Rxivist logo

A tail of two sides: Artificially doubled false positive rates in neuroimaging due to the sidedness choice with t-tests

By Gang Chen, Robert W. Cox, Daniel R. Glen, Justin K Rajendra, Richard C Reynolds, Paul A. Taylor

Posted 23 May 2018
bioRxiv DOI: 10.1101/328567 (published DOI: 10.1002/hbm.24399)

One-sided t-tests are widely used in neuroimaging data analysis. While such a test may be applicable when investigating specific regions and prior information about directionality is present, we argue here that it is often mis-applied, with severe consequences for false positive rate (FPR) control. Conceptually, a pair of one-sided t-tests conducted in tandem (e.g., to test separately for both positive and negative effects), effectively amounts to a two-sided t-test. However, replacing the two-sided test with a pair of one-sided tests without multiple comparisons correction essentially doubles the intended FPR of statements made about the same study; that is, the actual family-wise error (FWE) of results at the whole brain level would be 10% instead of the 5% intended by the researcher. Therefore, we strongly recommend that, unless otherwise explicitly justified, two-sided t-tests be applied instead of two simultaneous one-sided t-tests.

Download data

  • Downloaded 1,201 times
  • Download rankings, all-time:
    • Site-wide: 22,567
    • In neuroscience: 2,726
  • Year to date:
    • Site-wide: 143,058
  • Since beginning of last month:
    • Site-wide: 81,757

Altmetric data

Downloads over time

Distribution of downloads per paper, site-wide