Geographic Information Systems, spatial analysis, and HIV in Africa: A scoping review

Introduction Geographic Information Systems (GIS) and spatial analysis are emerging tools for global health, but it is unclear to what extent they have been applied to HIV research in Africa. To help inform researchers and program implementers, this scoping review documents the range and depth of published HIV-related GIS and spatial analysis research studies conducted in Africa. Methods A systematic literature search for articles related to GIS and spatial analysis was conducted through PubMed, EMBASE, and Web of Science databases. Using pre-specified inclusion criteria, articles were screened and key data were abstracted. Grounded, inductive analysis was conducted to organize studies into meaningful thematic areas. Results and discussion The search returned 773 unique articles, of which 65 were included in the final review. 15 different countries were represented. Over half of the included studies were published after 2014. Articles were categorized into the following non-mutually exclusive themes: (a) HIV geography, (b) HIV risk factors, and (c) HIV service implementation. Studies demonstrated a broad range of GIS and spatial analysis applications including characterizing geographic distribution of HIV, evaluating risk factors for HIV, and assessing and improving access to HIV care services. Conclusions GIS and spatial analysis have been widely applied to HIV-related research in Africa. The current literature reveals a diversity of themes and methodologies and a relatively young, but rapidly growing, evidence base.


Introduction
Geographic Information Systems (GIS) and spatial analysis are emerging tools for global health, but it is unclear to what extent they have been applied to HIV research in Africa. To help inform researchers and program implementers, this scoping review documents the range and depth of published HIV-related GIS and spatial analysis research studies conducted in Africa.

Methods
A systematic literature search for articles related to GIS and spatial analysis was conducted through PubMed, EMBASE, and Web of Science databases. Using pre-specified inclusion criteria, articles were screened and key data were abstracted. Grounded, inductive analysis was conducted to organize studies into meaningful thematic areas.

Results and discussion
The search returned 773 unique articles, of which 65 were included in the final review. 15 different countries were represented. Over half of the included studies were published after 2014. Articles were categorized into the following non-mutually exclusive themes: (a) HIV geography, (b) HIV risk factors, and (c) HIV service implementation. Studies demonstrated a broad range of GIS and spatial analysis applications including characterizing geographic distribution of HIV, evaluating risk factors for HIV, and assessing and improving access to HIV care services.

Conclusions
GIS and spatial analysis have been widely applied to HIV-related research in Africa. The current literature reveals a diversity of themes and methodologies and a relatively young, but rapidly growing, evidence base. a1111111111 a1111111111 a1111111111 a1111111111 a1111111111

Introduction
Geographic Information Systems (GIS) are systems designed to store, manage and display spatial data and aid in analysis and interpretation of spatial data [1,2]. Spatial analysis are techniques applied to geographic data to measure, interpret, and explore characteristics and associations [1][2][3][4]. Geographic data can include defining the locations of a point in space (known as point pattern data), quantifying values of a characteristic at a given location (geostatistical data), or describing an entire geographical region (area-level data) [1,[5][6][7]. With the rise and increasing accessibility of computing power and relevant software, GIS and spatial analysis are finding wider audiences and novel applications [3,[8][9][10].
Notably, GIS and spatial analysis can be powerful tools to understand, prevent, and help treat diseases. For example, as a visualization aid, GIS software can be used to map the geographic distribution of disease, associated risk factors, and services available for prevention and treatment. Furthermore, spatial analysis of this data can analyze risks for disease, epidemic trends over space and time, and disease hotspots [9,11]. Taken together, these tools can contribute to the design, planning, and allocation of global health resources for prevention and treatment services, as well as help to assess intervention impact.
HIV continues to be a major global health threat, particularly in Africa which accounts for 67% of the world's infections, highlighting a need for additional approaches to understand and mitigate the epidemic [12]. GIS and spatial analysis techniques have begun to be applied to HIV-related research in sub-Saharan Africa. However, it is unclear to what extent and depth these tools have been utilized. In this scoping review focused on providing an overview of a broad topic [13], we sought to systematically find and inductively summarize the literature on applications of GIS and spatial analysis to HIV-related research in Africa. Our goal is to help inform researchers, implementers, and policy makers on how GIS and spatial analysis tools have been used, and to explore underlying themes and methodologies. strategy). This search was developed through an iterative process of incorporating new terms and refining those included based on results returned and identification of relevant citations. Reviewers conducted electronic searches of PubMed, EMBASE, and Web of Science on November 16, 2017, with no restriction on date or language of publication. Potentially eligible articles known to paper authors were also included in the initial search.
Study identification and data collection. Two reviewers independently screened article titles and abstracts of all initial search results. Articles deemed eligible for inclusion by either reviewer underwent full text review by three authors to determine final inclusion with adjudication done by majority vote when necessary. Reviewers extracted data to a master table, capturing details about location, objectives, study design, data source and study population, spatial analysis methodologies, software used, and results.
Synthesis of results. Following data extraction, we categorized studies based on the primary content of the paper in an inductive, thematic analysis. Articles were permitted to sort into more than one category as relevant. We did not assess study quality due to study heterogeneity. studies identified 65 final articles for inclusion in the review. The reasons for exclusion were that the study was not based in Africa (1 reference), that the study was not primarily focused on HIV-related topics (10 references), and that the analysis did not involve spatial analysis (16 references).
Included studies spanned the continent in geographic focus. Eight papers were multi-country analyses, and 15 different countries were represented in single-country papers. Among single-country studies, 37 were based in Southern Africa, 16 in East Africa, and two each in West Africa and Central Africa. South Africa was the most common country of study with 17 papers. Dates of publication spanned from 2000 to 2017, with almost all (98%) published between 2006-2017 and over half published after 2014.

Study themes
Inductive thematic analysis found that GIS and spatial analysis were most commonly used to examine: (a) HIV geography, (b) HIV risk factors, and (c) HIV service implementation. We discuss methodological techniques in each theme and narratively and in tabular format highlight key findings below.

HIV spatial epidemiology
Twenty-seven articles sorted into this thematic category which focused on the spatial distribution of HIV. The following subcategories were identified and used to further characterize studies: (i) spatial cluster analysis, and (ii) smoothed mapping.
Spatial clustering analysis. One of the most common spatial analytic techniques for evaluating the geographic distribution of HIV was local cluster detection (Table 1), or the identification of HIV disease "hotspots". Clusters are areas with higher or lower numbers of events in a particular study region than expected if cases are randomly distributed [14].
Thirteen studies used the Kulldorff spatial scan statistic [15] to identify clustering of HIV infection, [16][17][18][19][20][21][22][23][24], HIV-specific and all-cause mortality [25][26][27], and sero-discordant partnerships [28]. In these studies, the spatial scan statistic identified statistically significant "hotspots" or "cold spots" by systematically scanning circular windows of varying sizes across the study area, comparing the number of cases within the window to those outside the window. The resulting data were used to identify areas of particular risk or concentrated need, and to compare characteristics inside and outside of clusters. SaTScan software was the most commonly used tool for identifying clusters (see S1 Table for software used in each study).
Clusters of high and/or low HIV prevalence were detected, independent of administrative boundaries, in 20 of 22 countries where the Kulldorff spatial scan statistic was applied, though the geographic size and relative risk strength of the clusters varied widely. Studies that compared clusters over time found that they tended to persist with little change in location [17,22,25]. In one study, areas contained within high-prevalence clusters had stable or increasing HIV prevalence over time, even when HIV significantly decreased in the rest of the country [17].
Less commonly-used spatial statistics included Local Moran's I and the τ statistic [29]. Both of these methods are global clustering statistics which measure the tendency for points to occur closer together in space by chance across the entire study area. In contrast, the Kulldorff spatial scan statistic identifies local clusters in a particular region. Local clusters can exist in either the absence or presence of global clustering [30]. In HIV epidemiology, global clustering statistics have been used to determine if HIV-positive people tend to live closer to each other than would be expected by chance. For example, a study in Uganda conducted a global clustering analysis and [31] found that a participant living in the same household as an HIV-positive 16,140 No spatial pattern for sero-discordancy independent of HIV prevalence patterns. HIV prevalence correlated with proportion of couples that were sero-discordant.

Mozambique
Kulldorff cluster detection to compare HIV over time.

GRABOWSKI ET AL. [31]
Uganda Clustering analysis to determine the likelihood that a participant living in the same household as an HIV-positive person, or within given distance rings from an HIV-positive person would also have HIV.
14,594 Strong clustering within households: sharing a household with an HIV-positive person increased likelihood of HIV by 3.2. Weaker clustering within 10-250m (1.2 times likelihood) and 250-500m (1.08 times likelihood).

MEE ET AL. [26]
South Africa Space-time Kulldorff cluster detection for HIV/TB mortality and non-HIV/TB mortality during decentralization of ART provision.

73,000
Two low-risk and one high-risk HIV/TB mortality clusters detected. Unclear link to ART decentralization.

MESSINA ET AL. [19]
Democratic Republic of Congo Kulldorff cluster detection of HIV by sex. 9,755 Detected clusters of HIV with evidence that spatial distribution and intensity varies by sex.

NAMOSHA ET AL. [25]
South Africa Kulldorff cluster detection for HIV mortality comparing pre-and post-ART roll-out.

86,175
Strong clusters persisted over time. Highmortality clusters in peri-urban communities near National Road.

SARTORIUS [27]
South Africa Comparison of identified high-mortality and low-mortality Kulldorff clusters.

1,110,166 personyears
Identified clusters and several risk factors that differed significantly between high and low clusters.

SCHAEFER ET AL. [24]
Zimbabwe Compare HIV service uptake and demographic characteristics inside and out of identified HIV Kulldorff clusters.
Two studies compared different mapping techniques, and the most comprehensive study found that a Bayesian geostatistical modeling approach was slightly more accurate than five other methods [32,34]. However, it was also noted that the choice of mapping method may have less impact on the accuracy of the HIV estimates than the underlying quality and sampling frame of the data used [32][33][34]. HIV estimates derived from smoothed maps and aggregated to the country level were found to align closely with more complex UNAIDS model data, regardless of mapping methodology [34].
Smoothed mapping techniques generated maps which estimate HIV prevalence in areas where data is not available from surveys [17,29,[32][33][34]42]. Studies that generated smoothed maps found high-risk areas that stretched across more than one administrative area [19, 29, 32-34, 36, 37], or high prevalence pockets in provinces that would not necessarily have been identified as high priority settings [29,32,36]. As Kleinschmidt et al note, a high prevalence area that they detected across two South African provinces could benefit from inter-provincial coordination of treatment and prevention service provision [36].
Detailed smoothed maps also enabled visual comparisons of the spatial patterns of HIV over time periods [17,29,34,35], or between various subgroups of the population, such as men and women or those living in sero-discordant partnerships and those not [19,28,36,39]. GIS software also helped validate HIV prevalence estimates from antenatal care surveillance against population-based survey estimates, which are more comprehensive, but more resource-intensive, by comparing data between the two sources at similar locations [43,44].

HIV risk factor analysis
Thirty-three studies sorted into this thematic subcategory which focused on the spatial epidemiology of HIV which is based on the concept that both HIV and the factors associated with infection vary spatially and have some degree of spatial dependence. Articles in this theme divided into three sub-topics based on underlying methods utilized: (i) spatial regression, (ii) joint spatial disease/risk factor modeling, and (iii) geography of risk factors. Spatial regression. Spatial analysis contributed to the evaluation of risk factors for HIV by testing the assumption that data is independent of geography (Table 3). Many studies noted that this independence assumption was violated if data displayed spatial dependence or correlation, meaning that data points in close proximity to each other were more alike than those further apart [18,21,27,40,[45][46][47][48][49][50]. To address this issue, thirteen studies included in their analysis spatial regressions with spatial random effects to represent the correlations within geographic areas and between geographically proximate areas [18,21,27,39,40,[45][46][47][48][49][50][51][52]. These analyses used BayesX [18,21,45,46,[48][49][50], WinBUGS [27,39,47]and OpenBUGS Bayesian software [40]. Studies that compared non-spatial and spatial regressions found that the spatial models had better fit [45][46][47]. Spatial models may have been more accurate in these studies HIV prevalence decreased in urban locations and increased in rural locations, but areas of high difference were relatively small.

CHANG ET AL. [41]
Uganda Bayesian modeling of percent and number of people living with HIV (PLHIV) per km 2 17,119 High HIV prevalence along Lake Victoria and patchy prevalence in district interior. Areas with highest number of PLHIV were inland in high population-density trading centers.

COBURN ET AL. [38]
Lesotho Inverse distance weighted (IDW) mapping combined with population density map to display the number of HIV-positive persons per km 2 .

7,099
Density of infection is significantly higher in urban areas, but the majority of HIVpositive people live dispersed in rural areas.
10 countries HIV prevalence within high-prevalence clusters either did not decline or increased, even if national prevalence declined.

CUADROS ET AL. [37]
Tanzania Kriging of HIV prevalence and male circumcision rates to assess their spatial correlation. Areas of low male circumcision overlap with areas high HIV prevalence, and vice versa.

16,140
No spatial pattern for sero-discordancy independent of HIV prevalence patterns.

KALIPENI & ZULU [34]
Continental IDW and kriging interpolation of international HIV prevalence for countrylevel estimates. Model epidemic curves for each country and project future trends.

1,442 sentinel sites over 18 years
Differences between UNAIDS estimates vs. kriging-and IDW-generated national estimates were statistically insignificant. Nearly all countries have reached maturity level of epidemic curve.

KLEINSCHMIDT ET AL. [36]
South Africa Bayesian kriging map of HIV prevalence among youth.

11,758
Variation in HIV prevalence independent of provincial boundaries, highest in the east and for women.

Continuity of HIV estimates across borders.
Certainty of estimates varied depending on total sampling size, total number of administrative units, distribution of survey clusters across area.

MESSINA ET AL. [19]
Democratic Republic of Congo IDW HIV prevalence maps by sex to create regional-level estimates.

9,755
Spatial variation in HIV, distribution and intensity varied by sex.

SARTORIUS ET AL. [40] South Africa
Bayesian kriging of all-cause and causespecific child mortality risk.

46,675
Two geographic foci of high mortality, matching areas of high HIV/TB mortality.

SARTORIUS ET AL. [39] South Africa
Bayesian kriging of all-cause and causespecific adult mortality risk.

104,969
Five geographic foci of high mortality, correlating to areas of high HIV/TB mortality.

SARTORIUS [27]
South Africa Bayesian kriging of age-specific all-cause and HIV/TB mortality risk.

1,110,166 person-years
Spatial distribution of all-cause mortality risk varied by age group, reflecting spatial trends in HIV/TB mortality.

SCHAEFER ET AL. [24]
Zimbabwe Kriging of HIV prevalence and uptake of HIV testing and counseling (HTC).

8,092
HIV prevalence higher in two urban areas for men and women, but HTC uptake lower in those areas and in one other.
(Continued ) because they accounted for underlying, unmeasured environmental, social, behavioral or other confounding factors relevant to HIV prevalence that were geographically determined. Seven studies used the results of spatial regression to map adjusted relative HIV risk over geographic areas, controlling for other factors like age, sex and employment. Maps of adjusted HIV risk visualized spatial patterns and heterogeneity, displaying relative risk by administrative area [21,[47][48][49][50] or over smoothed, continuous surfaces [45,46]. In Zambia, spatial patterns of HIV changed after adjusting for age and spatial correlation so that the two highestprevalence districts were no longer among the areas at highest risk [49]. These results, together with similar findings elsewhere [48][49][50], highlight the extent to which geographic location can provide insights on complex or poorly understood risks for HIV that would not be captured in a standard statistical model. Joint spatial disease/risk factor modeling. Four studies examined how the relationship between HIV and a risk factor or other sexually transmitted infection (STI) changed over geographic space (Table 3) [53][54][55][56]. Such spatial variation may have been due to unmeasured, underlying modifying or cofounding factors that varied across space, such as attitudes, cultures and preferences, or due to sampling variation [53]. Spatial analysis was used to evaluate and display how different risk factors [53,56] or other STIs [53][54][55] were more or less correlated with HIV across different geographic areas. Joint spatial modeling was conducted with Win-BUGS [54,55], R [53]or a combination of ArcGIS and STATA software [56].
These studies aimed to assist resource allocation decisions by identifying opportunities for geographic targeting of HIV prevention interventions [53,56]. They searched for geographic areas where particular risk factors were more relevant to HIV outcomes, for instance finding regions where age at first sex had a stronger association with HIV prevalence [53]. However, models that allowed the effect of each risk factor to differ by spatial location were not consistently better or worse fitting than models that held relationships constant over space [53,56]. There was some evidence that the impact of social risk factors relating to sexual habits varied more over space than the impact of demographic risk factors in a South African analysis [56]. All methods revealed within-country variations and were similar in accuracy, but Bayesian geostatistical approach slightly better.

TANSER ET AL. [23]
South Africa Kernel density smoothing to estimate spatial distribution of HIV.

12,221
Spatial variation in HIV prevalence with highest prevalence in urban settlements near the National Road.

ZULU ET AL. [29]
Malawi IDW HIV prevalence maps for eight years to compare trends over time.
19 ANCs for time trends Spatial variation independent of district boundaries, shifting spatial patterns over time.

GONESE ET AL. [44]
Zimbabwe Compare ANC surveillance with geographically proximate DHS data. Three studies compared spatial patterns of HIV and another STI, either HSV-2 [53,55] or syphilis [54], in order to address questions of population-level ecological associations between STIs. These studies examined the extent to which these STIs were spatially correlated, whether because the infections were mutual risk factors or because they had similar underlying, potentially unmeasured causes in Africa [53][54][55]. Joint spatial modeling revealed that syphilis prevalence did not predict HIV prevalence in South Africa, given their low degree of spatial correlation, and as such would not be suitable as a proxy measure to examine ecological differences in risky sexual behavior resulting in HIV infection [54].
Geography of HIV risk factors. GIS enabled the systematic measurement of spatial risk factors (Table 4). For example, GIS technology allowed for the characterization of a defined area around a study participant's community, such as the risk of schistosomiasis [57] or the

4,864 Geographic distribution and correlation of HIV and HSV-2
County-level Spatial model had best fit. HIV and HSV-2 significantly spatially correlated, with higher risk of both infections in regions around Lake Victoria in the west of the country.

4,864 Geographic variation in the effect of risk factors on HIV and HSV-2
County-level Risk factor variation across space was significant for HSV-2 but not for HIV. Visually, the effects of some demographic and social factors for HIV were stronger in some counties than others.

SPATIAL ANALYSIS OF RISK FACTORS ABIODUN ET AL [64]
Nigeria Early sexual debut Bayesian spatial Cox hazards model for spatial analysis of early sexual debut.

4,301
Northern states significantly earlier sexual debut after controlling for other factors.

AKULLIAN ET AL [65]
Kenya HIV stigma Describe spatial patterns of HIV stigma using difference of K-function cluster analysis and spatial regression.

373
Spatial trend and clustering in external stigma (blame) but not internal stigma (shame).

AKULLIAN ET AL [60]
Kenya Outside of low-MC clusters, females at greater risk than males, but inside low-MC clusters, males and females at equal risk.

Sero-discordant partnerships
Compare Kuldorff clusters of serodiscordant couples and HIV prevalence. Compare epidemiologic measures of discordancy inside and outside clusters. 16,140 No spatial pattern for serodiscordancy independent of HIV prevalence patterns. HIV prevalence correlated with proportion of couples that were sero-discordant.
(Continued ) density of refugee camps within a 25km radius [19]. Similarly, GIS tools were used to measure distances between a household's location and the nearest city [19,35], major road [19,58], body of water or refugee camp [19]. One study used geo-referenced mobile phone data from a cell service provider in Côte d'Ivoire to develop a predictive model of area HIV prevalence based on residents' travel and communication patterns [42]. Local cluster detection methods were used to identify clusters of HIV risk factors, for instance of condom non-use [59]. These risk factor clusters anticipated areas that might benefit from prevention services, and were compared to mapped clusters of HIV to display the overlap between risk and outcome [29,37]. Two studies compared incidence of new HIV infections inside and outside of risk factor clusters [28,37]. Three of the studies that identified clusters of high HIV prevalence [23,28] or mortality [27] also compared characteristics inside and outside of these clusters to identify associated risk factors.
Smoothed maps of risk factors showed geographic variation independent of administrative boundaries enabling new insights [28,37,[60][61][62]. Smoothed mapping of malaria endemicity enabled the estimation of malaria prevalence at every point where HIV prevalence had been measured [63]. Studies compared the spatial distribution of a risk factor to the spatial distribution of HIV to assess evidence of geographic concurrence [28,37,57,61]. Geographic overlap of particular risk factors and HIV enabled a community-or population-level analysis of their relationship, rather than individual-level analysis, and highlighted the extent to which location influences the HIV epidemic. Risk factors were also examined with spatial regression methods, such as early sexual initiation [64] or HIV stigma [65]. In Nigeria, for example, controlling for spatial correlations, demographic and knowledge covariates, and the non-linear effect of age revealed a clear North-South dichotomy in early sexual initiation [64]. Studies analyzing the geography of risk factors often used GIS software like ArcGIS [19,28,29,63], Idrisi [58], and SaTScan [23,59] to extract spatial data about the risk factor that was then incorporated into an analysis in statistical software like R, STATA, SPSS, and others. Alternatively, some analyses were conducted entirely in R [35,[60][61][62]65], BayesX [64] or STATA, using thefunctions in these software applications to handle both spatial and statistical analyses, such as the geonear function in STATA [57].

HIV services implementation
Twenty-three articles contributed to understanding relationships between geography and implementation of HIV services. Two subcategories were the (i) effect of distance to HIV services on health outcomes, and (ii) planning and evaluating HIV service provision.
Effect of distance. GIS technology allowed for measurement of geographic distances and distribution of services and enabled analysis of the effect of distance on health outcomes and service uptake (Table 5). For example, GIS tools enabled the calculation of distances between points, including both Euclidean straight-line distance [29,40,58,[66][67][68][69][70][71][72] and road distance [73][74][75][76] between households and clinics, providing objective measures of geographic burdens to health care access. Evidence was mixed on whether these measures were [77] or were not [66,75,78] correlated with self-reported measures of health care barriers. Studies also examined the impact of proximity to health care on presence of HIV risk hotspots [45,69] and HIV clusters [24,70,72,79].
In some studies, access for HIV patients appeared to be mediated through the public transportation network more than through road access [29,71,75], particularly for patients who were further away from the clinic [71]. This relationship may reflect the positive correlation between road transportation and HIV transmission observed in some settings [29,58]. There was furthermore evidence that the effect of distance on service access differed between those accessing mobile care compared with those attending static clinics [67,73]. HIV-positive patients did not necessarily transfer to a closer clinic when given the chance [66,76], and postpartum women who switched clinics returned to care faster and with better outcomes when they transferred to distant clinics than when they re-entered nearby clinics [80].
Service provision. GIS technology was used to examine the extent to which health care service expansion [60,70,82] and outreach to communities [68,73,76] increased access and uptake ( Table 6). Decentralization of HIV services was found to dramatically decrease the average distance traveled for ART among people living with HIV [68,76] and to diminish the

BASSETT ET AL. [73]
South Africa Distance to testing site HIV testing 4,701 Compare testing for HIV at mobile vs clinic-based sites Mobile testers more likely to test <1km or >5km from home than fixed-site testers.

BUEHLER ET AL. [72]
Mozambique Distance to ARV clinic HIV knowledge 3749 Clusters of high or low knowledge (Getis-Ord statistic) relative to clinic locations. Regression.
Clustering of higher HIV knowledge closer to facilities. Distance negatively associated with outcome.

CARLUCCI ET AL. [77]
Zambia Distance, road distance, travel time to clinic ARV adherence 424 Regression Measures of distance correlated with each other but not associated with outcome.

CLOUSE ET AL. [80]
South Africa Distance between clinic where woman initiated ART and clinic where she re-entered care Time to re-entry, CD4 count at reentry 300 Comparison of median and IQR values between women re-entering care in the same province vs. a different province Post-partum women who re-entered care in a different province had a higher median distance to their new facility, re-entered care faster and had better CD4 count outcomes.

COOKE ET AL. [68]
South Africa Distance to clinic ARV initiation 1,660 Regression Distance negatively associated with outcome.

ESCAMILLA ET AL. [69]
Zambia Distance to clinic PMTCT uptake 254 Uptake density (kernel density estimates) in relation to clinic locations. Regression.
Areas with high-density uptake were located near health centers. Distance negatively associated with outcome, with a 1.9km threshold.

GOLUB ET AL. [67]
Kenya Distance to male circumcision facility Male circumcision follow-up 1437 Regression Distance negatively associated with outcome for fixed facilities but not mobile facilities.

HOUBEN ET AL. [76]
Malawi Travel time and transfers declined, uptake increased as ART clinics opened. Proportion of patients not attending nearest clinic increased slightly.

JOHNSON ET AL. [71]
Malawi Distance from neighborhood to clinic Timely ARV initiation 15,734 Regression Distance negatively associated with outcome for one clinic but not the other, located next to central transport hub.

MEE ET AL. [74]
South Africa Distance to clinic Biomedical vs traditional health use 2,833 Visual analysis of outcome in relation to clinics. Regression.
No spatial patterns or significant association.

MUSENGE ET AL. [45]
South Africa Health facility presence, minimum distance to clinic HIV/TB mortality 6,692 Bayesian spatial regression with visual analysis of odds ratio map in relation to clinic locations.
Odds ratio hotspot is area furthest from clinics. Distance covariate not significant.

SIEDNER ET AL. [78]
Uganda Distance and route distance, travel time and cost to clinic HIV clinic attendance 188 Regression GPS distances negatively associated with outcome, self-reported measures not associated.

SARTORIUS ET AL. [40]
South Africa Distance to clinic All-cause and HIV/ TB mortality

46,675 Regression
Distance not significantly associated with outcome.

SCHAEFER ET AL. [24]
Zimbabwe Distance to clinic HIV testing 8,092 Compared distance and uptake inside and outside high and low clusters of HIV prevalence (Kuldorff spatial scan) Distance not associated with outcome. Those living in high-prevalence clusters had better access but lower uptake of HIV testing.

TANSER ET AL. [58]
South Africa Distance to road HIV prevalence 2,013 Regression Distance negatively associated with outcome.
(Continued ) variation in service uptake over space [60,70]. For example, as HIV testing became more widely available in Mozambique, clusters of high or low rates of testing disappeared, indicating that the geographic barriers to testing became less relevant [70]. Two studies used GIS tools to target HIV services and improve their efficiency [38,82]. One study mapped the density of HIV infection in Lesotho in order to calculate the geographic variation of coverage if the country was to optimize efficiency in treating 70% of its HIV-positive population [38]. In rural areas where most of the HIV-positive population lives, the authors found, coverage would be as low as 4%, whereas the densely populated urban areas

BASSETT ET AL. [73]
South Africa Compare the yield, geographic distribution and demographics of mobile vs clinic-based HIV testing services.

5327
Mobile testers differed from clinic testers in age, gender, and distance travelled to test. HIV prevalence at mobile sites differed by type of venue.

COBURN ET AL. [38]
Lesotho Map the density of HIV infection to compare coverage across districts under efficient vs. equitable resource allocation.

7099
Majority of HIV-positive people live in low-density rural areas that would receive low coverage in optimally efficient resource allocation. Coverage would range from 4% to 94% if areas with 5 infected people per km 2 (70% national coverage) were prioritized.

COOKE ET AL. [68]
South Africa Comparison of median distance traveled for ART over time.

7576
Median distance decreased from 34.2km to 3.1km when treatment was made available through all primary healthcare facilities.

HOUBEN ET AL. [76]
Malawi Track changes in travel time to the nearest clinic providing ART and clinic actually attended as services expanded between 2005 and 2009.

5411
Median travel time to the nearest and attended clinics fell, uptake increased, and the proportion not attending their nearest ART clinic increased slightly.

YAO AND MURRAY [82]
Mozambique would receive up to 94% coverage. Yao and Murray [82] used GIS technology along with optimization software Gurobi to optimize the location of HIV testing and counseling sites, so that the overall distance traveled by the entire population to their nearest testing site would be as short as possible. Software like ArcGIS [38,70,73,76], ArcView [60]and MapInfo [68]was used for visualization and extraction of distance data to be analyzed in statistical software like STATA and R.

Discussion
The review demonstrates the wide applications of GIS and spatial analysis to understanding HIV and supporting HIV care and prevention services in Africa. Most of the literature was published recently and rapid growth in the evidence-base was clear. We expect that GIS use and spatial analysis methodologies will continue to expand as more researchers and implementers develop knowledge and capacity in this growing field and increasingly recognize its usefulness. For example, in this review GIS tools were regularly used to display HIV prevalence, incidence or mortality data in order to communicate complex information about the epidemic in a clear and accessible way. The ability to identify spatial clusters and generate smoothed maps revealed variations within and across administrative regions, the boundaries of which often have little bearing on disease prevalence.
In the context of HIV risk factor analysis, spatial analysis reflected the likelihood that people living close to each other share common exposures and disease outcomes and that location was critical in including in HIV risk factor analyses. The findings that spatial regressions of HIV which accounted for geographic correlation in the data had better fit than non-spatial regressions highlight the benefit of adding spatial data to certain analyses.
Another recurring theme was the critical role of geography in informing access to HIVrelated health services in low-income settings with poor existing health care infrastructure. GIS and spatial analysis may allow for more efficient allocation of resources and appropriate response targeting in many African settings.
Of note, while this review was focused on more technologically and methodologically sophisticated applications, several articles not included in this review demonstrated the utility of simply visually displaying mapped data points, such as for plotting the locations of HIV services [83] and risky sexual behaviors [84][85][86][87][88]. Mapping the overlap of service provision and demand was useful for selecting priority areas for programmatic expansion [81,89,90]. Program implementation could benefit from dynamic, interactive and iteratively updated maps with verified health facility coordinates and infrastructure data [91] and simple mapping may be good entry point for more advanced geospatial techniques.
A number of gaps in the literature were also found by this review. West African countries were less represented, perhaps because they have lower HIV prevalence rates and less HIV research capacity. Central African and Horn countries were similarly less represented. Countries that did not include both HIV biomarkers and geographical coordinates in their DHS surveys also may be less likely to feature in spatial analysis studies because these were characteristics that determined inclusion in certain studies. Some important HIV related outcomes had minimal or no found spatial studies such as HIV medication shortages or stock-outs. Adherence and retention outcomes were also not substantially considered in relation to spatial factors. Finally, there was limited literature on HIV resistance and spatial analysis, potentially because these data are too sparse.
Interesting directions and issues in spatial research were also found. The novel use of cell phone technology in some studies points to an important future area of spatial research. Data independent of national boundaries/national data collection practices to facilitate cross-country comparisons and multi-country analyses would be a helpful future direction given the mobility of persons and complexity of social networks. Privacy concerns continue to be an important area of consideration as GIS tools become easier to use and more broadly implemented.

Limitations
Our search strategy may have missed studies relating to GIS or spatial analysis and HIV, particularly as the vocabulary for GIS is dispersed and evolving. There is some risk of publication bias, potentially favoring articles that detect spatial heterogeneity over those that do not, or favoring particular spatial analysis methods over others. This review was not intended to assess the quality, methodological rigor or risk of bias within studies included for review, nor did we attempt to conduct a meta-analysis. In part, this decision reflects the wide diversity of contexts, approaches, and methodologies that emerged through the review, which were difficult to synthesize. There are not currently any best practice or guidelines for reporting of geospatial studies, and this may be an area for future development.

Conclusions
This systematic review searched for and summarized evidence on the use of GIS and spatial analysis techniques for HIV in Africa. Our findings demonstrate the wide array of spatial approaches to HIV-related data. These applications include characterizing geographic distribution of HIV, evaluating HIV epidemiologic risk factors, and assessing and improving implementation of HIV services. The rapid growth and diversity of applications of GIS and spatial analysis to the field of HIV yields great potential for future insights and progress.
Supporting information S1