Abstract
INTRODUCTION: In low- and middle-income countries, self-reported data on chronic cardiometabolic conditions such as high blood pressure and diabetes are commonly used in large-scale epidemiologic studies because implementing objective measures is challenging in these contexts. However, existing evidence suggests that the sensitivity of such measures may be low, and performance may differ by factors such as age, education, or income. We sought to confirm these prior findings and assess bias due to the use of self-reported data in hypothetical epidemiologic studies considering high blood pressure and diabetes as exposures, outcomes, and confounders. METHODS: We used data from the Longitudinal Aging Study in India (analytic N = 55,392) to assess the performance of self-reported data on high blood pressure and diabetes compared with objective measures, overall and stratified by basic demographic factors. We then compared regression coefficients from models considering self-reported and objective high blood pressure and diabetes as exposures, outcomes, and confounders. In all models, we examined whether the mode of data collection (self-report or objective) for other key variables in the model affected results. RESULTS: The overall sensitivity of self-reported high blood pressure and diabetes was 0.514 and 0.570, respectively; specificity for the two conditions was 0.922 and 0.984. Sensitivity of both conditions increased with age, and was higher among women, those in urban settings, and those with higher educational attainment. Across almost all models considering high blood pressure and diabetes as either exposures or outcomes anti-conservative bias was observed when using self-reported vs. objective measures, regardless of the mode of data collection for other key variables. When high blood pressure and diabetes were considered as confounders, differences between using self-report and objective measures were minimal. DISCUSSION: Anti-conservative bias due to the use of self-reported measures of chronic cardiometabolic conditions in surveys conducted in low- and middle-income contexts may be common. Future studies may seek to quantify the magnitude of anticipated bias in existing data resources and use quantitative bias analysis to formally estimate the potential implications of misclassification.