Background: Interviewers can substantially affect self-reported data. This may be due to random variation in interviewers' ability to put respondents at ease or in how they frame questions. It may also be due to systematic differences such as social distance between interviewer and respondent (e.g., by age, gender, ethnicity) or different perceptions of what interviewers consider socially desirable responses. Exploration of such variation is limited, especially in stigmatized populations.
Methods: We analyzed data from a randomized controlled trial of HIV self-testing amongst 965 female sex workers (FSWs) in Zambian towns. In the trial, 16 interviewers were randomly assigned to respondents. We used hierarchical regression models to examine how interviewers may both affect responses on more and less sensitive topics, and confound associations between key risk factors and HIV self-test use.
Results: Model variance (ICC) at the interviewer level was over 15% for most topics. ICC was lower for socio-demographic and cognitively simple questions, and highest for sexual behaviour, substance use, violence and psychosocial wellbeing questions. Respondents reported significantly lower socioeconomic status and more sex-work related violence to female interviewers. Not accounting for interviewer identity in regressions predicting HIV self-test behaviour led to coefficients moving from non-significant to significant.
Conclusions: We found substantial interviewer-level effects for prevalence and associational outcomes among Zambian FSWs, particularly for sensitive questions. Our findings highlight the importance of careful training and response monitoring to minimize inter-interviewer variation, of considering social distance when selecting interviewers and of evaluating whether interviewers are driving key findings in self-reported data.
Trial registration: clinicaltrials.gov NCT02827240 . Registered 11 July 2016.
Keywords: Female sex workers; Gender; Gender-based violence; Interviewer; Validity; Zambia.