This is the first study to compare sentinel HIV seroprevalence surveys with both routine aggregated and individuated data. Our results show that from 2015 onwards, the PHDC provides a reliable source of individuated data for accurate and timely antenatal HIV surveillance at provincial, district and age-group levels. These are essential for a responsive health system to plan and evaluate programmes.
From 2015 onwards, the PHDC dataset enumerated more pregnancies than reflected in routine aggregate antenatal first visit registers (DHIS), supporting the completeness of these individuated data. The lower enumeration of pregnancies prior to 2015 is most likely due to limited availability of electronic data for linkage in earlier years, when electronic data systems were less widely established. As more public health facilities began using routine electronic data systems, patients with any contact with public health facilities would have some electronic record enabling inclusion in the PHDC. Given that the routine individuated data represent the entire population utilising public health services, these data are more representative of the population than survey data. It is also likely that linked individuated data are less prone to the biases of routine aggregated data such as duplication and lack of completeness, both numerically and within available records. A single unique identifier across the province prevents duplication of records where patients attend different facilities. Linkage of different electronic systems using a unique identifier also allows more opportunities for ascertainment of pregnancy and HIV status than fixed variables used in aggregate data.
Comparison of descriptive characteristics between PHDC cohort and survey cohort enabled validation of the PHDC cohort prior to estimating HIV prevalence. The median age of women in the PHDC cohort (26.9 years) was similar to that in the antenatal survey population (3, 6), supporting the close alignment in demographic profile of the two cohorts. PHDC, however, had a higher proportion of women with no prior evidence of pregnancy (60%) when compared to recorded gravidity in the survey (approximately 30%)(3, 6). This is likely due to less well-established electronic systems in earlier years for prior pregnancy ascertainment. Furthermore, the PHDC cohort had a higher proportion of patients from the Cape Metro district and lower proportion from the rural districts. This is also likely due to increased PHDC coverage in the Cape Metro, particularly in earlier years as electronic systems were better established in urban areas. The national surveys may further include oversampling of rural districts to generate accurate district-level estimates. We would suggest whole population surveillance negates a need for oversampling.
HIV prevalence estimates in both national and provincial antenatal surveys remained consistent between 2011 and 2015, with provincial surveys providing more precise estimates due to the larger sample size. From 2014 onwards, PHDC estimates are closely aligned to the survey. Given the wide confidence intervals for survey estimates, trends are difficult to infer. DHIS HIV prevalence estimates for 2011 and 2012 were consistent with survey estimates, as shown in a comparative study (4). This study, however, used HIV-service specific denominators from DHIS viz. PMTCT initial HIV test acceptance or refusal. From 2014 onwards, refusal of PMTCT initial test was no longer recorded hence first antenatal visit was used as the denominator, resulting in lower estimates than in earlier years. DHIS data, however, showed an increasing proportion of women over time to be presenting for their first antenatal visit and known to be living with HIV. This increase corresponded with a decreasing proportion of women accepting HIV testing at first antenatal visit. The lower proportion of test acceptance is most likely due to increasing awareness of HIV positive status, in keeping with a recently published modelling study(23). Additionally, some women with known HIV positive status may still retest in antenatal settings, as seen in a recent study in the Western Cape(24). The change in DHIS elements likely represents the changing HIV context with greater awareness of HIV status and wider roll-out of antiretroviral therapy (ART) over time(23). Variability presented in DHIS estimates is likely, in part, due to challenges in maintaining routine aggregate data systems where contextual factors, such as staff turnover, high workload, and social unrest impact data collection and consolidation from registers.
PHDC HIV prevalence estimates 2011 to 2013 were lower than survey estimates, most likely due to incomplete electronic data. From 2014 onwards, estimates were better aligned and more consistent over time. As PHDC estimates do not distinguish between pregnancies with or without antenatal care, prevalence was expected to be slightly higher than that estimated using DHIS and survey data, as these data are limited to pregnancies with antenatal care. Antenatal first visit coverage is however high in the Western Cape at 94% and therefore the contribution of pregnancies without prior antenatal care to HIV prevalence is small in the PHDC estimates(13, 25, 26). Reassuringly, PHDC estimates remained within the 95% confidence intervals of the national survey, and PHDC and DHIS estimates showed a positive correlation and low average difference. It should be noted, however, that average difference may mask the volatility of prevalence trends over time. Low average difference between survey and PHDC estimates suggest the PHDC may be reliably used to estimate antenatal HIV prevalence at provincial level. PHDC estimates over time were also similar to the Thembisa model estimates. Since the Thembisa model includes both private and public sector data, lower estimates than the PHDC were expected (reflecting lower HIV prevalence among private patients)(27).
At a more granular level, from 2015 onwards, estimates from PHDC were closely aligned to both DHIS and survey at district level, with disparities noted more in sparsely populated rural districts like Overberg. These disparities may reflect survey under-sampling in smaller rural districts as well as differences brought about by migrant populations such as workers(28). Age-group HIV prevalence estimates from the PHDC were closely aligned to the national survey from 2015 onwards, showing consistent HIV prevalence in all age groups over a 5-year period, with higher estimates in older age groups. Higher prevalence in older age groups is expected due to a combination of factors, including increased cumulative incidence with age, and use of ART extending life expectancy. Since DHIS cannot provide age-disaggregated estimates, comparisons with PHDC and survey could not be made. Furthermore, in recent years, estimates at sub-district level are only possible with the PHDC data as the national survey is underpowered at sub-district level. This again highlights the advantages of linked individuated data over both survey data and aggregate routine data in providing granular estimates, not limited to predetermined indicators.
Limitations
A comparative study of this nature is subject to several limitations. Firstly, routine data were validated against sentinel surveillance data, with national surveys serving as the most accurate HIV prevalence estimates or “gold standard”. These survey estimates are however based on smaller sample sizes, over a limited time period, and underpowered for granular estimates which may impact accuracy. PHDC data were less reliable before 2014 due to incomplete electronic data in earlier years. Estimates from each dataset are derived using differing numerators and denominators with differing levels of quality and completeness. Furthermore, both individuated and aggregated routine data are subject to various pitfalls such as capturing errors, administrative errors and consolidation errors which may impact quality and completeness of these data. Pregnancy ascertainment may differ between districts and sub-districts due to differing use of electronic information systems. Since detailed patient characteristics such as socio-economic status and education level are not captured routinely, in-depth comparisons with survey data were not possible. Contextual factors impacting on observed trends in routine data are diverse and require further investigation – these include clerical changes in capturing approaches, widespread impact of the COVID-19 pandemic on service utilisation and staffing, migration patterns etc. Survey and routine data sources evaluated in this study include only patients utilising public health care services, excluding those in the private sector and those without access to public health care. Lastly, while the Western Cape province has established individuated data systems, most other provinces in South Africa are still reliant on aggregated register-based data. The results of this study are therefore not representative of the whole country and highlight the need and potential to strengthen individuated routine information systems for improved surveillance.