Permutation based methods for comparing quality of life between observed treatments

Abstract

Quality of life is becoming an important outcome for the comparison of aggressive therapies. To measure quality of life (QOL), questionnaires have been designed that ask patients about symptoms and functionality in several aspects of daily life. Primary analyses of such questionnaires typically focus on a summary statistic, such as a sum score or a single global question. This avoids inflated type I errors or loss of power due to multiple testing of individual items. In return, specific questions and answers that initially mattered to the patient may unfortunately get buried. To avoid reduced specificity and interpretability for both patients and physicians, we propose to also analyse all original questions. In this paper, we seek to detect items of the QOL questionnaire that differ significantly over observed treatments even in the face of multiple testing. We sequentially build a model that combines features which additionally discriminate between treatments. To achieve this, we draw on insights gained in the field of statistical genetics where one is often confronted with a vast amount of predictors, e.g. of a genotypic nature. Specifically, we adopt a permutation based approach to evaluate the null distribution of the maximum of many correlated test statistics and use it to build a regression model that explains QOL differences between treatment arms. We apply the new methodology to analyse QOL data in an observational study of four different treatments of breast cancer. We discover that a single question captures most of the observed treatment differences in this population. Copyright © 2005 John Wiley & Sons, Ltd.