A scatter diagram is a chart that plots two variables against each other, one on the x-axis, one on the y-axis, to see whether they move together. Each point is one paired observation. The pattern of the cloud tells you whether a suspected cause and an effect are related, how strongly, and in which direction. It shows correlation, never proof of cause.
On the plant floor the scatter diagram answers a specific, recurring question: someone thinks X drives the defect, and you want to know if the data agrees before you spend money changing X. Oven temperature and seal failures. Line speed and label misalignment. Ambient humidity and print smearing. Plot the pairs, look at the cloud, and you get a fast, honest read on whether the hunch holds up. It is one of the seven basic quality tools and the natural next step after a fishbone diagram has narrowed the suspects.
What does a scatter diagram show you?
A scatter diagram shows the direction and strength of the relationship between two measured variables. Direction is whether the points trend up, down, or nowhere. Strength is how tightly the points hug that trend versus scattering into a shapeless blob. Both are read by eye first, and both matter: a strong relationship in the wrong direction is still a finding, and a weak one is a signal to stop chasing that variable.
Statisticians put a number on that tightness, the correlation coefficient, written r which runs from −1 (a perfect downward line) through 0 (no linear relationship) to +1 (a perfect upward line). You do not need to compute r to use a scatter diagram, and reading the cloud by eye is often enough to make a decision. But the number is useful for one thing: it keeps you honest about weak relationships that look more convincing than they are.
How do you build a scatter diagram?
You build a scatter diagram by collecting paired data for the two variables you suspect are related, plotting each pair as a point, and reading the resulting cloud. The discipline is in the pairing and the sampling, not the plotting.
- State the hypothesis in plain words. "Higher oven temperature causes more seal failures." Naming the suspected cause and effect up front keeps you from fishing for any relationship in the data afterward.
- Decide which variable is which. Put the suspected cause on the x-axis and the effect on the y-axis. It does not change the math, but it keeps the reading consistent across your team.
- Collect paired measurements. Each data point needs both values measured at the same time, on the same unit or batch. Thirty to fifty pairs is a reasonable starting range; fewer than about 25 and random noise can fake a pattern.
- Cover the real operating range. If oven temperature only varied by two degrees during your sample, the plot can not reveal a relationship that only shows up across ten degrees. Sample across the range the variable actually moves through.
- Plot every pair as one point. Do not average, do not bin, one observation, one dot. Averaging hides the spread that makes a scatter diagram worth building.
- Read direction, then strength. Up, down, or nowhere first; then how tightly the points hug the trend. Optionally compute r to pin down borderline cases.
- Watch for stratification. If the cloud looks like two separate clusters, a hidden variable, two machines, two shifts, two material lots, is probably mixed in. Color the points by that variable and re-read; the pattern often splits cleanly.
- Decide and verify. A relationship is a lead, not a conclusion. Confirm it by deliberately changing the suspected cause and watching the effect, ideally under statistical process control so you can tell a real shift from noise.
By the numbers. The American Society for Quality classifies the scatter diagram as one of the seven basic quality tools and defines it as a graph that displays pairs of numerical data, one variable on each axis, to look for a relationship between them. ASQ notes the scatter diagram shows correlation but does not by itself prove that one variable causes the other. See ASQ's scatter diagram resource and its overview of the seven basic quality tools.
How do you judge how strong the relationship is?
You judge strength by how tightly the points cluster around an imaginary line through the cloud. Points strung along a narrow band signal a strong relationship; the same general slope with points scattered widely on either side signals a weak one that may not survive a second sample. Three practical cues help you read it without any math. First, look at whether you could draw a single line that most points sit near, if several equally good lines fit, the relationship is weak. Second, watch the tails: a relationship that holds in the middle but falls apart at high or low values is often a sign the effect only kicks in past a threshold, which is useful to know. Third, check for a curve. A scatter diagram can reveal relationships that are not straight lines at all, an effect that climbs, plateaus, then climbs again, and a straight-line correlation coefficient will badly understate a real but curved relationship. When the pattern bends, trust your eyes over r.
What are the common mistakes with scatter diagrams?
Most bad scatter diagrams fail before a single point is plotted, in how the data was gathered.
- Too narrow a range. If the suspected cause barely moved during your sample, the plot cannot show a relationship that only appears across a wider swing. Sample deliberately across the full operating range.
- Mismatched pairs. Measuring the cause on one batch and the effect on another, or at different times, breaks the pairing. Every point must be two values from the same unit at the same moment.
- Too few points. A dozen points can form a convincing-looking line purely by chance. Get to a few dozen before you trust the shape.
- Hidden stratification. Mixing two machines or two shifts into one plot can hide a relationship or invent one. When the cloud splits into clusters, separate the groups and re-read.
- Stopping at the plot. The most expensive mistake is treating a correlation as a conclusion and changing the process on the strength of a cloud. The plot is a lead; the confirmation trial is the finding.
Why is correlation not the same as causation?
Correlation means two variables move together; causation means one makes the other happen, and a scatter diagram can only ever show the first. Two variables can track each other tightly for reasons that have nothing to do with one causing the other, and mistaking the plot for proof is how teams spend money adjusting the wrong knob.
| Why they correlate | What is really going on | Plant example |
|---|---|---|
| Direct cause | X genuinely drives Y | Higher oven temp actually weakens the seal material |
| Reverse cause | Y drives X, not the other way around | Slower line speed is a response to defects, not their cause |
| Lurking variable | A third factor drives both | Humid days raise both ambient temp and print smearing; heat is not the cause |
| Coincidence | No real link at all | Two unrelated trends that happen to rise over the same months |
The lurking variable is the one that catches good teams. Two effects that share a common cause will correlate beautifully with each other while neither causes the other. The defense is the same discipline that makes any root-cause work honest: once the scatter diagram points at a suspect, confirm it by changing the suspected cause on purpose and watching whether the effect follows. If it does, under controlled conditions, you have a cause worth acting on. If it does not, you just saved yourself an expensive false lead. That confirmation step is why the scatter diagram lives alongside the 5 whys and controlled experiments rather than replacing them.
Where does the scatter diagram fit among the quality tools?
The scatter diagram is the relationship tester in the toolkit, it comes after you have a suspect and before you commit to a fix. A typical chain runs from a Pareto chart that picks the biggest defect, to a fishbone that brainstorms causes, to a scatter diagram that tests the most plausible cause against real data, to a confirming trial tracked on a control chart. Each tool hands off to the next; the scatter diagram's job is to keep you from acting on a hunch that the numbers do not support. For the full set and when each one earns its place, see the seven basic quality tools.
All of it depends on paired data you can trust. A scatter diagram built on measurements taken at different times, on different units, or transcribed by hand at shift end is a scatter diagram built on sand. When process variables and outcomes are captured together in real time at the line, the pairs are genuine and the plot means something, which is the data foundation CLS built by moving off paper logging. Get the pairing right, read direction before strength, and never let a tidy cloud talk you out of the confirmation step.