Diabetes and all-cause mortality among middle-aged and older adults in China, England, Mexico, rural South Africa, and the USA: a population-based study of longitudinal aging cohorts

Study design and sample

We conducted a longitudinal analysis of population-based aging cohorts in five countries: China (China Health and Retirement Longitudinal Study (CHARLS)),9 England (English Longitudinal Study of Ageing (ELSA)),10 Mexico (Mexican Health and Aging Study (MHAS)),11 rural South Africa (Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa (HAALSI)),12 and the USA (Health and Retirement Study (HRS)).13 These cohorts are part of the HRS International Family of Studies, a network of longitudinal aging studies with similar sampling designs, eligibility, and assessment methods.14 The cohort inclusion criteria for this analysis were (1) availability of baseline and follow-up data from 2010 to 2020 and (2) collection of a blood-based diabetes biomarker at the baseline wave during this period. We chose 2010–2020 as our period of interest to maximize comparability between cohorts. The cohorts from China, England, and the USA were nationally representative of each country’s middle-aged and older population. The cohort from Mexico was representative from four states (a rural state, an urban state, a state with high migration, and a state with high presumed diabetes prevalence). The cohort from South Africa was representative of rural communities in Southern Africa. See online supplemental appendix 1 for details on the years of data collection and censoring by cohort.

Due to minor differences in the lower end of age eligibility between cohorts, we excluded individuals younger than 51 at baseline to ensure comparability. We also excluded respondents without follow-up information, with no available blood-based diabetes biomarker, or with missing data on prior diabetes diagnosis, gender, education, economic status, smoking status, body mass index (BMI), or survey weights. In the South Africa cohort, we excluded individuals with missing HIV status. Online supplemental appendix 2 shows participant flow diagrams for each cohort including numbers lost to follow-up.

Definition of diabetes

We defined diabetes as either (1) a history of self-reported diagnosis by a physician or healthcare worker or (2) an elevated blood-based biomarker meeting clinical criteria for diabetes.15 16 We used hemoglobin A1c (HbA1c) ≥6.5% (48 mmol/mol) as the biomarker threshold in all countries except China and South Africa, where we used fasting blood glucose ≥126 mg/dL (7.0 mmol/L) or random blood glucose ≥200 mg/dL (11.1 mmol/L). In China, plasma glucose was assessed using an enzymatic colorimetric test (92% of individuals were fasting). In the South African cohort, capillary glucose was assessed using a point-of-care analyzer (24% of individuals were fasting). In England, HbA1c was assessed using venous blood samples. In Mexico, HbA1c was assessed using a point-of-care analyzer certified by the National Glycohemoglobin Standardization Programme.17 In the USA, HbA1c was assessed using dried blood spots converted to whole blood equivalent values.18 Relevant question text and biomarker details are provided in online supplemental appendices 3 and 4.

Mortality ascertainment

All-cause mortality was captured in England by linking to the National Health Service Central Register (latest available data from April 2018). In other cohorts, all-cause mortality was captured during interviews with respondents’ spouses or other informants. In all cohorts, the month and year of death were available. If the date of death was unknown, it was estimated as the midpoint between waves in which an individual was known to be alive and had died. We measured survival time in years from the baseline interview as defined in this study to death, loss to follow-up, or the end of the follow-up period (May 2018 in England and December 2019 in the other countries), whichever came first.

Statistical analysis

Analyses were conducted within each cohort and accounted for survey weights and sampling design when available. We first calculated the overall and age-stratified diabetes prevalence at baseline. In calculating overall prevalence, we age-standardized to the WHO standard population. We then used Poisson regression with an offset for log-transformed person-years and robust standard errors to estimate differences in mortality rate ratios between individuals with diabetes (diagnosed or undiagnosed) and those without diabetes. Poisson models give similar results to Cox models when there are shorter follow-up intervals and have the advantage of directly estimating event rates.19 20 Both relative (mortality rate ratios) and absolute (mortality rate and mortality rate differences) measures are reported. Mortality rates and mortality rate differences are presented as the number of deaths per 1000 person-years.

We used prior evidence to develop a directed acyclic graph (DAG) showing our conceptual model of the relationship between diabetes and mortality (online supplemental appendix 5).21 While the DAG informed the selection of covariates, some confounders, such as genetic ancestry and physical activity, were unobserved and could not be included in our models. We adjusted for baseline covariates, including age (51–59 years, 60–69 years, and ≥70 years), gender (women vs men), education (less than upper secondary, upper secondary and vocational, and tertiary), smoking status (current vs not current smoker), BMI categories (underweight: BMI <18.5 kg/m²; normal weight: 18.5–24.9 kg/m²; overweight: 25.0–29.9 kg/m²; obese: ≥30.0 kg/m²), and economic status (tertiles). Economic status was defined as the annual income of an individual and their coresiding spouse or dependent children in high-income countries (England and the USA), and the annual household per-capita consumption in upper-middle-income countries (China, Mexico, and South Africa). Per-capita consumption is the preferred measure of living standard derived from surveys in the developing countries.22 In the South African cohort, we also adjusted for HIV status, given the high prevalence (23%) and known mortality association in this population.23

Models were fitted in the overall sample, by age category, and by gender. We also fit a model that separated individuals with diabetes into diagnosed and undiagnosed groups, comparing each to those without diabetes. This analysis aimed to evaluate differences in mortality between individuals with diagnosed diabetes and those with undiagnosed diabetes. Analyses were performed using Stata V.18.0.

Sensitivity analyses

We conducted three sensitivity analyses. First, we evaluated the sensitivity of our findings to the Poisson assumption of a constant hazard of mortality by fitting alternative survival models (Cox proportional hazard models and Gompertz parametric survival models).24 In these models, we used age as the timescale to allow for left truncation, given the current study included participants 51 years and older.25 Second, we estimated the association between diabetes and mortality using a slightly more restrictive epidemiological diabetes definition favored by the WHO to reduce misclassification of either (1) the self-reported use of a glucose-lowering medication or (2) an elevated biomarker meeting clinical criteria for diabetes.26 Finally, we performed an analysis without the adjustment for BMI categories given the potentially bidirectional relationship between diabetes and BMI.

Survey and respondent characteristics

Table 1 presents survey and respondent characteristics for the five cohorts. The final sample included 6251 individuals in China, 4819 in England, 1717 in Mexico, 3411 in South Africa, and 13 199 in the USA. Of the 29 397 total individuals, 4916 (16.7%) died during the study period. The median follow-up time ranged from 4.6 (IQR: 4.4–4.8) years in South Africa to 8.3 (IQR: 8.2–8.4) years in China. There were 191 782 total person-years of follow-up in the cohorts (China: 48 122 person-years; England: 24 536 person-years; Mexico: 11 192 person-years; South Africa: 14 722 person-years; and USA: 93 210 person-years).

There was considerable cross-country variation in some respondent characteristics, as illustrated in table 1. For example, while nine-tenths of individuals in China (89.9% (95% CI: 88.2% to 91.4%)), Mexico (88.8% (95% CI: 85.9% to 91.2%)), and South Africa (93.3% (95% CI: 92.4% to 94.1%)) had less than an upper secondary education, most individuals in England (69.2% (95% CI: 67.5% to 70.8%)) and the USA (86.1% (95% CI: 84.7% to 87.4%)) had an upper secondary education or greater. Current smoking ranged from 8.4% (95% CI: 7.5% to 9.3%) in South Africa to 31.0% (95% CI: 28.6% to 33.6%) in China. The prevalence of individuals who were obese ranged from 4.5% (95% CI: 3.9% to 5.3%) in China to 43.3% (95% CI: 42.3% to 44.4%) in the USA.

Diabetes prevalence

The age-standardized prevalence of diabetes was highest in Mexico (37.4% (95% CI: 33.4% to 41.5%), followed by the USA (21.8% (95% CI: 20.8% to 22.8%)), China (15.7% (95% CI: 14.3% to 17.2%)), South Africa (12.1% (95% CI: 11.1% to 13.3%)), and England (11.7% (95% CI: 10.6% to 13.0%)). Figure 1 shows the age-specific prevalence of diabetes by cohort at baseline. Among individuals with diabetes, the age-standardized proportion of those with diabetes who reported a prior diabetes diagnosis ranged from 46.4% (95% CI: 42.1% to 50.7%) in China to 86.1% (95% CI: 84.1% to 87.9%) in the USA (table 1).

Mortality rates

Adjusted all-cause mortality rates (per 1000 person-years) are presented in figure 2 and online supplemental appendix 6. In each cohort, mortality rates were higher among people with diabetes than those without diabetes. Across the cohorts, mortality rates among people with diabetes were highest in South Africa (57.5 (95% CI: 45.5 to 69.5)), followed by the USA (39.2 (95% CI: 36.1 to 42.4)), China (95% CI: 35.5 (95% CI: 28.6 to 42.4)), England (28.8 (95% CI: 22.1 to 35.6)), and Mexico (29.0 (95% CI: 19.0 to 39.0)).

Mortality rate ratios and mortality rate differences

The adjusted overall all-cause mortality rate ratios for people with diabetes versus those without diabetes ranged from 1.53 (95% CI: 1.39 to 1.68) in the USA to 2.02 (95% CI: 1.34 to 3.06) in Mexico (figure 3A). The adjusted mortality rate differences (per 1000 person-years) for people with diabetes versus those without diabetes ranged from 11.9 (95% CI: 4.8 to 18.9) in England to 24.6 (95% CI: 12.2 to 37.0) in South Africa. No significant differences were observed in adjusted mortality rate ratios or adjusted mortality rate differences by sex in cohorts. Mortality rate ratios appeared to decrease in older age groups in the cohorts from England and the USA (online supplemental appendix 7).

Figure 3B shows results when the diabetes classification was separated by diagnosed or undiagnosed status, compared with no diabetes. In general, there appeared to be a tendency among people diagnosed with diabetes to have higher mortality than people with undiagnosed diabetes. However, these differences were statistically significant only in Mexico, where people with diagnosed diabetes compared with undiagnosed diabetes had an adjusted mortality rate ratio of 1.95 (95% CI: 1.11 to 3.43), corresponding to an adjusted mortality rate difference of 22.6 (95% CI: −7.3 to 52.4) deaths per 1000 person-years (online supplemental appendix 8).

Sensitivity analyses

The results of the first sensitivity analysis using Cox and Gompertz models (online supplemental appendix 9) were very similar to the main results using Poisson regression models. In the second sensitivity analysis, using the slightly more restrictive epidemiological diabetes definition of either the use of a glucose-lowering medication (instead of self-reported diagnosis) or an elevated biomarker, we observed a slightly higher point estimate for the adjusted mortality rate ratios in the China cohort (1.86 vs 1.79) and slightly lower adjusted mortality rate ratios in the Mexico cohort (1.84 vs 2.02; online supplemental appendix 10). The third sensitivity analysis removing adjustment for BMI had the effect of slightly attenuating the mortality rate ratios and mortality rate differences compared with the main analysis (online supplemental appendix 11).