List randomization for eliciting HIV status and sexual behaviors in rural KwaZulu-Natal, South Africa: a randomized experiment using known true values for validation

Noah Haber, Guy Harling, Jessica Cohen, Tinofa Mutevedzi, Frank Tanser, Dickman Gareta, Kobus Herbst, Deenan Pillay, Till Bärnighausen, Günther Fink

List randomization (LR), a survey method intended to mitigate biases related to sensitive true/false questions, has received recent attention from researchers. However, tests of its validity are limited, with no study comparing LR-elicited results with individually known truths. We conducted a test of LR for HIV-related responses in a high HIV prevalence setting in KwaZulu-Natal. By using researcher-known HIV serostatus and HIV test refusal data, we were able to assess how LR and direct questionnaires perform against individual known truth. Participants were recruited from the participation list from the 2016 round of the Africa Health Research Institute demographic surveillance system, oversampling individuals who were HIV positive. Participants were randomized to two study arms. In Arm A, participants were presented five true/false statements, one of which was the sensitive item, the others non-sensitive. Participants were then asked how many of the five statements they believed were true. In Arm B, participants were asked about each statement individually. LR estimates used data from both arms, while direct estimates were generated from Arm B alone. We compared elicited responses to HIV testing and serostatus data collected through the demographic surveillance system. We enrolled 483 participants, 262 (54%) were randomly assigned to Arm A, and 221 (46%) to Arm B. LR estimated 56% (95% CI: 40 to 72%) of the population to be HIV-negative, compared to 47% (95% CI: 39 to 54%) using direct estimates; the population-estimate of the true value was 32% (95% CI: 28 to 36%). LR estimates yielded HIV test refusal percentages of 55% (95% CI: 37 to 73%) compared to 13% (95% CI: 8 to 17%) by direct estimation, and 15% (95% CI: 12 to 18%) based on observed past behavior. In this context, LR performed poorly when compared to known truth, and did not improve estimates over direct questioning methods when comparing with known truth. These results may reflect difficulties in implementation or comprehension of the LR approach, which is inherently complex. Adjustments to delivery procedures may improve LR's usefulness. Further investigation of the cognitive processes of participants in answering LR surveys is warranted.