Correlation vs causation practice is about matching a claim to the evidence. Two variables can move together without one causing the other. To answer a question well, identify the study design, describe the observed relationship, and explain what else could account for it. A strong correlation alone does not settle the causal question.
These eight original campus scenarios are fictional. Their numbers illustrate reasoning, not findings about real students or study products. Try each question before reading the answer. Your goal is to write a defensible sentence, not simply circle “correlation” whenever an example sounds uncertain.
A short checklist before the questions
Ask who chose the explanatory condition. Did students choose an activity themselves, did researchers merely record what happened, or did researchers randomly assign a condition? Then identify the outcome and when it was measured. If the variables are hard to separate, review these independent and dependent variable examples.
Next, consider an alternative explanation. A third variable could influence both measured variables. The outcome could influence the supposed cause. Selection or measurement could also create a misleading pattern. Penn State’s lesson on relationships between measurement variables explains these alternatives and why statistical significance alone does not establish causation.
Finally, write the strongest conclusion the design supports. Random assignment can support a causal comparison when the experiment is well conducted and analyzed. Random sampling answers a different question: how participants were selected from a population. A study can have one, both, or neither.
1. A note-sharing habit and higher quiz scores
Question: In a voluntary survey, students who regularly exchange their own lecture summaries report higher quiz scores. A club writes, “Sharing summaries raises everyone’s quiz score.” What needs to change?
Worked answer: The survey shows an association among respondents. Students chose whether to share summaries, so the groups may differ before sharing happens. For example, students who plan their work carefully might both share notes more often and prepare more consistently for quizzes.
A defensible replacement is: “Respondents who reported sharing summaries regularly also reported higher quiz scores; this survey does not establish that sharing caused the difference.” Mention that self-reported scores and voluntary participation can also limit the evidence.
Repair the reasoning: A possible confounder is an alternative explanation, not a fact you have discovered. Say planning habits “could” influence both variables. Do not replace an unsupported sharing claim with an equally unsupported claim that planning definitely explains the result.
2. More help requests, lower assignment scores
Question: Students sending more help requests during a programming assignment tend to receive lower scores. Does asking questions harm performance?
Worked answer: That conclusion is unsupported. Students who are already struggling may ask more questions. Difficulty with the material could contribute to both help requests and lower scores. If the score measures difficulty that began before the requests, the proposed causal direction is especially doubtful.
A better conclusion is: “Help-request frequency was associated with lower scores in this assignment, but the pattern may reflect students seeking help when they encounter difficulty.” This wording leaves open whether help improved their scores relative to what they otherwise would have achieved.
Design improvement: Measure prior performance and when difficulties began. Those details would clarify the timeline, though collecting them alone would not automatically remove all confounding. A student could benefit from help and still score below classmates who never needed it.
3. Random sampling does not assign a study method
Question: A college randomly samples 600 enrolled students. It asks whether they use printed or digital notes and records their recent grades. Printed-note users have a higher average. Does the random sample prove printed notes work better?
Worked answer: No. Random sampling concerns who enters the study. Students still chose their note format, so this is an observational comparison. Course difficulty, prior achievement, accessibility needs, or other factors could differ between format groups.
The sample may support an estimate for the college’s enrolled population if selection, response, and measurement are appropriate. It does not turn note-format choice into a randomized treatment. Nonresponse could still undermine representativeness even when the initial sample was random.
Exam sentence: “Random sampling can support population generalization, but without random assignment or another justified causal design, the observed grade difference does not establish an effect of note format.” Avoid writing that a large sample solves the issue. More observations do not automatically remove systematic differences.
4. A randomized experiment without enough results
Question: Volunteers are randomly assigned to review the same material with either practice questions or rereading. Study time is equal, and everyone takes the same test. The practice group averages two points higher. Can you declare the method effective?
Worked answer: Random assignment makes a causal comparison more credible, but the two-point difference alone is not enough to assess the evidence. You need group sizes, variation, the analysis, adherence, and information about missing results. A difference can arise through random variation.
If the experiment is well conducted and an appropriate analysis supports a treatment difference, the findings can provide evidence of a causal effect under these conditions. The prompt does not supply that analysis, so do not invent a p-value or claim statistical significance.
Scope check: The participants were volunteers. Even convincing evidence within the experiment would not automatically establish the same effect for every student, subject, or exam. Separate the causal question from how broadly the result applies.
5. Two changes introduced together
Question: One course section receives weekly practice quizzes and an extra review session. Another receives neither. Their final exam averages differ. The instructor credits the quizzes alone. What is the problem?
Worked answer: The quiz change is bundled with extra review. The comparison cannot isolate their separate effects. If students also chose their sections, additional differences between the groups may exist.
A narrower statement is: “The section receiving the combined program had a different average, but this comparison does not identify the effect of practice quizzes alone.” Notice that even if sections had been randomly assigned, the treatment being tested would still be the combined program.
Design improvement: To investigate quizzes separately, keep review-session access the same and randomly assign the quiz condition when appropriate. If the aim is to compare both components, a design with separate combinations could help, but it also needs enough independent groups and a suitable analysis.
6. A striking correlation coefficient
Question: Across twelve weeks, the number of campus event listings and the number of library bookings have a correlation of r = 0.94. Someone says event listings must cause bookings because the correlation is so high.
Worked answer: The reported coefficient describes a strong positive linear association in these observations. It does not reveal why the association exists. Both quantities might increase as the semester becomes busier, or another shared time pattern might influence them.
Write: “Event listings and library bookings increased together in this data set, but a common semester pattern could explain the association.” A plot of the observations and their timing would help you understand the pattern. Twelve successive weeks may also raise questions about dependence between observations.
Tempting mistake: Treating 0.94 as “94 percent caused by events.” A correlation coefficient is not a percentage of causation. Computing its square would not transform it into one either.
7. Before and after a planning workshop
Question: A group attends a planning workshop. Its average number of late submissions falls from four in the previous month to two in the next month. Is the workshop responsible?
Worked answer: The before-and-after change is real within the hypothetical data, but its cause is unresolved. The second month might contain fewer assignments, easier deadlines, or other support. Without a suitable comparison, those possibilities remain open.
The raw average fell by two submissions, a 50 percent decrease relative to four. That arithmetic describes the change; it does not identify its cause. You should also compare assignment opportunities. Two late submissions out of four assignments differs from two out of twenty.
Better reporting: “Average late submissions fell after the workshop, but this before-and-after comparison cannot isolate the workshop’s effect.” A contemporaneous comparison group and a clearly defined outcome could improve a future evaluation.
8. A weak linear relationship
Question: A fictional experiment varies the brightness of a study display. Performance is highest at medium brightness and lower at both extremes. The linear correlation is close to zero. Does that prove brightness has no effect?
Worked answer: No. A correlation near zero can coexist with a curved relationship. A straight-line summary may miss the pattern described in the prompt. Plotting performance against brightness would reveal information the single coefficient hides.
Whether the experiment provides convincing causal evidence still depends on assignment, execution, measurement, and analysis. Do not jump from “linear correlation is weak” to “no relationship,” and do not jump from a visible curve to a proven effect without checking the design.
Exam sentence: “A near-zero linear correlation does not rule out a nonlinear relationship; the experimental design and appropriate analysis are needed to evaluate the effect.” This question tests what the statistic measures as much as what a study can establish.
A three-sentence answer format
First describe the observed pattern. Second explain the design limitation or strength. Third give a conclusion with the appropriate scope. For an observational example, you might write: “The groups differed in average score. Students chose their groups, so prior preparation could influence both membership and performance. These data show an association but do not establish a causal effect.”
Practice that structure on two cases above without copying their wording. Then change one design detail, such as replacing student choice with random assignment, and explain how your answer changes. Check both what the new detail permits and what remains unknown.
Build a study sheet others can use
Organize a page into observed pattern, study design, possible alternative, and justified conclusion. Add your own fictional examples. Compare definitions with statistics course notes, or browse study documents for complementary practice. Check the source and course context before relying on a shared answer.
Students should upload their own notes to help other students. Label invented scenarios as invented and keep the difference between a possible explanation and a demonstrated finding clear.
Frequently asked questions
Does correlation mean there is no causation?
No. A causal relationship may exist. Correlation alone does not establish it or determine its direction.
Does statistical significance remove confounding?
No. A result can be statistically significant while the comparison still reflects systematic differences between groups.
Are randomized experiments the only causal evidence?
No. Other research designs can support causal inference under explicit assumptions. For introductory questions, evaluate the actual design rather than treating every observed association as proof.