[Paper Review] On the Distinction Between "Conditional Average Treatment Effects" (CATE) and "Individual Treatment Effects" (ITE) Under Ignorability Assumptions

Brian Vegetabile|arXiv (Cornell University)|Aug 10, 2021

Advanced Causal Inference Techniques14 references8 citations

TL;DR

This paper clarifies that under ignorability assumptions, methods estimating 'individual treatment effects' (ITE) actually estimate conditional average treatment effects (CATE), not true individual-level effects. It demonstrates that ITE and CATE are distinct estimands, and ignorability alone is insufficient to identify ITE, even in randomized trials, due to unobserved confounders or model misspecification.

ABSTRACT

Recent years have seen a swell in methods that focus on estimating "individual treatment effects". These methods are often focused on the estimation of heterogeneous treatment effects under ignorability assumptions. This paper hopes to draw attention to the fact that there is nothing necessarily "individual" about such effects under ignorability assumptions and isolating individual effects may require additional assumptions. Such individual effects, more often than not, are more precisely described as "conditional average treatment effects" and confusion between the two has the potential to hinder advances in personalized and individualized effect estimation.

Motivation & Objective

• To clarify the conceptual and statistical distinction between individual treatment effects (ITE) and conditional average treatment effects (CATE) under ignorability assumptions.
• To challenge the common conflation of ITE and CATE in machine learning and causal inference literature, especially in personalized medicine.
• To demonstrate that ignorability does not guarantee identification of true individual-level effects, even when conditioning on observed covariates.
• To highlight that unobserved confounders can lead to ITE estimates that differ in sign or magnitude from true individual effects.
• To advocate for greater precision in terminology and stronger assumptions when aiming for individual-level inference.

Proposed method

• Uses the Neyman-Rubin potential outcomes framework to formally define ITE and CATE as distinct estimands.
• Applies strong ignorability assumptions (i.e., A ⊥ Y(1), Y(0) | X) to show that only CATE is nonparametrically identified under these conditions.
• Constructs a counterexample with unobserved confounders (Z) to show that ITE can differ in sign from CATE, even when ignorability holds.
• Analyzes randomized controlled trials (RCTs) to show that multiple, potentially contradictory CATEs can be estimated depending on the conditioning set.
• Uses a quadratic outcome function with unobserved Z to illustrate that CATEs can have opposing trends (e.g., positive vs. negative quadratic relationships) when conditioned on different subsets of X.
• Proposes that estimating ITE requires additional assumptions beyond ignorability, such as within-individual repeated measures or case-crossover designs.

Experimental results

Research questions

• RQ1 Under what conditions can conditional average treatment effects (CATE) be identified under ignorability?
• RQ2 Why is it incorrect to equate estimated individual-level effects with CATE in observational studies?
• RQ3 How can ITE estimates differ in sign or magnitude from true individual effects even when ignorability holds?
• RQ4 Can multiple, conflicting CATEs be identified in a single RCT depending on the conditioning set?
• RQ5 What additional assumptions are required to identify true individual treatment effects (ITE) beyond ignorability?

Key findings

• Under ignorability, only CATE is nonparametrically identified; ITE is not identified without stronger assumptions.
• ITE estimates can have the opposite sign of true individual effects when unobserved confounders (Z) interact with treatment and observed covariates.
• In randomized trials, multiple CATEs can be estimated—e.g., E[Y(1)−Y(0)|X₁=x₁] and E[Y(1)−Y(0)|X₂=x₂]—with opposing functional forms (e.g., positive vs. negative quadratic).
• Even in RCTs, unobserved confounders prevent identification of ITE, showing that randomization alone does not ensure individual-level inference.
• The CATE is not equivalent to ITE, though they may be correlated; confusion between them can mislead personalized medicine applications.
• Sensitivity analyses or analytical bounds may be necessary to assess the plausible range of ITE given an estimated CATE, especially when unobserved confounding is suspected.