Voodoo-corrected effect sizes at local maxima using Selective Inference

[BrainStatsSam]

Title
Voodoo-corrected effect sizes at local maxima using selective inference

Presenter and Affiliation
Samuel J. Davenport, University of Oxford (Big Data Institute) (@BrainStatsSam)

Collaborators
Thomas E. Nichols, University of Oxford (Big Data Institute) (@nicholst)

Github Link:
https://github.com/BrainStatsSam/SIbootstrap

Website:
https://sjdavenport.github.io

Abstract (max. 200 words):
The voodoo correlation or double-dipping problem is fundamental to brain imaging: using the same data to identify voxels as significant and to measure effect size (at those voxels) leads to biased estimates [1].

If this problem is addressed at all, the typical solution is data-splitting, using the first half of the data to find significant regions and the second half of the data to calculate the effect sizes. This is unbiased, however the effect is detected with less power (and spatial accuracy) relative to no splitting and the estimates are more variable.

Our solution to this problem [2] is to employ resampling methods to correct for the bias allowing for accurate estimates of peak Cohen's d and R². These estimates are essential for use in power analyses. This approach allows for accurate inference on the location of significant effects, as it uses all (rather than half) of the subjects to estimate the locations of significant voxels.

In this demonstration, we will present the SIbootstrap package. We will implement it on fMRI data to illustrate how it can be used in practice to provide estimates of the effect and to calculate power.

Preferred Session
Lightning Talks - 1. Neuroscience toolkit
Or
Oral sessions and demos - 3. Demo: New advances in open neuroimaging methods

Additional Context
[1] E Vul, C Harris, P Winkielman, and H Pashler. Puzzlingly high correlations in fMRI
studies of emotion, personality, and social cognition. Perspectives on Psychological
Science, 4(3):274–290, 2009.
[2] Samuel J. Davenport and Thomas E. Nichols. Selective peak inference: Unbiased estima-
tion of raw and standardized effect size at local maxima. bioRxiv, 2019. doi: 10.1101/
500512. URL https://www.biorxiv.org/content/early/2019/05/02/500512.

Tag other Attendees
@AlexBowring, @TomMaullin, @asoroosh

[petyakindalova]

Sounds great! I would attend for sure!

[asoroosh]

Sounds great, I'd love to be there too.

[TomMaullin]

Nice - sounds like a pretty interesting talk! Will be there!

[habib61]

Interesting, I'd love to hear the talk

[marcopalma3]

It sounds interesting! I am looking forward to hear the talk!

[TimVanMourik]

Hi @BrainStatsSam and others, I’m happy to tell you that we’d like to host your lightning talk in the OSR in the neuroscience toolkit session. This will be a talk of 5 minutes + 5 minutes of questions. We’ll update the program in the ReadMe.md shortly. We’d much appreciate it if you could submit presentation material to the presentations folder by means of a Pull Request to this repository, preferably but not necessarily before the presentation.