Classifying Mouse Movements and Providing Help in Web Surveys
Conrad, Frederick G
MetadataShow full item record
Survey administrators go to great lengths to make sure survey questions are easy to understand for a broad range of respondents. Despite these efforts, respondents do not always understand what the questions ask of them. In interviewer-administrated surveys, interviewers can pick up on cues from the respondent that suggest they do not understand or know how to answer the question and can provide assistance as their training allows. However, due to the high costs of interviewer administration, many surveys are moving towards other survey modes (at least for some respondents) that do not include costly interviewers, and with that a valuable source for clarification is gone. In Web surveys, researchers have experimented with providing real-time assistance to respondents who take a long time to answer a question. Help provided in such a fashion has resulted in increased accuracy, but some respondents do not like the imposition of unsolicited help. There may be alternative ways to provide help that can refine or overcome the limitations to using response times. This dissertation is organized into three separate studies that each use a set of independently collected data to identify a set of indicators survey administrators can use to determine when a respondent is having difficulty answering a question and proposes alternative ways of providing real-time assistance that increase accuracy as well as user satisfaction. The first study identifies nine movements that respondents make with the mouse cursor while answering survey questions and hypothesizes, using exploratory analyses, which movements are related to difficulty. The second study confirms use of these movements and uses hierarchical modeling to identify four movements which are the most predictive. The third study tests three different of providing unsolicited help to respondents: text box, audio recording, and chat. Accuracy and respondent satisfaction are evaluated for each mode. There were no differences in accuracy across the three modes, but participants reported a preference for receiving help in a standard text box. These findings allow survey designers to identify difficult questions on a larger scale than previously possible and to increase accuracy by providing real-time assistance while maintaining respondent satisfaction.