Excess body weight increases the burden of age-associated chronic diseases and their associated health care expenditures

Sample descriptive statistics

Table 1 summarizes the breakdown of the variables according to BMI classes, showing large differences in demographic and clinical parameters such as age and prevalence of comorbidities. Hypertension and type 2 diabetes were the most common BMI-associated health conditions and their prevalence shows a strong increase with increasing BMI (p=0.0001 for all BMI categories with respect to normal weight individuals). In addition, the prevalence of dyslipidaemia, CVD and arthrosis were higher in individuals with overweight and obesity than in normal weight individuals in both younger (<55 yrs) and older (>55 yrs) patients (p=0.0001 for all BMI categories with respect to normal weight individuals). There was a clear negative association between BMI and the proportion of individuals with no chronic disease (p=0.0001 for all BMI categories with respect to normal weight individuals). In contrast, the proportion of individuals affected by 2 or more chronic diseases increased sharply with raising BMI.

BMI and healthcare costs

Table 2 and Table 3 present the coefficient estimates of indirect (i.e. obesity-associated diseases), direct and differential effects of BMI on outpatient and total health care costs for each age group. The tables report marginal effects for each BMI category within each age specific subsample, as well as the percentage differences of the marginal effect estimate with respect to the annual average expenditure of normo-weight individuals.

As shown in Figure 1, after adjusting for age, sex, geographic residence, and GPs heterogeneity, there was a J-shaped association between BMI and overall (direct and indirect) total health care expenditure, which was stronger among middle-aged and elderly individuals. Total health care expenditures among the middle-aged (45-64 yrs old) individuals with overweight, obesity, severe obesity and very severe obesity were 10%, 27%, 52% and 68% higher, respectively, than among those with BMI of 18.5 to 24.9 (p=0.0001) (third panel of Table 3). In absolute terms, outpatient costs were more strongly related to BMI among individuals aged 45 to 64 years. The annual mean costs among the overweight, obesity, severe obesity and very severe obesity patients were 76, 159, 237, and 310 euro higher, respectively, than in normo-weight individuals, which translates in a cost increase of about 18%, 38%, 57% and 75% for each BMI category, respectively (third panel of Table 2). In contrast, total costs were more strongly related to BMI among individuals aged 65+ years. The annual differential mean costs among the patients with overweight, obesity, severe obesity and very severe obesity were 75, 302, 719, and 790 euro higher, respectively, than in normal weight individuals (fourth panel of Table 3). Total overall (direct and indirect) costs were also significantly higher in underweight individuals than in normal weight individuals in all subsamples (Table 3). In particular, annual mean total costs among all underweight individuals were 138 euro higher than in the normo-weight subjects, which translates in a cost increase of 13% (first panel of Table 3).

Indirect and direct costs

The share of indirect costs within overall outpatient costs was the largest in overweight and obese men and women aged 45-64 years (Table 2). Moreover, the indirect costs of the underweight subjects were lower than those of the normo-weight individuals for each age group, and this differential was more pronounced in the elderly, amounting to 11% (Table 2).

In terms of total outpatient and inpatient health expenditure, direct costs were negative in the overweight and obesity groups, suggesting that after correcting for BMI-related pathologies, these patients on average had lower health care expenditures than normo-weight individuals. This direct cost differential turned positive for the category with severe and very severe obesity, and was particularly pronounced in the elderly (Table 3). Moreover, the total indirect costs in the underweight individuals were significantly lower than in the normoweight subjects, and this differential was particularly high in the elderly (fourth panel of Table 3). Finally, the total direct costs in the underweight men and women were substantially higher than in the elderly normo-weight individuals.

Relative effects of BMI and adiposity-associated comorbidities on health care costs

As shown in figure 1, based on SURE analysis, the indirect costs of excess body weight in individuals with overweight and obesity explain the great majority of overall BMI-related costs both in terms of outpatient and total health care expenditures. Moreover, as shown in figure 2, when hypertension, diabetes and CVD were accounted for, the age-adjusted relation between BMI and total and outpatient health care costs was to a large extent eliminated.

Strengths and limitations

It is important to highlight the strengths and limitations of this study. The use of clinical data (e.g. weights, heights, chronic disease diagnosis, test results, drug prescriptions, outpatient diagnostic tests, specialist visits and hospital admissions), collected for a large sample of patients and entered in an up-to-date computer-based database by trained GPs, is a major strength of this study. In addition, it is important to stress that the Italian National Health Service is a public and universalistic system, which provides substantially free health services for all citizens. According to OECD Health Data, in 2012 about 87% of medical services in Italy was publicly financed. This setting favours the external validity of the study and minimizes the selection problems related to presence of private insurance plans. Finally, the use of a multi-equation recursive model (SURE) to calculate the relative effects of BMI and adiposity-associated chronic diseases on health care costs represents another strength of this study, since it provides better and more efficient estimates of the cost of obesity and allows to study direct and indirect effects of BMI of health care expenditure. One major limitation of this study is the lack of alternative measures of adiposity, such as waist circumference or waist-hip ratio. Moreover, the analysis could potentially benefit by introducing covariates such as smoking or socio-economic status, enhancing the precision of the coefficient estimates, which are not available within this dataset. Finally, the estimates might suffer from a downward bias, as the control group includes individuals who, while not being affected by the diseases included in this analysis, may have had other pathologies, which may increase health expenditure with respect to the overall population.

Sources of data

The present analysis was based on data of 557,145 men and women, older than 18 years, who were observed at different times between 2004 and 2010, amounting to a total of 2,705,211 observations. The anthropometric and clinical data have been extracted from the Health Search/CSD Patient Database (HS), an Italian general practice registry that includes data obtained from computer-based patient records of a selected group of 700 general practitioners (GP), homogeneously distributed across all Italian regions, covering a patient population of over a 1.8 million between 2004-2010. The GPs voluntarily agreed to collect patient information and to attend specific training courses for data entry [13]. The HS database contains patient demographic data that are linked through the use of an encrypted patient code with their medical records (diagnoses, prescribed tests results), drug prescription information (medication name, date of filled prescription, and number of days' supply), hospital admissions, and date of death. To be considered for participation in epidemiological studies, GPs should meet “up-to-standard” quality criteria pertaining to the levels of coding, prevalence of well-known diseases, mortality rates, and years of recording [13]. The HS database complies with the European Union guidelines on the use of medical data for research, and has previously been demonstrated to be a valid data source for scientific research [14-16]. GPs collected this information on daily basis. However, in this study the records have been collapsed to obtain yearly aggregates. Finally, these data have been merged, at patient level, with data from the SiSSi (Simulazione Spesa Sanitaria Italiana – Simulation of the Italian Health Care Expenditure) project, which includes information on prices and tariffs for drugs, outpatient visits, diagnostic tests and hospitalization visits. By multiplying health care service utilization data from the HS with price and tariff data from the SiSSi project we obtained detailed information on public health care expenditure at the patient level [17-19].

Model structure

Obesity-related health care expenditure is generally studied with a cost-regression approach [20-30]. However, this method is likely to produce biased estimates, since it does not analyze the mechanisms at play between BMI and health care expenditures. Therefore, we have built a multi-equation recursive model within a Seemingly Unrelated Regression Estimator (SURE) approach to quantify: (1) indirect costs, resulting from the impact that BMI has on a set of chronic diseases commonly attributable to obesity, which in turn influence health expenditure; and (2) direct costs, defined as the residual impact of BMI independent of adiposity-associated chronic diseases, after adjusting for age, sex, geographic residence, and GPs heterogeneity. The statistical model consisted of eight equations, estimated simultaneously, of which the first 7 represent the sets of obesity-related chronic diseases (1) type 2 diabetes mellitus; (2) hypertension; (3) dyslipidemia; (4) cardiovascular diseases: coronary artery disease, ischemic stroke, congestive heart failure; (5) cancers: breast, colon, and prostate cancer, (6) arthrosis: hip and knee arthrosis; (7) depression, while the last equation modeled the health care expenditure as a function of the above chronic conditions plus BMI.

Dependent variables

The empirical model comprised 8 equations, and hence there were 8 dependent variables: 7 dummy variables indicating the presence of one of the selected diseases at patient level plus the per-capita health care expenditure variable. We used two definitions of health expenditure: “outpatient” health care expenditure and “total” health care expenditure (i.e. inpatient and outpatient medical costs). Precise data on health care expenditure were obtained from a merge of Health Search CSD-LPD database with data on costs and tariffs of drugs, outpatient visits, diagnostic tests, and hospitalizations from the SiSSI Project. Furthermore, since the Health Search CSD-LPD PCP's may underreport the events of hospital admissions, we reweighted the Health Search CSD-LPD hospitalization rates with the hospital admission data collected by the Italian National Institute of Statistics (ISTAT). The weighting procedure stratifies the patient population by region, age (18-44, 45-64 and 65 and older), gender and BMI level in order to match cell groups of individuals in the Health Search CSD-LPD and ISTAT datasets. Subsequently, average hospitalization costs for the same cell groups were computed from the Hospital Discharge dataset of the Italian Ministry of Health and assigned to the Health Search CSD-LPD population.

Covariates

BMI was calculated as weight in kilograms divided by the square of height in meters. Study participants were divided in 6 groups with: normal weight (BMI, 18.5-24.99 [the reference group]), underweight (BMI between 15 and 18.49), overweight (BMI, 25.0-29.99), obesity (BMI, 30.0-34.99), severe obesity (BMI, 35.0-39.99) or very severe obesity (BMI >40.0). Since health expenditure may differ due to GPs practice heterogeneity, we controlled for GPs behavior, including variables such as the average number of patient contacts, the average number of prevention visits, the number of overall patients registered with the GPs practice and the average unit price of the prescribed drugs. Finally, we accounted for patient heterogeneity by controlling for age, gender, and region of residence.

Statistical analysis

In econometrics, the seemingly unrelated regression equations (SURE) model is a generalization of a linear regression model that consists of several regression equations, each having its own dependent variable and potentially different sets of exogenous explanatory variables [20]. Analytically, the structure of a SURE model employed in this analysis can be represented by the following 8 equation model: $P_{jit}={\alpha }_{j0}+B_{j}BM{I}_{it-1}+{\delta }_j{X}_{it}+{\theta }_{j}G{P}_{it}+{\epsilon }_{jit}$$H{E}_{it}={\theta }_0+{\Sigma }_{j=1}^7{\lambda }_j{P}_{jit}+\pi BM{I}_{it-1}+\rho X_{it}+\sigma G{P}_{it}+{\eta }_{it}$ where j = 1,..,7 represents the sets of pathologies, i stands for individuals, t =2004,...,2010 for years, and where P is the set of seven pathologies' indicator dummy variables, HE denotes health expenditure in euro per year, BMI is the vector of body weight dummy variables categories observed in the year preceding the diagnosis of pathologies and health expenditure realization; X is a set of control variables including age, gender, region of residence and year time dummies; and GP is a vector of variables including physician characteristics. Finally, εjit and ηit are idiosincratic error terms. The assumption of the model is that error terms are independent across time, but may have cross-equation contemporaneous correlations. Thus we assume that E[ε_ir ε_is | X] = 0 whenever r ≠ s, whereas E[ε_ir ε_jr | X] = σ_ij. Although Ordinary Least Squares (OLS) estimate of our model parameters are consistent, generally they are not as efficient as the SURE method, which uses the feasible generalized least squares (FGLS) method with a specific form of the variance-covariance matrix. The analysis was conducted separately for the whole sample of patients as well as for subsamples by age classes (18-44 year olds, 45-64 year olds and 65+ year olds). The coefficient estimates were expressed as marginal effects. For the full sample, the reference category was a male, resident of the Piedmont region in North Italy, in the normal weight BMI category, not affected by any of the pathologies considered in the study, observed for the first time in 2004 and aged between 18 and 44. This last condition was not considered when dealing with the subsamples by ages. The parameter estimates described the impact of each covariate on individual health care expenditure. The statistical analysis was performed using Stata software version 13 (Stata Corps).