Repository navigation
This bugfix release improves the reliability of standard errors and confidence intervals in several estimators, and includes smaller fixes to model selection, summary output, and documentation.
What's Changed
- More accurate inference. Correct standard errors and confidence intervals for CausalForestDML's doubly robust ATE and ATT estimates with multiple outcomes, DRLearner estimates using supplied outcome variances, DynamicDML estimates involving multiple periods, and OrthoForest marginal effects with a treatment featurizer. By @kbattocchi in #1067, #1069, and #1077.
- More consistent model selection for multiple outcomes. Correct score normalization so that shifting the mean of one outcome does not artificially improve a model's score. By @kbattocchi in #1069.
- More robust population summaries. An invalid outcome/treatment column no longer prevents confidence intervals from being computed for the remaining columns. Summaries also render correctly with statsmodels 0.15 when standard errors are unavailable or LaTeX tables have different widths. By @kbattocchi in #1069 and #1079.
- Avoid a Cython build regression. Exclude Cython 3.3.0, which crashes during compilation, while allowing the fixed 3.3.1 release. By @kbattocchi in #1068 and #1078.
- Clearer, more accurate API documentation. Bring documented parameter names and defaults into line with the API, remove an unsupported DML constructor argument, and clarify the confounding parameters used in sensitivity analysis. By @VenishPaneliya in #1076, #1073, #1072, and #1080.
- Fix table rendering in the customer segmentation case study so its results are easier to read. By @kbattocchi in #1065.
- Make documentation examples less sensitive to harmless numerical differences, restore Python 3.9 Linux CI, and adapt a test to NumPy 2.5. By @kbattocchi in #1051, #1082, and #1079.
New Contributors
- @VenishPaneliya made their first contribution in #1076
Full Changelog: v0.17.0...v0.17.1