Lifetime Medical Costs of Obesity: Prevention No Cure for Increasing Health Expenditure

Background Obesity is a major cause of morbidity and mortality and is associated with high medical expenditures. It has been suggested that obesity prevention could result in cost savings. The objective of this study was to estimate the annual and lifetime medical costs attributable to obesity, to compare those to similar costs attributable to smoking, and to discuss the implications for prevention. Methods and Findings With a simulation model, lifetime health-care costs were estimated for a cohort of obese people aged 20 y at baseline. To assess the impact of obesity, comparisons were made with similar cohorts of smokers and “healthy-living” persons (defined as nonsmokers with a body mass index between 18.5 and 25). Except for relative risk values, all input parameters of the simulation model were based on data from The Netherlands. In sensitivity analyses the effects of epidemiologic parameters and cost definitions were assessed. Until age 56 y, annual health expenditure was highest for obese people. At older ages, smokers incurred higher costs. Because of differences in life expectancy, however, lifetime health expenditure was highest among healthy-living people and lowest for smokers. Obese individuals held an intermediate position. Alternative values of epidemiologic parameters and cost definitions did not alter these conclusions. Conclusions Although effective obesity prevention leads to a decrease in costs of obesity-related diseases, this decrease is offset by cost increases due to diseases unrelated to obesity in life-years gained. Obesity prevention may be an important and cost-effective way of improving public health, but it is not a cure for increasing health expenditures.


Introduction
The RIVM Chronic Disease Model (CDM) is a state transition Markov-type simulation model that describes how morbidity and mortality for several chronic diseases change over time in the Dutch population as a result of changes in epidemiological risk factors 1 2 . In general, the state transition model is a suitable and accepted model to describe demographic / epidemiological processes 3 . Disease experts and modelers have cooperated in building and testing of the CDM and several studies with different applications of the model have already been published [4][5][6][7][8][9][10] . The CDM has been formulated as a set of time-continuous differential and is implemented in the software package Mathemetica.

Basic structure
In the CDM different states are defined for the risk factor classes (e.g. never smokers, current smokers and former smokers) and states for the diseases included in the model (e.g. stroke: yes or no). In the starting year of the simulation period all persons are distributed over these states. Then, in time steps of 1-year, persons move from one state to another. These transitions are governed by so-called transition rates. E.g., class transition rates between the BMI states 'normal weight' and 'overweight' govern the change of the BMI distribution in the population, incidence rates between the states 'without diabetes' and 'with diabetes' govern the disease prevalence rates, and mortality rates from the state 'alive' to 'deceased' govern the surviving population numbers. For our calculations, we did not take into account transitions between risk factor classes over time. Thus, all cohorts are closed in the sense that no transitions occur between risk factor classes over the life-time. The transition rate is assumed independent from the preceding states and depends only on the present state defined by risk factor class, disease state, sex and age. The disease incidence rates depend on the risk factor class, using relative risk values. E.g., 'overweight' persons have higher diabetes risks than persons with 'normal weight'. For non diseased, the mortality rates depend on risk factor class, e.g., obese persons have higher mortality risks than persons with a normal weight. The mortality rates also depend on the disease states, but are conditional hereon independent on the risk factor values. E.g., the excess mortality risks of people with diabetes compared to people without diabetes are equal for all BMI states.
We assumed that all risk factors that are distinguished are independently distributed, e.g. we assumed that the distribution of smoking independent from BMI. All disease incidence risks were made dependent on these risk factors by multiplying the baseline risk with relative risk values specified by risk factor class. Moreover, we assumed for some disease pairs independent effects of one disease on the other. E.g., people with diabetes have higher risks of myocardial infarction compared to people without diabetes, independently from overweight and the other risk factors.
The main model outcome variables are incidence, prevalence and mortality numbers, specified by disease, age and gender. For the calculation of lifetime health care costs of the different cohort we used the number of survivors and disease prevalence numbers of the different cohorts.

Input data
Smoking classes distinguished in the CDM are never smokers, current smokers and former smokers. Body weight is modeled in three classes using Body Mass Index (BMI) as indicator: 18.5<BMI<25 (normal weight), 25 BMI< 30 (overweight), BMI 30 (obesity). Table A1 displays the diseases modeled in the CDM that are related to BMI and/or smoking.  11 . For the selection of smoking related diseases we followed the report by the Surgeons' General 12 . For all diseases related to smoking and/or obesity modeled in the CDM, age and sex specific incidence, prevalence and mortality rates were estimated using a three state transition model [13][14][15] . The higher mortality risks of patients compared to diseasefree persons were calculated from published relative survival proportions for types of cancer, and from incidence and prevalence figures for other chronic diseases, using an Incidence, Prevalence, Mortality (IPM) model 16 . Data on risk factor class prevalence were obtained from representative national or regional surveys Risk factor prevalence rates for smoking are based on data of STIVORO 17 . For obesity, data from the annual POLS survey from Statistics Netherlands are used 18 . Relative risks on morbidity and mortality for smoking and obesity are based on several observational studies  . Relative risks of the three BMI classes were calculated in three steps. First, a quadratic function was estimated to describe the non-linear relation between BMI and all cause mortality relative risks for different studies. The parameters of these functions were then plotted against age to estimate an age gradient. In a third step, average relative risks for the three different BMI classes were computed using the BMI distribution within these classes in the Netherlands. For the current and former smoking classes distinguished in the CDM, data were used from studies that reported relative risks for all current and/or all former smokers specified by gender and age. To estimate health care costs for the different cohorts data of the Costs of Illness in the Netherlands study were used 66 . In that study the total direct health care costs in the Netherlands of 2003 are uniquely attributed to disease categories specified by gender and age classes. All input data were specified by gender and age (see Appendix B).

Mathematical model equations
The a Since, in this application, we did not take into account transitions between risk factor classes we will focus exclusively on the transitions between disease states.

Model initialization part
The parameters calculated here are the baseline disease incidence rates, i.e. the incidence rate values for normal weight never smokers, the mortality rates for other causes of death.
Mortality rates from Statistics Netherlands for the year 2004 are attributed to risk factor classes to derive mortality rates specified by risk factor class. Assuming independence between risk factor class prevalence rates and multiplicative relative risks (i.e. no interaction on log-linear scale) we can write mortality rates for the different cohorts as (for notational simplicity, age and sex indices have been omitted in the notation throughout the paper): Baseline disease incidence rates and risk factor class specific disease incidence rates are calculated in the same fashion as mortality rates:    (8) and (9) into equation (7). In a similar fashion, relative risks for other causes mortality of for overweight and obesity can be derived. Given RR( oc | s j, ) and RR(oc | b k, ) the baseline other cause mortality rate can be found:

Model simulation part
Risk factors and diseases are linked through relative risks of disease incidence for each risk factor. That is, incidence rates for each risk factor class are found as relative risks times baseline incidence rate. The general assumption used is that conditional on the risk factors included, the disease event rates are independent. For the 'healthy living cohort' relative risks equal one. Formula (12) denotes the change over time in the prevalence rate of disease d for a cohort, homogeneous in its risk factor class prevalence, as a function of relative risks, incidence and mortality rates: The CDM describes disease prevalence numbers for each disease separately and it is assumed that the disease-specific attributed mortality rates are additive. Given the relations between disease specific attributed mortality, other causes mortality, disease prevalence rates and relative risks we can describe the change in population numbers needed to estimate life expectancy: The difference of the mortality rates for persons with and without the disease can be interpreted as the excess mortality rate for that disease. However, in a model with multiple diseases these excess mortality rates cannot be interpreted as mortality uniquely attributable to a disease, since the excess mortality rates can also be caused by other co-morbid chronic diseases, e.g. coronary heart disease being a complication of diabetes. Therefore, in the calculation of the prevalence rates excess mortality rates are used, while in the calculation of the number of survivors disease specific attributed mortality rates are used.