Causal Inference Is Not Just a Statistics Problem

assumptions
causal inference
epidemiology
statistical methods
study design
  • Core Argument: This article argues that causal inference relies primarily on study design and domain knowledge to establish untestable assumptions (like exchangeability and no unmeasured confounding), with statistical methods serving only to quantify the effect under those assumptions.
  • Design Trumps Analysis: The authors emphasize that a causal estimate is invalid if the study design fails to account for critical (especially unmeasured) confounders, regardless of the sophistication of the statistical modeling used.
  • The Role of Causal Diagrams: Identifying which variables to control is a causal question, not a statistical one, necessitating the use of Directed Acyclic Graphs (DAGs) and scientific expertise to avoid introducing bias by controlling for mediators or colliders.
  • Conclusion: Teaching and practicing causal inference requires prioritizing the formulation of a causal question and the design of the study before the application of statistical tools.
Published

23 January 2026

PubMed: Not Indexed (Journal of Statistics and Data Science Education) DOI: 10.1080/26939169.2023.2276446 Overview generated by: Gemini 2.5 Flash, 26/11/2025

Key Findings

This article argues that causal inference—the process of determining whether an exposure causes an outcome—is a challenge rooted primarily in study design and conceptual modeling, not just statistical analysis. While statistical methods are necessary for quantifying effects, they cannot salvage a poorly designed study or validate a flawed causal hypothesis.

Causal Inference is Not Statistical Regression

The authors emphasize the crucial distinction between prediction/association (a purely statistical exercise) and causal inference (a scientific exercise):

  1. Causality Requires Assumptions: Unlike association, causality requires making strong, often untestable assumptions about the data-generating process. These assumptions include positivity (everyone had a chance to receive the treatment), consistency (the treatment is well-defined), and exchangeability (the treated and untreated groups are comparable, usually requiring control for all confounders).
  2. Statistics Quantifies, Design Ensures Validity: Statistical methods (like regression, matching, or propensity scores) can only adjust for observed confounding variables under the assumption that all necessary confounders have been identified and measured without error. If a critical confounder is unmeasured (unmeasured confounding), the resulting causal estimate is likely biased, regardless of the sophistication of the statistics used.

The Primacy of Study Design

The article strongly aligns with the philosophy that “Design Trumps Analysis” (a concept attributed to Donald Rubin).

  • Need for Domain Knowledge: The process of identifying the correct causal model and the necessary variables to control (confounders) is a non-statistical process that relies entirely on domain-specific scientific knowledge (e.g., biology, epidemiology, medicine).
  • The “Which Variables to Control?” Problem: The decision of which covariates to include in a regression model is a causal question, not a statistical one. Including a variable that is actually a collider or a mediator can introduce bias where none existed, an error statisticians cannot prevent without external scientific guidance.
  • Randomized Controlled Trials (RCTs): RCTs are the gold standard for causal inference precisely because they use a design (randomization) to satisfy the assumption of exchangeability (balance all confounders, measured and unmeasured), thereby circumventing the statistical problem of controlling for confounders.

Conclusion and Education Focus

The conclusion stresses that teaching causal inference must go beyond simply running statistical models. Students must be trained to: * Formulate a Causal Question first. * Diagram the Causal Structure using tools like Directed Acyclic Graphs (DAGs). * Identify the Sources of Bias (confounding, selection bias, measurement error) based on their domain knowledge. * Choose a Design (experimental or observational) that minimizes these biases. * Use Statistics only to quantify the effect size within the context of the chosen design and stated causal assumptions.