This gave rise to the “partially adjusted” method, which is widely used in recent studies (Allison et al., 1999a; Mokdad et al., 2004). Using Levin’s formula for unadjusted relative risks, this approach calculates the relative risk adjusted for people in the subgroup. A single attributable fraction is obtained by apply this relative risk estimate to the prevalence of exposure of the entire population. The advantage of this approach is that it does not require the number of deaths in each subgroup to be known. However, another problem arises. Since Levin’s formula is only appropriate for unadjusted relative risks, the use of adjusted relative risk would lead to a biased estimate when there is confounding (Rockhill, 1998). The result is usually an overestimate, with the magnitude depending on the degree of confounding. Flegal and colleagues (2004a) conducted a study examining the effect of confounding by age and sex. The result was an overestimate as high as 17%. Besides, this method does not take effect modification into account. As the derivation cohort may differ from the entire population for different reasons such as the prevalence of obesity and the age structure (Flegal, 2005), further bias could occur. To tackle this problem, Graubard (2005) suggested that, in calculating attributable risk, information source on the prevalence of obesity should be collected from national survey data to ensure that it is representative for the entire population. Possible interactions between lifestyle factors and the confounders must also be considered (Goodman, 2005). Nonetheless, the “partially adjusted” method remains the most commonly used method due to its practicability.
Review and Discussion
It was not until after CDC published the 400,000 figure that Allison’s method began to draw questions. Flegal and colleagues (2004a) examined the methodology employed by CDC and found several debatable points. They found that even within the epidemiological approach, different data collection and calculation methods in the process could have a large effect on the outcome and, hence, the economic cost estimates of obesity. In the following paragraphs, different calculation factors and variables were discussed
When studying the impact of obesity, one must first ask the following question – What defines obesity? The three main methods of determining obesity are body mass index (BMI), weight-for-height chart, and measurement of body fat. Although body fat measurement is the most accurate for obesity, such data for the entire population is not likely to be obtainable. Some studies have used weight-for-height categories (Lew and Garfinkel, 1979, as cited in Banegas et al., 2003, p.203). However, self-reported weight and height often underestimate the prevalence of obesity compared to measured data (Bertakis and Azarit, 2005). Up to now, the BMI index is still the most accepted method because it is simple, quick, and cost effective; but even using BMI index, variability in the cutoff criteria would have a large effect on the overall cost estimate. Since different countries might have different BMI categories, the cutoff criteria for obesity would also be different. For instance, Banegas et al. (2003) and Finkelstein et al. (2004) used BMI ≥ 30kg/m2 to define obese while Allison et al. (1999b) used BMI ≥ 29kg/m2; The American Institute of Nutrition (1994, as cited in Wolf and Colditz, 1998, p.103) even recommended a cutoff value of BMI ≥ 25kg/m2 due to increased health risks in their country seen at that level. Wolf and Colditz (1998) found that the difference of applying BMI ≥ 27kg/m2 and BMI ≥ 30kg/m2 to the population in France would yield an enormous difference of $6.2 billion French francs. Furthermore, BMI measure does not consider frame size, body fat percentage, or muscle mass. People who are very masculine could be put into the obese category. Martin (2005) reported that half of all National Football League (NFL) in America would be counted as obese. As a result, if any of them were to die in a car accident, his death would be counted as attributable to obesity in CDC’s study. Since all estimates are inherently influenced by the BMI criteria, difference in BMI standards became one of the main problems in obtaining correct estimations.
The next key factor is data collection. Prevalence of obesity could be estimated by looking at a large population sample. One example of data source is the National Health and Nutrition Examination Survey (NHANES). Detailed health information, including measurement of height and weight, was collected by CDC from the U.S. population. The data were then followed up from time to time to find out whether survey participants were still living. Three versions of the NHANES included NHANES I (1971-1975), NHANES II (1976-1980), and NHANES III (1988-1994). Allison and colleagues (1999a) relied on NHANES III to estimate the prevalence of obesity. In obtaining the mortality risk of obesity, they used six population studies. Among these six studies, however, only the NHANES I was included. The average start date of these studies was 1963 and the average end date was 1983, which meant that the health status of the US population at the time of the study was not accurately reflected. They did not take into account the improvements in medical care and the overall ability to treat obesity related diseases throughout the years. In fact, Flegal et al. (2004b) reported a much lower risk of obesity when using data from 2002. In calculating the mortality risk of obesity, Flegal et al. (2004b) employed data from the NHANES series, which included NHANES I, II, and III. They then applied the relative risk to the up to date population profile obtained from NHANES 1999-2002. Apart from the decline in the risk of obesity, they also found no increased risk among people with a BMI under 35 in the most recent data. Therefore, the data used by Allison et al. (1999a) was not representative for the current target population. This possibly led to a significant overestimation of the economic cost due to obesity. In addition, the cohort group might also differ from the target population in terms of gender, age, social class, and demographics. For instance, there was found to be an over-representative of whites and Americans from upper socioeconomic classes due to data collection procedures. Both NHANES I and II excluded people who were older than a certain age, hospitalized, or in nursing homes (Flegal et al., 2004b). As a result, the population was not truly represented and the result was inaccurate.
Another issue is alternative causes of death in the obese population. In CDC’s investigation, validity of the study is undermined by the assumption that all excess mortality in obese people was purely caused by obesity (Martin, 2005). It failed to acknowledge that obesity may have served as a marker for other conditions that increase the risk of death, such as cardiovascular disease, cancer, and smoking. Adjustment must be made for these factors as they may be related to an alteration in risk of death. Failing to define factors associated with obesity and factors caused by obesity would lead to incorrect calculation of the cost estimate. Some have attempted to exclude participants with specific health conditions at baseline as well as smokers in order to counter this problem. It was believed that people who are excluded during this process were the ones at highest risk of death and were likely to die of factors unrelated to obesity. However, Flegal et al.(2004b) argued that this exclusion would make the mortality experience of the group different from the population, leading to a bias assuming an increase in the risk of death due to obesity. The relative proportions of the elderly, the death rates of the nonobese, as well as the prevalence of obesity would also be altered. Hence, different exclusion categories would give different cost estimates.
Estimates of deaths attributable to obesity are very sensitive to the precision of relative risk estimates in the elderly (Flegal et al., 2004b). This is because of the large number of deaths among the elderly. Researchers must therefore be aware of the fact that relative risk of mortality associated with obesity declines with age (Bender et al., 1999). Flegal et al. (2004b) even discovered potential protective effects of obesity to the elderly, possibly due to greater nutritional reserves in times of stress, lower risks of injury from falls, and lower rates of osteoporosis. If one is to restrict attention to people under the age of 70, for instance, the estimated deaths caused by obesity would be significantly smaller. Researchers must recognize the under-representation of the aged people in data source used as well as the change in proportion of elderly in the society. Flegal et al. (2004b) argued that failure in doing so have become common errors in many past studies. In order to adequately account for the different effects of age on mortality relative risk of obesity, age must be considered as a confounder or an effect modifier. It must be noted that a little error in the relative risk estimates could cause a 10-fold difference in the overall cost estimate. Flegal et al. (2004a) proved that a lower proportion of the elderly in the sample together with an overestimation of relative risk by 0.1 could lead to a notable 83% overestimation of deaths attributable to obesity. Researchers should, therefore, carry out sensitive analysis of relative risk of obesity in the elderly.
Additional factors such as life expectancy and social-economic factors could also lead to variations in cost estimates. The overall life expectancy in a country could be constantly changing. This would have an effect on the relative risk value. As Flegal et al. (2004a) suggested, a variation of 0.1 or 0.2 in the RR estimate could produce a very large bias in the overall estimated mortality attributable to obesity. Therefore, relative risk data should be obtained from a recent study that is large, well analyzed, and based on a national sample (Allison et al., 1999). Mortality of obese people could be deceptively high as overweight people are more likely to be inactive and of low socio-economic status. Hence, one must find out whether additional factors such as physical inactivity, lifestyle, weight fluctuation, poor diet, use of weight loss drugs, and inadequate access to health care services could have justified some of the excess mortality in obese people.
It was shown that the prevalence based approach remains the most feasible way in calculating deaths attributable to obesity. It helps to provide important information on the expenditures associated with obesity. However, one must keep in mind that this approach generates estimates only for a given period of time, usually a year. This means that it does not quantify or account for long term prevalence and consequences of obesity, from both health and economic point of view. Allison et al. (1999b) argued that quantitative cost estimates would be useful for health care policymakers to make appropriate policy decisions, allocate funds, and tackle health problems faced by a country. However, considering the complexity and a range of different calculation methods, uncertainty about the estimate value would make it difficult to use for public health purposes. Policymaker, public health groups, and the media must be careful when using these estimates. Quite often, exaggerations were found in those figures due to an attempt by researchers to raise profiles of the problem. If one is to consider the number of deaths caused by starvation, the problem of obesity would appear to be relative less significant. To claim it as a crisis that would soon overtake tobacco as the number one cause of preventable death seemed to be an overstatement. Nonetheless, it is a problem that the government must recognize as it increases the risk of a number of diseases, leading to increased treatment cost and other medical expenses. Up to date, scientists have yet to come up with a calculation method that gives a level of accuracy that policymakers could rely on. It is important to keep in mind that estimates by the PAF are generated by statistics, not science. As shown previously, little statistical flaws, which could be hard to notice, would largely affect the results. With the “partially adjusted” method likely to produce an overestimate, errors should be limited by controlling for confounding, employing up to date data, using population representative health information and BMI category, accounting for lower relative risk for the elderly, and recognizing additional factors such as life expectancy and social-economic issues. Future studies should aim to discover method to determine whether a certain disease is caused by obesity. Efforts should be put into better defining BMI cutoff criteria for different countries. More precise age-specific estimates of mortality relative risks for the elderly would also help to improve the accuracy of the overall estimate. As well, methods of calculating deaths attributable to obesity should be formulated to allow for variation in relative risks with age. Until these limitations were solved, government should only take such estimates as references rather than relying on them in the policy making process.
Conclusion
In conclusion, obesity appeared to be less damaging than what was reported by the CDC. Nonetheless, it is a significant health problem faced by a lot of countries. This paper has presented different methods used in estimating deaths attributable to obesity and how different calculations would affect the estimates. Due to feasibility, statistical estimation via population attributable fraction has been mostly used in past researches. Existing estimates using the “partially adjusted” method were likely to be biased due to flaws and limitations in the calculation process. Estimations from this approach should therefore be taken carefully. It was shown that the precision of estimates of deaths attributable to obesity depends on the estimates of relative risk; a small range of relative risk estimates could result in a 10-fold difference in the overall estimate. To obtain accurate estimates would require a lot more complex calculation method, one that is yet to be devised given present knowledge about the epidemiology of obesity. Corresponding to recommendations by Flegal et al. (2004a; 2004b), though, the current estimation could be enhanced by using up to date data representative for the target population, controlling for confounding, allowing for more categories of BMI, and employing more precise age-specific estimates of morality relative risks for the elderly.
References
Allison, D.B., Fontaine, K.R., Manson, J.E., Stevens, J., and Vanltallie, T.B. (1999a). Annual deaths attributable to obesity in the United States. Journal of the Amercian Medical Association, 282(16), 1530-1538.
Allison, D.B., Zannolli, R., and Narayan, V. (1999b). The Direct Health Care Costs of Obesity in the United States. American Journal of Public Health, 89(8), 1194-1199.
Banegas, J.R., Lopez-Garcia, E., Gutierrez-Fisac, J.L., Cuallar-Castillon, P., and Rodriguez-Artalejo, F. (2003). A simple estimate of mortality attributable to excess weight in the European Union. European Journal of Clinical Nutrition, 57, 201-208.
BBC (2004). The Science of Obesity, 22 November 2004, http://www.bbc.co.uk/print/science/hottopics/obesity/print.shtml.
Bender, R., Jockel, K.H., Trautner, C., Spraul, M., and Berger, M. (1999). Effect of age on excess mortality in obesity. Journal of the American Medical Association, 281, 1498-1504.
Bertakis, K.D. and Azarit, R. (2005). Obesity and the Use of Health Care Services. Obesity Research, 13(2), 372-379.
Boyko, E.J. (2005). Progress in the Estimation of Mortality due to Diabetes. Diabetes Care, 28, 2320-2321.
Finkelstein, E.A., Fiebelkorn, I.C., and Wang, G. (2004). State-Level Estimates of Annual Medical Expenditures Attributable to Obesity. Obesity Research, 12(1), 18-24.
Flegal, K.M., Graubard, B.I., and Williamson, D.F. (2004a). Methods of Calculating Deaths Attributable to Obesity. American Journal of Epidemiology, 160(4), 331-338.
Flegal, K.M., Williamson, D.F., Pamuk, E.R., and Rosenberg, H.M. (2004b). Estimating Deaths Attributable to Obesity in the United States. American Journal of Public Health, 94(9), 1486-1489.
Flegal, K.M. (2005). Methodological Issues when Estimating Lifestyle Factors. In Estimating the Contributions of Lifestyle-Related Factors to Preventable Death: A Workshop Summary (edited by H. Fineberg), pp.11-12. Washington: The National Academies Press.
Goodman, S. (2005). Attributable Risk in Epidemiology: Interpreting and Calculating Population Attributable Fractions. In Estimating the Contributions of Lifestyle-Related Factors to Preventable Death: A Workshop Summary (edited by H. Fineberg), pp.7-10. Washington: The National Academies Press.
Graubard, B. (2005). Calculating the Number of Deaths Attributable to a Risk Factor Using National Survey Data. In Estimating the Contributions of Lifestyle-Related Factors to Preventable Death: A Workshop Summary (edited by H. Fineberg), pp.12-13. Washington: The National Academies Press.
Levin, M.L. (1953). The occurrence of lung cancer in man. Acta Unio Internationalis Contra Cancrum, 19, 531-541.
Longley, R. (2006). Obesity Target of Renewed HHS Attack, Oct/06, http://usgovinfo.about.com/cs/healthmedical/a/hhsobesity.htm
Martin, M. (2005). Obesity Scandal, Sept 06, http://www.heartland.org/PrinterFriendly.cfm?theType=artId&theID=16860
Mokdad, A.H., Marks, J.S., Stroup, D.S., Gerberding, J.L. (2004). Actual causes of death in the United States, 2000. Journal of the American Medical Association, 291(10), 1238-1245.
Rockhill, B. (1998). Use and Misuse of Population Attributable Fractions. American Journal of Public Health, 88(1), 15-19.
Wolf, A.M. and Colditz, G.A. (1998). Current Estimates of the Economic Cost of Obesity in the United States. Obesity Research, 6(2), 97-106.