Spatial analysis of HIV infection and associated individual characteristics in Burundi: indications for effective prevention

The 2010 Burundi DHS was conducted by national authorities and ICF International between August 29, 2010 and January 30, 2011 following the standardized DHS methodology [5]. This methodology and crude results are described extensively elsewhere [17]. Briefly, the 2010 Burundi DHS used a stratified two-stage random cluster sampling design. Stratification was made at two levels: the provincial level (17 provinces) and the urban or rural area. At the first stage, 376 of the 8104 enumeration areas (i.e. groupings of households) identified in the 2008 national population and housing census [19] were selected with a probability proportional to their size (i.e. the number of households in each enumeration area). At the second stage, 24 households were selected in each enumeration area with an equal probability, leading to a total of 9024 eligible households. HIV tests were proposed to 50 % of the 9024 households. All women aged 15–49 years and men aged 15–59 years living in or having spent the previous night in one of these households were eligible for HIV testing. The Institutional Review Board of ICF International and the National Ethics Committee of Burundi approved the study protocol. After being provided with information, respondents aged 18 years or older and parents or guardians of minors aged 15–18 years gave their written consent to participate in the survey. Minors gave their oral assent.

Blood spot samples were collected from consenting household residents’ fingers and put on filter papers. Serologic screening for HIV infection was then performed on the dried blood spots at the Public Health National Institute in Bujumbura-Mairie using an enzyme-linked immunosorbent assay (ELISA; Vironostika HIV Uni-Form Ag/Ab, Biomérieux, Marcy l’Etoile, France). All positive samples and 10 % of the negative samples (for quality control) were further tested with a second ELISA (Enzignost HIV Integral II, Siemens, Erlangen, Germany). Samples with discordant results between both tests were reanalyzed using a line immunoassay (INNO-LIA HIV I/II Score, Innogenetics, Gent, Belgium).

Social and behavioral characteristics of residents were collected using a standardized questionnaire and included gender (woman or man), age (continuous variable), level of education (no formal education, primary, secondary, or superior), marital status (single, married, cohabiting, divorced, separated, or widowed), religion (Catholic, Protestant, Muslim, Adventist, Jehovah’s witness, other, or none), wealth quintile (poorest, poorer, middle, richer, or richest), sexual activity (never had sex, active in the last four weeks, or not active in the last four weeks), the number of extramarital sex partners in the last 12 months, and a history of sexually transmitted infection (STI) in the last 12 months (yes or no).

The geographic coordinates (latitude and longitude) of the 376 enumeration areas were collected using the geographic information system (GIS) and global positioning system (GPS) technologies. They were recorded at the center of the enumeration areas. In order to preserve the confidentiality of the respondents, the GPS latitude/longitude positions were randomly displaced (up to 2 km for the urban enumeration areas and up to 5 km for the rural enumeration areas, with 1 % of the rural enumeration areas being displaced up to 10 km).

For the present study, the analyses were restricted to the 15–49 year age group because HIV testing had not been performed in women aged 50–59 years (by contrast to men) in the 2010 Burundi DHS. It is worth noting that UNAIDS also reports adult data in the 15–49 year age group [1, 18].

All analyses used the individual-level data and their specific weights provided in the DHS databases (available from the DHS program website [5]). As usual, the weights took into account the survey design of the 2010 Burundi DHS and the proportion of respondents in each enumeration area. They were equal to the inverse of the probability for a given resident of being included in the survey.

First, we computed and mapped the crude estimations of HIV prevalence in the 17 provinces. We then used a non-spatial logistic regression model to assess the relationship between HIV infection and the provinces, taking the province with the lowest HIV prevalence as the reference category. These analyses were performed using the Stata software version 11 [20].

Second, we analyzed the spatial autocorrelation of HIV prevalence data by performing a global Moran test. Moran’s I statistic tests the null hypothesis that observed data at one location are independent of data at other locations. Its value ranges from −1 (data perfectly dispersed), 0 (data randomly dispersed) to 1 (data perfectly correlated). Because the Moran’s I statistic showed the existence of a significant spatial autocorrelation of our HIV prevalence data, we subsequently analyzed the data using a geostatistical approach which takes into account this spatial autocorrelation.

Third, we mapped HIV prevalence throughout the country independently of provincial boundaries using a Gaussian kernel density estimation with adaptive bandwidths and the specific prevR package of R software (R Core Development Team, April 10, 2014) [21, 22]. This approach is promoted by UNAIDS and has been used to estimate HIV prevalence at a sub-national level in various countries including Burundi [18, 23]. It allows generating a smoothed surface of HIV prevalence based on observed data. In our study, we interpolated HIV prevalence data in 449,065 points using the observed data in the 376 enumeration areas. As recommended by Larmarange and Bendaud [23], we set the number of observations at 500 so that the bandwidths adapt to capture this minimum number.

Finally, we identified the spatial clusters with high and low HIV prevalence using the Kulldorff spatial scan statistics (SaTScan software version 9.3) [24]. This method has been widely used in the last years, especially in the field of HIV/AIDS [7, 10, 12, 15, 25‐27]. It allows finding the location of areas with higher or lower numbers of HIV cases than expected under the hypothesis of uniform spatial distribution of cases by gradually scanning circular windows of various sizes across the study area. We assumed that the number of HIV cases in each circular window was an independent Bernoulli random variable. For the circular windows, we used a maximum radius of 15 Km for the detection of clusters with high HIV prevalence and of 50 Km for the detection of clusters with low prevalence. We chose these radii because high HIV prevalence was more likely in small, densely populated areas such as in Bujumbura-Mairie where the maximum distance between the centroid and the city’s boundaries is of 13 Km while low HIV prevalence was more likely in large, sparsely populated areas. We also used the default value of 50 % of the total study population for the maximal size of the clusters. The statistical significance of clusters was ascertained using the likelihood ratio test and its associated p-value obtained through 999 Monte Carlo simulations. The null hypothesis of uniform spatial distribution of HIV cases (no cluster) was rejected if the p-value was <0.05. When a cluster was identified, the strength of the clustering was estimated using the relative risk of excess HIV cases.

The association of social and behavioral characteristics with HIV infection (infected or not) was investigated using a spatial logistic regression model performed with the BayesX software version 2.1 [28]. This model allowed adjusting for the spatial and non-spatial random effects of provinces. The parameters were estimated using 400 Markov chain Monte Carlo simulations in restricted maximum likelihood regression models. Independent covariates associated with HIV infection with a conservative p-value of <0.2 in univariate analysis were subsequently tested in multivariate analysis [29]. A backward elimination procedure was used to determine the final model containing only the covariates significantly associated with HIV infection. The strength of associations was estimated using the posterior odds ratios (PORs) and their 95 % credible intervals (CrIs). Finally, the goodness-of-fit of models was assessed using the conditional Bayesian Information Criterion (BIC).

All statistical tests were interpreted at the 0.05 significance level.