Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Personalised Decision-Making without Counterfactuals

A. Philip Dawid, Stephen Senn

TL;DR

The paper challenges Mueller and Pearl's counterfactual, personalised decision-making framework by contrasting it with the traditional decision-theoretic approach. It formalises the MP perspective using potential outcomes, probabilities of benefit ${\rm PB}$ and harm ${\rm PH}$, and shows that exact identification is typically unattainable, with inference relying on interval bounds that can be narrowed under strong, often impractical assumptions. Through a detailed analysis, including the role of covariates, the concept of a preferred treatment (PT), and data fusion from experiments and observations, the authors demonstrate that, in general, decision theory yields clearer, more ethically defensible guidance and can, in special cases, match MP when PT is observed. They advocate exploiting PT within a decision-theoretic framework as a practical, robust alternative, cautioning against adopting counterfactual-based methods in routine practice. The work emphasizes realism about representativeness and the limits of data fusion, urging humility and careful consideration of clinical applicability.

Abstract

This article is a response to recent proposals by Pearl and others for a new approach to personalised treatment decisions, in contrast to the traditional one based on statistical decision theory. We argue that this approach is dangerously misguided and should not be used in practice.

Personalised Decision-Making without Counterfactuals

TL;DR

The paper challenges Mueller and Pearl's counterfactual, personalised decision-making framework by contrasting it with the traditional decision-theoretic approach. It formalises the MP perspective using potential outcomes, probabilities of benefit and harm , and shows that exact identification is typically unattainable, with inference relying on interval bounds that can be narrowed under strong, often impractical assumptions. Through a detailed analysis, including the role of covariates, the concept of a preferred treatment (PT), and data fusion from experiments and observations, the authors demonstrate that, in general, decision theory yields clearer, more ethically defensible guidance and can, in special cases, match MP when PT is observed. They advocate exploiting PT within a decision-theoretic framework as a practical, robust alternative, cautioning against adopting counterfactual-based methods in routine practice. The work emphasizes realism about representativeness and the limits of data fusion, urging humility and careful consideration of clinical applicability.

Abstract

This article is a response to recent proposals by Pearl and others for a new approach to personalised treatment decisions, in contrast to the traditional one based on statistical decision theory. We argue that this approach is dangerously misguided and should not be used in practice.
Paper Structure (35 sections, 2 theorems, 28 equations, 4 tables)

This paper contains 35 sections, 2 theorems, 28 equations, 4 tables.

Table of Contents

  1. Introduction
  2. Basic set-up
  3. Outline
  4. The approach of Mueller and Pearl
  5. Potential outcomes
  6. MP approach
  7. Decision-theoretic approach
  8. Single patient decision problem
  9. Unit selection
  10. Missing information
  11. Relation with MP
  12. Analysis
  13. Simplest case
  14. Benefit and harm
  15. Biased data
  16. ...and 20 more sections

Key Result

Theorem 3

Suppose we can identify the joint distribution of $X$ and $Y$ in the observational context, where $0<\Pr(X=1)<1$, and can also identify the distribution of $Y$ under either intervention $X \leftarrow x$$(x=0,1)$. Then, under conditions (eq:xint) and (eq:distcons), all the probabilities $\Pr(Y=y \mid

Theorems & Definitions (9)

  • Example 1
  • Remark 1
  • Remark 2
  • Theorem 3
  • Corollary 1
  • Remark 4
  • Example 2
  • Example 3
  • Example 4