In sub-Saharan Africa, surveillance of women attending antenatal care (ANC) is often used to measure prevalence and monitor trends in HIV infection. However, when applying ANC-based HIV prevalence estimates to the general population the following biases should be considered: only pregnant women are eligible for testing (structural bias) [1]; women who become pregnant and attend ANC facilities are sexually active and not using contraceptives (self selection bias) [1, 2]; attendance varies by factors associated with HIV [2]; and, HIV-infected women may be less likely to become pregnant [1–7].

Fertility among HIV-infected women in sub-Saharan Africa is lower than in HIV-uninfected women, except in women aged 15–19 years [7]. In young women the selective pressure of sexual debut on pregnancy and HIV infection resulted in higher fertility rates among the HIV infected [7]. Six studies conducted in high fertility populations in sub-Saharan Africa showed that HIV prevalence in pregnant women was lower than in women of reproductive age overall [6]. In populations of southern Africa with low fertility and extensive contraceptive use, bias due to the selection for pregnancy in ANC-based HIV prevalence estimates could be smaller [8]. This is because women who use modern methods of family planning may take more effective measures to avoid HIV infection [1], and sub-fertility in HIV-infected women will have a weaker effect [8].

Bias due to the purposive selection of ANC facilities should also be considered when applying ANC-based prevalence estimates to a population [1, 8]. Over-representation of ANC clinics in urban areas, where HIV prevalence is usually relatively high, may result in HIV prevalence levels being exaggerated [8]. However, evidence of urban and peri-urban based clinics attracting large numbers of women from rural areas, where HIV prevalence tends to be lower, would mitigate this [8].

Population-based surveys are more representative of the general population than ANC-based surveys as they include non-pregnant and non-ANC attending pregnant women, as well as men. However, limitations exist and of particular concern is the effect of non-response on HIV prevalence estimates.

Population-based surveys have been conducted in several sub-Saharan African countries [8–20], including South Africa where primary health care services are free, use of modern contraception is high and the total fertility rate is low [21]. It has been suggested that women using modern methods of family planning may take more effective measures to avoid HIV infection [1]. In KwaZulu Natal, South Africa, information collected by the Africa Centre Demographic Information System (ACDIS) [21–23], in 2001 shows 51.7% of women aged 15 to 49 years to have ever used a modern contraceptive method, the median age of first sexual intercourse to be 17.7 years, and fertility between 1980/84 and 2000/01 to have declined in women aged 18 years and older [21].

In this paper we present and compare HIV prevalence estimates from both population-based surveillance and ANC sentinel surveillance within an area of sub-Saharan Africa with high contraceptive use and low fertility. We identify and explore differences, and reasons for differences, between the two surveillance systems, with particular emphasis on unrepresentative selection of clinics in ANC sentinel surveillance and unrepresentative testing in population-based surveillance. Where estimates from the two surveillance systems differ we explore methods to reduce these differences.

The ACDIS is conducted in the rural sub-district of Hlabisa in northern KwaZulu-Natal, South Africa. It covers 435 square kilometres and a total resident population of 85123 (unpublished data as of January 2005). The 11284 homesteads within the area have been enumerated and mapped using a geographic information system (GIS). The area includes a formally designated urban township, peri-urban areas (settlements with a population density of more than 400 people per km²), and rural areas. The rural population live in scattered homesteads that are not concentrated in villages.

Population-based linked anonymous HIV testing was introduced within the ACDIS in July 2003. Sampling for testing is based upon information collected routinely through demographic surveillance [21–23]. All resident women aged 15 to 49 years and men aged 15 to 54 years, were eligible for annual HIV testing through a finger-prick blood sample on filter paper and approached for inclusion in the survey [24–27]. Additionally, 10% of non-resident members of households located within the study area, in above age groups, were randomly selected for testing. To facilitate a comparison with ANC-based estimates, only resident women were included in these analyses. A resident is an individual, reported by the household informant, who keeps their daily belongings, and who spends most nights, within the survey area [23, 24]. Results are those from the second annual HIV survey (January to December, 2005).

Ethical approval was received from the University of KwaZulu Natal (E029/2003). All individuals eligible for HIV testing were asked for written informed consent and informed about the potential risks to becoming aware of ones HIV status, about how and where HIV test results and post-test counselling may be accessed and, if found positive, how they may be referred to a local clinic for further screening and assessment of eligibility for antiretroviral treatment. The choice to provide a test sample and to access the HIV test result rests fully with the individual.

In December 2001, Hlabisa Health sub-district became the first rural district in South Africa to provide antiretroviral drugs for the prevention of HIV mother to child transmission. Between January and May 2005, alongside the Prevention of Mother-to-Child Transmission (PMTCT) programme, venous blood was taken for routine ANC laboratory tests from all women attending first ANC visits at all six government clinics delivering ANC within the ACDIS. Surplus blood from these samples was also used for anonymous unlinked HIV testing. Parity and age was linked to a woman's HIV test result. Results cannot be linked back to the individual as, apart from date of birth, no personal identifiers were collected.

For each participant the most likely clinic at which antenatal care was obtained was predicted on the basis of a GIS accessibility model that estimated travel time to the six government clinics within the surveillance area offering ANC. The model took into account the quality and distribution of the road network, barriers to movement and the likelihood of utilising public transport to access care [28]. The six clinics were categorised as mixed peri-urban/urban, mixed rural/peri-urban or rural respectively on the basis of their predicted constituent catchment populations.

As ANC sentinel surveillance does not collect residency information, ANC attendees were proportionally assigned to one of the three residency types (urban/peri-urban/rural) based on the underlying predicted catchment population of the clinic attended. To assess the reliability of the clinic accessibility model in predicting ante-natal attendance we compared the prediction of the model with reported ante-natal clinic usage amongst women ever reporting a pregnancy within the ACDIS.

Pearson chi² values and all confidence intervals presented are at the 95% level. STATA 9.0 (Stata Corp., College Station, Texas, USA) was used for univariate and multivariate analyses. To account for unrepresentative testing, population-based and ANC-based HIV prevalence estimates were standardised for age, age and location (clinic for ANC-based estimates and individual-residence for population-based estimates), and age and clinic catchment by applying the respective prevalence estimates to samples of women adjusted to proportionally match ACDIS population level data on all women aged 15 to 49 as of 1^st January 2005. Women reported to the ACDIS during twice yearly fieldworker visits as having been pregnant (regardless of outcome) during the period 1^st July 2004 and 30^th June 2005, who were also eligible for population-based testing, were identified to assist comparative analyses.

Unrepresentative testing by HIV status was analysed by linking records (based on a unique identifier allocated to all participants) between the first (July 2003 to December 2004) and second (January to December, 2005) population-based HIV surveys. The proportion of women with a negative HIV test result in the first survey who also consented to test in 2005 was applied to all women with a first survey test result from whom consent was sought in 2005.

The results of this paper show population-based estimates of HIV prevalence in women to be consistently lower than ANC-based estimates. Although there are several possible explanations for this difference (discussed below), one possible explanation that should first be considered is unrepresentative testing in the population-based survey

Within the population-based survey, women in the 25–29 and 30–34 age-groups presented both the highest HIV prevalence estimates and the lowest proportions agreeing to test. Women resident in the urban area were the overall group least likely to consent to test. Prevalence estimates among groups where the proportion of women contacted consenting to test is particularly low should be interpreted with caution. The proportions of women consenting to test are presented as of those contacted and not as of the full eligible population. Therefore, it may be likely that in terms of HIV status those contacted differ to those not contacted. Further analyses are necessary to explore this possible bias.

It is likely that within the ACDIS area, where provision of ANC clinics offering PMTCT services is comprehensive [29], many women will already be aware of their HIV status, and this may well influence their decision to agree to give a sample in population-based surveillance. Although testing consent in 2005 was lower amongst women with a previous HIV positive test result than amongst women with a previous negative test result, crudely adjusting for testing bias by HIV status had little effect. It should be noted that although test results are made available to participants [27] a low rate of uptake of test results [25] ensures a surveillance test result is no indication of an individual being aware of their HIV status.

A review of 20 national population-based surveys across sub-Saharan Africa suggests non-responders are likely to have a higher prevalence of HIV than responders and, by applying the most extreme scenario to account for such bias, an adjustment factor of 1.34 may be required [30]. Our analyses suggest prevalence among women testing in the first population-based survey but not the second was 1.3 times higher than that among women testing in both. However, it is unlikely that this figure is representative of all women refusing to test and therefore, it was not used to adjust the population-based estimate.

Although statistically similar, the overall population-based estimate for women of 25.2% (95% CIs: 24.0%, 26.4%) was lower than the estimate of 27.3% (95% CIs: 26.3%, 28.4%) from the first ACDIS population-based survey [24, 31] and the estimate of 30.4% (95% CIs: 24.7%, 36.7%) amongst 15–49 year old women resident in 2005 in a province wide population-based survey of KwaZulu Natal [32].

The overall ANC-based estimate of 37.7% (95% CIs: 34.9%, 40.6%), although statistically similar, falls mid-way between estimates of 35.3% (95% CIs: 33.7, 36.9%) amongst pregnant women tested within the Hlabisa sub-District PMTCT programme January to June 2005 [29, 33] and 39.1% (95% CIs: 36.8%, 41.4%) amongst 15–49 year old ANC clinic attendees in a province wide survey of KwaZulu Natal 2005 [34]. The estimate presented in this paper of 40.7% (95% CIs: 36.7%, 44.8%) amongst women attending the urban clinic is comparable to the estimate of 41.2% (95% CIs: 34.7%, 47.9%) for the same clinic in December 1998 [35].

The findings of this study suggest that where ANC coverage is high, population-based HIV surveillance systems under-estimate HIV prevalence due to unrepresentative testing by HIV status that results in unrepresentative testing by age and residence. The results also suggest that despite evidence of large numbers of women from rural areas (where HIV prevalence tends to be lower) attending urban and peri-urban clinics, in an area with high contraceptive use and low fertility, resulting in selection bias due to age of sexual debut, ANC sentinel surveillance over-estimates prevalence. Understanding how to appropriately adjust ANC sentinel surveillance estimates to represent HIV prevalence in general populations is important since ANC clinics continue to be a relatively cheap and timely source of data.

The findings of this study highlight the possible biases inherent to both surveillance systems and suggest that attention should be paid to unrepresentative HIV testing, particularly by age and residency location. Analysing population-based HIV prevalence estimates, and comparing them to ANC-based estimates, also highlights the importance of not assuming population-based estimates equate to a gold standard.