You're staring at a spreadsheet. Here's the thing — 34 mm, 12. 37 mm, 12.Also, 31 mm. Both columns describe the same parts. But column A has "Pass" and "Fail. Plus, " Column B has measurements like 12. But if you treat them the same way — same charts, same tests, same sample sizes — you're already in trouble That alone is useful..
The difference isn't academic. It's the line between seeing a problem clearly and missing it entirely Small thing, real impact..
Attribute data are counted. That's the short version. Now, variable data are measured. But the implications? They ripple through every control chart, every capability study, every decision you make about a process.
What Is Attribute Data
Attribute data lives in buckets. Here's the thing — 34 defects. That's it. On top of that, no "12. That said, scratch, dent, crack, none of the above. Good or bad. In real terms, no decimals. You count how many fall into each bucket. Go or no-go. " You either have a defect or you don't.
The Two Flavors
Binary attribute data — also called go/no-go or pass/fail — gives you two categories. Conforming or nonconforming. The part fits the gauge or it doesn't. The invoice is correct or it has an error. The customer renews or they churn.
Count attribute data — sometimes called Poisson-type data — tracks how many somethings show up. Defects per unit. Errors per invoice. Scratches per panel. Calls dropped per hour. The count can be zero, one, five, forty-two. But it's still a count. Discrete. Integer-only Turns out it matters..
You can't have 3.You can't have half a dropped call. And 7 scratches. Well, you can have a dropped call that reconnects, but that's a different definition — and definitions matter more than people admit Simple, but easy to overlook..
What Is Variable Data
Variable data lives on a continuum. Think about it: length. Weight. Temperature. Voltage. On top of that, cycle time. Torque. And fill volume. These are measurements. They carry resolution — theoretically infinite resolution, limited only by your instrument.
12.34 mm. 12.341 mm. 12.3412 mm. Each decimal place tells you something, assuming your gauge can actually see it And that's really what it comes down to..
Continuous vs. Discrete Measurements
Here's where people get tripped up. That's why variable data is conceptually continuous. But your measurement system makes it discrete in practice. A digital caliper reading to 0.Here's the thing — 01 mm gives you discrete steps: 12. And 34, 12. 35, 12.36. That's still variable data. The underlying property — length — is continuous. The resolution is just coarse Took long enough..
Contrast that with counting defects on a circuit board. Worth adding: the underlying property — number of defects — is inherently discrete. No amount of better measuring changes that.
This distinction matters when you pick statistical tools. But we'll get there.
Why It Matters / Why People Care
You might think: data is data. Run the numbers. Get the answer. But the type of data dictates everything downstream Worth keeping that in mind..
Information Content
Variable data carries more information per observation. A single measurement of 12.34 mm tells you where the part sits relative to spec, how far from target, which direction it's drifting. One pass/fail tells you... pass or fail. That's one bit of information. Literally That alone is useful..
To get the same statistical power from attribute data, you need larger sample sizes. Hundreds. Day to day, a capability study on variable data might need 30–50 parts. On the flip side, the equivalent attribute study? Sometimes much larger. Sometimes thousands Turns out it matters..
Sensitivity to Change
Variable data sees shifts earlier. You won't know until the defect rate climbs enough to be statistically detectable. Attribute data? Worth adding: a process mean drifting 0. 1 mm toward the upper spec limit shows up in your X-bar chart long before parts start failing. By then, you've been shipping marginal parts for a while It's one of those things that adds up..
This is why variable data is preferred for process control. Attribute data is often the only option for final inspection — especially when measurement is destructive, expensive, or slow.
Cost Trade-offs
Measuring takes time. Consider this: gauges cost money. On top of that, calibration takes effort. Counting is fast. A go/no-go gauge costs a fraction of a CMM. Sometimes attribute data is the only economically viable choice Most people skip this — try not to..
But — and this is the trap — the analysis cost of attribute data is higher. That's why more sophisticated models (logistic regression, Poisson regression, exact binomial intervals). You need more data. The savings at the gauge can vanish in the analyst's time.
How It Works (or How to Do It)
Let's walk through the practical workflow for each type. Not theory — what you actually do Worth keeping that in mind..
Working With Variable Data
Step 1: Verify the measurement system. Before you trust a single number, run a Gage R&R. If your measurement variation eats 30% of your tolerance, your data is noise. Fix the gauge first. Or accept wider limits. But don't pretend you have precision you don't.
Step 2: Check for normality. Most variable-data tools — Cp/Cpk, X-bar/R charts, t-tests — assume normality. Not because the universe is normal. Because the math is tractable. Plot a histogram. Run an Anderson-Darling test. If it's not normal, you have options: transform (Box-Cox, Johnson), use nonparametric methods, or fit a distribution (Weibull, lognormal, whatever fits).
Step 3: Choose your control chart.
- Individual measurements? X-mR chart.
- Subgroups of 2–10? X-bar/R.
- Subgroups larger than 10? X-bar/S.
- Rare events with variable data? Not a thing. Rare events are attribute territory.
Step 4: Calculate capability. Cp, Cpk, Pp, Ppk — know the difference. Cp/Cpk use within-subgroup variation (short-term). Pp/Ppk use total variation (long-term). If your process isn't stable, capability indices are fiction. Stability first. Always.
Working With Attribute Data
Step 1: Define the attribute clearly. "Defect" is not a definition. "Any scratch longer than 2 mm on surface A" is. "Late delivery" is not a definition. "Shipped more than 2 business days after promised date" is. Vague definitions = garbage data. This sounds obvious. It's the most common failure mode Easy to understand, harder to ignore..
Step 2: Decide: binomial or Poisson?
- Binomial: each unit is pass/fail. You're tracking proportion defective (p-chart) or number defective (np-chart).
- Poisson: each unit can have multiple defects. You're tracking defects per unit (u-chart) or total defects (c-chart).
The choice isn't philosophical. If a unit can have more than one defect, you're in Poisson territory. Worth adding: it's structural. If it's one shot — the widget works or it doesn't — you're binomial It's one of those things that adds up..
Step 3: Watch your sample sizes. Attribute charts need consistent or known-varying sample sizes. p-charts and u-charts handle varying sample sizes naturally. np-charts and c-charts need constant sample sizes. If your production batch sizes jump around, use p or u charts. Period.
Step 4: Be careful with capability. Attribute capability — DPMO, sigma level, % defective — requires large samples for precision. A sample of 50 with zero defects doesn't prove 0.1% defective. It proves you haven't
seen a defect yet. The upper confidence bound on your defect rate is roughly 6% (using the rule of three: 3/n). Practically speaking, if you need to demonstrate 100 PPM, you need zero defects in 30,000 units. Attribute capability is a sample-size game. Play it honestly or don't play.
Step 5: Beware the "defects vs. defectives" trap. A single unit with five scratches counts as one defective unit but five defects. If you feed defect counts into a p-chart (which expects defectives), your control limits will be wrong. If you feed defective counts into a u-chart (which expects defects), your limits will be wrong. This mistake survives in production dashboards for years. Check your numerator. Check your denominator. Match the chart to the count.
The Universal Rules (No Matter the Data Type)
Stability precedes capability. Always. A capability index on an unstable process is a random number with a fancy label. Run the control chart first. Look for signals: points outside limits, runs, trends, cycles, stratification. If the process isn't in statistical control, fix the special causes. Then calculate capability. The order is not negotiable And that's really what it comes down to. Simple as that..
Rational subgrouping is where studies live or die. Subgroups must capture only common-cause variation — the inherent noise of the process. Put two shifts in one subgroup and you've inflated within-subgroup variation, masking between-shift differences. Put one cavity of an 8-cavity mold in a subgroup and you've hidden cavity-to-cavity variation. Subgroup by the noise you want to filter. Subgroup by the signal you want to detect.
Control limits are not spec limits. Spec limits come from the customer (or engineering). Control limits come from the process. They answer different questions. Spec limits ask: "Is this unit good?" Control limits ask: "Is this process behaving like it did yesterday?" Confusing them leads to two classic errors: adjusting a stable process because a point is near a spec limit (tampering), or ignoring a process shift because everything is still inside spec limits (negligence).
Automation without understanding is dangerous. Software will happily calculate Cpk for a bimodal distribution. It will draw control limits for a subgroup size of one using an R-bar estimate from a moving range of 2 when the data is autocorrelated. It will run a p-chart on defect counts. The software doesn't know your process. You do. Validate the assumptions before you trust the output.
Document the operational definition. How you measure. How you sample. How you define a defect. How you handle rework. How you handle missing data. If it's not written down, it doesn't exist. And if it doesn't exist, the next shift — or the next engineer — will do it differently. That variation becomes your new noise floor.
Conclusion
Statistical Process Control is not a compliance exercise. So naturally, it is not a folder of charts for the auditor. It is a discipline for seeing your process clearly — separating the signal of real change from the noise of inherent variation.
Variable data gives you resolution. Attribute data gives you scale. Both demand the same rigor: defined terms, verified measurement, rational subgroups, stability before capability, and an operator who understands the math well enough to know when it lies It's one of those things that adds up..
The tools are standard. Now, the thinking is not. The difference between a dashboard that decorates a wall and a chart that stops a scrap event at 2:00 AM is the engineer who looked at the data, understood the context, and acted The details matter here. Still holds up..
Easier said than done, but still worth knowing.
Build the system. Here's the thing — investigate the signals. Day to day, trust the limits. That is the job.
