We suspect that when providing a feature that has mostly unique string values and some NaN values, then this might lead the ipw model to give just weights of 1.
This should be:

1. Checked.
2. If exists, to add a check and warning when it happens.