Article
Statistics & Probability
Erin E. Gabriel, Michael C. Sachs, Arvid Sjolander
Summary: Outcome-dependent sampling designs are common in various scientific fields. However, these designs often suffer from unmeasured confounding, which makes it difficult to identify causal effects. Nonparametric bounds can provide a solution to estimate causal effects without making untestable assumptions.
JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
(2022)
Article
Economics
Sung Jae Jun, Joris Pinkse
Summary: This paper discusses how to determine the optimal reserve price in auctions using maximum entropy estimation and proposes a maxmin decision rule that includes both the maximum entropy solution and other methods.
JOURNAL OF ECONOMETRICS
(2024)
Article
Biochemical Research Methods
Sara Mohammad-Taheri, Jeremy Zucker, Charles Tapley Hoyt, Karen Sachs, Vartika Tewari, Robert Ness, Olga Vitek
Summary: This article proposes a general and practical approach for estimating causal queries based on latent variable models. It proves the accuracy of the approach under specific conditions and demonstrates its broad applicability and practicality through synthetic and experimental case studies.
Article
Economics
Yukitoshi Matsushita, Taisuke Otsu, Keisuke Takahata
Summary: This article addresses the problem of estimating probability density ratios in causal inference and proposes a least square density ratio estimation method. The usefulness of this method is demonstrated through simulation studies and empirical examples.
JOURNAL OF BUSINESS & ECONOMIC STATISTICS
(2023)
Article
Biochemical Research Methods
Sara Mohammad-Taheri, Vartika Tewari, Rohan Kapre, Ehsan Rahiminasab, Karen Sachs, Charles Taply Hoyt, Jeremy Zucker, Olga Vitek
Summary: Causal query estimation in biomolecular networks often requires selecting a 'valid adjustment set' to eliminate bias. However, current methods based on graph criteria fail when models with the same topology have different data generation processes. To address this, we propose an approach that considers data nature, bias, variance, and cost to derive 'optimal adjustment sets' through learning from historical data and simulation. We demonstrate the utility of our approach in four biomolecular case studies with different topologies and data generation processes.
Article
Computer Science, Artificial Intelligence
Amir Feder, Nadav Oved, Uri Shalit, Roi Reichart
Summary: Understanding predictions made by deep neural networks is difficult, requiring the differentiation between correlation and causation. The proposed CausaLM framework aims to provide causal model explanations by using counterfactual language representation models.
COMPUTATIONAL LINGUISTICS
(2021)
Article
Physics, Multidisciplinary
Sergei Kalinin, Ayana Ghosh, Rama Vasudevan, Maxim Ziatdinov
Summary: The development of high-resolution imaging methods has provided rich information on the atomic structure and functionalities of solids, necessitating the adaptation of classical macroscopic definitions to understand local imaging data. This data can be used to construct statistical and physical models that generate observed structures. The availability of observational data opens pathways for exploring causal mechanisms underlying solid structure and functionality.
Article
Statistics & Probability
Nicola Gnecco, Nicolai Meinshausen, Jonas Peters, Sebastian Engelke
Summary: This study aims to connect causal inference and extreme value theory by defining the causal tail coefficient to analyze extremal dependence between two random variables. It shows that the causal tail coefficient can reveal causal structure when the distribution follows a linear structural causal model, proposing an efficient algorithm for causal structure estimation. The method consistently recovers causal order and is compared to other approaches in causal discovery on synthetic and real data, with the code available as an open-access R package.
ANNALS OF STATISTICS
(2021)
Article
Ecology
Suchinta Arif, Aaron MacNeil
Summary: Ecologists often rely on observational data to understand causal relationships, but predictive techniques are not suitable for drawing causal conclusions. Instead, valid causal inference methods such as the backdoor criterion can be used to determine causal relationships in observational studies.
Article
Multidisciplinary Sciences
Manuel Castro, Pedro Ribeiro Mendes Junior, Aurea Soriano-Vargas, Rafael de Oliveira Werneck, Maiara Moreira Goncalves, Leopoldo Lusquino Filho, Renato Moura, Marcelo Zampieri, Oscar Linares, Vitor Ferreira, Alexandre Ferreira, Alessandra Davolio, Denis Schiozer, Anderson Rocha
Summary: In this study, we propose using ensemble models (such as Random Forest) to assess the importance of input features in machine learning models, in order to establish causal relationships between variables. By analyzing oil field production data, we find that our results align with confirmed tracer information, demonstrating the effectiveness of our proposed methodology.
SCIENTIFIC REPORTS
(2023)
Article
Statistics & Probability
Michael C. Sachs, Gustav Jonzon, Arvid Sjolander, Erin E. Gabriel
Summary: A causal query is often unidentifiable from observed data, but symbolic bounds can still be derived based on the distribution of observed variables, which can be more valuable than statistical estimators derived from implausible assumptions. We develop a general approach for calculating symbolic bounds and prove its effectiveness in certain settings. Additionally, we demonstrate that our method can provide valid and potentially informative symbolic bounds in a larger range of problems.
JOURNAL OF COMPUTATIONAL AND GRAPHICAL STATISTICS
(2023)
Article
Mathematics, Interdisciplinary Applications
Arvid Sjolander, Erin E. Gabriel, Iuliana Ciocanea-Teodorescu
Summary: This article extends and applies existing methods of sensitivity analysis for residual confounding, allowing for arbitrary exposures and a wide range of causal effect measures. It demonstrates how the generalized sensitivity analysis can be easily implemented using standard software.
JOURNAL OF CAUSAL INFERENCE
(2022)
Article
Statistics & Probability
Yuqi Gu, David B. Dunson
Summary: This article introduces an identifiable multilayer discrete latent structure model, Bayesian Pyramids, for high-dimensional categorical data. The identifiability of Bayesian Pyramids is established by developing novel conditions on the deep latent directed graph. A Bayesian shrinkage estimation approach is proposed for the two-latent-layer model, and simulation results validate the identifiability and estimability of model parameters.
JOURNAL OF THE ROYAL STATISTICAL SOCIETY SERIES B-STATISTICAL METHODOLOGY
(2023)
Article
Sociology
Kenneth A. Frank, Qinyun Lin, Ran Xu, Spiro Maroulis, Anna Mueller
Summary: Social scientists must carefully consider the identification and expression of effects and inferences when informing policy or public action, as actions based on invalid inferences may not achieve desired results. To contribute to debates about causal inferences, this study reviews sensitivity analyses within omitted variables and potential outcomes frameworks and introduces the Impact Threshold for a Confounding Variable (ITCV) and the Robustness of Inference to Replacement (RIR) methodologies. These approaches are extended to incorporate benchmarks and address sampling variability and bias. Social scientists are encouraged to quantify the robustness of their inferences after utilizing the best available data and methods to draw initial causal inferences.
SOCIAL SCIENCE RESEARCH
(2023)
Article
Mathematics, Interdisciplinary Applications
Kara E. Rudolph, Nicholas T. Williams, Caleb H. Miles, Joseph Antonelli, Ivan Diaz
Summary: There has been a longstanding debate on whether nonparametric estimation using machine learning has any advantage over simpler, parametric approaches in finite sample estimation of causal effects. This study compares the performance of nonparametric and parametric estimators in estimating the average treatment effect across a large number of data-generating mechanisms. The results show that the two nonparametric estimators can significantly reduce bias compared to the parametric estimators in large-sample settings with interactions and nonlinearities, while maintaining comparable performance even in simple, small-sample settings.
JOURNAL OF CAUSAL INFERENCE
(2023)
