Military service and risk of subsequent drug use disorders among Swedish men

Abstract

Purpose

Environmental factors contribute substantially to risk for drug use disorders (DUD). The current study applies multiple methods to empirically test whether military service is associated with subsequent DUD, as previous findings are inconsistent.

Methods

Longitudinal Swedish national registry data on a cohort of male conscripts born 1972–1987 (maximum N = 485,900) were used to test the association between military service and subsequent registration for DUD. Cox proportional hazard models were used in preliminary analyses, followed by three methods that enable causal inference: propensity score models, co-relative models, and instrumental variable analysis.

Results

Across all methods, military service was causally associated with lower risk of DUD. Hazard ratios ranged from HR = 0.43 (95% confidence intervals [CI] 0.37; 0.50) in the instrumental variable analysis to 0.77 (0.75; 0.79) in the multivariate propensity score matching analysis. This effect diminished across time. In the model including a propensity score, HRs remained below 1 across the observation period, while confidence intervals included 1 after ~ 11 years in the co-relative analysis and after ~ 21 years in the instrumental variable analysis.

Conclusions

In this cohort of Swedish men, complementary methods indicate that military service conferred substantial but time-limited protection against subsequent DUD. The observed effect could be due to reduced opportunity for substance use during service, social cohesion experienced during and after service, and/or socioeconomic advantages among veterans. Additional research is necessary to clarify these protective mechanisms and determine how other environmental contexts can provide similar benefits.

Introduction

Drug use disorders (DUD) are a persistent worldwide public health concern. Individuals with DUD experience a wide range of negative medical and psychological consequences, including organ damage [1-3], increased risk for communicable diseases [4, 5], and psychiatric disorders such as major depression [6, 7]. Substance misuse also exacts steep interpersonal and economic [8, 9] tolls. Improvements to our understanding of factors that may reduce DUD risk have the capacity to inform prevention efforts.

Liability to DUD is moderately genetically influenced, with heritability estimates from twin studies in the range of 30–60% [10]. Environmental factors also play a critical role in DUD risk: A study of US male twins reported that environmental variance accounted for 27–78% of total liability across illicit substances [11], and a Swedish twin-family study reported estimates of 35–48% for an aggregate measure of illicit drugs [12]. Specific environmental exposures have been previously associated with substance use/misuse: social capital [13], neighborhood social deprivation [14], parental divorce or death [15], one’s own divorce [16, 17], and peer deviance [18-20] are generally associated with increased risk. Putatively protective factors have also been identified, including religiosity [21-23] and marriage [23-25]. Social support is generally conceived of as protective, with positive evidence from some studies [26-28] but not others [22, 29].

The relationship between military service and DUD is unclear. The prevalence of DUD among US veterans is widely perceived to be higher than in the civilian population [30]. However, direct comparisons between groups within the same study, while uncommon, reveal few differences [31]. Appropriate comparisons must consider that most active service members are young men, a demographic group among whom DUD is more prevalent in general [32]. Two reports using data from the National Surveys on Drug Use and Health found sporadic evidence of higher substance use, but not abuse/dependence, among veterans [33, 34]. Interpreting such findings is complicated by the fact that, in the US, military service is initiated by the individual, introducing potential bias.

Military service might instead lead to lower levels of DUD relative to service members’ unselected peers. For example, social cohesion, a central tenet of military training, is positively associated with well-being, individual performance, and readiness [35, 36], and negatively associated with lack of discipline [36] and psychological distress [35]. Oversight within the chain of command, in conjunction with periodic drug testing, might also lead to lower rates of DUD during active military service (termed “incapacitation” by Hjalmarsson, Lindquist [37]) but not protect against later problems. In contrast, combat deployment and/or post-traumatic stress disorder (PTSD) are likely key explanatory factors in studies that report increased substance use/misuse among service members or veterans [38-40], complicating efforts to determine whether military service per se is associated with DUD risk. A potential contributor to inconsistencies in the previous findings is that not all members of the military are combat-exposed.

In the current study, we aimed to clarify the association between military service and DUD using a cohort of Swedish men born 1972–1987. Our primary research question was whether service was associated with increased or decreased risk for DUD and the degree to which this association was likely causal. We were secondarily interested in the persistence of any observed effect across time, i.e., whether selected versus unselected men differed in the prevalence of DUD onset only briefly versus persistently. While serving in the military in Sweden does not involve combat exposure—as it did for approximately 18% of US service members in a recent large-scale cohort study [41]—the absence of combat-related deployments within this cohort allows us to specifically address the role of military service itself in impacting DUD risk. Our research question should therefore be considered within that context.

Observational epidemiologic studies do not traditionally consider causal inference. To address this limitation, we employ three complementary methodological approaches: (i) propensity score matching, in which we match men selected for service to those not selected based on potential confounders, and compare outcomes across groups; (ii) co-relative analysis, an extension of the co-twin control method [42], which accounts for genetic and familial environmental confounding factors that could jointly impact likelihood of military service and DUD; and (iii) instrumental variable (IV) analysis, which accounts for observed and unobserved confounders. In addition to providing insight to the potential for a causal effect, confidence in findings is increased where triangulation across multiple methods yields similar results [43, 44].

Materials and methods

Sample

We collected information on individuals from Swedish population-based registers with national coverage linking each person’s unique personal identification number which, to preserve confidentiality, was replaced with a serial number by Statistics Sweden. We secured ethical approval for this study from the Regional Ethical Review Board in Lund and participant consent was not required (No. 2008/409, 2012/795, and 2016/679). From the Swedish registers, we selected all male individuals born in Sweden between 1972 and 1987 who had not died or emigrated prior to age 18 (N = 779,755). Of these, 718,992 were included in the Swedish Military Conscription Register, which includes results of the conscription examinations undertaken shortly after turning 18. Women were excluded, because they were not subject to the conscription process, which would introduce issues with selection bias.

Outcome variable

The outcome of interest was registration for Drug Use Disorder (DUD), which applies to substances other than alcohol or nicotine. DUD registrations were identified using Swedish medical and mortality registries, criminal and suspicion registries, and the Prescribed Drug Register. ICD and criminal codes for DUD are provided in Supplemental Table 1.

Predictor variables

The primary predictor of interest was military service. Of the 718,992 individuals in the Swedish Military Conscription Register, 45.3% were selected for military service. This information was taken from Swedish tax-registers where compensation for persons participating in basic education (military service or civil duty) longer than 60 days is specifically registered.

Because we applied a propensity score method using the information from the Military Conscription Register, we included variables measured at the conscription examination (see Supplemental Table 1 and below). As shown in Fig. 1, the share of missing values among our selected variables was higher among individuals who did not perform military service than among individuals who did serve. Accordingly, we included 188,784 individuals who did not complete service (representing 48.0% of all males who did not do military service) and 297,116 individuals who did (91.3%).

Fig. 1.

Selection process for propensity score matching

Furthermore, we included as covariates a familial genetic risk score (FGRS) for DUD that was based on information from 1st to 5th-degree relatives [45], school grades at age 16, community deprivation, parental divorce, low parental education, and parental death. For a detailed definition of all variables; see Supplemental Table 1.

Statistical analyses

Cox proportional hazards models

We used a Cox proportional hazards model to investigate risk of DUD as a function of service, from date of conscription until end of follow-up (DUD registration, death, emigration, or 2015–12-31). This is the crude association between military service and DUD used for comparisons.

Propensity score matching

We created a propensity score for service, using a logistic regression model, based on variables accessible to those involved in the conscript selection process: resilience, IQ, height, weight, diastolic blood pressure, systolic blood pressure, muscular strength, and year of birth. We used the predicted probabilities from this model as the propensity for military service. We categorized this propensity into 25 equally sized groups and used a stratified Cox proportional hazards model, with a separate stratum for each of the 25 groups. This means that we are investigating the association between the true value of service and DUD among individuals with similar values of expected military service. Subsequent models included the covariates described above, which were not included in the propensity score as the military did not have access to these variables. Finally, we tested whether the effect of service varied during the follow-up period through the inclusion of an interaction term between a log-transformed measure of time (in months) and service.

Co-relative model

We employed a co-relative design to examine if the crude association between military service and DUD reflected confounding by familial risk factors. This approach is an extension of the co-twin control design [42, 46]. If HRs become less pronounced (i.e., approach 1) among pairs of higher relatedness, this indicates that familial confounding accounts for some proportion of any observed association between exposure and outcome. If no shifts in HRs are observed across groups, or if the HR remains significantly different from 1 even among highly related pairs, this suggests that a causal relationship contributes to the observed association. From the Swedish Multi-Generation Register, we identified all cousin, full- and half-sibling pairs. Using stratified Cox proportional hazards models, with a separate stratum for each relative pair, we investigated the risk of DUD as a function of service. As in the propensity score models, we further tested whether the observed effect shifted across time by including an interaction between log(time) and service. See Supplemental Material for additional details, including on competing models.

Instrumental variable analysis

We used an instrumental variable approach (IV) to control for unmeasured confounding using information on test officiator as an instrument, which was available from 1990 to 1997. At the end of the 2-day conscription examination, the conscripts met with a randomly assigned test officiator who placed the conscripts into service categories based on the conscript’s full range of test scores, interviews, and specific skills. Subjective judgement from the test officiator also contributed to the decision of which recruits to release from military service and which to select. Accordingly, some officiators had with higher placement rates into actual service than others. This randomness provides an instrument that could be used in further analyses, as demonstrated previously [37]. Consistent with the other methodological approaches, we assessed the persistence of the observed effect by evaluating the interaction between the instrument and log(time) in months. Additional details of our approach to these analyses and confirmation that officiator was suitable to use as an IV are reported in the Supplemental Material.

All statistical analyses were performed using SAS 9.4 [47].

Results

Descriptive statistics

Figure 1 describes the process by which Swedish men born 1972–1987 were selected for inclusion in the current analyses. We compared men who were selected for service (N = 297,116) to matched individuals who were not selected for service (N = 188,784). The prevalences of lifetime DUD registration were 3.33% (N = 9903) and 6.92% (N = 13,055) for those selected versus not selected, respectively.

Cox proportional hazard regressions

In a crude regression, military service was negatively associated with subsequent DUD registration (HR = 0.45, 95% confidence intervals [CI] 0.44; 0.46). This and the crude associations between other covariates and DUD are available in Supplemental Table 2. We subsequently tested whether the association between military service and subsequent DUD persisted when controlling for propensity score and found an attenuated but still pronounced association (HR = 0.78 [0.76; 0.81]; Table 1, Model 1). Controlling for a DUD registration prior to conscription, school grades, FGRS_DUD, and other covariates had little impact on the estimated effect of service (HR = 0.77 [0.75; 0.79]; Table 1, Model 2). In that model, prior DUD, FGRS_DUD, low school grades, community deprivation, and parental death were associated with increased risk of DUD. Parental divorce and low parental education were inversely associated with DUD; this contrasts with their direction of effect in univariate associations (see Supplemental Table 2), potentially due to statistical over-control.

Table 1.

Hazard ratios from univariate and multivariate models estimating the association between military service and subsequent registration for drug use disorder in Swedish men born 1972–1987

Predictor	Model 1		Model 2		Model 3
	HR	95% CI	HR	95% CI	HR	95% CI
Military service	0.78	0.76; 0.81	0.77	0.75; 0.79	0.45	0.38; 0.54
Prior DUD			6.58	5.51; 7.74	6.49	5.48; 7.69
FGRSDUD			1.34	1.33; 1.36	1.34	1.32; 1.36
Low school grades			1.77	1.74; 1.80	1.77	1.74; 1.79
Community deprivation			1.03	1.01; 1.04	1.03	1.01; 1.04
Parental divorce			0.88	0.81; 0.95	0.88	0.81; 0.96
Low parental education			0.89	0.87; 0.91	0.89	0.87; 0.91
Parental death			4.85	4.18; 5.61	4.85	4.19; 5.62
Military service × log(time)					1.13	1.08; 1.17

DUD drug use disorder, FGRS_DUD family genetic risk score for drug use disorder, HR hazard ratio, CI confidence intervals

Finally, we evaluated whether the effect of military service on DUD shifted across time by including an interaction term between service and the log of time, measured in months. As shown in Table 1, Model 3, the interaction term was significantly higher than 1 (HR = 1.13 [1.08; 1.17]), indicating that as time elapsed, service became less protective against subsequent DUD. This effect is illustrated in Fig. 2A, which shows that the putative protective effect of service declines rapidly for approximately 2 years after individuals complete their 2-day conscription test; the effect then attenuates more gradually, but confidence intervals remain below 1 (the null hypothesis) across the full observation period.

Fig. 2.

Depiction of interaction between time and military service on risk for subsequent drug use disorder. The x axis shows the time elapsed since the 2-day conscription test. Note that the x axis scale is in years for ease of interpretation, though statistical models included time at the level of the month. Hazard ratios for military service are depicted by the solid line, with 95% confidence intervals represented by dotted lines. A Estimates from the analysis that incorporates a propensity score. In B, based on the co-relative analysis, separate curves are shown for the full cohort (population), cousins, full siblings, half siblings, and monozygotic twins. The curve for monozygotic twins is based values from the predictive model (see text for full description). For ease of visualization, confidence intervals are presented only for monozygotic twins. C Estimates from the instrumental variable analysis

Co-relative analysis

We used co-relative models to account for familial confounding (genetic and shared environmental) factors that may jointly impact likelihood of selection for army service and DUD. We first used observed data to estimate HRs for DUD among relative pairs discordant for service (i.e., where one member of the pair was selected and the other was not). We next fit a predictive model where HRs were estimated as a function of genetic resemblance and compared the fit of this model to that which relied on observed associations within group. The AIC was slightly improved in our predictive model (615229.77 versus 615226.04). Figure 3 depicts HRs for both observed and predicted models across the population and each relative type, and demonstrates that, as increasing relatedness is accounted for, the inverse association between army service and subsequent DUD is attenuated. However, even among monozygotic twin pairs, within which familial confounders are most thoroughly accounted for, the predicted HR remains well below 1 (HR = 0.73 [0.64; 0.82]).

Fig. 3.

Risk of drug use disorder as a function of military service. Figure depicts results from co-relative analyses testing the association between military service and subsequent registration for drug use disorder. Hazard ratios are presented from both observed models and a “predicted” model in which estimates are based on genetic relatedness in each pair group and extrapolated to monozygotic (MZ) twin pairs. Additional detail is available in the Supplemental Material

As above, we further tested the interaction between the log of time and service. The HR of the interaction term was HR = 1.26 (1.23; 1.30); the shift in HR across time is depicted in Fig. 2B. Among monozygotic twins—the most thoroughly controlled group, with 100% genetic resemblance—the effect of service dissipated: By approximately 11 years after individuals completed the 2-day conscription test, the confidence intervals included 1.

Instrumental variable analysis

We assessed whether an instrumental variable (IV), officiator selection rate, was associated with DUD registration in the cohort tested between 1990 and 1997, when data on officiators were available. In step 1, we found that the IV was associated with army service (OR 1.05 [1.04; 1.07]). In step 2, the IV was not associated with DUD registration (HR = 0.998, [0.994; 1.001]). For step 3, we confirmed the association between military service and DUD (HR = 0.43 [0.41; 0.44]). Thus, the assumptions of the IV model were met. We then tested whether the predicted values from the logistic model in step 1 were associated with DUD registration, and found that they were (HR = 0.43 [0.37; 0.50]). The interpretation was consistent when we replaced the logistic model with a linear model more commonly used in IV analyses (β = − 0.025 [− 0.032; − 0.018]). As in our other analyses, we assessed the interaction between the instrument and the log of time. The HR of the interaction term was HR = 2.14 (1.71; 2.68). Approximately 21 years after the 2-day test, the HR no longer significantly differed from 1 (Fig. 2C).

Discussion

In the current study, we assessed whether military service was associated with risk for subsequent registration for DUD in a large cohort of Swedish conscripts. We employed complementary methodological approaches to improve causal inference: propensity score matching, co-relative analysis, and an instrumental variable model. All three methods provide support for a substantial causal protective effect of military service on later drug problems. The reduced risk of DUD onset dissipated over time, with the co-relative models and instrumental variable models projecting protective periods of approximately 11–21 years beyond the initial military assessment. These results suggest that features of military service—importantly, outside the context of combat exposure—could be useful for DUD prevention efforts.

The putative protective effect of military service on DUD onset that we observed suggests that, where increased rates of DUD are observed among veterans, they are unlikely to be due to their military service per se. Rather, such observations could be attributable to adverse experiences during service, including combat exposure [48-50], sexual assault [51-53], or the development of PTSD [54, 55]. Other than combat exposure, these parallel similar associations in civilian samples (e.g., [56-58]), underscoring the universality of risk for substance misuse.

We found that the effect of service dissipated over time, though the duration of the effect differed across methods. Several explanations for these shifts are plausible. First, members of the military are subject to behavioral monitoring that could substantially deter illicit substance use. In the US, this includes routine drug testing since the Vietnam era [59], which has coincided with reductions in substance use [60, 61]. This would account for markedly lower DUD risk in the initial years after assessment. Second, features of military service that convey protection against substance misuse might only be relevant for DUD onset during a particular phase in the life course—here, from the late teens through mid-adulthood. Third, these protective features might be beneficial only if their practice is deliberately maintained in the other contexts. For example, the social support experienced during military service might be difficult to independently cultivate in middle adulthood. Clarifying the temporally limited nature of the protective effect will require identification of the responsible features of military service and additional research on their impact over time.

The context of military service in the current study likely impacts our findings. First, undergoing the 2-day exam for service selection is compulsory in Sweden, while presentation to the military is a self-selective process in the United States. Both countries employ written and physical tests that impact the extent to which individuals selected for service are representative of their civilian counterparts. Second, while the Swedish military does deploy internationally, it has not been involved in active warfare for over 200 years, while members of the US military born during the same period as our Swedish cohort might have been deployed to multiple combat zones. Combat exposure has been previously associated with increased risk for substance misuse in some [40, 62, 63] but not all [38, 64] studies; this could impact the generalizability of the current study, since members of this cohort were most likely not exposed to combat.

Features of military service that could lead to reduced risk for DUD include social cohesion/support; exposure to and integration of rules and disciplined routines; and/or the development of skill sets that could improve long-term social functioning [65]. Skills developed in a military setting can be advantageous in other professional contexts [66, 67]. Importantly, non-military contexts have the capacity to offer similar experiences. For example, civic and religious organizations offer social cohesion; many civilian jobs provide structure for one’s day/life; and online, community, and governmental entities offer training opportunities for job and relationship skills. A notable difference between those options and the Swedish military is that all individuals are initially eligible for conscription and are compensated for service, while other opportunities require self-selection and may require a financial investment that could be prohibitive, especially for young adults. Other potential mechanisms by which military service may lead to lower DUD include higher rates of marriage [68] and more promising income and employment opportunities [69, 70] among veterans. Additional research is necessary to determine which aspect(s) of military service is most effective in reducing DUD risk and determining how other opportunities, perhaps offered in the workplace or community centers, could provide similar protection.

Limitations

Our findings must be considered in the context of several limitations. First, we focused on DUD risk within a cohort of male Swedish conscripts whose military experience may not generalize to military recruits in other countries. This is offset by several key features: (i) unparalleled access to unbiased, longitudinal data; (ii) our ability to jointly employ multiple methods that enable causal inference, which would not be possible in most other studies; and (iii) the absence of combat exposure among the sample, which in other studies may complicate interpretations of the role of military service per se. It is important to note that our findings are unlikely to reflect the relationship between military service and DUD among service members who have been combat-exposed, and who are therefore at increased risk of a range of negative mental health outcomes [41, 71, 72].

Second, our propensity score matching approach relied on measures that were available to the military at the point of conscription. Therefore, individuals for whom information was missing (e.g., IQ score, resilience score) were not eligible for inclusion in our study. Results might differ slightly had these individuals been included, as the prevalence of DUD was higher among individuals with missing values who were not selected for service (11.4%) compared to those who were selected (5.2%). Third, while our predictive model, which included monozygotic twins, is supported by fit statistics, observed discordant monozygotic twins were too rare to be sufficiently informative.

While all three methods employed in the current study improve causal inference, the data are observational. We are further unable to determine whether the effect of military service on DUD registration is attenuated by exposure to peer DUD, as we do not have access to information regarding individuals who served in the same unit. The current findings are also specific to substance use disorders other than AUD and nicotine dependence. Given the complex relationship between military service and AUD, including the potential for positive reinforcement of alcohol use [73, 74], addressing how the role of military service for subsequent AUD was outside the scope of the current analyses but remains an important topic for future research. In addition, we are likely to be capturing more severe cases of DUD through our use of registry data, and some individuals resist seeking medical care due to stigma, which would lead to underestimates of the rate of DUD. However, the observed rates are consistent with those in neighboring Norway [75-77].

Finally, our reference group—males who were eligible for conscription but not selected for service—consists of a mix of those who are employed (outside of the military) and those who are not. This raises the issue of the “healthy worker effect”, which posits that employed individuals have more favorable health-related outcomes than the overall population [78]: Our findings might have been attenuated had we used a more restricted reference group. As we cannot rule out the possibility of the bias this could introduce, the current results should be considered preliminary pending replication under analytic conditions that apply a different approach.

In summary, through the implementation of three independent methodological approaches that facilitate causal inference, we find robust support for a causal protective effect of military service on subsequent registration for DUD in a large male Swedish cohort. This effect is temporally limited, with no difference in risk across groups observed as early as 11 years beyond conscription. The reduced risk could be attributable to a range of features of military service, including reduced access to and/or enhanced negative repercussions of substance use during service, social support and cohesion, and advantageous social and economic outcomes among veterans. To the extent that some protective features of military service could be fostered in other settings, these findings provide insight to the causal role of environmental contexts on risk for DUD.

Data availability

The data used in these analyses are available through Statistics Sweden. Restrictions apply to the availability of the data.