Inter-Cohort Variation in the Consequences of U.S. Military Service for Men’s Mid- to Late-Life Body Mass Index Trajectories

Cohort birth years

Years turned 18 (historical period)

Age range in 1992

Age range in 2010

<1929

<1947 (WWII)

64 and older

82 and older

1929–1936

1947–1954 (Post-WWII/Korean War)

56–63

74–81

1937–1945

1955–1963 (Cold War)

47–55

65–73

1946–1956

1964–1974 (Vietnam War)

36–46

54–64

Notes: The 1992 HRS included subjects ages 51–61 who were born between 1931 and 1941. Older subjects in our analysis were initially recruited into AHEAD, which collected data from subjects ages 70 and older in 1993. The HRS and AHEAD were merged in 1998 at the same time additional birth cohorts were recruited into the study. HRS age-eligible subjects are over the age of 50 at the time of entry into the study. For details regarding the HRS study design see http://hrsonline.isr.umich.edu/sitedocs/DesignHistory.pdf.

It should be noted that the youngest subjects entered the HRS later in the study period and therefore contributed fewer person-period observations to the analysis. This is due to the design of the HRS, which initially included subjects who were ages 51–61 in 1992, who were born between 1931 and 1941, and its companion study, the AHEAD, which collected data from subjects ages 70 and older in 1993, who were born before 1924. Individuals born from 1924 to 1930 and after 1941 were not included in the HRS until it was merged with the AHEAD and moved to a steady-state design in 1998. At that point, two cohorts—one born 1924–1930 and the other born from 1942 to 1947—were added. Since 1998, additional new cohorts have been added every 6 years (Health and Retirement Study 2008).

Control and Potentially Mediating Variables

The analysis includes a broad range of control and potentially mediating variables, which prior research indicates are associated with both veteran status and BMI. The first set of control variables are retrospectively reported early-life characteristics that occurred prior to military service: race/ethnicity, early-life socioeconomic disadvantage, and early-life health. Race/ethnicity includes non-Hispanic White (reference), non-Hispanic Black, non-Hispanic other race, and Hispanic (of any race). Early-life socioeconomic disadvantage is an indexed scale, which is comprised of four dichotomously coded retrospective childhood variables: mother’s education and father’s education (<8 years = 1; ≥8 years = 0); father’s occupation when the respondent was age 16 (unskilled manual = 1, non-manual, skilled, and professional = 0); and overall family SES from birth to age 16 (poor = 1, not poor, including the family was “pretty well off financially,” “about average,” and “it varied” = 0). Per the procedure outlined in Wilmoth et al. (2010), respondents missing on any of these four variables were assigned to the zero category for that variable, which is a conservative approach to dealing with missing data because that category represents greater advantage and misclassification would bias toward the null. We summed these four variables and divided by the number of items answered to create an early-life disadvantage scale that ranges from 0 to 1, with higher values indicating more disadvantage. Because a relatively large proportion of respondents had missing values on at least one early-life disadvantage item, which is due primarily to father’s absence or attrition prior to the 1998 survey when these questions were asked for the first time, we also include a variable in the models that is equal to one for all individuals for whom at least one of these variables was missing and set to zero. Additionally, the analysis includes a measure of health from birth to age 16 years that measures poor health and missing relative to good health (reference).

The other variables, which reference mid- to late-life characteristics that potentially mediate the relationship between military service and BMI, are measured many years after military service has ended. All of these mid- to late-life variables, except education, are time-varying across the 18-year study period. Marital status includes four categories: married (reference), never married, divorced/separated, and widowed. Education is measured at entry into the HRS and includes high school or less (reference), some college or college graduate, and more than college. Household income is measured in dollars. Labor force status is recoded as a binary variable (in the labor force = 1). Two health behaviors are measured dichotomously: ever smoked cigarettes (yes = 1) and currently drinks alcohol (yes = 1).

Finally, we include three methodological controls. These are binary variables measuring ever had a proxy interview (yes = 1), dropping out of the study (yes = 1), and dying during the 18-year time span of the study (yes = 1).

Analysis Plan

After describing the sample, we estimate conditional growth curve models using the person-period file described above and the PROC MIXED procedure in SAS. In all of these trajectory analyses, we define time in terms of chronological age (as opposed to study duration) because we are interested in modeling age-related changes in BMI. The grand mean age for all sample members, which is equal to 67 years, is used to center age.

Models include terms for age and age squared to test for nonlinearity in the relationship between age and BMI. Additionally, we include terms that interact age and age squared with the veteran status and cohort measures to account for potential nonlinearity in the slope that estimates age-related changes in BMI in relation to our focal variables. The models control hierarchically for early-life and mid- to late-life characteristics, as well as methodological controls for proxy response, attrition, and death over the study period, although the coefficients for those controls are not shown (full models are available upon request). We focus on interpreting the coefficients for the effect of veteran status and cohort membership on BMI at the mean age of 67 years, and the effect of veteran status and cohort membership on age-related change in BMI. Positive coefficients indicate that men who are veterans or in a particular cohort have higher BMI relative to the reference group, whereas negative coefficients indicate the opposite.

To facilitate the interpretation of the fully-specified model, we present predicted age-related BMI trajectories by birth cohort and veteran status. The predicted values represent men with the following characteristics: non-Hispanic White; mean early-life disadvantage; good early-life health; high school graduate; out of the labor force; mean household income; married; never smoked; current non-drinker; never had a proxy interview; not lost to follow up; and did not die during the study period. Predicted values are only presented for the age ranges over which the birth cohorts are observed during the study period, which are shown in Table 7.1. The figure only presents predicted values at age 50 and older because the men must have been at least 50 to be eligible for inclusion in the HRS.

Results

Sample Description

Table 7.2 presents descriptive statistics overall and by veteran status. As seen in Table 7.2, the characteristics of veterans and non-veterans differ substantially. Overall, the mean BMI is 27.4, with veterans having slightly, but significantly, lower BMI than non-veterans (27.34 versus 27.49). The mean age of the sample is 67, but veterans are significantly older than non-veterans. The sample is fairly evenly distributed across the first three birth cohorts, with each containing less than one-third of the sample; as expected due to the design of the HRS (see discussion in Methods), the youngest cohort has a substantially smaller percentage (11 %). Consistent with the very high rates of participation in the military in and around WWII, veterans are over-represented among the oldest two cohorts, while non-veterans have relatively higher representation in the younger two cohorts.

Table 7.2

Total sample characteristics by veteran status^a

	Total sample		Non-veteran		Veteran
	%	Mean	%	Mean	%	Mean	p
Variable
BMI		27.40		27.49		27.34	***
Age (in years)		67.00		65.62		68.15	***
Veteran
Yes	54.52
No	45.48
Birth cohort
<1929	27.17		21.27		32.09		***
1929–1936	29.46		25.00		33.17
1937–1945	32.19		38.35		27.06
1946–1956	11.18		15.38		7.68
Race/ethnicity
White	77.13		69.54		85.02		***
Black	12.63		17.23	16.56	9.35
Other	2.14		3.00		1.42
Hispanic	8.10		12.77		4.21
Childhood disadvantage index (0–1)	0.34			0.37		0.32	***
Poor childhood health
Yes	4.37		5.41		3.50		***
No	95.63		94.59		96.50
Education
High school or less	58.05		64.81		52.41		***
Attended/graduated college	29.26		22.48		34.92
More than college	12.69		12.71		12.67
Labor force participation
Yes	31.97		36.72		28.01		***
No	68.03		63.28		71.99
Household income (in $1,000)	62.50			64.94		60.47	*
Marital status
Married/partnered	80.46		79.32		81.43		***
Divorced/separated	8.37		8.94		7.89
Widowed	8.29		8.06		8.48
Never married	2.88		3.68		2.20
Ever smoked
Yes	71.43		66.20		75.79		***
No	28.57		33.80		24.21
Currently drinks alcohol
Yes	57.66		54.51		60.29		***
No	42.34		45.49		39.71
Proxy report
Yes	23.17		27.34		19.69		***
No	76.83		72.66		80.31
Lost to follow-up
Yes	17.29		18.47		16.31		***
No	82.71		81.53		83.69
Died
Yes	23.70		22.81		24.44		***
No	76.30		77.19		75.56

Significance: *p < 0.05; **p < 0.01; ***p < 0.001

^aData are un-weighted

There are significant differences between veterans and non-veterans with respect to the early-life controls. Veterans are significantly more likely than non-veterans to be non-Hispanic White (85 % versus 69 %), which is expected on the basis of the policies that prevailed prior to the end of World War II (see Lutz 2013). Interestingly, the mean childhood disadvantage index is lower among veterans than non-veterans, but, as expected, a smaller percentage of veterans report having had poor childhood health.

There are also significant veteran status differences in the potentially mediating variables. Education levels are higher among veterans, with 35 % having attended or graduated from college (compared to 22 % among non-veterans). Veterans are more likely to be currently out of the labor force and have mean household incomes that are significantly lower than non-veterans. Veterans are slightly more likely to be currently married (81 %) than non-veterans (79 %). In addition, veterans are significantly more likely than non-veterans to have ever smoked (76 % versus 66 %) and to currently drink alcohol (60 % versus 55 %).

Finally, veterans and non-veterans differ with respect to the methodological controls. During the study period, veterans are less likely than non-veterans to have had a proxy report (19 % versus 27 %) or be lost to follow-up (16 % versus 18 %), but are somewhat more likely to have died during the course of the study (24 % versus 23 %).

BMI Trajectories in Later Life

Table 7.3 presents four models predicting BMI. Model 1, which only includes age, age squared, and veteran status, indicates that veterans have significantly lower BMI than non-veterans at the mean-centered age of 67. In addition, as expected, for all men, BMI decreases at an increasing rate with age.

Table 7.3

Growth curve models predicting BMI trajectories