HIV clustering and spatiotemporal patterns of HIV prevalence in Malawi
This study has provided a visually powerful and empirically derived, multi-scalar analysis of spatiotemporal variation in HIV prevalence among pregnant women attending ANCs in Malawi from 1994 to 2010. It goes beyond current broad characterizations at national, regional and urban/rural variation in HIV prevalence (Figure
2) to local variation at district and lower (continuous) levels (Figures
4,
5,
6,
7,
8,
9,
10,
11,
12, and
13). GIS tools allowed the generation of HIV prevalence estimates at scales where HIV data are not traditionally collected (district and continuous) from point HIV sentinel data at 19 ANCs for spatiotemporal pattern analysis and from 54 ANCs for use in multivariate analysis of possible drivers and their spatial variation. Observed widespread spatial variation in HIV prevalence reveal the HIV epidemic as an aggregation of several spatially defined sub-epidemics - national, regional, urban, rural, and local (clusters). Findings broadly call into question the use/effectiveness of one-size-fits-all interventions and policies under such circumstances.
The most prominent temporal trend was the general decline in HIV prevalence after rapid spread/intensification up to 1999. For Malawi, the trend positively reflects on the concerted planning and considerable human and financial resources (most from donor aid) applied to prevention and treatment efforts by government and other agencies over the past decade [
17]. However, the impressive rates of national decline in HIV prevalence among ANC-attending pregnant women are 1) slowing down from 1.1% annually betweeen1999 and 2010, to 0.91% and 0.88% annually for the periods 2003-2010 and 2005-2010, respectively (Figures
4-
11) tempered by a modest decline in the general population (from 12.0% in 2004 to 10.6% in 2010, 0.2% annually) based on MDHS data [
14] - a cautionary tale against complacency.
Another significant spatiotemporal trend was a slow but emergent spatial evening in HIV prevalence as the epidemic stabilized and then declined after an initial period of localized spatial heterogeneity and explosive spatial spread and intensification to a peak in 1999, as illustrated empirically in temporal patterns of global Moran’s I (Figure
3) and in various ways in Figures
4-
11. First, there was a general narrowing in the prevalence gap among the regional sub-epidemics (more so for the north and center, and between the urban/semi-urban and rural epidemics. Second, the negative autocorrelation evident in 1995 and 1996 indicated localized heterogeneity with high next to low prevalence pockets confirmed as spatial outliers in clustering patterns (Figures
10 and
11), which are generally associated with expansionary spatial processes and rapid spread [
54]. Subsequent increases in the level and significance of positive spatial autocorrelation (Figure
3) further indicate
relative spatial evening of the epidemic with districts of similar HIV prevalence more clustered together. The (low) decrease in total statistically significant clustering (12 core hotspot and coldspot cluster districts 2001 to 2010 relative to 14 in 1994 and 1999) also indicates spatial evening as districts that stand out from the rest decrease. The fact that the spatial and population center of gravity was firmly located in rural areas where the majority (85%) of Malawians lived [
49] and the HIV epidemic was more stable and lower in intensity (at least among pregnant women), may have provided the inertia that kept the national epidemic in relative check. The rural epidemic’s declining trend may thus be a key turning point nationally.
Identification of a hotspot cluster of 5-11 districts making up the HIV epicenter consistently located in the south was the most significant outcome of cluster analysis. Potential explanations include the long history of urbanization in the south forming a hub for HIV transmission. Cluster analysis revealed a hierarchy among the four biggest cities in the country in HIV prevalence relative to surrounding districts. Blantyre, the first city in Malawi and the commercial capital, anchored the HIV hotspot/epicenter every year of analysis, followed by Zomba (twice in a hotspot cluster), then Lilongwe (twice an HL outlier but part of a primary or secondary coldspot in most of the study years), and then Mzuzu city [Figures
10 and
11]. The Southern Region was also the most densely populated [
49], and had the highest levels of rural poverty [
50] and prevalence of syphilis [
17]. The 2010 MDHS also showed that women and men in the Southern Region had the highest percentage of those who had two or more sexual partners in the previous 12 months and mean number of sexual partners per lifetime, while men had the lowest percentage reporting using a condom during last sexual intercourse (22.9%) [
14]. Elevated labor migrancy, particularly returning Malawian mine workers from South Africa late1980s/early 1990s, may partly explain emergence of an HIV sub-epicenter from 1994 to 1999 (Figures
6-
9) among the northern districts of Nkhata Bay and Rumphi, and Mzuzu city [
32,
55].
Potential drivers of HIV prevalence for 2010 and their local spatial patterns
Geographic variables, particularly mean travel time to nearest road, had the most explanatory influence on HIV prevalence among the variables in the “best” regression model (Table
2), but the two geographic variables in the model were virtually uncorrelated and they appeared to capture subtly different aspects of “access”. The spatial variation of these two variables in relation to HIV prevalence (and hotspots) also differed. Main roads generally link places to main towns and cities and this variable mainly captured the urban/rural divide. Increasing distance from main roads reflected increasing distance from urban areas and the factors that elevated HIV prevalence there, while increasingly capturing more isolated and sparsely populated areas with limited mobility and sexual mixing and access to health services and conditions that are generally associated with lower risks of HIV infection. This explains the negative relationship with HIV prevalence. Cluster analysis of distance from main roads revealed only one statistically significant core cluster, a coldspot (districts exceptionally close to main roads and urban areas, hence high HIV risk areas) cluster of Blantyre city and neighboring Blantyre Rural and Chiradzulu districts. This suggests that distance to cities (or urban influences) is a major explaining factor for high HIV rates in this subset of the core HIV hotspot districts, although hotspot analysis with the Getis-OrdGi* statistic expanded the primary coldspot cluster to 8 districts which virtually matched the larger core HIV hotspot.
In contrast, core and secondary hotspots for mean travel time to nearest transport (30-44 age group), which varied positively with HIV prevalence, matched three rural HIV hotspot districts of Mulanje, Phalombe and Thyolo spatially. These districts (including Zomba Rural as a tertiary HH cluster) have hilly terrain, which is a physical obstacle to travel and access to health services. Further, high concentration of commercial farms (mainly tea) forced many people into smaller remaining areas, making for some of the highest population densities in relation to available arable land. The dense populations, including the predominantly male commercial farm workers (a high HIV risk group in Malawi [
56]), facilitated increased sexual mixing and risky sexual encounters, and HIV spread. These areas need increased access to HIV information and other services, including through mobile services.
Our findings are generally consistent with recent studies elsewhere in Africa. Tanser et al. [
1] also found an inverse significant relationship between HIV prevalence and distance to the main road in Kwazulu-Natal, South Africa. Travel time to nearest public transport captures a similar relationship in the inverse direction (longer/farther from transport, lower HIV rates). Studies reveal a mixed role of education, suggesting that it is mediated by complex factors. The 2004 and 2010 MDHS studies show the level of education generally positively associated with HIV prevalence [
14]. However, Moise and Kalipeni found a negative relationship for literacy rates above 20% in Zambia based on a spatial lag model. Given our more specific definition of education, and controlling for HIV testing, distance to main roads and time to main transport, the negative relation with the percentage of those who had attained this modest (senior primary education) suggests some level of education is good for AIDS prevention.
Further, we found that the percentage of people in the larger population (men and women) who had taken at least one HIV test ever was positively associated with HIV prevalence. The secondary hotspot clusters for HIV testing revealed by the Getis-OrdGi* statistic captured seven districts that closely matched the core HIV prevalence hotspot for the women’s ANC sample for 2010 (Figure
13, B4). This finding that levels of HIV testing were highest in districts that already had high HIV prevalence is plausible because the sense of risk and value of HIV testing would likely be heightened in such areas. Indeed, the 2011 Malawi Welfare Monitoring Survey found that the main reason given by 44% of non-tested respondents for not testing was that they did not feel at risk or in need of an HIV test [
51]. During Malawi’s 2010/11 fiscal year, 1.773 million people (28% of the sexually active population) were tested for HIV in Malawi [
12]. However, HIV testing was higher among urban residents who are generally wealthier and more educated, have better access to testing facilities, but also a higher risk of being infected than rural residents [
3,
6,
14]. This relationship explains the emergence of Lilongwe City as a core HL outlier and Zomba City as a core hotspot (Figure
12, A4). However, Blantyre City does not appear as an outlier probably because the core HIV cluster in the south is more extensive and includes Blantyre and several (rural) districts (Figure
12, HIV, A1). Further nuanced analysis may be needed on HIV testing behavior and its impact on HIV prevention.
Although the four variables included in the “best” model explained a relatively high proportion of the variance in HIV prevalence (68.8%, see Table
2), closer examination suggests that these variables were essentially proxies for underlying factors related to geographic/physical and socio-economic access or exposure to health services and other amenities, HIV prevention knowledge, personal resources, sexual networks and risky sex and other factors that influence HIV-related behavior and risk, mainly mediated by urban/rural residence or proximity and poverty/wealth and education. Further, some variables not included in the final model were interesting. For instance, the core and secondary clustering patterns (hotspots) for 2010 syphilis prevalence, highly significant in binary correlations (Table
1), were a near-perfect match with HIV hotspot districts (Figures
12 and
13), suggesting that syphilis treatment/management also needs special attention. Therefore, more detailed analysis needs to be done, focusing on local HIV variation at finer spatial scales and drivers of HIV
incidence to better inform policy, rather than on prevalence as was done in this study. Nevertheless, HIV incidence is often closely associated with prevalence, and as in Tanser et al. [
1], we also assumed that time lags between cause and effect would not significantly model outcomes given our emphasis on preliminary associations.
Limitations of the study
While this study contributes significantly to advancing spatiotemporal analysis of HIV/AIDS in Malawi, Africa, and the developing world generally, it has limitations. The longitudinal depth of HIV sentinel data and caution with the interpolation process including consistency across years allowed adequate assessment of spatiotemporal patterns/trends among pregnant women, but the small number (19) of sentinel ANCs, used in the spatiotemporal analysis means that the interpolated products and especially the derived district estimates are and should be treated as indicative, although the HIV data for the regression analysis were from a larger sample (54 points). Local spatial-statistical pattern (hotspot) analysis also came with uncertainties, including the choice of method or search distance to define a neighborhood. In this study, we empirically determined it as the distance at which global Moran’s I was highest for most years, 80 km. Further research is needed to determine the spatial limit of the spatial dependence of HIV on neighboring values at different spatial scales to better guide interventions. Finally, although our choice of an OLS regression model in the presence of spatial dependence breached the assumption of independence of measurements, OLS was sufficient for the study purpose of identifying indicative explanatory factors for observed spatial variation (as opposed to producing predictive models). Moreover, a corrective spatial lag model of HIV prevalence using standardized variables in the “best” OLS model based on spatial diagnostics was performed within the GEODA spatial statistical software [
53,
57]. Despite improvements in the explanatory power of the model and significance of coefficients for all but one of the four independent variables (the p-value for distance to roads fell slightly from 0.043 to 0.06), the spatial lag model did not change the directions of the variable relationships nor their relative importance.
f In the interest of space, we left out the spatial lag model findings.