Our knowledge about the course of the AIDS epidemic is largely based on national HIV surveillance systems using sentinel sites. In generalized epidemics, women attending antenatal clinics (ANCs) are the key population group for surveillance. National household surveys with biomarker collection have become increasingly popular: more than 20 countries in sub-Saharan Africa have conducted such surveys, but no country has conducted more than 1 survey to enable trends to be assessed from this source.1 Observational cohorts (longitudinal community studies without major interventions) are rare but provide a unique opportunity to assess long-term trends in the epidemic. The longest running community study is located in rural Uganda, which furnished a valuable perspective on data on prevalence changes based on the surveillance system.
Data from the ANC HIV sentinel surveillance system suggest that HIV prevalence was stabilizing in Tanzania at a time when the population had little access to antiretroviral therapy (ART).2 From ANC data alone, we cannot tell whether stabilizing prevalence is attributable to stable or declining incidence or is mainly attributable to deaths among HIV-infected persons exceeding new infections.3 Serial cross-sectional household surveys with HIV testing have a similar limitation.
This article presents data from a longitudinal community study in rural northwest Tanzania, which was initiated in 1994.4 The analysis presents updated estimates of HIV incidence and prevalence, using results from 4 surveys conducted over a decade, applying new estimation methods to obtain more robust incidence estimates, and describes how the HIV epidemic has unfolded in this population over the past decade.
The study area is located 20 km from Mwanza City (Tanzania's second largest city) and straddles the main road to Kenya. The entire population of Kisesa ward is regularly enumerated as part of an ongoing demographic surveillance system (DSS),4-6 which provides information on the size and structure of the population. The population grew from 19,350 in 1994, when surveillance started, to 26,330 in 2004. The contiguous area has been divided into 3 strata: “remote” rural villages, which accounted for 57% of the total in 1994 and 53% in 2004; roadside villages; and the central trading center. The last 2 strata have been grouped together in this analysis. Most of the population is Sukuma, the region's dominant ethnic group, and most are farmers.
Since 1994, when the DSS started, there have been 4 community-based surveys in which blood samples were taken for anonymous HIV testing, based on informed consent without result disclosure. These surveys also obtain information about HIV awareness, sexual behavior, and reproductive health.4-7
Awareness of HIV infection is high in this population: 98.2% of people have heard of HIV/AIDS, and 95.9% have accurate knowledge about modes of transmission. Since 1995, there have been a number of HIV prevention interventions in the study population. These include a school-based education program; mapping of high-risk areas for HIV transmission; the establishment of village HIV/AIDS action committees; health unit-based training; improved management of sexually transmitted diseases (STDs); condom promotion; and, since 2004, provision of voluntary counseling and testing (VCT) services in the ward. ART only became available after the last round of survey.
Epidemiologic serosurveys were carried out in 1994/1995, 1996/1997, 1999/2000, and 2003/2004. In each round, study participants provided information on behavioral and demographic characteristics and gave a blood sample for HIV testing. Residents of Kisesa ward aged older than 15 years were eligible for each survey. The upper age limit was 44 years in the first survey, increased to 46 years in the second survey, and had no upper age limit in the last 2 surveys. This allowed the oldest participants in the first survey to be included in all subsequent rounds. For each round, eligible residents were identified from the demographic surveillance data and invited to participate in the survey. In the first 3 rounds, villages were surveyed individually; in the fourth round, 2 villages were surveyed simultaneously. Respondents were asked to come to central locations in their village to be interviewed and to give blood for HIV testing. Respondents who did not attend were traced and encouraged to participate. Eligible respondents and their children were offered free medical care (but not ART) at temporary clinics, which were better supplied with essential drugs than the local dispensaries. Respondents gave verbal consent for interview and consent for their blood to be tested for HIV on the understanding that the results would not be disclosed to them and would not be linked to individuals on a named basis by the study staff. VCT was offered on-site in round 4; in earlier rounds, it was available from a counselor who followed the survey team.
All the surveys were given ethical approval by the Medical Research Coordinating Committee of the Tanzanian Ministry of Health.
HIV testing was carried out at a regional reference laboratory in Mwanza City. In the first survey, venous blood was collected; subsequently, dried blood spots from finger pricks were used. In the first 3 surveys, HIV testing was carried out using Vironostika HIV-MIXT (Organon, Teknika, Boxtel, The Netherlands) and Enzygnost HIV1/HIV2 (Dade Behring Marburg GmbH, Marburg, Germany). In the fourth round, Uniform 2 (BioMerieux, Boxtel, The Netherlands) and Enzygnost HIV1/HIV2 (Dade Behring Marburg GmbH) were used. In all survey rounds, samples for which enzyme-linked immunosorbent assay (ELISA) results were reactive were considered to be HIV-positive. Western blot tests were used to discriminate between discrepant results in round 1. In subsequent rounds, the 2 ELISA tests were repeated; if the results were still discrepant, the sample was excluded from analysis (27 samples).
Prevalence ratios are calculated for residents aged 15 to 44 years who participated in a particular survey round, generally excluding data for those aged 45 years and older, which are only available in later survey rounds. Individual test results can be matched on study numbers, which makes it possible to identify people who acquired HIV infection between 2 consecutive survey rounds and those who contributed person-years “at risk” but remained negative over the intersurvey intervals. Incidence rates are based on people who participated in 2 or more serosurveys. In addition, the incidence analysis includes young people who had attended only 1 survey and had been resident in Kisesa ward when the previous round was conducted but had been too young to attend the previous round. These young people, who were all aged younger than 15 years at the previous survey, were presumed to be uninfected at this earlier time point. The inclusion of these extra young people minimizes the degree of left-censoring in the 15- to 19-year-old age group, which would otherwise lead to an overestimate of incidence in this age group.
Survival analysis was used to calculate incidence rates. Person-years of exposure were calculated using time elapsed after the date of an individual's first negative HIV test result, except for the young residents who had been too young to participate in their village's previous serosurvey, for whom exposure is calculated as time elapsed from their 15th birthday. Exposed individuals who tested positive exited at the time of seroconversion. People who did not seroconvert are censored at the date of the last serosurvey they attended. Information on the serostatus of individuals older than 45 years of age is used to identify seroconverters and to compute person-years of exposure up to the age of 45 years, but the reported incidence rates are limited to individuals aged 15 to 44 years. We have also estimated the cumulated risk of acquiring HIV infection based on the life table proportion infected by the age of 45 years.
Overall HIV prevalence ratios and incidence rates were standardized using direct standardization for age and residence and are presented separately for male and female participants.
Because individuals can move between study villages, the intervals between consecutive survey rounds in which they participated may be quite wide (between 2 and 3.5 years) and an approximate date for seroconversion was imputed for each case, even for those who changed status after missing a survey round. This problem is common to most studies of HIV incidence and is often dealt with by taking the midpoint of the interval as the estimated date of seroconversion. This may not be an ideal solution when incidence is changing and the gap between tests is long and varies between individuals, however.8
To overcome the problems of interval censoring, we used a multiple imputation technique. Seroconverters were allocated a random date between their last negative test result date and their first positive test result date; the young Kisesa residents who tested positive at their first test were allocated a random date between the median date of the earlier survey in their village of residence (which they had been too young to attend) and their first test date. The random numbers were drawn in such a way as to produce a uniform distribution of seroconversion dates in each interval for the population as a whole. The process of random number allocation was repeated until the average 5-year age- and gender-specific incidence rates in each intersurvey interval and for both geographic strata did not change; this occurred after approximately 4500 runs. The random number allocation process was continued for an additional 5500 runs, with no further changes observed. The tabulated incidence rates and estimates of lifetime risk are based on the averages of these 10,000 repetitions.
The reported variance of the incidence rate estimates is the average variance for each run plus the variance in the incidence rates across all the runs.9 Uncertainty bounds for the estimated incidence rate were derived using the t-distribution.9 The serosurveys are village censuses of the entire adult population rather than sample surveys; thus, the “within-run” variance associated with each incidence ratio is 0. The uncertainty bounds for the incidence rates therefore represent only the uncertainty introduced by the imputation process.
Although the prevalence ratios are not affected by sampling or imputation errors, they may be affected by participation bias. We have estimated the possible effect of participation bias in 3 additive steps based on estimating the prevalence that would have been observed had nonattenders been included. First, we assumed that people who were negative at the last survey they attended had been negative in all the previous surveys for which they were eligible but did not attend and that those who tested positive in the last survey they attended were positive in all subsequent surveys that they missed. Second, it was assumed that eligible nonattenders with no evidence of seroconversion who had earlier tested negative continued to be negative at later surveys that they did not attend and that eligible nonattenders with no evidence of seroconversion who later tested positive were already positive at earlier surveys that they did not attend. Finally, those who never tested were assigned the directly measured prevalence ratio for the gender, age, and residence subgroup in each round for which they were eligible nonattenders. This last step has no effect on the specific prevalence rates in each subgroup but allows for the effect of the age, gender, and residence structure of the nonattenders on crude prevalence ratios. The bias limits given after each prevalence ratio estimate represent the highest and lowest values obtained by combining 1, 2, or 3 of these steps.
In total, 28,591 interviews and 28,523 HIV tests were carried out in the 4 surveys among adults aged 15 years or older. A total of 6448 adults (55% women) came to 2 or more surveys and were tested for HIV on both occasions. Table 1 gives the numbers attending each survey round. Four hundred thirty-two seroconverters were identified, of whom 46 were young Kisesa residents positive at their first test, and 56 had skipped a survey round when they had become infected.
Throughout the study, attendance rates are higher for women than for men and higher in the remote rural villages (attendance range for both genders: 70% to 86%) than in the roadside villages (range: 61% to 81%). The most common reason for nonattendance was temporary absence from home. Almost all those who completed the survey questionnaire also provided a blood sample for HIV testing. Change of residence out of the ward was the main reason for not attending more than 1 serosurvey: overall, 73% of those tested at serosurvey 1 tested again at some later date; however, among those who remained resident in the ward throughout the entire period of the study, 97% tested again. Similarly, although only 41% of those who tested at serosurvey 4 had tested previously, the proportion testing previously among those who had been age eligible and resident in the study area throughout was 98%.
When comparing prevalence trend and incidence levels, it should be noted that new infections are only one of many factors contributing to prevalence change. Prevalence is also affected by mortality of infected persons and by in- and out-migration of infected and uninfected individuals, and prevalence and incidence estimates are affected by changing participation rates. Figure 1 illustrates the magnitude of the main influences that determine overall prevalence change. Migratory changes dwarf all the others; however, in each of the intervals, overall in-migration of HIV-infected persons has been more or less balanced by out-migration of HIV-infected persons. This is not always the case in narrower age bands, however, in which it is also necessary to account for aging in and out of the age group. This can lead to unexpected relations between age-specific incidence and prevalence change in particular age groups.
In the 2003/2004 survey, HIV prevalence among adults aged 15 years or older was 7.5% for men and 8.2% for women compared with 6.9% for men and 7.8% for women in the previous survey. Data for the population aged 15 to 44 years, available for all 4 surveys, are shown in Table 2.
In the area as a whole, prevalence seems to be rising still for men but has declined slightly in recent years for women. Prevalence is lower in remote rural areas, but it has continued to rise in these areas for both genders, whereas in the area closest to the road and trading center, it has leveled off for men and fallen sharply for women. Overall, prevalence is little influenced by changes in the composition of the population between surveys; when standardized for age and residence, the standardized prevalence ratios (not shown) differed by <0.2% from the crude values.
HIV prevalence varies by age group, and the pattern of age variation differs between genders and between roadside and remote rural areas; furthermore, the age patterns change over time (Fig. 2). Prevalence is consistently higher across all ages in roadside residents compared with those living in more remote rural areas, but the gap has reduced over time. For women, prevalence peaks in the 25- to 34-year-old age group; however, for men, it continues to be high in the 35- to 44-year-old age group. It has increased over time among male and female residents younger than 20 years of age in both residence categories; however, for men, this increase has been recent (2000 to 2003), whereas for young women, the increase occurred mainly at the start of the study (1994 to 1997). In older age groups, prevalence was fairly steady in roadside villages in the first 3 surveys but fell at the most recent survey.
Incidence estimates are based on a total of 33,140 person-years of observation between the ages of 15 and 45 years, with 362 seroconversions observed between these ages. This includes 3690 person-years contributed by young residents in the interval before their first serosurvey attendance (31 seroconversions), and 2630 person-years contributed by those who skipped 1 or more serosurveys (32 seroconversions). The second panel of Table 2 shows crude incidence rates by gender, residence, and intersurvey interval. Incidence rose sharply between the first 2 intersurvey intervals in all areas, but roadside areas experienced a reversal of this trend in the last interval, with a particularly dramatic fall among women, in contrast to remote rural areas, where incidence has leveled out for men but continues to rise for women.
Figure 3 shows that women experience a younger incidence pattern than men, with higher rates than men in the 15- to 24-year-old age group and relatively little difference between incidence rates in the 15- to 24-year-old and 25- to 34-year-old age groups. This contrasts with the male age pattern, which generally peaks in the 25- to 34-year-old age group. The incidence age pattern has become less differentiated by age in rural areas, with little overall change in level; in urban areas, the pattern has become older during the recent decline; and differences between the areas have decreased.
Incomplete coverage of the study population might affect incidence estimates if individuals who attend at least 2 serosurveys are disproportionately concentrated in subgroups with unusually high or low incidence rates. The number of HIV-negative persons who attended only 1 of the first 3 serosurveys was 4338; of these, 894 were eligible to attend at least 1 other serosurvey. They contributed a total of 5008 person-years during the time that they were under observation according to the demographic surveillance. Had they experienced the gender-, age-, and residence-specific incidence rates observed in their groups, their crude incidence rate would have been 12.6 per 1000 over the study period as a whole compared with 10.9 per 1000 observed in those who attended 2 or more serosurveys. The difference arises because those who came to only 1 serosurvey are mainly found in the more mobile populations in the trading center and roadside villages. The inclusion of these extra person-years with their expected incidence would raise the overall incidence to 11.1 per 1000.
The overall effect of differences in incidence patterns is clearly seen in the cumulative probability of infection shown in the last panel of Table 2. This life table measure is independent of the structure of the population at risk and represents the probability that a person experiencing prevailing risks of infection between 15 and 45 years of age would have become infected by the age of 45 years. The contrasting age patterns between the genders generally cancel out by the age of 45 years, with cumulated infection being almost the same for men and women, except for the period between surveys 3 and 4, where women have an overall advantage over men in the roadside areas, whereas men have an advantage in rural areas. In recent times, overall levels in rural and roadside villages have converged as a result of this contrasting trend between the genders.
HIV prevalence among adults in Kisesa ward increased from 6.5% to 8.3% between 1994 and 2000 and has remained at the same level until 2003/2004. The incidence data do not show a similar time trend. The crude prevalence of HIV infection in Kisesa ward in 2003/2004 (7.5% for men and 8.8% for women) was reasonably close to the crude prevalence of HIV in Mwanza Region, estimated in the 2003/2004 national AIDS Impact Survey as 7.5% (4.1% to 10.8%) for men and 7.0% (3.6% to 10.4%) for women.10 Unlinked anonymous HIV testing carried out in Kisesa ANCs in 2000 produced an estimate of 4.6% in clinics serving remote rural areas and 11.3% in the roadside clinic.11 This is lower than the serosurvey 3 prevalence estimate for women in 2000 (5.7% in remote rural areas and 12.4% in the roadside clinic) because of lower fertility among HIV-infected women.
The assessment of national trends in HIV prevalence in Tanzania is hampered by lack of data from a sufficient number of clinics providing data continuously over a prolonged period. A general assessment of the trend based on the ANC-based HIV surveillance system suggested that the epidemic reached a peak at 8.1% HIV prevalence among adults aged 15 to 49 years in 1995 and had decreased to 6.5% in 2004.12 Our analysis suggests that trends in different parts of the country may have followed divergent paths.
The main factor influencing the accuracy with which our prevalence estimates reflect true prevalence in Kisesa ward is attendance at serosurveys. If survey participation is in any way associated with HIV status, this could bias our results. Our attempts to predict HIV status of nonattenders, using past or future known HIV status and predictions based on age, gender, and residence for those who never attended, have produced a range of crude prevalence estimates (see Table 2) that encompass the prevalence measured among those who did attend in every case, however. Our predictions for nonattenders were based solely on demographic characteristics available from the regular household censuses and not on behavioral data, because the latter were collected at the same time as serostatus was determined. Overall, the observed prevalence lies slightly closer to the lower bound of the predicted range (average deviation of 0.5) than to the upper bound (average deviation of 0.6), but there is no indication of a strong bias in either direction.
We found that failure to attend a follow-up serosurvey by eligible persons was more common in persons resident in the roadside and trading center area, who might be expected to have higher than average incidence. The person-years of missed follow-up for eligible persons who did not attend a follow-up survey (5008) compared with the person-years observed for those who did attend a subsequent survey (33,140) is relatively small, however, such that the downward bias in our incidence estimate is probably only on the order of 0.2 per 1000 person-years.
Multiple imputation of seroconversion dates does not suffer from the same bias as midpoint estimation but introduces random error, because each person is unlikely to have been allocated the true date of seroconversion. This random error does not have any effect on the number of seroconversions and has only a negligible effect on overall exposure time, but it could affect the classification by age group at seroconversion. The uncertainty range attributable to interval censoring (which indicates the effect of a particular seroconversion falling outside the age range of interest [15 to 44 years] or taking place in a later or earlier intersurvey period in the case of those who became infected in an interval that included a missed survey round) is narrow, suggesting that imputation has not had an appreciable effect on these results. A comparison of the different methods for estimating seroconversion dates found that the method we used performed less well than conditional mean imputation;13 however, that method needs additional assumptions about the distribution of seroconversion times by calendar year.
Midpoint imputation of incidence rates produced some extreme and improbable estimates for certain age and gender groupings, which further justifies the choice of an alternative imputation method.
Our analysis highlights the problems of relying on prevalence data to gauge the general trend of the epidemic. For example, in the most recent interval, prevalence rose for men and fell for women (by 0.6% and −0.5% points, respectively), whereas crude incidence rates and cumulated infection risk indicators were virtually identical for both genders. Prevalence data become even less reliable as an indicator of epidemic spread if ART roll-out is successful; hence, the importance of continuing to collect high-quality incidence data.
Our data show that the gap in incidence level (as measured by the lifetime risk of infection indicator) between rural and roadside communities is narrowing, mainly because women in rural areas have recently experienced higher infection risks. This information should prompt further research to try to discover whether HIV prevention messages are reaching these women, whether they are engaging in risky behavior, or whether this trend could be attributable to their increasingly becoming the sexual partners of choice of men from the roadside villages, of whom a relatively large number are infected.
The fact that incidence seems to be falling in roadside areas is an encouraging sign, but the continued gradual rise in incidence in “remote” rural areas is worrying, especially because most (66%) of the Kisesa population lives in these areas. There is an urgent need to promote and expand access to the existing HIV prevention efforts such as behavioral change, including abstinence and condom use, VCT services, and early diagnosis and treatment of STDs, and to ensure that these services reach the most remote rural parts of the country. The roll-out of ART affords an opportunity to strengthen prevention messages in a destigmatizing context of providing treatment, and it is equally important to ensure that these treatment services also reach rural areas.
ART started to become available in Kisesa ward in the last quarter of 2005, after the end of the fourth serosurvey, as a result of referrals by the Tanzania AIDS Monitoring Activities (TAZAMA) project (funded by the Global Fund for AIDS, Tuberculosis, and Malaria). The referrals were initiated after a short qualitative study into “perceived barriers to ART access”.14 When the results of the fifth serosurvey (for which field work was completed in July 2007) become available, we should be able to judge whether access to ART has had any impact on incidence trends.
The authors thank the Principal Secretary, Ministry of Health, and the Director General, National Institute for Medical Research, for permission to carry out the study. They are grateful to all the respondents for their valuable time and the information they furnished. They also thank the dedicated field workers.
1. García-Calleja JM, Gouws E, Ghys PD. National population based HIV prevalence
surveys in sub-Saharan Africa: results and implications for HIV and AIDS estimates. Sex Transm Infect
. 2006;82(Suppl 3):iii64-iii70.
2. National AIDS Control Programme. Surveillance of HIV and Syphilis Among Antenatal Clinic Enrollees 2003-2004
. Dar es Salaam, Tanzania
: The United Republic of Tanzania
Ministry of Health Tanzania
3. Wawer MJ, Serwadda D, Gray RH, et al. Trends in HIV-1 prevalence may not reflect trends in incidence in mature epidemics: data from the Rakai population-based cohort, Uganda. AIDS
4. Mwaluko G, Urassa M, Isingo R, et al. Trends in HIV and sexual behaviour in a longitudinal study in a rural population in Tanzania
, 1994-2000. AIDS
5. Boerma J, Urassa M, Senkoro K, et al. Spread of HIV infection in a rural area of Tanzania
6. Urassa M, Boerma J, Isingo R, et al. The impact of HIV/AIDS on mortality and household mobility in rural Tanzania
7. Bloom S, Urassa M, Isingo R, et al. Community effects on the risk of HIV infection in rural Tanzania
. Sex Transm Infect
8. Law CG, Brookmeyer R. Effect of mid-point imputation on the analysis of double censored data. Stat Med
9. Little RJA, Rubin DB. Statistical Analysis with Missing Data
. 2nd ed. Hoboken, NJ: John Wiley and Sons; 2002.
Commission for AIDS (TACAIDS), National Bureau of Statistics (NBS), ORC Macro. Tanzania HIV/AIDS Indicator Survey 2003-04
. Calverton, MD: TACAIDS, NBS, and Opinion Research Corporation (ORC) Macro; 2005.
11. Urassa M, Kumogola Y, Isingo R, et al. HIV prevalence
and sexual behaviour changes measured in an ante-natal clinic setting in northern Tanzania
. Sex Transm Infect
12. Somi GR, Matee MI, Swai RO, et al. Estimating and projecting HIV prevalence
and AIDS deaths in Tanzania
using antenatal surveillance data. BMC Public Health
13. Geskus RB. Methods for estimating the AIDS incubation time distribution when date of seroconversion is censored. Stat Med
14. Mshana G, Wamoyi J, Busza J, et al. Barriers to accessing antiretroviral therapy in Kisesa, Tanzania
: a qualitative study of early rural referrals to the national programme. AIDS Patient Care STDS