Baryarama, Fulgentius MSc, PhD*; Bunnell, Rebecca E ScD, Med*; Montana, Livia MA†; Hladik, Wolfgang MD, MSc*; Opio, Alex MDChB, MSc, PhD‡; Musinguzi, Joshua MBChB, MPH‡; Kirungi, Wilford MBChB, MSc‡; Waswa-Bright, Laban MStat§; Mermin, Jonathan H MD, MPH*
Current rapid expansion of HIV prevention, care, and treatment services requires accurate surveillance data to target HIV intervention efforts to at-risk populations. Although population-based surveys can provide excellent data for intervention planning, they are very expensive and require high-level human resources. Despite increased funding, resources remain insufficient to meet treatment needs and any diversion of resources poses difficult dilemmas. Routinely collected programmatic data that can be used for surveillance purposes have the advantage of being available for little or no extra cost, if existing health management information systems collect vital data and allow deduplicating of repeat clients. However, there are concerns about the generalizability of program data due to the potential participation bias associated with program data. Where the program uptake is high and records can be deduplicated, such bias may be minimal,1 and program data could be used as an additional data source for surveillance.
The percentage of the adult population receiving voluntary counseling and testing (VCT) is still low but increasing in most parts of sub-Saharan Africa. Use of VCT data is still limited by concerns of participation bias, the effects of which could vary by VCT setting. One study in Uganda compared HIV prevalence among women who were offered testing in a prevention of mother-to-child transmission (PMTCT) program with that obtained anonymously from women who declined PMTCT.1 HIV prevalence was higher among women consenting to participate in PMTCT compared with anonymously tested women, especially when the uptake of PMTCT was less than 70%. In a study in Britain, VCT was highly associated with higher risk behaviors, or subgroups at higher risk, and was recognized as a key strategy to reduce undiagnosed prevalent HIV infections.2 Likewise, in Burkina Faso, among women participating in PMTCT, the VCT acceptance rate was 18%. Accepting VCT was associated with the number of previous pregnancies and miscarriages and hence higher risk of HIV infection,3 although this may be confounded by age and exposure.
However, other studies have detected minimal or no differences in HIV prevalence between VCT clients and the general population. A study in Uganda comparing HIV prevalence trends from VCT to antenatal care (ANC) sites in the same towns found that prevalence among the 15- to 19-year-old VCT women was lower than ANC prevalence, VCT prevalence among women older than 24 years was higher than ANC prevalence, and VCT and ANC prevalence for the 20- to 24-year-old women was similar.4 Another study in Northern Uganda concluded that the prevalence based on VCT data collected as part of the PMTCT program could be used for HIV surveillance.5 Moreover, a population-based VCT study in rural southwestern Uganda found that certain higher risk subpopulations were underrepresented among VCT clients.6 In addition to increased availability of program data, VCT data might need to satisfy certain criteria before they can be used for routine surveillance of HIV. These criteria include data quality, representativeness, and sufficient knowledge of the magnitude and direction of bias as a result of self-selection.7
In a Joint United Nations Programme on HIV/AIDS (UNAIDS)/World Health Organization reference group meeting for improving estimates and projections, it was recommended that more detailed analysis of VCT versus population prevalence is needed before VCT data can be used for any surveillance purposes.8 This article compares HIV prevalence estimates from a national VCT program with those in a population-based national serosurvey conducted from August 2004 to January 2005 in Uganda. We also assess biases in VCT data and identify subpopulations of VCT data that generate plausible prevalence estimates that accurately represent prevalence estimates of those subpopulations.
Uganda HIV/AIDS Sero-Behavioral Survey
We analyzed data from the 2004-2005 Uganda HIV/AIDS Sero-Behavioral Survey (UHSBS), a nationally representative, population-based survey involving adults 15-59 years old. Data were collected on behavioral, social, and demographic indicators, and blood samples were obtained for testing of HIV, syphilis, herpes simplex virus 2, and hepatitis B virus. Separate informed consent was provided by respondents for interviews and blood sampling.
The survey collected information from 9529 households in 417 sample enumeration areas using a 2-stage cluster sample. A total of 11,454 women and 9905 men aged 15-59 years were eligible for individual interviews and blood sample collection. Of the eligible women and men, individual questionnaires were completed for 95% of women and 89% of men and blood specimens were collected from 90% of women and 84% of men. Testing for biomarkers was conducted using standard testing and quality control procedures. HIV testing was done using 2 HIV enzyme immunosorbent assays (Murex HIV.1.2.O; Abbott Diagnostics, Chicago, IL + Vironostika HIV Uniform II plus O; BioMerieux, Paris, France; for serum testing, same Vironostika + Genetic Systems rLAV enzyme immunosorbent assay (EIA); BioRad Laboratories, Hercules, CA), based on different antigens. Specimens with discordant test results were retested with the same EIAs, and, if still discordant, were resolved by Western blot (HIV Blot 1.3; Genelabs Diagnostics, Singapore, Malaysia). For quality control, all positive specimens and 5% of negative specimens were retested at a different laboratory using the same testing algorithm. Results of retesting for quality control indicated a very high overall concordance with the original test results, with retesting data suggesting a rise in overall HIV prevalence of less than 0.1%.
The test results for individuals were anonymously linked to individual and household questionnaire information through bar codes. The survey protocol was cleared by the Uganda National Council of Science and Technology and the Centers for Disease Control and Prevention and approved by the Institutional Review Boards of the Uganda Virus Research Institute and ORC Macro.
VCT data were routinely collected from all clients visiting the AIDS Information Centre (AIC), a nongovernmental organization formed in 1990 to offer VCT services.9 In 2004-2005, over the same UHSBS period, AIC had 8 stand-alone VCT centers that were not attached to or integrated into any previously established health unit and 152 health facility-based sites (indirect sites) mostly situated in government health centers and hospitals. The stand-alone sites operated 6 days a week and, in addition to HIV testing, provided all clients with syphilis screening, education on sexually transmitted infections and tuberculosis, and an opportunity for psychosocial support through post test clubs. There was cost sharing at stand-alone sites with an exemption policy for clients who were unable to pay, whereas services at most indirect sites were free. HIV testing for VCT clients was done on-site using a 3-assay serial rapid testing algorithm using Capillus HIV-1/HIV-2 (Trinity Biotech, Bray, Ireland) as a screening test, SeroCard (Trinity Biotec) as a confirmatory test, and Determine (Abbot Laboratories, Tokyo, Japan) as a “tiebreaker” in 2004. In 2005, SeroCard was replaced by Stat-Pack (ChemBio Diagnostics Systems Inc., Medford, NY) as a confirmatory test. The Uganda Virus Research Institute, a national HIV reference laboratory, provided quality assurance by retesting all specimens that reached a tiebreaker and a 3% random sample of the remaining specimens. For repeat testers, only the client's most recent visit was included in the analysis. Repeat testers were deduplicated using a unique client identification number that has been in use since 1998. Completeness of previous testing data was found at 97% in 2003. The analysis was restricted to clients aged 15-59 years, excluding clients who stated that illness was their reason for testing.
Geographical Information System Methods
Locations of the 160 VCT sites were obtained from the World Health Organization HealthMapper 4.2 database, and those missing from the HealthMapper database were georeferenced using parish-level data from the Uganda Bureau of Statistics. Place of residence of VCT clients was not available for this study; therefore, a straight-line distance buffer was drawn around each VCT site to approximate the catchment area. Buffers of 15 and 30 km radius were drawn around each VCT site to define the catchment areas. The UHSBS was carried out in 417 enumeration areas or clusters. Each cluster was georeferenced to the parish in which it was located, using the Uganda Bureau of Statistics parish-level geographic data file. Using ArcGIS 9.1, we then identified the UHSBS sample clusters, which fell within the 30- and 15-km buffers around each VCT site. Of the 417 UHSBS clusters, 83% were located within 30 km of the nearest VCT site and 56% within 15 km. The 30-km buffer was not used because it included almost all the survey populations.
VCT clients who reported illness as a reason for HIV testing were excluded from the analysis to minimize participation bias due to overrepresentation of those who suspected their HIV-positive status due to illness. Analyses were disaggregated by sex, 5-year age groups, and type of residence (rural or urban). Urban areas were defined as all recognized towns, municipalities, and cities in Uganda. All other areas were classified as rural. We used age-standardized HIV estimates with UHSBS as the standard population. VCT and UHSBS HIV prevalence point estimates were compared using 95% confidence intervals (CIs) for a proportion. Further comparisons were made with UHSBS HIV prevalence in the 15 km catchment radius, VCT clients at stand-alone sites, and VCT clients at health facility-based sites. Point estimates with overlapping CIs were interpreted as similar, whereas point estimates with nonoverlapping CIs were interpreted as significantly different.
UHSBS clusters were in all the parts of Uganda, although urban areas were oversampled. At the time of the survey, AIC had no VCT sites in large areas of central and western Uganda and in some isolated portions of northern Uganda (Fig. 1). Characteristics of the 75,640 eligible VCT clients and 18,525 UHSBS participants are presented in Table 1. The proportion of persons living in urban areas was 38% in the VCT population and 16% (unweighted) in the survey population. Geographical distribution by region was similar in the VCT and survey populations, except that the eastern region was overrepresented (VCT: 19%, UHSBS: 10%) and the western region was underrepresented (VCT: 6%, UHSBS: 11%) in the VCT population. The proportion by 5-year age groups showed underrepresentation of the 15- to 19-year age group (VCT: 16%, UHSBS: 22%) and overrepresentation of the 20- to 24-year age group (VCT: 26%, UHSBS: 16%) in the VCT population. Older age groups were underrepresented in the VCT population. Forty-eight percent of VCT clients had secondary or higher education compared with 23% of survey participants. VCT clients were more likely to have been previously tested for HIV (23%) than survey participants (13%). The characteristics of the UHSBS participants in survey clusters within 15-km catchment areas of the VCT sites were similar to those of all UHSBS participants, except that the catchment area participants were more likely to be urban residents (28% versus 16%) and more likely to be of secondary education or higher (29% versus 23%).
VCT clients at stand-alone sites were mostly of secondary education or higher (70%) compared with facility-based (40%) or UHSBS (23%) clients. VCT clients at stand-alone sites were also mostly never married (53%) compared with clients at facility-based (34%) or UHSBS (27%) sites. Previous HIV testing was more common among clients at stand-alone sites (42%) compared with facility-based (16%) or UHSBS (13%) sites.
Overall, HIV prevalence was 5.9%, 95% CI 5.6 to 6.2 in the UHSBS population and 8.8%, CI 8.7 to 9.1 in the VCT population (Table 2). Survey participants living within a 15 km radius of a VCT site had an HIV prevalence of 5.6%, CI 5.2 to 6.1, which was similar to the prevalence among all UHSBS participants, but significantly lower than that among VCT clients. Prevalence at stand-alone sites was 10.6%, CI 10.2 to 11.1 and was 8.3%, CI 8.1 to 8.5 at facility-based sites (Table 3).
HIV prevalence among UHSBS participants and among VCT clients was similar in urban areas (UHSBS: 9.7%, CI 8.6 to 10.7; VCT: 10.1%, CI 9.8 to 10.5) and for both urban men (UHSBS: 6.3%, CI 4.9 to 7.6; VCT: 7.1%, CI 6.6 to 7.5) and urban women (UHSBS: 12.2%, CI 10.6 to 13.7; VCT: 12.9%, CI 12.3 to 13.4). The prevalence was also similar for each of the 5-year age groups for both urban men and urban women, except for adolescent men aged 15-19 years (UHSBS: 0.3%, CI 0.0 to 1.0; VCT: 1.9%, CI 1.2 to 2.5). Further analysis showed (Table 3) that UHSBS HIV prevalence for urban residents (9.7%, CI 8.6 to 10.7) was similar to VCT prevalence at urban stand-alone sites (10.3%, CI 9.8 to 10.8) and for clients testing at urban health facility-based sites (10.0%, CI 9.5 to 10.4) for both men and women but showing more marked differences by age groups.
However, the VCT-based HIV prevalence estimates among rural residents were much higher than among rural survey participants (UHSBS: 5.2%, CI 4.8 to 5.5; VCT: 8.2%, CI 7.9 to 8.4), for both rural men (UHSBS: 4.5%, CI 4.0 to 5.0; VCT: 7.2%, CI 6.9 to 7.6) and rural women (UHSBS: 5.7%, CI 5.2 to 6.2; VCT: 8.9, CI 8.6 to 9.3), and in most 5-year age groups. Clients of rural residence tested at stand-alone sites had high prevalence (11.10%, CI 10.3 to 11.9) comparable to urban clients testing at either stand-alone (10.3%, CI 9.8% to 10.8%) or health facility-based sites (10.0%, CI 9.5% to 10.4%). Clients of rural residence tested at health facility-based VCT sites showed the lowest VCT prevalence (7.7%, CI 7.5 to 8.0), which was however higher than the rural UHSBS prevalence of 5.2%, CI 4.8 to 5.5.
In urban Uganda, HIV prevalence from a large general population-based survey was similar to prevalence among VCT clients who did not report illness as a reason for testing. This could be due to relatively high VCT coverage in urban areas arising from better access to VCT services and the stage of the HIV epidemic in Uganda.
Rural VCT clients had higher HIV prevalence than rural UHSBS participants. This led to a higher HIV prevalence estimate among all (urban and rural combined) VCT clients compared with the UHSBS prevalence estimate. The higher HIV prevalence among rural VCT clients compared with UHSBS clients suggests a self-selection bias toward higher proportions of high-risk rural VCT clients or could be a result of mobility of high-risk individuals in rural areas that were not captured in the UHSBS. The selection bias may be related to poor accessibility of VCT services in rural areas and low awareness and education levels. Hence, use of VCT data for a national estimate of HIV prevalence may require an adjustment factor for rural areas. From this analysis, the UHSBS:VCT rural prevalence ratio was 0.63, CI 0.53 to 0.72. With improvements in VCT coverage in rural areas, this adjustment factor may change over time, hence limiting the use of VCT data for monitoring trends in HIV prevalence.
The HIV prevalence at stand-alone VCT sites among rural clients was higher than the corresponding prevalence at rural facility-based VCT sites and similar to the urban prevalence at either stand-alone or facility-based sites. Although this could be a result of urban clients being misclassified as rural clients or peri-urban clients mostly classified as rural, it is likely that this is evidence of self-selection bias among rural clients. The observed lower rural serosurvey prevalence in the 15-km catchment areas (but higher rural VCT prevalence) than general serosurvey prevalence is further evidence of self-selection bias among rural VCT clients.
Interpretation of our results is subject to several limitations. First, one of the major limitations of VCT data is self-selection bias. Our analysis shows that these biases were more marked in rural areas than in urban centers. We minimized these biases by excluding clients who gave illness as a reason for testing in the VCT population and deduplicating VCT clients to retain only the client's most recent visit. This limitation implies that although this analysis shows that overall urban prevalence and urban prevalence for men and women separately derived from either stand-alone or facility-based VCT data are good estimates of the corresponding urban population prevalence in Uganda, further breakdowns by other sociodemographic characteristics may not provide good estimates. Moreover, urban VCT prevalence may be modified by intensive prevention and treatment interventions and various methods of delivery of VCT. Second, our results on age-specific comparisons were compromised by large CIs due to small numbers. Third, this analysis is subject to several data quality issues. For example, the address of residence up to parish level or distance from the VCT center was not available. Such spatial data would help to provide a better understanding of accessibility of VCT services. In addition, reliance on the unique identification number could not remove all repeat testers if people did not admit having previously tested. And removing clients who gave illness as reason for testing could potentially lower VCT prevalence below community-level prevalence. However, previous analyses found the unique identification number highly reliable and the exclusion of clients reporting illness a reasonable assumption.4,10 Last, though AIC is the largest provider of VCT in Uganda, it is not the only VCT provider. The other main VCT provider in Uganda is the Ministry of Health. Comparing the combined VCT data from AIC and the Ministry of Health would provide a more complete countrywide comparison.
Despite these limitations, our findings underscore the usefulness of VCT data, not only as a planning tool for targeting interventions for HIV prevention and providing a gateway for HIV/AIDS care through referrals but also as a tool to help monitor the HIV epidemic in urban areas in Uganda. Further research may tell how VCT data can be used to reliably track the epidemic in rural areas, where the majority of the population lives. Using VCT data for monitoring the HIV epidemic requires further evaluation in other settings and development of international guidelines for key VCT data collection questions and data quality control measures. Collection of geographical information system data to study the spatial distribution of VCT uptake can improve the understanding of geographical distribution and help monitor the trends in the epidemic.
ORC Macro and Institute for Global Health provided technical support.
© 2008 Lippincott Williams & Wilkins, Inc.