[FieldTrip] Fwd: Identify a window of best correlation between ERP and behavioural data

[Stephen Politzer-Ahles]

Hi Harold,

If you're only interested in identifying which timepoints have the best
correlation, multiple comparisons aren't really an issue; if you're not
worried about testing for significance, then the place where you see the
highest estimate is the highest. See e.g., Hauk et al. 2006 in NeuroImage,
and Smith & Kutas 2015 in Psychophysiology, for similar analyses.

If you're interested in finding a part of the waveform where the
correlation is significantly different from zero while also dealing with
multiple comparisons, you could use [spatio]temporal clustering (Maris &
Oostenveld 2007). Just make an event-related regression coefficient (as
described in Hauk et al.; just do a regression at each timepoint, and then
rather than plotting the ERP amplitude at each timepoint, plot the
regression estimate (*b*) at each timepoint, to derive a similar waveform)
and then test that waveform against 0 using temporal clustering.

Best,
Steve




---
Stephen Politzer-Ahles
University of Oxford
Language and Brain Lab
Faculty of Linguistics, Phonetics & Philology
http://users.ox.ac.uk/~cpgl0080/



> Message: 5
> Date: Tue, 26 Jan 2016 00:25:13 +0100
> From: Harold Cavendish <harold.cav89 at gmail.com>
> To: fieldtrip at science.ru.nl
> Subject: [FieldTrip] Fwd: Identify a window of best correlation
>         between ERP and behavioural data
> Message-ID:
>         <
> CALOTjvZXPNWHLefXrG6uss7uF3kkXie7qjiWqEcesOUqWay6qA at mail.gmail.com>
> Content-Type: text/plain; charset="utf-8"
>
> Dear FieldTrip users,
>
> I'm looking for the best method to analyse the relationship between ERP and
> behavioural data but all I've been able to find so far are methods to
> compare ERPs (or more generally time-series).
>
> Consider the following example: 20 participants performed a memory-related
> task in which they had to respond based on the difference between the
> target stimulus and the probe stimulus. EEG (50 electrodes, about 1000
> samples at 200 Hz) and responses were recorded; ERPs were then created
> based on epochs in which the responses were correct. Similarly, response
> accuracy was averaged over all successful epochs per each subject.
>
> The resulting data are therefore: 20 participants x 50 channels x 1000
> samples (ERP) and 20 averages of accuracy (one value per participant).
>
> My objective is to identify which portion of the ERP data best explains (or
> predicts) the changes in accuracy. In other words, in which time interval
> is the most significant difference between people who do well in the task
> and those who don't.
>
> So far I've been doing correlations between ERP at each time point and
> overall accuracy, which is unreliable due to multiple comparisons
> (correction methods are reportedly too conservative). Nevertheless, certain
> channels show quite strong correlations in multiple time intervals.
>
> How would you approach this problem, please?
>
> Many thanks!
>
> Best regards,
> Harold
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL: <
> http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20160126/b5769634/attachment-0001.html
> >
>
> ------------------------------
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20160126/2c0d7efb/attachment-0002.html>