Removal of Confounding Effects of Global Signal in Functional MRI Analyses 1 Adrien E. Desjardins, Kent A. Kiehl, and Peter F. Liddle Neuroimaging Laboratory, Department of Psychiatry, University of British Columbia, Vancouver, British Columbia, Canada V6T 2A1 Received June 13, 2000 Local signals obtained from BOLD fMRI are gener- ally confounded by global effects. In this paper, we make an essential distinction between global effects and the global signal. Global effects have a similar influence on local signals from a large proportion of cerebral voxels. They may reflect diffuse physiological processes or variations in scanner sensitivity and are difficult to measure directly. Global effects are often estimated from the global signal, which is the spatial average of local signals from all cerebral voxels. If the global signal is strongly correlated with experimental manipulations, meaningfully different results may be obtained whether or not global effects are modeled (G. K. Aguirre et al., 1998, NeuroImage, 8, 302–306). In particular, if local BOLD signals make a significant contribution to the global signal, analyses using ANCOVA or proportional scaling models may yield ar- tifactual deactivations. In this paper, we present a modification to the proportional scaling model that accounts for the contribution of local BOLD signals to the global signal. An event-related oddball stimulus paradigm and a block design working memory task were used to illustrate the efficacy of our model. © 2001 Academic Press INTRODUCTION Functional magnetic resonance imaging (fMRI) is used to identify local hemodynamic responses invoked by experimentally controlled stimuli. Typically, in the case of blood oxygen level-dependent (BOLD) fMRI, a large cerebral volume is imaged, and statistical anal- yses are performed on signals from subcomponents of this volume to test hypotheses regarding changes in neural activity. In this paper, we assume that analyses are performed with the general linear model (Friston et al., 1995b,c; Worsley and Friston, 1995). A variety of global effects can interfere with the detection of local BOLD signals. Their origins include underlying physiological processes, gross body move- ments, physiological movements (pulsations, swallow- ing, abdominal movements, breathing), and long-term instabilities of the scanner baseline (Friston, 1996; Kruggel, 1999). In an ANCOVA model, global effects are treated as additive effects that do not affect local BOLD signals or error variances. In a proportional scaling model, they are treated as gain effects that equally affect all local BOLD signals and error vari- ances. An assumption contained in both models is that global effects are not correlated with the covariates of interest. Global effects are often estimated from the global signal, the spatial average of local signals from all cerebral voxels. Note that the terms global effects and global signal are sometimes used interchangeably in the literature; the distinction is crucial for the pur- poses of this paper. If the global signal is strongly correlated with exper- imental manipulations, meaningfully different results may be obtained whether or not global effects are mod- eled (Aguirre et al., 1998). Similar problems have been encountered in analyses of positron emission tomogra- phy (PET) activation images (Andersson, 1997). Strong correlations between the global signal and experimen- tal manipulations can reflect the contribution of local BOLD signals to the global signal. If this contribution is significant, it is likely that the global signal is also strongly correlated with the covariates of interest. In the context of PET, Andersson (1997) proposed using a modified global signal calculated only from voxels that are weakly correlated with the covariates of interest. This method may not be appropriate for fMRI given the greater sensitivity of fMRI compared with PET. In this paper, we present a modification to the pro- portional scaling model that accounts for the contribu- tion of local BOLD signals to the global signal. In our model, an adjusted global signal replaces the global signal. The adjusted global signal is calculated by or- 1 The code for implementing adjusted proportional scaling in SPM 99 analyses is available on request from Peter F. Liddle, 2255 Wes- brook Mall, Department of Psychiatry, University of British Colum- bia, Vancouver, BC, Canada, V6T 2A1. Fax: (604) 822-7756. E-mail: liddle@interchange.ubc.ca. NeuroImage 13, 751–758 (2001) doi:10.1006/nimg.2000.0719, available online at http://www.idealibrary.com on 751 1053-8119/01 $35.00 Copyright © 2001 by Academic Press All rights of reproduction in any form reserved.