How skill assessments report scores: raw, scaled, percentile and cut score

A score report is a number wearing a label, and the label does most of the work. The same figure can mean a count of points, a position in a group, or a pass, depending on which score word sits beside it. Most misreadings start by skipping that word.

Read in order, the score words show what was done to a result before it reached you. The sections below follow that order, from the raw count to the decision rule at the end.

What to take away

From raw score to scaled score

A raw score is the count of correct answers or points awarded, before any transformation. It is the plainest number an assessment produces, and the least comparable. A raw score on one version of a test says little about a raw score on another.

A scaled score is a raw score converted onto a common scale so that results from different versions of an assessment can be compared. The conversion is part of the instrument's design.

A scaled score cannot be interpreted without knowing the scale. If a report gives only the number, ask which scale it sits on before comparing it with anything.

Percentage, percentile and percentage point

Percentage is a share of a whole, out of a hundred. Percentile is something else: the share of a comparison group that scored at or below this result. A percentile is a statement about a population, not about capability.

Moving from one comparison group to another changes the percentile without changing what the person can do. A strong result against a weak group and a modest result against a strong group can carry the same percentile.

A percentage point is the arithmetic difference between two percentages. The same trio of terms causes the same trouble in business reporting, and the metric words guide sets them out there.

Purely as arithmetic, with figures invented for the illustration: a move from a 40 percent result to a 44 percent result is a rise of four percentage points and a relative increase of ten percent. Both statements describe the same change and they sound very different, which is why the unit has to be stated.

Which group, or which standard?

Norm-referenced results are interpreted against how others performed. They answer the question of compared to whom. Criterion-referenced results are interpreted against a defined standard of performance, and they answer whether the person can do the thing.

If the question is whether someone is ready to do a job, a norm-referenced result answers a different question than the one asked. A percentile tells you rank. It does not tell you readiness.

The norm-referenced or criterion-referenced choice is therefore made before the score exists, when the instrument is designed. Check which one the report uses before reading anything into the number.

Cut score: where a number becomes a decision

A cut score is the threshold at which a decision changes. It is set by human judgment against a defined procedure, it is not discovered in the data, and the procedure used to set it is a fair thing to ask about.

Every threshold trades one error against the other. Moving a cut score to reduce false positives increases false negatives, always. A claim that a change improves accuracy without a cost should prompt the question of which error was traded away.

In the United States, the Equal Employment Opportunity Commission states that under Title VII employers are not permitted to adjust scores or use different cutoff scores on the basis of race, color, religion, sex, or national origin.

Choosing where the line falls is itself a decision, and the decision skills guide shows why reversibility sets the process around it. A line that is hard to undo deserves more scrutiny before it is adopted.

A pass mark presented as a property of the test is a judgment with its authorship removed. A person set it, using a documented procedure, and you are entitled to see that procedure.

Which instrument produced the score

Four instrument words explain a lot about how trustworthy a score is. Each records something different, and each has a weak point that the score itself hides.

The Office of Personnel Management describes work sample tests as ones that require applicants to perform tasks that mirror the tasks employees perform on the job. That closeness buys content validity directly, and the trade-off is real and should be made deliberately.

What the score is evidence for

A score supports an inference, and several kinds of validity name different inferences. They are not degrees of the same thing.

The third row is the one that matters for selection and the one least often evidenced. The fourth is a perception rather than evidence, and it is routinely presented as though it were the other three.

An unreliable instrument cannot be valid, because it is not measuring anything stably. A perfectly reliable one can be entirely invalid: it can measure the wrong thing, precisely, every time. Vendors quote reliability more often, because it is easier to produce.

Checks before you rely on a score

For the wider set of questions worth asking about any term before you rely on it, the same habit applies: name the term, then name what it was built to show.

Common questions

Is a percentile the same as a percentage?

No. A percentage is a share of a whole. A percentile is the share of a comparison group that scored at or below this result. One describes the test, the other describes the group, and mixing them up is a common reading error.

Is a high score always good evidence?

Only for the inference the instrument was built to support. Validity is a property of the instrument used for a particular purpose. The same test can be valid for placing someone in a course and invalid for deciding a promotion, so ask what the score is being used for.

Who decides where the pass mark sits?

A person, using a documented procedure, and the procedure should be available to you. A threshold presented as a property of the test is a judgment with its authorship removed.

Does a result expire?

In effect, yes, because capability changes and conditions change. Treating a result from a year ago as a current fact about a person is one of the commonest misuses, and it is not something the instrument itself can prevent. Words that carry two silent meanings make it worse.