Statistics and evidence
Correlation vs causation: what do the data show?
A pattern in paired data is a starting point. Check the variables, alternative explanations and study design before claiming a cause.
Correlation vs causation is a distinction you need whenever a worksheet, chart or headline suggests that one thing produces another. Two measurements can vary together without changing one being responsible for changing the other. The useful response is not simply to repeat “correlation does not imply causation”. Explain what the data show, name a plausible alternative explanation and identify what evidence would help distinguish the possibilities.
This guide uses original fictional school and library scenarios. The small datasets are invented for reasoning practice, not findings from real students, an app test or a scientific study. You can work with paper and a pencil. The focus is interpreting evidence, not calculating a correlation coefficient or running a significance test. If your assignment requires those calculations, keep them separate from the question of whether its study design supports a causal claim.
1. Correlation vs causation starts with the question
An association describes how variables occur together in a set of observations. A causal claim says that changing one would produce a change in another, under the relevant conditions. Suppose a school library records that students who visit more often borrow more books. That is a statement about an observed pattern. “Making every student visit twice more will increase borrowing” adds a claim about an intervention that the observation alone has not tested.
Write both variables precisely. Are visits counted per student per month, or are you comparing daily totals for the whole library? Is borrowing the number of books, the number of borrowers or the time spent reading? Those are different quantities. Name the unit represented by one observation, the period covered and the group studied. A conclusion about daily library totals cannot automatically become a claim about what an individual student will do.
“Describe the association” asks for a pattern; “explain why causation is not established” asks about the design and alternatives. “Suggest an investigation” asks what additional comparison could help. Mark a causal word such as “causes”, “improves” or “leads to” in the proposed conclusion. Then check whether the task actually supplies evidence about deliberately changing the suspected cause, rather than merely observing two measurements together.
2. Keep paired observations together and inspect the plot
Use this fictional set of six student records, written as monthly library visits and books borrowed: (1, 3), (2, 7), (3, 4), (4, 9), (5, 8), (6, 12). Each pair belongs to one student. Plot one point per pair, with visits horizontally and borrowed books vertically. The general pattern rises, although it is not perfectly ordered. A student with three visits borrowed fewer books than the student with two visits.
Do not sort the two columns independently: that would create pairings that were never observed. Check repeated values, missing records and units before describing the pattern. A scatter plot can show direction, shape, clusters and unusual points. The NIST explanation of scatter plots is a reference for inspecting these features. Our six records are new teaching examples; the photographs illustrate study contexts and do not encode the numerical answers.
Label the axes and choose a scale that displays the actual values clearly. Do not join these student points into a time sequence, because they are different people rather than successive stages for one person. If axis placement or scale is the difficulty, review the guide to plotting data with clear axes.
3. Separate describing, predicting and changing an outcome
For the six fictional records, a cautious description is: students with more recorded visits generally borrowed more books in this sample. It does not say that every extra visit adds a fixed number of books. It also does not establish a pattern for all students in the school. The dataset is small, its selection is unspecified and individual differences remain. Describe the group and period actually represented instead of silently expanding the claim.
Prediction is a different question. Visits might help estimate borrowing for similar students under similar circumstances, even if visits are not the sole cause of borrowing. A predictor can be useful because it shares information with other factors. But a relationship observed under one set of circumstances may change when the policy or population changes. Do not treat a prediction made from existing records as a tested promise about what a new intervention will accomplish.
An intervention question would ask what happens if comparable students are encouraged or required to make additional visits. Perhaps existing keen readers already choose to visit frequently; perhaps an assignment drives both visits and borrowing. Forcing a visit need not recreate those reasons. Keep separate sentences for what was observed, what might be predicted and what change is proposed.
4. Consider direction, common causes and chance
One possible explanation is that visiting creates more opportunities to borrow. Another is that students planning to borrow several books choose to visit more frequently. That reverses the proposed direction. You do not have to prove that the reverse explanation is correct to show why the original observation is insufficient. It is enough to identify a credible process that could also produce the observed pattern and explain how the current records fail to distinguish it.
A common cause is another possibility. A research assignment could prompt both more library visits and more book borrowing. In this scenario, assignment requirements would be a potential confounding factor because they could influence both quantities being compared. Merely naming an unrelated characteristic is not enough: explain the two connections. Different courses or different access to the library might matter, but you need a reason they could affect both the supposed cause and the outcome.
Chance, selective reporting and measurement problems can also produce misleading patterns. With a small set, an unusual student may strongly influence the picture. A headline may show only the class that fits its story, while excluding classes with different patterns. These possibilities do not prove the association is false. They explain why you should ask how observations were selected, how reliably variables were recorded and whether the relationship repeats before treating a pattern as a dependable explanation.
5. A coefficient cannot answer the causal question
If your course uses Pearson’s r, remember that it summarises a linear association between two numerical variables. Its sign describes direction, while its magnitude describes how closely the points follow a linear pattern. It does not measure the size of an intervention’s effect. Even a value close to positive or negative one cannot, on its own, tell you which variable causes which, or whether a third factor explains their relationship.
A value near zero does not rule out every kind of relationship. A curved pattern may have little linear correlation, so inspect the plot rather than relying only on the number. Outliers, the range of values sampled and combining different groups can also affect interpretation. Do not delete an inconvenient observation just to make r larger. Check whether it is a recording error, a different type of case or a genuine value that needs explanation.
Nor does “statistically significant”, when properly established, mean “causal”, “important” or “true for every individual”. A significance calculation concerns a particular statistical model and its assumptions. This guide has not calculated one for the invented records, and you should not invent a p-value from a visual trend. If the worksheet provides a coefficient or test result, report it accurately, then return to the study design to evaluate the stronger cause-and-effect statement.
6. Ask how the comparison was created
In an observational study, researchers record what happens without assigning the suspected causal condition. Such studies can reveal useful patterns and contribute to causal investigations, but a simple observational comparison needs careful treatment of alternatives. A larger sample can estimate an association more precisely while leaving the same confounding problem unresolved. More observations do not automatically repair a comparison between groups that differ in several important ways before the supposed treatment occurs.
A well-designed randomised experiment deliberately assigns conditions by chance and compares outcomes measured consistently. For a harmless classroom example, a teacher might randomly assign equivalent seedling pots to two approved watering schedules, keeping other planned conditions comparable. Random assignment helps balance other factors on average; it does not guarantee identical groups or flawless execution. The OpenStax chapter on experimental design introduces treatment groups, comparisons and the reasoning behind assignment.
Random sampling and random assignment solve different problems. Sampling concerns who or what enters a study; assignment concerns which condition included units receive. A random sample followed by students choosing their own treatment does not become a randomised experiment. Conversely, randomly assigning a small volunteer group does not guarantee results apply to everyone. Causal evidence can also come from carefully justified nonexperimental methods, but those need explicit assumptions and analysis beyond simply drawing a scatter plot.
7. Write a conclusion that matches the evidence
Build your answer around three connected statements: the observed pattern, a plausible alternative and the evidence needed to judge the causal claim. For the library example: in these six fictional records, more visits generally accompany more borrowing; a research assignment could increase both; the records do not show what would happen if visits alone were changed. This explains the limitation rather than using the slogan about correlation as a substitute for reasoning.
Choose language that matches the task: “is associated with” for the observed relationship, “could help explain” for an untested mechanism, and “the study does not establish” for a causal limit. Avoid saying “there is no causal relationship” merely because the design cannot establish one. Lack of sufficient evidence for a cause is different from evidence that no cause exists. Keep the possibility open while stating exactly what has and has not been investigated.
Trace the claim to its original context, not only a headline or a cropped chart. The guide to analysing a source helps distinguish an author’s statement from supporting evidence. Check the date, population, definitions and method. If your teacher gives a short fictional extract instead of a full study, evaluate the information supplied and identify what is missing. Do not invent details about randomisation, sample size or measurements to make the conclusion stronger.
8. Test the reasoning on a different school example
Try this new fictional claim: students who attend optional revision sessions have lower quiz scores, so the sessions must reduce learning. Before accepting it, ask who attends. Students who were already struggling might choose extra help. Prior difficulty could explain both attendance and later scores. The pattern alone does not show whether those students would have scored better, worse or the same without the sessions. Name the missing comparison in your answer rather than guessing the true effect.
Now consider a different proposed investigation: equivalent seedling pots are randomly assigned to two teacher-approved watering schedules and their growth is measured over the same period. That design addresses a causal question more directly than observing whichever pots happened to receive more water. You would still inspect consistent measurement, adherence, light, chance variation and the scope of the conclusion. Do not report a result: this is a plan, not an experiment that we conducted or evidence that either schedule works better.
Write your own explanation before using optional assistance. If permitted, Lirno’s study and practice workflow can help you turn a specific confusion into another question. AI can misread a study or reason incorrectly; school rules still apply, and Lirno guarantees neither correctness, grades, mastery nor permission. Check any proposed explanation against the actual design, then close the reference and explain why the evidence supports an association, a causal claim or neither.
