The different data sources and methods used to estimate HIV prevalence trends indicated overall falling HIV prevalence among young women in Zambia but also highlighted geographical differences. At provincial level, the ANC data showed substantial geographical variations in the prevalence trends, with declines ranging from between 10% and 68% among young urban women, and from stability in three provinces to 86% decline among young rural women. Although the point prevalence tended to differ between ANC data and population-based data sources, the overall results of the ANC were in agreement with the changes observed between the two population-based ZDHS. Furthermore, the more educated young pregnant women had substantial falling prevalence trends, whereas the less educated had almost stable HIV prevalence. Similar results were reported in the population-based studies in Chelstone and Kapiri Mposhi 
In our study we used HIV prevalence estimates among young women aged 15–24 years as an indicator of HIV incidence. The only study that has attempted to estimate incidence in Zambia directly was the “Microbicide clinical preparedness study”, which drew its sample population from two communities in Lusaka 
. The incidence estimates from this study (47 per 1000 person years) was much higher than national incidence estimates derived from mathematical modelling of pooled urban and rural ZDHS data (approximately 17 per 1000 and 12 per 1000 person years among 15–19 and 20–24 year old women, respectively) 
. This difference is obviously due in part to the participation and selection criteria of participants, or to the differences in incidence between urban and rural women. A comparison with the HIV prevalence estimates from ANC surveillance and from the other studies conducted in Lusaka thus indicates that women recruited for the cohort study may have had a higher incidence than women in the ANC surveillance or the general population. This was expected since only sexually active women were included, and probably women who joined the microbicide study were more likely to perceive themselves as being at a heightened risk of HIV infection.
Our study also showed that declines in HIV prevalence varied by urban-rural residence and educational attainment. The ANC data indicated that HIV prevalence decline among young urban women in Zambia had started by the mid-1990s, whereas declines among young rural women became clearly evident only after 2004. The later change among rural residents may reflect differences in intensity and outreach of prevention campaigns in rural versus urban areas. For both urban and rural young women with more than seven years of educational attainment, a sharp drop in HIV prevalence can be observed from the mid or late 1990s. These results are in line with other studies on the association between HIV prevalence and educational attainment in sub-Saharan Africa 
, including Zambia 
. In addition, a similar pattern of marked decrease in syphilis prevalence among educated women was observed in the same ANC data 
. A likely explanation for this change is that educated people, once equipped with knowledge about HIV from prevention campaigns, have been quicker in changing their sexual behaviour 
Provincial ANC data suggest substantial inter-provincial differences in magnitude and trends of HIV prevalence. Most provinces in Zambia had either a stable or a decreasing HIV prevalence among both urban and rural residents. Lusaka, Northern and North-Western provinces recorded consistent and significant declines in HIV prevalence among urban ANC attendees during the period. Multiple factors are likely to explain the geographical differentials in magnitude and trends of HIV infection in Zambia. Variation in coverage and intensity of preventive efforts is an example. Furthermore, cultural and socio-economic factors are likely to have contributed substantially 
. For example, the interplay of structural and individual factors in Lusaka (i.e. relatively higher intensity of HIV preventive programmes and educational attainment) may have fostered sexual behaviour change, resulting in the decline reported here. Better understanding of the factors underlining these differentials in magnitude and trends might be critical for proper guidance of future preventive efforts.
Comparison of ANC and ZDHS data reveals that the provincial estimates for young urban women were similar in terms of direction (except in three provinces) but slightly different in terms of magnitude of change. Among young rural women, the provincial estimates of change differed both in direction and magnitude. It is likely that the differences seen were partially attributable to the small sample sizes of young people at provincial level in the ZDHS, since the provincial estimates were at least more in line with the magnitude and direction of change among urban women aged 15–49 years. HIV prevalence trends for men and women tend to be parallel 
, as seen in the population-based trends in selected urban and rural population 
and other data sources, indicating that sexual risk taking among men in Zambia has declined during the same period (paper in progress); we believe that the apparent increase in HIV prevalence among young men in the ZDHS is also likely to be an artefact due to small sample sizes. However, if another round of the DHS shows differences in trend estimates for men and women, improved HIV surveillance of men should be considered.
In the Ndola data, the ANC-based estimate was similar to the crude HIV estimate of the overall general population but it underestimated the prevalence of females aged 15–29 years by 25%, possibly indicating that HIV infection may affect fertility at a young age 
. Another factor influencing fertility is contraceptive use. Modern contraceptive use among women in Zambia's general population has increased gradually from 8.9% in 1992, to 14.4% in 1996, to 22.6% in 2001/02 and to 32.7% in 2007 
. In a situation where use of contraceptives in the general population is very high, the use of HIV prevalence of young women as a proxy of incidence can be less reliable because women may be sexually active for many years without becoming pregnant, leading to selection biases in the antenatal surveillance data. Hormonal contraceptive use has been linked to increased risk of HIV infection 
. However, those who take contraceptives to prevent pregnancy might also take other precautions that put them at lower risk of HIV (e.g., condom use or having fewer sexual partners) than those who are sexually active and become pregnant. In line with this, educated women have been found to be both more likely to use contraceptives 
and to postpone sexual debut, leading to a strong association between educational attainment and reduced fertility 
. Since HIV prevalence declines were biased towards higher educational attainment in Chelstone, postponement of first pregnancy among women with high education seems the most plausible explanation for the finding that ANC-based HIV prevalence trends substantially underestimated the actual declines in HIV prevalence in the general population 
. The prevalence declines were also clearly biased towards higher education groups in the national ANC-based data presented here, so it is likely that the presented estimates underestimate actual trends in the population. Another potential bias affecting the trend estimates at ANC site level is changes in the coverage of rural clinics over time, leading to fluctuations in the number of urban and rural residents included at individual sites. Such changes in the outreach to rural residents are likely to explain the discrepancy observed in Kapiri Mposhi between the trend estimates obtained from ANC and population-based data 
Furthermore, we could not rule out the possibility that non-participating respondents or excluded non-household populations had a different risk of HIV infection from those who participated in the DHS. The sensitivity analyses showed that point prevalence estimates only increased by 1–2 percentage points in the most extreme scenario and that the magnitude of change between the 2001/02 and 2007 ZDHS was similar. We assumed that people who were absent had a lower risk of HIV than those who refused, and this is reasonable since data from the population-based surveys in Kapiri Mposhi and Chelstone indicated that the most common reason for absence among young people was school attendance. Being in school tends to be protective against HIV infection 
. Our sensitivity analysis is admittedly less sophisticated than that conducted by Bärnighausen et al using a Heckman-type selection model on the ZDHS 2007 data 
. However, the estimates calculated in the latter paper indicate a prevalence ratio of 3
7 among non-responders to participants, and this is highly unlikely. Our assumptions and results are likely to be more plausible.
Using PMTCT data to estimate HIV prevalence has inherent biases since women agreeing to participate in the PMTCT programme may be different from those refusing to participate 
. The differences may relate to the risk of infection and the quality of counselling in the programme. However, a study conducted in Uganda found that this bias was only important in the initial period of PMTCT; after a couple of months there was no significant difference between accepting and refusing women 
. This is consistent with the findings of the PMTCT study in Lusaka, since the acceptance rate for women attending ANC in the PMTCT sites increased (from 71% to 94% between July 2002 to December 2006) 
. Furthermore, the PMTCT- and ANC-based estimates from Lusaka were closer in 2006 than 2002. It has been suggested that in situations where ANC and PMTCT coverage are very high and routinely collected PMTCT data are complete and accurate, routine PMTCT may replace the ANC surveillance system 
. However, in most of sub-Saharan Africa, PMTCT data are still of poor quality 
Bias in HIV surveys may also arise as a result of the antiretroviral therapy (ART) programme, stigma and migration. Scaling-up of the ART programme to most parts of Zambia has resulted in increased survival time among those accessing the therapy, and this might in the long run have implications for the reliability of using HIV prevalence among young people to estimate incidence. This is because HIV-positive children may survive into their teens 
and become pregnant 
, thus distorting the assumption of recent infections among young people. However, since the ART programme was only implemented nationally in 2003 in Zambia, no such effect could yet be seen in 2008. HIV-related stigma could increase participation bias because some people may fear that others will discover their HIV status 
. Selection bias due to migration is possible in HIV trend studies since migrants who are at high risk of infection are less likely to participate, and this warrants further investigation 
In conclusion, the findings suggest that although there are convincing HIV incidence declines in Zambia, the overall prevalence trend estimates have masked differential trends by place and by educational attainment. This might not only suggest differential and dynamic sub-population epidemics but also the need for tailored prevention programmes. Focusing on country-level trends in epidemiological reports therefore seems to have critical limitations and may even be directly misleading for policy makers and local programme managers who should base their efforts on comprehensive knowledge of the different epidemiological contexts within Zambia.