Deployment and Alcohol Use in a Military Cohort: Use of Combined Methods to Account for Exposure-Related Covariates and Heterogeneous Response to Exposure

Abstract

Studies have shown that combat-area deployment is associated with increases in alcohol use; however, studying the influence of deployment on alcohol use faces 2 complications. First, the military considers a confluence of factors before determining whether to deploy a service member, creating a nonignorable exposure and unbalanced comparison groups that inevitably complicate inference about the role of deployment itself. Second, regression analysis assumes that a single effect estimate can approximate the population's change in postdeployment alcohol use, which ignores previous studies that have documented that respondents tend to exhibit heterogeneous postdeployment drinking behaviors. Therefore, we used propensity score matching to balance baseline covariates for the 2 comparison groups (deployed and nondeployed), followed by a variable-oriented difference-in-differences approach to account for the confounding and a person-oriented approach using a latent growth mixture model to account for the heterogeneous response to deployment in this prospective cohort study of the US Army National Guard (2009–2014). We observed a nonsignificant increase in estimated monthly drinks in the first year after deployment that regressed to predeployment drinking levels 2 years after deployment. We found a 4-class model that fit these data best, suggesting that common regression analyses likely conceal substantial interindividual heterogeneity in postdeployment alcohol-use behaviors.

Alcohol misuse is a widespread problem in the military. In a population-based survey of active duty personnel, about 20% of respondents reported past-month heavy alcohol use (defined as 5 or more drinks per typical occasion at least once per week), and 47% reported binge drinking (5 or more drinks per occasion for men or 4 or more for women, at least once in the past month) (1), compared with approximately 7% and 25%, respectively, among nonmilitary civilians ages 18 or older (2). Among both military personnel and civilians, alcohol use has been linked to aggressive behavior (3–5), injuries (6, 7), marital issues (4, 8–10), and psychiatric comorbidity (11). Preventive strategies against increases in alcohol use and early recognition of risk of abuse are therefore of great interest.

A large body of research has shown that combat-area deployment is associated with increases in alcohol use (1, 12, 13). Several coping models, such as the Tension Reduction Hypothesis (14, 15), affect regulation model (16), and alcohol-stress vulnerability model (17), have been proposed to suggest that some people exposed to stressful events may self-medicate with alcohol to reduce event-related psychopathology. A major shortcoming of these coping models is that the entire group of deployed personnel is assumed to respond similarly to the same stimuli, ignoring the large interindividual differences within the group. Because much literature has documented that stress can have both negative and positive effects on psychopathology and health behaviors (18, 19), the alcohol-stress vulnerability model (17) provides an alternative explanation for the association between combat-area deployment and alcohol use.

A life-course epidemiology of trauma suggests that the emergence of psychopathology represents a consequence of the complex accumulation and interaction of life experiences that range from social to biological factors that occur over the life span (20). Under the alcohol-stress vulnerability model, this complex accumulation and interaction of life experiences produces each person's vulnerability to alcohol use after stress such that the same stressful stimuli can elicit different patterns of psychopathology as a function of interindividual differences in vulnerability to stress (21, 22). Therefore, estimation of the effect of deployment on alcohol use is challenging for 2 reasons. First, if those factors that increase vulnerability to postdeployment alcohol use are associated with a service member being deployed, a selection bias is introduced. To reduce bias when inferring causal effects from observational data, we must assume that confounding factors (i.e., factors that cause both the exposure and the outcome of interest) are balanced between the 2 groups compared. However, the military considers a confluence of factors before determining whether to deploy a service member, including both operational factors (e.g., need for service members of a particular occupational specialty) and individual factors (e.g., time since last deployment, psychological fitness). As such, deployment is a nonignorable exposure, making it difficult to identify a reasonable group of nondeployed military personnel that has a similar balance of potential confounding factors as do deployed personnel. Unable to identify a defensible reference group of nondeploying service members, studies that have evaluated the association between deployment and alcohol use among service members have often evaluated no control group (23–25).

Second, using regression to identify the average effect of deployment on alcohol use is likely to conceal the different patterns of postdeployment alcohol-use behaviors hypothesized by the alcohol-stress vulnerability model (17). For example, the average causal risk difference is estimated as the average treatment effect for the treated (ATET) (i.e., proportion of respondents increasing their alcohol use after deployment minus the proportion of respondents decreasing their alcohol use after deployment). As such, a positive risk difference suggests that a greater proportion of deployers increased their alcohol use than decreased their alcohol use compared with nondeployers. Given the negative and positive effects of stress on psychopathology and health behaviors (18, 19), heterogeneity in these effects is likely and should be explored using a model that accommodates multiple effects of exposure. For this reason, a major shortcoming of most types of regression analysis is that they assume that respondents come from a single population, which ignores previous studies that have documented heterogeneous postdeployment drinking behaviors among respondents.

To our knowledge, no study has handled selection into deployment in a meaningful way and explored heterogeneity in the course of alcohol use during the postdeployment years. Here, we applied a modeling approach that addresses the nonignorable assignment of deployment and estimates the average effect of deployment on alcohol use for subsets of persons who share similar characteristics. This modeling approach is applied in 2 steps. First, we propose a propensity score–matched, difference-in-differences analytical strategy to identify the average effect of deployment on alcohol use over 2 years. The strength of this combined approach is that, whereas propensity score matching can account for observed factors that might confound the causal effect of the exposure on the outcome, difference-in-differences models address unobserved factors that are stable over time and might confound the causal effect. However, time-varying factors that are not explicitly adjusted for in the model can still confound the effect estimate under study. Second, we employ a latent growth mixture model (LGMM) to explore the distribution and identifying characteristics for subsets of persons who exhibit similar alcohol-use patterns after deployment (26, 27). And while we aimed to understand those characteristics that determine whether deployment has a negative or positive effect on alcohol use, such an analysis would require prior information about the factors that might modify alcohol-use patterns after deployment. Absent a strong understanding of these factors, we use LGMM to explore both the distribution of and the identifying characteristics for homogeneous subsets of deployers exhibiting heterogeneous patterns of alcohol use after deployment.

Whereas the propensity score–matched difference-in-differences analytical strategy aims to address the nonignorable assignment of deployment to estimate the average effect of deployment among the deployed, the LGMM aims to estimate the average effect for heterogeneous subsets of homogeneous deployers who share similar deployment and nondeployment characteristics. As such, our modeling approach extends previous work by combining methods for propensity score matching (28, 29), to control for confounding variables, with LGMM for visualizing longitudinal data to explore the groups of deployers exhibiting similar patterns of postdeployment patterns of alcohol use in a longitudinal sample of National Guard service members.

METHODS

Study population

Participants were a sample initially recruited as US National Guard service members from the state of Ohio (n = 2,616; age, mean = 30.7 (standard deviation, 9.5) years; 85.2% male; 87.8% non-Hispanic white) (30). Data were collected annually for waves 1 to 5 from 2009 to 2014. Retention of wave 1 participants was 80.5%, 68.7%, 60.6%, and 52.2% at waves 2–5, respectively. A second round of baseline interviews for new participants (n = 578) was also initiated in 2010 and 2011 to replenish the sample after loss to follow-up.

Regardless of the calendar year that a respondent entered the study, respondents with a baseline interview and 2 additional interviews were eligible for inclusion. Of the total 6,402 person-years observed, 45% of deployers (203 of 314 person-years) and 44% of nondeployers (2,646 of 6,031) met the eligibility criteria.

The study protocol was approved by the Ohio National Guard and the institutional review boards of the University Hospital Case Medical Center, University of Toledo, University of Michigan, Ann Arbor Veterans Administration Medical Center, Columbia University, Boston University, and the Office of Humans Research Protections of the US Army Medical Research and Materiel Command.

Measures

Deployment

At each wave, respondents were asked whether they had been deployed since their last interview and the location of any deployment. We categorized respondents who reported a deployment to either Iraq or Afghanistan during their most recent deployment as being deployed to a combat area, based on the overlap in dates between this study's data collection (i.e., 2009–2014) and Operation Enduring Freedom (2001 to present) and Operation Iraqi Freedom (2003–2011).

Alcohol use (Quantity-Frequency Index)

Past-month alcohol use was measured at each wave with a 2-item measure asking participants “Thinking about just the past 30 days, on how many days did you drink any alcoholic beverage?” and “On the days when you drank alcohol over the past 30 days, on average, how many drinks did you have each day?” The product of these 2 questions provided a measure of the estimated number of drinks consumed over the past month. Because this measure was right skewed, we applied a logarithmic transformation to the alcohol-use measure.

Potentially confounding variables

Twenty-seven potentially confounding sociodemographic, military, psychiatric, and general health variables were included in the model that estimated each respondent's propensity to deploy. See Web Table 1 (available at https://academic.oup.com/aje) for full list of variables included in the propensity score model. We gathered information for each of the time-stable potential confounding variables at baseline as well as information for each of the potential time-varying confounding variables at the wave immediately prior to deployment for deployers or immediately prior to the deployment year for their matched deployers among nondeployers.

Statistical analysis

Our analysis occurred in 2 phases, a matching phase and an analysis phase. First, we estimated the propensity score for deployment for each person at each wave and formed a matched data set by connecting each respondent deployed to a combat area with respondents who were not deployed to a combat area and who had the most similar propensity score from the cohort. Propensity scores (predicted probability of deployment) were estimated for each deployed respondent on their index date (i.e., deployment date) and used to match each deployer to a contemporaneous nondeployer. We estimated the predicted probabilities of deployment (propensity scores) for each eligible respondent using logistic regression, with the outcome being deployment between study intervals, and predictors being derived from the respondent characteristics preceding the index date for each respondent. This modeling approach allowed the propensity score to flexibly account for temporal changes in deployment environment. Within each cohort wave, each eligible deployer was matched to 4 nondeployers using both a nearest-neighbor and an optimal-matching strategy. In accordance with Rosenbaum and Rubin (31), we estimated the standardized mean difference in both the original data and the matched data to determine the best matching strategy. Specifically, a standardized mean difference of less than 0.2 after adjustment indicates balance on the measured covariates between the 2 groups after matching (32). We conducted the propensity score matching in R (R Foundation for Statistical Computing, Vienna, Austria), using the MatchIt package (33).

Second, we conducted both a variable-oriented and a person-oriented analysis to estimate the effect of combat-area deployment on alcohol use. Specifically, we estimated the risk difference using a mixed effects linear regression model that estimated the difference-in-differences of alcohol use among matched deployers and nondeployers (method 1), and we estimated the number and nature of growth curves using LGMM analysis to examine differences in the proportion of deployers and nondeployers assigned to each growth curve (method 2). Method 1 used a difference-in-differences specification with individual-level fixed effects to precisely estimate the effect of deployment on the change in monthly alcohol use:

$$\begin{array}{c}{R}_{\mathit{ijk}}={\mathrm{\beta }}_{0\mathit{jk}}+{\mathrm{\beta }}_1\left({\mathit{Deployed}}_{\mathit{jk}}\right)+{\mathrm{\beta }}_2\left({\mathit{Year}}_{\mathit{ijk}}\right)\hfill \\ +{\mathrm{\beta }}_3\left({\mathit{Deployed}}_{\mathit{jk}}{\times \mathit{Year}}_{\mathit{ijk}}\right)+{\mathrm{\mu }}_{\mathit{jk}}+e_{\mathit{ijk}}\hfill \end{array}$$

$R_{\mathit{ijk}}$ is the annual monthly alcohol-use rate at time i for person j in each propensity score–matched pair k. We used conditional likelihood, stratified propensity score–matched pairs, and estimated individual intercepts for each stratum, represented in this model as ${\mathrm{\beta }}_{0\mathit{jk}}$. ${\mathit{Deployed}}_k$ indicates the deployed-to-a-combat-area exposure versus the nondeployed. Year indicates the follow-up study interview year. The coefficient of interest (${\mathrm{\beta }}_3$) multiplies the interaction term, time × group, which is equal to 1 for deployers in the postdeployment period. We conducted the difference-in-differences analysis using SAS, version 9.2 (SAS Institute, Inc., Cary, North Carolina).

In method 2, we employed an LGMM to estimate alcohol-use trajectories using the Mplus statistical modeling program, version 7.11 (34). LGMM uses information about interindividual differences and intraindividual changes over time to identify distinct classes of individuals who follow like trajectories of a single outcome variable across multiple time points. Three steps were required to fit the final model. First, we used conventional polynomial growth models to test whether the data exhibited linear or quadratic growth, determining that these data exhibited a quadratic growth. Second, we fit a series of LGMM models to determine best fit, beginning with a 1-class model and progressing to a 6-class model. To allow for differences in baseline drinking levels, we let the intercept vary and held constant the variance of the slope. The best model fit prioritized parsimony, lowest Bayesian information criterion, lowest Akaike information criterion, significant (P < 0.05) adjusted Lo-Mendell-Rubin likelihood ratio test, and highest entropy.

RESULTS

Focusing first on the representativeness of the 203 deployed respondents who met the study inclusion criteria (i.e., had a baseline interview and two additional interviews) (Table 1), the included study sample of 203 deployers tended to be more likely to be married (55.7% vs. 38.8%; d = −0.34), were less likely to be either never married (36.0% vs. 47.9%; d = −0.24) or previously married (8.4% vs. 13.2%; d = 0.16), and reported more time in service (10.2 vs. 9.1 years; d = 0.25) than excluded deployers. Although no significant differences in percentage with a mental disorder were found (i.e., Cohen's d ≥ 0.2), a greater proportion of the included study sample met criteria for past-year posttraumatic stress disorder (14.3% vs. 8.3%) and a lesser proportion of the study sample met criteria for current alcohol abuse (29.1% vs. 33.0%).

Table 1.

Sensitivity Analysis: Descriptive Statistics for Key Variables Among Included and Excluded Deployers, Ohio Army National Guard Mental Health Initiative, 2009–2014

Characteristic	Included Deployers (n = 203)		Excluded Deployers (n = 121)		Cohen's d
	No.	%	No.	%
Age, yearsa	31.0 (9.1)		30.1 (9.8)		0.13
Male	189	93.1	105	86.8	−0.21
Marital status: Never married	73	36.0	58	47.9	−0.24
Marital status: Married	113	55.7	47	38.8	0.34
Marital status: Previously married	17	8.4	16	13.2	−0.16
Parent or primary caregiver	89	43.8	49	40.5	−0.07
Education: Some college or more	159	78.3	95	78.5	0.07
Annual income > $40,000	135	66.5	73	62.9	0.15
Currently employed	167	82.3	98	81.0	−0.03
Lifetime deployments: 0	81	39.9	58	48.3	−0.16
Lifetime deployments: 1	53	26.1	30	25.0	0.03
Lifetime deployments: ≥2	69	34.0	33	26.7	0.16
Time since last deployment, daysa	1,485.7 (1,028.9)		1,450.6 (593.2)		−0.09
Mean time in service, yearsa	10.2 (8.4)		9.1 (9.0)		0.25
Posttraumatic stress disorderb	29	14.3	10	8.3	0.19
Depressionc	3	5.9	0	0.0	0.03
Current alcohol dependenced	31	15.1	18	14.9	−0.06
Current alcohol abusee	59	29.1	39	33.0	−0.11

^a Values are presented as mean (standard deviation).

^bDiagnostic and Statistical Manual of Mental Disorders, Text Revision IV, criteria (39).

^c Major depression was diagnosed if 5 or more of 9 depressive symptom criteria were present at least “more than half the days” in the prior 2 weeks, and 1 of the symptoms was either depressed mood or anhedonia.

^d Three or more alcohol-dependence symptoms.

^e One or more alcohol-abuse symptoms.

There were 2,344 respondents (deployers = 203; nondeployers = 2,141) who met the study inclusion/exclusion criteria. The overall incidence of deployment was 8.7% per year. The distribution of propensity score estimates overlapped between the 2 groups: Deployers spanned from 0.02 to 0.48, and nondeployers from 0.00 to 0.52. This overlap indicated a region of common support and suggested that nondeployed matches could be identified for each deployed respondent (Web Figure 1). Next, we conducted a 1:4 match using both a nearest-neighbor strategy and an optimal-match strategy. After examination of the standardized mean differences after matching, we determined that the optimal match strategy provided the best match. Table 2 shows that the standardized mean differences after matching indicated that the treatment groups were balanced on the measured confounders (i.e., Cohen's d ≥ 0.2).

Table 2.

Descriptive Statistics for Key Variables in the Propensity Score Model Among the Deployed and Nondeployed Groups, Ohio Army National Guard Mental Health Initiative, 2009–2014

Characteristic	Deployed (n = 203)		Nondeployed (n = 812)		Cohen's d
	No.	%	No.	%
Age, yearsa	31.0 (9.1)		31.5 (9.5)		−0.04
Male	189	93.1	750	92.4	0.03
Marital status: Never married	73	36.0	305	37.6	−0.03
Marital status: Married	113	55.7	450	55.4	0.00
Marital status: Previously married	17	8.4	57	7.0	0.05
Parent or primary caregiver	89	43.8	371	45.7	−0.04
Education: Some college or more	159	78.3	669	82.4	−0.10
Annual income > $40,000	135	66.5	555	68.5	−0.01
Currently employed	167	82.3	685	84.4	−0.06
Lifetime deployments: 0	81	39.9	326	40.2	−0.01
Lifetime deployments: 1	53	26.1	217	26.7	−0.01
Lifetime deployments: ≥2	69	34.0	269	33.1	0.02
Time since last deployment, daysa	1,485.7 (1,028.9)		1,607.8 (1,230.4)		0.01
Mean time in service, yearsa	10.2 (8.4)		11.1 (8.7)		−0.06
Posttraumatic stress disorderb	29	14.3	99	12.2	0.06
Depressionc	3	5.9	12	0.9	0.05
Current alcohol dependenced	31	15.1	112	13.8	0.04
Current alcohol abusee	59	29.1	201	24.8	0.08

^a Values are presented as mean (standard deviation).

^bDiagnostic and Statistical Manual of Mental Disorders, Text Revision IV, criteria (39).

^c Major depression was diagnosed if 5 or more of 9 depressive symptom criteria were present at least “more than half the days” in the prior 2 weeks, and 1 of the symptoms was either depressed mood or anhedonia.

^d Three or more alcohol-dependence symptoms.

^e One or more alcohol-abuse symptoms.

Figure 1 shows that deployers, compared with their nondeployed matches, reported a higher mean number of monthly alcohol drinks before deployment and at 1 year after deployment but not at 2 years after deployment. Furthermore, this figure shows that, relative to predeployment, past-month alcohol use was not significantly different at either 1 year or 2 years afterward between matched deployers and nondeployers. Table 3 shows that the difference-in-differences general linear model did not find that the deployers’ monthly alcohol use differed between any of the study waves relative to their matched controls. Although we found that the mean number of alcoholic drinks monthly among deployers increased by 2.24 from predeployment to 1 year afterward, decreased by 6.37 from 1 year to 2 years, and decreased by 4.21 from before deployment to 2 years after, compared with their matched controls, large standard errors of the estimates resulted in all estimates being insignificant at α = 0.05. The large standard errors suggest that the differences in monthly alcohol consumption among the matched sets exhibited substantial heterogeneity. As such, we explored this heterogeneity using LGMM.

Figure 1.

Change in the annual average rate of monthly drinks by deployment status before and after deployment or index date, Ohio Army National Guard Mental Health Initiative, 2009–2014. Estimates result from difference-in-differences regression of the annual rate of monthly drinks between deployed (n = 203) and nondeployed (n = 812) service members before deployment or index date and at 1 and 2 years afterward.

Table 3.

Difference-in-Differences Generalized Linear Models Estimating the Number of Monthly Drinks^a by Deployment Status in a Matched Data Set of the Ohio Army National Guard Mental Health Initiative, 2009–2014

Deployment and Year Category	Coefficient	SE	P Value
Before Deployment or Index Date to 1 Year Afterward
Deployed	3.51	1.65	0.28
Year	1.12	1.66	0.50
Deployed × Year	2.24	3.30	0.62
1 Year to 2 Years After Deployment or Index Date
Deployed	1.34	1.42	0.52
Year	−3.03	1.43	0.03
Deployed × Year	−6.37	2.83	0.27
Before Deployment or Index Date to 2 Years Afterward
Deployed	0.29	1.38	0.84
Year	−2.02	1.38	0.14
Deployed × Year	−4.21	2.77	0.13

Abbreviation: SE, standard error.

^a Product of “Thinking about just the past 30 days, on how many days did you drink any alcoholic beverages?” and “On average, how many drinks did you have each day?”

Web Table 2 shows the fit indices of past-month alcohol use for the 1-class model through the 6-class model. The 4-class model produced a viable, theoretically defensible solution, with good fit statistics (Akaike information criterion = 9,625.3; Bayesian information criterion = 9,718.8; entropy = 0.83; adjusted Lo-Mendell-Rubin likelihood ratio test: P < 0.01). Figure 2 shows that the majority of respondents were assigned to a class with consistently high drinking (47.3%). Other classes were characterized by consistently low drinking (33.2%), high predeployment drinking that became low at 1 year afterward and increased at 2 years (high-decreasing; 6.9%), and low predeployment drinking that increased at 1 year and decreased at 2 years (low-increasing; 12.6%). Table 4 shows that a higher proportion of deployers than nondeployers were in the low-increasing class (14.3% v. 11.2%, respectively), whereas a higher proportion of nondeployers than deployers were in the high-decreasing class (6.4% v. 3.9%, respectively). In a post hoc analysis of deployment-related combat experiences (Web Table 3), deployers in the low-increasing class were found to have similar deployment experiences as their peers; however, deployers in the low-increasing class tended to be younger, with fewer years of service experience. Because the low-increasing class of deployers tended to be younger than their peers, and alcohol use tends to increase from late adolescents to early adulthood, the increased alcohol use 1 year after deployment among the low-increasing class could represent the influence of age or deployment on alcohol use. Whereas a typical age-related alcohol-use trajectory forms a long arc that changes slowly from one year to the next, the return to near predeployment drinking levels from 1 year to 2 years after deployment suggests a more dramatic change, suggesting that the postdeployment increase in drinking is more likely to represent the association between deployment and alcohol use over and above the influence of age.

Figure 2.

Past-month alcohol-use trajectory over 3 years among study participants in the Ohio Army National Guard Mental Health Initiative, 2009–2014. The 4 groups (and prevalence of group within sample) are depicted with the following symbols: square, high-stable (47.3%); circle, high-decreasing (6.9%); triangle, low-increasing (12.6%); ×, low-stable (33.2%).

Table 4.

Trajectory of Estimated Monthly Drinks by Deployment Status Among Participants in the Ohio Army National Guard Mental Health Initiative, 2009–2014

Trajectory	Deployed (n = 203)		Nondeployed (n = 812)
	No.	%	No.	%
High-stable	117	57.6	380	46.8
High-decreasing	8	3.9	52	6.4
Low-increasing	29	14.3	91	11.2
Low-stable	49	24.1	289	35.6

DISCUSSION

We have presented a framework for estimating the effect of combat-area deployment on alcohol use over time. The approach we presented used a propensity score–matching method to balance baseline covariates for the 2 comparison groups (deployed and nondeployed), followed by a variable-oriented approach based on mixed models to account for the confounding and a person-oriented approach based on LGMM to investigate the heterogeneous response to deployment.

The results of our variable-oriented analysis do not provide evidence of a significant positive association between combat-area deployment and monthly alcohol use. We obtained positive, although nonsignificant, estimands on the rate of change in alcohol use 1 year after a combat-area deployment relative to nondeployers. This increase in postdeployment alcohol use is consistent with earlier reports in Operation Enduring Freedom/Operation Iraqi Freedom populations (1, 12, 13, 23). In our study, however, the increase in alcohol use from predeployment to 1 year after deployment was offset by a subsequent decrease in past-month alcohol use during the subsequent year. Specifically, deployers reported drinking about 7 fewer alcoholic drinks on average between 1 year after deployment (23.0 monthly drinks) and 2 years after deployment (16.1 monthly drinks). To our knowledge, no previous study has looked at the drinking behaviors in Operation Enduring Freedom/Operation Iraqi Freedom veterans across 2 years after deployment; we believe this is the first study to observe that monthly alcohol use regressed to the mean in the period between 1 and 2 years after deployment. While this finding suggests that a critical period for prevention of postdeployment alcohol misuse is isolated to the year immediately after deployment, variable-oriented analyses can conceal the interindividual differences in postdeployment drinking trajectories that we clarified in the LGMM analysis.

Our person-oriented trajectory analysis addressed a key methodological limitation of previous studies using variable-oriented analyses. Although the majority of both the deployers and the nondeployers exhibited a stable trajectory, a substantial proportion of respondents exhibited either an increasing or a decreasing trajectory. That a 4-class solution fit these data best, consistent with other studies (35), exhibits an increasingly recognized finding that a mean alcohol-use trajectory reflects a mixture of multiple trajectories. We observed heterogeneity in alcohol-use trajectories among both the deployers and the nondeployers, which suggests that other factors, beyond the deployment itself, drive the interindividual changes in alcohol use over time. Because studies that have evaluated the influence of deployment on service members have often evaluated no control group (23–25), all the factors that might drive these interindividual changes in alcohol use over time are uncontrolled. The strength of using a difference-in-differences model, over an approach that does not use a control group, is that unobserved factors that are stable over time and might confound the causal effect are addressed; however, time-varying factors that are not explicitly adjusted for in the model can still confound the effect estimate under study. Therefore, previous studies that have examined alcohol use among deployed personnel, absent a control group, likely have conflated the effects of deployment with other unmeasured factors that are driving whether deployment has a negative or positive effect on alcohol use over time. Whereas variable-oriented analysis (i.e., difference-in-differences) assumes that the factors driving interindividual differences in alcohol-use trajectories are negligible or appropriately controlled for in the model, a person-oriented analysis models this heterogeneity.

To interpret an association as the ATET using a variable-oriented analysis, we most often assume that interindividual differences in response to an exposure are minimal or random; however, previous studies have demonstrated that persons exposed to deployment exhibit heterogeneous changes to psychiatric symptomology (36, 37). When theory does not support a consistent estimator, the ATET is the proportion of persons increasing their alcohol use minus the proportion of persons decreasing their alcohol use between the groups. Thus, we propose that alcohol use increased at a positive rate between baseline and 1 year afterward using a difference-in-differences approach because a greater proportion of persons were in an increasing trajectory than decreasing trajectory among deployers than nondeployers. Future psychiatric epidemiology studies should consider whether the ATET under study might conceal informative and knowable interindividual differences in how the outcome is patterned over time. Under the condition of a consistent estimator, difference-in-differences models address unobserved factors that are stable over time and might confound the causal effect, making this an ideal analytical strategy. However, when a population is likely to be composed of heterogeneous subsets of persons exhibiting homogeneous response patterns over time, LGMM provides the ideal analytical strategy to estimate the average effect for each subset.

Our study findings should be interpreted within the context of 2 limitations. First, we did not consider the influence of time-varying factors on alcohol use. However, previous study findings have suggested that early life experiences (e.g., childhood maltreatment) might have the greatest confounding effect on the relationship between exposure to potentially traumatic events and alcohol use (21, 22). As such, we prioritized addressing any confounding that would arise from a nonignorable treatment assignment and preexposure differences using propensity score matching (31) and fixed-effects modeling (38), respectively. Second, we experienced substantial censoring in our study sample. This censoring was a function of both the overall loss to follow-up among study participants and the exclusion criteria among deployed personnel. At this time, no analytical strategy exists to address the former when applying a propensity score–matched difference-in-differences analysis. As such, the development of a strategy to address censoring that uses, for example, inverse probability weights represents an important avenue for future research. In addition, our study's inclusion criteria required that deployers completed both the data collection immediately prior to their deployment and the 2 data collections immediately after their deployment. Because deployers excluded from this study were more likely than those included to be single, male, and have less time in service—and persons in the increasing alcohol-use trajectory had the lowest mean age and shortest time in service compared with the other alcohol-use trajectories—it is likely that a high proportion of excluded persons would have been classified into the increasing alcohol-use trajectory. Inclusion of these respondents might have resulted in a higher estimate of deployment on alcohol use using the difference-in-differences method.

In conclusion, we observed a nonsignificant increase in estimated monthly drinks in the first year after deployment that regressed to predeployment drinking levels 2 years after deployment. A greater proportion of deployers in the increasing trajectory than in the decreasing trajectory, compared with nondeployers, likely drove this increase in postdeployment alcohol use at 1 year after deployment. We used combined methods to address challenges to causal identification inherent in studies that examine the psychiatric sequelae of combat-area deployment, specifically nonignorable treatment assignment and heterogeneous response to exposure. Specifically, we used propensity score–matched difference-in-differences and LGMM analyses to address these. Whereas propensity score matching can account for observed factors that might confound the causal effect of the exposure on the outcome, difference-in-differences models address unobserved factors that are stable over time and might confound the causal effect. In addition, whereas difference-in-differences analysis estimates the ATET, an LGMM can investigate the ATET for a subset of persons that share similar characteristics. Future studies should consider using both variable-oriented and person-oriented analyses to better understand the distribution of response patterns that comprise the more often employed variable-oriented approach.