Participants
This research was conducted using the Taiwan MJ cohort resource—a longitudinal, population-based health dataset run by the MJ Health Management Institution, Taiwan(11). The details of the MJ cohort population and data collection are reported elsewhere(12–15). Briefly, the MJ cohort has enrolled about 600,000 Taiwanese individuals since 1994.
Participants
in the MJ cohort were healthy individuals who received health examinations including a self-reported questionnaire on medical, social and family history as well as demographic information and underwent a series of medical tests and physical examinations.
The details of participant selection for the present study are shown in Fig.
1. The database accumulated 290,279 participants who were born before 1977 and were first included in the MJ cohort between 1998 and 2006. We obtained all of the included participants’ data that had been collected from the year they were first included in MJ cohort to year 2017. Setting three years as an interval, we built a 7-wave longitudinal dataset (The interval of seventh wave has only 2 years, 2016–2017). If two or more measurements for one person were accessible within one interval, the measurement closest to the center of the interval was chosen. We excluded those who were less than 40 years old at 1998 (n = 51,731), less than two waves follow-up (n = 145,270), and who had ever been diagnosed at baseline with any of the following self-reported conditions: cancer, stroke (n = 3392). The final study sample included 89,886 participants.
Figure 1 Patient attrition and cohort selection. Inclusion and exclusion criteria show cohort selection for the MJ cohort dataset.</fig>
Death information was based on the National Death Registry obtained from the Ministry of Health and Welfare, Taiwan, and was linked to the MJ cohort. All deaths were identified from death certificates and confirmed by trained physicians. Follow-up time for mortality started at the date of each participant’s first measurement and was censored by October 31, 2019, the date when we linked death information to the MJ cohort. We categorized the cause of mortality into three leading causes: “cancer” (including all kinds of cancer), “cardiovascular disease”, and “respiratory disease.” All other causes of death were grouped into the category “Other.”
Outcome measures
This analysis focused on three separate outcome measures: (1) change of BMI across time, (2) hazard of all-cause mortality in distinct groups, and (3) hazard of cause-specific mortality in distinct groups.
Exposure
Body height and weight for each participant were measured at every follow-up visit and BMI was calculated.
According to the BMI classification by the World Health Organization (WHO), BMI above 30 should be classified as obesity, BMI between 25 and 30 should be classified as overweight, and BMI below 18.5 should be classified as underweight(
16).
Covariate variables
Age, smoking status, alcohol consumption, educational level, and physical activity at baseline were obtained from the self-reported questionnaire from the MJ cohort and were incorporated as covariates in the present study. Age was defined at the year of 1998. Smoking status and alcohol consumption were both classified into three categories: never smoker/drinker, former smoker/drinker, and current smoker/drinker. Educational level was classified into two categories: high school or less, and college or above. Physical activity was classified into three categories: seldom (exercise less than two hours a week), sometimes (exercise between two and five hours a week) and frequent (exercise more than five hours a week).
Statistical Methods
We separated the participants into 4 groups by gender (male or female) and age (40 to 60-year-old and more than 60 years old).
Descriptive statistics (mean, standard deviation, and percentages) were used to summarize the participants’ demographic and clinical characteristics. We used a group-based trajectory model with maximum likelihood estimation to identify distinct trajectories of changes in BMI(
17) using the SAS PROC TRAJ program (SAS Institute, Inc., Cary, North Carolina). As recommended, we estimated models with 2–5 trajectories by assuming linear, quadratic, and cubic patterns of change in BMI over time; the best-fitting model (the number of distinct trajectories and the patterns of change in BMI) was determined on the basis of Bayesian Information Criterion (BIC) scores(
18).
The hazard ratio (HR) of all-cause mortality was analyzed using a Cox proportional hazards model, including BMI trajectory groups, and covariates (age, smoking status, alcohol consumption, educational level and physical activity). In this study, the time from baseline was used as the time scale to parameterize the baseline hazard function(19) because different birth cohorts were observed at different ages. The analyses were performed using the SAS PROC PHREG program.
A cause-specific hazards model was used to assess the HR of cause-specific mortality. This model can be estimated by censoring participants with the competing event and then fitting the standard Cox proportional hazards model(20). The analyses were performed using the SAS PROC PHREG program.