Department of Biostatistics


Prospective Students



News & Events



Employment Opportunities

Resource Quicklinks

Computing Environment




Statistical Issues in Functional Magnetic Resonance Imaging

Colin McCulloch, Assistant Professor of Biostatistics

These are good times for statisticians in functional image analysis! The data is messy, the application is trendy, and the results are important. In this seminar, I will talk about the beginning of my journey into functional imaging. In one type of fMRI experiment, called a "box-car" design, a subject is scanned for several minutes during which time, half the time she performs some task (eg. looks at a black and white checkerboard on a monitor) and the other half she performs some other task (eg. looks at a black screen). The tasks are generally alternated for the duration of the experiment. Throughout the experiment high-resolution, but noisy, images representing blood flow in the brain are acquired and the goal is to quantify differences in these images taken under the stimulus state (checkerboard screen) and the control (black screen) state. Current analysis techniques generally employ a simple t-test comparing stimulus image intensities versus control intensities at every pixel in the image. However, correcting for multiple comparisons is difficult since each image contains around 100,000 pixels. Bonferroni won't help us here! The state of the art in fMRI analysis is to manually choose an effect size threshold that makes the effect images look "right"! Other non-parametric analysis methods have been devised, but the effect size threshold problem persists. I will discuss two Bayesian analyses I have performed, one on a box-car experiment and the other on a so-called single-trial experiment. Bayes seems to be the way to go with these data for several reasons I will discuss, one of which is that it effectively nullifies the multiple-comparison problem tormenting the field.

Return to Longitudinal/Survival Working Group List | Return to Home Page