A population-based study was conducted in an attempt to predict which child's epilepsy will remit. Use of data from a regional electroencephalography laboratory allowed identification of all children in Nova Scotia with epilepsy onset from 1977 through 1985 (excluding those with absence and "minor motor" seizures). Children were followed for an average of 7 years. On the basis of clinical characteristics, a multivariate analysis was used to develop a scoring scheme to predict remission (defined as off medication at the end of the follow-up period). Survival curve methods were used to estimate the duration of medication treatment for those with remission. Of the 504 eligible patients, approximately 70% became seizure free long enough to discontinue medication. Approximately 70% of those stopping medication a first time remained seizure free. At the end of follow-up, 55% of the total cohort were in remission. At diagnosis, the best predictors of remission were age < 12 years at onset, normal intelligence, no prior neonatal seizures, and fewer than 21 seizures before treatment. If predicted to have a remission, then, on the basis of survival curve analysis, 80% were without medication 100 months after diagnosis. After 12 months of treatment, prediction was enhanced by including a score for the number of seizures between 6 and 12 months on treatment. We conclude that approximately 55% of childhood epilepsy will remit. Our scoring system predicts reasonably accurately who will have a remission and when medication is likely to be discontinued.