Background and objective: In the area of oncology and cardiovascular disease, treatments often effect overall survival, progression free survival and other time-to-event outcomes. For such treatments, the evaluation of cost-effectiveness often implies an extrapolation of trial results to periods beyond the trial length. The choice of extrapolation function may have a substantial impact on the mean survival: in some of our projects, the mean survival using the log-normal and log-logistic distribution was more than 1.5 times larger than the mean survival using the weibull distribution. This triggered us to perform an evaluation of methods used for extrapolation in NICE submissions, in order to know which methods were accepted. METHODS: CEAs published between 2004 and 2008 by the NICE Technology Appraisal programme, which included failure-time outcome(s), were systematically reviewed with respect to curve fitting procedures used for extrapolation. RESULTS: In the HTA reports, exponential, weibull, log-logistic or log-normal curves were fitted. The distribution was chosen based on face-value, by comparing it with the Kaplan Meier Curve. The quality of the graphical methods is limited, especially because the three curves often have a comparable fit. In the reports, the proportional hazard assumption was used to compare the treatment arm with the comparator arm, often without assessing the validity of the assumption. CONCLUSION: The choice of methods used to extrapolate survival curves in HTA reports has been inadequately justified, and has under-estimated uncertainty. In our opinion, researchers should: assess the validity of proportional hazards and use different methodology when the assumption is violated; evaluate goodness-of-fit more appropriately. Consider using a generalized distribution, for which the weibull, log-logistic and log-normal are special cases