03 · What You Need to Know
Not Every Explanation Is a Causal Effect
Researchers often receive the advice to avoid “why” questions unless they are conducting an experiment. That advice is understandable because “why” can invite causal claims. Taken literally, however, it is too restrictive.
Qualitative methodological guidance explicitly recognizes explanatory questions and notes that qualitative inquiry may ask “how” and “why” in order to develop an in-depth understanding or explanation of a phenomenon. Qualitative research can investigate people's reasons, social processes, contextual conditions, meanings, and perceived influences without estimating a causal effect in the statistical or counterfactual sense.
The real issue is therefore not the word itself. It is what you intend “why” to mean and what claim you intend to make from the answer.
“Why” can ask for someone's reasons
Consider:
“Why do some students use generative AI when completing assignments even when their instructors discourage it?”
An interview study could investigate students' accounts of their decisions. Participants might discuss time pressure, uncertainty about expectations, perceptions of fairness, difficulty with the assignment, previous experiences, or beliefs about what counts as acceptable assistance.
The study could credibly analyze these as reported reasons, meanings, motivations, or considerations involved in decision-making.
That does not automatically demonstrate that any one of those factors caused AI use across the wider student population. What participants say influenced their behavior and what would happen to behavior if one factor were changed are different questions.
Reason or perceived explanation
What participants say motivated, influenced, enabled, constrained, or shaped their actions or experiences.
Causal effect
What difference in an outcome would result from changing an exposure, intervention, or condition, under a specified causal interpretation.
“Why” can ask how a process came about
Some why questions are explanatory without reducing the answer to a single variable causing an outcome.
Suppose a university introduces a new academic-integrity policy, yet faculty members apply it inconsistently. A researcher might ask why implementation differs among departments.
A qualitative case study could examine policy documents, interviews, meetings, institutional practices, departmental norms, and local interpretations. The resulting explanation might show how ambiguity in the policy interacts with disciplinary expectations, workload, leadership practices, and previous experiences.
Such an account can provide a substantive explanation of how the observed situation developed. Qualitative research guidance recognizes explanatory questions as a legitimate form of inquiry and qualitative case studies as capable of addressing how and why phenomena occur within context.
Again, the researcher should not automatically translate that contextual explanation into a quantified causal effect.
“Why” can also ask a genuinely causal question
Now consider:
“Why does chronic sleep deprivation reduce academic performance?”
This wording appears to presuppose that sleep deprivation reduces academic performance and asks for an explanation of that causal relationship. Depending on the intended answer, the study might need evidence about whether the effect exists, mechanisms through which it occurs, or both.
A causal research question makes a stronger evidentiary demand than a descriptive or associational one. Contemporary causal-inference literature distinguishes questions about the world as observed from questions about what would happen under alternative exposure conditions. Causal questions concern outcomes under such counterfactual conditions, not merely whether exposed and unexposed groups happen to differ.
This distinction becomes important whenever your “why” question is really asking: Would Y have been different if X had been different?
Observational does not automatically mean noncausal
There is an important nuance here. It is also too simplistic to say that only randomized experiments can address causal questions.
Modern causal-inference methods can be applied to observational data when researchers explicitly formulate a causal question, define the causal quantity of interest, use a design and analytical strategy appropriate to that question, and state and defend the assumptions required for causal interpretation. Recent methodological guidance specifically addresses causal inference from observational studies rather than treating observational research as inherently incapable of causal analysis.
Randomization has major advantages because, when successfully implemented, it helps create comparability between treatment conditions. But inability to randomize does not automatically transform every scientifically causal question into a purely associational one.
Watch Out
Do not infer that ordinary observational analysis becomes causal merely because sophisticated statistical methods are used. Causal interpretation requires an explicitly causal question, an appropriate design and estimand, defensible assumptions about issues such as confounding and selection, and analysis aligned with those assumptions.
A cross-sectional association usually cannot answer “what caused this?” by itself
Suppose a one-time survey finds that students reporting heavier social media use also report more depressive symptoms.
The data may support a statement that social media use and depressive symptoms were associated in the observed sample, assuming the analysis is appropriate. But several causal explanations remain possible. Social media use might affect depressive symptoms. Depressive symptoms might affect social media use. Other factors might influence both. Measurement or selection processes could also contribute to the observed association.
Simply asking participants about both variables at one point in time does not resolve those possibilities.
Methodological guidance on observational research therefore recommends clearly distinguishing descriptive or associational questions from causal-inference questions and using terminology consistent with the intended inference.
Participants' explanations are evidence about their perspectives
If participants say, “I used AI because I had three assignments due that week,” that statement is evidence that the participant identifies workload as a reason for using AI.
It is not, by itself, evidence that reducing assignment workload would causally reduce AI use across the student population.
This distinction can feel overly cautious until you consider cases in which people's explanations of their own behavior are incomplete, retrospective, socially desirable, or shaped by information unavailable to them. Self-reported reasons are valuable evidence for understanding perceptions and decision-making. Their evidentiary meaning should simply be stated accurately.
Qualitative research can go beyond collecting isolated opinions by examining patterns across accounts, comparing cases, considering context, drawing on multiple sources, and developing theoretically informed explanations. Its strength lies in the depth and contextualization of that analysis, not in pretending that an interview is a randomized intervention wearing a name badge.
Mechanism questions require their own evidentiary logic
Another form of “why” asks about mechanism:
“Why might retrieval practice improve long-term retention?”
This question goes beyond asking whether retrieval practice has an effect. It asks what processes could produce the effect.
Evidence that an intervention works does not automatically establish why it works. Conversely, evidence consistent with a proposed mechanism does not necessarily establish the total causal effect of the intervention.
Researchers should therefore distinguish questions about whether X causes Y from questions about how or through what mechanisms X may produce Y. Depending on the field, answering a mechanism question might require experiments, mediation analysis, longitudinal evidence, process tracing, qualitative evidence, physiological measures, or several complementary approaches.
Be careful when the question assumes the causal conclusion in advance
Compare these questions:
“Why does generative AI reduce students' critical-thinking ability?”
“How do students perceive generative AI as shaping their critical-thinking practices?”
“Is generative AI use associated with critical-thinking performance?”
“What is the causal effect of access to generative AI on critical-thinking performance?”
The first question assumes that AI reduces critical thinking before the study has established that relationship. The second investigates participants' perceptions. The third asks about an association. The fourth explicitly asks for a causal effect.
They are not stylistic alternatives. They make different empirical commitments.
If the causal premise is genuinely established by prior evidence and your study investigates the mechanism, the first type of question might be defensible. If not, it risks embedding the conclusion inside the question.
“Why” is not the only word that can imply causation
Researchers sometimes remove “why” but leave the causal claim untouched.
For example:
“What factors influence student achievement?”
“What is the impact of social media on academic performance?”
“How does AI affect learning?”
Words such as “affect,” “impact,” “increase,” “reduce,” and sometimes “influence” can imply causal relationships depending on context. Guidance on reporting observational research specifically cautions that causal language extends beyond the literal words “cause” and “causal.”
This is why simply replacing “why” with “what” does not solve an inference problem. The neighboring guide on words such as “impact,” “influence,” and “effect” in observational research questions considers that problem more directly.
Do not replace every “why” question with “what factors are associated with”
Overcorrection creates a different problem. Suppose your genuine research interest concerns how novice teachers understand the reasons they leave the profession.
Changing the question to “What factors are associated with teacher attrition?” does not merely make the wording safer. It creates a different study. You have moved from an interpretive question about reasons and experiences to an associational question requiring measured variables.
This is especially important when distinguishing qualitative from quantitative research questions. Qualitative research may legitimately investigate how and why participants understand events as they do, while quantitative causal inference asks a different kind of question about what would happen under alternative conditions.
The verb should match the level of inference
| What you want to know |
Possible wording |
What the evidence needs to support |
| Participants' reasons |
Why do students report using generative AI for assessed work? |
Credible evidence about participants' accounts, reasoning, and experiences |
| Contextual explanation |
How and why did implementation of the policy differ across departments? |
Evidence about processes, context, interactions, and plausible explanations |
| Association |
Is workload associated with generative AI use? |
Evidence about the relationship between measured workload and AI use |
| Prediction |
Which characteristics predict subsequent generative AI use? |
Evidence that the specified predictors provide useful out-of-sample or otherwise appropriately evaluated prediction |
| Causal effect |
Does reducing workload change students' use of generative AI? |
A design, estimand, assumptions, and analysis capable of supporting causal inference |
| Mechanism |
Through what processes might workload affect generative AI use? |
Evidence capable of illuminating the proposed causal pathway or mechanism |
The rows are not a hierarchy from weak to strong research. They answer different questions.
Sometimes the best solution is to make “why” more explicit
If the word “why” could be interpreted causally when that is not what you intend, you can often preserve the intellectual question while clarifying the object of explanation.
Instead of:
“Why do students use generative AI?”
you might ask:
“How do students explain their decisions to use generative AI for assessed coursework?”
Instead of:
“Why did the new policy fail?”
you might ask:
“How do faculty and administrators explain the difficulties encountered during implementation of the new policy?”
This does not make the questions intrinsically superior. It makes the intended evidentiary claim harder to misunderstand.
Your study should be capable of answering the version of “why” you choose
Ultimately, “why” is an answerability problem rather than a forbidden-word problem.
If you want participants' reasons, collect evidence capable of illuminating their reasoning. If you want a contextual explanation, choose a design that can examine processes and context. If you want a causal effect, formulate a causal estimand and use a design and analysis that can support causal inference. If you want a mechanism, collect evidence relevant to the proposed pathway.
This returns to the basic test of whether your research question can actually be answered by the proposed study. The question should not promise a stronger explanation than the evidence can provide.
