Box score representation

Why can a player's statistics look excellent while his team plays worse whenever he is on the floor?

Here is a player who averages twenty points, six rebounds and four assists. By every line in the newspaper he is the best thing on the roster. Yet when the coaching staff pulls the on-off numbers, the team outscores opponents while he sits and loses ground while he plays. Two honest measurements, pointing opposite ways.

The tempting response is to pick a winner — decide the box score lies, or decide the on-off number is noise. But before we take sides, a smaller question is more useful: what is a box score, exactly? Not what it means, but what it physically is.

A box score is a tally sheet. Someone sits courtside and, for each event that the scoring rules define as an event, writes a mark next to a name. So the first thing to ask is: which events does the rulebook define?

They are, almost without exception, the events that terminate a possession and are attributable to exactly one player. A shot goes in — one name. A rebound is secured — one name. A pass immediately precedes a made basket — one name. Notice what these have in common: each is discrete, each is instantaneous, and each has an unambiguous owner.

Now ask the reverse question. What happens in a basketball possession that is not discrete, not instantaneous, or not owned by one player? Quite a lot, it turns out. A screen that arrives half a second early. A help rotation that never becomes a block because the drive was abandoned on sight of it. Spacing — standing in a corner so that a defender cannot leave you, which is literally the act of doing nothing in a particular place. A closeout that forces a worse shot rather than no shot.

None of these can be tallied, and the reason is not laziness. They fail the ownership test. The value of a good screen is realised by the shooter, and the tally has already given that basket to him. There is no second mark to give.

So we have not found a flaw in the box score. We have found its shape. It is a representation, and every representation is a choice about what to preserve and what to discard. The box score preserves discrete, single-owner, possession-ending acts and discards everything continuous, joint, or preventive. Is a player whose contribution lives mostly in the discarded category badly served by it? He must be — not because the tally is wrong, but because he is largely invisible to the thing doing the tallying.

Then what is plus-minus doing differently? Ask what its raw material is. It records only one thing: the scoreboard, and who was on the floor. It never asks who did what. It cannot tell you the reason for anything.

And that is precisely why it captures what the box score cannot. It measures the outcome the game is actually decided by, and by refusing to attribute causes it never has to leave anything out. Screens, spacing, deterrence and help are all in the number, silently, because they moved the scoreboard.

The price is severe, and we should name it before it sounds like a solution. Attributing nothing means it cannot separate you from your four teammates. It is credit-blind. A modest player alongside four excellent ones posts a wonderful figure. And because it is a difference of two scores over a small sample, it is extremely noisy — a few hundred possessions can move it wildly. The adjusted variants (RAPM and its relatives) exist to fight exactly this: they use many lineup combinations to estimate each player's share, but they need enormous samples and still struggle to separate teammates who never play apart.

So the contradiction dissolves. We were never comparing two measurements of one quantity. We were comparing two encodings of a game — one that records named acts and cannot see the unnamed, one that records the result and cannot see who caused it.

Two corrections are worth making explicitly.

First, this is not the claim that box-score statistics are misleading. They are excellent at what they encode. Shooting efficiency from a large sample is a genuinely reliable measurement, far more so than any on-off figure over the same span. The failure mode is not the tally; it is treating the tally as a complete inventory of contribution when it was only ever an inventory of attributable contribution.

Second, a bad on-off number is not proof of hidden harm. It is weak evidence, and often it is evidence about someone else — who you happened to play with, which opponents were on the floor, whether the bench unit you anchored was outscoring another team's starters. This is why the honest analytical answer is neither number alone but a model that borrows from both, and why practitioners disagree about the right weighting rather than having settled it.

The general lesson outlives basketball. Whenever a metric surprises you, the productive question is rarely "is this true?" and almost always "what does this instrument physically record, and what is structurally invisible to it?"

• How adjusted plus-minus models (RAPM) use ridge regression to pull teammates apart, and why they need several seasons of data to stabilise.
• Tracking data as a third encoding: optical systems now record continuous position, which finally makes screens and spacing measurable in principle.
• The same representation problem in other fields — clinical outcome measures, developer productivity metrics, standardised test scores.

Every term the collection defines is gathered in the glossary.