Bootstrapping the P300 ERP
In the field of deception detection, we are expected to classify an
individual as guilty/knowledgeable or innocent/unknowledgeable, and
hence analyze the data within-subjects (specifically, within anindividual ). This analysis would use single trial EEG data, which
is often noisy and variable, leading to unreliable classification
decisions. Therefore, in these situations it is especially useful to use
simulation techniques like bootstrapping to aid intraindividual
diagnosis (Wasserman & Bockenholt, 1989). The bootstrapping process
involves selecting a set of single sweep ERP data points (with
replacement) and averaging them to produce a single ERP amplitude per
individual. Since the resampling of data points is done with
replacement, the averages produced in each run will be slightly
different from other runs. Within a subject, we can bootstrap with
replacement a sample of probe trials and calculate the average, and
similarly produce an average using irrelevant trials. Then, we compare
the difference between resampled probes and irrelevant averages to
estimate the CIT effect. In iterating this process n times, one can
estimate out of all iterations how many times the probe P300 average
exceeded the irrelevant P300 average, enabling investigators to create a
diagnostic index for classifying the subject as guilty/knowledgeable or
innocent/unknowledgeable. This diagnostic metric is known as the
bootstrap iteration score (BSITER) (e.g., Rosenfeld et al., 2015;
Rosenfeld & Donchin, 2015), and has been published in
psychophysiological experiments for several decades.