Implement causal impact estimation

[davidpomerenke]

The most intuitive estimators are those that predict a counterfactual and compare with it. These are especially nice because they give us the impact of every single event.

• Synthetic control. Deals with hidden confounding under some assumptions (geographic uniformity of the confounders). Has other unrealistic assumptions (causal impact does not transfer to other regions) that make it overly conservative in our case.
  • simple version in https://github.com/pymc-labs/CausalPy
  • more complicated version in https://github.com/google/CausalImpact (still have to understand the math behind it; in R)
  • Python version of it in https://github.com/WillianFuks/tfcausalimpact, it might not be exactly the same but just very similar concepts
• Interrupted time series. Does not deal with hidden confounding, thus tends to be overly optimistic about the impact of the events. This is a very simple concept
  • in https://github.com/pymc-labs/CausalPy
  • also simple to implement ourselves using any time-series prediction model, e.g. ARIMA
  • as a machine learning model, we can use random forests or boosted trees (from https://scikit-learn.org/stable/)

Generic causal inference estimators:

• Regression. This can be extended to time-series relatively easily.
  • Regression in Python: https://www.statsmodels.org/stable/index.html
  • There's probably better R libraries for regression
• Inverse propensity weighting. Predict the probability of events using any time-series prediction model, then do some fancy inverse weighting.
  • Implemented in https://www.pywhy.org/dowhy/v0.11.1/dowhy.causal_estimators.html#module-dowhy.causal_estimators.propensity_score_estimator. DoWhy is a bit cumbersome, maybe we can find another implementation or just copy the code that we actually need.
  • in R: https://github.com/ngreifer/WeightIt
• Doubly robust estimation. Combines the two above, and is in principle more robust. But in my master thesis I found it to be the opposite of robust.
  • Implemented in https://econml.azurewebsites.net/_autosummary/econml.dr.LinearDRLearner.html

Meta-learners: These are more fancy. They can be used with any machine learning models, which is cool. Double Machine Learning seems most promising to me, but I still have to read more about it. See e.g. https://matheusfacure.github.io/python-causality-handbook/22-Debiased-Orthogonal-Machine-Learning.html or https://causalml-book.org/. And causal random forests might also be relevant.

I suggest focusing on the first two for the first versions of the project, synthetic control and interrupted time series.

[davidpomerenke]

Alternative to CausalImpact: ConformalImpact. See https://www.linkedin.com/posts/activity-7180155807836053504-EaYk

[davidpomerenke]

• time-series
  • [x] interrupted time series
  • [ ] synthetic control
• generic
  • [ ] regression
  • [ ] inverse propensity weighting
  • [ ] doubly robust
• fancy
  • [ ] double machine learning
  • [ ] causal random forests

[davidpomerenke]

Inverse propensity weighting has been added to CausalPy https://github.com/pymc-labs/CausalPy/releases/tag/0.3.0