Test Impact of bound updating restrictions for grouped data
carolyner opened this issue · comments
When using sample_reduce method, updating of bounds are limited by the worst case for grouped targets. This means that when/how targets are grouped can have a big impact on updating of bounds.
Test this by branching and removing code that limits updating in get_new_bounds (lines 40-51), shown below. Also note that bounds would only be based on "deltas" when using the ratchet/gtbr method.
closing this issue: for the percentile method, the target bounds can update at different rates across goups. For the direct method, the bounds all narrow by the same percentage, but the step size depends on the initial bounds and potentially large step sizes may be required initially. Picking this issue up in another thread.