fMRI data analysis: Statistical challenges and solutions.

fMRI is a non-invasive method that captures brain activity in a sequence of images while par- ticipants perform one or more tasks. Changes in the blood-oxygen-level-dependent (BOLD) signal between images are used to localize task related areas in the brain. The signal is measured on a series of time points and a test statistic is calculated for each of a large number of volume units (voxels) across the brain using a general linear model. The end product is a statistical parametric map (SPM) which graphically displays the evidence for activation in each voxel using null hypothesis significance testing (NHST). fMRI is an increasingly popular technique in neuroimaging research, but, not unlike other methods, it is faced with certain specific pitfalls. We sketch an overview of these pitfalls and offer solutions where possible. To determine in which voxels the evidence for activation is larger than can be expected by chance, the SPMs are thresholded. Simultaneously performing a large number of statistical tests leads to an explosion of false positives, the so-called multiple testing problem. Therefore, thresholding is typically made very conservative using multiple comparisons corrections. Such conservatism boosts the ability to exclude activation when there is none (specificity) but dramatically reduces the ability or power to detect activation when it is present (sensitivity). We present two approaches to ensure the practical significance of the effects found in fMRI data. First, since NHST does not include testing against a functional relevant alternative, it cannot be implied that statistically significant results bear this property. Concerning functional relevance of results, there is growing awareness of the importance of effect sizes (ESs), because ESs quantify the amplitude of activation of the voxel during the task. ESs are also directly linked with the underlying neurological process of the BOLD signal. One way to ensure functional relevance of the detected brain activation is by including ESs in the hypothesis testing. We show that this offers a more balanced view on results, especially concerning functional relevance. A second approach consists of increasing sample sizes in the analyses. When working retrospec- tively, this approach requires integration of data across studies using meta-analyses. However, when it comes to reporting and retrieving the results of an fMRI study there are two options. The easiest approach is to merely report the location of the activation peaks in a table. This is currently the most prevailing approach. However, this implicates a massive loss of data as we ignore information about more than 99 % of the voxels. A more appropriate approach is to report entire SPM's but this requires a lot of organization and storage space. Due to this, popular meta-analyses only use the reported peak locations (these are called coordinate-based meta-analyses). Recent research has shown that toolboxes such as seed based d-mapping and ALE, when provided with enough studies, are fairly capable of balancing type I and II errors. Importantly, as meta-analyses based on fMRI data are only recently introduced, not all the tools that are associated with classic meta-analyses are yet available. For instance, fMRI studies might suffer from publication bias, where studies that fail to report certain results do not get published. While in classic meta-analysis methods for detecting publication bias have existed for 4 decades, no similar tools are present for the meta-analysis of fMRI studies. We propose two methods to assess publication bias, both based on methods available for a classic meta-analysis. The first one is an adaptation of the Fail-Safe N, which verifies the robustness of the results. It calculates the number of null studies (studies that do not report activation at a certain location) necessary to render the meta-analysis non-significant. The second one is based on the Egger regression test and determines the regions in the brain where the results of the meta-analysis are mainly driven by studies with small sample sizes. In summary, we sketch an overview of ongoing issues in fMRI data analysis, both on single studies and meta-analyses. We propose adaptations to improve these methods, both increasing power and practical significance.

http://ift.tt/2q5UQab