Biomedical Engineering Reference
In-Depth Information
1. Shuffling spike trains in time or in trials.
2. Applying functional interaction analysis of interest to shuffled
data.
3. Repeat steps 1 and 2 as many times as computationally feasi-
ble, and obtain a distribution of test statistics from functional
interaction analysis. Ten or one-thousand iterations are prefer-
able for defining probability distribution; however, even one
hundred iterations can establish significance at p
<
0. 05.
4. Establish a significance threshold by determining what val-
ues can be expected by chance at a probability of less than
p
0. 05 (1 in 20; more stringent thresholds can be used as
needed).
For instance, if one is interested in correlations in time, one
should compare test statistics derived from correlations to test
statistics derived from time-shuffled data (
Fig. 7.2A)
.Time
shuffling is appropriate to considering spike trains, which are a
series of spikes in time recorded by a data acquisition system cor-
responding to the timing of action potentials. To time-shuffle
a spike train, one can simply generate a random series of spikes
matched to the length of the spike train of interest. For instance,
to generate a spike train 10 s long at 10 Hz:
<
randSpiketrain = sort(rand(1,100)
*
10); % random
timestamps,10s @ 10 Hz
More sophisticated temporal distributions (Poisson, bursting)
can be generated by providing structure to the random data.
Experiments are commonly performed in 'trials', wherein
a set of conditions is repeated many times while neuronal
activity is tracked. In this scenario, trial-by-trial relationships
become important in considering functional interaction between
neurons; therefore, destroying this relationship for each neuron
preserves neurons' task modulation but decorrelates neurons'
trial-by-trial relationships. If one is interested in correlations over
trials, one should compare test statistics derived from functional
interaction analysis to test statistics derived from trial-shuffled
data (
Fig. 7.2B)
. To trial-shuffle data, spike trains associated
with one trial are switched with a randomly selected trial. This
process is also readily achieved in MATLAB with the command
randperm:
myData = rand(149, 10); % random perievent
matrix: 149 trials 10 bins
trialShuffledData = myData(randperm(size(myData,1)),
:); %shuffled data
These methods illustrate that data can easily be shuffled in
time or with respect to trials. As detailed above, one should
