Valid Inference for Treatment Effects under Multimodal Confounding

Martin Spindler, Philipp Bach, Victor Chernozhukov, Sven Klaassen, Jan Teichert-Kluge, Suhas Vijaykumar

Fifth Conference on Causal Learning and Reasoning (CLeaR 2026)

We develop a neural network architecture for the partially linear model in the double machine learning framework, with text and images as confounders, and we evaluate it on a semi-synthetic dataset with multimodal confounding.

Earlier version DoubleMLDeep: Estimation of Causal Effects with Multimodal Data (2024)

OpenReview

Behavioral Dynamics in Negotiations: How Specific Verbal Behaviors Relate to Negotiation Agreements and Negotiation Processes

Jan Niklas Engel, Jan Teichert-Kluge, Clara Sofie Hemshorn de Sánchez, Nale Lehmann-Willenbrock

Negotiation and Conflict Management Research 19 (2)

We model negotiations as complex adaptive systems and examine which verbal behaviors during a negotiation are associated with agreement or non-agreement.

DOI

Behavioral Dynamics in Negotiations: How Specific Verbal Behaviors Shape Outcomes and Processes

Jan Niklas Engel, Clara Sofie Hemshorn de Sánchez, Jan Teichert-Kluge

Academy of Management Proceedings 2025 (1), 12834

Using machine learning on 22,880 coded verbal behaviors from 40 face-to-face negotiation trainings, we identify the behaviors associated with agreement or impasse.

DOI

Pragmatic DML with AI-Learned Representations

Andres Aradillas Fernandez, Victor Chernozhukov, Carlos Cinelli, Sven Klaassen, Whitney K. Newey, Martin Spindler, Jan Teichert-Kluge, Suhas Vijaykumar

arXiv preprint

We study when AI-learned representations of text and images can be used as controls in causal analysis. With an imperfect representation, the target parameter is distorted by the product of two representation errors, and inference based on cross-fitted double machine learning is valid for the target defined by the representation.

arXiv

Adventures in Demand Analysis Using AI

Philipp Bach, Victor Chernozhukov, Sven Klaassen, Martin Spindler, Jan Teichert-Kluge, Suhas Vijaykumar

arXiv preprint · cemmap working paper

Using data on toy cars sold on Amazon, we represent each product by transformer-based embeddings that combine its text description and images with tabular covariates. With these embeddings, predictions of sales ranks and prices are more accurate and estimates of price elasticities are more credible.

arXivcemmap

DoubleMLDeep: Estimation of Causal Effects with Multimodal Data

Sven Klaassen, Jan Teichert-Kluge, Philipp Bach, Victor Chernozhukov, Martin Spindler, Suhas Vijaykumar

arXiv preprint

We propose a neural network architecture for double machine learning with text and images as confounders and a method to generate semi-synthetic data for evaluating estimates of causal effects under multimodal confounding.

Conference version Valid Inference for Treatment Effects under Multimodal Confounding (CLeaR 2026)

arXiv

DoubleMLDeep: Estimation of Causal Effects with Multimodal Data

Econometric Society ESIF Economics and AI+ML Meeting, Cornell University

Estimating Price Elasticities with Text and Images

Computational Economics Workshop, Australian National University

Estimating Price Elasticities with Text and Images

Econometrics Colloquium, University of Queensland

DoubleMLDeep: Estimation of Causal Effects with Multimodal Data

ISMS Marketing Science Conference (invited)

Price Elasticity Estimation Using Image and Text Data

Causal Data Science Meeting

DoubleML

Contributor to the open-source package for double machine learning in Python and R

DMLSim

Library for simulations and case studies with DoubleML and neural network learners

Causal Machine Learning with Deep Learning Approaches Using the DoubleML Framework

Moritz Sundermann, Jan Teichert-Kluge

Master’s thesis, University of Hamburg

PDF