The box score and what it hides
Points, rebounds and assists are counts. Before they mean anything they have to be adjusted for two things: how long the player was on the floor, and how fast his team plays. A team that takes more possessions per game produces more of every counting statistic, so comparing raw totals across teams compares tempo as much as ability.
Per-minute and per-possession versions solve most of that. They introduce their own distortion — a player who is excellent in twelve minutes may not be excellent in thirty-four — but they at least put two players on the same scale before the argument starts.
Assists are the least reliable entry in the box score, because what counts as one is a judgement made by a human scorer in real time. Rebounds are more solid but still team-dependent: a big who plays alongside another strong rebounder will post lower numbers for reasons that have nothing to do with him.
Shooting efficiency done properly
Field-goal percentage treats a three-point shot and a layup as equally valuable, which they are not. Two adjusted measures fix this. Effective field-goal percentage credits the extra point a three is worth. True shooting goes further and folds in free throws, which is the only way to fairly evaluate a player whose game is built on drawing fouls.
The practical use of true shooting is comparison against league average rather than against a fixed target. A guard who takes difficult shots at slightly above average efficiency while absorbing a large share of his team's attempts may be more valuable than a big finishing easy chances at a far higher rate.
That trade-off is captured by usage rate, which estimates the proportion of his team's possessions a player finishes. Efficiency and usage have to be read together; either one alone can be gamed by simply taking easier or fewer shots.
Advanced numbers and their limits
Plus-minus records the scoring margin while a player is on the floor. In a single game it is close to noise, because it depends entirely on who else was out there. Adjusted versions try to isolate the individual by accounting for teammates and opponents across a season, and they are genuinely informative, but they need a large sample before they settle.
All-in-one ratings combine box-score production, minutes and sometimes tracking data into a single number. They are useful for sorting a long list quickly and misleading when treated as a verdict. Each one embeds a view about what matters, and two reputable models can disagree by a wide margin about the same player.
The honest way to use any of this is as a prompt for a question rather than an answer to it. A number that surprises you is telling you to go and watch the player, not telling you who is better.