To assess the risk of skin cancer and other cancers among patients with atopic dermatitis. Register-based retrospective cohort study. Sweden. Patients A total of 15 666 hospitalized patients identified in the national Inpatient Register as having discharge diagnoses of atopic dermatitis between January 1, 1965, and December 31, 1999. Interventions The National Swedish Cancer Register coded malignant neoplasms during the entire period of study. Follow-up time was calculated from the date of entry in the cohort until the occurrence of a first cancer diagnosis, emigration, death, or the end of the observation period, whichever occurred first. Follow-up by means of record linkages to several nationwide registers, among them the National Swedish Cancer Register. Standardized incidence ratios (SIRs) (the ratios of numbers of observed patients with cancer to expected numbers of incident cases of cancer) estimated the risk of developing cancer relative to the risks in the age-, sex-, and calendar year- matched general Swedish population. After excluding the first year of follow-up, the risk of developing any cancer was increased by 13% (95% confidence interval [CI] of SIR, 1.01-1.25, based on 311 observed patients with cancer). Excess risks were observed for cancers of the esophagus (SIR, 3.5; 95% CI, 1.3-7.7; 6 patients), pancreas (SIR, 1.9; 95% CI, 1.0-3.4; 11 patients), brain (SIR, 1.6; 95% CI, 1.1-2.4; 27 patients), and lung (SIR, 2.0; 95% CI, 1.3-2.8; 31 patients) and for lymphoma (SIR, 2.0; 95% CI, 1.4-2.9; 29 patients). There was a nonsignificant 50% excess risk for nonmelanoma skin cancer (SIR, 1.5; 95% CI, 0.8-2.6; 12 patients), seemingly confined to men and to the first 10 years of follow-up. Malignant melanoma did not occur more often than expected. The observed risk elevations, all of borderline statistical significance, should be interpreted cautiously. We could not control for possible confounding by cases of cancer caused by smoking, and the combination of multiple significance testing and few observed patients may have generated chance findings.