The results of our analysis indicate stark geographical variation in HIV prevalencein most of the countries. The observed spatial variation in HIV prevalencehighlights a clustered HIV transmission across SSA within micro-epidemics ofdifferent scales. The map of HIV clustering reflects a landscape with‘valleys’ (areas with high HIV prevalence), ‘dams’ (areaswhere HIV found barriers to propagate efficiently), and ‘islands’ (smallisolated areas with characteristically either very high or very low HIVprevalence).
Our results indicate that only ~14% of the population across the countries resideswithin clusters of high HIV prevalence. The strength of the clustering tended to behigher in countries with low national HIV prevalence. For instance, the strongestclustering (highest RR) is found in a cluster in Senegal (RR = 6.69, HIVprevalence = 4.3%); the country with the lowest national HIV prevalence(0.7%). Our study revealed similar settings with localized epidemics at high HIVprevalence hidden in a map of low national HIV prevalence, such as in Burkina Faso,Congo, Sierra Leone and Ethiopia.
The strength of the clustering was smaller in countries with high national HIVprevalence, indicative of more diffusive epidemics. For instance, in Zambia andLesotho, the strength of the clustering was fairly small (RR = 1.74, andRR = 1.28, respectively). In Swaziland, no clusters with high HIVprevalence were identified. This result underlines how the HIV epidemics in thesehigh prevalence countries had percolated throughout much of the demography andgeography of these countries.
We also identified clusters with low HIV prevalence in most of the countries includedin our study. These clusters appear to reflect ‘dams’ where somebehavioral or biological protective factors appear to have slowed HIV transmissionin such populations, in contrast to their neighboring populations. In fact, weidentified settings with very low HIV prevalence even in countries with substantialHIV epidemics such as in Tanzania, Kenya and Malawi.
The topography of this infection poses a question about the drivers of such starkheterogeneities even at micro-geographic scales within countries. Male circumcision , the presence of other sexually transmitted infections (STIs) , tropical co-infections increasing HIV viral load , hormonal contraception , viral factors , and host genetics and immunology  vary across SSA, and are believed to influence HIV transmission risk.Behavioral factors such as concurrency , number of sexual partners , commercial sex , and coital frequency  appear also to vary across SSA, and may contribute to explaining theheterogeneities in prevalence. Preliminary statistical analyses of the DHS databases(not shown) indicated that it is challenging, if not a formidable task, todisentangle the contribution of the different factors in the clustering of theinfection. This is a consequence of the complex array of independent variables toconsider, and also because of the population sizes of the clusters’sub-samples which are not large enough to power meaningful multivariate regressionanalyses. Nevertheless, these preliminary analyses suggest that the scale anddistribution of the differences in the biological and behavioral factors, withinversus outside the clusters, may not be sufficient to explain the observed sharpcontours in the topography of HIV infection at the local level. We suspect thatthere is an additional dynamical factor that has strongly influenced the localecology of this infection even when the differences in the biological and behavioralfactors may not have been markedly large.
We hypothesize that the HIV epidemic among the general population in much of SSA isnot far from its epidemic (or sustainability) threshold. A generic feature of aninfection epidemic is that near the epidemic threshold, the prevalence dependsnon-linearly on the determinants of infection transmission, and that small changesin the epidemiological context can drive much larger changes in the prevalence ofthe infection . Figure 4 illustrates this dynamical effectfor the case of HIV infection. As can be seen in the figure, modest changes insexual risk behavior near the epidemic threshold could generate a substantialincrease in HIV prevalence. Conversely, beyond the region of epidemic threshold, thesame increase in sexual risk behavior could generate only a modest increase in HIVprevalence.
Accordingly, we hypothesize that an essential driver of the stark variability in HIVinfection transmission in SSA is that the epidemiology of this infection is not farfrom its epidemic threshold in the general population outside of conventionalhigh-risk groups. The conspicuously large clustering of HIV infection may notstrictly reflect conspicuously large variations in sexual risk behavior or thepresence (or absence) of specific biological co-factors in HIV transmission. Thevariability in sexual risk behavior or biological co-factors within the populationhas driven a much larger variability in HIV prevalence, thanks to the non-linearepidemic dynamics near the infection sustainability threshold. This hypothesis mayalso contribute to explaining the global variability in HIV infectious spread whereonly in SSA massive general-population HIV epidemics have occurred. In SSA, butnowhere else, the epidemiology of HIV infection has crossed, though not by far, theepidemic threshold of sustainability in the general population (Figure 4). That was enough to spark localized epidemics in the generalpopulation; and these epidemics, not far above the sustainability threshold,exhibited consequently high diversity in size at the micro-geographic scale(Figures 1 and 2).
Several study limitations could have affected our results. First, the selection ofthe DHS round for the different countries was constrained by the availability of HIVbiomarker information and geographical coordinates of each survey data point at anyparticular DHS round. This limited our ability to consider more countries in SSA foranalysis with more recent DHS rounds. Small clusters of HIV infection could havebeen missed if there is not enough sampling within their geographic setting. Giventhe multiple logistical difficulties in conducting the DHS, some of our measurescould have been influenced by inherent biases in the data such as the variability inresponse rates to HIV testing [28, 29].
Mobile individuals and high-risk subpopulations such as female sex workers, injectingdrug users, and men who have sex with men, may have been undersampled by the DHS.Clusters of HIV infection among such subpopulations may have been missed in ouranalysis. It is not clear though whether undersampling of such populations couldnecessarily affect our findings or not. Epidemics among high-risk subpopulationsshould lead to some infection onward transmission such as among spouses and clientsof female sex workers which are less likely to be undersampled in the DHS. Lastly,due to the cross-sectional nature of the data used in this study, some of theclusters identified here could reflect epidemics at different stages, rather thangenuine differences in epidemic sizes. The HIV epidemic in SSA, however, is in amature stage [25, 30], and therefore this potential limitation is probably not influencing ourresults. Moreover, we analyzed the clustering of the infection in four countriesthat had more than one DHS serological biomarker survey at different years, and noconsequential changes in the distribution of the clusters were observed (notshown).