50 / 62
CARDIOVASCULAR JOURNAL OF AFRICA • Volume 31, No 1, January/February 2020
48
AFRICA
cardiovascular co-morbidities, including hypertension, were
estimated using an outdated data set (1998 South African
Demographic Health Survey).
15
National prevalence of disease
can conceal important differences in prevalence in sub-national
areas.
16
In most high-income countries where data are available
to finer geographies such as counties, NCD-related studies
have shown substantial heterogeneity in the prevalence of these
diseases and associated risk factors between sub-regions within
a country.
17-19
The aim of this study was therefore to profile the variations in
hypertension prevalence between districts in South Africa after
controlling for the individual’s demographic, social, economic,
behavioural and environmental variables.
Methods
The 2008, 2010/11, 2012, 2014/15 and 2017 samples for adults
aged 15 years and above from the National income Dynamics
Study (NiDS) panel survey were used in the study. The survey
provides a large nationally representative sample that is stratified
by the country’s 52 districts.
The target population was adult (15+ years) individuals in
private households and residents in workers’ hostels, convents
and monasteries, but excluded other living quarters such as
students’ hostels, old-age homes, hospitals, prisons and military
barracks. The sampling technique employed in the panel study
is exhaustively discussed elsewhere.
20
The sample retained for the
study includes respondents who had at least two blood pressure
(BP) measurements taken at the time of the survey.
The outcome of interest was hypertension prevalence for
individuals with systolic/diastolic BP of more than 140/90
mmHg or on medication for hypertension. BP measurements
for each panel were taken twice from each survey respondent.
Valid BP measurements were determined according to previously
applied criteria
13,21
as follows: (1) if the second systolic or diastolic
BP differed by more than 5 mmHg, the first BP reading was
excluded; and (2) a set of BP readings (systolic and diastolic) was
retained in the data set if the systolic BP was 80 mmHg or higher
AND if the systolic BP was at least 15 mmHg higher than the
diastolic BP level. A final systolic/diastolic blood pressure was
calculated as the average of the valid BP measurements (Table 1).
Several risk factors known to be associated with hypertension
and recorded in the NiDS data were adjusted for in estimating
the prevalence of hypertension, using multilevel logistic
regression. Important factors at the individual level were (1)
demographic factors: age, gender and race (self-identification as
African, Coloured, white, Asian/Indian); (2) biological factors:
specifically body mass index (BMI); (3) behavioural factors:
alcohol use (never used and past/current user), smoking status
(never and past/current) and physical exercise (none or some
exercise); (4) social and economic factors: education level (
≤
primary school, high school, and post high school), employment
status (employed, unemployed or economically inactive), medical
cover status (membership subscription to a registered medical
aid provider), residency type (urban and traditional/farms), and
income tertile calculated from equivalised per-capita household
income (household income divided by square root of number
of people in household); and (5) one environmental factor:
the season (summer, autumn, winter and spring) when the BP
measurements were taken.
Subjects self-identifying as whites or Indians/Asians were
combined in the analysis as they had relatively smaller sample
sizes. The alcohol use variable was not available for wave 5
(2017), and so the last observed status (from previous waves) was
used, or indicated as unknown if the subject was not in previous
waves.
Ethics approval was granted by the Human Research and
Ethics Committee of the University of the Witwatersrand,
Johannesburg, South Africa
Statistical analysis
The prevalence of hypertension was estimated using the following
two statistical methods reporting results at the district level.
Design-based estimates used the survey’s post-stratification
weights. This first step of the analysis was to estimate the
prevalence of hypertension nationally and by the levels of
each explanatory variable on univariate basis, followed by the
estimation of the prevalence by districts.
A three-level analysis model was used where periodic (survey
waves or the repeated measurements) hypertension statuses (first
level) are nested in individual respondents (second level), nested
within districts (third level). The risk factors listed above were
adjusted for in the model.
•
Level-specific distribution of hypertension variance. This step
aimed at estimating the distribution of the hypertension prev-
alence variance between the three levels, and the proportion
of variance explained by the individual-level demographic,
behavioural and socio-economic risk factors. This involved
first fitting a multilevel model without the covariates (a null
model), which allowed partitioning the variance between the
hierarchical levels. This was followed by constructing a full
model that adjusted for the risk factors (covariates) stated
above. The variance structure was described by the variance
partition coefficient (VPC) and the level-specific change in
variance (
Δσ
2
). The VPC measures the proportion of vari-
ance explained by each level within the model, and the
Δσ
2
measures the proportion of change in variance for each level
between the null and the adjusted model. Together, these two
measures describe how much of the variation is explained by
the variables included in the model.
•
Association of hypertension prevalence with individual-level
risk factors. Using the fully adjusted model, odds ratios (OR)
and
p
-values were calculated for each risk factor in the model.
•
Estimation of adjusted hypertension prevalence at the district
units involved using the predicted individual probability for
hypertension in estimating the prevalence at district level.
The estimated prevalences and 95% confidence intervals (CI)
were graphically presented to profile districts. This procedure
allowed a visualisation of which units were significantly
different from the national prevalence. Further analysis was
Table 1. Sample by wave for participants 15 years and above
Wave
Valid hypertension data
Total sample
Percentage valid
2008
14 135
18 617
75.9
2010/11
15 128
21 943
68.9
2012
18 393
25 228
72.9
2014/15
22 526
28 460
79.1
2017
23 605
32 123
73.5
Total
93 787
126 371
74.2