Diabetes prevalence model
The model for diabetes prevalence used in this study was created using data from the NHANES III (1988–1994), and then fitted to data from the NHANES 1999–2002 as a validity check of the accuracy of the model’s projections. The resulting model was then used to project the number of individuals with diabetes in the US in 10-year increments into the future. We evaluated 10-year age classes at each 10-year interval. Our model has the following components:
Number of individuals with diabetesTime 2 = ∑ (number of individuals with diabetesTime 1i
+ incident cases
), where i equals each 10-year age group, and incident cases consist of: (1) persons converting from a disease-free state to having diabetes; (2) diabetic patients immigrating to the United States; and (3) persons with diabetes moving into the 20 to 29-year-old age class.
The percentage of persons with diabetes, which is calculated thus:
percentage diabetesTime 2 = (number of individuals with diabetesTime 2 / total populationTime 2) × 100.
The estimate of future diabetes is therefore based on this equation, including the number of individuals with diabetes in the previous time period, conversion to diabetes, migration, and mortality, rather than being a linear extrapolation of the change in diabetes prevalence from the known values of 1991 and 2001.
The NHANES is a programme of surveys conducted by the National Center for Health Statistics and designed to assess the health and nutritional status of adults and children in the United States. The survey is unique on a national level in that it combines interviews and physical examinations. The NHANES uses a complex multistage sampling design, making it representative of the non-institutionalised US population and allowing weighted estimates to be computed.
For this study we used several of the NHANES data sets. Specifically, we used the NHANES III (1988–1994) (unweighted n = 4,950) and the NHANES 1999–2002 (unweighted n = 3,804) to estimate among individuals of 20 years of age and older the prevalence of diagnosed diabetes, the total diabetes burden (diagnosed and undiagnosed diabetes), and the proportion of the population at risk of developing diabetes. Since mortality within the population affects future prevalence , we also used the cohort from the NHANES II mortality survey (1976–1992) (unweighted n = 3,916) to provide estimates of diabetes mortality. Computation of all analyses using the NHANES data sets to provide nationally representative estimates for the models was designed to account for the complex survey design and the appropriate sample weights. All analyses were conducted using SUDAAN software (Research Triangle Institute, Research Triangle Park, NC, USA).
Variables used in models
Prevalence of diabetes
Diabetes burden was assessed as diagnosed diabetes plus undiagnosed diabetes. Because of the substantial proportion of people with undetected diabetes, we focused on this formula for total diabetes, rather than using diagnosed diabetes to indicate diabetes burden in the population. Moreover, by focusing on diabetes as diagnosed and undiagnosed disease, we minimised the possible impact on future diabetes prevalence of changes in screening practices for diagnosing diabetes during an ensuing time period.
Diagnosed diabetes was assessed as individuals who answered yes to a question of whether a doctor had told them they had diabetes. Undiagnosed diabetes was estimated on the basis of individuals who said they had not had a previous diagnosis of diabetes, but who had fasting plasma glucose (FPG) > 7.0 mmol/l. Although, the diagnostic criteria for diabetes during the time between the NHANES II and the NHANES III changed from FPG > 7.8 mmol/l to FPG > 7.0 mmol/l, we used the newer criteria to gain an awareness of the total diabetes burden at each point in time using the same criteria .
Persons converting from a disease-free state to having diabetes
Although a variety of diabetes risk scores exist, most have been created from cross-sectional studies and have as their aim the identification of individuals with undiagnosed diabetes. Their ability, therefore, to make predictions on development of diabetes is unknown [13, 14, 16]. The risk score used in this study is based on one developed for the Atherosclerosis Risk in Communities (ARIC) cohort study . Among individuals without diagnosed diabetes or FPG > 7.0 mmol/l, we used a scoring strategy which includes: high waist circumference (>102 cm in men, >88 cm for women), raised blood pressure (>130/85 mmHg or antihypertensive medications), low HDL-cholesterol (<1.03 mmol/l for men, <1.29 mmol/l for women), high triacylglycerol (>1.7 mmol/l), BMI > 30 kg/m2, and hyperglycaemia. Each of the characteristics is worth 1 point except for hyperglycaemia, which can be worth 2 points if FPG is > 5.6 mmol/l or 5 points when FPG is >6.1 mmol/l. A score of >4 puts an individual at high risk of developing diabetes, whether diagnosed or undiagnosed. A score of <4 indicates that a person has a low risk of developing diabetes.
This particular risk score was chosen for several reasons. First, it has moderate sensitivity (68%) and specificity (75%). Second, it is computed in a reasonably straightforward manner without having to use coefficients from the ARIC cohort that may be specific to that cohort. Third, data and results provided in the study by Schmidt et al.  allowed for computation of the rate of development of diabetes in both the high-risk group and the low-risk group. The ratio of development of diabetes in the high-risk group versus the low-risk group was 4.5:1. Variables needed to compute this diabetes risk score are available only in the NHANES III and the NHANES 1999–2002.
Although the ARIC diabetes risk score did not specifically consider race or age in the computation , we computed conversion rates for 10-year age classes for three race/ethnic groups (non-Hispanic Whites, non-Hispanic Blacks and Hispanic individuals) by fitting age categories for the data from 1991 to 2001 and then fitting race/ethnicity on to the same time change. We did not compute specific sex-specific conversion rates because sex was already differentiated in several of the variables in the ARIC diabetes risk score .
Migration of persons with diabetes
Migration of individuals with or without diabetes into the population can also affect future diabetes prevalence. Recent projections have included migration within their models . Because we are looking at changes in diabetes prevalence among adults, migration of adults, particularly from ethnic minorities, could substantially affect the 10-year projections. We used data from the NHANES III to estimate migration of persons with diabetes in the 20 years and older age groups. The NHANES III measured how many years foreign-born immigrants had been in the US. Thus, we estimated the number of foreign-born individuals who had been in the country for 9 years or less for the total population as well as for different racial/ethnic groups. The NHANES III data allowed us to make estimates for non-Hispanic Whites, non-Hispanic Blacks and Hispanic individuals.
Persons with diabetes moving into the 20 to 29-year-old age class
For 2011, 2021 and 2031 the total number of persons with diabetes in the 20 to 29-year-old age class was estimated using a linear projection of the NHANES III and NHANES 1999–2002 data. The proportion of 20 to 29-year-olds with diabetes in each race/ethnic group was held constant at the proportions found in the NHANES 1999–2002 data at the later time intervals.
Mortality among individuals with diabetes
Diabetes mortality for the total population was based on data from the NHANES II mortality survey (1976–1992). This population-based cohort study was used to provide estimates of diabetes mortality, since mortality within the population affects future prevalence . Diabetes mortality was estimated as all-cause mortality among individuals with diabetes (either diagnosed or undiagnosed) at baseline, rather than as mortality with diabetes listed as the cause of death. This definition is more consistent with the potential impact of diabetes on future prevalence. Mortality estimates were computed separately for the total population by age classes.
The NHANES II mortality cohort is based on a sample of individuals aged 30 to 75, whereas we made diabetes estimates on individuals aged 20 years and older. Consequently, we assumed no deaths due to diabetes in the 20 to 29-year-old age group over the 10-year period.
Total population of 10-year age classes was estimated using data from NHANES III for 1991, NHANES 1999–2002 for 2001, and US Census Bureau, Middle Series projections for 2011, 2021 and 2031 . Total population of race/ethnic groups was also determined by 10-year age classes using the same sources of information.
In an effort to provide an estimate of future trends in diabetes and the population at high risk of developing diabetes, we employed the following procedure. We used the NHANES III data to fit a model to predict total diabetes in the NHANES 1999–2002. We used this strategy prior to making future projections, because it allowed us to develop and fit the model to an existing national estimate of diabetes prevalence. Because both the NHANES III and the NHANES 1999–2002 are based on multi-year data collection, we estimated a mid-point of 1991 and 2001 for the two surveys.
The number of persons with diabetes 10 years post-baseline was calculated for 10-year age classes by first adding baseline prevalence and incidence (the number of low-risk and number of high-risk persons who developed diabetes over the 10-year interval), then adding persons with diabetes who immigrated to the United States, and persons with diabetes who moved into the 20 to 29-year-old age class, and finally subtracting the number of diabetic subjects who died. Percentage of persons with diabetes was estimated for each time period by taking the total number of persons with diabetes and dividing by the expected total population, then multiplying by 100.
Varying model assumptions
Our initial predictions of future diabetes burden were based on the assumption of a constant proportion of individuals at high risk of diabetes at the levels present in the NHANES 1999–2002. To account for potential changes in the proportion of persons at high risk of diabetes, we also evaluated increases in the proportion of persons at high risk by 10, 20 and 30%, as well as estimates based on decreases in the proportion of persons at high risk by 10, 20 and 30%. Theoretically, it is unlikely that the proportion of persons at high risk will remain stable, because from NHANES III to NHANES 1999–2002 the proportion at high risk was seen to increase. Also, a major risk factor for diabetes, obesity, has increased substantially over a 40-year time period [10, 12]. We evaluated the effect of decreasing proportions at high risk, to account for the possibility that interventions to improve lifestyle of adults in the US may be effective.
In addition, to address the potential impact on mortality of healthcare interventions in management of diabetes, we examined potential reductions of 10, 20 and 30% in mortality among individuals with diabetes. Finally, we computed a model examining a combination of effects, assuming that lifestyle interventions would yield a 10% decrease of persons at high risk and healthcare interventions would yield a 10% decrease in mortality of persons with diabetes.