[jonathan-taylor]

Candidate for 0.3.0, so glmstar can build y with Outcome and find its columns by name.

Outcome(response, **columns). Each keyword names an output column, so an estimator can find it by name:

Y = Outcome(Surv('time', 'status'), offset='O', weight='w', strata='inst')
y = Y.fit_transform(D)                     # columns: stop, status, offset, weight, strata
X = MS([wildcard(exclude=Y)]).fit_transform(D)   # leaves out time, status, O, w, inst

• A keyword's value is a column of the data, taken as it is (a categorical such as strata keeps its labels), or a single-column feature fit as a term, e.g. R's offset(log(exposure)) as a derived_feature with a log transform. A feature with more than one column raises.
• The response can be a Surv, Binomial or Multinomial, or a plain column. Its columns come first, then the others in keyword order. A name used twice raises.
• transform(X, response=False) builds only the other columns, e.g. the offset of new data for predict(X, offset=...).
• With Binomial(..., output='proportion'), the trials become the column weight, multiplied by any weight= given (R's prior weights of a proportion).
• The given response object isn't modified: a copy is fit.

Breaking: Surv names its columns by role, start, stop and status (R's names, and lifelines' defaults for time-varying Cox), not after the data's columns. So Surv('time', 'status') now gives stop and status.

Also: polars and pyarrow output keeps categorical columns (as a polars Enum; pyarrow has none, so strings there).

Tests: tests/test_outcome.py. A Poisson GLM with offset=log(exposure) and a stratified PHReg with an offset match statsmodels fit directly on the data. Also covered: proportion weights, the offset of new data, errors, arrays, polars and pyarrow. The full suite gives 276 passed and 5 strict xfails.

Docs: an Outcome section on the responses page, executed locally.

For glmstar: default offset_id, weight_id and the CoxFamily ids to offset, weight, start, stop, status and strata, so CoxNet().fit(X, y) works directly.

🤖 Generated with Claude Code