Study population and design

This study used data from the National Health Examination Survey (NHES) III, a population-based survey conducted by the Health Systems Research Institute and the Bureau of Policy and Strategy, Ministry of Public Health, Thailand during January 15-April 15, 2004. A three-stage stratified probability sampling was adopted for data collection. Details in sampling method were already mentioned elsewhere [9]. The final representative samples were 39290 individuals of aged 15 and over. For health assessment, people were interviewed using structural questionnaires containing general questions adapted from the health state description questions used in the WHO Multi-Country Survey Study on Health and Responsiveness 2000-2001 [24]. Weight and height were measured using standard techniques [25].

The present study focused on older persons; subjects aged 60 years and older were included in the analysis (n = 19372). The records of the interviewed individuals had been linked using a unique personal identifier (not all records had personal identifiers) to 2004-2007 death records from vital registration, which had per cent completeness greater than 90 [26, 27]. According to the personal identifier completion rate, only 17504 from the total number of older persons (90.4%) could be used to link to vital registration. From this linkage, older persons who died from all causes, except accidents and assault (57 cases), and those without information on health risk behaviours were included. The final number of subjects left in this analysis was 15997 or 82.6% of the total number of older persons.

Measures

All-cause mortality was the outcome variable of interest. Survival time was calculated as the interval between interview date and date of death or until the date of December 31, 2007 if those older persons were still alive.

The main independent variable, body weight, was measured as BMI (weight in kilogram divided by square of height in meter, kg/m²). In this study, it was classified into 7 categories according to the recommendation by WHO expert consultation with some modification [28]: < 18.5 kg/m², 18.5-22.9 kg/m², 23.0-24.9 kg/m², 25.0-27.4 kg/m², 27.5-29.9 kg/m², 30.0-34.9 kg/m² and ≥ 35.0 kg/m². For other independent variables which were taken into account as potential confounding factors were demo-social variables, namely, age, marital status, residential area, and education, and health risk variables, including, diabetes, hypertension, hypercholesterolemia, smoking, fruit and vegetable consumption and physical activity.

For demo-social confounding variables, age was categorized as follows: "60-69", "70-79", and "80+". Marital status was grouped into 3 categories: "single", "married", and "widowed/divorced/separated". For the education variable, the highest level of educational attainment was used and grouped into: "no education", "elementary level", and "secondary level or higher".

For potential health risk confounding variables, diabetes was defined as a person whose fasting blood sugar was ≥ 126 mg/dL or who had previously been diagnosed and was taking hyperglycaemic lowering drugs or insulin [29]. A hypertensive person was defined as a person whose systolic blood pressure ≥ 140 mm Hg and diastolic blood pressure ≥ 90 mm Hg or who had been diagnosed by a physician and was taking antihypertensive drugs [30]. A person having hypercholesterolemia was defined as a person with total blood cholesterol of ≥ 240 mg/dl (5.7 mmol/L) or who was taking hypercholesterolemia-lowering drugs [31]. We classified smoking behaviour of samples into 3 groups: "non-smoker" referred to a person who never smoked, "past smoker" referred to a person who smoked 100 or more cigarettes during his/her lifetime but stop smoking at the time of interview, "current smoker" referred to a person who smoked 100 or more cigarettes during his/her lifetime and was still smoking at the time of the interview. Seven-day recall data regarding fruit and vegetable daily intake were assigned to 2 groups: "less than 5 servings per day", "5 servings or more per day". For physical activities, people were surveyed using the WHO Global Physical Activity Questionnaire and classified as either "active" or "inactive" [32].

Statistical analysis

In the initial analysis, age-standardized mortality rates by BMI as well as other health risk behaviours were calculated separately for men and women. Then, the association between BMI with all-cause mortality was analysed using the Cox proportional hazard model. In the models, the hazard of death during 4 years of follow-up was a function of BMI and a set of covariates. Three models were created to assess the relationship of BMI and mortality. The first one was unadjusted model or model without controlling by other covariates. The second was an adjusted model where demo-social and health risk variables were added to control for as potential confounding factors. The third model was analysed for samples excluding those who were current smokers or alcohol drinkers. Interaction between BMI and co-morbidity including diabetes or hypertension were tested using a criteria suggesting interaction of P-value <0.02. All analyses were performed with the Stata/SE 10.0 for Windows statistical software package (StataCorp LP, College Station, TX, USA).

This study was approved by the Institute for the Development of Human Research Protection (IHRP), Ministry of Public Health, Thailand.

Characteristics of the studied subjects

Age-adjusted all-cause mortality rate by BMI and other potential health risks

During the mean follow-up time of 3.8 years (60545.7 person-years), 1575 subjects (936 men and 639 women) died (9.9%). Median survival time among those who died was 2.25 years (95% confidence interval [CI], 2.12-2.37) and 2.31 years (95% CI, 2.14-2.43) for men and women respectively.

Association of BMI and all-cause mortality

In Cox proportional hazard models (Table 3), there was evidence of the association between BMI and all-cause mortality. A flat U-shaped associations with the highest risk of death in the lowest and the highest categories of BMI (<18.5 kg/m² and ≥ 35.0 kg/m², respectively) was observed in older men (Figure 1, upper panel) while a reverse J-shaped association with the highest risk of dying in the lowest BMI category (<18.5 kg/m²) was observed in older women (Figure 1, lower panel). Among older men, BMI of <18.5 kg/m² was strongly associated with increasing risk of death but BMI of ≥35.0 kg/m² was weakly associated with increasing risk of death. Conversely, the other BMI categories were strongly associated with reducing risk of dying in older men. Among older women, the only BMI of <18.5 kg/m² was strongly associated with increasing risk of death. The higher BMI groups were strongly associated with reducing risk of dying except for the highest BMI category (≥35.0 kg/m²).

		Males			Females
Variables	Unadjusted Model	Adjusted Modela	Adjusted Model with Excluding Current Smoking & drinkinga	Unadjusted Model	Adjusted Modela	Adjusted Model with Excluding Current Smoking & drinkinga
	(n = 7742)	(n = 7742)	(n = 4928)	(n = 8255)	(n = 8255)	(n = 7805)
Age (yr) in 2004 (Ref: 60-69)
70-79	1.87 (1.62-2.16)	1.69 (1.46-1.95)	1.61 (1.33-1.93)	2.07 (1.74-2.46)	1.82 (1.52-2.18)	1.89 (1.57-2.27)
80+	4.31 (3.58-5.19)	3.42 (2.79-4.18)	3.61 (2.83-4.62)	5.07 (4.07-6.33)	3.95 (3.09-5.06)	4.06 (3.15-5.23)
BMI (kg/m2) (Ref: 18.5-22.9)
< 18.5	1.48 (1.26-1.74)	1.34 (1.14-1.58)	1.71 (1.37-2.11)	1.39 (1.13-1.72)	1.29 (1.04-1.60)	1.33 (1.06-1.67)
23.0-24.9	0.78 (0.64-0.95)	0.79 (0.65-0.97)	0.78 (0.61-0.99)	0.70 (0.55-0.89)	0.70 (0.55-0,90)	0.71 (0.55-0.91)
25.0-27.4	0.80 (0.65-0.99)	0.81 (0.65-1.00)	0.79 (0.61-1.03)	0.77 (0.60-0.97)	0.79 (0.62-1.01)	0.80 (0.62-1.02)
27.5-29.9	0.67 (0.48-0.93)	0.67 (0.48-0.94)	0.63 (0.43-0.94)	0.56 (0.40-0.78)	0.57 (0.41-0.81)	0.58 (0.41-0.82)
30.0-34.9	0.56 (0.33-0.95)	0.60 (0.35-1.03)	0.61 (0.34-1.09)	0.54 (0.37-0.81)	0.58 (0.39-0.87)	0.60 (0.40-0.90)
≥ 35.0	1.96 (0.81-4.73)	1.87 (0.77-4.56)	2.02 (0.74-5.52)	0.82 (0.41-1.66)	0.78 (0.38-1.59)	0.71 (0.22-1.51)

^a Adjusted for marital status, urban/rural, education, living arrangement, diabetes, hypertension, smoking, and physical activity.

There was a slightly reduction in risk of dying after controlled by all covariates (adjusted model) compared to the unadjusted model. In this model, an interaction between health risk variables, such as BMI with diabetes or BMI with hypertension, as well as an interaction between demo-social variables, such as BMI with age or BMI with sex or BMI with education were tested, but no statistical significance at P-value <0.2 were found among those pairs. In adjusted model, the hazard ratio (HR) according to BMI <18.5 kg/m², 23.0-24.9 kg/m², 25.0-27.4 kg/m², 27.5-29.9 kg/m², 30.0-34.9 kg/m², and ≥35.0 kg/m² relatively to BMI 18.5-22.9 kg/m² was 1.34 (95% CI, 1.14-1.58), 0.79 (95% CI, 0.65-0.97), 0.81 (95% CI, 0.65-1.00), 0.67 (95% CI, 0.48-0.94), 0.60 (95% CI, 0.35-1.03), and 1.87 (95% CI, 0.77-4.56), respectively, for men, and was 1.29 (95% CI,1.04-1.60), 0.70 (95% CI, 0.55-0.90), 0.79 (95% CI, 0.62-1.01), 0.57 (95% CI, 0.41-0.81), 0.58 (95% CI, 0.39-0.87), and 0.78 (95% CI, 0.38-1.59), respectively, for women.

Additional analysis was done for model excluding those who were current smoking or drinking (36.3% of total men, 5.5% of total women). In this reducing model, the pattern of association of BMI with all-cause mortality controlling for other covariates did not substantially change.