States manage data collection through state health departments or a subcontracted entity, such as a medical examiner's office, where data are gathered and coded by trained abstractors. Data may be manually extracted from reports or imported electronically from other systems (eg, Bureau of Vital Statistics death certificate files). All data are reviewed by the abstractor to ensure accuracy of the codes and adherence to the NVDRS coding manual.20 The NVDRS has been described in further detail elsewhere.18,20
The data used for this analysis represent calendar years 2003 to 2008. Case identification was conducted from an NVDRS database that was updated through March 2010. Suicide decedents were identified by the final manner of death they were assigned by the data abstractors; abstractors make this determination using information from the death certificates (ie, the listed manner of death and the external cause of death codes) as well as from the coroner or medical examiner and law enforcement reports. Based on this initial criterion, 49 265 suicide decedents were identified from 17 states. Two additional inclusion criteria were also required. First, because this study focused on patterns of co-occurring risk factors, a case had to have circumstance information documented for at least 2 health-related or life-stress–related risk factors. This criterion yielded 36 756 cases (74.6%) overall. Second, states had to have at least two-thirds of their cases meet this first criterion to be included. States that did not meet this criterion might have had difficulties with capturing suicide circumstance information. In such situations, it is likely that circumstance data were not missing at random; therefore, a fair assessment of patterns for these states might have been problematic. As a result of this criterion, cases from 5 states (South Carolina, Georgia, Kentucky, California, and New Jersey) were excluded. The final sample included 28 703 cases from 12 NVDRS states.
The known health-related and life-stress–related risk factors used to assess the patterns included having a current depressed mood; a current MH condition (ie, a documented condition); a history of MH conditions (approximated by having a history of MH treatment); alcohol problems; other substance abuse problems; medical problems believed to have precipitated the suicide (eg, chronic pain or cancer); financial problems; job problems; criminal legal problems; a recent crisis (within 2 weeks); a recent exposure to a death (eg, recent death of a friend or family member); IP or other relationship problems; a history of perpetrating interpersonal violence; and a history of suicide attempts. We also included information on whether the decedent was suspected of using alcohol at the time of the incident and whether the decedent disclosed his or her intent to commit suicide. All of the variables were not mutually exclusive. Three known risk factors (ie, school problems, other legal problems, and violence victimization) were excluded because of low prevalence estimates. (Throughout this article, we use the term known risk factors because these items have been identified as risk factors for suicide in previous studies; in the NVDRS, these items are often referred to as“preceding circumstances of suicide” because the system is currently unable to assess relative risk.)
After categorization by the pattern of known risk factors, we also characterized the decedents by demographic factors (eg, race/ethnicity, sex, age, and marital status), military status (ie, whether the decedent ever served in the US military), and whether they accessed MH treatment prior to death. We also characterized the incidents by the type of location (eg, home, recreational area, or commercial area) and the type of mechanism and weapon used (eg, firearm, poisoning, hanging).
Latent class analysis (LCA) was used to identify distinct patterns of known risk factors for suicide among the decedents and to classify these decedents by these patterns. The LCA method is very useful in identifying distinct classes based on patterns of nonmutually exclusive categorical variables21 such as the known risk factors.21 To determine which model best explained the patterns of known risk factors (eg, a 3-class model identifies 3 classes [or patterns], a 4-class model identifies 4 classes, and so forth), we used the Bayesian information criterion (see footnote i of the eAppendix). We also considered the entropy value (see footnote ii of the eAppendix). Based on these criteria, we chose a 9-class model. Details on LCA and our model selection process are provided in the eAppendix and eTable. This analysis was first conducted for suicide decedents in each NVDRS data year to assess the stability of latent class structures (eg, whether the characteristics of each class identified were similar in each data year). Because the same classes (ie, distinct patterns) emerged in each data year, we then consolidated the data for the 2003 through 2008 data years. All LCAs were performed using Mplus version 5.1 statistical software (Muthen and Muthen, Los Angeles, California).
To show how each class had distinct patterns of known risk factors, we provided the class-specific item endorsement probabilities for each factor. We also assessed the proportion of each class by year and characterized each of the classes by the descriptive variables. The Cochran-Armitage test for trends was used to test for the presence of linear trends in the proportion of decedents in each latent class across data years examined in this study. The tests determine significant increases or decreases in class proportions using data from all years beyond a reference year. Because the NVDRS was expanding from 2003 through 2005, significant changes detected by these tests might have been an artifact of the system capturing more cases; therefore, Cochran-Armitage tests were conducted for each class using 3 starting points: 2003, 2004, and 2005. Each analysis used 2-tailed tests, and significance was determined at the P = .05 level.
The number of classes, the proportion of respondents assigned to each class, and the endorsement probabilities for each factor by class are presented in Table 1.
Class 1 exemplified MH conditions with alcohol problems (n = 2567 [8.9%]). This class endorsed high probabilities for MH-related risk factors, alcohol dependence, and suspected alcohol use at the time of death. Compared with other classes, this group also had the highest probabilities for having other substance abuse problems and previous suicide attempts.
Class 2 displayed MH conditions with recent crises (n = 4193 [14.6%]). This class had high probabilities for all MH-related risk factors as well as a moderately high probability for having a crisis within 2 weeks of death. This class differed from the first by having low probabilities for alcohol and substance abuse problems.
Class 3 had MH conditions only (n = 7138 [24.9%]). This class was the last of the 3 classes that endorsed all of the MH-related risk factors. In contrast to the first 2 classes, this class had low probabilities for all other factors.
Class 4 decedents had a current depressed mood with financial problems (n = 3574 [12.5%]). This class had a high probability for being recognized as having a depressed mood around the time of death; however, they had a low probability for having a documented MH condition. This class also had the highest probability for having experienced financial problems.
Class 5 exhibited alcohol problems with other life stresses (n = 1249 [4.4%]). These decedents had high probabilities for alcohol-related risk factors and moderate to high probabilities for a variety of life stresses (eg, a recent crisis, job problems, IP problems). They also had a moderately high probability for being identified as having a depressed mood but a low probability for having a documented MH condition.
Class 6 had medical problems with depressed mood (n = 2641 [9.2%]). For this class, the probability for having medical problems that were believed to have contributed to the suicide was 1.000, or 100%. Like classes 4 and 5, this class also had a moderately high probability for being identified as having a depressed mood but a low probability for having a documented MH condition.
Class 7 experienced recent crises with criminal legal problems (n = 1875 [6.5%]). This group had the highest probabilities for having a recent crisis and recent criminal problems. This group also had moderate probabilities for having relationship problems and perpetrating interpersonal violence (eg, domestic violence, violence among acquaintances).
Class 8 encountered IP problems with recent crises (n = 3015 [10.5%]). This group also had a moderately high probability for being suspected of alcohol use at the time of death.
Lastly, class 9 is characterized as being suspected of alcohol use at the time of death (n = 2451 [8.5%]). This class also had a moderately high probability for having history of alcohol dependence.
Across all years, class 3 was the most common class of decedents (Table 2). The Cochran-Armitage tests supported significant changes in the class proportions for 3 classes, regardless of which reference year (2003, 2004, or 2005) was used (comparisons were significant at the P = .05 level). The annual proportion of decedents allocated to class 2 significantly increased during the study period, while the annual proportions of decedents allocated to classes 6 and 7 steadily decreased in this time.
Many similarities were observed across the classes. For example, in each class, most decedents (≥73.9%) were of white non-Hispanic racial/ethnic status, most decedents (≥64.0%) were male, the most common mechanism or weapon used was a firearm (≥38.5%), and the most common location of death was at a house or an apartment (≥59.5%) (Table 3). Also, classes 1 through 3 (those identified with a current MH condition) had similar patterns of MH diagnoses, with depression or dysthymia listed as the most common diagnosis (75.2%-77.8%), bipolar disorder as the second most common diagnosis (13.5%-14.7%), and anxiety disorder as the third most common disorder (7.7%-9.1%).
Several differences were also discovered. Classes 1 through 3, those with documented MH conditions, had the highest proportions of female decedents (25.6%-36.1%). These 3 classes were the only classes who had decedents in MH treatment prior to death (≥75.9% were in treatment) and were the 3 classes with the highest proportion of decedents who died by poisoning (≥25.2%). Class 7 (those with recent crises and criminal legal problems) had the highest proportion of decedents younger than 20 years, whereas class 6 decedents (those with contributing medical problems) had the highest proportion of decedents aged 65 years or older. Class 6 also had the highest proportion of decedents who were widowed, divorced, or separated from their spouse and who had served in the military.
Suicide is a complex phenomenon often resulting from multiple risk factors.2,22 We found that distinct patterns among these factors exist. Similar to previous research, we found the following: the majority of decedents were non-Hispanic white males1,23; the most common place of death was at a residence24; the most common mechanism or weapon used was a firearm23,25; decedents with contributing medical problems were older7; those with criminal legal problems were younger26,27; and many decedents who had MH conditions also had substance abuse problems.28- 31 Furthermore, previous research found that females are more likely than males to access MH treatment32,33 and attempt suicide by poisoning or less lethal means.34- 37 Therefore, as expected, we found higher proportions of females and poisoning deaths among classes with higher proportions of decedents who accessed MH treatment.
Additionally, we found that more than 10 000 decedents who had MH issues (ie, who either had a documented condition or were known to have a sad or depressed mood) were experiencing other risk factors prior to death and that the proportions of decedents experiencing MH problems with a recent crisis (class 2) have been increasing in recent years, although more years of data might be needed to substantiate this trend. Furthermore, at least 75% of those in classes 1 through 3—classes that consisted of half the entire decedent population—were in MH treatment around the time of death, which suggests that more suicide prevention strategies are needed with treatment. The MH treatment alone might not sufficiently address both the MH condition and potential co-occurring life stressors, or these stressors might be disrupting the treatment process (eg, a recent crisis or substance abuse might deter compliance with treatment regimens). For those with only MH conditions, additional social support might still be needed to help monitor treatment and prevent suicidal behavior.
We also found that many decedents had a recent crisis with IP or other relationship problems in the absence of having known MH conditions, as indicated by classes 7 and 8. These findings suggest that there might be a need for services that help build strong coping skills to better handle relationship issues during times of crises, especially because a high proportion of the class 7 decedents perpetrated interpersonal violence prior to death. Furthermore, a high proportion of class 7 decedents had criminal legal problems, indicating that the court/legal system might be in a position to act as a gateway for multicomponent suicide and violence prevention strategies. Such strategies can incorporate individual counseling with family therapy, counseling on building positive relationships and coping skills, and assistance with accessing additional resources that can help reduce stress on relationships (eg, employment assistance programs).38
Similar to Flensborg-Madsen et al,39 we also found that alcohol-related problems can be common among suicide decedents in the absence of psychiatric comorbidity, as indicated by classes 5 and 9. Alcohol-related problems were the most transparent characteristics for these decedents. Both of these classes had moderate probabilities for being identified as having a sad or depressed mood, which could mean they had MH problems that were not yet diagnosed. The class 5 decedents had moderately high probabilities for having a recent crisis and other life stressors, suggesting that these decedents might have been using alcohol as a coping device, which warrants efforts to reduce this maladaptive coping behavior in favor of more socially adaptive ways of dealing with stress.
The link between medical illness and depression with suicide in older adults is well documented.40- 42 This study found that class 6 decedents (those with contributing medical problems) mostly comprised older adults. Nearly half of them were identified as having a depressed mood, which indicates a need for increasing social support and MH screening and treatment once diagnosed as having a severe or terminal medical condition. Class 6 deaths were also characterized by the highest proportion of widowed, divorced, or separated decedents and more than double the number of decedents with prior military service than any other class, both of which may indicate the need for additional support systems.
In light of these findings, many limitations of this study should be considered. First, the NVDRS data were not nationally representative. As the NVDRS expands to include all 50 states, the class structure, the class proportions, or both might change over time. Second, changes in class proportions over time might be a reflection of changes in the population demographic characteristics, which signifies that scientists must continue to observe patterns of risk factors if they hope to help prevent suicide. Third, the reported information on MH and medical health problems and substance abuse was obtained from coroners and medical examiners, family members, and friends of the victims. As such, it may be incomplete based on the knowledge level of the informant. Furthermore, many of these decedents might have been living with MH conditions but never received a diagnosis. Because of the potential underdiagnosis of MH conditions, the proportion of decedents placed in classes 4 through 9 might have been overestimated and the proportion estimates for classes 1 through 3 might have been underestimated. Fourth, 5 states had to be excluded because of lack of circumstance information. The excluded population did not largely differ from the study population by demographic characteristics such as race/ethnicity, sex, and age (significant differences were detected because large populations were compared, but the differences were slight); therefore, potential biases associated with population demographic characteristics were most likely minimal. However, our findings should still be generalized only to the population of the 12 NVDRS states included in the study. The fact that 5 states were excluded because of the lack of suicide circumstance information signifies the need for thorough reporting on suicide deaths by medicolegal death investigators to appropriately monitor suicide trends. Lastly, LCA can often lead to subjective interpretation of class patterns; however, the victim demographic and mechanism or weapon characteristics we observed across the classes were similar to the correlations observed in previous research. These characteristics helped validate the existence of the patterns we identified. More research is still needed to help understand why different populations are drawn into different patterns of risk factors for suicide.
Suicide is a complex problem that is typically precipitated by a variety of circumstances. With the use of LCA, this study identified patterns of known risk factors among decedents, which can hopefully provide more direction on how to develop or combine prevention efforts. For ongoing suicide mortality surveillance initiatives, this type of analysis also provides a more accurate and comprehensive picture of the surrounding circumstances experienced by suicide decedents. In this study, the common theme across most classes of decedents was a need for more connected medical, MH, substance abuse, court, and social services; one service alone might not sufficiently prevent this problem. Furthermore, all decedents needed more social support, even those who were currently receiving MH counseling or treatment. Initiatives that recruit family, friends, and community members to help those experiencing high levels of mental and/or life stress might further help prevent suicide.
Correspondence: Joseph Logan, PhD, Division of Violence Prevention, Etiology and Surveillance Branch, National Center for Injury Prevention and Control, Centers for Disease Control and Prevention, 4770 Buford Hwy, MS-F63, Atlanta, GA 30341-3724 (firstname.lastname@example.org).
Submitted for Publication: October 27, 2010; final revision received April 5, 2011; accepted April 17, 2011.
Author Contributions: Dr Logan had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Financial Disclosure: None reported.
Funding/Support: This study was supported by the Centers for Disease Control and Prevention.
Role of the Sponsor: The Centers for Disease Control and Prevention funding was provided for the design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, and approval of the manuscript.
Disclaimer: The findings and conclusions in this article are those of the authors and do not necessarily represent the views of the Centers for Disease Control and Prevention or the Agency for Toxic Substances and Disease Registry.