OBJECTIVES: When data from survival analysis are summarized in meta analysis, they are usually based on the number of events at the end of the study. This will bias the estimates of differences in survival unless the relative hazards are relatively constant (the proportional hazards assumption). The aim of this study was to explore this assumption by estimating the frequency of convergences and crossings of survival curves. METHODS: We reviewed all publications in Annals of Internal Medicine, British Medical Journal, JAMA, New England Journal of Medicine (NEJM) and The Lancet for 2007 and identified studies that included survival graphs. We extracted the following data from included studies: type of disease, type of exposure, sample size and number of events, maximum follow-up time, number and timing of survival curve convergences and crossings, and whether Cox regression and log-rank tests had been performed. RESULTS: Among 175 included studies, 35% had survival curve convergences and 47% crossings. 38% of the crossings occurred later than halfway through the study (40% for convergences). The proportion of crossings by type of disease was 46% for cardiovascular disease, 38% for cancer and 53% for other diseases. Among studies with survival curve crossings, Cox regression was performed in 66% and logrank-test in 70% of the studies. Only 31% of all the studies reported testing for proportional hazards when Cox regression had been employed. CONCLUSIONS: Survival curve convergences and crossings are common in medical research. Effectiveness estimates based on end of study results will likely be biased unless convergences and crossings are accounted for, and this bias will carry over to meta analyses of individual studies. Researchers frequently employ Cox modeling when the proportional hazard assumption is not met or use log rank tests when other test would be more appropriate.