Selective participation in a study can potentially skew estimates of the outcome of interest in a study population [1–5]. This is more likely to be the case if the circumstances that influence low participation are in some way related to the main outcome. Nonresponse in HIV serosurveys is mainly due to refusal to provide a blood sample for HIV testing or absenteeism of the sampled individual during the survey period. Several population-based HIV seroprevalence studies have reported varying nonresponse rates for HIV testing, ranging from as low as 5% among men in Rwanda to 56% in Lesotho [2, 3]. A moderate nonresponse rate (14.4%) for HIV testing for Kenya was reported in an earlier survey. From the studies carried out to date on this topic, it has been shown that the effect of participant nonresponse on HIV prevalence estimates vary by certain characteristics, such as gender and residence, among others. Yet in general, the overall effect on national estimates is small, unless the level of nonresponse is very high, as was the case in Lesotho [1–3, 6].
HIV/AIDS remains a highly stigmatized disease, with many people preferring either not to know their status or to keep it a secret [7, 8]. The preference of an individual not to participate in a serosurvey may partly be influenced by the fear of knowing his or her own HIV serostatus. On the other hand, those who know their status as positive may participate in a serosurvey in the hope that they can be helped, or they may choose not to participate as they see no immediate benefit. Personal perceived risk may be correlated with actual risk of HIV infection . Perceptions about HIV risk are unlikely to be random among individuals in a population; they are likely to vary by defined individual characteristics, such as race, religion, ethnicity, and past behaviors, including experience with drug use or sex work . For that reason, if participants who perceive themselves to be at a higher risk of contracting HIV do not participate in a serosurvey, then prevalence estimates may be biased downward and might affect the overall estimate.
Interviewers may fail to make contact with a sampled person for a number of reasons, including temporary absence, work patterns, inability to locate the household/structure in which the sampled person lives, and out-migration. Highly mobile individuals, such as long-distance truck drivers, security personnel, and migrant workers, often have a different level of exposure to the risk of HIV [11–14]. In highly mobile populations, many sampled individuals may not be contacted, even if a good random sample is drawn. If a population has a substantial proportion of highly mobile individuals who miss out on a seroprevalence study and yet are likely to be at a higher risk, the estimates are likely to be biased downward as less mobile and low risk individuals are overrepresented in the effective sample interviewed [2, 3]. On the other hand, if a majority of a community's residents are migrant workers who live away from their families, they are likely to be exposed to higher risks of HIV infection. To the extent that such individuals are overrepresented in a seroprevalence survey, estimates are likely to be biased upward.
The slum context
Although informal settlements in Nairobi city are home to more than 60% of Nairobi's population , the informal nature of housing is likely to lead to underrepresentation of the slum population in national surveys, given the difficulty involved in listing temporary housing structures. Until the project on which this paper is based was conducted, HIV prevalence in the informal settlements was unknown. Kenya has had at least two large population-based HIV testing surveys [16, 17]. The Kenya Demographic and Health Survey of 2003 put the HIV prevalence estimate for Nairobi province at 10%. Nyanza province had the highest prevalence rate at 15%, and the national prevalence rate was 6.7% . There were differences in HIV prevalence rates by age, gender, ethnicity, rural-urban residence, educational attainment, and wealth status. These differences have been observed in several other surveys in sub-Saharan countries [2, 3, 16]. A more recent survey, the Kenya AIDS Indicator Survey 2007, estimated the national prevalence to be 7% and Nairobi province's prevalence rate to be 9% . However, the national surveys are unable to provide HIV prevalence estimates for slums. Earlier behavioral research indicates that high-risk sexual practices are prevalent in the informal settlements of Nairobi [18, 19]. Furthermore, recent work using verbal autopsies to establish causes of death, without HIV status, showed that HIV/AIDS and tuberculosis accounted for more than 50% of the adult mortality burden in the slums .
The African Population and Health Research Center (APHRC), in partnership with the Kenya Medical Research Institute (KEMRI), carried out a survey to estimate the prevalence and risk factors for HIV in two informal settlements in Nairobi city. The two communities where the project was carried out are informal settlements characterized by poor housing, lack of clean water, poor sanitation, unemployment, poverty, and overcrowding. Viwandani slum is located very close to the city's industrial area and is home to many low-income youths working in the industries close by. Korogocho is a more established slum settlement with a high proportion of men living with their spouses and children. Korogocho residents are predominantly either very low-income earners or unemployed. Additionally, residents of Viwandani are relatively more educated than those of Korogocho.
The survey, like many community-based surveys, faced a challenge of nonresponse, with a sizeable proportion of sampled individuals being nonresponders (43%). The desire to understand the effect of nonresponse on prevalence estimates was the basis for this paper. We hypothesised that the HIV prevalence estimate in the survey was underestimated due to low participation of highly mobile community members. Specifically, this paper aimed to describe selective participation in the serosurvey by sociodemographic characteristics and also to estimate the bias in the estimates of HIV prevalence.