When two variables move together, the relationship does not tell you which variable caused the other—or whether either caused the other at all. Before turning an association into advice, check timing, plausible mechanisms, common causes and whether the proposed direction survives alternative explanations.

This matters because causal direction changes the solution. If struggling students ask for more help, a dataset may show that students receiving more help also have lower marks. Concluding that help caused the lower marks would reverse the likely pathway in at least some cases.

The tutoring graph that tells the wrong story

Imagine a fictional school records weekly support hours and examination scores. Students with more support hours tend to have lower scores. A hurried conclusion says, “Extra support reduces performance.”

Another explanation is immediately plausible: students who are struggling receive or seek more support. Lower performance influences support exposure. The observed relationship can therefore arise even when support is beneficial, neutral or harmful.

The association alone cannot distinguish those possibilities. The causal question needs information about why support was assigned, what happened before it began and what comparable students would have done without it.

Time order helps, but it does not finish the job

A cause must occur before its effect in the relevant pathway. If the supposed cause is measured after the outcome, the proposed direction is immediately suspect.

But measuring support before the final examination does not prove support caused the final score. Earlier weakness may have caused both the support and later performance. Timing removes some impossible stories; it does not remove every common cause.

Keep a timeline: prior attainment → support decision → learning during support → later assessment. Ask where each measured variable enters that sequence.

A third variable can produce the relationship

Suppose students with upcoming examinations both study more hours and report more stress. A positive association between study time and stress does not establish that studying creates all the stress. Examination proximity could increase both.

Likewise, hot weather can increase both cold-drink sales and use of air conditioning. Buying cold drinks does not thereby cause air conditioners to run.

The useful question is not “Can I imagine a third variable?” Almost anything can be imagined. Ask which alternative causes are plausible from subject knowledge and whether the design measured them adequately.

Interventions require a causal claim

If a school wants to know whether adding a practice routine will improve performance, it needs a causal question: what would happen if the routine were changed while relevant alternatives were addressed?

An observational association can motivate that investigation. It should not be promoted directly into “therefore make every student do more of X”. The intervention itself changes the system and may produce effects not visible in the original correlation.

The existing How Causal Inference Works guide develops the wider framework. This short article focuses on one failure: direction is not encoded in correlation.

A mechanism can make one direction more plausible

A causal explanation should describe a pathway. “Feedback improves revision” is incomplete. A more useful proposal is that feedback identifies a specific error, the learner changes the next attempt, and the corrected relationship transfers to a fresh task.

That mechanism suggests observations to check: Did the feedback identify the error? Did the next attempt change? Did improvement persist on a different question?

A plausible mechanism does not prove causation by itself. It makes the hypothesis more testable and prevents the explanation from remaining a decorative arrow between two correlated variables.

Reverse causality can coexist with a real forward effect

The support example need not have only one direction. Weak performance may lead to more support, and support may subsequently improve performance. Feedback loops are common in education, health, economics and social systems.

A simple cross-sectional correlation compresses that dynamic relationship into one number. The result can look weak, strong or even point in a direction that obscures the underlying loop.

When feedback is plausible, collect repeated observations and model the sequence carefully rather than forcing the system into a one-way story.

Prediction and causation are different achievements

A variable can predict an outcome without being a useful intervention target. Umbrellas predict rainy conditions because people carry them when rain is expected. Removing umbrellas does not stop the rain.

Similarly, a learning behaviour may be a useful warning signal even when changing that behaviour directly would not solve the underlying problem. Prediction asks whether information helps anticipate an outcome. Causation asks what would happen if the variable were changed.

Do not dismiss a predictor because it is not causal. Use it for the job it can perform, while keeping intervention claims separate.

The repair routine

Start by rewriting “X causes Y” as three competing possibilities: X causes Y; Y causes X; another factor influences both. Add feedback between X and Y when the system makes that plausible.

Build a timeline. Identify how participants entered different conditions. Ask what mechanism would transmit the proposed effect and what observations would differ under the competing explanations.

Then match the conclusion to the design. If the evidence establishes association only, write association. Do not hide causal language inside verbs such as “drives”, “leads to”, “improves” or “reduces”.

For readers comparing changing study routines, continue with Why Changing Every Study Habit at Once Can Fail. For wider research reasoning, use the World Knowledge Research Library.

Correlation can tell you where to look. Causal direction tells you what might happen if you act. Those are different jobs, and confusing them can make a well-intended solution push on the wrong part of the system.

Browse connected guides in the Parent Learning Support Directory.