India contributes to almost 60% of South Asia's HIV epidemic,1 with an estimated 2.4 million people living with HIV.2 The epidemic is concentrated in the 4 southern states of Karnataka, Maharashtra, Andhra Pradesh, and Tamil Nadu and 2 northeastern states of Manipur and Nagaland. The National AIDS Control Programme is currently in its third phase (2007–2012), and a key focus is on saturating targeted interventions among high-risk populations to halt and reverse the HIV epidemic,3 Currently, the epidemic seems to be stabilizing, with a steady decline in prevalence observed in the southern high-prevalence states.
In 2003, the Bill & Melinda Gates Foundation initiated a focused HIV prevention programme called “Avahan,” the India AIDS Initiative, directed at female sex workers (FSWs), high-risk men who have sex with men (including transgenders) (MSM-T), injecting drug users, and clients of FSWs, in the 6 high HIV prevalence states indicated above. In the southern states, where the HIV epidemic is predominantly because of heterosexual transmission,4 the main strategies used to address HIV prevention among FSW, MSM-T, and clients of FSW were promotion of safer sex behavior through peer-led outreach, with particular focus on promotion of condom use; enhanced sexually transmitted infection management, with distribution and social marketing of condoms; and enhancing the enabling environment for the adoption of safer sex practices.3
The initial impact evaluation plan of Avahan included serial cross-sectional studies in these high-risk populations, based on which sophisticated mathematical models were developed to estimate impact in them, and in the general population.5,6 The use of this modeling approach has been complicated by the lack of adequate baseline data7 and, so far, results have been published only on the impact on the groups targeted by the interventions in Karnataka state.8 A purely statistical approach based on HIV surveillance data among pregnant women was recently used to estimate the overall Avahan impact in the 6 Avahan states.9 The validity of this analysis was limited by the fact that, in many districts, the Avahan intervention coexists with other programs targeting the same high-risk populations, so it was difficult to disentangle the impact of the Avahan program from other efforts. However, in Karnataka state, the distinction between Avahan and non-Avahan districts is clear because the districts where Avahan and non-Avahan interventions are implemented are different. Of the 27 districts in Karnataka as of 2003, Avahan interventions were implemented in 18, and other interventions (non-Avahan) were implemented in the other 9 districts, mainly through nongovernmental organizations contracted by the state AIDS prevention society. The programs in these 9 districts were largely restricted to large towns, and a limited number of sex work sites were accessed. The average yearly programmatic expenditure was US $111,802 and US $24,000 for Avahan and non-Avahan districts, respectively, and the average per capita expenditure for Avahan and non-Avahan districts was US $29 and US $5.5, respectively. Interventions to enhance the enabling environment through community mobilization, advocacy, and addressing social inequality and other vulnerability reduction measures, were very limited in the non-Avahan intervention districts compared with the Avahan intervention districts.10 It is in this context that we developed a multilevel statistical model to assess the impact of Avahan.
Annual HIV Sentinel Surveillance Data
Individual-level data from the National AIDS Control Organization's (NACO's) annual HIV sentinel surveillance of pregnant women attending antenatal clinics (ANCs) were used for this analysis. Rural and urban antenatal clinic (ANC) sentinel surveillance sites consistently operational from 2003 to 2008 across all of the 27 districts of Karnataka were included. The detailed methodology of sentinel surveillance is discussed elsewhere.11 For this analysis, we restricted the sample to women younger than 25 years of age. This was done because HIV infection among these women is more likely to represent recent infection and is thus more likely to reflect recent changes in HIV transmission dynamics.11
Annual HIV sentinel surveillance data collected included place of residence (rural/urban), literacy, and type of clinic site [urban ANC site referred to as district hospital (DH) and rural ANC site referred to as first referral unit (FRU)]. HIV testing was done by unlinked anonymous testing of blood samples using a standard methodology.12
We developed a multilevel logistic regression model to evaluate the impact of Avahan in Karnataka between 2003 and 2008. The model considered individual women at level 1 and districts at level 2 and were developed using the “xtmelogit” command using Stata 10.0 IC for Windows (Statacorp LP, College Station, TX).
The HIV status of pregnant women younger than 25 years of age was used as the dependent variable. The type of intervention (Avahan/non-Avahan) in each district was the primary exposure variable. Place of residence and literacy were included as a priori independent variables, as they are known risk factors for HIV. Year of survey and type of ANC clinic site were also included as a priori variables. Year of survey was included to capture HIV time trends and type of ANC clinic was included, as the characteristics of women attending these clinics may differ. As HIV prevalence in 2003 was significantly higher in Avahan compared with non-Avahan districts (the highest prevalence districts were purposefully chosen for the intervention), the analysis was adjusted for baseline HIV prevalence (BHP) in the district in 2003 and its interaction with year.10 We used the interaction between year and type of intervention for testing the impact of the intervention using the likelihood ratio (LR) test of significance (with a 2-sided alpha threshold set at 0.05) before inclusion into the model. Year of surveillance, interaction of BHP with year of surveillance, and interaction of type of intervention with year of surveillance were included as categorical variables in the model because of a reversal of trend in HIV prevalence in the year 2008. To account for the different stages of the HIV epidemic in different districts, district-level (random) intercepts and district-level random coefficients for year of surveillance were included in the model.
Use of the Multilevel Model for Estimation of District-Wise Yearly HIV Prevalence
The following formula provides the details of the model: subscript d denotes the district, while subscript i denotes the woman; the γ are the district-level random effects, while the β are the fixed coefficients; Avahan stands for the type of intervention (Avahan is 1 for Avahan districts, 0 for non-Avahan); FRU stands for the type of clinic site, (1 for FRU vs. 0 for district hospital (DH); Year200Xdi is an indicator that takes on value 1, if woman i in district d participated in the study in year 200X; HIV2003d denotes the HIV prevalence in district d in 2003; Literacy denotes the literacy status of the women (1 for literate vs. 0 for illiterate); Locality is an indicator for place of residence of the woman participating in the study (1 for rural vs. 0 for urban).
Estimation of the Number of HIV Infections Averted in the General Population
The Avahan impact was assessed by calculating the number of HIV infections averted because of Avahan in the general population using the developed multilevel regression model. For each district, specific intercepts and coefficients corresponding to the year of the survey were computed adding fixed and district-specific random coefficients. For each district, year-wise average values were calculated for the individual covariates (place of residence, literacy, type of clinic site, and type of intervention) to solve the equation above. HIV prevalence from ANC sentinel surveillance data was adjusted for the general population by multiplying a correction factor using HIV prevalence from National Family Health Survey-3 (NFHS-3) data.13 This was calculated as the ratio of NFHS-3 HIV prevalence to ANC HIV prevalence for the year 2006 assuming that the ratio between adult female and adult male prevalences to ANC seroprevalence from NFHS was constant. Correction factors for each gender and for age groups 15–24 and 25–49 years were calculated separately (Table 1).
The estimated HIV prevalence for each district for each year was calculated using random coefficients for intercept and time and average value of individual covariates in the multilevel model, while attributing to them their specific values for BHP and type of intervention. The predicted number of HIV positive cases among pregnant women aged 15–24 years was calculated for each district for each year using district-specific projected HIV prevalence and the total number of pregnant women registered at the district level. These district-wise numbers of HIV-positive cases were added and then divided by the total number of surveyed women of age group 15–24 years at the state level, to obtain a predicted year-wise HIV prevalence for the whole state. A similar approach was used to create a counterfactual, with predicted year-wise HIV prevalence for the whole state in the absence of Avahan, calculated by putting Avahan = 0 and the interaction effect of type of intervention and year as 0 in the multilevel model.
For years 2003–2005 and 2007–2008, the projected population for age groups 15–24 years and 15–49 years was calculated considering exponential growth rate for years 2006 and 2011, respectively. Exponential growth rate was calculated using the age and sex-wise population for the year 2001, using census data and using age and the sex-wise projected population for years 2006 and 2011 from the Census of India.14 Then, the sex-specific number of HIV cases among people younger than 25 years of age in Karnataka was computed by multiplying the ANC HIV prevalence predicted by the model for each year by the sex-specific correction factor and by the sex-specific size of the population aged younger than 25 years of age. This was done for both the intervention and the counterfactual scenarios. The number of cases averted among the 15–24 years population was calculated by the difference in the number of HIV cases between these 2 models for each year. As Avahan could not have “caused” HIV cases, we set the number of cases averted to zero for the years 2004 and 2008, where HIV prevalence was higher in Avahan than in non-Avahan districts. The number of cases averted among 15–49 years was calculated by extrapolating the number of cases averted among 15–24 years population to the population of 15–49 years using 3 scenarios:
- Preferred scenario: percentage of population where HIV is prevented among those 25–49 years of age is 50% of the corresponding percentage in the 15–24 years population.
- Worst case scenario: percentage of population where HIV is prevented among those 25–49 years is 25% of the corresponding percentage in the 15–24 years population.
- Best case scenario: percentage of the population where HIV is prevented among those 25–49 years is 100% of the corresponding percentage in the 15–24 years population.
The number of HIV infections averted in people aged 15–49 years using the preferred scenario was considered as the estimated number of HIV infections averted, whereas the range was calculated using the worst and best case scenarios as lower and upper limits for the number of HIV infections averted in the 15–49 years population.
Overall, 129,345 women aged 15–49 years participated in HIV sentinel surveillance from 54 consistent rural (FRU) and urban (DH) ANC sites between the years 2003 and 2008 in 27 districts. Among these, 88,003 (68%) women were aged 15–24 years. Among these young ANC attendees, 72% were literate, 63% came from rural areas, and 51% attended FRU ANC clinics during the surveillance period. In univariate analysis using the χ2 test, HIV prevalence over the 6 years of observation was higher among urban women (1.24% for urban vs. 1.03% for rural, P = 0.004), women attending urban ANC sites (1.24% for district hospital vs. 0.99% for FRU, P = 0.001), and illiterate women (1.43% for illiterate women vs. 0.99% for literate women, P < 0.001).
HIV prevalence among young ANC women from all Karnataka districts significantly decreased between 2003 and 2008 (P < 0.001). This was the case in both Avahan and non-Avahan districts (Fig. 1). However, the decline was much faster in Avahan districts between 2003 (1.49%) and 2007 (0.61%), although HIV prevalence increased to 0.90% in these districts in 2008. In non-Avahan districts, HIV prevalence was 1.21% in 2003 and decreased to 0.69% in 2008, most of the decrease being in 2008.
The observed HIV prevalence and the predicted HIV prevalence in the presence of the Avahan intervention are similar (Fig. 2A). This shows that the model for predicted HIV prevalence fits well to the observed HIV prevalence. Figure 2B shows the predicted overall trend in HIV prevalence compared with its counterfactual when all districts are modeled as non-Avahan. The difference between HIV prevalence at baseline (2003) and at a given year was significantly larger for Avahan than for non-Avahan districts for years 2006 [adjusted odds ratio (AOR) for the interaction between year and Avahan = 0.55; 95% confidence interval (CI): 0.31 to 0.96] and 2007 (AOR for interaction = 0.41; 95% CI: 0.21 to 0.79) (Table 2). The overall P value for the −2 Log LR test (with 5 degrees of freedom) for the addition of all interaction terms between year and type of intervention was 0.046.
The area between the 2 curves in Figure 2B was used to estimate the number of cases averted. As shown in Table 3, in the general population aged 15–24 years, 23,285 HIV infections were averted between 2003 and 2008 (Table 3). Overall, 87,035 cases were averted (range, 55,160–150,784 cases) in the general population aged 15–49 years between 2003 and 2008 because of Avahan (Table 3). Fifty-five percent of these averted cases were among men.
The results of our study suggest that the Avahan intervention among high-risk groups had a significant impact on the reduction of HIV prevalence in the general population in Karnataka state. Indeed, the change in HIV prevalence in young ANC women between 2003 and 2008 was larger in Avahan than in non-Avahan districts with a significant −2 Log LR test (P = 0.046). This corresponds to a significantly stronger decrease in HIV prevalence in Avahan than in non-Avahan districts among young ANC women between 2003 and 2006 (AOR = 0.546, P = 0.035, Wald test) and between 2003 and 2007 (AOR = 0.407, P = 0.008, Wald test). The corresponding value for year 2005 was borderline significant (AOR = 0.673, P = 0.078, Wald test). Overall, an estimated 87,035 HIV infections were averted in the general population of Karnataka because of Avahan between 2003 and 2008.
Recent studies have shown evidence of population level impact of large-scale interventions against HIV in a concentrated epidemic such as in India.9,10,15 Our analysis is an extension of the work by Moses et al.10 We have added one more year of observation (2008) and constructed a statistical model to evaluate the impact of Avahan and to estimate the number of HIV cases averted by the intervention. Kumar et al15 used a quasi-experimental approach with condom gap as a measure of the intensity of targeted interventions in the 4 large southern states in India, with decline in HIV prevalence as the outcome. Ng et al9 assessed the impact of Avahan in all the 6 intervention states and estimated 41,683 HIV infections averted in Karnataka (95% CI: 12,299 to 81,496). Our estimate of 87,035 infections averted (range, 55,160–150,784) is higher, although not significantly different statistically. The 2 studies differ in the way that the baseline state of the epidemics in the districts was entered in the model. In our model, the BHP and its interaction with time were introduced. We also allowed the time trend in HIV prevalence to vary by district by having district level random intercepts and random coefficients for the year variables. It is likely that the epidemic trajectory of districts with higher prevalence at baseline will be such that HIV prevalence will decline faster compared to those with lower prevalence. This is an important consideration, since Avahan interventions were started in districts with higher prevalence in Karnataka (1.49% vs 1.21%). Inclusion of the interaction between time and BHP in the model controls for this effect and the variability in the effect of time from district to district. The model used by Ng et al controlled for BHP, but it is unclear if they controlled for the interaction between BHP and year. Furthermore our analysis is more straightforward, because there is a clear distinction between Avahan and non-Avahan districts, in Karnataka. Although there might be a small level of variability in program implementation and coverage across the Avahan districts, the main advantage of the model considered is its simplicity. In addition, it still reflects the reality of the differences between Avahan and non-Avahan districts quite well.
Our study has some limitations. First, this analysis does not take into account the nature of the Avahan intervention and its various components. Although the apparent impact on the general population is encouraging, the impact of Avahan would first occur among high-risk groups before the beneficial effect in the general population becomes obvious. Our analysis has the same limitation as Ng et al, in that it does not allow analysis of the actual mechanism, which leads to the impact. This needs a causal pathway analysis of the mechanism of impact.16 The mathematical modeling analysis being conducted as part of the overall Avahan evaluation will provide estimates of the number of HIV infections averted in the subpopulations targeted by the intervention (FSWs, their clients, MSM).5 Second, our analysis and those from other published articles so far, does not consider HIV transmission dynamics because only HIV prevalence (and not incidence) is used. Third, our analysis does not take into account the number of cases averted after 2008, which may be because of the cases prevented previously. This is important because the cumulative effect of interventions would likely increase over time. Fourth, the assumptions made for the computation of cases averted among 25–49 year olds is unverifiable. However, the assumptions are logical, as many women in the 25–49 age group could have had the infection acquired before Avahan started, and a smaller proportion of cases were averted among women of this age group compared with women aged 15–24 years. Thus, extrapolating the cases averted from 15–24 to 25–49 years directly would be inappropriate. Hence, we considered scenarios to provide a range of the estimated cases averted. Finally, the explanation of the reversal of trend observed in HIV prevalence for 2008 is unclear. This could be a random increase in the prevalence or an actual reversal of the trend. Similarly, the reasons for a sudden decline in the HIV prevalence in non-Avahan districts between 2007 and 2008 after a relative stability in the previous years are also unclear. However, the interventions in the non-Avahan districts started to scale-up in 2006 after the National AIDS Control Programme, phase III, was initiated. This could have contributed to this decline. Results of the sentinel surveillance data for the year 2010 (not yet available, and no ANC sentinel surveillance was conducted in Karnataka in 2009) will throw light on future trends.
Another important consideration is the cost of Avahan interventions compared with interventions in other districts. Clearly, the amount of money spent on the interventions in Avahan districts was higher and likely to have contributed to a faster decline in HIV prevalence in Avahan districts. It indeed seems to be important to have a sufficient intensity and scope for such interventions to have an impact in a country like India. The scale-up of interventions from 2006 onward in the non-Avahan districts (with consequent increased costs) was temporally associated with a decline in the HIV prevalence in these districts in 2008, which could be another indication of the importance of sufficient investment in HIV prevention programmes for them to be successful. The detailed cost and cost effectiveness analysis of Avahan is currently being conducted.
To conclude, despite the limitations noted, it seems that the Avahan intervention has had a significant impact on decreasing HIV prevalence in the general population of Karnataka state. Lessons learnt from this focused rapidly scaled-up HIV preventive intervention, may help with the design and implementation of targeted interventions elsewhere. These results make a strong case for implementing and scaling up such interventions in all concentrated and mixed HIV epidemics.
The authors would like to thank the National AIDS Control Organisation, New Delhi and Karnataka State AIDS Prevention Society for providing data on sentinel surveillance for this study.
1. Moses S, Blanchard JF, Kang H, et al.. AIDS in South Asia: understanding and responding to a heterogeneous epidemic. Washington, DC: The World Bank; 2006.
2. UNAIDS. Global Report: UNAIDS Report on the Global AIDS Epidemic 2010. Geneva, Switzerland: Joint United Nations Programme on HIV/AIDS; 2010.
3. National AIDS Control Organisation (NACO). Strategy and Implementation Plan: National AIDS Control Programme Phase III (2006-2011). New Delhi, India: Ministry of Health and Family Welfare; 2006:19–57.
4. Bill & Melinda Gates Foundation. Avahan
-The India AIDS Initiative: The Business of HIV prevention at Scale. New Delhi, India: Bill & Melinda Gates Foundation in India; 2008.
5. Boily MC, Lowndes CM, Vickerman P, et al.. Evaluating large-scale HIV prevention interventions: study design for an integrated mathematical modelling approach. Sex Transm Infect. 2007;83:582–589.
6. Chandrasekaran P, Dallabetta G, Loo V, et al.. For the Avahan
evaluation partners. Evaluation design for large scale HIV prevention programs: the case of Avahan
, the India AIDS Initiative. AIDS. 2008;22(suppl 5):S1–S15.
7. Lowndes CM, Alary M, Verma S, et al.. Assessment of intervention outcome in the absence of baseline data: reconstruction of condom use time trends using retrospective analysis of survey data. Sex Transm Infect. 2010;86(suppl 1):i49–i55.
8. Pickles M, Foss AM, Vickerman P, et al.. Interim modelling analysis to validate reported increases in condom use and assess HIV infections averted among female sex workers and clients in southern India following a targeted HIV prevention programme. Sex Transm Infect. 2010;86 (suppl 1):i33–i43.
9. Ng M, Gakidou E, Levin-Rector A, et al.. Assessment of population-level effect of Avahan
, an HIV-prevention initiative in India. Lancet. 2011;378:1643–1652.
10. Moses S, Ramesh BM, Nagelkerke NJD, et al.. Impact of an intensive HIV prevention programme for female sex workers on HIV prevalence among antenatal clinic attenders in Karnataka state, south India: an ecological analysis. AIDS. 2008;22(suppl 5):S101–S108.
11. National AIDS Control Organisation (NACO). HIV Sentinel Surveillance and HIV Estimation, 2006. Ministry of Health and Family Welfare; 2007:1–16.
12. UNAIDS/WHO. Guidelines for Conducting HIV Sentinel Serosurveys Among Pregnant Women and Other Groups. Geneva, Switzerland: WHO/UNAIDS; 2003.
13. International Institute of Population Sciences. National Family Health Survey (NFHS-III). Mumbai, India: International Institute of Population Sciences; 2005–06 September 2007.
14. Registrar General of India. Census of India. New Delhi, India: Office of the Registrar General & Census Commissioner, India; 2010. Available at: http://www.censusindia.gov.in/Census_Data_2001/Census_Data_Online
. Accessed May 25, 2012.
15. Kumar R, Mehendale S, Panda S, et al.. Impact of targeted interventions on heterosexual transmission of HIV in India. BMC Publ Health. 11:549.
16. Boerma T, de Zoysa I. Beyond accountability: learning from large-scale evaluations. Lancet. 2011;378:1610–1612.