Participants
Data for the current analysis were collected in Project EAT (Eating and Activity in Teens and Young Adults), a longitudinal study of weight-related health. In 1998–1999 (wave 1), a cross-sectional investigation of students at 31 public secondary schools in the metropolitan area of Minneapolis-St. Paul, Minnesota, U.S. was conducted. Public middle schools and high schools in the Minneapolis/St. Paul metropolitan area, serving socioeconomically and racially/ethnically diverse communities, were invited to participate in the study. Of 53 schools contacted, 31 agreed to take part. Participants aged 11 to 18 (mean age 14.9, n = 4746) completed surveys during class time, and height and weight measurements were taken by trained research staff in a private area [22, 23].
Given growing interest in weight-related health, a decision was made to follow-up with participants who provided sufficient contact information (n = 3672) at 5-year intervals. Follow-up assessments were completed by mail and online in 2003–04 (wave 2, n = 2516), 2008–09 (wave 3, n = 2287) and 2015–16 (wave 4, n = 1830) as participants progressed through adolescence and early adulthood [24]. At each wave participants completed a survey of around 100 questions, focused on diet, eating behaviours and other weight-related behaviours. At the final follow-up, wave 4, invitations to complete the survey were mailed along with a two-dollar bill to all participants who had responded to at least one previous follow-up survey. Survey invitation letters provided the web address and a unique password for completing the online version of the survey. Non-responders received up to six reminders through a combination of U.S. mail, email, and text messages. The final two mailed reminders included paper copies of the survey and FFQ, and all mailings provided the option to complete the survey by phone. Reminder emails were additionally mailed to participants that did not complete the survey and FFQ after logging into the online version. Participants were mailed a financial incentive ($50) following survey completion. The University of Minnesota’s Institutional Review Board Human Subjects Committee approved all protocols, with parental consent and participant assent obtained at wave 1 and participant consent obtained at each of the subsequent waves. In this study, given the focus on life transitions between waves of data collection, individuals were included in the analysis if they reported data at 2 or more waves of Project EAT (n = 2902), and therefore provided data that covered at least one period during which a transition may occur.
Exposures
To allow for longitudinal comparisons, each wave of the EAT study included identical measures of key constructs. We report the psychometric properties of measures using data collected for Wave 4. The estimates of item test-retest reliability, reported below, were determined in a subgroup of 103 participants who completed the Wave 4 survey twice within a period of one to 4 weeks. The EAT surveys included measures of each of the five major life transitions that was based on assessment of living arrangements, participation in education and employment, and having children. Below we describe the measures used to assess each transition and how these were used to create the exposure variable used in the analysis. In a small number of cases (less than 10% of the population for each exposure) individuals who had gone through a life transition returned to pre-transition state (e.g. they moved back into the parental home). As we were primarily interested in the first occurrence of a life transition, we coded any data points after the return to the pre-transition state as missing, thereby removing them from the analysis. For example, if a participant reported living with parents at wave 1, not living with parents at wave 2, but then reported living with parents again at wave 3, the variable “leaving the parental home” would be coded as missing at wave 3 and wave 4 for that participant. This only applied to a small number of participants and did not contribute greatly to missing data.
Leaving the parental home was measured by asking participants about their current living arrangements at Waves 2–4. Participants were asked to select all that applied of nine options (e.g., “my parent(s)”) in answering “During the past year, with whom did you live the majority of the time?”. As this item was not asked when participants were under age 18, we assumed all participants were living with their parent(s) at Wave 1 and at Wave 2 if they were still a high school student. Based on the selection of “my parents” (Test-retest agreement = 97%) from the list of provided response options, participants were categorized as ‘living with parents’ or ‘not living with parents’ at waves 2–4. The transition variable of ‘leaving the parental home’ was positive the first time participants reported not living with parents.
Leaving full-time education was based on data about current educational status collected at waves 2, 3 and 4 in response to the question: “Which of the following best describes your student status (for the majority of the past year)?” (Test-retest agreement = 95%). Responses included type of education (e.g. community or technical college, four-year college) and whether education was full or part-time. Since Wave 1 surveys were administered in school classrooms, we assumed that all participants were attending full-time education at this time point. Graduate studies (reported at waves 3 and 4) was included under part-time education since no information was available as to whether graduate study was full-time or part-time, and data from the U.S. Department of Education suggests that 64% of graduate students completed their studies as part-time students in 2015–16 [25]. Less than 8% of the sample reported graduate studies at Waves 3 and 4, so this assumption is not likely to strongly affect our results. The transition variable of ‘leaving full-time education’ was positive the first time that participants reported participation in part-time or no education.
Beginning full-time employment was based on data about current employment status, derived in response to the question: “How many hours a week do you work for pay?” (Test-retest r = 0.80). Response options differed for those still in high-school compared to subsequent questionnaires. Those still in high-school only reported hours of work up to a maximum of ‘20 or more hours a week’; we categorised these individuals as participating in part-time work since they were still in full-time education. For those who had completed high-school, we categorised full time work as 30 or more hours per week [26]. The transition variable of ‘beginning full-time employment’ was positive the first wave that participants reported full-time work.
Beginning cohabitation with a significant other was based on responses to the current living situation measure that is described above. Participants who selected “my husband/wife” (Test-retest agreement = 93%), “my partner of the opposite sex” (Test-retest agreement = 92%), or “my partner of the same sex” (Test-retest agreement = 100%) were categorized as cohabiting with a significant other. As above, since this question item was not asked when participants were under age 18, we assumed all individuals at wave 1 and those who were still high school students at wave 2 did not live with a partner. The transition variable of ‘beginning cohabitation’ was positive the first time that participants began to report living with a partner.
Becoming a parent was based on responses to survey questions that asked participants to report how many children they had at waves 2, 3, and 4 (Test-retest r = 0.98). This question was not asked at wave 1, and it was assumed that participants did not have any children at wave 1, unless further detailed data on child age collected at wave 4 indicated that they would have already had a biological child at wave 1 (n = 10). 'Becoming a parent' was positive the first time that individuals moved from having no children to having one or more child at a subsequent wave.
Outcomes
Fast food consumption frequency was reported at each wave in response to the question: “In the past week, how often did you eat something from a fast food restaurant (like McDonald’s, Burger King, etc.)?”. Response categories were converted to a continuous scale of times per week: never [recoded to 0 times per week], 1–2 times [1.5], 3–4 times [3.5], 5–6 times [5.5], 7 times [7], more than 7 times [10], as done previously [27]. This measure was developed for the Project EAT study and is comparable to established measures used by other studies such as the National Longitudinal Study of Adolescent Health [13].
Covariates
We included baseline sociodemographic variables that may be associated with both life transitions and underlying changes in fast food intake as time-invariant covariates. These include age, gender, race/ethnicity, parental socio-economic status (SES) and health status [9, 15]. Baseline age of the participant in years and months, and gender were reported at wave 1. Race/ethnicity was also reported at wave 1 and collapsed to give two categories: (1) White (non-Hispanic) and (2) People of Colour. Parental SES was primarily determined from parental education, defined by the highest level of education achieved by either parent, as reported by the participants at Wave 1. This measure was complimented by additional variables on family eligibility for public assistance (yes, no, and do not know), eligibility for free or reduced-cost school meals (yes, no, and do not know), and maternal and paternal employment status (full-time, part-time, not working, and do not know), as described previously [28], to give a 5-level variable from 1 (low SES) to 5 (high SES). Overall health status was reported at waves 1 and 2 in response to the question “How would you describe your health?”, with response options poor (1), fair (2), good (3) or excellent (4). In order to provide a longer-term measure of health rather than only a snapshot at the time the initial questionnaire was completed, responses were averaged across waves 1 and 2 to represent baseline health status.
Statistical analysis
Descriptive analyses were performed to evaluate sample population demographics and changes in fast food intake and occurrence of life transitions across the 4 waves. We tested for differences between the sample demographics of those included and excluded from our analysis (age, gender, race/ethnicity, parental SES and baseline health status) using linear regression and chi-squared tests. We calculated the phi coefficients of associations between pairs of transitions, to assess the likely influence of collinearity.
Latent growth curve models were used to test our research questions. Growth models were developed as a hierarchy of increasing complexity. First an unconditional growth model was estimated to examine overall growth trajectories and to test for individual variability in change over time. We tested the main effects of 5 time-invariant covariates: baseline age, gender, ethnicity, parental SES and baseline health status, including these in the model as predictors of the intercept, slope and quadratic slope of the growth curve (Fig. 1). Additional latent factors were then included to test for associations of each of the early adult life transitions with changes in fast food intake between pairs of waves, following Curran et al. [29], as shown in a path diagram in Fig. 1. Each of these latent factors (int2, int3 and int4) represents an additional intercept which is added to the growth curve at wave 2, 3 and 4 respectively. These additional intercepts are regressed onto the transition variables (T1-T5) and their variance constrained to 0 such that the regression coefficients produced represent a step increase or decrease in fast food intake associated with each transition. For each transition we compare those who went through the transition with a reference category of those whose status did not change between consecutive waves: they did not transition, or stayed in the transitioned state between 2 waves. Since the additional intercept between waves is regressed on all transition variables, the associations of each of these transitions are mutually adjusted.
To examine the associations between fast food intake and life transitions, we developed two separate growth curve models: in the first model, ‘Transitions across all waves’, the association between the transition and the change in fast food intake was constrained to be the same across all waves, providing an overall indication of change in fast food intake associated with that transition, regardless of the timing of the transition during early adulthood. We then removed this constraint, to assess ‘Transitions between pairs of waves’, allowing the association between the transition and fast food intake to vary over time, to look at the effect of a transition occurring between each contiguous pair of waves.
Since adolescents were recruited through schools we also tested whether school-level intraclass correlation coefficients (ICCs) might suggest use of multilevel analysis to adjust for clustering by school. However, ICCs were low (ICC of 0·02 for baseline fast food intake and lower at subsequent waves), which, with an average cluster size of 94, suggests little impact of clustering on these analyses. Therefore, adjustment for school-level clustering was not included.
All models were estimated in Mplus (version 8.3) using the maximum likelihood estimator with robust standard errors (MLR), which is robust to non-normality in the dependent variable. Missing data were addressed using the full information maximum likelihood estimation, under the Missing at Random assumption, with variables related to attrition included as time-invariant covariates in the model. Model fit was assessed using a panel of fit indices: Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), Root Mean Square Error Of Approximation (RMSEA), Comparative Fit Index (CFI), Standardized Root Mean Square Residual (SRMR).