Two People, One Total
A total can hide how it was reached, and a stable average can hide who changed places. Two small examples show what a score keeps and what it drops.
Browse the directoryTake a made-up ten-item scale, each item scored from 0 to 10, with no real instrument behind it. Person A answers 5 on every item. Person B answers 8 on five items and 2 on the other five. Both total 50. The totals are equal, and that is all the arithmetic shows. A is steady and B is split, and a total cannot reveal which pattern produced it. Whether the split matters depends on what the scale was built to measure, which the numbers themselves cannot tell you.
That is why the first step in reading any score, as the guide to reading a measure puts it, is to name the interpretation you want to make. A construct is the attribute of interest, such as perceived stress. An operation is the observation used to study it, such as a questionnaire total. Perceived stress, the number of stressful events and a physiological stress response are three different targets, and a questionnaire about one does not stand in for the others.
A score used to rank people needs a reference group. Take a fictional raw total of 24. In one group 18 of 100 people scored below 24, and in another 62 of 100 did. The same raw total sits at different positions.
The second example is about repeated measurement. In a task like the Stroop color-word task, people name the ink color of a word that may spell a different color. The extra time taken when the word and color disagree is the conflict cost, a difference between two response times. Suppose four fictional participants have conflict costs of 60, 70, 80 and 90 milliseconds in one session. In a second session their costs are 80, 60, 90 and 70. The average is 75 milliseconds both times. The order of the four people has changed.
In these two sessions the average held still, and the individual order did not. An unchanged average does not by itself show that individuals keep their places, and treating the first session's order as stable would go beyond what these numbers support. Reliability for ranking individuals asks whether people keep similar places across occasions, and this fictional task does not show that. Whether the group-level effect itself is dependable is a separate matter that two sessions cannot settle. Validity is a further question: whether a score supports the meaning you give it. A conflict cost estimates interference under a particular protocol. It does not read out a single inhibition capacity.