Learning from Interventions 2022
2022-02-14..2022-02-17
Berkeley, United States
[Algorithmic Aspects of Causal Inference 2022, Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026), Quantifying Uncertainty: Stochastic, Adversarial, and Beyond, Quantum Physics and Statistical Causal Models (Simons Institute workshop, 2022), Learning, Algorithm Design and Beyond Worst-Case Analysis, Meet the Fellows Welcome Event]
who it's for
causal inference researchers, applied machine learning researchers, biostatisticians and epidemiologists, policy and program evaluation researchers, computational biologists and geneticists, reinforcement learning practitioners
topics
causal representation learning; latent confounder identification; off-policy and offline reinforcement learning; sequential and adaptive experiment design; target trial emulation for observational data; randomized controlled trial estimators and adjustments; missing covariate imputation with indicators; genome-wide knockoff testing across populations; machine learning driven protein design; differentially private randomized experiments
| title | author | length |
|---|---|---|
| Causal Emergence: When Distortions In A Map Obscure The Territory | Frederick Eberhardt | 38m |
Argues that causal emergence is defined by finding a coarsening partition of a finite-state transition matrix that maximizes effective information, computed as… | ||
| Learning Causal Representations From Unknown Interventions | Kun Zhang | 28m |
Claims identifiability for latent causal variables that participate in changing mechanisms and gives concrete tests and algorithms to recover them. Uses the… | ||
| Online Reinforcement Learning With The Help Of Confounded Offline Data | Uri Shalit | 31m |
Two linked papers claim offline, confounded logs can speed online learning under concrete assumptions. For linear contextual bandits the method estimates… | ||
| Panel Discussion | Elias Bareinboim, Frederick Eberhardt, Kun Zhang, Uri Shalit | 31m |
Panelists argue that the core issues are practical: using mutual information or max-entropy to justify coarse-graining produces misleading identification, and… | ||
| Counterfactual Inference For Sequential Experiment Design | Raaz Dwivedi | 30m |
A policy-agnostic nearest-neighbor estimator treats each treatment as an N-by-T matrix and reduces counterfactuals to sequential matrix completion, proving… | ||
| Learning from RCTs in Public Health and Medicine | James Robins | 37m |
Argues observational data can be judged against randomized benchmarks by target trial emulation where each eligible person is nested into every trial they… | ||
| Panel Discussion | Raaz Dwivedi, James Robins | 21m |
Argues for concrete procedures: run multiple estimators and report why they disagree, specifically G-formula vs IPW vs doubly robust, publish de-identified… | ||
| To Adjust Or Not To Adjust? Estimating The Average Treatment Effect In Randomized Experiments... | Peng Ding | 32m |
Randomized trials with missing covariates (example: 27 covariates, seven with missingness, ~20% units incomplete) get an easy, provably good estimator: impute… | ||
| Searching For Causal Genetic Mechanisms Across Human Populations | Chiara Sabatti | 33m |
Proposes a concrete pipeline that builds knockoff framework variables per population, runs Lasso-style importance scores, and tests conditional independence… | ||
| Machine Learning-Based Design Of Proteins | Jennifer Listgarten | 32m |
Machine learning trains a predictive model on assay data for a seven‑amino‑acid insertion in the AAV capsid, then optimizes a distribution over sequences… | ||
| Panel Discussion | Chiara Sabatti, Jennifer Listgarten, Peng Ding | 32m |
Panelists argue that the missingness‑indicator approach fails under observational unconfoundedness and under conditioning on missingness patterns, with… | ||
| Learning And Testing Causal Models: A Property Testing Viewpoint | Arnab Bhattacharyya | 31m |
Frames causal testing as distribution property testing and gives tight sample trade-offs: trade-off between observational and interventional samples is… | ||
| Some Staged Tree Models For Learning From Interventions | Liam Solus | 29m |
Defines a restricted family of colored probability trees called CS-trees and proves a global Markov-type property: distributions in the model correspond… | ||
| Balancing Covariates In Randomized Experiments: The Gram--Schmidt Walk Design | Christopher Harshaw | 33m |
Claims a fundamental balance–robustness trade-off in randomized experiments and gives a constructive design, the Gram–Schmidt walk, that interpolates between… | ||
| Learning Abstract Causal Structure From Causal Interventions In Children And Artificial Agents | Alison Gopnik | 33m |
Claim: preschoolers extract abstract causal models from sparse statistics and improve them by active interventions on a simple blicket detector task. Methods… | ||
| Balancing Weights For Causal Effects With Panel Data: Some Recent Extensions To The Synthetic... | Avi Feller | 34m |
Presents concrete extensions to synthetic control: change the optimization to trade off separate versus pooled balancing with a tuning parameter nu, and… | ||
| Designing Adaptive Experiments For Policy Learning And Inference | Molly Offer-Westort | 32m |
Adaptive experiments address multi-armed bandit problems but require pre-specified design choices: use top-two Thompson Sampling with probability floors and… | ||
| Panel Discussion | Alison Gopnik, Avi Feller, Molly Offer-Westort | 25m |
Shows that regret‑optimal bandit algorithms lose precision on low arms but a quick practical fix is to inflate CIs by 1.5 so UCB/Thompson sampling become… | ||
| Panel Discussion | Arnab Bhattacharyya, Christopher Harshaw, Liam Solus | 22m |
Defines robustness as controlling average error over an L2 ball of potential outcomes and provides a design that guarantees an unbiased Horvitz-Thompson… | ||
| Causal Estimands And Interventions | Vanessa Didelez | 30m |
Proposes separable effects as an estimand that splits a randomized treatment into two manipulable components (one driving the disease pathway, one driving… | ||
| Unmeasured Spatial Confounding | Georgia Papadogeorgou | 27m |
Proposes that unmeasured confounders with spatial structure are identifiable and correctable by exploiting spatial correlation, using a distance-adjusted… | ||
| A Practitioner’S Guide To Intent-To-Treat Effects From Multisite (Blocked)... | Luke Miratrix | 27m |
Defines four concrete estimands (finite or super population crossed with person average vs site average) and catalogs about 15 estimators applied to 12 real… | ||
| Panel Discussion | Georgia Papadogeorgou, Luke Miratrix, Vanessa Didelez | 26m |
Panel compares concrete problems and pragmatic answers: spatial confounding needs assumptions about smoothness or tail behavior (non‑Gaussianity might help)… | ||
| Causal Inference For Socio-Economic And Engineering Systems | Anish Agarwal | 30m |
Phase‑three trial data on 1,162 patients with ADAS‑Cog scores over two years and four arms (three therapies plus placebo) is framed as a missing‑entry tensor… | ||
| Efficient Least Squares For Estimating Total Effects Under Linearity And Causal Sufficiency | Emilija Perkovic | 32m |
Proposes a reparameterization of linear SCMs over an MP-DAG into ordered bucket blocks and defines a plug‑in G regression estimator that estimates blockwise… | ||
| A Robust, Differentially Private Randomized Experiment For Evaluating Online Educational... | Hyunseung Kang | 28m |
Designs a randomized experiment that injects noise with RP-RCT and a two-spinner, sample-split procedure to detect protocol deviators and still estimate causal… | ||
| Panel Discussion | Anish Agarwal, Emilija Perkovic, Hyunseung Kang | 34m |
Panel covers three practical research threads: a classroom RCT that implements local differential privacy and analyzes how participant cheating breaks… | ||
| nothing matches | ||