There is very little that is known about breakoffs as a form of nonresponse. Without understanding the mechanisms generating them, breakoffs tend to be treated together with unitnonresponse. Understanding the unique and the common causes of breakoffs would allow for reduction of breakoffs through survey design, as well as for better informed imputation and weighting adjustment models. Studying breakoffs could also provide insight into factors affecting cooperation decisions in a manner that could not be done before – in mail surveys, for example, only a priori hypotheses have been possible due to a single binary outcome being observed. A framework for breakoff and associated response behaviors in web surveys is proposed. The framework incorporates respondent and survey related factors, the multiple response decisions that are possible in web surveys, and presents survey breakoff as a participation decision that is continuously reevaluated throughout the survey. An empirical test was conducted on two similar large scale web survey experiments with 6,000 completed surveys and a 12% overall breakoff rate. A discrete hazard survival model with page-varying and page-invariant covariates was used to estimate the risk of breakoff. The size of the task indicated by the number of questions in a page, comprehension (long questions), mapping (open-end response format), and topic & commitment (introductory screens) led to higher risk of breakoff. There was some indication that questions requiring more judgment processes to cause breakoffs, and no effect was found of retrieval-intensive and editing-intensive (sensitive or threatening) questions. The effect of respondent education was consistent with a hypothesis of cognitive ability. Contrary to previous beliefs, those who break off tend to be more careful respondents who may be finding various tasks too difficult, rather than being careless respondents who haphazardly break off.