A correlation describes how two measured variables vary together in the available data. It does not by itself show that one variable caused the other. A causal claim needs a defensible account of time order, comparison or counterfactual, confounding, selection, measurement and design. In student research, the responsible default is usually to describe an observed association, then explain plausible alternative explanations and the study’s limits. Stronger causal wording requires a design capable of supporting it, such as a well-conducted experiment or a carefully justified quasi-experimental evaluation—not merely a larger sample, a regression output or a small p value (USC Libraries, 2026; Greenland et al., 2016).
Association and cause answer different questions
Correlation asks whether higher or lower values of one measure tend to occur with higher or lower values of another. Causation asks what would have happened to the same unit if the proposed cause had been different, all else relevant being equal. The second question requires a counterfactual comparison that ordinary observational data often cannot provide.
This does not make correlation useless. Association can reveal patterns worth investigating, generate hypotheses, guide descriptive reporting and support prediction within a defined dataset. The error is moving from “these measures are related” to “one explains or produces the other” without the evidence needed for that step.
| Statement | What the evidence may show | What it does not establish alone |
| Areas with higher X also have higher Y | An area-level association in the included data. | That individuals with X necessarily have Y, or that X caused Y. |
| Students reporting more work hours report fewer study hours | A relationship in the respondent sample. | That paid work caused the lower study time. |
| Sales rose after a campaign began | A temporal pattern. | That the campaign caused the rise; seasonality, price or other changes may matter. |
| A model predicts an outcome reasonably well | Predictive performance in a defined dataset. | The causal mechanism, fairness or future performance elsewhere. |
The five alternative explanations to examine
Whenever an association is observed, examine at least five possibilities before writing a conclusion.
1. Confounding
A third variable may influence both measures. Ice-cream sales and heat-related hospital admissions may rise together because temperature affects both; sales do not therefore explain admissions. In student research, socioeconomic circumstances, age, location, season, baseline condition or organisational differences may affect more than one variable.
2. Reverse direction
The presumed outcome may influence the presumed cause. For example, performance may affect motivation as well as motivation affecting performance. Cross-sectional data often cannot establish which came first.
3. Selection and representation
The people or records observed may differ systematically from those not observed. A voluntary survey can over-represent those with a particular interest, availability or experience. A relationship in respondents may not apply to a wider population.
4. Measurement and data process
Variables may be measured with error, based on self-report or created by administrative processes. A label such as “engagement” may be an imperfect proxy, while a change in coding or reporting can create an apparent trend.
5. Chance and model assumptions
Random variation, multiple analyses, inappropriate model form, time dependence or clustering can create a misleading association. A p value does not identify which of these issues may be present; it is calculated under an entire statistical model (Greenland et al., 2016).
| Alternative explanation | Diagnostic question | Responsible action |
| Confounding | What variable could plausibly influence both X and Y? | Measure, discuss or design around it where possible. |
| Reverse direction | Could Y precede or affect X? | Avoid directional causal wording unless timing/design resolve it. |
| Selection | Who is included, excluded or non-responding? | State sample/data coverage and limit generalisation. |
| Measurement | What exactly do X and Y record? | Use source documentation; name measures accurately. |
| Data structure | Are records independent, time-ordered or clustered? | Use methods suited to the structure and explain limits. |
A causal-evidence ladder
Causal reasoning is not a binary label. Different designs provide different kinds of evidence. The ladder below is deliberately simplified: real-world validity depends on execution, context and assumptions, not on a design name alone.
| Design approach | Core strength | Principal causal limitation |
| Cross-sectional observational data | Can describe and explore associations at one point in time. | Timing, confounding and selection are often unresolved. |
| Longitudinal observational data | Can establish that some measures preceded others. | Confounding and attrition can remain. |
| Before/after series | Can show a change around an event. | Other simultaneous changes or trends may explain it. |
| Matched/comparison observational design | Makes observed groups more comparable. | Unmeasured differences may persist. |
| Quasi-experimental design | Uses a credible natural or policy comparison. | Validity depends on strong, testable assumptions. |
| Randomised experiment | Random allocation can address many confounders. | Adherence, measurement, external validity and implementation still matter. |
USC Libraries distinguishes descriptive quantitative studies, which establish associations, from experimental designs intended to establish causality. The distinction should guide wording in a report. In an observational project, say “was associated with”, “varied alongside” or “was higher among”, rather than “led to”, “determined” or “caused”, unless the design has a credible causal basis (USC Libraries, 2026).
Correlation coefficients and scatterplots
A correlation coefficient gives a numerical summary of a relationship’s direction and strength under a chosen form of association. It cannot show the whole structure. Always inspect a scatterplot or another display suited to the data. The same coefficient can mask a curved relation, distinct clusters, influential observations, a restricted range or a time trend.
| Observation in a scatterplot | Why a coefficient alone can mislead | Interpretation response |
| One extreme point drives the trend | The relationship may disappear without one case. | Investigate validity; report influence transparently. |
| Curved pattern | A linear coefficient may be small despite a clear non-linear pattern. | Describe the shape and consider a suitable model. |
| Separate clusters | An overall association may reflect group membership. | Examine context and potential confounders. |
| Time trend in both series | Correlation can arise because both change over time. | Address trend, seasonality and temporal dependence. |
| Narrow observed range | Association may be understated or unrepresentative. | Explain sample/measurement restrictions. |
Causal language as a quality control
Writing is part of research integrity. Compare these statements:
“The analysis found that working more hours lowers study time.”
“In the available survey, reported paid-work hours were negatively associated with reported study hours. The cross-sectional self-report design cannot establish direction or rule out other influences.”
The second statement is not weaker; it is precise. It distinguishes an observed relation from a causal conclusion and tells the reader what would be needed to strengthen the inference.
The Hypothesis Testing guide explains why a p value cannot turn an observational association into causal proof. The Statistics Project Methodology guide addresses the earlier choices of sampling, measurement and data route that shape causal limits before analysis begins.
Fictional planning application
The following scenario is fictional. Omar downloads a public dataset showing local-authority green-space coverage and a published wellbeing indicator. His scatterplot suggests that areas with more green space have higher average wellbeing scores. He initially writes, “Green space improves wellbeing.”
Omar pauses to apply the five explanations. The data are area-level, so they do not show individual exposure. Area income, housing, age structure, urban density and access to services could influence both green space and wellbeing. The period of the two indicators does not exactly match, and the administrative boundaries vary in size. His research question is therefore revised to an association question.
His final conclusion states that the included local authorities showed a positive area-level association, while identifying potential confounding and the ecological limit. He proposes that a future study would need individual-level exposure, time order and a credible comparison strategy to investigate a causal claim. The revised conclusion is more useful because it tells readers exactly what the data support.
Practical interpretation sequence
First, state the variables, unit and data source. Second, produce descriptive summaries and a scatterplot or appropriate comparison display. Third, identify the form of association, unusual observations, clusters and possible time patterns. Fourth, list plausible confounders, reverse-direction pathways and selection issues before fitting a model.
Fifth, report the estimate and uncertainty in plain language. Sixth, choose conclusion wording that matches the design. If a causal question is essential, redesign the study or obtain appropriate methods advice rather than attempting to upgrade an observational association through wording or software output.
Critical limitations and safeguards
Do not use correlation to make medical, legal, financial or individual risk judgments. Group-level data can stigmatise communities if interpreted as a description of every resident. Sensitive variables require careful ethical consideration, de-identification and disclosure control.
Do not exclude cases, variables or time periods merely because they weaken a preferred relationship. Preserve the data, document any justified cleaning decision and make analytical choices transparent. Causal claims influence decisions; when consequences are material, independent review, subject expertise and an appropriate causal design are essential.
FAQs
Q1. Does a strong correlation prove causation?
No. It may reflect confounding, reverse direction, selection, measurement or other data features. Strength of association does not substitute for causal design.
Q2. Can regression analysis prove cause and effect?
No. Regression can adjust for measured variables under stated assumptions, but it cannot automatically address unmeasured confounding, selection or incorrect measurement.
Q3. Why show a scatterplot if I have a correlation coefficient?
A plot reveals shape, clusters, unusual observations and time-related patterns that a single coefficient can conceal.
Q4. What wording is safe for observational findings?
Use “associated with”, “correlated with”, “varied alongside” or “was higher/lower in the available data”, then state the principal limits of the design.
References
Greenland, S., Senn, S.J., Rothman, K.J., Carlin, J.B., Poole, C., Goodman, S.N. and Altman, D.G. (2016) ‘Statistical Tests, P Values, Confidence Intervals, and Power: A Guide to Misinterpretations’, European Journal of Epidemiology, 31, pp. 337–350 (Accessed: 25 August 2026).
NIST/SEMATECH (2025) e-Handbook of Statistical Methods (Accessed: 25 August 2026).
Texas A&M University (2026) Choosing the Right Research Design (Accessed: 25 August 2026).
USC Libraries (2026) Quantitative Methods (Accessed: 25 August 2026).