Study Design, Population, and Data Source
We conducted a retrospective, longitudinal cohort study of veterans who served in Afghanistan (Operation Enduring Freedom; OEF) and Iraq (Operation Iraqi Freedom; OIF and Operation New Dawn; OND) who had used the VA healthcare system for a clinical visit between October 7, 2001 (the beginning of OEF) through December 31, 2011. We included 496,722 veterans (59,790 female and 436,932 male veterans) whose height and weight were recorded at the VA at least once after the end of their last deployment and whose first post-deployment outpatient encounter at the VA was at least 1 year prior to the end of the study period. Data for this study were derived by merging the OEF/OIF/OND Roster with three other VA national level administrative data sets: Demographics and Vital Signs (via the Corporate Data Warehouse), National Data Extract of pharmacy data via Decision Support System, and the National Patient Care Database of outpatient clinical encounters and associated clinical diagnoses. The study was approved by the Committee on Human Research, University of California, San Francisco, and the San Francisco VA Medical Center.
Outcome Measures
Body Mass Index (BMI)
We calculated BMI by dividing weight in kilograms (kg) by height in meters (m) squared (kg/m2) for each weight measurement, using the median of each patient’s recorded height measurements at all clinical visits during the study period. For our analysis, we used BMI measurements up to 3 years following the index weight measurement (after the last deployment); the average patient-level BMI was calculated for each 6 month interval starting at the index BMI measurement.
Only biologically plausible heights and weights were included in the analyses (> 70 lb and < 700 lb; > 46 in. and < 84 in.). To further improve data quality, all available weight measurements (both pre-deployment and post-deployment) were included in a linear mixed model of BMI over time (with a random intercept and slope for each veteran, adjusted for age and gender), in order to identify and exclude within-patient outliers (absolute value of conditional residual ≥ 10). BMI measurements were retained if they were both biologically plausible and not extreme outliers (16 ≤ BMI ≤ 52). We used the International Classification of adult underweight, overweight and obesity according to BMI.16
The number and frequency of BMI measurements varied across veterans. The percentage of the cohort with BMI data during each 6-month interval subsequent to the first interval was as follows: 6–12 months: 44 %, 12–18 months: 40 %, 18–24 months: 32 %, 24–30 months: 28 % and 30–36 months: 23 %. Veterans with PTSD had slightly more BMI measurements recorded than veterans without PTSD. We confirmed that the mean and median index BMI did not differ for those with only one measurement compared to those with more BMI measurements.
Pregnancy-related diagnostic codes17 were used to determine whether a female veteran was pregnant at each clinical encounter. All BMI measurements in the 6 months preceding and 12 months subsequent to any clinical encounter with a pregnancy-related code were excluded from the analysis. As a result, a total of 2,061 women were completely excluded and an additional 7,623 women had some BMI values excluded.
Predictor Variables
Age at first BMI measurement and gender were extracted from the VA demographics data file. Race, marital status, and military characteristics [end of last deployment date, armed forces branch (Army, Navy/Coast Guard or Air Force), rank, component type (Reserves or Active Duty), and number of deployments (whether deployed once versus multiple times)] were extracted from the OEF/OIF/OND Roster. Distance and type of the nearest VA facility was derived from the OEF/OIF/OND Roster by the VA Planning and System Support Group. Mental health diagnoses were identified using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnostic codes associated with VA clinical visits, and included depression (293.83, 296.20–296.25, 296.30–296.35, 300.4, and 311), PTSD (309.81), other anxiety disorders (300.00–300.09, 300.20–300.29, and 300.3), adjustment disorders (308, 309.0–309.9, excluding 309.81), alcohol use disorders (AUD; 305.00–305.03 and 303), and drug use disorders (DUD; 305.20–305.93 and 304). Veterans were also categorized by whether they had none, one, two, or three or more mental health diagnoses. Information about antidepressant medication use (including monoamine oxidase inhibitors, tricyclic antidepressants, selective serotonin reuptake inhibitors, atypical antidepressants, and serotonin-norepinephrine reuptake inhibitors) and antipsychotic medication use for greater than 30 days within each 6-month interval was obtained from VA outpatient pharmacy data.
Statistical Analysis
We compared baseline demographic, military and clinical characteristics in female and male veterans with and without PTSD diagnoses using the chi-square test for categorical variables and the t-test for continuous variables.
Growth mixture modeling (GMM) is a person-centered approach to modeling longitudinal trajectories that assumes that the data are from a heterogeneous population made up of a mixture of types or “classes” of individuals (latent trajectory classes).18 This approach allows for differences in patterns of change in BMI over time where different classes of individuals are allowed to vary around different mean intercepts and slopes. For example, there may be distinct latent BMI trajectory classes, such as individuals who are slowly increasing in weight, rapidly increasing in weight, slowly losing weight, etc. The relationship between BMI trajectory and mental health conditions was allowed to vary by latent class.
We used GMM to identify latent classes of BMI trajectories and to estimate class-specific mean initial BMI (intercept) and class-specific mean change in BMI (linear and quadratic slopes). Random intercepts and slopes were included and the random effect variances were assumed to be equal across classes. Women and men were modeled separately. We specified an analysis model that included age at first post-deployment BMI measurement and race/ethnicity as covariates for the purpose of identifying the number of trajectory classes. We selected a solution based on a combination of theory, previous research findings,6
,
19 model fit criteria,18
,
20
,
21 and parsimony. We performed multinomial logistic regressions using posterior probability-based multiple imputations (“pseudo-class” draws) to examine the association between the latent BMI trajectory classes and each mental health condition, separately. First we built models adjusted for age and race/ethnicity (as described above), and then we built models further adjusted for demographics (time from last deployment to index BMI, marital status, distance to and type of the nearest VA facility) and military characteristics (branch of service, rank, component type, and number of deployments), as well as antipsychotic medication use. We performed a sensitivity analysis that further adjusted for antidepressant medication. We also tested interactions between PTSD and age at baseline and between PTSD and race, given that both age and race are associated with BMI. Full-information maximum likelihood algorithm for handling data that is missing at random and missing completely at random was used for estimation of GMM. We assigned individuals to their most likely latent class based on posterior probabilities from the GMM models, in order to determine the probability of belonging to latent BMI trajectory class by mental health condition. Results were considered statistically significant at the p < 0.001 level, given the large sample size.22
,
23 We used SAS (v 9.3, SAS Institute Inc., Cary, NC) for chi-square, t-tests and linear mixed models and Mplus (v 6.12) for all GMM models.24