IndexError when a declared group has no trials

[kiante-fernandez]

EventModel._estim_probs_groups collects per-group results by appending over np.unique(groups):

for cur_group in data_groups:
    likes_events_group.append(self.estim_probs(...))

but then indexes that list by group id:

all_xreventprobs.data[np.ix_(groups == cur_group,
    range(likes_events_group[cur_group][1].shape[1]),
    self.channel_map[cur_group, :] >= 0)] = likes_events_group[cur_group][1]

If the group ids present in the data are not 0..n_groups-1 the two disagree, since the list is positional but the index is an id.

Reproduced with three declared groups (channel and time maps with 3 rows) over 4 trials:

• groups = [0, 0, 1, 2] works, likelihood -16.34
• groups = [0, 0, 2, 2], so group 1 is declared but empty, raises IndexError: list index out of range

This can happen when a level of a grouping variable is absent after trial rejection, or when a participant subset does not contain every condition.

The fix is to index by position, for example dict(zip(data_groups, likes_events_group)), or to error early with a clear message if a declared group has no trials.

Noticed while working on #244, where the per-trial scatter I added inherits the same pattern and so raises a few lines earlier. May be of interest to @rickdott given #309.

[GWeindel]

Right, nice catch! @jelmerborst you want to address this ? Otherwise I'll have time end of august/beginning of september

[MohammadYusif]

Happy to pick this up if @jelmerborst hasn't started. The mismatch is that likes_events_group is built positionally over np.unique(groups) while the assignment loop indexes it by group id, and group_lkh inherits the same problem so it silently stops lining up with group_labels as soon as a declared group has no trials. My plan is to keep the position and the group index apart in _estim_probs_groups and report a nan likelihood for a declared group without trials, but I'm equally happy to go the "error early" route instead if you'd rather not allow empty groups at all.