Survival analysis: caveats and pitfalls

Background: Survival analysis in clinical studies is important to assess the effectiveness of a given treatment and to understand the effect of various disease characteristics. A number of methods exist to estimate the survival rate and its standard error. However, one cannot be certain that these methods have been handled appropriately. The widespread use of computers has made it possible to carry out survival analysis without expert guidance, but using inappropriate methods can give rise to erroneous conclusions. The majority of the biomedical journals now recommend that a statistical review of each manuscript should be carried out by an experienced bio-statistician, in addition to obtaining expert referees' comments on the article. The problem is compounded in papers from third-world countries where bio-statisticians may not be available in all institutions to guide clinicians as to the selection of proper techniques. Methods: The present paper deals with the various techniques of survival analysis and their interpretation, using a modal data set of malignant upper-aerodigestive tract melanoma patients treated in the Regional Cancer Centre, Trivandrum since 1982. Results: The Kaplan-Meier method was found to be the most suitable for survival analysis. The median survival time is a better method of summarizing data than the mean. Rothman's method of estimation of the confidence limit is better than Peto's method as the confidence limit for survival probability tends to go beyond the range of 0-1.0 when calculated by Peto's method, especially when the sample size is small. Conclusion: The results from the present study suggest that survival analysis should be carried out by the Kaplan-Meier method. The median survival time should be provided wherever possible, rather than relying on mean survival. Confidence limits should be calculated as a measure of variability. A suitable rank test should be used to compare two or more survival curves, rather than a Z-test. Stratified analysis and Cox's model, when stratified analysis fails, can be used to define the impact of prognostic factors on survival.