Data
Individual participant health care costs related to obesity and overweight were obtained by linking two national databases, i.e. the Belgian Health Interview Survey (BHIS) 2013 and the national health insurance data compiled by the Intermutualistic Agency (IMA) 2013–2017. Linkage was performed by means of a National Registry Number. The BHIS was conducted between January and December 2013 among a representative sample of the Belgian population (N = 10,828) and comprises data on health status and related health behaviour and determinants. Respondents were recruited following a multistage sampling design, as described in detail elsewhere [11]. Interviews were performed using a face-to-face paper and pencil interviewing, supplemented with a self-administered questionnaire covering more sensitive topics [11]. Health insurance is compulsory in Belgium covering more than 99% of the population. The linked IMA database used for this study comprises aggregated reimbursed health care costs from 2013–2017 for all HIS participants including expenditures for 1) ambulatory care (pharmaceuticals excluded), 2) hospital care, and 3) reimbursed medicines purchased through public pharmacies. The linked IMA database readily included only information on hospital care variable costs (i.e. costs depending on the type of interventions performed during the hospital stay). However, in Belgium, the national health insurance also pays a fixed amount to the hospitals per admitted patient, depending on the type of hospital and treatment. Precise information on these costs was not directly available in the dataset. In order to estimate the fixed part of the total hospital care cost, the hospitalizations per patient per year were multiplied with the average annual 100% per diem cost publicly available by type of hospitalization (per diem costs available through: https://www.riziv.fgov.be/nl/themas/kost-terugbetaling/door-ziekenfonds/verzorging-ziekenhuizen/Paginas/verpleegdagprijzen-ziekenhuizen.aspx). Finally, we summed up the estimated fixed costs with the available variable hospital costs resulting in the total hospitalization costs used in this analysis.
The study included the adult population (age ≥ 18 years) who reported weight and height and for whom linkage with health insurance data was possible and were continuously insured from 2013–2017 (latest linkage available). People who deceased during the study period (from their participation to the BHIS until 31/12/2017) were excluded. The final study sample comprised 7,633 participants (Fig. 1).
Health care costs were analysed by BMI category calculated from self-reported weight and height obtained from the BHIS using the classification recommended by the World Health Organization, i.e., underweight (BMI < 18.5 kg/m2), normal weight (18.5 ≤ BMI ≤ 24.99 kg/m2), overweight (25 ≤ BMI < 30 kg/m2) and obesity (BMI ≥ 30 kg/m2) [12]. Socio-demographic variables taken into account included age, gender, household educational level (i.e. the highest educational level within the household), and income level (based on the calculated quintiles of the household income), as well as behavioural risk factors with respect to alcohol misuse, smoking, poor dietary quality and physical inactivity. Analyses also concerned information from the BHIS database on the prevalence of 23 major chronic conditions, i.e., asthma, chronic bronchitis, myocardial infarction, coronary heart disease, other serious heart disease, hypertension, high blood cholesterol, stroke, narrowing of blood vessels, arthritis (including rheumatoid arthritis and osteoarthritis), low back pain, neck pain, diabetes, allergy, stomach ulcer, cirrhosis of the liver, cancer, severe headache, urinary incontinence, serious gloom or depression, thyroid problems, and eye disease.
Absenteeism was reported in the BHIS as days absent from work during the 12 months prior to the BHIS interview queried by the following question: “Have you been absent from work during the past 12 months due to health problems? In doing so, take into account any conditions, injuries or other health problems you may have had and which resulted in an absence from work”. Followed by the question: “How many days in total have you been absent from work for the past 12 months due to health problems? If you are unable to indicate this number of days correctly, please give an estimate.”. The question was asked to working individuals only (N = 3,857) – individuals that stated to have a paid job at the moment of the interview.
Analysis
Analyses were conducted in R 4.0.5 [13] taking the design of the survey into account. The sampling design included stratification at the level of the provinces and clustering at the household level, as described in Demarest et al. [11]. Analysis of socio-demographic characteristics and healthcare costs per BMI category were performed using nominal logistic regression for comparison of proportions between BMI categories with normal weight as reference group. Confidence intervals (CI) were computed via the delta method, using the standard errors resulting from the survey analysis.
Health care costs
Overall health care costs and health care costs by type and by payment modality were calculated per BMI category. Univariate and multivariable regressions with negative binomial distribution and log link were used to explore the extent to which average yearly health care cost was associated with BMI category, socio-demographic characteristics and behavioural risk factors. The univariate model of health care cost in function of BMI-class allowed to estimate the unadjusted incremental health care costs and to evaluate statistical differences in average costs between individuals with underweight, overweight and obesity compared to normal weight individuals.
A “double-selection” approach was used for the selection of the variables to be included in the final multivariable model, based on backward elimination to identify significant variables at the 10% level [14]. The variables were identified in two steps, finding those that predict the dependent variable (costs) and those that predict the independent variables (BMI-class). The final linear regression included the covariates identified in either of the two steps. Candidate explanatory variables included age groups, gender, household educational level, household level of income and some behavioural risk factors such as smoking, alcohol misuse, unhealthy eating behaviour and physical inactivity. This use of double-selection is more likely to detect common causes of BMI-category and costs, and thereby results in more accurate inferences that also acknowledge the uncertainty in the selected variables.
Indirect costs—Cost of absenteeism
Cost of absenteeism was computed by multiplying the number of days absent from work by the national average labour cost per day. Using the costing year 2010 from Eurostat, the average Belgian labour cost per working day was estimated at €257 (monthly labour cost and assuming 18.8 working days per month (i.e., 52 weeks * 5 working days minus 24 days (legal holidays and agreed extra holidays) minus 10 public holidays) [15]. However, since respondents might have included weekend days in their answer, the total days absent from work was subtracted from the maximum number of working days per year, i.e., 226 days. If this difference was equal to or greater than zero, the answer was kept, else the maximum number of working days was used. The “double-selection” approach was performed also for the indirect costs, see above for more details.
Attributable cost of excess weight status.
The final regression model allowed to estimate the adjusted attributable costs and associated uncertainties of overweight and obesity compared to normal weight. Incremental costs were estimated at the individual level using the method of recycled predictions (also known as direct standardisation or g-computation) that allows to estimate the marginal effect from overweight and obesity on health care costs [16, 17]. The coefficients of the regression model were used to 1) predict health care costs for each respondent using the BMI from their reported weight and height; 2) predict health care cost assuming all respondents did have a normal BMI, keeping all other characteristics as observed; 3) calculate the individual incremental cost of obesity as the difference of an individual’s predicted costs assuming they were affected by obesity or overweight and assuming they had normal weight; and 4) calculate the attributable cost of obesity as the population survey-weighted average of the individual incremental cost. In order to jointly reflect prediction and survey uncertainty, means and CIs by BMI-classes were computed via bootstrapping with 1000 replicates and 1000 Monte Carlo simulations drawn per replication (for the survey design), leading to 1000*1000 interactions. In addition, total direct costs were calculated multiplying the average incremental cost by the proportion of individuals with overweight and obesity in the total adult population on the 1st of January 2018 (N = 9,074,575) [18]. According to the BHIS2018, 33.4% and 15.9% of the Belgian adult population were respectively affected by overweight and obesity. Total indirect costs were calculated multiplying the average incremental absenteeism cost by the proportion of individuals with overweight and obesity in the total population with a paid job according to BHIS2013 (N = 3,906,170) [18], namely 33.9% and 11.3% of the Belgian population with a paid job were respectively affected by overweight and obesity.
Relative contribution of chronic conditions
To investigate the relative importance of chronic conditions contributing to differences in health expenditure of persons with obesity and overweight compared to normal weight individuals, we evaluated how much of the attributable cost of excess weight can be attributed to each of 23 diseases. For this, we 1) extended the regression model for health care costs to also include the considered disease along with the covariates significant in a model with the disease as dependent variable; 2) used the model to predict health care cost assuming all respondents had a normal BMI, keeping all other characteristics (including disease status) as observed; 3) subtracted the predictions obtained in step 2 of the previous section from the obtained predictions; and 4) calculated how much of the attributable cost of obesity is due to the considered disease as the population survey-weighted average of the individual incremental cost obtained in the previous step, and dividing this by the average incremental cost of excess weight. For these analyses, the underweight population was omitted, considering that diseases related to underweight are commonly different from those related to excess weight. This method allowed to rank the diseases by their relative contribution to the incremental cost of obesity.