Combining the tools: Activation- and information-based fMRI analysis
We agree with most of Andreas Kleinschmidt's thoughtful comments and historical remarks (Kleinschmidt, 2007-this isuue). Activation-and information-based techniques contribute complementary tools to the fMRI analysis toolbox (see Table 1 ). They should be used in combination as motivated by neuroscientific inquiry. We should clarify that our somewhat provocative title "Analyzing for information, not activation, to exploit hi-res fMRI" (Kriegeskorte and Bandettini, 2007-this issue) was meant to
... ontrast these fundamentally different concepts, not to suggest that activation-based analysis (Worsley et al., 1992; Friston et al., 1994; Friston et al., 1995) has no place in the toolbox for hi-res fMRI. Hi-res fMRI: a different regime In his title, Kleinschmidt poses the question: "Different analysis solutions for different spatial resolutions?" We think this question merits serious consideration and would like to maintain that the spatial resolution of the measurement (among many other factors) does need to be taken into account in deciding the analysis strategy. Upon initial consideration, it may appear that moving to high resolution constitutes merely a quantitative change, with the same analyses equally applicable, but yielding more fine-grained maps. In fact, hi-res fMRI puts us in an altogether different regime in terms of both the neuroscientific questions to be addressed and the statistical analyses appropriate. Most fundamentally for the neuroscientist, hi-res fMRI promises access to columnar-level information. This motivates shifting the goal of analysis from the localization of activated functional regions to the characterization of their intrinsic representations, i.e. from activation to information. More practically, hi-res fMRI confronts us with the four challenges we describe (Kriegeskorte and Bandettini, 2007-this issue). As Kleinschmidt suggests, these challenges seem familiar from standard-resolution fMRI. But they take on a novel quality in hi-res fMRI because of their greater severity and combined effect. To recapitulate: at high resolution, fMRI patterns may not provide accurate images of neuronal activity patterns (challenge 1); the noise (challenge 2) and the number of voxels (challenge 3) are substantially greater; finally, Talairach or a cortex-based common space cannot accurately relate hi-res voxels between subjects for group analysis (challenge 4). Crucially, the conventional method of dealing with the milder versions of these challenges at standard resolution, i.e. smoothing or local averaging, would defy the purpose of hi-res fMRI: smoothing would decrease the effective resolution. To clarify the core of our argument for the synergy between hi-res fMRI and information-based analysis: 1. Smoothing removes fine-grained pattern information from the data and thus defies the main purpose of hi-res fMRI. 2. Without local combination of single-voxel signals (as provided by smoothing), the four challenges (already substantial at standard resolution) can prove prohibitive in hi-res fMRI. 3. Multivariate statistics summarizing local response-pattern information provide an alternative means of locally combining single-voxel signals without removing the neuroscientifically valuable fine-grained pattern information (as smoothing would). Combining the tools Independently of spatial resolution, we think that the widespread 2-step strategy, consisting in a mapping of the entire imaged volume (top 2 rows of Table 1 ) followed by selective region of interest (ROI) analysis (bottom 3 rows), continues to have great potential. Mapping provides a more exploratory, wider view of the data and can lead to the discovery of new regions involved in a given process. Selective ROI analysis provides a complementary, more hypothesis-driven view of a detail of the functional architecture, focusing statistical power to reveal a given region's functional properties. This 2-step approach can either be based on a single experiment (using independent contrasts for the mapping that defines the ROI and the selective ROI analysis) or on a localizer experiment (for defining the ROI) and a main experiment (to which selective ROI analysis is applied). These two variants are discussed in the exchange between Friston et al. (2006) and Saxe et al. (2006) .