A lot of papers have been looking at correlations between spikes and LFPs recently, to figure out how the two signals relate to each other (do LFPs drive spikes or vice versa? Is some of what we usually consider neural noise really coupling to the local network measured by the LFP?). One thing that has always bugged me about this is that most of the time the LFPs and spikes that are compared are derived from the same electrode, with the LFP obtained by low-passing and downsampling a wideband signal (usually in hardware) and the spikes by high-pass filtering, thresholding and sorting the wideband signal. So, if any small part of the spikes leak onto the LFP channel (spectral leakage) then some of the spike-LFP correlation is actually a spike-mangled-version-of-a-spike spurious correlation.
The central question is not whether these spurious correlations exist (they have to exist unless your spike waveforms are really weird), but what fraction of observed spike-LFP correlations are spurious (if it’s 5% it doesn’t matter but if it’s 95% spike-LFP correlations shouldn’t be looked at). Now to solve that question you must have some way of removing spike remnants from LFPs, then compare spike-LFP versus spike-despiked LFP correlations. So I came up with an algorithm to attack this problem. The idea is to assume a generative model for a wideband signal: the wideband signal is the sum of a low-frequency spike-free-LFP signal plus spike waveforms at the time of spikes plus noise. Then turn the Bayesian crank and solve for the spike-free-LFP. Turns out the Bayesian crank is really hard to turn here but it’s doable.
Theo (Zanos; a postdoc in our lab) then recorded some neurons in visual areas and then determined how important the spurious spike-LFP correlations were in a variety of analysis scenarios. The short answer is: if the spike-LFP metric you’re using is sensitive to the timing of the spikes and LFPs, for example a spike-triggered average of the LFP, then spectral leakage accounts for about 50% of the correlation you’re measuring (obviously depends on the exact metric you’re using). Which means that it could completely undermine your conclusions or be inconsequential depending on what you’re trying to prove/disprove. In the undermine category I’m thinking of a recent batch of papers from the Logothetis lab which try to predict spikes from LFPs; my feeling (haven’t tested this) is that any reported predictability of spikes from LFPs above 40 Hz (possibly lower) is BS. On the other hand when using metrics insensitive to time, for example comparing tuning curves of LFPs and spikes, then it’s more like 5%, it doesn’t matter unless you have some crazy-large spikes.
I should note that the despiking algorithm only works if you have access to a wideband signal; LFP signals recorded from LFP boards (for example from Plexon) are currently undespikable. IMHO get the capacity to record wideband signals rather than go with the LFP boards, not only will you be able to do more advanced LFP analyses, you will also be able to do your own filtering/thresholding to find spikes.
Theodoros P. Zanos, Patrick J. Mineault, and Christopher C. Pack
Removal of Spurious Correlations Between Spikes and Local Field Potentials
J Neurophysiol January 2011 105:474-486