03 · What You Need to Know
Asking Whether Something Happens Is Different From Asking How It Happens
Research questions can operate at different explanatory levels. A descriptive question may ask what occurs. An associational question may ask whether variables are related. An intervention question may ask whether changing one condition changes an outcome. A mechanistic question goes further by asking about the process or pathway through which the outcome may arise.
For example:
Outcome question: “Does formative feedback improve students' subsequent writing performance?”
Mechanism question: “To what extent is any improvement in subsequent writing performance associated with changes in students' revision behavior?”
These questions require overlapping but not identical evidence. The second requires meaningful measurement of the proposed intermediate process and a design capable of supporting whatever interpretation the researcher intends to make.
A Mechanism Is More Than Another Variable in the Model
Adding a third variable between an exposure and an outcome does not automatically establish a mechanism.
Suppose researchers observe that:
- AI-assisted feedback is associated with greater writing performance;
- AI-assisted feedback is associated with greater revision activity;
- revision activity is associated with writing performance.
Those patterns may be consistent with the hypothesis that revision activity helps explain the relationship. They do not, by themselves, establish that AI feedback caused greater revision and that the resulting revision caused the improvement in writing.
Mechanistic interpretation requires attention to temporal ordering, confounding, measurement, alternative pathways, assumptions, and study design. Mediation analysis can be useful, but statistical mediation and demonstrated causal mechanism are not interchangeable concepts.
Watch Out
A statistically significant indirect effect does not automatically prove the mechanism you proposed. Causal interpretations of mediation require assumptions and design considerations that should be justified rather than inferred from a statistical model alone.
Do Not Put the Mechanism in the Question as an Established Fact
Compare these two formulations:
“How does academic stress reduce achievement through sleep disruption?”
“Does sleep disruption help explain the association between academic stress and achievement?”
The first formulation presupposes both that academic stress reduces achievement and that sleep disruption is the pathway. If those relationships are precisely what the study is supposed to establish, the question has already embedded its preferred conclusion.
The second treats the proposed pathway as an empirical possibility.
This is closely related to the problem of asking “why” when the design cannot establish the implied explanation. Research questions should invite evidence to discriminate among possibilities rather than linguistically settle them in advance.
A Proposed Mechanism Can Remain in the Conceptual Framework
Not every theoretically important idea needs to appear in the main research-question sentence.
You may have a conceptual model proposing that an intervention influences an outcome through motivation, self-efficacy, cognitive load, social interaction, or another process. If the primary study is designed only to estimate the intervention's relationship with the outcome, the mechanism may appropriately remain part of the theoretical rationale and future research agenda.
Alternatively, you might formulate a primary outcome question and a secondary mechanistic question.
This can preserve clarity while acknowledging that establishing whether something works and understanding how it works are related but distinct objectives.
Mechanism Questions Require Mechanism-Relevant Evidence
If you explicitly ask whether X affects Y through M, you need evidence about M.
That sounds obvious, but mechanisms are sometimes introduced in research questions because a theory predicts them even though the study measures only X and Y.
Without data on the proposed pathway, the study may discuss the mechanism as a possible explanation, but it cannot empirically evaluate whether that pathway accounts for the observed result.
Even measuring the mediator may not be sufficient. Timing matters. If the proposed sequence is intervention → mediator → outcome, measurements should permit that sequence to be meaningfully evaluated rather than collecting all variables at a single undifferentiated time point and treating the arrows as established.
This is one reason temporal structure can matter in a research question, even when the exact dates remain in the protocol.
Mechanisms Can Be Qualitative as Well as Quantitative
Mechanism research is not limited to statistical mediation models.
Qualitative research can investigate how participants describe processes through which an intervention, policy, organizational change, or social condition produces consequences. Process evaluations can examine implementation, participant responses, contextual interactions, and plausible pathways that help explain observed outcomes.
The claims should still match the evidence. Participants' explanations can provide important evidence about perceived processes and experiences, but their accounts do not necessarily establish a causal mechanism in the same way that a design specifically capable of testing causal pathways might.
Mechanisms Are Particularly Important When the Same Outcome Could Arise in Different Ways
Two interventions can produce similar average outcomes through very different processes.
Imagine two educational technologies that both improve test performance. One may work because students receive more practice. Another may improve metacognitive monitoring. A third may simply increase time on task.
If researchers care about transferability, optimization, unintended consequences, or why an intervention succeeds for some people but not others, understanding the process can become scientifically important.
Mechanistic evidence may also help distinguish competing explanations that produce similar surface-level predictions.
A Mechanism Question Is Not Automatically Better Than an Outcome Question
Mechanistic questions can sound theoretically sophisticated, which creates a temptation to include a mediator in every conceptual model.
But additional complexity should earn its place.
If the study's primary uncertainty is whether an intervention produces an outcome at all, an elaborate mechanistic question may be premature. If prior evidence already establishes an outcome reasonably well but the process remains uncertain, mechanism may deserve greater emphasis.
The appropriate question depends on the state of knowledge and the purpose of the study.
This is another instance in which adding detail can make a research question worse rather than better if the additional element is not genuinely part of the inquiry.
Do Not Let a Preferred Theory Exclude Plausible Alternatives
A researcher may strongly expect that an intervention works through a particular pathway. That expectation can provide a useful hypothesis, but it should not prevent consideration of rival explanations.
Suppose students using an AI tutor perform better. The researcher hypothesizes that personalized feedback explains the improvement. Other possibilities might include increased practice time, novelty, greater motivation, access to additional examples, or differences in how students use the tool.
A study that measures only the preferred mechanism may have limited ability to distinguish among these explanations.
When several pathways are plausible and the evidence cannot discriminate among them, the question should not imply that one mechanism has already won the theoretical contest.
Mechanism Language Should Match the Strength of the Design
Terms such as “through,” “because of,” “explains,” and “mediates” can imply a causal pathway. Researchers should therefore be as careful with mechanism language as with words such as impact, influence, and effect in observational research questions.
If the design supports only associations among an exposure, possible mediator, and outcome, wording such as “is associated with,” “is consistent with,” or “to what extent does M account statistically for the association” may better reflect the evidence, depending on the analysis and assumptions.
Greater caution in wording does not make the research less interesting. It makes the inferential claim easier to defend.
