Sensitivity of PPI analysis to differences in noise reduction strategies

In some fields of fMRI data analysis, using correct methods for dealing with noise is crucial for achieving meaningful results. This paper provides a quantitative assessment of the effects of different preprocessing and noise filtering strategies on psychophysiological interactions (PPI) methods for analyzing fMRI data where noise management has not yet been established.

Both real and simulated fMRI data were used to assess these effects. Four regions of interest (ROIs) were chosen for the PPI analysis on the basis of their engagement during two tasks. PPI analysis was performed for 32 different preprocessing and analysis settings, which included data filtering with RETROICOR or no such filtering; different filtering of the ROI "seed" signal with a nuisance data-driven time series; and the involvement of these data-driven time series in the subsequent PPI GLM analysis. The extent of the statistically significant results was quantified at the group level using simple descriptive statistics. Simulated data were generated to assess statistical improvement of different filtering strategies.

We observed that different approaches for dealing with noise in PPI analysis yield differing results in real data. In simulated data, we found RETROICOR, seed signal filtering and the addition of data-driven covariates to the PPI design matrix significantly improves results.

We recommend the use of RETROICOR, and data-driven filtering of the whole data, or alternatively, seed signal filtering with data-driven signals and the addition of data-driven covariates to the PPI design matrix.