Improving Reproducibility by Incorporating Uncertainty

点击翻译按钮获取中文摘要

In recent years, stakeholders in the scientific and lay communities have raised alarms about a lack of reproducibility of scientific results. These stakeholders view the reproducibility crisis as a product of the behavior of researchers and editors. While these behaviors likely have an impact on reproducibility, there are credible reasons why studies of complex data should be expected to arrive at different estimates. First, each study is differentially susceptible to systematic biases, including confounding, selection bias and measurement error. These biases may be large drivers of the appearance of poor reproducibility. Second, many studies that have been criticized for lack of replication are small, and therefore subject to substantial random variability. In combination with selection forces emanating from significance testing, these small studies are likely to overestimate effects, further contributing to the appearance of poor reproducibility. To date, proposed solutions to the perceived reproducibility crisis have largely ignored these contributing factors. We propose to (a) use simulation-based quantitative bias analysis techniques to adjust for the influence of systematic errors on estimates of association and on summaries of an evidence base, and (b) use Bayesian statistical methods to synthesize prior information with estimates of association and summaries of an evidence base to reduce random variability. The premise of the proposed project is that the use of these informatics approaches will reduce the potential for systematic and random error to misleadingly portray research as poorly reproducible and will identify the most important limitations in an evidence base, which will optimize decisions regarding new data collection. The proposed informatics will be extended and applied in the context of two high profile, controversial topic areas with complex data, which will provide examples applicable to other topic areas. For both topic areas, many sources of potential bias have been identified in the surrounding discourse—as have powerful sources of prior information to temper uncertainty—but their influences on individual estimates of association and summaries of the evidence base have not been fully quantified. Quantitative adjustments for these errors using quantitative bias analysis and Bayesian methods—and for publication bias on meta-analytic summaries—would improve reproducibility. We will then extend and apply web-enabled informatics tools to implement the methods for any topic with a set of heterogeneous study results, allowing stakeholders without advanced analytic skills to tailor the underlying assumptions and see for themselves the impact on the summary results. By achieving our aims, this project will advance the use of research informatics to diminish the reproducibility crisis, help to speed consensus-building for any research topic, and productively channel research resources towards resolving the most influential sources of uncertainty in any topic area.