Does the Research Question Contain a Hidden Causal Assumption?

Is Your Question Asking About Causation Without Saying So?

Consider the question: “How does students' use of generative AI affect their critical-thinking skills?” It may initially look like a straightforward question about two variables. Yet “affect” does important methodological work. The question is not merely asking whether AI use and critical thinking occur together. It asks whether changing AI use would change critical thinking.

Causal assumptions often enter research questions through ordinary words such as “effect,” “impact,” “influence,” “improve,” “reduce,” “increase,” “lead to,” and “result in.” Researchers may use these terms conversationally while intending only to examine an association. The eventual study may then be designed to estimate a relationship even though the question asks for something stronger.

That distinction matters because descriptive, predictive, and causal questions target different kinds of knowledge. A statistical association can be scientifically valuable, but it does not automatically answer a causal question.

Causal Language Commits You to a Causal Question

How to Recognize the Causal Claim Hidden Inside a Question

Methodologists commonly distinguish descriptive, predictive, and causal research questions. A descriptive question characterizes what is observed. A predictive question asks what can be forecast from available information. A causal question asks how an outcome would differ if some exposure or condition were changed. This last type involves a counterfactual comparison: what would happen under one condition compared with what would happen under another condition for the relevant population.

The distinction is not merely semantic. The scientific question should guide the design and analysis, and causal terminology should correspond to a study intended and structured for causal inference.

Look Beyond the Word “Cause”

A question does not need to contain “cause” to be causal. Consider these formulations:

• What is the effect of formative feedback on examination performance?
• Does generative AI improve student writing?
• How does remote work influence employee productivity?
• Does social media use increase academic procrastination?
• What is the impact of simulation training on clinical competence?

Each asks, or can reasonably be read as asking, whether a change in one condition produces a change in another. Words such as “effect,” “improve,” “influence,” “increase,” and “impact” should therefore prompt a causal check.

Ask the Intervention Test

One practical way to expose a hidden causal question is to imagine intervening on the exposure.

Suppose your question is: “Does frequent use of generative AI reduce students' independent problem-solving ability?” Ask what you really want to know. Is it simply whether students who use AI more frequently tend to have different problem-solving scores? Or do you want to know whether the same relevant population would perform differently if its AI use were changed?

The second interpretation is causal. Causal-inference literature characterizes such questions through counterfactual contrasts: outcomes under one exposure or intervention condition are compared conceptually with outcomes under an alternative condition.

The questions may involve the same variables, but they do not ask for the same conclusion.

A Comparison Between Existing Groups Is Not Automatically a Causal Comparison

Suppose students who use an AI tutor obtain higher examination scores than students who do not. That observed difference does not by itself tell us what would have happened to the AI-using students had they not used the tutor.

The two groups may differ in prior achievement, motivation, instructor support, digital competence, socioeconomic resources, study habits, or other factors related to the outcome. Causal inference asks about outcomes under alternative exposure conditions, not merely differences between people who happened to experience different conditions.

This is why a causal question should also be examined for whether it assumes a comparison capable of addressing the intended contrast.

Temporal Order Matters, but It Is Not Enough

For X to cause Y, X must occur before the relevant change in Y. A study measuring both variables at approximately the same time may therefore struggle to establish the temporal sequence implied by some causal questions.

Yet showing that X came first does not establish causation on its own. Other explanations for the observed relationship still need consideration. Longitudinal data can strengthen the temporal information available, but longitudinal does not mean automatically causal.

Confounding Becomes Relevant Because the Question Is Causal

Confounding is fundamentally a causal concept. When the scientific aim is to estimate a causal effect, researchers need to consider variables and structures that could produce differences between the observed exposure groups while also relating to the outcome.

Merely adding many variables to a regression model is not equivalent to solving this problem. Which variables should be adjusted for depends on the causal structure and the estimand of interest. Contemporary causal-inference approaches therefore encourage researchers to define the causal question and theoretical estimand before selecting a statistical estimator.

A Randomized Experiment Is Not the Definition of a Causal Question

It is tempting to treat “experimental” and “causal” as synonyms, but they describe different things. Causality concerns the scientific question and inference. Randomization is a powerful design strategy for supporting causal inference because, when successfully implemented, it can make treatment groups comparable with respect to potential outcomes in expectation.

Some causal questions cannot ethically or practically be studied through randomized experiments. Observational studies may still be designed explicitly for causal inference, but they require careful specification of the causal contrast and assumptions that connect observed data to the desired causal effect.

Statistical Significance Does Not Turn an Association Into an Effect

A small p-value for the association between X and Y does not establish that changing X would change Y. Statistical significance addresses uncertainty under a specified statistical model and null hypothesis. It does not determine whether the estimated relationship has a causal interpretation.

Likewise, a large regression coefficient, strong correlation, sophisticated machine-learning model, or very large dataset cannot independently settle the causal question. The inferential interpretation depends on how the question, design, measurements, comparisons, and assumptions fit together.

The Question Should Identify the Causal Contrast Clearly Enough to Design Around It

“Does technology affect learning?” may express causal intent, but it is too underspecified to guide a convincing causal study. What technology? Used how? Compared with what alternative? For whom? What aspect of learning? Over what period?

Frameworks such as PICO can help specify causal-comparative questions by identifying the population, intervention or exposure, comparison, and outcome. The broader principle applies beyond clinical research: a causal question becomes more useful when the alternative conditions being compared are sufficiently explicit.

Do Not Remove Causal Language Merely to Avoid the Methodological Problem

If your substantive question is genuinely causal, changing “effect” to “association” does not solve it. That simply changes the question.

The correct response is to decide which question you actually care about. If you want an associational answer, frame and interpret the study accordingly. If you want a causal answer, retain the causal question and determine whether a defensible design can address it. If not, the project may need a different design, a more modest causal target, or a different question.

One Dataset, Two Very Different Questions

The decision is substantive, not cosmetic. “Is associated with” and “causes” should not be treated as interchangeable wording for the same research question.

Common Mistakes When Asking Causal Questions

Decide What Kind of Answer You Actually Want

Take the central relationship in your question and ask: “Do I merely want to know whether X and Y differ or occur together, or do I want to know what would happen to Y if X were changed?”

If the latter is your real scientific objective, acknowledge that you are asking a causal question. Then make the causal contrast explicit enough that the study can be designed around it.

This check should occur while you stress-test the research question before designing the study, not after the analysis has already been completed.

Check Whether Your Research Question Is Causal

Questions About Causal Language in Research Questions

Know Whether You Are Asking What Happens or What Would Happen

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.