Correlation vs. Causation
Correlation vs. Causation
What are correlation and causation?
- Moving together doesn’t mean one causes the other
- Suspect a third factor, reverse direction, or self-selected volunteers
- Split into cohorts and compare with the group not yet trained
The sentence training managers most want to write is 'Employees who took the training performed better.' The data often does show that. But what this sentence shows is only that training completion and performance moved together. Claiming that training raised performance requires separate proof. If you don't distinguish the two sentences, next year's budget goes to a course that doesn't work while the course that did work gets cut.
A correlation without causation falls into three cases. The first is a third factor: teams with more manager support send more people to training and also perform better. The second is reverse direction: high performers may have been the ones picked for training. The third is selection: people who volunteer are highly motivated to begin with, so they would likely have performed well even without the training. With a small sample, chance correlations are also common.
To get closer to causation, you need something to compare against. The strongest method is to randomly split participants and train only one group first, but in practice training in sequential cohorts and comparing with the group not yet trained works well enough. If that is also hard, measure the same metric before and after training, and match people with similar job level, tenure, and past performance. A correlation coefficient ranges from -1 to 1 and shows only how closely two values move together, not the size of any cause.
Where things go wrong in practice is the verb in the report. The data only says scores were 'higher,' but the sentence that reaches executives has changed to 'raised.' Reading a highly rated course as proof of strong on-the-job results is the same trap. The opposite is no answer either: saying nothing because causation can't be proven. Reports earn more trust when they separate what was confirmed from what was inferred, sentence by sentence, and, if there was no comparison group, say so and add how the next cohort will check it.
Inwoo LeeExpert in leadership, communication, and performance management
John KimAuthor of B2B Marketing Design and CEO of Nanum Consulting| Category | Correlation | Causation |
|---|---|---|
| Meaning | Two values move together | One causes the other |
| How to check | Compare two values side by side | An experiment with a comparison group, or a comparable design |
| The verb in the report | was higher with…, rose along with… | "X raised Y" or "Y grew because of X" |
| An example from training practice | Completers have higher performance ratings than non-completers | Of two classes split by lottery, the one trained first improved on the metric |
| Common pitfalls | A third factor, reverse direction, self-selection | Drawing conclusions from before-and-after differences alone, with no comparison group |
The principle that correlation does not imply causation is a basic proposition long emphasized in statistics and research methodology. It has become a shared norm for interpreting data rather than the claim of any one person. It is also widely held that confirming causation requires an experiment with a comparison group, or a comparable design.
Consider a hypothetical case. The talent development team at Company A, an insurer with 1,500 employees, had to report the results of a voluntary data analytics course. Among 120 completers, 45% received a top year-end rating, versus 22% of non-completers, and the draft stated, "The training doubled performance."
- When the HR development manager sent the draft up to the division head, a question came back: "Aren't the people who applied the ones who were already strong? What were their ratings last year?"
- The manager checked last year's ratings for the 120 completers and found that 41% already held top ratings. Even before the training, they were a group with nearly double the top-rating share of non-completers.
- The team lead rewrote the sentence. It now read "Completers had a higher share of top ratings," with the limitation noted below it: "Effect of applicant characteristics could not be ruled out."
- The following year, the team split the 240 applicants into a first-half class and a second-half class by lottery. The manager explained, "While the second-half class waits, they serve as the comparison group."
- At the end of June, they compared the number of analysis reports written by the two classes. The first-half class averaged 3.1 per person per month, while the second-half class, not yet trained, averaged 2.2.
- Based on this result, the division head made the course part of the regular program, and the wording in the report changed from "double" to "an increase of 0.9 per month." The following year's budget review passed without questions.