Study design and participants
NHANES is a national program designed to assess the health and nutritional status of both adults and children in the United States(9), and detailed information regarding the sampling design and data collection for NHANES can be found on the official website(10). Briefly, NHANES employs a stratified, multistage, probability sampling approach to collect nationally representative health-related data on the U.S. population. Comprehensive data collection methods were utilized, including completion of a household screener, a household interview, physical examinations and provision of biological samples. It has been conducted since 1999, with a survey cycle of every two years, continuously collecting demographic data, socioeconomic, dietary nutrition and health-related outcomes of the participants. NHANES was approved by the Ethics Review Committee of the National Center for Health Statistics. All participants provided informed consent.
In this analysis, we included 101,316 participants from 10 consecutive surveys of NHANES (1999–2000 to 2017–2018). We excluded participants who were ineligible for follow-up (n = 42,252) or younger than 20 years of age at baseline (n = 4,119), who had incomplete information on weight loss strategies (n = 60), who were underweight (i.e., baseline body mass index (BMI) < 18.5 kg/m2, n = 837) or had missing baseline BMI (n = 3,696), who had a positive lab pregnancy test or self-reported pregnancy or who were unable to determine pregnancy at the time of examination (n = 1,922). Ultimately, a total of 48,430 participants with a baseline age of 20 years and above, nonpregnant, and BMI ≥ 18.5 kg/m2 were included in the analysis. Figure S1 shows the inclusion screening process of the study participants.
Assessment of weight loss strategies
In the household interview, the participants were asked “During the past 12 months, have you tried to lose weight?” If the participants answered “yes”, then they were further asked “How did you try to lose weight?”. A variety of weight loss strategy options were provided. From 1999–2000 to 2017–2018, the NHANES provided a list of 14 to 21 options. The following 14 options were included in all surveys: (1) ate less food, (2) lowered calories, (3) ate less fat, (4) exercised, (5) skipped meals, (6) consumed diet foods or products, (7) used a liquid diet formula, (8) joined a weight loss program, (9) took prescription diet pills, (10) took nonprescription diet pills, (11) took laxatives or vomiting, (12) drank a lot of water, (13) followed a special diet, and (14) other strategies. Starting in 2005, the NHANES survey included 7 additional options, as follows: (1) reduced carbohydrate intake; (2) began or resumed a smoking habit; (3) increased intake of fruits, vegetables, and salads; (4) altered eating habits (e.g., no food consumption late at night); (5) reduced intake of sugar, candy, and sweets; and (6) reduced consumption of junk food or fast food; and (7) had weight loss surgery to lose weight. The strategies were not mutually exclusive, and the participants could choose one or more of them.
We included 14 weight loss strategies that were presented in all surveys in our main analysis. First, we counted the various numbers and percentages of participants who took from 0 to 14 cumulative weight loss strategies and the incidence of mortality. Based on trends between the number of weight loss strategies and mortality, participants were divided into five groups: 0 (no weight loss attempt), 1, 2, 3–4 and ≥ 5 (5 or more weight loss strategies) (Table S1). Then, to distinguish the association between various combinations of weight loss strategies and death, we clustered 14 weight loss strategies into 5 clusters using a clustering model (Figure S2).
Ascertainment of mortality
Deaths were identified by linking NHANES participants to death certificate records from the National Death Index (NDI) and tracked through December 31, 2019. We classified causes of death according to International Statistical Classification of Diseases, 10th revision (ICD-10) codes. The primary outcomes of our study were all-cause mortality, CVD mortality (codes I00-I09, I11, I13 and I20-I51) and cancer mortality (codes C00-C97).
Assessments of covariates
Information on covariates was obtained from the baseline information, including gender, age, race/ethnicity, education level, marital status, family income poverty ratio, weight change since last year, waist circumference, BMI, self-considered weight status, smoking status, drinking status, physical activity, total energy intakes, family history of prediabetes or diabetes and heart diseases, baseline history of diabetes, hypertension, high cholesterol, coronary heart disease and cancer, and history of medication for diabetes, hypertension and high cholesterol.
Data on BMI (kg/m2) and waist circumference (cm) were collected in the Mobile Examination Center (MEC) by trained health technicians. According to the classification standard of BMI by the National Institute of Health (NIH) and the World Health Organization (WHO)(11), we divided BMI into three groups: normal weight (18.5 to < 24.9 kg/m2), overweight (25.0 to < 29.9 kg/m2), and obesity (≥ 30.0 kg/m2). The family income poverty ratio was calculated by dividing family income by the poverty guidelines, specific to the appropriate year and participant’s state. We divided the ratio into three levels: low-income level (0 to ≤ 1.0), middle-income level (1.1 to ≤ 3.0), and high-income level (> 3.0)(12). Physical activity was assessed by investigating at least 10 minutes of moderate or vigorous activity in the past 30 days that resulted in sweating or increased breathing or heart rate, such as brisk walking, cycling, golfing, and dancing. Total energy intakes (kcal/d) was collected in person at the MEC in one 24-hour dietary recall interview. History of diseases was obtained by asking whether they had been informed of a disease by a doctor or health professional, and medication use was collected by self-report questionnaire or verbal interviews. Weight change since last year (pounds) was estimated as the difference value between current self-reported weight and self-reported weight one year ago.
Statistical analysis
Detailed information on missing covariates is presented in Table S2. To account for these missing values, the multiple imputation by chained equations (MICE) method was used, and 10 datasets were created through this imputation process. All variables, including the outcomes, were included in the multiple imputation model, ensuring a comprehensive imputation of missing values. Baseline characteristics are presented as the mean (standard deviation [SD]) for continuous variables and number (percentage [%]) for categorical variables.
We defined baseline as the time at which participants conducted their interviews. We calculated person-years from baseline to death, loss to follow-up, or December 31, 2019, whichever occurred first. The proportional hazards assumption was tested using Schoenfeld residuals. We used the Cox proportional hazards model to calculate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association of five groups of cumulative weight loss strategies (0, 1, 2, 3–4, ≥ 5) with all-cause and specific-cause mortality. Then, we used the latent class analysis (LCA)(13–15) clustering algorithm to cluster and combine 14 weight loss strategies into 5 clusters to explore the relationship between the combinations of two or more weight loss strategies and mortality after removing participants who used only one weight loss strategy.
We constructed two models to estimate associations. Model 1 adjusted for baseline gender (male or female) and age (years, continuous). Model 2 further adjusted for race/ethnicity (Hispanic, or non-Hispanic), education level (less than high school, high school or equivalent, or college or above), marital status (married, widowed, divorced, separated, or never married), family income poverty ratio (ratio, continuous), weight change since last year (pounds, continuous), waist circumference (cm, continuous), BMI (normal, overweight, or obesity), self-considered weight status (underweight, about the right weight, or overweight), smoking status (never, past, or current), drinking status (never, past or current), physical activity (yes or no), total energy intakes (kcal/d, continuous), family history of prediabetes or diabetes and heart diseases (yes or no), baseline history of diabetes, hypertension, high cholesterol, coronary heart disease and cancer (yes or no), and history of medication for diabetes, hypertension and high cholesterol (yes or no). We also calculated linear trends with cumulative weight loss strategies as continuous variables.
Using cumulative weight loss strategies as a continuous variable, we performed stratified analyses to explore the relationship between weight loss strategies and all-cause mortality in subgroups of gender (male or female), age (20 to ≤ 44, 45 to ≤ 59 or ≥ 60), education level (less than high school, high school or equivalent, or college or above), family income poverty ratio (> 3.0 or ≤ 3.0), smoking status (never, past, or current), drinking status (never, past or current), BMI (18.5 to < 24.9, 25.0 to < 29.9, or ≥ 30.0 kg/m2) and physical activity (yes or no), and the interaction effects were estimated.
We performed a series of sensitivity analyses to examine the robustness of the results. First, we excluded participants with missing covariates, restricting the analysis to complete cases. Second, to minimize the reverse causality bias, participants who had CVDs and cancer at baseline, who died within the first three years of follow-up or who did not report intentional weight loss but lost more than 10 pounds were excluded from the analysis. Third, we excluded participants with normal weight, and the analysis was restricted to overweight or obese participants (BMI ≥ 25.0 kg/m2) to reduce the health effects among normal weight people. Fourth, we used waist circumference to define obesity (men ≥ 102 cm, women ≥ 88 cm)(16), limiting the study to participants with abdominal obesity to analyze the association of weight loss strategies with mortality. Fifth, to reduce the impact of the limitations of the NHANES questionnaire settings on the results, we removed participants who adopted “other strategies” from the 14 weight loss strategies or included participants who additionally adopted any of the newly added 7 weight loss strategies since 2005 as “other strategies” for analysis. Finally, we further adjusted the survey rounds to reduce the effect of the birth cohort.
All statistical analyses were performed using R statistical software (version 4.3.0). We used the R package "mice" to perform efficient multiple imputation of missing values(17). All statistical tests were two-sided, and the test level α = 0.05, p < 0.05 was considered statistically significant.