Assessing the effect of toxicity on prognosis: methods of analysis and interpretation.

There has recently been an increased interest in and reporting of the association between toxicity and treatment outcome in cancer clinical trials. Such comparisons may be used to suggest mechanisms by which cytotoxic agents function in the cancer patient, especially regarding the importance of the effect of dose on the individual patient. However, analyses of "time-to-failure" outcomes such as survival by toxicity are subject to bias due to the time-dependence of both the predictor and outcome variables. In addition, interpretation of even appropriately conducted statistical analyses is problematic, as with analyses of survival by outcome variables such as response. The use of statistical methodology designed to avoid such biases in these comparisons is shown and the problems in the interpretation of results are discussed.