Confidence Intervals and Survival Estimates

Abstract

Kaplan-Meier methods generate an estimate of survival when follow-up is less than complete. Confidence intervals (CI) provide a range of values that is likely to contain the true result. We hypothesized that reporting of CI around survival estimates is not uniform. We searched all articles in randomly selected issues of Journal of Clinical Oncology (JCO), Cancer, and International Journal of Radiation Oncology Biology, and Physics (IJROBP) years 1999, 2001, and 2004. Articles reporting overall or median survival were reviewed to determine if CI were reported in text or graphical form. Data regarding disease site, intervention, study size, and type of study were collected. A total of 313 articles reported overall or median survival from 35 issues of the 3 journals (IJROBP 115, JCO 132, Cancer 66). CI were reported in 97/313 (31.0%) articles reviewed. CI reporting was more common in prospective clinical trials (38%) than retrospective reviews (25%), P = 0.018. If patients were treated with chemotherapy the article was more likely to report CI (no chemo 20% versus chemo 40%; P = 0.001). CI reporting varied with journal (IJROBP 20%, Cancer 29%, JCO 42%, P = 0.001). CI reporting increased according to year of publication (25% in 1999, 27% in 2001, and 41% in 2004, P = 0.03). A majority of studies in major oncology journals do not report CI with survival estimates. The likelihood of reporting CI is associated with study type (prospective versus retrospective), use of chemotherapy, journal, and year of publication.