Code repository & Versions

DoWhy is hosted on GitHub.

You can browse the code in a html-friendly format here.

v0.7: Better Refuters for unobserved confounders and placebo treatment

  • [Major] Faster backdoor identification with support for minimal adjustment, maximal adjustment or exhaustive search. More test coverage for identification.

  • [Major] Added new functionality of causal discovery [Experimental]. DoWhy now supports discovery algorithms from external libraries like CDT. [Example notebook]

  • [Major] Implemented ID algorithm for causal identification. [Experimental]

  • Added friendly text-based interpretation for DoWhy’s effect estimate.

  • Added a new estimation method, distance matching that relies on a distance metrics between inputs.

  • Heuristics to infer default parameters for refuters.

  • Inferring default strata automatically for propensity score stratification.

  • Added support for custom propensity models in propensity-based estimation methods.

  • Bug fixes for confidence intervals for linear regression. Better version of bootstrap method.

  • Allow effect estimation without need to refit the model for econml estimators

Big thanks to @AndrewC19, @ha2trinh, @siddhanthaldar, and @vojavocni

v0.6: Better Refuters for unobserved confounders and placebo treatment

  • [Major] Placebo refuter also works for IV methods

  • [Major] Moved matplotlib to an optional dependency. Can be installed using pip install dowhy[plotting]

  • [Major] A new method for generating unobserved confounder for refutation

  • Update to align with EconML’s new API

  • All refuters now support control and treatment values for continuous treatments

  • Better logging configuration

  • Dummyoutcomerefuter supports unobserved confounder

A big thanks to @arshiaarya, @n8sty, @moprescu and @vojavocni

v0.5-beta: Enhanced documentation and support for causal mediation

Installation

  • DoWhy can be installed on Conda now!

Code

  • Support for identification by mediation formula

  • Support for the front-door criterion

  • Linear estimation methods for mediation

  • Generalized backdoor criterion implementation using paths and d-separation

  • Added GLM estimators, including logistic regression

  • New API for interpreting causal models, estimates and refuters. First interpreter by @ErikHambardzumyan visualizes how the distribution of confounder changes

  • Friendlier error messages for propensity score stratification estimator when there is not enough data in a bin

  • Enhancements to the dummy outcome refuter with machine learned components–now can simulate non-zero effects too. Ready for alpha testing

Docs

Community

  • Created a contributors page with guidelines for contributing

  • Added allcontributors bot so that new contributors can added just after their pull requests are merged

A big thanks to @Tanmay-Kulkarni101, @ErikHambardzumyan, @Sid-darthvader for their contributions.

v0.4-beta: Powerful refutations and better support for heterogeneous treatment effects

  • DummyOutcomeRefuter now includes machine learning functions to increase power of the refutation.
    • In addition to generating a random dummy outcome, now you can generate a dummyOutcome that is an arbitrary function of confounders but always independent of treatment, and then test whether the estimated treatment effect is zero. This is inspired by ideas from the T-learner.

    • We also provide default machine learning-based methods to estimate such a dummyOutcome based on confounders. Of course, you can specify any custom ML method.

  • Added a new BootstrapRefuter that simulates the issue of measurement error with confounders. Rather than a simple bootstrap, you can generate bootstrap samples with noise on the values of the confounders and check how sensitive the estimate is.
    • The refuter supports custom selection of the confounders to add noise to.

  • All refuters now provide confidence intervals and a significance value.

  • Better support for heterogeneous effect libraries like EconML and CausalML
    • All CausalML methods can be called directly from DoWhy, in addition to all methods from EconML.

    • [Change to naming scheme for estimators] To achieve a consistent naming scheme for estimators, we suggest to prepend internal dowhy estimators with the string “dowhy”. For example, “backdoor.dowhy.propensity_score_matching”. Not a breaking change, so you can keep using the old naming scheme too.

    • EconML-specific: Since EconML assumes that effect modifiers are a subset of confounders, a warning is issued if a user specifies effect modifiers outside of confounders and tries to use EconML methods.

  • CI and Standard errors: Added bootstrap-based confidence intervals and standard errors for all methods. For linear regression estimator, also implemented the corresponding parametric forms.

  • Convenience functions for getting confidence intervals, standard errors and conditional treatment effects (CATE), that can be called after fitting the estimator if needed

  • Better coverage for tests. Also, tests are now seeded with a random seed, so more dependable tests.

Thanks to @Tanmay-Kulkarni101 and @Arshiaarya for their contributions!

v0.2-alpha: CATE estimation and integration with EconML

This release includes many major updates:

  • (BREAKING CHANGE) The CausalModel import is now simpler: “from dowhy import CausalModel”

  • Multivariate treatments are now supported.

  • Conditional Average Treatment Effects (CATE) can be estimated for any subset of the data. Includes integration with EconML–any method from EconML can be called using DoWhy through the estimate_effect method (see example notebook).

  • Other than CATE, specific target estimands like ATT and ATC are also supported for many of the estimation methods.

  • For reproducibility, you can specify a random seed for all refutation methods.

  • Multiple bug fixes and updates to the documentation.

Includes contributions from @j-chou, @ktmud, @jrfiedler, @shounak112358, @Lnk2past. Thank you all!

v0.1.1-alpha: First release

This is the first release of the library.