Scatter Plot & Scatterplot Matrix: Test One Relationship Properly Before You Act On It

The Scatter Plot and Scatterplot Matrix tool sits in the Analyse phase of RDMAICS. Its job is narrow and useful. It lets you study one relationship in detail, or screen three to six numeric variables together in a true pairwise matrix.

You reach for it after you have measured something. By the Analyse phase you usually hold a list of suspect causes. Some of those suspects are numeric. Cycle time, temperature, pressure, queue length, staffing level. A scatter plot is the fastest honest way to ask whether two of those numbers move together.

Use the single plot when you already have a candidate pair and you want detail. Use the matrix when you have several numeric columns and no clear favourite. The matrix is a screening device. It narrows the field. The single plot then does the careful work on the pair that survived screening.

The tool reports Pearson r, R squared, slope and a plain strength label. It can colour points by a grouping column and fit a separate trend line for each group. It also offers an AI review that explains the result in words.

Everything below is built on one case. A fictional manufacturing area runs two production lines, Line A and Line B. Each line runs day and night shifts. The team records cycle time and defect count for eighteen production runs, along with temperature and pressure.

This case is fictional. It is the sample data that loads automatically when you open the tool in Ci Flow, so you can follow every step without supplying your own file.

The team believes defects fall as cycle time rises. The theory is that slower runs allow more care. They want to know whether the data supports that belief, and whether the pattern holds on both lines.

The data table carries six columns. Two are marked with an asterisk because the tool requires them. Cycle time is the X variable. Defects is the Y variable. The remaining columns are optional context.

Cycle time rises in even steps across the eighteen rows, from 18 at row 1 to 45.2 at row 18. Defects move the other way, from 15 at row 1 down to 8 at row 18, with plenty of wobble on the way. Temperature cycles between 68 and 77 and repeats. Pressure climbs from 28 to 35.95 with small dips.

Two labelling columns matter. Group or stratum holds Line A for rows 1 to 9 and Line B for rows 10 to 18. Series or scenario alternates Night and Day. These are the columns that let you cut the picture apart later.

The input panel is short. You pick the X variable, the Y variable, the group column and the series column. You then choose what to colour by. In the screenshots, colour by is set to Group or stratum, regression by group is set to yes, and swap X and Y is set to no.

The result panel reports four headline numbers for all eighteen rows together. N is 18. Overall r is -0.9363. Overall R squared is 0.8766. The slope is -0.3283. The strength label reads Very strong.

The negative sign confirms the direction the team expected. As cycle time rises, defects fall. The slope of -0.3283 is the rate of that fall in the units of the data. R squared of 0.8766 tells you how much of the variation in defects is accounted for by the straight line through cycle time.

Then read the group table underneath, because it changes the story. Line A has N of 9, Pearson r of -0.7506, R squared of 0.5633 and slope of -0.2708. The strength label reads Strong. Line B has exactly the same four figures and the same label.

So within each line the relationship is weaker than the overall figure suggests. Overall r of -0.9363 against group r of -0.7506 on both lines. Overall R squared of 0.8766 against 0.5633 within each line. The pooled number is partly carrying the difference between the two lines, not the relationship inside them.

A scatter plot does not make a decision for you. It changes which question you ask next. In this case it supports three moves.

First, it keeps cycle time on the shortlist of factors worth testing. The direction is consistent and the within group relationship is labelled Strong on both lines. That is enough to justify a designed experiment or a controlled trial.

Second, it raises the line itself as a variable. Line B runs at longer cycle times and lower defect counts than Line A across the whole range. Rows 10 to 18 sit entirely above 32 on cycle time and at or below 11 on defects. That gap is worth investigating in its own right, because it may be equipment, method or people rather than cycle time.

Third, it tells you to be careful with the overall figure when you report results. If you quote 0.8766 without the group table beside it, you overstate how much of the defect variation cycle time explains inside a single line.

Most errors with this tool come from reading one number and stopping. The headline strength label is a summary, not a verdict. A few habits will keep you out of trouble.

Turning regression by group off is the quickest way to miss what is happening here. With one pooled line you see -0.9363 and move on. With two lines you see -0.7506 twice and an offset between the lines.

Swapping X and Y is another trap. The tool has a swap control, and the correlation is unchanged by swapping, but the slope is not. Keep the variable you can control on the X axis so the slope reads in a useful direction.

This tool will not tell you that cycle time causes defects. Pearson r measures how closely two numbers track a straight line. Nothing more. The -0.9363 in this case is consistent with cycle time driving defects, with defects driving cycle time, and with some third factor driving both. The data cannot choose between those.

It will not find non linear patterns for you. The r and the slope assume a straight line. If the true relationship curves or flattens past a threshold, the straight line fit will still return a number and a label, and the number will mislead you. You have to look at the plot.

It will not handle the confounding it reveals. The tool shows you that the two lines differ. It does not adjust for that difference, pool the groups correctly, or produce an estimate of the within line effect after removing the line effect. That is a job for a different method.

It will not judge whether eighteen rows is enough, whether the data was collected properly, or whether the measurement system is sound. N of 18 and 9 per group is small. Nothing in the output warns you about that. It will also not tell you whether the sampling covered the full operating range, and in this sample the temperature column simply repeats the same nine values twice.

Finally, it will not replace an experiment. The honest output of a good scatter analysis is a better hypothesis and a shorter list. In this case the list is cycle time and the line difference. Testing them is the next phase of work, not this one. The AI review will help you put the result into words, but the test still has to be run.