The Need for Statistical Evaluation of Surrogates
The unfortunate precedent of anti-arrhythmic drugs made it plain that a statistical evaluation of potential surrogates was required before they were used in clinical research. The seminal paper of Ross Prentice sparked interest in using statistical methods to validate potential surrogates (Prentice, 1989). Prentice suggested definitions and criteria for the validation of surrogates (see details in Section 3.2). At the surge of the AIDS epidemic, the impressive early therapeutic results obtained with zidovudine, and the pressure for accelerated approval of new antiviral therapies, led to the use of CD4+ T-lymphocyte counts as a surrogate endpoint for time to clinical events and overall survival (Lagakos and Hoth, 1992). Yet concerns were expressed about the limitations of CD4+ counts as a reliable predictor for clinical endpoints. Modeling of the effect of treatment and CD4+ counts on survival suggested that CD4+ counts did not fulfill the Prentice criteria (Lin et al., 1993; DeGruttola and Tu, 1994). CD4+ counts were also found to be an incomplete surrogate for clinical progression in individuals with asymptomatic HIV infection taking zidovudine (Choi et al., 1993). These analyses led to a great deal of skepticism about the evaluation and use of surrogate endpoints in general (DeGruttola et al., 1997). However, all these assessments were done using data from single trials, an approach that has serious intrinsic limitations (Molenberghs et al., 2002;
Alonso et al., 2004b, 2006). Shortly after the results of several trials carried out by the AIDS Clinical Trials Group were analyzed simultaneously and a meta-analytic approach was proposed to inform the evaluation of CD4+ as a surrogate endpoint (Daniels and Hughes, 1997).
The present book makes a clear distinction between surrogacy analyses in a single trial and those that require a meta-analysis of multiple trials. When a single trial is available, only individual-level surrogacy can be investigated, unless the trial is very large and can be broken down in smaller units such as countries, regions, or investigational sites. When a meta-analysis of several trials is available, a full evaluation of individual-level as well as trial-level surrogacy can be investigated (Gail et al., 2000). The basic concepts for such analyses have been covered in detail in a previous book (Burzykowski, Molen- berghs, and Buyse, 2005). Our purpose here is three-fold:
- 1. to build on the approach published previously to present more recent methodological developments,
- 2. to describe software available in SAS and R to implement the methods presented, and
- 3. to illustrate use of these methods on actual datasets that are available for readers to perform their own analyses.
This book will hopefully contribute to making the methods for surrogate evaluation readily available to anyone familiar with the SAS or R environment.