The confounding effect of response amplitude on MVPA performance measures

Multi-voxel pattern analysis (MVPA) is proving very powerful in the analysis of fMRI timeseries data, yielding surprising sensitivity, in many different contexts, to the response characteristics of neurons in a given brain region. However, MVPA yields a metric (classification performance) that does not readily lend itself to quantitative comparisons across experimental conditions, brain regions or people. This is because performance is influenced by a number of factors other than the sensitivity of neurons to the experimental manipulation. One such factor that varies widely but has been largely ignored in MVPA studies is the amplitude of the response being decoded. In a noisy system, it is expected that measured classification performance will decline with declining response amplitude, even if the underlying neuronal specificity is constant. We document the relationship between response amplitude and classification performance in the context of orientation decoding in the visual cortex. Flickering sine gratings were presented at each of two orthogonal orientations in a block design (multivariate experiment) or an event-related design (univariate experiment). Response amplitude was manipulated by varying stimulus contrast. Orientation classification performance in retinotopically defined occipital area V1 increased approximately linearly with the logarithm of stimulus contrast. As expected, univariate response amplitude also increased with contrast. Similar results were obtained in V2, V3 and V3A. Plotting classification performance against response amplitude gave a function with a compressive non-linearity that was well fit by a power function. Knowledge of this function potentially allows adjustment of classification performance to take account of the effect of response size, making comparisons across brain areas, categories or people more meaningful.