Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Position: The Causal Revolution Needs Scientific Pragmatism

Joshua Loftus

TL;DR

The paper addresses stagnation in the causal revolution by contrasting scientific perfectionism with an algorithmically oriented ML culture that favors automation. It argues for scientific pragmatism, treating causal models as tools for hypothetical reasoning and distinguishing $X \rightarrow Y$ causality from purely predictive relations, with interventions via $do(X)$ when appropriate. It further endorses value pluralism and human-centered design to justify using causal methods across diverse goals beyond predictive accuracy, emphasizing transparency about assumptions. Together, these ideas form a pragmatic framework to standardize causal explanations, broaden applicability in science and policy, and align progress with human-centered values.

Abstract

Causal models and methods have great promise, but their progress has been stalled. Proposals using causality get squeezed between two opposing worldviews. Scientific perfectionism--an insistence on only using "correct" models--slows the adoption of causal methods in knowledge generating applications. Pushing in the opposite direction, the academic discipline of computer science prefers algorithms with no or few assumptions, and technologies based on automation and scalability are often selected for economic and business applications. We argue that these system-centric inductive biases should be replaced with a human-centric philosophy we refer to as scientific pragmatism. The machine learning community must strike the right balance to make space for the causal revolution to prosper.

Position: The Causal Revolution Needs Scientific Pragmatism

TL;DR

The paper addresses stagnation in the causal revolution by contrasting scientific perfectionism with an algorithmically oriented ML culture that favors automation. It argues for scientific pragmatism, treating causal models as tools for hypothetical reasoning and distinguishing causality from purely predictive relations, with interventions via when appropriate. It further endorses value pluralism and human-centered design to justify using causal methods across diverse goals beyond predictive accuracy, emphasizing transparency about assumptions. Together, these ideas form a pragmatic framework to standardize causal explanations, broaden applicability in science and policy, and align progress with human-centered values.

Abstract

Causal models and methods have great promise, but their progress has been stalled. Proposals using causality get squeezed between two opposing worldviews. Scientific perfectionism--an insistence on only using "correct" models--slows the adoption of causal methods in knowledge generating applications. Pushing in the opposite direction, the academic discipline of computer science prefers algorithms with no or few assumptions, and technologies based on automation and scalability are often selected for economic and business applications. We argue that these system-centric inductive biases should be replaced with a human-centric philosophy we refer to as scientific pragmatism. The machine learning community must strike the right balance to make space for the causal revolution to prosper.
Paper Structure (12 sections, 1 figure)

This paper contains 12 sections, 1 figure.

Table of Contents

  1. Introduction
  2. Published Context
  3. Positions
  4. Speculation.
  5. Principles
  6. Scientific Pragmatism
  7. Value Pluralism
  8. Humanism or Anthropocentrism
  9. Conclusion
  10. Related Ideas
  11. Significance
  12. Summary

Figures (1)

  • Figure 1: Structural causal models (SCMs) represented as directed acyclic graphs (DAGs). Variables are the nodes of the graph and causal effects are represented by arrows. An intervention is an operation that modifies the graph in some way. Intervening on a variable means we erase the arrows pointing into that variable, set the value of that variable arbitrarily, and then propagate the new value along directed pathways pointing out of that variable. On the left, $X$ is a cause of $Y$, so intervening on $X$ will result in a change in $Y$. In the other cases, intervening on $X$ results in no change in $Y$. It is possible that the accuracy of some function $f(X)$ in predicting $Y$ is equal in all cases.