[gov3009-l] Applied Statistics - 10/14 - Cynthia Rudin

Hi everyone! This week at the Applied Statistics Workshop we will be welcoming* Cynthia Rudin*, Associate Professor of Statistics at MIT where she runs the Prediction Analysis Lab. She will be presenting work entitled* A Machine Learning Perspective on Causal Inference*. Please find the abstract below and on the website <http://projects.iq.harvard.edu/applied.stats.workshop-gov3009/presentations/peng-ding-harvard-sensitivity-analysis-without> . As usual, we will meet in CGIS Knafel Room 354 from noon to 1:30pm, and lunch will be provided. See you all there! To view previous Applied Statistics presentations, please visit the website <http://projects.iq.harvard.edu/applied.stats.workshop-gov3009/videos>. -- Aaron Kaufman Title: A Machine Learning Perspective on Causal Inference Abstract: Usually the terms "causal inference" and "machine learning" mix like oil and water. Machine learning models are often black box complicated functions that provide predictions without causal explanations. For causal inference, this kind of model is unacceptable. Maybe we can find ways to harness the predictive power of machine learning methods for the purpose of causal inference. I will discuss three very recent preliminary ideas, from the perspective of a machine learner: 1) Causal Falling Rule Lists (with Fulton Wang). This is a machine learning method that bridges the gap - it's nonlinear yet interpretable, and models causal effects. (More details below.) 2) The Factorized Self-Controlled Case Series Method: An Approach for Estimating the Effects of Many Drugs on Many Outcomes (with Ramin Moghaddass and David Madigan). We estimate the effects of many drugs on many outcomes simultaneously. This Bayesian hierarchical model is formulated with layers of latent factors, which substantially helps with both computation and interpretability. 3) Robust Testing for Causal Inference in Natural Experiments (with Md. Noor-E-Alam). We claim there is a major source of uncertainty that is ignored in matched pairs tests, which is how the matches were constructed by the experimenter. No matter which reasonably good experimenter conducts the test, the hypothesis test result still ought to hold. Our robust matched pairs tests use mixed-integer programming. ----- (More on Causal Falling Rule Lists) ---- A Causal Falling Rule List is a sequence of IF-THEN rules that specifies heterogeneous treatment effects. In this model, (a) the order of rules determines the treatment effect subgroup that a subject belongs to, (b) the treatment effect decreases monotonically down the list. For example, a Causal Falling Rule List might say that if a person is below 60 years, then they are in the highest treatment effect subgroup, such that administering a drug will result in a 20 unit increase in good cholesterol levels. Otherwise, if they are regular exercisers, then taking the drug will result in a 15 unit increase in cholesterol level. Finally, if they satisfy neither of these rules, they are in the default treatment subgroup, such that the drug will result in only a 2 unit increase. The collection of rules, their sequence, and the treatment effects are learned from data. ---------- -- Aaron R Kaufman PhD Candidate, Harvard University Department of Government 818.263.5583