Causal effects conditional on post-treatment variables

Causal effects conditional on post-treatment variables

by Dr. Mats Stensrud

Formal & Physical Sciences Mathematics PBMathematicsPBTProbability and statistics
🎙 Dr. Mats Stensrud 👥 8K 📅 March 6, 2026 ⏱ 48 min 👁 720 📄 seminar 🧭 2026-08-15
Available in: English (current) Français

Keywords

causal inferencesurvivor average causal effectseparable effectspost-treatment variablescollider bias

Summary

The seminar by Dr. Mats Stensrud addresses the challenge of defining and estimating causal effects when conditioning on post-treatment variables, such as survival status. He introduces the problem with examples from oncology, COVID-19, and education, highlighting the pitfalls of naive conditioning. He discusses several causal contrasts, including the survivor average causal effect (SACE) and separable effects, which decompose treatment into components. He emphasizes the importance of practical relevance and interventionist motivations. He presents identification assumptions, including dismissible component conditions, and discusses the role of unmeasured confounding, referencing proximal inference and the birthweight paradox. The talk concludes with a discussion of the implications for future treatment design and the need for careful causal reasoning.

113 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a high-value overview of complex causal inference concepts, clearly explaining the differences between various causal estimands and their practical implications. The argumentation is solid, grounded in formal definitions and assumptions, and supported by examples from applied literature. The speaker effectively motivates the need for separable effects by linking them to real-world questions about treatment improvement, making the theoretical discussion relevant.

Scientific Rigor, Source Quality, Title Accuracy

The speaker demonstrates scientific rigor by referencing key literature, including works by Robins, Richardson, and Pearl, and by clearly stating assumptions. The sources cited are appropriate and credible. The title accurately describes the content, which focuses on causal effects conditional on post-treatment variables. The talk is well-structured and technically precise, though it assumes a certain level of familiarity with causal inference.

139 words

Title / Content Match

The title accurately reflects the content, which focuses on causal effects conditional on post-treatment variables, with detailed discussion of survivor average causal effects and separable effects.

Quality & Reliability

8/10

The talk is a rigorous seminar by a researcher from EPFL, presenting formal causal inference concepts with clear definitions, assumptions, and references to established literature. The content is technically sound and well-structured, though it is a seminar rather than a peer-reviewed publication.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

  • Rubin, D. B. (2006). Causal inference — Rubin's work on causal inference, including discussion of post-treatment variables, is referenced in the talk.
  • Robins, J. M., & Richardson, T. S. (2010). Alternative graphical causal models — The talk references the work by Robins and Richardson on separable effects and interventionist mediation.

Dissenting Sources

  • Pearl, J. (2009). Causality — Pearl's perspective on causal inference may differ on the interpretation of certain effects, as discussed in the talk.

Contribution & Novelties

The talk provides a clear exposition of causal effects conditional on post-treatment variables, particularly focusing on the survivor average causal effect and separable effects. It offers a practical motivation for these concepts, linking them to real-world questions about treatment improvement. The discussion of identification assumptions and the role of unmeasured confounding adds depth to the understanding of these effects.

Pour aller plus loin :

  • Survivor average causal effect — Overview of the concept and its applications.
  • Causal inference — General introduction to causal inference methods.
  • Collider bias — Explanation of collider bias, relevant to the discussion of conditioning on post-treatment variables.

101 words

Radar Profile

The radar profile shows high scores in information quantity, quality, and technical level, with slightly lower but still strong reliability. This indicates a technically dense and informative seminar, suitable for an audience with background in causal inference.

Reliability 8/10

💬 No comments were provided for analysis.