Overinterpretation and misreporting of prognostic factor and biomarker studies in medical oncology.
Kempf E., de Beyer J., Porcher R., Nguyen L., Moons K., Altman D., Reitsma J., Hopewell S., Sauerbrei W., Collins G.
e14023 Background: To generate empirical evidence, classify and estimate the frequency of distorted presentation and overinterpretation of results in prognostic factor (PF) studies in oncology. Methods: We selected 17 oncology journals with an impact factor of 7 or more. PubMed was searched to identify primary clinical studies evaluating one or more PFs published in 2015. Each article was independently evaluated by 2 reviewers using a data extraction form that was extensively pilot-tested. We classified aspects of overinterpretation as misleading reporting, interpretation and extrapolation. Results: Our search identified 10 844 articles, 98 met eligibility criteria. The PF was assessed prospectively in 8 of 56 observational studies, and in 16 of 42 clinical trials. The median number of PFs per study was 2 (Q1-Q3, 1-5). Overall survival was used as the outcome in 77 studies. The median number of statistical analyses reported with regards to the PF effect assessment was 42.5 per study (Q1-Q3, 15.5-86.5). Thirty-three reports reported using 2 or more different multivariable models to assess the PF effect and 21 did not adjust. Misleading reporting included selective and incomplete reporting of the PF effect (n = 26 and n = 8, respectively). The conclusions focused solely on significant results in 80 reports, and in those where there was at least one NS result, 80% of studies focused their conclusions solely on the significant results. Misleading interpretation included not reporting the PF effect within a multivariable model (n = 23 and n = 54 in full-texts and abstracts, respectively). One out of 5 conclusions used linguistic spin with strong statements in both full-text and abstract. Linguistic spin of NS results was found in 28 Results sections in the full-text and in 18 abstract conclusions. The conclusions were inconsistent with the study findings in one out of five articles (both in the full-text and abstract). Discrepancies between the conclusions presented in the full-text and in the abstract were found in 18 reports. Conclusions: Our study provides insight into the level of reporting and overinterpretation of findings that were frequently inconsistent with the results in oncology journals with high impact factors.