Stressful life events and incident depression among U.S. military personnel

Abstract

Purpose:

Although stressful life events (i.e., stressors) and depression are often assumed to be linked, the relation between stressors and incident depression is rarely studied, particularly in the military. The National Guard is a part-time subset of the U.S. military for whom civilian life stressors may be particularly salient, due to the soldiers’ dual roles and frequent transitions between military and civilian life.

Methods:

We used a dynamic cohort study of National Guard members from 2010–2016 to investigate the relationship between recent stressful experiences (e.g., divorce) and incident depression, with an exploratory analysis of effect modification by income.

Results:

Respondents endorsing at least one of nine past-year stressful events (a time-varying exposure, lagged by one year) had almost twice the adjusted rate of incident depression compared to those with no stressful events (HR = 1.8; 95% CI: 1.4, 2.4). This association may be modified by income: among individuals making under $80,000 per year, those with past-year stressors had twice the rate of depression compared to those with no stressors, but among those making over $80,000, past-year stressors were associated with only 1.2 times the rate of depression.

Conclusion:

Stressful life events outside of deployment are important determinants of incident depression among National Guard servicemembers, but the effect of these events may be buffered by higher income.

INTRODUCTION

Despite the fairly comprehensive epidemiologic literature on depression prevalence in the United States (U.S.), incidence studies are comparatively lacking, among both general and military populations [1–4]. For example, a global systematic review of prevalence and incidence of Major Depressive Disorder (MDD) identified 116 prevalence studies but only four incidence studies [5]. While prevalence is a critical measure for understanding the burden of a condition at a given point in time, it provides us with limited insight into temporality and thus causation.

Further, while it has been demonstrated that stressful life events are associated with depression [6–11], there are few studies that have considered whether stressful life events are associated with incident depression in adulthood [12]. These studies are even more rare among military personnel, for whom the relation between stressors and depression may be particularly important. While the estimated burden of depression within the military varies across different studies, one meta-analysis among different samples estimated overall MDD prevalence to be 12% among currently deployed individuals, 13% among previously deployed individuals, and 5.7% among those who had never been deployed [2]. This burden contributes to a large overall cost related to mental health problems; psychiatric services alone cost the military health system over $4 billion between 2007–2012 [13].

Military reservists may experience additional challenges compared to Active Duty forces. For example, National Guard members (part-time soldiers) are often deployed unexpectedly and with a different unit than their usual training unit, which has been shown in prior work to be associated with post-deployment mental health problems [14]. National Guard members also typically experience differential access to health care and the frequent balance of civilian jobs on top of military engagement [15]. Further, deployment of the National Guard has increased in recent years, both to conflict areas and for humanitarian relief following events like civil unrest and natural disasters [16]. These factors together contribute to National Guard members facing both a potentially greater burden of stressor exposure and larger mental health burden compared to the Active Duty Component [15, 17, 18]. However, mental health indicators among the National Guard have not been studied to the extent that they have been in the Active Duty Component, and particularly not in the longitudinal fashion that is essential for establishing temporality [15, 19–21].

The lack of incidence studies of depression in general is driven in part by the relatively small number of longitudinal cohort studies within psychiatric epidemiology. Most of our knowledge in the field stems from large cross-sectional studies [3, 22, 23]. Although these studies are well-designed and representative of the U.S. population, they cannot, by definition, measure incidence. Augmenting these existing studies with longitudinal research is critical, not only for estimating incidence but also for understanding temporality among exposures, confounders, and outcomes, particularly in the case of psychiatric disorders for which comorbidity is common but temporal ordering often unclear [24].

Within military populations, incident depression studies are even more rare. One study that did measure incidence rates of depression among military personnel came from the in-person validation sub-study of the Ohio Army National Guard Mental Health Initiative (OHARNG-MHI), the cohort used in the current study. That study reported an incidence rate of 4.2 per 100 person-years for MDD [25]. Although a summary of traumatic event experience in adulthood (e.g., assault, witnessing death) was used as an exposure of interest, stressors such as divorce or unemployment were not included. These types of events are not considered traumatic, but they may cause depression. No studies to our knowledge to date have assessed U.S. servicemembers longitudinally across multiple years to determine rates of incident depression and how time-changing stressful events associate with depression. Further, this relation may be modified by current income. This hypothesis is informed by the conservation of resources theory, which suggests that the accumulation and retention of resources such as money or housing act as buffers for stressful events and resulting mental health problems, including depression [26].

Our primary aims for this study were to estimate a) the incidence rates of depression across four years of follow-up and b) the association between reporting stressful events during adult, civilian life and the rates of incident depression, using time-to-event analysis in a cohort of U.S. Army National Guard members. An exploratory aim was to assess potential effect modification of this relationship by income.

METHOD

Data source

We used data from the OHARNG-MHI, a cohort study that began in 2009 and for which the recruitment details have been previously described [27]. The baseline sample, and the Ohio Army National Guard in general, is representative of the U.S. Army National Guard population as a whole in terms of many demographic factors such as military rank, gender, and age [27, 28].

The first and primary cohort of the study (n=2,616 participants at baseline) was followed via telephone interviews approximately once per year. In order to mitigate attrition of sample size over time and related changes in demographics, smaller random samples from newer recruits to the Guard have replenished the original group of respondents each year, beginning in the third year of the study, creating a dynamic cohort study design [29].

There were three inclusion criteria for selection into our analytic sample for this study. First, participants had to be free of lifetime history of depression at the beginning of follow-up time, in order to measure incident depression. Second, participants had to be present for their year 2 (“wave” 2) interview, so that all individuals had depression data for waves 1 and 2, in order to ensure that no participants had depression prior to the start of follow-up at wave 3. Finally, individuals had to be present in at least one follow-up wave among waves 3–6, as the depression outcome started being counted at wave 3 (as described below). Thus, all participants contributed at least one full year of follow-up time at risk of depression.

The depression outcome was counted beginning at wave 3 because the time-varying past-year stressor exposure was first measured beginning at wave 2, and we lagged this exposure by one year, to preserve temporality and reduce the potential for reverse causation in between waves (e.g., depression may lead to job loss through absenteeism, and job loss is considered a stressor). Figure 1 shows a visual representation of the lagged study design. Since the time scale used in this analysis was time on study, “wave 1” in the figure refers to the baseline interview for all groups of participants, regardless of the calendar year during which they entered the study; “wave 2” refers to their second interview, and so on.

Fig 1.

Study design diagram of exposure and outcome assessment for lagged, time-varying stressors as they predict depression. b/w = between. w = wave.

In all, the analytic sample comprised 1,334 participants, including 1,121 from the primary cohort (those who joined the beginning of the study in 2009), 136 from the first supplemental (additional) cohort who enrolled in 2011–2012, and 77 from the second supplemental cohort who enrolled in in 2012–2013.

All participants provided informed consent. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.

Exposure

The exposure of interest was time-varying stressor experience that typically occurs in a context outside of the military, in civilian life (e.g., divorce). These events were ascertained at each follow-up interview, asked with reference to the past year. Appendix Table 1 lists the specific stressors that were ascertained. The baseline (wave 1) surveys only assessed lifetime—not past-year—stressors, thus stressors were measured beginning at wave 2 (see Figure 1 and for a visual representation of exposure and outcome assessment).

During the 2014–2015 interview year (the fifth total interview for the primary cohort, the third interview for the first supplemental cohort, and the fourth interview for the second supplemental cohort), the follow-up survey was cut in length as a result of a reduced budget. Consequently, only two out of the nine past-year stressor questions that were asked during the other follow-up interviews were assessed that year: job loss and home loss. This problem was addressed in two ways. For the two supplemental cohorts, the one-or-more-past-year stressor variable from 2014–2015 was multiply imputed (as described below), with the other cohorts providing the available non-missing values, since the time scale was time on study and thus the values came from different calendar years for each cohort (as can be seen in Figure 2). In this figure, the boxes that include asterisks represent the 2014–2015 interview with shorter surveys, and the dotted lines represent interview years for the supplemental cohorts that were imputed with multiple imputation using the other cohorts’ data for those years, due to shorter overall length of follow-up time given the later enrollment years. Specifically, for the first supplemental cohort (second row of the diagram), the past-year stressor variable for wave 3 was imputed using the two other “interview 3” or wave 3 past-year stressor variables from the other cohorts. Similarly, for the second supplemental cohort, the past-year stressor variable for wave 2 was imputed using the two other “interview 2” or wave 2 past-year stressor variables from the other cohorts. However, the stressor variable for the fifth wave of the primary cohort could not be imputed using other cohorts, as the primary cohort was the only group that contributed four full years of follow-up time in this sample (given the different enrollment years), as can be seen in Figure 2. In other words, there was no non-imputed data on those events for anyone else in that year. Thus, the past-year stressor variable for interview 5 (the final past-year stressor variable counted) was created with only the two available questions.

Fig 2.

Diagram of time-on-study design. * = 2014–2015, shorter interviews, only 2 stressors assessed. Dashed boxes = imputed data.

Confounders and modifier

Potential confounders and modifiers were chosen based on previous literature and hypothesized associations. Time-stable confounders were assessed at baseline and included (a) gender, (b) race/ethnicity (dichotomized into two categories due to small cell size of non-white races: white vs. black or “other” race, including Hispanic ethnicity), and (c) age category at baseline (categorized into 25–34 years and 35+ years vs. 18–24 years, based on survey response options).

Potential time-varying confounders included the following lagged variables (taking the value from the year before the exposure of interest, to preserve temporality): past-year head injury (which was asked only in reference to the respondent’s most recent deployment), generalized anxiety disorder (GAD), and posttraumatic stress disorder (PTSD). Due to a very small number of individuals with past-year head injuries and GAD, time-varying past-year PTSD was the only time-varying confounder included in the final models. PTSD was defined by either reporting a doctor’s diagnosis or meeting the Diagnostic and Statistical Manual for Mental Disorders version IV (DSM-IV) criteria [30], assessed in the surveys with the PTSD Check List-Civilian version [31].

The potential modifier of interest was baseline income level (split into categories of $40,000 per year or less, between $40,001-$80,000, and greater than $80,000, which were the survey response options).

Outcome

The outcome for this study was the first episode of depression among participants. Depression was defined using DSM-IV criteria [30] (identical to DSM-5 criteria), as measured with the 9-item Patient Health Questionnaire (PHQ-9) [32]. Any Depressive Disorder, which includes the DSM-IV categories of MDD and Depressive Disorder Not Otherwise Specified, was selected as the depression construct in this study due to higher sensitivity compared to MDD only (51% vs. 35%), without sacrificing specificity (83%), when validated against the in-person Structured Clinical Interview for DSM-IV Axis I Disorders, which was conducted on a random subsample of 500 members of the original cohort [33].

This depression construct is defined by reporting a period of at least two weeks in the past year with two or more co-occurring symptoms, where one of the symptoms is depressed mood or anhedonia (inability to feel pleasure), with a frequency of “more than half the days” or “nearly every day.” Having thoughts of self-harm or suicide is an exception to the frequency criteria, counting as a symptom when reported at any frequency.

This same definition was used to define history of depression at the start of follow-up for the exclusion criteria, including lifetime symptom ascertainment during baseline interviews.

Statistical analysis

First, we performed multiple imputation on missing data within the analytic sample, including missing data from loss to follow-up, using multivariable regressions with fully conditional specification, otherwise known as imputation by chained equations [34, 35], with five imputations. Using the first imputed dataset (after 20 burn-in imputations), we descriptively assessed the rate of incident depression over time, stratifying by exposures as well as potential modifiers and confounders.

After imputation, we transformed the data from a “wide” format in which each observation is a respondent, to a “long format” or person-year dataset, in which each observation is a year of follow-up [36]. We compared the person-year dataset to the wide dataset to check for errors, and ran all Cox-proportional hazards regressions on both versions of the dataset to check for consistency. The natural logs of the hazard ratios and 95% confidence interval (CI) limits from each Cox-proportional hazards regression were averaged across the five imputed datasets, and then exponentiated to arrive at the estimated hazard ratios and 95% CIs.

Time on study was used as the time scale due to the relatively short follow-up time compared to typical time-to-event analyses and the timing of interviews being approximately one year apart each. Due to the imputation on all missing data, no respondents were censored out of analysis from loss to follow-up; they were only administratively censored at the end of the study if they had not yet had depression. Exact event ties were used, accounting for every possible ordering of events, due to the uncertainty of exactly when depression onset within each year of follow-up (between interviews).

For our exploratory analysis, we assessed potential effect measure modification or interaction in two different ways. First, we ran adjusted models as described above, stratified by income level. Second, we ran the adjusted model with a multiplicative interaction term for time-varying stressors by income level (in addition to the main effects for each variable), to assess the Wald Chi square value for the interaction term.

RESULTS

Descriptive results

Table 1 shows the prevalence of all variables used in analyses. The sample was 88% male and 12% female. At baseline, over a third (38%) of respondents were between the ages of 18 and 24; 31% were between the ages of 25 and 34, and the remaining 32% were 35 or older. The majority of the sample (89%) were white. Thirty-six percent of respondents reported an annual income of $40,000 or less at baseline; 40% reported having an income between $40,001 and $80,000; and 24% reported more than $80,000. On average, 4.5% had past-year PTSD each follow-up year and just over a third (35%) reported one or more stressors during each follow-up year.

Table 1.

Sample characteristics: Prevalence of confounders, modifiers, and exposure of interest (n = 1,334).

	n	%
Male	1,173	88%
Female	161	12%
Age 18–24	503	38%
Age 25–34	410	31%
Age 35+	421	32%
White race	1,192	89%
Black and “other” race/ethnicity, including mixed races and Hispanic ethnicity	142	11%
Annual income less than or equal to $40,000 at baseline	484	36%
Annual income between $40,001 and $80,000 at baseline	529	40%
Annual income greater than $80,000 at baseline	321	24%
Past-year PTSD (average across follow-up waves)	60	4.5%
No past-year PTSD (average across follow-up waves)	1,274	96%
One or more stressors per year (average across follow-up waves)	466	35%
No stressors per year (average across follow-up waves)	868	65%

PTSD = posttraumatic stress disorder.

There were 265 cases of incident depression among 4,868 person-years of follow-up (1,134 individuals), which was an incidence of 20% across four years and an incidence rate of 5.4 cases per 100 person-years.

Crude and adjusted models

Table 2 shows the results from Cox-proportional hazards models estimating the relationships between reporting one or more past-year stressful events with time to incident depression during follow-up. The crude hazard ratio was 2.0 (95% CI: 1.5, 2.5). The adjusted hazard ratio was 1.8; (95% CI: 1.4, 2.4), when controlling for gender, race, age group, and past-year PTSD, indicating that individuals who reported at least one stressor in the past year had almost twice the rate of incident of depression the following year, compared to those who reported no past-year stressful events.

Table 2.

Relationship between one or more past-year stressful events (time-varying) and time to incident depression, crude and adjusted (n = 1,334).

	Hazard ratio	95% CI
1 or more past-year stressful events, crude	2.0	(1.5, 2.5)
1 or more past-year stressful events, adjusteda	1.8	(1.4, 2.4)

CI = confidence interval.

^aAdjusted for gender, race, age group, and time-varying past-year PTSD.

Effect measure modification

Figure 3 shows potential effect modification by past-year income at the baseline interview, graphed on the log scale. Among individuals making less than $40,000 per year, the adjusted hazard ratio for time-varying stressors was 1.9 (95% CI: 1.2, 2.9). Among those making between $40,001 and $80,000, this relationship was essentially the same (adjusted hazard ratio=2.1; 95% CI: 1.4, 3.2), but among those making over $80,000 per year, past-year stressors were associated with only 1.2 times the adjusted rate of incident depression (95% CI: 0.69, 2.2). In the adjusted model with the multiplicative interaction between income and past-year stressors, the Wald Chi Square value for the interaction term was 4.72 (p-value = 0.0944).

Fig 3.

Adjusted relationship between one or more past-year stressful events (time-varying) and time to incident depression, adjusted and stratified by baseline income category, graphed on the log scale. aHR = adjusted hazard ratio. Adjusted for gender, race, age group, and time-varying past-year PTSD.

DISCUSSION

This is the first study to our knowledge to assess the relationship between yearly stressors and rates of incident depression among U.S. military personnel. We found that stressors were associated with higher rates of incident depression, and that this relationship may be stronger among those making less than $80,000 per year.

There are few incidence studies of depression in general or military populations with which to compare our estimates, but we found higher incidence in our current study (20% across four years) compared to an earlier study by Rudenstine and colleagues using the same underlying cohort, which reported 11% with incident depression since the most recent deployment [37]. This difference is most likely due to the longer follow-up time in our study (which allows for more opportunity to develop depression), as well as two additional inclusion criteria in the earlier study, both of which likely contributed to a healthier sample in general: a) that respondents had no history of PTSD in addition to no history of depression (the authors also examined new-onset PTSD in the same set of analyses), and b) that respondents had previously been deployed. Individuals with physical and mental health problems often are not deployed to active conflict, resulting in potentially artificially lower rates of mental health problems when assessing only deployed personnel [38].

The study by Fink and colleagues from the sample of respondents from the OHARNG-MHI platform study who participated in clinician interviews also reported a slightly lower estimate of new-onset depression (4.2 new cases of MDD per 100 person-years) [25], compared the rate we found, 5.4 per 100 person-years. This difference is likely due to a) use of the in-person sample in the former study, which is smaller and may also represent a different type of sample compared to the overall cohort, though there is likely some overlap in participation between the two analytic samples; and b) the difference in assessment for depression (Fink et al. reported MDD only and used the in-person SCID, whereas we used the PHQ-9 and included Depression Not Otherwise Specified in addition to MDD).

For the relation between stressors and depression, a study by Galea and colleagues assessed past-year stressors including many of the stressors assessed in the present study (though only measured once, at baseline), among a sample of New York City residents in 2002 [12]. That study observed a higher incidence of MDD among those who reported one or more past-year stressors at baseline compared to no past-year stressors at baseline (17% vs. 14%), but no detectable relationship between this variable and incident depression in the final multivariable model.

Our preliminary finding of potential effect modification by income (although we did not observe a statistically significant interaction term between income level and stressors at the p < 0.05 level, likely due to our sample size) is consistent with the conservation of resources theory. This theory suggests that the accumulation and retention of resources such as money or housing act as buffers for stressful events and resulting mental health problems, including depression [26]. Therefore, soldiers with fewer resources (i.e., lower income) may have less of a buffer against ongoing stressors. The association between the stressors themselves and incident depression may also be explained using this theory. When resources are threatened or lost—through events such as divorce or job loss—stress occurs, which can lead to depression through the loss of these formerly protective buffers. The pathway from stressful life events to depression can also be explained biologically. Incidents of perceived stress can trigger repeated physiological responses via the nervous system, which—while critical for normal function and survival—may become overactivated with stress. The cost of this continued overactivation of regulatory systems over time, or “allostatic load”, can result in excess levels of glucocorticoids and other hormones [39–43]. In addition to potentially triggering physical health problems, McEwen et al. argue that this “wear and tear” of either overactivity or inefficient use of allostatic systems can lead to depression through changes in the brain [39–41, 44, 45].

One limitation of our study is the crude income measure used for the exploratory analysis, based on a survey question that only assessed income with reference to three categories, rather than exact income. Accordingly, the income strata used in this study may be arbitrary and contribute to imprecise estimates of the relationship between stressors and depression within the highest income category in particular, which has the smallest number of individuals. This may also be the reason we did not observe a statistically significant interaction between stressors and income. Future studies on this topic should measure exact income if possible, or replicate this finding in larger samples. Additional socioeconomic-related data such as number of dependents, cost of living, and overall wealth including savings, may also be warranted to further understand the potential effect modification.

A second limitation in our study is potential misclassification of our measures, as with any survey-based epidemiologic study. Misclassification of our exposures and covariates are likely to be non-differential by depression status, since this was a prospective study and depression had not yet occurred at the time of exposure and confounder assessment. A particular pattern of misclassification likely occurred for the past-year stressor value in study year (wave) 5 of the primary cohort, due to an overall shorter survey in the 2014–2015 interview year: only two types of past-year stressful events were assessed that year, compared to nine types of stressful events in other follow-up years. This change likely resulted in an underestimate of whether individuals had one or more stressful experiences in the fourth (and last) measure of time-varying past-year stressors for the primary cohort. However, this underestimate of stressful events in wave 5 is not expected to be related to depression status, as it was a systematic issue for all participants of the last year of stressor ascertainment in this study, regardless of depression status. Accordingly, this misclassification may have caused an underestimate of the effect for individuals who had incident depression in the very last year of the study (a small proportion of individuals).

Further, we do not expect misclassification of depression, our outcome, to be differential by our exposure, stressors. Accordingly, we expect that our estimates of the associations observed are likely under-estimated due to misclassification of depression. On the other hand, since all of the data in this study was obtained from one source, it is possible that we have dependent error (or “same-source bias”) in this study (respondents who were more or less likely to report stressors may also have been more or less likely, respectively, to report depression symptoms, which could bias our results away from the null (i.e., in the other direction as that expected from non-differential, independent misclassification) [46]. However, even if events and symptoms are misclassified with correlated error, perceived mental health symptoms and perceived life events are nonetheless of clinical interest both in the literature and to the military, as they affect functional health and predict retention and performance in the military [47].

A third overall limitation to this study is the potential for confounding from unmeasured constructs. There are potential confounders that were too rare in this sample to include in final models, but which may still contribute to unmeasured confounding. One such example is traumatic brain injury, which has been linked to depression [48–50], and could conceivably cause stressors such as job loss as a result of symptoms. However, brain injury affected very few individuals in our sample, at least in a measurable way (it is notoriously difficult to measure traumatic brain injury, and our survey only asked about symptoms with reference to an injury during the respondents’ most recent deployment). Similarly, we were unable to include incident PTSD in the past year as a confounder, due to the small number of respondents in our sample with new-onset PTSD each year. Thus, prevalent PTSD in the year prior to each depression assessment was included as a proxy for incident PTSD. Although prevalent PTSD was lagged by one year (and still rare), the temporality of which disorder might influence the other is less clear.

Nonetheless, this study was the first to our knowledge to describe how civilian life stressors are associated with time to depression onset among a military population. Future research should aim to replicate these findings, and if sample size allows, individual types of stressful events (e.g., divorce in the past year) instead of a summary variable should be assessed to understand whether particular types of events have stronger associations with depression onset.

Overall, our findings indicate the importance of considering stressful events that occur outside of military contexts when studying the mental health of personnel, despite the relative lack of such studies in the military mental health literature. These results may help inform potential interventions for depression in order to reduce the depression burden within the U.S. Army National Guard, a unique, under-studied, and potentially vulnerable population.