Study design and participants

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

The SLHAS used a stratified, multistage probability design to recruit a nationally representative sample of the noninstitutionalized adult population of Sri Lanka. Stratification involved two steps. In the first, following the conventional approach in Sri Lanka, all grama niladhari divisions (GNDs) (N = 14,104), which are the smallest administrative unit in Sri Lanka and primary sampling units (PSUs) in the SLHAS, were categorized by district (N = 25) and sector of residence (urban, rural, estate, rural/estate) into 57 preliminary strata. In a second innovative step, PSUs (GNDs) within each preliminary stratum were substratified into equally sized population quantiles after having ranked them using an index of area socioeconomic status (SES), generating 157 strata. This index was constructed by applying principal components analysis (PCA) to a set of GND-level indicators derived from the 2012 national population census obtained from the Department of Census and Statistics (DCS), such as the percentages of adults in different employment categories or the percentages of households using different cooking fuels, supplemented by a GND-level poverty index estimated by DCS using unpublished 2012 census data [21]. This was expected to reduce survey design effects, as many health indicators would be correlated with the GND SES index which would make stratification beneficial [22], and given previous research showing that SES area stratification performs better than traditional area stratification [23].

Within strata, individuals were sampled using multistage, probability cluster sampling. In the first stage, a minimum of two PSUs were selected from each stratum using probability proportionate to the size of their adult population, with disproportionately large PSUs picked with certainty as an implicit additional layer of stratification. At the second stage, four to six widely spaced households were identified in each PSU by systematically sampling the electoral register, or in rural areas where households were not distributed by street by using satellite maps showing the distribution of buildings to pick dispersed geolocations. Recruitment teams visited these households or the household nearest to the sampled geolocation, and then additional households at preset interhousehold intervals of two to four by walking in a predefined track, with larger intervals in more densely populated or urban PSUs. If they gained entry, recruiters enumerated all household residents and recorded data about the household and its residents, using a computer-assisted personal interviewing (CAPI) application running on iPad computer tablets. If the household gave consent, the CAPI software randomly selected one adult (≥18 years) using weighted probabilities that targeted a final equal distribution of respondents by sex and by age up to 69 years with oversampling of those aged ≥70 years. The study protocol excluded pregnant women and adults unable to give informed consent, and if the selected individual declined participation, the whole household was excluded, with the CAPI system preventing recruiters from selecting another individual.

Selected individuals were invited to attend a field clinic near their residence, typically a MOH clinic, where they were interviewed to assess their health and collect other individual and household information, and they underwent a medical examination and collection of biomarkers, including anthropometric measures and blood pressure. Individuals with mobility limitations were interviewed at home with a shorter examination that included blood pressure.

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

Survey completion

Data collection took place from November 9, 2018 to November 14, 2019, with field work staggered within districts to minimize seasonal bias. Out of 10,689 sampled households, recruiters contacted 10,247 (95.9%), of which 185 (1.8%) refused participation. Of the 10,062 individuals who agreed to participate, 6,627 were interviewed at a field clinic and 41 completed home interviews, giving an effective response rate of 65.0%. Response rates were higher in women (69%), known diabetics (73%), adults ≥45 years (74%), and rural areas (70%), and they were lower in Muslim women. Feedback by field recruiters indicate that younger adults were less likely to see the study as beneficial or relevant, whereas Muslim women in some communities were reluctant or discouraged for sociocultural reasons to attend field clinics alone or without spousal permission. The terrorist attacks in Sri Lanka in April 2019 caused a six-week disruption of field work leading to noncoverage of several PSUs, a fall in Muslim response rates, and the inability to survey one predominantly Muslim PSU owing to security conditions. The study team replaced the affected PSU with a substitute PSU from the same stratum, but it proved impossible to match the ethnic profile of the original PSU. Other PSUs that could not be covered were handled by merging 40 strata that were left with single PSUs with adjacent strata to yield 117 final strata.

Blood pressure measurement

All field clinic staff, who were medical or nursing graduates, were trained on the use of the study’s BP measurement protocol, including preparation of participants and using the SLHAS CAPI platform running on iPads. BP was measured after the participant rested for 10 minutes in a sitting position and with the arm resting at the level of the heart. All measurements were taken using the right arm unless specific conditions prohibited this. Two readings were obtained, 10 minutes apart, using an OMRON HEM-7320 Automatic BP Monitor, with a preformed flexible cuff allowing for arm circumferences of 17–36 cm. Accuracy of all devices was validated against a mercury column sphygmomanometer.

Other data collection

The participant interview, which also checked personal medical records and medicines that the participant brought with them, assessed whether the individual had evidence of prior diagnoses of HTN, diabetes mellitus, and other chronic conditions; symptoms consistent with angina or intermittent claudication (ROSE criteria) [24]; history of acute myocardial infarction (AMI) or stroke; and recorded medicines used in the previous two weeks. Individuals with medical records, symptoms, or self-reported history consistent with angina, intermittent claudication, AMI, or stroke were classified as having history of cardiovascular or peripheral vascular disease (CVD/PVD). Medicines were coded to the WHO Anatomical Therapeutic Chemical (ATC) classification, with antihypertensive medications defined as medicines in ATC classes C03A, C07–C08, C09A–C09D, C10BX03–C10BX04, C10BX06–C10BX07, C10BX09–C10BX15, and C09C–C09D.

Height was measured using a Seca 240 cm height measure to the nearest 0.1 cm. Weight was measured using an OMRON BF511 Body Composition Monitor to the nearest 0.1 kg. BMI was calculated as weight (kg) divided by the square of height (m). Waist circumference measurement was collected from participants using a Seca 200 cm tape measure at the level of natural indent of trunk during expiration.

Information on participant household assets was used to generate a SES ranking through PCA and to group all respondents by SES quantiles.

Hypertension definitions

Participants’ SBP and DBP were taken as the average of the two recordings. Hypertension (HTN) was defined in two ways. Following the JNC7 definition, individuals who had a SBP ≥140 mmHg or a DBP ≥90 mmHg or took any antihypertensive medications were categorized as hypertensive (HTN-JNC7). To define hypertension according to ACC/AHA 2017 definitions (HTN-ACC/AHA), a lower threshold of a systolic BP ≥130 mmHg or a diastolic BP ≥80 mmHg was used.

Prevalence estimation

Owing to its complex sampling design, SLHAS data are weighted to make population-level inferences. For all participants, the SLHAS generates a nonresponse weight using multilevel logistic regression that models the propensity to participate as a function of individual, household, area, and operational characteristics recorded during recruitment, including sex, age, language, whether self-reporting as diabetic, household SES, and operational details such as day of week. Ethnicity or whether the respondent was a known hypertensive were not recorded at recruitment and were not used.

The original study design required these weights to be calibrated so their sums matched population totals at the level of each stratum, district, and province for each of 14 age-sex groups (age groups: 18–29, 30–39, 40–49, 50–59, 60–69, 70–79, 80+), with population estimates sourced from the 2012 national census with adjustment for demographic change to 2019. This would not address potential bias from the observed underrepresentation of Muslims, which the nonresponse weighting did not consider. Consequently, this was revised to include three calibration steps done iteratively at increasing levels of aggregation (stratum > district > province > national): (i) population totals (stratum > national); (ii) age-sex totals (district > national); and (iii) ethnic totals (province > national), with weight trimming at selected points. As ethnicity is diverse and often contested in Sri Lanka, we defined Muslim ethnicity as those who self-described themselves as “Muslim,” “Moor,” or “Malay” when asked about ethnicity.

We estimated prevalence using these final weights, accounting for the clustered sampling design with a finite population correction and estimating variances using Taylor linearization. For generating estimates for international comparison, a second set of weights, recalibrated to match the WHO standard population [25], were also compiled. All analyses were performed using statistical software Stata version 17.0.

Evaluation of sample and weighting design

We evaluated several design features intended to improve precision and reduce bias, to inform future research. To assess area SES stratification, we examined the relationship with HTN using bivariate and categorical analysis, as well as localized regression. We evaluated the inclusion of ethnicity in the poststratification weighting by repeating the analysis without it. We examined the extent of spatial autocorrelation in HTN, which reduces precision when using cluster sampling, by using Stata’s loneway command to calculate intraclass correlation coefficients (ICCs), and we evaluated the overall study design by estimating design effects (DEFFs), which is simply the ratio of the actual sample size to the sample size that would yield the same precision when using simple random sampling.

Characteristics of participants

We excluded 3 SLHAS participants aged <18 years and 323 with missing data for taking antihypertensive medication or lacking either SBP or DBP measurements, leaving 6,342 (95.1%) participants for analysis from 298 PSUs. Their mean age (±SD) was 49.9 (±17.2), and 51.0% (3,235) were female, 16.9% (1,070) were previously diagnosed diabetics, and 23.6% (1,495) were self-reported as hypertensive or were taking antihypertensive medication. The sample had underrepresentation of younger adults and Muslims but, after weighting, matched the national population on age, sex, ethnicity, and area and household SES (Table 1).

Hypertension prevalence

The overall weighted prevalence of hypertension among all Sri Lankan adults was 28.2% (95% CI: 26.6–29.7) using the JNC7 definition, which almost doubled to 51.3% (95% CI: 49.6–53.1) using the ACC/AHA 2017 definition. For adults aged ≤69 years, the corresponding estimates were 24.7% (95% CI: 23.0–26.3) and 48.9% (95% CI: 47.0–50.8), respectively. Of those classified as hypertensive (HTN-JNC7), 53.4% (weighted estimate) were previously diagnosed or on antihypertensive medication, and the remaining 46.6% were undiagnosed. Using the ACC/AHA 2017 definition, the corresponding ratios were 31.2% and 68.8%, respectively. Table 2 summarizes the prevalence estimates by major characteristics. When reweighting to the WHO standard population, prevalence for all adults was 26.6% (95% CI: 25.0–28.2) for JNC-7 and 49.8% (95% CI: 48.0–51.6) for ACC/AHA 2017.

Hypertension prevalence varied similarly with both definitions across major characteristics. HTN-JNC7 prevalence increased almost linearly with age (1.4% per year) between 30 and 69 years, peaking in those aged 70–79 years (68.5%, 95% CI: 64.6–72.4), with evidence of plateauing and even a small fall in those aged ≥80 years. The increase in HTN-ACC/AHA with age was less steep, with the differences applying the two definitions being greatest in those aged 40–49 years, with prevalence doubling from 27.2% to 55.4%. With both definitions, overall prevalence was similar in both men (HTN-JNC7: 28.2%; HTN-ACC/AHA: 51.9%): and women (HTN-JNC7: 28.2%; HTN-ACC/AHA: 50.9%). However, the age-related increase was steeper in women from their 30s (Figure 1), with prevalence lower in women aged <40 years (HTN-JNC7: 7.8% vs. 12.6%) and higher in women aged ≥60 years (HTN-JNC7: 65.1% vs. 56.0%) than men of similar ages.

Across ethnic groups, prevalence was higher (p < 0.01) in Muslims, less educated adults, urban and rural locations, and Western and Central provinces, but these differences may be confounded by differences in age, ethnicity, and SES and require further investigation. HTN was highest in the richest SES quintile and the most developed area SES tertile. It was higher in those previously diagnosed as diabetic or taking diabetic medication, and prevalence also increased with BMI category, with 40.6% and 70.9% of those who were obese (BMI ≥30) being hypertensive according to the JNC7 and ACC/AHA 2017 definitions, respectively.

ACC/AHA 2017 guidelines recommend treatment with antihypertensive medication in individuals with Stage 1 HTN who have CVD or 10-year CVD risk >10%, whereas the WHO 2021 guidelines recommend medical therapy in those with diabetes or high CVD risk (undefined). It was beyond the scope of this analysis to comprehensively assess CVD status or to compute CVD risk, but the data yield a proxy measure that is history or medication consistent with CVD/PVD or prior diabetes diagnosis. Of the estimated 23.2% (95% CI: 22.1–24.5) of Sri Lankan adults classified as Stage 1 HTN, 16.6% (95% CI: 14.2–19.4) have a history or medication consistent with CVD/PVD or diabetes.

Assessment of study design elements

Our final sample consisted of 139 urban and 159 nonurban PSUs, with average cluster sizes of 13.8 and 27.8, respectively. Estimation of ICCs confirmed spatial autocorrelation of HTN prevalence and BP at the PSU level: ICCs were 0.039 for hypertension (HTN-JNC7), 0.050 for SBP, and 0.075 for DBP. Spatial autocorrelation was also greater in urban (ICC = 0.039) than nonurban (ICC = 0.029) PSUs despite smaller cluster sizes and increased interhousehold intervals.

The expectation that hypertension would be correlated with area SES was confirmed. HTN was significantly higher (p < 0.001) in the most developed area SES tertile (Table 2), and regression analysis confirmed a strong association of HTN with area SES centile (p < 0.001 for both HTN measures), whereas localized regression revealed a linear relationship between the 3rd and 10^th decile of area SES, increasing from 32% to 45% (HTN-JNC7). When ethnicity was excluded from the weighting process, HTN prevalence fell to 27.8% (HTN-JNC7) and 51.1% (HTN-ACC/AHA), with DEFFs of 1.97 and 1.85, respectively, compared with our actual final DEFFs of 1.96 and 1.98, respectively.

Strengths and limitations

The strengths of this study are that it used a large, nationally representative survey covering all demographic segments and districts in Sri Lanka; employed trained, dedicated field staff using standard procedures to take clinical measurements; undertook home examinations of respondents physically unable to attend a field clinic; and appropriately accounted for its complex survey design when making estimates. Poststratification weighting also ensured representativeness along multiple dimensions, and area SES stratification may have increased precision.

The overall SLHAS response rate of 65% is lower than reported by many health surveys in Sri Lanka. This can be explained by upweighting during recruitment of demographics anticipated to have lower response rates (e.g., young adults and men), disruption to survey operations caused by terrorist bombings, and rigorous operating procedures that prevented recruiters substituting another household member if the selected respondent refused, something that we have observed taking place in other Sri Lankan surveys. Nevertheless, our analysis specifically adjusted for nonresponse and observed biases, controlling for a wide range of factors, and our data were near complete for all participants, minimizing nonresponse bias.

A limitation of this analysis is that it cannot identity whether factors associated with higher prevalence are true risk factors. Further analysis using multivariate analysis that can simultaneously control for all covariates is required, and the authors plan to do this. In addition, the cross-sectional nature of the study further limits establishment of causality. However, the longitudinal nature of the SLHAS study does offer opportunities in future to identify what factors drive increases in hypertension. We also acknowledge that guidelines recommend measurement of blood pressure over more than one visit, which was not feasible given our study design and resources.