Defensive metrics promise to isolate a defender and cannot quite manage it

Why defence resists individual measurement

An offensive action is initiated by one player and can be credited to them, while a defensive outcome results from several players executing a shared assignment. Assigning the result to the nearest defender treats a coordinated failure as an individual one and vice versa. A defender who executed their responsibility correctly can be charged with a basket that occurred because a teammate rotated late.

Any metric built on proximity therefore encodes the scheme and the teammates into a number presented as individual. That is not a defect of any particular metric; it follows from the structure of the task being measured. Any measure that divides a jointly produced outcome into individual shares must adopt a rule for dividing it, and the rule is a choice.

The events that never happen

A defender whose reputation prevents an opponent from attacking a particular route produces no recorded event at all. Deterrence is a real defensive contribution and it is invisible to any system that counts what occurred rather than what was avoided. Tracking data can partially recover it by measuring where offences choose to attack, which is an improvement and remains indirect.

The best interior defenders are systematically undercounted by outcome-based measures for exactly this reason. Adjusting for it requires modelling counterfactual offensive choices, which reintroduces assumptions the metric was meant to avoid.

Assignment is not random

Coaches assign their strongest defenders to the most difficult opponents, so the raw outcomes those defenders concede are inflated by their assignments. A weaker defender hidden on a low-usage opponent will accumulate a flattering record without doing anything difficult. Correcting for assignment quality is possible in principle and requires deciding how difficulty is measured, which is another judgement.

Each correction improves the metric and adds a modelling choice that the resulting number no longer displays. A number carrying several buried judgements looks like a measurement and functions like an argument.

Sample size at the individual level

The number of possessions in which a specific defender is the primary responsible party is much smaller than a season possession total suggests. Estimates built from that subset are noisy, and differences between similarly rated defenders are frequently smaller than the noise. Rankings nonetheless present them in order, which conveys a resolution the underlying data does not support.

This is the same precision illusion that affects any ranking of closely matched entities. It is more consequential here because contract and role decisions are made on the resulting order. A decision made on a difference that the data cannot resolve is a decision made on noise, however carefully the table was constructed.

What the numbers are good for

Aggregate defensive measures work reasonably well at the team level, where the coordination that defeats individual attribution is contained within the unit. They also work for identifying extremes, since a defender who is genuinely far from the mean will show up despite the noise. They work poorly for the comparisons people most want to make, which involve players who are close together.

Recognising which questions a metric can answer is more useful than either adopting or dismissing it wholesale. The metrics are improving steadily, and the structural difficulty of dividing a shared task will not be improved away. Better tracking narrows the uncertainty around each estimate without changing the fact that the underlying activity is performed by five people at once.