Detecting Latency Differences in Event-Related BOLD Responses: Application to Words versus Nonwords and Initial versus Repeated Face Presentations R. N. A. Henson,* , † ,1 C. J. Price,* M. D. Rugg,† R. Turner,* and K. J. Friston* *Wellcome Department of Cognitive Neurology, Institute of Neurology, and †Institute of Cognitive Neuroscience and Department of Psychology, University College London, Queen Square, London WC1N 3BG, United Kingdom Received March 8, 2001 We introduce a new method for detecting differences in the latency of blood oxygenation level-dependent (BOLD) responses to brief events within the context of the General Linear Model. Using a first-order Taylor approximation in terms of the temporal derivative of a canonical hemodynamic response function, statistical parametric maps of differential latencies were esti- mated via the ratio of derivative to canonical parameter estimates. This method was applied to two example datasets: comparison of words versus nonwords in a lex- ical decision task and initial versus repeated presenta- tions of faces in a fame-judgment task. Tests across sub- jects revealed both magnitude and latency differences within several brain regions. This approach offers a computationally efficient means of detecting BOLD la- tency differences over the whole brain. Precise charac- terization of the hemodynamic latency and its interpre- tation in terms of underlying neural differences remain problematic, however. © 2002 Elsevier Science Key Words: event-related; fMRI; word; nonword; faces; BOLD; impulse; latency; deactivations. Several authors have argued that analysis of the latency (as well as the magnitude) of the blood oxygen- ation level-dependent (BOLD) impulse response may prove informative with regard to the neural/synaptic activity following brief stimulation (e.g., Menon et al., 1998; Kruggel and von Cramon, 1999; Miezin et al., 2000). The present work introduces a new whole-brain statistical technique for testing differences in the la- tency of the BOLD impulse response function within the context of the General Linear Model (GLM). We use data from a lexical decision task and a face fame- judgment task to demonstrate the ability of this tech- nique to detect latency differences within brain regions between two classes of stimuli: words versus nonwords and initial versus repeated presentations of famous faces. We rarely know the precise shape of the BOLD im- pulse response for a given brain region, but we can make an informed guess in light of knowledge of the canonical hemodynamic response function (HRF) de- rived from previous studies. To allow for some devia- tions about this canonical form, Friston et al. (1998) added further response functions derived from a first- order multivariate Taylor expansion of the canonical HRF. These functions included the partial derivative with respect to time (temporal derivative) and partial derivative with respect to duration (dispersion deriva- tive). The inclusion of this set of “basis” functions within the General Linear Model allows estimation of the contribution of each basis function (its parameter estimate), linear combination of which allows calcula- tion of the mean and standard error of the best-fitting event-related response. Friston et al. (1998) also pro- posed that tests of differences in the latency of re- sponses can be derived from knowledge of the standard error of the fitted response. The present work takes this proposal further by explicitly estimating response latency (relative to the canonical HRF) via the ratio of temporal derivative to canonical parameter estimates. A preliminary application of this proposal was reported by Henson et al. (1999a), and a related, more general proposal has been made recently by Liao et al. (2001). The canonical response function and its temporal derivative are shown in Fig. 1a. Assume that the real event-related BOLD response, F(t), as a function of poststimulus time, t, is delayed relative to the canoni- cal response, f(t), by a small amount dt, such that F t  = f  t + dt  , where  is a scaling factor. With a first-order Taylor expansion of the canonical response, f  t + dt   f  t  + f ' t  dt , 1 To whom correspondence should be addressed. E-mail: r.henson@ucl.ac.uk. NeuroImage 15, 83–97 (2002) doi:10.1006/nimg.2001.0940, available online at http://www.idealibrary.com on 83 1053-8119/02 $35.00 © 2002 Elsevier Science All rights reserved.