Multi-variate Analysis

Multivariate Pattern Analysis (MVPA) attempts to find patterns of brain activation. It differs from univariate analysis in that it does not summarise voxel activity into regional global activity. Instead it looks at the activity of each voxel within the region in terms of spatial patterns of activity. The assumption behind MVPA is that information is represented by such patterns and that these patterns can be used to predict stimulus features.

Representing regional brain activity as a matrix makes it easily comparable with other representational spaces. For example, we can compare activation matrices of different conditions in a dissimilarity matrix (DSM). However, the disadvantage is that topographical and anatomical information is lost in this DSM.

Another way to compare spatial patterns of the brain is to measure the similarity to stimulus feature spaces or behavioural judgements. As long as everything can be represented as a spatial pattern, they can be compared, which is one main advantage of MVPA.

Apart from comparing representational spaces (Representational Similarity Analysis), MVPA is often done to classify responses. Classification can be used to retrieve the pattern of responses to an experimental condition. With the help of a classifier we can determine regional differences in activation patterns. This means that it can learn how the spatial pattern to images of faces look like and how they look like for houses. After training on a specific training set, testing data is used to see how well the classifier performs. This performance is measured in classifcation accuracy (% of correctly classified images) or the classification frequency (how many time it was wrong and how often it was right). If the performance of the classifier for a particular region is high, this may indicate that this region is particularly effcient at separating faces from houses.

If you want to read more on MVPA check out Haxby et al.,2014.