Monitoring HIV/AIDS dynamics: an application of real time assessment methods for estimating prevalence at district levels in Uganda

Abstract

Background More than 70% of the HIV cases and >75% of the HIV-related deaths globally occur in Sub-Saharan Africa. In Uganda, 1.5 million people were living with HIV in 2019 with approximately 80% of them being adults aged 15-49 years. The general population prevalence of HIV in Uganda was at 6.2% (7.6% for females and 4.7% among males; regional prevalence varies from 3.1% to 8.0%) in 2016/17. Although the annual number of new infections has dropped significantly since 1990 due to the available effective interventions, the number of new infections remain high. In 2019, there were 53,000 new infections, 21,000 deaths although 87% of the PLHIV were on treatment.

To curb the further spread of the HIV virus, information on population subgroups who are more likely to contribute to new infections is required at more decentralised levels such as districts. Existing databases such as the population surveys and the Health Management Information System (HMIS) do not provide reliable data at district levels. Statistical tools can however be utilised to ensure availability of more timely, precise and reliable district level information to guide planning and decision-making.

Aim of the study The overall aim of this study was to identify and apply advanced statistical methodologies to estimate HIV prevalence and its predictors among adults aged 15-49 years at district levels in Uganda using facility HIV testing, population survey, community Lot Quality Assurance Sampling survey and Census data.

Methods The data sources used to address the specific study objectives were: the nationally representative population-based surveys conducted in 2011 and 2016 among adults aged 15-64 years, who consented to interview and blood draw for HIV testing; HIV testing data collected during service delivery at health facilities, which is reported to the National District Health Information System version 2; Lot Quality Assurance Sampling survey data conducted in 2016 and the National Population and Housing Census conducted in 2014.

The study assessed comparability of health facility and population survey testing data. The 2011 Uganda AIDS Indicator Survey (UAIS) data was stratified according to venue of the most recent HIV test, i.e. tested in a health facility or tested in a community setting. HIV prevalence ratio and the 95% confidence interval was computed using the KATZ methodology.

The Hybrid Prevalence Estimation (HPE) methodology was applied to obtain HIV prevalence estimates for districts in Uganda, through the following steps: (I) a multilevel logistic regression model was applied to the 2011 UAIS data to obtain the propensity of testing for HIV in a health facility. (II) Denominators for the health facility data were adjusted based on the population survey design. (III) District level average propensity to test for HIV in a health facility was then used to combine the adjusted health facility data and the population survey data of non-facility testers to obtain district HIV prevalence estimates for the 111 districts in Uganda. Percentage change in standard errors for the direct survey and HPEs was computed and assessed for improvement in efficiency of the HPEs. Consistency/agreement between HPEs and the direct survey estimates was assessed using the Bland Altman analysis.

Small Area Estimation techniques were applied and assessed for robustness in estimating district level HIV prevalence estimates. Four different approaches were applied namely (i) the direct survey method, where district HIV prevalence estimate was computed based on only sampled individuals from the district; (ii) the area level, Fay-Herriot (FH) model, applied to the 2016 population survey data with health facility antenatal data as the auxiliary variable. The methodology is a two-stage hierarchical model that links the outcome with possible predictors defined at area level; (iii) the Battese, Harter and Fuller (BHF) unit level model, was applied to the 2016 population survey data using the LQAS survey data as the source of auxiliary information. The BHF methodology uses auxiliary information at individual/unit level; and lastly the Spatial Fay-Herriot (SFH) model was applied, assuming a spatial correlation between districts.

The mean squared errors for BHF estimate was computed as a weighted combination of variance among the sample observations and variance of the estimates for the unsampled observations. Similarity and consistency of the direct survey, FH, SFH and BHF model estimates were assessed using regression methods. Percentage improvement in the standard errors were also computed.

Finally, participants’ distribution by the socio-demographic characteristics was analysed separately for the UAIS 2011 and UPHIA 2016 surveys. Descriptive analysis was also carried out to assess change in HIV prevalence between the two surveys. Pooled logistic regression was used to assess factors associated with HIV positivity while multivariable logistic decomposition model was used to determine the contribution of characteristics and coefficients of the characteristics to change in HIV prevalence between the two surveys. The Blinder-Oaxaca decomposition method was used to display the contribution of each characteristics or coefficient of the characteristics to change in HIV prevalence between the two surveys. All analyses were weighted using the population survey weights and analysis carried out using R version 3.5 and Stata Release 15.1.

Results Generally, testing in a health facility was associated with a higher HIV positivity compared to testing in a community setting (Prevalence Ratio=1.8, 95% CI: 1.4-2.2). There was a two-fold or higher increase in HIV positivity among health facility testers compared to nonfacility testers if they were male; aged 15-19 years, 30-39 years and 40-49 years; had no formal education; never married or were previously married; reported no sexual partner in the 12 months preceding the survey; not in formal employment and resident in urban areas.

The propensity to test in a health facility was generally higher among females across all regions of the country. It was also higher in Mid-Northern, North East and Kampala regions of the country. Of the 111 districts included in the analysis, 105 (95.5%) had narrower confidence intervals compared to confidence intervals from direct survey-based estimates. Overall, the HPE standard errors decreased by 28.9% (95% CI: 23.4-34.4) compared to survey-based Standard Errors (SE).

There was no overall difference between survey and HP estimates. Average difference for males was -0.01 units (95% CI: -0.05, 0.03) while for females was 0.00 units (95% CI: -0.06, 0.06). The mean difference between the HPE and DHIS2 HIV prevalence estimates was 0.01 units (95% CI: -0.05, 0.06). Average difference for males was -0.01 units (95% CI: -0.07, 0.06) while for females was 0.02 units (95% CI: -0.05, 0.09).

Although FH, SFH and BHF models and the direct survey estimates were similar, direct survey estimate (0.065) had a larger average SE of 0.055 compared to the FH estimate (0.054 (SE=0.020)), SFH estimate (0.054 (SE=0.021)), and BHF estimate (0.058 (SE=0.026)). Regression results showed that the BHF model estimates were strongly correlated with the direct survey estimates (𝛽1 = 0.73, 𝑟 2=0.891) compared to FH model against direct survey estimates (𝛽1 = 0.27, 𝑟 2=0.477).

Model and survey based estimates were similar for large survey sample sizes in the districts but significantly different for districts with small survey sample sizes. For example Nwoya district with survey sample of 44 individuals, had an HIV prevalence estimate of 0.13 (SE=0.68); 0.062 (0.014) and 0.083 (0.021) while Mbale district with a survey sample of 874 individuals had estimates of 0.053 (0.007); 0.053 (0.007); and 0.052 (0.004) for direct survey, FH and BHF models respectively. Overall average improvement in the precision of the estimates, were 37.5% and 33.1% for the BHF and FH model estimates respectively compared to the direct survey estimates.

Coefficient of variation of the direct survey estimates was generally larger compared to CV of FH and BHF model. If estimates with CV less than 20% were considered suitable for decision making, then only 21 (18%) of the districts would have reliable information for decision making based on direct survey estimates while 53 (47.3%) and 35 (50%) of the districts would have reliable information for decision making based on the FH and BHF models respectively.

HIV prevalence dropped by 1.3 percentage points between 2011 and 2016. A greater reduction in HIV prevalence was observed among Males (1.8); individuals aged 20-24 years (2.0); individuals with primary level of education (1.4); individuals with two or more sexual partners in the 12 months preceding the survey (2.7); and individuals who were divorced/separated (3.5). An increase in HIV prevalence between the 2011 and 2016 HIV prevalence dropped by 1.3 percentage points between 2011 and 2016. A greater reduction in HIV prevalence was observed among Males (1.8); individuals aged 20-24 years (2.0); individuals with primary level of education (1.4); individuals with two or more sexual partners in the 12 months preceding the survey (2.7); and individuals who were divorced/separated (3.5). An increase in HIV prevalence between the 2011 and 2016 survey was observed among individuals with no formal education (0.3), individuals with no sexual partners in the 12 months preceding the survey (0.3) and among individuals aged 40-49 years (1.6). Change in the composition of survey participants contributed to 54% increase in HIV prevalence between 2011 and 2016 time points, while the coefficients contributed to 154% decline in HIV positivity between the two survey time points. While the coefficients contributed to a larger reduction in HIV prevalence, the contribution of most of the variables was not statistically significant except for the age group 40-49years.

Conclusion Information for decision making for districts in Uganda was obtained. The findings show that it’s feasible to use existing population survey and routine facility HIV testing data either combined through the HPE methodology or as auxiliary variables in the FH area level SAE model to obtain more precise indicator estimates for district level planning and decision making. Availability of individual level data such as those obtained using LQAS surveys leads to greater increase in the precision and reliability of the estimates.

Higher risk of infection among individuals accessing health facilities especially male, age groups 15-19 years, never married and those who reported no sexual partners in the 12 months preceding the survey, implies screening protocols should focus on these risk categories to minimise missed opportunities of identifying those who are likely to be HIV positive.

These findings also imply that to use the datasets in complementarity, there is need to take into account HIV prevalence differences by socio-demographic characteristics. Additionally, the compositions of the population subgroups with higher risks of infection also contribute significantly to change in HIV prevalence. Intervention efforts should therefore focus on such subgroups to increase the impact on HIV prevention.

The HPE methodology presents dual benefits: (I) improves precision of the estimates and (II) enables computation of standard errors for routine health facility data whose denominator is difficult to ascertain directly.

Further research should consider incorporating more HIV risk factors in the models, using subgroup HIV prevalence differences as prior information in the models and applying mechanisms to deal with missing data while applying the methods.