Table shows the distribution of demographic and military characteristics ascertained from the DMDC; self-reported military, behavioral, and health characteristics; and survey response characteristics among responders and nonresponders. Overall, 71.6 percent (n = 54,960) of individuals completed the first follow-up survey. The response proportion was 10 or more percentage points above average (i.e., ≥81.6 percent) in the following subgroups: age ≥44 years, educational level of a bachelor's degree or higher; and rank of warrant or commissioned officer. The response proportion was 10 or more percentage points below average (≤61.6 percent) among individuals who were aged 17-24 years, no longer in the military at follow-up for reasons other than retirement, Marine Corps service members, missing smoking status, and those who reported having poor health or a major depressive disorder.
Table also includes the adjusted beta coefficients and P values for the full propensity score model for each stratum compared with its reference category. A coefficient <0 indicates that the multivariable-adjusted probability of response is smaller than the reference category (equivalent to an odds ratio [OR] <1). Conversely, coefficients >0 indicate that the multivariable adjusted probability of response is larger than the reference category (equivalent to OR >1). To describe an individual's estimated multivariable response probability, beta coefficients across the various characteristics are summed. Since the focus of this analysis was on development of the propensity scores, and not specific ORs, coefficients for each interaction term are not shown in Table ; the following example is provided for illustration. Women in the Army with less than a high school education (the reference category) were more likely to respond than men in the Army with less than a high school education (OR = exp[0.36] = 1.43). In contrast, women in the Army with a bachelor's degree were no more likely to respond than men in the Army with the same level of education (OR = exp[0.36 +(-0.33)] = 1.03).
Figure presents information about the distribution of propensity scores. The mean propensity score was 0.75, indicating that the average weight given to each follow-up responder was 1.33, while the minimum and maximum weights were 1.03 and 14.4, (corresponding to maximum and minimum propensity scores of 0.97 and 0.0696, respectively). The C statistic, a measure of the goodness of fit for the model, was equal to 0.71.
Propensity score statistics (N = 54,960 responders)
Tables , , and present ORs and 95 percent confidence intervals (CIs) for the complete case analysis (i.e., "unweighted" results, ignoring nonresponse) and weighted for nonresponse for the association between new-onset eating disorders (Table ), depression (Table ), and PTSD (Table ), respectively, and various exposures, including deployment experience, history of alcohol misuse, and smoking status. In the unweighted analyses, the following characteristics were associated with increased risks of new-onset eating disorders (Table ): a history of diagnosed mental disorders (men and women), being on a special diet for weight loss (men and women), being active duty (vs. Reserve/Guard, men only), and a history of major life stressors or alcohol misuse (both for men only). After weighting for nonresponse, there was little change in ORs for any strata and no change in the interpretation of results. In some cases, the 95 percent CIs were slightly wider for the nonresponse-weighted estimates. Associations between new-onset depression and deployment experience, smoking status, problem drinking, and PTSD at baseline among men and women were similar with and without weighting for nonresponse (Table ), although 95 percent CIs were again slightly wider in some cases after weighting for nonresponse.
Comparison of associations of new-onset eating disorders based on complete case results (ignoring nonresponse) and weighted for nonresponse
Comparison of associations of new-onset depression based on complete case results (ignoring nonresponse) and weighted for nonresponse
Comparison of associations of new-onset posttraumatic stress disorder based on complete case results (ignoring nonresponse) and weighted for nonresponse
As in the published study by Smith et al. [26
], ORs for the associations of deployment experience, gender, smoking status, problem drinking, military rank, and new-onset PTSD were stratified by service branch (Army, Air Force, Navy and Coast Guard, and Marines; Table ). The number of new-onset PTSD cases by service branch was 906 for Army, 184 for Air Force, 195 for Navy and Coast Guard, and 62 for Marine Corps. Particularly for Marines, the sample sizes were small and resulted in relatively imprecise OR estimates. In Army, Navy, and Coast Guard members, deployment without combat exposures was associated with statistically significant reductions in new-onset PTSD in the nonresponse weighted analyses, but not in the complete case analyses (nonresponse weighted analyses: Army, OR: 0.63, 95 percent CI: 0.44, 0.92; Navy/Coast Guard, OR: 0.48, 95 percent CI: 0.25, 0.90). Weighting for nonresponse in analyses of Marines resulted in a shift in ORs from greater than one to less than one, but the 95 percent CIs for both the unweighted and weighted analyses included the null value. For all service branches, the associations between deployment with combat exposure and risk of new-onset PTSD remained strong and positive after weighting. In the unweighted analyses, problem drinking in Marines was associated with a 73 percent increased risk of PTSD, which was of borderline statistical significance (OR = 1.73, 95 percent CI: 1.00, 2.99); after weighting for nonresponse, the point estimate was attenuated toward the null and the confidence limits widened to include 1.0 (OR: 1.60, 95 percent CI: 0.84, 3.04). Similarly, the association between enlisted rank and new-onset PTSD among Navy and Coast Guard members was no longer statistically significant (unweighted OR: 2.14, 95 percent CI: 1.16, 3.94; weighted for nonresponse OR: 1.99, 95 percent CI: 0.85, 1.68). Nevertheless, in both cases, the point estimates did not change substantially.
There was little difference in the distribution of self-reported health at follow-up, with and without weighting for nonresponse (data not shown). Adjusting for nonresponse resulted in a slightly greater proportion of individuals classified as having poor (0.9 percent weighted for nonresponse vs. 0.8 percent unweighted), fair (8.7 percent weighted for nonresponse vs. 8.1 percent unweighted), and good (35.4 percent weighted for nonresponse vs. 34.5 percent unweighted) health, and a smaller proportion classified as reporting very good (38.4 percent weighted for nonresponse vs. 39.2 percent unweighted) or excellent heath (15.3 percent weighted for nonresponse vs. 16.1 percent unweighted).