Data types and measurement scales
Variables vs attribute data, the four measurement scales, and why the type of data decides the chart and the test.
Every analysis starts with knowing what kind of data you have. The type of data decides which control chart, which capability method and which hypothesis test are valid.
Variables and attribute data
Variables data (also called continuous or measurement data) come from measuring: a diameter of 10.012 mm, a fill weight of 501.3 g, a cycle time of 42.7 s. Values can fall anywhere on a scale and carry a lot of information per reading.
Attribute data (also called discrete or count data) come from counting or classifying: pass or fail, number of scratches on a panel, number of late shipments. Each reading carries less information, so attribute methods need much larger samples to detect the same change.
- A part either passes the go/no-go gauge or not: attribute (defective).
- The number of blemishes on a painted door: attribute (defects).
- The measured bore diameter: variables.
The four measurement scales
- Nominal: categories with no order (machine A, B, C; defect type).
- Ordinal: ordered categories with uneven steps (poor, fair, good; severity 1 to 10).
- Interval: equal steps but no true zero (temperature in °F or °C).
- Ratio: equal steps and a true zero (length, weight, time, count).
Means and standard deviations make sense for interval and ratio data. For ordinal data, medians and nonparametric methods are safer.
Defects vs defectives
A defect is a single nonconformity. A defective (nonconforming unit) is a unit with at least one defect that makes it unacceptable. One unit can have three defects and still count as one defective. Defectives use the binomial distribution (p and np charts); defects use the Poisson distribution (c and u charts).
Questions and answers
Should I convert measurements into pass/fail?
Avoid it whenever you can. Turning a diameter into pass/fail throws away most of the information. You need far more parts to see a shift, and you lose the warning that parts are drifting toward a limit before they fail.
Is a count, like the number of defects, continuous?
No. Counts are discrete. When counts are large (dozens or more per sample) they often behave close enough to continuous data, but small counts need Poisson or binomial methods.
My data are percentages. Which type is that?
It depends on where the percentage came from. A percentage of defective units is attribute (binomial) data. A percentage concentration measured by an instrument is variables data.
