The core idea
Correlation summarises association between variables. Pearson’s correlation describes linear association and ranges from −1 to +1; a value near zero does not rule out a nonlinear relationship. A scatter plot can expose outliers and patterns hidden by one coefficient. Association does not establish that changing one variable will change the other. Reverse direction, common causes and selection can all produce an observed relationship.
Source and attribution [1]Using it in practice
Plot the data and check the unit of analysis before calculating a coefficient. Ask whether observations are independent and whether a small number of points dominate the result. Distinguish an employee-level association from a site-level one. Use the finding to formulate a testable explanation, then choose a design capable of addressing the causal question if that is what the decision requires.
An example, not a reported case
Worked example · illustrative
In a fictional set of sites, higher overtime and higher absence occur together. It would be premature to conclude that overtime alone causes absence: understaffing may affect both, and absence can itself create overtime. HR reviews timing, staffing changes and employee accounts before choosing an intervention. The correlation is a clue about a pattern, not an estimate of the benefit from cutting overtime.
What to watch for
Multiple exploratory correlations can produce chance findings. A statistically significant coefficient may still be practically small. Missing data, restricted ranges and aggregation can alter the result. Do not use an observed relationship to label an individual or to claim that a proposed intervention has already been validated.
Read the plot
Look for clusters, curvature and unusual observations. Check whether the axes use comparable periods and whether the population changed. A straight-line summary can be misleading when several different groups are mixed together.
Write a cautious conclusion
State which variables were associated, in which population and period, and what remains uncertain. Replace “X drives Y” with a description of the observed pattern unless the design supports a causal claim. Identify the next piece of evidence needed for a decision.
Plot the relationship before interpreting a coefficient
Check whether a pattern is roughly linear, whether outliers dominate and whether several different groups are mixed together. A single coefficient can conceal a curved relationship or opposing patterns across teams. Match the observations properly: monthly team averages answer a different question from individual records. Do not infer that every person behaves like the average of their site or department.
Generate competing explanations
Suppose teams with more training also have higher performance. Training may help, but stronger teams may receive more investment, better managers may influence both, or the work may differ. Write down these alternatives before recommending a programme. Ask about timing and what information would help distinguish them. Correlation is useful for identifying a relationship to investigate; it does not decide the direction of influence.
Avoid confusing a group pattern with an individual rule
An illustrative site-level association between engagement and retention does not mean the least engaged individual will leave, or that increasing a survey score will prevent an exit. Aggregation and group composition matter. Explain the level of analysis in the report. If the result could affect individuals, review whether the data and method support that use and whether a less intrusive group-level action would answer the business need.
Report what the result cannot establish
State the variables, population, period and uncertainty, alongside the size and direction of the observed relationship. Avoid dramatic language such as 'proves' or 'drives' when the design is observational. Use a follow-up analysis, qualitative evidence or a suitable intervention evaluation to investigate the mechanism. A weak correlation may also reflect poor measurement, restricted range or a non-linear pattern, so absence of a large coefficient is not a universal finding of no relationship.
Inspect the pattern before reporting a coefficient
Plot or otherwise examine the data for outliers, clusters and non-linear relationships. A single unusual site can strongly influence a small dataset, while two distinct groups can create an association that is absent within each group. Confirm that the observations are comparable and that the measures mean what their labels suggest. Correlation describes a pattern in the available data; it does not establish the sequence or mechanism needed for a causal explanation.
Develop competing explanations
If engagement and retention move together, better management may influence both, retention may affect the survey population, or satisfied employees may be more likely to respond. Use timing, additional evidence and an appropriate evaluation design to investigate. Do not convert the coefficient into a prediction of what will happen if a manager raises an engagement score. The useful next step is a testable hypothesis and a decision about what further evidence would justify action, not a causal claim dressed in statistical language.
Take it into your next conversation
Three useful questions.
- What question can this method answer, and what can it not establish?
- Are the comparison, observation period and assumptions defensible?
- What decision follows, and how will its consequences be reviewed?
Related terms
Go to the evidence
Sources & attribution
Further reading: Nadeem Khan and Dave Millner (2023), Introduction to People Analytics: A practical guide to data-driven HR, 2nd edition, PDF pp. 227, 244 and 294–295. Edition & reading notes ↗
[1] NIST: Scatter plots and association ↗
The core idea is an original summary of the cited work. Application notes, examples and sketches are our interpretations, not quotations or reproductions of the authors’ figures. Publisher records may require access to read the full original work.
Application guidance expanded on 26 September 2026. These practical suggestions and invented examples are editorial explanations; further-reading sources are not the original authors of the historical models unless expressly identified. Original overview review dates are retained.
Published 2026-09-20 · Reviewed 2026-09-22. Editorial approach