Functional magnetic resonance imaging (fMRI) has become a fundamental tool for studying brain function. However, the presence of serial correlations in fMRI data complicates data analysis, violates the statistical assumptions of analyses methods, and can lead to incorrect conclusions in fMRI studies.
In this paper, we show that conventional whitening procedures designed for data with longer repetition times (TRs) (>2 s) are inadequate for the increasing use of short-TR fMRI data. Furthermore, we comprehensively investigate the shortcomings of existing whitening methods and introduce an iterative whitening approach named “IDAR” (Iterative Data-adaptive Autoregressive model) to address these shortcomings. IDAR employs high-order autoregressive (AR) models with flexible and data-driven orders, offering the capability to model complex serial correlation structures in both short-TR and long-TR fMRI datasets.
Conventional whitening methods, such as AR(1), ARMA(1,1), and higher-order AR, were effective in reducing serial correlation in long-TR data but were largely ineffective in even reducing serial correlation in short-TR data. In contrast, IDAR significantly outperformed conventional methods in addressing serial correlation, power, and Type-I error for both long-TR and especially short-TR data. However, IDAR could not simultaneously address residual correlations and inflated Type-I error effectively.
This study highlights the urgent need to address the problem of serial correlation in short-TR (< 1 s) fMRI data, which are increasingly used in the field. Although IDAR can address this issue for a wide range of applications and datasets, the complexity of short-TR data necessitates continued exploration and innovative approaches. These efforts are essential to simultaneously reduce serial correlations and control Type-I error rates without compromising analytical power.