Abstract Reproducibility is generally regarded as a hallmark of scientific validity. It can be undermined by two very different factors, namely inflated false positive rates or inflated false negative rates. Here we investigate the role of the second factor, i.e. the degree to which true effects are not detected reliably. The availability of large public databases and also supercomputing allows us to tackle this problem quantitatively. Specifically, we estimated the reproducibility in task-based fMRI data over different samples randomly drawn from a large cohort of subjects obtained from the Human Connectome Project. We use the full cohort as a standard of reference to approximate true positive effects, and compute the fraction of those effects that was detected reliably using standard software packages at various smaller sample sizes. We found that with standard sample sizes this fraction was less than 25 percent. We conclude that inflated false negative rates are a major factor that undermine reproducibility. We introduce a new statistical inference algorithm based on a novel test statistic and show that it improves reproducibility without inflating false positive rates.

Inflated false negative rates undermine reproducibility in task-based fMRI