Accuracy of self-reported HIV testing history and awareness of HIV-positive 1 status among people living with HIV in four Sub-Saharan African countries

22 Background: In many countries in Sub-Saharan Africa, self-reported HIV testing history 23 and awareness of HIV-positive status from household surveys are used to estimate the 24 percentage of people living with HIV (PLHIV) who know their HIV status. Despite 25 widespread use, there is limited empirical information on the sensitivity of those self- 26 reports, which can be affected by non-disclosure. 27 Methods: Bayesian latent class models were used to estimate the sensitivity of self- 28 reported HIV testing history and awareness of HIV-positive status in four Population- 29 based HIV Impact Assessment surveys in Eswatini, Malawi, Tanzania, and Zambia. 30 Antiretroviral (ARV) metabolites biomarkers were used to identify persons on treatment 31 who did not accurately report their status. For those without ARV biomarkers, the 32 pooled estimate of non-disclosure among untreated persons was 1.48 higher than those 33 on treatment. 34 Results: Among PLHIV, the sensitivity of self-reported HIV testing history ranged 96% 35 to 99% across surveys. Sensitivity of self-reported awareness of HIV status varied from 36 91% to 97%. Non-disclosure was generally higher among men and those aged 15-24 37 years. Adjustments for imperfect sensitivity did not substantially influence estimates of 38 of PLHIV ever tested (difference <4%) but the proportion of PLHIV aware of their HIV- 39 positive status was higher than the unadjusted proportion (difference <8%). 40 Conclusions: Self-reported HIV testing histories in four Eastern and Southern African 41 countries are generally robust although adjustment for non-disclosure increases 42 estimated awareness of status. These findings can contribute to further refinements in 43 methods for monitoring progress along the HIV testing and treatment cascade.

estimated awareness of status. These findings can contribute to further refinements in 43 methods for monitoring progress along the HIV testing and treatment cascade. 44 Keywords: Sensitivity; Bayesian latent class; self-report; testing behaviors; HIV 45 disclosure; HIV/AIDS 46 47 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020. ;https://doi.org/10.1101https://doi.org/10. /2020   Consideration of the potential for measurement bias is needed when interpreting self-60 reported survey data. Studies have shown that self-reporting about sensitive 61 information, such as an individual's HIV testing history and HIV status, could be affected 62 by non-disclosure (6,9,10). For example, inconsistencies have been documented in 63 Kenya and Malawi between an individual's self-reported data and biomarkers for 64 metabolites of antiretrovirals (ARVs) and viral load suppression (5, 10). While previous 65 studies have sought to validate the accuracy of self-reported HIV status (10-12), 66 analyzing recent data on both non-disclosure of self-reported HIV testing history and 67 HIV status among PLHIV is key to improving the validity of these estimates. 68 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020. ;https://doi.org/10.1101https://doi.org/10. /2020 Surveys that collect both self-reported information and ARV biomarkers can be used to 69 assess the accuracy of self-reported HIV testing histories and HIV awareness status. In 70 this study, Bayesian latent class models are used to estimate the sensitivity of self-71 reported HIV testing history and awareness of HIV status among PLHIV based on the 72 presence of detectable ARVs (13). 73

Study population 75
The Population-based HIV Impact Assessment (PHIA) surveys are nationally 76 representative multistage household-based surveys designed to provide population-77 level information on the burden of HIV disease and to document the progress of HIV 78 programs (14-17). All four PHIA surveys with available microdata on PLHIV aged 15+

Self-reports and antiretroviral (ARV) status 83
Participants who reported having ever received the results of any HIV test were 84 classified as ever tested and received results (hereafter referred to as "ever tested"; see 85 Table S1). Participants who reported having received a positive test result after any HIV 86 test, were classified as aware of HIV-positive status. The specific laboratory algorithms 87 used to detect ARVs varied across surveys, although all were analyzed in the same 88 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020. ; https://doi.org/10. 1101 laboratory, and included drugs in the nationally recommended first-and second-line 89 regimens: efavirenz, lopinavir, and nevirapine (14-17). 90

Bayesian latent class models 91
Bayesian latent class models (18) were used to quantify the sensitivity of both self-92 reported HIV testing history and HIV status awareness among PLHIV. Cross-tabulations 93 of self-reports with ARV biomarkers provide empirical information on their sensitivity 94 among those with detectable ARVs ( Figure 1A). 95 With regard to participants with detectable ARVs, we assumed that: (1) they had been 96 tested for HIV, received their results, and were aware of their status; and (2) there were 97 no false positives in the detection of ARV metabolites, self-reported HIV testing history, 98 or awareness of HIV status. 99 As ARV metabolite data only provide information about the sensitivity of self-reports 100 among participants on ARVs, the ratio of non-disclosure for PLHIV without detectable 101 ARVs versus those with detectable ARVs was given a log-normal prior distribution with 102 a mean of log(1.48) (standard error: 4) to estimate the sensitivity of self-reports for 103 people without detectable ARVs. This prior was elicited by reviewing available studies 104 an meta-analyzing the evidence. The pooling of two studies conducted in rural 105 Mozambique and Malawi (19, 20) suggests that people not receiving ARVs are 1.48 106 more likely to not disclose their diagnosis. Additional analyses were conducted to 107 investigate the influence of this prior on our results. Equations and prior distributions are 108 presented in the Supplementary Materials (Table S2 and Text S1). 109 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020. ; https://doi.org/10. 1101 Given known biases in self-reported estimates of HIV status awareness, analysts often 110 manually reclassify individuals not aware of their status but with detectable ARVs -as in 111 published PHIA reports. Most surveys, however, do not collect ARV biomarkers and 112 only rely on self-reported information. To examine the impact of this partial adjustment, 113 we compared the unadjusted, ARV-reclassified (as in PHIA reports), and Bayesian-114 adjusted estimates of PLHIV aware. 115 Models were run separately for each country and for subgroup analyses (i.e. age, sex, 116 urban/rural and socio-economic status). Bayesian hierarchical models using Markov 117 Chain Monte Carlo (MCMC), implemented through the JAGS software (21) and the 118 rjags packages, were used to approximate the posterior densities (22,23). CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020.  Figure 2). For people without ARV metabolites, the estimated sensitivity was 2.4% 146 points (0.1-11.4%) lower than those with detectable ARVs in Tanzania.The differences 147 were smaller elsewhere. Detailed values can be found in Supplementary Table 3. 148

Self-reported awareness of HIV status 149
The sensitivity of self-reported awareness of HIV-positive status among participants with 150 ARV metabolites was 97.4% (96.7-98.0%) in Eswatini, 94.2% (93.0-95.4%) in Malawi, 151 92.3% (90.5-93.8%) in Tanzania, and 91.6% (90.1-92.9%) in Zambia (Figure 2A). The 152 estimated differences in sensitivity between PLHIV with ARV metabolites and those 153 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Differences by gender, age, rural/urban, and socioeconomic status 156
Among participants with detectable ARVs, women had 0.9-2.4% points higher 157 sensitivities of self-reported HIV testing history and HIV status awareness than men 158 ( Figure 2B). The estimated sensitivities were the lowest at age 15-24 years (94.7-97.2% 159 for HIV testing history and 83.9-91.9% for HIV status awareness) in all of the countries 160 ( Figure 2C). Participants residing in urban and rural areas had similar sensitivities 161 ( Figure S1A) and variations by socio-economic status (SES) were also small ( Figure  162 S1B and Figure S2). 163

Adjusted proportion of PLHIV ever tested and PLHIV aware of their status 164
Adjusting for imperfect sensitivity influenced the estimates of the self-reported 165 proportion ever tested for HIV less (largest difference between the adjusted and the 166 self-reports was 3.9% points in Tanzania

173
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020. ; https://doi.org/10. 1101 Discussion 174 Self-reported information on HIV testing and diagnosis are primary data sources used to 175 monitor trends in the HIV treatment and care continuum (2, 9). These same data have 176 also been proposed to estimate cross-sectional HIV incidence (24). In this study, we 177 leveraged ARV biomarkers from four household representative surveys in Eswatini, 178 Malawi, Tanzania and Zambia to estimate the sensitivity of self-reported HIV testing 179 history and awareness of HIV status among PLHIV. We found that self-reports of HIV 180 testing history have a high sensitivity (>96%) among PLHIV with detectable ARVs 181 across these settings. Self-reported awareness of HIV status had a marginally lower 182 sensitivity (>91%) in these same countries. In this study, we estimated the sensitivity of the self-reports alone but when ARV 192 biomarkers are available, presentation of cascade results from the surveys usually 193 adjust these self-reports by reclassifying PLHIV that do not disclose their status but for 194 which ARV metabolites are detected as "aware". We have found that this partial 195 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020. ;https://doi.org/10.1101https://doi.org/10. /2020 adjustment may be insufficient, especially if ART coverage is low in the surveyed 196 population and the ratio of non-disclosure among those not on ART is high (26)(27)(28)(29). To 197 accurately estimate awareness of status, results must also be adjusted for non-198 disclosure among PLHIV with undetectable ARVs. 199 Our results need to be interpreted considering certain study limitations. First, only four 200 PHIA surveys have publicly available micro-data, none of which are located in the West 201 and Central African regions, where non-disclosure could be higher (30). The PHIA 202 included here had some of the lowest levels of non-disclosure of these reviewed studies 203 suggesting that other settings could have lower sensitivities. Second, our study design 204 limited our assessment of the sensitivity of testing history to PLHIV, and findings should 205 not be extrapolated to people not living with HIV. Third, it is not possible to empirically 206 validate the sensitivity of self-reports among PLHIV without ARV metabolites. As such, 207 we had to use information from two previous studies that used medical records to inform 208 the non-disclosure ratio. Results could be sensitive to this non-disclosure ratio but the 209 high ART coverage in the four countries mitigates this influence ( Figure S3). Finally, the 210 specificity of self-reports was assumed to be 100% which could lead to overestimating 211 the proportion of PLHIV ever tested / aware of HIV-positive status. However, previous 212 study has shown a high specificity of self-reported HIV testing results (11) implying that 213 this assumption will likely have little impact on the outcomes. 214 Strengths of this study include the use of standardized survey and laboratory data (i.e. 215 detection of ARV metabolites). Second, the Bayesian latent class models propagate 216 uncertainty to our results by assuming prior distributions and generating posterior 217 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020. ; https://doi.org/10.1101/2020.09.16.20196105 doi: medRxiv preprint credible intervals. Finally, we examined sex, age, urban/rural, and SES differences in 218 the sensitivity of self-reports. 219 In conclusion, self-reported HIV testing histories have high sensitivities in the four 220 countries examined but self-reported awareness of HIV status are lower. Whenever 221 available, ARV biomarkers data can be used to adjust self-reports but such adjustments 222 may still underestimate diagnosis coverage, especially if ART coverage is low in that CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 18, 2020. ; https://doi.org/10. 1101