In this study, we investigate misreporting behavior in online surveys based on the field experiments in a large-scale online training program for rural teachers. We link the digitally recorded data with survey responses and integrate randomized controlled trials (RCTs) in the survey design. Noncognitive human capital is measured using both self-reported personality traits and proxies based on observed behaviors. Our results show that the impact of observed individual characteristics varies depending on the nature of the question and survey specifics. Unobserved heterogeneity affects both survey participation and response accuracy, resulting in sample selectivity. Noncognitive human capital inferred from observed behaviors consistently shows important influence on misreporting, while that measured by self-reported personality traits suffers from the same misreporting problem. However, behavior proxy may also capture factors external to survey respondents, and it is important to separate the effect of noncognitive human capital from the external impacts. Additionally, survey design affects misreporting. Therefore, improving the efficiency of survey such as by changing the saliency and optimizing the sequence of questions, can improve survey quality. These findings carry important implications for using survey data and for improving survey data quality.