Our methods were similar to Wang et al
] A societal approach to costs was used as was a three percent annual discount rate. The flow chart in Figure outlines the approach. First, we predicted the number of obese adult cases averted, as described in more detail below. Then we estimated costs associated with obesity and quality adjusted life-years beyond the age of 40. Note that labor productivity costs, medical costs and QALY
s were relevant for cost-effectiveness ratios (CER
); labor productivity costs and medical costs were relevant for net benefits (NB
Let us first examine CER. The numerator of the CER is the cost of the intervention less the total medical costs due to obesity (which are averted due to the intervention). The medical costs are known as direct costs, and they would have been expected to have been incurred by society had the obese cases not been averted. In the denominator are total QALYs gained.
The CER formula is
where subscript i
indicates male and female, respectively. C
represents the costs of the CATCH intervention in 2004 dollars, Ni
represents the number of adult obese cases averted due to CATCH, Ai
represents the averted medical costs when obese adults aged 40–64, inclusive, are instead non-obese adults; Qi
represents the additional QALY
s gained when obese adults are instead non-obese. The denominator is the additional QALY
s accruing to averted obese adults due to the CATCH intervention. If the CER
is less than approximately US$30,000, then we can consider the intervention cost-effective [43
]. This is based on valuing a year of full human life at US$30,000. Other valuations of life-years are 10-fold this amount [46
Now let us define net benefits (NB). We subtracted the intervention costs from the total averted medical costs and productivity costs between age 40 and 64, inclusive, for an average obese adult in comparison to an average non-obese adult. The NB formula is
where subscript i = m, f indicates male and female, respectively. Bi represents the value of labor productivity gains for adults who have averted obesity.
In equations (1) and (2), Ni
is predicted from data from the obesity progression model, as described below [40
The intervention costs of CATCH
Intervention costs are given in Table . As is standard in economics, the value of the training time is the hourly wage. Wage and salary information for CATCH staff was suppressed for confidentiality. All wages are in 2004 US$.
Intervention Costs, 2004 US$
Note that as in Wang et al
., we excluded classroom time from the intervention cost [35
]. CATCH increases the effectiveness of PE and classroom time without taking additional time away from other activities.
Predicting adulthood obesity based on child overweight
The number of adult obese cases, defined as having a BMI > 30kg/m2, averted cannot be observed from the trial because it ends in the fifth grade. We used a lifetime obesity progression model to estimate averted adulthood obesity. The process is outlined in Figure .
Our lifetime obesity progression model is
where subscript i
again indicates male and female, respectively, and j
represent at-risk or overweight. Ni
was defined above and Hi
represents the number of children in the fifth grade trial schools in El Paso [40
are the proportions of at-risk and overweight children in grades three (the beginning of the trial) and five (the end of the trial) in the control schools; P1ij3
are the proportions of at-risk and overweight children in grades three and five in the intervention schools. P3ij
captures the probabilities of obesity at age 21 to 29 conditional on being at-risk and conditional on being obese at age 11; P4ij
measures the probabilities of obesity at age 21 to 29 conditional on being not at-risk and conditional on being not obese at age 11. P5i
is the probability of obesity at age 40 conditional on being obese at age 21 to 29; P6i
is the probability of obesity at age 40 conditional on not being obese at age 21 to 29.
Table lists the conditional probabilities needed in (3) in expanded form along with their sources.
Conditional Probabilities Needed for Predicting Adulthood Obesity
In order to estimate the probability of obesity at age 40 conditional on being obese during ages 21–29, we linked 1992, 1987, and 1982 NHANES I Epidemiologic Followup Study (NHEFS) data with the original 1975 National Health and Nutrition Examination Survey (NHANES) I data [47
]. For the 1975 data, BMI is available by sex and age. We kept those aged 25–29 from the 1975 data. Whichever follow-up dataset placed the subject closest to 40 was used. Those aged 28 and 29 in 1975 were linked to 1987 data (they were 40 and 41 then); those aged 25–27 in 1975 were linked to 1992 data (they were aged 42–44 then). The 'svy' facility of STATA 7.0© was used to account for the complex sampling design of NHANES. Note that Wang et al
. use the same technique, but for females only [35
Medical costs averted (direct costs)
As in Wang et al
., we used medical costs parameters from the literature [35
Wang et al
. used medical cost data for obese women between 40–64 years of age, inclusive, from Gorsky [48
]. However, unlike in the Planet Health trial Wang et al
. used, we predict male adult obesity cases will be averted. Therefore, we took medical costs from a study due to Oster et al
., which includes obese men and women [49
]. Oster et al
. used NHANES III [50
] to estimate the costs associated with hyper-tension, hypercholesterolemia, type 2 diabetes mellitus, cardiovascular disease, and stroke [49
]. The age period for averted medical costs was 35 years old until death rather than 40–64 years of age as we would have preferred. If the BMI score is in a category >32.5 kg
in Oster et al
., then we considered the person to be obese. Recall that our definition is based on BMI being greater than 30 kg
. However, this was as close to our definition as possible given the existing literature.
In order to ensure comparability with Wang et al
., we also considered NB
using parameters for medical costs 40–64 years of age, inclusive. from Gorsky et al
. (see Table ) [48
]. Because Gorsky et al
. only estimated medical costs for females, using their estimates necessitated substituting medical costs for females for males [48
Net Benefits (NB) and Cost-Effectiveness Ratio (CER) US$ Per QALY saved
Labor productivity costs (indirect costs)
Equations (5, 6, and 7) in the appendix were used to estimate labor productivity costs. In order to estimate labor productivity costs averted, we estimated the number of sick days missed per year by obese adults in comparison to non-obese adults for persons aged 40–64, inclusive, or from the age of 40 until the person turns 65 years of age. We used median wages to place values on the lost time due to obesity-related illnesses for persons aged 40–64, inclusive. We also estimated the number of lost sick days for the obese and the non-obese using Poisson regression. The model controlled for age, age 40–64, smoking status, Hispanic ethnicity, and gender.
In addition to increased sick days, obese adults also have reduced life expectancy. Therefore, to assume that people aged 40 will live and work until they turn 65 years old would be to over-estimate labor productivity losses averted because more obese 40 year olds will die before 65 than non-obese 40 year olds. Therefore, life expectancy and mortality for obese and non-obese 40-year olds who die before 65 were calculated. We also estimated the life expectancy for those alive at 40 who die before 65 by gender for obese adults and for non-obese adults.
In order to project lost work days, we used 2002 National Health Interview Survey (NHIS) data. Because of the complex sampling design of the NHIS data, we estimated the model with STATA 7.0©, again using the 'svy' feature. As seen in Table , we included overall costs of work-loss estimates and Hispanic costs of work-loss estimates.
Peeters et al
. created life tables for both men and women by obesity status based on Framingham data [51
]. Thus, we were able to project the life expectancy at 40 for an obese person conditional on dying before 65 years of age.
In order to place a value on the sick days averted in our net benefit analysis, we used U.S. Department of Labor, Bureau of Labor Statistics Current Population Survey data [52
]. The data are for full-time workers only above 25 years of age for all workers, above 16 years of age for Hispanics. The median wage data is reported by week only. Therefore, in order to estimate the daily wage, the weekly wage was divided by five; in order to calculated the yearly wage, the weekly wage was multiplied by 52.
Quality-Adjusted Life-Years (QALYs)
Equation (4) in the appendix was used to estimate QALY
s in our context are the additional quality-adjusted life-years gained through avoiding adult obesity. Activity scales were used in QALY
to weight, or quality-adjust, years of life that may be added due to the intervention based on questions regarding their activity limitations, if any, and perceived health status [53
]. In our study, we estimated scales using the Centers for Disease Control and Prevention's activity scale matrix using 2002 NHIS data. Depending on a person's answer to NHIS survey questions, a health state value is assigned ranging from 0.10 (limited with poor health) up to 1.00 (no limitation with excellent health).
NHIS survey questions on self-assessed health and activity limitations were used. We again used life tables due to Peeters et al
. to project the life expectancy at 40 for an obese person [51
In order to determine the extent to which our results are dependent on the parameters we used, sensitivity analysis was conducted for both overall parameters and with parameters for Hispanics. All 48 parameters used in the analysis in Tables and were included in the sensitivity analysis (the Hispanic parameters in the lower part of Table replace the corresponding parameters in the upper part of the table). In order to avoid the problems of the infinite support in the normal distribution, the triangular distribution, which has a finite support, was assumed. The support of the triangular distribution was created from the 95th percentile confidence intervals of our 48 parameters. We conducted 1,000 independent simulations trials. Each simulation trial draws were made for each of the 48 parameters simultaneously, and CER and NB calculated (see Table ). Separate simulations, using the same method as above, were conducted for each of the 48 parameters, holding the other 47 parameters constant.
Parameters Used in the Sensitivity Analysis†