Prevalence of diabetes and pre-diabetes in Sri Lanka: a new global hotspot–estimates from the Sri Lanka Health and Ageing Survey 2018/2019

Study design and participants

We used data from the first wave of the Sri Lanka Health and Ageing Study (SLHAS) conducted from mid-November 2018 to mid-November 2019. The SLHAS is a national longitudinal cohort study approved by the Ministry of Health (MOH), Sri Lanka, and managed by a consortium of the Institute for Health Policy, University of Colombo, University of Peradeniya, University of Ruhuna, and the University of Rajarata.

The SLHAS Wave 1 used stratified, multistage probability sampling to recruit a nationally representative sample of the non-institutionalized adult (≥18 years) population of Sri Lanka, details of which have been described previously.12 The study design treated all 14 014 Grama Niladhari Divisions (GNDs)—the smallest administrative unit in Sri Lanka—as primary sampling units (PSUs). These were stratified by district, residential sector, and quantile of area socioeconomic status (ASES). The ASES ranking of GNDs was generated using principal component analysis (PCA) of social and economic indicators for each GND derived from the 2012 national population census and obtained from the Department of Census and Statistics (DCS).12 13 Two or more PSUs were then selected from each stratum by probability-proportionate-to-size. Within each PSU, households were sampled systematically, and within households, one eligible adult (defined as currently resident, age ≥18 years, not pregnant, and able to give informed consent) was randomly selected with age weighting designed to achieve a smooth age distribution. If the individual refused participation, the whole household was excluded.

Fieldwork was staggered within provinces to minimize seasonal bias. Participating subjects attended a nearby field clinic, typically a MOH clinic, in the morning, where they underwent an interview and examination, including blood tests. Individuals with mobility limitations were interviewed at home. All data were collected using a Computer-Assisted Personal Interviewing application running on computer tablets.

Out of 10 689 sampled households, 10 062 agreed to participate, 6627 selected adults attended a field clinic, and 41 completed home interviews, giving an effective response rate of 65.0%. Response rates were higher in women (69%), adults aged ≥45 years (74%), rural residents (70%), and known diabetics (73%), based on the analysis of data collected during field recruitment.

The Sri Lanka Medical Association Ethical Review Committee approved the study (ERC/18-022). Study information was provided to all participants, together with an official letter from MOH encouraging participation, and all participants gave informed written consent.

Measures and procedures

Participants were requested to fast for 12 hours prior to clinic attendance. On clinic arrival and having given informed consent, a venous blood sample for glucose was taken. Those who did not report being diabetic, had fasted, had provided an initial fasting blood glucose sample, and who consented underwent an oral glucose tolerance test (OGTT) by taking glucose monohydrate equivalent to 75 g of anhydrous glucose in 300 mL of water within a period of 10 min, followed by a second venous blood sample after 2 hours (±15 min). To standardize OGTT doses, the study commissioned prefilled bottles from a local manufacturer, and the time of each blood sample was logged, with subjects not meeting the stipulated time windows excluded.

Blood for glucose estimation was collected into sodium fluoride/potassium oxalate tubes and kept in a cooler at 4°C–8°C, before centrifuging at the field clinic to separate the plasma. Samples were stored within 6–10 hours of initial collection in a field freezer at −40°C to −20°C for transport to the study’s laboratory at the Medical Research Institute in Colombo. Glucose assay was performed by GOD-PAP (glucose oxidase-phenol amino phenazone) enzymatic method using a Rx Daytona+Clinical Chemistry Analyzer (Randox Laboratories, Crumlin, UK), with a coefficient of variation of 4.5% during the study period, which was within the acceptable range as evaluated by the manufacturer’s external quality assurance scheme (RIQAS, Randox International Quality Assessment Scheme).

For 79 subjects (67 of whom were not diagnosed as diabetics), a faulty batch of Vacutainer tubes resulted in insufficient sample to assay plasma glucose (PG). For these, the PG equivalent value was imputed by linear regression of PG values against serum glucose values that had been separately collected for participants attending the relevant session. Blood was also collected for glycated hemoglobin (HbA1c) assay, but limited funds meant that this was only performed for subjects with known diabetes, those who did not fast, and a small random subsample of others, with remaining samples biobanked for future testing.

Height was measured using a Seca 240 cm height measure (Seca, Hamburg, Germany) to the nearest 0.1 cm. Weight was measured using an OMRON BF511 Body Composition Monitor to the nearest 0.1 kg. Body mass index (BMI) was calculated as weight (kilogram) divided by the square of height (m²). BMI categories are defined using the WHO cut-offs in this paper (underweight <18.5; normal 18.5–24.9; overweight 25.0–29.9; obese ≥30 kg/m²), but we also report results using the WHO Asian BMI categories (underweight < 18.5; normal 18.5–22.9; overweight 23.0–24.9; obese ≥25 kg/m²). Waist circumference measurement was collected from participants using a Seca 200 cm tape measure at the level of the natural indent of the trunk during expiration.

The participant interview collected information about personal and household characteristics and reviewed personal medical records and medicines that the participants were requested to bring. Personal medical records consisted of written notes and other documents held by patients that they obtain from their providers, including personal logbooks that some clinics provide. Medicines were coded to the WHO Anatomical Therapeutic Chemical (ATC) classification, with antidiabetic medications identified as medicines in ATC classes A10A (insulins) and A10B (oral hypoglycemics). A measure of relative socioeconomic status (SES) was derived from PCA of participant-reported information on household assets,13 with the first principal components used to rank and group participants into SES quantile groups (online supplemental text S1).

Definitions

We defined individuals as having diagnosed diabetes if having a positive history as evidenced by self-report or explicit mention of a diagnosis in their medical notes or the use of antidiabetic medication. Following the 2022 American Diabetes Association (ADA) guidelines,14 we defined new cases of diabetes (undiagnosed diabetes) on the basis of satisfying either or both of the following criteria: (1) fasting plasma glucose (FPG) ≥126 mg/dL (7.0 mmol/L) and (2) 2-hour plasma glucose during OGTT (2-h PG) of ≥200 mg/dL (11.1 mmol/L); and we defined pre-diabetes as either impaired fasting glucose (IFG) (FPG of 100 mg/dL (5.6 mmol/L) to 125 mg/dL (6.9 mmol/L)) or impaired glucose tolerance (IGT) (2-h PG of 140 mg/dL (7.8 mmol/L) to 199 mg/dL (11.0 mmol/L)). We defined dysglycemia as either pre-diabetes or diabetes.

Prevalence estimation

The SLHAS Wave 1 data come with design weights to make the sample representative of the adult population of Sri Lanka. The basic procedure has been described previously.12 In brief, non-response weights were generated by modeling the propensity to participate as a function of characteristics recorded during recruitment, including whether self-reporting as diabetic. These were then adjusted to the age–sex population structure of strata, districts, and provinces to generate poststratification design weights. Finally, weights were calibrated in successive steps using iterative proportional fitting (IPF) to match the strata, age–sex, ethnic and sector structure of the population at district, province, and then national levels. This was done as Muslims were under-represented in the original sample, which could bias prevalence estimates,12 and since reference data on ethnic proportions were only available in aggregate at the district level, necessitating approaches such as IPF.15 For this study, the base weights for the eligible sample were again calibrated using IPF at increasing levels of aggregation from stratum to national level to match the 2019 population totals for strata, age–sex groups, sectors, and ethnicity. For age-standardized estimates, we recalibrated these to match the WHO standard population.16

We further modified the weights for the different subsamples involved in prevalence estimation. Out of 5487 subjects who did not have diagnosed diabetes, valid FPG values were obtained for 5351 (98%), with most exclusions due to failure to fast ≥9 hours. These are referred to as the FPG subsample. Valid 2-h PG values were obtained for 3653 subjects (67% of the FPG subsample) who lacked a previous diabetes diagnosis and had fasted, with losses mostly due to the insufficient clinic time or refusals to take the OGTT; these are referred to as the OGTT subsample. Naively combining these using their standard weights with the diagnosed diabetes subsample would lead to the overestimation of prevalence and ignore potential non-random test participation. Consequently, each subject’s weight was multiplied by a propensity weight for test participation, generated for each subsample by logistic modeling of participation as a function of individual characteristics including age, sex, BMI, SES, hypertension, and blood test values including total and low-density lipoprotein-cholesterol (and FPG in case of the OGTT subsample). Adapting the method used by NCD-RisC and by previous analyses of US National Health and Nutrition Examination Survey (NHANES) data,1 11 the resulting weights for the FPG and OGTT subsamples were recalibrated so that their sums matched the weight totals by age–sex groups and ethnicity for all subjects who did not have a prior diabetes diagnosis. This ensured that the final weighted samples matched the overall national population in 2019 and minimized biases arising from non-random test participation.

When making prevalence estimates at provincial and district levels, the weighting procedure described above was modified to generate province-specific and district-specific weights by matching to the age, sex, and ethnic profiles at the province or district level. Although we measured HbA1c in some subjects who had not fasted, we did not use HbA1c to classify diabetic status in these, as overall numbers were too small to support robust reweighting and because of known differences in sensitivity and specificity compared with OGTT and FPG.14

We estimated the prevalence using the final weights, accounting for the clustered sampling design with a finite population correction and estimating variances using Taylor linearization. We used Wald tests to test whether differences in proportions were significantly different from zero, using the df that reflected the complex sample design. A p-value <0.05 was considered statistically significant. All analyses were performed using Stata V.17.0 (Stata, College Station, Texas, USA), and we followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines in reporting the study.

Characteristics and classification of participants

We excluded three individuals aged <18 years, and four with missing data on their previous diabetic status, leaving 6661 participants (99%) for analysis. Their mean age (±SD) was 50.1 (±17.2) years, 3396 (51.0%) were women, and 603 (9.2%) were obese (BMI ≥30 kg/m²) (online supplemental table S1). Younger adults and Muslims were under-represented, but after weighting the sample matched the national population on age, sex, ethnicity, sector, and area and household SES quantiles (online supplemental table S1).

Subject selection and classification are summarized in online supplemental figure S1. Out of the total sample, 1170 (17.6%) self-reported a previous diabetes diagnosis or were taking antidiabetic medication, of whom 14 (1.2%) were taking insulin. Of the remaining sample, that is those who did not self-report a diabetes diagnosis and were not taking antidiabetic medication (n=5491), only 356 (6%) provided medical notes (compared with 68% of those who did self-report a diagnosis or were taking antidiabetic medication). From these, only an additional four individuals (1%) were identified as having a diabetes diagnosis solely based on their medical notes, yielding a total of 1174 individuals with a previous diagnosis. FPG was obtained for an additional 5351 subjects, of which OGTT was also obtained in 3653. This allowed diabetes status to be assessed in 6525 subjects using previous diagnosis and FPG and in 4827 subjects using previous diagnosis and OGTT (and FPG). FPG newly identified 271 subjects (4.1%) as diabetic, and OGTT 353 (5.3%), of which 147 were also classified as diabetic using FPG.

Participants in the OGTT differed significantly from those who were eligible but did not take the test when assessed using multivariate logit regression. Characteristics significantly associated with higher participation included being aged 30–49 years, living in rural or estate areas, speaking Tamil, being interviewed at home, being normal or overweight, having higher systolic blood pressure, and being a smoker (all p<0.01), while being aged 80 years or more was associated with lower participation (p<0.001).

Diagnosed diabetes

The estimated weighted crude prevalence of diagnosed diabetes in adults aged ≥18 years in 2019 was 14.3% (table 1). Prevalence increased with age, peaking at the age of 60–69 years (29.8%), with a decline at older ages. It was higher in women (15.3%) than in men (13.2%; p<0.02), higher in Muslims (29.0%) than in Sinhalese (12.4%; p<0.001) and Tamils (14.7%; p<0.001), and in urban (24.7%) than in rural (11.9%; p<0.001) and estate (4.6%; p<0.001) sectors. Prevalence increased with household SES and was highest in the Western, Northern, and Eastern provinces (online supplemental table S2).

Undiagnosed diabetes

The prevalence of undiagnosed diabetes in all adults was 4.2% using FPG and almost double using OGTT (7.5%). The combined prevalence of total undiagnosed diabetes using both FPG and OGTT was 8.7% (table 1). Total undiagnosed diabetes increased with age, peaking in those aged 60–69 years, and was higher in women than men (p<0.002). It increased with BMI, with differences between each BMI category highly significant (p<0.003). Prevalence was significantly higher in Muslims than Tamils (p<0.05) but otherwise did not differ significantly with ethnicity and household SES.

Total diabetes

The overall crude prevalence of diabetes, including both diagnosed and undiagnosed diabetes (detected by FPG or OGTT), in adults aged ≥18 years in 2019 was 23.0% (95% CI 21.2% to 24.7%) (table 1). Using FPG alone, total diabetes prevalence was 18.5% (95% CI 17.1% to 19.8%). Age-standardized prevalence was 21.8% (95% CI 20.1% to 23.5%) and 17.4% (95% CI 16.0% to 18.7%) if using FPG alone.

Crude prevalence increased with age, peaking at 43% in those aged 60–69 years and was higher (p<0.001) in women (25.5%) than in men (20.1%). The age-sex specific trajectories as estimated using restricted cubic splines were similar across sexes, except prevalence was higher in women aged 30–40 years and >50 years (figure 1). Point estimates by sex and age group are visualized in online supplemental figure S2 and tabulated in online supplemental table S3. Prevalence increased with BMI, being lower in the underweight (6.8%) than in those of normal weight (20.6%; p<0.001), and higher in the overweight and obese than those of normal weight (p<0.001 for both). The male and female gradients of diabetes with BMI were similar (figure 1), and in both sexes, the largest increase occurred at BMI levels below 23 kg/m², which is the upper cut-off for normal weight using the Asian BMI categorization.

Request permission

Prevalence was higher in Muslims (40.8%) than other ethnic groups (p<0.001), and in urban (35.8%) than the rural sectors (20.4%, p<0.001), which was higher than in the estate sector (9.6%, p=0.09). Prevalence increased with household SES from 19.9% in the poorest quintile to 28.9% in the richest quintile. However, prevalence increased more steeply in relation to area SES (figure 1) increasing from 14.6% in the least developed to 31.3% in the most developed tertile. Prevalence was highest in the most urbanized, economically developed southwestern region of the country, and in the northern and eastern regions (figure 2). Prevalence doubled at the provincial level from 16% in Uva to 29% in the Western Province (online supplemental table S2), and across districts from 8.0% in Matale to 34.1% in Colombo and 30.8% in Jaffna (online supplemental table S4).

Request permission

Proportion of undiagnosed total diabetes

Among people with diabetes, 37.7% (95% CI 34.0% to 41.4%) were undiagnosed. This was higher in age groups <50 years compared with older ones (p<0.05 for all), but not significantly different between men and women. Compared with the urban sector (31.1%), the proportions undiagnosed were higher in the rural (41.7%; p<0.01) and estate (52.1%; p<0.001) sectors. They were lower in the most developed area SES tertile (31.5%) than the middle (42.1%; p<0.05) and least developed (44.6%; p<0.01), but differences between household SES quintiles were not significant.

Pre-diabetes

Pre-diabetes assessed using IFG was 23.2% (95% CI 21.3% to 25.1%) and using IGT was 18.9% (95% CI 17.0% to 20.8%) in adults (≥18 years) in 2019 (table 2). IFG increased with age (18–29 years: 13.3% versus ≥80 years: 38.7%), but IGT exhibited no relationship with age. IFG was higher in men (25.2%) than in women (21.4%; p<0.01), but IGT was less in men (14.4%) than in women (22.9%; p<0.001). Both IFG and IGT increased with BMI, but the increase was steeper with IGT than with IFG.

Using both IFG and IGT, overall pre-diabetes prevalence was 30.5% (95% CI 28.2% to 32.7%) in adults. Prevalence was lower in those aged 18–29 years (21.7%) than older age groups and was similar in men (30.3%) and women (30.7%). Prevalence was lower in Tamils (25.6%) and Muslims (26.4%) than in Sinhalese (31.9%; p<0.02), but otherwise there were no significant differences across ethnic groups, sectors of residence, household SES quintiles, area SES tertiles, and provinces (online supplemental table S5). Overall prevalence increased with BMI and was significantly lower in the underweight (21.4%) than in those of normal weight (29.4%; p<0.01), and significantly higher in the obese (42.4%) than in the overweight (32.2%; p<0.003).

Overall dysglycemia

An estimated 53.4% (95% CI 50.8% to 56.1%) of Sri Lankan adults had dysglycemia in 2019 (table 2). Prevalence rose with age, peaking at 74.2% in those aged 60–79 years. It was higher in women (56.2%) than men (50.4%; p<0.02); in Muslims (67.2%) than Sinhalese (52.8%) and Tamils (47.2%) (p<0.001 for both); and in the urban (64.1%) than rural sector (51.6%; p<0.001), which was higher than the estate sector (32.2%; p<0.01) (online supplemental table S5). Dysglycemia prevalence increased with household SES quintile from 50.1% in the poorest quintile to 59.9% in the richest quintile. The increase with BMI was steeper from 28.2% in the underweight to 62.1% in the overweight category (p<0.001 across all comparisons).

Concordance of FPG and OGTT classification

In the OGTT subsample subjects that were not known diabetics and who had both FPG and OGTT results (n=3653), FPG newly diagnosed 200 and OGTT newly diagnosed 353 as diabetic. Having appropriately weighted each subsample, both tests concurred in classifying 3.3% of adults as undiagnosed diabetes prevalence, but the OGTT test separately identified three times the additional prevalence (4.1%) as FPG (1.2%), consistent with OGTT having higher sensitivity.10 11 Overall, total diabetes prevalence based on combining both FPG and OGTT (23.0%) was 4.5% higher than prevalence based only on FPG (18.5%), and 1.2% higher than relying on OGTT alone (21.8%).

When estimating prevalence using the sample of adults that had no prior diabetes diagnosis that was jointly assessed with both tests to identify undiagnosed diabetes, there was considerable concordance in classifying adults as undiagnosed diabetes or normal. Of those classified as having undiagnosed diabetes using FPG, 73.3% (95% CI 61.1% to 85.6%) would be classified as diabetic using OGTT, and of those classified as normal using FPG, 79.1% (95% CI 76.8% to 81.3%) would be classified as normal using OGTT. There was less concordance on IFG, with OGTT classifying only 32.2% (95% CI 27.3% to 37.1%) as IGT and 12.1% (95% CI 9.4% to 14.8%) as diabetic. However, in the latter group, the weighted mean HbA1c was 5.8% (40 mmol/mol) in subjects (n=135) with this data, within the normal range. This replicates similar findings from US NHANES data11 and helps explain the lower pre-diabetes prevalence and higher undiagnosed diabetes prevalence obtained with OGTT.