Spring 2020 complete list with abstracts
Tuesday, May 26, 2020: Ya Xu (LinkedIn)
"Causal Inference Challenges in Industry: A perspective from experiences at LinkedIn"
Discussant: Iavor Bojinov (Harvard)
Abstract: In this talk, we will briefly give some background how online controlled experiments are commonly used in industry, and introduce some challenges we face, and also some opportunities in novel applications.
[Video] [Speaker slides] [Discussant slides]Tuesday, May 19, 2020: Susan Murphy (Harvard University)
"Inference for Batched Bandits"
Discussant: Stefan Wager (Stanford University)
Abstract: As bandit algorithms are increasingly utilized in scientific studies and industrial applications, there is an associated increasing need for reliable inference methods based on the resulting adaptively collected data. In this work, we develop methods for inference on data collected in batches using a bandit algorithm. When there is no unique arm we prove that the ordinary least squares estimator(OLS) is not asymptotically normal on data collected using standard bandit algorithms. This is the case even when the bandit is constrained to select each arm with probabilities bounded away from 0 and 1. We show that this problem can be traced to the fact that the arm selection probabilities do not concentrate. We take advantage of the batched setting to develop a Batched OLS estimator (BOLS) that we prove is (1) asymptotically normal on data collected from both multi-arm and contextual bandits and (2) robust to nonstationarity in the baseline reward. This is joint work with Kelly Zhang and Lucas Janson.
[Video] [Paper] [Speaker slides] [Discussant slides]Tuesday, May 12, 2020: Ilya Shpitser (Johns Hopkins University)
"Identification and estimation in graphical models of missing data"
Discussant: Jin Tian (Iowa State University)
Abstract: Missing data is a pervasive problem in data analyses, resulting in datasets that contain censored realizations of a target distribution. Many approaches to inference on the target distribution using censored observed data rely on missing data models represented as a factorization with respect to a graph. We describe a simple characterization of all identified missing data models where the full data distribution factorizes with respect to a directed acyclic graph (DAG). We show how statistical inference may be performed within maximum likelihood and semi-parametric frameworks in this class of models. Finally, we discuss why identification of marginal target parameters in missing data is significantly more complicated and requires approaches more general than those developed in classical causal inference problems.
This is joint work with Rohit Bhattacharya, Razieh Nabi, Daniel Malinsky, Eric Tchetgen Tchetgen, and James M. Robins.
[Video] [Paper 1] [Paper 2] [Paper 3] [Speaker slides]Tuesday, May 5, 2020: Eric Tchetgen Tchetgen (Wharton)
"Selective Machine Learning of Doubly Robust Functionals"
Discussant: Stijn Vansteelandt (UGent)
Abstract: While model selection is a well-studied topic in parametric and nonparametric regression or density estimation, model selection of possibly high-dimensional nuisance parameters in semiparametric problems is far less developed. In this paper, we propose a selective machine learning framework for making inferences about a finite-dimensional functional defined on a semiparametric model, when the latter admits a doubly robust estimating function. We introduce two model selection criteria for bias reduction of functional of interest, each based on a novel definition of pseudo-risk for the functional that embodies this double robustness property and thus may be used to select the candidate model that is nearest to fulfilling this property even when all models are wrong. We establish an oracle property for a multi-fold cross-validation version of the new model selection criteria which states that our empirical criteria perform nearly as well as an oracle with a priori knowledge of the pseudo-risk for each candidate model. We also describe a smooth approximation to the selection criteria which allows for valid post-selection inference. Finally, we apply the approach to model selection of a semiparametric estimator of average treatment effect given an ensemble of candidate machine learners to account for confounding in an observational study. This is joint work with Yifan Cui.
[Video] [Paper] [Speaker slides]Tuesday, April 28, 2020: Edward Kennedy (Carnegie Mellon University)
"Optimal doubly robust estimation of heterogeneous causal effects"
Discussant: James Robins (Harvard University)
Abstract: Heterogeneous effect estimation has become a major enterprise in causal inference, with ramifications across medicine and social science, e.g., improving understanding of variation, as well as informing policy and optimizing treatment decisions. Many methods for estimating the conditional average treatment effect (CATE) have been proposed in recent years; however, there are important gaps in the literature, particularly on the theory side, vis-a-vis understanding if and when such methods can be optimal. These gaps are especially pronounced in settings where the CATE is more structured and less complex than the rest of the data-generating process. We contribute in several main ways. First, we study a two-stage doubly robust CATE estimator, similar to that proposed by van der Laan (2013) and used by others since. We give a generic model-free error bound, which, despite its generality, yields sharper results than those in the current literature. Second, we illustrate the results in nonparametric models with smoothness or sparsity, and give sufficient conditions for oracle efficiency, depending on nuisance structure. Underlying our error bound is a general oracle inequality for regression with estimated or imputed outcomes, which is of independent interest; this is the third main contribution. The fourth contribution is aimed at understanding the fundamental statistical limits of CATE estimation. To that end, we propose and study a local polynomial adaptation of the R-Learner and double-residual regression (Nie & Wager 2017, Robinson 1998). We show that this estimator is oracle efficient under weaker conditions than the first estimator, when using a specialized form of sample splitting and careful choices of tuning parameters. These conditions are the weakest currently found in the literature, and we conjecture that they are minimal in a minimax sense. We go on to give error bounds in the challenging regime where oracle rates cannot be achieved. We illustrate our results and study finite-sample properties with simulations.
[Video] [Paper] [Speaker slides]Tuesday, April 21, 2020: Elizabeth Ogburn (Johns Hopkins University)
"Social network dependence, unmeasured confounding, and the replication crisis"
Discussant: Ilya Shpitser (Johns Hopkins University)
[Video] [Paper] [Speaker slides]Tuesday, April 14, 2020: Elizabeth Tipton (Northwestern University)
"Will this Intervention Work in this Population? Designing Randomized Trials for Generalization"
Discussant: Andrew Gelman (Columbia University)
[Video] [Website: The Generalizer] [Paper] [Speaker slides]Tuesday, April 8, 2020: Hyunseung Kang (University of Wisconsin-Madison)
"Inferring Treatment Effects After Testing Instrument Strength in Linear Models" (w/ Nan Bi and Jonathan Taylor)
Discussant: Will Fithian (UC Berkeley)
Abstract: A common practice in IV studies is to check for instrument strength, i.e. its association to the treatment, with an F-test from regression. If the F-statistic is above some threshold, usually 10, the instrument is deemed to satisfy one of the three core IV assumptions and used to test for the treatment effect. However, in many cases, the inference on the treatment effect does not take into account the strength test done a priori. In this paper, we show that not accounting for this pretest can severely distort the distribution of the test statistic and propose a method to correct this distortion, producing valid inference. A key insight in our method is to frame the F-test as a randomized convex optimization problem and to leverage recent methods in selective inference. We prove that our method provides conditional and marginal Type I error control. We also extend our method to weak instrument settings. We conclude with a reanalysis of studies concerning the effect of education on earning where we show that not accounting for pre-testing can dramatically alter the original conclusion about education's effects.
[Video] [Paper] [Speaker slides]Tuesday, March 31, 2020: Dylan Small (Wharton)
"Testing an Elaborate Theory of a Causal Hypothesis" (w/ Bikram Karmakar)
Discussant: Peter Bühlmann (ETH Zurich)
Abstract: When R.A. Fisher was asked what can be done in observational studies to clarify the step from association to causation, he replied, “Make your theories elaborate” -- when constructing a causal hypothesis, envisage as many different consequences of its truth as possible and plan observational studies to discover whether each of these consequences is found to hold. William Cochran called “this multi-phasic attack…one of the most potent weapons in observational studies.” Statistical tests for the various pieces of the elaborate theory help to clarify how much the causal hypothesis is corroborated. In practice, the degree of corroboration of the causal hypothesis has been assessed by a verbal description of which of the several tests provides evidence for which of the several predictions. This verbal approach can miss quantitative patterns. We develop a quantitative approach to making statistical inference about the amount of the elaborate theory that is supported by evidence.
[Video] [Paper] [Speaker slides]