next up previous
Next: Variance Components Up: tr04ss1 Previous: tr04ss1

Introduction

The magnitude of the BOLD (Blood Oxygenation Level Dependent) effect in FMRI, a marker of neuronal activation, is often only of similar magnitude to the noise present in the measured signal. In order to increase power and to allow conclusions to be made about subject populations, it is common practice to combine data from multiple subjects. It is also common to take multiple sessions from each subject, again to increase sensitivity to activation, or for other experimental design reasons, such as tracking changes in function over time. Therefore it is important that inter-session variability present in FMRI data be understood, and in response, McGonigle et al. [18] presented an in-depth study of this issue.

In designing both multi-subject and single-subject multi-session studies, it is critical for the experimenter to have some idea of the relative sizes of within-session variance and inter-session variance. For example, if inter-session variance is large, it could be difficult to detect longitudinal experimental effects (e.g., in studies of learning [21] and post-stroke recovery [15]). If FMRI is to be used in pre-surgical mapping (e.g., [9]), which, by its nature will involve only a single subject, correct interpretation will be dependent on an appreciation of the potential uncertainty due simply to a session effect. In multi-subject studies, it is advantageous to have some idea of the expected inter-session variance, as this will contribute to the observed inter-subject variance.

In order to investigate how well a single session dataset from a single subject typified the subject's responses across multiple sessions, McGonigle and colleagues [18] carried out the same FMRI protocol on 33 separate days; on each day 3 paradigms were run (visual, motor and cognitive), and the variation in ``activation'' was studied. This paper drew three main conclusions: i) the use of ``voxel-counting'' on thresholded statistical maps was not an ideal way to examine reproducibility in FMRI; ii) a ``reasonably large'' number of repeated sessions was essential to properly estimate inter-session variability, and iii) the results of a single session on a single subject should be treated with care if nothing was known about inter-session variability.

While [18] noted the presence of between-session variability in their experiment, they did not attempt to systematically assess the causes of this variance. There are a number of potential contributors; physiological variance (subject), acquisition variance (scanner), and also differences in analysis methodology and implementation. As noted in the original paper, ``it is possible that spatial preprocessing (for example) may affect inter-session variance quite independently of underlying physical or physiological variability''. This view is supported by [19], where analysis methodology is shown to affect apparent intersession variance. Here we revisit the analysis of McGonigle's data and consider session variability in the light of the effects that different first-level processing methods can have.

Furthermore, some readers (e.g., [6,4]) have taken from [18] the simple broad-brush conclusion that there was a ``large amount of session variability''. One of the purposes of this paper is to address this misconception; for example, Section 4.3 shows that in fact inter-session variability was of similar magnitude to within-session variability in this dataset.

We start with a brief theoretical overview of the components of variance present in multiple-session data. We then describe the original data and analysis, as well as the new analyses carried out for this study, with explanation of the measures used in this study to assess session variability. We then present the variability results as found from this data, centring around the use of mixed-effects Z values in relevant voxels as the primary measure of interest. We also show qualitatively why it is dangerous to judge variability through the use of thresholded single-session images.


next up previous
Next: Variance Components Up: tr04ss1 Previous: tr04ss1