Re: Use of Archived Specimens in Evaluation of Prognostic and Predictive Biomarkers

Extract

In previous issues of the Journal, two commentaries ( 1 , 2 ) proposed schemes for the evaluation and clinical translation of cancer biomarker research, including assigning evidence weights to individual studies based on design features ( 1 , 2 ). We support these efforts, but we are concerned that the recommendations place undue emphasis on studies nested in clinical trials and penalize evidence from high-quality prospective observational studies without a valid basis.

As Simon et al. ( 2 ) noted, the ideal design for evaluating tumor marker associations is to incorporate biomarker knowledge into the randomization scheme of an experimental study. However, they and others ( 2 , 3 ) have delineated several reasons why such trials may be impractical.

These impracticalities leave us with a reality in which clinical translation of biomarker research must rely on evidence from nonrandomized studies. To improve the inferential process, Simon et al. ( 2 ) drew boundaries that distinguish the quality of evidence contributed by different types of epidemiological study [Table 1 in reference ( 2 )]. We disagree with their distinction between studies nested in clinical trials in which the biomarker was not incorporated into the randomization scheme (their category B) and prospective observational studies (their category C). Because neither design confers the inferential advantage of randomization upon comparisons made between biomarker categories, they can be equally susceptible to systematic bias such as confounding, misclassification, and differential attrition ( 4 ). Therefore, it is not appropriate to assess a uniform categorical penalty against evidence from nonrandomized research nested in a prospective cohort population as opposed to nonrandomized research nested in a trial population. Results under both designs must be carefully assessed for bias in the context of the studied association, with an appreciation that each design is capable of providing equally strong (or weak) evidence for a given hypothesis.