Why ask Why? Forward Causal Inference and Reverse Causal Questions

Andrew Gelman, Guido W. Imbens2013 · DOI

Abstract

The statistical and econometrics literature on causality is more focused on "effects of causes" than on "causes of effects." That is, in the standard approach it is natural to study the effect of a treatment, but it is not in general possible to define the causes of any particular outcome. This has led some researchers to dismiss the search for causes as "cocktail party chatter" that is outside the realm of science. We argue here that the search for causes can be understood within traditional statistical frameworks as a part of model checking and hypothesis generation. We argue that it can make sense to ask questions about the causes of effects, but the answers to these questions will be in terms of effects of causes.

Introduction

In modern statistics and econometrics, causal inference is almost synonymous with "forward causal inference"—the study of the effects of causes. Researchers ask well-defined questions like, "What is the effect of a specific drug on patient recovery?" or "How does a $100 donation impact a political candidate's vote share?" These questions are amenable to rigorous methodology, such as randomized controlled trials or instrumental variable designs.

However, much of human curiosity and scientific inquiry begins with the opposite: reverse causal questions, or the search for the causes of effects. We observe an anomaly—a cancer cluster in a specific town, a sudden drop in corporate sales, or a persistent wealth gap between nations—and ask "Why?" Historically, many statisticians have dismissed these queries as unscientific because they are often ill-posed; there is rarely a single, unique cause for a complex outcome. In this paper, Andrew Gelman and Guido Imbens argue that reverse causal inference is not only scientific but is the essential engine of discovery that drives the forward-looking models we value so highly.

The Forward-Reverse Dichotomy

The authors distinguish between two modes of thinking:

  1. Forward Causal Inference (Estimation): This is the estimation of the effect of a known treatment. It is the "What happens if...?" of science. It yields precise numbers, confidence intervals, and p-values.
  2. Reverse Causal Inference (Discovery): This is the search for potential causes of an observed effect. It is the "Why did this happen?" of science. It is exploratory, qualitative, and often subjective, but it identifies the variables that eventually become the subjects of forward inference.

While forward inference is about getting the right answer to a specific question, reverse inference is about finding the right questions to ask in the first place.

Reverse Causal Inference as Model Checking

The core contribution of the paper is a conceptual framework that bridges these two modes using the Potential Outcomes Framework (the Neyman-Rubin model). The authors argue that a "Why" question is essentially a realization that one's current model of the world is inadequate. They frame the process as a three-step cycle of model checking:

  1. The Anomaly: A researcher observes a pattern in the data (an outcome $Y$ associated with an attribute $Z$) that their current model cannot explain. For example, noticing that incumbents raise significantly more money than challengers.
  2. The "Why" Question: This is the formal recognition of a model failure. The researcher asks why $Y$ and $Z$ are correlated, implying that the current set of control variables $W$ is insufficient.
  3. The Resolution: The researcher searches for a new variable $V$ (a potential cause or omitted attribute). If, by including $V$ in a new forward model, the original anomaly is explained away, the "Why" question has been answered.

In this framework, an anomaly is a violation of conditional independence: $Y \not\perp Z | W$. The goal of reverse causal inference is to find a variable $V$ such that $Y \perp Z | W, V$. By framing "Why" as "What is wrong with my model?", the authors turn an informal query into a standard statistical task.

The Relativity of Anomalies

A key insight of the paper is that "Why" questions do not exist in a vacuum; they are relative to a baseline model (even an implicit one). To ask "Why is the sky blue?" is to compare the actual sky to an implicit model where the atmosphere does not scatter light.

This relativity explains why reverse causal questions often have multiple valid answers. For example, if we ask "Why do taller people earn more?", one researcher might point to "better childhood nutrition" (a manipulable cause), while another might point to "self-confidence as a proxy for height" (an omitted attribute). Both can be correct because they are addressing different gaps in different baseline models.

Case Studies in Discovery

The authors illustrate their framework using several classic examples:

  • Cancer Clusters: Using mortality maps to identify geographical anomalies. The "Why" leads researchers to investigate local environmental factors or demographic shifts.
  • Political Campaigns: Analyzing the funding gap between incumbents and challengers. This leads to forward-looking studies on the effects of name recognition or committee assignments.
  • Corporate Metrics: Citing statistician Kaiser Fung, the authors describe how sudden drops in sales at companies like P&G trigger a search for causes (e.g., a competitor's new campaign), which are then tested via forward-looking market analysis.

Conclusion and Future Directions

Gelman and Imbens conclude that "causes of effects" and "effects of causes" are not competing theories but different stages of the same scientific process. Reverse causal inference identifies the anomalies that motivate the experiments used in forward causal inference.

While reverse inference remains inherently exploratory and lacks the "neat" point estimates of forward models, it is the necessary precursor to rigorous science. The authors suggest that future research should focus on improving tools for graphical exploratory data analysis and automated model checking, helping researchers systematically identify the anomalies that trigger the most productive "Why" questions.

Summary updated

Abstract

The statistical and econometrics literature on causality is more focused on "effects of causes" than on "causes of effects." That is, in the standard approach it is natural to study the effect of a treatment, but it is not in general possible to define the causes of any particular outcome. This has led some researchers to dismiss the search for causes as "cocktail party chatter" that is outside the realm of science. We argue here that the search for causes can be understood within traditional statistical frameworks as a part of model checking and hypothesis generation. We argue that it can make sense to ask questions about the causes of effects, but the answers to these questions will be in terms of effects of causes.
BibTeX
@misc{gelman2013why,
  author = {Andrew Gelman and Guido W. Imbens},
  title = {{Why ask Why? Forward Causal Inference and Reverse Causal Questions}},
  year = {2013},
  doi = {10.3386/w19614},
}
Figure