Reader outcome
After reading this guide, a student should be able to decide whether Pearson’s product-moment correlation is suitable for a research question, inspect the relationship before calculating a coefficient, explain what the sign and magnitude of the coefficient mean, distinguish statistical evidence from practical importance, and write a cautious thesis result without implying causation.
What Pearson correlation measures
Pearson correlation summarises the direction and strength of a linear association between two numeric variables. The coefficient is usually written as r. It ranges from -1 to +1. A positive value means that larger values of one variable tend to occur with larger values of the other. A negative value means that larger values of one variable tend to occur with smaller values of the other. A value near zero means that the data do not show a strong linear pattern; it does not prove that the variables are unrelated in every possible way.
The coefficient can be expressed as the centred cross-product of the two variables divided by the product of their centred sums of squares. This standardisation makes the result unit-free. Changing a variable from centimetres to metres, for example, does not change r.
Pearson correlation is symmetric: the correlation of X with Y is the same as the correlation of Y with X. That symmetry is useful for association, but it also explains why the coefficient does not identify a dependent variable, a causal direction or a mechanism.
When it is an appropriate choice
Pearson correlation is a reasonable candidate when the research question concerns linear association between two quantitative variables measured on the same observational units. Examples include the association between study time and examination score, advertising expenditure and monthly sales, or temperature and energy consumption.
Before selecting it, check four issues.
- Measurement and pairing. Both variables must be numeric, and each pair must refer to the same case, person, organisation, location or time point.
- Linearity. A scatterplot should show an approximately straight-line pattern. A clear curve can produce a small Pearson coefficient even when the variables are strongly related.
- Influential observations. Extreme or high-leverage points can substantially change the coefficient. Inspect them, verify whether they are data errors, and report justified decisions rather than deleting inconvenient cases.
- Inference conditions. A p-value and conventional confidence interval depend on a sampling model. Independence, study design and distributional assumptions matter more for inference than for the mechanical calculation of r.
A statistically significant result does not repair a poor design. Repeated measurements from the same participant, clustered observations, time series, selection bias and range restriction can all require a different analysis or a more specialised model.
Step-by-step analysis workflow
1. Translate the research question
Write the question without statistical shorthand. For example: “Among first-year students in the sampled programme, is weekly independent study time linearly associated with final examination score?” Name the population, the two variables and the intended form of association.
2. Audit the data
Confirm units, coding, impossible values, duplicate cases and missing observations. Record how many complete pairs are available. Do not silently substitute zero for a missing value.
3. Plot the paired observations
Use a scatterplot with meaningful axis labels and units. Look for direction, curvature, clusters, ceiling or floor effects, and influential observations. A fitted straight line can help display the linear component, but the raw points must remain visible.
4. Calculate the coefficient and uncertainty
Calculate r. When the study supports statistical inference, also obtain a confidence interval and the test of the null hypothesis that the population correlation is zero. The confidence interval usually communicates uncertainty more effectively than a p-value alone.
5. Interpret in context
Describe direction first, then strength in relation to the field, measurement quality, sample and consequences. Avoid universal “small, medium, large” labels as if they were natural laws. A correlation that appears modest can be important in a high-stakes setting; a large correlation can still be unhelpful if the measures are biased or the design is weak.
6. Check robustness
Compare the conclusion with the scatterplot and, where justified, a rank correlation such as Spearman’s rho or Kendall’s tau. A large discrepancy can indicate outliers, a monotonic but nonlinear pattern, ties or distributional features that deserve explanation.
Worked example
Suppose 30 students provide weekly independent-study hours and receive a final examination score. The scatterplot shows an upward, approximately linear pattern with no single observation dominating the fit. The analysis produces r = .46 with a 95% confidence interval from .12 to .70 and a two-sided p-value of .010.
A defensible interpretation is that the sample shows a positive linear association: students reporting more independent study tended to have higher scores. The confidence interval indicates uncertainty about the population magnitude. The result does not show that increasing study time by itself will cause a particular score increase. Motivation, prior attainment, course attendance and measurement error could affect both variables.
The coefficient of determination for this two-variable linear association is r² = .212. It can be described as approximately 21% shared linear variation in this sample. It should not be described as “21% of examination performance is caused by study time.”
Reporting in a thesis
A concise results paragraph could read:
A scatterplot indicated an approximately linear positive relationship between weekly independent-study time and final examination score, with no single observation exerting obvious dominance. Pearson’s correlation showed a moderate positive association, r(28) = .46, 95% CI [.12, .70], p = .010. In this sample, students who reported more independent study tended to obtain higher scores; the observational design does not establish a causal effect.
Adapt the wording to the discipline’s reporting convention. State the number of complete pairs, missing-data rule, software and version when relevant. If the analysis was preregistered or weighted, clustered or adjusted, say so.
Common mistakes
- Reporting only “significant” or “not significant” without the coefficient, uncertainty and sample size.
- Treating a near-zero coefficient as proof of no relationship without checking for curvature.
- Using Pearson correlation for ordinal categories merely because they were coded with numbers.
- Removing an outlier solely because it weakens the result.
- Correlating repeated measures as if every row were independent.
- Running many correlations and highlighting only the smallest p-values without multiplicity control or a declared analysis plan.
- Interpreting association as causation or prediction accuracy.
Limitations and alternatives
Use Spearman or Kendall methods when the research question concerns a monotonic rank relationship, the scale is ordinal, or the data contain features that make a rank analysis more defensible. Use regression when the study has a directional outcome, covariates, interactions or prediction goals. Use repeated-measures, multilevel or time-series methods when observations are dependent. For nonlinear relationships, consider transformations, splines or models grounded in theory rather than forcing a straight line.
Source and verification notes
This guide is anchored to IBM’s documentation describing Pearson correlation as a measure of linear association and warning readers to inspect linearity and outliers; NIST’s definition and -1 to +1 range; and the R documentation for Pearson testing and Fisher-transformed confidence intervals. The numerical example is original and synthetic. No source wording has been copied as instructional prose.