Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2022 Oct 1.
Published in final edited form as: Psychol Trauma. 2021 Apr 22;13(7):725–729. doi: 10.1037/tra0001033

A Comorbid Mental Disorder Paradox: Using Causal Diagrams to Understand Associations Between Posttraumatic Stress Disorder and Suicide

Tammy Jiang 1, Meghan L Smith 1, Amy E Street 2,3, Vijaya L Seegulam 1,4, Laura Sampson 5, Eleanor J Murray 1, Matthew P Fox 1,6, Jaimie L Gradus 1,3
PMCID: PMC8564019  NIHMSID: NIHMS1709788  PMID: 34723565

Abstract

Objective:

Although some studies document that posttraumatic stress disorder (PTSD) increases suicide risk, other studies have produced the paradoxical finding that PTSD decreases suicide risk. We sought to understand methodologic biases that may explain these paradoxical findings through the use of directed acyclic graphs (DAGs).

Method:

DAGs are causal diagrams that visually encode a researcher’s assumptions about data generating mechanisms and assumed causal relations among variables. DAGs can connect theories to data and guide statistical choices made in study design and analysis. In this article, we describe DAGs and explain how they can be used to identify biases that may arise from inappropriate analytic decisions and data limitations.

Results:

We define a particular form of bias, collider bias, that is a likely explanation for why studies have found a supposedly protective association of PTSD with suicide. This protective association is interpreted by some researchers as evidence that PTSD reduces the risk of suicide. Collider bias may occur through inappropriate adjustment for a psychiatric comorbidity, such as adjustment for variables that are affected by PTSD and share common causes with suicide.

Conclusions:

We recommend that researchers collect longitudinal measurements of psychiatric comorbidities, which would help establish the temporal ordering of variables and avoid the biases discussed in this article. Furthermore, researchers could use DAGs to explore how results may be impacted by design and analytic decisions prior to execution.

Keywords: causal diagrams, directed acyclic graphs, posttraumatic stress disorder, psychiatric comorbidity, suicide

Several epidemiologic studies have documented an increased risk of suicide among persons with posttraumatic stress disorder (PTSD; Bullman & Kang, 1994; Gradus et al., 2010, 2015; Ilgen et al., 2010). For example, in a study of veterans receiving health care services from the Department of Veterans Affairs, male veterans with PTSD had 1.8 (95% CI [1.7, 2.0]) times the rate of suicide compared with male veterans without PTSD, and female veterans with PTSD had 3.5 (95% CI [2.5, 4.9]) times the rate of suicide compared with female veterans without PTSD, adjusting for age (Ilgen et al., 2010). A nested case–control study using longitudinal data in the general population in Denmark found that the rate of suicide death among persons with PTSD was 5.3 (95% CI [3.4, 8.1]) times that of persons without PTSD, adjusting for depression diagnoses that occurred before PTSD and demographic confounders (Gradus et al., 2010).

In contrast, other studies have reported a supposedly protective association of PTSD with suicide such that PTSD appears to reduce the risk of suicide (Britton et al., 2017; Conner et al., 2014; Desai et al., 2008; Shen et al., 2016; Zivin et al., 2007). In a study of individuals receiving treatment for depression in the Veterans Affairs health system, veterans who received a PTSD diagnosis had a lower rate of suicide than veterans without PTSD (hazard ratio, .77; 95% CI [.68, .87]; Zivin et al., 2007). A cross-sectional study of Veterans Health Administration users found that after adjusting for multiple other mental disorder diagnoses, PTSD was associated with a decreased odds of suicide (odds ratio, .77; 95% CI [.69, .86]; Conner et al., 2014). In further analyses to investigate which mental disorder had the largest impact on the association between PTSD and suicide, the authors found that regression adjustment for major depressive disorder led to the largest attenuation of the effect estimate (Conner et al., 2014). A study of male veterans discharged from inpatient units found that veterans with PTSD had a lower risk of suicide compared with veterans without PTSD, adjusting for demographic characteristics and psychiatric comorbidity (hazard ratio, .66; 95% CI [.56, .78]; Britton et al., 2017).

These counterintuitive findings that PTSD is associated with lower risk of suicide present a significant challenge to our understanding of the association between PTSD and suicide. PTSD can be a highly impairing condition with long term negative consequences (Kessler, 2000), and a protective effect of PTSD on suicide is inconsistent with this knowledge. Studies that have documented this protective association have hypothesized various underlying mechanisms including more intensive care and monitoring among persons with PTSD and increased resilience and stress tolerance among persons with PTSD. An alternative explanation, however, may relate to the methodologic challenges that arise due to the cooccurrence of various mental disorders (e.g., PTSD and depression). Incorrectly handling this cooccurrence during study design and analysis can lead to biased results.

The purpose of this study is to demonstrate how paradoxical findings may arise as a result of inappropriate incorporation of comorbid mental conditions in study design and analysis when examining the association between PTSD and suicide, with a focus on comorbid depression as an illustrative example. This paper builds on previous work that discusses possible etiologic and methodologic explanations for disparate findings in the PTSD and suicide literature (Gradus, 2017) by using causal diagrams, specifically, directed acyclic graphs (DAGs). Causal DAGs can help researchers (a) explicitly state their assumptions about the temporal ordering of disorders (and other variables) within studies, (b) determine whether or not it is appropriate to statistically adjust for comorbid conditions, and (c) identify types of biases that may arise from specific analytic choices. In this article, we introduce DAGs and their rules and describe how they can be used to identify variables that should and should not be considered for adjustment. We also explain the bias that may arise from restriction and statistical adjustment (e.g., stratification, controlling for a variable in regression analyses, matching). This article focuses on the challenges in studying the association between PTSD and suicide in the context of psychiatric comorbidity, but these challenges are applicable to the study of other mental disorders and suicide-related outcomes (e.g., suicide attempts, suicidal ideation) as well.

What Are DAGs?

Causal DAGs are graphical tools used to depict investigators’ understanding or knowledge of data generating mechanisms (i.e., the causal structure that created our data; Hernán et al., 2001; Pearl, 1995). To construct a DAG, expert knowledge is used to organize beliefs about etiology and represent them in a diagram. DAGs can help make theory underlying a statistical model known, so that assumptions and their implications can be evaluated. The DAGs that we use throughout this article can be found in Figure 1a–1e. A DAG is composed of variables (i.e., nodes or vertices) and arrows (i.e., edges). DAGs are “directed,” meaning that variables are linked by arrows (Greenland et al., 1999). The arrows in a causal DAG can be interpreted as assumed causal effects (either protective or harmful effects). For example, A → B implies that A may cause B. DAGs are “acyclic,” meaning one should not be able to trace a path from one variable to other variables following the direction of the arrows and eventually end up back at the original variable. This is an important rule because a variable cannot cause itself, either directly or through other variables (i.e., the future cannot affect the past). To properly depict feedback loops or a cycle between two variables in a DAG, one can draw the feedback by time points (e.g., times 0 and 1) such as PTSD0 → Social Support0 → PTSD1 → Social Support1. If two or more variables in a DAG share a cause, that cause is referred to as a “common cause.” For A ← C → B, A and B share a common cause C. To accurately represent the causal structure that created a given set of data, all common causes of any pair of variables shown in the graph must be included on the graph. This also holds for unknown/unmeasured variables, which may be labeled “U” (e.g., an unknown/unmeasured common cause). One way to think about DAGs is as nonparametric structural equation models (SEMs; Elwert, 2013; Rohrer, 2018). However, there are some key differences between DAGs and SEMs: (a) SEMs can encode assumptions regarding the functional form (i.e., linear, polynomial) of the relations among the variables, whereas an arrow in a DAG only indicates that one variable is assumed to cause another, without indicating the functional form of the relation; and (b) DAGs allow only single-headed arrows and do not permit double-headed arrows, which are a feature of SEMs (Rohrer, 2018).

Figure 1.

Figure 1

Open in a new tab

Directed Acyclic Graphs Representing the Effect of PTSD on Suicide

Understanding DAGs Using Applied Examples of PTSD, Comorbidity, and Suicide

When evaluating the association between PTSD and suicide, an important consideration is the role of PTSD comorbidity, because it is well known that PTSD can be highly comorbid with other psychiatric disorders such as depression (Brady et al., 2000). Such comorbidities are of crucial importance when evaluating the associations between PTSD and suicide because there may be different underlying causal scenarios; for example, comorbidities may act as confounders or mediators of the effect of PTSD on suicide. Figure 1a displays a simple DAG depicting confounding. Confounding is a type of bias that arises when there is an uncontrolled third variable that causes the exposure (i.e., independent variable) and the outcome (i.e., dependent variable), and this variable is not on the causal pathway between the exposure and outcome (i.e., does not occur temporally between the exposure and outcome). Stated another way, confounding of an exposure and outcome relation occurs when they have a common cause. In Figure 1a, PTSD and suicide share a common cause (i.e., a confounder) that could be an unknown or unmeasured environmental, social, or genetic factor. The confounder occurs before PTSD and suicide (i.e., is not on the causal pathway). For Figure 1a, if a researcher were to estimate the unadjusted (i.e., crude) association between PTSD and suicide, the apparent effect of PTSD would likely be biased because the effect of the confounder on suicide would be mixed with the actual effect of PTSD on suicide. The bias introduced by a confounder may lead to overestimation or underestimation of the effect of PTSD on suicide. In designing and analyzing a study, typical approaches the researcher could take to control confounding include restriction, stratification, regression adjustment, or matching.

Figure 1b displays a DAG depicting mediation, another scenario under which comorbidity may play a role in the PTSD and suicide association. A mediator is an intermediate variable that is on the causal pathway between exposure (e.g., PTSD) and outcome (e.g., suicide). In other words, a mediator occurs temporally between PTSD and suicide. PTSD affects the mediator, which in turn affects suicide. For a variable to be a mediator, it must have occurred after the exposure and before the outcome. Figure 1c displays a DAG specifically showing depression as a mediator of the PTSD and suicide association. For Figure 1c, the unadjusted association between PTSD and suicide (ignoring the mediator, depression) would be a valid estimate of the total effect of PTSD on suicide. However, if depression were adjusted (i.e., controlled) for in the absence of a causal mediation analysis, the estimate of the effect of PTSD on suicide would be falsely attenuated because the pathway by which PTSD affects suicide (i.e., through affecting depression) would become blocked after adjustment for depression. If a variable is an intermediate, it should not be treated as a confounder and should not be adjusted for through restriction, stratification, regression adjustment, or matching. Although it goes beyond the scope of the present article, a more appropriate analytic approach to assessing mediation would be a causal mediation analysis (Valeri & Vanderweele, 2013; VanderWeele, 2015).

Collider Bias as an Explanation for the Paradoxical Protective Association Between PTSD and Suicide

Building on our previous example of a mediator on the causal pathway between PTSD and suicide, we will now turn to an example of a methodological issue called “collider bias” which can create a spurious protective association through the inappropriate adjustment of a third variable. Figure 1d displays a DAG that we will use to explain this form of bias. In this DAG, there is a variable called a “collider.” A collider (also called a “common effect”) is a shared effect of more than one cause. In this DAG, the collider is a shared effect of two other variables, PTSD and U. The variable U (an unknown/unmeasured risk factor) is a cause of both the collider and suicide. Colliders are specific to the causal effect and causal pathways of interest. For example, in the DAG in Figure 1d, suicide is a common effect of PTSD and U but should not be considered a collider on the causal pathway of interest from PTSD to suicide. For Figure 1d, the researcher will obtain the correct result if they estimate the unadjusted total effect of PTSD on suicide, ignoring the collider. However, if a researcher controls for the collider via any of the previously stated techniques (including restriction), the researcher may obtain a biased estimate of the effect of PTSD on suicide. For a more detailed introduction to collider bias, we refer readers to other resources (Cole et al., 2010; Greenland, 2003; Hernán et al., 2004; Schisterman et al., 2009).

Suppose you are a researcher interested in determining whether PTSD has an association with suicide above and beyond the effect of depression, and you know that lack of social support also plays an important role. You hypothesize that PTSD is associated with an increased risk of suicide. You draw the DAG in Figure 1e to depict what you believe to be the nature of these associations. The DAG displays that PTSD and lack of social support can affect depression, and that depression, PTSD, and lack of social support affect suicide risk. In this scenario, depression is both a mediator and collider of the relation between PTSD and suicide because depression is on the causal pathway between PTSD and suicide and depression is also a common effect of PTSD and lack of social support. You proceed with analyses, despite not having data on social support, to examine the association between PTSD and suicide, restricting to persons with depression. To your surprise, you observe a protective association between PTSD and suicide among persons with depression. What clues in the DAG could explain this paradoxical finding? According to the DAG, if someone is depressed but did not have PTSD, it would increase the chances that their depression is attributable to lacking social support (as depression must occur as a result of some cause depicted in the DAG). When assessing the effect of PTSD on suicide exclusively among persons with depression, we are comparing persons who developed depression because of PTSD (index group) with persons who got depression because of lack of social support (reference group). If the association between lack of social support and suicide is stronger than the association between PTSD and suicide, then there would be a higher risk of suicide in the reference group (persons with only depression) than the index group (persons with both PTSD and depression). The use of a reference group with a high risk of suicide could create a spurious protective association between PTSD and suicide among persons with depression. Thus, depression should not be adjusted for. To help prevent such spurious findings, researchers can draw DAGs prior to performing a statistical analysis to evaluate potential analytic approaches. By drawing a DAG, researchers can identify variables that are affected by the exposure (PTSD) and that share common causes with the outcome (suicide) that should not be adjusted for because adjusting for them will create collider bias.

Researchers often adjust for comorbid mental disorders (e.g., depression) when examining the association between PTSD and suicide, but this may be inappropriate depending on the underlying causal scenario. A crucial part of the solution to the issues presented above would be to collect longitudinal data and to take into consideration the timing of comorbid disorders relative to PTSD and suicide in the data. In addition, drawing DAGs prior to study design and analysis can help researchers evaluate the impact of adjustment decisions prior to execution. It is appropriate to adjust for depression that occurred before PTSD and suicide to control for confounding. It is not appropriate to control for depression that occurs after PTSD using traditional methods because depression may be on the causal pathway between PTSD and suicide. Depression may be a mediator of the effect of PTSD on suicide and, if this is the case, controlling for depression would remove part of the causal effect of PTSD on suicide (e.g., effect estimates will be biased toward a null effect). Furthermore, as previously explained, depression could additionally be a collider, and thus controlling for it could induce a spurious (negative) effect of PTSD on suicide. If researchers have repeated measurements and there is time-varying feedback between variables, then g-methods such as marginal structural models should be used in place of traditional adjustment or restriction (Cole & Hernán, 2008; Hernán et al., 2000, 2001; Hernán & Robins, 2020; Robins et al., 2000). If researchers are unable to establish the temporal ordering of variables, such as in cross-sectional study designs, and subsequently a paradoxical association is found, then such findings should be met with caution because they may be vulnerable to bias from adjusting for a mediator and/or collider. DAGs are a useful tool for illustrating the ordering of variables in the data and their associations, identifying variables that should and should not be adjusted for, and predicting the kinds of biases that may arise from inappropriate adjustment for variables.

Conclusion

When depression is affected by PTSD and shares common causes with suicide (a scenario that is not difficult to imagine), adjustment for depression in statistical analyses or restriction of the sample to only depressed persons can create the paradoxical finding that PTSD reduces the risk of suicide. This can hinder our ability to gain an accurate understanding of the effect of PTSD on suicide risk, particularly in the context of pervasive psychiatric comorbidity. An accurate understanding of these effects is crucial for prevention and treatment strategies aimed at reducing the burden of suicide. Methodologic errors that lead to a belief in a protective association between PTSD and suicide could have negative clinical and research consequences. In clinical practice, it may create the impression that PTSD does not always need evidence-based treatment or that screening for suicide among persons with PTSD is unnecessary.

In this article we have described a methodologic issue, collider bias, that may provide an explanation for some studies’ paradoxical finding that PTSD is supposedly protective against suicide. We used causal diagrams to illustrate how a protective association between PTSD and suicide can be observed under particular circumstances. There are other well-known examples of collider bias reversing the direction of an effect such that harmful exposures appear protective. For example, the “birth weight paradox” refers to the finding that low-birth-weight infants born to mothers who smoke have lower infant mortality than low-birth-weight infants born to nonsmokers (Hernández-Díaz et al., 2006). This inverse association is attributable to bias from stratification on a variable (birth weight) that is affected by the exposure of interest (maternal smoking) and that shares common causes (e.g., birth defects) with the outcome (infant mortality; Hernández-Díaz et al., 2006). Similarly, the “obesity paradox” refers to the finding that obesity reduces mortality among people with cardiovascular disease, cancer, diabetes, respiratory disease, renal disease, and other conditions (Banack & Kaufman, 2013, 2014; Banack & Stokes, 2017). This protective association can arise as a result of bias from adjusting for a variable (e.g., cardiovascular disease) that is affected by the exposure of interest (obesity) and that shares common causes (e.g., genetic, physiologic, and behavioral factors) with the outcome (mortality). In this article, we propose a “comorbid mental disorder paradox” as another example of the bias that can arise from adjusting for a collider (e.g., depression) that is affected by the exposure of interest (PTSD) and that shares common causes (e.g., social support) with the outcome (suicide).

Collider bias is likely to arise in the study of other mental disorders and suicide-related outcomes. Although our article used depression as an example of a collider for which adjustment could induce bias, comorbid mental disorders (e.g., substance use, anxiety, etc.) could also be colliders because they are affected by PTSD and share common causes (e.g., social isolation, poor physical health) with suicide. Furthermore, we focused on suicide death as the outcome, but the methodologic issues described in this article may also be relevant to the study of other suicide-related outcomes, including suicidal ideation and nonfatal suicide attempts, given the overlap in the exposures of interest and confounding variables. Lack of data on the timing of mental disorders and inappropriate adjustment for mediators and colliders could lead to biased effect estimates in the study of any mental disorder with suicide-related outcomes.

In summary, we have shown how DAGs can be used as a tool to demonstrate different relations between variables both overall and specific to the association between PTSD and suicide, and how crucial longitudinal data and consideration of causal structure are to obtaining accurate estimates of effects. DAGs allow us to encode our assumptions about data-generating mechanisms and the causal webs underlying our studies, which enables the identification of potential common causes (i.e., confounders) that we should adjust for, mediators that we should not adjust for (without causal mediation analysis), and colliders that, if adjusted for, can lead to spurious protective associations.

Clinical Impact Statement.

To obtain accurate estimates of the effect of PTSD on suicide, it is crucial to use longitudinal data and consider the causal structure of the data. Lack of data on the timing of mental disorders and inappropriate adjustment for mediators and colliders may lead to biased effect estimates. Methodologic errors that lead to a belief in a protective association between PTSD and suicide may have negative clinical and research consequences.

Acknowledgments

This work was supported by National Institute of Mental Health Grant R01MH109507 (PI: Jaimie L. Gradus) and National Institute of Mental Health Grant 1R01MH110453-01A1 (PI: Jaimie L. Gradus).

References

  1. Banack HR, & Kaufman JS (2013). The “obesity paradox” explained. Epidemiology, 24(3), 461–462. 10.1097/EDE.0b013e31828c776c [DOI] [PubMed] [Google Scholar]
  2. Banack HR, & Kaufman JS (2014). The obesity paradox: Understanding the effect of obesity on mortality among individuals with cardiovascular disease. Preventive Medicine, 62, 96–102. 10.1016/j.ypmed.2014.02.003 [DOI] [PubMed] [Google Scholar]
  3. Banack HR, & Stokes A. (2017). The ‘obesity paradox’ may not be a paradox at all. International Journal of Obesity, 41(8), 1162–1163. 10.1038/ijo.2017.99 [DOI] [PubMed] [Google Scholar]
  4. Brady KT, Killeen TK, Brewerton T, & Lucerini S. (2000). Comorbidity of psychiatric disorders and posttraumatic stress disorder. The Journal of Clinical Psychiatry, 61(Suppl. 7), 22–32. [PubMed] [Google Scholar]
  5. Britton PC, Bohnert KM, Ilgen MA, Kane C, Stephens B, & Pigeon WR (2017). Suicide mortality among male veterans discharged from Veterans Health Administration acute psychiatric units from 2005 to 2010. Social Psychiatry and Psychiatric Epidemiology, 52(9), 1081–1087. 10.1007/s00127-017-1377-x [DOI] [PubMed] [Google Scholar]
  6. Bullman TA, & Kang HK (1994). Posttraumatic stress disorder and the risk of traumatic deaths among Vietnam veterans. Journal of Nervous and Mental Disease, 182(11), 604–610. 10.1097/00005053-199411000-00002 [DOI] [PubMed] [Google Scholar]
  7. Cole SR, & Hernán MA (2008). Constructing inverse probability weights for marginal structural models. American Journal of Epidemiology, 168(6), 656–664. 10.1093/aje/kwn164 [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Cole SR, Platt RW, Schisterman EF, Chu H, Westreich D, Richardson D, & Poole C. (2010). Illustrating bias due to conditioning on a collider. International Journal of Epidemiology, 39(2), 417–420. 10.1093/ije/dyp334 [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Conner KR, Bossarte RM, He H, Arora J, Lu N, Tu XM, & Katz IR (2014). Posttraumatic stress disorder and suicide in 5.9 million individuals receiving care in the veterans health administration health system. Journal of Affective Disorders, 166, 1–5. 10.1016/j.jad.2014.04.067 [DOI] [PubMed] [Google Scholar]
  10. Desai RA, Dausey D, & Rosenheck RA (2008). Suicide among discharged psychiatric inpatients in the Department of Veterans Affairs. Military Medicine, 173(8), 721–728. 10.7205/MILMED.173.8.721 [DOI] [PubMed] [Google Scholar]
  11. Elwert F. (2013). Graphical causal models. In Morgan SL(Ed.), Handbook of causal analysis for social research (pp. 245–273). Springer. 10.1007/978-94-007-6094-3_13 [DOI] [Google Scholar]
  12. Gradus JL (2017). PTSD and death from suicide. PTSD Research Quarterly, 28(4), 1–8. [Google Scholar]
  13. Gradus JL, Antonsen S, Svensson E, Lash TL, Resick PA, & Hansen JG (2015). Trauma, comorbidity, and mortality following diagnoses of severe stress and adjustment disorders: A nationwide cohort study. American Journal of Epidemiology, 182(5), 451–458. 10.1093/aje/kwv066 [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Gradus JL, Qin P, Lincoln AK, Miller M, Lawler E, Sorensen HT, & Lash TL (2010). Posttraumatic stress disorder and completed suicide. American Journal of Epidemiology, 171(6), 721–727. 10.1093/aje/kwp456 [DOI] [PubMed] [Google Scholar]
  15. Greenland S. (2003). Quantifying biases in causal models: Classical confounding vs collider-stratification bias. Epidemiology, 14(3), 300–306. 10.1097/01.EDE.0000042804.12056.6C [DOI] [PubMed] [Google Scholar]
  16. Greenland S, Pearl J, & Robins JM (1999). Causal diagrams for epidemiologic research. Epidemiology, 10(1), 37–48. 10.1097/00001648-199901000-00008 [DOI] [PubMed] [Google Scholar]
  17. Hernán MA, & Robins JM (2020). Causal inference: What if. Chapman & Hall/CRC. [Google Scholar]
  18. Hernán MA, Brumback B, & Robins JM (2001). Marginal structural models to estimate the joint causal effect of nonrandomized treatments Journal of the American Statistical Association, 96(454), 440–448. 10.1198/016214501753168154 [DOI] [Google Scholar]
  19. Hernán MA, Brumback B, & Robins JM (2000). Marginal structural models to estimate the causal effect of zidovudine on the survival of HIV-positive men. Epidemiology, 11(5), 561–570. 10.1097/00001648-200009000-00012 [DOI] [PubMed] [Google Scholar]
  20. Hernández-Díaz S, Schisterman EF, & Hernán MA (2006). The birth weight “paradox” uncovered? American Journal of Epidemiology, 164(11), 1115–1120. 10.1093/aje/kwj275 [DOI] [PubMed] [Google Scholar]
  21. Hernán MA, Hernández-Díaz S, Robins JM, & Hernán T. (2004). A structural approach to selection bias. Epidemiology, 15(5), 615–625. 10.1097/01.ede.0000135174.63482.43 [DOI] [PubMed] [Google Scholar]
  22. Ilgen MA, Bohnert A, Ignacio I, McCarthy JF, Valenstein MM, Kim HM, & Blow FC (2010). Psychiatric diagnoses and risk of suicide in veterans. Archives of General Psychiatry, 67(11), 1152–1158. 10.1001/archgenpsychiatry.2010.129 [DOI] [PubMed] [Google Scholar]
  23. Kessler RC (2000). Posttraumatic stress disorder: The burden to the individual and to society. The Journal of Clinical Psychiatry, 61(Suppl. 5), 4–12. [PubMed] [Google Scholar]
  24. Pearl J. (1995). Causal diagrams for empirical research. Biometrika, 82(4), 669–688. 10.1093/biomet/82.4.669 [DOI] [Google Scholar]
  25. Robins JM, Hernán MA, & Brumback B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550–560. 10.1097/00001648-200009000-00011 [DOI] [PubMed] [Google Scholar]
  26. Rohrer JM (2018). Thinking clearly about correlations and causation: Graphical causal models for observational data. Advances in Methods and Practices in Psychological Science, 1(1), 27–42. 10.1177/2515245917745629 [DOI] [Google Scholar]
  27. Schisterman EF, Cole SR, & Platt RW (2009). Overadjustment bias and unnecessary adjustment in epidemiologic studies. Epidemiology, 20(4), 488–495. 10.1097/EDE.0b013e3181a819a1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Shen Y-C, Cunha JM, & Williams TV (2016). Time-varying associations of suicide with deployments, mental health conditions, and stressful life events among current and former U.S. military personnel: A retrospective multivariate analysis. The Lancet. Psychiatry, 3(11), 1039–1048. 10.1016/S2215-0366(16)30304-2 [DOI] [PubMed] [Google Scholar]
  29. Valeri L, & Vanderweele TJ (2013). Mediation analysis allowing for exposure-mediator interactions and causal interpretation: Theoretical assumptions and implementation with SAS and SPSS macros. Psychological Methods, 18(2), 137–150. 10.1037/a0031034 [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. VanderWeele T. (2015). Explanation in causal inference: Methods for mediation and interaction. Oxford University Press. http://ebookcentralproquest.com/lib/bu/detail.action?docID=1920742 [Google Scholar]
  31. Zivin K, Kim HM, McCarthy JF, Austin KL, Hoggatt KJ, Walters H, & Valenstein M. (2007). Suicide mortality among individuals receiving treatment for depression in the Veterans Affairs Health System: Associations with patient and treatment setting characteristics. American Journal of Public Health, 97(12), 2193–2198. 10.2105/AJPH.2007.115477 [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES