Outlier Analysis: Trace Every Unusual Value Back to Its Source Row

Outlier Analysis sits in the Analyse phase of RDMAICS. Its job is narrow and useful. It traces unusual observations to their source row, screens them within the correct process stratum, and compares results without deleting the original data.

You reach for it when a mean looks wrong, when a standard deviation looks larger than the process feels, or when a single value is dragging a chart. Before you test a hypothesis or fit a model, you want to know which observations deserve a conversation with the people who run the process.

The tool does not clean your data for you. It builds a register of candidates and gives you a way to test how much they matter. That distinction runs through everything below.

This case is fictional. It is the sample data that loads with the tool, and every number quoted in this guide comes from that dataset or from the result the tool calculated.

The dataset holds 24 observations of cycle time, labelled OBS-001 through OBS-024. Twelve belong to the Day group and twelve belong to the Night group. Each row carries a time or sequence value from 1 to 24, so the order of the observations is preserved.

Day values cluster between 47 and 52, with one exception. OBS-011 records 120. Night values cluster between 68 and 73, again with one exception. OBS-023 records 35. Both of those rows carry the note Investigate source record.

Notice what the two exceptions have in common and what they do not. Both sit far from their neighbours. But 35 would look unremarkable if you pooled all 24 values together, because it falls between the Day cluster and the Night cluster. That is the reason stratum matters.

The input panel asks one framing question first: what should the screen help you decide? In this case the answer is set to which observations require investigation. That question shapes how you read the output. You are not asking which values to delete.

Then you map the columns. Value is set to Measurement. Group is set to Group or stratum, which separates Day from Night. Time is set to Time or sequence. Observation is set to Observation or ID, and this is the field that makes traceability possible. Without it, a flag is a number. With it, a flag is a row you can go and look up.

Screening method is left on Automatic, which the tool describes as robust screening recommended. That runs two screens together: Tukey 1.5 times the interquartile range, plus a modified Z-score with a threshold of absolute value greater than 3.5.

The first output is the within-stratum outlier diagnostic. Two boxplots sit side by side, Day with n equal to 12 and Night with n equal to 12. The axis runs from 29.9 to 125.1. Whiskers stop at the most extreme value inside 1.5 times the IQR. Every dot stays visible, and red rings mark the flagged observations. Nothing disappears from the picture.

The statistical results table gives you the arithmetic. Day has N of 12, mean of 55.4167 and standard deviation of 20.3892. Day IQR fences run from 45.375 to 54.375. Day median is 50 and MAD is 1. One Day observation is flagged. Night has N of 12, mean of 67.5 and standard deviation of 10.3353. Night fences run from 65.625 to 74.625. Night median is 70 and MAD is 1. One Night observation is flagged.

Compare the mean and the median within each stratum. Day mean is 55.4167 but Day median is 50. That gap is the signature of a single large value pulling the average. Night mean is 67.5 and Night median is 70, pulled the other way by the low value.

Both screens agree. IQR outliers are OBS-011 at 120 and OBS-023 at 35. Modified Z outliers are the same two records. The executive interpretation states that 2 of 24 observations were flagged within 2 process strata using the automatic robust screen.

Below the plot sits the outlier investigation register. It lists OBS-011, Day, 120, flagged by IQR and Modified Z. It lists OBS-023, Night, 35, flagged by the same two screens. Each row offers three decisions: Investigate, Keep, or Exclude from sensitivity only. Each row also has a field for a documented reason.

The register states plainly that the original worksheet remains unchanged, and that Exclude applies only to the sensitivity comparison below. This is the design choice that makes the tool safe to use. You never edit the source.

The comparison shows four figures. Original mean is 61.4583 and sensitivity mean is 61.4583. Original SD is 16.9705 and sensitivity SD is 16.9705. They match because no exclusion has been applied yet. Once you mark a record as excluded from sensitivity only, the right-hand pair moves and the left-hand pair does not. The size of that movement is your evidence about whether the point changes your conclusion.

The decision-ready conclusion states that potentially unusual observations exist, and asks you to record a process or measurement reason, then use the sensitivity comparison before deciding how the points should affect inference. That is the order of work.

The next actions follow the same logic. Verify data entry, measurement-system and assignable-cause explanations first. Use the register to keep, investigate or exclude without altering the source worksheet. If groups represent different process populations, continue to analyse them separately.

For this case, that last point carries weight. Day centres near 50 and Night centres near 70. Those are not two samples of one process. Treating them as one would hide the difference and would also hide OBS-023, because 35 only looks strange against the Night cluster.

The tool also offers an AI review that explains the result in plain language if you want a second reading of the output.

The first mistake is pooling. If you leave the Group field empty, both strata collapse into one distribution and the fences widen. OBS-023 at 35 may survive that screen unflagged. Set the stratum before you set anything else.

The second mistake is deleting. A flag is a question, not a verdict. The tool repeats the phrase investigate, never delete automatically against each stratum for a reason. A value of 120 may be a transcription error, a changeover, a different product, or a real event worth understanding.

The third mistake is skipping the documented reason. An empty reason field turns a defensible analysis into an unexplained one. If a reviewer asks why 120 was excluded six months from now, the register is your answer.

Outlier Analysis will not tell you why a value is unusual. It flags OBS-011 at 120 and OBS-023 at 35. It cannot tell you whether the clock was stopped late, whether the operator was new, or whether the row was typed twice. That answer lives in the process, not in the dataset.

It will not decide whether to exclude a point. The register offers Keep, Investigate and Exclude from sensitivity only, and the choice stays with you. The tool records the choice and shows the consequence through the sensitivity comparison. It does not make the judgement.

It will not validate your strata. If you declare Day and Night as groups, the tool screens within those groups. It cannot tell you whether a third stratum, such as product type or shift lead, would explain the pattern better. Twelve observations per group is reported as adequate for screening, and that is a statement about screening only, not about the strength of any conclusion you draw.

It will not prove causation, confirm stability over time, or replace a control chart. It uses the sequence field to preserve order but does not test for trends or shifts. And it will not fix a measurement system. If your gauge is unreliable, every screen below it inherits that problem.

Use it for what it does well. It gives you a short, traceable, stratum-aware list of observations that deserve a conversation, and a safe way to test how much they change the answer.