Poster F99, Tuesday, March 28, 8:00 – 10:00 am, Pacific Concourse
Minimizing researcher bias and improving statistical power in the analysis of Event-Related Potentials with condition inference random forests (cForest)
Francesco Usai1, Antoine Tremblay1,2, Kiera O'Neil1, Aaron J. Newman1; 1Dalhousie University, 2Saint Mary's University
In event-related potential (ERP) data analysis, two critical choices concern how to cut up the time and space (electrodes) continua into windows and regions of interest for analysis. Choosing time windows and electrodes of interest either a priori or “based on visual inspection” risks a lack of sensitivity to true effects, and/or false positives owing to researcher bias. Once possible solution to this problem is conditional inference random forests (cForest), a technique that considers the entire sensor topography through time as a function of other variables, in a single model. We applied cForest to analyse ERPs from a language task, and compared results with those of linear mixed-effects (LME) modelling. ERPs were recorded from 12 people prior to, immediately after, and 2 weeks after they learned vocabulary in a new language, in response to spoken words that matched or did not match a preceding picture. In both post-training sessions LME on mean amplitude from 300-500 ms revealed an expected N400 mismatch effect over central-parietal electrodes (with time window and electrodes selected a priori). Without any prespecification of time or electrodes of interest, cForest identified the same effects. Moreover, cForest revealed changes in the scalp topography of the N400 that varied over time, and with learning success. These results indicate that cForest can both replicates effects found via typical ERP analysis, and also reveal features of the data that other methods may not be sensitive to, in a way that eliminates researcher bias in selecting time windows or electrodes of interest.
Topic Area: METHODS: Neuroimaging