CRELES (Costa Rican Longevity and Healthy Aging Study), a three-wave nationally representative longitudinal study, is the main source of data for this research (n = 2827). The baseline interview was conducted between 2004 and 2006, the second wave between 2006 and 2007, and the third wave between 2008 and 2009. Mortality was tracked up to October 31, 2017, by linking the CRELES dataset with the National Death Registry. More details on this survey have been previously published [44, 45].
Self-reports were used to define the diabetes status of individuals. The corresponding items in the questionnaire were: “Has a medical doctor ever told you that you have diabetes or high blood sugar levels?” and “How old were you when you were first told that you had diabetes”. To avoid the potential inclusion of individuals with type 1 diabetes, only individuals reporting a diagnosis at 30 years or older were classified as having the condition. This criterion has been used in other population studies [22, 23, 45, 49].
Sociodemographic variables included in regression models were education, income, sex and age. Education was a dichotomous variable for incomplete/complete primary school; complete primary was defined as six or more years of formal education. Income was a low/high dichotomous variable, with a cut-off point of 50,000 colones (Costa Rican currency) per individual per month. This was equivalent to approximately 100 United States Dollars (USD) during the wave 1 time period (2004–2006). Income was defined as the older person’s own income, if not married, or the couple’s mean monthly income, if married. This income cut-off has been used in similar studies with CRELES [8, 30].
Risk factors included in regression models were body mass index (BMI) and waist circumference, which are indicators of general and abdominal obesity, respectively. As for BMI, individuals were classified as underweight (< 18.5 kg/m2), normal weight (18.5–24.9), overweight (25.0–29.9), or obese (≥30.0 kg/m2) [56]. Waist circumference categories for men were normal (< 94 cm), increased risk of metabolic complications (94–101), and substantially increased risk of metabolic complications (≥102 cm). Waist circumference categories for women were normal (< 80 cm), increased risk of metabolic complications (80–87), and substantially increased risk of metabolic complications (≥88 cm) [56].
Behavioral health risks included in analyses were smoking, alcohol consumption, total daily energy intake, and regular physical activity. Individuals living with a smoking partner were classified as passive smokers, when they were not active smokers themselves. Smoking behavior refers to 100 or more cigarettes or cigars during participants’ lives. Categories were defined as never smoked, former passive or active smoker, current passive smoker, and current active smoker. Alcohol consumption refers to alcoholic beverages ever consumed during individuals’ lives. Categories were defined as never consumed alcohol, former consumer, and current alcohol consumer.
Total daily energy intake was estimated from a 10-min Food Frequency Questionnaire, with a 3000 kcal/day cut-off. This standard cut point is associated with differential risk of cardiovascular disease [9] and has been used in similar population studies [30, 38, 43]. Regular physical activity was defined as three or more days per week of exercise routines or other physical rigorous activities like sports, jogging, dancing, or heavy work during the 12 months preceding the baseline interview.
Geriatric syndromes included in the models were polypharmacy, functional dependency and geriatric depression. Polypharmacy was defined as 5 or more medications daily. Functional dependency was defined as limitations in 7 or more of the 14 activities of daily living (ADL) and the instrumental activities of daily living (IADL). ADL were crossing the bedroom from side to side, bathing, self-feeding, going to bed, toileting, nail trimming, walking, climbing stairs, pushing objects, and raising arms. IADL were cooking, handling money, shopping, and taking medications. The Geriatric Depression Scale (Short Form) which includes 15 items [47] was used to classify participants according to their depression level. Normal was defined as less than 6 depression symptoms; mild depression as 6 to 9 and depression as 10 to 15 depression symptoms.
Health condition variables included in the mortality analyses were previous diagnoses of diabetes, hypertension, dyslipidemia, cardiovascular disease, cancer, and lung disease. Blood pressure was measured twice during the interview, yielding two measures of systolic pressure and two of diastolic pressure. Following Méndez-Chacón et al. [30] individuals were classified as hypertensive if they either had a previous medical diagnosis; or had blood pressure of 140/90 or higher in three out of the four systolic and diastolic measures during the interview; or if they were taking antihypertensive medications.
Dyslipidemia was defined as having at least one of hypercholesterolemia or hypertriglyceridemia. Hypercholesterolemia was defined as Total/HDL ratio of 5.92 or greater. Hypertriglyceridemia was defined as 150 mg/dl or greater. Individuals who were not fasting when their blood sample was taken have missing information on these biomarkers. Cardiovascular disease was defined as the diagnosis of at least one of myocardial infarction, ischemic heart disease without infarction, or stroke.
Data analyses and estimations were conducted with STATA computer software [52]. The analyses include descriptive statistics, multiple regression models, longitudinal regression models, and Lee-Carter stochastic population projections.
Projection of diabetes prevalence in the elderly
To project the expected size of this diabetic elderly population, we first estimated its rates of prevalence, incidence and mortality. The main input for prevalence projections was the incidence and mortality rates computed for 5-year age groups ending in the open 95+ group. In the estimations of the prevalence of diabetes in the elderly, migration was not taken into consideration; only incidence and mortality were allowed to affect prevalence projections. These projections are therefore under the assumptions of null migration and constant incidence and mortality along time.
The size of the diabetic population was computed 5 years back from the year 2006 to the year 2001, and then 5 years back again from the year 2001 to 1996. Growth rates between these time-points were used to project the size of the diabetic population assuming linear growth. Projections were estimated with the following formula:
$$ {{\mathrm{N}}_{\mathrm{x}}}^{\mathrm{d}}=\left[{{\mathrm{N}}_{\mathrm{x}+\mathrm{t}}}^{\mathrm{d}}-\left({\mathrm{N}}_{\mathrm{x}}{\ast}_{\mathrm{t}}{\mathrm{d}}_{\mathrm{x}}\right)\right]/\left(1{-}_{\mathrm{t}}{\mathrm{d}}_{\mathrm{x}}{\ast}_{\mathrm{t}}{\mathrm{q}}_{\mathrm{x}}\right) $$
where:
Nx: Total population at age x
Nxd: Diabetic population at age x
Nxnd: Non-diabetic population at age x
tdx: Diabetes incidence rate for the population aged x to x + t
tqxd: Probability of dying for the diabetic population aged x to x + t
Diabetic population size for each age-group (Nxd) was estimated based on age-specific prevalence rates and the official total population size (Nx) in 2006 [12]. Prevalence rates for the elderly (60+) are this study’s computations from multiple regression models. For the younger adult population (30–59) prevalence rates are the five-year age-group national estimates reported by another study in Costa Rica [39].
Incidence rates in the above formulas (tdx) are this study’s computations from longitudinal regression models for ages 30+ using the respective information on the age of diagnosis (and assuming the absence of selection survival). Death rates for the diabetic population (tmxd) for ages 30–59 are assumed to be the same as all-cause mortality in the general population from the National Death Registry [12]. Death rates for the diabetic population for ages 60+ are this study’s estimates from longitudinal regression models of all-cause mortality in the diabetic elderly. The age-specific probabilities of dying for the diabetic population (tqx) were derived from the tmxd using the relations and separation factors in Coale and Demeny [14] West model life tables.
Based on our population projections, the doubling time of the diabetic elderly population was estimated. Doubling time refers to the number of years it takes for the diabetic elderly population to double in size in Costa Rica under the assumption of a constant linear growth rate. Doubling time of the diabetic elderly population was also estimated for six hypothetical scenarios. Each scenario was under the assumptions of a linear population growth and a constant age-pattern of the incidence as observed in CRELES. For each scenario age-specific incidence rates were allowed to increase or decrease by 25, 50, and 75%, respectively.
Impact of diabetes on future healthcare costs
Future costs of health care for the elderly population were estimated using this study projected prevalence of diabetes and previously published estimations of current individual use of healthcare services [45]. Hospitalizations and outpatient consultations were modeled using two-part models, a common approach in health economics [18]. Costs were recorded based on the mean volume of utilization of hospitalizations and outpatient visits over a calendar year to avoid bias due to seasonality in the patterns of use of these services.
The costs of health care services used in this research are from the provider’s perspective, reported by the Costa Rican Social Security Fund (CCSS, for its Spanish acronym). Monthly current costs are publicly available at http://www.ccss.sa.cr. These monthly costs for each health care service are estimated by the CCSS as the total expenses incurred in one specific service divided by the total production in the same service over 1 month. Costs are reported in this study in United States Dollars from the year 2011 (2011 USD). We assumed that current patterns of healthcare services utilization remain constant in the future both in the diabetic as in the non-diabetic elderly.
Impact of diabetes on future life expectancy
The Lee-Carter method for mortality forecasting was used to forecast all-cause-mortality in this study. The Lee-Carter (LC) procedure is a stochastic model. It, therefore, allows for the quantification of uncertainty in the estimates. This method has been widely adopted in the USA, and it has also been used in the G7 countries and Australia [5]. Modifications and extensions have also been proposed to the method [6, 28].
Historical mortality rates for the 70s decade was not used in the time-series for mortality forecasting purposes. A significant decrease in infant mortality occurred during the 60s and 70s in Costa Rica. Using the 70s-decade mortality rates as an input for the estimation would have implied that such drastic mortality declines would have a chance to repeat in the future, which is not reasonable. Historical data on mortality from 1980 to 2010 was used to forecast 25 years of mortality up to the year 2035. Forecasts were estimated for (1) all-cause mortality and for (2) non-diabetes-related mortality. Forecasts were estimated using the LCFIT software [51], which produces a set of forecasted rates as an output. Median mortality rates by age-group and calendar year, as well as their corresponding confidence intervals, were estimated from the set of forecasted rates.
Using forecasted mortality rates to estimate life tables, two approaches were used to project the impact of diabetes in terms of life lost. They were based on estimations of life expectancy at birth (e0) and life expectancy at age 60 (e60).
In our first approach, we estimated e0 and one hypothetical scenario. We forecasted the total life expectancy at birth, and the total life expectancy at birth that would result from all mortality causes except diabetes.
In our second approach, we estimated e60, a commonly used indicator of longevity. We forecasted the total life expectancy at age 60, and two hypothetical scenarios. Our first hypothetical e60 scenario was the total e60 that would result from all mortality causes except diabetes. Our second hypothetical e60 scenario was the total e60 that would result from removing diabetes mortality and adding diabetes-caused mortality from a longitudinal competing-risks model. In this competing-risks model, the mortality hazard was computed as a function of diabetes-caused mortality, and the competing event was mortality due to any other cause. Forecasts were estimated with the assumption that the observed pattern and level of diabetes-caused mortality remains constant in the future.
Years of life lost to diabetes were estimated as the difference between the total life expectancy at age 60, minus the life expectancy at age 60 that would result from removing diabetes mortality and adding diabetes-caused mortality from the longitudinal competing risks model.