Elias Bareinboim

A General Algorithm for Deciding Transportability of Experimental Results

Abstract Generalizing empirical findings to new environments, settings, or populations is essential in most scientific explorations. This article treats a particular problem of generalizability, called “transportability”, defined as a license to transfer information learned in experimental studies to a different population, on which only observational studies can be conducted. Given a set of assumptions concerning commonalities and differences between the two populations, Pearl and Bareinboim [1] derived sufficient conditions that permit such transfer to take place. This article summarizes their findings and supplements them with an effective procedure for deciding when and how transportability is feasible. It establishes a necessary and sufficient condition for deciding when causal effects in the target population are estimable from both the statistical information available and the causal information transferred from the experiments. The article further provides a complete algorithm for computing the transport formula, that is, a way of combining observational and experimental information to synthesize bias-free estimate of the desired causal relation. Finally, the article examines the differences between transportability and other variants of generalizability.

Causal inference and the data-fusion problem

We review concepts, principles, and tools that unify current approaches to causal analysis and attend to new challenges presented by big data. In particular, we address the problem of data fusion—piecing together multiple datasets collected under heterogeneous conditions (i.e., different populations, regimes, and sampling methods) to obtain valid answers to queries of interest. The availability of multiple heterogeneous datasets presents new opportunities to big data analysts, because the knowledge that can be acquired from combined data would not be possible from any individual source alone. However, the biases that emerge in heterogeneous environments require new analytical tools. Some of these biases, including confounding, sampling selection, and cross-population biases, have been addressed in isolation, largely in restricted parametric models. We here present a general, nonparametric framework for handling these biases and, ultimately, a theoretical solution to the problem of data fusion in causal inference tasks.

Analyzing marginal cases in differential shotgun proteomics

SUMMARY We present an approach to statistically pinpoint differentially expressed proteins that have quantitation values near the quantitation threshold and are not identified in all replicates (marginal cases). Our method uses a Bayesian strategy to combine parametric statistics with an empirical distribution built from the reproducibility quality of the technical replicates. AVAILABILITY The software is freely available for academic use at http://pcarvalho.com/patternlab.

Descents and nodal load in scale-free networks

The load of a node in a network is the total traffic going through it when every node pair sustains a uniform bidirectional traffic between them on shortest paths. We express nodal load in terms of the more elementary notion of a nodes descents in breadth-first-search [(BFS) or shortest-path] trees and study both the descent and nodal-load distributions in the case of scale-free networks. Our treatment is both semianalytical (combining a generating-function formalism with simulation-derived BFS branching probabilities) and computational for the descent distribution; it is exclusively computational in the case of the load distribution. Our main result is that the load distribution, even though it can be disguised as a power law through subtle (but inappropriate) binning of the raw data, is in fact a succession of sharply delineated probability peaks, each of which can be clearly interpreted as a function of the underlying BFS descents. This find is in stark contrast with previously held belief, based on which a power law of exponent -2.2 was conjectured to be valid regardless of the exponent of the power-law distribution of node degrees.

A Graphical Criterion for Effect Identification in Equivalence Classes of Causal Diagrams

Computing the effects of interventions from observational data is an important task encountered in many data-driven sciences. The problem is addressed by identifying the post-interventional distribution with an expression that involves only quantities estimable from the pre-interventional distribution over observed variables, given some knowledge about the causal structure. In this work, we relax the requirement of having a fully specified causal structure and study the identifiability of effects with a singleton intervention (X), supposing that the structure is known only up to an equivalence class of causal diagrams, which is the output of standard structural learning algorithms (e.g., FCI). We derive a necessary and sufficient graphical criterion for the identifiability of the effect of X on all observed variables. We further establish a sufficient graphical criterion to identify the effect of X on a subset of the observed variables, and prove that it is strictly more powerful than the current state-of-the-art result on this problem.

Transfer Learning in Multi-Armed Bandits: A Causal Approach.

We leverage causal inference tools to support a principled and more robust transfer of knowledge in reinforcement learning (RL) settings. In particular, we tackle the problem of transferring knowledge across bandit agents in settings where causal effects cannot be identified by Pearls {do-calculus} nor standard off-policy learning techniques. Our new identification strategy combines two steps -- first, deriving bounds over the arms distribution based on structural knowledge; second, incorporating these bounds in a novel bandit algorithm, B-kl-UCB. Simulations demonstrate that our strategy is consistently more efficient than the current (non-causal) state-of-the-art methods.

Guest editorial: special issue on causal discovery

In the big data era, data have become an important source of automatically generated knowledge which, in turn, is able to help to change the system to achieve a certain objective, to make predictions under interventions, or to assist decision making in various fields. To this end, it is crucial that the knowledge encodes causal information. As a consequence, more and more data mining researchers are interested in causal discovery. In acknowledging the increasing interest in causal discovery in the data mining community, and to facilitate further development of this research area, in August 2016 we organized the Workshop on Causal Discovery with ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD 2016). The aim of the workshop is to provide a forum for data mining researchers and researchers in causal modeling and inference to communicate, understand each others’ problems, and to create stronger synergy between the research communities to solve real, large-scale problems. This KDD workshop follows the success of the Causal Discovery Workshop held in conjunction

Preface to the ACM TIST Special Issue on Causal Discovery and Inference

Causality plays an important role in explanation, prediction, policy making, and control in many fields of empirical sciences, medicine, and engineering. Traditionally, causal relationships are estimated using randomized controlled experiments. However, conducting such experiments usually is expensive or even impossible in many real-world scenarios. Therefore, there has been a growing interest in principles and tools for inferring cause-and-effect relationships from passive observations or partially controlled experiments. These efforts have led to significant progress in various fields in past decades, including computer science, statistics, and philosophy. Reasoning with causal relationships involves both deductive and inductive tasks. The deductive component asks what can be inferred when the researcher is in possession of certain knowledge (usually in the form of a causal graph or features thereof). The inductive component asks how aspects of the underlying causal process can be inferred from data when the researcher is willing to make only generic assumptions about the generative process (e.g., causal faithfulness, linearity). These are complementary and strongly intertwined tasks. Recently, with the rapid accumulation of a huge volume of data, the field of causal discovery and inference is seeing exciting opportunities as well as greater challenges. This special issue aims at reporting the progress in practical methodologies, efficient implementations, and applications of causal discovery and inference. The first part of this special issue consists of three articles related to the problem of estimating causal effects in specific subgroups of a given population, from a combination of data and assumptions about the underlying data-generating model. “Sharp Bounds on Survivor Average Causal Effects When the Outcome Is Binary and Truncated by Death,” by N. Shan, X. Dong, P. Xu, and J. Guo, derives nonparametric bounds on the survivor average causal effect (SACE), which is the effect in the subgroup of individuals who would have recovered independently of the treatment. The article uses linear programming techniques to derive bounds for this quantity. “Semiparametric Inference of Complier Average Causal Effect with Nonignorable Missing Outcomes,” by H. Chen, P. Ding, Z. Geng, and X. Zhou, considers another quantity of interest, namely the complier average causal effect (CACE). It is the effect for those who would take the treatment if and only if assigned treatment. This article extends previous work by Chen et al. [2009] and Imai [2009] for computing the CACE with missing and continuous responses. “Bounds on Direct and Indirect Effects of Treatment on a Continuous Endpoint,” by P. Luo and Z. Geng, derives bounds on the controlled direct effect, natural direct effect,

Generalizing causal knowledge: theory and algorithms

This article is a short summary of the full dissertation thesis that was defended in 2014 at the University of California, Los Angeles.