🎯 Quick Answer

What: Correlation vs causation is the difference between "these two things move together" and "this one makes that one happen." Any two metrics that move together are carrying four possible stories: the first drives the second, the second drives the first, a third variable drives both, or the pattern is chance. The chart alone cannot tell you which.

Why it matters: Operations act on charts. When a correlation is read as a cause and nobody tests it, the fix aims at a passenger while the real driver keeps working. The money is spent, the metric does not move, and the belief usually survives the failure.

How to apply: Treat every pattern as a candidate, not a conclusion. Stratify the data by the suspected factor, compare the groups, and run the appropriate hypothesis test before any cause is written into a finding. A cause that cannot say how it was verified is still an opinion.

The payoff: Improvement budgets land on drivers instead of passengers, and the beliefs that survive testing become assets, because now they are known rather than assumed.

The chart hung in the production office for over a year, and it was true.

Two lines, month by month: overtime hours and the defect rate, rising and falling together. Everyone read it the same way. Tired people make mistakes. The chart settled arguments, shaped a budget, and eventually became a policy: overtime was capped.

Defects did not move. Deliveries slipped, because the hours had been real work. And when a small team finally pulled the chart apart, the story underneath came apart with it. The defects were not living in the overtime weeks at all. They were living in expedited orders, whichever week those landed in, because rush jobs quietly skipped a setup verification step. Rush jobs also generated the overtime. The two lines moved together because a third thing was pushing both, and the third thing was on nobody's chart.

That is the correlation trap, and versions of it sit on the wall of almost every operation. This note is about how to read such a chart honestly: what a correlation actually proves, what it never can, and the small discipline that separates the two.

What a correlation actually says

A correlation says one thing: these move together. Said precisely, that is valuable. It narrows a search, nominates a suspect, tells you where looking is likely to pay. In Lean Six Sigma work, a strong pattern in the data is one of the best reasons to put a candidate on the cause list in the first place, alongside the ones the team gathers by walking the process.

What a correlation never says is why. Correlation vs causation is not a statistics course distinction; it is the difference between a lead and a verdict. Treating the lead as the verdict is how an operation ends up capping overtime to fix a defect problem.

The four explanations

Any two metrics that move together are carrying four possible stories, and all four fit the same chart.

The first drives the second. Sometimes true, and it is always the story that gets told, because it was the reason anyone drew the chart.

The second drives the first. Arrows point backward more often than teams expect. Errors create rework backlogs, and backlogs pull in the extra hands that then get blamed for the errors. Ask what order things actually happened in; the calendar kills more theories than the math does.

A third variable drives both. The quiet third thing. Volume, season, mix, a scheduling practice, an expediting habit: something that pushes both lines at once and appears on neither. This is the most common resolution of a beloved operational chart, and the least visible, precisely because the third thing was never charted.

Chance. Two lines can simply keep company for a while. The shorter the history and the more charts an operation draws, the more often this happens. Somebody somewhere has a convincing chart correlating two things that have never met.

One chart, four stories. Reading the chart harder does not pick between them. Only a test does.

How to test a pattern

The correlation vs causation question is settled by splitting the data, not by staring at it. The test is not exotic, and it does not require anyone to distrust their own numbers.

Stratify the results by the suspected factor and by the plausible third things, and let the groups argue. In the overtime story, one split settled it: expedited orders against routine ones. Defects tracked the expedites regardless of overtime, and the fatigue story ended that afternoon. In a warehouse version of the same trap, quiet weeks with a high share of temporary staff showed normal error rates, and busy weeks ran high with temps or without, which retired a two-year-old belief about temps and errors in one afternoon.

Where the stakes justify it, the comparison graduates into a formal hypothesis test, which asks whether the difference between the groups is bigger than the process's own ordinary variation. That discipline has its own field note. The point here is smaller and comes first: whatever the tool, a causal claim must survive one honest attempt to break it before anyone spends money on it.

The standard worth adopting is a single line of paper. No cause reaches a finding without a note saying how it was verified. A root cause that cannot fill that line in is still a hypothesis, however many meetings it has survived.

Where this sits in a project

In the DMAIC arc, this discipline is the heart of the Analyze phase, where correlation vs causation stops being a slogan and becomes a work step. Measure ends with an honest picture of what the process does, and the wastes and delays it carries. Analyze asks why it does it, and the why is exactly where correlations pour in, because patterns are cheap and verdicts are not. A team that lets moves-together stand in for causes exits Analyze with a confident list of passengers. A team that tests before it writes exits with fewer causes and better ones, and the Improve phase spends its budget on things that are actually driving.

The habit scales down to a sentence. When a pattern is offered as an explanation, in a review, on a chart, in a corridor, ask what else moved. It is four words, it offends nobody, and it is the cheapest analysis step in the entire methodology.

I write one of these every week, drawn from what actually happens on projects rather than what the methodology says should. You can join the newsletter here.

FAQ

What is the difference between correlation and causation?

Correlation means two things move together: when one is high, the other tends to be high, or low. Causation means one of them actually makes the other happen. Correlation is a fact about a chart; causation is a claim about the world. Every causal relationship produces a correlation, but most correlations are not causal: the arrow may point the other way, a third variable may drive both, or the pattern may be chance.

Can a correlation ever prove causation?

Not by itself, no matter how strong or how long it has held. What moves a correlation toward a causal finding is testing: splitting the data by the suspected factor and the plausible third variables, checking the time order, running a hypothesis test against the process's ordinary variation, and, where possible, changing the suspected cause deliberately and watching whether the effect follows. A correlation that survives all of that has earned the word "cause."

What is a third variable, in plain words?

Something that pushes both of the things you charted, while appearing on neither axis. Volume is the classic one in operations: a busy season can raise overtime, temporary staffing, congestion, expediting and error rates all at once, making any two of them correlate beautifully with each other. The third variable is why the most convincing chart in the building can still be telling the wrong story.

How do I test a correlation in an operation without a statistician?

Start with stratification, which needs no software: split the results by the suspected factor and compare the groups. Then check the calendar: did the supposed cause actually come first? Then look for the third thing: what else changed at the same time, and what happens to the pattern inside weeks where that third thing was flat? Most operational correlations die or survive on those three checks alone. The formal hypothesis test comes last, when the decision is big enough to deserve it.