The reliability of young children's self reports of psychiatric information is a concern of epidemiologists and clinicians alike. This paper explores the determinants of test-retest reliability in a sample of children from the general population using reliability coefficients constructed from a kappa statistic. Age, cognitive ability, and gender are related to consistency of reports in a test-retest paradigm. Controlling for age, cognitive ability and gender, children report more reliably on observable behaviors, and less reliably on questions involving unspecified time, reflections of one's own thoughts, and comparison of themselves with others. The reliability of reports of emotions lies between these two extremes. Surprisingly, sentence length of up to 40 words and psychiatric impairment of the child as measured by the Child Global Assessment Scale did not influence reliability. As might be expected, parents' reports of their children are more reliable than their children's reports.