1. "Doubly Robust Confidence Sequences for Sequential Causal Inference" (Wauby-Smith et al. 2021): The data I typically work with comes in a stream--customers taking actions on a website. However, the traditional experimental analysis tools (e.g. t-tests) don't take advantage of this. Last month I was thinking about how we could improve the privacy of our data analysis systems if we recursively analysed it as it came in. This new working paper by Wauby-Smith et al. asks how we can make valid causal inferences as data streams in.
2. "Why did the Distribution Change?" (Budhathoki 2021): This piece got me pretty excited. A pretty straightforward development of graphical causal models for causal discovery. Though they do save the biggest caveat for the last paragraph of the paper:
2. "Why did the Distribution Change?" (Budhathoki 2021): This piece got me pretty excited. A pretty straightforward development of graphical causal models for causal discovery. Though they do save the biggest caveat for the last paragraph of the paper:
Finally, our attribution proposal requires a causal graph, which may not be identifiable from observational data. If the causal graph is not identifiable, the Shapley values will not be identifiable as well. Therefore, the question on the robustness of our attribution proposal to causal graph misspecification deserves further research.
There’s an efficient frontier between growth rates and certainty, and at some level frauds are all an attempt to imagine the company existing beyond that frontier.