Spring 2023 complete list with abstracts
Tuesday, June 6, 2023: Erin Gabriel (University of Copenhagen)
- Title: Derivation and usefulness of tight symbolic causal bounds for measures of benefit in observational and imperfect randomized studies with ordinal outcomes
- Discussant: Michael Fay (NIH/NIAID)
- Abstract: A causal query (estimand) will commonly only be identifiable from observed data with the assumption of no unmeasured confounders. However, some valuable estimands, like the probability of benefit, are not identifiable even in perfect randomized trials. Regardless of what causes the lack of identifiability, it may still be possible to derive bounds on the query based on the observable variable distribution. Bounds, numeric or symbolic, can often be more valuable than a statistical estimator derived under implausible assumptions. Symbolic bounds, however, provide a measure of uncertainty and information loss due to the lack of an identifiable estimand even in the absence of data. We develop and describe a general approach for the computation of symbolic bounds and characterize a class of settings in which our method is guaranteed to provide tight valid bounds. This is particularly useful for estimands, such as the probability of benefit, that are not identifiable even in perfect randomized trials but are valuable measures of treatment impact. We demonstrate our bounding method by deriving bounds for the probability of benefit, and variations of it, in imperfect randomized trials and observational settings with unmeasured confounders, extending the settings where these measures can be used. In confounded observational settings, we derive general sharp bounds for an ordinal outcome with an arbitrary number of levels.
[Video]Tuesday, May 30, 2023
- Student speaker 1: Benedicte Colnet (INRIA)
- Title: Risk ratio, odds ratio, risk difference... Which causal measure is easier to generalize?
- Abstract: There are many measures to report so-called treatment or causal effect: absolute difference, ratio, odds ratio, number needed to treat, and so on. The choice of a measure, e.g. absolute versus relative, is often debated because it leads to different appreciations of the same phenomenon; but it also implies different heterogeneity of treatment effect. In addition some measures -- but not all -- have appealing properties such as collapsibility, matching the intuition of a population summary. We review common measures and their pros and cons typically brought forward. Doing so, we clarify notions of collapsibility and treatment effect heterogeneity, unifying different existing definitions. Our main contribution is to propose to reverse the thinking: rather than starting from the measure, we start from a non-parametric generative model of the outcome. Depending on the nature of the outcome, some causal measures disentangle treatment modulations from baseline risk. Therefore, our analysis outlines an understanding what heterogeneity and homogeneity of treatment effect mean, not through the lens of the measure, but through the lens of the covariates. Our goal is the generalization of causal measures. We show that different sets of covariates are needed to generalize an effect to a different target population depending on (i) the causal measure of interest, (ii) the nature of the outcome, and (iii) the generalization's method itself (generalizing either conditional outcome or local effects).
[Slides] [Video]
- Student speaker 2: Keegan Harris (CMU)
- Title: Strategyproof Decision-Making in Panel Data Settings
- Abstract: We consider the classical problem of decision-making using panel data, in which a decision-maker gets noisy, repeated measurements of multiple units (or agents). We consider a setup where there is a pre-intervention period, when the principal observes the outcomes of each unit, after which the principal uses these observations to assign a treatment to each unit. Unlike this classical setting, we permit the units generating the panel data to be strategic, i.e. units may modify their pre-intervention outcomes in order to receive a more desirable intervention. We first identify a necessary and sufficient condition under which such a strategyproof intervention policy exists, and provide a strategyproof mechanism with a simple closed form when one does exist. Along the way, we prove impossibility results for strategic multiclass classification, which may be of independent interest. When there are two interventions, we establish that there always exists a strategyproof mechanism, and provide an algorithm for learning such a mechanism. For 3+ interventions, we provide an algorithm for learning a strategyproof mechanism if there exists a sufficiently large gap in principal rewards between different interventions. Finally, we empirically evaluate our model using panel data collected from product sales over 18 months. We find that our methods compare favorably to baselines which do not take strategic interactions into consideration, even in the presence of model misspecification.
[Slides] [Video]Tuesday, May 23, 2023: Richard Samworth (University of Cambridge)
- Title: Optimal nonparametric testing of Missing Completely At Random, and its connections to compatibility
- Discussant: Yixin Wang (University of Michigan)
- Abstract: Given a set of incomplete observations, we study the nonparametric problem of testing whether data are Missing Completely At Random (MCAR). Our first contribution is to characterise precisely the set of alternatives that can be distinguished from the MCAR null hypothesis. This reveals interesting and novel links to the theory of Fréchet classes (in particular, compatible distributions) and linear programming, that allow us to propose MCAR tests that are consistent against all detectable alternatives. We define an incompatibility index as a natural measure of ease of detectability, establish its key properties, and show how it can be computed exactly in some cases and bounded in others. Moreover, we prove that our tests can attain the minimax separation rate according to this measure, up to logarithmic factors. Our methodology does not require any complete cases to be effective, and is available in the R package MCARtest.
[Recording] [Slides]Tuesday, May 9, 2023: Yuhao Wang (Tsinghua University)
- Title: Root-n-consistent estimators for average treatment effect with minimal sparsity
- Discussant: Rajarshi Mukherjee (Harvard University)
- Abstract: This talk is about root-n-consistent estimation of average treatment effects with high dimensional confounders under minimal sparsity conditions. The entire talk is splitted into two parts. In part I, we introduce a debiased inverse propensity score weighting (DIPW) scheme for average treatment effect estimation that delivers root-n-consistent estimates when the propensity score follows a sparse logistic regression model; and the outcome regression functions are permitted to be arbitrarily complex. We further demonstrate how confidence intervals centred on our estimates may be constructed. Our theoretical results quantify the price to pay for permitting the regression functions to be unestimable, which shows up as an inflation of the variance of the estimator compared to the semiparametric efficient variance by a constant factor, under mild conditions. We also show that when outcome regressions can be estimated faster than a slow o(1 / log n) rate. Finally, we show how propensity score models with more general link functions may be handled within our framework. This is based on a joint work with Professor Rajen Shah from the University of Cambridge.
In part II, we introduce an extension of DIPW estimator, which we call double calibration (DCal) estimator. By calibrating both the outcome regression model and the propensity model, DCal is guaranteed to deliver root-n-consistent estimation when either the propensity score model or the outcome regression model follows a sparse generalized linear model, and the other nuisance function can be arbitrarily complex. In other words, DCal is agnostic to the identity of the sparse nuisance function. We further discuss some extensions to semiparametric estimation of a single regression coefficient in high dimensional partially linear models. This is based on a joint work with Professor Lin Liu from Shanghai Jiao Tong University.
[Video] [Slides] [Discussant slides] [Related paper #1, #2]Tuesday, May 2, 2023: M. (Thijs) van Ommen (Utrecht University)
- Title: Graphical Representations for Algebraic Constraints of Linear Structural Equations Models
- Discussant: Rohit Bhattacharya (Williams College)
- Abstract: For any directed acyclic graph, the set of observational distributions realizable by that graph can be described in terms of conditional independence constraints. When latent confounders are considered, conditional independence constraints no longer suffice for this purpose. We study linear structural equation models, where the constraints take the form of polynomial (in)equalities on the observed covariance matrix. However, these polynomials can be exponentially large, making them impractical for many purposes. We present a graphical notation for many of these polynomial constraints.
[Video] [Slides] [Discussant slides]Tuesday, April 25, 2023: Matthew Gentzkow (Stanford University)
- Title: Causal Interpretation of Structural IV Estimands
- Discussant: Peter Hull (Brown University)
- Abstract: We study the causal interpretation of instrumental variables (IV) estimands of nonlinear, multivariate structural models with respect to rich forms of model misspecification. We focus on guaranteeing that the researcher's estimator is sharp-zero consistent, meaning that the researcher concludes that the endogneous variable has no causal effect on the outcome whenever this is actually the case. Sharp-zero consistency generally requires a condition on the researcher's estimator that we call strong exclusion. When a researcher has access to excluded, exogenous variables, strong exclusion can often be achieved by appropriate choice of estimator and instruments. Failure of strong exclusion can lead to large bias in estimates of causal effects in realistic situations. Our results cover many settings of interest including models of differentiated goods demand with endogenous prices and models of production with endogenous inputs.Tuesday, April 18, 2023: Philipp Bach and Sven Klaassen (University of Hamburg)
- Title: (Tutorial) DoubleML - A state-of-the-art framework for double machine learning in Python and R
- Abstract: The Python and R packages DoubleML implement the double/debiased machine learning framework of Chernozhukov et al. (2018) for causal machine learning. This talk serves as an introduction to the double machine learning framework and as a tutorial for the implementation in Python and R. The double machine learning framework consists of three key ingredients: Neyman orthogonality, high-quality machine learning estimation and sample splitting. In DoubleML, estimation of nuisance components can be performed by various state-of-the-art machine learning methods that are available in the mlr3 ecosystem for R and scikit-learn for Python, respectively. The package allows users to perform inference in a variety of causal models, including partially linear and interactive regression models and their extensions to instrumental variable estimation. The object-oriented implementation of DoubleML enables a high flexibility for the model specification and makes it easily extendable. We demonstrate how users of DoubleML can perform valid inference based on machine learning methods in code examples with simulated and real data. Moreover, we offer an outlook on current and future extensions of the package, including for example quantile treatment effects and conditional average treatment effects.
[Video] [Slides] [Website] [Paper]Tuesday, April 11, 2023: Niels Richard Hansen (University of Copenhagen)
- Title: Cyclic graphical models and causal learning
- Discussant: Patrick Forré (University of Amsterdam)
- Abstract: Directed Graphs (DGs) can be used both formally and informally to represent and communicate causal relations. The formal mathematical theory is particularly well developed for Directed Acyclic Graphs (DAGs) to support structural causal models, do-calculus, identification theory and causal learning. It is natural to interpret DGs with cycles as allowing for feedback mechanisms, but this can be formalized by different incompatible mathematical theories. In the first part of the talk I will survey two competing theories: equilibrium models and dynamic models. In the second part of the talk I will focus on so-called local independence models induced by dynamic models, and their graphical representation via DGs and Directed Mixed Graphs (DMGs). I will show how equivalence classes of DMGs can be represented in terms of maximal elements, which, in turn, can be learned from data via conditional local independence testing. The theory will be illustrated by an application to neuron spike data for multiple neurons.
[Recording] [Slides]Tuesday, April 4, 2023: Interview with Philip Dawid
- Interviewer: Vanessa Didelez (BIPS Leibniz Institute)
[Video]