Interviewer observations of household characteristics can be useful variables for constructing post-survey nonresponse adjustments, particularly if the observations are good proxies of key variables collected later in the survey. The U.S. Census Bureau and other survey data-collection organizations now systematically design interviewer observations with this purpose in mind. However, because these observations are typically estimates or judgments made by the interviewers, they may be incorrect. The effectiveness of nonresponse adjustments using these observations is hampered by such errors. Thus, knowledge about both household- and interviewer-level factors that affect the accuracy of interviewer observations is important for survey researchers in their efforts to design interviewer observations. To date, only one published study of a face-to-face household survey has attempted to identify these factors. This paper attempts to help fill this gap in knowledge by presenting multilevel models of the accuracy of a housing-unit observation recorded by interviewers in the National Survey of Family Growth (NSFG): whether or not young children under the age of 15 are present in the household. The presence of children in the household is an important correlate of key variables in the NSFG and many other socio-economic surveys, and is related to survey participation in general. The accuracy of this observation was found to be a function of selected household-level factors and interactions between household- and interviewer-level factors. Implications of these results for practice and directions for future research are discussed.