Scoring Anchors

Scoring anchors are the raw values that a preference scale maps to its end points, usually 0 and 100, together with the rule that converts values between and beyond them. They turn a measurement with a unit and a direction, such as minutes per case where lower is better, into a score that can be weighted. The anchors are value judgments chosen by the assessors, not properties of the measurement.

State the end points, the rule, and the cap

Suppose clerk handling time is anchored at 7 minutes for 0 points and 5 minutes for 100. A linear rule gives (7 − minutes) / 2 × 100, so 6.5 minutes scores 25 and 5.4 minutes scores 80. Linear means equal improvements earn equal points anywhere in the range; if the first minute saved matters more than the second, a curved rule would be needed, and it should be stated. Values beyond an anchor are capped, so 4.5 minutes scores 100, not 125. A measurement so poor that the option is unacceptable belongs to a gate, not to a negative score.

Anchors are scoring definitions. They do not replace a target or an acceptance condition: a six-minute workflow target can be a preference in one evaluation and a gate in another, and the anchors stay what they are.

Fixed anchors versus relative scaling

The UK multi-criteria analysis manual distinguishes relative scales, where the least preferred option present scores 0 and the most preferred 100, from fixed scales, where 0 can be the lowest acceptable value and 100 the maximum feasible. Under relative scaling, adding or removing an option can move the best or worst anchor, and then the scores of unchanged options change too: with options at 5.4 and 5.8 minutes, the first scores 100; add an option at 5.2 minutes and the first drops to (5.8 − 5.4) / (5.8 − 5.2) × 100 ≈ 67 although nothing about it changed, while the option at 5.8 stays at 0. An option added between the current best and worst leaves the existing scores alone. Relative scores also cannot say whether any option is good enough. Fixed anchors keep a score’s meaning stable and allow comparison with a standard.

Ordinal inputs need care. Star ratings, ranks, and maturity levels order options without guaranteeing equal steps, so multiplying stars by weights presumes an interval meaning that was never defined. Give each level a description or a measured basis before scoring it. This scaling is also different from feature normalization in machine learning, which rescales inputs for a model rather than expressing a preference.

Reference: UK Government: Multi-criteria analysis manual. Figures here are illustrative.