Various forms of nonsampling error contributed to the total coverage error in the 1990 Decennial Census. The census relies largely on self response as a means of collecting critical data to determine the count of the population. Interviewers conduct personal visits to complete the basic enumeration. Fundamental to the success of these tasks is the design of the questionnaire, in particular the roster of household members. In the census, the rostering of household members is governed by a complicated set of rules and instructions which are provided for respondents on the self-administered census questionnaire. Coverage errors can result if respondents or interviewers ignore, do not understand, or misapply the census rules of residence, or if the census rules do not fit actual living situations. This paper examines what we know about respondents and interviewers as sources of errors of omission or erroneous inclusion of persons in the census. We examine the cognitive requirements posed by the task of compiling a household roster, and consider the implications of the task for the design of the questionnaire. In planning for the 2000 Decennial Census, research is underway to attempt to reduce coverage error through the development of improved methods and strategies for obtaining household rosters. This paper presents some results from past and current research, and assesses the potential role that questionnaire design can have in reducing census coverage error.