Selective correlations; not voodoo

Area of Science:

• Neuroimaging
• Statistical Neuroscience
• Brain Imaging Analysis

Background:

• "Voodoo" correlations, or exceptionally high correlations in selected brain regions, are a known issue in neuroimaging.
• This problem arises from estimating quantities of interest from the same data used for selection, a statistical challenge known as selective inference.
• Current remedies like data splitting have drawbacks, including discarding valuable data.

Purpose of the Study:

• To develop a novel statistical method for constructing confidence intervals (CIs) that address selective inference in neuroimaging.
• To ensure good reproducibility prospects by controlling the False Coverage Rate (FCR) even when selection and estimation use the same data.
• To provide a more informative and powerful alternative to data-splitting methods.

Main Methods:

• Adaptation of recent developments in selective inference to create new confidence intervals.
• Control of the expected proportion of non-covered correlations in selected voxels (False Coverage Rate - FCR).
• Development of a "confidence calibration plot" for clear and interpretable reporting of results.

Main Results:

• Proposed confidence intervals control the FCR in realistic social neuroscience simulations.
• The selective intervals attenuate the impression of highly biased observed correlations by extending towards zero.
• The method demonstrated considerable selection bias in a loss-aversion study, highlighting the need for such corrections.
• Selective intervals showed more power and were more informative than data-splitting, as no data was discarded.

Conclusions:

• The proposed selective inference method provides reliable confidence intervals in neuroimaging, mitigating "voodoo" correlations.
• This approach offers improved statistical rigor and reproducibility compared to traditional data-splitting techniques.
• The accompanying software package facilitates the computation and application of these advanced statistical intervals in neuroimaging research.