Detecting Model Mis-specification in fMRI using Scan Statistics

starting at time t, where K(t) is a weighting function. Note that it is possible to use any kernel function here (e.g. Uniform or Gaussian) that sums to 1. Under H0: Γ=0, the statistic Yw follows a normal distribution with mean 0 for all w, t. The window that yields the largest value, gives the strongest evidence of model misfit (See Fig. 2). This value is compared to the maximum that would be obtained if the residuals were iid Normal, as expected under the null hypothesis. Detecting Model Mis-specification in fMRI using Scan Statistics Ji Meng Loh1, Tor D. Wager2 and Martin A. Lindquist1 1Department of Statistics, Columbia University 2Department of Psychology, Columbia University