[BUG] set_annotations fails for data with a meas_date prior to 1970

Try reading this example_file (800MB) and check out raw.info['meas_date'] --> it will be negative.

That’s because the data has been anonymized (cc @agramfort ) to some time prior 1970.

That’s an issue, because we represent raw.info['meas_date'] in unix time, which is 0 at 1970, and thus negative for all prior years (strictly speaking, it’s even undefined)

A negative meas_date will raise an error when we make an annotation object from scratch and try to set it with raw.set_annotations().

That is because of this code:

in set_annotations: https://github.com/mne-tools/mne-python/blob/368a490f9d7a298050ee04c2aa7882629c9fab21/mne/io/base.py#L792

–> time_of_first_sample will be negative if meas_date is negative

then later we do: https://github.com/mne-tools/mne-python/blob/368a490f9d7a298050ee04c2aa7882629c9fab21/mne/io/base.py#L798-L801

i.e., pass a negative tmin to the crop function, which then raises an error: https://github.com/mne-tools/mne-python/blob/368a490f9d7a298050ee04c2aa7882629c9fab21/mne/annotations.py#L363-L364

cc @massich

[sappelhoff]

@massich see: https://github.com/mne-tools/mne-bids/issues/193#issue-436776387

I don’t know about how anonymization is currently handled in MNE, but I think the way we SHOULD handle this is to allow for an anonymization by a certain offset in days … this would then be applied to all data recordings of a subject, so that the relative time distances keep their logic.

the anonymization_offset_days would be set to a sensible default but could also be given as an input to raw.annonymize(anonymization_offset_days=36500)

I briefly discussed this with @larsoner at the last MNE sprint, maybe he wants to weigh in

[agramfort]

If there is no limitations with fif files to have negative time stamps we should handle these files smoothly