[jonathan-taylor]

For 0.4.0: offsets of multi-response models, with as many columns as the full response.

Outcome(['y1', 'y2', 'y3'], offset='O')
# -> y1, y2, y3, offset                      one offset, the same for every response

Outcome(['y1', 'y2', 'y3'], offset={'y1': 'O1', 'y2': 'O2', 'y3': log_exposure})
# -> y1, y2, y3, offset[y1], offset[y2], offset[y3]

Outcome(Multinomial('kind'), offset={'lo': 'O_lo', 'mid': 'O_mid', 'hi': 'O_hi'})
# -> kind, offset[lo], offset[mid], offset[hi]

• A dict gives one column for each response. Its keys are the responses: the columns of a list of columns, or the classes of a Multinomial. For a multinomial the keys are its levels whether its output is labels or one-hot, so glmstar can take either form; for a Multinomial of counts, they are its count columns. Each value is a column or a single-column feature, as before.
• Keys rather than positions, so an offset can't end up in the wrong column silently. The offset[...] columns follow the order of the responses, whatever the dict's order. A missing or unknown key raises, as does a dict for a response with a single linear predictor (a column, Surv, Binomial).
• A list of columns is now a response of Outcome: the responses of a multi-response model.
• Unchanged: a single column (offset='O') stays one column, offset, which glmstar broadcasts to every response. transform(X, response=False) gives the (n, k) offsets of new data. wildcard(exclude=outcome) leaves out every source column.

For glmstar: a column offset is broadcast to every response, and columns offset[...] are one per response, in the response's order. If both are present, glmstar should raise an error.

Tests: six new tests in tests/test_outcome.py: dict order vs response order, a feature as a value, offsets of new data, multinomial labels and one-hot, counts, errors, polars. The full suite gives 282 passed and 5 strict xfails, and the docs page executes.

🤖 Generated with Claude Code