CFE 2020: Start Registration
View Submission - CMStatistics
Title: Improving the replicability of spatial extent inference via effect size thresholding Authors:  Simon Vandekar - Vanderbilt University (United States) [presenting]
Abstract: The classical approach for testing statistical images using spatial extent inference (SEI) thresholds the statistical image based on a probability threshold (the p-value). This approach has an unfortunate consequence on the replicability of neuroimaging because the target set of the image is affected by the sample size -- larger studies have more power to detect smaller effects. Here, we use the preprocessed (ABIDE) data set, interactive visualizations, and a fully reproducible analysis pipeline to argue for thresholding statistical images by effect sizes instead of probability values. Using a constant effect size threshold means that the p-value threshold naturally scales with the sample size to ensure that the target set is similar across repetitions of the study that use different sample sizes. Because the statistical threshold depends on the sample size, robust inference procedures must be used to ensure that the procedure maintains accurate error rates at an arbitrary p-value cluster forming threshold; for this reason, SEI via Gaussian Random Field approximations is not a valid inference procedure. Future work may investigate how effect size thresholding affects SEI power in small sample sizes and meta-analytic results.