Cohort studies, disease registries and other sources of administrative data are increasingly available as sources of information on chronic disease processes. The status of individuals, however, can often only be determined at periodic assessment times and analyses may yield invalid inferences if there is an association between the disease process and the assessment process. Much attention has been directed at methods for investigating and correcting for dependent right-censoring, but relatively little attention has been given to the general study of dependent assessment processes. We consider the use of multi-state models for joint characterization of the disease and observation process and discuss estimability issues under non-independent observation schemes. This framework includes failure time models, recurrent events, and other multi-state processes widely employed in life history analysis. The utility of this framework is illustrated using data from registries of individuals in a rheumatology clinic. The types of auxiliary information that enable checking of typical independence assumptions are also discussed, along with the value of tracing studies.