Data collection in Six Sigma: how to do it properly
A badly taken measurement isn't half a measurement. It's worse than none, because it gives you just enough confidence to decide wrongly.
Data collection is the phase that holds up any Six Sigma project. It's the M for Measure in the DMAIC cycle, and how well you do it decides whether the analysis finds the real cause or an invented one.
Enter your email and you will get an alert when something new is published. Unsubscribe anytime.
And the rule is harsh, but it's the one there is: a badly taken measurement is worse than no measurement. Because it gives you exactly the confidence you need to make a wrong decision.
What you decide before measuring
What exactly you're going to measure. "Defects" is not an operational definition. You have to agree what counts as a defect and what doesn't, to the point where two people measuring the same thing come out with the same number.
How much and when. A sample taken only on the morning shift, or only in the first hour, describes a process that doesn't exist. The variation between shifts, between operators and between raw material batches is usually bigger than the one you're hunting.
Who measures and with what. If the instrument isn't calibrated, or each person reads the scale their own way, what you're looking at is the variation of your measurement system. Not of your process.
| Common bias | How it gets in | How to avoid it |
|---|---|---|
| Convenience sample | you measure what's easy to measure | sampling defined in advance |
| Observer effect | people work differently when watched | measure several days, unannounced |
| Ambiguous definition | everyone counts differently | operational definition in writing |
| Rounded data | "about twenty minutes" | record the value, not the impression |
| Only the bad gets logged | faults recorded, normal running isn't | record the correct process too |
The last one is the important one
In almost every plant I've walked into, the records are biased towards failure. The breakdown gets logged, the stoppage, the defect. When everything is fine, nothing gets logged.
And what does that leave you with? You can count how many problems there were. But you have nothing to compare them against. And without comparison there's no cause analysis worth the name.
The bridge to 2026
This is the phase that's changed most, and it strikes me as the biggest gift technology has given Six Sigma.
When the record depends on a person writing a number on a sheet while doing their job, you have three guaranteed problems: it gets logged late, it gets logged rounded, and only the eye-catching stuff gets logged.
And note, that's nobody's negligence. It's that this person has another task in hand and yours is the second one.
When it's the line itself doing the recording — times, stoppages with their reason, an image of every piece, the operator's reaction time — the three biases disappear at once. You have a hundred per cent of the population instead of a sample. You have the normal data as well as the anomalous. And you have it the moment it happens.
The Measure phase stops being a three-week project and becomes your process's default state. Everything that comes after it in DMAIC rests on that.