Many analyses of sleep EEG data are effectively agnostic to the polarity of the EEG signal. That is, you could flip the signal (i.e. multiply every sample value by -1) and still obtain equivalent results, e.g. from most spectral analyses. For certain analyses that consider the phase of a signal, however, it will in fact matter that the polarity of the signal is correct.
As described in this vignette on the Luna website, we recently discovered some issues with the polarity of EEG signals in some NSRR studies. Specifically, they were flipped or upside-down: positive-to-negative and vice versa. We tracked the source of the issue and considered some approaches to flag whether a sleep EEG channel is likely to be flipped. These include looking at correlations between channels, at the coupling between spindles and slow oscillations, and an analysis that considers the small but characteristic asymmetries in the NREM sleep EEG.
The NSRR is in the process of fixing affected studies and will post an update in due course. Note that signals other than the EEG may have been affected by this issue.
The bottom line: if you are performing analyses that might depend on a signal's polarity (e.g. phase relationships in EEG data), then you should attend to this issue!
Shaun Purcell, firstname.lastname@example.org