Interviewer biases in medical survey data: The example of blood pressure measurements

• Published in: PNAS nexus Vol. 3; no. 3; p. pgae109
• Main Authors: Geldsetzer, Pascal, Chang, Andrew Young, Meijer, Erik, Sudharsanan, Nikkil, Charu, Vivek, Kramlinger, Peter, Haarburger, Richard
• Format: Journal Article
• Language: English
• Published: England Oxford University Press 01.03.2024
• Subjects: Analysis; Blood pressure; Case studies; Hypertension; Measurement; Surveys; World health; Germany; South Africa; California; Indonesia; India; blood pressure; interviewer effects; measurement error; health survey; hypertension
• ISSN: 2752-6542; 2752-6542
• DOI: 10.1093/pnasnexus/pgae109

Health agencies rely upon survey-based physical measures to estimate the prevalence of key global health indicators such as hypertension. Such measures are usually collected by nonhealthcare worker personnel and are potentially subject to measurement error due to variations in interviewer technique and setting, termed "interviewer effects." In the context of physical measurements, particularly in low- and middle-income countries, interviewer-induced biases have not yet been examined. Using blood pressure as a case study, we aimed to determine the relative contribution of interviewer effects on the total variance of blood pressure measurements in three large nationally representative health surveys from the Global South. We utilized 169,681 observations between 2008 and 2019 from three health surveys (Indonesia Family Life Survey, National Income Dynamics Study of South Africa, and Longitudinal Aging Study in India). In a linear mixed model, we modeled systolic blood pressure as a continuous dependent variable and interviewer effects as random effects alongside individual factors as covariates. To quantify the interviewer effect-induced uncertainty in hypertension prevalence, we utilized a bootstrap approach comparing subsamples of observed blood pressure measurements to their adjusted counterparts. Our analysis revealed that the proportion of variation contributed by interviewers to blood pressure measurements was statistically significant but small: % depending on the cohort. Thus, hypertension prevalence estimates were not substantially impacted at national scales. However, individual extreme interviewers could account for measurement divergences as high as 12%. Thus, highly biased interviewers could have important impacts on hypertension estimates at the subdistrict level.