ART coverage and viral load suppression rates as correlates to HIV positivity in Kenya; Spatial-temporal analyses 2015-17

High antiretroviral therapy (ART) coverage and high rates of viral load suppression (VLS) should reduce transmission of HIV, and ultimately, HIV incidence and the number of new HIV diagnoses out of the number tested (HIV positivity). We used 3 years of HIV program data in Kenya to assess whether trends in the number of new HIV diagnoses were associated with ART coverage and VLS rates and spatial-temporally auto-correlated at county level. Methods We analyzed routine program county-level aggregate data on ART coverage and VLS (proportion of persons on ART with VL<1000 copies/mL) from 3 years (2015-2017). We examined the association between ART coverage and VLS rates to HIV positivity by fitting spatial and spatial-temporal semi-parametric Poisson regression models using R-Integrated Nested Laplace Approximation (INLA) package. We used the extended Cochran-Mantel-Haenszel stratified test of association to test for trend for rates across years and Kruskal-Wallis equality-of-populations nonparametric rank test to compare medians for continuous variables. We fit a structural equation model to assess direct and total effects between the two exogenous covariates to adjusted newly HIV-diagnosed as the endogenous variable adjusting for clustering by 47 counties. Finally, we mapped adjusted HIV positivity using QGIS version 3.2.


Introduction
High antiretroviral therapy (ART) coverage and high rates of viral load suppression (VLS) should reduce transmission of HIV, and ultimately, HIV incidence and the number of new HIV diagnoses out of the number tested (HIV positivity). We used 3 years of HIV program data in Kenya to assess whether trends in the number of new HIV diagnoses were associated with ART coverage and VLS rates and spatial-temporally auto-correlated at county level.

Methods
We analyzed routine program county-level aggregate data on ART coverage and VLS (proportion of persons on ART with VL<1000 copies/mL) from 3 years (2015-2017). We examined the association between ART coverage and VLS rates to HIV positivity by fitting spatial and spatial-temporal semi-parametric Poisson regression models using R-Integrated Nested Laplace Approximation (INLA) package. We used the extended Cochran-Mantel-Haenszel stratified test of association to test for trend for rates across years and Kruskal-Wallis equality-of-populations nonparametric rank test to compare medians for continuous variables. We fit a structural equation model to assess direct and total effects between the two exogenous covariates to adjusted newly HIV-diagnosed as the endogenous variable adjusting for clustering by 47 counties. Finally, we mapped adjusted HIV positivity using QGIS version 3.2.

Results and discussion
A spatial-temporal model with covariates was better in explaining geographical variation in HIV positivity (deviance information criterion (DIC) 381.2), than either a non-temporal spatial model (DIC 418.6) or temporal model without covariates (DIC 449.2). Overall, the adjusted HIV positivity decreased over 3 years from median of 2.9% in 2015, [interquartile range (IQR): 1.9-3.4] to 1.5% in 2017, IQR(1.3-2.0), p=0.032. While adjusting for clustering and covariance, VLS had a direct effect on HIV positivity rates p=0.004, but ART coverage did not, p=0.843.

Conclusions
From 2015-2017, there has been improved ART coverage and sustained VL coverage and suppression rates. We have observed a general decline of rates of HIV positivity associated with VLS rates. To assess the trends and impact of implementation of scaled-up care and treatment, spatial-temporal analyses help to identify geographic areas that need focused interventions.

Background
HIV epidemic control will be attained when persons living with HIV/AIDS (PLHIV) have been identified, put on antiretroviral therapy (ART) t and are virally suppressed. This cascade in HIV diagnoses, engagement in care, initiation on ART and the impact of treatment resulting in viral load suppression (VLS, proportion of persons on ART with VL < 1000 copies/mL)) has been proposed as the UNAIDS fast-track targets referred to as the 90-90-90 or 95-95-95 [1,2]. It is expected that by the year 2030, 95% of PLHIV will be diagnosed, 95% of them linked to HIV treatment and 95% will be virally suppressed. High be the driver of the epidemic [4].
Exploring the association between geographic location and programmatic HIV targets such as 90-90-90 is important to help focus interventions [5]. Based on this chronological relationship from HIV diagnoses, linkage to care, initiation of ART after linkage and VLS, we postulated that ART coverage and VLS can be analyzed as covariates in a structural equation model to explain observed differences in new HIV diagnoses as one of the impact indicators for HIV epidemic control. Such structural equation models provide a quick snapshot of associations between correlates to a phenomenon of interest [6][7][8] such as in our case, variation in rates of HIV positivity. Covariates measured at the individual level have been demonstrated as associated with access to HIV care, ART uptake, adherence and better outcomes. These include; age, sex, location, having experienced previous illness or health conditions or symptoms of disease, disclosure and a supportive family [9][10][11]. These are useful covariates but may not be easy to summarize and analyze at geographical level.
We used 3 years of HIV program data in Kenya to assess whether trends in the number of new HIV diagnoses were associated with ART coverage and VLS rates, and explored spatial-temporal autocorrelation at county-level.

Program setting and data sources
In Kenya, HIV program planning is done using the county level as the planning unit. There are 47 counties in Kenya with wide ranging HIV burden from the highest in Homabay County at 26.0%, which is about 4.5 times higher than the national prevalence to < 1% in Wajir County [12]. The estimated number of PLHIV in Kenya was 1,517,707, 1,587,840, 1,493,381 while ART coverage 61.3, 68.6 and 74.6% in 2015, 2016 and 2017, respectively [13]. We aggregated ART coverage and VLS data for 3 years (2015-2017). By 2017, there were 10 VL testing laboratories in Kenya. These laboratories receive dried blood spot (DBS) samples from HIV care and treatment sites from all over the country. The hub laboratories are organized such that they receive samples from the facilities closest to them. HIV testing and treatment data were downloaded from the data for accountability, transparency and impact (DATIM) repository < https://www.datim.org/>. These data come from all geographic regions providing HIV care and treatment services in the country.
Shape polygons used in the manuscript were downloaded from a public repository < http://data.ilri.org/geoportal/catalog/main/home.page>.

Measures
We defined ART coverage as the estimated ratio of persons currently on ART out of the estimated number of PLHIV for a particular year according to national HIV estimates in Kenya [14]. Viral load was measured in copies/mL of viral load ribonucleic acid (HIV RNA) measured after 6 of follow-up after initiation of ART according to Kenyan guidelines [15].
VLS rates were directly calculated from PLHIV with VL < 1000 cells per mL out of persons on ART with a VL result. The number of HIV-positives identified during the reporting period were recorded, and the crude positivity rate was defined as the proportion of tests performed that were reactive. Fitted estimates were derived after spatial-temporal analyses.

Statistical analyses
We examined the association between ART coverage and VLS rates to HIV positivity by fitting spatial and spatial-temporal semi-parametric Poisson regression models using R-Integrated Nested Laplace Approximation (INLA) package [16, 17] and mapped adjusted HIV positivity using QGIS version 3.2. To assess spatial relationships, we fitted these semiparametric Poisson regression models: (1) spatial-temporal model without covariates; (2) a spatial non-temporal model with covariates; and (3) a spatial-temporal model with covariates as proposed by Blangiardo, Cameletti and Rue [18]. For each of these spatial models we used Bayesian Deviance Information Criterion (DIC) according to Spiegelhalter et al., [19,20] to assess the strength of the fit. Since Bayesian analyses are based on an assumed probability model. Appropriateness of these models can be assessed using DIC.
We used the extended Cochran-Mantel-Haenszel stratified test of association to test for trend across the three years for fitted rates of HIV positivity, Kruskal-Wallis equality-ofpopulations nonparametric rank test to compare medians for continuous variables and a structural equation model implemented in Stata version 14.1, (Stata Corp, College Station, TX, USA) to assess direct effects between the two exogenous covariates to fitted newly HIV-diagnosed as the endogenous variable (ε 1 ) adjusting for clustering by 47 counties. We assessed for goodness of fit for the model using the standardized root mean squared residual and coefficient of determination. We also used the equation-level goodness of fit to assess for correlation between endogenous variables and their predictions using the Bentler-Raykov squared multiple correlation coefficient. We then looked for modification indices to ascertain that the model was correctly constructed. We reported direct and total effects and described associations using a path diagram.

Results
Overall summary of covariates, their rates and outcomes are presented in Table 1.     We acknowledge that ecological analyses are flaunt with biases due to inherent individuallevel characteristics that may be sufficiently accounted for. Our data are aggregated at county level therefore it is not possible to do more granular analyses e.g., by age and other classifications for example by HIV testing strategies, sex etc. Although similar spatial-temporal reduction in transmission rates among infants has been demonstrated in a prevention of mother to child transmission (PMTCT) setting [27], VLS rates for mothers was not a factor in these analyses, and neither was it possible to relate the impact of virally suppressed mothers and PMTCT. Finally, we assumed that the numbers reported as newly diagnosed were indeed new diagnoses yet retesting of previously diagnosed may happen to some unknown extent. Although these results may be triangulated with population-level HIV impact assessments (PHIAs) whenever available, PHIAs are conducted every 5 years yet HIV programs in the era of epidemic control need fast and robust analyses.

Conclusions
Our analyses demonstrate the possibility of using programmatic data in a spatial context to describe outcomes that are associated with HIV epidemic control.     Figure 3b shows the direct relationship between correlates and HIV positivity rates.