Spatial analysis of factors associated with HIV infection among young people in Uganda, 2011

This cross-sectional study utilized secondary data from the 2011 Uganda Demographic and Health Survey (UDHS) and AIDS Indicator Survey (AIS) where data was extracted for participants aged 15 to 24 years. In the primary study, respondents were selected using a two-stage sampling process from stratified urban and rural areas. The first stage involved the selection of a number of Enumeration Areas (EAs) selected from a list of DHS clusters created from a recent population census [32, 33]. The second stage involved selecting households from a household list in the selected EAs where all household members of reproductive age 15-49 (females) and 15-54 (males) were selected [32, 33].

Ethical clearance for this study was obtained from the University of the Witwatersrand Ethics Committee on Human Subjects (M120856). Permission was also sought from the MEASURE DHS ICF International to use the data for secondary analysis.

HIV data was obtained from the UAIS conducted in 2011. Blood collection through finger pricking or venous flow for HIV testing was provided voluntarily by individuals whose households were selected during the survey. Home-based rapid tests and dried blood spots were used for individuals who consented to venous flow and those who preferred finger- pricking respectively. Subsequent testing was performed using Murex and Vironostika Uniform II O to confirm sero-status and ANILAB to resolve discordant results [8].

The sample used in this study was acquired through the merging and appending of five datasets downloaded from the MEASURE DHS website. These were; the standard individual dataset containing socio-demographic information for males and females, a household member dataset with information on all household members, AIS dataset which had indicators for effective monitoring of HIV, and geo-referenced information on the households that participated in the survey. A final dataset which included respondents who were between 15-24 years of age having a GPS co-ordinate and HIV result was used in the final analysis.

The dependent variable was HIV sero-status categorized as being HIV positive/negative. The independent variables examined were socio-demographic, biological and behavioural. Socio-demographic variables were gender (male/female), place of residence (urban/rural), education level (no education/primary/secondary/higher), religion (Christian/Muslim/Traditional/None), marital status (Never married/married/separated/divorced/widowed) and Circumcision (yes/no). Behavioural factors included alcohol consumption in the past 12 months (yes/no), multiple sexual partners defined as having more than one sexual partner in the past 12 months (yes/no), condom use in the past one year (yes/no), transactional sex defined as engaging in sex in exchange for goods and money (yes/no) and coital debut (none/8-10/11-14/15-19/20-24). Young coital debut was defined as age at first sexual encounter below 20 years of age. Biological factors which included the presence of a Sexually Transmitted Infection (STI) or its symptom in the past 12 months (yes/no) were examined. Social Economic Status (SES) was assessed using the wealth quintile. Lowest, second/middle/fourth and highest was categorized as low, middle and high SES respectively. Media exposure looked at respondents who were not, irregularly and regularly exposed to radio, television or newspapers respectively.

Data was analysed using STATA 12 [34] and BayesX [35] software for non-spatial and spatial analysis respectively. Population size adjusting sampling weights at cluster level were also included during the analysis. Descriptive measures were used to summarize the overall characteristics of the study participants in the study area using frequencies and percentages for categorical variables and median (interquartile range) for continuous variables. Likewise, proportions of HIV sero-positive were also reported after adjusting for sampling weights. The statistical significance of apparent associations between potential risk factors and HIV prevalence was explored using chi-square and independent t-test for categorical and continuous variables respectively. Unadjusted and adjusted multiple logistic regression analyses were used to determine associations between the outcome and risk factors. Significant associations (p-value <0.05) from the univariate analysis were included into the final multivariable model. A multiple variable logistic regression analysis was performed without a spatial component to determine factors associated with HIV prevalence as well as adjusting for any potential confounding. This analysis was restricted to participants whose cluster centroid coordinates were collected during the survey and also consented to HIV testing and received the results.

To identify and detect clusters of HIV in the study area, a Poisson-based spatial scan statistic was employed to adjust for the underlying populations in each survey cluster using the spatial scan statistic function of the SaTScan™ software version 9.0 [27]. The spatial scan statistic identified these geographical clusters by taking into consideration the rates of nearby clusters across multiple spatial scales, minimizing the potential for error resulting from the small sample sizes within each individual cluster of households [22]. The Kulldorff's SaTScan program has been widely used in public health research for applying a simple statistic in identifying spatial clusters based on geographic coordinates [36]. The spatial scan method uses a circular window which moves across the map and at each position; the radius of the circular window varies repeatedly from zero up to a set maximum radius of 50 which restricts the maximum size of the window from exceeding 50% of the total study population [37]. In this study, high rate where the observed cases exceeded the expected cases and low rate clusters where the expected cases exceeded observed cases were scanned [36, 37]. The null hypothesis of no clusters was rejected when the p-value was less than or equal to 0.05.

Spatial binomial logistic regression was undertaken via Bayesian estimation based Markov chain Monte Carlo (MCMC) simulation using the BayesX software version 2.1 [35] to adjust for non-spatial and spatial random effects in the model. The spatial random effect model accounts for heterogeneity across spatial units that occur in geo-referenced data [35]. Significant risk factors from the non-spatial model were included as fixed effects for analysis in conjunction with the spatial components in the Bayesian modelling. Posterior Odds Ratios (POR) and their 95% Bayesian credible intervals (lower bound of 2.5% and 97.5%) as well as the spatial (structured) and non-spatial (unstructured) random effect models were reported. POR estimation was achieved by taking into account the autocorrelation in the structure of the data and the regional ambiguity of HIV infection in Uganda [28, 29]. All statistical tests were performed using two-sided tests at the 0.05 level of significance.

The results from the Bayesian and cluster analysis were superimposed to produce a map that displayed low and high rate geographical clusters of HIV in the study area. High (red) and low (green) HIV risk areas were identified as shown in Figure 1. Figure 2 depicted the variations of HIV in different regions in Uganda from the BayesX output. A standard Geographical Information System (GIS) programme [38], Quantum GIS was used to translate the outputs into maps that depicted the distribution of HIV prevalence in Uganda.

Table 1 provides a summary of the socio-demographic and behavioural characteristics of the HIV-positive participants in the study population. The mean (sd) age of this study participants was approximately 19 (2.86) years. Slightly more than half of the study participants were female (n = 309, 58%). Majority of the participants resided in the rural areas (n = 386, 72%). Compared to urban areas, the proportion of HIV positive participants was higher in the rural areas (5%) although this difference was not significant. Slightly more than half (57%) were of low socio-economic status and had an irregular exposure to media sources (n = 328, 74%). Basic literacy levels were shown by a large number of participants possessing primary school (n = 314, 59%) education and fewer individuals progressing beyond secondary school (n = 22, 6%).

Majority of the participants were married (n = 319, 60%) and reported a mean coital debut age of 16 years. Circumcision rate among the young male HIV positive participants was high (n = 145, 78%). Multiple sexual partnerships among HIV positive participants (n = 371, 71%) were rife in the study population and majority of the participants used condoms during sexual encounters (n = 307, 81%). Slightly above half of the participants reported the presence of an STI or its symptoms (n = 268, 54%). 90% of the HIV positive participants reported having engaged in transactional sex. Alcohol use was rampant in this study population as more than half of the study participants reported being inebriated during sexual intercourse (n = 258, 68%).

The spatial cluster analysis detected one significant primary (Table 2) and fifteen tertiary clusters. The most likely cluster comprised of thirty two locations situated around the Central and Eastern regions of Uganda. In this area, 70 cases were detected during the study period, while the number of expected cases was 37. An Odds Ratio [OR] (2.14, 95% CI 1.63–2.81) was also estimated which implied that locations in the primary cluster were 2.14 times more likely to be HIV infected than those outside. This geographical cluster with a radius of 74.11 km covered a great part of central Uganda. The tertiary clusters were located towards the north-east, east and south-east of Uganda, as shown in Figure 1.

This section presents the results from the univariate and multivariable analysis of the factors predisposing participants to HIV infection in Uganda. The multivariable analysis presents only the variables that were significantly associated with HIV infection in the bivariate analysis in Table 3.

Young people who were previously [OR = 5.62; 95% CI (1.52-20.73), p = <0.01] and currently married [OR = 3.64; 95% CI (1.23-10.72), p = 0.01] had a higher risk of HIV infection relative to those who were never married. Higher education (tertiary) was protective against HIV infection by decreasing the likelihood significantly by 68% [OR = 0.34; 95% CI (0.17-0.68), p = 0.02]. As expected, circumcision among the males reduced the chances of acquiring HIV infection by 34% compared to uncircumcised males [AOR = 0.66; 95% CI (0.44-0.99), p = 0.02]. Risky sexual behaviour was associated with HIV infection before controlling for demographic factors. A young coital debut (below 14 years) increased the risk of HIV among young people by 58% and this risk was seen to decrease as the age at sex debut increased. Having more than one sexual partner significantly increased the risk of HIV infection [OR = 1.94; 95% CI (1.42-2.65), p = <0.01]. Transactional sex was positively associated with increased risk of HIV infection [OR = 4.14; 95% CI (1.31-13.13), p = 0.02]. Participants who reported engaging in sex while inebriated increased their chances of acquiring HIV infection by 69% when compared with those who did not use alcohol [AOR = 1.69; 95% CI (1.57-2.45), p = 0.02]. STIs are a risk factor for HIV infection. The results from this study indicate that young persons aged 15–24 years, who reported having a sexually transmitted infection in the past 12 months, were approximately twice as likely associated with HIV infection compared with their counterparts who did not report a sexually transmitted infection during that same period of time [AOR = 2.21; 95% CI (1.35-3.60), p = <0.01]. Condom use is critical for the prevention of HIV and AIDS and other sexually transmitted infections and provided 72% protection against HIV infection relative to non-users [AOR = 0.28; 95% CI (0.15-0.55), p = <0.01].

This section presents results from the multivariable analysis of the factors associated with HIV infection in Uganda after controlling for the random and spatial effects respectively. The multivariable analysis presents only the variables that were significantly associated with HIV infection in the univariate analysis (Table 4).

Previously married participants [POR = 3.23; 95% BCI (1.23-8.44), p = <0.01] had higher risks of HIV infection relative to those who were never married. Adjusting for random and spatial effects provided a protective effect against HIV infection in married participants reducing the chance of HIV infection by approximately 61% [POR = 0.39; 95% BCI (0.18-0.72), p = <0.01]. Circumcision among males reduced HIV infection by 36% [POR = 0.64; 95% BCI (0.48-0.84), p = <0.01]. Risky sexual behaviour after controlling for spatial random effects was associated with HIV infection among the young people in Uganda as well. Participants who reported engaging in sex while inebriated increased their chances of acquiring HIV infection by 32% when compared with those who did not use alcohol [POR = 1.32; 95% BCI (1.11-1.63), p = 0.02]. Participants who reported presence of an STI in the past 12 months, had an increased chance of HIV infection by 72% compared to those with no symptoms [POR = 1.72; 95% BCI (1.49-1.95), p = <0.01]. Condom use significantly reduced the chance of HIV infection by 46% [POR = 0.54; 95% BCI (0.41-0.69), p = <0.01]. Mapping was subsequently done based on the posterior estimates from the multivariable binomial regression spatial model as shown in Figures 1 and 2. Figure 1 illustrates the areas perceived as high (red) and low (green) areas for HIV infection among young people in Uganda. This was achieved by overlaying the results from the cluster and Bayesian analysis. As expected, majority of the areas in Uganda are seen as having a high HIV prevalence. The map in Figure 2 shows the estimated posterior regional odds of HIV infection after adjusting for geographical locations. Spatial analysis revealed the regions depicted in orange and red colours with a significantly high HIV prevalence while yellow-coloured regions showed significantly lower HIV prevalence with odd ratios less than 1.

The major strength of this study was use of spatial analytical techniques had advantages over standard statistical techniques to identify geographical variations of HIV prevalence in Uganda. This may be of public health significance in the fight against the spread of HIV/AIDS not only in Uganda but in other countries gravely affected by this scourge. Cluster analysis using the Scan Statistic method adjusts for population density and reduces selection bias as the clusters are explored without subsequent knowledge of their location, size or time period [51]. The use of the Bayesian approach, by adding a spatial random effect, reduces bias and inaccurate conclusions that would arise from ignoring spatial auto-correlation present in the associated factors [29]. As inherent with all cross-sectional studies, this study could neither establish temporality nor causality of the observed associations with the outcome. Self-reporting of sexual behaviours could have introduced recall or social desirability bias. Performing cluster analysis in areas of Uganda where the survey communities were further apart may possibly render the spatial scan statistic less robust. Finally, lack of geocodes in sampled areas failed to depict a correct representation of cluster analysis results with the Western most area of Uganda devoid of any spatial HIV clusters.