The effect of interview length on wave nonresponse in a longitudinal survey was studied, controlling for respondent-specific characteristics known to affect survey response. Data from the National Longitudinal Survey of Youth were used, for a sample of over 10,000 people who were 14 to 22 years old when first interviewed in 1979. These individuals have been interviewed annually every year since, yielding 16 years of data. Over the years, the length of the interview has varied. A transition probability model is estimated using hazard equations. With personal, demographic, and environmental factors known to influence survey response and several measures of respondent attitude and cooperation held constant, longer interview length was associated with sample retention. With the hypothesis that interview length may proxy for some uncontrolled dimension of respondent cooperation, an alternative measure to interview length, the number of questions asked, was constructed. Reestimating the hazards with this variable generates similar findings. It is conjectured that survey length, whether measured in minutes or in the number of questions asked, measures the salience or applicability of the survey to the respondent. Those with characteristics most important to the content of the survey have the longest interviews and are also the most interested. It is proposed that survey instruments be designed with questions applicable to all respondents, focusing less on the average length of the interview and more on the range of potential interview lengths. An appendix presents variable definitions and tables of nonresponse estimation hazard by subgroups. (Contains five tables, seven figures, five appendix tables, and eight references.) (SLD)