Survival analysis: caveats and pitfalls

Background: Survival analysis in clinical studies is important to assess th e effectiveness of a given treatment and to understand the effect of variou s disease characteristics. A number of methods exist to estimate the surviv al rate and its standard error. However, one cannot be certain that these m ethods have been handled appropriately. The widespread use of computers has made it possible to carry out survival analysis without expert guidance, b ut 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-statisticia n, in addition to obtaining expert referees' comments on the article. The p roblem is compounded in papers from third-world countries where bio-statist icians may not be available in all institutions to guide clinicians as to t he selection of proper techniques. Methods: The present paper deals with the various techniques of survival an alysis 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 surv ival analysis. The median survival time is a better method of summarizing d ata than the mean. Rothman's method of estimation of the confidence limit i s better than Peto's method as the confidence limit for survival probabilit y tends to go beyond the range of 0-1.0 when calculated by Peto's method, e specially when the sample size is small. Conclusion: The results from the present study suggest that survival analys is should be carried out by the Kaplan-Meier method. The median survival ti me should be provided wherever possible, rather than relying on mean surviv al. Confidence limits should be calculated as a measure of variability. A s uitable rank test should be used to compare two or more survival curves, ra ther than a Z-test. Stratified analysis and Cox's model, when stratified an alysis fails, can be used to define the impact of prognostic factors on sur vival.